1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2a4ce40a9SAlex Elder 33d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h> 43d14c5d2SYehuda Sadeh 53d14c5d2SYehuda Sadeh #include <linux/module.h> 63d14c5d2SYehuda Sadeh #include <linux/err.h> 73d14c5d2SYehuda Sadeh #include <linux/highmem.h> 83d14c5d2SYehuda Sadeh #include <linux/mm.h> 93d14c5d2SYehuda Sadeh #include <linux/pagemap.h> 103d14c5d2SYehuda Sadeh #include <linux/slab.h> 113d14c5d2SYehuda Sadeh #include <linux/uaccess.h> 123d14c5d2SYehuda Sadeh #ifdef CONFIG_BLOCK 133d14c5d2SYehuda Sadeh #include <linux/bio.h> 143d14c5d2SYehuda Sadeh #endif 153d14c5d2SYehuda Sadeh 168cb441c0SIlya Dryomov #include <linux/ceph/ceph_features.h> 173d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h> 183d14c5d2SYehuda Sadeh #include <linux/ceph/osd_client.h> 193d14c5d2SYehuda Sadeh #include <linux/ceph/messenger.h> 203d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h> 213d14c5d2SYehuda Sadeh #include <linux/ceph/auth.h> 223d14c5d2SYehuda Sadeh #include <linux/ceph/pagelist.h> 2308c1ac50SIlya Dryomov #include <linux/ceph/striper.h> 243d14c5d2SYehuda Sadeh 253d14c5d2SYehuda Sadeh #define OSD_OPREPLY_FRONT_LEN 512 263d14c5d2SYehuda Sadeh 275522ae0bSAlex Elder static struct kmem_cache *ceph_osd_request_cache; 285522ae0bSAlex Elder 293d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops; 303d14c5d2SYehuda Sadeh 313d14c5d2SYehuda Sadeh /* 323d14c5d2SYehuda Sadeh * Implement client access to distributed object storage cluster. 333d14c5d2SYehuda Sadeh * 343d14c5d2SYehuda Sadeh * All data objects are stored within a cluster/cloud of OSDs, or 353d14c5d2SYehuda Sadeh * "object storage devices." (Note that Ceph OSDs have _nothing_ to 363d14c5d2SYehuda Sadeh * do with the T10 OSD extensions to SCSI.) Ceph OSDs are simply 373d14c5d2SYehuda Sadeh * remote daemons serving up and coordinating consistent and safe 383d14c5d2SYehuda Sadeh * access to storage. 393d14c5d2SYehuda Sadeh * 403d14c5d2SYehuda Sadeh * Cluster membership and the mapping of data objects onto storage devices 413d14c5d2SYehuda Sadeh * are described by the osd map. 423d14c5d2SYehuda Sadeh * 433d14c5d2SYehuda Sadeh * We keep track of pending OSD requests (read, write), resubmit 443d14c5d2SYehuda Sadeh * requests to different OSDs when the cluster topology/data layout 453d14c5d2SYehuda Sadeh * change, or retry the affected requests when the communications 463d14c5d2SYehuda Sadeh * channel with an OSD is reset. 473d14c5d2SYehuda Sadeh */ 483d14c5d2SYehuda Sadeh 495aea3dcdSIlya Dryomov static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req); 505aea3dcdSIlya Dryomov static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req); 51922dab61SIlya Dryomov static void link_linger(struct ceph_osd *osd, 52922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq); 53922dab61SIlya Dryomov static void unlink_linger(struct ceph_osd *osd, 54922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq); 55a02a946dSIlya Dryomov static void clear_backoffs(struct ceph_osd *osd); 565aea3dcdSIlya Dryomov 575aea3dcdSIlya Dryomov #if 1 585aea3dcdSIlya Dryomov static inline bool rwsem_is_wrlocked(struct rw_semaphore *sem) 595aea3dcdSIlya Dryomov { 605aea3dcdSIlya Dryomov bool wrlocked = true; 615aea3dcdSIlya Dryomov 625aea3dcdSIlya Dryomov if (unlikely(down_read_trylock(sem))) { 635aea3dcdSIlya Dryomov wrlocked = false; 645aea3dcdSIlya Dryomov up_read(sem); 655aea3dcdSIlya Dryomov } 665aea3dcdSIlya Dryomov 675aea3dcdSIlya Dryomov return wrlocked; 685aea3dcdSIlya Dryomov } 695aea3dcdSIlya Dryomov static inline void verify_osdc_locked(struct ceph_osd_client *osdc) 705aea3dcdSIlya Dryomov { 715aea3dcdSIlya Dryomov WARN_ON(!rwsem_is_locked(&osdc->lock)); 725aea3dcdSIlya Dryomov } 735aea3dcdSIlya Dryomov static inline void verify_osdc_wrlocked(struct ceph_osd_client *osdc) 745aea3dcdSIlya Dryomov { 755aea3dcdSIlya Dryomov WARN_ON(!rwsem_is_wrlocked(&osdc->lock)); 765aea3dcdSIlya Dryomov } 775aea3dcdSIlya Dryomov static inline void verify_osd_locked(struct ceph_osd *osd) 785aea3dcdSIlya Dryomov { 795aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 805aea3dcdSIlya Dryomov 815aea3dcdSIlya Dryomov WARN_ON(!(mutex_is_locked(&osd->lock) && 825aea3dcdSIlya Dryomov rwsem_is_locked(&osdc->lock)) && 835aea3dcdSIlya Dryomov !rwsem_is_wrlocked(&osdc->lock)); 845aea3dcdSIlya Dryomov } 85922dab61SIlya Dryomov static inline void verify_lreq_locked(struct ceph_osd_linger_request *lreq) 86922dab61SIlya Dryomov { 87922dab61SIlya Dryomov WARN_ON(!mutex_is_locked(&lreq->lock)); 88922dab61SIlya Dryomov } 895aea3dcdSIlya Dryomov #else 905aea3dcdSIlya Dryomov static inline void verify_osdc_locked(struct ceph_osd_client *osdc) { } 915aea3dcdSIlya Dryomov static inline void verify_osdc_wrlocked(struct ceph_osd_client *osdc) { } 925aea3dcdSIlya Dryomov static inline void verify_osd_locked(struct ceph_osd *osd) { } 93922dab61SIlya Dryomov static inline void verify_lreq_locked(struct ceph_osd_linger_request *lreq) { } 945aea3dcdSIlya Dryomov #endif 955aea3dcdSIlya Dryomov 963d14c5d2SYehuda Sadeh /* 973d14c5d2SYehuda Sadeh * calculate the mapping of a file extent onto an object, and fill out the 983d14c5d2SYehuda Sadeh * request accordingly. shorten extent as necessary if it crosses an 993d14c5d2SYehuda Sadeh * object boundary. 1003d14c5d2SYehuda Sadeh * 1013d14c5d2SYehuda Sadeh * fill osd op in request message. 1023d14c5d2SYehuda Sadeh */ 103dbe0fc41SAlex Elder static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen, 104a19dadfbSAlex Elder u64 *objnum, u64 *objoff, u64 *objlen) 1053d14c5d2SYehuda Sadeh { 10660e56f13SAlex Elder u64 orig_len = *plen; 107dccbf080SIlya Dryomov u32 xlen; 1083d14c5d2SYehuda Sadeh 10960e56f13SAlex Elder /* object extent? */ 110dccbf080SIlya Dryomov ceph_calc_file_object_mapping(layout, off, orig_len, objnum, 111dccbf080SIlya Dryomov objoff, &xlen); 112dccbf080SIlya Dryomov *objlen = xlen; 11375d1c941SAlex Elder if (*objlen < orig_len) { 11475d1c941SAlex Elder *plen = *objlen; 11560e56f13SAlex Elder dout(" skipping last %llu, final file extent %llu~%llu\n", 11660e56f13SAlex Elder orig_len - *plen, off, *plen); 11760e56f13SAlex Elder } 11860e56f13SAlex Elder 11975d1c941SAlex Elder dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen); 1203ff5f385SAlex Elder return 0; 1213d14c5d2SYehuda Sadeh } 1223d14c5d2SYehuda Sadeh 123c54d47bfSAlex Elder static void ceph_osd_data_init(struct ceph_osd_data *osd_data) 124c54d47bfSAlex Elder { 125c54d47bfSAlex Elder memset(osd_data, 0, sizeof (*osd_data)); 126c54d47bfSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_NONE; 127c54d47bfSAlex Elder } 128c54d47bfSAlex Elder 12989486833SIlya Dryomov /* 13089486833SIlya Dryomov * Consumes @pages if @own_pages is true. 13189486833SIlya Dryomov */ 132a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 13343bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 13443bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 13543bfe5deSAlex Elder { 13643bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 13743bfe5deSAlex Elder osd_data->pages = pages; 13843bfe5deSAlex Elder osd_data->length = length; 13943bfe5deSAlex Elder osd_data->alignment = alignment; 14043bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 14143bfe5deSAlex Elder osd_data->own_pages = own_pages; 14243bfe5deSAlex Elder } 14343bfe5deSAlex Elder 14489486833SIlya Dryomov /* 14589486833SIlya Dryomov * Consumes a ref on @pagelist. 14689486833SIlya Dryomov */ 147a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 14843bfe5deSAlex Elder struct ceph_pagelist *pagelist) 14943bfe5deSAlex Elder { 15043bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 15143bfe5deSAlex Elder osd_data->pagelist = pagelist; 15243bfe5deSAlex Elder } 15343bfe5deSAlex Elder 15443bfe5deSAlex Elder #ifdef CONFIG_BLOCK 155a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 1565359a17dSIlya Dryomov struct ceph_bio_iter *bio_pos, 1575359a17dSIlya Dryomov u32 bio_length) 15843bfe5deSAlex Elder { 15943bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 1605359a17dSIlya Dryomov osd_data->bio_pos = *bio_pos; 16143bfe5deSAlex Elder osd_data->bio_length = bio_length; 16243bfe5deSAlex Elder } 16343bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 16443bfe5deSAlex Elder 165b9e281c2SIlya Dryomov static void ceph_osd_data_bvecs_init(struct ceph_osd_data *osd_data, 1660010f705SIlya Dryomov struct ceph_bvec_iter *bvec_pos, 1670010f705SIlya Dryomov u32 num_bvecs) 168b9e281c2SIlya Dryomov { 169b9e281c2SIlya Dryomov osd_data->type = CEPH_OSD_DATA_TYPE_BVECS; 170b9e281c2SIlya Dryomov osd_data->bvec_pos = *bvec_pos; 1710010f705SIlya Dryomov osd_data->num_bvecs = num_bvecs; 172b9e281c2SIlya Dryomov } 173b9e281c2SIlya Dryomov 17449719778SAlex Elder static struct ceph_osd_data * 17549719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 17649719778SAlex Elder { 17749719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 17849719778SAlex Elder 17949719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 18049719778SAlex Elder } 18149719778SAlex Elder 182a4ce40a9SAlex Elder struct ceph_osd_data * 183a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 184406e2c9fSAlex Elder unsigned int which) 185a4ce40a9SAlex Elder { 186863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 187a4ce40a9SAlex Elder } 188a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 189a4ce40a9SAlex Elder 19049719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 19149719778SAlex Elder unsigned int which, struct page **pages, 19249719778SAlex Elder u64 length, u32 alignment, 19349719778SAlex Elder bool pages_from_pool, bool own_pages) 19449719778SAlex Elder { 19549719778SAlex Elder struct ceph_osd_data *osd_data; 19649719778SAlex Elder 19749719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 19849719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 19949719778SAlex Elder pages_from_pool, own_pages); 20049719778SAlex Elder } 20149719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 20249719778SAlex Elder 203a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 204406e2c9fSAlex Elder unsigned int which, struct page **pages, 205406e2c9fSAlex Elder u64 length, u32 alignment, 206a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 207a4ce40a9SAlex Elder { 208a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 209a4ce40a9SAlex Elder 210863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 211a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 212a4ce40a9SAlex Elder pages_from_pool, own_pages); 213a4ce40a9SAlex Elder } 214a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 215a4ce40a9SAlex Elder 216a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 217406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 218a4ce40a9SAlex Elder { 219a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 220a4ce40a9SAlex Elder 221863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 222a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 223a4ce40a9SAlex Elder } 224a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 225a4ce40a9SAlex Elder 226a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 227a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 2285359a17dSIlya Dryomov unsigned int which, 2295359a17dSIlya Dryomov struct ceph_bio_iter *bio_pos, 2305359a17dSIlya Dryomov u32 bio_length) 231a4ce40a9SAlex Elder { 232a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 233863c7eb5SAlex Elder 234863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2355359a17dSIlya Dryomov ceph_osd_data_bio_init(osd_data, bio_pos, bio_length); 236a4ce40a9SAlex Elder } 237a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 238a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 239a4ce40a9SAlex Elder 2400010f705SIlya Dryomov void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request *osd_req, 2410010f705SIlya Dryomov unsigned int which, 2420010f705SIlya Dryomov struct bio_vec *bvecs, u32 num_bvecs, 2430010f705SIlya Dryomov u32 bytes) 2440010f705SIlya Dryomov { 2450010f705SIlya Dryomov struct ceph_osd_data *osd_data; 2460010f705SIlya Dryomov struct ceph_bvec_iter it = { 2470010f705SIlya Dryomov .bvecs = bvecs, 2480010f705SIlya Dryomov .iter = { .bi_size = bytes }, 2490010f705SIlya Dryomov }; 2500010f705SIlya Dryomov 2510010f705SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2520010f705SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, &it, num_bvecs); 2530010f705SIlya Dryomov } 2540010f705SIlya Dryomov EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvecs); 2550010f705SIlya Dryomov 256b9e281c2SIlya Dryomov void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req, 257b9e281c2SIlya Dryomov unsigned int which, 258b9e281c2SIlya Dryomov struct ceph_bvec_iter *bvec_pos) 259b9e281c2SIlya Dryomov { 260b9e281c2SIlya Dryomov struct ceph_osd_data *osd_data; 261b9e281c2SIlya Dryomov 262b9e281c2SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2630010f705SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, bvec_pos, 0); 264b9e281c2SIlya Dryomov } 265b9e281c2SIlya Dryomov EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvec_pos); 266b9e281c2SIlya Dryomov 267a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 268a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 269a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 270a4ce40a9SAlex Elder { 271a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 272a4ce40a9SAlex Elder 273863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 274a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 275a4ce40a9SAlex Elder } 276a4ce40a9SAlex Elder 27704017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 27804017e29SAlex Elder struct ceph_osd_request *osd_req, 27904017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 28004017e29SAlex Elder { 28104017e29SAlex Elder struct ceph_osd_data *osd_data; 28204017e29SAlex Elder 283863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 28404017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 285bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += pagelist->length; 286bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += pagelist->length; 28704017e29SAlex Elder } 28804017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 28904017e29SAlex Elder 2906c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2916c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2926c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2936c57b554SAlex Elder { 2946c57b554SAlex Elder struct ceph_osd_data *osd_data; 2956c57b554SAlex Elder 2966c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2976c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2986c57b554SAlex Elder pages_from_pool, own_pages); 299bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += length; 300bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += length; 3016c57b554SAlex Elder } 3026c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 3036c57b554SAlex Elder 304b9e281c2SIlya Dryomov void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req, 305b9e281c2SIlya Dryomov unsigned int which, 3060010f705SIlya Dryomov struct bio_vec *bvecs, u32 num_bvecs, 3070010f705SIlya Dryomov u32 bytes) 308b9e281c2SIlya Dryomov { 309b9e281c2SIlya Dryomov struct ceph_osd_data *osd_data; 310b9e281c2SIlya Dryomov struct ceph_bvec_iter it = { 311b9e281c2SIlya Dryomov .bvecs = bvecs, 312b9e281c2SIlya Dryomov .iter = { .bi_size = bytes }, 313b9e281c2SIlya Dryomov }; 314b9e281c2SIlya Dryomov 315b9e281c2SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, cls, request_data); 3160010f705SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, &it, num_bvecs); 317b9e281c2SIlya Dryomov osd_req->r_ops[which].cls.indata_len += bytes; 318b9e281c2SIlya Dryomov osd_req->r_ops[which].indata_len += bytes; 319b9e281c2SIlya Dryomov } 320b9e281c2SIlya Dryomov EXPORT_SYMBOL(osd_req_op_cls_request_data_bvecs); 321b9e281c2SIlya Dryomov 322a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 323a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 324a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 325a4ce40a9SAlex Elder { 326a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 327a4ce40a9SAlex Elder 328863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 329a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 330a4ce40a9SAlex Elder pages_from_pool, own_pages); 331a4ce40a9SAlex Elder } 332a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 333a4ce40a9SAlex Elder 33423c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 33523c08a9cSAlex Elder { 33623c08a9cSAlex Elder switch (osd_data->type) { 33723c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 33823c08a9cSAlex Elder return 0; 33923c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 34023c08a9cSAlex Elder return osd_data->length; 34123c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 34223c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 34323c08a9cSAlex Elder #ifdef CONFIG_BLOCK 34423c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 34523c08a9cSAlex Elder return (u64)osd_data->bio_length; 34623c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 347b9e281c2SIlya Dryomov case CEPH_OSD_DATA_TYPE_BVECS: 348b9e281c2SIlya Dryomov return osd_data->bvec_pos.iter.bi_size; 34923c08a9cSAlex Elder default: 35023c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 35123c08a9cSAlex Elder return 0; 35223c08a9cSAlex Elder } 35323c08a9cSAlex Elder } 35423c08a9cSAlex Elder 355c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 356c54d47bfSAlex Elder { 3575476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 358c54d47bfSAlex Elder int num_pages; 359c54d47bfSAlex Elder 360c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 361c54d47bfSAlex Elder (u64)osd_data->length); 362c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 36389486833SIlya Dryomov } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 36489486833SIlya Dryomov ceph_pagelist_release(osd_data->pagelist); 365c54d47bfSAlex Elder } 3665476492fSAlex Elder ceph_osd_data_init(osd_data); 3675476492fSAlex Elder } 3685476492fSAlex Elder 3695476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 3705476492fSAlex Elder unsigned int which) 3715476492fSAlex Elder { 3725476492fSAlex Elder struct ceph_osd_req_op *op; 3735476492fSAlex Elder 3745476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 3755476492fSAlex Elder op = &osd_req->r_ops[which]; 3765476492fSAlex Elder 3775476492fSAlex Elder switch (op->op) { 3785476492fSAlex Elder case CEPH_OSD_OP_READ: 3795476492fSAlex Elder case CEPH_OSD_OP_WRITE: 380e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 3815476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 3825476492fSAlex Elder break; 3835476492fSAlex Elder case CEPH_OSD_OP_CALL: 3845476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 38504017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 3865476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 3875476492fSAlex Elder break; 388d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 389d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 390d74b50beSYan, Zheng ceph_osd_data_release(&op->xattr.osd_data); 391d74b50beSYan, Zheng break; 39266ba609fSYan, Zheng case CEPH_OSD_OP_STAT: 39366ba609fSYan, Zheng ceph_osd_data_release(&op->raw_data_in); 39466ba609fSYan, Zheng break; 395922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 396922dab61SIlya Dryomov ceph_osd_data_release(&op->notify_ack.request_data); 397922dab61SIlya Dryomov break; 39819079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 39919079203SIlya Dryomov ceph_osd_data_release(&op->notify.request_data); 40019079203SIlya Dryomov ceph_osd_data_release(&op->notify.response_data); 40119079203SIlya Dryomov break; 402a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 403a4ed38d7SDouglas Fuller ceph_osd_data_release(&op->list_watchers.response_data); 404a4ed38d7SDouglas Fuller break; 40578beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 40623ddf9beSLuis Henriques ceph_osd_data_release(&op->copy_from.osd_data); 40723ddf9beSLuis Henriques break; 4085476492fSAlex Elder default: 4095476492fSAlex Elder break; 4105476492fSAlex Elder } 411c54d47bfSAlex Elder } 412c54d47bfSAlex Elder 4133d14c5d2SYehuda Sadeh /* 41463244fa1SIlya Dryomov * Assumes @t is zero-initialized. 41563244fa1SIlya Dryomov */ 41663244fa1SIlya Dryomov static void target_init(struct ceph_osd_request_target *t) 41763244fa1SIlya Dryomov { 41863244fa1SIlya Dryomov ceph_oid_init(&t->base_oid); 41963244fa1SIlya Dryomov ceph_oloc_init(&t->base_oloc); 42063244fa1SIlya Dryomov ceph_oid_init(&t->target_oid); 42163244fa1SIlya Dryomov ceph_oloc_init(&t->target_oloc); 42263244fa1SIlya Dryomov 42363244fa1SIlya Dryomov ceph_osds_init(&t->acting); 42463244fa1SIlya Dryomov ceph_osds_init(&t->up); 42563244fa1SIlya Dryomov t->size = -1; 42663244fa1SIlya Dryomov t->min_size = -1; 42763244fa1SIlya Dryomov 42863244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 42963244fa1SIlya Dryomov } 43063244fa1SIlya Dryomov 431922dab61SIlya Dryomov static void target_copy(struct ceph_osd_request_target *dest, 432922dab61SIlya Dryomov const struct ceph_osd_request_target *src) 433922dab61SIlya Dryomov { 434922dab61SIlya Dryomov ceph_oid_copy(&dest->base_oid, &src->base_oid); 435922dab61SIlya Dryomov ceph_oloc_copy(&dest->base_oloc, &src->base_oloc); 436922dab61SIlya Dryomov ceph_oid_copy(&dest->target_oid, &src->target_oid); 437922dab61SIlya Dryomov ceph_oloc_copy(&dest->target_oloc, &src->target_oloc); 438922dab61SIlya Dryomov 439922dab61SIlya Dryomov dest->pgid = src->pgid; /* struct */ 440dc98ff72SIlya Dryomov dest->spgid = src->spgid; /* struct */ 441922dab61SIlya Dryomov dest->pg_num = src->pg_num; 442922dab61SIlya Dryomov dest->pg_num_mask = src->pg_num_mask; 443922dab61SIlya Dryomov ceph_osds_copy(&dest->acting, &src->acting); 444922dab61SIlya Dryomov ceph_osds_copy(&dest->up, &src->up); 445922dab61SIlya Dryomov dest->size = src->size; 446922dab61SIlya Dryomov dest->min_size = src->min_size; 447922dab61SIlya Dryomov dest->sort_bitwise = src->sort_bitwise; 4482f3fead6SIlya Dryomov dest->recovery_deletes = src->recovery_deletes; 449922dab61SIlya Dryomov 450922dab61SIlya Dryomov dest->flags = src->flags; 4517ed286f3SIlya Dryomov dest->used_replica = src->used_replica; 452922dab61SIlya Dryomov dest->paused = src->paused; 453922dab61SIlya Dryomov 45404c7d789SIlya Dryomov dest->epoch = src->epoch; 455dc93e0e2SIlya Dryomov dest->last_force_resend = src->last_force_resend; 456dc93e0e2SIlya Dryomov 457922dab61SIlya Dryomov dest->osd = src->osd; 458922dab61SIlya Dryomov } 459922dab61SIlya Dryomov 46063244fa1SIlya Dryomov static void target_destroy(struct ceph_osd_request_target *t) 46163244fa1SIlya Dryomov { 46263244fa1SIlya Dryomov ceph_oid_destroy(&t->base_oid); 46330c156d9SYan, Zheng ceph_oloc_destroy(&t->base_oloc); 46463244fa1SIlya Dryomov ceph_oid_destroy(&t->target_oid); 46530c156d9SYan, Zheng ceph_oloc_destroy(&t->target_oloc); 46663244fa1SIlya Dryomov } 46763244fa1SIlya Dryomov 46863244fa1SIlya Dryomov /* 4693d14c5d2SYehuda Sadeh * requests 4703d14c5d2SYehuda Sadeh */ 4713540bfdbSIlya Dryomov static void request_release_checks(struct ceph_osd_request *req) 4723540bfdbSIlya Dryomov { 4733540bfdbSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 4744609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_mc_node)); 47594e85771SIlya Dryomov WARN_ON(!list_empty(&req->r_private_item)); 4763540bfdbSIlya Dryomov WARN_ON(req->r_osd); 4773540bfdbSIlya Dryomov } 4783540bfdbSIlya Dryomov 4799e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 4803d14c5d2SYehuda Sadeh { 4819e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 4829e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 4835476492fSAlex Elder unsigned int which; 4843d14c5d2SYehuda Sadeh 4859e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 4869e94af20SIlya Dryomov req->r_request, req->r_reply); 4873540bfdbSIlya Dryomov request_release_checks(req); 4889e94af20SIlya Dryomov 4893d14c5d2SYehuda Sadeh if (req->r_request) 4903d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 4915aea3dcdSIlya Dryomov if (req->r_reply) 492ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 4930fff87ecSAlex Elder 4945476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 4955476492fSAlex Elder osd_req_op_data_release(req, which); 4960fff87ecSAlex Elder 497a66dd383SIlya Dryomov target_destroy(&req->r_t); 4983d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 499d30291b9SIlya Dryomov 5003d14c5d2SYehuda Sadeh if (req->r_mempool) 5013d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 5023f1af42aSIlya Dryomov else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS) 5035522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 5043f1af42aSIlya Dryomov else 5053f1af42aSIlya Dryomov kfree(req); 5063d14c5d2SYehuda Sadeh } 5079e94af20SIlya Dryomov 5089e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 5099e94af20SIlya Dryomov { 5109e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 5112c935bc5SPeter Zijlstra kref_read(&req->r_kref)); 5129e94af20SIlya Dryomov kref_get(&req->r_kref); 5139e94af20SIlya Dryomov } 5149e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 5159e94af20SIlya Dryomov 5169e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 5179e94af20SIlya Dryomov { 5183ed97d63SIlya Dryomov if (req) { 5199e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 5202c935bc5SPeter Zijlstra kref_read(&req->r_kref)); 5219e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 5229e94af20SIlya Dryomov } 5233ed97d63SIlya Dryomov } 5249e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 5253d14c5d2SYehuda Sadeh 5263540bfdbSIlya Dryomov static void request_init(struct ceph_osd_request *req) 5273540bfdbSIlya Dryomov { 528042f6498SJeff Layton /* req only, each op is zeroed in osd_req_op_init() */ 5293540bfdbSIlya Dryomov memset(req, 0, sizeof(*req)); 5303540bfdbSIlya Dryomov 5313540bfdbSIlya Dryomov kref_init(&req->r_kref); 5323540bfdbSIlya Dryomov init_completion(&req->r_completion); 5333540bfdbSIlya Dryomov RB_CLEAR_NODE(&req->r_node); 5344609245eSIlya Dryomov RB_CLEAR_NODE(&req->r_mc_node); 53594e85771SIlya Dryomov INIT_LIST_HEAD(&req->r_private_item); 5363540bfdbSIlya Dryomov 5373540bfdbSIlya Dryomov target_init(&req->r_t); 5383540bfdbSIlya Dryomov } 5393540bfdbSIlya Dryomov 540922dab61SIlya Dryomov /* 541922dab61SIlya Dryomov * This is ugly, but it allows us to reuse linger registration and ping 542922dab61SIlya Dryomov * requests, keeping the structure of the code around send_linger{_ping}() 543922dab61SIlya Dryomov * reasonable. Setting up a min_nr=2 mempool for each linger request 544922dab61SIlya Dryomov * and dealing with copying ops (this blasts req only, watch op remains 545922dab61SIlya Dryomov * intact) isn't any better. 546922dab61SIlya Dryomov */ 547922dab61SIlya Dryomov static void request_reinit(struct ceph_osd_request *req) 548922dab61SIlya Dryomov { 549922dab61SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 550922dab61SIlya Dryomov bool mempool = req->r_mempool; 551922dab61SIlya Dryomov unsigned int num_ops = req->r_num_ops; 552922dab61SIlya Dryomov u64 snapid = req->r_snapid; 553922dab61SIlya Dryomov struct ceph_snap_context *snapc = req->r_snapc; 554922dab61SIlya Dryomov bool linger = req->r_linger; 555922dab61SIlya Dryomov struct ceph_msg *request_msg = req->r_request; 556922dab61SIlya Dryomov struct ceph_msg *reply_msg = req->r_reply; 557922dab61SIlya Dryomov 558922dab61SIlya Dryomov dout("%s req %p\n", __func__, req); 5592c935bc5SPeter Zijlstra WARN_ON(kref_read(&req->r_kref) != 1); 560922dab61SIlya Dryomov request_release_checks(req); 561922dab61SIlya Dryomov 5622c935bc5SPeter Zijlstra WARN_ON(kref_read(&request_msg->kref) != 1); 5632c935bc5SPeter Zijlstra WARN_ON(kref_read(&reply_msg->kref) != 1); 564922dab61SIlya Dryomov target_destroy(&req->r_t); 565922dab61SIlya Dryomov 566922dab61SIlya Dryomov request_init(req); 567922dab61SIlya Dryomov req->r_osdc = osdc; 568922dab61SIlya Dryomov req->r_mempool = mempool; 569922dab61SIlya Dryomov req->r_num_ops = num_ops; 570922dab61SIlya Dryomov req->r_snapid = snapid; 571922dab61SIlya Dryomov req->r_snapc = snapc; 572922dab61SIlya Dryomov req->r_linger = linger; 573922dab61SIlya Dryomov req->r_request = request_msg; 574922dab61SIlya Dryomov req->r_reply = reply_msg; 575922dab61SIlya Dryomov } 576922dab61SIlya Dryomov 5773d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 5783d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 5791b83bef2SSage Weil unsigned int num_ops, 5803d14c5d2SYehuda Sadeh bool use_mempool, 58154a54007SAlex Elder gfp_t gfp_flags) 5823d14c5d2SYehuda Sadeh { 5833d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5843d14c5d2SYehuda Sadeh 5853d14c5d2SYehuda Sadeh if (use_mempool) { 5863f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_SLAB_OPS); 5873d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 5883f1af42aSIlya Dryomov } else if (num_ops <= CEPH_OSD_SLAB_OPS) { 5893f1af42aSIlya Dryomov req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags); 5903d14c5d2SYehuda Sadeh } else { 5913f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_MAX_OPS); 592acafe7e3SKees Cook req = kmalloc(struct_size(req, r_ops, num_ops), gfp_flags); 5933d14c5d2SYehuda Sadeh } 5943f1af42aSIlya Dryomov if (unlikely(!req)) 5953d14c5d2SYehuda Sadeh return NULL; 5963d14c5d2SYehuda Sadeh 5973540bfdbSIlya Dryomov request_init(req); 5983d14c5d2SYehuda Sadeh req->r_osdc = osdc; 5993d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 60079528734SAlex Elder req->r_num_ops = num_ops; 60184127282SIlya Dryomov req->r_snapid = CEPH_NOSNAP; 60284127282SIlya Dryomov req->r_snapc = ceph_get_snap_context(snapc); 6033d14c5d2SYehuda Sadeh 60413d1ad16SIlya Dryomov dout("%s req %p\n", __func__, req); 60513d1ad16SIlya Dryomov return req; 6063f1af42aSIlya Dryomov } 60713d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_request); 6083f1af42aSIlya Dryomov 6092e59ffd1SIlya Dryomov static int ceph_oloc_encoding_size(const struct ceph_object_locator *oloc) 61030c156d9SYan, Zheng { 61130c156d9SYan, Zheng return 8 + 4 + 4 + 4 + (oloc->pool_ns ? oloc->pool_ns->len : 0); 61230c156d9SYan, Zheng } 61330c156d9SYan, Zheng 6140d9c1ab3SIlya Dryomov static int __ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp, 6150d9c1ab3SIlya Dryomov int num_request_data_items, 6160d9c1ab3SIlya Dryomov int num_reply_data_items) 61713d1ad16SIlya Dryomov { 61813d1ad16SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 61913d1ad16SIlya Dryomov struct ceph_msg *msg; 62013d1ad16SIlya Dryomov int msg_size; 6213d14c5d2SYehuda Sadeh 6220d9c1ab3SIlya Dryomov WARN_ON(req->r_request || req->r_reply); 623d30291b9SIlya Dryomov WARN_ON(ceph_oid_empty(&req->r_base_oid)); 62430c156d9SYan, Zheng WARN_ON(ceph_oloc_empty(&req->r_base_oloc)); 625d30291b9SIlya Dryomov 62613d1ad16SIlya Dryomov /* create request message */ 6278cb441c0SIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 6288cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 6298cb441c0SIlya Dryomov msg_size += 4 + 4 + 4; /* hash, osdmap_epoch, flags */ 6308cb441c0SIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 6318cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid); /* reqid */ 6328cb441c0SIlya Dryomov msg_size += sizeof(struct ceph_blkin_trace_info); /* trace */ 6338cb441c0SIlya Dryomov msg_size += 4 + sizeof(struct ceph_timespec); /* client_inc, mtime */ 63430c156d9SYan, Zheng msg_size += CEPH_ENCODING_START_BLK_LEN + 63530c156d9SYan, Zheng ceph_oloc_encoding_size(&req->r_base_oloc); /* oloc */ 63613d1ad16SIlya Dryomov msg_size += 4 + req->r_base_oid.name_len; /* oid */ 63713d1ad16SIlya Dryomov msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op); 638ae458f5aSIlya Dryomov msg_size += 8; /* snapid */ 639ae458f5aSIlya Dryomov msg_size += 8; /* snap_seq */ 64013d1ad16SIlya Dryomov msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0); 6418cb441c0SIlya Dryomov msg_size += 4 + 8; /* retry_attempt, features */ 642ae458f5aSIlya Dryomov 64313d1ad16SIlya Dryomov if (req->r_mempool) 6440d9c1ab3SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op, msg_size, 6450d9c1ab3SIlya Dryomov num_request_data_items); 6463d14c5d2SYehuda Sadeh else 6470d9c1ab3SIlya Dryomov msg = ceph_msg_new2(CEPH_MSG_OSD_OP, msg_size, 6480d9c1ab3SIlya Dryomov num_request_data_items, gfp, true); 64913d1ad16SIlya Dryomov if (!msg) 65013d1ad16SIlya Dryomov return -ENOMEM; 6513d14c5d2SYehuda Sadeh 6523d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 6533d14c5d2SYehuda Sadeh req->r_request = msg; 6543d14c5d2SYehuda Sadeh 65513d1ad16SIlya Dryomov /* create reply message */ 65613d1ad16SIlya Dryomov msg_size = OSD_OPREPLY_FRONT_LEN; 657711da55dSIlya Dryomov msg_size += req->r_base_oid.name_len; 658711da55dSIlya Dryomov msg_size += req->r_num_ops * sizeof(struct ceph_osd_op); 65913d1ad16SIlya Dryomov 66013d1ad16SIlya Dryomov if (req->r_mempool) 6610d9c1ab3SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op_reply, msg_size, 6620d9c1ab3SIlya Dryomov num_reply_data_items); 66313d1ad16SIlya Dryomov else 6640d9c1ab3SIlya Dryomov msg = ceph_msg_new2(CEPH_MSG_OSD_OPREPLY, msg_size, 6650d9c1ab3SIlya Dryomov num_reply_data_items, gfp, true); 66613d1ad16SIlya Dryomov if (!msg) 66713d1ad16SIlya Dryomov return -ENOMEM; 66813d1ad16SIlya Dryomov 66913d1ad16SIlya Dryomov req->r_reply = msg; 67013d1ad16SIlya Dryomov 67113d1ad16SIlya Dryomov return 0; 67213d1ad16SIlya Dryomov } 6733d14c5d2SYehuda Sadeh 674a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 675a8dd0a37SAlex Elder { 676a8dd0a37SAlex Elder switch (opcode) { 67770b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 67870b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 67970b5bfa3SIlya Dryomov #undef GENERATE_CASE 680a8dd0a37SAlex Elder default: 681a8dd0a37SAlex Elder return false; 682a8dd0a37SAlex Elder } 683a8dd0a37SAlex Elder } 684a8dd0a37SAlex Elder 6850d9c1ab3SIlya Dryomov static void get_num_data_items(struct ceph_osd_request *req, 6860d9c1ab3SIlya Dryomov int *num_request_data_items, 6870d9c1ab3SIlya Dryomov int *num_reply_data_items) 6880d9c1ab3SIlya Dryomov { 6890d9c1ab3SIlya Dryomov struct ceph_osd_req_op *op; 6900d9c1ab3SIlya Dryomov 6910d9c1ab3SIlya Dryomov *num_request_data_items = 0; 6920d9c1ab3SIlya Dryomov *num_reply_data_items = 0; 6930d9c1ab3SIlya Dryomov 6940d9c1ab3SIlya Dryomov for (op = req->r_ops; op != &req->r_ops[req->r_num_ops]; op++) { 6950d9c1ab3SIlya Dryomov switch (op->op) { 6960d9c1ab3SIlya Dryomov /* request */ 6970d9c1ab3SIlya Dryomov case CEPH_OSD_OP_WRITE: 6980d9c1ab3SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 6990d9c1ab3SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 7000d9c1ab3SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 7010d9c1ab3SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 70278beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 7030d9c1ab3SIlya Dryomov *num_request_data_items += 1; 7040d9c1ab3SIlya Dryomov break; 7050d9c1ab3SIlya Dryomov 7060d9c1ab3SIlya Dryomov /* reply */ 7070d9c1ab3SIlya Dryomov case CEPH_OSD_OP_STAT: 7080d9c1ab3SIlya Dryomov case CEPH_OSD_OP_READ: 7090d9c1ab3SIlya Dryomov case CEPH_OSD_OP_LIST_WATCHERS: 7100d9c1ab3SIlya Dryomov *num_reply_data_items += 1; 7110d9c1ab3SIlya Dryomov break; 7120d9c1ab3SIlya Dryomov 7130d9c1ab3SIlya Dryomov /* both */ 7140d9c1ab3SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 7150d9c1ab3SIlya Dryomov *num_request_data_items += 1; 7160d9c1ab3SIlya Dryomov *num_reply_data_items += 1; 7170d9c1ab3SIlya Dryomov break; 7180d9c1ab3SIlya Dryomov case CEPH_OSD_OP_CALL: 7190d9c1ab3SIlya Dryomov *num_request_data_items += 2; 7200d9c1ab3SIlya Dryomov *num_reply_data_items += 1; 7210d9c1ab3SIlya Dryomov break; 7220d9c1ab3SIlya Dryomov 7230d9c1ab3SIlya Dryomov default: 7240d9c1ab3SIlya Dryomov WARN_ON(!osd_req_opcode_valid(op->op)); 7250d9c1ab3SIlya Dryomov break; 7260d9c1ab3SIlya Dryomov } 7270d9c1ab3SIlya Dryomov } 7280d9c1ab3SIlya Dryomov } 7290d9c1ab3SIlya Dryomov 7300d9c1ab3SIlya Dryomov /* 7310d9c1ab3SIlya Dryomov * oid, oloc and OSD op opcode(s) must be filled in before this function 7320d9c1ab3SIlya Dryomov * is called. 7330d9c1ab3SIlya Dryomov */ 7340d9c1ab3SIlya Dryomov int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp) 7350d9c1ab3SIlya Dryomov { 7360d9c1ab3SIlya Dryomov int num_request_data_items, num_reply_data_items; 7370d9c1ab3SIlya Dryomov 7380d9c1ab3SIlya Dryomov get_num_data_items(req, &num_request_data_items, &num_reply_data_items); 7390d9c1ab3SIlya Dryomov return __ceph_osdc_alloc_messages(req, gfp, num_request_data_items, 7400d9c1ab3SIlya Dryomov num_reply_data_items); 7410d9c1ab3SIlya Dryomov } 7420d9c1ab3SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_messages); 7430d9c1ab3SIlya Dryomov 74433803f33SAlex Elder /* 74533803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 74633803f33SAlex Elder * other information associated with them. It also serves as a 74733803f33SAlex Elder * common init routine for all the other init functions, below. 74833803f33SAlex Elder */ 749042f6498SJeff Layton struct ceph_osd_req_op * 750042f6498SJeff Layton osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 751144cba14SYan, Zheng u16 opcode, u32 flags) 75233803f33SAlex Elder { 753c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 754c99d2d4aSAlex Elder 755c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 75633803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 75733803f33SAlex Elder 758c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 75933803f33SAlex Elder memset(op, 0, sizeof (*op)); 76033803f33SAlex Elder op->op = opcode; 761144cba14SYan, Zheng op->flags = flags; 762c99d2d4aSAlex Elder 763c99d2d4aSAlex Elder return op; 76433803f33SAlex Elder } 76549719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 76649719778SAlex Elder 767c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 768c99d2d4aSAlex Elder unsigned int which, u16 opcode, 76933803f33SAlex Elder u64 offset, u64 length, 77033803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 77133803f33SAlex Elder { 772042f6498SJeff Layton struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, 773144cba14SYan, Zheng opcode, 0); 77433803f33SAlex Elder size_t payload_len = 0; 77533803f33SAlex Elder 776ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 777e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 778e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_TRUNCATE); 77933803f33SAlex Elder 78033803f33SAlex Elder op->extent.offset = offset; 78133803f33SAlex Elder op->extent.length = length; 78233803f33SAlex Elder op->extent.truncate_size = truncate_size; 78333803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 784e30b7577SIlya Dryomov if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 78533803f33SAlex Elder payload_len += length; 78633803f33SAlex Elder 787de2aa102SIlya Dryomov op->indata_len = payload_len; 78833803f33SAlex Elder } 78933803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 79033803f33SAlex Elder 791c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 792c99d2d4aSAlex Elder unsigned int which, u64 length) 793e5975c7cSAlex Elder { 794c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 795c99d2d4aSAlex Elder u64 previous; 796c99d2d4aSAlex Elder 797c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 798c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 799c99d2d4aSAlex Elder previous = op->extent.length; 800e5975c7cSAlex Elder 801e5975c7cSAlex Elder if (length == previous) 802e5975c7cSAlex Elder return; /* Nothing to do */ 803e5975c7cSAlex Elder BUG_ON(length > previous); 804e5975c7cSAlex Elder 805e5975c7cSAlex Elder op->extent.length = length; 806d641df81SYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 807de2aa102SIlya Dryomov op->indata_len -= previous - length; 808e5975c7cSAlex Elder } 809e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 810e5975c7cSAlex Elder 8112c63f49aSYan, Zheng void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 8122c63f49aSYan, Zheng unsigned int which, u64 offset_inc) 8132c63f49aSYan, Zheng { 8142c63f49aSYan, Zheng struct ceph_osd_req_op *op, *prev_op; 8152c63f49aSYan, Zheng 8162c63f49aSYan, Zheng BUG_ON(which + 1 >= osd_req->r_num_ops); 8172c63f49aSYan, Zheng 8182c63f49aSYan, Zheng prev_op = &osd_req->r_ops[which]; 819042f6498SJeff Layton op = osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags); 8202c63f49aSYan, Zheng /* dup previous one */ 8212c63f49aSYan, Zheng op->indata_len = prev_op->indata_len; 8222c63f49aSYan, Zheng op->outdata_len = prev_op->outdata_len; 8232c63f49aSYan, Zheng op->extent = prev_op->extent; 8242c63f49aSYan, Zheng /* adjust offset */ 8252c63f49aSYan, Zheng op->extent.offset += offset_inc; 8262c63f49aSYan, Zheng op->extent.length -= offset_inc; 8272c63f49aSYan, Zheng 8282c63f49aSYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 8292c63f49aSYan, Zheng op->indata_len -= offset_inc; 8302c63f49aSYan, Zheng } 8312c63f49aSYan, Zheng EXPORT_SYMBOL(osd_req_op_extent_dup_last); 8322c63f49aSYan, Zheng 833fe943d50SChengguang Xu int osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 83424639ce5SIlya Dryomov const char *class, const char *method) 83533803f33SAlex Elder { 83624639ce5SIlya Dryomov struct ceph_osd_req_op *op; 8375f562df5SAlex Elder struct ceph_pagelist *pagelist; 83833803f33SAlex Elder size_t payload_len = 0; 83933803f33SAlex Elder size_t size; 8404766815bSDavid Disseldorp int ret; 84133803f33SAlex Elder 842042f6498SJeff Layton op = osd_req_op_init(osd_req, which, CEPH_OSD_OP_CALL, 0); 84333803f33SAlex Elder 84433165d47SIlya Dryomov pagelist = ceph_pagelist_alloc(GFP_NOFS); 845fe943d50SChengguang Xu if (!pagelist) 846fe943d50SChengguang Xu return -ENOMEM; 847fe943d50SChengguang Xu 84833803f33SAlex Elder op->cls.class_name = class; 84933803f33SAlex Elder size = strlen(class); 85033803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 85133803f33SAlex Elder op->cls.class_len = size; 8524766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, class, size); 8534766815bSDavid Disseldorp if (ret) 8544766815bSDavid Disseldorp goto err_pagelist_free; 85533803f33SAlex Elder payload_len += size; 85633803f33SAlex Elder 85733803f33SAlex Elder op->cls.method_name = method; 85833803f33SAlex Elder size = strlen(method); 85933803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 86033803f33SAlex Elder op->cls.method_len = size; 8614766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, method, size); 8624766815bSDavid Disseldorp if (ret) 8634766815bSDavid Disseldorp goto err_pagelist_free; 86433803f33SAlex Elder payload_len += size; 86533803f33SAlex Elder 866a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 867de2aa102SIlya Dryomov op->indata_len = payload_len; 868fe943d50SChengguang Xu return 0; 8694766815bSDavid Disseldorp 8704766815bSDavid Disseldorp err_pagelist_free: 8714766815bSDavid Disseldorp ceph_pagelist_release(pagelist); 8724766815bSDavid Disseldorp return ret; 87333803f33SAlex Elder } 87433803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 8758c042b0dSAlex Elder 876d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 877d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 878d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 879d74b50beSYan, Zheng { 880042f6498SJeff Layton struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, 881144cba14SYan, Zheng opcode, 0); 882d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 883d74b50beSYan, Zheng size_t payload_len; 8844766815bSDavid Disseldorp int ret; 885d74b50beSYan, Zheng 886d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 887d74b50beSYan, Zheng 88833165d47SIlya Dryomov pagelist = ceph_pagelist_alloc(GFP_NOFS); 889d74b50beSYan, Zheng if (!pagelist) 890d74b50beSYan, Zheng return -ENOMEM; 891d74b50beSYan, Zheng 892d74b50beSYan, Zheng payload_len = strlen(name); 893d74b50beSYan, Zheng op->xattr.name_len = payload_len; 8944766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, name, payload_len); 8954766815bSDavid Disseldorp if (ret) 8964766815bSDavid Disseldorp goto err_pagelist_free; 897d74b50beSYan, Zheng 898d74b50beSYan, Zheng op->xattr.value_len = size; 8994766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, value, size); 9004766815bSDavid Disseldorp if (ret) 9014766815bSDavid Disseldorp goto err_pagelist_free; 902d74b50beSYan, Zheng payload_len += size; 903d74b50beSYan, Zheng 904d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 905d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 906d74b50beSYan, Zheng 907d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 908de2aa102SIlya Dryomov op->indata_len = payload_len; 909d74b50beSYan, Zheng return 0; 9104766815bSDavid Disseldorp 9114766815bSDavid Disseldorp err_pagelist_free: 9124766815bSDavid Disseldorp ceph_pagelist_release(pagelist); 9134766815bSDavid Disseldorp return ret; 914d74b50beSYan, Zheng } 915d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 916d74b50beSYan, Zheng 917922dab61SIlya Dryomov /* 918922dab61SIlya Dryomov * @watch_opcode: CEPH_OSD_WATCH_OP_* 919922dab61SIlya Dryomov */ 920922dab61SIlya Dryomov static void osd_req_op_watch_init(struct ceph_osd_request *req, int which, 921922dab61SIlya Dryomov u64 cookie, u8 watch_opcode) 92233803f33SAlex Elder { 923922dab61SIlya Dryomov struct ceph_osd_req_op *op; 92433803f33SAlex Elder 925042f6498SJeff Layton op = osd_req_op_init(req, which, CEPH_OSD_OP_WATCH, 0); 92633803f33SAlex Elder op->watch.cookie = cookie; 927922dab61SIlya Dryomov op->watch.op = watch_opcode; 928922dab61SIlya Dryomov op->watch.gen = 0; 92933803f33SAlex Elder } 93033803f33SAlex Elder 931d3798accSIlya Dryomov /* 932d3798accSIlya Dryomov * @flags: CEPH_OSD_OP_ALLOC_HINT_FLAG_* 933d3798accSIlya Dryomov */ 934c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 935c647b8a8SIlya Dryomov unsigned int which, 936c647b8a8SIlya Dryomov u64 expected_object_size, 937d3798accSIlya Dryomov u64 expected_write_size, 938d3798accSIlya Dryomov u32 flags) 939c647b8a8SIlya Dryomov { 940042f6498SJeff Layton struct ceph_osd_req_op *op; 941c647b8a8SIlya Dryomov 942042f6498SJeff Layton op = osd_req_op_init(osd_req, which, CEPH_OSD_OP_SETALLOCHINT, 0); 943c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 944c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 945d3798accSIlya Dryomov op->alloc_hint.flags = flags; 946c647b8a8SIlya Dryomov 947c647b8a8SIlya Dryomov /* 948c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 949c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 950c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 951c647b8a8SIlya Dryomov */ 952c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 953c647b8a8SIlya Dryomov } 954c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 955c647b8a8SIlya Dryomov 95690af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 957ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 958ec9123c5SAlex Elder { 959ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 960ec9123c5SAlex Elder 961ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 962ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 963ec9123c5SAlex Elder if (length) 96490af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 965e8862740SIlya Dryomov length, osd_data->alignment, false); 966ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 967ec9123c5SAlex Elder BUG_ON(!length); 96890af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 969ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 970ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 9715359a17dSIlya Dryomov ceph_msg_data_add_bio(msg, &osd_data->bio_pos, length); 972ec9123c5SAlex Elder #endif 973b9e281c2SIlya Dryomov } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BVECS) { 974b9e281c2SIlya Dryomov ceph_msg_data_add_bvecs(msg, &osd_data->bvec_pos); 975ec9123c5SAlex Elder } else { 976ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 977ec9123c5SAlex Elder } 978ec9123c5SAlex Elder } 979ec9123c5SAlex Elder 980bb873b53SIlya Dryomov static u32 osd_req_encode_op(struct ceph_osd_op *dst, 981bb873b53SIlya Dryomov const struct ceph_osd_req_op *src) 9823d14c5d2SYehuda Sadeh { 983065a68f9SAlex Elder switch (src->op) { 984fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 985fbfab539SAlex Elder break; 9863d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 9873d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 988e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 989ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 990ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 991175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 992175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 9933d14c5d2SYehuda Sadeh dst->extent.truncate_size = 9943d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 9953d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 9963d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 9973d14c5d2SYehuda Sadeh break; 9983d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 9993d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 10003d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 1001bb873b53SIlya Dryomov dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 10023d14c5d2SYehuda Sadeh break; 1003a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 1004a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 1005922dab61SIlya Dryomov dst->watch.ver = cpu_to_le64(0); 1006922dab61SIlya Dryomov dst->watch.op = src->watch.op; 1007922dab61SIlya Dryomov dst->watch.gen = cpu_to_le32(src->watch.gen); 1008922dab61SIlya Dryomov break; 1009922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 1010a40c4f10SYehuda Sadeh break; 101119079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 101219079203SIlya Dryomov dst->notify.cookie = cpu_to_le64(src->notify.cookie); 101319079203SIlya Dryomov break; 1014a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 1015a4ed38d7SDouglas Fuller break; 1016c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 1017c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 1018c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 1019c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 1020c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 1021d3798accSIlya Dryomov dst->alloc_hint.flags = cpu_to_le32(src->alloc_hint.flags); 1022c647b8a8SIlya Dryomov break; 1023d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 1024d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 1025d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 1026d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 1027d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 1028d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 1029d74b50beSYan, Zheng break; 1030864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 1031864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 1032864e9197SYan, Zheng break; 103378beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 103423ddf9beSLuis Henriques dst->copy_from.snapid = cpu_to_le64(src->copy_from.snapid); 103523ddf9beSLuis Henriques dst->copy_from.src_version = 103623ddf9beSLuis Henriques cpu_to_le64(src->copy_from.src_version); 103723ddf9beSLuis Henriques dst->copy_from.flags = src->copy_from.flags; 103823ddf9beSLuis Henriques dst->copy_from.src_fadvise_flags = 103923ddf9beSLuis Henriques cpu_to_le32(src->copy_from.src_fadvise_flags); 104023ddf9beSLuis Henriques break; 10413d14c5d2SYehuda Sadeh default: 10424c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 10438f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 10444c46459cSAlex Elder WARN_ON(1); 1045a8dd0a37SAlex Elder 1046a8dd0a37SAlex Elder return 0; 10473d14c5d2SYehuda Sadeh } 10487b25bf5fSIlya Dryomov 1049a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 10507b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 1051de2aa102SIlya Dryomov dst->payload_len = cpu_to_le32(src->indata_len); 1052175face2SAlex Elder 1053bb873b53SIlya Dryomov return src->indata_len; 10543d14c5d2SYehuda Sadeh } 10553d14c5d2SYehuda Sadeh 10563d14c5d2SYehuda Sadeh /* 10573d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 10583d14c5d2SYehuda Sadeh * extent as needed. 10593d14c5d2SYehuda Sadeh * 10603d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 10613d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 10623d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 10633d14c5d2SYehuda Sadeh */ 10643d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 10653d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 10663d14c5d2SYehuda Sadeh struct ceph_vino vino, 1067715e4cd4SYan, Zheng u64 off, u64 *plen, 1068715e4cd4SYan, Zheng unsigned int which, int num_ops, 10693d14c5d2SYehuda Sadeh int opcode, int flags, 10703d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 10713d14c5d2SYehuda Sadeh u32 truncate_seq, 10723d14c5d2SYehuda Sadeh u64 truncate_size, 1073153e5167SAlex Elder bool use_mempool) 10743d14c5d2SYehuda Sadeh { 10753d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 107675d1c941SAlex Elder u64 objnum = 0; 107775d1c941SAlex Elder u64 objoff = 0; 107875d1c941SAlex Elder u64 objlen = 0; 10796816282dSSage Weil int r; 10803d14c5d2SYehuda Sadeh 1081ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 1082864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 1083864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 10843d14c5d2SYehuda Sadeh 1085acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 1086ae7ca4a3SAlex Elder GFP_NOFS); 108713d1ad16SIlya Dryomov if (!req) { 108813d1ad16SIlya Dryomov r = -ENOMEM; 108913d1ad16SIlya Dryomov goto fail; 109013d1ad16SIlya Dryomov } 109179528734SAlex Elder 10923d14c5d2SYehuda Sadeh /* calculate max write size */ 1093a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 109413d1ad16SIlya Dryomov if (r) 109513d1ad16SIlya Dryomov goto fail; 1096a19dadfbSAlex Elder 1097864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 1098144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 1099864e9197SYan, Zheng } else { 11007627151eSYan, Zheng u32 object_size = layout->object_size; 1101864e9197SYan, Zheng u32 object_base = off - objoff; 1102ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 1103d18d1e28SAlex Elder if (truncate_size <= object_base) { 1104d18d1e28SAlex Elder truncate_size = 0; 1105d18d1e28SAlex Elder } else { 1106d18d1e28SAlex Elder truncate_size -= object_base; 1107d18d1e28SAlex Elder if (truncate_size > object_size) 1108d18d1e28SAlex Elder truncate_size = object_size; 1109d18d1e28SAlex Elder } 1110ccca4e37SYan, Zheng } 1111715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 1112b0270324SAlex Elder truncate_size, truncate_seq); 1113864e9197SYan, Zheng } 1114d18d1e28SAlex Elder 11157627151eSYan, Zheng req->r_base_oloc.pool = layout->pool_id; 111630c156d9SYan, Zheng req->r_base_oloc.pool_ns = ceph_try_get_string(layout->pool_ns); 1117d30291b9SIlya Dryomov ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum); 111822d2cfdfSIlya Dryomov req->r_flags = flags | osdc->client->options->read_from_replica; 1119dbe0fc41SAlex Elder 1120bb873b53SIlya Dryomov req->r_snapid = vino.snap; 1121bb873b53SIlya Dryomov if (flags & CEPH_OSD_FLAG_WRITE) 1122bb873b53SIlya Dryomov req->r_data_offset = off; 1123bb873b53SIlya Dryomov 11240d9c1ab3SIlya Dryomov if (num_ops > 1) 11250d9c1ab3SIlya Dryomov /* 11260d9c1ab3SIlya Dryomov * This is a special case for ceph_writepages_start(), but it 11270d9c1ab3SIlya Dryomov * also covers ceph_uninline_data(). If more multi-op request 11280d9c1ab3SIlya Dryomov * use cases emerge, we will need a separate helper. 11290d9c1ab3SIlya Dryomov */ 11300d9c1ab3SIlya Dryomov r = __ceph_osdc_alloc_messages(req, GFP_NOFS, num_ops, 0); 11310d9c1ab3SIlya Dryomov else 113213d1ad16SIlya Dryomov r = ceph_osdc_alloc_messages(req, GFP_NOFS); 113313d1ad16SIlya Dryomov if (r) 113413d1ad16SIlya Dryomov goto fail; 113513d1ad16SIlya Dryomov 11363d14c5d2SYehuda Sadeh return req; 113713d1ad16SIlya Dryomov 113813d1ad16SIlya Dryomov fail: 113913d1ad16SIlya Dryomov ceph_osdc_put_request(req); 114013d1ad16SIlya Dryomov return ERR_PTR(r); 11413d14c5d2SYehuda Sadeh } 11423d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 11433d14c5d2SYehuda Sadeh 11443d14c5d2SYehuda Sadeh /* 11453d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 11463d14c5d2SYehuda Sadeh */ 1147fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node) 11484609245eSIlya Dryomov DEFINE_RB_FUNCS(request_mc, struct ceph_osd_request, r_tid, r_mc_node) 11493d14c5d2SYehuda Sadeh 115066850df5SIlya Dryomov /* 115166850df5SIlya Dryomov * Call @fn on each OSD request as long as @fn returns 0. 115266850df5SIlya Dryomov */ 115366850df5SIlya Dryomov static void for_each_request(struct ceph_osd_client *osdc, 115466850df5SIlya Dryomov int (*fn)(struct ceph_osd_request *req, void *arg), 115566850df5SIlya Dryomov void *arg) 115666850df5SIlya Dryomov { 115766850df5SIlya Dryomov struct rb_node *n, *p; 115866850df5SIlya Dryomov 115966850df5SIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 116066850df5SIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 116166850df5SIlya Dryomov 116266850df5SIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 116366850df5SIlya Dryomov struct ceph_osd_request *req = 116466850df5SIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 116566850df5SIlya Dryomov 116666850df5SIlya Dryomov p = rb_next(p); 116766850df5SIlya Dryomov if (fn(req, arg)) 116866850df5SIlya Dryomov return; 116966850df5SIlya Dryomov } 117066850df5SIlya Dryomov } 117166850df5SIlya Dryomov 117266850df5SIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 117366850df5SIlya Dryomov struct ceph_osd_request *req = 117466850df5SIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 117566850df5SIlya Dryomov 117666850df5SIlya Dryomov p = rb_next(p); 117766850df5SIlya Dryomov if (fn(req, arg)) 117866850df5SIlya Dryomov return; 117966850df5SIlya Dryomov } 118066850df5SIlya Dryomov } 118166850df5SIlya Dryomov 11820247a0cfSIlya Dryomov static bool osd_homeless(struct ceph_osd *osd) 11830247a0cfSIlya Dryomov { 11840247a0cfSIlya Dryomov return osd->o_osd == CEPH_HOMELESS_OSD; 11850247a0cfSIlya Dryomov } 11860247a0cfSIlya Dryomov 11875aea3dcdSIlya Dryomov static bool osd_registered(struct ceph_osd *osd) 11883d14c5d2SYehuda Sadeh { 11895aea3dcdSIlya Dryomov verify_osdc_locked(osd->o_osdc); 11903d14c5d2SYehuda Sadeh 11915aea3dcdSIlya Dryomov return !RB_EMPTY_NODE(&osd->o_node); 11923d14c5d2SYehuda Sadeh } 11933d14c5d2SYehuda Sadeh 11943d14c5d2SYehuda Sadeh /* 11950247a0cfSIlya Dryomov * Assumes @osd is zero-initialized. 11960247a0cfSIlya Dryomov */ 11970247a0cfSIlya Dryomov static void osd_init(struct ceph_osd *osd) 11980247a0cfSIlya Dryomov { 119902113a0fSElena Reshetova refcount_set(&osd->o_ref, 1); 12000247a0cfSIlya Dryomov RB_CLEAR_NODE(&osd->o_node); 12015aea3dcdSIlya Dryomov osd->o_requests = RB_ROOT; 1202922dab61SIlya Dryomov osd->o_linger_requests = RB_ROOT; 1203a02a946dSIlya Dryomov osd->o_backoff_mappings = RB_ROOT; 1204a02a946dSIlya Dryomov osd->o_backoffs_by_id = RB_ROOT; 12050247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_osd_lru); 12060247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_keepalive_item); 12070247a0cfSIlya Dryomov osd->o_incarnation = 1; 12085aea3dcdSIlya Dryomov mutex_init(&osd->lock); 12090247a0cfSIlya Dryomov } 12100247a0cfSIlya Dryomov 12110247a0cfSIlya Dryomov static void osd_cleanup(struct ceph_osd *osd) 12120247a0cfSIlya Dryomov { 12130247a0cfSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&osd->o_node)); 12145aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 1215922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 1216a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoff_mappings)); 1217a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoffs_by_id)); 12180247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_osd_lru)); 12190247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_keepalive_item)); 12200247a0cfSIlya Dryomov 12210247a0cfSIlya Dryomov if (osd->o_auth.authorizer) { 12220247a0cfSIlya Dryomov WARN_ON(osd_homeless(osd)); 12230247a0cfSIlya Dryomov ceph_auth_destroy_authorizer(osd->o_auth.authorizer); 12240247a0cfSIlya Dryomov } 12250247a0cfSIlya Dryomov } 12260247a0cfSIlya Dryomov 12270247a0cfSIlya Dryomov /* 12283d14c5d2SYehuda Sadeh * Track open sessions with osds. 12293d14c5d2SYehuda Sadeh */ 1230e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 12313d14c5d2SYehuda Sadeh { 12323d14c5d2SYehuda Sadeh struct ceph_osd *osd; 12333d14c5d2SYehuda Sadeh 12340247a0cfSIlya Dryomov WARN_ON(onum == CEPH_HOMELESS_OSD); 12350247a0cfSIlya Dryomov 12367a28f59bSIlya Dryomov osd = kzalloc(sizeof(*osd), GFP_NOIO | __GFP_NOFAIL); 12370247a0cfSIlya Dryomov osd_init(osd); 12383d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1239e10006f8SAlex Elder osd->o_osd = onum; 12403d14c5d2SYehuda Sadeh 1241b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 12423d14c5d2SYehuda Sadeh 12433d14c5d2SYehuda Sadeh return osd; 12443d14c5d2SYehuda Sadeh } 12453d14c5d2SYehuda Sadeh 12463d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 12473d14c5d2SYehuda Sadeh { 124802113a0fSElena Reshetova if (refcount_inc_not_zero(&osd->o_ref)) { 124902113a0fSElena Reshetova dout("get_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref)-1, 125002113a0fSElena Reshetova refcount_read(&osd->o_ref)); 12513d14c5d2SYehuda Sadeh return osd; 12523d14c5d2SYehuda Sadeh } else { 12533d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 12543d14c5d2SYehuda Sadeh return NULL; 12553d14c5d2SYehuda Sadeh } 12563d14c5d2SYehuda Sadeh } 12573d14c5d2SYehuda Sadeh 12583d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 12593d14c5d2SYehuda Sadeh { 126002113a0fSElena Reshetova dout("put_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref), 126102113a0fSElena Reshetova refcount_read(&osd->o_ref) - 1); 126202113a0fSElena Reshetova if (refcount_dec_and_test(&osd->o_ref)) { 12630247a0cfSIlya Dryomov osd_cleanup(osd); 12643d14c5d2SYehuda Sadeh kfree(osd); 12653d14c5d2SYehuda Sadeh } 12663d14c5d2SYehuda Sadeh } 12673d14c5d2SYehuda Sadeh 1268fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node) 1269fcd00b68SIlya Dryomov 12709dd2845cSIlya Dryomov static void __move_osd_to_lru(struct ceph_osd *osd) 12713d14c5d2SYehuda Sadeh { 12729dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 12739dd2845cSIlya Dryomov 12749dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12753d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1276bbf37ec3SIlya Dryomov 12779dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 12783d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 12799dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 12809dd2845cSIlya Dryomov 1281a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 12823d14c5d2SYehuda Sadeh } 12833d14c5d2SYehuda Sadeh 12849dd2845cSIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd *osd) 1285bbf37ec3SIlya Dryomov { 12865aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1287922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) 12889dd2845cSIlya Dryomov __move_osd_to_lru(osd); 1289bbf37ec3SIlya Dryomov } 1290bbf37ec3SIlya Dryomov 12913d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 12923d14c5d2SYehuda Sadeh { 12939dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 12949dd2845cSIlya Dryomov 12959dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12969dd2845cSIlya Dryomov 12979dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 12983d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 12993d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 13009dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 13013d14c5d2SYehuda Sadeh } 13023d14c5d2SYehuda Sadeh 13033d14c5d2SYehuda Sadeh /* 13045aea3dcdSIlya Dryomov * Close the connection and assign any leftover requests to the 13055aea3dcdSIlya Dryomov * homeless session. 13065aea3dcdSIlya Dryomov */ 13075aea3dcdSIlya Dryomov static void close_osd(struct ceph_osd *osd) 13085aea3dcdSIlya Dryomov { 13095aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 13105aea3dcdSIlya Dryomov struct rb_node *n; 13115aea3dcdSIlya Dryomov 13125aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 13135aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 13145aea3dcdSIlya Dryomov 13155aea3dcdSIlya Dryomov ceph_con_close(&osd->o_con); 13165aea3dcdSIlya Dryomov 13175aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 13185aea3dcdSIlya Dryomov struct ceph_osd_request *req = 13195aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 13205aea3dcdSIlya Dryomov 13215aea3dcdSIlya Dryomov n = rb_next(n); /* unlink_request() */ 13225aea3dcdSIlya Dryomov 13235aea3dcdSIlya Dryomov dout(" reassigning req %p tid %llu\n", req, req->r_tid); 13245aea3dcdSIlya Dryomov unlink_request(osd, req); 13255aea3dcdSIlya Dryomov link_request(&osdc->homeless_osd, req); 13265aea3dcdSIlya Dryomov } 1327922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 1328922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 1329922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 1330922dab61SIlya Dryomov 1331922dab61SIlya Dryomov n = rb_next(n); /* unlink_linger() */ 1332922dab61SIlya Dryomov 1333922dab61SIlya Dryomov dout(" reassigning lreq %p linger_id %llu\n", lreq, 1334922dab61SIlya Dryomov lreq->linger_id); 1335922dab61SIlya Dryomov unlink_linger(osd, lreq); 1336922dab61SIlya Dryomov link_linger(&osdc->homeless_osd, lreq); 1337922dab61SIlya Dryomov } 1338a02a946dSIlya Dryomov clear_backoffs(osd); 13395aea3dcdSIlya Dryomov 13405aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 13415aea3dcdSIlya Dryomov erase_osd(&osdc->osds, osd); 13425aea3dcdSIlya Dryomov put_osd(osd); 13435aea3dcdSIlya Dryomov } 13445aea3dcdSIlya Dryomov 13455aea3dcdSIlya Dryomov /* 13463d14c5d2SYehuda Sadeh * reset osd connect 13473d14c5d2SYehuda Sadeh */ 13485aea3dcdSIlya Dryomov static int reopen_osd(struct ceph_osd *osd) 13493d14c5d2SYehuda Sadeh { 1350c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 13513d14c5d2SYehuda Sadeh 13525aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 13535aea3dcdSIlya Dryomov 13545aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1355922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) { 13565aea3dcdSIlya Dryomov close_osd(osd); 1357c3acb181SAlex Elder return -ENODEV; 1358c3acb181SAlex Elder } 1359c3acb181SAlex Elder 13605aea3dcdSIlya Dryomov peer_addr = &osd->o_osdc->osdmap->osd_addr[osd->o_osd]; 1361c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 13623d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 13635aea3dcdSIlya Dryomov struct rb_node *n; 1364c3acb181SAlex Elder 13653d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 13660b4af2e8SIlya Dryomov "letting msgr retry\n"); 13673d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 13685aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 13695aea3dcdSIlya Dryomov struct ceph_osd_request *req = 13705aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 13713d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 13725aea3dcdSIlya Dryomov } 1373c3acb181SAlex Elder 1374c3acb181SAlex Elder return -EAGAIN; 13753d14c5d2SYehuda Sadeh } 1376c3acb181SAlex Elder 1377c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1378c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1379c3acb181SAlex Elder osd->o_incarnation++; 1380c3acb181SAlex Elder 1381c3acb181SAlex Elder return 0; 13823d14c5d2SYehuda Sadeh } 13833d14c5d2SYehuda Sadeh 13845aea3dcdSIlya Dryomov static struct ceph_osd *lookup_create_osd(struct ceph_osd_client *osdc, int o, 13855aea3dcdSIlya Dryomov bool wrlocked) 13863d14c5d2SYehuda Sadeh { 13875aea3dcdSIlya Dryomov struct ceph_osd *osd; 13885aea3dcdSIlya Dryomov 13895aea3dcdSIlya Dryomov if (wrlocked) 13905aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 13915aea3dcdSIlya Dryomov else 13925aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 13935aea3dcdSIlya Dryomov 13945aea3dcdSIlya Dryomov if (o != CEPH_HOMELESS_OSD) 13955aea3dcdSIlya Dryomov osd = lookup_osd(&osdc->osds, o); 13965aea3dcdSIlya Dryomov else 13975aea3dcdSIlya Dryomov osd = &osdc->homeless_osd; 13985aea3dcdSIlya Dryomov if (!osd) { 13995aea3dcdSIlya Dryomov if (!wrlocked) 14005aea3dcdSIlya Dryomov return ERR_PTR(-EAGAIN); 14015aea3dcdSIlya Dryomov 14025aea3dcdSIlya Dryomov osd = create_osd(osdc, o); 14035aea3dcdSIlya Dryomov insert_osd(&osdc->osds, osd); 14045aea3dcdSIlya Dryomov ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, 14055aea3dcdSIlya Dryomov &osdc->osdmap->osd_addr[osd->o_osd]); 14065aea3dcdSIlya Dryomov } 14075aea3dcdSIlya Dryomov 14085aea3dcdSIlya Dryomov dout("%s osdc %p osd%d -> osd %p\n", __func__, osdc, o, osd); 14095aea3dcdSIlya Dryomov return osd; 1410a40c4f10SYehuda Sadeh } 1411a40c4f10SYehuda Sadeh 14123d14c5d2SYehuda Sadeh /* 14135aea3dcdSIlya Dryomov * Create request <-> OSD session relation. 14145aea3dcdSIlya Dryomov * 14155aea3dcdSIlya Dryomov * @req has to be assigned a tid, @osd may be homeless. 14163d14c5d2SYehuda Sadeh */ 14175aea3dcdSIlya Dryomov static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req) 14183d14c5d2SYehuda Sadeh { 14195aea3dcdSIlya Dryomov verify_osd_locked(osd); 14205aea3dcdSIlya Dryomov WARN_ON(!req->r_tid || req->r_osd); 14215aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 142235f9f8a0SSage Weil req, req->r_tid); 14235aea3dcdSIlya Dryomov 14245aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 14255aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 14265aea3dcdSIlya Dryomov else 14275aea3dcdSIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 14285aea3dcdSIlya Dryomov 14295aea3dcdSIlya Dryomov get_osd(osd); 14305aea3dcdSIlya Dryomov insert_request(&osd->o_requests, req); 14315aea3dcdSIlya Dryomov req->r_osd = osd; 143235f9f8a0SSage Weil } 143335f9f8a0SSage Weil 14345aea3dcdSIlya Dryomov static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req) 14353d14c5d2SYehuda Sadeh { 14365aea3dcdSIlya Dryomov verify_osd_locked(osd); 14375aea3dcdSIlya Dryomov WARN_ON(req->r_osd != osd); 14385aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 14395aea3dcdSIlya Dryomov req, req->r_tid); 14405aea3dcdSIlya Dryomov 14415aea3dcdSIlya Dryomov req->r_osd = NULL; 14425aea3dcdSIlya Dryomov erase_request(&osd->o_requests, req); 14435aea3dcdSIlya Dryomov put_osd(osd); 14445aea3dcdSIlya Dryomov 14455aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 14465aea3dcdSIlya Dryomov maybe_move_osd_to_lru(osd); 14475aea3dcdSIlya Dryomov else 14485aea3dcdSIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 14493d14c5d2SYehuda Sadeh } 14503d14c5d2SYehuda Sadeh 145163244fa1SIlya Dryomov static bool __pool_full(struct ceph_pg_pool_info *pi) 145263244fa1SIlya Dryomov { 145363244fa1SIlya Dryomov return pi->flags & CEPH_POOL_FLAG_FULL; 145463244fa1SIlya Dryomov } 145563244fa1SIlya Dryomov 145642c1b124SIlya Dryomov static bool have_pool_full(struct ceph_osd_client *osdc) 145742c1b124SIlya Dryomov { 145842c1b124SIlya Dryomov struct rb_node *n; 145942c1b124SIlya Dryomov 146042c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 146142c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 146242c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 146342c1b124SIlya Dryomov 146442c1b124SIlya Dryomov if (__pool_full(pi)) 146542c1b124SIlya Dryomov return true; 146642c1b124SIlya Dryomov } 146742c1b124SIlya Dryomov 146842c1b124SIlya Dryomov return false; 146942c1b124SIlya Dryomov } 147042c1b124SIlya Dryomov 14715aea3dcdSIlya Dryomov static bool pool_full(struct ceph_osd_client *osdc, s64 pool_id) 14725aea3dcdSIlya Dryomov { 14735aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 14745aea3dcdSIlya Dryomov 14755aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 14765aea3dcdSIlya Dryomov if (!pi) 14775aea3dcdSIlya Dryomov return false; 14785aea3dcdSIlya Dryomov 14795aea3dcdSIlya Dryomov return __pool_full(pi); 14805aea3dcdSIlya Dryomov } 14815aea3dcdSIlya Dryomov 14823d14c5d2SYehuda Sadeh /* 1483d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1484d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1485d29adb34SJosh Durgin */ 148663244fa1SIlya Dryomov static bool target_should_be_paused(struct ceph_osd_client *osdc, 148763244fa1SIlya Dryomov const struct ceph_osd_request_target *t, 148863244fa1SIlya Dryomov struct ceph_pg_pool_info *pi) 148963244fa1SIlya Dryomov { 1490b7ec35b3SIlya Dryomov bool pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 1491b7ec35b3SIlya Dryomov bool pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 1492b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 149363244fa1SIlya Dryomov __pool_full(pi); 149463244fa1SIlya Dryomov 14956d637a54SIlya Dryomov WARN_ON(pi->id != t->target_oloc.pool); 149658eb7932SJeff Layton return ((t->flags & CEPH_OSD_FLAG_READ) && pauserd) || 149758eb7932SJeff Layton ((t->flags & CEPH_OSD_FLAG_WRITE) && pausewr) || 149858eb7932SJeff Layton (osdc->osdmap->epoch < osdc->epoch_barrier); 149963244fa1SIlya Dryomov } 150063244fa1SIlya Dryomov 1501117d96a0SIlya Dryomov static int pick_random_replica(const struct ceph_osds *acting) 1502117d96a0SIlya Dryomov { 1503117d96a0SIlya Dryomov int i = prandom_u32() % acting->size; 1504117d96a0SIlya Dryomov 1505117d96a0SIlya Dryomov dout("%s picked osd%d, primary osd%d\n", __func__, 1506117d96a0SIlya Dryomov acting->osds[i], acting->primary); 1507117d96a0SIlya Dryomov return i; 1508117d96a0SIlya Dryomov } 1509117d96a0SIlya Dryomov 1510117d96a0SIlya Dryomov /* 1511117d96a0SIlya Dryomov * Picks the closest replica based on client's location given by 1512117d96a0SIlya Dryomov * crush_location option. Prefers the primary if the locality is 1513117d96a0SIlya Dryomov * the same. 1514117d96a0SIlya Dryomov */ 1515117d96a0SIlya Dryomov static int pick_closest_replica(struct ceph_osd_client *osdc, 1516117d96a0SIlya Dryomov const struct ceph_osds *acting) 1517117d96a0SIlya Dryomov { 1518117d96a0SIlya Dryomov struct ceph_options *opt = osdc->client->options; 1519117d96a0SIlya Dryomov int best_i, best_locality; 1520117d96a0SIlya Dryomov int i = 0, locality; 1521117d96a0SIlya Dryomov 1522117d96a0SIlya Dryomov do { 1523117d96a0SIlya Dryomov locality = ceph_get_crush_locality(osdc->osdmap, 1524117d96a0SIlya Dryomov acting->osds[i], 1525117d96a0SIlya Dryomov &opt->crush_locs); 1526117d96a0SIlya Dryomov if (i == 0 || 1527117d96a0SIlya Dryomov (locality >= 0 && best_locality < 0) || 1528117d96a0SIlya Dryomov (locality >= 0 && best_locality >= 0 && 1529117d96a0SIlya Dryomov locality < best_locality)) { 1530117d96a0SIlya Dryomov best_i = i; 1531117d96a0SIlya Dryomov best_locality = locality; 1532117d96a0SIlya Dryomov } 1533117d96a0SIlya Dryomov } while (++i < acting->size); 1534117d96a0SIlya Dryomov 1535117d96a0SIlya Dryomov dout("%s picked osd%d with locality %d, primary osd%d\n", __func__, 1536117d96a0SIlya Dryomov acting->osds[best_i], best_locality, acting->primary); 1537117d96a0SIlya Dryomov return best_i; 1538117d96a0SIlya Dryomov } 1539117d96a0SIlya Dryomov 154063244fa1SIlya Dryomov enum calc_target_result { 154163244fa1SIlya Dryomov CALC_TARGET_NO_ACTION = 0, 154263244fa1SIlya Dryomov CALC_TARGET_NEED_RESEND, 154363244fa1SIlya Dryomov CALC_TARGET_POOL_DNE, 154463244fa1SIlya Dryomov }; 154563244fa1SIlya Dryomov 154663244fa1SIlya Dryomov static enum calc_target_result calc_target(struct ceph_osd_client *osdc, 154763244fa1SIlya Dryomov struct ceph_osd_request_target *t, 154863244fa1SIlya Dryomov bool any_change) 154963244fa1SIlya Dryomov { 155063244fa1SIlya Dryomov struct ceph_pg_pool_info *pi; 155163244fa1SIlya Dryomov struct ceph_pg pgid, last_pgid; 155263244fa1SIlya Dryomov struct ceph_osds up, acting; 1553117d96a0SIlya Dryomov bool is_read = t->flags & CEPH_OSD_FLAG_READ; 1554117d96a0SIlya Dryomov bool is_write = t->flags & CEPH_OSD_FLAG_WRITE; 155563244fa1SIlya Dryomov bool force_resend = false; 155684ed45dfSIlya Dryomov bool unpaused = false; 1557a5613724SIlya Dryomov bool legacy_change = false; 15587de030d6SIlya Dryomov bool split = false; 1559b7ec35b3SIlya Dryomov bool sort_bitwise = ceph_osdmap_flag(osdc, CEPH_OSDMAP_SORTBITWISE); 1560ae78dd81SIlya Dryomov bool recovery_deletes = ceph_osdmap_flag(osdc, 1561ae78dd81SIlya Dryomov CEPH_OSDMAP_RECOVERY_DELETES); 156263244fa1SIlya Dryomov enum calc_target_result ct_res; 156363244fa1SIlya Dryomov 156404c7d789SIlya Dryomov t->epoch = osdc->osdmap->epoch; 156563244fa1SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool); 156663244fa1SIlya Dryomov if (!pi) { 156763244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 156863244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 156963244fa1SIlya Dryomov goto out; 157063244fa1SIlya Dryomov } 157163244fa1SIlya Dryomov 157263244fa1SIlya Dryomov if (osdc->osdmap->epoch == pi->last_force_request_resend) { 1573dc93e0e2SIlya Dryomov if (t->last_force_resend < pi->last_force_request_resend) { 1574dc93e0e2SIlya Dryomov t->last_force_resend = pi->last_force_request_resend; 157563244fa1SIlya Dryomov force_resend = true; 1576dc93e0e2SIlya Dryomov } else if (t->last_force_resend == 0) { 157763244fa1SIlya Dryomov force_resend = true; 157863244fa1SIlya Dryomov } 157963244fa1SIlya Dryomov } 158063244fa1SIlya Dryomov 1581db098ec4SIlya Dryomov /* apply tiering */ 1582db098ec4SIlya Dryomov ceph_oid_copy(&t->target_oid, &t->base_oid); 1583db098ec4SIlya Dryomov ceph_oloc_copy(&t->target_oloc, &t->base_oloc); 1584db098ec4SIlya Dryomov if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 1585117d96a0SIlya Dryomov if (is_read && pi->read_tier >= 0) 158663244fa1SIlya Dryomov t->target_oloc.pool = pi->read_tier; 1587117d96a0SIlya Dryomov if (is_write && pi->write_tier >= 0) 158863244fa1SIlya Dryomov t->target_oloc.pool = pi->write_tier; 15896d637a54SIlya Dryomov 15906d637a54SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool); 15916d637a54SIlya Dryomov if (!pi) { 15926d637a54SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 15936d637a54SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 15946d637a54SIlya Dryomov goto out; 15956d637a54SIlya Dryomov } 159663244fa1SIlya Dryomov } 159763244fa1SIlya Dryomov 1598a86f009fSIlya Dryomov __ceph_object_locator_to_pg(pi, &t->target_oid, &t->target_oloc, &pgid); 159963244fa1SIlya Dryomov last_pgid.pool = pgid.pool; 160063244fa1SIlya Dryomov last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask); 160163244fa1SIlya Dryomov 1602df28152dSIlya Dryomov ceph_pg_to_up_acting_osds(osdc->osdmap, pi, &pgid, &up, &acting); 160363244fa1SIlya Dryomov if (any_change && 160463244fa1SIlya Dryomov ceph_is_new_interval(&t->acting, 160563244fa1SIlya Dryomov &acting, 160663244fa1SIlya Dryomov &t->up, 160763244fa1SIlya Dryomov &up, 160863244fa1SIlya Dryomov t->size, 160963244fa1SIlya Dryomov pi->size, 161063244fa1SIlya Dryomov t->min_size, 161163244fa1SIlya Dryomov pi->min_size, 161263244fa1SIlya Dryomov t->pg_num, 161363244fa1SIlya Dryomov pi->pg_num, 161463244fa1SIlya Dryomov t->sort_bitwise, 161563244fa1SIlya Dryomov sort_bitwise, 1616ae78dd81SIlya Dryomov t->recovery_deletes, 1617ae78dd81SIlya Dryomov recovery_deletes, 161863244fa1SIlya Dryomov &last_pgid)) 161963244fa1SIlya Dryomov force_resend = true; 162063244fa1SIlya Dryomov 162163244fa1SIlya Dryomov if (t->paused && !target_should_be_paused(osdc, t, pi)) { 162263244fa1SIlya Dryomov t->paused = false; 162384ed45dfSIlya Dryomov unpaused = true; 162463244fa1SIlya Dryomov } 162584ed45dfSIlya Dryomov legacy_change = ceph_pg_compare(&t->pgid, &pgid) || 1626117d96a0SIlya Dryomov ceph_osds_changed(&t->acting, &acting, 1627117d96a0SIlya Dryomov t->used_replica || any_change); 16287de030d6SIlya Dryomov if (t->pg_num) 16297de030d6SIlya Dryomov split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num); 163063244fa1SIlya Dryomov 16317de030d6SIlya Dryomov if (legacy_change || force_resend || split) { 163263244fa1SIlya Dryomov t->pgid = pgid; /* struct */ 1633df28152dSIlya Dryomov ceph_pg_to_primary_shard(osdc->osdmap, pi, &pgid, &t->spgid); 163463244fa1SIlya Dryomov ceph_osds_copy(&t->acting, &acting); 163563244fa1SIlya Dryomov ceph_osds_copy(&t->up, &up); 163663244fa1SIlya Dryomov t->size = pi->size; 163763244fa1SIlya Dryomov t->min_size = pi->min_size; 163863244fa1SIlya Dryomov t->pg_num = pi->pg_num; 163963244fa1SIlya Dryomov t->pg_num_mask = pi->pg_num_mask; 164063244fa1SIlya Dryomov t->sort_bitwise = sort_bitwise; 1641ae78dd81SIlya Dryomov t->recovery_deletes = recovery_deletes; 164263244fa1SIlya Dryomov 1643117d96a0SIlya Dryomov if ((t->flags & (CEPH_OSD_FLAG_BALANCE_READS | 1644117d96a0SIlya Dryomov CEPH_OSD_FLAG_LOCALIZE_READS)) && 1645117d96a0SIlya Dryomov !is_write && pi->type == CEPH_POOL_TYPE_REP && 1646117d96a0SIlya Dryomov acting.size > 1) { 1647117d96a0SIlya Dryomov int pos; 1648117d96a0SIlya Dryomov 1649117d96a0SIlya Dryomov WARN_ON(!is_read || acting.osds[0] != acting.primary); 1650117d96a0SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_BALANCE_READS) { 1651117d96a0SIlya Dryomov pos = pick_random_replica(&acting); 1652117d96a0SIlya Dryomov } else { 1653117d96a0SIlya Dryomov pos = pick_closest_replica(osdc, &acting); 1654117d96a0SIlya Dryomov } 1655117d96a0SIlya Dryomov t->osd = acting.osds[pos]; 1656117d96a0SIlya Dryomov t->used_replica = pos > 0; 1657117d96a0SIlya Dryomov } else { 165863244fa1SIlya Dryomov t->osd = acting.primary; 1659117d96a0SIlya Dryomov t->used_replica = false; 1660117d96a0SIlya Dryomov } 166163244fa1SIlya Dryomov } 166263244fa1SIlya Dryomov 1663a5613724SIlya Dryomov if (unpaused || legacy_change || force_resend || split) 166484ed45dfSIlya Dryomov ct_res = CALC_TARGET_NEED_RESEND; 166584ed45dfSIlya Dryomov else 166684ed45dfSIlya Dryomov ct_res = CALC_TARGET_NO_ACTION; 166784ed45dfSIlya Dryomov 166863244fa1SIlya Dryomov out: 1669a5613724SIlya Dryomov dout("%s t %p -> %d%d%d%d ct_res %d osd%d\n", __func__, t, unpaused, 1670a5613724SIlya Dryomov legacy_change, force_resend, split, ct_res, t->osd); 167163244fa1SIlya Dryomov return ct_res; 167263244fa1SIlya Dryomov } 167363244fa1SIlya Dryomov 1674a02a946dSIlya Dryomov static struct ceph_spg_mapping *alloc_spg_mapping(void) 1675a02a946dSIlya Dryomov { 1676a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1677a02a946dSIlya Dryomov 1678a02a946dSIlya Dryomov spg = kmalloc(sizeof(*spg), GFP_NOIO); 1679a02a946dSIlya Dryomov if (!spg) 1680a02a946dSIlya Dryomov return NULL; 1681a02a946dSIlya Dryomov 1682a02a946dSIlya Dryomov RB_CLEAR_NODE(&spg->node); 1683a02a946dSIlya Dryomov spg->backoffs = RB_ROOT; 1684a02a946dSIlya Dryomov return spg; 1685a02a946dSIlya Dryomov } 1686a02a946dSIlya Dryomov 1687a02a946dSIlya Dryomov static void free_spg_mapping(struct ceph_spg_mapping *spg) 1688a02a946dSIlya Dryomov { 1689a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&spg->node)); 1690a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&spg->backoffs)); 1691a02a946dSIlya Dryomov 1692a02a946dSIlya Dryomov kfree(spg); 1693a02a946dSIlya Dryomov } 1694a02a946dSIlya Dryomov 1695a02a946dSIlya Dryomov /* 1696a02a946dSIlya Dryomov * rbtree of ceph_spg_mapping for handling map<spg_t, ...>, similar to 1697a02a946dSIlya Dryomov * ceph_pg_mapping. Used to track OSD backoffs -- a backoff [range] is 1698a02a946dSIlya Dryomov * defined only within a specific spgid; it does not pass anything to 1699a02a946dSIlya Dryomov * children on split, or to another primary. 1700a02a946dSIlya Dryomov */ 1701a02a946dSIlya Dryomov DEFINE_RB_FUNCS2(spg_mapping, struct ceph_spg_mapping, spgid, ceph_spg_compare, 1702a02a946dSIlya Dryomov RB_BYPTR, const struct ceph_spg *, node) 1703a02a946dSIlya Dryomov 1704a02a946dSIlya Dryomov static u64 hoid_get_bitwise_key(const struct ceph_hobject_id *hoid) 1705a02a946dSIlya Dryomov { 1706a02a946dSIlya Dryomov return hoid->is_max ? 0x100000000ull : hoid->hash_reverse_bits; 1707a02a946dSIlya Dryomov } 1708a02a946dSIlya Dryomov 1709a02a946dSIlya Dryomov static void hoid_get_effective_key(const struct ceph_hobject_id *hoid, 1710a02a946dSIlya Dryomov void **pkey, size_t *pkey_len) 1711a02a946dSIlya Dryomov { 1712a02a946dSIlya Dryomov if (hoid->key_len) { 1713a02a946dSIlya Dryomov *pkey = hoid->key; 1714a02a946dSIlya Dryomov *pkey_len = hoid->key_len; 1715a02a946dSIlya Dryomov } else { 1716a02a946dSIlya Dryomov *pkey = hoid->oid; 1717a02a946dSIlya Dryomov *pkey_len = hoid->oid_len; 1718a02a946dSIlya Dryomov } 1719a02a946dSIlya Dryomov } 1720a02a946dSIlya Dryomov 1721a02a946dSIlya Dryomov static int compare_names(const void *name1, size_t name1_len, 1722a02a946dSIlya Dryomov const void *name2, size_t name2_len) 1723a02a946dSIlya Dryomov { 1724a02a946dSIlya Dryomov int ret; 1725a02a946dSIlya Dryomov 1726a02a946dSIlya Dryomov ret = memcmp(name1, name2, min(name1_len, name2_len)); 1727a02a946dSIlya Dryomov if (!ret) { 1728a02a946dSIlya Dryomov if (name1_len < name2_len) 1729a02a946dSIlya Dryomov ret = -1; 1730a02a946dSIlya Dryomov else if (name1_len > name2_len) 1731a02a946dSIlya Dryomov ret = 1; 1732a02a946dSIlya Dryomov } 1733a02a946dSIlya Dryomov return ret; 1734a02a946dSIlya Dryomov } 1735a02a946dSIlya Dryomov 1736a02a946dSIlya Dryomov static int hoid_compare(const struct ceph_hobject_id *lhs, 1737a02a946dSIlya Dryomov const struct ceph_hobject_id *rhs) 1738a02a946dSIlya Dryomov { 1739a02a946dSIlya Dryomov void *effective_key1, *effective_key2; 1740a02a946dSIlya Dryomov size_t effective_key1_len, effective_key2_len; 1741a02a946dSIlya Dryomov int ret; 1742a02a946dSIlya Dryomov 1743a02a946dSIlya Dryomov if (lhs->is_max < rhs->is_max) 1744a02a946dSIlya Dryomov return -1; 1745a02a946dSIlya Dryomov if (lhs->is_max > rhs->is_max) 1746a02a946dSIlya Dryomov return 1; 1747a02a946dSIlya Dryomov 1748a02a946dSIlya Dryomov if (lhs->pool < rhs->pool) 1749a02a946dSIlya Dryomov return -1; 1750a02a946dSIlya Dryomov if (lhs->pool > rhs->pool) 1751a02a946dSIlya Dryomov return 1; 1752a02a946dSIlya Dryomov 1753a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) < hoid_get_bitwise_key(rhs)) 1754a02a946dSIlya Dryomov return -1; 1755a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) > hoid_get_bitwise_key(rhs)) 1756a02a946dSIlya Dryomov return 1; 1757a02a946dSIlya Dryomov 1758a02a946dSIlya Dryomov ret = compare_names(lhs->nspace, lhs->nspace_len, 1759a02a946dSIlya Dryomov rhs->nspace, rhs->nspace_len); 1760a02a946dSIlya Dryomov if (ret) 1761a02a946dSIlya Dryomov return ret; 1762a02a946dSIlya Dryomov 1763a02a946dSIlya Dryomov hoid_get_effective_key(lhs, &effective_key1, &effective_key1_len); 1764a02a946dSIlya Dryomov hoid_get_effective_key(rhs, &effective_key2, &effective_key2_len); 1765a02a946dSIlya Dryomov ret = compare_names(effective_key1, effective_key1_len, 1766a02a946dSIlya Dryomov effective_key2, effective_key2_len); 1767a02a946dSIlya Dryomov if (ret) 1768a02a946dSIlya Dryomov return ret; 1769a02a946dSIlya Dryomov 1770a02a946dSIlya Dryomov ret = compare_names(lhs->oid, lhs->oid_len, rhs->oid, rhs->oid_len); 1771a02a946dSIlya Dryomov if (ret) 1772a02a946dSIlya Dryomov return ret; 1773a02a946dSIlya Dryomov 1774a02a946dSIlya Dryomov if (lhs->snapid < rhs->snapid) 1775a02a946dSIlya Dryomov return -1; 1776a02a946dSIlya Dryomov if (lhs->snapid > rhs->snapid) 1777a02a946dSIlya Dryomov return 1; 1778a02a946dSIlya Dryomov 1779a02a946dSIlya Dryomov return 0; 1780a02a946dSIlya Dryomov } 1781a02a946dSIlya Dryomov 1782a02a946dSIlya Dryomov /* 1783a02a946dSIlya Dryomov * For decoding ->begin and ->end of MOSDBackoff only -- no MIN/MAX 1784a02a946dSIlya Dryomov * compat stuff here. 1785a02a946dSIlya Dryomov * 1786a02a946dSIlya Dryomov * Assumes @hoid is zero-initialized. 1787a02a946dSIlya Dryomov */ 1788a02a946dSIlya Dryomov static int decode_hoid(void **p, void *end, struct ceph_hobject_id *hoid) 1789a02a946dSIlya Dryomov { 1790a02a946dSIlya Dryomov u8 struct_v; 1791a02a946dSIlya Dryomov u32 struct_len; 1792a02a946dSIlya Dryomov int ret; 1793a02a946dSIlya Dryomov 1794a02a946dSIlya Dryomov ret = ceph_start_decoding(p, end, 4, "hobject_t", &struct_v, 1795a02a946dSIlya Dryomov &struct_len); 1796a02a946dSIlya Dryomov if (ret) 1797a02a946dSIlya Dryomov return ret; 1798a02a946dSIlya Dryomov 1799a02a946dSIlya Dryomov if (struct_v < 4) { 1800a02a946dSIlya Dryomov pr_err("got struct_v %d < 4 of hobject_t\n", struct_v); 1801a02a946dSIlya Dryomov goto e_inval; 1802a02a946dSIlya Dryomov } 1803a02a946dSIlya Dryomov 1804a02a946dSIlya Dryomov hoid->key = ceph_extract_encoded_string(p, end, &hoid->key_len, 1805a02a946dSIlya Dryomov GFP_NOIO); 1806a02a946dSIlya Dryomov if (IS_ERR(hoid->key)) { 1807a02a946dSIlya Dryomov ret = PTR_ERR(hoid->key); 1808a02a946dSIlya Dryomov hoid->key = NULL; 1809a02a946dSIlya Dryomov return ret; 1810a02a946dSIlya Dryomov } 1811a02a946dSIlya Dryomov 1812a02a946dSIlya Dryomov hoid->oid = ceph_extract_encoded_string(p, end, &hoid->oid_len, 1813a02a946dSIlya Dryomov GFP_NOIO); 1814a02a946dSIlya Dryomov if (IS_ERR(hoid->oid)) { 1815a02a946dSIlya Dryomov ret = PTR_ERR(hoid->oid); 1816a02a946dSIlya Dryomov hoid->oid = NULL; 1817a02a946dSIlya Dryomov return ret; 1818a02a946dSIlya Dryomov } 1819a02a946dSIlya Dryomov 1820a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->snapid, e_inval); 1821a02a946dSIlya Dryomov ceph_decode_32_safe(p, end, hoid->hash, e_inval); 1822a02a946dSIlya Dryomov ceph_decode_8_safe(p, end, hoid->is_max, e_inval); 1823a02a946dSIlya Dryomov 1824a02a946dSIlya Dryomov hoid->nspace = ceph_extract_encoded_string(p, end, &hoid->nspace_len, 1825a02a946dSIlya Dryomov GFP_NOIO); 1826a02a946dSIlya Dryomov if (IS_ERR(hoid->nspace)) { 1827a02a946dSIlya Dryomov ret = PTR_ERR(hoid->nspace); 1828a02a946dSIlya Dryomov hoid->nspace = NULL; 1829a02a946dSIlya Dryomov return ret; 1830a02a946dSIlya Dryomov } 1831a02a946dSIlya Dryomov 1832a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->pool, e_inval); 1833a02a946dSIlya Dryomov 1834a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1835a02a946dSIlya Dryomov return 0; 1836a02a946dSIlya Dryomov 1837a02a946dSIlya Dryomov e_inval: 1838a02a946dSIlya Dryomov return -EINVAL; 1839a02a946dSIlya Dryomov } 1840a02a946dSIlya Dryomov 1841a02a946dSIlya Dryomov static int hoid_encoding_size(const struct ceph_hobject_id *hoid) 1842a02a946dSIlya Dryomov { 1843a02a946dSIlya Dryomov return 8 + 4 + 1 + 8 + /* snapid, hash, is_max, pool */ 1844a02a946dSIlya Dryomov 4 + hoid->key_len + 4 + hoid->oid_len + 4 + hoid->nspace_len; 1845a02a946dSIlya Dryomov } 1846a02a946dSIlya Dryomov 1847a02a946dSIlya Dryomov static void encode_hoid(void **p, void *end, const struct ceph_hobject_id *hoid) 1848a02a946dSIlya Dryomov { 1849a02a946dSIlya Dryomov ceph_start_encoding(p, 4, 3, hoid_encoding_size(hoid)); 1850a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->key, hoid->key_len); 1851a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->oid, hoid->oid_len); 1852a02a946dSIlya Dryomov ceph_encode_64(p, hoid->snapid); 1853a02a946dSIlya Dryomov ceph_encode_32(p, hoid->hash); 1854a02a946dSIlya Dryomov ceph_encode_8(p, hoid->is_max); 1855a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->nspace, hoid->nspace_len); 1856a02a946dSIlya Dryomov ceph_encode_64(p, hoid->pool); 1857a02a946dSIlya Dryomov } 1858a02a946dSIlya Dryomov 1859a02a946dSIlya Dryomov static void free_hoid(struct ceph_hobject_id *hoid) 1860a02a946dSIlya Dryomov { 1861a02a946dSIlya Dryomov if (hoid) { 1862a02a946dSIlya Dryomov kfree(hoid->key); 1863a02a946dSIlya Dryomov kfree(hoid->oid); 1864a02a946dSIlya Dryomov kfree(hoid->nspace); 1865a02a946dSIlya Dryomov kfree(hoid); 1866a02a946dSIlya Dryomov } 1867a02a946dSIlya Dryomov } 1868a02a946dSIlya Dryomov 1869a02a946dSIlya Dryomov static struct ceph_osd_backoff *alloc_backoff(void) 1870a02a946dSIlya Dryomov { 1871a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1872a02a946dSIlya Dryomov 1873a02a946dSIlya Dryomov backoff = kzalloc(sizeof(*backoff), GFP_NOIO); 1874a02a946dSIlya Dryomov if (!backoff) 1875a02a946dSIlya Dryomov return NULL; 1876a02a946dSIlya Dryomov 1877a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->spg_node); 1878a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->id_node); 1879a02a946dSIlya Dryomov return backoff; 1880a02a946dSIlya Dryomov } 1881a02a946dSIlya Dryomov 1882a02a946dSIlya Dryomov static void free_backoff(struct ceph_osd_backoff *backoff) 1883a02a946dSIlya Dryomov { 1884a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->spg_node)); 1885a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->id_node)); 1886a02a946dSIlya Dryomov 1887a02a946dSIlya Dryomov free_hoid(backoff->begin); 1888a02a946dSIlya Dryomov free_hoid(backoff->end); 1889a02a946dSIlya Dryomov kfree(backoff); 1890a02a946dSIlya Dryomov } 1891a02a946dSIlya Dryomov 1892a02a946dSIlya Dryomov /* 1893a02a946dSIlya Dryomov * Within a specific spgid, backoffs are managed by ->begin hoid. 1894a02a946dSIlya Dryomov */ 1895a02a946dSIlya Dryomov DEFINE_RB_INSDEL_FUNCS2(backoff, struct ceph_osd_backoff, begin, hoid_compare, 1896a02a946dSIlya Dryomov RB_BYVAL, spg_node); 1897a02a946dSIlya Dryomov 1898a02a946dSIlya Dryomov static struct ceph_osd_backoff *lookup_containing_backoff(struct rb_root *root, 1899a02a946dSIlya Dryomov const struct ceph_hobject_id *hoid) 1900a02a946dSIlya Dryomov { 1901a02a946dSIlya Dryomov struct rb_node *n = root->rb_node; 1902a02a946dSIlya Dryomov 1903a02a946dSIlya Dryomov while (n) { 1904a02a946dSIlya Dryomov struct ceph_osd_backoff *cur = 1905a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_backoff, spg_node); 1906a02a946dSIlya Dryomov int cmp; 1907a02a946dSIlya Dryomov 1908a02a946dSIlya Dryomov cmp = hoid_compare(hoid, cur->begin); 1909a02a946dSIlya Dryomov if (cmp < 0) { 1910a02a946dSIlya Dryomov n = n->rb_left; 1911a02a946dSIlya Dryomov } else if (cmp > 0) { 1912a02a946dSIlya Dryomov if (hoid_compare(hoid, cur->end) < 0) 1913a02a946dSIlya Dryomov return cur; 1914a02a946dSIlya Dryomov 1915a02a946dSIlya Dryomov n = n->rb_right; 1916a02a946dSIlya Dryomov } else { 1917a02a946dSIlya Dryomov return cur; 1918a02a946dSIlya Dryomov } 1919a02a946dSIlya Dryomov } 1920a02a946dSIlya Dryomov 1921a02a946dSIlya Dryomov return NULL; 1922a02a946dSIlya Dryomov } 1923a02a946dSIlya Dryomov 1924a02a946dSIlya Dryomov /* 1925a02a946dSIlya Dryomov * Each backoff has a unique id within its OSD session. 1926a02a946dSIlya Dryomov */ 1927a02a946dSIlya Dryomov DEFINE_RB_FUNCS(backoff_by_id, struct ceph_osd_backoff, id, id_node) 1928a02a946dSIlya Dryomov 1929a02a946dSIlya Dryomov static void clear_backoffs(struct ceph_osd *osd) 1930a02a946dSIlya Dryomov { 1931a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&osd->o_backoff_mappings)) { 1932a02a946dSIlya Dryomov struct ceph_spg_mapping *spg = 1933a02a946dSIlya Dryomov rb_entry(rb_first(&osd->o_backoff_mappings), 1934a02a946dSIlya Dryomov struct ceph_spg_mapping, node); 1935a02a946dSIlya Dryomov 1936a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&spg->backoffs)) { 1937a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff = 1938a02a946dSIlya Dryomov rb_entry(rb_first(&spg->backoffs), 1939a02a946dSIlya Dryomov struct ceph_osd_backoff, spg_node); 1940a02a946dSIlya Dryomov 1941a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 1942a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 1943a02a946dSIlya Dryomov free_backoff(backoff); 1944a02a946dSIlya Dryomov } 1945a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 1946a02a946dSIlya Dryomov free_spg_mapping(spg); 1947a02a946dSIlya Dryomov } 1948a02a946dSIlya Dryomov } 1949a02a946dSIlya Dryomov 1950a02a946dSIlya Dryomov /* 1951a02a946dSIlya Dryomov * Set up a temporary, non-owning view into @t. 1952a02a946dSIlya Dryomov */ 1953a02a946dSIlya Dryomov static void hoid_fill_from_target(struct ceph_hobject_id *hoid, 1954a02a946dSIlya Dryomov const struct ceph_osd_request_target *t) 1955a02a946dSIlya Dryomov { 1956a02a946dSIlya Dryomov hoid->key = NULL; 1957a02a946dSIlya Dryomov hoid->key_len = 0; 1958a02a946dSIlya Dryomov hoid->oid = t->target_oid.name; 1959a02a946dSIlya Dryomov hoid->oid_len = t->target_oid.name_len; 1960a02a946dSIlya Dryomov hoid->snapid = CEPH_NOSNAP; 1961a02a946dSIlya Dryomov hoid->hash = t->pgid.seed; 1962a02a946dSIlya Dryomov hoid->is_max = false; 1963a02a946dSIlya Dryomov if (t->target_oloc.pool_ns) { 1964a02a946dSIlya Dryomov hoid->nspace = t->target_oloc.pool_ns->str; 1965a02a946dSIlya Dryomov hoid->nspace_len = t->target_oloc.pool_ns->len; 1966a02a946dSIlya Dryomov } else { 1967a02a946dSIlya Dryomov hoid->nspace = NULL; 1968a02a946dSIlya Dryomov hoid->nspace_len = 0; 1969a02a946dSIlya Dryomov } 1970a02a946dSIlya Dryomov hoid->pool = t->target_oloc.pool; 1971a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1972a02a946dSIlya Dryomov } 1973a02a946dSIlya Dryomov 1974a02a946dSIlya Dryomov static bool should_plug_request(struct ceph_osd_request *req) 1975a02a946dSIlya Dryomov { 1976a02a946dSIlya Dryomov struct ceph_osd *osd = req->r_osd; 1977a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1978a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1979a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 1980a02a946dSIlya Dryomov 1981a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &req->r_t.spgid); 1982a02a946dSIlya Dryomov if (!spg) 1983a02a946dSIlya Dryomov return false; 1984a02a946dSIlya Dryomov 1985a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, &req->r_t); 1986a02a946dSIlya Dryomov backoff = lookup_containing_backoff(&spg->backoffs, &hoid); 1987a02a946dSIlya Dryomov if (!backoff) 1988a02a946dSIlya Dryomov return false; 1989a02a946dSIlya Dryomov 1990a02a946dSIlya Dryomov dout("%s req %p tid %llu backoff osd%d spgid %llu.%xs%d id %llu\n", 1991a02a946dSIlya Dryomov __func__, req, req->r_tid, osd->o_osd, backoff->spgid.pgid.pool, 1992a02a946dSIlya Dryomov backoff->spgid.pgid.seed, backoff->spgid.shard, backoff->id); 1993a02a946dSIlya Dryomov return true; 1994a02a946dSIlya Dryomov } 1995a02a946dSIlya Dryomov 19960d9c1ab3SIlya Dryomov /* 19970d9c1ab3SIlya Dryomov * Keep get_num_data_items() in sync with this function. 19980d9c1ab3SIlya Dryomov */ 199998c4bfe9SIlya Dryomov static void setup_request_data(struct ceph_osd_request *req) 20003d14c5d2SYehuda Sadeh { 200198c4bfe9SIlya Dryomov struct ceph_msg *request_msg = req->r_request; 200298c4bfe9SIlya Dryomov struct ceph_msg *reply_msg = req->r_reply; 200398c4bfe9SIlya Dryomov struct ceph_osd_req_op *op; 20043d14c5d2SYehuda Sadeh 200598c4bfe9SIlya Dryomov if (req->r_request->num_data_items || req->r_reply->num_data_items) 2006bb873b53SIlya Dryomov return; 20073d14c5d2SYehuda Sadeh 200898c4bfe9SIlya Dryomov WARN_ON(request_msg->data_length || reply_msg->data_length); 200998c4bfe9SIlya Dryomov for (op = req->r_ops; op != &req->r_ops[req->r_num_ops]; op++) { 2010bb873b53SIlya Dryomov switch (op->op) { 2011bb873b53SIlya Dryomov /* request */ 2012bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITE: 2013bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 2014bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->extent.length); 201598c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 201698c4bfe9SIlya Dryomov &op->extent.osd_data); 2017bb873b53SIlya Dryomov break; 2018bb873b53SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 2019bb873b53SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 2020bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->xattr.name_len + 2021bb873b53SIlya Dryomov op->xattr.value_len); 202298c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 202398c4bfe9SIlya Dryomov &op->xattr.osd_data); 2024bb873b53SIlya Dryomov break; 2025922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 202698c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 2027922dab61SIlya Dryomov &op->notify_ack.request_data); 2028922dab61SIlya Dryomov break; 202978beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 203023ddf9beSLuis Henriques ceph_osdc_msg_data_add(request_msg, 203123ddf9beSLuis Henriques &op->copy_from.osd_data); 203223ddf9beSLuis Henriques break; 2033bb873b53SIlya Dryomov 2034bb873b53SIlya Dryomov /* reply */ 2035bb873b53SIlya Dryomov case CEPH_OSD_OP_STAT: 203698c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2037bb873b53SIlya Dryomov &op->raw_data_in); 2038bb873b53SIlya Dryomov break; 2039bb873b53SIlya Dryomov case CEPH_OSD_OP_READ: 204098c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2041bb873b53SIlya Dryomov &op->extent.osd_data); 2042bb873b53SIlya Dryomov break; 2043a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 204498c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2045a4ed38d7SDouglas Fuller &op->list_watchers.response_data); 2046a4ed38d7SDouglas Fuller break; 2047bb873b53SIlya Dryomov 2048bb873b53SIlya Dryomov /* both */ 2049bb873b53SIlya Dryomov case CEPH_OSD_OP_CALL: 2050bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->cls.class_len + 2051bb873b53SIlya Dryomov op->cls.method_len + 2052bb873b53SIlya Dryomov op->cls.indata_len); 205398c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 205498c4bfe9SIlya Dryomov &op->cls.request_info); 2055bb873b53SIlya Dryomov /* optional, can be NONE */ 205698c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 205798c4bfe9SIlya Dryomov &op->cls.request_data); 2058bb873b53SIlya Dryomov /* optional, can be NONE */ 205998c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2060bb873b53SIlya Dryomov &op->cls.response_data); 2061bb873b53SIlya Dryomov break; 206219079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 206398c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 206419079203SIlya Dryomov &op->notify.request_data); 206598c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 206619079203SIlya Dryomov &op->notify.response_data); 206719079203SIlya Dryomov break; 2068bb873b53SIlya Dryomov } 2069bb873b53SIlya Dryomov } 2070bb873b53SIlya Dryomov } 2071bb873b53SIlya Dryomov 20722e59ffd1SIlya Dryomov static void encode_pgid(void **p, const struct ceph_pg *pgid) 20732e59ffd1SIlya Dryomov { 20742e59ffd1SIlya Dryomov ceph_encode_8(p, 1); 20752e59ffd1SIlya Dryomov ceph_encode_64(p, pgid->pool); 20762e59ffd1SIlya Dryomov ceph_encode_32(p, pgid->seed); 20772e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 20782e59ffd1SIlya Dryomov } 20792e59ffd1SIlya Dryomov 20808cb441c0SIlya Dryomov static void encode_spgid(void **p, const struct ceph_spg *spgid) 20818cb441c0SIlya Dryomov { 20828cb441c0SIlya Dryomov ceph_start_encoding(p, 1, 1, CEPH_PGID_ENCODING_LEN + 1); 20838cb441c0SIlya Dryomov encode_pgid(p, &spgid->pgid); 20848cb441c0SIlya Dryomov ceph_encode_8(p, spgid->shard); 20858cb441c0SIlya Dryomov } 20868cb441c0SIlya Dryomov 20872e59ffd1SIlya Dryomov static void encode_oloc(void **p, void *end, 20882e59ffd1SIlya Dryomov const struct ceph_object_locator *oloc) 20892e59ffd1SIlya Dryomov { 20902e59ffd1SIlya Dryomov ceph_start_encoding(p, 5, 4, ceph_oloc_encoding_size(oloc)); 20912e59ffd1SIlya Dryomov ceph_encode_64(p, oloc->pool); 20922e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 20932e59ffd1SIlya Dryomov ceph_encode_32(p, 0); /* key len */ 20942e59ffd1SIlya Dryomov if (oloc->pool_ns) 20952e59ffd1SIlya Dryomov ceph_encode_string(p, end, oloc->pool_ns->str, 20962e59ffd1SIlya Dryomov oloc->pool_ns->len); 20972e59ffd1SIlya Dryomov else 20982e59ffd1SIlya Dryomov ceph_encode_32(p, 0); 20992e59ffd1SIlya Dryomov } 21002e59ffd1SIlya Dryomov 21018cb441c0SIlya Dryomov static void encode_request_partial(struct ceph_osd_request *req, 21028cb441c0SIlya Dryomov struct ceph_msg *msg) 2103bb873b53SIlya Dryomov { 2104bb873b53SIlya Dryomov void *p = msg->front.iov_base; 2105bb873b53SIlya Dryomov void *const end = p + msg->front_alloc_len; 2106bb873b53SIlya Dryomov u32 data_len = 0; 2107bb873b53SIlya Dryomov int i; 2108bb873b53SIlya Dryomov 2109bb873b53SIlya Dryomov if (req->r_flags & CEPH_OSD_FLAG_WRITE) { 2110bb873b53SIlya Dryomov /* snapshots aren't writeable */ 2111bb873b53SIlya Dryomov WARN_ON(req->r_snapid != CEPH_NOSNAP); 2112bb873b53SIlya Dryomov } else { 2113bb873b53SIlya Dryomov WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec || 2114bb873b53SIlya Dryomov req->r_data_offset || req->r_snapc); 2115bb873b53SIlya Dryomov } 2116bb873b53SIlya Dryomov 211798c4bfe9SIlya Dryomov setup_request_data(req); 2118bb873b53SIlya Dryomov 21198cb441c0SIlya Dryomov encode_spgid(&p, &req->r_t.spgid); /* actual spg */ 21208cb441c0SIlya Dryomov ceph_encode_32(&p, req->r_t.pgid.seed); /* raw hash */ 2121bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_osdc->osdmap->epoch); 2122bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_flags); 21238cb441c0SIlya Dryomov 21248cb441c0SIlya Dryomov /* reqid */ 21258cb441c0SIlya Dryomov ceph_start_encoding(&p, 2, 2, sizeof(struct ceph_osd_reqid)); 21268cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_osd_reqid)); 21278cb441c0SIlya Dryomov p += sizeof(struct ceph_osd_reqid); 21288cb441c0SIlya Dryomov 21298cb441c0SIlya Dryomov /* trace */ 21308cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_blkin_trace_info)); 21318cb441c0SIlya Dryomov p += sizeof(struct ceph_blkin_trace_info); 21328cb441c0SIlya Dryomov 21338cb441c0SIlya Dryomov ceph_encode_32(&p, 0); /* client_inc, always 0 */ 2134fac02ddfSArnd Bergmann ceph_encode_timespec64(p, &req->r_mtime); 2135bb873b53SIlya Dryomov p += sizeof(struct ceph_timespec); 2136aa26d662SJeff Layton 21372e59ffd1SIlya Dryomov encode_oloc(&p, end, &req->r_t.target_oloc); 21382e59ffd1SIlya Dryomov ceph_encode_string(&p, end, req->r_t.target_oid.name, 21392e59ffd1SIlya Dryomov req->r_t.target_oid.name_len); 2140bb873b53SIlya Dryomov 2141bb873b53SIlya Dryomov /* ops, can imply data */ 2142bb873b53SIlya Dryomov ceph_encode_16(&p, req->r_num_ops); 2143bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 2144bb873b53SIlya Dryomov data_len += osd_req_encode_op(p, &req->r_ops[i]); 2145bb873b53SIlya Dryomov p += sizeof(struct ceph_osd_op); 2146bb873b53SIlya Dryomov } 2147bb873b53SIlya Dryomov 2148bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapid); /* snapid */ 2149bb873b53SIlya Dryomov if (req->r_snapc) { 2150bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->seq); 2151bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_snapc->num_snaps); 2152bb873b53SIlya Dryomov for (i = 0; i < req->r_snapc->num_snaps; i++) 2153bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->snaps[i]); 2154bb873b53SIlya Dryomov } else { 2155bb873b53SIlya Dryomov ceph_encode_64(&p, 0); /* snap_seq */ 2156bb873b53SIlya Dryomov ceph_encode_32(&p, 0); /* snaps len */ 2157bb873b53SIlya Dryomov } 2158bb873b53SIlya Dryomov 2159bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_attempts); /* retry_attempt */ 2160986e8989SIlya Dryomov BUG_ON(p > end - 8); /* space for features */ 2161bb873b53SIlya Dryomov 21628cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(8); /* MOSDOp v8 */ 21638cb441c0SIlya Dryomov /* front_len is finalized in encode_request_finish() */ 2164986e8989SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 2165986e8989SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 2166bb873b53SIlya Dryomov msg->hdr.data_len = cpu_to_le32(data_len); 2167bb873b53SIlya Dryomov /* 2168bb873b53SIlya Dryomov * The header "data_off" is a hint to the receiver allowing it 2169bb873b53SIlya Dryomov * to align received data into its buffers such that there's no 2170bb873b53SIlya Dryomov * need to re-copy it before writing it to disk (direct I/O). 2171bb873b53SIlya Dryomov */ 2172bb873b53SIlya Dryomov msg->hdr.data_off = cpu_to_le16(req->r_data_offset); 2173bb873b53SIlya Dryomov 21748cb441c0SIlya Dryomov dout("%s req %p msg %p oid %s oid_len %d\n", __func__, req, msg, 21758cb441c0SIlya Dryomov req->r_t.target_oid.name, req->r_t.target_oid.name_len); 21768cb441c0SIlya Dryomov } 21778cb441c0SIlya Dryomov 21788cb441c0SIlya Dryomov static void encode_request_finish(struct ceph_msg *msg) 21798cb441c0SIlya Dryomov { 21808cb441c0SIlya Dryomov void *p = msg->front.iov_base; 2181986e8989SIlya Dryomov void *const partial_end = p + msg->front.iov_len; 21828cb441c0SIlya Dryomov void *const end = p + msg->front_alloc_len; 21838cb441c0SIlya Dryomov 21848cb441c0SIlya Dryomov if (CEPH_HAVE_FEATURE(msg->con->peer_features, RESEND_ON_SPLIT)) { 21858cb441c0SIlya Dryomov /* luminous OSD -- encode features and be done */ 2186986e8989SIlya Dryomov p = partial_end; 21878cb441c0SIlya Dryomov ceph_encode_64(&p, msg->con->peer_features); 21888cb441c0SIlya Dryomov } else { 21898cb441c0SIlya Dryomov struct { 21908cb441c0SIlya Dryomov char spgid[CEPH_ENCODING_START_BLK_LEN + 21918cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1]; 21928cb441c0SIlya Dryomov __le32 hash; 21938cb441c0SIlya Dryomov __le32 epoch; 21948cb441c0SIlya Dryomov __le32 flags; 21958cb441c0SIlya Dryomov char reqid[CEPH_ENCODING_START_BLK_LEN + 21968cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid)]; 21978cb441c0SIlya Dryomov char trace[sizeof(struct ceph_blkin_trace_info)]; 21988cb441c0SIlya Dryomov __le32 client_inc; 21998cb441c0SIlya Dryomov struct ceph_timespec mtime; 22008cb441c0SIlya Dryomov } __packed head; 22018cb441c0SIlya Dryomov struct ceph_pg pgid; 22028cb441c0SIlya Dryomov void *oloc, *oid, *tail; 22038cb441c0SIlya Dryomov int oloc_len, oid_len, tail_len; 22048cb441c0SIlya Dryomov int len; 22058cb441c0SIlya Dryomov 22068cb441c0SIlya Dryomov /* 22078cb441c0SIlya Dryomov * Pre-luminous OSD -- reencode v8 into v4 using @head 22088cb441c0SIlya Dryomov * as a temporary buffer. Encode the raw PG; the rest 22098cb441c0SIlya Dryomov * is just a matter of moving oloc, oid and tail blobs 22108cb441c0SIlya Dryomov * around. 22118cb441c0SIlya Dryomov */ 22128cb441c0SIlya Dryomov memcpy(&head, p, sizeof(head)); 22138cb441c0SIlya Dryomov p += sizeof(head); 22148cb441c0SIlya Dryomov 22158cb441c0SIlya Dryomov oloc = p; 22168cb441c0SIlya Dryomov p += CEPH_ENCODING_START_BLK_LEN; 22178cb441c0SIlya Dryomov pgid.pool = ceph_decode_64(&p); 22188cb441c0SIlya Dryomov p += 4 + 4; /* preferred, key len */ 22198cb441c0SIlya Dryomov len = ceph_decode_32(&p); 22208cb441c0SIlya Dryomov p += len; /* nspace */ 22218cb441c0SIlya Dryomov oloc_len = p - oloc; 22228cb441c0SIlya Dryomov 22238cb441c0SIlya Dryomov oid = p; 22248cb441c0SIlya Dryomov len = ceph_decode_32(&p); 22258cb441c0SIlya Dryomov p += len; 22268cb441c0SIlya Dryomov oid_len = p - oid; 22278cb441c0SIlya Dryomov 22288cb441c0SIlya Dryomov tail = p; 2229986e8989SIlya Dryomov tail_len = partial_end - p; 22308cb441c0SIlya Dryomov 22318cb441c0SIlya Dryomov p = msg->front.iov_base; 22328cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.client_inc, sizeof(head.client_inc)); 22338cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.epoch, sizeof(head.epoch)); 22348cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.flags, sizeof(head.flags)); 22358cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.mtime, sizeof(head.mtime)); 22368cb441c0SIlya Dryomov 22378cb441c0SIlya Dryomov /* reassert_version */ 22388cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_eversion)); 22398cb441c0SIlya Dryomov p += sizeof(struct ceph_eversion); 22408cb441c0SIlya Dryomov 22418cb441c0SIlya Dryomov BUG_ON(p >= oloc); 22428cb441c0SIlya Dryomov memmove(p, oloc, oloc_len); 22438cb441c0SIlya Dryomov p += oloc_len; 22448cb441c0SIlya Dryomov 22458cb441c0SIlya Dryomov pgid.seed = le32_to_cpu(head.hash); 22468cb441c0SIlya Dryomov encode_pgid(&p, &pgid); /* raw pg */ 22478cb441c0SIlya Dryomov 22488cb441c0SIlya Dryomov BUG_ON(p >= oid); 22498cb441c0SIlya Dryomov memmove(p, oid, oid_len); 22508cb441c0SIlya Dryomov p += oid_len; 22518cb441c0SIlya Dryomov 22528cb441c0SIlya Dryomov /* tail -- ops, snapid, snapc, retry_attempt */ 22538cb441c0SIlya Dryomov BUG_ON(p >= tail); 22548cb441c0SIlya Dryomov memmove(p, tail, tail_len); 22558cb441c0SIlya Dryomov p += tail_len; 22568cb441c0SIlya Dryomov 22578cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */ 22588cb441c0SIlya Dryomov } 22598cb441c0SIlya Dryomov 22608cb441c0SIlya Dryomov BUG_ON(p > end); 22618cb441c0SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 22628cb441c0SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 22638cb441c0SIlya Dryomov 22648cb441c0SIlya Dryomov dout("%s msg %p tid %llu %u+%u+%u v%d\n", __func__, msg, 22658cb441c0SIlya Dryomov le64_to_cpu(msg->hdr.tid), le32_to_cpu(msg->hdr.front_len), 22668cb441c0SIlya Dryomov le32_to_cpu(msg->hdr.middle_len), le32_to_cpu(msg->hdr.data_len), 22678cb441c0SIlya Dryomov le16_to_cpu(msg->hdr.version)); 2268bb873b53SIlya Dryomov } 2269bb873b53SIlya Dryomov 2270bb873b53SIlya Dryomov /* 2271bb873b53SIlya Dryomov * @req has to be assigned a tid and registered. 2272bb873b53SIlya Dryomov */ 2273bb873b53SIlya Dryomov static void send_request(struct ceph_osd_request *req) 2274bb873b53SIlya Dryomov { 2275bb873b53SIlya Dryomov struct ceph_osd *osd = req->r_osd; 2276bb873b53SIlya Dryomov 22775aea3dcdSIlya Dryomov verify_osd_locked(osd); 2278bb873b53SIlya Dryomov WARN_ON(osd->o_osd != req->r_t.osd); 2279bb873b53SIlya Dryomov 2280a02a946dSIlya Dryomov /* backoff? */ 2281a02a946dSIlya Dryomov if (should_plug_request(req)) 2282a02a946dSIlya Dryomov return; 2283a02a946dSIlya Dryomov 22845aea3dcdSIlya Dryomov /* 22855aea3dcdSIlya Dryomov * We may have a previously queued request message hanging 22865aea3dcdSIlya Dryomov * around. Cancel it to avoid corrupting the msgr. 22875aea3dcdSIlya Dryomov */ 22885aea3dcdSIlya Dryomov if (req->r_sent) 22895aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 22905aea3dcdSIlya Dryomov 2291bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR; 2292bb873b53SIlya Dryomov if (req->r_attempts) 2293bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_RETRY; 2294bb873b53SIlya Dryomov else 2295bb873b53SIlya Dryomov WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY); 2296bb873b53SIlya Dryomov 22978cb441c0SIlya Dryomov encode_request_partial(req, req->r_request); 2298bb873b53SIlya Dryomov 229904c7d789SIlya Dryomov dout("%s req %p tid %llu to pgid %llu.%x spgid %llu.%xs%d osd%d e%u flags 0x%x attempt %d\n", 2300bb873b53SIlya Dryomov __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed, 2301dc98ff72SIlya Dryomov req->r_t.spgid.pgid.pool, req->r_t.spgid.pgid.seed, 230204c7d789SIlya Dryomov req->r_t.spgid.shard, osd->o_osd, req->r_t.epoch, req->r_flags, 230304c7d789SIlya Dryomov req->r_attempts); 2304bb873b53SIlya Dryomov 2305bb873b53SIlya Dryomov req->r_t.paused = false; 23063d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 2307bb873b53SIlya Dryomov req->r_attempts++; 23083d14c5d2SYehuda Sadeh 2309bb873b53SIlya Dryomov req->r_sent = osd->o_incarnation; 2310bb873b53SIlya Dryomov req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 2311bb873b53SIlya Dryomov ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request)); 23123d14c5d2SYehuda Sadeh } 23133d14c5d2SYehuda Sadeh 231442c1b124SIlya Dryomov static void maybe_request_map(struct ceph_osd_client *osdc) 231542c1b124SIlya Dryomov { 231642c1b124SIlya Dryomov bool continuous = false; 231742c1b124SIlya Dryomov 23185aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 231942c1b124SIlya Dryomov WARN_ON(!osdc->osdmap->epoch); 232042c1b124SIlya Dryomov 2321b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2322b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD) || 2323b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 232442c1b124SIlya Dryomov dout("%s osdc %p continuous\n", __func__, osdc); 232542c1b124SIlya Dryomov continuous = true; 232642c1b124SIlya Dryomov } else { 232742c1b124SIlya Dryomov dout("%s osdc %p onetime\n", __func__, osdc); 232842c1b124SIlya Dryomov } 232942c1b124SIlya Dryomov 233042c1b124SIlya Dryomov if (ceph_monc_want_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 233142c1b124SIlya Dryomov osdc->osdmap->epoch + 1, continuous)) 233242c1b124SIlya Dryomov ceph_monc_renew_subs(&osdc->client->monc); 233342c1b124SIlya Dryomov } 233442c1b124SIlya Dryomov 2335a1f4020aSJeff Layton static void complete_request(struct ceph_osd_request *req, int err); 23364609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req); 23374609245eSIlya Dryomov 23385aea3dcdSIlya Dryomov static void __submit_request(struct ceph_osd_request *req, bool wrlocked) 23390bbfdfe8SIlya Dryomov { 23405aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 23415aea3dcdSIlya Dryomov struct ceph_osd *osd; 23424609245eSIlya Dryomov enum calc_target_result ct_res; 234366850df5SIlya Dryomov int err = 0; 23445aea3dcdSIlya Dryomov bool need_send = false; 23455aea3dcdSIlya Dryomov bool promoted = false; 23460bbfdfe8SIlya Dryomov 2347b18b9550SIlya Dryomov WARN_ON(req->r_tid); 23485aea3dcdSIlya Dryomov dout("%s req %p wrlocked %d\n", __func__, req, wrlocked); 23495aea3dcdSIlya Dryomov 23505aea3dcdSIlya Dryomov again: 23518edf84baSIlya Dryomov ct_res = calc_target(osdc, &req->r_t, false); 23524609245eSIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE && !wrlocked) 23534609245eSIlya Dryomov goto promote; 23544609245eSIlya Dryomov 23555aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, wrlocked); 23565aea3dcdSIlya Dryomov if (IS_ERR(osd)) { 23575aea3dcdSIlya Dryomov WARN_ON(PTR_ERR(osd) != -EAGAIN || wrlocked); 23585aea3dcdSIlya Dryomov goto promote; 23595aea3dcdSIlya Dryomov } 23605aea3dcdSIlya Dryomov 236166850df5SIlya Dryomov if (osdc->abort_err) { 236266850df5SIlya Dryomov dout("req %p abort_err %d\n", req, osdc->abort_err); 236366850df5SIlya Dryomov err = osdc->abort_err; 236466850df5SIlya Dryomov } else if (osdc->osdmap->epoch < osdc->epoch_barrier) { 236558eb7932SJeff Layton dout("req %p epoch %u barrier %u\n", req, osdc->osdmap->epoch, 236658eb7932SJeff Layton osdc->epoch_barrier); 236758eb7932SJeff Layton req->r_t.paused = true; 236858eb7932SJeff Layton maybe_request_map(osdc); 236958eb7932SJeff Layton } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 2370b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 23715aea3dcdSIlya Dryomov dout("req %p pausewr\n", req); 23725aea3dcdSIlya Dryomov req->r_t.paused = true; 23735aea3dcdSIlya Dryomov maybe_request_map(osdc); 23745aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_READ) && 2375b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 23765aea3dcdSIlya Dryomov dout("req %p pauserd\n", req); 23775aea3dcdSIlya Dryomov req->r_t.paused = true; 23785aea3dcdSIlya Dryomov maybe_request_map(osdc); 23795aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 23805aea3dcdSIlya Dryomov !(req->r_flags & (CEPH_OSD_FLAG_FULL_TRY | 23815aea3dcdSIlya Dryomov CEPH_OSD_FLAG_FULL_FORCE)) && 2382b7ec35b3SIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 23835aea3dcdSIlya Dryomov pool_full(osdc, req->r_t.base_oloc.pool))) { 23845aea3dcdSIlya Dryomov dout("req %p full/pool_full\n", req); 238502b2f549SDongsheng Yang if (ceph_test_opt(osdc->client, ABORT_ON_FULL)) { 238629e87820SIlya Dryomov err = -ENOSPC; 238729e87820SIlya Dryomov } else { 23885aea3dcdSIlya Dryomov pr_warn_ratelimited("FULL or reached pool quota\n"); 23895aea3dcdSIlya Dryomov req->r_t.paused = true; 23905aea3dcdSIlya Dryomov maybe_request_map(osdc); 239129e87820SIlya Dryomov } 23925aea3dcdSIlya Dryomov } else if (!osd_homeless(osd)) { 23935aea3dcdSIlya Dryomov need_send = true; 23940bbfdfe8SIlya Dryomov } else { 23955aea3dcdSIlya Dryomov maybe_request_map(osdc); 23960bbfdfe8SIlya Dryomov } 23970bbfdfe8SIlya Dryomov 23985aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 23995aea3dcdSIlya Dryomov /* 24005aea3dcdSIlya Dryomov * Assign the tid atomically with send_request() to protect 24015aea3dcdSIlya Dryomov * multiple writes to the same object from racing with each 24025aea3dcdSIlya Dryomov * other, resulting in out of order ops on the OSDs. 24035aea3dcdSIlya Dryomov */ 24045aea3dcdSIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 24055aea3dcdSIlya Dryomov link_request(osd, req); 24065aea3dcdSIlya Dryomov if (need_send) 24075aea3dcdSIlya Dryomov send_request(req); 240866850df5SIlya Dryomov else if (err) 240966850df5SIlya Dryomov complete_request(req, err); 24105aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 24115aea3dcdSIlya Dryomov 24126001567cSIlya Dryomov if (!err && ct_res == CALC_TARGET_POOL_DNE) 24134609245eSIlya Dryomov send_map_check(req); 24144609245eSIlya Dryomov 24155aea3dcdSIlya Dryomov if (promoted) 24165aea3dcdSIlya Dryomov downgrade_write(&osdc->lock); 24175aea3dcdSIlya Dryomov return; 24185aea3dcdSIlya Dryomov 24195aea3dcdSIlya Dryomov promote: 24205aea3dcdSIlya Dryomov up_read(&osdc->lock); 24215aea3dcdSIlya Dryomov down_write(&osdc->lock); 24225aea3dcdSIlya Dryomov wrlocked = true; 24235aea3dcdSIlya Dryomov promoted = true; 24245aea3dcdSIlya Dryomov goto again; 24250bbfdfe8SIlya Dryomov } 24260bbfdfe8SIlya Dryomov 24275aea3dcdSIlya Dryomov static void account_request(struct ceph_osd_request *req) 24285aea3dcdSIlya Dryomov { 242954ea0046SIlya Dryomov WARN_ON(req->r_flags & (CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK)); 2430b18b9550SIlya Dryomov WARN_ON(!(req->r_flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE))); 24315aea3dcdSIlya Dryomov 2432b18b9550SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_ONDISK; 243322d2cfdfSIlya Dryomov atomic_inc(&req->r_osdc->num_requests); 24347cc5e38fSIlya Dryomov 24357cc5e38fSIlya Dryomov req->r_start_stamp = jiffies; 243697e27aaaSXiubo Li req->r_start_latency = ktime_get(); 24375aea3dcdSIlya Dryomov } 24385aea3dcdSIlya Dryomov 24395aea3dcdSIlya Dryomov static void submit_request(struct ceph_osd_request *req, bool wrlocked) 24405aea3dcdSIlya Dryomov { 24415aea3dcdSIlya Dryomov ceph_osdc_get_request(req); 24425aea3dcdSIlya Dryomov account_request(req); 24435aea3dcdSIlya Dryomov __submit_request(req, wrlocked); 24445aea3dcdSIlya Dryomov } 24455aea3dcdSIlya Dryomov 244645ee2c1dSIlya Dryomov static void finish_request(struct ceph_osd_request *req) 24475aea3dcdSIlya Dryomov { 24485aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 24495aea3dcdSIlya Dryomov 24504609245eSIlya Dryomov WARN_ON(lookup_request_mc(&osdc->map_checks, req->r_tid)); 245104c7d789SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 245204c7d789SIlya Dryomov 245397e27aaaSXiubo Li req->r_end_latency = ktime_get(); 245497e27aaaSXiubo Li 245504c7d789SIlya Dryomov if (req->r_osd) 245604c7d789SIlya Dryomov unlink_request(req->r_osd, req); 24575aea3dcdSIlya Dryomov atomic_dec(&osdc->num_requests); 24585aea3dcdSIlya Dryomov 24595aea3dcdSIlya Dryomov /* 24605aea3dcdSIlya Dryomov * If an OSD has failed or returned and a request has been sent 24615aea3dcdSIlya Dryomov * twice, it's possible to get a reply and end up here while the 24625aea3dcdSIlya Dryomov * request message is queued for delivery. We will ignore the 24635aea3dcdSIlya Dryomov * reply, so not a big deal, but better to try and catch it. 24645aea3dcdSIlya Dryomov */ 24655aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 24665aea3dcdSIlya Dryomov ceph_msg_revoke_incoming(req->r_reply); 24675aea3dcdSIlya Dryomov } 24685aea3dcdSIlya Dryomov 2469fe5da05eSIlya Dryomov static void __complete_request(struct ceph_osd_request *req) 2470fe5da05eSIlya Dryomov { 2471d75f773cSSakari Ailus dout("%s req %p tid %llu cb %ps result %d\n", __func__, req, 2472b18b9550SIlya Dryomov req->r_tid, req->r_callback, req->r_result); 247326df726bSIlya Dryomov 247426df726bSIlya Dryomov if (req->r_callback) 2475fe5da05eSIlya Dryomov req->r_callback(req); 247626df726bSIlya Dryomov complete_all(&req->r_completion); 247726df726bSIlya Dryomov ceph_osdc_put_request(req); 2478b18b9550SIlya Dryomov } 2479fe5da05eSIlya Dryomov 248088bc1922SIlya Dryomov static void complete_request_workfn(struct work_struct *work) 248188bc1922SIlya Dryomov { 248288bc1922SIlya Dryomov struct ceph_osd_request *req = 248388bc1922SIlya Dryomov container_of(work, struct ceph_osd_request, r_complete_work); 248488bc1922SIlya Dryomov 248588bc1922SIlya Dryomov __complete_request(req); 24860bbfdfe8SIlya Dryomov } 24870bbfdfe8SIlya Dryomov 24884609245eSIlya Dryomov /* 2489b18b9550SIlya Dryomov * This is open-coded in handle_reply(). 24904609245eSIlya Dryomov */ 24914609245eSIlya Dryomov static void complete_request(struct ceph_osd_request *req, int err) 24924609245eSIlya Dryomov { 24934609245eSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 24944609245eSIlya Dryomov 24954609245eSIlya Dryomov req->r_result = err; 249645ee2c1dSIlya Dryomov finish_request(req); 249788bc1922SIlya Dryomov 249888bc1922SIlya Dryomov INIT_WORK(&req->r_complete_work, complete_request_workfn); 249988bc1922SIlya Dryomov queue_work(req->r_osdc->completion_wq, &req->r_complete_work); 25004609245eSIlya Dryomov } 25014609245eSIlya Dryomov 25024609245eSIlya Dryomov static void cancel_map_check(struct ceph_osd_request *req) 25034609245eSIlya Dryomov { 25044609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 25054609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 25064609245eSIlya Dryomov 25074609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 25084609245eSIlya Dryomov 25094609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 25104609245eSIlya Dryomov if (!lookup_req) 25114609245eSIlya Dryomov return; 25124609245eSIlya Dryomov 25134609245eSIlya Dryomov WARN_ON(lookup_req != req); 25144609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 25154609245eSIlya Dryomov ceph_osdc_put_request(req); 25164609245eSIlya Dryomov } 25174609245eSIlya Dryomov 25185aea3dcdSIlya Dryomov static void cancel_request(struct ceph_osd_request *req) 25195aea3dcdSIlya Dryomov { 25205aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 25215aea3dcdSIlya Dryomov 25224609245eSIlya Dryomov cancel_map_check(req); 252345ee2c1dSIlya Dryomov finish_request(req); 2524b18b9550SIlya Dryomov complete_all(&req->r_completion); 2525c297eb42SIlya Dryomov ceph_osdc_put_request(req); 25265aea3dcdSIlya Dryomov } 25275aea3dcdSIlya Dryomov 25287cc5e38fSIlya Dryomov static void abort_request(struct ceph_osd_request *req, int err) 25297cc5e38fSIlya Dryomov { 25307cc5e38fSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 25317cc5e38fSIlya Dryomov 25327cc5e38fSIlya Dryomov cancel_map_check(req); 25337cc5e38fSIlya Dryomov complete_request(req, err); 25347cc5e38fSIlya Dryomov } 25357cc5e38fSIlya Dryomov 253666850df5SIlya Dryomov static int abort_fn(struct ceph_osd_request *req, void *arg) 253766850df5SIlya Dryomov { 253866850df5SIlya Dryomov int err = *(int *)arg; 253966850df5SIlya Dryomov 254066850df5SIlya Dryomov abort_request(req, err); 254166850df5SIlya Dryomov return 0; /* continue iteration */ 254266850df5SIlya Dryomov } 254366850df5SIlya Dryomov 254466850df5SIlya Dryomov /* 254566850df5SIlya Dryomov * Abort all in-flight requests with @err and arrange for all future 254666850df5SIlya Dryomov * requests to be failed immediately. 254766850df5SIlya Dryomov */ 254866850df5SIlya Dryomov void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err) 254966850df5SIlya Dryomov { 255066850df5SIlya Dryomov dout("%s osdc %p err %d\n", __func__, osdc, err); 255166850df5SIlya Dryomov down_write(&osdc->lock); 255266850df5SIlya Dryomov for_each_request(osdc, abort_fn, &err); 255366850df5SIlya Dryomov osdc->abort_err = err; 255466850df5SIlya Dryomov up_write(&osdc->lock); 255566850df5SIlya Dryomov } 255666850df5SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_abort_requests); 255766850df5SIlya Dryomov 25582cef0ba8SYan, Zheng void ceph_osdc_clear_abort_err(struct ceph_osd_client *osdc) 25592cef0ba8SYan, Zheng { 25602cef0ba8SYan, Zheng down_write(&osdc->lock); 25612cef0ba8SYan, Zheng osdc->abort_err = 0; 25622cef0ba8SYan, Zheng up_write(&osdc->lock); 25632cef0ba8SYan, Zheng } 25642cef0ba8SYan, Zheng EXPORT_SYMBOL(ceph_osdc_clear_abort_err); 25652cef0ba8SYan, Zheng 256658eb7932SJeff Layton static void update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 256758eb7932SJeff Layton { 256858eb7932SJeff Layton if (likely(eb > osdc->epoch_barrier)) { 256958eb7932SJeff Layton dout("updating epoch_barrier from %u to %u\n", 257058eb7932SJeff Layton osdc->epoch_barrier, eb); 257158eb7932SJeff Layton osdc->epoch_barrier = eb; 257258eb7932SJeff Layton /* Request map if we're not to the barrier yet */ 257358eb7932SJeff Layton if (eb > osdc->osdmap->epoch) 257458eb7932SJeff Layton maybe_request_map(osdc); 257558eb7932SJeff Layton } 257658eb7932SJeff Layton } 257758eb7932SJeff Layton 257858eb7932SJeff Layton void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 257958eb7932SJeff Layton { 258058eb7932SJeff Layton down_read(&osdc->lock); 258158eb7932SJeff Layton if (unlikely(eb > osdc->epoch_barrier)) { 258258eb7932SJeff Layton up_read(&osdc->lock); 258358eb7932SJeff Layton down_write(&osdc->lock); 258458eb7932SJeff Layton update_epoch_barrier(osdc, eb); 258558eb7932SJeff Layton up_write(&osdc->lock); 258658eb7932SJeff Layton } else { 258758eb7932SJeff Layton up_read(&osdc->lock); 258858eb7932SJeff Layton } 258958eb7932SJeff Layton } 259058eb7932SJeff Layton EXPORT_SYMBOL(ceph_osdc_update_epoch_barrier); 259158eb7932SJeff Layton 2592fc36d0a4SJeff Layton /* 25934eea0fefSIlya Dryomov * We can end up releasing caps as a result of abort_request(). 25944eea0fefSIlya Dryomov * In that case, we probably want to ensure that the cap release message 25954eea0fefSIlya Dryomov * has an updated epoch barrier in it, so set the epoch barrier prior to 25964eea0fefSIlya Dryomov * aborting the first request. 25974eea0fefSIlya Dryomov */ 25984eea0fefSIlya Dryomov static int abort_on_full_fn(struct ceph_osd_request *req, void *arg) 25994eea0fefSIlya Dryomov { 26004eea0fefSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 26014eea0fefSIlya Dryomov bool *victims = arg; 26024eea0fefSIlya Dryomov 2603c843d13cSIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 26044eea0fefSIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2605690f951dSIlya Dryomov pool_full(osdc, req->r_t.base_oloc.pool))) { 26064eea0fefSIlya Dryomov if (!*victims) { 26074eea0fefSIlya Dryomov update_epoch_barrier(osdc, osdc->osdmap->epoch); 26084eea0fefSIlya Dryomov *victims = true; 26094eea0fefSIlya Dryomov } 26104eea0fefSIlya Dryomov abort_request(req, -ENOSPC); 26114eea0fefSIlya Dryomov } 26124eea0fefSIlya Dryomov 26134eea0fefSIlya Dryomov return 0; /* continue iteration */ 26144eea0fefSIlya Dryomov } 26154eea0fefSIlya Dryomov 26164eea0fefSIlya Dryomov /* 2617fc36d0a4SJeff Layton * Drop all pending requests that are stalled waiting on a full condition to 261858eb7932SJeff Layton * clear, and complete them with ENOSPC as the return code. Set the 261958eb7932SJeff Layton * osdc->epoch_barrier to the latest map epoch that we've seen if any were 262058eb7932SJeff Layton * cancelled. 2621fc36d0a4SJeff Layton */ 2622fc36d0a4SJeff Layton static void ceph_osdc_abort_on_full(struct ceph_osd_client *osdc) 2623fc36d0a4SJeff Layton { 262458eb7932SJeff Layton bool victims = false; 2625fc36d0a4SJeff Layton 262602b2f549SDongsheng Yang if (ceph_test_opt(osdc->client, ABORT_ON_FULL) && 2627c843d13cSIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || have_pool_full(osdc))) 26284eea0fefSIlya Dryomov for_each_request(osdc, abort_on_full_fn, &victims); 2629fc36d0a4SJeff Layton } 2630fc36d0a4SJeff Layton 26314609245eSIlya Dryomov static void check_pool_dne(struct ceph_osd_request *req) 26324609245eSIlya Dryomov { 26334609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 26344609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 26354609245eSIlya Dryomov 26364609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 26374609245eSIlya Dryomov WARN_ON(!map->epoch); 26384609245eSIlya Dryomov 26394609245eSIlya Dryomov if (req->r_attempts) { 26404609245eSIlya Dryomov /* 26414609245eSIlya Dryomov * We sent a request earlier, which means that 26424609245eSIlya Dryomov * previously the pool existed, and now it does not 26434609245eSIlya Dryomov * (i.e., it was deleted). 26444609245eSIlya Dryomov */ 26454609245eSIlya Dryomov req->r_map_dne_bound = map->epoch; 26464609245eSIlya Dryomov dout("%s req %p tid %llu pool disappeared\n", __func__, req, 26474609245eSIlya Dryomov req->r_tid); 26484609245eSIlya Dryomov } else { 26494609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u have %u\n", __func__, 26504609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, map->epoch); 26514609245eSIlya Dryomov } 26524609245eSIlya Dryomov 26534609245eSIlya Dryomov if (req->r_map_dne_bound) { 26544609245eSIlya Dryomov if (map->epoch >= req->r_map_dne_bound) { 26554609245eSIlya Dryomov /* we had a new enough map */ 26564609245eSIlya Dryomov pr_info_ratelimited("tid %llu pool does not exist\n", 26574609245eSIlya Dryomov req->r_tid); 26584609245eSIlya Dryomov complete_request(req, -ENOENT); 26594609245eSIlya Dryomov } 26604609245eSIlya Dryomov } else { 26614609245eSIlya Dryomov send_map_check(req); 26624609245eSIlya Dryomov } 26634609245eSIlya Dryomov } 26644609245eSIlya Dryomov 26654609245eSIlya Dryomov static void map_check_cb(struct ceph_mon_generic_request *greq) 26664609245eSIlya Dryomov { 26674609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 26684609245eSIlya Dryomov struct ceph_osd_request *req; 26694609245eSIlya Dryomov u64 tid = greq->private_data; 26704609245eSIlya Dryomov 26714609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 26724609245eSIlya Dryomov 26734609245eSIlya Dryomov down_write(&osdc->lock); 26744609245eSIlya Dryomov req = lookup_request_mc(&osdc->map_checks, tid); 26754609245eSIlya Dryomov if (!req) { 26764609245eSIlya Dryomov dout("%s tid %llu dne\n", __func__, tid); 26774609245eSIlya Dryomov goto out_unlock; 26784609245eSIlya Dryomov } 26794609245eSIlya Dryomov 26804609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u newest %llu\n", __func__, 26814609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, greq->u.newest); 26824609245eSIlya Dryomov if (!req->r_map_dne_bound) 26834609245eSIlya Dryomov req->r_map_dne_bound = greq->u.newest; 26844609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 26854609245eSIlya Dryomov check_pool_dne(req); 26864609245eSIlya Dryomov 26874609245eSIlya Dryomov ceph_osdc_put_request(req); 26884609245eSIlya Dryomov out_unlock: 26894609245eSIlya Dryomov up_write(&osdc->lock); 26904609245eSIlya Dryomov } 26914609245eSIlya Dryomov 26924609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req) 26934609245eSIlya Dryomov { 26944609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 26954609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 26964609245eSIlya Dryomov int ret; 26974609245eSIlya Dryomov 26984609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 26994609245eSIlya Dryomov 27004609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 27014609245eSIlya Dryomov if (lookup_req) { 27024609245eSIlya Dryomov WARN_ON(lookup_req != req); 27034609245eSIlya Dryomov return; 27044609245eSIlya Dryomov } 27054609245eSIlya Dryomov 27064609245eSIlya Dryomov ceph_osdc_get_request(req); 27074609245eSIlya Dryomov insert_request_mc(&osdc->map_checks, req); 27084609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 27094609245eSIlya Dryomov map_check_cb, req->r_tid); 27104609245eSIlya Dryomov WARN_ON(ret); 27114609245eSIlya Dryomov } 27124609245eSIlya Dryomov 27130bbfdfe8SIlya Dryomov /* 2714922dab61SIlya Dryomov * lingering requests, watch/notify v2 infrastructure 2715922dab61SIlya Dryomov */ 2716922dab61SIlya Dryomov static void linger_release(struct kref *kref) 2717922dab61SIlya Dryomov { 2718922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 2719922dab61SIlya Dryomov container_of(kref, struct ceph_osd_linger_request, kref); 2720922dab61SIlya Dryomov 2721922dab61SIlya Dryomov dout("%s lreq %p reg_req %p ping_req %p\n", __func__, lreq, 2722922dab61SIlya Dryomov lreq->reg_req, lreq->ping_req); 2723922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->node)); 2724922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->osdc_node)); 27254609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->mc_node)); 2726922dab61SIlya Dryomov WARN_ON(!list_empty(&lreq->scan_item)); 2727b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lreq->pending_lworks)); 2728922dab61SIlya Dryomov WARN_ON(lreq->osd); 2729922dab61SIlya Dryomov 2730922dab61SIlya Dryomov if (lreq->reg_req) 2731922dab61SIlya Dryomov ceph_osdc_put_request(lreq->reg_req); 2732922dab61SIlya Dryomov if (lreq->ping_req) 2733922dab61SIlya Dryomov ceph_osdc_put_request(lreq->ping_req); 2734922dab61SIlya Dryomov target_destroy(&lreq->t); 2735922dab61SIlya Dryomov kfree(lreq); 2736922dab61SIlya Dryomov } 2737922dab61SIlya Dryomov 2738922dab61SIlya Dryomov static void linger_put(struct ceph_osd_linger_request *lreq) 2739922dab61SIlya Dryomov { 2740922dab61SIlya Dryomov if (lreq) 2741922dab61SIlya Dryomov kref_put(&lreq->kref, linger_release); 2742922dab61SIlya Dryomov } 2743922dab61SIlya Dryomov 2744922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2745922dab61SIlya Dryomov linger_get(struct ceph_osd_linger_request *lreq) 2746922dab61SIlya Dryomov { 2747922dab61SIlya Dryomov kref_get(&lreq->kref); 2748922dab61SIlya Dryomov return lreq; 2749922dab61SIlya Dryomov } 2750922dab61SIlya Dryomov 2751922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2752922dab61SIlya Dryomov linger_alloc(struct ceph_osd_client *osdc) 2753922dab61SIlya Dryomov { 2754922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2755922dab61SIlya Dryomov 2756922dab61SIlya Dryomov lreq = kzalloc(sizeof(*lreq), GFP_NOIO); 2757922dab61SIlya Dryomov if (!lreq) 2758922dab61SIlya Dryomov return NULL; 2759922dab61SIlya Dryomov 2760922dab61SIlya Dryomov kref_init(&lreq->kref); 2761922dab61SIlya Dryomov mutex_init(&lreq->lock); 2762922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->node); 2763922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->osdc_node); 27644609245eSIlya Dryomov RB_CLEAR_NODE(&lreq->mc_node); 2765922dab61SIlya Dryomov INIT_LIST_HEAD(&lreq->scan_item); 2766b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lreq->pending_lworks); 2767922dab61SIlya Dryomov init_completion(&lreq->reg_commit_wait); 276819079203SIlya Dryomov init_completion(&lreq->notify_finish_wait); 2769922dab61SIlya Dryomov 2770922dab61SIlya Dryomov lreq->osdc = osdc; 2771922dab61SIlya Dryomov target_init(&lreq->t); 2772922dab61SIlya Dryomov 2773922dab61SIlya Dryomov dout("%s lreq %p\n", __func__, lreq); 2774922dab61SIlya Dryomov return lreq; 2775922dab61SIlya Dryomov } 2776922dab61SIlya Dryomov 2777922dab61SIlya Dryomov DEFINE_RB_INSDEL_FUNCS(linger, struct ceph_osd_linger_request, linger_id, node) 2778922dab61SIlya Dryomov DEFINE_RB_FUNCS(linger_osdc, struct ceph_osd_linger_request, linger_id, osdc_node) 27794609245eSIlya Dryomov DEFINE_RB_FUNCS(linger_mc, struct ceph_osd_linger_request, linger_id, mc_node) 2780922dab61SIlya Dryomov 2781922dab61SIlya Dryomov /* 2782922dab61SIlya Dryomov * Create linger request <-> OSD session relation. 2783922dab61SIlya Dryomov * 2784922dab61SIlya Dryomov * @lreq has to be registered, @osd may be homeless. 2785922dab61SIlya Dryomov */ 2786922dab61SIlya Dryomov static void link_linger(struct ceph_osd *osd, 2787922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2788922dab61SIlya Dryomov { 2789922dab61SIlya Dryomov verify_osd_locked(osd); 2790922dab61SIlya Dryomov WARN_ON(!lreq->linger_id || lreq->osd); 2791922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2792922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2793922dab61SIlya Dryomov 2794922dab61SIlya Dryomov if (!osd_homeless(osd)) 2795922dab61SIlya Dryomov __remove_osd_from_lru(osd); 2796922dab61SIlya Dryomov else 2797922dab61SIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 2798922dab61SIlya Dryomov 2799922dab61SIlya Dryomov get_osd(osd); 2800922dab61SIlya Dryomov insert_linger(&osd->o_linger_requests, lreq); 2801922dab61SIlya Dryomov lreq->osd = osd; 2802922dab61SIlya Dryomov } 2803922dab61SIlya Dryomov 2804922dab61SIlya Dryomov static void unlink_linger(struct ceph_osd *osd, 2805922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2806922dab61SIlya Dryomov { 2807922dab61SIlya Dryomov verify_osd_locked(osd); 2808922dab61SIlya Dryomov WARN_ON(lreq->osd != osd); 2809922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2810922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2811922dab61SIlya Dryomov 2812922dab61SIlya Dryomov lreq->osd = NULL; 2813922dab61SIlya Dryomov erase_linger(&osd->o_linger_requests, lreq); 2814922dab61SIlya Dryomov put_osd(osd); 2815922dab61SIlya Dryomov 2816922dab61SIlya Dryomov if (!osd_homeless(osd)) 2817922dab61SIlya Dryomov maybe_move_osd_to_lru(osd); 2818922dab61SIlya Dryomov else 2819922dab61SIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 2820922dab61SIlya Dryomov } 2821922dab61SIlya Dryomov 2822922dab61SIlya Dryomov static bool __linger_registered(struct ceph_osd_linger_request *lreq) 2823922dab61SIlya Dryomov { 2824922dab61SIlya Dryomov verify_osdc_locked(lreq->osdc); 2825922dab61SIlya Dryomov 2826922dab61SIlya Dryomov return !RB_EMPTY_NODE(&lreq->osdc_node); 2827922dab61SIlya Dryomov } 2828922dab61SIlya Dryomov 2829922dab61SIlya Dryomov static bool linger_registered(struct ceph_osd_linger_request *lreq) 2830922dab61SIlya Dryomov { 2831922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2832922dab61SIlya Dryomov bool registered; 2833922dab61SIlya Dryomov 2834922dab61SIlya Dryomov down_read(&osdc->lock); 2835922dab61SIlya Dryomov registered = __linger_registered(lreq); 2836922dab61SIlya Dryomov up_read(&osdc->lock); 2837922dab61SIlya Dryomov 2838922dab61SIlya Dryomov return registered; 2839922dab61SIlya Dryomov } 2840922dab61SIlya Dryomov 2841922dab61SIlya Dryomov static void linger_register(struct ceph_osd_linger_request *lreq) 2842922dab61SIlya Dryomov { 2843922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2844922dab61SIlya Dryomov 2845922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2846922dab61SIlya Dryomov WARN_ON(lreq->linger_id); 2847922dab61SIlya Dryomov 2848922dab61SIlya Dryomov linger_get(lreq); 2849922dab61SIlya Dryomov lreq->linger_id = ++osdc->last_linger_id; 2850922dab61SIlya Dryomov insert_linger_osdc(&osdc->linger_requests, lreq); 2851922dab61SIlya Dryomov } 2852922dab61SIlya Dryomov 2853922dab61SIlya Dryomov static void linger_unregister(struct ceph_osd_linger_request *lreq) 2854922dab61SIlya Dryomov { 2855922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2856922dab61SIlya Dryomov 2857922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2858922dab61SIlya Dryomov 2859922dab61SIlya Dryomov erase_linger_osdc(&osdc->linger_requests, lreq); 2860922dab61SIlya Dryomov linger_put(lreq); 2861922dab61SIlya Dryomov } 2862922dab61SIlya Dryomov 2863922dab61SIlya Dryomov static void cancel_linger_request(struct ceph_osd_request *req) 2864922dab61SIlya Dryomov { 2865922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2866922dab61SIlya Dryomov 2867922dab61SIlya Dryomov WARN_ON(!req->r_linger); 2868922dab61SIlya Dryomov cancel_request(req); 2869922dab61SIlya Dryomov linger_put(lreq); 2870922dab61SIlya Dryomov } 2871922dab61SIlya Dryomov 2872922dab61SIlya Dryomov struct linger_work { 2873922dab61SIlya Dryomov struct work_struct work; 2874922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2875b07d3c4bSIlya Dryomov struct list_head pending_item; 2876b07d3c4bSIlya Dryomov unsigned long queued_stamp; 2877922dab61SIlya Dryomov 2878922dab61SIlya Dryomov union { 2879922dab61SIlya Dryomov struct { 2880922dab61SIlya Dryomov u64 notify_id; 2881922dab61SIlya Dryomov u64 notifier_id; 2882922dab61SIlya Dryomov void *payload; /* points into @msg front */ 2883922dab61SIlya Dryomov size_t payload_len; 2884922dab61SIlya Dryomov 2885922dab61SIlya Dryomov struct ceph_msg *msg; /* for ceph_msg_put() */ 2886922dab61SIlya Dryomov } notify; 2887922dab61SIlya Dryomov struct { 2888922dab61SIlya Dryomov int err; 2889922dab61SIlya Dryomov } error; 2890922dab61SIlya Dryomov }; 2891922dab61SIlya Dryomov }; 2892922dab61SIlya Dryomov 2893922dab61SIlya Dryomov static struct linger_work *lwork_alloc(struct ceph_osd_linger_request *lreq, 2894922dab61SIlya Dryomov work_func_t workfn) 2895922dab61SIlya Dryomov { 2896922dab61SIlya Dryomov struct linger_work *lwork; 2897922dab61SIlya Dryomov 2898922dab61SIlya Dryomov lwork = kzalloc(sizeof(*lwork), GFP_NOIO); 2899922dab61SIlya Dryomov if (!lwork) 2900922dab61SIlya Dryomov return NULL; 2901922dab61SIlya Dryomov 2902922dab61SIlya Dryomov INIT_WORK(&lwork->work, workfn); 2903b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lwork->pending_item); 2904922dab61SIlya Dryomov lwork->lreq = linger_get(lreq); 2905922dab61SIlya Dryomov 2906922dab61SIlya Dryomov return lwork; 2907922dab61SIlya Dryomov } 2908922dab61SIlya Dryomov 2909922dab61SIlya Dryomov static void lwork_free(struct linger_work *lwork) 2910922dab61SIlya Dryomov { 2911922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2912922dab61SIlya Dryomov 2913b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 2914b07d3c4bSIlya Dryomov list_del(&lwork->pending_item); 2915b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 2916b07d3c4bSIlya Dryomov 2917922dab61SIlya Dryomov linger_put(lreq); 2918922dab61SIlya Dryomov kfree(lwork); 2919922dab61SIlya Dryomov } 2920922dab61SIlya Dryomov 2921922dab61SIlya Dryomov static void lwork_queue(struct linger_work *lwork) 2922922dab61SIlya Dryomov { 2923922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2924922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2925922dab61SIlya Dryomov 2926922dab61SIlya Dryomov verify_lreq_locked(lreq); 2927b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lwork->pending_item)); 2928b07d3c4bSIlya Dryomov 2929b07d3c4bSIlya Dryomov lwork->queued_stamp = jiffies; 2930b07d3c4bSIlya Dryomov list_add_tail(&lwork->pending_item, &lreq->pending_lworks); 2931922dab61SIlya Dryomov queue_work(osdc->notify_wq, &lwork->work); 2932922dab61SIlya Dryomov } 2933922dab61SIlya Dryomov 2934922dab61SIlya Dryomov static void do_watch_notify(struct work_struct *w) 2935922dab61SIlya Dryomov { 2936922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2937922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2938922dab61SIlya Dryomov 2939922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2940922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2941922dab61SIlya Dryomov goto out; 2942922dab61SIlya Dryomov } 2943922dab61SIlya Dryomov 294419079203SIlya Dryomov WARN_ON(!lreq->is_watch); 2945922dab61SIlya Dryomov dout("%s lreq %p notify_id %llu notifier_id %llu payload_len %zu\n", 2946922dab61SIlya Dryomov __func__, lreq, lwork->notify.notify_id, lwork->notify.notifier_id, 2947922dab61SIlya Dryomov lwork->notify.payload_len); 2948922dab61SIlya Dryomov lreq->wcb(lreq->data, lwork->notify.notify_id, lreq->linger_id, 2949922dab61SIlya Dryomov lwork->notify.notifier_id, lwork->notify.payload, 2950922dab61SIlya Dryomov lwork->notify.payload_len); 2951922dab61SIlya Dryomov 2952922dab61SIlya Dryomov out: 2953922dab61SIlya Dryomov ceph_msg_put(lwork->notify.msg); 2954922dab61SIlya Dryomov lwork_free(lwork); 2955922dab61SIlya Dryomov } 2956922dab61SIlya Dryomov 2957922dab61SIlya Dryomov static void do_watch_error(struct work_struct *w) 2958922dab61SIlya Dryomov { 2959922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2960922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2961922dab61SIlya Dryomov 2962922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2963922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2964922dab61SIlya Dryomov goto out; 2965922dab61SIlya Dryomov } 2966922dab61SIlya Dryomov 2967922dab61SIlya Dryomov dout("%s lreq %p err %d\n", __func__, lreq, lwork->error.err); 2968922dab61SIlya Dryomov lreq->errcb(lreq->data, lreq->linger_id, lwork->error.err); 2969922dab61SIlya Dryomov 2970922dab61SIlya Dryomov out: 2971922dab61SIlya Dryomov lwork_free(lwork); 2972922dab61SIlya Dryomov } 2973922dab61SIlya Dryomov 2974922dab61SIlya Dryomov static void queue_watch_error(struct ceph_osd_linger_request *lreq) 2975922dab61SIlya Dryomov { 2976922dab61SIlya Dryomov struct linger_work *lwork; 2977922dab61SIlya Dryomov 2978922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_error); 2979922dab61SIlya Dryomov if (!lwork) { 2980922dab61SIlya Dryomov pr_err("failed to allocate error-lwork\n"); 2981922dab61SIlya Dryomov return; 2982922dab61SIlya Dryomov } 2983922dab61SIlya Dryomov 2984922dab61SIlya Dryomov lwork->error.err = lreq->last_error; 2985922dab61SIlya Dryomov lwork_queue(lwork); 2986922dab61SIlya Dryomov } 2987922dab61SIlya Dryomov 2988922dab61SIlya Dryomov static void linger_reg_commit_complete(struct ceph_osd_linger_request *lreq, 2989922dab61SIlya Dryomov int result) 2990922dab61SIlya Dryomov { 2991922dab61SIlya Dryomov if (!completion_done(&lreq->reg_commit_wait)) { 2992922dab61SIlya Dryomov lreq->reg_commit_error = (result <= 0 ? result : 0); 2993922dab61SIlya Dryomov complete_all(&lreq->reg_commit_wait); 2994922dab61SIlya Dryomov } 2995922dab61SIlya Dryomov } 2996922dab61SIlya Dryomov 2997922dab61SIlya Dryomov static void linger_commit_cb(struct ceph_osd_request *req) 2998922dab61SIlya Dryomov { 2999922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 3000922dab61SIlya Dryomov 3001922dab61SIlya Dryomov mutex_lock(&lreq->lock); 3002922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d\n", __func__, lreq, 3003922dab61SIlya Dryomov lreq->linger_id, req->r_result); 3004922dab61SIlya Dryomov linger_reg_commit_complete(lreq, req->r_result); 3005922dab61SIlya Dryomov lreq->committed = true; 3006922dab61SIlya Dryomov 300719079203SIlya Dryomov if (!lreq->is_watch) { 300819079203SIlya Dryomov struct ceph_osd_data *osd_data = 300919079203SIlya Dryomov osd_req_op_data(req, 0, notify, response_data); 301019079203SIlya Dryomov void *p = page_address(osd_data->pages[0]); 301119079203SIlya Dryomov 301219079203SIlya Dryomov WARN_ON(req->r_ops[0].op != CEPH_OSD_OP_NOTIFY || 301319079203SIlya Dryomov osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 301419079203SIlya Dryomov 301519079203SIlya Dryomov /* make note of the notify_id */ 301619079203SIlya Dryomov if (req->r_ops[0].outdata_len >= sizeof(u64)) { 301719079203SIlya Dryomov lreq->notify_id = ceph_decode_64(&p); 301819079203SIlya Dryomov dout("lreq %p notify_id %llu\n", lreq, 301919079203SIlya Dryomov lreq->notify_id); 302019079203SIlya Dryomov } else { 302119079203SIlya Dryomov dout("lreq %p no notify_id\n", lreq); 302219079203SIlya Dryomov } 302319079203SIlya Dryomov } 302419079203SIlya Dryomov 3025922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3026922dab61SIlya Dryomov linger_put(lreq); 3027922dab61SIlya Dryomov } 3028922dab61SIlya Dryomov 3029922dab61SIlya Dryomov static int normalize_watch_error(int err) 3030922dab61SIlya Dryomov { 3031922dab61SIlya Dryomov /* 3032922dab61SIlya Dryomov * Translate ENOENT -> ENOTCONN so that a delete->disconnection 3033922dab61SIlya Dryomov * notification and a failure to reconnect because we raced with 3034922dab61SIlya Dryomov * the delete appear the same to the user. 3035922dab61SIlya Dryomov */ 3036922dab61SIlya Dryomov if (err == -ENOENT) 3037922dab61SIlya Dryomov err = -ENOTCONN; 3038922dab61SIlya Dryomov 3039922dab61SIlya Dryomov return err; 3040922dab61SIlya Dryomov } 3041922dab61SIlya Dryomov 3042922dab61SIlya Dryomov static void linger_reconnect_cb(struct ceph_osd_request *req) 3043922dab61SIlya Dryomov { 3044922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 3045922dab61SIlya Dryomov 3046922dab61SIlya Dryomov mutex_lock(&lreq->lock); 3047922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d last_error %d\n", __func__, 3048922dab61SIlya Dryomov lreq, lreq->linger_id, req->r_result, lreq->last_error); 3049922dab61SIlya Dryomov if (req->r_result < 0) { 3050922dab61SIlya Dryomov if (!lreq->last_error) { 3051922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 3052922dab61SIlya Dryomov queue_watch_error(lreq); 3053922dab61SIlya Dryomov } 3054922dab61SIlya Dryomov } 3055922dab61SIlya Dryomov 3056922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3057922dab61SIlya Dryomov linger_put(lreq); 3058922dab61SIlya Dryomov } 3059922dab61SIlya Dryomov 3060922dab61SIlya Dryomov static void send_linger(struct ceph_osd_linger_request *lreq) 3061922dab61SIlya Dryomov { 3062922dab61SIlya Dryomov struct ceph_osd_request *req = lreq->reg_req; 3063922dab61SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[0]; 3064922dab61SIlya Dryomov 3065922dab61SIlya Dryomov verify_osdc_wrlocked(req->r_osdc); 3066922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 3067922dab61SIlya Dryomov 3068922dab61SIlya Dryomov if (req->r_osd) 3069922dab61SIlya Dryomov cancel_linger_request(req); 3070922dab61SIlya Dryomov 3071922dab61SIlya Dryomov request_reinit(req); 30725133ba8fSIlya Dryomov target_copy(&req->r_t, &lreq->t); 3073922dab61SIlya Dryomov req->r_mtime = lreq->mtime; 3074922dab61SIlya Dryomov 3075922dab61SIlya Dryomov mutex_lock(&lreq->lock); 307619079203SIlya Dryomov if (lreq->is_watch && lreq->committed) { 3077922dab61SIlya Dryomov WARN_ON(op->op != CEPH_OSD_OP_WATCH || 3078922dab61SIlya Dryomov op->watch.cookie != lreq->linger_id); 3079922dab61SIlya Dryomov op->watch.op = CEPH_OSD_WATCH_OP_RECONNECT; 3080922dab61SIlya Dryomov op->watch.gen = ++lreq->register_gen; 3081922dab61SIlya Dryomov dout("lreq %p reconnect register_gen %u\n", lreq, 3082922dab61SIlya Dryomov op->watch.gen); 3083922dab61SIlya Dryomov req->r_callback = linger_reconnect_cb; 3084922dab61SIlya Dryomov } else { 308519079203SIlya Dryomov if (!lreq->is_watch) 308619079203SIlya Dryomov lreq->notify_id = 0; 308719079203SIlya Dryomov else 3088922dab61SIlya Dryomov WARN_ON(op->watch.op != CEPH_OSD_WATCH_OP_WATCH); 3089922dab61SIlya Dryomov dout("lreq %p register\n", lreq); 3090922dab61SIlya Dryomov req->r_callback = linger_commit_cb; 3091922dab61SIlya Dryomov } 3092922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3093922dab61SIlya Dryomov 3094922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 3095922dab61SIlya Dryomov req->r_linger = true; 3096922dab61SIlya Dryomov 3097922dab61SIlya Dryomov submit_request(req, true); 3098922dab61SIlya Dryomov } 3099922dab61SIlya Dryomov 3100922dab61SIlya Dryomov static void linger_ping_cb(struct ceph_osd_request *req) 3101922dab61SIlya Dryomov { 3102922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 3103922dab61SIlya Dryomov 3104922dab61SIlya Dryomov mutex_lock(&lreq->lock); 3105922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d ping_sent %lu last_error %d\n", 3106922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, req->r_result, lreq->ping_sent, 3107922dab61SIlya Dryomov lreq->last_error); 3108922dab61SIlya Dryomov if (lreq->register_gen == req->r_ops[0].watch.gen) { 3109b07d3c4bSIlya Dryomov if (!req->r_result) { 3110b07d3c4bSIlya Dryomov lreq->watch_valid_thru = lreq->ping_sent; 3111b07d3c4bSIlya Dryomov } else if (!lreq->last_error) { 3112922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 3113922dab61SIlya Dryomov queue_watch_error(lreq); 3114922dab61SIlya Dryomov } 3115922dab61SIlya Dryomov } else { 3116922dab61SIlya Dryomov dout("lreq %p register_gen %u ignoring old pong %u\n", lreq, 3117922dab61SIlya Dryomov lreq->register_gen, req->r_ops[0].watch.gen); 3118922dab61SIlya Dryomov } 3119922dab61SIlya Dryomov 3120922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3121922dab61SIlya Dryomov linger_put(lreq); 3122922dab61SIlya Dryomov } 3123922dab61SIlya Dryomov 3124922dab61SIlya Dryomov static void send_linger_ping(struct ceph_osd_linger_request *lreq) 3125922dab61SIlya Dryomov { 3126922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3127922dab61SIlya Dryomov struct ceph_osd_request *req = lreq->ping_req; 3128922dab61SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[0]; 3129922dab61SIlya Dryomov 3130b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 3131922dab61SIlya Dryomov dout("%s PAUSERD\n", __func__); 3132922dab61SIlya Dryomov return; 3133922dab61SIlya Dryomov } 3134922dab61SIlya Dryomov 3135922dab61SIlya Dryomov lreq->ping_sent = jiffies; 3136922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu ping_sent %lu register_gen %u\n", 3137922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, lreq->ping_sent, 3138922dab61SIlya Dryomov lreq->register_gen); 3139922dab61SIlya Dryomov 3140922dab61SIlya Dryomov if (req->r_osd) 3141922dab61SIlya Dryomov cancel_linger_request(req); 3142922dab61SIlya Dryomov 3143922dab61SIlya Dryomov request_reinit(req); 3144922dab61SIlya Dryomov target_copy(&req->r_t, &lreq->t); 3145922dab61SIlya Dryomov 3146922dab61SIlya Dryomov WARN_ON(op->op != CEPH_OSD_OP_WATCH || 3147922dab61SIlya Dryomov op->watch.cookie != lreq->linger_id || 3148922dab61SIlya Dryomov op->watch.op != CEPH_OSD_WATCH_OP_PING); 3149922dab61SIlya Dryomov op->watch.gen = lreq->register_gen; 3150922dab61SIlya Dryomov req->r_callback = linger_ping_cb; 3151922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 3152922dab61SIlya Dryomov req->r_linger = true; 3153922dab61SIlya Dryomov 3154922dab61SIlya Dryomov ceph_osdc_get_request(req); 3155922dab61SIlya Dryomov account_request(req); 3156922dab61SIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 3157922dab61SIlya Dryomov link_request(lreq->osd, req); 3158922dab61SIlya Dryomov send_request(req); 3159922dab61SIlya Dryomov } 3160922dab61SIlya Dryomov 3161922dab61SIlya Dryomov static void linger_submit(struct ceph_osd_linger_request *lreq) 3162922dab61SIlya Dryomov { 3163922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3164922dab61SIlya Dryomov struct ceph_osd *osd; 3165922dab61SIlya Dryomov 316681c65213SIlya Dryomov down_write(&osdc->lock); 316781c65213SIlya Dryomov linger_register(lreq); 316881c65213SIlya Dryomov if (lreq->is_watch) { 316981c65213SIlya Dryomov lreq->reg_req->r_ops[0].watch.cookie = lreq->linger_id; 317081c65213SIlya Dryomov lreq->ping_req->r_ops[0].watch.cookie = lreq->linger_id; 317181c65213SIlya Dryomov } else { 317281c65213SIlya Dryomov lreq->reg_req->r_ops[0].notify.cookie = lreq->linger_id; 317381c65213SIlya Dryomov } 317481c65213SIlya Dryomov 31758edf84baSIlya Dryomov calc_target(osdc, &lreq->t, false); 3176922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3177922dab61SIlya Dryomov link_linger(osd, lreq); 3178922dab61SIlya Dryomov 3179922dab61SIlya Dryomov send_linger(lreq); 318081c65213SIlya Dryomov up_write(&osdc->lock); 3181922dab61SIlya Dryomov } 3182922dab61SIlya Dryomov 31834609245eSIlya Dryomov static void cancel_linger_map_check(struct ceph_osd_linger_request *lreq) 31844609245eSIlya Dryomov { 31854609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 31864609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 31874609245eSIlya Dryomov 31884609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 31894609245eSIlya Dryomov 31904609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 31914609245eSIlya Dryomov lreq->linger_id); 31924609245eSIlya Dryomov if (!lookup_lreq) 31934609245eSIlya Dryomov return; 31944609245eSIlya Dryomov 31954609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 31964609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 31974609245eSIlya Dryomov linger_put(lreq); 31984609245eSIlya Dryomov } 31994609245eSIlya Dryomov 3200922dab61SIlya Dryomov /* 3201922dab61SIlya Dryomov * @lreq has to be both registered and linked. 3202922dab61SIlya Dryomov */ 3203922dab61SIlya Dryomov static void __linger_cancel(struct ceph_osd_linger_request *lreq) 3204922dab61SIlya Dryomov { 320519079203SIlya Dryomov if (lreq->is_watch && lreq->ping_req->r_osd) 3206922dab61SIlya Dryomov cancel_linger_request(lreq->ping_req); 3207922dab61SIlya Dryomov if (lreq->reg_req->r_osd) 3208922dab61SIlya Dryomov cancel_linger_request(lreq->reg_req); 32094609245eSIlya Dryomov cancel_linger_map_check(lreq); 3210922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3211922dab61SIlya Dryomov linger_unregister(lreq); 3212922dab61SIlya Dryomov } 3213922dab61SIlya Dryomov 3214922dab61SIlya Dryomov static void linger_cancel(struct ceph_osd_linger_request *lreq) 3215922dab61SIlya Dryomov { 3216922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3217922dab61SIlya Dryomov 3218922dab61SIlya Dryomov down_write(&osdc->lock); 3219922dab61SIlya Dryomov if (__linger_registered(lreq)) 3220922dab61SIlya Dryomov __linger_cancel(lreq); 3221922dab61SIlya Dryomov up_write(&osdc->lock); 3222922dab61SIlya Dryomov } 3223922dab61SIlya Dryomov 32244609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq); 32254609245eSIlya Dryomov 32264609245eSIlya Dryomov static void check_linger_pool_dne(struct ceph_osd_linger_request *lreq) 32274609245eSIlya Dryomov { 32284609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 32294609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 32304609245eSIlya Dryomov 32314609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 32324609245eSIlya Dryomov WARN_ON(!map->epoch); 32334609245eSIlya Dryomov 32344609245eSIlya Dryomov if (lreq->register_gen) { 32354609245eSIlya Dryomov lreq->map_dne_bound = map->epoch; 32364609245eSIlya Dryomov dout("%s lreq %p linger_id %llu pool disappeared\n", __func__, 32374609245eSIlya Dryomov lreq, lreq->linger_id); 32384609245eSIlya Dryomov } else { 32394609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u have %u\n", 32404609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 32414609245eSIlya Dryomov map->epoch); 32424609245eSIlya Dryomov } 32434609245eSIlya Dryomov 32444609245eSIlya Dryomov if (lreq->map_dne_bound) { 32454609245eSIlya Dryomov if (map->epoch >= lreq->map_dne_bound) { 32464609245eSIlya Dryomov /* we had a new enough map */ 32474609245eSIlya Dryomov pr_info("linger_id %llu pool does not exist\n", 32484609245eSIlya Dryomov lreq->linger_id); 32494609245eSIlya Dryomov linger_reg_commit_complete(lreq, -ENOENT); 32504609245eSIlya Dryomov __linger_cancel(lreq); 32514609245eSIlya Dryomov } 32524609245eSIlya Dryomov } else { 32534609245eSIlya Dryomov send_linger_map_check(lreq); 32544609245eSIlya Dryomov } 32554609245eSIlya Dryomov } 32564609245eSIlya Dryomov 32574609245eSIlya Dryomov static void linger_map_check_cb(struct ceph_mon_generic_request *greq) 32584609245eSIlya Dryomov { 32594609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 32604609245eSIlya Dryomov struct ceph_osd_linger_request *lreq; 32614609245eSIlya Dryomov u64 linger_id = greq->private_data; 32624609245eSIlya Dryomov 32634609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 32644609245eSIlya Dryomov 32654609245eSIlya Dryomov down_write(&osdc->lock); 32664609245eSIlya Dryomov lreq = lookup_linger_mc(&osdc->linger_map_checks, linger_id); 32674609245eSIlya Dryomov if (!lreq) { 32684609245eSIlya Dryomov dout("%s linger_id %llu dne\n", __func__, linger_id); 32694609245eSIlya Dryomov goto out_unlock; 32704609245eSIlya Dryomov } 32714609245eSIlya Dryomov 32724609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u newest %llu\n", 32734609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 32744609245eSIlya Dryomov greq->u.newest); 32754609245eSIlya Dryomov if (!lreq->map_dne_bound) 32764609245eSIlya Dryomov lreq->map_dne_bound = greq->u.newest; 32774609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 32784609245eSIlya Dryomov check_linger_pool_dne(lreq); 32794609245eSIlya Dryomov 32804609245eSIlya Dryomov linger_put(lreq); 32814609245eSIlya Dryomov out_unlock: 32824609245eSIlya Dryomov up_write(&osdc->lock); 32834609245eSIlya Dryomov } 32844609245eSIlya Dryomov 32854609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq) 32864609245eSIlya Dryomov { 32874609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 32884609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 32894609245eSIlya Dryomov int ret; 32904609245eSIlya Dryomov 32914609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 32924609245eSIlya Dryomov 32934609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 32944609245eSIlya Dryomov lreq->linger_id); 32954609245eSIlya Dryomov if (lookup_lreq) { 32964609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 32974609245eSIlya Dryomov return; 32984609245eSIlya Dryomov } 32994609245eSIlya Dryomov 33004609245eSIlya Dryomov linger_get(lreq); 33014609245eSIlya Dryomov insert_linger_mc(&osdc->linger_map_checks, lreq); 33024609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 33034609245eSIlya Dryomov linger_map_check_cb, lreq->linger_id); 33044609245eSIlya Dryomov WARN_ON(ret); 33054609245eSIlya Dryomov } 33064609245eSIlya Dryomov 3307922dab61SIlya Dryomov static int linger_reg_commit_wait(struct ceph_osd_linger_request *lreq) 3308922dab61SIlya Dryomov { 3309922dab61SIlya Dryomov int ret; 3310922dab61SIlya Dryomov 3311922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 3312922dab61SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->reg_commit_wait); 3313922dab61SIlya Dryomov return ret ?: lreq->reg_commit_error; 3314922dab61SIlya Dryomov } 3315922dab61SIlya Dryomov 331619079203SIlya Dryomov static int linger_notify_finish_wait(struct ceph_osd_linger_request *lreq) 331719079203SIlya Dryomov { 331819079203SIlya Dryomov int ret; 331919079203SIlya Dryomov 332019079203SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 332119079203SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->notify_finish_wait); 332219079203SIlya Dryomov return ret ?: lreq->notify_finish_error; 332319079203SIlya Dryomov } 332419079203SIlya Dryomov 3325922dab61SIlya Dryomov /* 3326fbca9635SIlya Dryomov * Timeout callback, called every N seconds. When 1 or more OSD 3327fbca9635SIlya Dryomov * requests has been active for more than N seconds, we send a keepalive 3328fbca9635SIlya Dryomov * (tag + timestamp) to its OSD to ensure any communications channel 3329fbca9635SIlya Dryomov * reset is detected. 33303d14c5d2SYehuda Sadeh */ 33313d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 33323d14c5d2SYehuda Sadeh { 33333d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 33343d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 3335a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 33365aea3dcdSIlya Dryomov unsigned long cutoff = jiffies - opts->osd_keepalive_timeout; 33377cc5e38fSIlya Dryomov unsigned long expiry_cutoff = jiffies - opts->osd_request_timeout; 33385aea3dcdSIlya Dryomov LIST_HEAD(slow_osds); 33395aea3dcdSIlya Dryomov struct rb_node *n, *p; 33403d14c5d2SYehuda Sadeh 33415aea3dcdSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 33425aea3dcdSIlya Dryomov down_write(&osdc->lock); 33433d14c5d2SYehuda Sadeh 33443d14c5d2SYehuda Sadeh /* 33453d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 33463d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 33473d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 33483d14c5d2SYehuda Sadeh */ 33495aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 33505aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 33515aea3dcdSIlya Dryomov bool found = false; 33523d14c5d2SYehuda Sadeh 33537cc5e38fSIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 33545aea3dcdSIlya Dryomov struct ceph_osd_request *req = 33555aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 33565aea3dcdSIlya Dryomov 33577cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 33587cc5e38fSIlya Dryomov 33595aea3dcdSIlya Dryomov if (time_before(req->r_stamp, cutoff)) { 33605aea3dcdSIlya Dryomov dout(" req %p tid %llu on osd%d is laggy\n", 33615aea3dcdSIlya Dryomov req, req->r_tid, osd->o_osd); 33625aea3dcdSIlya Dryomov found = true; 33635aea3dcdSIlya Dryomov } 33647cc5e38fSIlya Dryomov if (opts->osd_request_timeout && 33657cc5e38fSIlya Dryomov time_before(req->r_start_stamp, expiry_cutoff)) { 33667cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 33677cc5e38fSIlya Dryomov req->r_tid, osd->o_osd); 33687cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 33697cc5e38fSIlya Dryomov } 33705aea3dcdSIlya Dryomov } 3371922dab61SIlya Dryomov for (p = rb_first(&osd->o_linger_requests); p; p = rb_next(p)) { 3372922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3373922dab61SIlya Dryomov rb_entry(p, struct ceph_osd_linger_request, node); 3374922dab61SIlya Dryomov 3375922dab61SIlya Dryomov dout(" lreq %p linger_id %llu is served by osd%d\n", 3376922dab61SIlya Dryomov lreq, lreq->linger_id, osd->o_osd); 3377922dab61SIlya Dryomov found = true; 3378922dab61SIlya Dryomov 3379922dab61SIlya Dryomov mutex_lock(&lreq->lock); 338019079203SIlya Dryomov if (lreq->is_watch && lreq->committed && !lreq->last_error) 3381922dab61SIlya Dryomov send_linger_ping(lreq); 3382922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3383922dab61SIlya Dryomov } 33845aea3dcdSIlya Dryomov 33855aea3dcdSIlya Dryomov if (found) 33863d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 33873d14c5d2SYehuda Sadeh } 33885aea3dcdSIlya Dryomov 33897cc5e38fSIlya Dryomov if (opts->osd_request_timeout) { 33907cc5e38fSIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 33917cc5e38fSIlya Dryomov struct ceph_osd_request *req = 33927cc5e38fSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 33937cc5e38fSIlya Dryomov 33947cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 33957cc5e38fSIlya Dryomov 33967cc5e38fSIlya Dryomov if (time_before(req->r_start_stamp, expiry_cutoff)) { 33977cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 33987cc5e38fSIlya Dryomov req->r_tid, osdc->homeless_osd.o_osd); 33997cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 34007cc5e38fSIlya Dryomov } 34017cc5e38fSIlya Dryomov } 34027cc5e38fSIlya Dryomov } 34037cc5e38fSIlya Dryomov 34045aea3dcdSIlya Dryomov if (atomic_read(&osdc->num_homeless) || !list_empty(&slow_osds)) 34055aea3dcdSIlya Dryomov maybe_request_map(osdc); 34065aea3dcdSIlya Dryomov 34073d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 34085aea3dcdSIlya Dryomov struct ceph_osd *osd = list_first_entry(&slow_osds, 34095aea3dcdSIlya Dryomov struct ceph_osd, 34103d14c5d2SYehuda Sadeh o_keepalive_item); 34113d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 34123d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 34133d14c5d2SYehuda Sadeh } 34143d14c5d2SYehuda Sadeh 34155aea3dcdSIlya Dryomov up_write(&osdc->lock); 3416fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 3417fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 34183d14c5d2SYehuda Sadeh } 34193d14c5d2SYehuda Sadeh 34203d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 34213d14c5d2SYehuda Sadeh { 34223d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 34233d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 34243d14c5d2SYehuda Sadeh osds_timeout_work.work); 3425a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 342642a2c09fSIlya Dryomov struct ceph_osd *osd, *nosd; 34273d14c5d2SYehuda Sadeh 342842a2c09fSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 34295aea3dcdSIlya Dryomov down_write(&osdc->lock); 343042a2c09fSIlya Dryomov list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 343142a2c09fSIlya Dryomov if (time_before(jiffies, osd->lru_ttl)) 343242a2c09fSIlya Dryomov break; 343342a2c09fSIlya Dryomov 34345aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 3435922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 34365aea3dcdSIlya Dryomov close_osd(osd); 343742a2c09fSIlya Dryomov } 343842a2c09fSIlya Dryomov 34395aea3dcdSIlya Dryomov up_write(&osdc->lock); 34403d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 34413d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 34423d14c5d2SYehuda Sadeh } 34433d14c5d2SYehuda Sadeh 3444205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 3445205ee118SIlya Dryomov struct ceph_object_locator *oloc) 3446205ee118SIlya Dryomov { 3447205ee118SIlya Dryomov u8 struct_v, struct_cv; 3448205ee118SIlya Dryomov u32 len; 3449205ee118SIlya Dryomov void *struct_end; 3450205ee118SIlya Dryomov int ret = 0; 3451205ee118SIlya Dryomov 3452205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3453205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3454205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3455205ee118SIlya Dryomov if (struct_v < 3) { 3456205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 3457205ee118SIlya Dryomov struct_v, struct_cv); 3458205ee118SIlya Dryomov goto e_inval; 3459205ee118SIlya Dryomov } 3460205ee118SIlya Dryomov if (struct_cv > 6) { 3461205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 3462205ee118SIlya Dryomov struct_v, struct_cv); 3463205ee118SIlya Dryomov goto e_inval; 3464205ee118SIlya Dryomov } 3465205ee118SIlya Dryomov len = ceph_decode_32(p); 3466205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3467205ee118SIlya Dryomov struct_end = *p + len; 3468205ee118SIlya Dryomov 3469205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 3470205ee118SIlya Dryomov *p += 4; /* skip preferred */ 3471205ee118SIlya Dryomov 3472205ee118SIlya Dryomov len = ceph_decode_32(p); 3473205ee118SIlya Dryomov if (len > 0) { 3474205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 3475205ee118SIlya Dryomov goto e_inval; 3476205ee118SIlya Dryomov } 3477205ee118SIlya Dryomov 3478205ee118SIlya Dryomov if (struct_v >= 5) { 3479cd08e0a2SYan, Zheng bool changed = false; 3480cd08e0a2SYan, Zheng 3481205ee118SIlya Dryomov len = ceph_decode_32(p); 3482205ee118SIlya Dryomov if (len > 0) { 348330c156d9SYan, Zheng ceph_decode_need(p, end, len, e_inval); 3484cd08e0a2SYan, Zheng if (!oloc->pool_ns || 3485cd08e0a2SYan, Zheng ceph_compare_string(oloc->pool_ns, *p, len)) 3486cd08e0a2SYan, Zheng changed = true; 348730c156d9SYan, Zheng *p += len; 3488cd08e0a2SYan, Zheng } else { 3489cd08e0a2SYan, Zheng if (oloc->pool_ns) 3490cd08e0a2SYan, Zheng changed = true; 3491cd08e0a2SYan, Zheng } 3492cd08e0a2SYan, Zheng if (changed) { 3493cd08e0a2SYan, Zheng /* redirect changes namespace */ 3494cd08e0a2SYan, Zheng pr_warn("ceph_object_locator::nspace is changed\n"); 3495cd08e0a2SYan, Zheng goto e_inval; 3496205ee118SIlya Dryomov } 3497205ee118SIlya Dryomov } 3498205ee118SIlya Dryomov 3499205ee118SIlya Dryomov if (struct_v >= 6) { 3500205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 3501205ee118SIlya Dryomov if (hash != -1) { 3502205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 3503205ee118SIlya Dryomov goto e_inval; 3504205ee118SIlya Dryomov } 3505205ee118SIlya Dryomov } 3506205ee118SIlya Dryomov 3507205ee118SIlya Dryomov /* skip the rest */ 3508205ee118SIlya Dryomov *p = struct_end; 3509205ee118SIlya Dryomov out: 3510205ee118SIlya Dryomov return ret; 3511205ee118SIlya Dryomov 3512205ee118SIlya Dryomov e_inval: 3513205ee118SIlya Dryomov ret = -EINVAL; 3514205ee118SIlya Dryomov goto out; 3515205ee118SIlya Dryomov } 3516205ee118SIlya Dryomov 3517205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 3518205ee118SIlya Dryomov struct ceph_request_redirect *redir) 3519205ee118SIlya Dryomov { 3520205ee118SIlya Dryomov u8 struct_v, struct_cv; 3521205ee118SIlya Dryomov u32 len; 3522205ee118SIlya Dryomov void *struct_end; 3523205ee118SIlya Dryomov int ret; 3524205ee118SIlya Dryomov 3525205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3526205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3527205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3528205ee118SIlya Dryomov if (struct_cv > 1) { 3529205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 3530205ee118SIlya Dryomov struct_v, struct_cv); 3531205ee118SIlya Dryomov goto e_inval; 3532205ee118SIlya Dryomov } 3533205ee118SIlya Dryomov len = ceph_decode_32(p); 3534205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3535205ee118SIlya Dryomov struct_end = *p + len; 3536205ee118SIlya Dryomov 3537205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 3538205ee118SIlya Dryomov if (ret) 3539205ee118SIlya Dryomov goto out; 3540205ee118SIlya Dryomov 3541205ee118SIlya Dryomov len = ceph_decode_32(p); 3542205ee118SIlya Dryomov if (len > 0) { 3543205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 3544205ee118SIlya Dryomov goto e_inval; 3545205ee118SIlya Dryomov } 3546205ee118SIlya Dryomov 3547205ee118SIlya Dryomov /* skip the rest */ 3548205ee118SIlya Dryomov *p = struct_end; 3549205ee118SIlya Dryomov out: 3550205ee118SIlya Dryomov return ret; 3551205ee118SIlya Dryomov 3552205ee118SIlya Dryomov e_inval: 3553205ee118SIlya Dryomov ret = -EINVAL; 3554205ee118SIlya Dryomov goto out; 3555205ee118SIlya Dryomov } 3556205ee118SIlya Dryomov 3557fe5da05eSIlya Dryomov struct MOSDOpReply { 3558fe5da05eSIlya Dryomov struct ceph_pg pgid; 3559fe5da05eSIlya Dryomov u64 flags; 3560fe5da05eSIlya Dryomov int result; 3561fe5da05eSIlya Dryomov u32 epoch; 3562fe5da05eSIlya Dryomov int num_ops; 3563fe5da05eSIlya Dryomov u32 outdata_len[CEPH_OSD_MAX_OPS]; 3564fe5da05eSIlya Dryomov s32 rval[CEPH_OSD_MAX_OPS]; 3565fe5da05eSIlya Dryomov int retry_attempt; 3566fe5da05eSIlya Dryomov struct ceph_eversion replay_version; 3567fe5da05eSIlya Dryomov u64 user_version; 3568fe5da05eSIlya Dryomov struct ceph_request_redirect redirect; 3569fe5da05eSIlya Dryomov }; 357025845472SSage Weil 3571fe5da05eSIlya Dryomov static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m) 35723d14c5d2SYehuda Sadeh { 3573fe5da05eSIlya Dryomov void *p = msg->front.iov_base; 3574fe5da05eSIlya Dryomov void *const end = p + msg->front.iov_len; 3575fe5da05eSIlya Dryomov u16 version = le16_to_cpu(msg->hdr.version); 3576fe5da05eSIlya Dryomov struct ceph_eversion bad_replay_version; 3577b0b31a8fSIlya Dryomov u8 decode_redir; 3578fe5da05eSIlya Dryomov u32 len; 3579fe5da05eSIlya Dryomov int ret; 3580fe5da05eSIlya Dryomov int i; 35813d14c5d2SYehuda Sadeh 3582fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, len, e_inval); 3583fe5da05eSIlya Dryomov ceph_decode_need(&p, end, len, e_inval); 3584fe5da05eSIlya Dryomov p += len; /* skip oid */ 35851b83bef2SSage Weil 3586fe5da05eSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->pgid); 3587fe5da05eSIlya Dryomov if (ret) 3588fe5da05eSIlya Dryomov return ret; 35891b83bef2SSage Weil 3590fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->flags, e_inval); 3591fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->result, e_inval); 3592fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval); 3593fe5da05eSIlya Dryomov memcpy(&bad_replay_version, p, sizeof(bad_replay_version)); 3594fe5da05eSIlya Dryomov p += sizeof(bad_replay_version); 3595fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->epoch, e_inval); 35961b83bef2SSage Weil 3597fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->num_ops, e_inval); 3598fe5da05eSIlya Dryomov if (m->num_ops > ARRAY_SIZE(m->outdata_len)) 3599fe5da05eSIlya Dryomov goto e_inval; 36001b83bef2SSage Weil 3601fe5da05eSIlya Dryomov ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op), 3602fe5da05eSIlya Dryomov e_inval); 3603fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) { 36041b83bef2SSage Weil struct ceph_osd_op *op = p; 36051b83bef2SSage Weil 3606fe5da05eSIlya Dryomov m->outdata_len[i] = le32_to_cpu(op->payload_len); 36071b83bef2SSage Weil p += sizeof(*op); 36081b83bef2SSage Weil } 3609fe5da05eSIlya Dryomov 3610fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval); 3611fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) 3612fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->rval[i], e_inval); 3613fe5da05eSIlya Dryomov 3614fe5da05eSIlya Dryomov if (version >= 5) { 3615fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval); 3616fe5da05eSIlya Dryomov memcpy(&m->replay_version, p, sizeof(m->replay_version)); 3617fe5da05eSIlya Dryomov p += sizeof(m->replay_version); 3618fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->user_version, e_inval); 3619fe5da05eSIlya Dryomov } else { 3620fe5da05eSIlya Dryomov m->replay_version = bad_replay_version; /* struct */ 3621fe5da05eSIlya Dryomov m->user_version = le64_to_cpu(m->replay_version.version); 36221b83bef2SSage Weil } 36231b83bef2SSage Weil 3624fe5da05eSIlya Dryomov if (version >= 6) { 3625fe5da05eSIlya Dryomov if (version >= 7) 3626fe5da05eSIlya Dryomov ceph_decode_8_safe(&p, end, decode_redir, e_inval); 3627b0b31a8fSIlya Dryomov else 3628b0b31a8fSIlya Dryomov decode_redir = 1; 3629b0b31a8fSIlya Dryomov } else { 3630b0b31a8fSIlya Dryomov decode_redir = 0; 3631b0b31a8fSIlya Dryomov } 3632b0b31a8fSIlya Dryomov 3633b0b31a8fSIlya Dryomov if (decode_redir) { 3634fe5da05eSIlya Dryomov ret = ceph_redirect_decode(&p, end, &m->redirect); 3635fe5da05eSIlya Dryomov if (ret) 3636fe5da05eSIlya Dryomov return ret; 3637205ee118SIlya Dryomov } else { 3638fe5da05eSIlya Dryomov ceph_oloc_init(&m->redirect.oloc); 3639205ee118SIlya Dryomov } 3640205ee118SIlya Dryomov 3641fe5da05eSIlya Dryomov return 0; 3642205ee118SIlya Dryomov 3643fe5da05eSIlya Dryomov e_inval: 3644fe5da05eSIlya Dryomov return -EINVAL; 3645fe5da05eSIlya Dryomov } 3646fe5da05eSIlya Dryomov 3647fe5da05eSIlya Dryomov /* 3648b18b9550SIlya Dryomov * Handle MOSDOpReply. Set ->r_result and call the callback if it is 3649b18b9550SIlya Dryomov * specified. 3650fe5da05eSIlya Dryomov */ 36515aea3dcdSIlya Dryomov static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg) 3652fe5da05eSIlya Dryomov { 36535aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 3654fe5da05eSIlya Dryomov struct ceph_osd_request *req; 3655fe5da05eSIlya Dryomov struct MOSDOpReply m; 3656fe5da05eSIlya Dryomov u64 tid = le64_to_cpu(msg->hdr.tid); 3657fe5da05eSIlya Dryomov u32 data_len = 0; 3658fe5da05eSIlya Dryomov int ret; 3659fe5da05eSIlya Dryomov int i; 3660fe5da05eSIlya Dryomov 3661fe5da05eSIlya Dryomov dout("%s msg %p tid %llu\n", __func__, msg, tid); 3662fe5da05eSIlya Dryomov 36635aea3dcdSIlya Dryomov down_read(&osdc->lock); 36645aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 36655aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 36665aea3dcdSIlya Dryomov goto out_unlock_osdc; 3667fe5da05eSIlya Dryomov } 36685aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 36695aea3dcdSIlya Dryomov 36705aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 36715aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 36725aea3dcdSIlya Dryomov if (!req) { 36735aea3dcdSIlya Dryomov dout("%s osd%d tid %llu unknown\n", __func__, osd->o_osd, tid); 36745aea3dcdSIlya Dryomov goto out_unlock_session; 36755aea3dcdSIlya Dryomov } 3676fe5da05eSIlya Dryomov 3677cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = req->r_t.target_oloc.pool_ns; 3678fe5da05eSIlya Dryomov ret = decode_MOSDOpReply(msg, &m); 3679cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = NULL; 3680fe5da05eSIlya Dryomov if (ret) { 3681fe5da05eSIlya Dryomov pr_err("failed to decode MOSDOpReply for tid %llu: %d\n", 3682fe5da05eSIlya Dryomov req->r_tid, ret); 3683fe5da05eSIlya Dryomov ceph_msg_dump(msg); 3684fe5da05eSIlya Dryomov goto fail_request; 3685fe5da05eSIlya Dryomov } 3686fe5da05eSIlya Dryomov dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n", 3687fe5da05eSIlya Dryomov __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed, 3688fe5da05eSIlya Dryomov m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch), 3689fe5da05eSIlya Dryomov le64_to_cpu(m.replay_version.version), m.user_version); 3690fe5da05eSIlya Dryomov 3691fe5da05eSIlya Dryomov if (m.retry_attempt >= 0) { 3692fe5da05eSIlya Dryomov if (m.retry_attempt != req->r_attempts - 1) { 3693fe5da05eSIlya Dryomov dout("req %p tid %llu retry_attempt %d != %d, ignoring\n", 3694fe5da05eSIlya Dryomov req, req->r_tid, m.retry_attempt, 3695fe5da05eSIlya Dryomov req->r_attempts - 1); 36965aea3dcdSIlya Dryomov goto out_unlock_session; 3697fe5da05eSIlya Dryomov } 3698fe5da05eSIlya Dryomov } else { 3699fe5da05eSIlya Dryomov WARN_ON(1); /* MOSDOpReply v4 is assumed */ 3700fe5da05eSIlya Dryomov } 3701fe5da05eSIlya Dryomov 3702fe5da05eSIlya Dryomov if (!ceph_oloc_empty(&m.redirect.oloc)) { 3703fe5da05eSIlya Dryomov dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid, 3704fe5da05eSIlya Dryomov m.redirect.oloc.pool); 37055aea3dcdSIlya Dryomov unlink_request(osd, req); 37065aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 3707205ee118SIlya Dryomov 370830c156d9SYan, Zheng /* 370930c156d9SYan, Zheng * Not ceph_oloc_copy() - changing pool_ns is not 371030c156d9SYan, Zheng * supported. 371130c156d9SYan, Zheng */ 371230c156d9SYan, Zheng req->r_t.target_oloc.pool = m.redirect.oloc.pool; 3713890bd0f8SJerry Lee req->r_flags |= CEPH_OSD_FLAG_REDIRECTED | 3714890bd0f8SJerry Lee CEPH_OSD_FLAG_IGNORE_OVERLAY | 3715890bd0f8SJerry Lee CEPH_OSD_FLAG_IGNORE_CACHE; 37165aea3dcdSIlya Dryomov req->r_tid = 0; 37175aea3dcdSIlya Dryomov __submit_request(req, false); 37185aea3dcdSIlya Dryomov goto out_unlock_osdc; 3719205ee118SIlya Dryomov } 3720205ee118SIlya Dryomov 3721117d96a0SIlya Dryomov if (m.result == -EAGAIN) { 3722117d96a0SIlya Dryomov dout("req %p tid %llu EAGAIN\n", req, req->r_tid); 3723117d96a0SIlya Dryomov unlink_request(osd, req); 3724117d96a0SIlya Dryomov mutex_unlock(&osd->lock); 3725117d96a0SIlya Dryomov 3726117d96a0SIlya Dryomov /* 3727117d96a0SIlya Dryomov * The object is missing on the replica or not (yet) 3728117d96a0SIlya Dryomov * readable. Clear pgid to force a resend to the primary 3729117d96a0SIlya Dryomov * via legacy_change. 3730117d96a0SIlya Dryomov */ 3731117d96a0SIlya Dryomov req->r_t.pgid.pool = 0; 3732117d96a0SIlya Dryomov req->r_t.pgid.seed = 0; 3733117d96a0SIlya Dryomov WARN_ON(!req->r_t.used_replica); 3734117d96a0SIlya Dryomov req->r_flags &= ~(CEPH_OSD_FLAG_BALANCE_READS | 3735117d96a0SIlya Dryomov CEPH_OSD_FLAG_LOCALIZE_READS); 3736117d96a0SIlya Dryomov req->r_tid = 0; 3737117d96a0SIlya Dryomov __submit_request(req, false); 3738117d96a0SIlya Dryomov goto out_unlock_osdc; 3739117d96a0SIlya Dryomov } 3740117d96a0SIlya Dryomov 3741fe5da05eSIlya Dryomov if (m.num_ops != req->r_num_ops) { 3742fe5da05eSIlya Dryomov pr_err("num_ops %d != %d for tid %llu\n", m.num_ops, 3743fe5da05eSIlya Dryomov req->r_num_ops, req->r_tid); 3744fe5da05eSIlya Dryomov goto fail_request; 3745fe5da05eSIlya Dryomov } 3746fe5da05eSIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 3747fe5da05eSIlya Dryomov dout(" req %p tid %llu op %d rval %d len %u\n", req, 3748fe5da05eSIlya Dryomov req->r_tid, i, m.rval[i], m.outdata_len[i]); 3749fe5da05eSIlya Dryomov req->r_ops[i].rval = m.rval[i]; 3750fe5da05eSIlya Dryomov req->r_ops[i].outdata_len = m.outdata_len[i]; 3751fe5da05eSIlya Dryomov data_len += m.outdata_len[i]; 3752fe5da05eSIlya Dryomov } 3753fe5da05eSIlya Dryomov if (data_len != le32_to_cpu(msg->hdr.data_len)) { 3754fe5da05eSIlya Dryomov pr_err("sum of lens %u != %u for tid %llu\n", data_len, 3755fe5da05eSIlya Dryomov le32_to_cpu(msg->hdr.data_len), req->r_tid); 3756fe5da05eSIlya Dryomov goto fail_request; 3757fe5da05eSIlya Dryomov } 3758b18b9550SIlya Dryomov dout("%s req %p tid %llu result %d data_len %u\n", __func__, 3759b18b9550SIlya Dryomov req, req->r_tid, m.result, data_len); 37603d14c5d2SYehuda Sadeh 3761b18b9550SIlya Dryomov /* 3762b18b9550SIlya Dryomov * Since we only ever request ONDISK, we should only ever get 3763b18b9550SIlya Dryomov * one (type of) reply back. 3764b18b9550SIlya Dryomov */ 3765b18b9550SIlya Dryomov WARN_ON(!(m.flags & CEPH_OSD_FLAG_ONDISK)); 3766fe5da05eSIlya Dryomov req->r_result = m.result ?: data_len; 376745ee2c1dSIlya Dryomov finish_request(req); 37685aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 37695aea3dcdSIlya Dryomov up_read(&osdc->lock); 37703d14c5d2SYehuda Sadeh 3771fe5da05eSIlya Dryomov __complete_request(req); 37723d14c5d2SYehuda Sadeh return; 3773fe5da05eSIlya Dryomov 3774fe5da05eSIlya Dryomov fail_request: 37754609245eSIlya Dryomov complete_request(req, -EIO); 37765aea3dcdSIlya Dryomov out_unlock_session: 37775aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 37785aea3dcdSIlya Dryomov out_unlock_osdc: 37795aea3dcdSIlya Dryomov up_read(&osdc->lock); 37803d14c5d2SYehuda Sadeh } 37813d14c5d2SYehuda Sadeh 378242c1b124SIlya Dryomov static void set_pool_was_full(struct ceph_osd_client *osdc) 378342c1b124SIlya Dryomov { 378442c1b124SIlya Dryomov struct rb_node *n; 378542c1b124SIlya Dryomov 378642c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 378742c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 378842c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 378942c1b124SIlya Dryomov 379042c1b124SIlya Dryomov pi->was_full = __pool_full(pi); 379142c1b124SIlya Dryomov } 379242c1b124SIlya Dryomov } 379342c1b124SIlya Dryomov 37945aea3dcdSIlya Dryomov static bool pool_cleared_full(struct ceph_osd_client *osdc, s64 pool_id) 37953d14c5d2SYehuda Sadeh { 37965aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 37973d14c5d2SYehuda Sadeh 37985aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 37995aea3dcdSIlya Dryomov if (!pi) 38005aea3dcdSIlya Dryomov return false; 38013d14c5d2SYehuda Sadeh 38025aea3dcdSIlya Dryomov return pi->was_full && !__pool_full(pi); 38033d14c5d2SYehuda Sadeh } 38043d14c5d2SYehuda Sadeh 3805922dab61SIlya Dryomov static enum calc_target_result 3806922dab61SIlya Dryomov recalc_linger_target(struct ceph_osd_linger_request *lreq) 3807922dab61SIlya Dryomov { 3808922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3809922dab61SIlya Dryomov enum calc_target_result ct_res; 3810922dab61SIlya Dryomov 38118edf84baSIlya Dryomov ct_res = calc_target(osdc, &lreq->t, true); 3812922dab61SIlya Dryomov if (ct_res == CALC_TARGET_NEED_RESEND) { 3813922dab61SIlya Dryomov struct ceph_osd *osd; 3814922dab61SIlya Dryomov 3815922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3816922dab61SIlya Dryomov if (osd != lreq->osd) { 3817922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3818922dab61SIlya Dryomov link_linger(osd, lreq); 3819922dab61SIlya Dryomov } 3820922dab61SIlya Dryomov } 3821922dab61SIlya Dryomov 3822922dab61SIlya Dryomov return ct_res; 3823922dab61SIlya Dryomov } 3824922dab61SIlya Dryomov 38253d14c5d2SYehuda Sadeh /* 38265aea3dcdSIlya Dryomov * Requeue requests whose mapping to an OSD has changed. 38273d14c5d2SYehuda Sadeh */ 38285aea3dcdSIlya Dryomov static void scan_requests(struct ceph_osd *osd, 38295aea3dcdSIlya Dryomov bool force_resend, 38305aea3dcdSIlya Dryomov bool cleared_full, 38315aea3dcdSIlya Dryomov bool check_pool_cleared_full, 38325aea3dcdSIlya Dryomov struct rb_root *need_resend, 38335aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 38343d14c5d2SYehuda Sadeh { 38355aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 38365aea3dcdSIlya Dryomov struct rb_node *n; 38375aea3dcdSIlya Dryomov bool force_resend_writes; 38383d14c5d2SYehuda Sadeh 3839922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 3840922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3841922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 3842922dab61SIlya Dryomov enum calc_target_result ct_res; 3843922dab61SIlya Dryomov 3844922dab61SIlya Dryomov n = rb_next(n); /* recalc_linger_target() */ 3845922dab61SIlya Dryomov 3846922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, 3847922dab61SIlya Dryomov lreq->linger_id); 3848922dab61SIlya Dryomov ct_res = recalc_linger_target(lreq); 3849922dab61SIlya Dryomov switch (ct_res) { 3850922dab61SIlya Dryomov case CALC_TARGET_NO_ACTION: 3851922dab61SIlya Dryomov force_resend_writes = cleared_full || 3852922dab61SIlya Dryomov (check_pool_cleared_full && 3853922dab61SIlya Dryomov pool_cleared_full(osdc, lreq->t.base_oloc.pool)); 3854922dab61SIlya Dryomov if (!force_resend && !force_resend_writes) 3855922dab61SIlya Dryomov break; 3856922dab61SIlya Dryomov 3857922dab61SIlya Dryomov /* fall through */ 3858922dab61SIlya Dryomov case CALC_TARGET_NEED_RESEND: 38594609245eSIlya Dryomov cancel_linger_map_check(lreq); 3860922dab61SIlya Dryomov /* 3861922dab61SIlya Dryomov * scan_requests() for the previous epoch(s) 3862922dab61SIlya Dryomov * may have already added it to the list, since 3863922dab61SIlya Dryomov * it's not unlinked here. 3864922dab61SIlya Dryomov */ 3865922dab61SIlya Dryomov if (list_empty(&lreq->scan_item)) 3866922dab61SIlya Dryomov list_add_tail(&lreq->scan_item, need_resend_linger); 3867922dab61SIlya Dryomov break; 3868922dab61SIlya Dryomov case CALC_TARGET_POOL_DNE: 3869a10bcb19SIlya Dryomov list_del_init(&lreq->scan_item); 38704609245eSIlya Dryomov check_linger_pool_dne(lreq); 3871922dab61SIlya Dryomov break; 3872922dab61SIlya Dryomov } 3873922dab61SIlya Dryomov } 3874922dab61SIlya Dryomov 38755aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 38765aea3dcdSIlya Dryomov struct ceph_osd_request *req = 38775aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 38785aea3dcdSIlya Dryomov enum calc_target_result ct_res; 3879ab60b16dSAlex Elder 38804609245eSIlya Dryomov n = rb_next(n); /* unlink_request(), check_pool_dne() */ 3881ab60b16dSAlex Elder 38825aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 38838edf84baSIlya Dryomov ct_res = calc_target(osdc, &req->r_t, false); 38845aea3dcdSIlya Dryomov switch (ct_res) { 38855aea3dcdSIlya Dryomov case CALC_TARGET_NO_ACTION: 38865aea3dcdSIlya Dryomov force_resend_writes = cleared_full || 38875aea3dcdSIlya Dryomov (check_pool_cleared_full && 38885aea3dcdSIlya Dryomov pool_cleared_full(osdc, req->r_t.base_oloc.pool)); 38895aea3dcdSIlya Dryomov if (!force_resend && 38905aea3dcdSIlya Dryomov (!(req->r_flags & CEPH_OSD_FLAG_WRITE) || 38915aea3dcdSIlya Dryomov !force_resend_writes)) 38925aea3dcdSIlya Dryomov break; 3893a40c4f10SYehuda Sadeh 38945aea3dcdSIlya Dryomov /* fall through */ 38955aea3dcdSIlya Dryomov case CALC_TARGET_NEED_RESEND: 38964609245eSIlya Dryomov cancel_map_check(req); 38975aea3dcdSIlya Dryomov unlink_request(osd, req); 38985aea3dcdSIlya Dryomov insert_request(need_resend, req); 38995aea3dcdSIlya Dryomov break; 39005aea3dcdSIlya Dryomov case CALC_TARGET_POOL_DNE: 39014609245eSIlya Dryomov check_pool_dne(req); 39025aea3dcdSIlya Dryomov break; 3903a40c4f10SYehuda Sadeh } 39043d14c5d2SYehuda Sadeh } 39053d14c5d2SYehuda Sadeh } 39066f6c7006SSage Weil 390742c1b124SIlya Dryomov static int handle_one_map(struct ceph_osd_client *osdc, 39085aea3dcdSIlya Dryomov void *p, void *end, bool incremental, 39095aea3dcdSIlya Dryomov struct rb_root *need_resend, 39105aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 391142c1b124SIlya Dryomov { 391242c1b124SIlya Dryomov struct ceph_osdmap *newmap; 391342c1b124SIlya Dryomov struct rb_node *n; 391442c1b124SIlya Dryomov bool skipped_map = false; 391542c1b124SIlya Dryomov bool was_full; 391642c1b124SIlya Dryomov 3917b7ec35b3SIlya Dryomov was_full = ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 391842c1b124SIlya Dryomov set_pool_was_full(osdc); 391942c1b124SIlya Dryomov 392042c1b124SIlya Dryomov if (incremental) 392142c1b124SIlya Dryomov newmap = osdmap_apply_incremental(&p, end, osdc->osdmap); 392242c1b124SIlya Dryomov else 392342c1b124SIlya Dryomov newmap = ceph_osdmap_decode(&p, end); 392442c1b124SIlya Dryomov if (IS_ERR(newmap)) 392542c1b124SIlya Dryomov return PTR_ERR(newmap); 392642c1b124SIlya Dryomov 392742c1b124SIlya Dryomov if (newmap != osdc->osdmap) { 392842c1b124SIlya Dryomov /* 392942c1b124SIlya Dryomov * Preserve ->was_full before destroying the old map. 393042c1b124SIlya Dryomov * For pools that weren't in the old map, ->was_full 393142c1b124SIlya Dryomov * should be false. 393242c1b124SIlya Dryomov */ 393342c1b124SIlya Dryomov for (n = rb_first(&newmap->pg_pools); n; n = rb_next(n)) { 393442c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 393542c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 393642c1b124SIlya Dryomov struct ceph_pg_pool_info *old_pi; 393742c1b124SIlya Dryomov 393842c1b124SIlya Dryomov old_pi = ceph_pg_pool_by_id(osdc->osdmap, pi->id); 393942c1b124SIlya Dryomov if (old_pi) 394042c1b124SIlya Dryomov pi->was_full = old_pi->was_full; 394142c1b124SIlya Dryomov else 394242c1b124SIlya Dryomov WARN_ON(pi->was_full); 394342c1b124SIlya Dryomov } 394442c1b124SIlya Dryomov 394542c1b124SIlya Dryomov if (osdc->osdmap->epoch && 394642c1b124SIlya Dryomov osdc->osdmap->epoch + 1 < newmap->epoch) { 394742c1b124SIlya Dryomov WARN_ON(incremental); 394842c1b124SIlya Dryomov skipped_map = true; 394942c1b124SIlya Dryomov } 395042c1b124SIlya Dryomov 395142c1b124SIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 395242c1b124SIlya Dryomov osdc->osdmap = newmap; 395342c1b124SIlya Dryomov } 395442c1b124SIlya Dryomov 3955b7ec35b3SIlya Dryomov was_full &= !ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 39565aea3dcdSIlya Dryomov scan_requests(&osdc->homeless_osd, skipped_map, was_full, true, 39575aea3dcdSIlya Dryomov need_resend, need_resend_linger); 39585aea3dcdSIlya Dryomov 39595aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; ) { 39605aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 39615aea3dcdSIlya Dryomov 39625aea3dcdSIlya Dryomov n = rb_next(n); /* close_osd() */ 39635aea3dcdSIlya Dryomov 39645aea3dcdSIlya Dryomov scan_requests(osd, skipped_map, was_full, true, need_resend, 39655aea3dcdSIlya Dryomov need_resend_linger); 39665aea3dcdSIlya Dryomov if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 39675aea3dcdSIlya Dryomov memcmp(&osd->o_con.peer_addr, 39685aea3dcdSIlya Dryomov ceph_osd_addr(osdc->osdmap, osd->o_osd), 39695aea3dcdSIlya Dryomov sizeof(struct ceph_entity_addr))) 39705aea3dcdSIlya Dryomov close_osd(osd); 39715aea3dcdSIlya Dryomov } 397242c1b124SIlya Dryomov 397342c1b124SIlya Dryomov return 0; 397442c1b124SIlya Dryomov } 39756f6c7006SSage Weil 39765aea3dcdSIlya Dryomov static void kick_requests(struct ceph_osd_client *osdc, 39775aea3dcdSIlya Dryomov struct rb_root *need_resend, 39785aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 39795aea3dcdSIlya Dryomov { 3980922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq, *nlreq; 398104c7d789SIlya Dryomov enum calc_target_result ct_res; 39825aea3dcdSIlya Dryomov struct rb_node *n; 39835aea3dcdSIlya Dryomov 398404c7d789SIlya Dryomov /* make sure need_resend targets reflect latest map */ 398504c7d789SIlya Dryomov for (n = rb_first(need_resend); n; ) { 398604c7d789SIlya Dryomov struct ceph_osd_request *req = 398704c7d789SIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 398804c7d789SIlya Dryomov 398904c7d789SIlya Dryomov n = rb_next(n); 399004c7d789SIlya Dryomov 399104c7d789SIlya Dryomov if (req->r_t.epoch < osdc->osdmap->epoch) { 39928edf84baSIlya Dryomov ct_res = calc_target(osdc, &req->r_t, false); 399304c7d789SIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE) { 399404c7d789SIlya Dryomov erase_request(need_resend, req); 399504c7d789SIlya Dryomov check_pool_dne(req); 399604c7d789SIlya Dryomov } 399704c7d789SIlya Dryomov } 399804c7d789SIlya Dryomov } 399904c7d789SIlya Dryomov 40005aea3dcdSIlya Dryomov for (n = rb_first(need_resend); n; ) { 40015aea3dcdSIlya Dryomov struct ceph_osd_request *req = 40025aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 40035aea3dcdSIlya Dryomov struct ceph_osd *osd; 40045aea3dcdSIlya Dryomov 40055aea3dcdSIlya Dryomov n = rb_next(n); 40065aea3dcdSIlya Dryomov erase_request(need_resend, req); /* before link_request() */ 40075aea3dcdSIlya Dryomov 40085aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, true); 40095aea3dcdSIlya Dryomov link_request(osd, req); 40105aea3dcdSIlya Dryomov if (!req->r_linger) { 40115aea3dcdSIlya Dryomov if (!osd_homeless(osd) && !req->r_t.paused) 40125aea3dcdSIlya Dryomov send_request(req); 4013922dab61SIlya Dryomov } else { 4014922dab61SIlya Dryomov cancel_linger_request(req); 40155aea3dcdSIlya Dryomov } 40165aea3dcdSIlya Dryomov } 4017922dab61SIlya Dryomov 4018922dab61SIlya Dryomov list_for_each_entry_safe(lreq, nlreq, need_resend_linger, scan_item) { 4019922dab61SIlya Dryomov if (!osd_homeless(lreq->osd)) 4020922dab61SIlya Dryomov send_linger(lreq); 4021922dab61SIlya Dryomov 4022922dab61SIlya Dryomov list_del_init(&lreq->scan_item); 4023922dab61SIlya Dryomov } 40245aea3dcdSIlya Dryomov } 40255aea3dcdSIlya Dryomov 40263d14c5d2SYehuda Sadeh /* 40273d14c5d2SYehuda Sadeh * Process updated osd map. 40283d14c5d2SYehuda Sadeh * 40293d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 40303d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 40313d14c5d2SYehuda Sadeh * off to different OSDs as needed. 40323d14c5d2SYehuda Sadeh */ 40333d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 40343d14c5d2SYehuda Sadeh { 403542c1b124SIlya Dryomov void *p = msg->front.iov_base; 403642c1b124SIlya Dryomov void *const end = p + msg->front.iov_len; 40373d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 40383d14c5d2SYehuda Sadeh u32 epoch; 40393d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 40405aea3dcdSIlya Dryomov struct rb_root need_resend = RB_ROOT; 40415aea3dcdSIlya Dryomov LIST_HEAD(need_resend_linger); 404242c1b124SIlya Dryomov bool handled_incremental = false; 404342c1b124SIlya Dryomov bool was_pauserd, was_pausewr; 404442c1b124SIlya Dryomov bool pauserd, pausewr; 404542c1b124SIlya Dryomov int err; 40463d14c5d2SYehuda Sadeh 404742c1b124SIlya Dryomov dout("%s have %u\n", __func__, osdc->osdmap->epoch); 40485aea3dcdSIlya Dryomov down_write(&osdc->lock); 40493d14c5d2SYehuda Sadeh 40503d14c5d2SYehuda Sadeh /* verify fsid */ 40513d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 40523d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 40533d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 405442c1b124SIlya Dryomov goto bad; 40553d14c5d2SYehuda Sadeh 4056b7ec35b3SIlya Dryomov was_pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 4057b7ec35b3SIlya Dryomov was_pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 4058b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 405942c1b124SIlya Dryomov have_pool_full(osdc); 40609a1ea2dbSJosh Durgin 40613d14c5d2SYehuda Sadeh /* incremental maps */ 40623d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 40633d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 40643d14c5d2SYehuda Sadeh while (nr_maps > 0) { 40653d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 40663d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 40673d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 40683d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 406942c1b124SIlya Dryomov if (osdc->osdmap->epoch && 407042c1b124SIlya Dryomov osdc->osdmap->epoch + 1 == epoch) { 40713d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 40723d14c5d2SYehuda Sadeh epoch, maplen); 40735aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, true, 40745aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 407542c1b124SIlya Dryomov if (err) 40763d14c5d2SYehuda Sadeh goto bad; 407742c1b124SIlya Dryomov handled_incremental = true; 40783d14c5d2SYehuda Sadeh } else { 40793d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 40803d14c5d2SYehuda Sadeh epoch, maplen); 40813d14c5d2SYehuda Sadeh } 408242c1b124SIlya Dryomov p += maplen; 40833d14c5d2SYehuda Sadeh nr_maps--; 40843d14c5d2SYehuda Sadeh } 408542c1b124SIlya Dryomov if (handled_incremental) 40863d14c5d2SYehuda Sadeh goto done; 40873d14c5d2SYehuda Sadeh 40883d14c5d2SYehuda Sadeh /* full maps */ 40893d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 40903d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 40913d14c5d2SYehuda Sadeh while (nr_maps) { 40923d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 40933d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 40943d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 40953d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 40963d14c5d2SYehuda Sadeh if (nr_maps > 1) { 40973d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 40983d14c5d2SYehuda Sadeh epoch, maplen); 4099e5253a7bSIlya Dryomov } else if (osdc->osdmap->epoch >= epoch) { 41003d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 41013d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 41023d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 41033d14c5d2SYehuda Sadeh } else { 41043d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 41055aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, false, 41065aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 410742c1b124SIlya Dryomov if (err) 41083d14c5d2SYehuda Sadeh goto bad; 41093d14c5d2SYehuda Sadeh } 41103d14c5d2SYehuda Sadeh p += maplen; 41113d14c5d2SYehuda Sadeh nr_maps--; 41123d14c5d2SYehuda Sadeh } 41133d14c5d2SYehuda Sadeh 41143d14c5d2SYehuda Sadeh done: 4115cd634fb6SSage Weil /* 4116cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 4117cd634fb6SSage Weil * we find out when we are no longer full and stop returning 4118cd634fb6SSage Weil * ENOSPC. 4119cd634fb6SSage Weil */ 4120b7ec35b3SIlya Dryomov pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 4121b7ec35b3SIlya Dryomov pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 4122b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 412342c1b124SIlya Dryomov have_pool_full(osdc); 412458eb7932SJeff Layton if (was_pauserd || was_pausewr || pauserd || pausewr || 412558eb7932SJeff Layton osdc->osdmap->epoch < osdc->epoch_barrier) 412642c1b124SIlya Dryomov maybe_request_map(osdc); 4127cd634fb6SSage Weil 41285aea3dcdSIlya Dryomov kick_requests(osdc, &need_resend, &need_resend_linger); 412942c1b124SIlya Dryomov 4130fc36d0a4SJeff Layton ceph_osdc_abort_on_full(osdc); 413142c1b124SIlya Dryomov ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 413242c1b124SIlya Dryomov osdc->osdmap->epoch); 41335aea3dcdSIlya Dryomov up_write(&osdc->lock); 41343d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 41353d14c5d2SYehuda Sadeh return; 41363d14c5d2SYehuda Sadeh 41373d14c5d2SYehuda Sadeh bad: 41383d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 41393d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 41405aea3dcdSIlya Dryomov up_write(&osdc->lock); 41415aea3dcdSIlya Dryomov } 41425aea3dcdSIlya Dryomov 41435aea3dcdSIlya Dryomov /* 41445aea3dcdSIlya Dryomov * Resubmit requests pending on the given osd. 41455aea3dcdSIlya Dryomov */ 41465aea3dcdSIlya Dryomov static void kick_osd_requests(struct ceph_osd *osd) 41475aea3dcdSIlya Dryomov { 41485aea3dcdSIlya Dryomov struct rb_node *n; 41495aea3dcdSIlya Dryomov 4150a02a946dSIlya Dryomov clear_backoffs(osd); 4151a02a946dSIlya Dryomov 4152922dab61SIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 41535aea3dcdSIlya Dryomov struct ceph_osd_request *req = 41545aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 41555aea3dcdSIlya Dryomov 4156922dab61SIlya Dryomov n = rb_next(n); /* cancel_linger_request() */ 4157922dab61SIlya Dryomov 41585aea3dcdSIlya Dryomov if (!req->r_linger) { 41595aea3dcdSIlya Dryomov if (!req->r_t.paused) 41605aea3dcdSIlya Dryomov send_request(req); 4161922dab61SIlya Dryomov } else { 4162922dab61SIlya Dryomov cancel_linger_request(req); 41635aea3dcdSIlya Dryomov } 41645aea3dcdSIlya Dryomov } 4165922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; n = rb_next(n)) { 4166922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 4167922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 4168922dab61SIlya Dryomov 4169922dab61SIlya Dryomov send_linger(lreq); 4170922dab61SIlya Dryomov } 41715aea3dcdSIlya Dryomov } 41725aea3dcdSIlya Dryomov 41735aea3dcdSIlya Dryomov /* 41745aea3dcdSIlya Dryomov * If the osd connection drops, we need to resubmit all requests. 41755aea3dcdSIlya Dryomov */ 41765aea3dcdSIlya Dryomov static void osd_fault(struct ceph_connection *con) 41775aea3dcdSIlya Dryomov { 41785aea3dcdSIlya Dryomov struct ceph_osd *osd = con->private; 41795aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 41805aea3dcdSIlya Dryomov 41815aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 41825aea3dcdSIlya Dryomov 41835aea3dcdSIlya Dryomov down_write(&osdc->lock); 41845aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 41855aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 41865aea3dcdSIlya Dryomov goto out_unlock; 41875aea3dcdSIlya Dryomov } 41885aea3dcdSIlya Dryomov 41895aea3dcdSIlya Dryomov if (!reopen_osd(osd)) 41905aea3dcdSIlya Dryomov kick_osd_requests(osd); 41915aea3dcdSIlya Dryomov maybe_request_map(osdc); 41925aea3dcdSIlya Dryomov 41935aea3dcdSIlya Dryomov out_unlock: 41945aea3dcdSIlya Dryomov up_write(&osdc->lock); 41953d14c5d2SYehuda Sadeh } 41963d14c5d2SYehuda Sadeh 4197a02a946dSIlya Dryomov struct MOSDBackoff { 4198a02a946dSIlya Dryomov struct ceph_spg spgid; 4199a02a946dSIlya Dryomov u32 map_epoch; 4200a02a946dSIlya Dryomov u8 op; 4201a02a946dSIlya Dryomov u64 id; 4202a02a946dSIlya Dryomov struct ceph_hobject_id *begin; 4203a02a946dSIlya Dryomov struct ceph_hobject_id *end; 4204a02a946dSIlya Dryomov }; 4205a02a946dSIlya Dryomov 4206a02a946dSIlya Dryomov static int decode_MOSDBackoff(const struct ceph_msg *msg, struct MOSDBackoff *m) 4207a02a946dSIlya Dryomov { 4208a02a946dSIlya Dryomov void *p = msg->front.iov_base; 4209a02a946dSIlya Dryomov void *const end = p + msg->front.iov_len; 4210a02a946dSIlya Dryomov u8 struct_v; 4211a02a946dSIlya Dryomov u32 struct_len; 4212a02a946dSIlya Dryomov int ret; 4213a02a946dSIlya Dryomov 4214a02a946dSIlya Dryomov ret = ceph_start_decoding(&p, end, 1, "spg_t", &struct_v, &struct_len); 4215a02a946dSIlya Dryomov if (ret) 4216a02a946dSIlya Dryomov return ret; 4217a02a946dSIlya Dryomov 4218a02a946dSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->spgid.pgid); 4219a02a946dSIlya Dryomov if (ret) 4220a02a946dSIlya Dryomov return ret; 4221a02a946dSIlya Dryomov 4222a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->spgid.shard, e_inval); 4223a02a946dSIlya Dryomov ceph_decode_32_safe(&p, end, m->map_epoch, e_inval); 4224a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->op, e_inval); 4225a02a946dSIlya Dryomov ceph_decode_64_safe(&p, end, m->id, e_inval); 4226a02a946dSIlya Dryomov 4227a02a946dSIlya Dryomov m->begin = kzalloc(sizeof(*m->begin), GFP_NOIO); 4228a02a946dSIlya Dryomov if (!m->begin) 4229a02a946dSIlya Dryomov return -ENOMEM; 4230a02a946dSIlya Dryomov 4231a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->begin); 4232a02a946dSIlya Dryomov if (ret) { 4233a02a946dSIlya Dryomov free_hoid(m->begin); 4234a02a946dSIlya Dryomov return ret; 4235a02a946dSIlya Dryomov } 4236a02a946dSIlya Dryomov 4237a02a946dSIlya Dryomov m->end = kzalloc(sizeof(*m->end), GFP_NOIO); 4238a02a946dSIlya Dryomov if (!m->end) { 4239a02a946dSIlya Dryomov free_hoid(m->begin); 4240a02a946dSIlya Dryomov return -ENOMEM; 4241a02a946dSIlya Dryomov } 4242a02a946dSIlya Dryomov 4243a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->end); 4244a02a946dSIlya Dryomov if (ret) { 4245a02a946dSIlya Dryomov free_hoid(m->begin); 4246a02a946dSIlya Dryomov free_hoid(m->end); 4247a02a946dSIlya Dryomov return ret; 4248a02a946dSIlya Dryomov } 4249a02a946dSIlya Dryomov 4250a02a946dSIlya Dryomov return 0; 4251a02a946dSIlya Dryomov 4252a02a946dSIlya Dryomov e_inval: 4253a02a946dSIlya Dryomov return -EINVAL; 4254a02a946dSIlya Dryomov } 4255a02a946dSIlya Dryomov 4256a02a946dSIlya Dryomov static struct ceph_msg *create_backoff_message( 4257a02a946dSIlya Dryomov const struct ceph_osd_backoff *backoff, 4258a02a946dSIlya Dryomov u32 map_epoch) 4259a02a946dSIlya Dryomov { 4260a02a946dSIlya Dryomov struct ceph_msg *msg; 4261a02a946dSIlya Dryomov void *p, *end; 4262a02a946dSIlya Dryomov int msg_size; 4263a02a946dSIlya Dryomov 4264a02a946dSIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 4265a02a946dSIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 4266a02a946dSIlya Dryomov msg_size += 4 + 1 + 8; /* map_epoch, op, id */ 4267a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4268a02a946dSIlya Dryomov hoid_encoding_size(backoff->begin); 4269a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4270a02a946dSIlya Dryomov hoid_encoding_size(backoff->end); 4271a02a946dSIlya Dryomov 4272a02a946dSIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_BACKOFF, msg_size, GFP_NOIO, true); 4273a02a946dSIlya Dryomov if (!msg) 4274a02a946dSIlya Dryomov return NULL; 4275a02a946dSIlya Dryomov 4276a02a946dSIlya Dryomov p = msg->front.iov_base; 4277a02a946dSIlya Dryomov end = p + msg->front_alloc_len; 4278a02a946dSIlya Dryomov 4279a02a946dSIlya Dryomov encode_spgid(&p, &backoff->spgid); 4280a02a946dSIlya Dryomov ceph_encode_32(&p, map_epoch); 4281a02a946dSIlya Dryomov ceph_encode_8(&p, CEPH_OSD_BACKOFF_OP_ACK_BLOCK); 4282a02a946dSIlya Dryomov ceph_encode_64(&p, backoff->id); 4283a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->begin); 4284a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->end); 4285a02a946dSIlya Dryomov BUG_ON(p != end); 4286a02a946dSIlya Dryomov 4287a02a946dSIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 4288a02a946dSIlya Dryomov msg->hdr.version = cpu_to_le16(1); /* MOSDBackoff v1 */ 4289a02a946dSIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 4290a02a946dSIlya Dryomov 4291a02a946dSIlya Dryomov return msg; 4292a02a946dSIlya Dryomov } 4293a02a946dSIlya Dryomov 4294a02a946dSIlya Dryomov static void handle_backoff_block(struct ceph_osd *osd, struct MOSDBackoff *m) 4295a02a946dSIlya Dryomov { 4296a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4297a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4298a02a946dSIlya Dryomov struct ceph_msg *msg; 4299a02a946dSIlya Dryomov 4300a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4301a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4302a02a946dSIlya Dryomov 4303a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &m->spgid); 4304a02a946dSIlya Dryomov if (!spg) { 4305a02a946dSIlya Dryomov spg = alloc_spg_mapping(); 4306a02a946dSIlya Dryomov if (!spg) { 4307a02a946dSIlya Dryomov pr_err("%s failed to allocate spg\n", __func__); 4308a02a946dSIlya Dryomov return; 4309a02a946dSIlya Dryomov } 4310a02a946dSIlya Dryomov spg->spgid = m->spgid; /* struct */ 4311a02a946dSIlya Dryomov insert_spg_mapping(&osd->o_backoff_mappings, spg); 4312a02a946dSIlya Dryomov } 4313a02a946dSIlya Dryomov 4314a02a946dSIlya Dryomov backoff = alloc_backoff(); 4315a02a946dSIlya Dryomov if (!backoff) { 4316a02a946dSIlya Dryomov pr_err("%s failed to allocate backoff\n", __func__); 4317a02a946dSIlya Dryomov return; 4318a02a946dSIlya Dryomov } 4319a02a946dSIlya Dryomov backoff->spgid = m->spgid; /* struct */ 4320a02a946dSIlya Dryomov backoff->id = m->id; 4321a02a946dSIlya Dryomov backoff->begin = m->begin; 4322a02a946dSIlya Dryomov m->begin = NULL; /* backoff now owns this */ 4323a02a946dSIlya Dryomov backoff->end = m->end; 4324a02a946dSIlya Dryomov m->end = NULL; /* ditto */ 4325a02a946dSIlya Dryomov 4326a02a946dSIlya Dryomov insert_backoff(&spg->backoffs, backoff); 4327a02a946dSIlya Dryomov insert_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4328a02a946dSIlya Dryomov 4329a02a946dSIlya Dryomov /* 4330a02a946dSIlya Dryomov * Ack with original backoff's epoch so that the OSD can 4331a02a946dSIlya Dryomov * discard this if there was a PG split. 4332a02a946dSIlya Dryomov */ 4333a02a946dSIlya Dryomov msg = create_backoff_message(backoff, m->map_epoch); 4334a02a946dSIlya Dryomov if (!msg) { 4335a02a946dSIlya Dryomov pr_err("%s failed to allocate msg\n", __func__); 4336a02a946dSIlya Dryomov return; 4337a02a946dSIlya Dryomov } 4338a02a946dSIlya Dryomov ceph_con_send(&osd->o_con, msg); 4339a02a946dSIlya Dryomov } 4340a02a946dSIlya Dryomov 4341a02a946dSIlya Dryomov static bool target_contained_by(const struct ceph_osd_request_target *t, 4342a02a946dSIlya Dryomov const struct ceph_hobject_id *begin, 4343a02a946dSIlya Dryomov const struct ceph_hobject_id *end) 4344a02a946dSIlya Dryomov { 4345a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 4346a02a946dSIlya Dryomov int cmp; 4347a02a946dSIlya Dryomov 4348a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, t); 4349a02a946dSIlya Dryomov cmp = hoid_compare(&hoid, begin); 4350a02a946dSIlya Dryomov return !cmp || (cmp > 0 && hoid_compare(&hoid, end) < 0); 4351a02a946dSIlya Dryomov } 4352a02a946dSIlya Dryomov 4353a02a946dSIlya Dryomov static void handle_backoff_unblock(struct ceph_osd *osd, 4354a02a946dSIlya Dryomov const struct MOSDBackoff *m) 4355a02a946dSIlya Dryomov { 4356a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4357a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4358a02a946dSIlya Dryomov struct rb_node *n; 4359a02a946dSIlya Dryomov 4360a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4361a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4362a02a946dSIlya Dryomov 4363a02a946dSIlya Dryomov backoff = lookup_backoff_by_id(&osd->o_backoffs_by_id, m->id); 4364a02a946dSIlya Dryomov if (!backoff) { 4365a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu backoff dne\n", 4366a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4367a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4368a02a946dSIlya Dryomov return; 4369a02a946dSIlya Dryomov } 4370a02a946dSIlya Dryomov 4371a02a946dSIlya Dryomov if (hoid_compare(backoff->begin, m->begin) && 4372a02a946dSIlya Dryomov hoid_compare(backoff->end, m->end)) { 4373a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu bad range?\n", 4374a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4375a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4376a02a946dSIlya Dryomov /* unblock it anyway... */ 4377a02a946dSIlya Dryomov } 4378a02a946dSIlya Dryomov 4379a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &backoff->spgid); 4380a02a946dSIlya Dryomov BUG_ON(!spg); 4381a02a946dSIlya Dryomov 4382a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 4383a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4384a02a946dSIlya Dryomov free_backoff(backoff); 4385a02a946dSIlya Dryomov 4386a02a946dSIlya Dryomov if (RB_EMPTY_ROOT(&spg->backoffs)) { 4387a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 4388a02a946dSIlya Dryomov free_spg_mapping(spg); 4389a02a946dSIlya Dryomov } 4390a02a946dSIlya Dryomov 4391a02a946dSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 4392a02a946dSIlya Dryomov struct ceph_osd_request *req = 4393a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 4394a02a946dSIlya Dryomov 4395a02a946dSIlya Dryomov if (!ceph_spg_compare(&req->r_t.spgid, &m->spgid)) { 4396a02a946dSIlya Dryomov /* 4397a02a946dSIlya Dryomov * Match against @m, not @backoff -- the PG may 4398a02a946dSIlya Dryomov * have split on the OSD. 4399a02a946dSIlya Dryomov */ 4400a02a946dSIlya Dryomov if (target_contained_by(&req->r_t, m->begin, m->end)) { 4401a02a946dSIlya Dryomov /* 4402a02a946dSIlya Dryomov * If no other installed backoff applies, 4403a02a946dSIlya Dryomov * resend. 4404a02a946dSIlya Dryomov */ 4405a02a946dSIlya Dryomov send_request(req); 4406a02a946dSIlya Dryomov } 4407a02a946dSIlya Dryomov } 4408a02a946dSIlya Dryomov } 4409a02a946dSIlya Dryomov } 4410a02a946dSIlya Dryomov 4411a02a946dSIlya Dryomov static void handle_backoff(struct ceph_osd *osd, struct ceph_msg *msg) 4412a02a946dSIlya Dryomov { 4413a02a946dSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 4414a02a946dSIlya Dryomov struct MOSDBackoff m; 4415a02a946dSIlya Dryomov int ret; 4416a02a946dSIlya Dryomov 4417a02a946dSIlya Dryomov down_read(&osdc->lock); 4418a02a946dSIlya Dryomov if (!osd_registered(osd)) { 4419a02a946dSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 4420a02a946dSIlya Dryomov up_read(&osdc->lock); 4421a02a946dSIlya Dryomov return; 4422a02a946dSIlya Dryomov } 4423a02a946dSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 4424a02a946dSIlya Dryomov 4425a02a946dSIlya Dryomov mutex_lock(&osd->lock); 4426a02a946dSIlya Dryomov ret = decode_MOSDBackoff(msg, &m); 4427a02a946dSIlya Dryomov if (ret) { 4428a02a946dSIlya Dryomov pr_err("failed to decode MOSDBackoff: %d\n", ret); 4429a02a946dSIlya Dryomov ceph_msg_dump(msg); 4430a02a946dSIlya Dryomov goto out_unlock; 4431a02a946dSIlya Dryomov } 4432a02a946dSIlya Dryomov 4433a02a946dSIlya Dryomov switch (m.op) { 4434a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_BLOCK: 4435a02a946dSIlya Dryomov handle_backoff_block(osd, &m); 4436a02a946dSIlya Dryomov break; 4437a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_UNBLOCK: 4438a02a946dSIlya Dryomov handle_backoff_unblock(osd, &m); 4439a02a946dSIlya Dryomov break; 4440a02a946dSIlya Dryomov default: 4441a02a946dSIlya Dryomov pr_err("%s osd%d unknown op %d\n", __func__, osd->o_osd, m.op); 4442a02a946dSIlya Dryomov } 4443a02a946dSIlya Dryomov 4444a02a946dSIlya Dryomov free_hoid(m.begin); 4445a02a946dSIlya Dryomov free_hoid(m.end); 4446a02a946dSIlya Dryomov 4447a02a946dSIlya Dryomov out_unlock: 4448a02a946dSIlya Dryomov mutex_unlock(&osd->lock); 4449a02a946dSIlya Dryomov up_read(&osdc->lock); 4450a02a946dSIlya Dryomov } 4451a02a946dSIlya Dryomov 44523d14c5d2SYehuda Sadeh /* 4453a40c4f10SYehuda Sadeh * Process osd watch notifications 4454a40c4f10SYehuda Sadeh */ 44553c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 44563c663bbdSAlex Elder struct ceph_msg *msg) 4457a40c4f10SYehuda Sadeh { 4458922dab61SIlya Dryomov void *p = msg->front.iov_base; 4459922dab61SIlya Dryomov void *const end = p + msg->front.iov_len; 4460922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4461922dab61SIlya Dryomov struct linger_work *lwork; 4462922dab61SIlya Dryomov u8 proto_ver, opcode; 4463922dab61SIlya Dryomov u64 cookie, notify_id; 4464922dab61SIlya Dryomov u64 notifier_id = 0; 446519079203SIlya Dryomov s32 return_code = 0; 4466922dab61SIlya Dryomov void *payload = NULL; 4467922dab61SIlya Dryomov u32 payload_len = 0; 4468a40c4f10SYehuda Sadeh 4469a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 4470a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 4471a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 4472922dab61SIlya Dryomov p += 8; /* skip ver */ 4473a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 4474a40c4f10SYehuda Sadeh 4475922dab61SIlya Dryomov if (proto_ver >= 1) { 4476922dab61SIlya Dryomov ceph_decode_32_safe(&p, end, payload_len, bad); 4477922dab61SIlya Dryomov ceph_decode_need(&p, end, payload_len, bad); 4478922dab61SIlya Dryomov payload = p; 4479922dab61SIlya Dryomov p += payload_len; 4480a40c4f10SYehuda Sadeh } 4481a40c4f10SYehuda Sadeh 4482922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 2) 448319079203SIlya Dryomov ceph_decode_32_safe(&p, end, return_code, bad); 4484922dab61SIlya Dryomov 4485922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 3) 4486922dab61SIlya Dryomov ceph_decode_64_safe(&p, end, notifier_id, bad); 4487922dab61SIlya Dryomov 4488922dab61SIlya Dryomov down_read(&osdc->lock); 4489922dab61SIlya Dryomov lreq = lookup_linger_osdc(&osdc->linger_requests, cookie); 4490922dab61SIlya Dryomov if (!lreq) { 4491922dab61SIlya Dryomov dout("%s opcode %d cookie %llu dne\n", __func__, opcode, 4492922dab61SIlya Dryomov cookie); 4493922dab61SIlya Dryomov goto out_unlock_osdc; 4494922dab61SIlya Dryomov } 4495922dab61SIlya Dryomov 4496922dab61SIlya Dryomov mutex_lock(&lreq->lock); 449719079203SIlya Dryomov dout("%s opcode %d cookie %llu lreq %p is_watch %d\n", __func__, 449819079203SIlya Dryomov opcode, cookie, lreq, lreq->is_watch); 4499922dab61SIlya Dryomov if (opcode == CEPH_WATCH_EVENT_DISCONNECT) { 4500922dab61SIlya Dryomov if (!lreq->last_error) { 4501922dab61SIlya Dryomov lreq->last_error = -ENOTCONN; 4502922dab61SIlya Dryomov queue_watch_error(lreq); 4503922dab61SIlya Dryomov } 450419079203SIlya Dryomov } else if (!lreq->is_watch) { 450519079203SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY_COMPLETE */ 450619079203SIlya Dryomov if (lreq->notify_id && lreq->notify_id != notify_id) { 450719079203SIlya Dryomov dout("lreq %p notify_id %llu != %llu, ignoring\n", lreq, 450819079203SIlya Dryomov lreq->notify_id, notify_id); 450919079203SIlya Dryomov } else if (!completion_done(&lreq->notify_finish_wait)) { 451019079203SIlya Dryomov struct ceph_msg_data *data = 45110d9c1ab3SIlya Dryomov msg->num_data_items ? &msg->data[0] : NULL; 451219079203SIlya Dryomov 451319079203SIlya Dryomov if (data) { 451419079203SIlya Dryomov if (lreq->preply_pages) { 451519079203SIlya Dryomov WARN_ON(data->type != 451619079203SIlya Dryomov CEPH_MSG_DATA_PAGES); 451719079203SIlya Dryomov *lreq->preply_pages = data->pages; 451819079203SIlya Dryomov *lreq->preply_len = data->length; 4519e8862740SIlya Dryomov data->own_pages = false; 452019079203SIlya Dryomov } 452119079203SIlya Dryomov } 452219079203SIlya Dryomov lreq->notify_finish_error = return_code; 452319079203SIlya Dryomov complete_all(&lreq->notify_finish_wait); 452419079203SIlya Dryomov } 4525922dab61SIlya Dryomov } else { 4526922dab61SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY */ 4527922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_notify); 4528922dab61SIlya Dryomov if (!lwork) { 4529922dab61SIlya Dryomov pr_err("failed to allocate notify-lwork\n"); 4530922dab61SIlya Dryomov goto out_unlock_lreq; 4531922dab61SIlya Dryomov } 4532922dab61SIlya Dryomov 4533922dab61SIlya Dryomov lwork->notify.notify_id = notify_id; 4534922dab61SIlya Dryomov lwork->notify.notifier_id = notifier_id; 4535922dab61SIlya Dryomov lwork->notify.payload = payload; 4536922dab61SIlya Dryomov lwork->notify.payload_len = payload_len; 4537922dab61SIlya Dryomov lwork->notify.msg = ceph_msg_get(msg); 4538922dab61SIlya Dryomov lwork_queue(lwork); 4539922dab61SIlya Dryomov } 4540922dab61SIlya Dryomov 4541922dab61SIlya Dryomov out_unlock_lreq: 4542922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 4543922dab61SIlya Dryomov out_unlock_osdc: 4544922dab61SIlya Dryomov up_read(&osdc->lock); 4545a40c4f10SYehuda Sadeh return; 4546a40c4f10SYehuda Sadeh 4547a40c4f10SYehuda Sadeh bad: 4548a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 4549a40c4f10SYehuda Sadeh } 4550a40c4f10SYehuda Sadeh 4551a40c4f10SYehuda Sadeh /* 45523d14c5d2SYehuda Sadeh * Register request, send initial attempt. 45533d14c5d2SYehuda Sadeh */ 45543d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 45553d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 45563d14c5d2SYehuda Sadeh bool nofail) 45573d14c5d2SYehuda Sadeh { 45585aea3dcdSIlya Dryomov down_read(&osdc->lock); 45595aea3dcdSIlya Dryomov submit_request(req, false); 45605aea3dcdSIlya Dryomov up_read(&osdc->lock); 45613d14c5d2SYehuda Sadeh 45625aea3dcdSIlya Dryomov return 0; 45633d14c5d2SYehuda Sadeh } 45643d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 45653d14c5d2SYehuda Sadeh 45663d14c5d2SYehuda Sadeh /* 4567c297eb42SIlya Dryomov * Unregister a registered request. The request is not completed: 4568c297eb42SIlya Dryomov * ->r_result isn't set and __complete_request() isn't called. 4569c9f9b93dSIlya Dryomov */ 4570c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 4571c9f9b93dSIlya Dryomov { 4572c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 4573c9f9b93dSIlya Dryomov 45745aea3dcdSIlya Dryomov down_write(&osdc->lock); 45755aea3dcdSIlya Dryomov if (req->r_osd) 45765aea3dcdSIlya Dryomov cancel_request(req); 45775aea3dcdSIlya Dryomov up_write(&osdc->lock); 4578c9f9b93dSIlya Dryomov } 4579c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 4580c9f9b93dSIlya Dryomov 4581c9f9b93dSIlya Dryomov /* 458242b06965SIlya Dryomov * @timeout: in jiffies, 0 means "wait forever" 458342b06965SIlya Dryomov */ 458442b06965SIlya Dryomov static int wait_request_timeout(struct ceph_osd_request *req, 458542b06965SIlya Dryomov unsigned long timeout) 458642b06965SIlya Dryomov { 458742b06965SIlya Dryomov long left; 458842b06965SIlya Dryomov 458942b06965SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 45900e76abf2SYan, Zheng left = wait_for_completion_killable_timeout(&req->r_completion, 459142b06965SIlya Dryomov ceph_timeout_jiffies(timeout)); 459242b06965SIlya Dryomov if (left <= 0) { 459342b06965SIlya Dryomov left = left ?: -ETIMEDOUT; 459442b06965SIlya Dryomov ceph_osdc_cancel_request(req); 459542b06965SIlya Dryomov } else { 459642b06965SIlya Dryomov left = req->r_result; /* completed */ 459742b06965SIlya Dryomov } 459842b06965SIlya Dryomov 459942b06965SIlya Dryomov return left; 460042b06965SIlya Dryomov } 460142b06965SIlya Dryomov 460242b06965SIlya Dryomov /* 46033d14c5d2SYehuda Sadeh * wait for a request to complete 46043d14c5d2SYehuda Sadeh */ 46053d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 46063d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 46073d14c5d2SYehuda Sadeh { 460842b06965SIlya Dryomov return wait_request_timeout(req, 0); 46093d14c5d2SYehuda Sadeh } 46103d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 46113d14c5d2SYehuda Sadeh 46123d14c5d2SYehuda Sadeh /* 46133d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 46143d14c5d2SYehuda Sadeh */ 46153d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 46163d14c5d2SYehuda Sadeh { 46175aea3dcdSIlya Dryomov struct rb_node *n, *p; 46185aea3dcdSIlya Dryomov u64 last_tid = atomic64_read(&osdc->last_tid); 46193d14c5d2SYehuda Sadeh 46205aea3dcdSIlya Dryomov again: 46215aea3dcdSIlya Dryomov down_read(&osdc->lock); 46225aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 46235aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 46245aea3dcdSIlya Dryomov 46255aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 46265aea3dcdSIlya Dryomov for (p = rb_first(&osd->o_requests); p; p = rb_next(p)) { 46275aea3dcdSIlya Dryomov struct ceph_osd_request *req = 46285aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 46295aea3dcdSIlya Dryomov 46303d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 46313d14c5d2SYehuda Sadeh break; 46323d14c5d2SYehuda Sadeh 46335aea3dcdSIlya Dryomov if (!(req->r_flags & CEPH_OSD_FLAG_WRITE)) 46343d14c5d2SYehuda Sadeh continue; 46353d14c5d2SYehuda Sadeh 46363d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 46375aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 46385aea3dcdSIlya Dryomov up_read(&osdc->lock); 46395aea3dcdSIlya Dryomov dout("%s waiting on req %p tid %llu last_tid %llu\n", 46405aea3dcdSIlya Dryomov __func__, req, req->r_tid, last_tid); 4641b18b9550SIlya Dryomov wait_for_completion(&req->r_completion); 46423d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 46435aea3dcdSIlya Dryomov goto again; 46443d14c5d2SYehuda Sadeh } 46455aea3dcdSIlya Dryomov 46465aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 46475aea3dcdSIlya Dryomov } 46485aea3dcdSIlya Dryomov 46495aea3dcdSIlya Dryomov up_read(&osdc->lock); 46505aea3dcdSIlya Dryomov dout("%s done last_tid %llu\n", __func__, last_tid); 46513d14c5d2SYehuda Sadeh } 46523d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 46533d14c5d2SYehuda Sadeh 4654922dab61SIlya Dryomov static struct ceph_osd_request * 4655922dab61SIlya Dryomov alloc_linger_request(struct ceph_osd_linger_request *lreq) 4656922dab61SIlya Dryomov { 4657922dab61SIlya Dryomov struct ceph_osd_request *req; 4658922dab61SIlya Dryomov 4659922dab61SIlya Dryomov req = ceph_osdc_alloc_request(lreq->osdc, NULL, 1, false, GFP_NOIO); 4660922dab61SIlya Dryomov if (!req) 4661922dab61SIlya Dryomov return NULL; 4662922dab61SIlya Dryomov 4663922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4664922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 4665922dab61SIlya Dryomov return req; 4666922dab61SIlya Dryomov } 4667922dab61SIlya Dryomov 466839e58c34SIlya Dryomov static struct ceph_osd_request * 466939e58c34SIlya Dryomov alloc_watch_request(struct ceph_osd_linger_request *lreq, u8 watch_opcode) 467039e58c34SIlya Dryomov { 467139e58c34SIlya Dryomov struct ceph_osd_request *req; 467239e58c34SIlya Dryomov 467339e58c34SIlya Dryomov req = alloc_linger_request(lreq); 467439e58c34SIlya Dryomov if (!req) 467539e58c34SIlya Dryomov return NULL; 467639e58c34SIlya Dryomov 467739e58c34SIlya Dryomov /* 467839e58c34SIlya Dryomov * Pass 0 for cookie because we don't know it yet, it will be 467939e58c34SIlya Dryomov * filled in by linger_submit(). 468039e58c34SIlya Dryomov */ 468139e58c34SIlya Dryomov osd_req_op_watch_init(req, 0, 0, watch_opcode); 468226f887e0SIlya Dryomov 468326f887e0SIlya Dryomov if (ceph_osdc_alloc_messages(req, GFP_NOIO)) { 468426f887e0SIlya Dryomov ceph_osdc_put_request(req); 468526f887e0SIlya Dryomov return NULL; 468626f887e0SIlya Dryomov } 468726f887e0SIlya Dryomov 468839e58c34SIlya Dryomov return req; 468939e58c34SIlya Dryomov } 469039e58c34SIlya Dryomov 4691922dab61SIlya Dryomov /* 4692922dab61SIlya Dryomov * Returns a handle, caller owns a ref. 4693922dab61SIlya Dryomov */ 4694922dab61SIlya Dryomov struct ceph_osd_linger_request * 4695922dab61SIlya Dryomov ceph_osdc_watch(struct ceph_osd_client *osdc, 4696922dab61SIlya Dryomov struct ceph_object_id *oid, 4697922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4698922dab61SIlya Dryomov rados_watchcb2_t wcb, 4699922dab61SIlya Dryomov rados_watcherrcb_t errcb, 4700922dab61SIlya Dryomov void *data) 4701922dab61SIlya Dryomov { 4702922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4703922dab61SIlya Dryomov int ret; 4704922dab61SIlya Dryomov 4705922dab61SIlya Dryomov lreq = linger_alloc(osdc); 4706922dab61SIlya Dryomov if (!lreq) 4707922dab61SIlya Dryomov return ERR_PTR(-ENOMEM); 4708922dab61SIlya Dryomov 470919079203SIlya Dryomov lreq->is_watch = true; 4710922dab61SIlya Dryomov lreq->wcb = wcb; 4711922dab61SIlya Dryomov lreq->errcb = errcb; 4712922dab61SIlya Dryomov lreq->data = data; 4713b07d3c4bSIlya Dryomov lreq->watch_valid_thru = jiffies; 4714922dab61SIlya Dryomov 4715922dab61SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 4716922dab61SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 471754ea0046SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_WRITE; 4718fac02ddfSArnd Bergmann ktime_get_real_ts64(&lreq->mtime); 4719922dab61SIlya Dryomov 472039e58c34SIlya Dryomov lreq->reg_req = alloc_watch_request(lreq, CEPH_OSD_WATCH_OP_WATCH); 4721922dab61SIlya Dryomov if (!lreq->reg_req) { 4722922dab61SIlya Dryomov ret = -ENOMEM; 4723922dab61SIlya Dryomov goto err_put_lreq; 4724922dab61SIlya Dryomov } 4725922dab61SIlya Dryomov 472639e58c34SIlya Dryomov lreq->ping_req = alloc_watch_request(lreq, CEPH_OSD_WATCH_OP_PING); 4727922dab61SIlya Dryomov if (!lreq->ping_req) { 4728922dab61SIlya Dryomov ret = -ENOMEM; 4729922dab61SIlya Dryomov goto err_put_lreq; 4730922dab61SIlya Dryomov } 4731922dab61SIlya Dryomov 473281c65213SIlya Dryomov linger_submit(lreq); 4733922dab61SIlya Dryomov ret = linger_reg_commit_wait(lreq); 4734922dab61SIlya Dryomov if (ret) { 4735922dab61SIlya Dryomov linger_cancel(lreq); 4736922dab61SIlya Dryomov goto err_put_lreq; 4737922dab61SIlya Dryomov } 4738922dab61SIlya Dryomov 4739922dab61SIlya Dryomov return lreq; 4740922dab61SIlya Dryomov 4741922dab61SIlya Dryomov err_put_lreq: 4742922dab61SIlya Dryomov linger_put(lreq); 4743922dab61SIlya Dryomov return ERR_PTR(ret); 4744922dab61SIlya Dryomov } 4745922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_watch); 4746922dab61SIlya Dryomov 4747922dab61SIlya Dryomov /* 4748922dab61SIlya Dryomov * Releases a ref. 4749922dab61SIlya Dryomov * 4750922dab61SIlya Dryomov * Times out after mount_timeout to preserve rbd unmap behaviour 4751922dab61SIlya Dryomov * introduced in 2894e1d76974 ("rbd: timeout watch teardown on unmap 4752922dab61SIlya Dryomov * with mount_timeout"). 4753922dab61SIlya Dryomov */ 4754922dab61SIlya Dryomov int ceph_osdc_unwatch(struct ceph_osd_client *osdc, 4755922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 4756922dab61SIlya Dryomov { 4757922dab61SIlya Dryomov struct ceph_options *opts = osdc->client->options; 4758922dab61SIlya Dryomov struct ceph_osd_request *req; 4759922dab61SIlya Dryomov int ret; 4760922dab61SIlya Dryomov 4761922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4762922dab61SIlya Dryomov if (!req) 4763922dab61SIlya Dryomov return -ENOMEM; 4764922dab61SIlya Dryomov 4765922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4766922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 476754ea0046SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_WRITE; 4768fac02ddfSArnd Bergmann ktime_get_real_ts64(&req->r_mtime); 4769922dab61SIlya Dryomov osd_req_op_watch_init(req, 0, lreq->linger_id, 4770922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_UNWATCH); 4771922dab61SIlya Dryomov 4772922dab61SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4773922dab61SIlya Dryomov if (ret) 4774922dab61SIlya Dryomov goto out_put_req; 4775922dab61SIlya Dryomov 4776922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4777922dab61SIlya Dryomov linger_cancel(lreq); 4778922dab61SIlya Dryomov linger_put(lreq); 4779922dab61SIlya Dryomov ret = wait_request_timeout(req, opts->mount_timeout); 4780922dab61SIlya Dryomov 4781922dab61SIlya Dryomov out_put_req: 4782922dab61SIlya Dryomov ceph_osdc_put_request(req); 4783922dab61SIlya Dryomov return ret; 4784922dab61SIlya Dryomov } 4785922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_unwatch); 4786922dab61SIlya Dryomov 4787922dab61SIlya Dryomov static int osd_req_op_notify_ack_init(struct ceph_osd_request *req, int which, 4788922dab61SIlya Dryomov u64 notify_id, u64 cookie, void *payload, 47896d54228fSIlya Dryomov u32 payload_len) 4790922dab61SIlya Dryomov { 4791922dab61SIlya Dryomov struct ceph_osd_req_op *op; 4792922dab61SIlya Dryomov struct ceph_pagelist *pl; 4793922dab61SIlya Dryomov int ret; 4794922dab61SIlya Dryomov 4795042f6498SJeff Layton op = osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY_ACK, 0); 4796922dab61SIlya Dryomov 479733165d47SIlya Dryomov pl = ceph_pagelist_alloc(GFP_NOIO); 4798922dab61SIlya Dryomov if (!pl) 4799922dab61SIlya Dryomov return -ENOMEM; 4800922dab61SIlya Dryomov 4801922dab61SIlya Dryomov ret = ceph_pagelist_encode_64(pl, notify_id); 4802922dab61SIlya Dryomov ret |= ceph_pagelist_encode_64(pl, cookie); 4803922dab61SIlya Dryomov if (payload) { 4804922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 4805922dab61SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 4806922dab61SIlya Dryomov } else { 4807922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, 0); 4808922dab61SIlya Dryomov } 4809922dab61SIlya Dryomov if (ret) { 4810922dab61SIlya Dryomov ceph_pagelist_release(pl); 4811922dab61SIlya Dryomov return -ENOMEM; 4812922dab61SIlya Dryomov } 4813922dab61SIlya Dryomov 4814922dab61SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify_ack.request_data, pl); 4815922dab61SIlya Dryomov op->indata_len = pl->length; 4816922dab61SIlya Dryomov return 0; 4817922dab61SIlya Dryomov } 4818922dab61SIlya Dryomov 4819922dab61SIlya Dryomov int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, 4820922dab61SIlya Dryomov struct ceph_object_id *oid, 4821922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4822922dab61SIlya Dryomov u64 notify_id, 4823922dab61SIlya Dryomov u64 cookie, 4824922dab61SIlya Dryomov void *payload, 48256d54228fSIlya Dryomov u32 payload_len) 4826922dab61SIlya Dryomov { 4827922dab61SIlya Dryomov struct ceph_osd_request *req; 4828922dab61SIlya Dryomov int ret; 4829922dab61SIlya Dryomov 4830922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4831922dab61SIlya Dryomov if (!req) 4832922dab61SIlya Dryomov return -ENOMEM; 4833922dab61SIlya Dryomov 4834922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, oid); 4835922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, oloc); 4836922dab61SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_READ; 4837922dab61SIlya Dryomov 483826f887e0SIlya Dryomov ret = osd_req_op_notify_ack_init(req, 0, notify_id, cookie, payload, 483926f887e0SIlya Dryomov payload_len); 4840922dab61SIlya Dryomov if (ret) 4841922dab61SIlya Dryomov goto out_put_req; 4842922dab61SIlya Dryomov 484326f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4844922dab61SIlya Dryomov if (ret) 4845922dab61SIlya Dryomov goto out_put_req; 4846922dab61SIlya Dryomov 4847922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4848922dab61SIlya Dryomov ret = ceph_osdc_wait_request(osdc, req); 4849922dab61SIlya Dryomov 4850922dab61SIlya Dryomov out_put_req: 4851922dab61SIlya Dryomov ceph_osdc_put_request(req); 4852922dab61SIlya Dryomov return ret; 4853922dab61SIlya Dryomov } 4854922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify_ack); 4855922dab61SIlya Dryomov 485619079203SIlya Dryomov static int osd_req_op_notify_init(struct ceph_osd_request *req, int which, 485719079203SIlya Dryomov u64 cookie, u32 prot_ver, u32 timeout, 48586d54228fSIlya Dryomov void *payload, u32 payload_len) 485919079203SIlya Dryomov { 486019079203SIlya Dryomov struct ceph_osd_req_op *op; 486119079203SIlya Dryomov struct ceph_pagelist *pl; 486219079203SIlya Dryomov int ret; 486319079203SIlya Dryomov 4864042f6498SJeff Layton op = osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY, 0); 486519079203SIlya Dryomov op->notify.cookie = cookie; 486619079203SIlya Dryomov 486733165d47SIlya Dryomov pl = ceph_pagelist_alloc(GFP_NOIO); 486819079203SIlya Dryomov if (!pl) 486919079203SIlya Dryomov return -ENOMEM; 487019079203SIlya Dryomov 487119079203SIlya Dryomov ret = ceph_pagelist_encode_32(pl, 1); /* prot_ver */ 487219079203SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, timeout); 487319079203SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 487419079203SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 487519079203SIlya Dryomov if (ret) { 487619079203SIlya Dryomov ceph_pagelist_release(pl); 487719079203SIlya Dryomov return -ENOMEM; 487819079203SIlya Dryomov } 487919079203SIlya Dryomov 488019079203SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify.request_data, pl); 488119079203SIlya Dryomov op->indata_len = pl->length; 488219079203SIlya Dryomov return 0; 488319079203SIlya Dryomov } 488419079203SIlya Dryomov 488519079203SIlya Dryomov /* 488619079203SIlya Dryomov * @timeout: in seconds 488719079203SIlya Dryomov * 488819079203SIlya Dryomov * @preply_{pages,len} are initialized both on success and error. 488919079203SIlya Dryomov * The caller is responsible for: 489019079203SIlya Dryomov * 489119079203SIlya Dryomov * ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len)) 489219079203SIlya Dryomov */ 489319079203SIlya Dryomov int ceph_osdc_notify(struct ceph_osd_client *osdc, 489419079203SIlya Dryomov struct ceph_object_id *oid, 489519079203SIlya Dryomov struct ceph_object_locator *oloc, 489619079203SIlya Dryomov void *payload, 48976d54228fSIlya Dryomov u32 payload_len, 489819079203SIlya Dryomov u32 timeout, 489919079203SIlya Dryomov struct page ***preply_pages, 490019079203SIlya Dryomov size_t *preply_len) 490119079203SIlya Dryomov { 490219079203SIlya Dryomov struct ceph_osd_linger_request *lreq; 490319079203SIlya Dryomov struct page **pages; 490419079203SIlya Dryomov int ret; 490519079203SIlya Dryomov 490619079203SIlya Dryomov WARN_ON(!timeout); 490719079203SIlya Dryomov if (preply_pages) { 490819079203SIlya Dryomov *preply_pages = NULL; 490919079203SIlya Dryomov *preply_len = 0; 491019079203SIlya Dryomov } 491119079203SIlya Dryomov 491219079203SIlya Dryomov lreq = linger_alloc(osdc); 491319079203SIlya Dryomov if (!lreq) 491419079203SIlya Dryomov return -ENOMEM; 491519079203SIlya Dryomov 491619079203SIlya Dryomov lreq->preply_pages = preply_pages; 491719079203SIlya Dryomov lreq->preply_len = preply_len; 491819079203SIlya Dryomov 491919079203SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 492019079203SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 492119079203SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_READ; 492219079203SIlya Dryomov 492319079203SIlya Dryomov lreq->reg_req = alloc_linger_request(lreq); 492419079203SIlya Dryomov if (!lreq->reg_req) { 492519079203SIlya Dryomov ret = -ENOMEM; 492619079203SIlya Dryomov goto out_put_lreq; 492719079203SIlya Dryomov } 492819079203SIlya Dryomov 492981c65213SIlya Dryomov /* 493081c65213SIlya Dryomov * Pass 0 for cookie because we don't know it yet, it will be 493181c65213SIlya Dryomov * filled in by linger_submit(). 493281c65213SIlya Dryomov */ 493381c65213SIlya Dryomov ret = osd_req_op_notify_init(lreq->reg_req, 0, 0, 1, timeout, 493481c65213SIlya Dryomov payload, payload_len); 493581c65213SIlya Dryomov if (ret) 493681c65213SIlya Dryomov goto out_put_lreq; 493781c65213SIlya Dryomov 493819079203SIlya Dryomov /* for notify_id */ 493919079203SIlya Dryomov pages = ceph_alloc_page_vector(1, GFP_NOIO); 494019079203SIlya Dryomov if (IS_ERR(pages)) { 494119079203SIlya Dryomov ret = PTR_ERR(pages); 494219079203SIlya Dryomov goto out_put_lreq; 494319079203SIlya Dryomov } 494419079203SIlya Dryomov ceph_osd_data_pages_init(osd_req_op_data(lreq->reg_req, 0, notify, 494519079203SIlya Dryomov response_data), 494619079203SIlya Dryomov pages, PAGE_SIZE, 0, false, true); 494719079203SIlya Dryomov 494826f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(lreq->reg_req, GFP_NOIO); 494926f887e0SIlya Dryomov if (ret) 495026f887e0SIlya Dryomov goto out_put_lreq; 495126f887e0SIlya Dryomov 495281c65213SIlya Dryomov linger_submit(lreq); 495319079203SIlya Dryomov ret = linger_reg_commit_wait(lreq); 495419079203SIlya Dryomov if (!ret) 495519079203SIlya Dryomov ret = linger_notify_finish_wait(lreq); 495619079203SIlya Dryomov else 495719079203SIlya Dryomov dout("lreq %p failed to initiate notify %d\n", lreq, ret); 495819079203SIlya Dryomov 495919079203SIlya Dryomov linger_cancel(lreq); 496019079203SIlya Dryomov out_put_lreq: 496119079203SIlya Dryomov linger_put(lreq); 496219079203SIlya Dryomov return ret; 496319079203SIlya Dryomov } 496419079203SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify); 496519079203SIlya Dryomov 49663d14c5d2SYehuda Sadeh /* 4967b07d3c4bSIlya Dryomov * Return the number of milliseconds since the watch was last 4968b07d3c4bSIlya Dryomov * confirmed, or an error. If there is an error, the watch is no 4969b07d3c4bSIlya Dryomov * longer valid, and should be destroyed with ceph_osdc_unwatch(). 4970b07d3c4bSIlya Dryomov */ 4971b07d3c4bSIlya Dryomov int ceph_osdc_watch_check(struct ceph_osd_client *osdc, 4972b07d3c4bSIlya Dryomov struct ceph_osd_linger_request *lreq) 4973b07d3c4bSIlya Dryomov { 4974b07d3c4bSIlya Dryomov unsigned long stamp, age; 4975b07d3c4bSIlya Dryomov int ret; 4976b07d3c4bSIlya Dryomov 4977b07d3c4bSIlya Dryomov down_read(&osdc->lock); 4978b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 4979b07d3c4bSIlya Dryomov stamp = lreq->watch_valid_thru; 4980b07d3c4bSIlya Dryomov if (!list_empty(&lreq->pending_lworks)) { 4981b07d3c4bSIlya Dryomov struct linger_work *lwork = 4982b07d3c4bSIlya Dryomov list_first_entry(&lreq->pending_lworks, 4983b07d3c4bSIlya Dryomov struct linger_work, 4984b07d3c4bSIlya Dryomov pending_item); 4985b07d3c4bSIlya Dryomov 4986b07d3c4bSIlya Dryomov if (time_before(lwork->queued_stamp, stamp)) 4987b07d3c4bSIlya Dryomov stamp = lwork->queued_stamp; 4988b07d3c4bSIlya Dryomov } 4989b07d3c4bSIlya Dryomov age = jiffies - stamp; 4990b07d3c4bSIlya Dryomov dout("%s lreq %p linger_id %llu age %lu last_error %d\n", __func__, 4991b07d3c4bSIlya Dryomov lreq, lreq->linger_id, age, lreq->last_error); 4992b07d3c4bSIlya Dryomov /* we are truncating to msecs, so return a safe upper bound */ 4993b07d3c4bSIlya Dryomov ret = lreq->last_error ?: 1 + jiffies_to_msecs(age); 4994b07d3c4bSIlya Dryomov 4995b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 4996b07d3c4bSIlya Dryomov up_read(&osdc->lock); 4997b07d3c4bSIlya Dryomov return ret; 4998b07d3c4bSIlya Dryomov } 4999b07d3c4bSIlya Dryomov 5000a4ed38d7SDouglas Fuller static int decode_watcher(void **p, void *end, struct ceph_watch_item *item) 5001a4ed38d7SDouglas Fuller { 5002a4ed38d7SDouglas Fuller u8 struct_v; 5003a4ed38d7SDouglas Fuller u32 struct_len; 5004a4ed38d7SDouglas Fuller int ret; 5005a4ed38d7SDouglas Fuller 5006a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 2, "watch_item_t", 5007a4ed38d7SDouglas Fuller &struct_v, &struct_len); 5008a4ed38d7SDouglas Fuller if (ret) 500951fc7ab4SJeff Layton goto bad; 5010a4ed38d7SDouglas Fuller 501151fc7ab4SJeff Layton ret = -EINVAL; 501251fc7ab4SJeff Layton ceph_decode_copy_safe(p, end, &item->name, sizeof(item->name), bad); 501351fc7ab4SJeff Layton ceph_decode_64_safe(p, end, item->cookie, bad); 501451fc7ab4SJeff Layton ceph_decode_skip_32(p, end, bad); /* skip timeout seconds */ 501551fc7ab4SJeff Layton 5016a4ed38d7SDouglas Fuller if (struct_v >= 2) { 501751fc7ab4SJeff Layton ret = ceph_decode_entity_addr(p, end, &item->addr); 501851fc7ab4SJeff Layton if (ret) 501951fc7ab4SJeff Layton goto bad; 502051fc7ab4SJeff Layton } else { 502151fc7ab4SJeff Layton ret = 0; 5022a4ed38d7SDouglas Fuller } 5023a4ed38d7SDouglas Fuller 5024a4ed38d7SDouglas Fuller dout("%s %s%llu cookie %llu addr %s\n", __func__, 5025a4ed38d7SDouglas Fuller ENTITY_NAME(item->name), item->cookie, 5026b726ec97SJeff Layton ceph_pr_addr(&item->addr)); 502751fc7ab4SJeff Layton bad: 502851fc7ab4SJeff Layton return ret; 5029a4ed38d7SDouglas Fuller } 5030a4ed38d7SDouglas Fuller 5031a4ed38d7SDouglas Fuller static int decode_watchers(void **p, void *end, 5032a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 5033a4ed38d7SDouglas Fuller u32 *num_watchers) 5034a4ed38d7SDouglas Fuller { 5035a4ed38d7SDouglas Fuller u8 struct_v; 5036a4ed38d7SDouglas Fuller u32 struct_len; 5037a4ed38d7SDouglas Fuller int i; 5038a4ed38d7SDouglas Fuller int ret; 5039a4ed38d7SDouglas Fuller 5040a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 1, "obj_list_watch_response_t", 5041a4ed38d7SDouglas Fuller &struct_v, &struct_len); 5042a4ed38d7SDouglas Fuller if (ret) 5043a4ed38d7SDouglas Fuller return ret; 5044a4ed38d7SDouglas Fuller 5045a4ed38d7SDouglas Fuller *num_watchers = ceph_decode_32(p); 5046a4ed38d7SDouglas Fuller *watchers = kcalloc(*num_watchers, sizeof(**watchers), GFP_NOIO); 5047a4ed38d7SDouglas Fuller if (!*watchers) 5048a4ed38d7SDouglas Fuller return -ENOMEM; 5049a4ed38d7SDouglas Fuller 5050a4ed38d7SDouglas Fuller for (i = 0; i < *num_watchers; i++) { 5051a4ed38d7SDouglas Fuller ret = decode_watcher(p, end, *watchers + i); 5052a4ed38d7SDouglas Fuller if (ret) { 5053a4ed38d7SDouglas Fuller kfree(*watchers); 5054a4ed38d7SDouglas Fuller return ret; 5055a4ed38d7SDouglas Fuller } 5056a4ed38d7SDouglas Fuller } 5057a4ed38d7SDouglas Fuller 5058a4ed38d7SDouglas Fuller return 0; 5059a4ed38d7SDouglas Fuller } 5060a4ed38d7SDouglas Fuller 5061a4ed38d7SDouglas Fuller /* 5062a4ed38d7SDouglas Fuller * On success, the caller is responsible for: 5063a4ed38d7SDouglas Fuller * 5064a4ed38d7SDouglas Fuller * kfree(watchers); 5065a4ed38d7SDouglas Fuller */ 5066a4ed38d7SDouglas Fuller int ceph_osdc_list_watchers(struct ceph_osd_client *osdc, 5067a4ed38d7SDouglas Fuller struct ceph_object_id *oid, 5068a4ed38d7SDouglas Fuller struct ceph_object_locator *oloc, 5069a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 5070a4ed38d7SDouglas Fuller u32 *num_watchers) 5071a4ed38d7SDouglas Fuller { 5072a4ed38d7SDouglas Fuller struct ceph_osd_request *req; 5073a4ed38d7SDouglas Fuller struct page **pages; 5074a4ed38d7SDouglas Fuller int ret; 5075a4ed38d7SDouglas Fuller 5076a4ed38d7SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 5077a4ed38d7SDouglas Fuller if (!req) 5078a4ed38d7SDouglas Fuller return -ENOMEM; 5079a4ed38d7SDouglas Fuller 5080a4ed38d7SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 5081a4ed38d7SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 5082a4ed38d7SDouglas Fuller req->r_flags = CEPH_OSD_FLAG_READ; 5083a4ed38d7SDouglas Fuller 5084a4ed38d7SDouglas Fuller pages = ceph_alloc_page_vector(1, GFP_NOIO); 5085a4ed38d7SDouglas Fuller if (IS_ERR(pages)) { 5086a4ed38d7SDouglas Fuller ret = PTR_ERR(pages); 5087a4ed38d7SDouglas Fuller goto out_put_req; 5088a4ed38d7SDouglas Fuller } 5089a4ed38d7SDouglas Fuller 5090a4ed38d7SDouglas Fuller osd_req_op_init(req, 0, CEPH_OSD_OP_LIST_WATCHERS, 0); 5091a4ed38d7SDouglas Fuller ceph_osd_data_pages_init(osd_req_op_data(req, 0, list_watchers, 5092a4ed38d7SDouglas Fuller response_data), 5093a4ed38d7SDouglas Fuller pages, PAGE_SIZE, 0, false, true); 5094a4ed38d7SDouglas Fuller 509526f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 509626f887e0SIlya Dryomov if (ret) 509726f887e0SIlya Dryomov goto out_put_req; 509826f887e0SIlya Dryomov 5099a4ed38d7SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 5100a4ed38d7SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 5101a4ed38d7SDouglas Fuller if (ret >= 0) { 5102a4ed38d7SDouglas Fuller void *p = page_address(pages[0]); 5103a4ed38d7SDouglas Fuller void *const end = p + req->r_ops[0].outdata_len; 5104a4ed38d7SDouglas Fuller 5105a4ed38d7SDouglas Fuller ret = decode_watchers(&p, end, watchers, num_watchers); 5106a4ed38d7SDouglas Fuller } 5107a4ed38d7SDouglas Fuller 5108a4ed38d7SDouglas Fuller out_put_req: 5109a4ed38d7SDouglas Fuller ceph_osdc_put_request(req); 5110a4ed38d7SDouglas Fuller return ret; 5111a4ed38d7SDouglas Fuller } 5112a4ed38d7SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_list_watchers); 5113a4ed38d7SDouglas Fuller 5114b07d3c4bSIlya Dryomov /* 5115dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 5116dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 5117dd935f44SJosh Durgin */ 5118f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 5119dd935f44SJosh Durgin { 512099d16943SIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 5121dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 5122dd935f44SJosh Durgin } 5123dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 5124dd935f44SJosh Durgin 51257cca78c9SIlya Dryomov void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc) 51267cca78c9SIlya Dryomov { 51277cca78c9SIlya Dryomov down_read(&osdc->lock); 51287cca78c9SIlya Dryomov maybe_request_map(osdc); 51297cca78c9SIlya Dryomov up_read(&osdc->lock); 51307cca78c9SIlya Dryomov } 51317cca78c9SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_maybe_request_map); 5132dd935f44SJosh Durgin 5133dd935f44SJosh Durgin /* 5134428a7158SDouglas Fuller * Execute an OSD class method on an object. 5135428a7158SDouglas Fuller * 5136428a7158SDouglas Fuller * @flags: CEPH_OSD_FLAG_* 51372544a020SIlya Dryomov * @resp_len: in/out param for reply length 5138428a7158SDouglas Fuller */ 5139428a7158SDouglas Fuller int ceph_osdc_call(struct ceph_osd_client *osdc, 5140428a7158SDouglas Fuller struct ceph_object_id *oid, 5141428a7158SDouglas Fuller struct ceph_object_locator *oloc, 5142428a7158SDouglas Fuller const char *class, const char *method, 5143428a7158SDouglas Fuller unsigned int flags, 5144428a7158SDouglas Fuller struct page *req_page, size_t req_len, 514568ada915SIlya Dryomov struct page **resp_pages, size_t *resp_len) 5146428a7158SDouglas Fuller { 5147428a7158SDouglas Fuller struct ceph_osd_request *req; 5148428a7158SDouglas Fuller int ret; 5149428a7158SDouglas Fuller 515068ada915SIlya Dryomov if (req_len > PAGE_SIZE) 51512544a020SIlya Dryomov return -E2BIG; 51522544a020SIlya Dryomov 5153428a7158SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 5154428a7158SDouglas Fuller if (!req) 5155428a7158SDouglas Fuller return -ENOMEM; 5156428a7158SDouglas Fuller 5157428a7158SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 5158428a7158SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 5159428a7158SDouglas Fuller req->r_flags = flags; 5160428a7158SDouglas Fuller 516124639ce5SIlya Dryomov ret = osd_req_op_cls_init(req, 0, class, method); 5162fe943d50SChengguang Xu if (ret) 5163fe943d50SChengguang Xu goto out_put_req; 5164fe943d50SChengguang Xu 5165428a7158SDouglas Fuller if (req_page) 5166428a7158SDouglas Fuller osd_req_op_cls_request_data_pages(req, 0, &req_page, req_len, 5167428a7158SDouglas Fuller 0, false, false); 516868ada915SIlya Dryomov if (resp_pages) 516968ada915SIlya Dryomov osd_req_op_cls_response_data_pages(req, 0, resp_pages, 51702544a020SIlya Dryomov *resp_len, 0, false, false); 5171428a7158SDouglas Fuller 517226f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 517326f887e0SIlya Dryomov if (ret) 517426f887e0SIlya Dryomov goto out_put_req; 517526f887e0SIlya Dryomov 5176428a7158SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 5177428a7158SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 5178428a7158SDouglas Fuller if (ret >= 0) { 5179428a7158SDouglas Fuller ret = req->r_ops[0].rval; 518068ada915SIlya Dryomov if (resp_pages) 5181428a7158SDouglas Fuller *resp_len = req->r_ops[0].outdata_len; 5182428a7158SDouglas Fuller } 5183428a7158SDouglas Fuller 5184428a7158SDouglas Fuller out_put_req: 5185428a7158SDouglas Fuller ceph_osdc_put_request(req); 5186428a7158SDouglas Fuller return ret; 5187428a7158SDouglas Fuller } 5188428a7158SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_call); 5189428a7158SDouglas Fuller 5190428a7158SDouglas Fuller /* 5191120a75eaSYan, Zheng * reset all osd connections 5192120a75eaSYan, Zheng */ 5193120a75eaSYan, Zheng void ceph_osdc_reopen_osds(struct ceph_osd_client *osdc) 5194120a75eaSYan, Zheng { 5195120a75eaSYan, Zheng struct rb_node *n; 5196120a75eaSYan, Zheng 5197120a75eaSYan, Zheng down_write(&osdc->lock); 5198120a75eaSYan, Zheng for (n = rb_first(&osdc->osds); n; ) { 5199120a75eaSYan, Zheng struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 5200120a75eaSYan, Zheng 5201120a75eaSYan, Zheng n = rb_next(n); 5202120a75eaSYan, Zheng if (!reopen_osd(osd)) 5203120a75eaSYan, Zheng kick_osd_requests(osd); 5204120a75eaSYan, Zheng } 5205120a75eaSYan, Zheng up_write(&osdc->lock); 5206120a75eaSYan, Zheng } 5207120a75eaSYan, Zheng 5208120a75eaSYan, Zheng /* 52093d14c5d2SYehuda Sadeh * init, shutdown 52103d14c5d2SYehuda Sadeh */ 52113d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 52123d14c5d2SYehuda Sadeh { 52133d14c5d2SYehuda Sadeh int err; 52143d14c5d2SYehuda Sadeh 52153d14c5d2SYehuda Sadeh dout("init\n"); 52163d14c5d2SYehuda Sadeh osdc->client = client; 52175aea3dcdSIlya Dryomov init_rwsem(&osdc->lock); 52183d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 52193d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 52209dd2845cSIlya Dryomov spin_lock_init(&osdc->osd_lru_lock); 52215aea3dcdSIlya Dryomov osd_init(&osdc->homeless_osd); 52225aea3dcdSIlya Dryomov osdc->homeless_osd.o_osdc = osdc; 52235aea3dcdSIlya Dryomov osdc->homeless_osd.o_osd = CEPH_HOMELESS_OSD; 5224264048afSIlya Dryomov osdc->last_linger_id = CEPH_LINGER_ID_START; 5225922dab61SIlya Dryomov osdc->linger_requests = RB_ROOT; 52264609245eSIlya Dryomov osdc->map_checks = RB_ROOT; 52274609245eSIlya Dryomov osdc->linger_map_checks = RB_ROOT; 52283d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 52293d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 52303d14c5d2SYehuda Sadeh 52313d14c5d2SYehuda Sadeh err = -ENOMEM; 5232e5253a7bSIlya Dryomov osdc->osdmap = ceph_osdmap_alloc(); 5233e5253a7bSIlya Dryomov if (!osdc->osdmap) 5234e5253a7bSIlya Dryomov goto out; 5235e5253a7bSIlya Dryomov 52369e767adbSIlya Dryomov osdc->req_mempool = mempool_create_slab_pool(10, 52379e767adbSIlya Dryomov ceph_osd_request_cache); 52383d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 5239e5253a7bSIlya Dryomov goto out_map; 52403d14c5d2SYehuda Sadeh 5241d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 52420d9c1ab3SIlya Dryomov PAGE_SIZE, CEPH_OSD_SLAB_OPS, 10, "osd_op"); 52433d14c5d2SYehuda Sadeh if (err < 0) 52443d14c5d2SYehuda Sadeh goto out_mempool; 5245d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 52460d9c1ab3SIlya Dryomov PAGE_SIZE, CEPH_OSD_SLAB_OPS, 10, 52470d9c1ab3SIlya Dryomov "osd_op_reply"); 52483d14c5d2SYehuda Sadeh if (err < 0) 52493d14c5d2SYehuda Sadeh goto out_msgpool; 5250a40c4f10SYehuda Sadeh 5251dbcae088SDan Carpenter err = -ENOMEM; 5252a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 5253dbcae088SDan Carpenter if (!osdc->notify_wq) 5254c172ec5cSIlya Dryomov goto out_msgpool_reply; 5255c172ec5cSIlya Dryomov 525688bc1922SIlya Dryomov osdc->completion_wq = create_singlethread_workqueue("ceph-completion"); 525788bc1922SIlya Dryomov if (!osdc->completion_wq) 525888bc1922SIlya Dryomov goto out_notify_wq; 525988bc1922SIlya Dryomov 5260fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 5261fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 5262b37ee1b9SIlya Dryomov schedule_delayed_work(&osdc->osds_timeout_work, 5263b37ee1b9SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 5264b37ee1b9SIlya Dryomov 52653d14c5d2SYehuda Sadeh return 0; 52663d14c5d2SYehuda Sadeh 526788bc1922SIlya Dryomov out_notify_wq: 526888bc1922SIlya Dryomov destroy_workqueue(osdc->notify_wq); 5269c172ec5cSIlya Dryomov out_msgpool_reply: 5270c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 52713d14c5d2SYehuda Sadeh out_msgpool: 52723d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 52733d14c5d2SYehuda Sadeh out_mempool: 52743d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 5275e5253a7bSIlya Dryomov out_map: 5276e5253a7bSIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 52773d14c5d2SYehuda Sadeh out: 52783d14c5d2SYehuda Sadeh return err; 52793d14c5d2SYehuda Sadeh } 52803d14c5d2SYehuda Sadeh 52813d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 52823d14c5d2SYehuda Sadeh { 528388bc1922SIlya Dryomov destroy_workqueue(osdc->completion_wq); 5284a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 52853d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 52863d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 528742a2c09fSIlya Dryomov 52885aea3dcdSIlya Dryomov down_write(&osdc->lock); 528942a2c09fSIlya Dryomov while (!RB_EMPTY_ROOT(&osdc->osds)) { 529042a2c09fSIlya Dryomov struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 529142a2c09fSIlya Dryomov struct ceph_osd, o_node); 52925aea3dcdSIlya Dryomov close_osd(osd); 529342a2c09fSIlya Dryomov } 52945aea3dcdSIlya Dryomov up_write(&osdc->lock); 529502113a0fSElena Reshetova WARN_ON(refcount_read(&osdc->homeless_osd.o_ref) != 1); 52965aea3dcdSIlya Dryomov osd_cleanup(&osdc->homeless_osd); 52975aea3dcdSIlya Dryomov 52985aea3dcdSIlya Dryomov WARN_ON(!list_empty(&osdc->osd_lru)); 5299922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_requests)); 53004609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->map_checks)); 53014609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_map_checks)); 53025aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_requests)); 53035aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_homeless)); 530442a2c09fSIlya Dryomov 53053d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 53063d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 53073d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 53083d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 53093d14c5d2SYehuda Sadeh } 53103d14c5d2SYehuda Sadeh 531123ddf9beSLuis Henriques static int osd_req_op_copy_from_init(struct ceph_osd_request *req, 531223ddf9beSLuis Henriques u64 src_snapid, u64 src_version, 531323ddf9beSLuis Henriques struct ceph_object_id *src_oid, 531423ddf9beSLuis Henriques struct ceph_object_locator *src_oloc, 531523ddf9beSLuis Henriques u32 src_fadvise_flags, 531623ddf9beSLuis Henriques u32 dst_fadvise_flags, 531778beb0ffSLuis Henriques u32 truncate_seq, u64 truncate_size, 531823ddf9beSLuis Henriques u8 copy_from_flags) 531923ddf9beSLuis Henriques { 532023ddf9beSLuis Henriques struct ceph_osd_req_op *op; 532123ddf9beSLuis Henriques struct page **pages; 532223ddf9beSLuis Henriques void *p, *end; 532323ddf9beSLuis Henriques 532423ddf9beSLuis Henriques pages = ceph_alloc_page_vector(1, GFP_KERNEL); 532523ddf9beSLuis Henriques if (IS_ERR(pages)) 532623ddf9beSLuis Henriques return PTR_ERR(pages); 532723ddf9beSLuis Henriques 5328042f6498SJeff Layton op = osd_req_op_init(req, 0, CEPH_OSD_OP_COPY_FROM2, 532978beb0ffSLuis Henriques dst_fadvise_flags); 533023ddf9beSLuis Henriques op->copy_from.snapid = src_snapid; 533123ddf9beSLuis Henriques op->copy_from.src_version = src_version; 533223ddf9beSLuis Henriques op->copy_from.flags = copy_from_flags; 533323ddf9beSLuis Henriques op->copy_from.src_fadvise_flags = src_fadvise_flags; 533423ddf9beSLuis Henriques 533523ddf9beSLuis Henriques p = page_address(pages[0]); 533623ddf9beSLuis Henriques end = p + PAGE_SIZE; 533723ddf9beSLuis Henriques ceph_encode_string(&p, end, src_oid->name, src_oid->name_len); 533823ddf9beSLuis Henriques encode_oloc(&p, end, src_oloc); 533978beb0ffSLuis Henriques ceph_encode_32(&p, truncate_seq); 534078beb0ffSLuis Henriques ceph_encode_64(&p, truncate_size); 534123ddf9beSLuis Henriques op->indata_len = PAGE_SIZE - (end - p); 534223ddf9beSLuis Henriques 534323ddf9beSLuis Henriques ceph_osd_data_pages_init(&op->copy_from.osd_data, pages, 534423ddf9beSLuis Henriques op->indata_len, 0, false, true); 534523ddf9beSLuis Henriques return 0; 534623ddf9beSLuis Henriques } 534723ddf9beSLuis Henriques 534823ddf9beSLuis Henriques int ceph_osdc_copy_from(struct ceph_osd_client *osdc, 534923ddf9beSLuis Henriques u64 src_snapid, u64 src_version, 535023ddf9beSLuis Henriques struct ceph_object_id *src_oid, 535123ddf9beSLuis Henriques struct ceph_object_locator *src_oloc, 535223ddf9beSLuis Henriques u32 src_fadvise_flags, 535323ddf9beSLuis Henriques struct ceph_object_id *dst_oid, 535423ddf9beSLuis Henriques struct ceph_object_locator *dst_oloc, 535523ddf9beSLuis Henriques u32 dst_fadvise_flags, 535678beb0ffSLuis Henriques u32 truncate_seq, u64 truncate_size, 535723ddf9beSLuis Henriques u8 copy_from_flags) 535823ddf9beSLuis Henriques { 535923ddf9beSLuis Henriques struct ceph_osd_request *req; 536023ddf9beSLuis Henriques int ret; 536123ddf9beSLuis Henriques 536223ddf9beSLuis Henriques req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL); 536323ddf9beSLuis Henriques if (!req) 536423ddf9beSLuis Henriques return -ENOMEM; 536523ddf9beSLuis Henriques 536623ddf9beSLuis Henriques req->r_flags = CEPH_OSD_FLAG_WRITE; 536723ddf9beSLuis Henriques 536823ddf9beSLuis Henriques ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc); 536923ddf9beSLuis Henriques ceph_oid_copy(&req->r_t.base_oid, dst_oid); 537023ddf9beSLuis Henriques 537123ddf9beSLuis Henriques ret = osd_req_op_copy_from_init(req, src_snapid, src_version, src_oid, 537223ddf9beSLuis Henriques src_oloc, src_fadvise_flags, 537378beb0ffSLuis Henriques dst_fadvise_flags, truncate_seq, 537478beb0ffSLuis Henriques truncate_size, copy_from_flags); 537523ddf9beSLuis Henriques if (ret) 537623ddf9beSLuis Henriques goto out; 537723ddf9beSLuis Henriques 537823ddf9beSLuis Henriques ret = ceph_osdc_alloc_messages(req, GFP_KERNEL); 537923ddf9beSLuis Henriques if (ret) 538023ddf9beSLuis Henriques goto out; 538123ddf9beSLuis Henriques 538223ddf9beSLuis Henriques ceph_osdc_start_request(osdc, req, false); 538323ddf9beSLuis Henriques ret = ceph_osdc_wait_request(osdc, req); 538423ddf9beSLuis Henriques 538523ddf9beSLuis Henriques out: 538623ddf9beSLuis Henriques ceph_osdc_put_request(req); 538723ddf9beSLuis Henriques return ret; 538823ddf9beSLuis Henriques } 538923ddf9beSLuis Henriques EXPORT_SYMBOL(ceph_osdc_copy_from); 539023ddf9beSLuis Henriques 539157a35dfbSChengguang Xu int __init ceph_osdc_setup(void) 53925522ae0bSAlex Elder { 53933f1af42aSIlya Dryomov size_t size = sizeof(struct ceph_osd_request) + 53943f1af42aSIlya Dryomov CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op); 53953f1af42aSIlya Dryomov 53965522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 53973f1af42aSIlya Dryomov ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size, 53983f1af42aSIlya Dryomov 0, 0, NULL); 53995522ae0bSAlex Elder 54005522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 54015522ae0bSAlex Elder } 54025522ae0bSAlex Elder 54035522ae0bSAlex Elder void ceph_osdc_cleanup(void) 54045522ae0bSAlex Elder { 54055522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 54065522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 54075522ae0bSAlex Elder ceph_osd_request_cache = NULL; 54085522ae0bSAlex Elder } 54095522ae0bSAlex Elder 54103d14c5d2SYehuda Sadeh /* 54113d14c5d2SYehuda Sadeh * handle incoming message 54123d14c5d2SYehuda Sadeh */ 54133d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 54143d14c5d2SYehuda Sadeh { 54153d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 54165aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 54173d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 54183d14c5d2SYehuda Sadeh 54193d14c5d2SYehuda Sadeh switch (type) { 54203d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 54213d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 54223d14c5d2SYehuda Sadeh break; 54233d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 54245aea3dcdSIlya Dryomov handle_reply(osd, msg); 54253d14c5d2SYehuda Sadeh break; 5426a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5427a02a946dSIlya Dryomov handle_backoff(osd, msg); 5428a02a946dSIlya Dryomov break; 5429a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 5430a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 5431a40c4f10SYehuda Sadeh break; 54323d14c5d2SYehuda Sadeh 54333d14c5d2SYehuda Sadeh default: 54343d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 54353d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 54363d14c5d2SYehuda Sadeh } 54375aea3dcdSIlya Dryomov 54383d14c5d2SYehuda Sadeh ceph_msg_put(msg); 54393d14c5d2SYehuda Sadeh } 54403d14c5d2SYehuda Sadeh 54413d14c5d2SYehuda Sadeh /* 5442d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 5443d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 5444d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 54453d14c5d2SYehuda Sadeh */ 54463d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 54473d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 54483d14c5d2SYehuda Sadeh int *skip) 54493d14c5d2SYehuda Sadeh { 54503d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 54513d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 54525aea3dcdSIlya Dryomov struct ceph_msg *m = NULL; 54533d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 54543f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 54553d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 54565aea3dcdSIlya Dryomov u64 tid = le64_to_cpu(hdr->tid); 54573d14c5d2SYehuda Sadeh 54585aea3dcdSIlya Dryomov down_read(&osdc->lock); 54595aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 54605aea3dcdSIlya Dryomov dout("%s osd%d unknown, skipping\n", __func__, osd->o_osd); 54615aea3dcdSIlya Dryomov *skip = 1; 54625aea3dcdSIlya Dryomov goto out_unlock_osdc; 54635aea3dcdSIlya Dryomov } 54645aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(hdr->src.num)); 54655aea3dcdSIlya Dryomov 54665aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 54675aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 54683d14c5d2SYehuda Sadeh if (!req) { 5469cd8140c6SIlya Dryomov dout("%s osd%d tid %llu unknown, skipping\n", __func__, 5470cd8140c6SIlya Dryomov osd->o_osd, tid); 5471d15f9d69SIlya Dryomov *skip = 1; 54725aea3dcdSIlya Dryomov goto out_unlock_session; 54733d14c5d2SYehuda Sadeh } 54743d14c5d2SYehuda Sadeh 54758921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 54763d14c5d2SYehuda Sadeh 5477f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 5478d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 5479d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 5480d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 54813f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 54823f0a4ac5SIlya Dryomov false); 54833d14c5d2SYehuda Sadeh if (!m) 54845aea3dcdSIlya Dryomov goto out_unlock_session; 54853d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 54863d14c5d2SYehuda Sadeh req->r_reply = m; 54873d14c5d2SYehuda Sadeh } 54883d14c5d2SYehuda Sadeh 5489d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 5490d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 5491d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 5492d15f9d69SIlya Dryomov req->r_reply->data_length); 54933d14c5d2SYehuda Sadeh m = NULL; 5494d15f9d69SIlya Dryomov *skip = 1; 54955aea3dcdSIlya Dryomov goto out_unlock_session; 54963d14c5d2SYehuda Sadeh } 5497d15f9d69SIlya Dryomov 5498d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 54993d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 55003d14c5d2SYehuda Sadeh 55015aea3dcdSIlya Dryomov out_unlock_session: 55025aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 55035aea3dcdSIlya Dryomov out_unlock_osdc: 55045aea3dcdSIlya Dryomov up_read(&osdc->lock); 55053d14c5d2SYehuda Sadeh return m; 55063d14c5d2SYehuda Sadeh } 55073d14c5d2SYehuda Sadeh 550819079203SIlya Dryomov static struct ceph_msg *alloc_msg_with_page_vector(struct ceph_msg_header *hdr) 550919079203SIlya Dryomov { 551019079203SIlya Dryomov struct ceph_msg *m; 551119079203SIlya Dryomov int type = le16_to_cpu(hdr->type); 551219079203SIlya Dryomov u32 front_len = le32_to_cpu(hdr->front_len); 551319079203SIlya Dryomov u32 data_len = le32_to_cpu(hdr->data_len); 551419079203SIlya Dryomov 55150d9c1ab3SIlya Dryomov m = ceph_msg_new2(type, front_len, 1, GFP_NOIO, false); 551619079203SIlya Dryomov if (!m) 551719079203SIlya Dryomov return NULL; 551819079203SIlya Dryomov 551919079203SIlya Dryomov if (data_len) { 552019079203SIlya Dryomov struct page **pages; 552119079203SIlya Dryomov 552219079203SIlya Dryomov pages = ceph_alloc_page_vector(calc_pages_for(0, data_len), 552319079203SIlya Dryomov GFP_NOIO); 5524c22e853aSWei Yongjun if (IS_ERR(pages)) { 552519079203SIlya Dryomov ceph_msg_put(m); 552619079203SIlya Dryomov return NULL; 552719079203SIlya Dryomov } 552819079203SIlya Dryomov 5529e8862740SIlya Dryomov ceph_msg_data_add_pages(m, pages, data_len, 0, true); 553019079203SIlya Dryomov } 553119079203SIlya Dryomov 553219079203SIlya Dryomov return m; 553319079203SIlya Dryomov } 553419079203SIlya Dryomov 55353d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 55363d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 55373d14c5d2SYehuda Sadeh int *skip) 55383d14c5d2SYehuda Sadeh { 55393d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 55403d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 55413d14c5d2SYehuda Sadeh 55421c20f2d2SAlex Elder *skip = 0; 55433d14c5d2SYehuda Sadeh switch (type) { 55443d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 5545a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5546a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 554719079203SIlya Dryomov return alloc_msg_with_page_vector(hdr); 55483d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 55493d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 55503d14c5d2SYehuda Sadeh default: 55515aea3dcdSIlya Dryomov pr_warn("%s osd%d unknown msg type %d, skipping\n", __func__, 55525aea3dcdSIlya Dryomov osd->o_osd, type); 55533d14c5d2SYehuda Sadeh *skip = 1; 55543d14c5d2SYehuda Sadeh return NULL; 55553d14c5d2SYehuda Sadeh } 55563d14c5d2SYehuda Sadeh } 55573d14c5d2SYehuda Sadeh 55583d14c5d2SYehuda Sadeh /* 55593d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 55603d14c5d2SYehuda Sadeh */ 55613d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 55623d14c5d2SYehuda Sadeh { 55633d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 55643d14c5d2SYehuda Sadeh if (get_osd(osd)) 55653d14c5d2SYehuda Sadeh return con; 55663d14c5d2SYehuda Sadeh return NULL; 55673d14c5d2SYehuda Sadeh } 55683d14c5d2SYehuda Sadeh 55693d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 55703d14c5d2SYehuda Sadeh { 55713d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 55723d14c5d2SYehuda Sadeh put_osd(osd); 55733d14c5d2SYehuda Sadeh } 55743d14c5d2SYehuda Sadeh 55753d14c5d2SYehuda Sadeh /* 55763d14c5d2SYehuda Sadeh * authentication 55773d14c5d2SYehuda Sadeh */ 5578a3530df3SAlex Elder /* 5579a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 5580a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 5581a3530df3SAlex Elder */ 5582a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 55838f43fb53SAlex Elder int *proto, int force_new) 55843d14c5d2SYehuda Sadeh { 55853d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 55863d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 55873d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 558874f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 55893d14c5d2SYehuda Sadeh 559074f1869fSAlex Elder if (force_new && auth->authorizer) { 55916c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(auth->authorizer); 559274f1869fSAlex Elder auth->authorizer = NULL; 55933d14c5d2SYehuda Sadeh } 559427859f97SSage Weil if (!auth->authorizer) { 559527859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 5596a3530df3SAlex Elder auth); 55973d14c5d2SYehuda Sadeh if (ret) 5598a3530df3SAlex Elder return ERR_PTR(ret); 559927859f97SSage Weil } else { 560027859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 56010bed9b5cSSage Weil auth); 56020bed9b5cSSage Weil if (ret) 56030bed9b5cSSage Weil return ERR_PTR(ret); 56043d14c5d2SYehuda Sadeh } 56053d14c5d2SYehuda Sadeh *proto = ac->protocol; 560674f1869fSAlex Elder 5607a3530df3SAlex Elder return auth; 56083d14c5d2SYehuda Sadeh } 56093d14c5d2SYehuda Sadeh 56106daca13dSIlya Dryomov static int add_authorizer_challenge(struct ceph_connection *con, 56116daca13dSIlya Dryomov void *challenge_buf, int challenge_buf_len) 56126daca13dSIlya Dryomov { 56136daca13dSIlya Dryomov struct ceph_osd *o = con->private; 56146daca13dSIlya Dryomov struct ceph_osd_client *osdc = o->o_osdc; 56156daca13dSIlya Dryomov struct ceph_auth_client *ac = osdc->client->monc.auth; 56166daca13dSIlya Dryomov 56176daca13dSIlya Dryomov return ceph_auth_add_authorizer_challenge(ac, o->o_auth.authorizer, 56186daca13dSIlya Dryomov challenge_buf, challenge_buf_len); 56196daca13dSIlya Dryomov } 56203d14c5d2SYehuda Sadeh 56210dde5848SIlya Dryomov static int verify_authorizer_reply(struct ceph_connection *con) 56223d14c5d2SYehuda Sadeh { 56233d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 56243d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 56253d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 56263d14c5d2SYehuda Sadeh 56270dde5848SIlya Dryomov return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer); 56283d14c5d2SYehuda Sadeh } 56293d14c5d2SYehuda Sadeh 56303d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 56313d14c5d2SYehuda Sadeh { 56323d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 56333d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 56343d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 56353d14c5d2SYehuda Sadeh 563627859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 56373d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 56383d14c5d2SYehuda Sadeh } 56393d14c5d2SYehuda Sadeh 56408cb441c0SIlya Dryomov static void osd_reencode_message(struct ceph_msg *msg) 56418cb441c0SIlya Dryomov { 5642914902afSIlya Dryomov int type = le16_to_cpu(msg->hdr.type); 5643914902afSIlya Dryomov 5644914902afSIlya Dryomov if (type == CEPH_MSG_OSD_OP) 56458cb441c0SIlya Dryomov encode_request_finish(msg); 56468cb441c0SIlya Dryomov } 56478cb441c0SIlya Dryomov 564879dbd1baSIlya Dryomov static int osd_sign_message(struct ceph_msg *msg) 564933d07337SYan, Zheng { 565079dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 565133d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 565279dbd1baSIlya Dryomov 565333d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 565433d07337SYan, Zheng } 565533d07337SYan, Zheng 565679dbd1baSIlya Dryomov static int osd_check_message_signature(struct ceph_msg *msg) 565733d07337SYan, Zheng { 565879dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 565933d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 566079dbd1baSIlya Dryomov 566133d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 566233d07337SYan, Zheng } 566333d07337SYan, Zheng 56643d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 56653d14c5d2SYehuda Sadeh .get = get_osd_con, 56663d14c5d2SYehuda Sadeh .put = put_osd_con, 56673d14c5d2SYehuda Sadeh .dispatch = dispatch, 56683d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 56696daca13dSIlya Dryomov .add_authorizer_challenge = add_authorizer_challenge, 56703d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 56713d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 56723d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 56738cb441c0SIlya Dryomov .reencode_message = osd_reencode_message, 567479dbd1baSIlya Dryomov .sign_message = osd_sign_message, 567579dbd1baSIlya Dryomov .check_message_signature = osd_check_message_signature, 56765aea3dcdSIlya Dryomov .fault = osd_fault, 56773d14c5d2SYehuda Sadeh }; 5678