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 5403d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 5413d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 5421b83bef2SSage Weil unsigned int num_ops, 5433d14c5d2SYehuda Sadeh bool use_mempool, 54454a54007SAlex Elder gfp_t gfp_flags) 5453d14c5d2SYehuda Sadeh { 5463d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5473d14c5d2SYehuda Sadeh 5483d14c5d2SYehuda Sadeh if (use_mempool) { 5493f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_SLAB_OPS); 5503d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 5513f1af42aSIlya Dryomov } else if (num_ops <= CEPH_OSD_SLAB_OPS) { 5523f1af42aSIlya Dryomov req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags); 5533d14c5d2SYehuda Sadeh } else { 5543f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_MAX_OPS); 555acafe7e3SKees Cook req = kmalloc(struct_size(req, r_ops, num_ops), gfp_flags); 5563d14c5d2SYehuda Sadeh } 5573f1af42aSIlya Dryomov if (unlikely(!req)) 5583d14c5d2SYehuda Sadeh return NULL; 5593d14c5d2SYehuda Sadeh 5603540bfdbSIlya Dryomov request_init(req); 5613d14c5d2SYehuda Sadeh req->r_osdc = osdc; 5623d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 56379528734SAlex Elder req->r_num_ops = num_ops; 56484127282SIlya Dryomov req->r_snapid = CEPH_NOSNAP; 56584127282SIlya Dryomov req->r_snapc = ceph_get_snap_context(snapc); 5663d14c5d2SYehuda Sadeh 56713d1ad16SIlya Dryomov dout("%s req %p\n", __func__, req); 56813d1ad16SIlya Dryomov return req; 5693f1af42aSIlya Dryomov } 57013d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_request); 5713f1af42aSIlya Dryomov 5722e59ffd1SIlya Dryomov static int ceph_oloc_encoding_size(const struct ceph_object_locator *oloc) 57330c156d9SYan, Zheng { 57430c156d9SYan, Zheng return 8 + 4 + 4 + 4 + (oloc->pool_ns ? oloc->pool_ns->len : 0); 57530c156d9SYan, Zheng } 57630c156d9SYan, Zheng 5770d9c1ab3SIlya Dryomov static int __ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp, 5780d9c1ab3SIlya Dryomov int num_request_data_items, 5790d9c1ab3SIlya Dryomov int num_reply_data_items) 58013d1ad16SIlya Dryomov { 58113d1ad16SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 58213d1ad16SIlya Dryomov struct ceph_msg *msg; 58313d1ad16SIlya Dryomov int msg_size; 5843d14c5d2SYehuda Sadeh 5850d9c1ab3SIlya Dryomov WARN_ON(req->r_request || req->r_reply); 586d30291b9SIlya Dryomov WARN_ON(ceph_oid_empty(&req->r_base_oid)); 58730c156d9SYan, Zheng WARN_ON(ceph_oloc_empty(&req->r_base_oloc)); 588d30291b9SIlya Dryomov 58913d1ad16SIlya Dryomov /* create request message */ 5908cb441c0SIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 5918cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 5928cb441c0SIlya Dryomov msg_size += 4 + 4 + 4; /* hash, osdmap_epoch, flags */ 5938cb441c0SIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 5948cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid); /* reqid */ 5958cb441c0SIlya Dryomov msg_size += sizeof(struct ceph_blkin_trace_info); /* trace */ 5968cb441c0SIlya Dryomov msg_size += 4 + sizeof(struct ceph_timespec); /* client_inc, mtime */ 59730c156d9SYan, Zheng msg_size += CEPH_ENCODING_START_BLK_LEN + 59830c156d9SYan, Zheng ceph_oloc_encoding_size(&req->r_base_oloc); /* oloc */ 59913d1ad16SIlya Dryomov msg_size += 4 + req->r_base_oid.name_len; /* oid */ 60013d1ad16SIlya Dryomov msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op); 601ae458f5aSIlya Dryomov msg_size += 8; /* snapid */ 602ae458f5aSIlya Dryomov msg_size += 8; /* snap_seq */ 60313d1ad16SIlya Dryomov msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0); 6048cb441c0SIlya Dryomov msg_size += 4 + 8; /* retry_attempt, features */ 605ae458f5aSIlya Dryomov 60613d1ad16SIlya Dryomov if (req->r_mempool) 6070d9c1ab3SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op, msg_size, 6080d9c1ab3SIlya Dryomov num_request_data_items); 6093d14c5d2SYehuda Sadeh else 6100d9c1ab3SIlya Dryomov msg = ceph_msg_new2(CEPH_MSG_OSD_OP, msg_size, 6110d9c1ab3SIlya Dryomov num_request_data_items, gfp, true); 61213d1ad16SIlya Dryomov if (!msg) 61313d1ad16SIlya Dryomov return -ENOMEM; 6143d14c5d2SYehuda Sadeh 6153d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 6163d14c5d2SYehuda Sadeh req->r_request = msg; 6173d14c5d2SYehuda Sadeh 61813d1ad16SIlya Dryomov /* create reply message */ 61913d1ad16SIlya Dryomov msg_size = OSD_OPREPLY_FRONT_LEN; 620711da55dSIlya Dryomov msg_size += req->r_base_oid.name_len; 621711da55dSIlya Dryomov msg_size += req->r_num_ops * sizeof(struct ceph_osd_op); 62213d1ad16SIlya Dryomov 62313d1ad16SIlya Dryomov if (req->r_mempool) 6240d9c1ab3SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op_reply, msg_size, 6250d9c1ab3SIlya Dryomov num_reply_data_items); 62613d1ad16SIlya Dryomov else 6270d9c1ab3SIlya Dryomov msg = ceph_msg_new2(CEPH_MSG_OSD_OPREPLY, msg_size, 6280d9c1ab3SIlya Dryomov num_reply_data_items, gfp, true); 62913d1ad16SIlya Dryomov if (!msg) 63013d1ad16SIlya Dryomov return -ENOMEM; 63113d1ad16SIlya Dryomov 63213d1ad16SIlya Dryomov req->r_reply = msg; 63313d1ad16SIlya Dryomov 63413d1ad16SIlya Dryomov return 0; 63513d1ad16SIlya Dryomov } 6363d14c5d2SYehuda Sadeh 637a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 638a8dd0a37SAlex Elder { 639a8dd0a37SAlex Elder switch (opcode) { 64070b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 64170b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 64270b5bfa3SIlya Dryomov #undef GENERATE_CASE 643a8dd0a37SAlex Elder default: 644a8dd0a37SAlex Elder return false; 645a8dd0a37SAlex Elder } 646a8dd0a37SAlex Elder } 647a8dd0a37SAlex Elder 6480d9c1ab3SIlya Dryomov static void get_num_data_items(struct ceph_osd_request *req, 6490d9c1ab3SIlya Dryomov int *num_request_data_items, 6500d9c1ab3SIlya Dryomov int *num_reply_data_items) 6510d9c1ab3SIlya Dryomov { 6520d9c1ab3SIlya Dryomov struct ceph_osd_req_op *op; 6530d9c1ab3SIlya Dryomov 6540d9c1ab3SIlya Dryomov *num_request_data_items = 0; 6550d9c1ab3SIlya Dryomov *num_reply_data_items = 0; 6560d9c1ab3SIlya Dryomov 6570d9c1ab3SIlya Dryomov for (op = req->r_ops; op != &req->r_ops[req->r_num_ops]; op++) { 6580d9c1ab3SIlya Dryomov switch (op->op) { 6590d9c1ab3SIlya Dryomov /* request */ 6600d9c1ab3SIlya Dryomov case CEPH_OSD_OP_WRITE: 6610d9c1ab3SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 6620d9c1ab3SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 6630d9c1ab3SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 6640d9c1ab3SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 66578beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 6660d9c1ab3SIlya Dryomov *num_request_data_items += 1; 6670d9c1ab3SIlya Dryomov break; 6680d9c1ab3SIlya Dryomov 6690d9c1ab3SIlya Dryomov /* reply */ 6700d9c1ab3SIlya Dryomov case CEPH_OSD_OP_STAT: 6710d9c1ab3SIlya Dryomov case CEPH_OSD_OP_READ: 6720d9c1ab3SIlya Dryomov case CEPH_OSD_OP_LIST_WATCHERS: 6730d9c1ab3SIlya Dryomov *num_reply_data_items += 1; 6740d9c1ab3SIlya Dryomov break; 6750d9c1ab3SIlya Dryomov 6760d9c1ab3SIlya Dryomov /* both */ 6770d9c1ab3SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 6780d9c1ab3SIlya Dryomov *num_request_data_items += 1; 6790d9c1ab3SIlya Dryomov *num_reply_data_items += 1; 6800d9c1ab3SIlya Dryomov break; 6810d9c1ab3SIlya Dryomov case CEPH_OSD_OP_CALL: 6820d9c1ab3SIlya Dryomov *num_request_data_items += 2; 6830d9c1ab3SIlya Dryomov *num_reply_data_items += 1; 6840d9c1ab3SIlya Dryomov break; 6850d9c1ab3SIlya Dryomov 6860d9c1ab3SIlya Dryomov default: 6870d9c1ab3SIlya Dryomov WARN_ON(!osd_req_opcode_valid(op->op)); 6880d9c1ab3SIlya Dryomov break; 6890d9c1ab3SIlya Dryomov } 6900d9c1ab3SIlya Dryomov } 6910d9c1ab3SIlya Dryomov } 6920d9c1ab3SIlya Dryomov 6930d9c1ab3SIlya Dryomov /* 6940d9c1ab3SIlya Dryomov * oid, oloc and OSD op opcode(s) must be filled in before this function 6950d9c1ab3SIlya Dryomov * is called. 6960d9c1ab3SIlya Dryomov */ 6970d9c1ab3SIlya Dryomov int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp) 6980d9c1ab3SIlya Dryomov { 6990d9c1ab3SIlya Dryomov int num_request_data_items, num_reply_data_items; 7000d9c1ab3SIlya Dryomov 7010d9c1ab3SIlya Dryomov get_num_data_items(req, &num_request_data_items, &num_reply_data_items); 7020d9c1ab3SIlya Dryomov return __ceph_osdc_alloc_messages(req, gfp, num_request_data_items, 7030d9c1ab3SIlya Dryomov num_reply_data_items); 7040d9c1ab3SIlya Dryomov } 7050d9c1ab3SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_messages); 7060d9c1ab3SIlya Dryomov 70733803f33SAlex Elder /* 70833803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 70933803f33SAlex Elder * other information associated with them. It also serves as a 71033803f33SAlex Elder * common init routine for all the other init functions, below. 71133803f33SAlex Elder */ 712042f6498SJeff Layton struct ceph_osd_req_op * 713042f6498SJeff Layton osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 714144cba14SYan, Zheng u16 opcode, u32 flags) 71533803f33SAlex Elder { 716c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 717c99d2d4aSAlex Elder 718c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 71933803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 72033803f33SAlex Elder 721c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 72233803f33SAlex Elder memset(op, 0, sizeof (*op)); 72333803f33SAlex Elder op->op = opcode; 724144cba14SYan, Zheng op->flags = flags; 725c99d2d4aSAlex Elder 726c99d2d4aSAlex Elder return op; 72733803f33SAlex Elder } 72849719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 72949719778SAlex Elder 730c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 731c99d2d4aSAlex Elder unsigned int which, u16 opcode, 73233803f33SAlex Elder u64 offset, u64 length, 73333803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 73433803f33SAlex Elder { 735042f6498SJeff Layton struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, 736144cba14SYan, Zheng opcode, 0); 73733803f33SAlex Elder size_t payload_len = 0; 73833803f33SAlex Elder 739ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 740e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 741e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_TRUNCATE); 74233803f33SAlex Elder 74333803f33SAlex Elder op->extent.offset = offset; 74433803f33SAlex Elder op->extent.length = length; 74533803f33SAlex Elder op->extent.truncate_size = truncate_size; 74633803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 747e30b7577SIlya Dryomov if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 74833803f33SAlex Elder payload_len += length; 74933803f33SAlex Elder 750de2aa102SIlya Dryomov op->indata_len = payload_len; 75133803f33SAlex Elder } 75233803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 75333803f33SAlex Elder 754c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 755c99d2d4aSAlex Elder unsigned int which, u64 length) 756e5975c7cSAlex Elder { 757c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 758c99d2d4aSAlex Elder u64 previous; 759c99d2d4aSAlex Elder 760c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 761c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 762c99d2d4aSAlex Elder previous = op->extent.length; 763e5975c7cSAlex Elder 764e5975c7cSAlex Elder if (length == previous) 765e5975c7cSAlex Elder return; /* Nothing to do */ 766e5975c7cSAlex Elder BUG_ON(length > previous); 767e5975c7cSAlex Elder 768e5975c7cSAlex Elder op->extent.length = length; 769d641df81SYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 770de2aa102SIlya Dryomov op->indata_len -= previous - length; 771e5975c7cSAlex Elder } 772e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 773e5975c7cSAlex Elder 7742c63f49aSYan, Zheng void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 7752c63f49aSYan, Zheng unsigned int which, u64 offset_inc) 7762c63f49aSYan, Zheng { 7772c63f49aSYan, Zheng struct ceph_osd_req_op *op, *prev_op; 7782c63f49aSYan, Zheng 7792c63f49aSYan, Zheng BUG_ON(which + 1 >= osd_req->r_num_ops); 7802c63f49aSYan, Zheng 7812c63f49aSYan, Zheng prev_op = &osd_req->r_ops[which]; 782042f6498SJeff Layton op = osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags); 7832c63f49aSYan, Zheng /* dup previous one */ 7842c63f49aSYan, Zheng op->indata_len = prev_op->indata_len; 7852c63f49aSYan, Zheng op->outdata_len = prev_op->outdata_len; 7862c63f49aSYan, Zheng op->extent = prev_op->extent; 7872c63f49aSYan, Zheng /* adjust offset */ 7882c63f49aSYan, Zheng op->extent.offset += offset_inc; 7892c63f49aSYan, Zheng op->extent.length -= offset_inc; 7902c63f49aSYan, Zheng 7912c63f49aSYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 7922c63f49aSYan, Zheng op->indata_len -= offset_inc; 7932c63f49aSYan, Zheng } 7942c63f49aSYan, Zheng EXPORT_SYMBOL(osd_req_op_extent_dup_last); 7952c63f49aSYan, Zheng 796fe943d50SChengguang Xu int osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 79724639ce5SIlya Dryomov const char *class, const char *method) 79833803f33SAlex Elder { 79924639ce5SIlya Dryomov struct ceph_osd_req_op *op; 8005f562df5SAlex Elder struct ceph_pagelist *pagelist; 80133803f33SAlex Elder size_t payload_len = 0; 80233803f33SAlex Elder size_t size; 8034766815bSDavid Disseldorp int ret; 80433803f33SAlex Elder 805042f6498SJeff Layton op = osd_req_op_init(osd_req, which, CEPH_OSD_OP_CALL, 0); 80633803f33SAlex Elder 80733165d47SIlya Dryomov pagelist = ceph_pagelist_alloc(GFP_NOFS); 808fe943d50SChengguang Xu if (!pagelist) 809fe943d50SChengguang Xu return -ENOMEM; 810fe943d50SChengguang Xu 81133803f33SAlex Elder op->cls.class_name = class; 81233803f33SAlex Elder size = strlen(class); 81333803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 81433803f33SAlex Elder op->cls.class_len = size; 8154766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, class, size); 8164766815bSDavid Disseldorp if (ret) 8174766815bSDavid Disseldorp goto err_pagelist_free; 81833803f33SAlex Elder payload_len += size; 81933803f33SAlex Elder 82033803f33SAlex Elder op->cls.method_name = method; 82133803f33SAlex Elder size = strlen(method); 82233803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 82333803f33SAlex Elder op->cls.method_len = size; 8244766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, method, size); 8254766815bSDavid Disseldorp if (ret) 8264766815bSDavid Disseldorp goto err_pagelist_free; 82733803f33SAlex Elder payload_len += size; 82833803f33SAlex Elder 829a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 830de2aa102SIlya Dryomov op->indata_len = payload_len; 831fe943d50SChengguang Xu return 0; 8324766815bSDavid Disseldorp 8334766815bSDavid Disseldorp err_pagelist_free: 8344766815bSDavid Disseldorp ceph_pagelist_release(pagelist); 8354766815bSDavid Disseldorp return ret; 83633803f33SAlex Elder } 83733803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 8388c042b0dSAlex Elder 839d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 840d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 841d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 842d74b50beSYan, Zheng { 843042f6498SJeff Layton struct ceph_osd_req_op *op = osd_req_op_init(osd_req, which, 844144cba14SYan, Zheng opcode, 0); 845d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 846d74b50beSYan, Zheng size_t payload_len; 8474766815bSDavid Disseldorp int ret; 848d74b50beSYan, Zheng 849d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 850d74b50beSYan, Zheng 85133165d47SIlya Dryomov pagelist = ceph_pagelist_alloc(GFP_NOFS); 852d74b50beSYan, Zheng if (!pagelist) 853d74b50beSYan, Zheng return -ENOMEM; 854d74b50beSYan, Zheng 855d74b50beSYan, Zheng payload_len = strlen(name); 856d74b50beSYan, Zheng op->xattr.name_len = payload_len; 8574766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, name, payload_len); 8584766815bSDavid Disseldorp if (ret) 8594766815bSDavid Disseldorp goto err_pagelist_free; 860d74b50beSYan, Zheng 861d74b50beSYan, Zheng op->xattr.value_len = size; 8624766815bSDavid Disseldorp ret = ceph_pagelist_append(pagelist, value, size); 8634766815bSDavid Disseldorp if (ret) 8644766815bSDavid Disseldorp goto err_pagelist_free; 865d74b50beSYan, Zheng payload_len += size; 866d74b50beSYan, Zheng 867d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 868d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 869d74b50beSYan, Zheng 870d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 871de2aa102SIlya Dryomov op->indata_len = payload_len; 872d74b50beSYan, Zheng return 0; 8734766815bSDavid Disseldorp 8744766815bSDavid Disseldorp err_pagelist_free: 8754766815bSDavid Disseldorp ceph_pagelist_release(pagelist); 8764766815bSDavid Disseldorp return ret; 877d74b50beSYan, Zheng } 878d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 879d74b50beSYan, Zheng 880922dab61SIlya Dryomov /* 881922dab61SIlya Dryomov * @watch_opcode: CEPH_OSD_WATCH_OP_* 882922dab61SIlya Dryomov */ 883922dab61SIlya Dryomov static void osd_req_op_watch_init(struct ceph_osd_request *req, int which, 88475dbb685SIlya Dryomov u8 watch_opcode, u64 cookie, u32 gen) 88533803f33SAlex Elder { 886922dab61SIlya Dryomov struct ceph_osd_req_op *op; 88733803f33SAlex Elder 888042f6498SJeff Layton op = osd_req_op_init(req, which, CEPH_OSD_OP_WATCH, 0); 88933803f33SAlex Elder op->watch.cookie = cookie; 890922dab61SIlya Dryomov op->watch.op = watch_opcode; 89175dbb685SIlya Dryomov op->watch.gen = gen; 89275dbb685SIlya Dryomov } 89375dbb685SIlya Dryomov 89475dbb685SIlya Dryomov /* 89575dbb685SIlya Dryomov * prot_ver, timeout and notify payload (may be empty) should already be 89675dbb685SIlya Dryomov * encoded in @request_pl 89775dbb685SIlya Dryomov */ 89875dbb685SIlya Dryomov static void osd_req_op_notify_init(struct ceph_osd_request *req, int which, 89975dbb685SIlya Dryomov u64 cookie, struct ceph_pagelist *request_pl) 90075dbb685SIlya Dryomov { 90175dbb685SIlya Dryomov struct ceph_osd_req_op *op; 90275dbb685SIlya Dryomov 90375dbb685SIlya Dryomov op = osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY, 0); 90475dbb685SIlya Dryomov op->notify.cookie = cookie; 90575dbb685SIlya Dryomov 90675dbb685SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify.request_data, request_pl); 90775dbb685SIlya Dryomov op->indata_len = request_pl->length; 90833803f33SAlex Elder } 90933803f33SAlex Elder 910d3798accSIlya Dryomov /* 911d3798accSIlya Dryomov * @flags: CEPH_OSD_OP_ALLOC_HINT_FLAG_* 912d3798accSIlya Dryomov */ 913c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 914c647b8a8SIlya Dryomov unsigned int which, 915c647b8a8SIlya Dryomov u64 expected_object_size, 916d3798accSIlya Dryomov u64 expected_write_size, 917d3798accSIlya Dryomov u32 flags) 918c647b8a8SIlya Dryomov { 919042f6498SJeff Layton struct ceph_osd_req_op *op; 920c647b8a8SIlya Dryomov 921042f6498SJeff Layton op = osd_req_op_init(osd_req, which, CEPH_OSD_OP_SETALLOCHINT, 0); 922c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 923c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 924d3798accSIlya Dryomov op->alloc_hint.flags = flags; 925c647b8a8SIlya Dryomov 926c647b8a8SIlya Dryomov /* 927c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 928c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 929c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 930c647b8a8SIlya Dryomov */ 931c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 932c647b8a8SIlya Dryomov } 933c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 934c647b8a8SIlya Dryomov 93590af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 936ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 937ec9123c5SAlex Elder { 938ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 939ec9123c5SAlex Elder 940ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 941ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 942ec9123c5SAlex Elder if (length) 94390af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 944e8862740SIlya Dryomov length, osd_data->alignment, false); 945ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 946ec9123c5SAlex Elder BUG_ON(!length); 94790af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 948ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 949ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 9505359a17dSIlya Dryomov ceph_msg_data_add_bio(msg, &osd_data->bio_pos, length); 951ec9123c5SAlex Elder #endif 952b9e281c2SIlya Dryomov } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BVECS) { 953b9e281c2SIlya Dryomov ceph_msg_data_add_bvecs(msg, &osd_data->bvec_pos); 954ec9123c5SAlex Elder } else { 955ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 956ec9123c5SAlex Elder } 957ec9123c5SAlex Elder } 958ec9123c5SAlex Elder 959bb873b53SIlya Dryomov static u32 osd_req_encode_op(struct ceph_osd_op *dst, 960bb873b53SIlya Dryomov const struct ceph_osd_req_op *src) 9613d14c5d2SYehuda Sadeh { 962065a68f9SAlex Elder switch (src->op) { 963fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 964fbfab539SAlex Elder break; 9653d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 9663d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 967e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 968ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 969ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 970175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 971175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 9723d14c5d2SYehuda Sadeh dst->extent.truncate_size = 9733d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 9743d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 9753d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 9763d14c5d2SYehuda Sadeh break; 9773d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 9783d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 9793d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 980bb873b53SIlya Dryomov dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 9813d14c5d2SYehuda Sadeh break; 982a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 983a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 984922dab61SIlya Dryomov dst->watch.ver = cpu_to_le64(0); 985922dab61SIlya Dryomov dst->watch.op = src->watch.op; 986922dab61SIlya Dryomov dst->watch.gen = cpu_to_le32(src->watch.gen); 987922dab61SIlya Dryomov break; 988922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 989a40c4f10SYehuda Sadeh break; 99019079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 99119079203SIlya Dryomov dst->notify.cookie = cpu_to_le64(src->notify.cookie); 99219079203SIlya Dryomov break; 993a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 994a4ed38d7SDouglas Fuller break; 995c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 996c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 997c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 998c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 999c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 1000d3798accSIlya Dryomov dst->alloc_hint.flags = cpu_to_le32(src->alloc_hint.flags); 1001c647b8a8SIlya Dryomov break; 1002d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 1003d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 1004d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 1005d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 1006d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 1007d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 1008d74b50beSYan, Zheng break; 1009864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 1010864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 1011864e9197SYan, Zheng break; 101278beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 101323ddf9beSLuis Henriques dst->copy_from.snapid = cpu_to_le64(src->copy_from.snapid); 101423ddf9beSLuis Henriques dst->copy_from.src_version = 101523ddf9beSLuis Henriques cpu_to_le64(src->copy_from.src_version); 101623ddf9beSLuis Henriques dst->copy_from.flags = src->copy_from.flags; 101723ddf9beSLuis Henriques dst->copy_from.src_fadvise_flags = 101823ddf9beSLuis Henriques cpu_to_le32(src->copy_from.src_fadvise_flags); 101923ddf9beSLuis Henriques break; 10203d14c5d2SYehuda Sadeh default: 10214c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 10228f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 10234c46459cSAlex Elder WARN_ON(1); 1024a8dd0a37SAlex Elder 1025a8dd0a37SAlex Elder return 0; 10263d14c5d2SYehuda Sadeh } 10277b25bf5fSIlya Dryomov 1028a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 10297b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 1030de2aa102SIlya Dryomov dst->payload_len = cpu_to_le32(src->indata_len); 1031175face2SAlex Elder 1032bb873b53SIlya Dryomov return src->indata_len; 10333d14c5d2SYehuda Sadeh } 10343d14c5d2SYehuda Sadeh 10353d14c5d2SYehuda Sadeh /* 10363d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 10373d14c5d2SYehuda Sadeh * extent as needed. 10383d14c5d2SYehuda Sadeh * 10393d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 10403d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 10413d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 10423d14c5d2SYehuda Sadeh */ 10433d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 10443d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 10453d14c5d2SYehuda Sadeh struct ceph_vino vino, 1046715e4cd4SYan, Zheng u64 off, u64 *plen, 1047715e4cd4SYan, Zheng unsigned int which, int num_ops, 10483d14c5d2SYehuda Sadeh int opcode, int flags, 10493d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 10503d14c5d2SYehuda Sadeh u32 truncate_seq, 10513d14c5d2SYehuda Sadeh u64 truncate_size, 1052153e5167SAlex Elder bool use_mempool) 10533d14c5d2SYehuda Sadeh { 10543d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 105575d1c941SAlex Elder u64 objnum = 0; 105675d1c941SAlex Elder u64 objoff = 0; 105775d1c941SAlex Elder u64 objlen = 0; 10586816282dSSage Weil int r; 10593d14c5d2SYehuda Sadeh 1060ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 1061864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 1062864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 10633d14c5d2SYehuda Sadeh 1064acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 1065ae7ca4a3SAlex Elder GFP_NOFS); 106613d1ad16SIlya Dryomov if (!req) { 106713d1ad16SIlya Dryomov r = -ENOMEM; 106813d1ad16SIlya Dryomov goto fail; 106913d1ad16SIlya Dryomov } 107079528734SAlex Elder 10713d14c5d2SYehuda Sadeh /* calculate max write size */ 1072a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 107313d1ad16SIlya Dryomov if (r) 107413d1ad16SIlya Dryomov goto fail; 1075a19dadfbSAlex Elder 1076864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 1077144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 1078864e9197SYan, Zheng } else { 10797627151eSYan, Zheng u32 object_size = layout->object_size; 1080864e9197SYan, Zheng u32 object_base = off - objoff; 1081ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 1082d18d1e28SAlex Elder if (truncate_size <= object_base) { 1083d18d1e28SAlex Elder truncate_size = 0; 1084d18d1e28SAlex Elder } else { 1085d18d1e28SAlex Elder truncate_size -= object_base; 1086d18d1e28SAlex Elder if (truncate_size > object_size) 1087d18d1e28SAlex Elder truncate_size = object_size; 1088d18d1e28SAlex Elder } 1089ccca4e37SYan, Zheng } 1090715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 1091b0270324SAlex Elder truncate_size, truncate_seq); 1092864e9197SYan, Zheng } 1093d18d1e28SAlex Elder 10947627151eSYan, Zheng req->r_base_oloc.pool = layout->pool_id; 109530c156d9SYan, Zheng req->r_base_oloc.pool_ns = ceph_try_get_string(layout->pool_ns); 1096d30291b9SIlya Dryomov ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum); 109722d2cfdfSIlya Dryomov req->r_flags = flags | osdc->client->options->read_from_replica; 1098dbe0fc41SAlex Elder 1099bb873b53SIlya Dryomov req->r_snapid = vino.snap; 1100bb873b53SIlya Dryomov if (flags & CEPH_OSD_FLAG_WRITE) 1101bb873b53SIlya Dryomov req->r_data_offset = off; 1102bb873b53SIlya Dryomov 11030d9c1ab3SIlya Dryomov if (num_ops > 1) 11040d9c1ab3SIlya Dryomov /* 11050d9c1ab3SIlya Dryomov * This is a special case for ceph_writepages_start(), but it 11060d9c1ab3SIlya Dryomov * also covers ceph_uninline_data(). If more multi-op request 11070d9c1ab3SIlya Dryomov * use cases emerge, we will need a separate helper. 11080d9c1ab3SIlya Dryomov */ 11090d9c1ab3SIlya Dryomov r = __ceph_osdc_alloc_messages(req, GFP_NOFS, num_ops, 0); 11100d9c1ab3SIlya Dryomov else 111113d1ad16SIlya Dryomov r = ceph_osdc_alloc_messages(req, GFP_NOFS); 111213d1ad16SIlya Dryomov if (r) 111313d1ad16SIlya Dryomov goto fail; 111413d1ad16SIlya Dryomov 11153d14c5d2SYehuda Sadeh return req; 111613d1ad16SIlya Dryomov 111713d1ad16SIlya Dryomov fail: 111813d1ad16SIlya Dryomov ceph_osdc_put_request(req); 111913d1ad16SIlya Dryomov return ERR_PTR(r); 11203d14c5d2SYehuda Sadeh } 11213d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 11223d14c5d2SYehuda Sadeh 11233d14c5d2SYehuda Sadeh /* 11243d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 11253d14c5d2SYehuda Sadeh */ 1126fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node) 11274609245eSIlya Dryomov DEFINE_RB_FUNCS(request_mc, struct ceph_osd_request, r_tid, r_mc_node) 11283d14c5d2SYehuda Sadeh 112966850df5SIlya Dryomov /* 113066850df5SIlya Dryomov * Call @fn on each OSD request as long as @fn returns 0. 113166850df5SIlya Dryomov */ 113266850df5SIlya Dryomov static void for_each_request(struct ceph_osd_client *osdc, 113366850df5SIlya Dryomov int (*fn)(struct ceph_osd_request *req, void *arg), 113466850df5SIlya Dryomov void *arg) 113566850df5SIlya Dryomov { 113666850df5SIlya Dryomov struct rb_node *n, *p; 113766850df5SIlya Dryomov 113866850df5SIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 113966850df5SIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 114066850df5SIlya Dryomov 114166850df5SIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 114266850df5SIlya Dryomov struct ceph_osd_request *req = 114366850df5SIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 114466850df5SIlya Dryomov 114566850df5SIlya Dryomov p = rb_next(p); 114666850df5SIlya Dryomov if (fn(req, arg)) 114766850df5SIlya Dryomov return; 114866850df5SIlya Dryomov } 114966850df5SIlya Dryomov } 115066850df5SIlya Dryomov 115166850df5SIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 115266850df5SIlya Dryomov struct ceph_osd_request *req = 115366850df5SIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 115466850df5SIlya Dryomov 115566850df5SIlya Dryomov p = rb_next(p); 115666850df5SIlya Dryomov if (fn(req, arg)) 115766850df5SIlya Dryomov return; 115866850df5SIlya Dryomov } 115966850df5SIlya Dryomov } 116066850df5SIlya Dryomov 11610247a0cfSIlya Dryomov static bool osd_homeless(struct ceph_osd *osd) 11620247a0cfSIlya Dryomov { 11630247a0cfSIlya Dryomov return osd->o_osd == CEPH_HOMELESS_OSD; 11640247a0cfSIlya Dryomov } 11650247a0cfSIlya Dryomov 11665aea3dcdSIlya Dryomov static bool osd_registered(struct ceph_osd *osd) 11673d14c5d2SYehuda Sadeh { 11685aea3dcdSIlya Dryomov verify_osdc_locked(osd->o_osdc); 11693d14c5d2SYehuda Sadeh 11705aea3dcdSIlya Dryomov return !RB_EMPTY_NODE(&osd->o_node); 11713d14c5d2SYehuda Sadeh } 11723d14c5d2SYehuda Sadeh 11733d14c5d2SYehuda Sadeh /* 11740247a0cfSIlya Dryomov * Assumes @osd is zero-initialized. 11750247a0cfSIlya Dryomov */ 11760247a0cfSIlya Dryomov static void osd_init(struct ceph_osd *osd) 11770247a0cfSIlya Dryomov { 117802113a0fSElena Reshetova refcount_set(&osd->o_ref, 1); 11790247a0cfSIlya Dryomov RB_CLEAR_NODE(&osd->o_node); 11805aea3dcdSIlya Dryomov osd->o_requests = RB_ROOT; 1181922dab61SIlya Dryomov osd->o_linger_requests = RB_ROOT; 1182a02a946dSIlya Dryomov osd->o_backoff_mappings = RB_ROOT; 1183a02a946dSIlya Dryomov osd->o_backoffs_by_id = RB_ROOT; 11840247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_osd_lru); 11850247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_keepalive_item); 11860247a0cfSIlya Dryomov osd->o_incarnation = 1; 11875aea3dcdSIlya Dryomov mutex_init(&osd->lock); 11880247a0cfSIlya Dryomov } 11890247a0cfSIlya Dryomov 11900247a0cfSIlya Dryomov static void osd_cleanup(struct ceph_osd *osd) 11910247a0cfSIlya Dryomov { 11920247a0cfSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&osd->o_node)); 11935aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 1194922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 1195a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoff_mappings)); 1196a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoffs_by_id)); 11970247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_osd_lru)); 11980247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_keepalive_item)); 11990247a0cfSIlya Dryomov 12000247a0cfSIlya Dryomov if (osd->o_auth.authorizer) { 12010247a0cfSIlya Dryomov WARN_ON(osd_homeless(osd)); 12020247a0cfSIlya Dryomov ceph_auth_destroy_authorizer(osd->o_auth.authorizer); 12030247a0cfSIlya Dryomov } 12040247a0cfSIlya Dryomov } 12050247a0cfSIlya Dryomov 12060247a0cfSIlya Dryomov /* 12073d14c5d2SYehuda Sadeh * Track open sessions with osds. 12083d14c5d2SYehuda Sadeh */ 1209e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 12103d14c5d2SYehuda Sadeh { 12113d14c5d2SYehuda Sadeh struct ceph_osd *osd; 12123d14c5d2SYehuda Sadeh 12130247a0cfSIlya Dryomov WARN_ON(onum == CEPH_HOMELESS_OSD); 12140247a0cfSIlya Dryomov 12157a28f59bSIlya Dryomov osd = kzalloc(sizeof(*osd), GFP_NOIO | __GFP_NOFAIL); 12160247a0cfSIlya Dryomov osd_init(osd); 12173d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1218e10006f8SAlex Elder osd->o_osd = onum; 12193d14c5d2SYehuda Sadeh 1220b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 12213d14c5d2SYehuda Sadeh 12223d14c5d2SYehuda Sadeh return osd; 12233d14c5d2SYehuda Sadeh } 12243d14c5d2SYehuda Sadeh 12253d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 12263d14c5d2SYehuda Sadeh { 122702113a0fSElena Reshetova if (refcount_inc_not_zero(&osd->o_ref)) { 122802113a0fSElena Reshetova dout("get_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref)-1, 122902113a0fSElena Reshetova refcount_read(&osd->o_ref)); 12303d14c5d2SYehuda Sadeh return osd; 12313d14c5d2SYehuda Sadeh } else { 12323d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 12333d14c5d2SYehuda Sadeh return NULL; 12343d14c5d2SYehuda Sadeh } 12353d14c5d2SYehuda Sadeh } 12363d14c5d2SYehuda Sadeh 12373d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 12383d14c5d2SYehuda Sadeh { 123902113a0fSElena Reshetova dout("put_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref), 124002113a0fSElena Reshetova refcount_read(&osd->o_ref) - 1); 124102113a0fSElena Reshetova if (refcount_dec_and_test(&osd->o_ref)) { 12420247a0cfSIlya Dryomov osd_cleanup(osd); 12433d14c5d2SYehuda Sadeh kfree(osd); 12443d14c5d2SYehuda Sadeh } 12453d14c5d2SYehuda Sadeh } 12463d14c5d2SYehuda Sadeh 1247fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node) 1248fcd00b68SIlya Dryomov 12499dd2845cSIlya Dryomov static void __move_osd_to_lru(struct ceph_osd *osd) 12503d14c5d2SYehuda Sadeh { 12519dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 12529dd2845cSIlya Dryomov 12539dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12543d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1255bbf37ec3SIlya Dryomov 12569dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 12573d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 12589dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 12599dd2845cSIlya Dryomov 1260a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 12613d14c5d2SYehuda Sadeh } 12623d14c5d2SYehuda Sadeh 12639dd2845cSIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd *osd) 1264bbf37ec3SIlya Dryomov { 12655aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1266922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) 12679dd2845cSIlya Dryomov __move_osd_to_lru(osd); 1268bbf37ec3SIlya Dryomov } 1269bbf37ec3SIlya Dryomov 12703d14c5d2SYehuda Sadeh static void __remove_osd_from_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); 12759dd2845cSIlya Dryomov 12769dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 12773d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 12783d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 12799dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 12803d14c5d2SYehuda Sadeh } 12813d14c5d2SYehuda Sadeh 12823d14c5d2SYehuda Sadeh /* 12835aea3dcdSIlya Dryomov * Close the connection and assign any leftover requests to the 12845aea3dcdSIlya Dryomov * homeless session. 12855aea3dcdSIlya Dryomov */ 12865aea3dcdSIlya Dryomov static void close_osd(struct ceph_osd *osd) 12875aea3dcdSIlya Dryomov { 12885aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 12895aea3dcdSIlya Dryomov struct rb_node *n; 12905aea3dcdSIlya Dryomov 12915aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 12925aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12935aea3dcdSIlya Dryomov 12945aea3dcdSIlya Dryomov ceph_con_close(&osd->o_con); 12955aea3dcdSIlya Dryomov 12965aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 12975aea3dcdSIlya Dryomov struct ceph_osd_request *req = 12985aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 12995aea3dcdSIlya Dryomov 13005aea3dcdSIlya Dryomov n = rb_next(n); /* unlink_request() */ 13015aea3dcdSIlya Dryomov 13025aea3dcdSIlya Dryomov dout(" reassigning req %p tid %llu\n", req, req->r_tid); 13035aea3dcdSIlya Dryomov unlink_request(osd, req); 13045aea3dcdSIlya Dryomov link_request(&osdc->homeless_osd, req); 13055aea3dcdSIlya Dryomov } 1306922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 1307922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 1308922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 1309922dab61SIlya Dryomov 1310922dab61SIlya Dryomov n = rb_next(n); /* unlink_linger() */ 1311922dab61SIlya Dryomov 1312922dab61SIlya Dryomov dout(" reassigning lreq %p linger_id %llu\n", lreq, 1313922dab61SIlya Dryomov lreq->linger_id); 1314922dab61SIlya Dryomov unlink_linger(osd, lreq); 1315922dab61SIlya Dryomov link_linger(&osdc->homeless_osd, lreq); 1316922dab61SIlya Dryomov } 1317a02a946dSIlya Dryomov clear_backoffs(osd); 13185aea3dcdSIlya Dryomov 13195aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 13205aea3dcdSIlya Dryomov erase_osd(&osdc->osds, osd); 13215aea3dcdSIlya Dryomov put_osd(osd); 13225aea3dcdSIlya Dryomov } 13235aea3dcdSIlya Dryomov 13245aea3dcdSIlya Dryomov /* 13253d14c5d2SYehuda Sadeh * reset osd connect 13263d14c5d2SYehuda Sadeh */ 13275aea3dcdSIlya Dryomov static int reopen_osd(struct ceph_osd *osd) 13283d14c5d2SYehuda Sadeh { 1329c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 13303d14c5d2SYehuda Sadeh 13315aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 13325aea3dcdSIlya Dryomov 13335aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1334922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) { 13355aea3dcdSIlya Dryomov close_osd(osd); 1336c3acb181SAlex Elder return -ENODEV; 1337c3acb181SAlex Elder } 1338c3acb181SAlex Elder 13395aea3dcdSIlya Dryomov peer_addr = &osd->o_osdc->osdmap->osd_addr[osd->o_osd]; 1340c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 13413d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 13425aea3dcdSIlya Dryomov struct rb_node *n; 1343c3acb181SAlex Elder 13443d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 13450b4af2e8SIlya Dryomov "letting msgr retry\n"); 13463d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 13475aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 13485aea3dcdSIlya Dryomov struct ceph_osd_request *req = 13495aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 13503d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 13515aea3dcdSIlya Dryomov } 1352c3acb181SAlex Elder 1353c3acb181SAlex Elder return -EAGAIN; 13543d14c5d2SYehuda Sadeh } 1355c3acb181SAlex Elder 1356c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1357c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1358c3acb181SAlex Elder osd->o_incarnation++; 1359c3acb181SAlex Elder 1360c3acb181SAlex Elder return 0; 13613d14c5d2SYehuda Sadeh } 13623d14c5d2SYehuda Sadeh 13635aea3dcdSIlya Dryomov static struct ceph_osd *lookup_create_osd(struct ceph_osd_client *osdc, int o, 13645aea3dcdSIlya Dryomov bool wrlocked) 13653d14c5d2SYehuda Sadeh { 13665aea3dcdSIlya Dryomov struct ceph_osd *osd; 13675aea3dcdSIlya Dryomov 13685aea3dcdSIlya Dryomov if (wrlocked) 13695aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 13705aea3dcdSIlya Dryomov else 13715aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 13725aea3dcdSIlya Dryomov 13735aea3dcdSIlya Dryomov if (o != CEPH_HOMELESS_OSD) 13745aea3dcdSIlya Dryomov osd = lookup_osd(&osdc->osds, o); 13755aea3dcdSIlya Dryomov else 13765aea3dcdSIlya Dryomov osd = &osdc->homeless_osd; 13775aea3dcdSIlya Dryomov if (!osd) { 13785aea3dcdSIlya Dryomov if (!wrlocked) 13795aea3dcdSIlya Dryomov return ERR_PTR(-EAGAIN); 13805aea3dcdSIlya Dryomov 13815aea3dcdSIlya Dryomov osd = create_osd(osdc, o); 13825aea3dcdSIlya Dryomov insert_osd(&osdc->osds, osd); 13835aea3dcdSIlya Dryomov ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, 13845aea3dcdSIlya Dryomov &osdc->osdmap->osd_addr[osd->o_osd]); 13855aea3dcdSIlya Dryomov } 13865aea3dcdSIlya Dryomov 13875aea3dcdSIlya Dryomov dout("%s osdc %p osd%d -> osd %p\n", __func__, osdc, o, osd); 13885aea3dcdSIlya Dryomov return osd; 1389a40c4f10SYehuda Sadeh } 1390a40c4f10SYehuda Sadeh 13913d14c5d2SYehuda Sadeh /* 13925aea3dcdSIlya Dryomov * Create request <-> OSD session relation. 13935aea3dcdSIlya Dryomov * 13945aea3dcdSIlya Dryomov * @req has to be assigned a tid, @osd may be homeless. 13953d14c5d2SYehuda Sadeh */ 13965aea3dcdSIlya Dryomov static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req) 13973d14c5d2SYehuda Sadeh { 13985aea3dcdSIlya Dryomov verify_osd_locked(osd); 13995aea3dcdSIlya Dryomov WARN_ON(!req->r_tid || req->r_osd); 14005aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 140135f9f8a0SSage Weil req, req->r_tid); 14025aea3dcdSIlya Dryomov 14035aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 14045aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 14055aea3dcdSIlya Dryomov else 14065aea3dcdSIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 14075aea3dcdSIlya Dryomov 14085aea3dcdSIlya Dryomov get_osd(osd); 14095aea3dcdSIlya Dryomov insert_request(&osd->o_requests, req); 14105aea3dcdSIlya Dryomov req->r_osd = osd; 141135f9f8a0SSage Weil } 141235f9f8a0SSage Weil 14135aea3dcdSIlya Dryomov static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req) 14143d14c5d2SYehuda Sadeh { 14155aea3dcdSIlya Dryomov verify_osd_locked(osd); 14165aea3dcdSIlya Dryomov WARN_ON(req->r_osd != osd); 14175aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 14185aea3dcdSIlya Dryomov req, req->r_tid); 14195aea3dcdSIlya Dryomov 14205aea3dcdSIlya Dryomov req->r_osd = NULL; 14215aea3dcdSIlya Dryomov erase_request(&osd->o_requests, req); 14225aea3dcdSIlya Dryomov put_osd(osd); 14235aea3dcdSIlya Dryomov 14245aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 14255aea3dcdSIlya Dryomov maybe_move_osd_to_lru(osd); 14265aea3dcdSIlya Dryomov else 14275aea3dcdSIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 14283d14c5d2SYehuda Sadeh } 14293d14c5d2SYehuda Sadeh 143063244fa1SIlya Dryomov static bool __pool_full(struct ceph_pg_pool_info *pi) 143163244fa1SIlya Dryomov { 143263244fa1SIlya Dryomov return pi->flags & CEPH_POOL_FLAG_FULL; 143363244fa1SIlya Dryomov } 143463244fa1SIlya Dryomov 143542c1b124SIlya Dryomov static bool have_pool_full(struct ceph_osd_client *osdc) 143642c1b124SIlya Dryomov { 143742c1b124SIlya Dryomov struct rb_node *n; 143842c1b124SIlya Dryomov 143942c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 144042c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 144142c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 144242c1b124SIlya Dryomov 144342c1b124SIlya Dryomov if (__pool_full(pi)) 144442c1b124SIlya Dryomov return true; 144542c1b124SIlya Dryomov } 144642c1b124SIlya Dryomov 144742c1b124SIlya Dryomov return false; 144842c1b124SIlya Dryomov } 144942c1b124SIlya Dryomov 14505aea3dcdSIlya Dryomov static bool pool_full(struct ceph_osd_client *osdc, s64 pool_id) 14515aea3dcdSIlya Dryomov { 14525aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 14535aea3dcdSIlya Dryomov 14545aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 14555aea3dcdSIlya Dryomov if (!pi) 14565aea3dcdSIlya Dryomov return false; 14575aea3dcdSIlya Dryomov 14585aea3dcdSIlya Dryomov return __pool_full(pi); 14595aea3dcdSIlya Dryomov } 14605aea3dcdSIlya Dryomov 14613d14c5d2SYehuda Sadeh /* 1462d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1463d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1464d29adb34SJosh Durgin */ 146563244fa1SIlya Dryomov static bool target_should_be_paused(struct ceph_osd_client *osdc, 146663244fa1SIlya Dryomov const struct ceph_osd_request_target *t, 146763244fa1SIlya Dryomov struct ceph_pg_pool_info *pi) 146863244fa1SIlya Dryomov { 1469b7ec35b3SIlya Dryomov bool pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 1470b7ec35b3SIlya Dryomov bool pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 1471b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 147263244fa1SIlya Dryomov __pool_full(pi); 147363244fa1SIlya Dryomov 14746d637a54SIlya Dryomov WARN_ON(pi->id != t->target_oloc.pool); 147558eb7932SJeff Layton return ((t->flags & CEPH_OSD_FLAG_READ) && pauserd) || 147658eb7932SJeff Layton ((t->flags & CEPH_OSD_FLAG_WRITE) && pausewr) || 147758eb7932SJeff Layton (osdc->osdmap->epoch < osdc->epoch_barrier); 147863244fa1SIlya Dryomov } 147963244fa1SIlya Dryomov 1480117d96a0SIlya Dryomov static int pick_random_replica(const struct ceph_osds *acting) 1481117d96a0SIlya Dryomov { 1482117d96a0SIlya Dryomov int i = prandom_u32() % acting->size; 1483117d96a0SIlya Dryomov 1484117d96a0SIlya Dryomov dout("%s picked osd%d, primary osd%d\n", __func__, 1485117d96a0SIlya Dryomov acting->osds[i], acting->primary); 1486117d96a0SIlya Dryomov return i; 1487117d96a0SIlya Dryomov } 1488117d96a0SIlya Dryomov 1489117d96a0SIlya Dryomov /* 1490117d96a0SIlya Dryomov * Picks the closest replica based on client's location given by 1491117d96a0SIlya Dryomov * crush_location option. Prefers the primary if the locality is 1492117d96a0SIlya Dryomov * the same. 1493117d96a0SIlya Dryomov */ 1494117d96a0SIlya Dryomov static int pick_closest_replica(struct ceph_osd_client *osdc, 1495117d96a0SIlya Dryomov const struct ceph_osds *acting) 1496117d96a0SIlya Dryomov { 1497117d96a0SIlya Dryomov struct ceph_options *opt = osdc->client->options; 1498117d96a0SIlya Dryomov int best_i, best_locality; 1499117d96a0SIlya Dryomov int i = 0, locality; 1500117d96a0SIlya Dryomov 1501117d96a0SIlya Dryomov do { 1502117d96a0SIlya Dryomov locality = ceph_get_crush_locality(osdc->osdmap, 1503117d96a0SIlya Dryomov acting->osds[i], 1504117d96a0SIlya Dryomov &opt->crush_locs); 1505117d96a0SIlya Dryomov if (i == 0 || 1506117d96a0SIlya Dryomov (locality >= 0 && best_locality < 0) || 1507117d96a0SIlya Dryomov (locality >= 0 && best_locality >= 0 && 1508117d96a0SIlya Dryomov locality < best_locality)) { 1509117d96a0SIlya Dryomov best_i = i; 1510117d96a0SIlya Dryomov best_locality = locality; 1511117d96a0SIlya Dryomov } 1512117d96a0SIlya Dryomov } while (++i < acting->size); 1513117d96a0SIlya Dryomov 1514117d96a0SIlya Dryomov dout("%s picked osd%d with locality %d, primary osd%d\n", __func__, 1515117d96a0SIlya Dryomov acting->osds[best_i], best_locality, acting->primary); 1516117d96a0SIlya Dryomov return best_i; 1517117d96a0SIlya Dryomov } 1518117d96a0SIlya Dryomov 151963244fa1SIlya Dryomov enum calc_target_result { 152063244fa1SIlya Dryomov CALC_TARGET_NO_ACTION = 0, 152163244fa1SIlya Dryomov CALC_TARGET_NEED_RESEND, 152263244fa1SIlya Dryomov CALC_TARGET_POOL_DNE, 152363244fa1SIlya Dryomov }; 152463244fa1SIlya Dryomov 152563244fa1SIlya Dryomov static enum calc_target_result calc_target(struct ceph_osd_client *osdc, 152663244fa1SIlya Dryomov struct ceph_osd_request_target *t, 152763244fa1SIlya Dryomov bool any_change) 152863244fa1SIlya Dryomov { 152963244fa1SIlya Dryomov struct ceph_pg_pool_info *pi; 153063244fa1SIlya Dryomov struct ceph_pg pgid, last_pgid; 153163244fa1SIlya Dryomov struct ceph_osds up, acting; 1532117d96a0SIlya Dryomov bool is_read = t->flags & CEPH_OSD_FLAG_READ; 1533117d96a0SIlya Dryomov bool is_write = t->flags & CEPH_OSD_FLAG_WRITE; 153463244fa1SIlya Dryomov bool force_resend = false; 153584ed45dfSIlya Dryomov bool unpaused = false; 1536a5613724SIlya Dryomov bool legacy_change = false; 15377de030d6SIlya Dryomov bool split = false; 1538b7ec35b3SIlya Dryomov bool sort_bitwise = ceph_osdmap_flag(osdc, CEPH_OSDMAP_SORTBITWISE); 1539ae78dd81SIlya Dryomov bool recovery_deletes = ceph_osdmap_flag(osdc, 1540ae78dd81SIlya Dryomov CEPH_OSDMAP_RECOVERY_DELETES); 154163244fa1SIlya Dryomov enum calc_target_result ct_res; 154263244fa1SIlya Dryomov 154304c7d789SIlya Dryomov t->epoch = osdc->osdmap->epoch; 154463244fa1SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool); 154563244fa1SIlya Dryomov if (!pi) { 154663244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 154763244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 154863244fa1SIlya Dryomov goto out; 154963244fa1SIlya Dryomov } 155063244fa1SIlya Dryomov 155163244fa1SIlya Dryomov if (osdc->osdmap->epoch == pi->last_force_request_resend) { 1552dc93e0e2SIlya Dryomov if (t->last_force_resend < pi->last_force_request_resend) { 1553dc93e0e2SIlya Dryomov t->last_force_resend = pi->last_force_request_resend; 155463244fa1SIlya Dryomov force_resend = true; 1555dc93e0e2SIlya Dryomov } else if (t->last_force_resend == 0) { 155663244fa1SIlya Dryomov force_resend = true; 155763244fa1SIlya Dryomov } 155863244fa1SIlya Dryomov } 155963244fa1SIlya Dryomov 1560db098ec4SIlya Dryomov /* apply tiering */ 1561db098ec4SIlya Dryomov ceph_oid_copy(&t->target_oid, &t->base_oid); 1562db098ec4SIlya Dryomov ceph_oloc_copy(&t->target_oloc, &t->base_oloc); 1563db098ec4SIlya Dryomov if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 1564117d96a0SIlya Dryomov if (is_read && pi->read_tier >= 0) 156563244fa1SIlya Dryomov t->target_oloc.pool = pi->read_tier; 1566117d96a0SIlya Dryomov if (is_write && pi->write_tier >= 0) 156763244fa1SIlya Dryomov t->target_oloc.pool = pi->write_tier; 15686d637a54SIlya Dryomov 15696d637a54SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool); 15706d637a54SIlya Dryomov if (!pi) { 15716d637a54SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 15726d637a54SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 15736d637a54SIlya Dryomov goto out; 15746d637a54SIlya Dryomov } 157563244fa1SIlya Dryomov } 157663244fa1SIlya Dryomov 1577a86f009fSIlya Dryomov __ceph_object_locator_to_pg(pi, &t->target_oid, &t->target_oloc, &pgid); 157863244fa1SIlya Dryomov last_pgid.pool = pgid.pool; 157963244fa1SIlya Dryomov last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask); 158063244fa1SIlya Dryomov 1581df28152dSIlya Dryomov ceph_pg_to_up_acting_osds(osdc->osdmap, pi, &pgid, &up, &acting); 158263244fa1SIlya Dryomov if (any_change && 158363244fa1SIlya Dryomov ceph_is_new_interval(&t->acting, 158463244fa1SIlya Dryomov &acting, 158563244fa1SIlya Dryomov &t->up, 158663244fa1SIlya Dryomov &up, 158763244fa1SIlya Dryomov t->size, 158863244fa1SIlya Dryomov pi->size, 158963244fa1SIlya Dryomov t->min_size, 159063244fa1SIlya Dryomov pi->min_size, 159163244fa1SIlya Dryomov t->pg_num, 159263244fa1SIlya Dryomov pi->pg_num, 159363244fa1SIlya Dryomov t->sort_bitwise, 159463244fa1SIlya Dryomov sort_bitwise, 1595ae78dd81SIlya Dryomov t->recovery_deletes, 1596ae78dd81SIlya Dryomov recovery_deletes, 159763244fa1SIlya Dryomov &last_pgid)) 159863244fa1SIlya Dryomov force_resend = true; 159963244fa1SIlya Dryomov 160063244fa1SIlya Dryomov if (t->paused && !target_should_be_paused(osdc, t, pi)) { 160163244fa1SIlya Dryomov t->paused = false; 160284ed45dfSIlya Dryomov unpaused = true; 160363244fa1SIlya Dryomov } 160484ed45dfSIlya Dryomov legacy_change = ceph_pg_compare(&t->pgid, &pgid) || 1605117d96a0SIlya Dryomov ceph_osds_changed(&t->acting, &acting, 1606117d96a0SIlya Dryomov t->used_replica || any_change); 16077de030d6SIlya Dryomov if (t->pg_num) 16087de030d6SIlya Dryomov split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num); 160963244fa1SIlya Dryomov 16107de030d6SIlya Dryomov if (legacy_change || force_resend || split) { 161163244fa1SIlya Dryomov t->pgid = pgid; /* struct */ 1612df28152dSIlya Dryomov ceph_pg_to_primary_shard(osdc->osdmap, pi, &pgid, &t->spgid); 161363244fa1SIlya Dryomov ceph_osds_copy(&t->acting, &acting); 161463244fa1SIlya Dryomov ceph_osds_copy(&t->up, &up); 161563244fa1SIlya Dryomov t->size = pi->size; 161663244fa1SIlya Dryomov t->min_size = pi->min_size; 161763244fa1SIlya Dryomov t->pg_num = pi->pg_num; 161863244fa1SIlya Dryomov t->pg_num_mask = pi->pg_num_mask; 161963244fa1SIlya Dryomov t->sort_bitwise = sort_bitwise; 1620ae78dd81SIlya Dryomov t->recovery_deletes = recovery_deletes; 162163244fa1SIlya Dryomov 1622117d96a0SIlya Dryomov if ((t->flags & (CEPH_OSD_FLAG_BALANCE_READS | 1623117d96a0SIlya Dryomov CEPH_OSD_FLAG_LOCALIZE_READS)) && 1624117d96a0SIlya Dryomov !is_write && pi->type == CEPH_POOL_TYPE_REP && 1625117d96a0SIlya Dryomov acting.size > 1) { 1626117d96a0SIlya Dryomov int pos; 1627117d96a0SIlya Dryomov 1628117d96a0SIlya Dryomov WARN_ON(!is_read || acting.osds[0] != acting.primary); 1629117d96a0SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_BALANCE_READS) { 1630117d96a0SIlya Dryomov pos = pick_random_replica(&acting); 1631117d96a0SIlya Dryomov } else { 1632117d96a0SIlya Dryomov pos = pick_closest_replica(osdc, &acting); 1633117d96a0SIlya Dryomov } 1634117d96a0SIlya Dryomov t->osd = acting.osds[pos]; 1635117d96a0SIlya Dryomov t->used_replica = pos > 0; 1636117d96a0SIlya Dryomov } else { 163763244fa1SIlya Dryomov t->osd = acting.primary; 1638117d96a0SIlya Dryomov t->used_replica = false; 1639117d96a0SIlya Dryomov } 164063244fa1SIlya Dryomov } 164163244fa1SIlya Dryomov 1642a5613724SIlya Dryomov if (unpaused || legacy_change || force_resend || split) 164384ed45dfSIlya Dryomov ct_res = CALC_TARGET_NEED_RESEND; 164484ed45dfSIlya Dryomov else 164584ed45dfSIlya Dryomov ct_res = CALC_TARGET_NO_ACTION; 164684ed45dfSIlya Dryomov 164763244fa1SIlya Dryomov out: 1648a5613724SIlya Dryomov dout("%s t %p -> %d%d%d%d ct_res %d osd%d\n", __func__, t, unpaused, 1649a5613724SIlya Dryomov legacy_change, force_resend, split, ct_res, t->osd); 165063244fa1SIlya Dryomov return ct_res; 165163244fa1SIlya Dryomov } 165263244fa1SIlya Dryomov 1653a02a946dSIlya Dryomov static struct ceph_spg_mapping *alloc_spg_mapping(void) 1654a02a946dSIlya Dryomov { 1655a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1656a02a946dSIlya Dryomov 1657a02a946dSIlya Dryomov spg = kmalloc(sizeof(*spg), GFP_NOIO); 1658a02a946dSIlya Dryomov if (!spg) 1659a02a946dSIlya Dryomov return NULL; 1660a02a946dSIlya Dryomov 1661a02a946dSIlya Dryomov RB_CLEAR_NODE(&spg->node); 1662a02a946dSIlya Dryomov spg->backoffs = RB_ROOT; 1663a02a946dSIlya Dryomov return spg; 1664a02a946dSIlya Dryomov } 1665a02a946dSIlya Dryomov 1666a02a946dSIlya Dryomov static void free_spg_mapping(struct ceph_spg_mapping *spg) 1667a02a946dSIlya Dryomov { 1668a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&spg->node)); 1669a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&spg->backoffs)); 1670a02a946dSIlya Dryomov 1671a02a946dSIlya Dryomov kfree(spg); 1672a02a946dSIlya Dryomov } 1673a02a946dSIlya Dryomov 1674a02a946dSIlya Dryomov /* 1675a02a946dSIlya Dryomov * rbtree of ceph_spg_mapping for handling map<spg_t, ...>, similar to 1676a02a946dSIlya Dryomov * ceph_pg_mapping. Used to track OSD backoffs -- a backoff [range] is 1677a02a946dSIlya Dryomov * defined only within a specific spgid; it does not pass anything to 1678a02a946dSIlya Dryomov * children on split, or to another primary. 1679a02a946dSIlya Dryomov */ 1680a02a946dSIlya Dryomov DEFINE_RB_FUNCS2(spg_mapping, struct ceph_spg_mapping, spgid, ceph_spg_compare, 1681a02a946dSIlya Dryomov RB_BYPTR, const struct ceph_spg *, node) 1682a02a946dSIlya Dryomov 1683a02a946dSIlya Dryomov static u64 hoid_get_bitwise_key(const struct ceph_hobject_id *hoid) 1684a02a946dSIlya Dryomov { 1685a02a946dSIlya Dryomov return hoid->is_max ? 0x100000000ull : hoid->hash_reverse_bits; 1686a02a946dSIlya Dryomov } 1687a02a946dSIlya Dryomov 1688a02a946dSIlya Dryomov static void hoid_get_effective_key(const struct ceph_hobject_id *hoid, 1689a02a946dSIlya Dryomov void **pkey, size_t *pkey_len) 1690a02a946dSIlya Dryomov { 1691a02a946dSIlya Dryomov if (hoid->key_len) { 1692a02a946dSIlya Dryomov *pkey = hoid->key; 1693a02a946dSIlya Dryomov *pkey_len = hoid->key_len; 1694a02a946dSIlya Dryomov } else { 1695a02a946dSIlya Dryomov *pkey = hoid->oid; 1696a02a946dSIlya Dryomov *pkey_len = hoid->oid_len; 1697a02a946dSIlya Dryomov } 1698a02a946dSIlya Dryomov } 1699a02a946dSIlya Dryomov 1700a02a946dSIlya Dryomov static int compare_names(const void *name1, size_t name1_len, 1701a02a946dSIlya Dryomov const void *name2, size_t name2_len) 1702a02a946dSIlya Dryomov { 1703a02a946dSIlya Dryomov int ret; 1704a02a946dSIlya Dryomov 1705a02a946dSIlya Dryomov ret = memcmp(name1, name2, min(name1_len, name2_len)); 1706a02a946dSIlya Dryomov if (!ret) { 1707a02a946dSIlya Dryomov if (name1_len < name2_len) 1708a02a946dSIlya Dryomov ret = -1; 1709a02a946dSIlya Dryomov else if (name1_len > name2_len) 1710a02a946dSIlya Dryomov ret = 1; 1711a02a946dSIlya Dryomov } 1712a02a946dSIlya Dryomov return ret; 1713a02a946dSIlya Dryomov } 1714a02a946dSIlya Dryomov 1715a02a946dSIlya Dryomov static int hoid_compare(const struct ceph_hobject_id *lhs, 1716a02a946dSIlya Dryomov const struct ceph_hobject_id *rhs) 1717a02a946dSIlya Dryomov { 1718a02a946dSIlya Dryomov void *effective_key1, *effective_key2; 1719a02a946dSIlya Dryomov size_t effective_key1_len, effective_key2_len; 1720a02a946dSIlya Dryomov int ret; 1721a02a946dSIlya Dryomov 1722a02a946dSIlya Dryomov if (lhs->is_max < rhs->is_max) 1723a02a946dSIlya Dryomov return -1; 1724a02a946dSIlya Dryomov if (lhs->is_max > rhs->is_max) 1725a02a946dSIlya Dryomov return 1; 1726a02a946dSIlya Dryomov 1727a02a946dSIlya Dryomov if (lhs->pool < rhs->pool) 1728a02a946dSIlya Dryomov return -1; 1729a02a946dSIlya Dryomov if (lhs->pool > rhs->pool) 1730a02a946dSIlya Dryomov return 1; 1731a02a946dSIlya Dryomov 1732a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) < hoid_get_bitwise_key(rhs)) 1733a02a946dSIlya Dryomov return -1; 1734a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) > hoid_get_bitwise_key(rhs)) 1735a02a946dSIlya Dryomov return 1; 1736a02a946dSIlya Dryomov 1737a02a946dSIlya Dryomov ret = compare_names(lhs->nspace, lhs->nspace_len, 1738a02a946dSIlya Dryomov rhs->nspace, rhs->nspace_len); 1739a02a946dSIlya Dryomov if (ret) 1740a02a946dSIlya Dryomov return ret; 1741a02a946dSIlya Dryomov 1742a02a946dSIlya Dryomov hoid_get_effective_key(lhs, &effective_key1, &effective_key1_len); 1743a02a946dSIlya Dryomov hoid_get_effective_key(rhs, &effective_key2, &effective_key2_len); 1744a02a946dSIlya Dryomov ret = compare_names(effective_key1, effective_key1_len, 1745a02a946dSIlya Dryomov effective_key2, effective_key2_len); 1746a02a946dSIlya Dryomov if (ret) 1747a02a946dSIlya Dryomov return ret; 1748a02a946dSIlya Dryomov 1749a02a946dSIlya Dryomov ret = compare_names(lhs->oid, lhs->oid_len, rhs->oid, rhs->oid_len); 1750a02a946dSIlya Dryomov if (ret) 1751a02a946dSIlya Dryomov return ret; 1752a02a946dSIlya Dryomov 1753a02a946dSIlya Dryomov if (lhs->snapid < rhs->snapid) 1754a02a946dSIlya Dryomov return -1; 1755a02a946dSIlya Dryomov if (lhs->snapid > rhs->snapid) 1756a02a946dSIlya Dryomov return 1; 1757a02a946dSIlya Dryomov 1758a02a946dSIlya Dryomov return 0; 1759a02a946dSIlya Dryomov } 1760a02a946dSIlya Dryomov 1761a02a946dSIlya Dryomov /* 1762a02a946dSIlya Dryomov * For decoding ->begin and ->end of MOSDBackoff only -- no MIN/MAX 1763a02a946dSIlya Dryomov * compat stuff here. 1764a02a946dSIlya Dryomov * 1765a02a946dSIlya Dryomov * Assumes @hoid is zero-initialized. 1766a02a946dSIlya Dryomov */ 1767a02a946dSIlya Dryomov static int decode_hoid(void **p, void *end, struct ceph_hobject_id *hoid) 1768a02a946dSIlya Dryomov { 1769a02a946dSIlya Dryomov u8 struct_v; 1770a02a946dSIlya Dryomov u32 struct_len; 1771a02a946dSIlya Dryomov int ret; 1772a02a946dSIlya Dryomov 1773a02a946dSIlya Dryomov ret = ceph_start_decoding(p, end, 4, "hobject_t", &struct_v, 1774a02a946dSIlya Dryomov &struct_len); 1775a02a946dSIlya Dryomov if (ret) 1776a02a946dSIlya Dryomov return ret; 1777a02a946dSIlya Dryomov 1778a02a946dSIlya Dryomov if (struct_v < 4) { 1779a02a946dSIlya Dryomov pr_err("got struct_v %d < 4 of hobject_t\n", struct_v); 1780a02a946dSIlya Dryomov goto e_inval; 1781a02a946dSIlya Dryomov } 1782a02a946dSIlya Dryomov 1783a02a946dSIlya Dryomov hoid->key = ceph_extract_encoded_string(p, end, &hoid->key_len, 1784a02a946dSIlya Dryomov GFP_NOIO); 1785a02a946dSIlya Dryomov if (IS_ERR(hoid->key)) { 1786a02a946dSIlya Dryomov ret = PTR_ERR(hoid->key); 1787a02a946dSIlya Dryomov hoid->key = NULL; 1788a02a946dSIlya Dryomov return ret; 1789a02a946dSIlya Dryomov } 1790a02a946dSIlya Dryomov 1791a02a946dSIlya Dryomov hoid->oid = ceph_extract_encoded_string(p, end, &hoid->oid_len, 1792a02a946dSIlya Dryomov GFP_NOIO); 1793a02a946dSIlya Dryomov if (IS_ERR(hoid->oid)) { 1794a02a946dSIlya Dryomov ret = PTR_ERR(hoid->oid); 1795a02a946dSIlya Dryomov hoid->oid = NULL; 1796a02a946dSIlya Dryomov return ret; 1797a02a946dSIlya Dryomov } 1798a02a946dSIlya Dryomov 1799a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->snapid, e_inval); 1800a02a946dSIlya Dryomov ceph_decode_32_safe(p, end, hoid->hash, e_inval); 1801a02a946dSIlya Dryomov ceph_decode_8_safe(p, end, hoid->is_max, e_inval); 1802a02a946dSIlya Dryomov 1803a02a946dSIlya Dryomov hoid->nspace = ceph_extract_encoded_string(p, end, &hoid->nspace_len, 1804a02a946dSIlya Dryomov GFP_NOIO); 1805a02a946dSIlya Dryomov if (IS_ERR(hoid->nspace)) { 1806a02a946dSIlya Dryomov ret = PTR_ERR(hoid->nspace); 1807a02a946dSIlya Dryomov hoid->nspace = NULL; 1808a02a946dSIlya Dryomov return ret; 1809a02a946dSIlya Dryomov } 1810a02a946dSIlya Dryomov 1811a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->pool, e_inval); 1812a02a946dSIlya Dryomov 1813a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1814a02a946dSIlya Dryomov return 0; 1815a02a946dSIlya Dryomov 1816a02a946dSIlya Dryomov e_inval: 1817a02a946dSIlya Dryomov return -EINVAL; 1818a02a946dSIlya Dryomov } 1819a02a946dSIlya Dryomov 1820a02a946dSIlya Dryomov static int hoid_encoding_size(const struct ceph_hobject_id *hoid) 1821a02a946dSIlya Dryomov { 1822a02a946dSIlya Dryomov return 8 + 4 + 1 + 8 + /* snapid, hash, is_max, pool */ 1823a02a946dSIlya Dryomov 4 + hoid->key_len + 4 + hoid->oid_len + 4 + hoid->nspace_len; 1824a02a946dSIlya Dryomov } 1825a02a946dSIlya Dryomov 1826a02a946dSIlya Dryomov static void encode_hoid(void **p, void *end, const struct ceph_hobject_id *hoid) 1827a02a946dSIlya Dryomov { 1828a02a946dSIlya Dryomov ceph_start_encoding(p, 4, 3, hoid_encoding_size(hoid)); 1829a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->key, hoid->key_len); 1830a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->oid, hoid->oid_len); 1831a02a946dSIlya Dryomov ceph_encode_64(p, hoid->snapid); 1832a02a946dSIlya Dryomov ceph_encode_32(p, hoid->hash); 1833a02a946dSIlya Dryomov ceph_encode_8(p, hoid->is_max); 1834a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->nspace, hoid->nspace_len); 1835a02a946dSIlya Dryomov ceph_encode_64(p, hoid->pool); 1836a02a946dSIlya Dryomov } 1837a02a946dSIlya Dryomov 1838a02a946dSIlya Dryomov static void free_hoid(struct ceph_hobject_id *hoid) 1839a02a946dSIlya Dryomov { 1840a02a946dSIlya Dryomov if (hoid) { 1841a02a946dSIlya Dryomov kfree(hoid->key); 1842a02a946dSIlya Dryomov kfree(hoid->oid); 1843a02a946dSIlya Dryomov kfree(hoid->nspace); 1844a02a946dSIlya Dryomov kfree(hoid); 1845a02a946dSIlya Dryomov } 1846a02a946dSIlya Dryomov } 1847a02a946dSIlya Dryomov 1848a02a946dSIlya Dryomov static struct ceph_osd_backoff *alloc_backoff(void) 1849a02a946dSIlya Dryomov { 1850a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1851a02a946dSIlya Dryomov 1852a02a946dSIlya Dryomov backoff = kzalloc(sizeof(*backoff), GFP_NOIO); 1853a02a946dSIlya Dryomov if (!backoff) 1854a02a946dSIlya Dryomov return NULL; 1855a02a946dSIlya Dryomov 1856a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->spg_node); 1857a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->id_node); 1858a02a946dSIlya Dryomov return backoff; 1859a02a946dSIlya Dryomov } 1860a02a946dSIlya Dryomov 1861a02a946dSIlya Dryomov static void free_backoff(struct ceph_osd_backoff *backoff) 1862a02a946dSIlya Dryomov { 1863a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->spg_node)); 1864a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->id_node)); 1865a02a946dSIlya Dryomov 1866a02a946dSIlya Dryomov free_hoid(backoff->begin); 1867a02a946dSIlya Dryomov free_hoid(backoff->end); 1868a02a946dSIlya Dryomov kfree(backoff); 1869a02a946dSIlya Dryomov } 1870a02a946dSIlya Dryomov 1871a02a946dSIlya Dryomov /* 1872a02a946dSIlya Dryomov * Within a specific spgid, backoffs are managed by ->begin hoid. 1873a02a946dSIlya Dryomov */ 1874a02a946dSIlya Dryomov DEFINE_RB_INSDEL_FUNCS2(backoff, struct ceph_osd_backoff, begin, hoid_compare, 1875a02a946dSIlya Dryomov RB_BYVAL, spg_node); 1876a02a946dSIlya Dryomov 1877a02a946dSIlya Dryomov static struct ceph_osd_backoff *lookup_containing_backoff(struct rb_root *root, 1878a02a946dSIlya Dryomov const struct ceph_hobject_id *hoid) 1879a02a946dSIlya Dryomov { 1880a02a946dSIlya Dryomov struct rb_node *n = root->rb_node; 1881a02a946dSIlya Dryomov 1882a02a946dSIlya Dryomov while (n) { 1883a02a946dSIlya Dryomov struct ceph_osd_backoff *cur = 1884a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_backoff, spg_node); 1885a02a946dSIlya Dryomov int cmp; 1886a02a946dSIlya Dryomov 1887a02a946dSIlya Dryomov cmp = hoid_compare(hoid, cur->begin); 1888a02a946dSIlya Dryomov if (cmp < 0) { 1889a02a946dSIlya Dryomov n = n->rb_left; 1890a02a946dSIlya Dryomov } else if (cmp > 0) { 1891a02a946dSIlya Dryomov if (hoid_compare(hoid, cur->end) < 0) 1892a02a946dSIlya Dryomov return cur; 1893a02a946dSIlya Dryomov 1894a02a946dSIlya Dryomov n = n->rb_right; 1895a02a946dSIlya Dryomov } else { 1896a02a946dSIlya Dryomov return cur; 1897a02a946dSIlya Dryomov } 1898a02a946dSIlya Dryomov } 1899a02a946dSIlya Dryomov 1900a02a946dSIlya Dryomov return NULL; 1901a02a946dSIlya Dryomov } 1902a02a946dSIlya Dryomov 1903a02a946dSIlya Dryomov /* 1904a02a946dSIlya Dryomov * Each backoff has a unique id within its OSD session. 1905a02a946dSIlya Dryomov */ 1906a02a946dSIlya Dryomov DEFINE_RB_FUNCS(backoff_by_id, struct ceph_osd_backoff, id, id_node) 1907a02a946dSIlya Dryomov 1908a02a946dSIlya Dryomov static void clear_backoffs(struct ceph_osd *osd) 1909a02a946dSIlya Dryomov { 1910a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&osd->o_backoff_mappings)) { 1911a02a946dSIlya Dryomov struct ceph_spg_mapping *spg = 1912a02a946dSIlya Dryomov rb_entry(rb_first(&osd->o_backoff_mappings), 1913a02a946dSIlya Dryomov struct ceph_spg_mapping, node); 1914a02a946dSIlya Dryomov 1915a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&spg->backoffs)) { 1916a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff = 1917a02a946dSIlya Dryomov rb_entry(rb_first(&spg->backoffs), 1918a02a946dSIlya Dryomov struct ceph_osd_backoff, spg_node); 1919a02a946dSIlya Dryomov 1920a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 1921a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 1922a02a946dSIlya Dryomov free_backoff(backoff); 1923a02a946dSIlya Dryomov } 1924a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 1925a02a946dSIlya Dryomov free_spg_mapping(spg); 1926a02a946dSIlya Dryomov } 1927a02a946dSIlya Dryomov } 1928a02a946dSIlya Dryomov 1929a02a946dSIlya Dryomov /* 1930a02a946dSIlya Dryomov * Set up a temporary, non-owning view into @t. 1931a02a946dSIlya Dryomov */ 1932a02a946dSIlya Dryomov static void hoid_fill_from_target(struct ceph_hobject_id *hoid, 1933a02a946dSIlya Dryomov const struct ceph_osd_request_target *t) 1934a02a946dSIlya Dryomov { 1935a02a946dSIlya Dryomov hoid->key = NULL; 1936a02a946dSIlya Dryomov hoid->key_len = 0; 1937a02a946dSIlya Dryomov hoid->oid = t->target_oid.name; 1938a02a946dSIlya Dryomov hoid->oid_len = t->target_oid.name_len; 1939a02a946dSIlya Dryomov hoid->snapid = CEPH_NOSNAP; 1940a02a946dSIlya Dryomov hoid->hash = t->pgid.seed; 1941a02a946dSIlya Dryomov hoid->is_max = false; 1942a02a946dSIlya Dryomov if (t->target_oloc.pool_ns) { 1943a02a946dSIlya Dryomov hoid->nspace = t->target_oloc.pool_ns->str; 1944a02a946dSIlya Dryomov hoid->nspace_len = t->target_oloc.pool_ns->len; 1945a02a946dSIlya Dryomov } else { 1946a02a946dSIlya Dryomov hoid->nspace = NULL; 1947a02a946dSIlya Dryomov hoid->nspace_len = 0; 1948a02a946dSIlya Dryomov } 1949a02a946dSIlya Dryomov hoid->pool = t->target_oloc.pool; 1950a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1951a02a946dSIlya Dryomov } 1952a02a946dSIlya Dryomov 1953a02a946dSIlya Dryomov static bool should_plug_request(struct ceph_osd_request *req) 1954a02a946dSIlya Dryomov { 1955a02a946dSIlya Dryomov struct ceph_osd *osd = req->r_osd; 1956a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1957a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1958a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 1959a02a946dSIlya Dryomov 1960a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &req->r_t.spgid); 1961a02a946dSIlya Dryomov if (!spg) 1962a02a946dSIlya Dryomov return false; 1963a02a946dSIlya Dryomov 1964a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, &req->r_t); 1965a02a946dSIlya Dryomov backoff = lookup_containing_backoff(&spg->backoffs, &hoid); 1966a02a946dSIlya Dryomov if (!backoff) 1967a02a946dSIlya Dryomov return false; 1968a02a946dSIlya Dryomov 1969a02a946dSIlya Dryomov dout("%s req %p tid %llu backoff osd%d spgid %llu.%xs%d id %llu\n", 1970a02a946dSIlya Dryomov __func__, req, req->r_tid, osd->o_osd, backoff->spgid.pgid.pool, 1971a02a946dSIlya Dryomov backoff->spgid.pgid.seed, backoff->spgid.shard, backoff->id); 1972a02a946dSIlya Dryomov return true; 1973a02a946dSIlya Dryomov } 1974a02a946dSIlya Dryomov 19750d9c1ab3SIlya Dryomov /* 19760d9c1ab3SIlya Dryomov * Keep get_num_data_items() in sync with this function. 19770d9c1ab3SIlya Dryomov */ 197898c4bfe9SIlya Dryomov static void setup_request_data(struct ceph_osd_request *req) 19793d14c5d2SYehuda Sadeh { 198098c4bfe9SIlya Dryomov struct ceph_msg *request_msg = req->r_request; 198198c4bfe9SIlya Dryomov struct ceph_msg *reply_msg = req->r_reply; 198298c4bfe9SIlya Dryomov struct ceph_osd_req_op *op; 19833d14c5d2SYehuda Sadeh 198498c4bfe9SIlya Dryomov if (req->r_request->num_data_items || req->r_reply->num_data_items) 1985bb873b53SIlya Dryomov return; 19863d14c5d2SYehuda Sadeh 198798c4bfe9SIlya Dryomov WARN_ON(request_msg->data_length || reply_msg->data_length); 198898c4bfe9SIlya Dryomov for (op = req->r_ops; op != &req->r_ops[req->r_num_ops]; op++) { 1989bb873b53SIlya Dryomov switch (op->op) { 1990bb873b53SIlya Dryomov /* request */ 1991bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITE: 1992bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 1993bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->extent.length); 199498c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 199598c4bfe9SIlya Dryomov &op->extent.osd_data); 1996bb873b53SIlya Dryomov break; 1997bb873b53SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 1998bb873b53SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 1999bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->xattr.name_len + 2000bb873b53SIlya Dryomov op->xattr.value_len); 200198c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 200298c4bfe9SIlya Dryomov &op->xattr.osd_data); 2003bb873b53SIlya Dryomov break; 2004922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 200598c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 2006922dab61SIlya Dryomov &op->notify_ack.request_data); 2007922dab61SIlya Dryomov break; 200878beb0ffSLuis Henriques case CEPH_OSD_OP_COPY_FROM2: 200923ddf9beSLuis Henriques ceph_osdc_msg_data_add(request_msg, 201023ddf9beSLuis Henriques &op->copy_from.osd_data); 201123ddf9beSLuis Henriques break; 2012bb873b53SIlya Dryomov 2013bb873b53SIlya Dryomov /* reply */ 2014bb873b53SIlya Dryomov case CEPH_OSD_OP_STAT: 201598c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2016bb873b53SIlya Dryomov &op->raw_data_in); 2017bb873b53SIlya Dryomov break; 2018bb873b53SIlya Dryomov case CEPH_OSD_OP_READ: 201998c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2020bb873b53SIlya Dryomov &op->extent.osd_data); 2021bb873b53SIlya Dryomov break; 2022a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 202398c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2024a4ed38d7SDouglas Fuller &op->list_watchers.response_data); 2025a4ed38d7SDouglas Fuller break; 2026bb873b53SIlya Dryomov 2027bb873b53SIlya Dryomov /* both */ 2028bb873b53SIlya Dryomov case CEPH_OSD_OP_CALL: 2029bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->cls.class_len + 2030bb873b53SIlya Dryomov op->cls.method_len + 2031bb873b53SIlya Dryomov op->cls.indata_len); 203298c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 203398c4bfe9SIlya Dryomov &op->cls.request_info); 2034bb873b53SIlya Dryomov /* optional, can be NONE */ 203598c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 203698c4bfe9SIlya Dryomov &op->cls.request_data); 2037bb873b53SIlya Dryomov /* optional, can be NONE */ 203898c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 2039bb873b53SIlya Dryomov &op->cls.response_data); 2040bb873b53SIlya Dryomov break; 204119079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 204298c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(request_msg, 204319079203SIlya Dryomov &op->notify.request_data); 204498c4bfe9SIlya Dryomov ceph_osdc_msg_data_add(reply_msg, 204519079203SIlya Dryomov &op->notify.response_data); 204619079203SIlya Dryomov break; 2047bb873b53SIlya Dryomov } 2048bb873b53SIlya Dryomov } 2049bb873b53SIlya Dryomov } 2050bb873b53SIlya Dryomov 20512e59ffd1SIlya Dryomov static void encode_pgid(void **p, const struct ceph_pg *pgid) 20522e59ffd1SIlya Dryomov { 20532e59ffd1SIlya Dryomov ceph_encode_8(p, 1); 20542e59ffd1SIlya Dryomov ceph_encode_64(p, pgid->pool); 20552e59ffd1SIlya Dryomov ceph_encode_32(p, pgid->seed); 20562e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 20572e59ffd1SIlya Dryomov } 20582e59ffd1SIlya Dryomov 20598cb441c0SIlya Dryomov static void encode_spgid(void **p, const struct ceph_spg *spgid) 20608cb441c0SIlya Dryomov { 20618cb441c0SIlya Dryomov ceph_start_encoding(p, 1, 1, CEPH_PGID_ENCODING_LEN + 1); 20628cb441c0SIlya Dryomov encode_pgid(p, &spgid->pgid); 20638cb441c0SIlya Dryomov ceph_encode_8(p, spgid->shard); 20648cb441c0SIlya Dryomov } 20658cb441c0SIlya Dryomov 20662e59ffd1SIlya Dryomov static void encode_oloc(void **p, void *end, 20672e59ffd1SIlya Dryomov const struct ceph_object_locator *oloc) 20682e59ffd1SIlya Dryomov { 20692e59ffd1SIlya Dryomov ceph_start_encoding(p, 5, 4, ceph_oloc_encoding_size(oloc)); 20702e59ffd1SIlya Dryomov ceph_encode_64(p, oloc->pool); 20712e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 20722e59ffd1SIlya Dryomov ceph_encode_32(p, 0); /* key len */ 20732e59ffd1SIlya Dryomov if (oloc->pool_ns) 20742e59ffd1SIlya Dryomov ceph_encode_string(p, end, oloc->pool_ns->str, 20752e59ffd1SIlya Dryomov oloc->pool_ns->len); 20762e59ffd1SIlya Dryomov else 20772e59ffd1SIlya Dryomov ceph_encode_32(p, 0); 20782e59ffd1SIlya Dryomov } 20792e59ffd1SIlya Dryomov 20808cb441c0SIlya Dryomov static void encode_request_partial(struct ceph_osd_request *req, 20818cb441c0SIlya Dryomov struct ceph_msg *msg) 2082bb873b53SIlya Dryomov { 2083bb873b53SIlya Dryomov void *p = msg->front.iov_base; 2084bb873b53SIlya Dryomov void *const end = p + msg->front_alloc_len; 2085bb873b53SIlya Dryomov u32 data_len = 0; 2086bb873b53SIlya Dryomov int i; 2087bb873b53SIlya Dryomov 2088bb873b53SIlya Dryomov if (req->r_flags & CEPH_OSD_FLAG_WRITE) { 2089bb873b53SIlya Dryomov /* snapshots aren't writeable */ 2090bb873b53SIlya Dryomov WARN_ON(req->r_snapid != CEPH_NOSNAP); 2091bb873b53SIlya Dryomov } else { 2092bb873b53SIlya Dryomov WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec || 2093bb873b53SIlya Dryomov req->r_data_offset || req->r_snapc); 2094bb873b53SIlya Dryomov } 2095bb873b53SIlya Dryomov 209698c4bfe9SIlya Dryomov setup_request_data(req); 2097bb873b53SIlya Dryomov 20988cb441c0SIlya Dryomov encode_spgid(&p, &req->r_t.spgid); /* actual spg */ 20998cb441c0SIlya Dryomov ceph_encode_32(&p, req->r_t.pgid.seed); /* raw hash */ 2100bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_osdc->osdmap->epoch); 2101bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_flags); 21028cb441c0SIlya Dryomov 21038cb441c0SIlya Dryomov /* reqid */ 21048cb441c0SIlya Dryomov ceph_start_encoding(&p, 2, 2, sizeof(struct ceph_osd_reqid)); 21058cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_osd_reqid)); 21068cb441c0SIlya Dryomov p += sizeof(struct ceph_osd_reqid); 21078cb441c0SIlya Dryomov 21088cb441c0SIlya Dryomov /* trace */ 21098cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_blkin_trace_info)); 21108cb441c0SIlya Dryomov p += sizeof(struct ceph_blkin_trace_info); 21118cb441c0SIlya Dryomov 21128cb441c0SIlya Dryomov ceph_encode_32(&p, 0); /* client_inc, always 0 */ 2113fac02ddfSArnd Bergmann ceph_encode_timespec64(p, &req->r_mtime); 2114bb873b53SIlya Dryomov p += sizeof(struct ceph_timespec); 2115aa26d662SJeff Layton 21162e59ffd1SIlya Dryomov encode_oloc(&p, end, &req->r_t.target_oloc); 21172e59ffd1SIlya Dryomov ceph_encode_string(&p, end, req->r_t.target_oid.name, 21182e59ffd1SIlya Dryomov req->r_t.target_oid.name_len); 2119bb873b53SIlya Dryomov 2120bb873b53SIlya Dryomov /* ops, can imply data */ 2121bb873b53SIlya Dryomov ceph_encode_16(&p, req->r_num_ops); 2122bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 2123bb873b53SIlya Dryomov data_len += osd_req_encode_op(p, &req->r_ops[i]); 2124bb873b53SIlya Dryomov p += sizeof(struct ceph_osd_op); 2125bb873b53SIlya Dryomov } 2126bb873b53SIlya Dryomov 2127bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapid); /* snapid */ 2128bb873b53SIlya Dryomov if (req->r_snapc) { 2129bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->seq); 2130bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_snapc->num_snaps); 2131bb873b53SIlya Dryomov for (i = 0; i < req->r_snapc->num_snaps; i++) 2132bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->snaps[i]); 2133bb873b53SIlya Dryomov } else { 2134bb873b53SIlya Dryomov ceph_encode_64(&p, 0); /* snap_seq */ 2135bb873b53SIlya Dryomov ceph_encode_32(&p, 0); /* snaps len */ 2136bb873b53SIlya Dryomov } 2137bb873b53SIlya Dryomov 2138bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_attempts); /* retry_attempt */ 2139986e8989SIlya Dryomov BUG_ON(p > end - 8); /* space for features */ 2140bb873b53SIlya Dryomov 21418cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(8); /* MOSDOp v8 */ 21428cb441c0SIlya Dryomov /* front_len is finalized in encode_request_finish() */ 2143986e8989SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 2144986e8989SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 2145bb873b53SIlya Dryomov msg->hdr.data_len = cpu_to_le32(data_len); 2146bb873b53SIlya Dryomov /* 2147bb873b53SIlya Dryomov * The header "data_off" is a hint to the receiver allowing it 2148bb873b53SIlya Dryomov * to align received data into its buffers such that there's no 2149bb873b53SIlya Dryomov * need to re-copy it before writing it to disk (direct I/O). 2150bb873b53SIlya Dryomov */ 2151bb873b53SIlya Dryomov msg->hdr.data_off = cpu_to_le16(req->r_data_offset); 2152bb873b53SIlya Dryomov 21538cb441c0SIlya Dryomov dout("%s req %p msg %p oid %s oid_len %d\n", __func__, req, msg, 21548cb441c0SIlya Dryomov req->r_t.target_oid.name, req->r_t.target_oid.name_len); 21558cb441c0SIlya Dryomov } 21568cb441c0SIlya Dryomov 21578cb441c0SIlya Dryomov static void encode_request_finish(struct ceph_msg *msg) 21588cb441c0SIlya Dryomov { 21598cb441c0SIlya Dryomov void *p = msg->front.iov_base; 2160986e8989SIlya Dryomov void *const partial_end = p + msg->front.iov_len; 21618cb441c0SIlya Dryomov void *const end = p + msg->front_alloc_len; 21628cb441c0SIlya Dryomov 21638cb441c0SIlya Dryomov if (CEPH_HAVE_FEATURE(msg->con->peer_features, RESEND_ON_SPLIT)) { 21648cb441c0SIlya Dryomov /* luminous OSD -- encode features and be done */ 2165986e8989SIlya Dryomov p = partial_end; 21668cb441c0SIlya Dryomov ceph_encode_64(&p, msg->con->peer_features); 21678cb441c0SIlya Dryomov } else { 21688cb441c0SIlya Dryomov struct { 21698cb441c0SIlya Dryomov char spgid[CEPH_ENCODING_START_BLK_LEN + 21708cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1]; 21718cb441c0SIlya Dryomov __le32 hash; 21728cb441c0SIlya Dryomov __le32 epoch; 21738cb441c0SIlya Dryomov __le32 flags; 21748cb441c0SIlya Dryomov char reqid[CEPH_ENCODING_START_BLK_LEN + 21758cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid)]; 21768cb441c0SIlya Dryomov char trace[sizeof(struct ceph_blkin_trace_info)]; 21778cb441c0SIlya Dryomov __le32 client_inc; 21788cb441c0SIlya Dryomov struct ceph_timespec mtime; 21798cb441c0SIlya Dryomov } __packed head; 21808cb441c0SIlya Dryomov struct ceph_pg pgid; 21818cb441c0SIlya Dryomov void *oloc, *oid, *tail; 21828cb441c0SIlya Dryomov int oloc_len, oid_len, tail_len; 21838cb441c0SIlya Dryomov int len; 21848cb441c0SIlya Dryomov 21858cb441c0SIlya Dryomov /* 21868cb441c0SIlya Dryomov * Pre-luminous OSD -- reencode v8 into v4 using @head 21878cb441c0SIlya Dryomov * as a temporary buffer. Encode the raw PG; the rest 21888cb441c0SIlya Dryomov * is just a matter of moving oloc, oid and tail blobs 21898cb441c0SIlya Dryomov * around. 21908cb441c0SIlya Dryomov */ 21918cb441c0SIlya Dryomov memcpy(&head, p, sizeof(head)); 21928cb441c0SIlya Dryomov p += sizeof(head); 21938cb441c0SIlya Dryomov 21948cb441c0SIlya Dryomov oloc = p; 21958cb441c0SIlya Dryomov p += CEPH_ENCODING_START_BLK_LEN; 21968cb441c0SIlya Dryomov pgid.pool = ceph_decode_64(&p); 21978cb441c0SIlya Dryomov p += 4 + 4; /* preferred, key len */ 21988cb441c0SIlya Dryomov len = ceph_decode_32(&p); 21998cb441c0SIlya Dryomov p += len; /* nspace */ 22008cb441c0SIlya Dryomov oloc_len = p - oloc; 22018cb441c0SIlya Dryomov 22028cb441c0SIlya Dryomov oid = p; 22038cb441c0SIlya Dryomov len = ceph_decode_32(&p); 22048cb441c0SIlya Dryomov p += len; 22058cb441c0SIlya Dryomov oid_len = p - oid; 22068cb441c0SIlya Dryomov 22078cb441c0SIlya Dryomov tail = p; 2208986e8989SIlya Dryomov tail_len = partial_end - p; 22098cb441c0SIlya Dryomov 22108cb441c0SIlya Dryomov p = msg->front.iov_base; 22118cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.client_inc, sizeof(head.client_inc)); 22128cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.epoch, sizeof(head.epoch)); 22138cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.flags, sizeof(head.flags)); 22148cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.mtime, sizeof(head.mtime)); 22158cb441c0SIlya Dryomov 22168cb441c0SIlya Dryomov /* reassert_version */ 22178cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_eversion)); 22188cb441c0SIlya Dryomov p += sizeof(struct ceph_eversion); 22198cb441c0SIlya Dryomov 22208cb441c0SIlya Dryomov BUG_ON(p >= oloc); 22218cb441c0SIlya Dryomov memmove(p, oloc, oloc_len); 22228cb441c0SIlya Dryomov p += oloc_len; 22238cb441c0SIlya Dryomov 22248cb441c0SIlya Dryomov pgid.seed = le32_to_cpu(head.hash); 22258cb441c0SIlya Dryomov encode_pgid(&p, &pgid); /* raw pg */ 22268cb441c0SIlya Dryomov 22278cb441c0SIlya Dryomov BUG_ON(p >= oid); 22288cb441c0SIlya Dryomov memmove(p, oid, oid_len); 22298cb441c0SIlya Dryomov p += oid_len; 22308cb441c0SIlya Dryomov 22318cb441c0SIlya Dryomov /* tail -- ops, snapid, snapc, retry_attempt */ 22328cb441c0SIlya Dryomov BUG_ON(p >= tail); 22338cb441c0SIlya Dryomov memmove(p, tail, tail_len); 22348cb441c0SIlya Dryomov p += tail_len; 22358cb441c0SIlya Dryomov 22368cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */ 22378cb441c0SIlya Dryomov } 22388cb441c0SIlya Dryomov 22398cb441c0SIlya Dryomov BUG_ON(p > end); 22408cb441c0SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 22418cb441c0SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 22428cb441c0SIlya Dryomov 22438cb441c0SIlya Dryomov dout("%s msg %p tid %llu %u+%u+%u v%d\n", __func__, msg, 22448cb441c0SIlya Dryomov le64_to_cpu(msg->hdr.tid), le32_to_cpu(msg->hdr.front_len), 22458cb441c0SIlya Dryomov le32_to_cpu(msg->hdr.middle_len), le32_to_cpu(msg->hdr.data_len), 22468cb441c0SIlya Dryomov le16_to_cpu(msg->hdr.version)); 2247bb873b53SIlya Dryomov } 2248bb873b53SIlya Dryomov 2249bb873b53SIlya Dryomov /* 2250bb873b53SIlya Dryomov * @req has to be assigned a tid and registered. 2251bb873b53SIlya Dryomov */ 2252bb873b53SIlya Dryomov static void send_request(struct ceph_osd_request *req) 2253bb873b53SIlya Dryomov { 2254bb873b53SIlya Dryomov struct ceph_osd *osd = req->r_osd; 2255bb873b53SIlya Dryomov 22565aea3dcdSIlya Dryomov verify_osd_locked(osd); 2257bb873b53SIlya Dryomov WARN_ON(osd->o_osd != req->r_t.osd); 2258bb873b53SIlya Dryomov 2259a02a946dSIlya Dryomov /* backoff? */ 2260a02a946dSIlya Dryomov if (should_plug_request(req)) 2261a02a946dSIlya Dryomov return; 2262a02a946dSIlya Dryomov 22635aea3dcdSIlya Dryomov /* 22645aea3dcdSIlya Dryomov * We may have a previously queued request message hanging 22655aea3dcdSIlya Dryomov * around. Cancel it to avoid corrupting the msgr. 22665aea3dcdSIlya Dryomov */ 22675aea3dcdSIlya Dryomov if (req->r_sent) 22685aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 22695aea3dcdSIlya Dryomov 2270bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR; 2271bb873b53SIlya Dryomov if (req->r_attempts) 2272bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_RETRY; 2273bb873b53SIlya Dryomov else 2274bb873b53SIlya Dryomov WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY); 2275bb873b53SIlya Dryomov 22768cb441c0SIlya Dryomov encode_request_partial(req, req->r_request); 2277bb873b53SIlya Dryomov 227804c7d789SIlya 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", 2279bb873b53SIlya Dryomov __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed, 2280dc98ff72SIlya Dryomov req->r_t.spgid.pgid.pool, req->r_t.spgid.pgid.seed, 228104c7d789SIlya Dryomov req->r_t.spgid.shard, osd->o_osd, req->r_t.epoch, req->r_flags, 228204c7d789SIlya Dryomov req->r_attempts); 2283bb873b53SIlya Dryomov 2284bb873b53SIlya Dryomov req->r_t.paused = false; 22853d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 2286bb873b53SIlya Dryomov req->r_attempts++; 22873d14c5d2SYehuda Sadeh 2288bb873b53SIlya Dryomov req->r_sent = osd->o_incarnation; 2289bb873b53SIlya Dryomov req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 2290bb873b53SIlya Dryomov ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request)); 22913d14c5d2SYehuda Sadeh } 22923d14c5d2SYehuda Sadeh 229342c1b124SIlya Dryomov static void maybe_request_map(struct ceph_osd_client *osdc) 229442c1b124SIlya Dryomov { 229542c1b124SIlya Dryomov bool continuous = false; 229642c1b124SIlya Dryomov 22975aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 229842c1b124SIlya Dryomov WARN_ON(!osdc->osdmap->epoch); 229942c1b124SIlya Dryomov 2300b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2301b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD) || 2302b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 230342c1b124SIlya Dryomov dout("%s osdc %p continuous\n", __func__, osdc); 230442c1b124SIlya Dryomov continuous = true; 230542c1b124SIlya Dryomov } else { 230642c1b124SIlya Dryomov dout("%s osdc %p onetime\n", __func__, osdc); 230742c1b124SIlya Dryomov } 230842c1b124SIlya Dryomov 230942c1b124SIlya Dryomov if (ceph_monc_want_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 231042c1b124SIlya Dryomov osdc->osdmap->epoch + 1, continuous)) 231142c1b124SIlya Dryomov ceph_monc_renew_subs(&osdc->client->monc); 231242c1b124SIlya Dryomov } 231342c1b124SIlya Dryomov 2314a1f4020aSJeff Layton static void complete_request(struct ceph_osd_request *req, int err); 23154609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req); 23164609245eSIlya Dryomov 23175aea3dcdSIlya Dryomov static void __submit_request(struct ceph_osd_request *req, bool wrlocked) 23180bbfdfe8SIlya Dryomov { 23195aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 23205aea3dcdSIlya Dryomov struct ceph_osd *osd; 23214609245eSIlya Dryomov enum calc_target_result ct_res; 232266850df5SIlya Dryomov int err = 0; 23235aea3dcdSIlya Dryomov bool need_send = false; 23245aea3dcdSIlya Dryomov bool promoted = false; 23250bbfdfe8SIlya Dryomov 2326b18b9550SIlya Dryomov WARN_ON(req->r_tid); 23275aea3dcdSIlya Dryomov dout("%s req %p wrlocked %d\n", __func__, req, wrlocked); 23285aea3dcdSIlya Dryomov 23295aea3dcdSIlya Dryomov again: 23308edf84baSIlya Dryomov ct_res = calc_target(osdc, &req->r_t, false); 23314609245eSIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE && !wrlocked) 23324609245eSIlya Dryomov goto promote; 23334609245eSIlya Dryomov 23345aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, wrlocked); 23355aea3dcdSIlya Dryomov if (IS_ERR(osd)) { 23365aea3dcdSIlya Dryomov WARN_ON(PTR_ERR(osd) != -EAGAIN || wrlocked); 23375aea3dcdSIlya Dryomov goto promote; 23385aea3dcdSIlya Dryomov } 23395aea3dcdSIlya Dryomov 234066850df5SIlya Dryomov if (osdc->abort_err) { 234166850df5SIlya Dryomov dout("req %p abort_err %d\n", req, osdc->abort_err); 234266850df5SIlya Dryomov err = osdc->abort_err; 234366850df5SIlya Dryomov } else if (osdc->osdmap->epoch < osdc->epoch_barrier) { 234458eb7932SJeff Layton dout("req %p epoch %u barrier %u\n", req, osdc->osdmap->epoch, 234558eb7932SJeff Layton osdc->epoch_barrier); 234658eb7932SJeff Layton req->r_t.paused = true; 234758eb7932SJeff Layton maybe_request_map(osdc); 234858eb7932SJeff Layton } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 2349b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 23505aea3dcdSIlya Dryomov dout("req %p pausewr\n", req); 23515aea3dcdSIlya Dryomov req->r_t.paused = true; 23525aea3dcdSIlya Dryomov maybe_request_map(osdc); 23535aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_READ) && 2354b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 23555aea3dcdSIlya Dryomov dout("req %p pauserd\n", req); 23565aea3dcdSIlya Dryomov req->r_t.paused = true; 23575aea3dcdSIlya Dryomov maybe_request_map(osdc); 23585aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 23595aea3dcdSIlya Dryomov !(req->r_flags & (CEPH_OSD_FLAG_FULL_TRY | 23605aea3dcdSIlya Dryomov CEPH_OSD_FLAG_FULL_FORCE)) && 2361b7ec35b3SIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 23625aea3dcdSIlya Dryomov pool_full(osdc, req->r_t.base_oloc.pool))) { 23635aea3dcdSIlya Dryomov dout("req %p full/pool_full\n", req); 236402b2f549SDongsheng Yang if (ceph_test_opt(osdc->client, ABORT_ON_FULL)) { 236529e87820SIlya Dryomov err = -ENOSPC; 236629e87820SIlya Dryomov } else { 2367dc9b0dc4SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL)) 2368dc9b0dc4SIlya Dryomov pr_warn_ratelimited("cluster is full (osdmap FULL)\n"); 2369dc9b0dc4SIlya Dryomov else 2370dc9b0dc4SIlya Dryomov pr_warn_ratelimited("pool %lld is full or reached quota\n", 2371dc9b0dc4SIlya Dryomov req->r_t.base_oloc.pool); 23725aea3dcdSIlya Dryomov req->r_t.paused = true; 23735aea3dcdSIlya Dryomov maybe_request_map(osdc); 237429e87820SIlya Dryomov } 23755aea3dcdSIlya Dryomov } else if (!osd_homeless(osd)) { 23765aea3dcdSIlya Dryomov need_send = true; 23770bbfdfe8SIlya Dryomov } else { 23785aea3dcdSIlya Dryomov maybe_request_map(osdc); 23790bbfdfe8SIlya Dryomov } 23800bbfdfe8SIlya Dryomov 23815aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 23825aea3dcdSIlya Dryomov /* 23835aea3dcdSIlya Dryomov * Assign the tid atomically with send_request() to protect 23845aea3dcdSIlya Dryomov * multiple writes to the same object from racing with each 23855aea3dcdSIlya Dryomov * other, resulting in out of order ops on the OSDs. 23865aea3dcdSIlya Dryomov */ 23875aea3dcdSIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 23885aea3dcdSIlya Dryomov link_request(osd, req); 23895aea3dcdSIlya Dryomov if (need_send) 23905aea3dcdSIlya Dryomov send_request(req); 239166850df5SIlya Dryomov else if (err) 239266850df5SIlya Dryomov complete_request(req, err); 23935aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 23945aea3dcdSIlya Dryomov 23956001567cSIlya Dryomov if (!err && ct_res == CALC_TARGET_POOL_DNE) 23964609245eSIlya Dryomov send_map_check(req); 23974609245eSIlya Dryomov 23985aea3dcdSIlya Dryomov if (promoted) 23995aea3dcdSIlya Dryomov downgrade_write(&osdc->lock); 24005aea3dcdSIlya Dryomov return; 24015aea3dcdSIlya Dryomov 24025aea3dcdSIlya Dryomov promote: 24035aea3dcdSIlya Dryomov up_read(&osdc->lock); 24045aea3dcdSIlya Dryomov down_write(&osdc->lock); 24055aea3dcdSIlya Dryomov wrlocked = true; 24065aea3dcdSIlya Dryomov promoted = true; 24075aea3dcdSIlya Dryomov goto again; 24080bbfdfe8SIlya Dryomov } 24090bbfdfe8SIlya Dryomov 24105aea3dcdSIlya Dryomov static void account_request(struct ceph_osd_request *req) 24115aea3dcdSIlya Dryomov { 241254ea0046SIlya Dryomov WARN_ON(req->r_flags & (CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK)); 2413b18b9550SIlya Dryomov WARN_ON(!(req->r_flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE))); 24145aea3dcdSIlya Dryomov 2415b18b9550SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_ONDISK; 241622d2cfdfSIlya Dryomov atomic_inc(&req->r_osdc->num_requests); 24177cc5e38fSIlya Dryomov 24187cc5e38fSIlya Dryomov req->r_start_stamp = jiffies; 241997e27aaaSXiubo Li req->r_start_latency = ktime_get(); 24205aea3dcdSIlya Dryomov } 24215aea3dcdSIlya Dryomov 24225aea3dcdSIlya Dryomov static void submit_request(struct ceph_osd_request *req, bool wrlocked) 24235aea3dcdSIlya Dryomov { 24245aea3dcdSIlya Dryomov ceph_osdc_get_request(req); 24255aea3dcdSIlya Dryomov account_request(req); 24265aea3dcdSIlya Dryomov __submit_request(req, wrlocked); 24275aea3dcdSIlya Dryomov } 24285aea3dcdSIlya Dryomov 242945ee2c1dSIlya Dryomov static void finish_request(struct ceph_osd_request *req) 24305aea3dcdSIlya Dryomov { 24315aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 24325aea3dcdSIlya Dryomov 24334609245eSIlya Dryomov WARN_ON(lookup_request_mc(&osdc->map_checks, req->r_tid)); 243404c7d789SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 243504c7d789SIlya Dryomov 243697e27aaaSXiubo Li req->r_end_latency = ktime_get(); 243797e27aaaSXiubo Li 243804c7d789SIlya Dryomov if (req->r_osd) 243904c7d789SIlya Dryomov unlink_request(req->r_osd, req); 24405aea3dcdSIlya Dryomov atomic_dec(&osdc->num_requests); 24415aea3dcdSIlya Dryomov 24425aea3dcdSIlya Dryomov /* 24435aea3dcdSIlya Dryomov * If an OSD has failed or returned and a request has been sent 24445aea3dcdSIlya Dryomov * twice, it's possible to get a reply and end up here while the 24455aea3dcdSIlya Dryomov * request message is queued for delivery. We will ignore the 24465aea3dcdSIlya Dryomov * reply, so not a big deal, but better to try and catch it. 24475aea3dcdSIlya Dryomov */ 24485aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 24495aea3dcdSIlya Dryomov ceph_msg_revoke_incoming(req->r_reply); 24505aea3dcdSIlya Dryomov } 24515aea3dcdSIlya Dryomov 2452fe5da05eSIlya Dryomov static void __complete_request(struct ceph_osd_request *req) 2453fe5da05eSIlya Dryomov { 2454d75f773cSSakari Ailus dout("%s req %p tid %llu cb %ps result %d\n", __func__, req, 2455b18b9550SIlya Dryomov req->r_tid, req->r_callback, req->r_result); 245626df726bSIlya Dryomov 245726df726bSIlya Dryomov if (req->r_callback) 2458fe5da05eSIlya Dryomov req->r_callback(req); 245926df726bSIlya Dryomov complete_all(&req->r_completion); 246026df726bSIlya Dryomov ceph_osdc_put_request(req); 2461b18b9550SIlya Dryomov } 2462fe5da05eSIlya Dryomov 246388bc1922SIlya Dryomov static void complete_request_workfn(struct work_struct *work) 246488bc1922SIlya Dryomov { 246588bc1922SIlya Dryomov struct ceph_osd_request *req = 246688bc1922SIlya Dryomov container_of(work, struct ceph_osd_request, r_complete_work); 246788bc1922SIlya Dryomov 246888bc1922SIlya Dryomov __complete_request(req); 24690bbfdfe8SIlya Dryomov } 24700bbfdfe8SIlya Dryomov 24714609245eSIlya Dryomov /* 2472b18b9550SIlya Dryomov * This is open-coded in handle_reply(). 24734609245eSIlya Dryomov */ 24744609245eSIlya Dryomov static void complete_request(struct ceph_osd_request *req, int err) 24754609245eSIlya Dryomov { 24764609245eSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 24774609245eSIlya Dryomov 24784609245eSIlya Dryomov req->r_result = err; 247945ee2c1dSIlya Dryomov finish_request(req); 248088bc1922SIlya Dryomov 248188bc1922SIlya Dryomov INIT_WORK(&req->r_complete_work, complete_request_workfn); 248288bc1922SIlya Dryomov queue_work(req->r_osdc->completion_wq, &req->r_complete_work); 24834609245eSIlya Dryomov } 24844609245eSIlya Dryomov 24854609245eSIlya Dryomov static void cancel_map_check(struct ceph_osd_request *req) 24864609245eSIlya Dryomov { 24874609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 24884609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 24894609245eSIlya Dryomov 24904609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 24914609245eSIlya Dryomov 24924609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 24934609245eSIlya Dryomov if (!lookup_req) 24944609245eSIlya Dryomov return; 24954609245eSIlya Dryomov 24964609245eSIlya Dryomov WARN_ON(lookup_req != req); 24974609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 24984609245eSIlya Dryomov ceph_osdc_put_request(req); 24994609245eSIlya Dryomov } 25004609245eSIlya Dryomov 25015aea3dcdSIlya Dryomov static void cancel_request(struct ceph_osd_request *req) 25025aea3dcdSIlya Dryomov { 25035aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 25045aea3dcdSIlya Dryomov 25054609245eSIlya Dryomov cancel_map_check(req); 250645ee2c1dSIlya Dryomov finish_request(req); 2507b18b9550SIlya Dryomov complete_all(&req->r_completion); 2508c297eb42SIlya Dryomov ceph_osdc_put_request(req); 25095aea3dcdSIlya Dryomov } 25105aea3dcdSIlya Dryomov 25117cc5e38fSIlya Dryomov static void abort_request(struct ceph_osd_request *req, int err) 25127cc5e38fSIlya Dryomov { 25137cc5e38fSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 25147cc5e38fSIlya Dryomov 25157cc5e38fSIlya Dryomov cancel_map_check(req); 25167cc5e38fSIlya Dryomov complete_request(req, err); 25177cc5e38fSIlya Dryomov } 25187cc5e38fSIlya Dryomov 251966850df5SIlya Dryomov static int abort_fn(struct ceph_osd_request *req, void *arg) 252066850df5SIlya Dryomov { 252166850df5SIlya Dryomov int err = *(int *)arg; 252266850df5SIlya Dryomov 252366850df5SIlya Dryomov abort_request(req, err); 252466850df5SIlya Dryomov return 0; /* continue iteration */ 252566850df5SIlya Dryomov } 252666850df5SIlya Dryomov 252766850df5SIlya Dryomov /* 252866850df5SIlya Dryomov * Abort all in-flight requests with @err and arrange for all future 252966850df5SIlya Dryomov * requests to be failed immediately. 253066850df5SIlya Dryomov */ 253166850df5SIlya Dryomov void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err) 253266850df5SIlya Dryomov { 253366850df5SIlya Dryomov dout("%s osdc %p err %d\n", __func__, osdc, err); 253466850df5SIlya Dryomov down_write(&osdc->lock); 253566850df5SIlya Dryomov for_each_request(osdc, abort_fn, &err); 253666850df5SIlya Dryomov osdc->abort_err = err; 253766850df5SIlya Dryomov up_write(&osdc->lock); 253866850df5SIlya Dryomov } 253966850df5SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_abort_requests); 254066850df5SIlya Dryomov 25412cef0ba8SYan, Zheng void ceph_osdc_clear_abort_err(struct ceph_osd_client *osdc) 25422cef0ba8SYan, Zheng { 25432cef0ba8SYan, Zheng down_write(&osdc->lock); 25442cef0ba8SYan, Zheng osdc->abort_err = 0; 25452cef0ba8SYan, Zheng up_write(&osdc->lock); 25462cef0ba8SYan, Zheng } 25472cef0ba8SYan, Zheng EXPORT_SYMBOL(ceph_osdc_clear_abort_err); 25482cef0ba8SYan, Zheng 254958eb7932SJeff Layton static void update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 255058eb7932SJeff Layton { 255158eb7932SJeff Layton if (likely(eb > osdc->epoch_barrier)) { 255258eb7932SJeff Layton dout("updating epoch_barrier from %u to %u\n", 255358eb7932SJeff Layton osdc->epoch_barrier, eb); 255458eb7932SJeff Layton osdc->epoch_barrier = eb; 255558eb7932SJeff Layton /* Request map if we're not to the barrier yet */ 255658eb7932SJeff Layton if (eb > osdc->osdmap->epoch) 255758eb7932SJeff Layton maybe_request_map(osdc); 255858eb7932SJeff Layton } 255958eb7932SJeff Layton } 256058eb7932SJeff Layton 256158eb7932SJeff Layton void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 256258eb7932SJeff Layton { 256358eb7932SJeff Layton down_read(&osdc->lock); 256458eb7932SJeff Layton if (unlikely(eb > osdc->epoch_barrier)) { 256558eb7932SJeff Layton up_read(&osdc->lock); 256658eb7932SJeff Layton down_write(&osdc->lock); 256758eb7932SJeff Layton update_epoch_barrier(osdc, eb); 256858eb7932SJeff Layton up_write(&osdc->lock); 256958eb7932SJeff Layton } else { 257058eb7932SJeff Layton up_read(&osdc->lock); 257158eb7932SJeff Layton } 257258eb7932SJeff Layton } 257358eb7932SJeff Layton EXPORT_SYMBOL(ceph_osdc_update_epoch_barrier); 257458eb7932SJeff Layton 2575fc36d0a4SJeff Layton /* 25764eea0fefSIlya Dryomov * We can end up releasing caps as a result of abort_request(). 25774eea0fefSIlya Dryomov * In that case, we probably want to ensure that the cap release message 25784eea0fefSIlya Dryomov * has an updated epoch barrier in it, so set the epoch barrier prior to 25794eea0fefSIlya Dryomov * aborting the first request. 25804eea0fefSIlya Dryomov */ 25814eea0fefSIlya Dryomov static int abort_on_full_fn(struct ceph_osd_request *req, void *arg) 25824eea0fefSIlya Dryomov { 25834eea0fefSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 25844eea0fefSIlya Dryomov bool *victims = arg; 25854eea0fefSIlya Dryomov 2586c843d13cSIlya Dryomov if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 25874eea0fefSIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2588690f951dSIlya Dryomov pool_full(osdc, req->r_t.base_oloc.pool))) { 25894eea0fefSIlya Dryomov if (!*victims) { 25904eea0fefSIlya Dryomov update_epoch_barrier(osdc, osdc->osdmap->epoch); 25914eea0fefSIlya Dryomov *victims = true; 25924eea0fefSIlya Dryomov } 25934eea0fefSIlya Dryomov abort_request(req, -ENOSPC); 25944eea0fefSIlya Dryomov } 25954eea0fefSIlya Dryomov 25964eea0fefSIlya Dryomov return 0; /* continue iteration */ 25974eea0fefSIlya Dryomov } 25984eea0fefSIlya Dryomov 25994eea0fefSIlya Dryomov /* 2600fc36d0a4SJeff Layton * Drop all pending requests that are stalled waiting on a full condition to 260158eb7932SJeff Layton * clear, and complete them with ENOSPC as the return code. Set the 260258eb7932SJeff Layton * osdc->epoch_barrier to the latest map epoch that we've seen if any were 260358eb7932SJeff Layton * cancelled. 2604fc36d0a4SJeff Layton */ 2605fc36d0a4SJeff Layton static void ceph_osdc_abort_on_full(struct ceph_osd_client *osdc) 2606fc36d0a4SJeff Layton { 260758eb7932SJeff Layton bool victims = false; 2608fc36d0a4SJeff Layton 260902b2f549SDongsheng Yang if (ceph_test_opt(osdc->client, ABORT_ON_FULL) && 2610c843d13cSIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || have_pool_full(osdc))) 26114eea0fefSIlya Dryomov for_each_request(osdc, abort_on_full_fn, &victims); 2612fc36d0a4SJeff Layton } 2613fc36d0a4SJeff Layton 26144609245eSIlya Dryomov static void check_pool_dne(struct ceph_osd_request *req) 26154609245eSIlya Dryomov { 26164609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 26174609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 26184609245eSIlya Dryomov 26194609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 26204609245eSIlya Dryomov WARN_ON(!map->epoch); 26214609245eSIlya Dryomov 26224609245eSIlya Dryomov if (req->r_attempts) { 26234609245eSIlya Dryomov /* 26244609245eSIlya Dryomov * We sent a request earlier, which means that 26254609245eSIlya Dryomov * previously the pool existed, and now it does not 26264609245eSIlya Dryomov * (i.e., it was deleted). 26274609245eSIlya Dryomov */ 26284609245eSIlya Dryomov req->r_map_dne_bound = map->epoch; 26294609245eSIlya Dryomov dout("%s req %p tid %llu pool disappeared\n", __func__, req, 26304609245eSIlya Dryomov req->r_tid); 26314609245eSIlya Dryomov } else { 26324609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u have %u\n", __func__, 26334609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, map->epoch); 26344609245eSIlya Dryomov } 26354609245eSIlya Dryomov 26364609245eSIlya Dryomov if (req->r_map_dne_bound) { 26374609245eSIlya Dryomov if (map->epoch >= req->r_map_dne_bound) { 26384609245eSIlya Dryomov /* we had a new enough map */ 26394609245eSIlya Dryomov pr_info_ratelimited("tid %llu pool does not exist\n", 26404609245eSIlya Dryomov req->r_tid); 26414609245eSIlya Dryomov complete_request(req, -ENOENT); 26424609245eSIlya Dryomov } 26434609245eSIlya Dryomov } else { 26444609245eSIlya Dryomov send_map_check(req); 26454609245eSIlya Dryomov } 26464609245eSIlya Dryomov } 26474609245eSIlya Dryomov 26484609245eSIlya Dryomov static void map_check_cb(struct ceph_mon_generic_request *greq) 26494609245eSIlya Dryomov { 26504609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 26514609245eSIlya Dryomov struct ceph_osd_request *req; 26524609245eSIlya Dryomov u64 tid = greq->private_data; 26534609245eSIlya Dryomov 26544609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 26554609245eSIlya Dryomov 26564609245eSIlya Dryomov down_write(&osdc->lock); 26574609245eSIlya Dryomov req = lookup_request_mc(&osdc->map_checks, tid); 26584609245eSIlya Dryomov if (!req) { 26594609245eSIlya Dryomov dout("%s tid %llu dne\n", __func__, tid); 26604609245eSIlya Dryomov goto out_unlock; 26614609245eSIlya Dryomov } 26624609245eSIlya Dryomov 26634609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u newest %llu\n", __func__, 26644609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, greq->u.newest); 26654609245eSIlya Dryomov if (!req->r_map_dne_bound) 26664609245eSIlya Dryomov req->r_map_dne_bound = greq->u.newest; 26674609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 26684609245eSIlya Dryomov check_pool_dne(req); 26694609245eSIlya Dryomov 26704609245eSIlya Dryomov ceph_osdc_put_request(req); 26714609245eSIlya Dryomov out_unlock: 26724609245eSIlya Dryomov up_write(&osdc->lock); 26734609245eSIlya Dryomov } 26744609245eSIlya Dryomov 26754609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req) 26764609245eSIlya Dryomov { 26774609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 26784609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 26794609245eSIlya Dryomov int ret; 26804609245eSIlya Dryomov 26814609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 26824609245eSIlya Dryomov 26834609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 26844609245eSIlya Dryomov if (lookup_req) { 26854609245eSIlya Dryomov WARN_ON(lookup_req != req); 26864609245eSIlya Dryomov return; 26874609245eSIlya Dryomov } 26884609245eSIlya Dryomov 26894609245eSIlya Dryomov ceph_osdc_get_request(req); 26904609245eSIlya Dryomov insert_request_mc(&osdc->map_checks, req); 26914609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 26924609245eSIlya Dryomov map_check_cb, req->r_tid); 26934609245eSIlya Dryomov WARN_ON(ret); 26944609245eSIlya Dryomov } 26954609245eSIlya Dryomov 26960bbfdfe8SIlya Dryomov /* 2697922dab61SIlya Dryomov * lingering requests, watch/notify v2 infrastructure 2698922dab61SIlya Dryomov */ 2699922dab61SIlya Dryomov static void linger_release(struct kref *kref) 2700922dab61SIlya Dryomov { 2701922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 2702922dab61SIlya Dryomov container_of(kref, struct ceph_osd_linger_request, kref); 2703922dab61SIlya Dryomov 2704922dab61SIlya Dryomov dout("%s lreq %p reg_req %p ping_req %p\n", __func__, lreq, 2705922dab61SIlya Dryomov lreq->reg_req, lreq->ping_req); 2706922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->node)); 2707922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->osdc_node)); 27084609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->mc_node)); 2709922dab61SIlya Dryomov WARN_ON(!list_empty(&lreq->scan_item)); 2710b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lreq->pending_lworks)); 2711922dab61SIlya Dryomov WARN_ON(lreq->osd); 2712922dab61SIlya Dryomov 271375dbb685SIlya Dryomov if (lreq->request_pl) 271475dbb685SIlya Dryomov ceph_pagelist_release(lreq->request_pl); 271575dbb685SIlya Dryomov if (lreq->notify_id_pages) 271675dbb685SIlya Dryomov ceph_release_page_vector(lreq->notify_id_pages, 1); 271775dbb685SIlya Dryomov 2718922dab61SIlya Dryomov ceph_osdc_put_request(lreq->reg_req); 2719922dab61SIlya Dryomov ceph_osdc_put_request(lreq->ping_req); 2720922dab61SIlya Dryomov target_destroy(&lreq->t); 2721922dab61SIlya Dryomov kfree(lreq); 2722922dab61SIlya Dryomov } 2723922dab61SIlya Dryomov 2724922dab61SIlya Dryomov static void linger_put(struct ceph_osd_linger_request *lreq) 2725922dab61SIlya Dryomov { 2726922dab61SIlya Dryomov if (lreq) 2727922dab61SIlya Dryomov kref_put(&lreq->kref, linger_release); 2728922dab61SIlya Dryomov } 2729922dab61SIlya Dryomov 2730922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2731922dab61SIlya Dryomov linger_get(struct ceph_osd_linger_request *lreq) 2732922dab61SIlya Dryomov { 2733922dab61SIlya Dryomov kref_get(&lreq->kref); 2734922dab61SIlya Dryomov return lreq; 2735922dab61SIlya Dryomov } 2736922dab61SIlya Dryomov 2737922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2738922dab61SIlya Dryomov linger_alloc(struct ceph_osd_client *osdc) 2739922dab61SIlya Dryomov { 2740922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2741922dab61SIlya Dryomov 2742922dab61SIlya Dryomov lreq = kzalloc(sizeof(*lreq), GFP_NOIO); 2743922dab61SIlya Dryomov if (!lreq) 2744922dab61SIlya Dryomov return NULL; 2745922dab61SIlya Dryomov 2746922dab61SIlya Dryomov kref_init(&lreq->kref); 2747922dab61SIlya Dryomov mutex_init(&lreq->lock); 2748922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->node); 2749922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->osdc_node); 27504609245eSIlya Dryomov RB_CLEAR_NODE(&lreq->mc_node); 2751922dab61SIlya Dryomov INIT_LIST_HEAD(&lreq->scan_item); 2752b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lreq->pending_lworks); 2753922dab61SIlya Dryomov init_completion(&lreq->reg_commit_wait); 275419079203SIlya Dryomov init_completion(&lreq->notify_finish_wait); 2755922dab61SIlya Dryomov 2756922dab61SIlya Dryomov lreq->osdc = osdc; 2757922dab61SIlya Dryomov target_init(&lreq->t); 2758922dab61SIlya Dryomov 2759922dab61SIlya Dryomov dout("%s lreq %p\n", __func__, lreq); 2760922dab61SIlya Dryomov return lreq; 2761922dab61SIlya Dryomov } 2762922dab61SIlya Dryomov 2763922dab61SIlya Dryomov DEFINE_RB_INSDEL_FUNCS(linger, struct ceph_osd_linger_request, linger_id, node) 2764922dab61SIlya Dryomov DEFINE_RB_FUNCS(linger_osdc, struct ceph_osd_linger_request, linger_id, osdc_node) 27654609245eSIlya Dryomov DEFINE_RB_FUNCS(linger_mc, struct ceph_osd_linger_request, linger_id, mc_node) 2766922dab61SIlya Dryomov 2767922dab61SIlya Dryomov /* 2768922dab61SIlya Dryomov * Create linger request <-> OSD session relation. 2769922dab61SIlya Dryomov * 2770922dab61SIlya Dryomov * @lreq has to be registered, @osd may be homeless. 2771922dab61SIlya Dryomov */ 2772922dab61SIlya Dryomov static void link_linger(struct ceph_osd *osd, 2773922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2774922dab61SIlya Dryomov { 2775922dab61SIlya Dryomov verify_osd_locked(osd); 2776922dab61SIlya Dryomov WARN_ON(!lreq->linger_id || lreq->osd); 2777922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2778922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2779922dab61SIlya Dryomov 2780922dab61SIlya Dryomov if (!osd_homeless(osd)) 2781922dab61SIlya Dryomov __remove_osd_from_lru(osd); 2782922dab61SIlya Dryomov else 2783922dab61SIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 2784922dab61SIlya Dryomov 2785922dab61SIlya Dryomov get_osd(osd); 2786922dab61SIlya Dryomov insert_linger(&osd->o_linger_requests, lreq); 2787922dab61SIlya Dryomov lreq->osd = osd; 2788922dab61SIlya Dryomov } 2789922dab61SIlya Dryomov 2790922dab61SIlya Dryomov static void unlink_linger(struct ceph_osd *osd, 2791922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2792922dab61SIlya Dryomov { 2793922dab61SIlya Dryomov verify_osd_locked(osd); 2794922dab61SIlya Dryomov WARN_ON(lreq->osd != osd); 2795922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2796922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2797922dab61SIlya Dryomov 2798922dab61SIlya Dryomov lreq->osd = NULL; 2799922dab61SIlya Dryomov erase_linger(&osd->o_linger_requests, lreq); 2800922dab61SIlya Dryomov put_osd(osd); 2801922dab61SIlya Dryomov 2802922dab61SIlya Dryomov if (!osd_homeless(osd)) 2803922dab61SIlya Dryomov maybe_move_osd_to_lru(osd); 2804922dab61SIlya Dryomov else 2805922dab61SIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 2806922dab61SIlya Dryomov } 2807922dab61SIlya Dryomov 2808922dab61SIlya Dryomov static bool __linger_registered(struct ceph_osd_linger_request *lreq) 2809922dab61SIlya Dryomov { 2810922dab61SIlya Dryomov verify_osdc_locked(lreq->osdc); 2811922dab61SIlya Dryomov 2812922dab61SIlya Dryomov return !RB_EMPTY_NODE(&lreq->osdc_node); 2813922dab61SIlya Dryomov } 2814922dab61SIlya Dryomov 2815922dab61SIlya Dryomov static bool linger_registered(struct ceph_osd_linger_request *lreq) 2816922dab61SIlya Dryomov { 2817922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2818922dab61SIlya Dryomov bool registered; 2819922dab61SIlya Dryomov 2820922dab61SIlya Dryomov down_read(&osdc->lock); 2821922dab61SIlya Dryomov registered = __linger_registered(lreq); 2822922dab61SIlya Dryomov up_read(&osdc->lock); 2823922dab61SIlya Dryomov 2824922dab61SIlya Dryomov return registered; 2825922dab61SIlya Dryomov } 2826922dab61SIlya Dryomov 2827922dab61SIlya Dryomov static void linger_register(struct ceph_osd_linger_request *lreq) 2828922dab61SIlya Dryomov { 2829922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2830922dab61SIlya Dryomov 2831922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2832922dab61SIlya Dryomov WARN_ON(lreq->linger_id); 2833922dab61SIlya Dryomov 2834922dab61SIlya Dryomov linger_get(lreq); 2835922dab61SIlya Dryomov lreq->linger_id = ++osdc->last_linger_id; 2836922dab61SIlya Dryomov insert_linger_osdc(&osdc->linger_requests, lreq); 2837922dab61SIlya Dryomov } 2838922dab61SIlya Dryomov 2839922dab61SIlya Dryomov static void linger_unregister(struct ceph_osd_linger_request *lreq) 2840922dab61SIlya Dryomov { 2841922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2842922dab61SIlya Dryomov 2843922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2844922dab61SIlya Dryomov 2845922dab61SIlya Dryomov erase_linger_osdc(&osdc->linger_requests, lreq); 2846922dab61SIlya Dryomov linger_put(lreq); 2847922dab61SIlya Dryomov } 2848922dab61SIlya Dryomov 2849922dab61SIlya Dryomov static void cancel_linger_request(struct ceph_osd_request *req) 2850922dab61SIlya Dryomov { 2851922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2852922dab61SIlya Dryomov 2853922dab61SIlya Dryomov WARN_ON(!req->r_linger); 2854922dab61SIlya Dryomov cancel_request(req); 2855922dab61SIlya Dryomov linger_put(lreq); 2856922dab61SIlya Dryomov } 2857922dab61SIlya Dryomov 2858922dab61SIlya Dryomov struct linger_work { 2859922dab61SIlya Dryomov struct work_struct work; 2860922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2861b07d3c4bSIlya Dryomov struct list_head pending_item; 2862b07d3c4bSIlya Dryomov unsigned long queued_stamp; 2863922dab61SIlya Dryomov 2864922dab61SIlya Dryomov union { 2865922dab61SIlya Dryomov struct { 2866922dab61SIlya Dryomov u64 notify_id; 2867922dab61SIlya Dryomov u64 notifier_id; 2868922dab61SIlya Dryomov void *payload; /* points into @msg front */ 2869922dab61SIlya Dryomov size_t payload_len; 2870922dab61SIlya Dryomov 2871922dab61SIlya Dryomov struct ceph_msg *msg; /* for ceph_msg_put() */ 2872922dab61SIlya Dryomov } notify; 2873922dab61SIlya Dryomov struct { 2874922dab61SIlya Dryomov int err; 2875922dab61SIlya Dryomov } error; 2876922dab61SIlya Dryomov }; 2877922dab61SIlya Dryomov }; 2878922dab61SIlya Dryomov 2879922dab61SIlya Dryomov static struct linger_work *lwork_alloc(struct ceph_osd_linger_request *lreq, 2880922dab61SIlya Dryomov work_func_t workfn) 2881922dab61SIlya Dryomov { 2882922dab61SIlya Dryomov struct linger_work *lwork; 2883922dab61SIlya Dryomov 2884922dab61SIlya Dryomov lwork = kzalloc(sizeof(*lwork), GFP_NOIO); 2885922dab61SIlya Dryomov if (!lwork) 2886922dab61SIlya Dryomov return NULL; 2887922dab61SIlya Dryomov 2888922dab61SIlya Dryomov INIT_WORK(&lwork->work, workfn); 2889b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lwork->pending_item); 2890922dab61SIlya Dryomov lwork->lreq = linger_get(lreq); 2891922dab61SIlya Dryomov 2892922dab61SIlya Dryomov return lwork; 2893922dab61SIlya Dryomov } 2894922dab61SIlya Dryomov 2895922dab61SIlya Dryomov static void lwork_free(struct linger_work *lwork) 2896922dab61SIlya Dryomov { 2897922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2898922dab61SIlya Dryomov 2899b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 2900b07d3c4bSIlya Dryomov list_del(&lwork->pending_item); 2901b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 2902b07d3c4bSIlya Dryomov 2903922dab61SIlya Dryomov linger_put(lreq); 2904922dab61SIlya Dryomov kfree(lwork); 2905922dab61SIlya Dryomov } 2906922dab61SIlya Dryomov 2907922dab61SIlya Dryomov static void lwork_queue(struct linger_work *lwork) 2908922dab61SIlya Dryomov { 2909922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2910922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2911922dab61SIlya Dryomov 2912922dab61SIlya Dryomov verify_lreq_locked(lreq); 2913b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lwork->pending_item)); 2914b07d3c4bSIlya Dryomov 2915b07d3c4bSIlya Dryomov lwork->queued_stamp = jiffies; 2916b07d3c4bSIlya Dryomov list_add_tail(&lwork->pending_item, &lreq->pending_lworks); 2917922dab61SIlya Dryomov queue_work(osdc->notify_wq, &lwork->work); 2918922dab61SIlya Dryomov } 2919922dab61SIlya Dryomov 2920922dab61SIlya Dryomov static void do_watch_notify(struct work_struct *w) 2921922dab61SIlya Dryomov { 2922922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2923922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2924922dab61SIlya Dryomov 2925922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2926922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2927922dab61SIlya Dryomov goto out; 2928922dab61SIlya Dryomov } 2929922dab61SIlya Dryomov 293019079203SIlya Dryomov WARN_ON(!lreq->is_watch); 2931922dab61SIlya Dryomov dout("%s lreq %p notify_id %llu notifier_id %llu payload_len %zu\n", 2932922dab61SIlya Dryomov __func__, lreq, lwork->notify.notify_id, lwork->notify.notifier_id, 2933922dab61SIlya Dryomov lwork->notify.payload_len); 2934922dab61SIlya Dryomov lreq->wcb(lreq->data, lwork->notify.notify_id, lreq->linger_id, 2935922dab61SIlya Dryomov lwork->notify.notifier_id, lwork->notify.payload, 2936922dab61SIlya Dryomov lwork->notify.payload_len); 2937922dab61SIlya Dryomov 2938922dab61SIlya Dryomov out: 2939922dab61SIlya Dryomov ceph_msg_put(lwork->notify.msg); 2940922dab61SIlya Dryomov lwork_free(lwork); 2941922dab61SIlya Dryomov } 2942922dab61SIlya Dryomov 2943922dab61SIlya Dryomov static void do_watch_error(struct work_struct *w) 2944922dab61SIlya Dryomov { 2945922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2946922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2947922dab61SIlya Dryomov 2948922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2949922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2950922dab61SIlya Dryomov goto out; 2951922dab61SIlya Dryomov } 2952922dab61SIlya Dryomov 2953922dab61SIlya Dryomov dout("%s lreq %p err %d\n", __func__, lreq, lwork->error.err); 2954922dab61SIlya Dryomov lreq->errcb(lreq->data, lreq->linger_id, lwork->error.err); 2955922dab61SIlya Dryomov 2956922dab61SIlya Dryomov out: 2957922dab61SIlya Dryomov lwork_free(lwork); 2958922dab61SIlya Dryomov } 2959922dab61SIlya Dryomov 2960922dab61SIlya Dryomov static void queue_watch_error(struct ceph_osd_linger_request *lreq) 2961922dab61SIlya Dryomov { 2962922dab61SIlya Dryomov struct linger_work *lwork; 2963922dab61SIlya Dryomov 2964922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_error); 2965922dab61SIlya Dryomov if (!lwork) { 2966922dab61SIlya Dryomov pr_err("failed to allocate error-lwork\n"); 2967922dab61SIlya Dryomov return; 2968922dab61SIlya Dryomov } 2969922dab61SIlya Dryomov 2970922dab61SIlya Dryomov lwork->error.err = lreq->last_error; 2971922dab61SIlya Dryomov lwork_queue(lwork); 2972922dab61SIlya Dryomov } 2973922dab61SIlya Dryomov 2974922dab61SIlya Dryomov static void linger_reg_commit_complete(struct ceph_osd_linger_request *lreq, 2975922dab61SIlya Dryomov int result) 2976922dab61SIlya Dryomov { 2977922dab61SIlya Dryomov if (!completion_done(&lreq->reg_commit_wait)) { 2978922dab61SIlya Dryomov lreq->reg_commit_error = (result <= 0 ? result : 0); 2979922dab61SIlya Dryomov complete_all(&lreq->reg_commit_wait); 2980922dab61SIlya Dryomov } 2981922dab61SIlya Dryomov } 2982922dab61SIlya Dryomov 2983922dab61SIlya Dryomov static void linger_commit_cb(struct ceph_osd_request *req) 2984922dab61SIlya Dryomov { 2985922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2986922dab61SIlya Dryomov 2987922dab61SIlya Dryomov mutex_lock(&lreq->lock); 298875dbb685SIlya Dryomov if (req != lreq->reg_req) { 298975dbb685SIlya Dryomov dout("%s lreq %p linger_id %llu unknown req (%p != %p)\n", 299075dbb685SIlya Dryomov __func__, lreq, lreq->linger_id, req, lreq->reg_req); 299175dbb685SIlya Dryomov goto out; 299275dbb685SIlya Dryomov } 299375dbb685SIlya Dryomov 2994922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d\n", __func__, lreq, 2995922dab61SIlya Dryomov lreq->linger_id, req->r_result); 2996922dab61SIlya Dryomov linger_reg_commit_complete(lreq, req->r_result); 2997922dab61SIlya Dryomov lreq->committed = true; 2998922dab61SIlya Dryomov 299919079203SIlya Dryomov if (!lreq->is_watch) { 300019079203SIlya Dryomov struct ceph_osd_data *osd_data = 300119079203SIlya Dryomov osd_req_op_data(req, 0, notify, response_data); 300219079203SIlya Dryomov void *p = page_address(osd_data->pages[0]); 300319079203SIlya Dryomov 300419079203SIlya Dryomov WARN_ON(req->r_ops[0].op != CEPH_OSD_OP_NOTIFY || 300519079203SIlya Dryomov osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 300619079203SIlya Dryomov 300719079203SIlya Dryomov /* make note of the notify_id */ 300819079203SIlya Dryomov if (req->r_ops[0].outdata_len >= sizeof(u64)) { 300919079203SIlya Dryomov lreq->notify_id = ceph_decode_64(&p); 301019079203SIlya Dryomov dout("lreq %p notify_id %llu\n", lreq, 301119079203SIlya Dryomov lreq->notify_id); 301219079203SIlya Dryomov } else { 301319079203SIlya Dryomov dout("lreq %p no notify_id\n", lreq); 301419079203SIlya Dryomov } 301519079203SIlya Dryomov } 301619079203SIlya Dryomov 301775dbb685SIlya Dryomov out: 3018922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3019922dab61SIlya Dryomov linger_put(lreq); 3020922dab61SIlya Dryomov } 3021922dab61SIlya Dryomov 3022922dab61SIlya Dryomov static int normalize_watch_error(int err) 3023922dab61SIlya Dryomov { 3024922dab61SIlya Dryomov /* 3025922dab61SIlya Dryomov * Translate ENOENT -> ENOTCONN so that a delete->disconnection 3026922dab61SIlya Dryomov * notification and a failure to reconnect because we raced with 3027922dab61SIlya Dryomov * the delete appear the same to the user. 3028922dab61SIlya Dryomov */ 3029922dab61SIlya Dryomov if (err == -ENOENT) 3030922dab61SIlya Dryomov err = -ENOTCONN; 3031922dab61SIlya Dryomov 3032922dab61SIlya Dryomov return err; 3033922dab61SIlya Dryomov } 3034922dab61SIlya Dryomov 3035922dab61SIlya Dryomov static void linger_reconnect_cb(struct ceph_osd_request *req) 3036922dab61SIlya Dryomov { 3037922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 3038922dab61SIlya Dryomov 3039922dab61SIlya Dryomov mutex_lock(&lreq->lock); 304075dbb685SIlya Dryomov if (req != lreq->reg_req) { 304175dbb685SIlya Dryomov dout("%s lreq %p linger_id %llu unknown req (%p != %p)\n", 304275dbb685SIlya Dryomov __func__, lreq, lreq->linger_id, req, lreq->reg_req); 304375dbb685SIlya Dryomov goto out; 304475dbb685SIlya Dryomov } 304575dbb685SIlya Dryomov 3046922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d last_error %d\n", __func__, 3047922dab61SIlya Dryomov lreq, lreq->linger_id, req->r_result, lreq->last_error); 3048922dab61SIlya Dryomov if (req->r_result < 0) { 3049922dab61SIlya Dryomov if (!lreq->last_error) { 3050922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 3051922dab61SIlya Dryomov queue_watch_error(lreq); 3052922dab61SIlya Dryomov } 3053922dab61SIlya Dryomov } 3054922dab61SIlya Dryomov 305575dbb685SIlya Dryomov out: 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 { 306275dbb685SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 306375dbb685SIlya Dryomov struct ceph_osd_request *req; 306475dbb685SIlya Dryomov int ret; 3065922dab61SIlya Dryomov 306675dbb685SIlya Dryomov verify_osdc_wrlocked(osdc); 306775dbb685SIlya Dryomov mutex_lock(&lreq->lock); 3068922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 3069922dab61SIlya Dryomov 307075dbb685SIlya Dryomov if (lreq->reg_req) { 307175dbb685SIlya Dryomov if (lreq->reg_req->r_osd) 307275dbb685SIlya Dryomov cancel_linger_request(lreq->reg_req); 307375dbb685SIlya Dryomov ceph_osdc_put_request(lreq->reg_req); 307475dbb685SIlya Dryomov } 3075922dab61SIlya Dryomov 307675dbb685SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, true, GFP_NOIO); 307775dbb685SIlya Dryomov BUG_ON(!req); 307875dbb685SIlya Dryomov 30795133ba8fSIlya Dryomov target_copy(&req->r_t, &lreq->t); 3080922dab61SIlya Dryomov req->r_mtime = lreq->mtime; 3081922dab61SIlya Dryomov 308219079203SIlya Dryomov if (lreq->is_watch && lreq->committed) { 308375dbb685SIlya Dryomov osd_req_op_watch_init(req, 0, CEPH_OSD_WATCH_OP_RECONNECT, 308475dbb685SIlya Dryomov lreq->linger_id, ++lreq->register_gen); 3085922dab61SIlya Dryomov dout("lreq %p reconnect register_gen %u\n", lreq, 308675dbb685SIlya Dryomov req->r_ops[0].watch.gen); 3087922dab61SIlya Dryomov req->r_callback = linger_reconnect_cb; 3088922dab61SIlya Dryomov } else { 308975dbb685SIlya Dryomov if (lreq->is_watch) { 309075dbb685SIlya Dryomov osd_req_op_watch_init(req, 0, CEPH_OSD_WATCH_OP_WATCH, 309175dbb685SIlya Dryomov lreq->linger_id, 0); 309275dbb685SIlya Dryomov } else { 309319079203SIlya Dryomov lreq->notify_id = 0; 309475dbb685SIlya Dryomov 309575dbb685SIlya Dryomov refcount_inc(&lreq->request_pl->refcnt); 309675dbb685SIlya Dryomov osd_req_op_notify_init(req, 0, lreq->linger_id, 309775dbb685SIlya Dryomov lreq->request_pl); 309875dbb685SIlya Dryomov ceph_osd_data_pages_init( 309975dbb685SIlya Dryomov osd_req_op_data(req, 0, notify, response_data), 310075dbb685SIlya Dryomov lreq->notify_id_pages, PAGE_SIZE, 0, false, false); 310175dbb685SIlya Dryomov } 3102922dab61SIlya Dryomov dout("lreq %p register\n", lreq); 3103922dab61SIlya Dryomov req->r_callback = linger_commit_cb; 3104922dab61SIlya Dryomov } 310575dbb685SIlya Dryomov 310675dbb685SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 310775dbb685SIlya Dryomov BUG_ON(ret); 3108922dab61SIlya Dryomov 3109922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 3110922dab61SIlya Dryomov req->r_linger = true; 311175dbb685SIlya Dryomov lreq->reg_req = req; 311275dbb685SIlya Dryomov mutex_unlock(&lreq->lock); 3113922dab61SIlya Dryomov 3114922dab61SIlya Dryomov submit_request(req, true); 3115922dab61SIlya Dryomov } 3116922dab61SIlya Dryomov 3117922dab61SIlya Dryomov static void linger_ping_cb(struct ceph_osd_request *req) 3118922dab61SIlya Dryomov { 3119922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 3120922dab61SIlya Dryomov 3121922dab61SIlya Dryomov mutex_lock(&lreq->lock); 312275dbb685SIlya Dryomov if (req != lreq->ping_req) { 312375dbb685SIlya Dryomov dout("%s lreq %p linger_id %llu unknown req (%p != %p)\n", 312475dbb685SIlya Dryomov __func__, lreq, lreq->linger_id, req, lreq->ping_req); 312575dbb685SIlya Dryomov goto out; 312675dbb685SIlya Dryomov } 312775dbb685SIlya Dryomov 3128922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d ping_sent %lu last_error %d\n", 3129922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, req->r_result, lreq->ping_sent, 3130922dab61SIlya Dryomov lreq->last_error); 3131922dab61SIlya Dryomov if (lreq->register_gen == req->r_ops[0].watch.gen) { 3132b07d3c4bSIlya Dryomov if (!req->r_result) { 3133b07d3c4bSIlya Dryomov lreq->watch_valid_thru = lreq->ping_sent; 3134b07d3c4bSIlya Dryomov } else if (!lreq->last_error) { 3135922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 3136922dab61SIlya Dryomov queue_watch_error(lreq); 3137922dab61SIlya Dryomov } 3138922dab61SIlya Dryomov } else { 3139922dab61SIlya Dryomov dout("lreq %p register_gen %u ignoring old pong %u\n", lreq, 3140922dab61SIlya Dryomov lreq->register_gen, req->r_ops[0].watch.gen); 3141922dab61SIlya Dryomov } 3142922dab61SIlya Dryomov 314375dbb685SIlya Dryomov out: 3144922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3145922dab61SIlya Dryomov linger_put(lreq); 3146922dab61SIlya Dryomov } 3147922dab61SIlya Dryomov 3148922dab61SIlya Dryomov static void send_linger_ping(struct ceph_osd_linger_request *lreq) 3149922dab61SIlya Dryomov { 3150922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 315175dbb685SIlya Dryomov struct ceph_osd_request *req; 315275dbb685SIlya Dryomov int ret; 3153922dab61SIlya Dryomov 3154b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 3155922dab61SIlya Dryomov dout("%s PAUSERD\n", __func__); 3156922dab61SIlya Dryomov return; 3157922dab61SIlya Dryomov } 3158922dab61SIlya Dryomov 3159922dab61SIlya Dryomov lreq->ping_sent = jiffies; 3160922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu ping_sent %lu register_gen %u\n", 3161922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, lreq->ping_sent, 3162922dab61SIlya Dryomov lreq->register_gen); 3163922dab61SIlya Dryomov 316475dbb685SIlya Dryomov if (lreq->ping_req) { 316575dbb685SIlya Dryomov if (lreq->ping_req->r_osd) 316675dbb685SIlya Dryomov cancel_linger_request(lreq->ping_req); 316775dbb685SIlya Dryomov ceph_osdc_put_request(lreq->ping_req); 316875dbb685SIlya Dryomov } 3169922dab61SIlya Dryomov 317075dbb685SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, true, GFP_NOIO); 317175dbb685SIlya Dryomov BUG_ON(!req); 317275dbb685SIlya Dryomov 3173922dab61SIlya Dryomov target_copy(&req->r_t, &lreq->t); 317475dbb685SIlya Dryomov osd_req_op_watch_init(req, 0, CEPH_OSD_WATCH_OP_PING, lreq->linger_id, 317575dbb685SIlya Dryomov lreq->register_gen); 3176922dab61SIlya Dryomov req->r_callback = linger_ping_cb; 317775dbb685SIlya Dryomov 317875dbb685SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 317975dbb685SIlya Dryomov BUG_ON(ret); 318075dbb685SIlya Dryomov 3181922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 3182922dab61SIlya Dryomov req->r_linger = true; 318375dbb685SIlya Dryomov lreq->ping_req = req; 3184922dab61SIlya Dryomov 3185922dab61SIlya Dryomov ceph_osdc_get_request(req); 3186922dab61SIlya Dryomov account_request(req); 3187922dab61SIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 3188922dab61SIlya Dryomov link_request(lreq->osd, req); 3189922dab61SIlya Dryomov send_request(req); 3190922dab61SIlya Dryomov } 3191922dab61SIlya Dryomov 3192922dab61SIlya Dryomov static void linger_submit(struct ceph_osd_linger_request *lreq) 3193922dab61SIlya Dryomov { 3194922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3195922dab61SIlya Dryomov struct ceph_osd *osd; 3196922dab61SIlya Dryomov 319781c65213SIlya Dryomov down_write(&osdc->lock); 319881c65213SIlya Dryomov linger_register(lreq); 319981c65213SIlya Dryomov 32008edf84baSIlya Dryomov calc_target(osdc, &lreq->t, false); 3201922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3202922dab61SIlya Dryomov link_linger(osd, lreq); 3203922dab61SIlya Dryomov 3204922dab61SIlya Dryomov send_linger(lreq); 320581c65213SIlya Dryomov up_write(&osdc->lock); 3206922dab61SIlya Dryomov } 3207922dab61SIlya Dryomov 32084609245eSIlya Dryomov static void cancel_linger_map_check(struct ceph_osd_linger_request *lreq) 32094609245eSIlya Dryomov { 32104609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 32114609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 32124609245eSIlya Dryomov 32134609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 32144609245eSIlya Dryomov 32154609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 32164609245eSIlya Dryomov lreq->linger_id); 32174609245eSIlya Dryomov if (!lookup_lreq) 32184609245eSIlya Dryomov return; 32194609245eSIlya Dryomov 32204609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 32214609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 32224609245eSIlya Dryomov linger_put(lreq); 32234609245eSIlya Dryomov } 32244609245eSIlya Dryomov 3225922dab61SIlya Dryomov /* 3226922dab61SIlya Dryomov * @lreq has to be both registered and linked. 3227922dab61SIlya Dryomov */ 3228922dab61SIlya Dryomov static void __linger_cancel(struct ceph_osd_linger_request *lreq) 3229922dab61SIlya Dryomov { 323075dbb685SIlya Dryomov if (lreq->ping_req && lreq->ping_req->r_osd) 3231922dab61SIlya Dryomov cancel_linger_request(lreq->ping_req); 323275dbb685SIlya Dryomov if (lreq->reg_req && lreq->reg_req->r_osd) 3233922dab61SIlya Dryomov cancel_linger_request(lreq->reg_req); 32344609245eSIlya Dryomov cancel_linger_map_check(lreq); 3235922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3236922dab61SIlya Dryomov linger_unregister(lreq); 3237922dab61SIlya Dryomov } 3238922dab61SIlya Dryomov 3239922dab61SIlya Dryomov static void linger_cancel(struct ceph_osd_linger_request *lreq) 3240922dab61SIlya Dryomov { 3241922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3242922dab61SIlya Dryomov 3243922dab61SIlya Dryomov down_write(&osdc->lock); 3244922dab61SIlya Dryomov if (__linger_registered(lreq)) 3245922dab61SIlya Dryomov __linger_cancel(lreq); 3246922dab61SIlya Dryomov up_write(&osdc->lock); 3247922dab61SIlya Dryomov } 3248922dab61SIlya Dryomov 32494609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq); 32504609245eSIlya Dryomov 32514609245eSIlya Dryomov static void check_linger_pool_dne(struct ceph_osd_linger_request *lreq) 32524609245eSIlya Dryomov { 32534609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 32544609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 32554609245eSIlya Dryomov 32564609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 32574609245eSIlya Dryomov WARN_ON(!map->epoch); 32584609245eSIlya Dryomov 32594609245eSIlya Dryomov if (lreq->register_gen) { 32604609245eSIlya Dryomov lreq->map_dne_bound = map->epoch; 32614609245eSIlya Dryomov dout("%s lreq %p linger_id %llu pool disappeared\n", __func__, 32624609245eSIlya Dryomov lreq, lreq->linger_id); 32634609245eSIlya Dryomov } else { 32644609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u have %u\n", 32654609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 32664609245eSIlya Dryomov map->epoch); 32674609245eSIlya Dryomov } 32684609245eSIlya Dryomov 32694609245eSIlya Dryomov if (lreq->map_dne_bound) { 32704609245eSIlya Dryomov if (map->epoch >= lreq->map_dne_bound) { 32714609245eSIlya Dryomov /* we had a new enough map */ 32724609245eSIlya Dryomov pr_info("linger_id %llu pool does not exist\n", 32734609245eSIlya Dryomov lreq->linger_id); 32744609245eSIlya Dryomov linger_reg_commit_complete(lreq, -ENOENT); 32754609245eSIlya Dryomov __linger_cancel(lreq); 32764609245eSIlya Dryomov } 32774609245eSIlya Dryomov } else { 32784609245eSIlya Dryomov send_linger_map_check(lreq); 32794609245eSIlya Dryomov } 32804609245eSIlya Dryomov } 32814609245eSIlya Dryomov 32824609245eSIlya Dryomov static void linger_map_check_cb(struct ceph_mon_generic_request *greq) 32834609245eSIlya Dryomov { 32844609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 32854609245eSIlya Dryomov struct ceph_osd_linger_request *lreq; 32864609245eSIlya Dryomov u64 linger_id = greq->private_data; 32874609245eSIlya Dryomov 32884609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 32894609245eSIlya Dryomov 32904609245eSIlya Dryomov down_write(&osdc->lock); 32914609245eSIlya Dryomov lreq = lookup_linger_mc(&osdc->linger_map_checks, linger_id); 32924609245eSIlya Dryomov if (!lreq) { 32934609245eSIlya Dryomov dout("%s linger_id %llu dne\n", __func__, linger_id); 32944609245eSIlya Dryomov goto out_unlock; 32954609245eSIlya Dryomov } 32964609245eSIlya Dryomov 32974609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u newest %llu\n", 32984609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 32994609245eSIlya Dryomov greq->u.newest); 33004609245eSIlya Dryomov if (!lreq->map_dne_bound) 33014609245eSIlya Dryomov lreq->map_dne_bound = greq->u.newest; 33024609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 33034609245eSIlya Dryomov check_linger_pool_dne(lreq); 33044609245eSIlya Dryomov 33054609245eSIlya Dryomov linger_put(lreq); 33064609245eSIlya Dryomov out_unlock: 33074609245eSIlya Dryomov up_write(&osdc->lock); 33084609245eSIlya Dryomov } 33094609245eSIlya Dryomov 33104609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq) 33114609245eSIlya Dryomov { 33124609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 33134609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 33144609245eSIlya Dryomov int ret; 33154609245eSIlya Dryomov 33164609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 33174609245eSIlya Dryomov 33184609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 33194609245eSIlya Dryomov lreq->linger_id); 33204609245eSIlya Dryomov if (lookup_lreq) { 33214609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 33224609245eSIlya Dryomov return; 33234609245eSIlya Dryomov } 33244609245eSIlya Dryomov 33254609245eSIlya Dryomov linger_get(lreq); 33264609245eSIlya Dryomov insert_linger_mc(&osdc->linger_map_checks, lreq); 33274609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 33284609245eSIlya Dryomov linger_map_check_cb, lreq->linger_id); 33294609245eSIlya Dryomov WARN_ON(ret); 33304609245eSIlya Dryomov } 33314609245eSIlya Dryomov 3332922dab61SIlya Dryomov static int linger_reg_commit_wait(struct ceph_osd_linger_request *lreq) 3333922dab61SIlya Dryomov { 3334922dab61SIlya Dryomov int ret; 3335922dab61SIlya Dryomov 3336922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 3337922dab61SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->reg_commit_wait); 3338922dab61SIlya Dryomov return ret ?: lreq->reg_commit_error; 3339922dab61SIlya Dryomov } 3340922dab61SIlya Dryomov 334119079203SIlya Dryomov static int linger_notify_finish_wait(struct ceph_osd_linger_request *lreq) 334219079203SIlya Dryomov { 334319079203SIlya Dryomov int ret; 334419079203SIlya Dryomov 334519079203SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 334619079203SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->notify_finish_wait); 334719079203SIlya Dryomov return ret ?: lreq->notify_finish_error; 334819079203SIlya Dryomov } 334919079203SIlya Dryomov 3350922dab61SIlya Dryomov /* 3351fbca9635SIlya Dryomov * Timeout callback, called every N seconds. When 1 or more OSD 3352fbca9635SIlya Dryomov * requests has been active for more than N seconds, we send a keepalive 3353fbca9635SIlya Dryomov * (tag + timestamp) to its OSD to ensure any communications channel 3354fbca9635SIlya Dryomov * reset is detected. 33553d14c5d2SYehuda Sadeh */ 33563d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 33573d14c5d2SYehuda Sadeh { 33583d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 33593d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 3360a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 33615aea3dcdSIlya Dryomov unsigned long cutoff = jiffies - opts->osd_keepalive_timeout; 33627cc5e38fSIlya Dryomov unsigned long expiry_cutoff = jiffies - opts->osd_request_timeout; 33635aea3dcdSIlya Dryomov LIST_HEAD(slow_osds); 33645aea3dcdSIlya Dryomov struct rb_node *n, *p; 33653d14c5d2SYehuda Sadeh 33665aea3dcdSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 33675aea3dcdSIlya Dryomov down_write(&osdc->lock); 33683d14c5d2SYehuda Sadeh 33693d14c5d2SYehuda Sadeh /* 33703d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 33713d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 33723d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 33733d14c5d2SYehuda Sadeh */ 33745aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 33755aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 33765aea3dcdSIlya Dryomov bool found = false; 33773d14c5d2SYehuda Sadeh 33787cc5e38fSIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 33795aea3dcdSIlya Dryomov struct ceph_osd_request *req = 33805aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 33815aea3dcdSIlya Dryomov 33827cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 33837cc5e38fSIlya Dryomov 33845aea3dcdSIlya Dryomov if (time_before(req->r_stamp, cutoff)) { 33855aea3dcdSIlya Dryomov dout(" req %p tid %llu on osd%d is laggy\n", 33865aea3dcdSIlya Dryomov req, req->r_tid, osd->o_osd); 33875aea3dcdSIlya Dryomov found = true; 33885aea3dcdSIlya Dryomov } 33897cc5e38fSIlya Dryomov if (opts->osd_request_timeout && 33907cc5e38fSIlya Dryomov time_before(req->r_start_stamp, expiry_cutoff)) { 33917cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 33927cc5e38fSIlya Dryomov req->r_tid, osd->o_osd); 33937cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 33947cc5e38fSIlya Dryomov } 33955aea3dcdSIlya Dryomov } 3396922dab61SIlya Dryomov for (p = rb_first(&osd->o_linger_requests); p; p = rb_next(p)) { 3397922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3398922dab61SIlya Dryomov rb_entry(p, struct ceph_osd_linger_request, node); 3399922dab61SIlya Dryomov 3400922dab61SIlya Dryomov dout(" lreq %p linger_id %llu is served by osd%d\n", 3401922dab61SIlya Dryomov lreq, lreq->linger_id, osd->o_osd); 3402922dab61SIlya Dryomov found = true; 3403922dab61SIlya Dryomov 3404922dab61SIlya Dryomov mutex_lock(&lreq->lock); 340519079203SIlya Dryomov if (lreq->is_watch && lreq->committed && !lreq->last_error) 3406922dab61SIlya Dryomov send_linger_ping(lreq); 3407922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3408922dab61SIlya Dryomov } 34095aea3dcdSIlya Dryomov 34105aea3dcdSIlya Dryomov if (found) 34113d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 34123d14c5d2SYehuda Sadeh } 34135aea3dcdSIlya Dryomov 34147cc5e38fSIlya Dryomov if (opts->osd_request_timeout) { 34157cc5e38fSIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 34167cc5e38fSIlya Dryomov struct ceph_osd_request *req = 34177cc5e38fSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 34187cc5e38fSIlya Dryomov 34197cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 34207cc5e38fSIlya Dryomov 34217cc5e38fSIlya Dryomov if (time_before(req->r_start_stamp, expiry_cutoff)) { 34227cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 34237cc5e38fSIlya Dryomov req->r_tid, osdc->homeless_osd.o_osd); 34247cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 34257cc5e38fSIlya Dryomov } 34267cc5e38fSIlya Dryomov } 34277cc5e38fSIlya Dryomov } 34287cc5e38fSIlya Dryomov 34295aea3dcdSIlya Dryomov if (atomic_read(&osdc->num_homeless) || !list_empty(&slow_osds)) 34305aea3dcdSIlya Dryomov maybe_request_map(osdc); 34315aea3dcdSIlya Dryomov 34323d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 34335aea3dcdSIlya Dryomov struct ceph_osd *osd = list_first_entry(&slow_osds, 34345aea3dcdSIlya Dryomov struct ceph_osd, 34353d14c5d2SYehuda Sadeh o_keepalive_item); 34363d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 34373d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 34383d14c5d2SYehuda Sadeh } 34393d14c5d2SYehuda Sadeh 34405aea3dcdSIlya Dryomov up_write(&osdc->lock); 3441fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 3442fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 34433d14c5d2SYehuda Sadeh } 34443d14c5d2SYehuda Sadeh 34453d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 34463d14c5d2SYehuda Sadeh { 34473d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 34483d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 34493d14c5d2SYehuda Sadeh osds_timeout_work.work); 3450a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 345142a2c09fSIlya Dryomov struct ceph_osd *osd, *nosd; 34523d14c5d2SYehuda Sadeh 345342a2c09fSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 34545aea3dcdSIlya Dryomov down_write(&osdc->lock); 345542a2c09fSIlya Dryomov list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 345642a2c09fSIlya Dryomov if (time_before(jiffies, osd->lru_ttl)) 345742a2c09fSIlya Dryomov break; 345842a2c09fSIlya Dryomov 34595aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 3460922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 34615aea3dcdSIlya Dryomov close_osd(osd); 346242a2c09fSIlya Dryomov } 346342a2c09fSIlya Dryomov 34645aea3dcdSIlya Dryomov up_write(&osdc->lock); 34653d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 34663d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 34673d14c5d2SYehuda Sadeh } 34683d14c5d2SYehuda Sadeh 3469205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 3470205ee118SIlya Dryomov struct ceph_object_locator *oloc) 3471205ee118SIlya Dryomov { 3472205ee118SIlya Dryomov u8 struct_v, struct_cv; 3473205ee118SIlya Dryomov u32 len; 3474205ee118SIlya Dryomov void *struct_end; 3475205ee118SIlya Dryomov int ret = 0; 3476205ee118SIlya Dryomov 3477205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3478205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3479205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3480205ee118SIlya Dryomov if (struct_v < 3) { 3481205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 3482205ee118SIlya Dryomov struct_v, struct_cv); 3483205ee118SIlya Dryomov goto e_inval; 3484205ee118SIlya Dryomov } 3485205ee118SIlya Dryomov if (struct_cv > 6) { 3486205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 3487205ee118SIlya Dryomov struct_v, struct_cv); 3488205ee118SIlya Dryomov goto e_inval; 3489205ee118SIlya Dryomov } 3490205ee118SIlya Dryomov len = ceph_decode_32(p); 3491205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3492205ee118SIlya Dryomov struct_end = *p + len; 3493205ee118SIlya Dryomov 3494205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 3495205ee118SIlya Dryomov *p += 4; /* skip preferred */ 3496205ee118SIlya Dryomov 3497205ee118SIlya Dryomov len = ceph_decode_32(p); 3498205ee118SIlya Dryomov if (len > 0) { 3499205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 3500205ee118SIlya Dryomov goto e_inval; 3501205ee118SIlya Dryomov } 3502205ee118SIlya Dryomov 3503205ee118SIlya Dryomov if (struct_v >= 5) { 3504cd08e0a2SYan, Zheng bool changed = false; 3505cd08e0a2SYan, Zheng 3506205ee118SIlya Dryomov len = ceph_decode_32(p); 3507205ee118SIlya Dryomov if (len > 0) { 350830c156d9SYan, Zheng ceph_decode_need(p, end, len, e_inval); 3509cd08e0a2SYan, Zheng if (!oloc->pool_ns || 3510cd08e0a2SYan, Zheng ceph_compare_string(oloc->pool_ns, *p, len)) 3511cd08e0a2SYan, Zheng changed = true; 351230c156d9SYan, Zheng *p += len; 3513cd08e0a2SYan, Zheng } else { 3514cd08e0a2SYan, Zheng if (oloc->pool_ns) 3515cd08e0a2SYan, Zheng changed = true; 3516cd08e0a2SYan, Zheng } 3517cd08e0a2SYan, Zheng if (changed) { 3518cd08e0a2SYan, Zheng /* redirect changes namespace */ 3519cd08e0a2SYan, Zheng pr_warn("ceph_object_locator::nspace is changed\n"); 3520cd08e0a2SYan, Zheng goto e_inval; 3521205ee118SIlya Dryomov } 3522205ee118SIlya Dryomov } 3523205ee118SIlya Dryomov 3524205ee118SIlya Dryomov if (struct_v >= 6) { 3525205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 3526205ee118SIlya Dryomov if (hash != -1) { 3527205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 3528205ee118SIlya Dryomov goto e_inval; 3529205ee118SIlya Dryomov } 3530205ee118SIlya Dryomov } 3531205ee118SIlya Dryomov 3532205ee118SIlya Dryomov /* skip the rest */ 3533205ee118SIlya Dryomov *p = struct_end; 3534205ee118SIlya Dryomov out: 3535205ee118SIlya Dryomov return ret; 3536205ee118SIlya Dryomov 3537205ee118SIlya Dryomov e_inval: 3538205ee118SIlya Dryomov ret = -EINVAL; 3539205ee118SIlya Dryomov goto out; 3540205ee118SIlya Dryomov } 3541205ee118SIlya Dryomov 3542205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 3543205ee118SIlya Dryomov struct ceph_request_redirect *redir) 3544205ee118SIlya Dryomov { 3545205ee118SIlya Dryomov u8 struct_v, struct_cv; 3546205ee118SIlya Dryomov u32 len; 3547205ee118SIlya Dryomov void *struct_end; 3548205ee118SIlya Dryomov int ret; 3549205ee118SIlya Dryomov 3550205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3551205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3552205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3553205ee118SIlya Dryomov if (struct_cv > 1) { 3554205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 3555205ee118SIlya Dryomov struct_v, struct_cv); 3556205ee118SIlya Dryomov goto e_inval; 3557205ee118SIlya Dryomov } 3558205ee118SIlya Dryomov len = ceph_decode_32(p); 3559205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3560205ee118SIlya Dryomov struct_end = *p + len; 3561205ee118SIlya Dryomov 3562205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 3563205ee118SIlya Dryomov if (ret) 3564205ee118SIlya Dryomov goto out; 3565205ee118SIlya Dryomov 3566205ee118SIlya Dryomov len = ceph_decode_32(p); 3567205ee118SIlya Dryomov if (len > 0) { 3568205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 3569205ee118SIlya Dryomov goto e_inval; 3570205ee118SIlya Dryomov } 3571205ee118SIlya Dryomov 3572205ee118SIlya Dryomov /* skip the rest */ 3573205ee118SIlya Dryomov *p = struct_end; 3574205ee118SIlya Dryomov out: 3575205ee118SIlya Dryomov return ret; 3576205ee118SIlya Dryomov 3577205ee118SIlya Dryomov e_inval: 3578205ee118SIlya Dryomov ret = -EINVAL; 3579205ee118SIlya Dryomov goto out; 3580205ee118SIlya Dryomov } 3581205ee118SIlya Dryomov 3582fe5da05eSIlya Dryomov struct MOSDOpReply { 3583fe5da05eSIlya Dryomov struct ceph_pg pgid; 3584fe5da05eSIlya Dryomov u64 flags; 3585fe5da05eSIlya Dryomov int result; 3586fe5da05eSIlya Dryomov u32 epoch; 3587fe5da05eSIlya Dryomov int num_ops; 3588fe5da05eSIlya Dryomov u32 outdata_len[CEPH_OSD_MAX_OPS]; 3589fe5da05eSIlya Dryomov s32 rval[CEPH_OSD_MAX_OPS]; 3590fe5da05eSIlya Dryomov int retry_attempt; 3591fe5da05eSIlya Dryomov struct ceph_eversion replay_version; 3592fe5da05eSIlya Dryomov u64 user_version; 3593fe5da05eSIlya Dryomov struct ceph_request_redirect redirect; 3594fe5da05eSIlya Dryomov }; 359525845472SSage Weil 3596fe5da05eSIlya Dryomov static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m) 35973d14c5d2SYehuda Sadeh { 3598fe5da05eSIlya Dryomov void *p = msg->front.iov_base; 3599fe5da05eSIlya Dryomov void *const end = p + msg->front.iov_len; 3600fe5da05eSIlya Dryomov u16 version = le16_to_cpu(msg->hdr.version); 3601fe5da05eSIlya Dryomov struct ceph_eversion bad_replay_version; 3602b0b31a8fSIlya Dryomov u8 decode_redir; 3603fe5da05eSIlya Dryomov u32 len; 3604fe5da05eSIlya Dryomov int ret; 3605fe5da05eSIlya Dryomov int i; 36063d14c5d2SYehuda Sadeh 3607fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, len, e_inval); 3608fe5da05eSIlya Dryomov ceph_decode_need(&p, end, len, e_inval); 3609fe5da05eSIlya Dryomov p += len; /* skip oid */ 36101b83bef2SSage Weil 3611fe5da05eSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->pgid); 3612fe5da05eSIlya Dryomov if (ret) 3613fe5da05eSIlya Dryomov return ret; 36141b83bef2SSage Weil 3615fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->flags, e_inval); 3616fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->result, e_inval); 3617fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval); 3618fe5da05eSIlya Dryomov memcpy(&bad_replay_version, p, sizeof(bad_replay_version)); 3619fe5da05eSIlya Dryomov p += sizeof(bad_replay_version); 3620fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->epoch, e_inval); 36211b83bef2SSage Weil 3622fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->num_ops, e_inval); 3623fe5da05eSIlya Dryomov if (m->num_ops > ARRAY_SIZE(m->outdata_len)) 3624fe5da05eSIlya Dryomov goto e_inval; 36251b83bef2SSage Weil 3626fe5da05eSIlya Dryomov ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op), 3627fe5da05eSIlya Dryomov e_inval); 3628fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) { 36291b83bef2SSage Weil struct ceph_osd_op *op = p; 36301b83bef2SSage Weil 3631fe5da05eSIlya Dryomov m->outdata_len[i] = le32_to_cpu(op->payload_len); 36321b83bef2SSage Weil p += sizeof(*op); 36331b83bef2SSage Weil } 3634fe5da05eSIlya Dryomov 3635fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval); 3636fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) 3637fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->rval[i], e_inval); 3638fe5da05eSIlya Dryomov 3639fe5da05eSIlya Dryomov if (version >= 5) { 3640fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval); 3641fe5da05eSIlya Dryomov memcpy(&m->replay_version, p, sizeof(m->replay_version)); 3642fe5da05eSIlya Dryomov p += sizeof(m->replay_version); 3643fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->user_version, e_inval); 3644fe5da05eSIlya Dryomov } else { 3645fe5da05eSIlya Dryomov m->replay_version = bad_replay_version; /* struct */ 3646fe5da05eSIlya Dryomov m->user_version = le64_to_cpu(m->replay_version.version); 36471b83bef2SSage Weil } 36481b83bef2SSage Weil 3649fe5da05eSIlya Dryomov if (version >= 6) { 3650fe5da05eSIlya Dryomov if (version >= 7) 3651fe5da05eSIlya Dryomov ceph_decode_8_safe(&p, end, decode_redir, e_inval); 3652b0b31a8fSIlya Dryomov else 3653b0b31a8fSIlya Dryomov decode_redir = 1; 3654b0b31a8fSIlya Dryomov } else { 3655b0b31a8fSIlya Dryomov decode_redir = 0; 3656b0b31a8fSIlya Dryomov } 3657b0b31a8fSIlya Dryomov 3658b0b31a8fSIlya Dryomov if (decode_redir) { 3659fe5da05eSIlya Dryomov ret = ceph_redirect_decode(&p, end, &m->redirect); 3660fe5da05eSIlya Dryomov if (ret) 3661fe5da05eSIlya Dryomov return ret; 3662205ee118SIlya Dryomov } else { 3663fe5da05eSIlya Dryomov ceph_oloc_init(&m->redirect.oloc); 3664205ee118SIlya Dryomov } 3665205ee118SIlya Dryomov 3666fe5da05eSIlya Dryomov return 0; 3667205ee118SIlya Dryomov 3668fe5da05eSIlya Dryomov e_inval: 3669fe5da05eSIlya Dryomov return -EINVAL; 3670fe5da05eSIlya Dryomov } 3671fe5da05eSIlya Dryomov 3672fe5da05eSIlya Dryomov /* 3673b18b9550SIlya Dryomov * Handle MOSDOpReply. Set ->r_result and call the callback if it is 3674b18b9550SIlya Dryomov * specified. 3675fe5da05eSIlya Dryomov */ 36765aea3dcdSIlya Dryomov static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg) 3677fe5da05eSIlya Dryomov { 36785aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 3679fe5da05eSIlya Dryomov struct ceph_osd_request *req; 3680fe5da05eSIlya Dryomov struct MOSDOpReply m; 3681fe5da05eSIlya Dryomov u64 tid = le64_to_cpu(msg->hdr.tid); 3682fe5da05eSIlya Dryomov u32 data_len = 0; 3683fe5da05eSIlya Dryomov int ret; 3684fe5da05eSIlya Dryomov int i; 3685fe5da05eSIlya Dryomov 3686fe5da05eSIlya Dryomov dout("%s msg %p tid %llu\n", __func__, msg, tid); 3687fe5da05eSIlya Dryomov 36885aea3dcdSIlya Dryomov down_read(&osdc->lock); 36895aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 36905aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 36915aea3dcdSIlya Dryomov goto out_unlock_osdc; 3692fe5da05eSIlya Dryomov } 36935aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 36945aea3dcdSIlya Dryomov 36955aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 36965aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 36975aea3dcdSIlya Dryomov if (!req) { 36985aea3dcdSIlya Dryomov dout("%s osd%d tid %llu unknown\n", __func__, osd->o_osd, tid); 36995aea3dcdSIlya Dryomov goto out_unlock_session; 37005aea3dcdSIlya Dryomov } 3701fe5da05eSIlya Dryomov 3702cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = req->r_t.target_oloc.pool_ns; 3703fe5da05eSIlya Dryomov ret = decode_MOSDOpReply(msg, &m); 3704cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = NULL; 3705fe5da05eSIlya Dryomov if (ret) { 3706fe5da05eSIlya Dryomov pr_err("failed to decode MOSDOpReply for tid %llu: %d\n", 3707fe5da05eSIlya Dryomov req->r_tid, ret); 3708fe5da05eSIlya Dryomov ceph_msg_dump(msg); 3709fe5da05eSIlya Dryomov goto fail_request; 3710fe5da05eSIlya Dryomov } 3711fe5da05eSIlya Dryomov dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n", 3712fe5da05eSIlya Dryomov __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed, 3713fe5da05eSIlya Dryomov m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch), 3714fe5da05eSIlya Dryomov le64_to_cpu(m.replay_version.version), m.user_version); 3715fe5da05eSIlya Dryomov 3716fe5da05eSIlya Dryomov if (m.retry_attempt >= 0) { 3717fe5da05eSIlya Dryomov if (m.retry_attempt != req->r_attempts - 1) { 3718fe5da05eSIlya Dryomov dout("req %p tid %llu retry_attempt %d != %d, ignoring\n", 3719fe5da05eSIlya Dryomov req, req->r_tid, m.retry_attempt, 3720fe5da05eSIlya Dryomov req->r_attempts - 1); 37215aea3dcdSIlya Dryomov goto out_unlock_session; 3722fe5da05eSIlya Dryomov } 3723fe5da05eSIlya Dryomov } else { 3724fe5da05eSIlya Dryomov WARN_ON(1); /* MOSDOpReply v4 is assumed */ 3725fe5da05eSIlya Dryomov } 3726fe5da05eSIlya Dryomov 3727fe5da05eSIlya Dryomov if (!ceph_oloc_empty(&m.redirect.oloc)) { 3728fe5da05eSIlya Dryomov dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid, 3729fe5da05eSIlya Dryomov m.redirect.oloc.pool); 37305aea3dcdSIlya Dryomov unlink_request(osd, req); 37315aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 3732205ee118SIlya Dryomov 373330c156d9SYan, Zheng /* 373430c156d9SYan, Zheng * Not ceph_oloc_copy() - changing pool_ns is not 373530c156d9SYan, Zheng * supported. 373630c156d9SYan, Zheng */ 373730c156d9SYan, Zheng req->r_t.target_oloc.pool = m.redirect.oloc.pool; 3738890bd0f8SJerry Lee req->r_flags |= CEPH_OSD_FLAG_REDIRECTED | 3739890bd0f8SJerry Lee CEPH_OSD_FLAG_IGNORE_OVERLAY | 3740890bd0f8SJerry Lee CEPH_OSD_FLAG_IGNORE_CACHE; 37415aea3dcdSIlya Dryomov req->r_tid = 0; 37425aea3dcdSIlya Dryomov __submit_request(req, false); 37435aea3dcdSIlya Dryomov goto out_unlock_osdc; 3744205ee118SIlya Dryomov } 3745205ee118SIlya Dryomov 3746117d96a0SIlya Dryomov if (m.result == -EAGAIN) { 3747117d96a0SIlya Dryomov dout("req %p tid %llu EAGAIN\n", req, req->r_tid); 3748117d96a0SIlya Dryomov unlink_request(osd, req); 3749117d96a0SIlya Dryomov mutex_unlock(&osd->lock); 3750117d96a0SIlya Dryomov 3751117d96a0SIlya Dryomov /* 3752117d96a0SIlya Dryomov * The object is missing on the replica or not (yet) 3753117d96a0SIlya Dryomov * readable. Clear pgid to force a resend to the primary 3754117d96a0SIlya Dryomov * via legacy_change. 3755117d96a0SIlya Dryomov */ 3756117d96a0SIlya Dryomov req->r_t.pgid.pool = 0; 3757117d96a0SIlya Dryomov req->r_t.pgid.seed = 0; 3758117d96a0SIlya Dryomov WARN_ON(!req->r_t.used_replica); 3759117d96a0SIlya Dryomov req->r_flags &= ~(CEPH_OSD_FLAG_BALANCE_READS | 3760117d96a0SIlya Dryomov CEPH_OSD_FLAG_LOCALIZE_READS); 3761117d96a0SIlya Dryomov req->r_tid = 0; 3762117d96a0SIlya Dryomov __submit_request(req, false); 3763117d96a0SIlya Dryomov goto out_unlock_osdc; 3764117d96a0SIlya Dryomov } 3765117d96a0SIlya Dryomov 3766fe5da05eSIlya Dryomov if (m.num_ops != req->r_num_ops) { 3767fe5da05eSIlya Dryomov pr_err("num_ops %d != %d for tid %llu\n", m.num_ops, 3768fe5da05eSIlya Dryomov req->r_num_ops, req->r_tid); 3769fe5da05eSIlya Dryomov goto fail_request; 3770fe5da05eSIlya Dryomov } 3771fe5da05eSIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 3772fe5da05eSIlya Dryomov dout(" req %p tid %llu op %d rval %d len %u\n", req, 3773fe5da05eSIlya Dryomov req->r_tid, i, m.rval[i], m.outdata_len[i]); 3774fe5da05eSIlya Dryomov req->r_ops[i].rval = m.rval[i]; 3775fe5da05eSIlya Dryomov req->r_ops[i].outdata_len = m.outdata_len[i]; 3776fe5da05eSIlya Dryomov data_len += m.outdata_len[i]; 3777fe5da05eSIlya Dryomov } 3778fe5da05eSIlya Dryomov if (data_len != le32_to_cpu(msg->hdr.data_len)) { 3779fe5da05eSIlya Dryomov pr_err("sum of lens %u != %u for tid %llu\n", data_len, 3780fe5da05eSIlya Dryomov le32_to_cpu(msg->hdr.data_len), req->r_tid); 3781fe5da05eSIlya Dryomov goto fail_request; 3782fe5da05eSIlya Dryomov } 3783b18b9550SIlya Dryomov dout("%s req %p tid %llu result %d data_len %u\n", __func__, 3784b18b9550SIlya Dryomov req, req->r_tid, m.result, data_len); 37853d14c5d2SYehuda Sadeh 3786b18b9550SIlya Dryomov /* 3787b18b9550SIlya Dryomov * Since we only ever request ONDISK, we should only ever get 3788b18b9550SIlya Dryomov * one (type of) reply back. 3789b18b9550SIlya Dryomov */ 3790b18b9550SIlya Dryomov WARN_ON(!(m.flags & CEPH_OSD_FLAG_ONDISK)); 3791fe5da05eSIlya Dryomov req->r_result = m.result ?: data_len; 379245ee2c1dSIlya Dryomov finish_request(req); 37935aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 37945aea3dcdSIlya Dryomov up_read(&osdc->lock); 37953d14c5d2SYehuda Sadeh 3796fe5da05eSIlya Dryomov __complete_request(req); 37973d14c5d2SYehuda Sadeh return; 3798fe5da05eSIlya Dryomov 3799fe5da05eSIlya Dryomov fail_request: 38004609245eSIlya Dryomov complete_request(req, -EIO); 38015aea3dcdSIlya Dryomov out_unlock_session: 38025aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 38035aea3dcdSIlya Dryomov out_unlock_osdc: 38045aea3dcdSIlya Dryomov up_read(&osdc->lock); 38053d14c5d2SYehuda Sadeh } 38063d14c5d2SYehuda Sadeh 380742c1b124SIlya Dryomov static void set_pool_was_full(struct ceph_osd_client *osdc) 380842c1b124SIlya Dryomov { 380942c1b124SIlya Dryomov struct rb_node *n; 381042c1b124SIlya Dryomov 381142c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 381242c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 381342c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 381442c1b124SIlya Dryomov 381542c1b124SIlya Dryomov pi->was_full = __pool_full(pi); 381642c1b124SIlya Dryomov } 381742c1b124SIlya Dryomov } 381842c1b124SIlya Dryomov 38195aea3dcdSIlya Dryomov static bool pool_cleared_full(struct ceph_osd_client *osdc, s64 pool_id) 38203d14c5d2SYehuda Sadeh { 38215aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 38223d14c5d2SYehuda Sadeh 38235aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 38245aea3dcdSIlya Dryomov if (!pi) 38255aea3dcdSIlya Dryomov return false; 38263d14c5d2SYehuda Sadeh 38275aea3dcdSIlya Dryomov return pi->was_full && !__pool_full(pi); 38283d14c5d2SYehuda Sadeh } 38293d14c5d2SYehuda Sadeh 3830922dab61SIlya Dryomov static enum calc_target_result 3831922dab61SIlya Dryomov recalc_linger_target(struct ceph_osd_linger_request *lreq) 3832922dab61SIlya Dryomov { 3833922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3834922dab61SIlya Dryomov enum calc_target_result ct_res; 3835922dab61SIlya Dryomov 38368edf84baSIlya Dryomov ct_res = calc_target(osdc, &lreq->t, true); 3837922dab61SIlya Dryomov if (ct_res == CALC_TARGET_NEED_RESEND) { 3838922dab61SIlya Dryomov struct ceph_osd *osd; 3839922dab61SIlya Dryomov 3840922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3841922dab61SIlya Dryomov if (osd != lreq->osd) { 3842922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3843922dab61SIlya Dryomov link_linger(osd, lreq); 3844922dab61SIlya Dryomov } 3845922dab61SIlya Dryomov } 3846922dab61SIlya Dryomov 3847922dab61SIlya Dryomov return ct_res; 3848922dab61SIlya Dryomov } 3849922dab61SIlya Dryomov 38503d14c5d2SYehuda Sadeh /* 38515aea3dcdSIlya Dryomov * Requeue requests whose mapping to an OSD has changed. 38523d14c5d2SYehuda Sadeh */ 38535aea3dcdSIlya Dryomov static void scan_requests(struct ceph_osd *osd, 38545aea3dcdSIlya Dryomov bool force_resend, 38555aea3dcdSIlya Dryomov bool cleared_full, 38565aea3dcdSIlya Dryomov bool check_pool_cleared_full, 38575aea3dcdSIlya Dryomov struct rb_root *need_resend, 38585aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 38593d14c5d2SYehuda Sadeh { 38605aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 38615aea3dcdSIlya Dryomov struct rb_node *n; 38625aea3dcdSIlya Dryomov bool force_resend_writes; 38633d14c5d2SYehuda Sadeh 3864922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 3865922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3866922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 3867922dab61SIlya Dryomov enum calc_target_result ct_res; 3868922dab61SIlya Dryomov 3869922dab61SIlya Dryomov n = rb_next(n); /* recalc_linger_target() */ 3870922dab61SIlya Dryomov 3871922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, 3872922dab61SIlya Dryomov lreq->linger_id); 3873922dab61SIlya Dryomov ct_res = recalc_linger_target(lreq); 3874922dab61SIlya Dryomov switch (ct_res) { 3875922dab61SIlya Dryomov case CALC_TARGET_NO_ACTION: 3876922dab61SIlya Dryomov force_resend_writes = cleared_full || 3877922dab61SIlya Dryomov (check_pool_cleared_full && 3878922dab61SIlya Dryomov pool_cleared_full(osdc, lreq->t.base_oloc.pool)); 3879922dab61SIlya Dryomov if (!force_resend && !force_resend_writes) 3880922dab61SIlya Dryomov break; 3881922dab61SIlya Dryomov 3882df561f66SGustavo A. R. Silva fallthrough; 3883922dab61SIlya Dryomov case CALC_TARGET_NEED_RESEND: 38844609245eSIlya Dryomov cancel_linger_map_check(lreq); 3885922dab61SIlya Dryomov /* 3886922dab61SIlya Dryomov * scan_requests() for the previous epoch(s) 3887922dab61SIlya Dryomov * may have already added it to the list, since 3888922dab61SIlya Dryomov * it's not unlinked here. 3889922dab61SIlya Dryomov */ 3890922dab61SIlya Dryomov if (list_empty(&lreq->scan_item)) 3891922dab61SIlya Dryomov list_add_tail(&lreq->scan_item, need_resend_linger); 3892922dab61SIlya Dryomov break; 3893922dab61SIlya Dryomov case CALC_TARGET_POOL_DNE: 3894a10bcb19SIlya Dryomov list_del_init(&lreq->scan_item); 38954609245eSIlya Dryomov check_linger_pool_dne(lreq); 3896922dab61SIlya Dryomov break; 3897922dab61SIlya Dryomov } 3898922dab61SIlya Dryomov } 3899922dab61SIlya Dryomov 39005aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 39015aea3dcdSIlya Dryomov struct ceph_osd_request *req = 39025aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 39035aea3dcdSIlya Dryomov enum calc_target_result ct_res; 3904ab60b16dSAlex Elder 39054609245eSIlya Dryomov n = rb_next(n); /* unlink_request(), check_pool_dne() */ 3906ab60b16dSAlex Elder 39075aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 39088edf84baSIlya Dryomov ct_res = calc_target(osdc, &req->r_t, false); 39095aea3dcdSIlya Dryomov switch (ct_res) { 39105aea3dcdSIlya Dryomov case CALC_TARGET_NO_ACTION: 39115aea3dcdSIlya Dryomov force_resend_writes = cleared_full || 39125aea3dcdSIlya Dryomov (check_pool_cleared_full && 39135aea3dcdSIlya Dryomov pool_cleared_full(osdc, req->r_t.base_oloc.pool)); 39145aea3dcdSIlya Dryomov if (!force_resend && 39155aea3dcdSIlya Dryomov (!(req->r_flags & CEPH_OSD_FLAG_WRITE) || 39165aea3dcdSIlya Dryomov !force_resend_writes)) 39175aea3dcdSIlya Dryomov break; 3918a40c4f10SYehuda Sadeh 3919df561f66SGustavo A. R. Silva fallthrough; 39205aea3dcdSIlya Dryomov case CALC_TARGET_NEED_RESEND: 39214609245eSIlya Dryomov cancel_map_check(req); 39225aea3dcdSIlya Dryomov unlink_request(osd, req); 39235aea3dcdSIlya Dryomov insert_request(need_resend, req); 39245aea3dcdSIlya Dryomov break; 39255aea3dcdSIlya Dryomov case CALC_TARGET_POOL_DNE: 39264609245eSIlya Dryomov check_pool_dne(req); 39275aea3dcdSIlya Dryomov break; 3928a40c4f10SYehuda Sadeh } 39293d14c5d2SYehuda Sadeh } 39303d14c5d2SYehuda Sadeh } 39316f6c7006SSage Weil 393242c1b124SIlya Dryomov static int handle_one_map(struct ceph_osd_client *osdc, 39335aea3dcdSIlya Dryomov void *p, void *end, bool incremental, 39345aea3dcdSIlya Dryomov struct rb_root *need_resend, 39355aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 393642c1b124SIlya Dryomov { 393742c1b124SIlya Dryomov struct ceph_osdmap *newmap; 393842c1b124SIlya Dryomov struct rb_node *n; 393942c1b124SIlya Dryomov bool skipped_map = false; 394042c1b124SIlya Dryomov bool was_full; 394142c1b124SIlya Dryomov 3942b7ec35b3SIlya Dryomov was_full = ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 394342c1b124SIlya Dryomov set_pool_was_full(osdc); 394442c1b124SIlya Dryomov 394542c1b124SIlya Dryomov if (incremental) 3946cd1a677cSIlya Dryomov newmap = osdmap_apply_incremental(&p, end, 3947cd1a677cSIlya Dryomov ceph_msgr2(osdc->client), 3948cd1a677cSIlya Dryomov osdc->osdmap); 394942c1b124SIlya Dryomov else 3950cd1a677cSIlya Dryomov newmap = ceph_osdmap_decode(&p, end, ceph_msgr2(osdc->client)); 395142c1b124SIlya Dryomov if (IS_ERR(newmap)) 395242c1b124SIlya Dryomov return PTR_ERR(newmap); 395342c1b124SIlya Dryomov 395442c1b124SIlya Dryomov if (newmap != osdc->osdmap) { 395542c1b124SIlya Dryomov /* 395642c1b124SIlya Dryomov * Preserve ->was_full before destroying the old map. 395742c1b124SIlya Dryomov * For pools that weren't in the old map, ->was_full 395842c1b124SIlya Dryomov * should be false. 395942c1b124SIlya Dryomov */ 396042c1b124SIlya Dryomov for (n = rb_first(&newmap->pg_pools); n; n = rb_next(n)) { 396142c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 396242c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 396342c1b124SIlya Dryomov struct ceph_pg_pool_info *old_pi; 396442c1b124SIlya Dryomov 396542c1b124SIlya Dryomov old_pi = ceph_pg_pool_by_id(osdc->osdmap, pi->id); 396642c1b124SIlya Dryomov if (old_pi) 396742c1b124SIlya Dryomov pi->was_full = old_pi->was_full; 396842c1b124SIlya Dryomov else 396942c1b124SIlya Dryomov WARN_ON(pi->was_full); 397042c1b124SIlya Dryomov } 397142c1b124SIlya Dryomov 397242c1b124SIlya Dryomov if (osdc->osdmap->epoch && 397342c1b124SIlya Dryomov osdc->osdmap->epoch + 1 < newmap->epoch) { 397442c1b124SIlya Dryomov WARN_ON(incremental); 397542c1b124SIlya Dryomov skipped_map = true; 397642c1b124SIlya Dryomov } 397742c1b124SIlya Dryomov 397842c1b124SIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 397942c1b124SIlya Dryomov osdc->osdmap = newmap; 398042c1b124SIlya Dryomov } 398142c1b124SIlya Dryomov 3982b7ec35b3SIlya Dryomov was_full &= !ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 39835aea3dcdSIlya Dryomov scan_requests(&osdc->homeless_osd, skipped_map, was_full, true, 39845aea3dcdSIlya Dryomov need_resend, need_resend_linger); 39855aea3dcdSIlya Dryomov 39865aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; ) { 39875aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 39885aea3dcdSIlya Dryomov 39895aea3dcdSIlya Dryomov n = rb_next(n); /* close_osd() */ 39905aea3dcdSIlya Dryomov 39915aea3dcdSIlya Dryomov scan_requests(osd, skipped_map, was_full, true, need_resend, 39925aea3dcdSIlya Dryomov need_resend_linger); 39935aea3dcdSIlya Dryomov if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 39945aea3dcdSIlya Dryomov memcmp(&osd->o_con.peer_addr, 39955aea3dcdSIlya Dryomov ceph_osd_addr(osdc->osdmap, osd->o_osd), 39965aea3dcdSIlya Dryomov sizeof(struct ceph_entity_addr))) 39975aea3dcdSIlya Dryomov close_osd(osd); 39985aea3dcdSIlya Dryomov } 399942c1b124SIlya Dryomov 400042c1b124SIlya Dryomov return 0; 400142c1b124SIlya Dryomov } 40026f6c7006SSage Weil 40035aea3dcdSIlya Dryomov static void kick_requests(struct ceph_osd_client *osdc, 40045aea3dcdSIlya Dryomov struct rb_root *need_resend, 40055aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 40065aea3dcdSIlya Dryomov { 4007922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq, *nlreq; 400804c7d789SIlya Dryomov enum calc_target_result ct_res; 40095aea3dcdSIlya Dryomov struct rb_node *n; 40105aea3dcdSIlya Dryomov 401104c7d789SIlya Dryomov /* make sure need_resend targets reflect latest map */ 401204c7d789SIlya Dryomov for (n = rb_first(need_resend); n; ) { 401304c7d789SIlya Dryomov struct ceph_osd_request *req = 401404c7d789SIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 401504c7d789SIlya Dryomov 401604c7d789SIlya Dryomov n = rb_next(n); 401704c7d789SIlya Dryomov 401804c7d789SIlya Dryomov if (req->r_t.epoch < osdc->osdmap->epoch) { 40198edf84baSIlya Dryomov ct_res = calc_target(osdc, &req->r_t, false); 402004c7d789SIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE) { 402104c7d789SIlya Dryomov erase_request(need_resend, req); 402204c7d789SIlya Dryomov check_pool_dne(req); 402304c7d789SIlya Dryomov } 402404c7d789SIlya Dryomov } 402504c7d789SIlya Dryomov } 402604c7d789SIlya Dryomov 40275aea3dcdSIlya Dryomov for (n = rb_first(need_resend); n; ) { 40285aea3dcdSIlya Dryomov struct ceph_osd_request *req = 40295aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 40305aea3dcdSIlya Dryomov struct ceph_osd *osd; 40315aea3dcdSIlya Dryomov 40325aea3dcdSIlya Dryomov n = rb_next(n); 40335aea3dcdSIlya Dryomov erase_request(need_resend, req); /* before link_request() */ 40345aea3dcdSIlya Dryomov 40355aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, true); 40365aea3dcdSIlya Dryomov link_request(osd, req); 40375aea3dcdSIlya Dryomov if (!req->r_linger) { 40385aea3dcdSIlya Dryomov if (!osd_homeless(osd) && !req->r_t.paused) 40395aea3dcdSIlya Dryomov send_request(req); 4040922dab61SIlya Dryomov } else { 4041922dab61SIlya Dryomov cancel_linger_request(req); 40425aea3dcdSIlya Dryomov } 40435aea3dcdSIlya Dryomov } 4044922dab61SIlya Dryomov 4045922dab61SIlya Dryomov list_for_each_entry_safe(lreq, nlreq, need_resend_linger, scan_item) { 4046922dab61SIlya Dryomov if (!osd_homeless(lreq->osd)) 4047922dab61SIlya Dryomov send_linger(lreq); 4048922dab61SIlya Dryomov 4049922dab61SIlya Dryomov list_del_init(&lreq->scan_item); 4050922dab61SIlya Dryomov } 40515aea3dcdSIlya Dryomov } 40525aea3dcdSIlya Dryomov 40533d14c5d2SYehuda Sadeh /* 40543d14c5d2SYehuda Sadeh * Process updated osd map. 40553d14c5d2SYehuda Sadeh * 40563d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 40573d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 40583d14c5d2SYehuda Sadeh * off to different OSDs as needed. 40593d14c5d2SYehuda Sadeh */ 40603d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 40613d14c5d2SYehuda Sadeh { 406242c1b124SIlya Dryomov void *p = msg->front.iov_base; 406342c1b124SIlya Dryomov void *const end = p + msg->front.iov_len; 40643d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 40653d14c5d2SYehuda Sadeh u32 epoch; 40663d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 40675aea3dcdSIlya Dryomov struct rb_root need_resend = RB_ROOT; 40685aea3dcdSIlya Dryomov LIST_HEAD(need_resend_linger); 406942c1b124SIlya Dryomov bool handled_incremental = false; 407042c1b124SIlya Dryomov bool was_pauserd, was_pausewr; 407142c1b124SIlya Dryomov bool pauserd, pausewr; 407242c1b124SIlya Dryomov int err; 40733d14c5d2SYehuda Sadeh 407442c1b124SIlya Dryomov dout("%s have %u\n", __func__, osdc->osdmap->epoch); 40755aea3dcdSIlya Dryomov down_write(&osdc->lock); 40763d14c5d2SYehuda Sadeh 40773d14c5d2SYehuda Sadeh /* verify fsid */ 40783d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 40793d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 40803d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 408142c1b124SIlya Dryomov goto bad; 40823d14c5d2SYehuda Sadeh 4083b7ec35b3SIlya Dryomov was_pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 4084b7ec35b3SIlya Dryomov was_pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 4085b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 408642c1b124SIlya Dryomov have_pool_full(osdc); 40879a1ea2dbSJosh Durgin 40883d14c5d2SYehuda Sadeh /* incremental maps */ 40893d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 40903d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 40913d14c5d2SYehuda Sadeh while (nr_maps > 0) { 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); 409642c1b124SIlya Dryomov if (osdc->osdmap->epoch && 409742c1b124SIlya Dryomov osdc->osdmap->epoch + 1 == epoch) { 40983d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 40993d14c5d2SYehuda Sadeh epoch, maplen); 41005aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, true, 41015aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 410242c1b124SIlya Dryomov if (err) 41033d14c5d2SYehuda Sadeh goto bad; 410442c1b124SIlya Dryomov handled_incremental = true; 41053d14c5d2SYehuda Sadeh } else { 41063d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 41073d14c5d2SYehuda Sadeh epoch, maplen); 41083d14c5d2SYehuda Sadeh } 410942c1b124SIlya Dryomov p += maplen; 41103d14c5d2SYehuda Sadeh nr_maps--; 41113d14c5d2SYehuda Sadeh } 411242c1b124SIlya Dryomov if (handled_incremental) 41133d14c5d2SYehuda Sadeh goto done; 41143d14c5d2SYehuda Sadeh 41153d14c5d2SYehuda Sadeh /* full maps */ 41163d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 41173d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 41183d14c5d2SYehuda Sadeh while (nr_maps) { 41193d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 41203d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 41213d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 41223d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 41233d14c5d2SYehuda Sadeh if (nr_maps > 1) { 41243d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 41253d14c5d2SYehuda Sadeh epoch, maplen); 4126e5253a7bSIlya Dryomov } else if (osdc->osdmap->epoch >= epoch) { 41273d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 41283d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 41293d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 41303d14c5d2SYehuda Sadeh } else { 41313d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 41325aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, false, 41335aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 413442c1b124SIlya Dryomov if (err) 41353d14c5d2SYehuda Sadeh goto bad; 41363d14c5d2SYehuda Sadeh } 41373d14c5d2SYehuda Sadeh p += maplen; 41383d14c5d2SYehuda Sadeh nr_maps--; 41393d14c5d2SYehuda Sadeh } 41403d14c5d2SYehuda Sadeh 41413d14c5d2SYehuda Sadeh done: 4142cd634fb6SSage Weil /* 4143cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 4144cd634fb6SSage Weil * we find out when we are no longer full and stop returning 4145cd634fb6SSage Weil * ENOSPC. 4146cd634fb6SSage Weil */ 4147b7ec35b3SIlya Dryomov pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 4148b7ec35b3SIlya Dryomov pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 4149b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 415042c1b124SIlya Dryomov have_pool_full(osdc); 415158eb7932SJeff Layton if (was_pauserd || was_pausewr || pauserd || pausewr || 415258eb7932SJeff Layton osdc->osdmap->epoch < osdc->epoch_barrier) 415342c1b124SIlya Dryomov maybe_request_map(osdc); 4154cd634fb6SSage Weil 41555aea3dcdSIlya Dryomov kick_requests(osdc, &need_resend, &need_resend_linger); 415642c1b124SIlya Dryomov 4157fc36d0a4SJeff Layton ceph_osdc_abort_on_full(osdc); 415842c1b124SIlya Dryomov ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 415942c1b124SIlya Dryomov osdc->osdmap->epoch); 41605aea3dcdSIlya Dryomov up_write(&osdc->lock); 41613d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 41623d14c5d2SYehuda Sadeh return; 41633d14c5d2SYehuda Sadeh 41643d14c5d2SYehuda Sadeh bad: 41653d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 41663d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 41675aea3dcdSIlya Dryomov up_write(&osdc->lock); 41685aea3dcdSIlya Dryomov } 41695aea3dcdSIlya Dryomov 41705aea3dcdSIlya Dryomov /* 41715aea3dcdSIlya Dryomov * Resubmit requests pending on the given osd. 41725aea3dcdSIlya Dryomov */ 41735aea3dcdSIlya Dryomov static void kick_osd_requests(struct ceph_osd *osd) 41745aea3dcdSIlya Dryomov { 41755aea3dcdSIlya Dryomov struct rb_node *n; 41765aea3dcdSIlya Dryomov 4177a02a946dSIlya Dryomov clear_backoffs(osd); 4178a02a946dSIlya Dryomov 4179922dab61SIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 41805aea3dcdSIlya Dryomov struct ceph_osd_request *req = 41815aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 41825aea3dcdSIlya Dryomov 4183922dab61SIlya Dryomov n = rb_next(n); /* cancel_linger_request() */ 4184922dab61SIlya Dryomov 41855aea3dcdSIlya Dryomov if (!req->r_linger) { 41865aea3dcdSIlya Dryomov if (!req->r_t.paused) 41875aea3dcdSIlya Dryomov send_request(req); 4188922dab61SIlya Dryomov } else { 4189922dab61SIlya Dryomov cancel_linger_request(req); 41905aea3dcdSIlya Dryomov } 41915aea3dcdSIlya Dryomov } 4192922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; n = rb_next(n)) { 4193922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 4194922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 4195922dab61SIlya Dryomov 4196922dab61SIlya Dryomov send_linger(lreq); 4197922dab61SIlya Dryomov } 41985aea3dcdSIlya Dryomov } 41995aea3dcdSIlya Dryomov 42005aea3dcdSIlya Dryomov /* 42015aea3dcdSIlya Dryomov * If the osd connection drops, we need to resubmit all requests. 42025aea3dcdSIlya Dryomov */ 42035aea3dcdSIlya Dryomov static void osd_fault(struct ceph_connection *con) 42045aea3dcdSIlya Dryomov { 42055aea3dcdSIlya Dryomov struct ceph_osd *osd = con->private; 42065aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 42075aea3dcdSIlya Dryomov 42085aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 42095aea3dcdSIlya Dryomov 42105aea3dcdSIlya Dryomov down_write(&osdc->lock); 42115aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 42125aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 42135aea3dcdSIlya Dryomov goto out_unlock; 42145aea3dcdSIlya Dryomov } 42155aea3dcdSIlya Dryomov 42165aea3dcdSIlya Dryomov if (!reopen_osd(osd)) 42175aea3dcdSIlya Dryomov kick_osd_requests(osd); 42185aea3dcdSIlya Dryomov maybe_request_map(osdc); 42195aea3dcdSIlya Dryomov 42205aea3dcdSIlya Dryomov out_unlock: 42215aea3dcdSIlya Dryomov up_write(&osdc->lock); 42223d14c5d2SYehuda Sadeh } 42233d14c5d2SYehuda Sadeh 4224a02a946dSIlya Dryomov struct MOSDBackoff { 4225a02a946dSIlya Dryomov struct ceph_spg spgid; 4226a02a946dSIlya Dryomov u32 map_epoch; 4227a02a946dSIlya Dryomov u8 op; 4228a02a946dSIlya Dryomov u64 id; 4229a02a946dSIlya Dryomov struct ceph_hobject_id *begin; 4230a02a946dSIlya Dryomov struct ceph_hobject_id *end; 4231a02a946dSIlya Dryomov }; 4232a02a946dSIlya Dryomov 4233a02a946dSIlya Dryomov static int decode_MOSDBackoff(const struct ceph_msg *msg, struct MOSDBackoff *m) 4234a02a946dSIlya Dryomov { 4235a02a946dSIlya Dryomov void *p = msg->front.iov_base; 4236a02a946dSIlya Dryomov void *const end = p + msg->front.iov_len; 4237a02a946dSIlya Dryomov u8 struct_v; 4238a02a946dSIlya Dryomov u32 struct_len; 4239a02a946dSIlya Dryomov int ret; 4240a02a946dSIlya Dryomov 4241a02a946dSIlya Dryomov ret = ceph_start_decoding(&p, end, 1, "spg_t", &struct_v, &struct_len); 4242a02a946dSIlya Dryomov if (ret) 4243a02a946dSIlya Dryomov return ret; 4244a02a946dSIlya Dryomov 4245a02a946dSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->spgid.pgid); 4246a02a946dSIlya Dryomov if (ret) 4247a02a946dSIlya Dryomov return ret; 4248a02a946dSIlya Dryomov 4249a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->spgid.shard, e_inval); 4250a02a946dSIlya Dryomov ceph_decode_32_safe(&p, end, m->map_epoch, e_inval); 4251a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->op, e_inval); 4252a02a946dSIlya Dryomov ceph_decode_64_safe(&p, end, m->id, e_inval); 4253a02a946dSIlya Dryomov 4254a02a946dSIlya Dryomov m->begin = kzalloc(sizeof(*m->begin), GFP_NOIO); 4255a02a946dSIlya Dryomov if (!m->begin) 4256a02a946dSIlya Dryomov return -ENOMEM; 4257a02a946dSIlya Dryomov 4258a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->begin); 4259a02a946dSIlya Dryomov if (ret) { 4260a02a946dSIlya Dryomov free_hoid(m->begin); 4261a02a946dSIlya Dryomov return ret; 4262a02a946dSIlya Dryomov } 4263a02a946dSIlya Dryomov 4264a02a946dSIlya Dryomov m->end = kzalloc(sizeof(*m->end), GFP_NOIO); 4265a02a946dSIlya Dryomov if (!m->end) { 4266a02a946dSIlya Dryomov free_hoid(m->begin); 4267a02a946dSIlya Dryomov return -ENOMEM; 4268a02a946dSIlya Dryomov } 4269a02a946dSIlya Dryomov 4270a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->end); 4271a02a946dSIlya Dryomov if (ret) { 4272a02a946dSIlya Dryomov free_hoid(m->begin); 4273a02a946dSIlya Dryomov free_hoid(m->end); 4274a02a946dSIlya Dryomov return ret; 4275a02a946dSIlya Dryomov } 4276a02a946dSIlya Dryomov 4277a02a946dSIlya Dryomov return 0; 4278a02a946dSIlya Dryomov 4279a02a946dSIlya Dryomov e_inval: 4280a02a946dSIlya Dryomov return -EINVAL; 4281a02a946dSIlya Dryomov } 4282a02a946dSIlya Dryomov 4283a02a946dSIlya Dryomov static struct ceph_msg *create_backoff_message( 4284a02a946dSIlya Dryomov const struct ceph_osd_backoff *backoff, 4285a02a946dSIlya Dryomov u32 map_epoch) 4286a02a946dSIlya Dryomov { 4287a02a946dSIlya Dryomov struct ceph_msg *msg; 4288a02a946dSIlya Dryomov void *p, *end; 4289a02a946dSIlya Dryomov int msg_size; 4290a02a946dSIlya Dryomov 4291a02a946dSIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 4292a02a946dSIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 4293a02a946dSIlya Dryomov msg_size += 4 + 1 + 8; /* map_epoch, op, id */ 4294a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4295a02a946dSIlya Dryomov hoid_encoding_size(backoff->begin); 4296a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4297a02a946dSIlya Dryomov hoid_encoding_size(backoff->end); 4298a02a946dSIlya Dryomov 4299a02a946dSIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_BACKOFF, msg_size, GFP_NOIO, true); 4300a02a946dSIlya Dryomov if (!msg) 4301a02a946dSIlya Dryomov return NULL; 4302a02a946dSIlya Dryomov 4303a02a946dSIlya Dryomov p = msg->front.iov_base; 4304a02a946dSIlya Dryomov end = p + msg->front_alloc_len; 4305a02a946dSIlya Dryomov 4306a02a946dSIlya Dryomov encode_spgid(&p, &backoff->spgid); 4307a02a946dSIlya Dryomov ceph_encode_32(&p, map_epoch); 4308a02a946dSIlya Dryomov ceph_encode_8(&p, CEPH_OSD_BACKOFF_OP_ACK_BLOCK); 4309a02a946dSIlya Dryomov ceph_encode_64(&p, backoff->id); 4310a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->begin); 4311a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->end); 4312a02a946dSIlya Dryomov BUG_ON(p != end); 4313a02a946dSIlya Dryomov 4314a02a946dSIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 4315a02a946dSIlya Dryomov msg->hdr.version = cpu_to_le16(1); /* MOSDBackoff v1 */ 4316a02a946dSIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 4317a02a946dSIlya Dryomov 4318a02a946dSIlya Dryomov return msg; 4319a02a946dSIlya Dryomov } 4320a02a946dSIlya Dryomov 4321a02a946dSIlya Dryomov static void handle_backoff_block(struct ceph_osd *osd, struct MOSDBackoff *m) 4322a02a946dSIlya Dryomov { 4323a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4324a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4325a02a946dSIlya Dryomov struct ceph_msg *msg; 4326a02a946dSIlya Dryomov 4327a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4328a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4329a02a946dSIlya Dryomov 4330a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &m->spgid); 4331a02a946dSIlya Dryomov if (!spg) { 4332a02a946dSIlya Dryomov spg = alloc_spg_mapping(); 4333a02a946dSIlya Dryomov if (!spg) { 4334a02a946dSIlya Dryomov pr_err("%s failed to allocate spg\n", __func__); 4335a02a946dSIlya Dryomov return; 4336a02a946dSIlya Dryomov } 4337a02a946dSIlya Dryomov spg->spgid = m->spgid; /* struct */ 4338a02a946dSIlya Dryomov insert_spg_mapping(&osd->o_backoff_mappings, spg); 4339a02a946dSIlya Dryomov } 4340a02a946dSIlya Dryomov 4341a02a946dSIlya Dryomov backoff = alloc_backoff(); 4342a02a946dSIlya Dryomov if (!backoff) { 4343a02a946dSIlya Dryomov pr_err("%s failed to allocate backoff\n", __func__); 4344a02a946dSIlya Dryomov return; 4345a02a946dSIlya Dryomov } 4346a02a946dSIlya Dryomov backoff->spgid = m->spgid; /* struct */ 4347a02a946dSIlya Dryomov backoff->id = m->id; 4348a02a946dSIlya Dryomov backoff->begin = m->begin; 4349a02a946dSIlya Dryomov m->begin = NULL; /* backoff now owns this */ 4350a02a946dSIlya Dryomov backoff->end = m->end; 4351a02a946dSIlya Dryomov m->end = NULL; /* ditto */ 4352a02a946dSIlya Dryomov 4353a02a946dSIlya Dryomov insert_backoff(&spg->backoffs, backoff); 4354a02a946dSIlya Dryomov insert_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4355a02a946dSIlya Dryomov 4356a02a946dSIlya Dryomov /* 4357a02a946dSIlya Dryomov * Ack with original backoff's epoch so that the OSD can 4358a02a946dSIlya Dryomov * discard this if there was a PG split. 4359a02a946dSIlya Dryomov */ 4360a02a946dSIlya Dryomov msg = create_backoff_message(backoff, m->map_epoch); 4361a02a946dSIlya Dryomov if (!msg) { 4362a02a946dSIlya Dryomov pr_err("%s failed to allocate msg\n", __func__); 4363a02a946dSIlya Dryomov return; 4364a02a946dSIlya Dryomov } 4365a02a946dSIlya Dryomov ceph_con_send(&osd->o_con, msg); 4366a02a946dSIlya Dryomov } 4367a02a946dSIlya Dryomov 4368a02a946dSIlya Dryomov static bool target_contained_by(const struct ceph_osd_request_target *t, 4369a02a946dSIlya Dryomov const struct ceph_hobject_id *begin, 4370a02a946dSIlya Dryomov const struct ceph_hobject_id *end) 4371a02a946dSIlya Dryomov { 4372a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 4373a02a946dSIlya Dryomov int cmp; 4374a02a946dSIlya Dryomov 4375a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, t); 4376a02a946dSIlya Dryomov cmp = hoid_compare(&hoid, begin); 4377a02a946dSIlya Dryomov return !cmp || (cmp > 0 && hoid_compare(&hoid, end) < 0); 4378a02a946dSIlya Dryomov } 4379a02a946dSIlya Dryomov 4380a02a946dSIlya Dryomov static void handle_backoff_unblock(struct ceph_osd *osd, 4381a02a946dSIlya Dryomov const struct MOSDBackoff *m) 4382a02a946dSIlya Dryomov { 4383a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4384a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4385a02a946dSIlya Dryomov struct rb_node *n; 4386a02a946dSIlya Dryomov 4387a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4388a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4389a02a946dSIlya Dryomov 4390a02a946dSIlya Dryomov backoff = lookup_backoff_by_id(&osd->o_backoffs_by_id, m->id); 4391a02a946dSIlya Dryomov if (!backoff) { 4392a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu backoff dne\n", 4393a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4394a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4395a02a946dSIlya Dryomov return; 4396a02a946dSIlya Dryomov } 4397a02a946dSIlya Dryomov 4398a02a946dSIlya Dryomov if (hoid_compare(backoff->begin, m->begin) && 4399a02a946dSIlya Dryomov hoid_compare(backoff->end, m->end)) { 4400a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu bad range?\n", 4401a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4402a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4403a02a946dSIlya Dryomov /* unblock it anyway... */ 4404a02a946dSIlya Dryomov } 4405a02a946dSIlya Dryomov 4406a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &backoff->spgid); 4407a02a946dSIlya Dryomov BUG_ON(!spg); 4408a02a946dSIlya Dryomov 4409a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 4410a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4411a02a946dSIlya Dryomov free_backoff(backoff); 4412a02a946dSIlya Dryomov 4413a02a946dSIlya Dryomov if (RB_EMPTY_ROOT(&spg->backoffs)) { 4414a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 4415a02a946dSIlya Dryomov free_spg_mapping(spg); 4416a02a946dSIlya Dryomov } 4417a02a946dSIlya Dryomov 4418a02a946dSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 4419a02a946dSIlya Dryomov struct ceph_osd_request *req = 4420a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 4421a02a946dSIlya Dryomov 4422a02a946dSIlya Dryomov if (!ceph_spg_compare(&req->r_t.spgid, &m->spgid)) { 4423a02a946dSIlya Dryomov /* 4424a02a946dSIlya Dryomov * Match against @m, not @backoff -- the PG may 4425a02a946dSIlya Dryomov * have split on the OSD. 4426a02a946dSIlya Dryomov */ 4427a02a946dSIlya Dryomov if (target_contained_by(&req->r_t, m->begin, m->end)) { 4428a02a946dSIlya Dryomov /* 4429a02a946dSIlya Dryomov * If no other installed backoff applies, 4430a02a946dSIlya Dryomov * resend. 4431a02a946dSIlya Dryomov */ 4432a02a946dSIlya Dryomov send_request(req); 4433a02a946dSIlya Dryomov } 4434a02a946dSIlya Dryomov } 4435a02a946dSIlya Dryomov } 4436a02a946dSIlya Dryomov } 4437a02a946dSIlya Dryomov 4438a02a946dSIlya Dryomov static void handle_backoff(struct ceph_osd *osd, struct ceph_msg *msg) 4439a02a946dSIlya Dryomov { 4440a02a946dSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 4441a02a946dSIlya Dryomov struct MOSDBackoff m; 4442a02a946dSIlya Dryomov int ret; 4443a02a946dSIlya Dryomov 4444a02a946dSIlya Dryomov down_read(&osdc->lock); 4445a02a946dSIlya Dryomov if (!osd_registered(osd)) { 4446a02a946dSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 4447a02a946dSIlya Dryomov up_read(&osdc->lock); 4448a02a946dSIlya Dryomov return; 4449a02a946dSIlya Dryomov } 4450a02a946dSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 4451a02a946dSIlya Dryomov 4452a02a946dSIlya Dryomov mutex_lock(&osd->lock); 4453a02a946dSIlya Dryomov ret = decode_MOSDBackoff(msg, &m); 4454a02a946dSIlya Dryomov if (ret) { 4455a02a946dSIlya Dryomov pr_err("failed to decode MOSDBackoff: %d\n", ret); 4456a02a946dSIlya Dryomov ceph_msg_dump(msg); 4457a02a946dSIlya Dryomov goto out_unlock; 4458a02a946dSIlya Dryomov } 4459a02a946dSIlya Dryomov 4460a02a946dSIlya Dryomov switch (m.op) { 4461a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_BLOCK: 4462a02a946dSIlya Dryomov handle_backoff_block(osd, &m); 4463a02a946dSIlya Dryomov break; 4464a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_UNBLOCK: 4465a02a946dSIlya Dryomov handle_backoff_unblock(osd, &m); 4466a02a946dSIlya Dryomov break; 4467a02a946dSIlya Dryomov default: 4468a02a946dSIlya Dryomov pr_err("%s osd%d unknown op %d\n", __func__, osd->o_osd, m.op); 4469a02a946dSIlya Dryomov } 4470a02a946dSIlya Dryomov 4471a02a946dSIlya Dryomov free_hoid(m.begin); 4472a02a946dSIlya Dryomov free_hoid(m.end); 4473a02a946dSIlya Dryomov 4474a02a946dSIlya Dryomov out_unlock: 4475a02a946dSIlya Dryomov mutex_unlock(&osd->lock); 4476a02a946dSIlya Dryomov up_read(&osdc->lock); 4477a02a946dSIlya Dryomov } 4478a02a946dSIlya Dryomov 44793d14c5d2SYehuda Sadeh /* 4480a40c4f10SYehuda Sadeh * Process osd watch notifications 4481a40c4f10SYehuda Sadeh */ 44823c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 44833c663bbdSAlex Elder struct ceph_msg *msg) 4484a40c4f10SYehuda Sadeh { 4485922dab61SIlya Dryomov void *p = msg->front.iov_base; 4486922dab61SIlya Dryomov void *const end = p + msg->front.iov_len; 4487922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4488922dab61SIlya Dryomov struct linger_work *lwork; 4489922dab61SIlya Dryomov u8 proto_ver, opcode; 4490922dab61SIlya Dryomov u64 cookie, notify_id; 4491922dab61SIlya Dryomov u64 notifier_id = 0; 449219079203SIlya Dryomov s32 return_code = 0; 4493922dab61SIlya Dryomov void *payload = NULL; 4494922dab61SIlya Dryomov u32 payload_len = 0; 4495a40c4f10SYehuda Sadeh 4496a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 4497a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 4498a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 4499922dab61SIlya Dryomov p += 8; /* skip ver */ 4500a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 4501a40c4f10SYehuda Sadeh 4502922dab61SIlya Dryomov if (proto_ver >= 1) { 4503922dab61SIlya Dryomov ceph_decode_32_safe(&p, end, payload_len, bad); 4504922dab61SIlya Dryomov ceph_decode_need(&p, end, payload_len, bad); 4505922dab61SIlya Dryomov payload = p; 4506922dab61SIlya Dryomov p += payload_len; 4507a40c4f10SYehuda Sadeh } 4508a40c4f10SYehuda Sadeh 4509922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 2) 451019079203SIlya Dryomov ceph_decode_32_safe(&p, end, return_code, bad); 4511922dab61SIlya Dryomov 4512922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 3) 4513922dab61SIlya Dryomov ceph_decode_64_safe(&p, end, notifier_id, bad); 4514922dab61SIlya Dryomov 4515922dab61SIlya Dryomov down_read(&osdc->lock); 4516922dab61SIlya Dryomov lreq = lookup_linger_osdc(&osdc->linger_requests, cookie); 4517922dab61SIlya Dryomov if (!lreq) { 4518922dab61SIlya Dryomov dout("%s opcode %d cookie %llu dne\n", __func__, opcode, 4519922dab61SIlya Dryomov cookie); 4520922dab61SIlya Dryomov goto out_unlock_osdc; 4521922dab61SIlya Dryomov } 4522922dab61SIlya Dryomov 4523922dab61SIlya Dryomov mutex_lock(&lreq->lock); 452419079203SIlya Dryomov dout("%s opcode %d cookie %llu lreq %p is_watch %d\n", __func__, 452519079203SIlya Dryomov opcode, cookie, lreq, lreq->is_watch); 4526922dab61SIlya Dryomov if (opcode == CEPH_WATCH_EVENT_DISCONNECT) { 4527922dab61SIlya Dryomov if (!lreq->last_error) { 4528922dab61SIlya Dryomov lreq->last_error = -ENOTCONN; 4529922dab61SIlya Dryomov queue_watch_error(lreq); 4530922dab61SIlya Dryomov } 453119079203SIlya Dryomov } else if (!lreq->is_watch) { 453219079203SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY_COMPLETE */ 453319079203SIlya Dryomov if (lreq->notify_id && lreq->notify_id != notify_id) { 453419079203SIlya Dryomov dout("lreq %p notify_id %llu != %llu, ignoring\n", lreq, 453519079203SIlya Dryomov lreq->notify_id, notify_id); 453619079203SIlya Dryomov } else if (!completion_done(&lreq->notify_finish_wait)) { 453719079203SIlya Dryomov struct ceph_msg_data *data = 45380d9c1ab3SIlya Dryomov msg->num_data_items ? &msg->data[0] : NULL; 453919079203SIlya Dryomov 454019079203SIlya Dryomov if (data) { 454119079203SIlya Dryomov if (lreq->preply_pages) { 454219079203SIlya Dryomov WARN_ON(data->type != 454319079203SIlya Dryomov CEPH_MSG_DATA_PAGES); 454419079203SIlya Dryomov *lreq->preply_pages = data->pages; 454519079203SIlya Dryomov *lreq->preply_len = data->length; 4546e8862740SIlya Dryomov data->own_pages = false; 454719079203SIlya Dryomov } 454819079203SIlya Dryomov } 454919079203SIlya Dryomov lreq->notify_finish_error = return_code; 455019079203SIlya Dryomov complete_all(&lreq->notify_finish_wait); 455119079203SIlya Dryomov } 4552922dab61SIlya Dryomov } else { 4553922dab61SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY */ 4554922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_notify); 4555922dab61SIlya Dryomov if (!lwork) { 4556922dab61SIlya Dryomov pr_err("failed to allocate notify-lwork\n"); 4557922dab61SIlya Dryomov goto out_unlock_lreq; 4558922dab61SIlya Dryomov } 4559922dab61SIlya Dryomov 4560922dab61SIlya Dryomov lwork->notify.notify_id = notify_id; 4561922dab61SIlya Dryomov lwork->notify.notifier_id = notifier_id; 4562922dab61SIlya Dryomov lwork->notify.payload = payload; 4563922dab61SIlya Dryomov lwork->notify.payload_len = payload_len; 4564922dab61SIlya Dryomov lwork->notify.msg = ceph_msg_get(msg); 4565922dab61SIlya Dryomov lwork_queue(lwork); 4566922dab61SIlya Dryomov } 4567922dab61SIlya Dryomov 4568922dab61SIlya Dryomov out_unlock_lreq: 4569922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 4570922dab61SIlya Dryomov out_unlock_osdc: 4571922dab61SIlya Dryomov up_read(&osdc->lock); 4572a40c4f10SYehuda Sadeh return; 4573a40c4f10SYehuda Sadeh 4574a40c4f10SYehuda Sadeh bad: 4575a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 4576a40c4f10SYehuda Sadeh } 4577a40c4f10SYehuda Sadeh 4578a40c4f10SYehuda Sadeh /* 45793d14c5d2SYehuda Sadeh * Register request, send initial attempt. 45803d14c5d2SYehuda Sadeh */ 45813d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 45823d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 45833d14c5d2SYehuda Sadeh bool nofail) 45843d14c5d2SYehuda Sadeh { 45855aea3dcdSIlya Dryomov down_read(&osdc->lock); 45865aea3dcdSIlya Dryomov submit_request(req, false); 45875aea3dcdSIlya Dryomov up_read(&osdc->lock); 45883d14c5d2SYehuda Sadeh 45895aea3dcdSIlya Dryomov return 0; 45903d14c5d2SYehuda Sadeh } 45913d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 45923d14c5d2SYehuda Sadeh 45933d14c5d2SYehuda Sadeh /* 4594*d0bb883cSIlya Dryomov * Unregister request. If @req was registered, it isn't completed: 4595*d0bb883cSIlya Dryomov * r_result isn't set and __complete_request() isn't invoked. 4596*d0bb883cSIlya Dryomov * 4597*d0bb883cSIlya Dryomov * If @req wasn't registered, this call may have raced with 4598*d0bb883cSIlya Dryomov * handle_reply(), in which case r_result would already be set and 4599*d0bb883cSIlya Dryomov * __complete_request() would be getting invoked, possibly even 4600*d0bb883cSIlya Dryomov * concurrently with this call. 4601c9f9b93dSIlya Dryomov */ 4602c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 4603c9f9b93dSIlya Dryomov { 4604c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 4605c9f9b93dSIlya Dryomov 46065aea3dcdSIlya Dryomov down_write(&osdc->lock); 46075aea3dcdSIlya Dryomov if (req->r_osd) 46085aea3dcdSIlya Dryomov cancel_request(req); 46095aea3dcdSIlya Dryomov up_write(&osdc->lock); 4610c9f9b93dSIlya Dryomov } 4611c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 4612c9f9b93dSIlya Dryomov 4613c9f9b93dSIlya Dryomov /* 461442b06965SIlya Dryomov * @timeout: in jiffies, 0 means "wait forever" 461542b06965SIlya Dryomov */ 461642b06965SIlya Dryomov static int wait_request_timeout(struct ceph_osd_request *req, 461742b06965SIlya Dryomov unsigned long timeout) 461842b06965SIlya Dryomov { 461942b06965SIlya Dryomov long left; 462042b06965SIlya Dryomov 462142b06965SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 46220e76abf2SYan, Zheng left = wait_for_completion_killable_timeout(&req->r_completion, 462342b06965SIlya Dryomov ceph_timeout_jiffies(timeout)); 462442b06965SIlya Dryomov if (left <= 0) { 462542b06965SIlya Dryomov left = left ?: -ETIMEDOUT; 462642b06965SIlya Dryomov ceph_osdc_cancel_request(req); 462742b06965SIlya Dryomov } else { 462842b06965SIlya Dryomov left = req->r_result; /* completed */ 462942b06965SIlya Dryomov } 463042b06965SIlya Dryomov 463142b06965SIlya Dryomov return left; 463242b06965SIlya Dryomov } 463342b06965SIlya Dryomov 463442b06965SIlya Dryomov /* 46353d14c5d2SYehuda Sadeh * wait for a request to complete 46363d14c5d2SYehuda Sadeh */ 46373d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 46383d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 46393d14c5d2SYehuda Sadeh { 464042b06965SIlya Dryomov return wait_request_timeout(req, 0); 46413d14c5d2SYehuda Sadeh } 46423d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 46433d14c5d2SYehuda Sadeh 46443d14c5d2SYehuda Sadeh /* 46453d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 46463d14c5d2SYehuda Sadeh */ 46473d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 46483d14c5d2SYehuda Sadeh { 46495aea3dcdSIlya Dryomov struct rb_node *n, *p; 46505aea3dcdSIlya Dryomov u64 last_tid = atomic64_read(&osdc->last_tid); 46513d14c5d2SYehuda Sadeh 46525aea3dcdSIlya Dryomov again: 46535aea3dcdSIlya Dryomov down_read(&osdc->lock); 46545aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 46555aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 46565aea3dcdSIlya Dryomov 46575aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 46585aea3dcdSIlya Dryomov for (p = rb_first(&osd->o_requests); p; p = rb_next(p)) { 46595aea3dcdSIlya Dryomov struct ceph_osd_request *req = 46605aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 46615aea3dcdSIlya Dryomov 46623d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 46633d14c5d2SYehuda Sadeh break; 46643d14c5d2SYehuda Sadeh 46655aea3dcdSIlya Dryomov if (!(req->r_flags & CEPH_OSD_FLAG_WRITE)) 46663d14c5d2SYehuda Sadeh continue; 46673d14c5d2SYehuda Sadeh 46683d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 46695aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 46705aea3dcdSIlya Dryomov up_read(&osdc->lock); 46715aea3dcdSIlya Dryomov dout("%s waiting on req %p tid %llu last_tid %llu\n", 46725aea3dcdSIlya Dryomov __func__, req, req->r_tid, last_tid); 4673b18b9550SIlya Dryomov wait_for_completion(&req->r_completion); 46743d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 46755aea3dcdSIlya Dryomov goto again; 46763d14c5d2SYehuda Sadeh } 46775aea3dcdSIlya Dryomov 46785aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 46795aea3dcdSIlya Dryomov } 46805aea3dcdSIlya Dryomov 46815aea3dcdSIlya Dryomov up_read(&osdc->lock); 46825aea3dcdSIlya Dryomov dout("%s done last_tid %llu\n", __func__, last_tid); 46833d14c5d2SYehuda Sadeh } 46843d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 46853d14c5d2SYehuda Sadeh 4686922dab61SIlya Dryomov /* 4687922dab61SIlya Dryomov * Returns a handle, caller owns a ref. 4688922dab61SIlya Dryomov */ 4689922dab61SIlya Dryomov struct ceph_osd_linger_request * 4690922dab61SIlya Dryomov ceph_osdc_watch(struct ceph_osd_client *osdc, 4691922dab61SIlya Dryomov struct ceph_object_id *oid, 4692922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4693922dab61SIlya Dryomov rados_watchcb2_t wcb, 4694922dab61SIlya Dryomov rados_watcherrcb_t errcb, 4695922dab61SIlya Dryomov void *data) 4696922dab61SIlya Dryomov { 4697922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4698922dab61SIlya Dryomov int ret; 4699922dab61SIlya Dryomov 4700922dab61SIlya Dryomov lreq = linger_alloc(osdc); 4701922dab61SIlya Dryomov if (!lreq) 4702922dab61SIlya Dryomov return ERR_PTR(-ENOMEM); 4703922dab61SIlya Dryomov 470419079203SIlya Dryomov lreq->is_watch = true; 4705922dab61SIlya Dryomov lreq->wcb = wcb; 4706922dab61SIlya Dryomov lreq->errcb = errcb; 4707922dab61SIlya Dryomov lreq->data = data; 4708b07d3c4bSIlya Dryomov lreq->watch_valid_thru = jiffies; 4709922dab61SIlya Dryomov 4710922dab61SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 4711922dab61SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 471254ea0046SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_WRITE; 4713fac02ddfSArnd Bergmann ktime_get_real_ts64(&lreq->mtime); 4714922dab61SIlya Dryomov 471581c65213SIlya Dryomov linger_submit(lreq); 4716922dab61SIlya Dryomov ret = linger_reg_commit_wait(lreq); 4717922dab61SIlya Dryomov if (ret) { 4718922dab61SIlya Dryomov linger_cancel(lreq); 4719922dab61SIlya Dryomov goto err_put_lreq; 4720922dab61SIlya Dryomov } 4721922dab61SIlya Dryomov 4722922dab61SIlya Dryomov return lreq; 4723922dab61SIlya Dryomov 4724922dab61SIlya Dryomov err_put_lreq: 4725922dab61SIlya Dryomov linger_put(lreq); 4726922dab61SIlya Dryomov return ERR_PTR(ret); 4727922dab61SIlya Dryomov } 4728922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_watch); 4729922dab61SIlya Dryomov 4730922dab61SIlya Dryomov /* 4731922dab61SIlya Dryomov * Releases a ref. 4732922dab61SIlya Dryomov * 4733922dab61SIlya Dryomov * Times out after mount_timeout to preserve rbd unmap behaviour 4734922dab61SIlya Dryomov * introduced in 2894e1d76974 ("rbd: timeout watch teardown on unmap 4735922dab61SIlya Dryomov * with mount_timeout"). 4736922dab61SIlya Dryomov */ 4737922dab61SIlya Dryomov int ceph_osdc_unwatch(struct ceph_osd_client *osdc, 4738922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 4739922dab61SIlya Dryomov { 4740922dab61SIlya Dryomov struct ceph_options *opts = osdc->client->options; 4741922dab61SIlya Dryomov struct ceph_osd_request *req; 4742922dab61SIlya Dryomov int ret; 4743922dab61SIlya Dryomov 4744922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4745922dab61SIlya Dryomov if (!req) 4746922dab61SIlya Dryomov return -ENOMEM; 4747922dab61SIlya Dryomov 4748922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4749922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 475054ea0046SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_WRITE; 4751fac02ddfSArnd Bergmann ktime_get_real_ts64(&req->r_mtime); 475275dbb685SIlya Dryomov osd_req_op_watch_init(req, 0, CEPH_OSD_WATCH_OP_UNWATCH, 475375dbb685SIlya Dryomov lreq->linger_id, 0); 4754922dab61SIlya Dryomov 4755922dab61SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4756922dab61SIlya Dryomov if (ret) 4757922dab61SIlya Dryomov goto out_put_req; 4758922dab61SIlya Dryomov 4759922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4760922dab61SIlya Dryomov linger_cancel(lreq); 4761922dab61SIlya Dryomov linger_put(lreq); 4762922dab61SIlya Dryomov ret = wait_request_timeout(req, opts->mount_timeout); 4763922dab61SIlya Dryomov 4764922dab61SIlya Dryomov out_put_req: 4765922dab61SIlya Dryomov ceph_osdc_put_request(req); 4766922dab61SIlya Dryomov return ret; 4767922dab61SIlya Dryomov } 4768922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_unwatch); 4769922dab61SIlya Dryomov 4770922dab61SIlya Dryomov static int osd_req_op_notify_ack_init(struct ceph_osd_request *req, int which, 4771922dab61SIlya Dryomov u64 notify_id, u64 cookie, void *payload, 47726d54228fSIlya Dryomov u32 payload_len) 4773922dab61SIlya Dryomov { 4774922dab61SIlya Dryomov struct ceph_osd_req_op *op; 4775922dab61SIlya Dryomov struct ceph_pagelist *pl; 4776922dab61SIlya Dryomov int ret; 4777922dab61SIlya Dryomov 4778042f6498SJeff Layton op = osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY_ACK, 0); 4779922dab61SIlya Dryomov 478033165d47SIlya Dryomov pl = ceph_pagelist_alloc(GFP_NOIO); 4781922dab61SIlya Dryomov if (!pl) 4782922dab61SIlya Dryomov return -ENOMEM; 4783922dab61SIlya Dryomov 4784922dab61SIlya Dryomov ret = ceph_pagelist_encode_64(pl, notify_id); 4785922dab61SIlya Dryomov ret |= ceph_pagelist_encode_64(pl, cookie); 4786922dab61SIlya Dryomov if (payload) { 4787922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 4788922dab61SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 4789922dab61SIlya Dryomov } else { 4790922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, 0); 4791922dab61SIlya Dryomov } 4792922dab61SIlya Dryomov if (ret) { 4793922dab61SIlya Dryomov ceph_pagelist_release(pl); 4794922dab61SIlya Dryomov return -ENOMEM; 4795922dab61SIlya Dryomov } 4796922dab61SIlya Dryomov 4797922dab61SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify_ack.request_data, pl); 4798922dab61SIlya Dryomov op->indata_len = pl->length; 4799922dab61SIlya Dryomov return 0; 4800922dab61SIlya Dryomov } 4801922dab61SIlya Dryomov 4802922dab61SIlya Dryomov int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, 4803922dab61SIlya Dryomov struct ceph_object_id *oid, 4804922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4805922dab61SIlya Dryomov u64 notify_id, 4806922dab61SIlya Dryomov u64 cookie, 4807922dab61SIlya Dryomov void *payload, 48086d54228fSIlya Dryomov u32 payload_len) 4809922dab61SIlya Dryomov { 4810922dab61SIlya Dryomov struct ceph_osd_request *req; 4811922dab61SIlya Dryomov int ret; 4812922dab61SIlya Dryomov 4813922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4814922dab61SIlya Dryomov if (!req) 4815922dab61SIlya Dryomov return -ENOMEM; 4816922dab61SIlya Dryomov 4817922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, oid); 4818922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, oloc); 4819922dab61SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_READ; 4820922dab61SIlya Dryomov 482126f887e0SIlya Dryomov ret = osd_req_op_notify_ack_init(req, 0, notify_id, cookie, payload, 482226f887e0SIlya Dryomov payload_len); 4823922dab61SIlya Dryomov if (ret) 4824922dab61SIlya Dryomov goto out_put_req; 4825922dab61SIlya Dryomov 482626f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4827922dab61SIlya Dryomov if (ret) 4828922dab61SIlya Dryomov goto out_put_req; 4829922dab61SIlya Dryomov 4830922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4831922dab61SIlya Dryomov ret = ceph_osdc_wait_request(osdc, req); 4832922dab61SIlya Dryomov 4833922dab61SIlya Dryomov out_put_req: 4834922dab61SIlya Dryomov ceph_osdc_put_request(req); 4835922dab61SIlya Dryomov return ret; 4836922dab61SIlya Dryomov } 4837922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify_ack); 4838922dab61SIlya Dryomov 483919079203SIlya Dryomov /* 484019079203SIlya Dryomov * @timeout: in seconds 484119079203SIlya Dryomov * 484219079203SIlya Dryomov * @preply_{pages,len} are initialized both on success and error. 484319079203SIlya Dryomov * The caller is responsible for: 484419079203SIlya Dryomov * 484519079203SIlya Dryomov * ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len)) 484619079203SIlya Dryomov */ 484719079203SIlya Dryomov int ceph_osdc_notify(struct ceph_osd_client *osdc, 484819079203SIlya Dryomov struct ceph_object_id *oid, 484919079203SIlya Dryomov struct ceph_object_locator *oloc, 485019079203SIlya Dryomov void *payload, 48516d54228fSIlya Dryomov u32 payload_len, 485219079203SIlya Dryomov u32 timeout, 485319079203SIlya Dryomov struct page ***preply_pages, 485419079203SIlya Dryomov size_t *preply_len) 485519079203SIlya Dryomov { 485619079203SIlya Dryomov struct ceph_osd_linger_request *lreq; 485719079203SIlya Dryomov int ret; 485819079203SIlya Dryomov 485919079203SIlya Dryomov WARN_ON(!timeout); 486019079203SIlya Dryomov if (preply_pages) { 486119079203SIlya Dryomov *preply_pages = NULL; 486219079203SIlya Dryomov *preply_len = 0; 486319079203SIlya Dryomov } 486419079203SIlya Dryomov 486519079203SIlya Dryomov lreq = linger_alloc(osdc); 486619079203SIlya Dryomov if (!lreq) 486719079203SIlya Dryomov return -ENOMEM; 486819079203SIlya Dryomov 486975dbb685SIlya Dryomov lreq->request_pl = ceph_pagelist_alloc(GFP_NOIO); 487075dbb685SIlya Dryomov if (!lreq->request_pl) { 487175dbb685SIlya Dryomov ret = -ENOMEM; 487275dbb685SIlya Dryomov goto out_put_lreq; 487375dbb685SIlya Dryomov } 487475dbb685SIlya Dryomov 487575dbb685SIlya Dryomov ret = ceph_pagelist_encode_32(lreq->request_pl, 1); /* prot_ver */ 487675dbb685SIlya Dryomov ret |= ceph_pagelist_encode_32(lreq->request_pl, timeout); 487775dbb685SIlya Dryomov ret |= ceph_pagelist_encode_32(lreq->request_pl, payload_len); 487875dbb685SIlya Dryomov ret |= ceph_pagelist_append(lreq->request_pl, payload, payload_len); 487975dbb685SIlya Dryomov if (ret) { 488075dbb685SIlya Dryomov ret = -ENOMEM; 488175dbb685SIlya Dryomov goto out_put_lreq; 488275dbb685SIlya Dryomov } 488375dbb685SIlya Dryomov 488475dbb685SIlya Dryomov /* for notify_id */ 488575dbb685SIlya Dryomov lreq->notify_id_pages = ceph_alloc_page_vector(1, GFP_NOIO); 488675dbb685SIlya Dryomov if (IS_ERR(lreq->notify_id_pages)) { 488775dbb685SIlya Dryomov ret = PTR_ERR(lreq->notify_id_pages); 488875dbb685SIlya Dryomov lreq->notify_id_pages = NULL; 488975dbb685SIlya Dryomov goto out_put_lreq; 489075dbb685SIlya Dryomov } 489175dbb685SIlya Dryomov 489219079203SIlya Dryomov lreq->preply_pages = preply_pages; 489319079203SIlya Dryomov lreq->preply_len = preply_len; 489419079203SIlya Dryomov 489519079203SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 489619079203SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 489719079203SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_READ; 489819079203SIlya Dryomov 489981c65213SIlya Dryomov linger_submit(lreq); 490019079203SIlya Dryomov ret = linger_reg_commit_wait(lreq); 490119079203SIlya Dryomov if (!ret) 490219079203SIlya Dryomov ret = linger_notify_finish_wait(lreq); 490319079203SIlya Dryomov else 490419079203SIlya Dryomov dout("lreq %p failed to initiate notify %d\n", lreq, ret); 490519079203SIlya Dryomov 490619079203SIlya Dryomov linger_cancel(lreq); 490719079203SIlya Dryomov out_put_lreq: 490819079203SIlya Dryomov linger_put(lreq); 490919079203SIlya Dryomov return ret; 491019079203SIlya Dryomov } 491119079203SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify); 491219079203SIlya Dryomov 49133d14c5d2SYehuda Sadeh /* 4914b07d3c4bSIlya Dryomov * Return the number of milliseconds since the watch was last 4915b07d3c4bSIlya Dryomov * confirmed, or an error. If there is an error, the watch is no 4916b07d3c4bSIlya Dryomov * longer valid, and should be destroyed with ceph_osdc_unwatch(). 4917b07d3c4bSIlya Dryomov */ 4918b07d3c4bSIlya Dryomov int ceph_osdc_watch_check(struct ceph_osd_client *osdc, 4919b07d3c4bSIlya Dryomov struct ceph_osd_linger_request *lreq) 4920b07d3c4bSIlya Dryomov { 4921b07d3c4bSIlya Dryomov unsigned long stamp, age; 4922b07d3c4bSIlya Dryomov int ret; 4923b07d3c4bSIlya Dryomov 4924b07d3c4bSIlya Dryomov down_read(&osdc->lock); 4925b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 4926b07d3c4bSIlya Dryomov stamp = lreq->watch_valid_thru; 4927b07d3c4bSIlya Dryomov if (!list_empty(&lreq->pending_lworks)) { 4928b07d3c4bSIlya Dryomov struct linger_work *lwork = 4929b07d3c4bSIlya Dryomov list_first_entry(&lreq->pending_lworks, 4930b07d3c4bSIlya Dryomov struct linger_work, 4931b07d3c4bSIlya Dryomov pending_item); 4932b07d3c4bSIlya Dryomov 4933b07d3c4bSIlya Dryomov if (time_before(lwork->queued_stamp, stamp)) 4934b07d3c4bSIlya Dryomov stamp = lwork->queued_stamp; 4935b07d3c4bSIlya Dryomov } 4936b07d3c4bSIlya Dryomov age = jiffies - stamp; 4937b07d3c4bSIlya Dryomov dout("%s lreq %p linger_id %llu age %lu last_error %d\n", __func__, 4938b07d3c4bSIlya Dryomov lreq, lreq->linger_id, age, lreq->last_error); 4939b07d3c4bSIlya Dryomov /* we are truncating to msecs, so return a safe upper bound */ 4940b07d3c4bSIlya Dryomov ret = lreq->last_error ?: 1 + jiffies_to_msecs(age); 4941b07d3c4bSIlya Dryomov 4942b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 4943b07d3c4bSIlya Dryomov up_read(&osdc->lock); 4944b07d3c4bSIlya Dryomov return ret; 4945b07d3c4bSIlya Dryomov } 4946b07d3c4bSIlya Dryomov 4947a4ed38d7SDouglas Fuller static int decode_watcher(void **p, void *end, struct ceph_watch_item *item) 4948a4ed38d7SDouglas Fuller { 4949a4ed38d7SDouglas Fuller u8 struct_v; 4950a4ed38d7SDouglas Fuller u32 struct_len; 4951a4ed38d7SDouglas Fuller int ret; 4952a4ed38d7SDouglas Fuller 4953a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 2, "watch_item_t", 4954a4ed38d7SDouglas Fuller &struct_v, &struct_len); 4955a4ed38d7SDouglas Fuller if (ret) 495651fc7ab4SJeff Layton goto bad; 4957a4ed38d7SDouglas Fuller 495851fc7ab4SJeff Layton ret = -EINVAL; 495951fc7ab4SJeff Layton ceph_decode_copy_safe(p, end, &item->name, sizeof(item->name), bad); 496051fc7ab4SJeff Layton ceph_decode_64_safe(p, end, item->cookie, bad); 496151fc7ab4SJeff Layton ceph_decode_skip_32(p, end, bad); /* skip timeout seconds */ 496251fc7ab4SJeff Layton 4963a4ed38d7SDouglas Fuller if (struct_v >= 2) { 496451fc7ab4SJeff Layton ret = ceph_decode_entity_addr(p, end, &item->addr); 496551fc7ab4SJeff Layton if (ret) 496651fc7ab4SJeff Layton goto bad; 496751fc7ab4SJeff Layton } else { 496851fc7ab4SJeff Layton ret = 0; 4969a4ed38d7SDouglas Fuller } 4970a4ed38d7SDouglas Fuller 4971a4ed38d7SDouglas Fuller dout("%s %s%llu cookie %llu addr %s\n", __func__, 4972a4ed38d7SDouglas Fuller ENTITY_NAME(item->name), item->cookie, 4973b726ec97SJeff Layton ceph_pr_addr(&item->addr)); 497451fc7ab4SJeff Layton bad: 497551fc7ab4SJeff Layton return ret; 4976a4ed38d7SDouglas Fuller } 4977a4ed38d7SDouglas Fuller 4978a4ed38d7SDouglas Fuller static int decode_watchers(void **p, void *end, 4979a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 4980a4ed38d7SDouglas Fuller u32 *num_watchers) 4981a4ed38d7SDouglas Fuller { 4982a4ed38d7SDouglas Fuller u8 struct_v; 4983a4ed38d7SDouglas Fuller u32 struct_len; 4984a4ed38d7SDouglas Fuller int i; 4985a4ed38d7SDouglas Fuller int ret; 4986a4ed38d7SDouglas Fuller 4987a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 1, "obj_list_watch_response_t", 4988a4ed38d7SDouglas Fuller &struct_v, &struct_len); 4989a4ed38d7SDouglas Fuller if (ret) 4990a4ed38d7SDouglas Fuller return ret; 4991a4ed38d7SDouglas Fuller 4992a4ed38d7SDouglas Fuller *num_watchers = ceph_decode_32(p); 4993a4ed38d7SDouglas Fuller *watchers = kcalloc(*num_watchers, sizeof(**watchers), GFP_NOIO); 4994a4ed38d7SDouglas Fuller if (!*watchers) 4995a4ed38d7SDouglas Fuller return -ENOMEM; 4996a4ed38d7SDouglas Fuller 4997a4ed38d7SDouglas Fuller for (i = 0; i < *num_watchers; i++) { 4998a4ed38d7SDouglas Fuller ret = decode_watcher(p, end, *watchers + i); 4999a4ed38d7SDouglas Fuller if (ret) { 5000a4ed38d7SDouglas Fuller kfree(*watchers); 5001a4ed38d7SDouglas Fuller return ret; 5002a4ed38d7SDouglas Fuller } 5003a4ed38d7SDouglas Fuller } 5004a4ed38d7SDouglas Fuller 5005a4ed38d7SDouglas Fuller return 0; 5006a4ed38d7SDouglas Fuller } 5007a4ed38d7SDouglas Fuller 5008a4ed38d7SDouglas Fuller /* 5009a4ed38d7SDouglas Fuller * On success, the caller is responsible for: 5010a4ed38d7SDouglas Fuller * 5011a4ed38d7SDouglas Fuller * kfree(watchers); 5012a4ed38d7SDouglas Fuller */ 5013a4ed38d7SDouglas Fuller int ceph_osdc_list_watchers(struct ceph_osd_client *osdc, 5014a4ed38d7SDouglas Fuller struct ceph_object_id *oid, 5015a4ed38d7SDouglas Fuller struct ceph_object_locator *oloc, 5016a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 5017a4ed38d7SDouglas Fuller u32 *num_watchers) 5018a4ed38d7SDouglas Fuller { 5019a4ed38d7SDouglas Fuller struct ceph_osd_request *req; 5020a4ed38d7SDouglas Fuller struct page **pages; 5021a4ed38d7SDouglas Fuller int ret; 5022a4ed38d7SDouglas Fuller 5023a4ed38d7SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 5024a4ed38d7SDouglas Fuller if (!req) 5025a4ed38d7SDouglas Fuller return -ENOMEM; 5026a4ed38d7SDouglas Fuller 5027a4ed38d7SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 5028a4ed38d7SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 5029a4ed38d7SDouglas Fuller req->r_flags = CEPH_OSD_FLAG_READ; 5030a4ed38d7SDouglas Fuller 5031a4ed38d7SDouglas Fuller pages = ceph_alloc_page_vector(1, GFP_NOIO); 5032a4ed38d7SDouglas Fuller if (IS_ERR(pages)) { 5033a4ed38d7SDouglas Fuller ret = PTR_ERR(pages); 5034a4ed38d7SDouglas Fuller goto out_put_req; 5035a4ed38d7SDouglas Fuller } 5036a4ed38d7SDouglas Fuller 5037a4ed38d7SDouglas Fuller osd_req_op_init(req, 0, CEPH_OSD_OP_LIST_WATCHERS, 0); 5038a4ed38d7SDouglas Fuller ceph_osd_data_pages_init(osd_req_op_data(req, 0, list_watchers, 5039a4ed38d7SDouglas Fuller response_data), 5040a4ed38d7SDouglas Fuller pages, PAGE_SIZE, 0, false, true); 5041a4ed38d7SDouglas Fuller 504226f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 504326f887e0SIlya Dryomov if (ret) 504426f887e0SIlya Dryomov goto out_put_req; 504526f887e0SIlya Dryomov 5046a4ed38d7SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 5047a4ed38d7SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 5048a4ed38d7SDouglas Fuller if (ret >= 0) { 5049a4ed38d7SDouglas Fuller void *p = page_address(pages[0]); 5050a4ed38d7SDouglas Fuller void *const end = p + req->r_ops[0].outdata_len; 5051a4ed38d7SDouglas Fuller 5052a4ed38d7SDouglas Fuller ret = decode_watchers(&p, end, watchers, num_watchers); 5053a4ed38d7SDouglas Fuller } 5054a4ed38d7SDouglas Fuller 5055a4ed38d7SDouglas Fuller out_put_req: 5056a4ed38d7SDouglas Fuller ceph_osdc_put_request(req); 5057a4ed38d7SDouglas Fuller return ret; 5058a4ed38d7SDouglas Fuller } 5059a4ed38d7SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_list_watchers); 5060a4ed38d7SDouglas Fuller 5061b07d3c4bSIlya Dryomov /* 5062dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 5063dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 5064dd935f44SJosh Durgin */ 5065f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 5066dd935f44SJosh Durgin { 506799d16943SIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 5068dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 5069dd935f44SJosh Durgin } 5070dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 5071dd935f44SJosh Durgin 50727cca78c9SIlya Dryomov void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc) 50737cca78c9SIlya Dryomov { 50747cca78c9SIlya Dryomov down_read(&osdc->lock); 50757cca78c9SIlya Dryomov maybe_request_map(osdc); 50767cca78c9SIlya Dryomov up_read(&osdc->lock); 50777cca78c9SIlya Dryomov } 50787cca78c9SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_maybe_request_map); 5079dd935f44SJosh Durgin 5080dd935f44SJosh Durgin /* 5081428a7158SDouglas Fuller * Execute an OSD class method on an object. 5082428a7158SDouglas Fuller * 5083428a7158SDouglas Fuller * @flags: CEPH_OSD_FLAG_* 50842544a020SIlya Dryomov * @resp_len: in/out param for reply length 5085428a7158SDouglas Fuller */ 5086428a7158SDouglas Fuller int ceph_osdc_call(struct ceph_osd_client *osdc, 5087428a7158SDouglas Fuller struct ceph_object_id *oid, 5088428a7158SDouglas Fuller struct ceph_object_locator *oloc, 5089428a7158SDouglas Fuller const char *class, const char *method, 5090428a7158SDouglas Fuller unsigned int flags, 5091428a7158SDouglas Fuller struct page *req_page, size_t req_len, 509268ada915SIlya Dryomov struct page **resp_pages, size_t *resp_len) 5093428a7158SDouglas Fuller { 5094428a7158SDouglas Fuller struct ceph_osd_request *req; 5095428a7158SDouglas Fuller int ret; 5096428a7158SDouglas Fuller 509768ada915SIlya Dryomov if (req_len > PAGE_SIZE) 50982544a020SIlya Dryomov return -E2BIG; 50992544a020SIlya Dryomov 5100428a7158SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 5101428a7158SDouglas Fuller if (!req) 5102428a7158SDouglas Fuller return -ENOMEM; 5103428a7158SDouglas Fuller 5104428a7158SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 5105428a7158SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 5106428a7158SDouglas Fuller req->r_flags = flags; 5107428a7158SDouglas Fuller 510824639ce5SIlya Dryomov ret = osd_req_op_cls_init(req, 0, class, method); 5109fe943d50SChengguang Xu if (ret) 5110fe943d50SChengguang Xu goto out_put_req; 5111fe943d50SChengguang Xu 5112428a7158SDouglas Fuller if (req_page) 5113428a7158SDouglas Fuller osd_req_op_cls_request_data_pages(req, 0, &req_page, req_len, 5114428a7158SDouglas Fuller 0, false, false); 511568ada915SIlya Dryomov if (resp_pages) 511668ada915SIlya Dryomov osd_req_op_cls_response_data_pages(req, 0, resp_pages, 51172544a020SIlya Dryomov *resp_len, 0, false, false); 5118428a7158SDouglas Fuller 511926f887e0SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 512026f887e0SIlya Dryomov if (ret) 512126f887e0SIlya Dryomov goto out_put_req; 512226f887e0SIlya Dryomov 5123428a7158SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 5124428a7158SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 5125428a7158SDouglas Fuller if (ret >= 0) { 5126428a7158SDouglas Fuller ret = req->r_ops[0].rval; 512768ada915SIlya Dryomov if (resp_pages) 5128428a7158SDouglas Fuller *resp_len = req->r_ops[0].outdata_len; 5129428a7158SDouglas Fuller } 5130428a7158SDouglas Fuller 5131428a7158SDouglas Fuller out_put_req: 5132428a7158SDouglas Fuller ceph_osdc_put_request(req); 5133428a7158SDouglas Fuller return ret; 5134428a7158SDouglas Fuller } 5135428a7158SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_call); 5136428a7158SDouglas Fuller 5137428a7158SDouglas Fuller /* 5138120a75eaSYan, Zheng * reset all osd connections 5139120a75eaSYan, Zheng */ 5140120a75eaSYan, Zheng void ceph_osdc_reopen_osds(struct ceph_osd_client *osdc) 5141120a75eaSYan, Zheng { 5142120a75eaSYan, Zheng struct rb_node *n; 5143120a75eaSYan, Zheng 5144120a75eaSYan, Zheng down_write(&osdc->lock); 5145120a75eaSYan, Zheng for (n = rb_first(&osdc->osds); n; ) { 5146120a75eaSYan, Zheng struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 5147120a75eaSYan, Zheng 5148120a75eaSYan, Zheng n = rb_next(n); 5149120a75eaSYan, Zheng if (!reopen_osd(osd)) 5150120a75eaSYan, Zheng kick_osd_requests(osd); 5151120a75eaSYan, Zheng } 5152120a75eaSYan, Zheng up_write(&osdc->lock); 5153120a75eaSYan, Zheng } 5154120a75eaSYan, Zheng 5155120a75eaSYan, Zheng /* 51563d14c5d2SYehuda Sadeh * init, shutdown 51573d14c5d2SYehuda Sadeh */ 51583d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 51593d14c5d2SYehuda Sadeh { 51603d14c5d2SYehuda Sadeh int err; 51613d14c5d2SYehuda Sadeh 51623d14c5d2SYehuda Sadeh dout("init\n"); 51633d14c5d2SYehuda Sadeh osdc->client = client; 51645aea3dcdSIlya Dryomov init_rwsem(&osdc->lock); 51653d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 51663d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 51679dd2845cSIlya Dryomov spin_lock_init(&osdc->osd_lru_lock); 51685aea3dcdSIlya Dryomov osd_init(&osdc->homeless_osd); 51695aea3dcdSIlya Dryomov osdc->homeless_osd.o_osdc = osdc; 51705aea3dcdSIlya Dryomov osdc->homeless_osd.o_osd = CEPH_HOMELESS_OSD; 5171264048afSIlya Dryomov osdc->last_linger_id = CEPH_LINGER_ID_START; 5172922dab61SIlya Dryomov osdc->linger_requests = RB_ROOT; 51734609245eSIlya Dryomov osdc->map_checks = RB_ROOT; 51744609245eSIlya Dryomov osdc->linger_map_checks = RB_ROOT; 51753d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 51763d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 51773d14c5d2SYehuda Sadeh 51783d14c5d2SYehuda Sadeh err = -ENOMEM; 5179e5253a7bSIlya Dryomov osdc->osdmap = ceph_osdmap_alloc(); 5180e5253a7bSIlya Dryomov if (!osdc->osdmap) 5181e5253a7bSIlya Dryomov goto out; 5182e5253a7bSIlya Dryomov 51839e767adbSIlya Dryomov osdc->req_mempool = mempool_create_slab_pool(10, 51849e767adbSIlya Dryomov ceph_osd_request_cache); 51853d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 5186e5253a7bSIlya Dryomov goto out_map; 51873d14c5d2SYehuda Sadeh 5188d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 51890d9c1ab3SIlya Dryomov PAGE_SIZE, CEPH_OSD_SLAB_OPS, 10, "osd_op"); 51903d14c5d2SYehuda Sadeh if (err < 0) 51913d14c5d2SYehuda Sadeh goto out_mempool; 5192d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 51930d9c1ab3SIlya Dryomov PAGE_SIZE, CEPH_OSD_SLAB_OPS, 10, 51940d9c1ab3SIlya Dryomov "osd_op_reply"); 51953d14c5d2SYehuda Sadeh if (err < 0) 51963d14c5d2SYehuda Sadeh goto out_msgpool; 5197a40c4f10SYehuda Sadeh 5198dbcae088SDan Carpenter err = -ENOMEM; 5199a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 5200dbcae088SDan Carpenter if (!osdc->notify_wq) 5201c172ec5cSIlya Dryomov goto out_msgpool_reply; 5202c172ec5cSIlya Dryomov 520388bc1922SIlya Dryomov osdc->completion_wq = create_singlethread_workqueue("ceph-completion"); 520488bc1922SIlya Dryomov if (!osdc->completion_wq) 520588bc1922SIlya Dryomov goto out_notify_wq; 520688bc1922SIlya Dryomov 5207fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 5208fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 5209b37ee1b9SIlya Dryomov schedule_delayed_work(&osdc->osds_timeout_work, 5210b37ee1b9SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 5211b37ee1b9SIlya Dryomov 52123d14c5d2SYehuda Sadeh return 0; 52133d14c5d2SYehuda Sadeh 521488bc1922SIlya Dryomov out_notify_wq: 521588bc1922SIlya Dryomov destroy_workqueue(osdc->notify_wq); 5216c172ec5cSIlya Dryomov out_msgpool_reply: 5217c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 52183d14c5d2SYehuda Sadeh out_msgpool: 52193d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 52203d14c5d2SYehuda Sadeh out_mempool: 52213d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 5222e5253a7bSIlya Dryomov out_map: 5223e5253a7bSIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 52243d14c5d2SYehuda Sadeh out: 52253d14c5d2SYehuda Sadeh return err; 52263d14c5d2SYehuda Sadeh } 52273d14c5d2SYehuda Sadeh 52283d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 52293d14c5d2SYehuda Sadeh { 523088bc1922SIlya Dryomov destroy_workqueue(osdc->completion_wq); 5231a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 52323d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 52333d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 523442a2c09fSIlya Dryomov 52355aea3dcdSIlya Dryomov down_write(&osdc->lock); 523642a2c09fSIlya Dryomov while (!RB_EMPTY_ROOT(&osdc->osds)) { 523742a2c09fSIlya Dryomov struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 523842a2c09fSIlya Dryomov struct ceph_osd, o_node); 52395aea3dcdSIlya Dryomov close_osd(osd); 524042a2c09fSIlya Dryomov } 52415aea3dcdSIlya Dryomov up_write(&osdc->lock); 524202113a0fSElena Reshetova WARN_ON(refcount_read(&osdc->homeless_osd.o_ref) != 1); 52435aea3dcdSIlya Dryomov osd_cleanup(&osdc->homeless_osd); 52445aea3dcdSIlya Dryomov 52455aea3dcdSIlya Dryomov WARN_ON(!list_empty(&osdc->osd_lru)); 5246922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_requests)); 52474609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->map_checks)); 52484609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_map_checks)); 52495aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_requests)); 52505aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_homeless)); 525142a2c09fSIlya Dryomov 52523d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 52533d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 52543d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 52553d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 52563d14c5d2SYehuda Sadeh } 52573d14c5d2SYehuda Sadeh 5258aca39d9eSLuís Henriques int osd_req_op_copy_from_init(struct ceph_osd_request *req, 525923ddf9beSLuis Henriques u64 src_snapid, u64 src_version, 526023ddf9beSLuis Henriques struct ceph_object_id *src_oid, 526123ddf9beSLuis Henriques struct ceph_object_locator *src_oloc, 526223ddf9beSLuis Henriques u32 src_fadvise_flags, 526323ddf9beSLuis Henriques u32 dst_fadvise_flags, 526478beb0ffSLuis Henriques u32 truncate_seq, u64 truncate_size, 526523ddf9beSLuis Henriques u8 copy_from_flags) 526623ddf9beSLuis Henriques { 526723ddf9beSLuis Henriques struct ceph_osd_req_op *op; 526823ddf9beSLuis Henriques struct page **pages; 526923ddf9beSLuis Henriques void *p, *end; 527023ddf9beSLuis Henriques 527123ddf9beSLuis Henriques pages = ceph_alloc_page_vector(1, GFP_KERNEL); 527223ddf9beSLuis Henriques if (IS_ERR(pages)) 527323ddf9beSLuis Henriques return PTR_ERR(pages); 527423ddf9beSLuis Henriques 5275042f6498SJeff Layton op = osd_req_op_init(req, 0, CEPH_OSD_OP_COPY_FROM2, 527678beb0ffSLuis Henriques dst_fadvise_flags); 527723ddf9beSLuis Henriques op->copy_from.snapid = src_snapid; 527823ddf9beSLuis Henriques op->copy_from.src_version = src_version; 527923ddf9beSLuis Henriques op->copy_from.flags = copy_from_flags; 528023ddf9beSLuis Henriques op->copy_from.src_fadvise_flags = src_fadvise_flags; 528123ddf9beSLuis Henriques 528223ddf9beSLuis Henriques p = page_address(pages[0]); 528323ddf9beSLuis Henriques end = p + PAGE_SIZE; 528423ddf9beSLuis Henriques ceph_encode_string(&p, end, src_oid->name, src_oid->name_len); 528523ddf9beSLuis Henriques encode_oloc(&p, end, src_oloc); 528678beb0ffSLuis Henriques ceph_encode_32(&p, truncate_seq); 528778beb0ffSLuis Henriques ceph_encode_64(&p, truncate_size); 528823ddf9beSLuis Henriques op->indata_len = PAGE_SIZE - (end - p); 528923ddf9beSLuis Henriques 529023ddf9beSLuis Henriques ceph_osd_data_pages_init(&op->copy_from.osd_data, pages, 529123ddf9beSLuis Henriques op->indata_len, 0, false, true); 529223ddf9beSLuis Henriques return 0; 529323ddf9beSLuis Henriques } 5294aca39d9eSLuís Henriques EXPORT_SYMBOL(osd_req_op_copy_from_init); 529523ddf9beSLuis Henriques 529657a35dfbSChengguang Xu int __init ceph_osdc_setup(void) 52975522ae0bSAlex Elder { 52983f1af42aSIlya Dryomov size_t size = sizeof(struct ceph_osd_request) + 52993f1af42aSIlya Dryomov CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op); 53003f1af42aSIlya Dryomov 53015522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 53023f1af42aSIlya Dryomov ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size, 53033f1af42aSIlya Dryomov 0, 0, NULL); 53045522ae0bSAlex Elder 53055522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 53065522ae0bSAlex Elder } 53075522ae0bSAlex Elder 53085522ae0bSAlex Elder void ceph_osdc_cleanup(void) 53095522ae0bSAlex Elder { 53105522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 53115522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 53125522ae0bSAlex Elder ceph_osd_request_cache = NULL; 53135522ae0bSAlex Elder } 53145522ae0bSAlex Elder 53153d14c5d2SYehuda Sadeh /* 53163d14c5d2SYehuda Sadeh * handle incoming message 53173d14c5d2SYehuda Sadeh */ 53184972cf60SIlya Dryomov static void osd_dispatch(struct ceph_connection *con, struct ceph_msg *msg) 53193d14c5d2SYehuda Sadeh { 53203d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 53215aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 53223d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 53233d14c5d2SYehuda Sadeh 53243d14c5d2SYehuda Sadeh switch (type) { 53253d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 53263d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 53273d14c5d2SYehuda Sadeh break; 53283d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 53295aea3dcdSIlya Dryomov handle_reply(osd, msg); 53303d14c5d2SYehuda Sadeh break; 5331a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5332a02a946dSIlya Dryomov handle_backoff(osd, msg); 5333a02a946dSIlya Dryomov break; 5334a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 5335a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 5336a40c4f10SYehuda Sadeh break; 53373d14c5d2SYehuda Sadeh 53383d14c5d2SYehuda Sadeh default: 53393d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 53403d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 53413d14c5d2SYehuda Sadeh } 53425aea3dcdSIlya Dryomov 53433d14c5d2SYehuda Sadeh ceph_msg_put(msg); 53443d14c5d2SYehuda Sadeh } 53453d14c5d2SYehuda Sadeh 53463d14c5d2SYehuda Sadeh /* 5347d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 5348d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 5349d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 53503d14c5d2SYehuda Sadeh */ 53513d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 53523d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 53533d14c5d2SYehuda Sadeh int *skip) 53543d14c5d2SYehuda Sadeh { 53553d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 53563d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 53575aea3dcdSIlya Dryomov struct ceph_msg *m = NULL; 53583d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 53593f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 53603d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 53615aea3dcdSIlya Dryomov u64 tid = le64_to_cpu(hdr->tid); 53623d14c5d2SYehuda Sadeh 53635aea3dcdSIlya Dryomov down_read(&osdc->lock); 53645aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 53655aea3dcdSIlya Dryomov dout("%s osd%d unknown, skipping\n", __func__, osd->o_osd); 53665aea3dcdSIlya Dryomov *skip = 1; 53675aea3dcdSIlya Dryomov goto out_unlock_osdc; 53685aea3dcdSIlya Dryomov } 53695aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(hdr->src.num)); 53705aea3dcdSIlya Dryomov 53715aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 53725aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 53733d14c5d2SYehuda Sadeh if (!req) { 5374cd8140c6SIlya Dryomov dout("%s osd%d tid %llu unknown, skipping\n", __func__, 5375cd8140c6SIlya Dryomov osd->o_osd, tid); 5376d15f9d69SIlya Dryomov *skip = 1; 53775aea3dcdSIlya Dryomov goto out_unlock_session; 53783d14c5d2SYehuda Sadeh } 53793d14c5d2SYehuda Sadeh 53808921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 53813d14c5d2SYehuda Sadeh 5382f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 5383d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 5384d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 5385d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 53863f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 53873f0a4ac5SIlya Dryomov false); 53883d14c5d2SYehuda Sadeh if (!m) 53895aea3dcdSIlya Dryomov goto out_unlock_session; 53903d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 53913d14c5d2SYehuda Sadeh req->r_reply = m; 53923d14c5d2SYehuda Sadeh } 53933d14c5d2SYehuda Sadeh 5394d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 5395d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 5396d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 5397d15f9d69SIlya Dryomov req->r_reply->data_length); 53983d14c5d2SYehuda Sadeh m = NULL; 5399d15f9d69SIlya Dryomov *skip = 1; 54005aea3dcdSIlya Dryomov goto out_unlock_session; 54013d14c5d2SYehuda Sadeh } 5402d15f9d69SIlya Dryomov 5403d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 54043d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 54053d14c5d2SYehuda Sadeh 54065aea3dcdSIlya Dryomov out_unlock_session: 54075aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 54085aea3dcdSIlya Dryomov out_unlock_osdc: 54095aea3dcdSIlya Dryomov up_read(&osdc->lock); 54103d14c5d2SYehuda Sadeh return m; 54113d14c5d2SYehuda Sadeh } 54123d14c5d2SYehuda Sadeh 541319079203SIlya Dryomov static struct ceph_msg *alloc_msg_with_page_vector(struct ceph_msg_header *hdr) 541419079203SIlya Dryomov { 541519079203SIlya Dryomov struct ceph_msg *m; 541619079203SIlya Dryomov int type = le16_to_cpu(hdr->type); 541719079203SIlya Dryomov u32 front_len = le32_to_cpu(hdr->front_len); 541819079203SIlya Dryomov u32 data_len = le32_to_cpu(hdr->data_len); 541919079203SIlya Dryomov 54200d9c1ab3SIlya Dryomov m = ceph_msg_new2(type, front_len, 1, GFP_NOIO, false); 542119079203SIlya Dryomov if (!m) 542219079203SIlya Dryomov return NULL; 542319079203SIlya Dryomov 542419079203SIlya Dryomov if (data_len) { 542519079203SIlya Dryomov struct page **pages; 542619079203SIlya Dryomov 542719079203SIlya Dryomov pages = ceph_alloc_page_vector(calc_pages_for(0, data_len), 542819079203SIlya Dryomov GFP_NOIO); 5429c22e853aSWei Yongjun if (IS_ERR(pages)) { 543019079203SIlya Dryomov ceph_msg_put(m); 543119079203SIlya Dryomov return NULL; 543219079203SIlya Dryomov } 543319079203SIlya Dryomov 5434e8862740SIlya Dryomov ceph_msg_data_add_pages(m, pages, data_len, 0, true); 543519079203SIlya Dryomov } 543619079203SIlya Dryomov 543719079203SIlya Dryomov return m; 543819079203SIlya Dryomov } 543919079203SIlya Dryomov 54404972cf60SIlya Dryomov static struct ceph_msg *osd_alloc_msg(struct ceph_connection *con, 54413d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 54423d14c5d2SYehuda Sadeh int *skip) 54433d14c5d2SYehuda Sadeh { 54443d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 54453d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 54463d14c5d2SYehuda Sadeh 54471c20f2d2SAlex Elder *skip = 0; 54483d14c5d2SYehuda Sadeh switch (type) { 54493d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 5450a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5451a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 545219079203SIlya Dryomov return alloc_msg_with_page_vector(hdr); 54533d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 54543d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 54553d14c5d2SYehuda Sadeh default: 54565aea3dcdSIlya Dryomov pr_warn("%s osd%d unknown msg type %d, skipping\n", __func__, 54575aea3dcdSIlya Dryomov osd->o_osd, type); 54583d14c5d2SYehuda Sadeh *skip = 1; 54593d14c5d2SYehuda Sadeh return NULL; 54603d14c5d2SYehuda Sadeh } 54613d14c5d2SYehuda Sadeh } 54623d14c5d2SYehuda Sadeh 54633d14c5d2SYehuda Sadeh /* 54643d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 54653d14c5d2SYehuda Sadeh */ 54664972cf60SIlya Dryomov static struct ceph_connection *osd_get_con(struct ceph_connection *con) 54673d14c5d2SYehuda Sadeh { 54683d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 54693d14c5d2SYehuda Sadeh if (get_osd(osd)) 54703d14c5d2SYehuda Sadeh return con; 54713d14c5d2SYehuda Sadeh return NULL; 54723d14c5d2SYehuda Sadeh } 54733d14c5d2SYehuda Sadeh 54744972cf60SIlya Dryomov static void osd_put_con(struct ceph_connection *con) 54753d14c5d2SYehuda Sadeh { 54763d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 54773d14c5d2SYehuda Sadeh put_osd(osd); 54783d14c5d2SYehuda Sadeh } 54793d14c5d2SYehuda Sadeh 54803d14c5d2SYehuda Sadeh /* 54813d14c5d2SYehuda Sadeh * authentication 54823d14c5d2SYehuda Sadeh */ 5483cd1a677cSIlya Dryomov 5484a3530df3SAlex Elder /* 5485a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 5486a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 5487a3530df3SAlex Elder */ 54884972cf60SIlya Dryomov static struct ceph_auth_handshake * 54894972cf60SIlya Dryomov osd_get_authorizer(struct ceph_connection *con, int *proto, int force_new) 54903d14c5d2SYehuda Sadeh { 54913d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 54923d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 54933d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 549474f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 5495ce287162SIlya Dryomov int ret; 54963d14c5d2SYehuda Sadeh 5497ce287162SIlya Dryomov ret = __ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_OSD, 5498ce287162SIlya Dryomov force_new, proto, NULL, NULL); 54993d14c5d2SYehuda Sadeh if (ret) 5500a3530df3SAlex Elder return ERR_PTR(ret); 550174f1869fSAlex Elder 5502a3530df3SAlex Elder return auth; 55033d14c5d2SYehuda Sadeh } 55043d14c5d2SYehuda Sadeh 55054972cf60SIlya Dryomov static int osd_add_authorizer_challenge(struct ceph_connection *con, 55066daca13dSIlya Dryomov void *challenge_buf, int challenge_buf_len) 55076daca13dSIlya Dryomov { 55086daca13dSIlya Dryomov struct ceph_osd *o = con->private; 55096daca13dSIlya Dryomov struct ceph_osd_client *osdc = o->o_osdc; 55106daca13dSIlya Dryomov struct ceph_auth_client *ac = osdc->client->monc.auth; 55116daca13dSIlya Dryomov 55126daca13dSIlya Dryomov return ceph_auth_add_authorizer_challenge(ac, o->o_auth.authorizer, 55136daca13dSIlya Dryomov challenge_buf, challenge_buf_len); 55146daca13dSIlya Dryomov } 55153d14c5d2SYehuda Sadeh 55164972cf60SIlya Dryomov static int osd_verify_authorizer_reply(struct ceph_connection *con) 55173d14c5d2SYehuda Sadeh { 55183d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 55193d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 55203d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 5521285ea34fSIlya Dryomov struct ceph_auth_handshake *auth = &o->o_auth; 55223d14c5d2SYehuda Sadeh 5523285ea34fSIlya Dryomov return ceph_auth_verify_authorizer_reply(ac, auth->authorizer, 5524285ea34fSIlya Dryomov auth->authorizer_reply_buf, auth->authorizer_reply_buf_len, 5525285ea34fSIlya Dryomov NULL, NULL, NULL, NULL); 55263d14c5d2SYehuda Sadeh } 55273d14c5d2SYehuda Sadeh 55284972cf60SIlya Dryomov static int osd_invalidate_authorizer(struct ceph_connection *con) 55293d14c5d2SYehuda Sadeh { 55303d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 55313d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 55323d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 55333d14c5d2SYehuda Sadeh 553427859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 55353d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 55363d14c5d2SYehuda Sadeh } 55373d14c5d2SYehuda Sadeh 5538cd1a677cSIlya Dryomov static int osd_get_auth_request(struct ceph_connection *con, 5539cd1a677cSIlya Dryomov void *buf, int *buf_len, 5540cd1a677cSIlya Dryomov void **authorizer, int *authorizer_len) 5541cd1a677cSIlya Dryomov { 5542cd1a677cSIlya Dryomov struct ceph_osd *o = con->private; 5543cd1a677cSIlya Dryomov struct ceph_auth_client *ac = o->o_osdc->client->monc.auth; 5544cd1a677cSIlya Dryomov struct ceph_auth_handshake *auth = &o->o_auth; 5545cd1a677cSIlya Dryomov int ret; 5546cd1a677cSIlya Dryomov 5547cd1a677cSIlya Dryomov ret = ceph_auth_get_authorizer(ac, auth, CEPH_ENTITY_TYPE_OSD, 5548cd1a677cSIlya Dryomov buf, buf_len); 5549cd1a677cSIlya Dryomov if (ret) 5550cd1a677cSIlya Dryomov return ret; 5551cd1a677cSIlya Dryomov 5552cd1a677cSIlya Dryomov *authorizer = auth->authorizer_buf; 5553cd1a677cSIlya Dryomov *authorizer_len = auth->authorizer_buf_len; 5554cd1a677cSIlya Dryomov return 0; 5555cd1a677cSIlya Dryomov } 5556cd1a677cSIlya Dryomov 5557cd1a677cSIlya Dryomov static int osd_handle_auth_reply_more(struct ceph_connection *con, 5558cd1a677cSIlya Dryomov void *reply, int reply_len, 5559cd1a677cSIlya Dryomov void *buf, int *buf_len, 5560cd1a677cSIlya Dryomov void **authorizer, int *authorizer_len) 5561cd1a677cSIlya Dryomov { 5562cd1a677cSIlya Dryomov struct ceph_osd *o = con->private; 5563cd1a677cSIlya Dryomov struct ceph_auth_client *ac = o->o_osdc->client->monc.auth; 5564cd1a677cSIlya Dryomov struct ceph_auth_handshake *auth = &o->o_auth; 5565cd1a677cSIlya Dryomov int ret; 5566cd1a677cSIlya Dryomov 5567cd1a677cSIlya Dryomov ret = ceph_auth_handle_svc_reply_more(ac, auth, reply, reply_len, 5568cd1a677cSIlya Dryomov buf, buf_len); 5569cd1a677cSIlya Dryomov if (ret) 5570cd1a677cSIlya Dryomov return ret; 5571cd1a677cSIlya Dryomov 5572cd1a677cSIlya Dryomov *authorizer = auth->authorizer_buf; 5573cd1a677cSIlya Dryomov *authorizer_len = auth->authorizer_buf_len; 5574cd1a677cSIlya Dryomov return 0; 5575cd1a677cSIlya Dryomov } 5576cd1a677cSIlya Dryomov 5577cd1a677cSIlya Dryomov static int osd_handle_auth_done(struct ceph_connection *con, 5578cd1a677cSIlya Dryomov u64 global_id, void *reply, int reply_len, 5579cd1a677cSIlya Dryomov u8 *session_key, int *session_key_len, 5580cd1a677cSIlya Dryomov u8 *con_secret, int *con_secret_len) 5581cd1a677cSIlya Dryomov { 5582cd1a677cSIlya Dryomov struct ceph_osd *o = con->private; 5583cd1a677cSIlya Dryomov struct ceph_auth_client *ac = o->o_osdc->client->monc.auth; 5584cd1a677cSIlya Dryomov struct ceph_auth_handshake *auth = &o->o_auth; 5585cd1a677cSIlya Dryomov 5586cd1a677cSIlya Dryomov return ceph_auth_handle_svc_reply_done(ac, auth, reply, reply_len, 5587cd1a677cSIlya Dryomov session_key, session_key_len, 5588cd1a677cSIlya Dryomov con_secret, con_secret_len); 5589cd1a677cSIlya Dryomov } 5590cd1a677cSIlya Dryomov 5591cd1a677cSIlya Dryomov static int osd_handle_auth_bad_method(struct ceph_connection *con, 5592cd1a677cSIlya Dryomov int used_proto, int result, 5593cd1a677cSIlya Dryomov const int *allowed_protos, int proto_cnt, 5594cd1a677cSIlya Dryomov const int *allowed_modes, int mode_cnt) 5595cd1a677cSIlya Dryomov { 5596cd1a677cSIlya Dryomov struct ceph_osd *o = con->private; 5597cd1a677cSIlya Dryomov struct ceph_mon_client *monc = &o->o_osdc->client->monc; 5598cd1a677cSIlya Dryomov int ret; 5599cd1a677cSIlya Dryomov 5600cd1a677cSIlya Dryomov if (ceph_auth_handle_bad_authorizer(monc->auth, CEPH_ENTITY_TYPE_OSD, 5601cd1a677cSIlya Dryomov used_proto, result, 5602cd1a677cSIlya Dryomov allowed_protos, proto_cnt, 5603cd1a677cSIlya Dryomov allowed_modes, mode_cnt)) { 5604cd1a677cSIlya Dryomov ret = ceph_monc_validate_auth(monc); 5605cd1a677cSIlya Dryomov if (ret) 5606cd1a677cSIlya Dryomov return ret; 5607cd1a677cSIlya Dryomov } 5608cd1a677cSIlya Dryomov 5609cd1a677cSIlya Dryomov return -EACCES; 5610cd1a677cSIlya Dryomov } 5611cd1a677cSIlya Dryomov 56128cb441c0SIlya Dryomov static void osd_reencode_message(struct ceph_msg *msg) 56138cb441c0SIlya Dryomov { 5614914902afSIlya Dryomov int type = le16_to_cpu(msg->hdr.type); 5615914902afSIlya Dryomov 5616914902afSIlya Dryomov if (type == CEPH_MSG_OSD_OP) 56178cb441c0SIlya Dryomov encode_request_finish(msg); 56188cb441c0SIlya Dryomov } 56198cb441c0SIlya Dryomov 562079dbd1baSIlya Dryomov static int osd_sign_message(struct ceph_msg *msg) 562133d07337SYan, Zheng { 562279dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 562333d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 562479dbd1baSIlya Dryomov 562533d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 562633d07337SYan, Zheng } 562733d07337SYan, Zheng 562879dbd1baSIlya Dryomov static int osd_check_message_signature(struct ceph_msg *msg) 562933d07337SYan, Zheng { 563079dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 563133d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 563279dbd1baSIlya Dryomov 563333d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 563433d07337SYan, Zheng } 563533d07337SYan, Zheng 56363d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 56374972cf60SIlya Dryomov .get = osd_get_con, 56384972cf60SIlya Dryomov .put = osd_put_con, 56394972cf60SIlya Dryomov .alloc_msg = osd_alloc_msg, 56404972cf60SIlya Dryomov .dispatch = osd_dispatch, 56414972cf60SIlya Dryomov .fault = osd_fault, 56428cb441c0SIlya Dryomov .reencode_message = osd_reencode_message, 56434972cf60SIlya Dryomov .get_authorizer = osd_get_authorizer, 56444972cf60SIlya Dryomov .add_authorizer_challenge = osd_add_authorizer_challenge, 56454972cf60SIlya Dryomov .verify_authorizer_reply = osd_verify_authorizer_reply, 56464972cf60SIlya Dryomov .invalidate_authorizer = osd_invalidate_authorizer, 564779dbd1baSIlya Dryomov .sign_message = osd_sign_message, 564879dbd1baSIlya Dryomov .check_message_signature = osd_check_message_signature, 5649cd1a677cSIlya Dryomov .get_auth_request = osd_get_auth_request, 5650cd1a677cSIlya Dryomov .handle_auth_reply_more = osd_handle_auth_reply_more, 5651cd1a677cSIlya Dryomov .handle_auth_done = osd_handle_auth_done, 5652cd1a677cSIlya Dryomov .handle_auth_bad_method = osd_handle_auth_bad_method, 56533d14c5d2SYehuda Sadeh }; 5654