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 129a4ce40a9SAlex Elder static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data, 13043bfe5deSAlex Elder struct page **pages, u64 length, u32 alignment, 13143bfe5deSAlex Elder bool pages_from_pool, bool own_pages) 13243bfe5deSAlex Elder { 13343bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGES; 13443bfe5deSAlex Elder osd_data->pages = pages; 13543bfe5deSAlex Elder osd_data->length = length; 13643bfe5deSAlex Elder osd_data->alignment = alignment; 13743bfe5deSAlex Elder osd_data->pages_from_pool = pages_from_pool; 13843bfe5deSAlex Elder osd_data->own_pages = own_pages; 13943bfe5deSAlex Elder } 14043bfe5deSAlex Elder 141a4ce40a9SAlex Elder static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data, 14243bfe5deSAlex Elder struct ceph_pagelist *pagelist) 14343bfe5deSAlex Elder { 14443bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST; 14543bfe5deSAlex Elder osd_data->pagelist = pagelist; 14643bfe5deSAlex Elder } 14743bfe5deSAlex Elder 14843bfe5deSAlex Elder #ifdef CONFIG_BLOCK 149a4ce40a9SAlex Elder static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data, 1505359a17dSIlya Dryomov struct ceph_bio_iter *bio_pos, 1515359a17dSIlya Dryomov u32 bio_length) 15243bfe5deSAlex Elder { 15343bfe5deSAlex Elder osd_data->type = CEPH_OSD_DATA_TYPE_BIO; 1545359a17dSIlya Dryomov osd_data->bio_pos = *bio_pos; 15543bfe5deSAlex Elder osd_data->bio_length = bio_length; 15643bfe5deSAlex Elder } 15743bfe5deSAlex Elder #endif /* CONFIG_BLOCK */ 15843bfe5deSAlex Elder 159b9e281c2SIlya Dryomov static void ceph_osd_data_bvecs_init(struct ceph_osd_data *osd_data, 1600010f705SIlya Dryomov struct ceph_bvec_iter *bvec_pos, 1610010f705SIlya Dryomov u32 num_bvecs) 162b9e281c2SIlya Dryomov { 163b9e281c2SIlya Dryomov osd_data->type = CEPH_OSD_DATA_TYPE_BVECS; 164b9e281c2SIlya Dryomov osd_data->bvec_pos = *bvec_pos; 1650010f705SIlya Dryomov osd_data->num_bvecs = num_bvecs; 166b9e281c2SIlya Dryomov } 167b9e281c2SIlya Dryomov 168863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 169863c7eb5SAlex Elder ({ \ 1708a703a38SIoana Ciornei struct ceph_osd_request *__oreq = (oreq); \ 1718a703a38SIoana Ciornei unsigned int __whch = (whch); \ 1728a703a38SIoana Ciornei BUG_ON(__whch >= __oreq->r_num_ops); \ 1738a703a38SIoana Ciornei &__oreq->r_ops[__whch].typ.fld; \ 174863c7eb5SAlex Elder }) 175863c7eb5SAlex Elder 17649719778SAlex Elder static struct ceph_osd_data * 17749719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 17849719778SAlex Elder { 17949719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 18049719778SAlex Elder 18149719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 18249719778SAlex Elder } 18349719778SAlex Elder 184a4ce40a9SAlex Elder struct ceph_osd_data * 185a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 186406e2c9fSAlex Elder unsigned int which) 187a4ce40a9SAlex Elder { 188863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 189a4ce40a9SAlex Elder } 190a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 191a4ce40a9SAlex Elder 19249719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 19349719778SAlex Elder unsigned int which, struct page **pages, 19449719778SAlex Elder u64 length, u32 alignment, 19549719778SAlex Elder bool pages_from_pool, bool own_pages) 19649719778SAlex Elder { 19749719778SAlex Elder struct ceph_osd_data *osd_data; 19849719778SAlex Elder 19949719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 20049719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 20149719778SAlex Elder pages_from_pool, own_pages); 20249719778SAlex Elder } 20349719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 20449719778SAlex Elder 205a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 206406e2c9fSAlex Elder unsigned int which, struct page **pages, 207406e2c9fSAlex Elder u64 length, u32 alignment, 208a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 209a4ce40a9SAlex Elder { 210a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 211a4ce40a9SAlex Elder 212863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 213a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 214a4ce40a9SAlex Elder pages_from_pool, own_pages); 215a4ce40a9SAlex Elder } 216a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 217a4ce40a9SAlex Elder 218a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 219406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 220a4ce40a9SAlex Elder { 221a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 222a4ce40a9SAlex Elder 223863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 224a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 225a4ce40a9SAlex Elder } 226a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 227a4ce40a9SAlex Elder 228a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 229a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 2305359a17dSIlya Dryomov unsigned int which, 2315359a17dSIlya Dryomov struct ceph_bio_iter *bio_pos, 2325359a17dSIlya Dryomov u32 bio_length) 233a4ce40a9SAlex Elder { 234a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 235863c7eb5SAlex Elder 236863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2375359a17dSIlya Dryomov ceph_osd_data_bio_init(osd_data, bio_pos, bio_length); 238a4ce40a9SAlex Elder } 239a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 240a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 241a4ce40a9SAlex Elder 2420010f705SIlya Dryomov void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request *osd_req, 2430010f705SIlya Dryomov unsigned int which, 2440010f705SIlya Dryomov struct bio_vec *bvecs, u32 num_bvecs, 2450010f705SIlya Dryomov u32 bytes) 2460010f705SIlya Dryomov { 2470010f705SIlya Dryomov struct ceph_osd_data *osd_data; 2480010f705SIlya Dryomov struct ceph_bvec_iter it = { 2490010f705SIlya Dryomov .bvecs = bvecs, 2500010f705SIlya Dryomov .iter = { .bi_size = bytes }, 2510010f705SIlya Dryomov }; 2520010f705SIlya Dryomov 2530010f705SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2540010f705SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, &it, num_bvecs); 2550010f705SIlya Dryomov } 2560010f705SIlya Dryomov EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvecs); 2570010f705SIlya Dryomov 258b9e281c2SIlya Dryomov void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req, 259b9e281c2SIlya Dryomov unsigned int which, 260b9e281c2SIlya Dryomov struct ceph_bvec_iter *bvec_pos) 261b9e281c2SIlya Dryomov { 262b9e281c2SIlya Dryomov struct ceph_osd_data *osd_data; 263b9e281c2SIlya Dryomov 264b9e281c2SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2650010f705SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, bvec_pos, 0); 266b9e281c2SIlya Dryomov } 267b9e281c2SIlya Dryomov EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvec_pos); 268b9e281c2SIlya Dryomov 269a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 270a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 271a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 272a4ce40a9SAlex Elder { 273a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 274a4ce40a9SAlex Elder 275863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 276a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 277a4ce40a9SAlex Elder } 278a4ce40a9SAlex Elder 27904017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 28004017e29SAlex Elder struct ceph_osd_request *osd_req, 28104017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 28204017e29SAlex Elder { 28304017e29SAlex Elder struct ceph_osd_data *osd_data; 28404017e29SAlex Elder 285863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 28604017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 287bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += pagelist->length; 288bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += pagelist->length; 28904017e29SAlex Elder } 29004017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 29104017e29SAlex Elder 2926c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2936c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2946c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2956c57b554SAlex Elder { 2966c57b554SAlex Elder struct ceph_osd_data *osd_data; 2976c57b554SAlex Elder 2986c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2996c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 3006c57b554SAlex Elder pages_from_pool, own_pages); 301bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += length; 302bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += length; 3036c57b554SAlex Elder } 3046c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 3056c57b554SAlex Elder 306b9e281c2SIlya Dryomov void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req, 307b9e281c2SIlya Dryomov unsigned int which, 3080010f705SIlya Dryomov struct bio_vec *bvecs, u32 num_bvecs, 3090010f705SIlya Dryomov u32 bytes) 310b9e281c2SIlya Dryomov { 311b9e281c2SIlya Dryomov struct ceph_osd_data *osd_data; 312b9e281c2SIlya Dryomov struct ceph_bvec_iter it = { 313b9e281c2SIlya Dryomov .bvecs = bvecs, 314b9e281c2SIlya Dryomov .iter = { .bi_size = bytes }, 315b9e281c2SIlya Dryomov }; 316b9e281c2SIlya Dryomov 317b9e281c2SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, cls, request_data); 3180010f705SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, &it, num_bvecs); 319b9e281c2SIlya Dryomov osd_req->r_ops[which].cls.indata_len += bytes; 320b9e281c2SIlya Dryomov osd_req->r_ops[which].indata_len += bytes; 321b9e281c2SIlya Dryomov } 322b9e281c2SIlya Dryomov EXPORT_SYMBOL(osd_req_op_cls_request_data_bvecs); 323b9e281c2SIlya Dryomov 324a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 325a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 326a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 327a4ce40a9SAlex Elder { 328a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 329a4ce40a9SAlex Elder 330863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 331a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 332a4ce40a9SAlex Elder pages_from_pool, own_pages); 333a4ce40a9SAlex Elder } 334a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 335a4ce40a9SAlex Elder 33623c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 33723c08a9cSAlex Elder { 33823c08a9cSAlex Elder switch (osd_data->type) { 33923c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 34023c08a9cSAlex Elder return 0; 34123c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 34223c08a9cSAlex Elder return osd_data->length; 34323c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 34423c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 34523c08a9cSAlex Elder #ifdef CONFIG_BLOCK 34623c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 34723c08a9cSAlex Elder return (u64)osd_data->bio_length; 34823c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 349b9e281c2SIlya Dryomov case CEPH_OSD_DATA_TYPE_BVECS: 350b9e281c2SIlya Dryomov return osd_data->bvec_pos.iter.bi_size; 35123c08a9cSAlex Elder default: 35223c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 35323c08a9cSAlex Elder return 0; 35423c08a9cSAlex Elder } 35523c08a9cSAlex Elder } 35623c08a9cSAlex Elder 357c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 358c54d47bfSAlex Elder { 3595476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 360c54d47bfSAlex Elder int num_pages; 361c54d47bfSAlex Elder 362c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 363c54d47bfSAlex Elder (u64)osd_data->length); 364c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 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; 4055476492fSAlex Elder default: 4065476492fSAlex Elder break; 4075476492fSAlex Elder } 408c54d47bfSAlex Elder } 409c54d47bfSAlex Elder 4103d14c5d2SYehuda Sadeh /* 41163244fa1SIlya Dryomov * Assumes @t is zero-initialized. 41263244fa1SIlya Dryomov */ 41363244fa1SIlya Dryomov static void target_init(struct ceph_osd_request_target *t) 41463244fa1SIlya Dryomov { 41563244fa1SIlya Dryomov ceph_oid_init(&t->base_oid); 41663244fa1SIlya Dryomov ceph_oloc_init(&t->base_oloc); 41763244fa1SIlya Dryomov ceph_oid_init(&t->target_oid); 41863244fa1SIlya Dryomov ceph_oloc_init(&t->target_oloc); 41963244fa1SIlya Dryomov 42063244fa1SIlya Dryomov ceph_osds_init(&t->acting); 42163244fa1SIlya Dryomov ceph_osds_init(&t->up); 42263244fa1SIlya Dryomov t->size = -1; 42363244fa1SIlya Dryomov t->min_size = -1; 42463244fa1SIlya Dryomov 42563244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 42663244fa1SIlya Dryomov } 42763244fa1SIlya Dryomov 428922dab61SIlya Dryomov static void target_copy(struct ceph_osd_request_target *dest, 429922dab61SIlya Dryomov const struct ceph_osd_request_target *src) 430922dab61SIlya Dryomov { 431922dab61SIlya Dryomov ceph_oid_copy(&dest->base_oid, &src->base_oid); 432922dab61SIlya Dryomov ceph_oloc_copy(&dest->base_oloc, &src->base_oloc); 433922dab61SIlya Dryomov ceph_oid_copy(&dest->target_oid, &src->target_oid); 434922dab61SIlya Dryomov ceph_oloc_copy(&dest->target_oloc, &src->target_oloc); 435922dab61SIlya Dryomov 436922dab61SIlya Dryomov dest->pgid = src->pgid; /* struct */ 437dc98ff72SIlya Dryomov dest->spgid = src->spgid; /* struct */ 438922dab61SIlya Dryomov dest->pg_num = src->pg_num; 439922dab61SIlya Dryomov dest->pg_num_mask = src->pg_num_mask; 440922dab61SIlya Dryomov ceph_osds_copy(&dest->acting, &src->acting); 441922dab61SIlya Dryomov ceph_osds_copy(&dest->up, &src->up); 442922dab61SIlya Dryomov dest->size = src->size; 443922dab61SIlya Dryomov dest->min_size = src->min_size; 444922dab61SIlya Dryomov dest->sort_bitwise = src->sort_bitwise; 445922dab61SIlya Dryomov 446922dab61SIlya Dryomov dest->flags = src->flags; 447922dab61SIlya Dryomov dest->paused = src->paused; 448922dab61SIlya Dryomov 44904c7d789SIlya Dryomov dest->epoch = src->epoch; 450dc93e0e2SIlya Dryomov dest->last_force_resend = src->last_force_resend; 451dc93e0e2SIlya Dryomov 452922dab61SIlya Dryomov dest->osd = src->osd; 453922dab61SIlya Dryomov } 454922dab61SIlya Dryomov 45563244fa1SIlya Dryomov static void target_destroy(struct ceph_osd_request_target *t) 45663244fa1SIlya Dryomov { 45763244fa1SIlya Dryomov ceph_oid_destroy(&t->base_oid); 45830c156d9SYan, Zheng ceph_oloc_destroy(&t->base_oloc); 45963244fa1SIlya Dryomov ceph_oid_destroy(&t->target_oid); 46030c156d9SYan, Zheng ceph_oloc_destroy(&t->target_oloc); 46163244fa1SIlya Dryomov } 46263244fa1SIlya Dryomov 46363244fa1SIlya Dryomov /* 4643d14c5d2SYehuda Sadeh * requests 4653d14c5d2SYehuda Sadeh */ 4663540bfdbSIlya Dryomov static void request_release_checks(struct ceph_osd_request *req) 4673540bfdbSIlya Dryomov { 4683540bfdbSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 4694609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_mc_node)); 4703540bfdbSIlya Dryomov WARN_ON(!list_empty(&req->r_unsafe_item)); 4713540bfdbSIlya Dryomov WARN_ON(req->r_osd); 4723540bfdbSIlya Dryomov } 4733540bfdbSIlya Dryomov 4749e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 4753d14c5d2SYehuda Sadeh { 4769e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 4779e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 4785476492fSAlex Elder unsigned int which; 4793d14c5d2SYehuda Sadeh 4809e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 4819e94af20SIlya Dryomov req->r_request, req->r_reply); 4823540bfdbSIlya Dryomov request_release_checks(req); 4839e94af20SIlya Dryomov 4843d14c5d2SYehuda Sadeh if (req->r_request) 4853d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 4865aea3dcdSIlya Dryomov if (req->r_reply) 487ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 4880fff87ecSAlex Elder 4895476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 4905476492fSAlex Elder osd_req_op_data_release(req, which); 4910fff87ecSAlex Elder 492a66dd383SIlya Dryomov target_destroy(&req->r_t); 4933d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 494d30291b9SIlya Dryomov 4953d14c5d2SYehuda Sadeh if (req->r_mempool) 4963d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 4973f1af42aSIlya Dryomov else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS) 4985522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 4993f1af42aSIlya Dryomov else 5003f1af42aSIlya Dryomov kfree(req); 5013d14c5d2SYehuda Sadeh } 5029e94af20SIlya Dryomov 5039e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 5049e94af20SIlya Dryomov { 5059e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 5062c935bc5SPeter Zijlstra kref_read(&req->r_kref)); 5079e94af20SIlya Dryomov kref_get(&req->r_kref); 5089e94af20SIlya Dryomov } 5099e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 5109e94af20SIlya Dryomov 5119e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 5129e94af20SIlya Dryomov { 5133ed97d63SIlya Dryomov if (req) { 5149e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 5152c935bc5SPeter Zijlstra kref_read(&req->r_kref)); 5169e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 5179e94af20SIlya Dryomov } 5183ed97d63SIlya Dryomov } 5199e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 5203d14c5d2SYehuda Sadeh 5213540bfdbSIlya Dryomov static void request_init(struct ceph_osd_request *req) 5223540bfdbSIlya Dryomov { 5233540bfdbSIlya Dryomov /* req only, each op is zeroed in _osd_req_op_init() */ 5243540bfdbSIlya Dryomov memset(req, 0, sizeof(*req)); 5253540bfdbSIlya Dryomov 5263540bfdbSIlya Dryomov kref_init(&req->r_kref); 5273540bfdbSIlya Dryomov init_completion(&req->r_completion); 5283540bfdbSIlya Dryomov RB_CLEAR_NODE(&req->r_node); 5294609245eSIlya Dryomov RB_CLEAR_NODE(&req->r_mc_node); 5303540bfdbSIlya Dryomov INIT_LIST_HEAD(&req->r_unsafe_item); 5313540bfdbSIlya Dryomov 5323540bfdbSIlya Dryomov target_init(&req->r_t); 5333540bfdbSIlya Dryomov } 5343540bfdbSIlya Dryomov 535922dab61SIlya Dryomov /* 536922dab61SIlya Dryomov * This is ugly, but it allows us to reuse linger registration and ping 537922dab61SIlya Dryomov * requests, keeping the structure of the code around send_linger{_ping}() 538922dab61SIlya Dryomov * reasonable. Setting up a min_nr=2 mempool for each linger request 539922dab61SIlya Dryomov * and dealing with copying ops (this blasts req only, watch op remains 540922dab61SIlya Dryomov * intact) isn't any better. 541922dab61SIlya Dryomov */ 542922dab61SIlya Dryomov static void request_reinit(struct ceph_osd_request *req) 543922dab61SIlya Dryomov { 544922dab61SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 545922dab61SIlya Dryomov bool mempool = req->r_mempool; 546922dab61SIlya Dryomov unsigned int num_ops = req->r_num_ops; 547922dab61SIlya Dryomov u64 snapid = req->r_snapid; 548922dab61SIlya Dryomov struct ceph_snap_context *snapc = req->r_snapc; 549922dab61SIlya Dryomov bool linger = req->r_linger; 550922dab61SIlya Dryomov struct ceph_msg *request_msg = req->r_request; 551922dab61SIlya Dryomov struct ceph_msg *reply_msg = req->r_reply; 552922dab61SIlya Dryomov 553922dab61SIlya Dryomov dout("%s req %p\n", __func__, req); 5542c935bc5SPeter Zijlstra WARN_ON(kref_read(&req->r_kref) != 1); 555922dab61SIlya Dryomov request_release_checks(req); 556922dab61SIlya Dryomov 5572c935bc5SPeter Zijlstra WARN_ON(kref_read(&request_msg->kref) != 1); 5582c935bc5SPeter Zijlstra WARN_ON(kref_read(&reply_msg->kref) != 1); 559922dab61SIlya Dryomov target_destroy(&req->r_t); 560922dab61SIlya Dryomov 561922dab61SIlya Dryomov request_init(req); 562922dab61SIlya Dryomov req->r_osdc = osdc; 563922dab61SIlya Dryomov req->r_mempool = mempool; 564922dab61SIlya Dryomov req->r_num_ops = num_ops; 565922dab61SIlya Dryomov req->r_snapid = snapid; 566922dab61SIlya Dryomov req->r_snapc = snapc; 567922dab61SIlya Dryomov req->r_linger = linger; 568922dab61SIlya Dryomov req->r_request = request_msg; 569922dab61SIlya Dryomov req->r_reply = reply_msg; 570922dab61SIlya Dryomov } 571922dab61SIlya Dryomov 5723d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 5733d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 5741b83bef2SSage Weil unsigned int num_ops, 5753d14c5d2SYehuda Sadeh bool use_mempool, 57654a54007SAlex Elder gfp_t gfp_flags) 5773d14c5d2SYehuda Sadeh { 5783d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5793d14c5d2SYehuda Sadeh 5803d14c5d2SYehuda Sadeh if (use_mempool) { 5813f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_SLAB_OPS); 5823d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 5833f1af42aSIlya Dryomov } else if (num_ops <= CEPH_OSD_SLAB_OPS) { 5843f1af42aSIlya Dryomov req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags); 5853d14c5d2SYehuda Sadeh } else { 5863f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_MAX_OPS); 5873f1af42aSIlya Dryomov req = kmalloc(sizeof(*req) + num_ops * sizeof(req->r_ops[0]), 5883f1af42aSIlya Dryomov gfp_flags); 5893d14c5d2SYehuda Sadeh } 5903f1af42aSIlya Dryomov if (unlikely(!req)) 5913d14c5d2SYehuda Sadeh return NULL; 5923d14c5d2SYehuda Sadeh 5933540bfdbSIlya Dryomov request_init(req); 5943d14c5d2SYehuda Sadeh req->r_osdc = osdc; 5953d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 59679528734SAlex Elder req->r_num_ops = num_ops; 59784127282SIlya Dryomov req->r_snapid = CEPH_NOSNAP; 59884127282SIlya Dryomov req->r_snapc = ceph_get_snap_context(snapc); 5993d14c5d2SYehuda Sadeh 60013d1ad16SIlya Dryomov dout("%s req %p\n", __func__, req); 60113d1ad16SIlya Dryomov return req; 6023f1af42aSIlya Dryomov } 60313d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_request); 6043f1af42aSIlya Dryomov 6052e59ffd1SIlya Dryomov static int ceph_oloc_encoding_size(const struct ceph_object_locator *oloc) 60630c156d9SYan, Zheng { 60730c156d9SYan, Zheng return 8 + 4 + 4 + 4 + (oloc->pool_ns ? oloc->pool_ns->len : 0); 60830c156d9SYan, Zheng } 60930c156d9SYan, Zheng 61013d1ad16SIlya Dryomov int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp) 61113d1ad16SIlya Dryomov { 61213d1ad16SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 61313d1ad16SIlya Dryomov struct ceph_msg *msg; 61413d1ad16SIlya Dryomov int msg_size; 6153d14c5d2SYehuda Sadeh 616d30291b9SIlya Dryomov WARN_ON(ceph_oid_empty(&req->r_base_oid)); 61730c156d9SYan, Zheng WARN_ON(ceph_oloc_empty(&req->r_base_oloc)); 618d30291b9SIlya Dryomov 61913d1ad16SIlya Dryomov /* create request message */ 6208cb441c0SIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 6218cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 6228cb441c0SIlya Dryomov msg_size += 4 + 4 + 4; /* hash, osdmap_epoch, flags */ 6238cb441c0SIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 6248cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid); /* reqid */ 6258cb441c0SIlya Dryomov msg_size += sizeof(struct ceph_blkin_trace_info); /* trace */ 6268cb441c0SIlya Dryomov msg_size += 4 + sizeof(struct ceph_timespec); /* client_inc, mtime */ 62730c156d9SYan, Zheng msg_size += CEPH_ENCODING_START_BLK_LEN + 62830c156d9SYan, Zheng ceph_oloc_encoding_size(&req->r_base_oloc); /* oloc */ 62913d1ad16SIlya Dryomov msg_size += 4 + req->r_base_oid.name_len; /* oid */ 63013d1ad16SIlya Dryomov msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op); 631ae458f5aSIlya Dryomov msg_size += 8; /* snapid */ 632ae458f5aSIlya Dryomov msg_size += 8; /* snap_seq */ 63313d1ad16SIlya Dryomov msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0); 6348cb441c0SIlya Dryomov msg_size += 4 + 8; /* retry_attempt, features */ 635ae458f5aSIlya Dryomov 63613d1ad16SIlya Dryomov if (req->r_mempool) 6373d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 6383d14c5d2SYehuda Sadeh else 63913d1ad16SIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp, true); 64013d1ad16SIlya Dryomov if (!msg) 64113d1ad16SIlya Dryomov return -ENOMEM; 6423d14c5d2SYehuda Sadeh 6433d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 6443d14c5d2SYehuda Sadeh req->r_request = msg; 6453d14c5d2SYehuda Sadeh 64613d1ad16SIlya Dryomov /* create reply message */ 64713d1ad16SIlya Dryomov msg_size = OSD_OPREPLY_FRONT_LEN; 648711da55dSIlya Dryomov msg_size += req->r_base_oid.name_len; 649711da55dSIlya Dryomov msg_size += req->r_num_ops * sizeof(struct ceph_osd_op); 65013d1ad16SIlya Dryomov 65113d1ad16SIlya Dryomov if (req->r_mempool) 65213d1ad16SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 65313d1ad16SIlya Dryomov else 65413d1ad16SIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size, gfp, true); 65513d1ad16SIlya Dryomov if (!msg) 65613d1ad16SIlya Dryomov return -ENOMEM; 65713d1ad16SIlya Dryomov 65813d1ad16SIlya Dryomov req->r_reply = msg; 65913d1ad16SIlya Dryomov 66013d1ad16SIlya Dryomov return 0; 66113d1ad16SIlya Dryomov } 66213d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_messages); 6633d14c5d2SYehuda Sadeh 664a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 665a8dd0a37SAlex Elder { 666a8dd0a37SAlex Elder switch (opcode) { 66770b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 66870b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 66970b5bfa3SIlya Dryomov #undef GENERATE_CASE 670a8dd0a37SAlex Elder default: 671a8dd0a37SAlex Elder return false; 672a8dd0a37SAlex Elder } 673a8dd0a37SAlex Elder } 674a8dd0a37SAlex Elder 67533803f33SAlex Elder /* 67633803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 67733803f33SAlex Elder * other information associated with them. It also serves as a 67833803f33SAlex Elder * common init routine for all the other init functions, below. 67933803f33SAlex Elder */ 680c99d2d4aSAlex Elder static struct ceph_osd_req_op * 68149719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 682144cba14SYan, Zheng u16 opcode, u32 flags) 68333803f33SAlex Elder { 684c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 685c99d2d4aSAlex Elder 686c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 68733803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 68833803f33SAlex Elder 689c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 69033803f33SAlex Elder memset(op, 0, sizeof (*op)); 69133803f33SAlex Elder op->op = opcode; 692144cba14SYan, Zheng op->flags = flags; 693c99d2d4aSAlex Elder 694c99d2d4aSAlex Elder return op; 69533803f33SAlex Elder } 69633803f33SAlex Elder 69749719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 698144cba14SYan, Zheng unsigned int which, u16 opcode, u32 flags) 69949719778SAlex Elder { 700144cba14SYan, Zheng (void)_osd_req_op_init(osd_req, which, opcode, flags); 70149719778SAlex Elder } 70249719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 70349719778SAlex Elder 704c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 705c99d2d4aSAlex Elder unsigned int which, u16 opcode, 70633803f33SAlex Elder u64 offset, u64 length, 70733803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 70833803f33SAlex Elder { 709144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 710144cba14SYan, Zheng opcode, 0); 71133803f33SAlex Elder size_t payload_len = 0; 71233803f33SAlex Elder 713ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 714e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 715e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_TRUNCATE); 71633803f33SAlex Elder 71733803f33SAlex Elder op->extent.offset = offset; 71833803f33SAlex Elder op->extent.length = length; 71933803f33SAlex Elder op->extent.truncate_size = truncate_size; 72033803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 721e30b7577SIlya Dryomov if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 72233803f33SAlex Elder payload_len += length; 72333803f33SAlex Elder 724de2aa102SIlya Dryomov op->indata_len = payload_len; 72533803f33SAlex Elder } 72633803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 72733803f33SAlex Elder 728c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 729c99d2d4aSAlex Elder unsigned int which, u64 length) 730e5975c7cSAlex Elder { 731c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 732c99d2d4aSAlex Elder u64 previous; 733c99d2d4aSAlex Elder 734c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 735c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 736c99d2d4aSAlex Elder previous = op->extent.length; 737e5975c7cSAlex Elder 738e5975c7cSAlex Elder if (length == previous) 739e5975c7cSAlex Elder return; /* Nothing to do */ 740e5975c7cSAlex Elder BUG_ON(length > previous); 741e5975c7cSAlex Elder 742e5975c7cSAlex Elder op->extent.length = length; 743d641df81SYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 744de2aa102SIlya Dryomov op->indata_len -= previous - length; 745e5975c7cSAlex Elder } 746e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 747e5975c7cSAlex Elder 7482c63f49aSYan, Zheng void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 7492c63f49aSYan, Zheng unsigned int which, u64 offset_inc) 7502c63f49aSYan, Zheng { 7512c63f49aSYan, Zheng struct ceph_osd_req_op *op, *prev_op; 7522c63f49aSYan, Zheng 7532c63f49aSYan, Zheng BUG_ON(which + 1 >= osd_req->r_num_ops); 7542c63f49aSYan, Zheng 7552c63f49aSYan, Zheng prev_op = &osd_req->r_ops[which]; 7562c63f49aSYan, Zheng op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags); 7572c63f49aSYan, Zheng /* dup previous one */ 7582c63f49aSYan, Zheng op->indata_len = prev_op->indata_len; 7592c63f49aSYan, Zheng op->outdata_len = prev_op->outdata_len; 7602c63f49aSYan, Zheng op->extent = prev_op->extent; 7612c63f49aSYan, Zheng /* adjust offset */ 7622c63f49aSYan, Zheng op->extent.offset += offset_inc; 7632c63f49aSYan, Zheng op->extent.length -= offset_inc; 7642c63f49aSYan, Zheng 7652c63f49aSYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 7662c63f49aSYan, Zheng op->indata_len -= offset_inc; 7672c63f49aSYan, Zheng } 7682c63f49aSYan, Zheng EXPORT_SYMBOL(osd_req_op_extent_dup_last); 7692c63f49aSYan, Zheng 770fe943d50SChengguang Xu int osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 77104017e29SAlex Elder u16 opcode, const char *class, const char *method) 77233803f33SAlex Elder { 773144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 774144cba14SYan, Zheng opcode, 0); 7755f562df5SAlex Elder struct ceph_pagelist *pagelist; 77633803f33SAlex Elder size_t payload_len = 0; 77733803f33SAlex Elder size_t size; 77833803f33SAlex Elder 77933803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 78033803f33SAlex Elder 7815f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 782fe943d50SChengguang Xu if (!pagelist) 783fe943d50SChengguang Xu return -ENOMEM; 784fe943d50SChengguang Xu 7855f562df5SAlex Elder ceph_pagelist_init(pagelist); 7865f562df5SAlex Elder 78733803f33SAlex Elder op->cls.class_name = class; 78833803f33SAlex Elder size = strlen(class); 78933803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 79033803f33SAlex Elder op->cls.class_len = size; 7915f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 79233803f33SAlex Elder payload_len += size; 79333803f33SAlex Elder 79433803f33SAlex Elder op->cls.method_name = method; 79533803f33SAlex Elder size = strlen(method); 79633803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 79733803f33SAlex Elder op->cls.method_len = size; 7985f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 79933803f33SAlex Elder payload_len += size; 80033803f33SAlex Elder 801a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 8025f562df5SAlex Elder 803de2aa102SIlya Dryomov op->indata_len = payload_len; 804fe943d50SChengguang Xu return 0; 80533803f33SAlex Elder } 80633803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 8078c042b0dSAlex Elder 808d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 809d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 810d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 811d74b50beSYan, Zheng { 812144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 813144cba14SYan, Zheng opcode, 0); 814d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 815d74b50beSYan, Zheng size_t payload_len; 816d74b50beSYan, Zheng 817d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 818d74b50beSYan, Zheng 819d74b50beSYan, Zheng pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); 820d74b50beSYan, Zheng if (!pagelist) 821d74b50beSYan, Zheng return -ENOMEM; 822d74b50beSYan, Zheng 823d74b50beSYan, Zheng ceph_pagelist_init(pagelist); 824d74b50beSYan, Zheng 825d74b50beSYan, Zheng payload_len = strlen(name); 826d74b50beSYan, Zheng op->xattr.name_len = payload_len; 827d74b50beSYan, Zheng ceph_pagelist_append(pagelist, name, payload_len); 828d74b50beSYan, Zheng 829d74b50beSYan, Zheng op->xattr.value_len = size; 830d74b50beSYan, Zheng ceph_pagelist_append(pagelist, value, size); 831d74b50beSYan, Zheng payload_len += size; 832d74b50beSYan, Zheng 833d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 834d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 835d74b50beSYan, Zheng 836d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 837de2aa102SIlya Dryomov op->indata_len = payload_len; 838d74b50beSYan, Zheng return 0; 839d74b50beSYan, Zheng } 840d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 841d74b50beSYan, Zheng 842922dab61SIlya Dryomov /* 843922dab61SIlya Dryomov * @watch_opcode: CEPH_OSD_WATCH_OP_* 844922dab61SIlya Dryomov */ 845922dab61SIlya Dryomov static void osd_req_op_watch_init(struct ceph_osd_request *req, int which, 846922dab61SIlya Dryomov u64 cookie, u8 watch_opcode) 84733803f33SAlex Elder { 848922dab61SIlya Dryomov struct ceph_osd_req_op *op; 84933803f33SAlex Elder 850922dab61SIlya Dryomov op = _osd_req_op_init(req, which, CEPH_OSD_OP_WATCH, 0); 85133803f33SAlex Elder op->watch.cookie = cookie; 852922dab61SIlya Dryomov op->watch.op = watch_opcode; 853922dab61SIlya Dryomov op->watch.gen = 0; 85433803f33SAlex Elder } 85533803f33SAlex Elder 856c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 857c647b8a8SIlya Dryomov unsigned int which, 858c647b8a8SIlya Dryomov u64 expected_object_size, 859c647b8a8SIlya Dryomov u64 expected_write_size) 860c647b8a8SIlya Dryomov { 861c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 862144cba14SYan, Zheng CEPH_OSD_OP_SETALLOCHINT, 863144cba14SYan, Zheng 0); 864c647b8a8SIlya Dryomov 865c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 866c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 867c647b8a8SIlya Dryomov 868c647b8a8SIlya Dryomov /* 869c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 870c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 871c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 872c647b8a8SIlya Dryomov */ 873c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 874c647b8a8SIlya Dryomov } 875c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 876c647b8a8SIlya Dryomov 87790af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 878ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 879ec9123c5SAlex Elder { 880ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 881ec9123c5SAlex Elder 882ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 883ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 884ec9123c5SAlex Elder if (length) 88590af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 886ec9123c5SAlex Elder length, osd_data->alignment); 887ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 888ec9123c5SAlex Elder BUG_ON(!length); 88990af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 890ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 891ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 8925359a17dSIlya Dryomov ceph_msg_data_add_bio(msg, &osd_data->bio_pos, length); 893ec9123c5SAlex Elder #endif 894b9e281c2SIlya Dryomov } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BVECS) { 895b9e281c2SIlya Dryomov ceph_msg_data_add_bvecs(msg, &osd_data->bvec_pos); 896ec9123c5SAlex Elder } else { 897ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 898ec9123c5SAlex Elder } 899ec9123c5SAlex Elder } 900ec9123c5SAlex Elder 901bb873b53SIlya Dryomov static u32 osd_req_encode_op(struct ceph_osd_op *dst, 902bb873b53SIlya Dryomov const struct ceph_osd_req_op *src) 9033d14c5d2SYehuda Sadeh { 904a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 905a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 906a8dd0a37SAlex Elder 907a8dd0a37SAlex Elder return 0; 908a8dd0a37SAlex Elder } 9093d14c5d2SYehuda Sadeh 910065a68f9SAlex Elder switch (src->op) { 911fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 912fbfab539SAlex Elder break; 9133d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 9143d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 915e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 916ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 917ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 918175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 919175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 9203d14c5d2SYehuda Sadeh dst->extent.truncate_size = 9213d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 9223d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 9233d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 9243d14c5d2SYehuda Sadeh break; 9253d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 9263d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 9273d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 928bb873b53SIlya Dryomov dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 9293d14c5d2SYehuda Sadeh break; 930a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 931a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 932922dab61SIlya Dryomov dst->watch.ver = cpu_to_le64(0); 933922dab61SIlya Dryomov dst->watch.op = src->watch.op; 934922dab61SIlya Dryomov dst->watch.gen = cpu_to_le32(src->watch.gen); 935922dab61SIlya Dryomov break; 936922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 937a40c4f10SYehuda Sadeh break; 93819079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 93919079203SIlya Dryomov dst->notify.cookie = cpu_to_le64(src->notify.cookie); 94019079203SIlya Dryomov break; 941a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 942a4ed38d7SDouglas Fuller break; 943c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 944c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 945c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 946c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 947c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 948c647b8a8SIlya Dryomov break; 949d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 950d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 951d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 952d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 953d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 954d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 955d74b50beSYan, Zheng break; 956864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 957864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 958864e9197SYan, Zheng break; 9593d14c5d2SYehuda Sadeh default: 9604c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 9618f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 9624c46459cSAlex Elder WARN_ON(1); 963a8dd0a37SAlex Elder 964a8dd0a37SAlex Elder return 0; 9653d14c5d2SYehuda Sadeh } 9667b25bf5fSIlya Dryomov 967a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 9687b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 969de2aa102SIlya Dryomov dst->payload_len = cpu_to_le32(src->indata_len); 970175face2SAlex Elder 971bb873b53SIlya Dryomov return src->indata_len; 9723d14c5d2SYehuda Sadeh } 9733d14c5d2SYehuda Sadeh 9743d14c5d2SYehuda Sadeh /* 9753d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 9763d14c5d2SYehuda Sadeh * extent as needed. 9773d14c5d2SYehuda Sadeh * 9783d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 9793d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 9803d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 9813d14c5d2SYehuda Sadeh */ 9823d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 9833d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 9843d14c5d2SYehuda Sadeh struct ceph_vino vino, 985715e4cd4SYan, Zheng u64 off, u64 *plen, 986715e4cd4SYan, Zheng unsigned int which, int num_ops, 9873d14c5d2SYehuda Sadeh int opcode, int flags, 9883d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 9893d14c5d2SYehuda Sadeh u32 truncate_seq, 9903d14c5d2SYehuda Sadeh u64 truncate_size, 991153e5167SAlex Elder bool use_mempool) 9923d14c5d2SYehuda Sadeh { 9933d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 99475d1c941SAlex Elder u64 objnum = 0; 99575d1c941SAlex Elder u64 objoff = 0; 99675d1c941SAlex Elder u64 objlen = 0; 9976816282dSSage Weil int r; 9983d14c5d2SYehuda Sadeh 999ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 1000864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 1001864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 10023d14c5d2SYehuda Sadeh 1003acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 1004ae7ca4a3SAlex Elder GFP_NOFS); 100513d1ad16SIlya Dryomov if (!req) { 100613d1ad16SIlya Dryomov r = -ENOMEM; 100713d1ad16SIlya Dryomov goto fail; 100813d1ad16SIlya Dryomov } 100979528734SAlex Elder 10103d14c5d2SYehuda Sadeh /* calculate max write size */ 1011a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 101213d1ad16SIlya Dryomov if (r) 101313d1ad16SIlya Dryomov goto fail; 1014a19dadfbSAlex Elder 1015864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 1016144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 1017864e9197SYan, Zheng } else { 10187627151eSYan, Zheng u32 object_size = layout->object_size; 1019864e9197SYan, Zheng u32 object_base = off - objoff; 1020ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 1021d18d1e28SAlex Elder if (truncate_size <= object_base) { 1022d18d1e28SAlex Elder truncate_size = 0; 1023d18d1e28SAlex Elder } else { 1024d18d1e28SAlex Elder truncate_size -= object_base; 1025d18d1e28SAlex Elder if (truncate_size > object_size) 1026d18d1e28SAlex Elder truncate_size = object_size; 1027d18d1e28SAlex Elder } 1028ccca4e37SYan, Zheng } 1029715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 1030b0270324SAlex Elder truncate_size, truncate_seq); 1031864e9197SYan, Zheng } 1032d18d1e28SAlex Elder 1033a1f4020aSJeff Layton req->r_abort_on_full = true; 1034bb873b53SIlya Dryomov req->r_flags = flags; 10357627151eSYan, Zheng req->r_base_oloc.pool = layout->pool_id; 103630c156d9SYan, Zheng req->r_base_oloc.pool_ns = ceph_try_get_string(layout->pool_ns); 1037d30291b9SIlya Dryomov ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum); 1038dbe0fc41SAlex Elder 1039bb873b53SIlya Dryomov req->r_snapid = vino.snap; 1040bb873b53SIlya Dryomov if (flags & CEPH_OSD_FLAG_WRITE) 1041bb873b53SIlya Dryomov req->r_data_offset = off; 1042bb873b53SIlya Dryomov 104313d1ad16SIlya Dryomov r = ceph_osdc_alloc_messages(req, GFP_NOFS); 104413d1ad16SIlya Dryomov if (r) 104513d1ad16SIlya Dryomov goto fail; 104613d1ad16SIlya Dryomov 10473d14c5d2SYehuda Sadeh return req; 104813d1ad16SIlya Dryomov 104913d1ad16SIlya Dryomov fail: 105013d1ad16SIlya Dryomov ceph_osdc_put_request(req); 105113d1ad16SIlya Dryomov return ERR_PTR(r); 10523d14c5d2SYehuda Sadeh } 10533d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 10543d14c5d2SYehuda Sadeh 10553d14c5d2SYehuda Sadeh /* 10563d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 10573d14c5d2SYehuda Sadeh */ 1058fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node) 10594609245eSIlya Dryomov DEFINE_RB_FUNCS(request_mc, struct ceph_osd_request, r_tid, r_mc_node) 10603d14c5d2SYehuda Sadeh 106166850df5SIlya Dryomov /* 106266850df5SIlya Dryomov * Call @fn on each OSD request as long as @fn returns 0. 106366850df5SIlya Dryomov */ 106466850df5SIlya Dryomov static void for_each_request(struct ceph_osd_client *osdc, 106566850df5SIlya Dryomov int (*fn)(struct ceph_osd_request *req, void *arg), 106666850df5SIlya Dryomov void *arg) 106766850df5SIlya Dryomov { 106866850df5SIlya Dryomov struct rb_node *n, *p; 106966850df5SIlya Dryomov 107066850df5SIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 107166850df5SIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 107266850df5SIlya Dryomov 107366850df5SIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 107466850df5SIlya Dryomov struct ceph_osd_request *req = 107566850df5SIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 107666850df5SIlya Dryomov 107766850df5SIlya Dryomov p = rb_next(p); 107866850df5SIlya Dryomov if (fn(req, arg)) 107966850df5SIlya Dryomov return; 108066850df5SIlya Dryomov } 108166850df5SIlya Dryomov } 108266850df5SIlya Dryomov 108366850df5SIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 108466850df5SIlya Dryomov struct ceph_osd_request *req = 108566850df5SIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 108666850df5SIlya Dryomov 108766850df5SIlya Dryomov p = rb_next(p); 108866850df5SIlya Dryomov if (fn(req, arg)) 108966850df5SIlya Dryomov return; 109066850df5SIlya Dryomov } 109166850df5SIlya Dryomov } 109266850df5SIlya Dryomov 10930247a0cfSIlya Dryomov static bool osd_homeless(struct ceph_osd *osd) 10940247a0cfSIlya Dryomov { 10950247a0cfSIlya Dryomov return osd->o_osd == CEPH_HOMELESS_OSD; 10960247a0cfSIlya Dryomov } 10970247a0cfSIlya Dryomov 10985aea3dcdSIlya Dryomov static bool osd_registered(struct ceph_osd *osd) 10993d14c5d2SYehuda Sadeh { 11005aea3dcdSIlya Dryomov verify_osdc_locked(osd->o_osdc); 11013d14c5d2SYehuda Sadeh 11025aea3dcdSIlya Dryomov return !RB_EMPTY_NODE(&osd->o_node); 11033d14c5d2SYehuda Sadeh } 11043d14c5d2SYehuda Sadeh 11053d14c5d2SYehuda Sadeh /* 11060247a0cfSIlya Dryomov * Assumes @osd is zero-initialized. 11070247a0cfSIlya Dryomov */ 11080247a0cfSIlya Dryomov static void osd_init(struct ceph_osd *osd) 11090247a0cfSIlya Dryomov { 111002113a0fSElena Reshetova refcount_set(&osd->o_ref, 1); 11110247a0cfSIlya Dryomov RB_CLEAR_NODE(&osd->o_node); 11125aea3dcdSIlya Dryomov osd->o_requests = RB_ROOT; 1113922dab61SIlya Dryomov osd->o_linger_requests = RB_ROOT; 1114a02a946dSIlya Dryomov osd->o_backoff_mappings = RB_ROOT; 1115a02a946dSIlya Dryomov osd->o_backoffs_by_id = RB_ROOT; 11160247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_osd_lru); 11170247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_keepalive_item); 11180247a0cfSIlya Dryomov osd->o_incarnation = 1; 11195aea3dcdSIlya Dryomov mutex_init(&osd->lock); 11200247a0cfSIlya Dryomov } 11210247a0cfSIlya Dryomov 11220247a0cfSIlya Dryomov static void osd_cleanup(struct ceph_osd *osd) 11230247a0cfSIlya Dryomov { 11240247a0cfSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&osd->o_node)); 11255aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 1126922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 1127a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoff_mappings)); 1128a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoffs_by_id)); 11290247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_osd_lru)); 11300247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_keepalive_item)); 11310247a0cfSIlya Dryomov 11320247a0cfSIlya Dryomov if (osd->o_auth.authorizer) { 11330247a0cfSIlya Dryomov WARN_ON(osd_homeless(osd)); 11340247a0cfSIlya Dryomov ceph_auth_destroy_authorizer(osd->o_auth.authorizer); 11350247a0cfSIlya Dryomov } 11360247a0cfSIlya Dryomov } 11370247a0cfSIlya Dryomov 11380247a0cfSIlya Dryomov /* 11393d14c5d2SYehuda Sadeh * Track open sessions with osds. 11403d14c5d2SYehuda Sadeh */ 1141e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 11423d14c5d2SYehuda Sadeh { 11433d14c5d2SYehuda Sadeh struct ceph_osd *osd; 11443d14c5d2SYehuda Sadeh 11450247a0cfSIlya Dryomov WARN_ON(onum == CEPH_HOMELESS_OSD); 11460247a0cfSIlya Dryomov 11477a28f59bSIlya Dryomov osd = kzalloc(sizeof(*osd), GFP_NOIO | __GFP_NOFAIL); 11480247a0cfSIlya Dryomov osd_init(osd); 11493d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1150e10006f8SAlex Elder osd->o_osd = onum; 11513d14c5d2SYehuda Sadeh 1152b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 11533d14c5d2SYehuda Sadeh 11543d14c5d2SYehuda Sadeh return osd; 11553d14c5d2SYehuda Sadeh } 11563d14c5d2SYehuda Sadeh 11573d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 11583d14c5d2SYehuda Sadeh { 115902113a0fSElena Reshetova if (refcount_inc_not_zero(&osd->o_ref)) { 116002113a0fSElena Reshetova dout("get_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref)-1, 116102113a0fSElena Reshetova refcount_read(&osd->o_ref)); 11623d14c5d2SYehuda Sadeh return osd; 11633d14c5d2SYehuda Sadeh } else { 11643d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 11653d14c5d2SYehuda Sadeh return NULL; 11663d14c5d2SYehuda Sadeh } 11673d14c5d2SYehuda Sadeh } 11683d14c5d2SYehuda Sadeh 11693d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 11703d14c5d2SYehuda Sadeh { 117102113a0fSElena Reshetova dout("put_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref), 117202113a0fSElena Reshetova refcount_read(&osd->o_ref) - 1); 117302113a0fSElena Reshetova if (refcount_dec_and_test(&osd->o_ref)) { 11740247a0cfSIlya Dryomov osd_cleanup(osd); 11753d14c5d2SYehuda Sadeh kfree(osd); 11763d14c5d2SYehuda Sadeh } 11773d14c5d2SYehuda Sadeh } 11783d14c5d2SYehuda Sadeh 1179fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node) 1180fcd00b68SIlya Dryomov 11819dd2845cSIlya Dryomov static void __move_osd_to_lru(struct ceph_osd *osd) 11823d14c5d2SYehuda Sadeh { 11839dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 11849dd2845cSIlya Dryomov 11859dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 11863d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1187bbf37ec3SIlya Dryomov 11889dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 11893d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 11909dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 11919dd2845cSIlya Dryomov 1192a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 11933d14c5d2SYehuda Sadeh } 11943d14c5d2SYehuda Sadeh 11959dd2845cSIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd *osd) 1196bbf37ec3SIlya Dryomov { 11975aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1198922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) 11999dd2845cSIlya Dryomov __move_osd_to_lru(osd); 1200bbf37ec3SIlya Dryomov } 1201bbf37ec3SIlya Dryomov 12023d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 12033d14c5d2SYehuda Sadeh { 12049dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 12059dd2845cSIlya Dryomov 12069dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12079dd2845cSIlya Dryomov 12089dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 12093d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 12103d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 12119dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 12123d14c5d2SYehuda Sadeh } 12133d14c5d2SYehuda Sadeh 12143d14c5d2SYehuda Sadeh /* 12155aea3dcdSIlya Dryomov * Close the connection and assign any leftover requests to the 12165aea3dcdSIlya Dryomov * homeless session. 12175aea3dcdSIlya Dryomov */ 12185aea3dcdSIlya Dryomov static void close_osd(struct ceph_osd *osd) 12195aea3dcdSIlya Dryomov { 12205aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 12215aea3dcdSIlya Dryomov struct rb_node *n; 12225aea3dcdSIlya Dryomov 12235aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 12245aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12255aea3dcdSIlya Dryomov 12265aea3dcdSIlya Dryomov ceph_con_close(&osd->o_con); 12275aea3dcdSIlya Dryomov 12285aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 12295aea3dcdSIlya Dryomov struct ceph_osd_request *req = 12305aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 12315aea3dcdSIlya Dryomov 12325aea3dcdSIlya Dryomov n = rb_next(n); /* unlink_request() */ 12335aea3dcdSIlya Dryomov 12345aea3dcdSIlya Dryomov dout(" reassigning req %p tid %llu\n", req, req->r_tid); 12355aea3dcdSIlya Dryomov unlink_request(osd, req); 12365aea3dcdSIlya Dryomov link_request(&osdc->homeless_osd, req); 12375aea3dcdSIlya Dryomov } 1238922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 1239922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 1240922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 1241922dab61SIlya Dryomov 1242922dab61SIlya Dryomov n = rb_next(n); /* unlink_linger() */ 1243922dab61SIlya Dryomov 1244922dab61SIlya Dryomov dout(" reassigning lreq %p linger_id %llu\n", lreq, 1245922dab61SIlya Dryomov lreq->linger_id); 1246922dab61SIlya Dryomov unlink_linger(osd, lreq); 1247922dab61SIlya Dryomov link_linger(&osdc->homeless_osd, lreq); 1248922dab61SIlya Dryomov } 1249a02a946dSIlya Dryomov clear_backoffs(osd); 12505aea3dcdSIlya Dryomov 12515aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 12525aea3dcdSIlya Dryomov erase_osd(&osdc->osds, osd); 12535aea3dcdSIlya Dryomov put_osd(osd); 12545aea3dcdSIlya Dryomov } 12555aea3dcdSIlya Dryomov 12565aea3dcdSIlya Dryomov /* 12573d14c5d2SYehuda Sadeh * reset osd connect 12583d14c5d2SYehuda Sadeh */ 12595aea3dcdSIlya Dryomov static int reopen_osd(struct ceph_osd *osd) 12603d14c5d2SYehuda Sadeh { 1261c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 12623d14c5d2SYehuda Sadeh 12635aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12645aea3dcdSIlya Dryomov 12655aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1266922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) { 12675aea3dcdSIlya Dryomov close_osd(osd); 1268c3acb181SAlex Elder return -ENODEV; 1269c3acb181SAlex Elder } 1270c3acb181SAlex Elder 12715aea3dcdSIlya Dryomov peer_addr = &osd->o_osdc->osdmap->osd_addr[osd->o_osd]; 1272c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 12733d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 12745aea3dcdSIlya Dryomov struct rb_node *n; 1275c3acb181SAlex Elder 12763d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 12770b4af2e8SIlya Dryomov "letting msgr retry\n"); 12783d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 12795aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 12805aea3dcdSIlya Dryomov struct ceph_osd_request *req = 12815aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 12823d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 12835aea3dcdSIlya Dryomov } 1284c3acb181SAlex Elder 1285c3acb181SAlex Elder return -EAGAIN; 12863d14c5d2SYehuda Sadeh } 1287c3acb181SAlex Elder 1288c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1289c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1290c3acb181SAlex Elder osd->o_incarnation++; 1291c3acb181SAlex Elder 1292c3acb181SAlex Elder return 0; 12933d14c5d2SYehuda Sadeh } 12943d14c5d2SYehuda Sadeh 12955aea3dcdSIlya Dryomov static struct ceph_osd *lookup_create_osd(struct ceph_osd_client *osdc, int o, 12965aea3dcdSIlya Dryomov bool wrlocked) 12973d14c5d2SYehuda Sadeh { 12985aea3dcdSIlya Dryomov struct ceph_osd *osd; 12995aea3dcdSIlya Dryomov 13005aea3dcdSIlya Dryomov if (wrlocked) 13015aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 13025aea3dcdSIlya Dryomov else 13035aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 13045aea3dcdSIlya Dryomov 13055aea3dcdSIlya Dryomov if (o != CEPH_HOMELESS_OSD) 13065aea3dcdSIlya Dryomov osd = lookup_osd(&osdc->osds, o); 13075aea3dcdSIlya Dryomov else 13085aea3dcdSIlya Dryomov osd = &osdc->homeless_osd; 13095aea3dcdSIlya Dryomov if (!osd) { 13105aea3dcdSIlya Dryomov if (!wrlocked) 13115aea3dcdSIlya Dryomov return ERR_PTR(-EAGAIN); 13125aea3dcdSIlya Dryomov 13135aea3dcdSIlya Dryomov osd = create_osd(osdc, o); 13145aea3dcdSIlya Dryomov insert_osd(&osdc->osds, osd); 13155aea3dcdSIlya Dryomov ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, 13165aea3dcdSIlya Dryomov &osdc->osdmap->osd_addr[osd->o_osd]); 13175aea3dcdSIlya Dryomov } 13185aea3dcdSIlya Dryomov 13195aea3dcdSIlya Dryomov dout("%s osdc %p osd%d -> osd %p\n", __func__, osdc, o, osd); 13205aea3dcdSIlya Dryomov return osd; 1321a40c4f10SYehuda Sadeh } 1322a40c4f10SYehuda Sadeh 13233d14c5d2SYehuda Sadeh /* 13245aea3dcdSIlya Dryomov * Create request <-> OSD session relation. 13255aea3dcdSIlya Dryomov * 13265aea3dcdSIlya Dryomov * @req has to be assigned a tid, @osd may be homeless. 13273d14c5d2SYehuda Sadeh */ 13285aea3dcdSIlya Dryomov static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req) 13293d14c5d2SYehuda Sadeh { 13305aea3dcdSIlya Dryomov verify_osd_locked(osd); 13315aea3dcdSIlya Dryomov WARN_ON(!req->r_tid || req->r_osd); 13325aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 133335f9f8a0SSage Weil req, req->r_tid); 13345aea3dcdSIlya Dryomov 13355aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 13365aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 13375aea3dcdSIlya Dryomov else 13385aea3dcdSIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 13395aea3dcdSIlya Dryomov 13405aea3dcdSIlya Dryomov get_osd(osd); 13415aea3dcdSIlya Dryomov insert_request(&osd->o_requests, req); 13425aea3dcdSIlya Dryomov req->r_osd = osd; 134335f9f8a0SSage Weil } 134435f9f8a0SSage Weil 13455aea3dcdSIlya Dryomov static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req) 13463d14c5d2SYehuda Sadeh { 13475aea3dcdSIlya Dryomov verify_osd_locked(osd); 13485aea3dcdSIlya Dryomov WARN_ON(req->r_osd != osd); 13495aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 13505aea3dcdSIlya Dryomov req, req->r_tid); 13515aea3dcdSIlya Dryomov 13525aea3dcdSIlya Dryomov req->r_osd = NULL; 13535aea3dcdSIlya Dryomov erase_request(&osd->o_requests, req); 13545aea3dcdSIlya Dryomov put_osd(osd); 13555aea3dcdSIlya Dryomov 13565aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 13575aea3dcdSIlya Dryomov maybe_move_osd_to_lru(osd); 13585aea3dcdSIlya Dryomov else 13595aea3dcdSIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 13603d14c5d2SYehuda Sadeh } 13613d14c5d2SYehuda Sadeh 136263244fa1SIlya Dryomov static bool __pool_full(struct ceph_pg_pool_info *pi) 136363244fa1SIlya Dryomov { 136463244fa1SIlya Dryomov return pi->flags & CEPH_POOL_FLAG_FULL; 136563244fa1SIlya Dryomov } 136663244fa1SIlya Dryomov 136742c1b124SIlya Dryomov static bool have_pool_full(struct ceph_osd_client *osdc) 136842c1b124SIlya Dryomov { 136942c1b124SIlya Dryomov struct rb_node *n; 137042c1b124SIlya Dryomov 137142c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 137242c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 137342c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 137442c1b124SIlya Dryomov 137542c1b124SIlya Dryomov if (__pool_full(pi)) 137642c1b124SIlya Dryomov return true; 137742c1b124SIlya Dryomov } 137842c1b124SIlya Dryomov 137942c1b124SIlya Dryomov return false; 138042c1b124SIlya Dryomov } 138142c1b124SIlya Dryomov 13825aea3dcdSIlya Dryomov static bool pool_full(struct ceph_osd_client *osdc, s64 pool_id) 13835aea3dcdSIlya Dryomov { 13845aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 13855aea3dcdSIlya Dryomov 13865aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 13875aea3dcdSIlya Dryomov if (!pi) 13885aea3dcdSIlya Dryomov return false; 13895aea3dcdSIlya Dryomov 13905aea3dcdSIlya Dryomov return __pool_full(pi); 13915aea3dcdSIlya Dryomov } 13925aea3dcdSIlya Dryomov 13933d14c5d2SYehuda Sadeh /* 1394d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1395d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1396d29adb34SJosh Durgin */ 139763244fa1SIlya Dryomov static bool target_should_be_paused(struct ceph_osd_client *osdc, 139863244fa1SIlya Dryomov const struct ceph_osd_request_target *t, 139963244fa1SIlya Dryomov struct ceph_pg_pool_info *pi) 140063244fa1SIlya Dryomov { 1401b7ec35b3SIlya Dryomov bool pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 1402b7ec35b3SIlya Dryomov bool pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 1403b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 140463244fa1SIlya Dryomov __pool_full(pi); 140563244fa1SIlya Dryomov 14066d637a54SIlya Dryomov WARN_ON(pi->id != t->target_oloc.pool); 140758eb7932SJeff Layton return ((t->flags & CEPH_OSD_FLAG_READ) && pauserd) || 140858eb7932SJeff Layton ((t->flags & CEPH_OSD_FLAG_WRITE) && pausewr) || 140958eb7932SJeff Layton (osdc->osdmap->epoch < osdc->epoch_barrier); 141063244fa1SIlya Dryomov } 141163244fa1SIlya Dryomov 141263244fa1SIlya Dryomov enum calc_target_result { 141363244fa1SIlya Dryomov CALC_TARGET_NO_ACTION = 0, 141463244fa1SIlya Dryomov CALC_TARGET_NEED_RESEND, 141563244fa1SIlya Dryomov CALC_TARGET_POOL_DNE, 141663244fa1SIlya Dryomov }; 141763244fa1SIlya Dryomov 141863244fa1SIlya Dryomov static enum calc_target_result calc_target(struct ceph_osd_client *osdc, 141963244fa1SIlya Dryomov struct ceph_osd_request_target *t, 14207de030d6SIlya Dryomov struct ceph_connection *con, 142163244fa1SIlya Dryomov bool any_change) 142263244fa1SIlya Dryomov { 142363244fa1SIlya Dryomov struct ceph_pg_pool_info *pi; 142463244fa1SIlya Dryomov struct ceph_pg pgid, last_pgid; 142563244fa1SIlya Dryomov struct ceph_osds up, acting; 142663244fa1SIlya Dryomov bool force_resend = false; 142784ed45dfSIlya Dryomov bool unpaused = false; 142884ed45dfSIlya Dryomov bool legacy_change; 14297de030d6SIlya Dryomov bool split = false; 1430b7ec35b3SIlya Dryomov bool sort_bitwise = ceph_osdmap_flag(osdc, CEPH_OSDMAP_SORTBITWISE); 1431ae78dd81SIlya Dryomov bool recovery_deletes = ceph_osdmap_flag(osdc, 1432ae78dd81SIlya Dryomov CEPH_OSDMAP_RECOVERY_DELETES); 143363244fa1SIlya Dryomov enum calc_target_result ct_res; 143463244fa1SIlya Dryomov int ret; 143563244fa1SIlya Dryomov 143604c7d789SIlya Dryomov t->epoch = osdc->osdmap->epoch; 143763244fa1SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool); 143863244fa1SIlya Dryomov if (!pi) { 143963244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 144063244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 144163244fa1SIlya Dryomov goto out; 144263244fa1SIlya Dryomov } 144363244fa1SIlya Dryomov 144463244fa1SIlya Dryomov if (osdc->osdmap->epoch == pi->last_force_request_resend) { 1445dc93e0e2SIlya Dryomov if (t->last_force_resend < pi->last_force_request_resend) { 1446dc93e0e2SIlya Dryomov t->last_force_resend = pi->last_force_request_resend; 144763244fa1SIlya Dryomov force_resend = true; 1448dc93e0e2SIlya Dryomov } else if (t->last_force_resend == 0) { 144963244fa1SIlya Dryomov force_resend = true; 145063244fa1SIlya Dryomov } 145163244fa1SIlya Dryomov } 145263244fa1SIlya Dryomov 1453db098ec4SIlya Dryomov /* apply tiering */ 1454db098ec4SIlya Dryomov ceph_oid_copy(&t->target_oid, &t->base_oid); 1455db098ec4SIlya Dryomov ceph_oloc_copy(&t->target_oloc, &t->base_oloc); 1456db098ec4SIlya Dryomov if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 145763244fa1SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_READ && pi->read_tier >= 0) 145863244fa1SIlya Dryomov t->target_oloc.pool = pi->read_tier; 145963244fa1SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_WRITE && pi->write_tier >= 0) 146063244fa1SIlya Dryomov t->target_oloc.pool = pi->write_tier; 14616d637a54SIlya Dryomov 14626d637a54SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool); 14636d637a54SIlya Dryomov if (!pi) { 14646d637a54SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 14656d637a54SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 14666d637a54SIlya Dryomov goto out; 14676d637a54SIlya Dryomov } 146863244fa1SIlya Dryomov } 146963244fa1SIlya Dryomov 1470df28152dSIlya Dryomov ret = __ceph_object_locator_to_pg(pi, &t->target_oid, &t->target_oloc, 1471df28152dSIlya Dryomov &pgid); 147263244fa1SIlya Dryomov if (ret) { 147363244fa1SIlya Dryomov WARN_ON(ret != -ENOENT); 147463244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 147563244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 147663244fa1SIlya Dryomov goto out; 147763244fa1SIlya Dryomov } 147863244fa1SIlya Dryomov last_pgid.pool = pgid.pool; 147963244fa1SIlya Dryomov last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask); 148063244fa1SIlya Dryomov 1481df28152dSIlya Dryomov ceph_pg_to_up_acting_osds(osdc->osdmap, pi, &pgid, &up, &acting); 148263244fa1SIlya Dryomov if (any_change && 148363244fa1SIlya Dryomov ceph_is_new_interval(&t->acting, 148463244fa1SIlya Dryomov &acting, 148563244fa1SIlya Dryomov &t->up, 148663244fa1SIlya Dryomov &up, 148763244fa1SIlya Dryomov t->size, 148863244fa1SIlya Dryomov pi->size, 148963244fa1SIlya Dryomov t->min_size, 149063244fa1SIlya Dryomov pi->min_size, 149163244fa1SIlya Dryomov t->pg_num, 149263244fa1SIlya Dryomov pi->pg_num, 149363244fa1SIlya Dryomov t->sort_bitwise, 149463244fa1SIlya Dryomov sort_bitwise, 1495ae78dd81SIlya Dryomov t->recovery_deletes, 1496ae78dd81SIlya Dryomov recovery_deletes, 149763244fa1SIlya Dryomov &last_pgid)) 149863244fa1SIlya Dryomov force_resend = true; 149963244fa1SIlya Dryomov 150063244fa1SIlya Dryomov if (t->paused && !target_should_be_paused(osdc, t, pi)) { 150163244fa1SIlya Dryomov t->paused = false; 150284ed45dfSIlya Dryomov unpaused = true; 150363244fa1SIlya Dryomov } 150484ed45dfSIlya Dryomov legacy_change = ceph_pg_compare(&t->pgid, &pgid) || 150584ed45dfSIlya Dryomov ceph_osds_changed(&t->acting, &acting, any_change); 15067de030d6SIlya Dryomov if (t->pg_num) 15077de030d6SIlya Dryomov split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num); 150863244fa1SIlya Dryomov 15097de030d6SIlya Dryomov if (legacy_change || force_resend || split) { 151063244fa1SIlya Dryomov t->pgid = pgid; /* struct */ 1511df28152dSIlya Dryomov ceph_pg_to_primary_shard(osdc->osdmap, pi, &pgid, &t->spgid); 151263244fa1SIlya Dryomov ceph_osds_copy(&t->acting, &acting); 151363244fa1SIlya Dryomov ceph_osds_copy(&t->up, &up); 151463244fa1SIlya Dryomov t->size = pi->size; 151563244fa1SIlya Dryomov t->min_size = pi->min_size; 151663244fa1SIlya Dryomov t->pg_num = pi->pg_num; 151763244fa1SIlya Dryomov t->pg_num_mask = pi->pg_num_mask; 151863244fa1SIlya Dryomov t->sort_bitwise = sort_bitwise; 1519ae78dd81SIlya Dryomov t->recovery_deletes = recovery_deletes; 152063244fa1SIlya Dryomov 152163244fa1SIlya Dryomov t->osd = acting.primary; 152263244fa1SIlya Dryomov } 152363244fa1SIlya Dryomov 15247de030d6SIlya Dryomov if (unpaused || legacy_change || force_resend || 15257de030d6SIlya Dryomov (split && con && CEPH_HAVE_FEATURE(con->peer_features, 15267de030d6SIlya Dryomov RESEND_ON_SPLIT))) 152784ed45dfSIlya Dryomov ct_res = CALC_TARGET_NEED_RESEND; 152884ed45dfSIlya Dryomov else 152984ed45dfSIlya Dryomov ct_res = CALC_TARGET_NO_ACTION; 153084ed45dfSIlya Dryomov 153163244fa1SIlya Dryomov out: 153263244fa1SIlya Dryomov dout("%s t %p -> ct_res %d osd %d\n", __func__, t, ct_res, t->osd); 153363244fa1SIlya Dryomov return ct_res; 153463244fa1SIlya Dryomov } 153563244fa1SIlya Dryomov 1536a02a946dSIlya Dryomov static struct ceph_spg_mapping *alloc_spg_mapping(void) 1537a02a946dSIlya Dryomov { 1538a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1539a02a946dSIlya Dryomov 1540a02a946dSIlya Dryomov spg = kmalloc(sizeof(*spg), GFP_NOIO); 1541a02a946dSIlya Dryomov if (!spg) 1542a02a946dSIlya Dryomov return NULL; 1543a02a946dSIlya Dryomov 1544a02a946dSIlya Dryomov RB_CLEAR_NODE(&spg->node); 1545a02a946dSIlya Dryomov spg->backoffs = RB_ROOT; 1546a02a946dSIlya Dryomov return spg; 1547a02a946dSIlya Dryomov } 1548a02a946dSIlya Dryomov 1549a02a946dSIlya Dryomov static void free_spg_mapping(struct ceph_spg_mapping *spg) 1550a02a946dSIlya Dryomov { 1551a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&spg->node)); 1552a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&spg->backoffs)); 1553a02a946dSIlya Dryomov 1554a02a946dSIlya Dryomov kfree(spg); 1555a02a946dSIlya Dryomov } 1556a02a946dSIlya Dryomov 1557a02a946dSIlya Dryomov /* 1558a02a946dSIlya Dryomov * rbtree of ceph_spg_mapping for handling map<spg_t, ...>, similar to 1559a02a946dSIlya Dryomov * ceph_pg_mapping. Used to track OSD backoffs -- a backoff [range] is 1560a02a946dSIlya Dryomov * defined only within a specific spgid; it does not pass anything to 1561a02a946dSIlya Dryomov * children on split, or to another primary. 1562a02a946dSIlya Dryomov */ 1563a02a946dSIlya Dryomov DEFINE_RB_FUNCS2(spg_mapping, struct ceph_spg_mapping, spgid, ceph_spg_compare, 1564a02a946dSIlya Dryomov RB_BYPTR, const struct ceph_spg *, node) 1565a02a946dSIlya Dryomov 1566a02a946dSIlya Dryomov static u64 hoid_get_bitwise_key(const struct ceph_hobject_id *hoid) 1567a02a946dSIlya Dryomov { 1568a02a946dSIlya Dryomov return hoid->is_max ? 0x100000000ull : hoid->hash_reverse_bits; 1569a02a946dSIlya Dryomov } 1570a02a946dSIlya Dryomov 1571a02a946dSIlya Dryomov static void hoid_get_effective_key(const struct ceph_hobject_id *hoid, 1572a02a946dSIlya Dryomov void **pkey, size_t *pkey_len) 1573a02a946dSIlya Dryomov { 1574a02a946dSIlya Dryomov if (hoid->key_len) { 1575a02a946dSIlya Dryomov *pkey = hoid->key; 1576a02a946dSIlya Dryomov *pkey_len = hoid->key_len; 1577a02a946dSIlya Dryomov } else { 1578a02a946dSIlya Dryomov *pkey = hoid->oid; 1579a02a946dSIlya Dryomov *pkey_len = hoid->oid_len; 1580a02a946dSIlya Dryomov } 1581a02a946dSIlya Dryomov } 1582a02a946dSIlya Dryomov 1583a02a946dSIlya Dryomov static int compare_names(const void *name1, size_t name1_len, 1584a02a946dSIlya Dryomov const void *name2, size_t name2_len) 1585a02a946dSIlya Dryomov { 1586a02a946dSIlya Dryomov int ret; 1587a02a946dSIlya Dryomov 1588a02a946dSIlya Dryomov ret = memcmp(name1, name2, min(name1_len, name2_len)); 1589a02a946dSIlya Dryomov if (!ret) { 1590a02a946dSIlya Dryomov if (name1_len < name2_len) 1591a02a946dSIlya Dryomov ret = -1; 1592a02a946dSIlya Dryomov else if (name1_len > name2_len) 1593a02a946dSIlya Dryomov ret = 1; 1594a02a946dSIlya Dryomov } 1595a02a946dSIlya Dryomov return ret; 1596a02a946dSIlya Dryomov } 1597a02a946dSIlya Dryomov 1598a02a946dSIlya Dryomov static int hoid_compare(const struct ceph_hobject_id *lhs, 1599a02a946dSIlya Dryomov const struct ceph_hobject_id *rhs) 1600a02a946dSIlya Dryomov { 1601a02a946dSIlya Dryomov void *effective_key1, *effective_key2; 1602a02a946dSIlya Dryomov size_t effective_key1_len, effective_key2_len; 1603a02a946dSIlya Dryomov int ret; 1604a02a946dSIlya Dryomov 1605a02a946dSIlya Dryomov if (lhs->is_max < rhs->is_max) 1606a02a946dSIlya Dryomov return -1; 1607a02a946dSIlya Dryomov if (lhs->is_max > rhs->is_max) 1608a02a946dSIlya Dryomov return 1; 1609a02a946dSIlya Dryomov 1610a02a946dSIlya Dryomov if (lhs->pool < rhs->pool) 1611a02a946dSIlya Dryomov return -1; 1612a02a946dSIlya Dryomov if (lhs->pool > rhs->pool) 1613a02a946dSIlya Dryomov return 1; 1614a02a946dSIlya Dryomov 1615a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) < hoid_get_bitwise_key(rhs)) 1616a02a946dSIlya Dryomov return -1; 1617a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) > hoid_get_bitwise_key(rhs)) 1618a02a946dSIlya Dryomov return 1; 1619a02a946dSIlya Dryomov 1620a02a946dSIlya Dryomov ret = compare_names(lhs->nspace, lhs->nspace_len, 1621a02a946dSIlya Dryomov rhs->nspace, rhs->nspace_len); 1622a02a946dSIlya Dryomov if (ret) 1623a02a946dSIlya Dryomov return ret; 1624a02a946dSIlya Dryomov 1625a02a946dSIlya Dryomov hoid_get_effective_key(lhs, &effective_key1, &effective_key1_len); 1626a02a946dSIlya Dryomov hoid_get_effective_key(rhs, &effective_key2, &effective_key2_len); 1627a02a946dSIlya Dryomov ret = compare_names(effective_key1, effective_key1_len, 1628a02a946dSIlya Dryomov effective_key2, effective_key2_len); 1629a02a946dSIlya Dryomov if (ret) 1630a02a946dSIlya Dryomov return ret; 1631a02a946dSIlya Dryomov 1632a02a946dSIlya Dryomov ret = compare_names(lhs->oid, lhs->oid_len, rhs->oid, rhs->oid_len); 1633a02a946dSIlya Dryomov if (ret) 1634a02a946dSIlya Dryomov return ret; 1635a02a946dSIlya Dryomov 1636a02a946dSIlya Dryomov if (lhs->snapid < rhs->snapid) 1637a02a946dSIlya Dryomov return -1; 1638a02a946dSIlya Dryomov if (lhs->snapid > rhs->snapid) 1639a02a946dSIlya Dryomov return 1; 1640a02a946dSIlya Dryomov 1641a02a946dSIlya Dryomov return 0; 1642a02a946dSIlya Dryomov } 1643a02a946dSIlya Dryomov 1644a02a946dSIlya Dryomov /* 1645a02a946dSIlya Dryomov * For decoding ->begin and ->end of MOSDBackoff only -- no MIN/MAX 1646a02a946dSIlya Dryomov * compat stuff here. 1647a02a946dSIlya Dryomov * 1648a02a946dSIlya Dryomov * Assumes @hoid is zero-initialized. 1649a02a946dSIlya Dryomov */ 1650a02a946dSIlya Dryomov static int decode_hoid(void **p, void *end, struct ceph_hobject_id *hoid) 1651a02a946dSIlya Dryomov { 1652a02a946dSIlya Dryomov u8 struct_v; 1653a02a946dSIlya Dryomov u32 struct_len; 1654a02a946dSIlya Dryomov int ret; 1655a02a946dSIlya Dryomov 1656a02a946dSIlya Dryomov ret = ceph_start_decoding(p, end, 4, "hobject_t", &struct_v, 1657a02a946dSIlya Dryomov &struct_len); 1658a02a946dSIlya Dryomov if (ret) 1659a02a946dSIlya Dryomov return ret; 1660a02a946dSIlya Dryomov 1661a02a946dSIlya Dryomov if (struct_v < 4) { 1662a02a946dSIlya Dryomov pr_err("got struct_v %d < 4 of hobject_t\n", struct_v); 1663a02a946dSIlya Dryomov goto e_inval; 1664a02a946dSIlya Dryomov } 1665a02a946dSIlya Dryomov 1666a02a946dSIlya Dryomov hoid->key = ceph_extract_encoded_string(p, end, &hoid->key_len, 1667a02a946dSIlya Dryomov GFP_NOIO); 1668a02a946dSIlya Dryomov if (IS_ERR(hoid->key)) { 1669a02a946dSIlya Dryomov ret = PTR_ERR(hoid->key); 1670a02a946dSIlya Dryomov hoid->key = NULL; 1671a02a946dSIlya Dryomov return ret; 1672a02a946dSIlya Dryomov } 1673a02a946dSIlya Dryomov 1674a02a946dSIlya Dryomov hoid->oid = ceph_extract_encoded_string(p, end, &hoid->oid_len, 1675a02a946dSIlya Dryomov GFP_NOIO); 1676a02a946dSIlya Dryomov if (IS_ERR(hoid->oid)) { 1677a02a946dSIlya Dryomov ret = PTR_ERR(hoid->oid); 1678a02a946dSIlya Dryomov hoid->oid = NULL; 1679a02a946dSIlya Dryomov return ret; 1680a02a946dSIlya Dryomov } 1681a02a946dSIlya Dryomov 1682a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->snapid, e_inval); 1683a02a946dSIlya Dryomov ceph_decode_32_safe(p, end, hoid->hash, e_inval); 1684a02a946dSIlya Dryomov ceph_decode_8_safe(p, end, hoid->is_max, e_inval); 1685a02a946dSIlya Dryomov 1686a02a946dSIlya Dryomov hoid->nspace = ceph_extract_encoded_string(p, end, &hoid->nspace_len, 1687a02a946dSIlya Dryomov GFP_NOIO); 1688a02a946dSIlya Dryomov if (IS_ERR(hoid->nspace)) { 1689a02a946dSIlya Dryomov ret = PTR_ERR(hoid->nspace); 1690a02a946dSIlya Dryomov hoid->nspace = NULL; 1691a02a946dSIlya Dryomov return ret; 1692a02a946dSIlya Dryomov } 1693a02a946dSIlya Dryomov 1694a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->pool, e_inval); 1695a02a946dSIlya Dryomov 1696a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1697a02a946dSIlya Dryomov return 0; 1698a02a946dSIlya Dryomov 1699a02a946dSIlya Dryomov e_inval: 1700a02a946dSIlya Dryomov return -EINVAL; 1701a02a946dSIlya Dryomov } 1702a02a946dSIlya Dryomov 1703a02a946dSIlya Dryomov static int hoid_encoding_size(const struct ceph_hobject_id *hoid) 1704a02a946dSIlya Dryomov { 1705a02a946dSIlya Dryomov return 8 + 4 + 1 + 8 + /* snapid, hash, is_max, pool */ 1706a02a946dSIlya Dryomov 4 + hoid->key_len + 4 + hoid->oid_len + 4 + hoid->nspace_len; 1707a02a946dSIlya Dryomov } 1708a02a946dSIlya Dryomov 1709a02a946dSIlya Dryomov static void encode_hoid(void **p, void *end, const struct ceph_hobject_id *hoid) 1710a02a946dSIlya Dryomov { 1711a02a946dSIlya Dryomov ceph_start_encoding(p, 4, 3, hoid_encoding_size(hoid)); 1712a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->key, hoid->key_len); 1713a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->oid, hoid->oid_len); 1714a02a946dSIlya Dryomov ceph_encode_64(p, hoid->snapid); 1715a02a946dSIlya Dryomov ceph_encode_32(p, hoid->hash); 1716a02a946dSIlya Dryomov ceph_encode_8(p, hoid->is_max); 1717a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->nspace, hoid->nspace_len); 1718a02a946dSIlya Dryomov ceph_encode_64(p, hoid->pool); 1719a02a946dSIlya Dryomov } 1720a02a946dSIlya Dryomov 1721a02a946dSIlya Dryomov static void free_hoid(struct ceph_hobject_id *hoid) 1722a02a946dSIlya Dryomov { 1723a02a946dSIlya Dryomov if (hoid) { 1724a02a946dSIlya Dryomov kfree(hoid->key); 1725a02a946dSIlya Dryomov kfree(hoid->oid); 1726a02a946dSIlya Dryomov kfree(hoid->nspace); 1727a02a946dSIlya Dryomov kfree(hoid); 1728a02a946dSIlya Dryomov } 1729a02a946dSIlya Dryomov } 1730a02a946dSIlya Dryomov 1731a02a946dSIlya Dryomov static struct ceph_osd_backoff *alloc_backoff(void) 1732a02a946dSIlya Dryomov { 1733a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1734a02a946dSIlya Dryomov 1735a02a946dSIlya Dryomov backoff = kzalloc(sizeof(*backoff), GFP_NOIO); 1736a02a946dSIlya Dryomov if (!backoff) 1737a02a946dSIlya Dryomov return NULL; 1738a02a946dSIlya Dryomov 1739a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->spg_node); 1740a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->id_node); 1741a02a946dSIlya Dryomov return backoff; 1742a02a946dSIlya Dryomov } 1743a02a946dSIlya Dryomov 1744a02a946dSIlya Dryomov static void free_backoff(struct ceph_osd_backoff *backoff) 1745a02a946dSIlya Dryomov { 1746a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->spg_node)); 1747a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->id_node)); 1748a02a946dSIlya Dryomov 1749a02a946dSIlya Dryomov free_hoid(backoff->begin); 1750a02a946dSIlya Dryomov free_hoid(backoff->end); 1751a02a946dSIlya Dryomov kfree(backoff); 1752a02a946dSIlya Dryomov } 1753a02a946dSIlya Dryomov 1754a02a946dSIlya Dryomov /* 1755a02a946dSIlya Dryomov * Within a specific spgid, backoffs are managed by ->begin hoid. 1756a02a946dSIlya Dryomov */ 1757a02a946dSIlya Dryomov DEFINE_RB_INSDEL_FUNCS2(backoff, struct ceph_osd_backoff, begin, hoid_compare, 1758a02a946dSIlya Dryomov RB_BYVAL, spg_node); 1759a02a946dSIlya Dryomov 1760a02a946dSIlya Dryomov static struct ceph_osd_backoff *lookup_containing_backoff(struct rb_root *root, 1761a02a946dSIlya Dryomov const struct ceph_hobject_id *hoid) 1762a02a946dSIlya Dryomov { 1763a02a946dSIlya Dryomov struct rb_node *n = root->rb_node; 1764a02a946dSIlya Dryomov 1765a02a946dSIlya Dryomov while (n) { 1766a02a946dSIlya Dryomov struct ceph_osd_backoff *cur = 1767a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_backoff, spg_node); 1768a02a946dSIlya Dryomov int cmp; 1769a02a946dSIlya Dryomov 1770a02a946dSIlya Dryomov cmp = hoid_compare(hoid, cur->begin); 1771a02a946dSIlya Dryomov if (cmp < 0) { 1772a02a946dSIlya Dryomov n = n->rb_left; 1773a02a946dSIlya Dryomov } else if (cmp > 0) { 1774a02a946dSIlya Dryomov if (hoid_compare(hoid, cur->end) < 0) 1775a02a946dSIlya Dryomov return cur; 1776a02a946dSIlya Dryomov 1777a02a946dSIlya Dryomov n = n->rb_right; 1778a02a946dSIlya Dryomov } else { 1779a02a946dSIlya Dryomov return cur; 1780a02a946dSIlya Dryomov } 1781a02a946dSIlya Dryomov } 1782a02a946dSIlya Dryomov 1783a02a946dSIlya Dryomov return NULL; 1784a02a946dSIlya Dryomov } 1785a02a946dSIlya Dryomov 1786a02a946dSIlya Dryomov /* 1787a02a946dSIlya Dryomov * Each backoff has a unique id within its OSD session. 1788a02a946dSIlya Dryomov */ 1789a02a946dSIlya Dryomov DEFINE_RB_FUNCS(backoff_by_id, struct ceph_osd_backoff, id, id_node) 1790a02a946dSIlya Dryomov 1791a02a946dSIlya Dryomov static void clear_backoffs(struct ceph_osd *osd) 1792a02a946dSIlya Dryomov { 1793a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&osd->o_backoff_mappings)) { 1794a02a946dSIlya Dryomov struct ceph_spg_mapping *spg = 1795a02a946dSIlya Dryomov rb_entry(rb_first(&osd->o_backoff_mappings), 1796a02a946dSIlya Dryomov struct ceph_spg_mapping, node); 1797a02a946dSIlya Dryomov 1798a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&spg->backoffs)) { 1799a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff = 1800a02a946dSIlya Dryomov rb_entry(rb_first(&spg->backoffs), 1801a02a946dSIlya Dryomov struct ceph_osd_backoff, spg_node); 1802a02a946dSIlya Dryomov 1803a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 1804a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 1805a02a946dSIlya Dryomov free_backoff(backoff); 1806a02a946dSIlya Dryomov } 1807a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 1808a02a946dSIlya Dryomov free_spg_mapping(spg); 1809a02a946dSIlya Dryomov } 1810a02a946dSIlya Dryomov } 1811a02a946dSIlya Dryomov 1812a02a946dSIlya Dryomov /* 1813a02a946dSIlya Dryomov * Set up a temporary, non-owning view into @t. 1814a02a946dSIlya Dryomov */ 1815a02a946dSIlya Dryomov static void hoid_fill_from_target(struct ceph_hobject_id *hoid, 1816a02a946dSIlya Dryomov const struct ceph_osd_request_target *t) 1817a02a946dSIlya Dryomov { 1818a02a946dSIlya Dryomov hoid->key = NULL; 1819a02a946dSIlya Dryomov hoid->key_len = 0; 1820a02a946dSIlya Dryomov hoid->oid = t->target_oid.name; 1821a02a946dSIlya Dryomov hoid->oid_len = t->target_oid.name_len; 1822a02a946dSIlya Dryomov hoid->snapid = CEPH_NOSNAP; 1823a02a946dSIlya Dryomov hoid->hash = t->pgid.seed; 1824a02a946dSIlya Dryomov hoid->is_max = false; 1825a02a946dSIlya Dryomov if (t->target_oloc.pool_ns) { 1826a02a946dSIlya Dryomov hoid->nspace = t->target_oloc.pool_ns->str; 1827a02a946dSIlya Dryomov hoid->nspace_len = t->target_oloc.pool_ns->len; 1828a02a946dSIlya Dryomov } else { 1829a02a946dSIlya Dryomov hoid->nspace = NULL; 1830a02a946dSIlya Dryomov hoid->nspace_len = 0; 1831a02a946dSIlya Dryomov } 1832a02a946dSIlya Dryomov hoid->pool = t->target_oloc.pool; 1833a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1834a02a946dSIlya Dryomov } 1835a02a946dSIlya Dryomov 1836a02a946dSIlya Dryomov static bool should_plug_request(struct ceph_osd_request *req) 1837a02a946dSIlya Dryomov { 1838a02a946dSIlya Dryomov struct ceph_osd *osd = req->r_osd; 1839a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1840a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1841a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 1842a02a946dSIlya Dryomov 1843a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &req->r_t.spgid); 1844a02a946dSIlya Dryomov if (!spg) 1845a02a946dSIlya Dryomov return false; 1846a02a946dSIlya Dryomov 1847a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, &req->r_t); 1848a02a946dSIlya Dryomov backoff = lookup_containing_backoff(&spg->backoffs, &hoid); 1849a02a946dSIlya Dryomov if (!backoff) 1850a02a946dSIlya Dryomov return false; 1851a02a946dSIlya Dryomov 1852a02a946dSIlya Dryomov dout("%s req %p tid %llu backoff osd%d spgid %llu.%xs%d id %llu\n", 1853a02a946dSIlya Dryomov __func__, req, req->r_tid, osd->o_osd, backoff->spgid.pgid.pool, 1854a02a946dSIlya Dryomov backoff->spgid.pgid.seed, backoff->spgid.shard, backoff->id); 1855a02a946dSIlya Dryomov return true; 1856a02a946dSIlya Dryomov } 1857a02a946dSIlya Dryomov 1858bb873b53SIlya Dryomov static void setup_request_data(struct ceph_osd_request *req, 1859bb873b53SIlya Dryomov struct ceph_msg *msg) 18603d14c5d2SYehuda Sadeh { 1861bb873b53SIlya Dryomov u32 data_len = 0; 1862bb873b53SIlya Dryomov int i; 18633d14c5d2SYehuda Sadeh 1864bb873b53SIlya Dryomov if (!list_empty(&msg->data)) 1865bb873b53SIlya Dryomov return; 18663d14c5d2SYehuda Sadeh 1867bb873b53SIlya Dryomov WARN_ON(msg->data_length); 1868bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 1869bb873b53SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[i]; 1870bb873b53SIlya Dryomov 1871bb873b53SIlya Dryomov switch (op->op) { 1872bb873b53SIlya Dryomov /* request */ 1873bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITE: 1874bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 1875bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->extent.length); 1876bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->extent.osd_data); 1877bb873b53SIlya Dryomov break; 1878bb873b53SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 1879bb873b53SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 1880bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->xattr.name_len + 1881bb873b53SIlya Dryomov op->xattr.value_len); 1882bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->xattr.osd_data); 1883bb873b53SIlya Dryomov break; 1884922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 1885922dab61SIlya Dryomov ceph_osdc_msg_data_add(msg, 1886922dab61SIlya Dryomov &op->notify_ack.request_data); 1887922dab61SIlya Dryomov break; 1888bb873b53SIlya Dryomov 1889bb873b53SIlya Dryomov /* reply */ 1890bb873b53SIlya Dryomov case CEPH_OSD_OP_STAT: 1891bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1892bb873b53SIlya Dryomov &op->raw_data_in); 1893bb873b53SIlya Dryomov break; 1894bb873b53SIlya Dryomov case CEPH_OSD_OP_READ: 1895bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1896bb873b53SIlya Dryomov &op->extent.osd_data); 1897bb873b53SIlya Dryomov break; 1898a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 1899a4ed38d7SDouglas Fuller ceph_osdc_msg_data_add(req->r_reply, 1900a4ed38d7SDouglas Fuller &op->list_watchers.response_data); 1901a4ed38d7SDouglas Fuller break; 1902bb873b53SIlya Dryomov 1903bb873b53SIlya Dryomov /* both */ 1904bb873b53SIlya Dryomov case CEPH_OSD_OP_CALL: 1905bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->cls.class_len + 1906bb873b53SIlya Dryomov op->cls.method_len + 1907bb873b53SIlya Dryomov op->cls.indata_len); 1908bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->cls.request_info); 1909bb873b53SIlya Dryomov /* optional, can be NONE */ 1910bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->cls.request_data); 1911bb873b53SIlya Dryomov /* optional, can be NONE */ 1912bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1913bb873b53SIlya Dryomov &op->cls.response_data); 1914bb873b53SIlya Dryomov break; 191519079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 191619079203SIlya Dryomov ceph_osdc_msg_data_add(msg, 191719079203SIlya Dryomov &op->notify.request_data); 191819079203SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 191919079203SIlya Dryomov &op->notify.response_data); 192019079203SIlya Dryomov break; 1921bb873b53SIlya Dryomov } 1922bb873b53SIlya Dryomov 1923bb873b53SIlya Dryomov data_len += op->indata_len; 1924bb873b53SIlya Dryomov } 1925bb873b53SIlya Dryomov 1926bb873b53SIlya Dryomov WARN_ON(data_len != msg->data_length); 1927bb873b53SIlya Dryomov } 1928bb873b53SIlya Dryomov 19292e59ffd1SIlya Dryomov static void encode_pgid(void **p, const struct ceph_pg *pgid) 19302e59ffd1SIlya Dryomov { 19312e59ffd1SIlya Dryomov ceph_encode_8(p, 1); 19322e59ffd1SIlya Dryomov ceph_encode_64(p, pgid->pool); 19332e59ffd1SIlya Dryomov ceph_encode_32(p, pgid->seed); 19342e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 19352e59ffd1SIlya Dryomov } 19362e59ffd1SIlya Dryomov 19378cb441c0SIlya Dryomov static void encode_spgid(void **p, const struct ceph_spg *spgid) 19388cb441c0SIlya Dryomov { 19398cb441c0SIlya Dryomov ceph_start_encoding(p, 1, 1, CEPH_PGID_ENCODING_LEN + 1); 19408cb441c0SIlya Dryomov encode_pgid(p, &spgid->pgid); 19418cb441c0SIlya Dryomov ceph_encode_8(p, spgid->shard); 19428cb441c0SIlya Dryomov } 19438cb441c0SIlya Dryomov 19442e59ffd1SIlya Dryomov static void encode_oloc(void **p, void *end, 19452e59ffd1SIlya Dryomov const struct ceph_object_locator *oloc) 19462e59ffd1SIlya Dryomov { 19472e59ffd1SIlya Dryomov ceph_start_encoding(p, 5, 4, ceph_oloc_encoding_size(oloc)); 19482e59ffd1SIlya Dryomov ceph_encode_64(p, oloc->pool); 19492e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 19502e59ffd1SIlya Dryomov ceph_encode_32(p, 0); /* key len */ 19512e59ffd1SIlya Dryomov if (oloc->pool_ns) 19522e59ffd1SIlya Dryomov ceph_encode_string(p, end, oloc->pool_ns->str, 19532e59ffd1SIlya Dryomov oloc->pool_ns->len); 19542e59ffd1SIlya Dryomov else 19552e59ffd1SIlya Dryomov ceph_encode_32(p, 0); 19562e59ffd1SIlya Dryomov } 19572e59ffd1SIlya Dryomov 19588cb441c0SIlya Dryomov static void encode_request_partial(struct ceph_osd_request *req, 19598cb441c0SIlya Dryomov struct ceph_msg *msg) 1960bb873b53SIlya Dryomov { 1961bb873b53SIlya Dryomov void *p = msg->front.iov_base; 1962bb873b53SIlya Dryomov void *const end = p + msg->front_alloc_len; 1963bb873b53SIlya Dryomov u32 data_len = 0; 1964bb873b53SIlya Dryomov int i; 1965bb873b53SIlya Dryomov 1966bb873b53SIlya Dryomov if (req->r_flags & CEPH_OSD_FLAG_WRITE) { 1967bb873b53SIlya Dryomov /* snapshots aren't writeable */ 1968bb873b53SIlya Dryomov WARN_ON(req->r_snapid != CEPH_NOSNAP); 1969bb873b53SIlya Dryomov } else { 1970bb873b53SIlya Dryomov WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec || 1971bb873b53SIlya Dryomov req->r_data_offset || req->r_snapc); 1972bb873b53SIlya Dryomov } 1973bb873b53SIlya Dryomov 1974bb873b53SIlya Dryomov setup_request_data(req, msg); 1975bb873b53SIlya Dryomov 19768cb441c0SIlya Dryomov encode_spgid(&p, &req->r_t.spgid); /* actual spg */ 19778cb441c0SIlya Dryomov ceph_encode_32(&p, req->r_t.pgid.seed); /* raw hash */ 1978bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_osdc->osdmap->epoch); 1979bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_flags); 19808cb441c0SIlya Dryomov 19818cb441c0SIlya Dryomov /* reqid */ 19828cb441c0SIlya Dryomov ceph_start_encoding(&p, 2, 2, sizeof(struct ceph_osd_reqid)); 19838cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_osd_reqid)); 19848cb441c0SIlya Dryomov p += sizeof(struct ceph_osd_reqid); 19858cb441c0SIlya Dryomov 19868cb441c0SIlya Dryomov /* trace */ 19878cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_blkin_trace_info)); 19888cb441c0SIlya Dryomov p += sizeof(struct ceph_blkin_trace_info); 19898cb441c0SIlya Dryomov 19908cb441c0SIlya Dryomov ceph_encode_32(&p, 0); /* client_inc, always 0 */ 1991bb873b53SIlya Dryomov ceph_encode_timespec(p, &req->r_mtime); 1992bb873b53SIlya Dryomov p += sizeof(struct ceph_timespec); 1993aa26d662SJeff Layton 19942e59ffd1SIlya Dryomov encode_oloc(&p, end, &req->r_t.target_oloc); 19952e59ffd1SIlya Dryomov ceph_encode_string(&p, end, req->r_t.target_oid.name, 19962e59ffd1SIlya Dryomov req->r_t.target_oid.name_len); 1997bb873b53SIlya Dryomov 1998bb873b53SIlya Dryomov /* ops, can imply data */ 1999bb873b53SIlya Dryomov ceph_encode_16(&p, req->r_num_ops); 2000bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 2001bb873b53SIlya Dryomov data_len += osd_req_encode_op(p, &req->r_ops[i]); 2002bb873b53SIlya Dryomov p += sizeof(struct ceph_osd_op); 2003bb873b53SIlya Dryomov } 2004bb873b53SIlya Dryomov 2005bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapid); /* snapid */ 2006bb873b53SIlya Dryomov if (req->r_snapc) { 2007bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->seq); 2008bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_snapc->num_snaps); 2009bb873b53SIlya Dryomov for (i = 0; i < req->r_snapc->num_snaps; i++) 2010bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->snaps[i]); 2011bb873b53SIlya Dryomov } else { 2012bb873b53SIlya Dryomov ceph_encode_64(&p, 0); /* snap_seq */ 2013bb873b53SIlya Dryomov ceph_encode_32(&p, 0); /* snaps len */ 2014bb873b53SIlya Dryomov } 2015bb873b53SIlya Dryomov 2016bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_attempts); /* retry_attempt */ 2017986e8989SIlya Dryomov BUG_ON(p > end - 8); /* space for features */ 2018bb873b53SIlya Dryomov 20198cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(8); /* MOSDOp v8 */ 20208cb441c0SIlya Dryomov /* front_len is finalized in encode_request_finish() */ 2021986e8989SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 2022986e8989SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 2023bb873b53SIlya Dryomov msg->hdr.data_len = cpu_to_le32(data_len); 2024bb873b53SIlya Dryomov /* 2025bb873b53SIlya Dryomov * The header "data_off" is a hint to the receiver allowing it 2026bb873b53SIlya Dryomov * to align received data into its buffers such that there's no 2027bb873b53SIlya Dryomov * need to re-copy it before writing it to disk (direct I/O). 2028bb873b53SIlya Dryomov */ 2029bb873b53SIlya Dryomov msg->hdr.data_off = cpu_to_le16(req->r_data_offset); 2030bb873b53SIlya Dryomov 20318cb441c0SIlya Dryomov dout("%s req %p msg %p oid %s oid_len %d\n", __func__, req, msg, 20328cb441c0SIlya Dryomov req->r_t.target_oid.name, req->r_t.target_oid.name_len); 20338cb441c0SIlya Dryomov } 20348cb441c0SIlya Dryomov 20358cb441c0SIlya Dryomov static void encode_request_finish(struct ceph_msg *msg) 20368cb441c0SIlya Dryomov { 20378cb441c0SIlya Dryomov void *p = msg->front.iov_base; 2038986e8989SIlya Dryomov void *const partial_end = p + msg->front.iov_len; 20398cb441c0SIlya Dryomov void *const end = p + msg->front_alloc_len; 20408cb441c0SIlya Dryomov 20418cb441c0SIlya Dryomov if (CEPH_HAVE_FEATURE(msg->con->peer_features, RESEND_ON_SPLIT)) { 20428cb441c0SIlya Dryomov /* luminous OSD -- encode features and be done */ 2043986e8989SIlya Dryomov p = partial_end; 20448cb441c0SIlya Dryomov ceph_encode_64(&p, msg->con->peer_features); 20458cb441c0SIlya Dryomov } else { 20468cb441c0SIlya Dryomov struct { 20478cb441c0SIlya Dryomov char spgid[CEPH_ENCODING_START_BLK_LEN + 20488cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1]; 20498cb441c0SIlya Dryomov __le32 hash; 20508cb441c0SIlya Dryomov __le32 epoch; 20518cb441c0SIlya Dryomov __le32 flags; 20528cb441c0SIlya Dryomov char reqid[CEPH_ENCODING_START_BLK_LEN + 20538cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid)]; 20548cb441c0SIlya Dryomov char trace[sizeof(struct ceph_blkin_trace_info)]; 20558cb441c0SIlya Dryomov __le32 client_inc; 20568cb441c0SIlya Dryomov struct ceph_timespec mtime; 20578cb441c0SIlya Dryomov } __packed head; 20588cb441c0SIlya Dryomov struct ceph_pg pgid; 20598cb441c0SIlya Dryomov void *oloc, *oid, *tail; 20608cb441c0SIlya Dryomov int oloc_len, oid_len, tail_len; 20618cb441c0SIlya Dryomov int len; 20628cb441c0SIlya Dryomov 20638cb441c0SIlya Dryomov /* 20648cb441c0SIlya Dryomov * Pre-luminous OSD -- reencode v8 into v4 using @head 20658cb441c0SIlya Dryomov * as a temporary buffer. Encode the raw PG; the rest 20668cb441c0SIlya Dryomov * is just a matter of moving oloc, oid and tail blobs 20678cb441c0SIlya Dryomov * around. 20688cb441c0SIlya Dryomov */ 20698cb441c0SIlya Dryomov memcpy(&head, p, sizeof(head)); 20708cb441c0SIlya Dryomov p += sizeof(head); 20718cb441c0SIlya Dryomov 20728cb441c0SIlya Dryomov oloc = p; 20738cb441c0SIlya Dryomov p += CEPH_ENCODING_START_BLK_LEN; 20748cb441c0SIlya Dryomov pgid.pool = ceph_decode_64(&p); 20758cb441c0SIlya Dryomov p += 4 + 4; /* preferred, key len */ 20768cb441c0SIlya Dryomov len = ceph_decode_32(&p); 20778cb441c0SIlya Dryomov p += len; /* nspace */ 20788cb441c0SIlya Dryomov oloc_len = p - oloc; 20798cb441c0SIlya Dryomov 20808cb441c0SIlya Dryomov oid = p; 20818cb441c0SIlya Dryomov len = ceph_decode_32(&p); 20828cb441c0SIlya Dryomov p += len; 20838cb441c0SIlya Dryomov oid_len = p - oid; 20848cb441c0SIlya Dryomov 20858cb441c0SIlya Dryomov tail = p; 2086986e8989SIlya Dryomov tail_len = partial_end - p; 20878cb441c0SIlya Dryomov 20888cb441c0SIlya Dryomov p = msg->front.iov_base; 20898cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.client_inc, sizeof(head.client_inc)); 20908cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.epoch, sizeof(head.epoch)); 20918cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.flags, sizeof(head.flags)); 20928cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.mtime, sizeof(head.mtime)); 20938cb441c0SIlya Dryomov 20948cb441c0SIlya Dryomov /* reassert_version */ 20958cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_eversion)); 20968cb441c0SIlya Dryomov p += sizeof(struct ceph_eversion); 20978cb441c0SIlya Dryomov 20988cb441c0SIlya Dryomov BUG_ON(p >= oloc); 20998cb441c0SIlya Dryomov memmove(p, oloc, oloc_len); 21008cb441c0SIlya Dryomov p += oloc_len; 21018cb441c0SIlya Dryomov 21028cb441c0SIlya Dryomov pgid.seed = le32_to_cpu(head.hash); 21038cb441c0SIlya Dryomov encode_pgid(&p, &pgid); /* raw pg */ 21048cb441c0SIlya Dryomov 21058cb441c0SIlya Dryomov BUG_ON(p >= oid); 21068cb441c0SIlya Dryomov memmove(p, oid, oid_len); 21078cb441c0SIlya Dryomov p += oid_len; 21088cb441c0SIlya Dryomov 21098cb441c0SIlya Dryomov /* tail -- ops, snapid, snapc, retry_attempt */ 21108cb441c0SIlya Dryomov BUG_ON(p >= tail); 21118cb441c0SIlya Dryomov memmove(p, tail, tail_len); 21128cb441c0SIlya Dryomov p += tail_len; 21138cb441c0SIlya Dryomov 21148cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */ 21158cb441c0SIlya Dryomov } 21168cb441c0SIlya Dryomov 21178cb441c0SIlya Dryomov BUG_ON(p > end); 21188cb441c0SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 21198cb441c0SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 21208cb441c0SIlya Dryomov 21218cb441c0SIlya Dryomov dout("%s msg %p tid %llu %u+%u+%u v%d\n", __func__, msg, 21228cb441c0SIlya Dryomov le64_to_cpu(msg->hdr.tid), le32_to_cpu(msg->hdr.front_len), 21238cb441c0SIlya Dryomov le32_to_cpu(msg->hdr.middle_len), le32_to_cpu(msg->hdr.data_len), 21248cb441c0SIlya Dryomov le16_to_cpu(msg->hdr.version)); 2125bb873b53SIlya Dryomov } 2126bb873b53SIlya Dryomov 2127bb873b53SIlya Dryomov /* 2128bb873b53SIlya Dryomov * @req has to be assigned a tid and registered. 2129bb873b53SIlya Dryomov */ 2130bb873b53SIlya Dryomov static void send_request(struct ceph_osd_request *req) 2131bb873b53SIlya Dryomov { 2132bb873b53SIlya Dryomov struct ceph_osd *osd = req->r_osd; 2133bb873b53SIlya Dryomov 21345aea3dcdSIlya Dryomov verify_osd_locked(osd); 2135bb873b53SIlya Dryomov WARN_ON(osd->o_osd != req->r_t.osd); 2136bb873b53SIlya Dryomov 2137a02a946dSIlya Dryomov /* backoff? */ 2138a02a946dSIlya Dryomov if (should_plug_request(req)) 2139a02a946dSIlya Dryomov return; 2140a02a946dSIlya Dryomov 21415aea3dcdSIlya Dryomov /* 21425aea3dcdSIlya Dryomov * We may have a previously queued request message hanging 21435aea3dcdSIlya Dryomov * around. Cancel it to avoid corrupting the msgr. 21445aea3dcdSIlya Dryomov */ 21455aea3dcdSIlya Dryomov if (req->r_sent) 21465aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 21475aea3dcdSIlya Dryomov 2148bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR; 2149bb873b53SIlya Dryomov if (req->r_attempts) 2150bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_RETRY; 2151bb873b53SIlya Dryomov else 2152bb873b53SIlya Dryomov WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY); 2153bb873b53SIlya Dryomov 21548cb441c0SIlya Dryomov encode_request_partial(req, req->r_request); 2155bb873b53SIlya Dryomov 215604c7d789SIlya 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", 2157bb873b53SIlya Dryomov __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed, 2158dc98ff72SIlya Dryomov req->r_t.spgid.pgid.pool, req->r_t.spgid.pgid.seed, 215904c7d789SIlya Dryomov req->r_t.spgid.shard, osd->o_osd, req->r_t.epoch, req->r_flags, 216004c7d789SIlya Dryomov req->r_attempts); 2161bb873b53SIlya Dryomov 2162bb873b53SIlya Dryomov req->r_t.paused = false; 21633d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 2164bb873b53SIlya Dryomov req->r_attempts++; 21653d14c5d2SYehuda Sadeh 2166bb873b53SIlya Dryomov req->r_sent = osd->o_incarnation; 2167bb873b53SIlya Dryomov req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 2168bb873b53SIlya Dryomov ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request)); 21693d14c5d2SYehuda Sadeh } 21703d14c5d2SYehuda Sadeh 217142c1b124SIlya Dryomov static void maybe_request_map(struct ceph_osd_client *osdc) 217242c1b124SIlya Dryomov { 217342c1b124SIlya Dryomov bool continuous = false; 217442c1b124SIlya Dryomov 21755aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 217642c1b124SIlya Dryomov WARN_ON(!osdc->osdmap->epoch); 217742c1b124SIlya Dryomov 2178b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2179b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD) || 2180b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 218142c1b124SIlya Dryomov dout("%s osdc %p continuous\n", __func__, osdc); 218242c1b124SIlya Dryomov continuous = true; 218342c1b124SIlya Dryomov } else { 218442c1b124SIlya Dryomov dout("%s osdc %p onetime\n", __func__, osdc); 218542c1b124SIlya Dryomov } 218642c1b124SIlya Dryomov 218742c1b124SIlya Dryomov if (ceph_monc_want_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 218842c1b124SIlya Dryomov osdc->osdmap->epoch + 1, continuous)) 218942c1b124SIlya Dryomov ceph_monc_renew_subs(&osdc->client->monc); 219042c1b124SIlya Dryomov } 219142c1b124SIlya Dryomov 2192a1f4020aSJeff Layton static void complete_request(struct ceph_osd_request *req, int err); 21934609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req); 21944609245eSIlya Dryomov 21955aea3dcdSIlya Dryomov static void __submit_request(struct ceph_osd_request *req, bool wrlocked) 21960bbfdfe8SIlya Dryomov { 21975aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 21985aea3dcdSIlya Dryomov struct ceph_osd *osd; 21994609245eSIlya Dryomov enum calc_target_result ct_res; 220066850df5SIlya Dryomov int err = 0; 22015aea3dcdSIlya Dryomov bool need_send = false; 22025aea3dcdSIlya Dryomov bool promoted = false; 22030bbfdfe8SIlya Dryomov 2204b18b9550SIlya Dryomov WARN_ON(req->r_tid); 22055aea3dcdSIlya Dryomov dout("%s req %p wrlocked %d\n", __func__, req, wrlocked); 22065aea3dcdSIlya Dryomov 22075aea3dcdSIlya Dryomov again: 22087de030d6SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, NULL, false); 22094609245eSIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE && !wrlocked) 22104609245eSIlya Dryomov goto promote; 22114609245eSIlya Dryomov 22125aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, wrlocked); 22135aea3dcdSIlya Dryomov if (IS_ERR(osd)) { 22145aea3dcdSIlya Dryomov WARN_ON(PTR_ERR(osd) != -EAGAIN || wrlocked); 22155aea3dcdSIlya Dryomov goto promote; 22165aea3dcdSIlya Dryomov } 22175aea3dcdSIlya Dryomov 221866850df5SIlya Dryomov if (osdc->abort_err) { 221966850df5SIlya Dryomov dout("req %p abort_err %d\n", req, osdc->abort_err); 222066850df5SIlya Dryomov err = osdc->abort_err; 222166850df5SIlya Dryomov } else if (osdc->osdmap->epoch < osdc->epoch_barrier) { 222258eb7932SJeff Layton dout("req %p epoch %u barrier %u\n", req, osdc->osdmap->epoch, 222358eb7932SJeff Layton osdc->epoch_barrier); 222458eb7932SJeff Layton req->r_t.paused = true; 222558eb7932SJeff Layton maybe_request_map(osdc); 222658eb7932SJeff Layton } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 2227b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 22285aea3dcdSIlya Dryomov dout("req %p pausewr\n", req); 22295aea3dcdSIlya Dryomov req->r_t.paused = true; 22305aea3dcdSIlya Dryomov maybe_request_map(osdc); 22315aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_READ) && 2232b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 22335aea3dcdSIlya Dryomov dout("req %p pauserd\n", req); 22345aea3dcdSIlya Dryomov req->r_t.paused = true; 22355aea3dcdSIlya Dryomov maybe_request_map(osdc); 22365aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 22375aea3dcdSIlya Dryomov !(req->r_flags & (CEPH_OSD_FLAG_FULL_TRY | 22385aea3dcdSIlya Dryomov CEPH_OSD_FLAG_FULL_FORCE)) && 2239b7ec35b3SIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 22405aea3dcdSIlya Dryomov pool_full(osdc, req->r_t.base_oloc.pool))) { 22415aea3dcdSIlya Dryomov dout("req %p full/pool_full\n", req); 22425aea3dcdSIlya Dryomov pr_warn_ratelimited("FULL or reached pool quota\n"); 22435aea3dcdSIlya Dryomov req->r_t.paused = true; 22445aea3dcdSIlya Dryomov maybe_request_map(osdc); 2245a1f4020aSJeff Layton if (req->r_abort_on_full) 224666850df5SIlya Dryomov err = -ENOSPC; 22475aea3dcdSIlya Dryomov } else if (!osd_homeless(osd)) { 22485aea3dcdSIlya Dryomov need_send = true; 22490bbfdfe8SIlya Dryomov } else { 22505aea3dcdSIlya Dryomov maybe_request_map(osdc); 22510bbfdfe8SIlya Dryomov } 22520bbfdfe8SIlya Dryomov 22535aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 22545aea3dcdSIlya Dryomov /* 22555aea3dcdSIlya Dryomov * Assign the tid atomically with send_request() to protect 22565aea3dcdSIlya Dryomov * multiple writes to the same object from racing with each 22575aea3dcdSIlya Dryomov * other, resulting in out of order ops on the OSDs. 22585aea3dcdSIlya Dryomov */ 22595aea3dcdSIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 22605aea3dcdSIlya Dryomov link_request(osd, req); 22615aea3dcdSIlya Dryomov if (need_send) 22625aea3dcdSIlya Dryomov send_request(req); 226366850df5SIlya Dryomov else if (err) 226466850df5SIlya Dryomov complete_request(req, err); 22655aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 22665aea3dcdSIlya Dryomov 22674609245eSIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE) 22684609245eSIlya Dryomov send_map_check(req); 22694609245eSIlya Dryomov 22705aea3dcdSIlya Dryomov if (promoted) 22715aea3dcdSIlya Dryomov downgrade_write(&osdc->lock); 22725aea3dcdSIlya Dryomov return; 22735aea3dcdSIlya Dryomov 22745aea3dcdSIlya Dryomov promote: 22755aea3dcdSIlya Dryomov up_read(&osdc->lock); 22765aea3dcdSIlya Dryomov down_write(&osdc->lock); 22775aea3dcdSIlya Dryomov wrlocked = true; 22785aea3dcdSIlya Dryomov promoted = true; 22795aea3dcdSIlya Dryomov goto again; 22800bbfdfe8SIlya Dryomov } 22810bbfdfe8SIlya Dryomov 22825aea3dcdSIlya Dryomov static void account_request(struct ceph_osd_request *req) 22835aea3dcdSIlya Dryomov { 228454ea0046SIlya Dryomov WARN_ON(req->r_flags & (CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK)); 2285b18b9550SIlya Dryomov WARN_ON(!(req->r_flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE))); 22865aea3dcdSIlya Dryomov 2287b18b9550SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_ONDISK; 22885aea3dcdSIlya Dryomov atomic_inc(&req->r_osdc->num_requests); 22897cc5e38fSIlya Dryomov 22907cc5e38fSIlya Dryomov req->r_start_stamp = jiffies; 22915aea3dcdSIlya Dryomov } 22925aea3dcdSIlya Dryomov 22935aea3dcdSIlya Dryomov static void submit_request(struct ceph_osd_request *req, bool wrlocked) 22945aea3dcdSIlya Dryomov { 22955aea3dcdSIlya Dryomov ceph_osdc_get_request(req); 22965aea3dcdSIlya Dryomov account_request(req); 22975aea3dcdSIlya Dryomov __submit_request(req, wrlocked); 22985aea3dcdSIlya Dryomov } 22995aea3dcdSIlya Dryomov 230045ee2c1dSIlya Dryomov static void finish_request(struct ceph_osd_request *req) 23015aea3dcdSIlya Dryomov { 23025aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 23035aea3dcdSIlya Dryomov 23044609245eSIlya Dryomov WARN_ON(lookup_request_mc(&osdc->map_checks, req->r_tid)); 230504c7d789SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 230604c7d789SIlya Dryomov 230704c7d789SIlya Dryomov if (req->r_osd) 230804c7d789SIlya Dryomov unlink_request(req->r_osd, req); 23095aea3dcdSIlya Dryomov atomic_dec(&osdc->num_requests); 23105aea3dcdSIlya Dryomov 23115aea3dcdSIlya Dryomov /* 23125aea3dcdSIlya Dryomov * If an OSD has failed or returned and a request has been sent 23135aea3dcdSIlya Dryomov * twice, it's possible to get a reply and end up here while the 23145aea3dcdSIlya Dryomov * request message is queued for delivery. We will ignore the 23155aea3dcdSIlya Dryomov * reply, so not a big deal, but better to try and catch it. 23165aea3dcdSIlya Dryomov */ 23175aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 23185aea3dcdSIlya Dryomov ceph_msg_revoke_incoming(req->r_reply); 23195aea3dcdSIlya Dryomov } 23205aea3dcdSIlya Dryomov 2321fe5da05eSIlya Dryomov static void __complete_request(struct ceph_osd_request *req) 2322fe5da05eSIlya Dryomov { 2323b18b9550SIlya Dryomov dout("%s req %p tid %llu cb %pf result %d\n", __func__, req, 2324b18b9550SIlya Dryomov req->r_tid, req->r_callback, req->r_result); 232526df726bSIlya Dryomov 232626df726bSIlya Dryomov if (req->r_callback) 2327fe5da05eSIlya Dryomov req->r_callback(req); 232826df726bSIlya Dryomov complete_all(&req->r_completion); 232926df726bSIlya Dryomov ceph_osdc_put_request(req); 2330fe5da05eSIlya Dryomov } 2331fe5da05eSIlya Dryomov 23324609245eSIlya Dryomov /* 2333b18b9550SIlya Dryomov * This is open-coded in handle_reply(). 23344609245eSIlya Dryomov */ 23354609245eSIlya Dryomov static void complete_request(struct ceph_osd_request *req, int err) 23364609245eSIlya Dryomov { 23374609245eSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 23384609245eSIlya Dryomov 23394609245eSIlya Dryomov req->r_result = err; 234045ee2c1dSIlya Dryomov finish_request(req); 23414609245eSIlya Dryomov __complete_request(req); 23424609245eSIlya Dryomov } 23434609245eSIlya Dryomov 23444609245eSIlya Dryomov static void cancel_map_check(struct ceph_osd_request *req) 23454609245eSIlya Dryomov { 23464609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 23474609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 23484609245eSIlya Dryomov 23494609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 23504609245eSIlya Dryomov 23514609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 23524609245eSIlya Dryomov if (!lookup_req) 23534609245eSIlya Dryomov return; 23544609245eSIlya Dryomov 23554609245eSIlya Dryomov WARN_ON(lookup_req != req); 23564609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 23574609245eSIlya Dryomov ceph_osdc_put_request(req); 23584609245eSIlya Dryomov } 23594609245eSIlya Dryomov 23605aea3dcdSIlya Dryomov static void cancel_request(struct ceph_osd_request *req) 23615aea3dcdSIlya Dryomov { 23625aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 23635aea3dcdSIlya Dryomov 23644609245eSIlya Dryomov cancel_map_check(req); 236545ee2c1dSIlya Dryomov finish_request(req); 2366b18b9550SIlya Dryomov complete_all(&req->r_completion); 2367c297eb42SIlya Dryomov ceph_osdc_put_request(req); 23685aea3dcdSIlya Dryomov } 23695aea3dcdSIlya Dryomov 23707cc5e38fSIlya Dryomov static void abort_request(struct ceph_osd_request *req, int err) 23717cc5e38fSIlya Dryomov { 23727cc5e38fSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 23737cc5e38fSIlya Dryomov 23747cc5e38fSIlya Dryomov cancel_map_check(req); 23757cc5e38fSIlya Dryomov complete_request(req, err); 23767cc5e38fSIlya Dryomov } 23777cc5e38fSIlya Dryomov 237866850df5SIlya Dryomov static int abort_fn(struct ceph_osd_request *req, void *arg) 237966850df5SIlya Dryomov { 238066850df5SIlya Dryomov int err = *(int *)arg; 238166850df5SIlya Dryomov 238266850df5SIlya Dryomov abort_request(req, err); 238366850df5SIlya Dryomov return 0; /* continue iteration */ 238466850df5SIlya Dryomov } 238566850df5SIlya Dryomov 238666850df5SIlya Dryomov /* 238766850df5SIlya Dryomov * Abort all in-flight requests with @err and arrange for all future 238866850df5SIlya Dryomov * requests to be failed immediately. 238966850df5SIlya Dryomov */ 239066850df5SIlya Dryomov void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err) 239166850df5SIlya Dryomov { 239266850df5SIlya Dryomov dout("%s osdc %p err %d\n", __func__, osdc, err); 239366850df5SIlya Dryomov down_write(&osdc->lock); 239466850df5SIlya Dryomov for_each_request(osdc, abort_fn, &err); 239566850df5SIlya Dryomov osdc->abort_err = err; 239666850df5SIlya Dryomov up_write(&osdc->lock); 239766850df5SIlya Dryomov } 239866850df5SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_abort_requests); 239966850df5SIlya Dryomov 240058eb7932SJeff Layton static void update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 240158eb7932SJeff Layton { 240258eb7932SJeff Layton if (likely(eb > osdc->epoch_barrier)) { 240358eb7932SJeff Layton dout("updating epoch_barrier from %u to %u\n", 240458eb7932SJeff Layton osdc->epoch_barrier, eb); 240558eb7932SJeff Layton osdc->epoch_barrier = eb; 240658eb7932SJeff Layton /* Request map if we're not to the barrier yet */ 240758eb7932SJeff Layton if (eb > osdc->osdmap->epoch) 240858eb7932SJeff Layton maybe_request_map(osdc); 240958eb7932SJeff Layton } 241058eb7932SJeff Layton } 241158eb7932SJeff Layton 241258eb7932SJeff Layton void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 241358eb7932SJeff Layton { 241458eb7932SJeff Layton down_read(&osdc->lock); 241558eb7932SJeff Layton if (unlikely(eb > osdc->epoch_barrier)) { 241658eb7932SJeff Layton up_read(&osdc->lock); 241758eb7932SJeff Layton down_write(&osdc->lock); 241858eb7932SJeff Layton update_epoch_barrier(osdc, eb); 241958eb7932SJeff Layton up_write(&osdc->lock); 242058eb7932SJeff Layton } else { 242158eb7932SJeff Layton up_read(&osdc->lock); 242258eb7932SJeff Layton } 242358eb7932SJeff Layton } 242458eb7932SJeff Layton EXPORT_SYMBOL(ceph_osdc_update_epoch_barrier); 242558eb7932SJeff Layton 2426fc36d0a4SJeff Layton /* 2427fc36d0a4SJeff Layton * Drop all pending requests that are stalled waiting on a full condition to 242858eb7932SJeff Layton * clear, and complete them with ENOSPC as the return code. Set the 242958eb7932SJeff Layton * osdc->epoch_barrier to the latest map epoch that we've seen if any were 243058eb7932SJeff Layton * cancelled. 2431fc36d0a4SJeff Layton */ 2432fc36d0a4SJeff Layton static void ceph_osdc_abort_on_full(struct ceph_osd_client *osdc) 2433fc36d0a4SJeff Layton { 2434fc36d0a4SJeff Layton struct rb_node *n; 243558eb7932SJeff Layton bool victims = false; 2436fc36d0a4SJeff Layton 2437fc36d0a4SJeff Layton dout("enter abort_on_full\n"); 2438fc36d0a4SJeff Layton 2439fc36d0a4SJeff Layton if (!ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) && !have_pool_full(osdc)) 2440fc36d0a4SJeff Layton goto out; 2441fc36d0a4SJeff Layton 244258eb7932SJeff Layton /* Scan list and see if there is anything to abort */ 244358eb7932SJeff Layton for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 244458eb7932SJeff Layton struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 244558eb7932SJeff Layton struct rb_node *m; 244658eb7932SJeff Layton 244758eb7932SJeff Layton m = rb_first(&osd->o_requests); 244858eb7932SJeff Layton while (m) { 244958eb7932SJeff Layton struct ceph_osd_request *req = rb_entry(m, 245058eb7932SJeff Layton struct ceph_osd_request, r_node); 245158eb7932SJeff Layton m = rb_next(m); 245258eb7932SJeff Layton 245358eb7932SJeff Layton if (req->r_abort_on_full) { 245458eb7932SJeff Layton victims = true; 245558eb7932SJeff Layton break; 245658eb7932SJeff Layton } 245758eb7932SJeff Layton } 245858eb7932SJeff Layton if (victims) 245958eb7932SJeff Layton break; 246058eb7932SJeff Layton } 246158eb7932SJeff Layton 246258eb7932SJeff Layton if (!victims) 246358eb7932SJeff Layton goto out; 246458eb7932SJeff Layton 246558eb7932SJeff Layton /* 246658eb7932SJeff Layton * Update the barrier to current epoch if it's behind that point, 246758eb7932SJeff Layton * since we know we have some calls to be aborted in the tree. 246858eb7932SJeff Layton */ 246958eb7932SJeff Layton update_epoch_barrier(osdc, osdc->osdmap->epoch); 247058eb7932SJeff Layton 2471fc36d0a4SJeff Layton for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 2472fc36d0a4SJeff Layton struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 2473fc36d0a4SJeff Layton struct rb_node *m; 2474fc36d0a4SJeff Layton 2475fc36d0a4SJeff Layton m = rb_first(&osd->o_requests); 2476fc36d0a4SJeff Layton while (m) { 2477fc36d0a4SJeff Layton struct ceph_osd_request *req = rb_entry(m, 2478fc36d0a4SJeff Layton struct ceph_osd_request, r_node); 2479fc36d0a4SJeff Layton m = rb_next(m); 2480fc36d0a4SJeff Layton 2481fc36d0a4SJeff Layton if (req->r_abort_on_full && 2482fc36d0a4SJeff Layton (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2483fc36d0a4SJeff Layton pool_full(osdc, req->r_t.target_oloc.pool))) 2484fc36d0a4SJeff Layton abort_request(req, -ENOSPC); 2485fc36d0a4SJeff Layton } 2486fc36d0a4SJeff Layton } 2487fc36d0a4SJeff Layton out: 248858eb7932SJeff Layton dout("return abort_on_full barrier=%u\n", osdc->epoch_barrier); 2489fc36d0a4SJeff Layton } 2490fc36d0a4SJeff Layton 24914609245eSIlya Dryomov static void check_pool_dne(struct ceph_osd_request *req) 24924609245eSIlya Dryomov { 24934609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 24944609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 24954609245eSIlya Dryomov 24964609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 24974609245eSIlya Dryomov WARN_ON(!map->epoch); 24984609245eSIlya Dryomov 24994609245eSIlya Dryomov if (req->r_attempts) { 25004609245eSIlya Dryomov /* 25014609245eSIlya Dryomov * We sent a request earlier, which means that 25024609245eSIlya Dryomov * previously the pool existed, and now it does not 25034609245eSIlya Dryomov * (i.e., it was deleted). 25044609245eSIlya Dryomov */ 25054609245eSIlya Dryomov req->r_map_dne_bound = map->epoch; 25064609245eSIlya Dryomov dout("%s req %p tid %llu pool disappeared\n", __func__, req, 25074609245eSIlya Dryomov req->r_tid); 25084609245eSIlya Dryomov } else { 25094609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u have %u\n", __func__, 25104609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, map->epoch); 25114609245eSIlya Dryomov } 25124609245eSIlya Dryomov 25134609245eSIlya Dryomov if (req->r_map_dne_bound) { 25144609245eSIlya Dryomov if (map->epoch >= req->r_map_dne_bound) { 25154609245eSIlya Dryomov /* we had a new enough map */ 25164609245eSIlya Dryomov pr_info_ratelimited("tid %llu pool does not exist\n", 25174609245eSIlya Dryomov req->r_tid); 25184609245eSIlya Dryomov complete_request(req, -ENOENT); 25194609245eSIlya Dryomov } 25204609245eSIlya Dryomov } else { 25214609245eSIlya Dryomov send_map_check(req); 25224609245eSIlya Dryomov } 25234609245eSIlya Dryomov } 25244609245eSIlya Dryomov 25254609245eSIlya Dryomov static void map_check_cb(struct ceph_mon_generic_request *greq) 25264609245eSIlya Dryomov { 25274609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 25284609245eSIlya Dryomov struct ceph_osd_request *req; 25294609245eSIlya Dryomov u64 tid = greq->private_data; 25304609245eSIlya Dryomov 25314609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 25324609245eSIlya Dryomov 25334609245eSIlya Dryomov down_write(&osdc->lock); 25344609245eSIlya Dryomov req = lookup_request_mc(&osdc->map_checks, tid); 25354609245eSIlya Dryomov if (!req) { 25364609245eSIlya Dryomov dout("%s tid %llu dne\n", __func__, tid); 25374609245eSIlya Dryomov goto out_unlock; 25384609245eSIlya Dryomov } 25394609245eSIlya Dryomov 25404609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u newest %llu\n", __func__, 25414609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, greq->u.newest); 25424609245eSIlya Dryomov if (!req->r_map_dne_bound) 25434609245eSIlya Dryomov req->r_map_dne_bound = greq->u.newest; 25444609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 25454609245eSIlya Dryomov check_pool_dne(req); 25464609245eSIlya Dryomov 25474609245eSIlya Dryomov ceph_osdc_put_request(req); 25484609245eSIlya Dryomov out_unlock: 25494609245eSIlya Dryomov up_write(&osdc->lock); 25504609245eSIlya Dryomov } 25514609245eSIlya Dryomov 25524609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req) 25534609245eSIlya Dryomov { 25544609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 25554609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 25564609245eSIlya Dryomov int ret; 25574609245eSIlya Dryomov 25584609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 25594609245eSIlya Dryomov 25604609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 25614609245eSIlya Dryomov if (lookup_req) { 25624609245eSIlya Dryomov WARN_ON(lookup_req != req); 25634609245eSIlya Dryomov return; 25644609245eSIlya Dryomov } 25654609245eSIlya Dryomov 25664609245eSIlya Dryomov ceph_osdc_get_request(req); 25674609245eSIlya Dryomov insert_request_mc(&osdc->map_checks, req); 25684609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 25694609245eSIlya Dryomov map_check_cb, req->r_tid); 25704609245eSIlya Dryomov WARN_ON(ret); 25714609245eSIlya Dryomov } 25724609245eSIlya Dryomov 25730bbfdfe8SIlya Dryomov /* 2574922dab61SIlya Dryomov * lingering requests, watch/notify v2 infrastructure 2575922dab61SIlya Dryomov */ 2576922dab61SIlya Dryomov static void linger_release(struct kref *kref) 2577922dab61SIlya Dryomov { 2578922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 2579922dab61SIlya Dryomov container_of(kref, struct ceph_osd_linger_request, kref); 2580922dab61SIlya Dryomov 2581922dab61SIlya Dryomov dout("%s lreq %p reg_req %p ping_req %p\n", __func__, lreq, 2582922dab61SIlya Dryomov lreq->reg_req, lreq->ping_req); 2583922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->node)); 2584922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->osdc_node)); 25854609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->mc_node)); 2586922dab61SIlya Dryomov WARN_ON(!list_empty(&lreq->scan_item)); 2587b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lreq->pending_lworks)); 2588922dab61SIlya Dryomov WARN_ON(lreq->osd); 2589922dab61SIlya Dryomov 2590922dab61SIlya Dryomov if (lreq->reg_req) 2591922dab61SIlya Dryomov ceph_osdc_put_request(lreq->reg_req); 2592922dab61SIlya Dryomov if (lreq->ping_req) 2593922dab61SIlya Dryomov ceph_osdc_put_request(lreq->ping_req); 2594922dab61SIlya Dryomov target_destroy(&lreq->t); 2595922dab61SIlya Dryomov kfree(lreq); 2596922dab61SIlya Dryomov } 2597922dab61SIlya Dryomov 2598922dab61SIlya Dryomov static void linger_put(struct ceph_osd_linger_request *lreq) 2599922dab61SIlya Dryomov { 2600922dab61SIlya Dryomov if (lreq) 2601922dab61SIlya Dryomov kref_put(&lreq->kref, linger_release); 2602922dab61SIlya Dryomov } 2603922dab61SIlya Dryomov 2604922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2605922dab61SIlya Dryomov linger_get(struct ceph_osd_linger_request *lreq) 2606922dab61SIlya Dryomov { 2607922dab61SIlya Dryomov kref_get(&lreq->kref); 2608922dab61SIlya Dryomov return lreq; 2609922dab61SIlya Dryomov } 2610922dab61SIlya Dryomov 2611922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2612922dab61SIlya Dryomov linger_alloc(struct ceph_osd_client *osdc) 2613922dab61SIlya Dryomov { 2614922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2615922dab61SIlya Dryomov 2616922dab61SIlya Dryomov lreq = kzalloc(sizeof(*lreq), GFP_NOIO); 2617922dab61SIlya Dryomov if (!lreq) 2618922dab61SIlya Dryomov return NULL; 2619922dab61SIlya Dryomov 2620922dab61SIlya Dryomov kref_init(&lreq->kref); 2621922dab61SIlya Dryomov mutex_init(&lreq->lock); 2622922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->node); 2623922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->osdc_node); 26244609245eSIlya Dryomov RB_CLEAR_NODE(&lreq->mc_node); 2625922dab61SIlya Dryomov INIT_LIST_HEAD(&lreq->scan_item); 2626b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lreq->pending_lworks); 2627922dab61SIlya Dryomov init_completion(&lreq->reg_commit_wait); 262819079203SIlya Dryomov init_completion(&lreq->notify_finish_wait); 2629922dab61SIlya Dryomov 2630922dab61SIlya Dryomov lreq->osdc = osdc; 2631922dab61SIlya Dryomov target_init(&lreq->t); 2632922dab61SIlya Dryomov 2633922dab61SIlya Dryomov dout("%s lreq %p\n", __func__, lreq); 2634922dab61SIlya Dryomov return lreq; 2635922dab61SIlya Dryomov } 2636922dab61SIlya Dryomov 2637922dab61SIlya Dryomov DEFINE_RB_INSDEL_FUNCS(linger, struct ceph_osd_linger_request, linger_id, node) 2638922dab61SIlya Dryomov DEFINE_RB_FUNCS(linger_osdc, struct ceph_osd_linger_request, linger_id, osdc_node) 26394609245eSIlya Dryomov DEFINE_RB_FUNCS(linger_mc, struct ceph_osd_linger_request, linger_id, mc_node) 2640922dab61SIlya Dryomov 2641922dab61SIlya Dryomov /* 2642922dab61SIlya Dryomov * Create linger request <-> OSD session relation. 2643922dab61SIlya Dryomov * 2644922dab61SIlya Dryomov * @lreq has to be registered, @osd may be homeless. 2645922dab61SIlya Dryomov */ 2646922dab61SIlya Dryomov static void link_linger(struct ceph_osd *osd, 2647922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2648922dab61SIlya Dryomov { 2649922dab61SIlya Dryomov verify_osd_locked(osd); 2650922dab61SIlya Dryomov WARN_ON(!lreq->linger_id || lreq->osd); 2651922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2652922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2653922dab61SIlya Dryomov 2654922dab61SIlya Dryomov if (!osd_homeless(osd)) 2655922dab61SIlya Dryomov __remove_osd_from_lru(osd); 2656922dab61SIlya Dryomov else 2657922dab61SIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 2658922dab61SIlya Dryomov 2659922dab61SIlya Dryomov get_osd(osd); 2660922dab61SIlya Dryomov insert_linger(&osd->o_linger_requests, lreq); 2661922dab61SIlya Dryomov lreq->osd = osd; 2662922dab61SIlya Dryomov } 2663922dab61SIlya Dryomov 2664922dab61SIlya Dryomov static void unlink_linger(struct ceph_osd *osd, 2665922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2666922dab61SIlya Dryomov { 2667922dab61SIlya Dryomov verify_osd_locked(osd); 2668922dab61SIlya Dryomov WARN_ON(lreq->osd != osd); 2669922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2670922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2671922dab61SIlya Dryomov 2672922dab61SIlya Dryomov lreq->osd = NULL; 2673922dab61SIlya Dryomov erase_linger(&osd->o_linger_requests, lreq); 2674922dab61SIlya Dryomov put_osd(osd); 2675922dab61SIlya Dryomov 2676922dab61SIlya Dryomov if (!osd_homeless(osd)) 2677922dab61SIlya Dryomov maybe_move_osd_to_lru(osd); 2678922dab61SIlya Dryomov else 2679922dab61SIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 2680922dab61SIlya Dryomov } 2681922dab61SIlya Dryomov 2682922dab61SIlya Dryomov static bool __linger_registered(struct ceph_osd_linger_request *lreq) 2683922dab61SIlya Dryomov { 2684922dab61SIlya Dryomov verify_osdc_locked(lreq->osdc); 2685922dab61SIlya Dryomov 2686922dab61SIlya Dryomov return !RB_EMPTY_NODE(&lreq->osdc_node); 2687922dab61SIlya Dryomov } 2688922dab61SIlya Dryomov 2689922dab61SIlya Dryomov static bool linger_registered(struct ceph_osd_linger_request *lreq) 2690922dab61SIlya Dryomov { 2691922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2692922dab61SIlya Dryomov bool registered; 2693922dab61SIlya Dryomov 2694922dab61SIlya Dryomov down_read(&osdc->lock); 2695922dab61SIlya Dryomov registered = __linger_registered(lreq); 2696922dab61SIlya Dryomov up_read(&osdc->lock); 2697922dab61SIlya Dryomov 2698922dab61SIlya Dryomov return registered; 2699922dab61SIlya Dryomov } 2700922dab61SIlya Dryomov 2701922dab61SIlya Dryomov static void linger_register(struct ceph_osd_linger_request *lreq) 2702922dab61SIlya Dryomov { 2703922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2704922dab61SIlya Dryomov 2705922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2706922dab61SIlya Dryomov WARN_ON(lreq->linger_id); 2707922dab61SIlya Dryomov 2708922dab61SIlya Dryomov linger_get(lreq); 2709922dab61SIlya Dryomov lreq->linger_id = ++osdc->last_linger_id; 2710922dab61SIlya Dryomov insert_linger_osdc(&osdc->linger_requests, lreq); 2711922dab61SIlya Dryomov } 2712922dab61SIlya Dryomov 2713922dab61SIlya Dryomov static void linger_unregister(struct ceph_osd_linger_request *lreq) 2714922dab61SIlya Dryomov { 2715922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2716922dab61SIlya Dryomov 2717922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2718922dab61SIlya Dryomov 2719922dab61SIlya Dryomov erase_linger_osdc(&osdc->linger_requests, lreq); 2720922dab61SIlya Dryomov linger_put(lreq); 2721922dab61SIlya Dryomov } 2722922dab61SIlya Dryomov 2723922dab61SIlya Dryomov static void cancel_linger_request(struct ceph_osd_request *req) 2724922dab61SIlya Dryomov { 2725922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2726922dab61SIlya Dryomov 2727922dab61SIlya Dryomov WARN_ON(!req->r_linger); 2728922dab61SIlya Dryomov cancel_request(req); 2729922dab61SIlya Dryomov linger_put(lreq); 2730922dab61SIlya Dryomov } 2731922dab61SIlya Dryomov 2732922dab61SIlya Dryomov struct linger_work { 2733922dab61SIlya Dryomov struct work_struct work; 2734922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2735b07d3c4bSIlya Dryomov struct list_head pending_item; 2736b07d3c4bSIlya Dryomov unsigned long queued_stamp; 2737922dab61SIlya Dryomov 2738922dab61SIlya Dryomov union { 2739922dab61SIlya Dryomov struct { 2740922dab61SIlya Dryomov u64 notify_id; 2741922dab61SIlya Dryomov u64 notifier_id; 2742922dab61SIlya Dryomov void *payload; /* points into @msg front */ 2743922dab61SIlya Dryomov size_t payload_len; 2744922dab61SIlya Dryomov 2745922dab61SIlya Dryomov struct ceph_msg *msg; /* for ceph_msg_put() */ 2746922dab61SIlya Dryomov } notify; 2747922dab61SIlya Dryomov struct { 2748922dab61SIlya Dryomov int err; 2749922dab61SIlya Dryomov } error; 2750922dab61SIlya Dryomov }; 2751922dab61SIlya Dryomov }; 2752922dab61SIlya Dryomov 2753922dab61SIlya Dryomov static struct linger_work *lwork_alloc(struct ceph_osd_linger_request *lreq, 2754922dab61SIlya Dryomov work_func_t workfn) 2755922dab61SIlya Dryomov { 2756922dab61SIlya Dryomov struct linger_work *lwork; 2757922dab61SIlya Dryomov 2758922dab61SIlya Dryomov lwork = kzalloc(sizeof(*lwork), GFP_NOIO); 2759922dab61SIlya Dryomov if (!lwork) 2760922dab61SIlya Dryomov return NULL; 2761922dab61SIlya Dryomov 2762922dab61SIlya Dryomov INIT_WORK(&lwork->work, workfn); 2763b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lwork->pending_item); 2764922dab61SIlya Dryomov lwork->lreq = linger_get(lreq); 2765922dab61SIlya Dryomov 2766922dab61SIlya Dryomov return lwork; 2767922dab61SIlya Dryomov } 2768922dab61SIlya Dryomov 2769922dab61SIlya Dryomov static void lwork_free(struct linger_work *lwork) 2770922dab61SIlya Dryomov { 2771922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2772922dab61SIlya Dryomov 2773b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 2774b07d3c4bSIlya Dryomov list_del(&lwork->pending_item); 2775b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 2776b07d3c4bSIlya Dryomov 2777922dab61SIlya Dryomov linger_put(lreq); 2778922dab61SIlya Dryomov kfree(lwork); 2779922dab61SIlya Dryomov } 2780922dab61SIlya Dryomov 2781922dab61SIlya Dryomov static void lwork_queue(struct linger_work *lwork) 2782922dab61SIlya Dryomov { 2783922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2784922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2785922dab61SIlya Dryomov 2786922dab61SIlya Dryomov verify_lreq_locked(lreq); 2787b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lwork->pending_item)); 2788b07d3c4bSIlya Dryomov 2789b07d3c4bSIlya Dryomov lwork->queued_stamp = jiffies; 2790b07d3c4bSIlya Dryomov list_add_tail(&lwork->pending_item, &lreq->pending_lworks); 2791922dab61SIlya Dryomov queue_work(osdc->notify_wq, &lwork->work); 2792922dab61SIlya Dryomov } 2793922dab61SIlya Dryomov 2794922dab61SIlya Dryomov static void do_watch_notify(struct work_struct *w) 2795922dab61SIlya Dryomov { 2796922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2797922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2798922dab61SIlya Dryomov 2799922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2800922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2801922dab61SIlya Dryomov goto out; 2802922dab61SIlya Dryomov } 2803922dab61SIlya Dryomov 280419079203SIlya Dryomov WARN_ON(!lreq->is_watch); 2805922dab61SIlya Dryomov dout("%s lreq %p notify_id %llu notifier_id %llu payload_len %zu\n", 2806922dab61SIlya Dryomov __func__, lreq, lwork->notify.notify_id, lwork->notify.notifier_id, 2807922dab61SIlya Dryomov lwork->notify.payload_len); 2808922dab61SIlya Dryomov lreq->wcb(lreq->data, lwork->notify.notify_id, lreq->linger_id, 2809922dab61SIlya Dryomov lwork->notify.notifier_id, lwork->notify.payload, 2810922dab61SIlya Dryomov lwork->notify.payload_len); 2811922dab61SIlya Dryomov 2812922dab61SIlya Dryomov out: 2813922dab61SIlya Dryomov ceph_msg_put(lwork->notify.msg); 2814922dab61SIlya Dryomov lwork_free(lwork); 2815922dab61SIlya Dryomov } 2816922dab61SIlya Dryomov 2817922dab61SIlya Dryomov static void do_watch_error(struct work_struct *w) 2818922dab61SIlya Dryomov { 2819922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2820922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2821922dab61SIlya Dryomov 2822922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2823922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2824922dab61SIlya Dryomov goto out; 2825922dab61SIlya Dryomov } 2826922dab61SIlya Dryomov 2827922dab61SIlya Dryomov dout("%s lreq %p err %d\n", __func__, lreq, lwork->error.err); 2828922dab61SIlya Dryomov lreq->errcb(lreq->data, lreq->linger_id, lwork->error.err); 2829922dab61SIlya Dryomov 2830922dab61SIlya Dryomov out: 2831922dab61SIlya Dryomov lwork_free(lwork); 2832922dab61SIlya Dryomov } 2833922dab61SIlya Dryomov 2834922dab61SIlya Dryomov static void queue_watch_error(struct ceph_osd_linger_request *lreq) 2835922dab61SIlya Dryomov { 2836922dab61SIlya Dryomov struct linger_work *lwork; 2837922dab61SIlya Dryomov 2838922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_error); 2839922dab61SIlya Dryomov if (!lwork) { 2840922dab61SIlya Dryomov pr_err("failed to allocate error-lwork\n"); 2841922dab61SIlya Dryomov return; 2842922dab61SIlya Dryomov } 2843922dab61SIlya Dryomov 2844922dab61SIlya Dryomov lwork->error.err = lreq->last_error; 2845922dab61SIlya Dryomov lwork_queue(lwork); 2846922dab61SIlya Dryomov } 2847922dab61SIlya Dryomov 2848922dab61SIlya Dryomov static void linger_reg_commit_complete(struct ceph_osd_linger_request *lreq, 2849922dab61SIlya Dryomov int result) 2850922dab61SIlya Dryomov { 2851922dab61SIlya Dryomov if (!completion_done(&lreq->reg_commit_wait)) { 2852922dab61SIlya Dryomov lreq->reg_commit_error = (result <= 0 ? result : 0); 2853922dab61SIlya Dryomov complete_all(&lreq->reg_commit_wait); 2854922dab61SIlya Dryomov } 2855922dab61SIlya Dryomov } 2856922dab61SIlya Dryomov 2857922dab61SIlya Dryomov static void linger_commit_cb(struct ceph_osd_request *req) 2858922dab61SIlya Dryomov { 2859922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2860922dab61SIlya Dryomov 2861922dab61SIlya Dryomov mutex_lock(&lreq->lock); 2862922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d\n", __func__, lreq, 2863922dab61SIlya Dryomov lreq->linger_id, req->r_result); 2864922dab61SIlya Dryomov linger_reg_commit_complete(lreq, req->r_result); 2865922dab61SIlya Dryomov lreq->committed = true; 2866922dab61SIlya Dryomov 286719079203SIlya Dryomov if (!lreq->is_watch) { 286819079203SIlya Dryomov struct ceph_osd_data *osd_data = 286919079203SIlya Dryomov osd_req_op_data(req, 0, notify, response_data); 287019079203SIlya Dryomov void *p = page_address(osd_data->pages[0]); 287119079203SIlya Dryomov 287219079203SIlya Dryomov WARN_ON(req->r_ops[0].op != CEPH_OSD_OP_NOTIFY || 287319079203SIlya Dryomov osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 287419079203SIlya Dryomov 287519079203SIlya Dryomov /* make note of the notify_id */ 287619079203SIlya Dryomov if (req->r_ops[0].outdata_len >= sizeof(u64)) { 287719079203SIlya Dryomov lreq->notify_id = ceph_decode_64(&p); 287819079203SIlya Dryomov dout("lreq %p notify_id %llu\n", lreq, 287919079203SIlya Dryomov lreq->notify_id); 288019079203SIlya Dryomov } else { 288119079203SIlya Dryomov dout("lreq %p no notify_id\n", lreq); 288219079203SIlya Dryomov } 288319079203SIlya Dryomov } 288419079203SIlya Dryomov 2885922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2886922dab61SIlya Dryomov linger_put(lreq); 2887922dab61SIlya Dryomov } 2888922dab61SIlya Dryomov 2889922dab61SIlya Dryomov static int normalize_watch_error(int err) 2890922dab61SIlya Dryomov { 2891922dab61SIlya Dryomov /* 2892922dab61SIlya Dryomov * Translate ENOENT -> ENOTCONN so that a delete->disconnection 2893922dab61SIlya Dryomov * notification and a failure to reconnect because we raced with 2894922dab61SIlya Dryomov * the delete appear the same to the user. 2895922dab61SIlya Dryomov */ 2896922dab61SIlya Dryomov if (err == -ENOENT) 2897922dab61SIlya Dryomov err = -ENOTCONN; 2898922dab61SIlya Dryomov 2899922dab61SIlya Dryomov return err; 2900922dab61SIlya Dryomov } 2901922dab61SIlya Dryomov 2902922dab61SIlya Dryomov static void linger_reconnect_cb(struct ceph_osd_request *req) 2903922dab61SIlya Dryomov { 2904922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2905922dab61SIlya Dryomov 2906922dab61SIlya Dryomov mutex_lock(&lreq->lock); 2907922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d last_error %d\n", __func__, 2908922dab61SIlya Dryomov lreq, lreq->linger_id, req->r_result, lreq->last_error); 2909922dab61SIlya Dryomov if (req->r_result < 0) { 2910922dab61SIlya Dryomov if (!lreq->last_error) { 2911922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 2912922dab61SIlya Dryomov queue_watch_error(lreq); 2913922dab61SIlya Dryomov } 2914922dab61SIlya Dryomov } 2915922dab61SIlya Dryomov 2916922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2917922dab61SIlya Dryomov linger_put(lreq); 2918922dab61SIlya Dryomov } 2919922dab61SIlya Dryomov 2920922dab61SIlya Dryomov static void send_linger(struct ceph_osd_linger_request *lreq) 2921922dab61SIlya Dryomov { 2922922dab61SIlya Dryomov struct ceph_osd_request *req = lreq->reg_req; 2923922dab61SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[0]; 2924922dab61SIlya Dryomov 2925922dab61SIlya Dryomov verify_osdc_wrlocked(req->r_osdc); 2926922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 2927922dab61SIlya Dryomov 2928922dab61SIlya Dryomov if (req->r_osd) 2929922dab61SIlya Dryomov cancel_linger_request(req); 2930922dab61SIlya Dryomov 2931922dab61SIlya Dryomov request_reinit(req); 2932922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 2933922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 2934922dab61SIlya Dryomov req->r_flags = lreq->t.flags; 2935922dab61SIlya Dryomov req->r_mtime = lreq->mtime; 2936922dab61SIlya Dryomov 2937922dab61SIlya Dryomov mutex_lock(&lreq->lock); 293819079203SIlya Dryomov if (lreq->is_watch && lreq->committed) { 2939922dab61SIlya Dryomov WARN_ON(op->op != CEPH_OSD_OP_WATCH || 2940922dab61SIlya Dryomov op->watch.cookie != lreq->linger_id); 2941922dab61SIlya Dryomov op->watch.op = CEPH_OSD_WATCH_OP_RECONNECT; 2942922dab61SIlya Dryomov op->watch.gen = ++lreq->register_gen; 2943922dab61SIlya Dryomov dout("lreq %p reconnect register_gen %u\n", lreq, 2944922dab61SIlya Dryomov op->watch.gen); 2945922dab61SIlya Dryomov req->r_callback = linger_reconnect_cb; 2946922dab61SIlya Dryomov } else { 294719079203SIlya Dryomov if (!lreq->is_watch) 294819079203SIlya Dryomov lreq->notify_id = 0; 294919079203SIlya Dryomov else 2950922dab61SIlya Dryomov WARN_ON(op->watch.op != CEPH_OSD_WATCH_OP_WATCH); 2951922dab61SIlya Dryomov dout("lreq %p register\n", lreq); 2952922dab61SIlya Dryomov req->r_callback = linger_commit_cb; 2953922dab61SIlya Dryomov } 2954922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2955922dab61SIlya Dryomov 2956922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 2957922dab61SIlya Dryomov req->r_linger = true; 2958922dab61SIlya Dryomov 2959922dab61SIlya Dryomov submit_request(req, true); 2960922dab61SIlya Dryomov } 2961922dab61SIlya Dryomov 2962922dab61SIlya Dryomov static void linger_ping_cb(struct ceph_osd_request *req) 2963922dab61SIlya Dryomov { 2964922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2965922dab61SIlya Dryomov 2966922dab61SIlya Dryomov mutex_lock(&lreq->lock); 2967922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d ping_sent %lu last_error %d\n", 2968922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, req->r_result, lreq->ping_sent, 2969922dab61SIlya Dryomov lreq->last_error); 2970922dab61SIlya Dryomov if (lreq->register_gen == req->r_ops[0].watch.gen) { 2971b07d3c4bSIlya Dryomov if (!req->r_result) { 2972b07d3c4bSIlya Dryomov lreq->watch_valid_thru = lreq->ping_sent; 2973b07d3c4bSIlya Dryomov } else if (!lreq->last_error) { 2974922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 2975922dab61SIlya Dryomov queue_watch_error(lreq); 2976922dab61SIlya Dryomov } 2977922dab61SIlya Dryomov } else { 2978922dab61SIlya Dryomov dout("lreq %p register_gen %u ignoring old pong %u\n", lreq, 2979922dab61SIlya Dryomov lreq->register_gen, req->r_ops[0].watch.gen); 2980922dab61SIlya Dryomov } 2981922dab61SIlya Dryomov 2982922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2983922dab61SIlya Dryomov linger_put(lreq); 2984922dab61SIlya Dryomov } 2985922dab61SIlya Dryomov 2986922dab61SIlya Dryomov static void send_linger_ping(struct ceph_osd_linger_request *lreq) 2987922dab61SIlya Dryomov { 2988922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2989922dab61SIlya Dryomov struct ceph_osd_request *req = lreq->ping_req; 2990922dab61SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[0]; 2991922dab61SIlya Dryomov 2992b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 2993922dab61SIlya Dryomov dout("%s PAUSERD\n", __func__); 2994922dab61SIlya Dryomov return; 2995922dab61SIlya Dryomov } 2996922dab61SIlya Dryomov 2997922dab61SIlya Dryomov lreq->ping_sent = jiffies; 2998922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu ping_sent %lu register_gen %u\n", 2999922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, lreq->ping_sent, 3000922dab61SIlya Dryomov lreq->register_gen); 3001922dab61SIlya Dryomov 3002922dab61SIlya Dryomov if (req->r_osd) 3003922dab61SIlya Dryomov cancel_linger_request(req); 3004922dab61SIlya Dryomov 3005922dab61SIlya Dryomov request_reinit(req); 3006922dab61SIlya Dryomov target_copy(&req->r_t, &lreq->t); 3007922dab61SIlya Dryomov 3008922dab61SIlya Dryomov WARN_ON(op->op != CEPH_OSD_OP_WATCH || 3009922dab61SIlya Dryomov op->watch.cookie != lreq->linger_id || 3010922dab61SIlya Dryomov op->watch.op != CEPH_OSD_WATCH_OP_PING); 3011922dab61SIlya Dryomov op->watch.gen = lreq->register_gen; 3012922dab61SIlya Dryomov req->r_callback = linger_ping_cb; 3013922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 3014922dab61SIlya Dryomov req->r_linger = true; 3015922dab61SIlya Dryomov 3016922dab61SIlya Dryomov ceph_osdc_get_request(req); 3017922dab61SIlya Dryomov account_request(req); 3018922dab61SIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 3019922dab61SIlya Dryomov link_request(lreq->osd, req); 3020922dab61SIlya Dryomov send_request(req); 3021922dab61SIlya Dryomov } 3022922dab61SIlya Dryomov 3023922dab61SIlya Dryomov static void linger_submit(struct ceph_osd_linger_request *lreq) 3024922dab61SIlya Dryomov { 3025922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3026922dab61SIlya Dryomov struct ceph_osd *osd; 3027922dab61SIlya Dryomov 30287de030d6SIlya Dryomov calc_target(osdc, &lreq->t, NULL, false); 3029922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3030922dab61SIlya Dryomov link_linger(osd, lreq); 3031922dab61SIlya Dryomov 3032922dab61SIlya Dryomov send_linger(lreq); 3033922dab61SIlya Dryomov } 3034922dab61SIlya Dryomov 30354609245eSIlya Dryomov static void cancel_linger_map_check(struct ceph_osd_linger_request *lreq) 30364609245eSIlya Dryomov { 30374609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 30384609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 30394609245eSIlya Dryomov 30404609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 30414609245eSIlya Dryomov 30424609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 30434609245eSIlya Dryomov lreq->linger_id); 30444609245eSIlya Dryomov if (!lookup_lreq) 30454609245eSIlya Dryomov return; 30464609245eSIlya Dryomov 30474609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 30484609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 30494609245eSIlya Dryomov linger_put(lreq); 30504609245eSIlya Dryomov } 30514609245eSIlya Dryomov 3052922dab61SIlya Dryomov /* 3053922dab61SIlya Dryomov * @lreq has to be both registered and linked. 3054922dab61SIlya Dryomov */ 3055922dab61SIlya Dryomov static void __linger_cancel(struct ceph_osd_linger_request *lreq) 3056922dab61SIlya Dryomov { 305719079203SIlya Dryomov if (lreq->is_watch && lreq->ping_req->r_osd) 3058922dab61SIlya Dryomov cancel_linger_request(lreq->ping_req); 3059922dab61SIlya Dryomov if (lreq->reg_req->r_osd) 3060922dab61SIlya Dryomov cancel_linger_request(lreq->reg_req); 30614609245eSIlya Dryomov cancel_linger_map_check(lreq); 3062922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3063922dab61SIlya Dryomov linger_unregister(lreq); 3064922dab61SIlya Dryomov } 3065922dab61SIlya Dryomov 3066922dab61SIlya Dryomov static void linger_cancel(struct ceph_osd_linger_request *lreq) 3067922dab61SIlya Dryomov { 3068922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3069922dab61SIlya Dryomov 3070922dab61SIlya Dryomov down_write(&osdc->lock); 3071922dab61SIlya Dryomov if (__linger_registered(lreq)) 3072922dab61SIlya Dryomov __linger_cancel(lreq); 3073922dab61SIlya Dryomov up_write(&osdc->lock); 3074922dab61SIlya Dryomov } 3075922dab61SIlya Dryomov 30764609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq); 30774609245eSIlya Dryomov 30784609245eSIlya Dryomov static void check_linger_pool_dne(struct ceph_osd_linger_request *lreq) 30794609245eSIlya Dryomov { 30804609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 30814609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 30824609245eSIlya Dryomov 30834609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 30844609245eSIlya Dryomov WARN_ON(!map->epoch); 30854609245eSIlya Dryomov 30864609245eSIlya Dryomov if (lreq->register_gen) { 30874609245eSIlya Dryomov lreq->map_dne_bound = map->epoch; 30884609245eSIlya Dryomov dout("%s lreq %p linger_id %llu pool disappeared\n", __func__, 30894609245eSIlya Dryomov lreq, lreq->linger_id); 30904609245eSIlya Dryomov } else { 30914609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u have %u\n", 30924609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 30934609245eSIlya Dryomov map->epoch); 30944609245eSIlya Dryomov } 30954609245eSIlya Dryomov 30964609245eSIlya Dryomov if (lreq->map_dne_bound) { 30974609245eSIlya Dryomov if (map->epoch >= lreq->map_dne_bound) { 30984609245eSIlya Dryomov /* we had a new enough map */ 30994609245eSIlya Dryomov pr_info("linger_id %llu pool does not exist\n", 31004609245eSIlya Dryomov lreq->linger_id); 31014609245eSIlya Dryomov linger_reg_commit_complete(lreq, -ENOENT); 31024609245eSIlya Dryomov __linger_cancel(lreq); 31034609245eSIlya Dryomov } 31044609245eSIlya Dryomov } else { 31054609245eSIlya Dryomov send_linger_map_check(lreq); 31064609245eSIlya Dryomov } 31074609245eSIlya Dryomov } 31084609245eSIlya Dryomov 31094609245eSIlya Dryomov static void linger_map_check_cb(struct ceph_mon_generic_request *greq) 31104609245eSIlya Dryomov { 31114609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 31124609245eSIlya Dryomov struct ceph_osd_linger_request *lreq; 31134609245eSIlya Dryomov u64 linger_id = greq->private_data; 31144609245eSIlya Dryomov 31154609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 31164609245eSIlya Dryomov 31174609245eSIlya Dryomov down_write(&osdc->lock); 31184609245eSIlya Dryomov lreq = lookup_linger_mc(&osdc->linger_map_checks, linger_id); 31194609245eSIlya Dryomov if (!lreq) { 31204609245eSIlya Dryomov dout("%s linger_id %llu dne\n", __func__, linger_id); 31214609245eSIlya Dryomov goto out_unlock; 31224609245eSIlya Dryomov } 31234609245eSIlya Dryomov 31244609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u newest %llu\n", 31254609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 31264609245eSIlya Dryomov greq->u.newest); 31274609245eSIlya Dryomov if (!lreq->map_dne_bound) 31284609245eSIlya Dryomov lreq->map_dne_bound = greq->u.newest; 31294609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 31304609245eSIlya Dryomov check_linger_pool_dne(lreq); 31314609245eSIlya Dryomov 31324609245eSIlya Dryomov linger_put(lreq); 31334609245eSIlya Dryomov out_unlock: 31344609245eSIlya Dryomov up_write(&osdc->lock); 31354609245eSIlya Dryomov } 31364609245eSIlya Dryomov 31374609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq) 31384609245eSIlya Dryomov { 31394609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 31404609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 31414609245eSIlya Dryomov int ret; 31424609245eSIlya Dryomov 31434609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 31444609245eSIlya Dryomov 31454609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 31464609245eSIlya Dryomov lreq->linger_id); 31474609245eSIlya Dryomov if (lookup_lreq) { 31484609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 31494609245eSIlya Dryomov return; 31504609245eSIlya Dryomov } 31514609245eSIlya Dryomov 31524609245eSIlya Dryomov linger_get(lreq); 31534609245eSIlya Dryomov insert_linger_mc(&osdc->linger_map_checks, lreq); 31544609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 31554609245eSIlya Dryomov linger_map_check_cb, lreq->linger_id); 31564609245eSIlya Dryomov WARN_ON(ret); 31574609245eSIlya Dryomov } 31584609245eSIlya Dryomov 3159922dab61SIlya Dryomov static int linger_reg_commit_wait(struct ceph_osd_linger_request *lreq) 3160922dab61SIlya Dryomov { 3161922dab61SIlya Dryomov int ret; 3162922dab61SIlya Dryomov 3163922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 3164922dab61SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->reg_commit_wait); 3165922dab61SIlya Dryomov return ret ?: lreq->reg_commit_error; 3166922dab61SIlya Dryomov } 3167922dab61SIlya Dryomov 316819079203SIlya Dryomov static int linger_notify_finish_wait(struct ceph_osd_linger_request *lreq) 316919079203SIlya Dryomov { 317019079203SIlya Dryomov int ret; 317119079203SIlya Dryomov 317219079203SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 317319079203SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->notify_finish_wait); 317419079203SIlya Dryomov return ret ?: lreq->notify_finish_error; 317519079203SIlya Dryomov } 317619079203SIlya Dryomov 3177922dab61SIlya Dryomov /* 3178fbca9635SIlya Dryomov * Timeout callback, called every N seconds. When 1 or more OSD 3179fbca9635SIlya Dryomov * requests has been active for more than N seconds, we send a keepalive 3180fbca9635SIlya Dryomov * (tag + timestamp) to its OSD to ensure any communications channel 3181fbca9635SIlya Dryomov * reset is detected. 31823d14c5d2SYehuda Sadeh */ 31833d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 31843d14c5d2SYehuda Sadeh { 31853d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 31863d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 3187a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 31885aea3dcdSIlya Dryomov unsigned long cutoff = jiffies - opts->osd_keepalive_timeout; 31897cc5e38fSIlya Dryomov unsigned long expiry_cutoff = jiffies - opts->osd_request_timeout; 31905aea3dcdSIlya Dryomov LIST_HEAD(slow_osds); 31915aea3dcdSIlya Dryomov struct rb_node *n, *p; 31923d14c5d2SYehuda Sadeh 31935aea3dcdSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 31945aea3dcdSIlya Dryomov down_write(&osdc->lock); 31953d14c5d2SYehuda Sadeh 31963d14c5d2SYehuda Sadeh /* 31973d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 31983d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 31993d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 32003d14c5d2SYehuda Sadeh */ 32015aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 32025aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 32035aea3dcdSIlya Dryomov bool found = false; 32043d14c5d2SYehuda Sadeh 32057cc5e38fSIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 32065aea3dcdSIlya Dryomov struct ceph_osd_request *req = 32075aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 32085aea3dcdSIlya Dryomov 32097cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 32107cc5e38fSIlya Dryomov 32115aea3dcdSIlya Dryomov if (time_before(req->r_stamp, cutoff)) { 32125aea3dcdSIlya Dryomov dout(" req %p tid %llu on osd%d is laggy\n", 32135aea3dcdSIlya Dryomov req, req->r_tid, osd->o_osd); 32145aea3dcdSIlya Dryomov found = true; 32155aea3dcdSIlya Dryomov } 32167cc5e38fSIlya Dryomov if (opts->osd_request_timeout && 32177cc5e38fSIlya Dryomov time_before(req->r_start_stamp, expiry_cutoff)) { 32187cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 32197cc5e38fSIlya Dryomov req->r_tid, osd->o_osd); 32207cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 32217cc5e38fSIlya Dryomov } 32225aea3dcdSIlya Dryomov } 3223922dab61SIlya Dryomov for (p = rb_first(&osd->o_linger_requests); p; p = rb_next(p)) { 3224922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3225922dab61SIlya Dryomov rb_entry(p, struct ceph_osd_linger_request, node); 3226922dab61SIlya Dryomov 3227922dab61SIlya Dryomov dout(" lreq %p linger_id %llu is served by osd%d\n", 3228922dab61SIlya Dryomov lreq, lreq->linger_id, osd->o_osd); 3229922dab61SIlya Dryomov found = true; 3230922dab61SIlya Dryomov 3231922dab61SIlya Dryomov mutex_lock(&lreq->lock); 323219079203SIlya Dryomov if (lreq->is_watch && lreq->committed && !lreq->last_error) 3233922dab61SIlya Dryomov send_linger_ping(lreq); 3234922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3235922dab61SIlya Dryomov } 32365aea3dcdSIlya Dryomov 32375aea3dcdSIlya Dryomov if (found) 32383d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 32393d14c5d2SYehuda Sadeh } 32405aea3dcdSIlya Dryomov 32417cc5e38fSIlya Dryomov if (opts->osd_request_timeout) { 32427cc5e38fSIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 32437cc5e38fSIlya Dryomov struct ceph_osd_request *req = 32447cc5e38fSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 32457cc5e38fSIlya Dryomov 32467cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 32477cc5e38fSIlya Dryomov 32487cc5e38fSIlya Dryomov if (time_before(req->r_start_stamp, expiry_cutoff)) { 32497cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 32507cc5e38fSIlya Dryomov req->r_tid, osdc->homeless_osd.o_osd); 32517cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 32527cc5e38fSIlya Dryomov } 32537cc5e38fSIlya Dryomov } 32547cc5e38fSIlya Dryomov } 32557cc5e38fSIlya Dryomov 32565aea3dcdSIlya Dryomov if (atomic_read(&osdc->num_homeless) || !list_empty(&slow_osds)) 32575aea3dcdSIlya Dryomov maybe_request_map(osdc); 32585aea3dcdSIlya Dryomov 32593d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 32605aea3dcdSIlya Dryomov struct ceph_osd *osd = list_first_entry(&slow_osds, 32615aea3dcdSIlya Dryomov struct ceph_osd, 32623d14c5d2SYehuda Sadeh o_keepalive_item); 32633d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 32643d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 32653d14c5d2SYehuda Sadeh } 32663d14c5d2SYehuda Sadeh 32675aea3dcdSIlya Dryomov up_write(&osdc->lock); 3268fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 3269fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 32703d14c5d2SYehuda Sadeh } 32713d14c5d2SYehuda Sadeh 32723d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 32733d14c5d2SYehuda Sadeh { 32743d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 32753d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 32763d14c5d2SYehuda Sadeh osds_timeout_work.work); 3277a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 327842a2c09fSIlya Dryomov struct ceph_osd *osd, *nosd; 32793d14c5d2SYehuda Sadeh 328042a2c09fSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 32815aea3dcdSIlya Dryomov down_write(&osdc->lock); 328242a2c09fSIlya Dryomov list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 328342a2c09fSIlya Dryomov if (time_before(jiffies, osd->lru_ttl)) 328442a2c09fSIlya Dryomov break; 328542a2c09fSIlya Dryomov 32865aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 3287922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 32885aea3dcdSIlya Dryomov close_osd(osd); 328942a2c09fSIlya Dryomov } 329042a2c09fSIlya Dryomov 32915aea3dcdSIlya Dryomov up_write(&osdc->lock); 32923d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 32933d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 32943d14c5d2SYehuda Sadeh } 32953d14c5d2SYehuda Sadeh 3296205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 3297205ee118SIlya Dryomov struct ceph_object_locator *oloc) 3298205ee118SIlya Dryomov { 3299205ee118SIlya Dryomov u8 struct_v, struct_cv; 3300205ee118SIlya Dryomov u32 len; 3301205ee118SIlya Dryomov void *struct_end; 3302205ee118SIlya Dryomov int ret = 0; 3303205ee118SIlya Dryomov 3304205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3305205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3306205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3307205ee118SIlya Dryomov if (struct_v < 3) { 3308205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 3309205ee118SIlya Dryomov struct_v, struct_cv); 3310205ee118SIlya Dryomov goto e_inval; 3311205ee118SIlya Dryomov } 3312205ee118SIlya Dryomov if (struct_cv > 6) { 3313205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 3314205ee118SIlya Dryomov struct_v, struct_cv); 3315205ee118SIlya Dryomov goto e_inval; 3316205ee118SIlya Dryomov } 3317205ee118SIlya Dryomov len = ceph_decode_32(p); 3318205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3319205ee118SIlya Dryomov struct_end = *p + len; 3320205ee118SIlya Dryomov 3321205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 3322205ee118SIlya Dryomov *p += 4; /* skip preferred */ 3323205ee118SIlya Dryomov 3324205ee118SIlya Dryomov len = ceph_decode_32(p); 3325205ee118SIlya Dryomov if (len > 0) { 3326205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 3327205ee118SIlya Dryomov goto e_inval; 3328205ee118SIlya Dryomov } 3329205ee118SIlya Dryomov 3330205ee118SIlya Dryomov if (struct_v >= 5) { 3331cd08e0a2SYan, Zheng bool changed = false; 3332cd08e0a2SYan, Zheng 3333205ee118SIlya Dryomov len = ceph_decode_32(p); 3334205ee118SIlya Dryomov if (len > 0) { 333530c156d9SYan, Zheng ceph_decode_need(p, end, len, e_inval); 3336cd08e0a2SYan, Zheng if (!oloc->pool_ns || 3337cd08e0a2SYan, Zheng ceph_compare_string(oloc->pool_ns, *p, len)) 3338cd08e0a2SYan, Zheng changed = true; 333930c156d9SYan, Zheng *p += len; 3340cd08e0a2SYan, Zheng } else { 3341cd08e0a2SYan, Zheng if (oloc->pool_ns) 3342cd08e0a2SYan, Zheng changed = true; 3343cd08e0a2SYan, Zheng } 3344cd08e0a2SYan, Zheng if (changed) { 3345cd08e0a2SYan, Zheng /* redirect changes namespace */ 3346cd08e0a2SYan, Zheng pr_warn("ceph_object_locator::nspace is changed\n"); 3347cd08e0a2SYan, Zheng goto e_inval; 3348205ee118SIlya Dryomov } 3349205ee118SIlya Dryomov } 3350205ee118SIlya Dryomov 3351205ee118SIlya Dryomov if (struct_v >= 6) { 3352205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 3353205ee118SIlya Dryomov if (hash != -1) { 3354205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 3355205ee118SIlya Dryomov goto e_inval; 3356205ee118SIlya Dryomov } 3357205ee118SIlya Dryomov } 3358205ee118SIlya Dryomov 3359205ee118SIlya Dryomov /* skip the rest */ 3360205ee118SIlya Dryomov *p = struct_end; 3361205ee118SIlya Dryomov out: 3362205ee118SIlya Dryomov return ret; 3363205ee118SIlya Dryomov 3364205ee118SIlya Dryomov e_inval: 3365205ee118SIlya Dryomov ret = -EINVAL; 3366205ee118SIlya Dryomov goto out; 3367205ee118SIlya Dryomov } 3368205ee118SIlya Dryomov 3369205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 3370205ee118SIlya Dryomov struct ceph_request_redirect *redir) 3371205ee118SIlya Dryomov { 3372205ee118SIlya Dryomov u8 struct_v, struct_cv; 3373205ee118SIlya Dryomov u32 len; 3374205ee118SIlya Dryomov void *struct_end; 3375205ee118SIlya Dryomov int ret; 3376205ee118SIlya Dryomov 3377205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3378205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3379205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3380205ee118SIlya Dryomov if (struct_cv > 1) { 3381205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 3382205ee118SIlya Dryomov struct_v, struct_cv); 3383205ee118SIlya Dryomov goto e_inval; 3384205ee118SIlya Dryomov } 3385205ee118SIlya Dryomov len = ceph_decode_32(p); 3386205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3387205ee118SIlya Dryomov struct_end = *p + len; 3388205ee118SIlya Dryomov 3389205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 3390205ee118SIlya Dryomov if (ret) 3391205ee118SIlya Dryomov goto out; 3392205ee118SIlya Dryomov 3393205ee118SIlya Dryomov len = ceph_decode_32(p); 3394205ee118SIlya Dryomov if (len > 0) { 3395205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 3396205ee118SIlya Dryomov goto e_inval; 3397205ee118SIlya Dryomov } 3398205ee118SIlya Dryomov 3399205ee118SIlya Dryomov len = ceph_decode_32(p); 3400205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 3401205ee118SIlya Dryomov 3402205ee118SIlya Dryomov /* skip the rest */ 3403205ee118SIlya Dryomov *p = struct_end; 3404205ee118SIlya Dryomov out: 3405205ee118SIlya Dryomov return ret; 3406205ee118SIlya Dryomov 3407205ee118SIlya Dryomov e_inval: 3408205ee118SIlya Dryomov ret = -EINVAL; 3409205ee118SIlya Dryomov goto out; 3410205ee118SIlya Dryomov } 3411205ee118SIlya Dryomov 3412fe5da05eSIlya Dryomov struct MOSDOpReply { 3413fe5da05eSIlya Dryomov struct ceph_pg pgid; 3414fe5da05eSIlya Dryomov u64 flags; 3415fe5da05eSIlya Dryomov int result; 3416fe5da05eSIlya Dryomov u32 epoch; 3417fe5da05eSIlya Dryomov int num_ops; 3418fe5da05eSIlya Dryomov u32 outdata_len[CEPH_OSD_MAX_OPS]; 3419fe5da05eSIlya Dryomov s32 rval[CEPH_OSD_MAX_OPS]; 3420fe5da05eSIlya Dryomov int retry_attempt; 3421fe5da05eSIlya Dryomov struct ceph_eversion replay_version; 3422fe5da05eSIlya Dryomov u64 user_version; 3423fe5da05eSIlya Dryomov struct ceph_request_redirect redirect; 3424fe5da05eSIlya Dryomov }; 342525845472SSage Weil 3426fe5da05eSIlya Dryomov static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m) 34273d14c5d2SYehuda Sadeh { 3428fe5da05eSIlya Dryomov void *p = msg->front.iov_base; 3429fe5da05eSIlya Dryomov void *const end = p + msg->front.iov_len; 3430fe5da05eSIlya Dryomov u16 version = le16_to_cpu(msg->hdr.version); 3431fe5da05eSIlya Dryomov struct ceph_eversion bad_replay_version; 3432b0b31a8fSIlya Dryomov u8 decode_redir; 3433fe5da05eSIlya Dryomov u32 len; 3434fe5da05eSIlya Dryomov int ret; 3435fe5da05eSIlya Dryomov int i; 34363d14c5d2SYehuda Sadeh 3437fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, len, e_inval); 3438fe5da05eSIlya Dryomov ceph_decode_need(&p, end, len, e_inval); 3439fe5da05eSIlya Dryomov p += len; /* skip oid */ 34401b83bef2SSage Weil 3441fe5da05eSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->pgid); 3442fe5da05eSIlya Dryomov if (ret) 3443fe5da05eSIlya Dryomov return ret; 34441b83bef2SSage Weil 3445fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->flags, e_inval); 3446fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->result, e_inval); 3447fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval); 3448fe5da05eSIlya Dryomov memcpy(&bad_replay_version, p, sizeof(bad_replay_version)); 3449fe5da05eSIlya Dryomov p += sizeof(bad_replay_version); 3450fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->epoch, e_inval); 34511b83bef2SSage Weil 3452fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->num_ops, e_inval); 3453fe5da05eSIlya Dryomov if (m->num_ops > ARRAY_SIZE(m->outdata_len)) 3454fe5da05eSIlya Dryomov goto e_inval; 34551b83bef2SSage Weil 3456fe5da05eSIlya Dryomov ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op), 3457fe5da05eSIlya Dryomov e_inval); 3458fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) { 34591b83bef2SSage Weil struct ceph_osd_op *op = p; 34601b83bef2SSage Weil 3461fe5da05eSIlya Dryomov m->outdata_len[i] = le32_to_cpu(op->payload_len); 34621b83bef2SSage Weil p += sizeof(*op); 34631b83bef2SSage Weil } 3464fe5da05eSIlya Dryomov 3465fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval); 3466fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) 3467fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->rval[i], e_inval); 3468fe5da05eSIlya Dryomov 3469fe5da05eSIlya Dryomov if (version >= 5) { 3470fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval); 3471fe5da05eSIlya Dryomov memcpy(&m->replay_version, p, sizeof(m->replay_version)); 3472fe5da05eSIlya Dryomov p += sizeof(m->replay_version); 3473fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->user_version, e_inval); 3474fe5da05eSIlya Dryomov } else { 3475fe5da05eSIlya Dryomov m->replay_version = bad_replay_version; /* struct */ 3476fe5da05eSIlya Dryomov m->user_version = le64_to_cpu(m->replay_version.version); 34771b83bef2SSage Weil } 34781b83bef2SSage Weil 3479fe5da05eSIlya Dryomov if (version >= 6) { 3480fe5da05eSIlya Dryomov if (version >= 7) 3481fe5da05eSIlya Dryomov ceph_decode_8_safe(&p, end, decode_redir, e_inval); 3482b0b31a8fSIlya Dryomov else 3483b0b31a8fSIlya Dryomov decode_redir = 1; 3484b0b31a8fSIlya Dryomov } else { 3485b0b31a8fSIlya Dryomov decode_redir = 0; 3486b0b31a8fSIlya Dryomov } 3487b0b31a8fSIlya Dryomov 3488b0b31a8fSIlya Dryomov if (decode_redir) { 3489fe5da05eSIlya Dryomov ret = ceph_redirect_decode(&p, end, &m->redirect); 3490fe5da05eSIlya Dryomov if (ret) 3491fe5da05eSIlya Dryomov return ret; 3492205ee118SIlya Dryomov } else { 3493fe5da05eSIlya Dryomov ceph_oloc_init(&m->redirect.oloc); 3494205ee118SIlya Dryomov } 3495205ee118SIlya Dryomov 3496fe5da05eSIlya Dryomov return 0; 3497205ee118SIlya Dryomov 3498fe5da05eSIlya Dryomov e_inval: 3499fe5da05eSIlya Dryomov return -EINVAL; 3500fe5da05eSIlya Dryomov } 3501fe5da05eSIlya Dryomov 3502fe5da05eSIlya Dryomov /* 3503b18b9550SIlya Dryomov * Handle MOSDOpReply. Set ->r_result and call the callback if it is 3504b18b9550SIlya Dryomov * specified. 3505fe5da05eSIlya Dryomov */ 35065aea3dcdSIlya Dryomov static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg) 3507fe5da05eSIlya Dryomov { 35085aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 3509fe5da05eSIlya Dryomov struct ceph_osd_request *req; 3510fe5da05eSIlya Dryomov struct MOSDOpReply m; 3511fe5da05eSIlya Dryomov u64 tid = le64_to_cpu(msg->hdr.tid); 3512fe5da05eSIlya Dryomov u32 data_len = 0; 3513fe5da05eSIlya Dryomov int ret; 3514fe5da05eSIlya Dryomov int i; 3515fe5da05eSIlya Dryomov 3516fe5da05eSIlya Dryomov dout("%s msg %p tid %llu\n", __func__, msg, tid); 3517fe5da05eSIlya Dryomov 35185aea3dcdSIlya Dryomov down_read(&osdc->lock); 35195aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 35205aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 35215aea3dcdSIlya Dryomov goto out_unlock_osdc; 3522fe5da05eSIlya Dryomov } 35235aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 35245aea3dcdSIlya Dryomov 35255aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 35265aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 35275aea3dcdSIlya Dryomov if (!req) { 35285aea3dcdSIlya Dryomov dout("%s osd%d tid %llu unknown\n", __func__, osd->o_osd, tid); 35295aea3dcdSIlya Dryomov goto out_unlock_session; 35305aea3dcdSIlya Dryomov } 3531fe5da05eSIlya Dryomov 3532cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = req->r_t.target_oloc.pool_ns; 3533fe5da05eSIlya Dryomov ret = decode_MOSDOpReply(msg, &m); 3534cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = NULL; 3535fe5da05eSIlya Dryomov if (ret) { 3536fe5da05eSIlya Dryomov pr_err("failed to decode MOSDOpReply for tid %llu: %d\n", 3537fe5da05eSIlya Dryomov req->r_tid, ret); 3538fe5da05eSIlya Dryomov ceph_msg_dump(msg); 3539fe5da05eSIlya Dryomov goto fail_request; 3540fe5da05eSIlya Dryomov } 3541fe5da05eSIlya Dryomov dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n", 3542fe5da05eSIlya Dryomov __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed, 3543fe5da05eSIlya Dryomov m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch), 3544fe5da05eSIlya Dryomov le64_to_cpu(m.replay_version.version), m.user_version); 3545fe5da05eSIlya Dryomov 3546fe5da05eSIlya Dryomov if (m.retry_attempt >= 0) { 3547fe5da05eSIlya Dryomov if (m.retry_attempt != req->r_attempts - 1) { 3548fe5da05eSIlya Dryomov dout("req %p tid %llu retry_attempt %d != %d, ignoring\n", 3549fe5da05eSIlya Dryomov req, req->r_tid, m.retry_attempt, 3550fe5da05eSIlya Dryomov req->r_attempts - 1); 35515aea3dcdSIlya Dryomov goto out_unlock_session; 3552fe5da05eSIlya Dryomov } 3553fe5da05eSIlya Dryomov } else { 3554fe5da05eSIlya Dryomov WARN_ON(1); /* MOSDOpReply v4 is assumed */ 3555fe5da05eSIlya Dryomov } 3556fe5da05eSIlya Dryomov 3557fe5da05eSIlya Dryomov if (!ceph_oloc_empty(&m.redirect.oloc)) { 3558fe5da05eSIlya Dryomov dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid, 3559fe5da05eSIlya Dryomov m.redirect.oloc.pool); 35605aea3dcdSIlya Dryomov unlink_request(osd, req); 35615aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 3562205ee118SIlya Dryomov 356330c156d9SYan, Zheng /* 356430c156d9SYan, Zheng * Not ceph_oloc_copy() - changing pool_ns is not 356530c156d9SYan, Zheng * supported. 356630c156d9SYan, Zheng */ 356730c156d9SYan, Zheng req->r_t.target_oloc.pool = m.redirect.oloc.pool; 35685aea3dcdSIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_REDIRECTED; 35695aea3dcdSIlya Dryomov req->r_tid = 0; 35705aea3dcdSIlya Dryomov __submit_request(req, false); 35715aea3dcdSIlya Dryomov goto out_unlock_osdc; 3572205ee118SIlya Dryomov } 3573205ee118SIlya Dryomov 3574fe5da05eSIlya Dryomov if (m.num_ops != req->r_num_ops) { 3575fe5da05eSIlya Dryomov pr_err("num_ops %d != %d for tid %llu\n", m.num_ops, 3576fe5da05eSIlya Dryomov req->r_num_ops, req->r_tid); 3577fe5da05eSIlya Dryomov goto fail_request; 3578fe5da05eSIlya Dryomov } 3579fe5da05eSIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 3580fe5da05eSIlya Dryomov dout(" req %p tid %llu op %d rval %d len %u\n", req, 3581fe5da05eSIlya Dryomov req->r_tid, i, m.rval[i], m.outdata_len[i]); 3582fe5da05eSIlya Dryomov req->r_ops[i].rval = m.rval[i]; 3583fe5da05eSIlya Dryomov req->r_ops[i].outdata_len = m.outdata_len[i]; 3584fe5da05eSIlya Dryomov data_len += m.outdata_len[i]; 3585fe5da05eSIlya Dryomov } 3586fe5da05eSIlya Dryomov if (data_len != le32_to_cpu(msg->hdr.data_len)) { 3587fe5da05eSIlya Dryomov pr_err("sum of lens %u != %u for tid %llu\n", data_len, 3588fe5da05eSIlya Dryomov le32_to_cpu(msg->hdr.data_len), req->r_tid); 3589fe5da05eSIlya Dryomov goto fail_request; 3590fe5da05eSIlya Dryomov } 3591b18b9550SIlya Dryomov dout("%s req %p tid %llu result %d data_len %u\n", __func__, 3592b18b9550SIlya Dryomov req, req->r_tid, m.result, data_len); 35933d14c5d2SYehuda Sadeh 3594b18b9550SIlya Dryomov /* 3595b18b9550SIlya Dryomov * Since we only ever request ONDISK, we should only ever get 3596b18b9550SIlya Dryomov * one (type of) reply back. 3597b18b9550SIlya Dryomov */ 3598b18b9550SIlya Dryomov WARN_ON(!(m.flags & CEPH_OSD_FLAG_ONDISK)); 3599fe5da05eSIlya Dryomov req->r_result = m.result ?: data_len; 360045ee2c1dSIlya Dryomov finish_request(req); 36015aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 36025aea3dcdSIlya Dryomov up_read(&osdc->lock); 36033d14c5d2SYehuda Sadeh 3604fe5da05eSIlya Dryomov __complete_request(req); 36053d14c5d2SYehuda Sadeh return; 3606fe5da05eSIlya Dryomov 3607fe5da05eSIlya Dryomov fail_request: 36084609245eSIlya Dryomov complete_request(req, -EIO); 36095aea3dcdSIlya Dryomov out_unlock_session: 36105aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 36115aea3dcdSIlya Dryomov out_unlock_osdc: 36125aea3dcdSIlya Dryomov up_read(&osdc->lock); 36133d14c5d2SYehuda Sadeh } 36143d14c5d2SYehuda Sadeh 361542c1b124SIlya Dryomov static void set_pool_was_full(struct ceph_osd_client *osdc) 361642c1b124SIlya Dryomov { 361742c1b124SIlya Dryomov struct rb_node *n; 361842c1b124SIlya Dryomov 361942c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 362042c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 362142c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 362242c1b124SIlya Dryomov 362342c1b124SIlya Dryomov pi->was_full = __pool_full(pi); 362442c1b124SIlya Dryomov } 362542c1b124SIlya Dryomov } 362642c1b124SIlya Dryomov 36275aea3dcdSIlya Dryomov static bool pool_cleared_full(struct ceph_osd_client *osdc, s64 pool_id) 36283d14c5d2SYehuda Sadeh { 36295aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 36303d14c5d2SYehuda Sadeh 36315aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 36325aea3dcdSIlya Dryomov if (!pi) 36335aea3dcdSIlya Dryomov return false; 36343d14c5d2SYehuda Sadeh 36355aea3dcdSIlya Dryomov return pi->was_full && !__pool_full(pi); 36363d14c5d2SYehuda Sadeh } 36373d14c5d2SYehuda Sadeh 3638922dab61SIlya Dryomov static enum calc_target_result 3639922dab61SIlya Dryomov recalc_linger_target(struct ceph_osd_linger_request *lreq) 3640922dab61SIlya Dryomov { 3641922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3642922dab61SIlya Dryomov enum calc_target_result ct_res; 3643922dab61SIlya Dryomov 36447de030d6SIlya Dryomov ct_res = calc_target(osdc, &lreq->t, NULL, true); 3645922dab61SIlya Dryomov if (ct_res == CALC_TARGET_NEED_RESEND) { 3646922dab61SIlya Dryomov struct ceph_osd *osd; 3647922dab61SIlya Dryomov 3648922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3649922dab61SIlya Dryomov if (osd != lreq->osd) { 3650922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3651922dab61SIlya Dryomov link_linger(osd, lreq); 3652922dab61SIlya Dryomov } 3653922dab61SIlya Dryomov } 3654922dab61SIlya Dryomov 3655922dab61SIlya Dryomov return ct_res; 3656922dab61SIlya Dryomov } 3657922dab61SIlya Dryomov 36583d14c5d2SYehuda Sadeh /* 36595aea3dcdSIlya Dryomov * Requeue requests whose mapping to an OSD has changed. 36603d14c5d2SYehuda Sadeh */ 36615aea3dcdSIlya Dryomov static void scan_requests(struct ceph_osd *osd, 36625aea3dcdSIlya Dryomov bool force_resend, 36635aea3dcdSIlya Dryomov bool cleared_full, 36645aea3dcdSIlya Dryomov bool check_pool_cleared_full, 36655aea3dcdSIlya Dryomov struct rb_root *need_resend, 36665aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 36673d14c5d2SYehuda Sadeh { 36685aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 36695aea3dcdSIlya Dryomov struct rb_node *n; 36705aea3dcdSIlya Dryomov bool force_resend_writes; 36713d14c5d2SYehuda Sadeh 3672922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 3673922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3674922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 3675922dab61SIlya Dryomov enum calc_target_result ct_res; 3676922dab61SIlya Dryomov 3677922dab61SIlya Dryomov n = rb_next(n); /* recalc_linger_target() */ 3678922dab61SIlya Dryomov 3679922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, 3680922dab61SIlya Dryomov lreq->linger_id); 3681922dab61SIlya Dryomov ct_res = recalc_linger_target(lreq); 3682922dab61SIlya Dryomov switch (ct_res) { 3683922dab61SIlya Dryomov case CALC_TARGET_NO_ACTION: 3684922dab61SIlya Dryomov force_resend_writes = cleared_full || 3685922dab61SIlya Dryomov (check_pool_cleared_full && 3686922dab61SIlya Dryomov pool_cleared_full(osdc, lreq->t.base_oloc.pool)); 3687922dab61SIlya Dryomov if (!force_resend && !force_resend_writes) 3688922dab61SIlya Dryomov break; 3689922dab61SIlya Dryomov 3690922dab61SIlya Dryomov /* fall through */ 3691922dab61SIlya Dryomov case CALC_TARGET_NEED_RESEND: 36924609245eSIlya Dryomov cancel_linger_map_check(lreq); 3693922dab61SIlya Dryomov /* 3694922dab61SIlya Dryomov * scan_requests() for the previous epoch(s) 3695922dab61SIlya Dryomov * may have already added it to the list, since 3696922dab61SIlya Dryomov * it's not unlinked here. 3697922dab61SIlya Dryomov */ 3698922dab61SIlya Dryomov if (list_empty(&lreq->scan_item)) 3699922dab61SIlya Dryomov list_add_tail(&lreq->scan_item, need_resend_linger); 3700922dab61SIlya Dryomov break; 3701922dab61SIlya Dryomov case CALC_TARGET_POOL_DNE: 3702a10bcb19SIlya Dryomov list_del_init(&lreq->scan_item); 37034609245eSIlya Dryomov check_linger_pool_dne(lreq); 3704922dab61SIlya Dryomov break; 3705922dab61SIlya Dryomov } 3706922dab61SIlya Dryomov } 3707922dab61SIlya Dryomov 37085aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 37095aea3dcdSIlya Dryomov struct ceph_osd_request *req = 37105aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 37115aea3dcdSIlya Dryomov enum calc_target_result ct_res; 3712ab60b16dSAlex Elder 37134609245eSIlya Dryomov n = rb_next(n); /* unlink_request(), check_pool_dne() */ 3714ab60b16dSAlex Elder 37155aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 37167de030d6SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, &req->r_osd->o_con, 37177de030d6SIlya Dryomov false); 37185aea3dcdSIlya Dryomov switch (ct_res) { 37195aea3dcdSIlya Dryomov case CALC_TARGET_NO_ACTION: 37205aea3dcdSIlya Dryomov force_resend_writes = cleared_full || 37215aea3dcdSIlya Dryomov (check_pool_cleared_full && 37225aea3dcdSIlya Dryomov pool_cleared_full(osdc, req->r_t.base_oloc.pool)); 37235aea3dcdSIlya Dryomov if (!force_resend && 37245aea3dcdSIlya Dryomov (!(req->r_flags & CEPH_OSD_FLAG_WRITE) || 37255aea3dcdSIlya Dryomov !force_resend_writes)) 37265aea3dcdSIlya Dryomov break; 3727a40c4f10SYehuda Sadeh 37285aea3dcdSIlya Dryomov /* fall through */ 37295aea3dcdSIlya Dryomov case CALC_TARGET_NEED_RESEND: 37304609245eSIlya Dryomov cancel_map_check(req); 37315aea3dcdSIlya Dryomov unlink_request(osd, req); 37325aea3dcdSIlya Dryomov insert_request(need_resend, req); 37335aea3dcdSIlya Dryomov break; 37345aea3dcdSIlya Dryomov case CALC_TARGET_POOL_DNE: 37354609245eSIlya Dryomov check_pool_dne(req); 37365aea3dcdSIlya Dryomov break; 3737a40c4f10SYehuda Sadeh } 37383d14c5d2SYehuda Sadeh } 37393d14c5d2SYehuda Sadeh } 37406f6c7006SSage Weil 374142c1b124SIlya Dryomov static int handle_one_map(struct ceph_osd_client *osdc, 37425aea3dcdSIlya Dryomov void *p, void *end, bool incremental, 37435aea3dcdSIlya Dryomov struct rb_root *need_resend, 37445aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 374542c1b124SIlya Dryomov { 374642c1b124SIlya Dryomov struct ceph_osdmap *newmap; 374742c1b124SIlya Dryomov struct rb_node *n; 374842c1b124SIlya Dryomov bool skipped_map = false; 374942c1b124SIlya Dryomov bool was_full; 375042c1b124SIlya Dryomov 3751b7ec35b3SIlya Dryomov was_full = ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 375242c1b124SIlya Dryomov set_pool_was_full(osdc); 375342c1b124SIlya Dryomov 375442c1b124SIlya Dryomov if (incremental) 375542c1b124SIlya Dryomov newmap = osdmap_apply_incremental(&p, end, osdc->osdmap); 375642c1b124SIlya Dryomov else 375742c1b124SIlya Dryomov newmap = ceph_osdmap_decode(&p, end); 375842c1b124SIlya Dryomov if (IS_ERR(newmap)) 375942c1b124SIlya Dryomov return PTR_ERR(newmap); 376042c1b124SIlya Dryomov 376142c1b124SIlya Dryomov if (newmap != osdc->osdmap) { 376242c1b124SIlya Dryomov /* 376342c1b124SIlya Dryomov * Preserve ->was_full before destroying the old map. 376442c1b124SIlya Dryomov * For pools that weren't in the old map, ->was_full 376542c1b124SIlya Dryomov * should be false. 376642c1b124SIlya Dryomov */ 376742c1b124SIlya Dryomov for (n = rb_first(&newmap->pg_pools); n; n = rb_next(n)) { 376842c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 376942c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 377042c1b124SIlya Dryomov struct ceph_pg_pool_info *old_pi; 377142c1b124SIlya Dryomov 377242c1b124SIlya Dryomov old_pi = ceph_pg_pool_by_id(osdc->osdmap, pi->id); 377342c1b124SIlya Dryomov if (old_pi) 377442c1b124SIlya Dryomov pi->was_full = old_pi->was_full; 377542c1b124SIlya Dryomov else 377642c1b124SIlya Dryomov WARN_ON(pi->was_full); 377742c1b124SIlya Dryomov } 377842c1b124SIlya Dryomov 377942c1b124SIlya Dryomov if (osdc->osdmap->epoch && 378042c1b124SIlya Dryomov osdc->osdmap->epoch + 1 < newmap->epoch) { 378142c1b124SIlya Dryomov WARN_ON(incremental); 378242c1b124SIlya Dryomov skipped_map = true; 378342c1b124SIlya Dryomov } 378442c1b124SIlya Dryomov 378542c1b124SIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 378642c1b124SIlya Dryomov osdc->osdmap = newmap; 378742c1b124SIlya Dryomov } 378842c1b124SIlya Dryomov 3789b7ec35b3SIlya Dryomov was_full &= !ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 37905aea3dcdSIlya Dryomov scan_requests(&osdc->homeless_osd, skipped_map, was_full, true, 37915aea3dcdSIlya Dryomov need_resend, need_resend_linger); 37925aea3dcdSIlya Dryomov 37935aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; ) { 37945aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 37955aea3dcdSIlya Dryomov 37965aea3dcdSIlya Dryomov n = rb_next(n); /* close_osd() */ 37975aea3dcdSIlya Dryomov 37985aea3dcdSIlya Dryomov scan_requests(osd, skipped_map, was_full, true, need_resend, 37995aea3dcdSIlya Dryomov need_resend_linger); 38005aea3dcdSIlya Dryomov if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 38015aea3dcdSIlya Dryomov memcmp(&osd->o_con.peer_addr, 38025aea3dcdSIlya Dryomov ceph_osd_addr(osdc->osdmap, osd->o_osd), 38035aea3dcdSIlya Dryomov sizeof(struct ceph_entity_addr))) 38045aea3dcdSIlya Dryomov close_osd(osd); 38055aea3dcdSIlya Dryomov } 380642c1b124SIlya Dryomov 380742c1b124SIlya Dryomov return 0; 380842c1b124SIlya Dryomov } 38096f6c7006SSage Weil 38105aea3dcdSIlya Dryomov static void kick_requests(struct ceph_osd_client *osdc, 38115aea3dcdSIlya Dryomov struct rb_root *need_resend, 38125aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 38135aea3dcdSIlya Dryomov { 3814922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq, *nlreq; 381504c7d789SIlya Dryomov enum calc_target_result ct_res; 38165aea3dcdSIlya Dryomov struct rb_node *n; 38175aea3dcdSIlya Dryomov 381804c7d789SIlya Dryomov /* make sure need_resend targets reflect latest map */ 381904c7d789SIlya Dryomov for (n = rb_first(need_resend); n; ) { 382004c7d789SIlya Dryomov struct ceph_osd_request *req = 382104c7d789SIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 382204c7d789SIlya Dryomov 382304c7d789SIlya Dryomov n = rb_next(n); 382404c7d789SIlya Dryomov 382504c7d789SIlya Dryomov if (req->r_t.epoch < osdc->osdmap->epoch) { 382604c7d789SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, NULL, false); 382704c7d789SIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE) { 382804c7d789SIlya Dryomov erase_request(need_resend, req); 382904c7d789SIlya Dryomov check_pool_dne(req); 383004c7d789SIlya Dryomov } 383104c7d789SIlya Dryomov } 383204c7d789SIlya Dryomov } 383304c7d789SIlya Dryomov 38345aea3dcdSIlya Dryomov for (n = rb_first(need_resend); n; ) { 38355aea3dcdSIlya Dryomov struct ceph_osd_request *req = 38365aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 38375aea3dcdSIlya Dryomov struct ceph_osd *osd; 38385aea3dcdSIlya Dryomov 38395aea3dcdSIlya Dryomov n = rb_next(n); 38405aea3dcdSIlya Dryomov erase_request(need_resend, req); /* before link_request() */ 38415aea3dcdSIlya Dryomov 38425aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, true); 38435aea3dcdSIlya Dryomov link_request(osd, req); 38445aea3dcdSIlya Dryomov if (!req->r_linger) { 38455aea3dcdSIlya Dryomov if (!osd_homeless(osd) && !req->r_t.paused) 38465aea3dcdSIlya Dryomov send_request(req); 3847922dab61SIlya Dryomov } else { 3848922dab61SIlya Dryomov cancel_linger_request(req); 38495aea3dcdSIlya Dryomov } 38505aea3dcdSIlya Dryomov } 3851922dab61SIlya Dryomov 3852922dab61SIlya Dryomov list_for_each_entry_safe(lreq, nlreq, need_resend_linger, scan_item) { 3853922dab61SIlya Dryomov if (!osd_homeless(lreq->osd)) 3854922dab61SIlya Dryomov send_linger(lreq); 3855922dab61SIlya Dryomov 3856922dab61SIlya Dryomov list_del_init(&lreq->scan_item); 3857922dab61SIlya Dryomov } 38585aea3dcdSIlya Dryomov } 38595aea3dcdSIlya Dryomov 38603d14c5d2SYehuda Sadeh /* 38613d14c5d2SYehuda Sadeh * Process updated osd map. 38623d14c5d2SYehuda Sadeh * 38633d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 38643d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 38653d14c5d2SYehuda Sadeh * off to different OSDs as needed. 38663d14c5d2SYehuda Sadeh */ 38673d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 38683d14c5d2SYehuda Sadeh { 386942c1b124SIlya Dryomov void *p = msg->front.iov_base; 387042c1b124SIlya Dryomov void *const end = p + msg->front.iov_len; 38713d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 38723d14c5d2SYehuda Sadeh u32 epoch; 38733d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 38745aea3dcdSIlya Dryomov struct rb_root need_resend = RB_ROOT; 38755aea3dcdSIlya Dryomov LIST_HEAD(need_resend_linger); 387642c1b124SIlya Dryomov bool handled_incremental = false; 387742c1b124SIlya Dryomov bool was_pauserd, was_pausewr; 387842c1b124SIlya Dryomov bool pauserd, pausewr; 387942c1b124SIlya Dryomov int err; 38803d14c5d2SYehuda Sadeh 388142c1b124SIlya Dryomov dout("%s have %u\n", __func__, osdc->osdmap->epoch); 38825aea3dcdSIlya Dryomov down_write(&osdc->lock); 38833d14c5d2SYehuda Sadeh 38843d14c5d2SYehuda Sadeh /* verify fsid */ 38853d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 38863d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 38873d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 388842c1b124SIlya Dryomov goto bad; 38893d14c5d2SYehuda Sadeh 3890b7ec35b3SIlya Dryomov was_pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 3891b7ec35b3SIlya Dryomov was_pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 3892b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 389342c1b124SIlya Dryomov have_pool_full(osdc); 38949a1ea2dbSJosh Durgin 38953d14c5d2SYehuda Sadeh /* incremental maps */ 38963d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 38973d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 38983d14c5d2SYehuda Sadeh while (nr_maps > 0) { 38993d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 39003d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 39013d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 39023d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 390342c1b124SIlya Dryomov if (osdc->osdmap->epoch && 390442c1b124SIlya Dryomov osdc->osdmap->epoch + 1 == epoch) { 39053d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 39063d14c5d2SYehuda Sadeh epoch, maplen); 39075aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, true, 39085aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 390942c1b124SIlya Dryomov if (err) 39103d14c5d2SYehuda Sadeh goto bad; 391142c1b124SIlya Dryomov handled_incremental = true; 39123d14c5d2SYehuda Sadeh } else { 39133d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 39143d14c5d2SYehuda Sadeh epoch, maplen); 39153d14c5d2SYehuda Sadeh } 391642c1b124SIlya Dryomov p += maplen; 39173d14c5d2SYehuda Sadeh nr_maps--; 39183d14c5d2SYehuda Sadeh } 391942c1b124SIlya Dryomov if (handled_incremental) 39203d14c5d2SYehuda Sadeh goto done; 39213d14c5d2SYehuda Sadeh 39223d14c5d2SYehuda Sadeh /* full maps */ 39233d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 39243d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 39253d14c5d2SYehuda Sadeh while (nr_maps) { 39263d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 39273d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 39283d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 39293d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 39303d14c5d2SYehuda Sadeh if (nr_maps > 1) { 39313d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 39323d14c5d2SYehuda Sadeh epoch, maplen); 3933e5253a7bSIlya Dryomov } else if (osdc->osdmap->epoch >= epoch) { 39343d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 39353d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 39363d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 39373d14c5d2SYehuda Sadeh } else { 39383d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 39395aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, false, 39405aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 394142c1b124SIlya Dryomov if (err) 39423d14c5d2SYehuda Sadeh goto bad; 39433d14c5d2SYehuda Sadeh } 39443d14c5d2SYehuda Sadeh p += maplen; 39453d14c5d2SYehuda Sadeh nr_maps--; 39463d14c5d2SYehuda Sadeh } 39473d14c5d2SYehuda Sadeh 39483d14c5d2SYehuda Sadeh done: 3949cd634fb6SSage Weil /* 3950cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 3951cd634fb6SSage Weil * we find out when we are no longer full and stop returning 3952cd634fb6SSage Weil * ENOSPC. 3953cd634fb6SSage Weil */ 3954b7ec35b3SIlya Dryomov pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 3955b7ec35b3SIlya Dryomov pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 3956b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 395742c1b124SIlya Dryomov have_pool_full(osdc); 395858eb7932SJeff Layton if (was_pauserd || was_pausewr || pauserd || pausewr || 395958eb7932SJeff Layton osdc->osdmap->epoch < osdc->epoch_barrier) 396042c1b124SIlya Dryomov maybe_request_map(osdc); 3961cd634fb6SSage Weil 39625aea3dcdSIlya Dryomov kick_requests(osdc, &need_resend, &need_resend_linger); 396342c1b124SIlya Dryomov 3964fc36d0a4SJeff Layton ceph_osdc_abort_on_full(osdc); 396542c1b124SIlya Dryomov ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 396642c1b124SIlya Dryomov osdc->osdmap->epoch); 39675aea3dcdSIlya Dryomov up_write(&osdc->lock); 39683d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 39693d14c5d2SYehuda Sadeh return; 39703d14c5d2SYehuda Sadeh 39713d14c5d2SYehuda Sadeh bad: 39723d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 39733d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 39745aea3dcdSIlya Dryomov up_write(&osdc->lock); 39755aea3dcdSIlya Dryomov } 39765aea3dcdSIlya Dryomov 39775aea3dcdSIlya Dryomov /* 39785aea3dcdSIlya Dryomov * Resubmit requests pending on the given osd. 39795aea3dcdSIlya Dryomov */ 39805aea3dcdSIlya Dryomov static void kick_osd_requests(struct ceph_osd *osd) 39815aea3dcdSIlya Dryomov { 39825aea3dcdSIlya Dryomov struct rb_node *n; 39835aea3dcdSIlya Dryomov 3984a02a946dSIlya Dryomov clear_backoffs(osd); 3985a02a946dSIlya Dryomov 3986922dab61SIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 39875aea3dcdSIlya Dryomov struct ceph_osd_request *req = 39885aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 39895aea3dcdSIlya Dryomov 3990922dab61SIlya Dryomov n = rb_next(n); /* cancel_linger_request() */ 3991922dab61SIlya Dryomov 39925aea3dcdSIlya Dryomov if (!req->r_linger) { 39935aea3dcdSIlya Dryomov if (!req->r_t.paused) 39945aea3dcdSIlya Dryomov send_request(req); 3995922dab61SIlya Dryomov } else { 3996922dab61SIlya Dryomov cancel_linger_request(req); 39975aea3dcdSIlya Dryomov } 39985aea3dcdSIlya Dryomov } 3999922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; n = rb_next(n)) { 4000922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 4001922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 4002922dab61SIlya Dryomov 4003922dab61SIlya Dryomov send_linger(lreq); 4004922dab61SIlya Dryomov } 40055aea3dcdSIlya Dryomov } 40065aea3dcdSIlya Dryomov 40075aea3dcdSIlya Dryomov /* 40085aea3dcdSIlya Dryomov * If the osd connection drops, we need to resubmit all requests. 40095aea3dcdSIlya Dryomov */ 40105aea3dcdSIlya Dryomov static void osd_fault(struct ceph_connection *con) 40115aea3dcdSIlya Dryomov { 40125aea3dcdSIlya Dryomov struct ceph_osd *osd = con->private; 40135aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 40145aea3dcdSIlya Dryomov 40155aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 40165aea3dcdSIlya Dryomov 40175aea3dcdSIlya Dryomov down_write(&osdc->lock); 40185aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 40195aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 40205aea3dcdSIlya Dryomov goto out_unlock; 40215aea3dcdSIlya Dryomov } 40225aea3dcdSIlya Dryomov 40235aea3dcdSIlya Dryomov if (!reopen_osd(osd)) 40245aea3dcdSIlya Dryomov kick_osd_requests(osd); 40255aea3dcdSIlya Dryomov maybe_request_map(osdc); 40265aea3dcdSIlya Dryomov 40275aea3dcdSIlya Dryomov out_unlock: 40285aea3dcdSIlya Dryomov up_write(&osdc->lock); 40293d14c5d2SYehuda Sadeh } 40303d14c5d2SYehuda Sadeh 4031a02a946dSIlya Dryomov struct MOSDBackoff { 4032a02a946dSIlya Dryomov struct ceph_spg spgid; 4033a02a946dSIlya Dryomov u32 map_epoch; 4034a02a946dSIlya Dryomov u8 op; 4035a02a946dSIlya Dryomov u64 id; 4036a02a946dSIlya Dryomov struct ceph_hobject_id *begin; 4037a02a946dSIlya Dryomov struct ceph_hobject_id *end; 4038a02a946dSIlya Dryomov }; 4039a02a946dSIlya Dryomov 4040a02a946dSIlya Dryomov static int decode_MOSDBackoff(const struct ceph_msg *msg, struct MOSDBackoff *m) 4041a02a946dSIlya Dryomov { 4042a02a946dSIlya Dryomov void *p = msg->front.iov_base; 4043a02a946dSIlya Dryomov void *const end = p + msg->front.iov_len; 4044a02a946dSIlya Dryomov u8 struct_v; 4045a02a946dSIlya Dryomov u32 struct_len; 4046a02a946dSIlya Dryomov int ret; 4047a02a946dSIlya Dryomov 4048a02a946dSIlya Dryomov ret = ceph_start_decoding(&p, end, 1, "spg_t", &struct_v, &struct_len); 4049a02a946dSIlya Dryomov if (ret) 4050a02a946dSIlya Dryomov return ret; 4051a02a946dSIlya Dryomov 4052a02a946dSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->spgid.pgid); 4053a02a946dSIlya Dryomov if (ret) 4054a02a946dSIlya Dryomov return ret; 4055a02a946dSIlya Dryomov 4056a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->spgid.shard, e_inval); 4057a02a946dSIlya Dryomov ceph_decode_32_safe(&p, end, m->map_epoch, e_inval); 4058a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->op, e_inval); 4059a02a946dSIlya Dryomov ceph_decode_64_safe(&p, end, m->id, e_inval); 4060a02a946dSIlya Dryomov 4061a02a946dSIlya Dryomov m->begin = kzalloc(sizeof(*m->begin), GFP_NOIO); 4062a02a946dSIlya Dryomov if (!m->begin) 4063a02a946dSIlya Dryomov return -ENOMEM; 4064a02a946dSIlya Dryomov 4065a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->begin); 4066a02a946dSIlya Dryomov if (ret) { 4067a02a946dSIlya Dryomov free_hoid(m->begin); 4068a02a946dSIlya Dryomov return ret; 4069a02a946dSIlya Dryomov } 4070a02a946dSIlya Dryomov 4071a02a946dSIlya Dryomov m->end = kzalloc(sizeof(*m->end), GFP_NOIO); 4072a02a946dSIlya Dryomov if (!m->end) { 4073a02a946dSIlya Dryomov free_hoid(m->begin); 4074a02a946dSIlya Dryomov return -ENOMEM; 4075a02a946dSIlya Dryomov } 4076a02a946dSIlya Dryomov 4077a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->end); 4078a02a946dSIlya Dryomov if (ret) { 4079a02a946dSIlya Dryomov free_hoid(m->begin); 4080a02a946dSIlya Dryomov free_hoid(m->end); 4081a02a946dSIlya Dryomov return ret; 4082a02a946dSIlya Dryomov } 4083a02a946dSIlya Dryomov 4084a02a946dSIlya Dryomov return 0; 4085a02a946dSIlya Dryomov 4086a02a946dSIlya Dryomov e_inval: 4087a02a946dSIlya Dryomov return -EINVAL; 4088a02a946dSIlya Dryomov } 4089a02a946dSIlya Dryomov 4090a02a946dSIlya Dryomov static struct ceph_msg *create_backoff_message( 4091a02a946dSIlya Dryomov const struct ceph_osd_backoff *backoff, 4092a02a946dSIlya Dryomov u32 map_epoch) 4093a02a946dSIlya Dryomov { 4094a02a946dSIlya Dryomov struct ceph_msg *msg; 4095a02a946dSIlya Dryomov void *p, *end; 4096a02a946dSIlya Dryomov int msg_size; 4097a02a946dSIlya Dryomov 4098a02a946dSIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 4099a02a946dSIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 4100a02a946dSIlya Dryomov msg_size += 4 + 1 + 8; /* map_epoch, op, id */ 4101a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4102a02a946dSIlya Dryomov hoid_encoding_size(backoff->begin); 4103a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4104a02a946dSIlya Dryomov hoid_encoding_size(backoff->end); 4105a02a946dSIlya Dryomov 4106a02a946dSIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_BACKOFF, msg_size, GFP_NOIO, true); 4107a02a946dSIlya Dryomov if (!msg) 4108a02a946dSIlya Dryomov return NULL; 4109a02a946dSIlya Dryomov 4110a02a946dSIlya Dryomov p = msg->front.iov_base; 4111a02a946dSIlya Dryomov end = p + msg->front_alloc_len; 4112a02a946dSIlya Dryomov 4113a02a946dSIlya Dryomov encode_spgid(&p, &backoff->spgid); 4114a02a946dSIlya Dryomov ceph_encode_32(&p, map_epoch); 4115a02a946dSIlya Dryomov ceph_encode_8(&p, CEPH_OSD_BACKOFF_OP_ACK_BLOCK); 4116a02a946dSIlya Dryomov ceph_encode_64(&p, backoff->id); 4117a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->begin); 4118a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->end); 4119a02a946dSIlya Dryomov BUG_ON(p != end); 4120a02a946dSIlya Dryomov 4121a02a946dSIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 4122a02a946dSIlya Dryomov msg->hdr.version = cpu_to_le16(1); /* MOSDBackoff v1 */ 4123a02a946dSIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 4124a02a946dSIlya Dryomov 4125a02a946dSIlya Dryomov return msg; 4126a02a946dSIlya Dryomov } 4127a02a946dSIlya Dryomov 4128a02a946dSIlya Dryomov static void handle_backoff_block(struct ceph_osd *osd, struct MOSDBackoff *m) 4129a02a946dSIlya Dryomov { 4130a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4131a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4132a02a946dSIlya Dryomov struct ceph_msg *msg; 4133a02a946dSIlya Dryomov 4134a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4135a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4136a02a946dSIlya Dryomov 4137a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &m->spgid); 4138a02a946dSIlya Dryomov if (!spg) { 4139a02a946dSIlya Dryomov spg = alloc_spg_mapping(); 4140a02a946dSIlya Dryomov if (!spg) { 4141a02a946dSIlya Dryomov pr_err("%s failed to allocate spg\n", __func__); 4142a02a946dSIlya Dryomov return; 4143a02a946dSIlya Dryomov } 4144a02a946dSIlya Dryomov spg->spgid = m->spgid; /* struct */ 4145a02a946dSIlya Dryomov insert_spg_mapping(&osd->o_backoff_mappings, spg); 4146a02a946dSIlya Dryomov } 4147a02a946dSIlya Dryomov 4148a02a946dSIlya Dryomov backoff = alloc_backoff(); 4149a02a946dSIlya Dryomov if (!backoff) { 4150a02a946dSIlya Dryomov pr_err("%s failed to allocate backoff\n", __func__); 4151a02a946dSIlya Dryomov return; 4152a02a946dSIlya Dryomov } 4153a02a946dSIlya Dryomov backoff->spgid = m->spgid; /* struct */ 4154a02a946dSIlya Dryomov backoff->id = m->id; 4155a02a946dSIlya Dryomov backoff->begin = m->begin; 4156a02a946dSIlya Dryomov m->begin = NULL; /* backoff now owns this */ 4157a02a946dSIlya Dryomov backoff->end = m->end; 4158a02a946dSIlya Dryomov m->end = NULL; /* ditto */ 4159a02a946dSIlya Dryomov 4160a02a946dSIlya Dryomov insert_backoff(&spg->backoffs, backoff); 4161a02a946dSIlya Dryomov insert_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4162a02a946dSIlya Dryomov 4163a02a946dSIlya Dryomov /* 4164a02a946dSIlya Dryomov * Ack with original backoff's epoch so that the OSD can 4165a02a946dSIlya Dryomov * discard this if there was a PG split. 4166a02a946dSIlya Dryomov */ 4167a02a946dSIlya Dryomov msg = create_backoff_message(backoff, m->map_epoch); 4168a02a946dSIlya Dryomov if (!msg) { 4169a02a946dSIlya Dryomov pr_err("%s failed to allocate msg\n", __func__); 4170a02a946dSIlya Dryomov return; 4171a02a946dSIlya Dryomov } 4172a02a946dSIlya Dryomov ceph_con_send(&osd->o_con, msg); 4173a02a946dSIlya Dryomov } 4174a02a946dSIlya Dryomov 4175a02a946dSIlya Dryomov static bool target_contained_by(const struct ceph_osd_request_target *t, 4176a02a946dSIlya Dryomov const struct ceph_hobject_id *begin, 4177a02a946dSIlya Dryomov const struct ceph_hobject_id *end) 4178a02a946dSIlya Dryomov { 4179a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 4180a02a946dSIlya Dryomov int cmp; 4181a02a946dSIlya Dryomov 4182a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, t); 4183a02a946dSIlya Dryomov cmp = hoid_compare(&hoid, begin); 4184a02a946dSIlya Dryomov return !cmp || (cmp > 0 && hoid_compare(&hoid, end) < 0); 4185a02a946dSIlya Dryomov } 4186a02a946dSIlya Dryomov 4187a02a946dSIlya Dryomov static void handle_backoff_unblock(struct ceph_osd *osd, 4188a02a946dSIlya Dryomov const struct MOSDBackoff *m) 4189a02a946dSIlya Dryomov { 4190a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4191a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4192a02a946dSIlya Dryomov struct rb_node *n; 4193a02a946dSIlya Dryomov 4194a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4195a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4196a02a946dSIlya Dryomov 4197a02a946dSIlya Dryomov backoff = lookup_backoff_by_id(&osd->o_backoffs_by_id, m->id); 4198a02a946dSIlya Dryomov if (!backoff) { 4199a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu backoff dne\n", 4200a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4201a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4202a02a946dSIlya Dryomov return; 4203a02a946dSIlya Dryomov } 4204a02a946dSIlya Dryomov 4205a02a946dSIlya Dryomov if (hoid_compare(backoff->begin, m->begin) && 4206a02a946dSIlya Dryomov hoid_compare(backoff->end, m->end)) { 4207a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu bad range?\n", 4208a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4209a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4210a02a946dSIlya Dryomov /* unblock it anyway... */ 4211a02a946dSIlya Dryomov } 4212a02a946dSIlya Dryomov 4213a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &backoff->spgid); 4214a02a946dSIlya Dryomov BUG_ON(!spg); 4215a02a946dSIlya Dryomov 4216a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 4217a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4218a02a946dSIlya Dryomov free_backoff(backoff); 4219a02a946dSIlya Dryomov 4220a02a946dSIlya Dryomov if (RB_EMPTY_ROOT(&spg->backoffs)) { 4221a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 4222a02a946dSIlya Dryomov free_spg_mapping(spg); 4223a02a946dSIlya Dryomov } 4224a02a946dSIlya Dryomov 4225a02a946dSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 4226a02a946dSIlya Dryomov struct ceph_osd_request *req = 4227a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 4228a02a946dSIlya Dryomov 4229a02a946dSIlya Dryomov if (!ceph_spg_compare(&req->r_t.spgid, &m->spgid)) { 4230a02a946dSIlya Dryomov /* 4231a02a946dSIlya Dryomov * Match against @m, not @backoff -- the PG may 4232a02a946dSIlya Dryomov * have split on the OSD. 4233a02a946dSIlya Dryomov */ 4234a02a946dSIlya Dryomov if (target_contained_by(&req->r_t, m->begin, m->end)) { 4235a02a946dSIlya Dryomov /* 4236a02a946dSIlya Dryomov * If no other installed backoff applies, 4237a02a946dSIlya Dryomov * resend. 4238a02a946dSIlya Dryomov */ 4239a02a946dSIlya Dryomov send_request(req); 4240a02a946dSIlya Dryomov } 4241a02a946dSIlya Dryomov } 4242a02a946dSIlya Dryomov } 4243a02a946dSIlya Dryomov } 4244a02a946dSIlya Dryomov 4245a02a946dSIlya Dryomov static void handle_backoff(struct ceph_osd *osd, struct ceph_msg *msg) 4246a02a946dSIlya Dryomov { 4247a02a946dSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 4248a02a946dSIlya Dryomov struct MOSDBackoff m; 4249a02a946dSIlya Dryomov int ret; 4250a02a946dSIlya Dryomov 4251a02a946dSIlya Dryomov down_read(&osdc->lock); 4252a02a946dSIlya Dryomov if (!osd_registered(osd)) { 4253a02a946dSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 4254a02a946dSIlya Dryomov up_read(&osdc->lock); 4255a02a946dSIlya Dryomov return; 4256a02a946dSIlya Dryomov } 4257a02a946dSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 4258a02a946dSIlya Dryomov 4259a02a946dSIlya Dryomov mutex_lock(&osd->lock); 4260a02a946dSIlya Dryomov ret = decode_MOSDBackoff(msg, &m); 4261a02a946dSIlya Dryomov if (ret) { 4262a02a946dSIlya Dryomov pr_err("failed to decode MOSDBackoff: %d\n", ret); 4263a02a946dSIlya Dryomov ceph_msg_dump(msg); 4264a02a946dSIlya Dryomov goto out_unlock; 4265a02a946dSIlya Dryomov } 4266a02a946dSIlya Dryomov 4267a02a946dSIlya Dryomov switch (m.op) { 4268a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_BLOCK: 4269a02a946dSIlya Dryomov handle_backoff_block(osd, &m); 4270a02a946dSIlya Dryomov break; 4271a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_UNBLOCK: 4272a02a946dSIlya Dryomov handle_backoff_unblock(osd, &m); 4273a02a946dSIlya Dryomov break; 4274a02a946dSIlya Dryomov default: 4275a02a946dSIlya Dryomov pr_err("%s osd%d unknown op %d\n", __func__, osd->o_osd, m.op); 4276a02a946dSIlya Dryomov } 4277a02a946dSIlya Dryomov 4278a02a946dSIlya Dryomov free_hoid(m.begin); 4279a02a946dSIlya Dryomov free_hoid(m.end); 4280a02a946dSIlya Dryomov 4281a02a946dSIlya Dryomov out_unlock: 4282a02a946dSIlya Dryomov mutex_unlock(&osd->lock); 4283a02a946dSIlya Dryomov up_read(&osdc->lock); 4284a02a946dSIlya Dryomov } 4285a02a946dSIlya Dryomov 42863d14c5d2SYehuda Sadeh /* 4287a40c4f10SYehuda Sadeh * Process osd watch notifications 4288a40c4f10SYehuda Sadeh */ 42893c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 42903c663bbdSAlex Elder struct ceph_msg *msg) 4291a40c4f10SYehuda Sadeh { 4292922dab61SIlya Dryomov void *p = msg->front.iov_base; 4293922dab61SIlya Dryomov void *const end = p + msg->front.iov_len; 4294922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4295922dab61SIlya Dryomov struct linger_work *lwork; 4296922dab61SIlya Dryomov u8 proto_ver, opcode; 4297922dab61SIlya Dryomov u64 cookie, notify_id; 4298922dab61SIlya Dryomov u64 notifier_id = 0; 429919079203SIlya Dryomov s32 return_code = 0; 4300922dab61SIlya Dryomov void *payload = NULL; 4301922dab61SIlya Dryomov u32 payload_len = 0; 4302a40c4f10SYehuda Sadeh 4303a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 4304a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 4305a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 4306922dab61SIlya Dryomov p += 8; /* skip ver */ 4307a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 4308a40c4f10SYehuda Sadeh 4309922dab61SIlya Dryomov if (proto_ver >= 1) { 4310922dab61SIlya Dryomov ceph_decode_32_safe(&p, end, payload_len, bad); 4311922dab61SIlya Dryomov ceph_decode_need(&p, end, payload_len, bad); 4312922dab61SIlya Dryomov payload = p; 4313922dab61SIlya Dryomov p += payload_len; 4314a40c4f10SYehuda Sadeh } 4315a40c4f10SYehuda Sadeh 4316922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 2) 431719079203SIlya Dryomov ceph_decode_32_safe(&p, end, return_code, bad); 4318922dab61SIlya Dryomov 4319922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 3) 4320922dab61SIlya Dryomov ceph_decode_64_safe(&p, end, notifier_id, bad); 4321922dab61SIlya Dryomov 4322922dab61SIlya Dryomov down_read(&osdc->lock); 4323922dab61SIlya Dryomov lreq = lookup_linger_osdc(&osdc->linger_requests, cookie); 4324922dab61SIlya Dryomov if (!lreq) { 4325922dab61SIlya Dryomov dout("%s opcode %d cookie %llu dne\n", __func__, opcode, 4326922dab61SIlya Dryomov cookie); 4327922dab61SIlya Dryomov goto out_unlock_osdc; 4328922dab61SIlya Dryomov } 4329922dab61SIlya Dryomov 4330922dab61SIlya Dryomov mutex_lock(&lreq->lock); 433119079203SIlya Dryomov dout("%s opcode %d cookie %llu lreq %p is_watch %d\n", __func__, 433219079203SIlya Dryomov opcode, cookie, lreq, lreq->is_watch); 4333922dab61SIlya Dryomov if (opcode == CEPH_WATCH_EVENT_DISCONNECT) { 4334922dab61SIlya Dryomov if (!lreq->last_error) { 4335922dab61SIlya Dryomov lreq->last_error = -ENOTCONN; 4336922dab61SIlya Dryomov queue_watch_error(lreq); 4337922dab61SIlya Dryomov } 433819079203SIlya Dryomov } else if (!lreq->is_watch) { 433919079203SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY_COMPLETE */ 434019079203SIlya Dryomov if (lreq->notify_id && lreq->notify_id != notify_id) { 434119079203SIlya Dryomov dout("lreq %p notify_id %llu != %llu, ignoring\n", lreq, 434219079203SIlya Dryomov lreq->notify_id, notify_id); 434319079203SIlya Dryomov } else if (!completion_done(&lreq->notify_finish_wait)) { 434419079203SIlya Dryomov struct ceph_msg_data *data = 434519079203SIlya Dryomov list_first_entry_or_null(&msg->data, 434619079203SIlya Dryomov struct ceph_msg_data, 434719079203SIlya Dryomov links); 434819079203SIlya Dryomov 434919079203SIlya Dryomov if (data) { 435019079203SIlya Dryomov if (lreq->preply_pages) { 435119079203SIlya Dryomov WARN_ON(data->type != 435219079203SIlya Dryomov CEPH_MSG_DATA_PAGES); 435319079203SIlya Dryomov *lreq->preply_pages = data->pages; 435419079203SIlya Dryomov *lreq->preply_len = data->length; 435519079203SIlya Dryomov } else { 435619079203SIlya Dryomov ceph_release_page_vector(data->pages, 435719079203SIlya Dryomov calc_pages_for(0, data->length)); 435819079203SIlya Dryomov } 435919079203SIlya Dryomov } 436019079203SIlya Dryomov lreq->notify_finish_error = return_code; 436119079203SIlya Dryomov complete_all(&lreq->notify_finish_wait); 436219079203SIlya Dryomov } 4363922dab61SIlya Dryomov } else { 4364922dab61SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY */ 4365922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_notify); 4366922dab61SIlya Dryomov if (!lwork) { 4367922dab61SIlya Dryomov pr_err("failed to allocate notify-lwork\n"); 4368922dab61SIlya Dryomov goto out_unlock_lreq; 4369922dab61SIlya Dryomov } 4370922dab61SIlya Dryomov 4371922dab61SIlya Dryomov lwork->notify.notify_id = notify_id; 4372922dab61SIlya Dryomov lwork->notify.notifier_id = notifier_id; 4373922dab61SIlya Dryomov lwork->notify.payload = payload; 4374922dab61SIlya Dryomov lwork->notify.payload_len = payload_len; 4375922dab61SIlya Dryomov lwork->notify.msg = ceph_msg_get(msg); 4376922dab61SIlya Dryomov lwork_queue(lwork); 4377922dab61SIlya Dryomov } 4378922dab61SIlya Dryomov 4379922dab61SIlya Dryomov out_unlock_lreq: 4380922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 4381922dab61SIlya Dryomov out_unlock_osdc: 4382922dab61SIlya Dryomov up_read(&osdc->lock); 4383a40c4f10SYehuda Sadeh return; 4384a40c4f10SYehuda Sadeh 4385a40c4f10SYehuda Sadeh bad: 4386a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 4387a40c4f10SYehuda Sadeh } 4388a40c4f10SYehuda Sadeh 4389a40c4f10SYehuda Sadeh /* 43903d14c5d2SYehuda Sadeh * Register request, send initial attempt. 43913d14c5d2SYehuda Sadeh */ 43923d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 43933d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 43943d14c5d2SYehuda Sadeh bool nofail) 43953d14c5d2SYehuda Sadeh { 43965aea3dcdSIlya Dryomov down_read(&osdc->lock); 43975aea3dcdSIlya Dryomov submit_request(req, false); 43985aea3dcdSIlya Dryomov up_read(&osdc->lock); 43993d14c5d2SYehuda Sadeh 44005aea3dcdSIlya Dryomov return 0; 44013d14c5d2SYehuda Sadeh } 44023d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 44033d14c5d2SYehuda Sadeh 44043d14c5d2SYehuda Sadeh /* 4405c297eb42SIlya Dryomov * Unregister a registered request. The request is not completed: 4406c297eb42SIlya Dryomov * ->r_result isn't set and __complete_request() isn't called. 4407c9f9b93dSIlya Dryomov */ 4408c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 4409c9f9b93dSIlya Dryomov { 4410c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 4411c9f9b93dSIlya Dryomov 44125aea3dcdSIlya Dryomov down_write(&osdc->lock); 44135aea3dcdSIlya Dryomov if (req->r_osd) 44145aea3dcdSIlya Dryomov cancel_request(req); 44155aea3dcdSIlya Dryomov up_write(&osdc->lock); 4416c9f9b93dSIlya Dryomov } 4417c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 4418c9f9b93dSIlya Dryomov 4419c9f9b93dSIlya Dryomov /* 442042b06965SIlya Dryomov * @timeout: in jiffies, 0 means "wait forever" 442142b06965SIlya Dryomov */ 442242b06965SIlya Dryomov static int wait_request_timeout(struct ceph_osd_request *req, 442342b06965SIlya Dryomov unsigned long timeout) 442442b06965SIlya Dryomov { 442542b06965SIlya Dryomov long left; 442642b06965SIlya Dryomov 442742b06965SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 44280e76abf2SYan, Zheng left = wait_for_completion_killable_timeout(&req->r_completion, 442942b06965SIlya Dryomov ceph_timeout_jiffies(timeout)); 443042b06965SIlya Dryomov if (left <= 0) { 443142b06965SIlya Dryomov left = left ?: -ETIMEDOUT; 443242b06965SIlya Dryomov ceph_osdc_cancel_request(req); 443342b06965SIlya Dryomov } else { 443442b06965SIlya Dryomov left = req->r_result; /* completed */ 443542b06965SIlya Dryomov } 443642b06965SIlya Dryomov 443742b06965SIlya Dryomov return left; 443842b06965SIlya Dryomov } 443942b06965SIlya Dryomov 444042b06965SIlya Dryomov /* 44413d14c5d2SYehuda Sadeh * wait for a request to complete 44423d14c5d2SYehuda Sadeh */ 44433d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 44443d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 44453d14c5d2SYehuda Sadeh { 444642b06965SIlya Dryomov return wait_request_timeout(req, 0); 44473d14c5d2SYehuda Sadeh } 44483d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 44493d14c5d2SYehuda Sadeh 44503d14c5d2SYehuda Sadeh /* 44513d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 44523d14c5d2SYehuda Sadeh */ 44533d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 44543d14c5d2SYehuda Sadeh { 44555aea3dcdSIlya Dryomov struct rb_node *n, *p; 44565aea3dcdSIlya Dryomov u64 last_tid = atomic64_read(&osdc->last_tid); 44573d14c5d2SYehuda Sadeh 44585aea3dcdSIlya Dryomov again: 44595aea3dcdSIlya Dryomov down_read(&osdc->lock); 44605aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 44615aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 44625aea3dcdSIlya Dryomov 44635aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 44645aea3dcdSIlya Dryomov for (p = rb_first(&osd->o_requests); p; p = rb_next(p)) { 44655aea3dcdSIlya Dryomov struct ceph_osd_request *req = 44665aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 44675aea3dcdSIlya Dryomov 44683d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 44693d14c5d2SYehuda Sadeh break; 44703d14c5d2SYehuda Sadeh 44715aea3dcdSIlya Dryomov if (!(req->r_flags & CEPH_OSD_FLAG_WRITE)) 44723d14c5d2SYehuda Sadeh continue; 44733d14c5d2SYehuda Sadeh 44743d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 44755aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 44765aea3dcdSIlya Dryomov up_read(&osdc->lock); 44775aea3dcdSIlya Dryomov dout("%s waiting on req %p tid %llu last_tid %llu\n", 44785aea3dcdSIlya Dryomov __func__, req, req->r_tid, last_tid); 4479b18b9550SIlya Dryomov wait_for_completion(&req->r_completion); 44803d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 44815aea3dcdSIlya Dryomov goto again; 44823d14c5d2SYehuda Sadeh } 44835aea3dcdSIlya Dryomov 44845aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 44855aea3dcdSIlya Dryomov } 44865aea3dcdSIlya Dryomov 44875aea3dcdSIlya Dryomov up_read(&osdc->lock); 44885aea3dcdSIlya Dryomov dout("%s done last_tid %llu\n", __func__, last_tid); 44893d14c5d2SYehuda Sadeh } 44903d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 44913d14c5d2SYehuda Sadeh 4492922dab61SIlya Dryomov static struct ceph_osd_request * 4493922dab61SIlya Dryomov alloc_linger_request(struct ceph_osd_linger_request *lreq) 4494922dab61SIlya Dryomov { 4495922dab61SIlya Dryomov struct ceph_osd_request *req; 4496922dab61SIlya Dryomov 4497922dab61SIlya Dryomov req = ceph_osdc_alloc_request(lreq->osdc, NULL, 1, false, GFP_NOIO); 4498922dab61SIlya Dryomov if (!req) 4499922dab61SIlya Dryomov return NULL; 4500922dab61SIlya Dryomov 4501922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4502922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 4503922dab61SIlya Dryomov 4504922dab61SIlya Dryomov if (ceph_osdc_alloc_messages(req, GFP_NOIO)) { 4505922dab61SIlya Dryomov ceph_osdc_put_request(req); 4506922dab61SIlya Dryomov return NULL; 4507922dab61SIlya Dryomov } 4508922dab61SIlya Dryomov 4509922dab61SIlya Dryomov return req; 4510922dab61SIlya Dryomov } 4511922dab61SIlya Dryomov 4512922dab61SIlya Dryomov /* 4513922dab61SIlya Dryomov * Returns a handle, caller owns a ref. 4514922dab61SIlya Dryomov */ 4515922dab61SIlya Dryomov struct ceph_osd_linger_request * 4516922dab61SIlya Dryomov ceph_osdc_watch(struct ceph_osd_client *osdc, 4517922dab61SIlya Dryomov struct ceph_object_id *oid, 4518922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4519922dab61SIlya Dryomov rados_watchcb2_t wcb, 4520922dab61SIlya Dryomov rados_watcherrcb_t errcb, 4521922dab61SIlya Dryomov void *data) 4522922dab61SIlya Dryomov { 4523922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4524922dab61SIlya Dryomov int ret; 4525922dab61SIlya Dryomov 4526922dab61SIlya Dryomov lreq = linger_alloc(osdc); 4527922dab61SIlya Dryomov if (!lreq) 4528922dab61SIlya Dryomov return ERR_PTR(-ENOMEM); 4529922dab61SIlya Dryomov 453019079203SIlya Dryomov lreq->is_watch = true; 4531922dab61SIlya Dryomov lreq->wcb = wcb; 4532922dab61SIlya Dryomov lreq->errcb = errcb; 4533922dab61SIlya Dryomov lreq->data = data; 4534b07d3c4bSIlya Dryomov lreq->watch_valid_thru = jiffies; 4535922dab61SIlya Dryomov 4536922dab61SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 4537922dab61SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 453854ea0046SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_WRITE; 45391134e091SDeepa Dinamani ktime_get_real_ts(&lreq->mtime); 4540922dab61SIlya Dryomov 4541922dab61SIlya Dryomov lreq->reg_req = alloc_linger_request(lreq); 4542922dab61SIlya Dryomov if (!lreq->reg_req) { 4543922dab61SIlya Dryomov ret = -ENOMEM; 4544922dab61SIlya Dryomov goto err_put_lreq; 4545922dab61SIlya Dryomov } 4546922dab61SIlya Dryomov 4547922dab61SIlya Dryomov lreq->ping_req = alloc_linger_request(lreq); 4548922dab61SIlya Dryomov if (!lreq->ping_req) { 4549922dab61SIlya Dryomov ret = -ENOMEM; 4550922dab61SIlya Dryomov goto err_put_lreq; 4551922dab61SIlya Dryomov } 4552922dab61SIlya Dryomov 4553922dab61SIlya Dryomov down_write(&osdc->lock); 4554922dab61SIlya Dryomov linger_register(lreq); /* before osd_req_op_* */ 4555922dab61SIlya Dryomov osd_req_op_watch_init(lreq->reg_req, 0, lreq->linger_id, 4556922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_WATCH); 4557922dab61SIlya Dryomov osd_req_op_watch_init(lreq->ping_req, 0, lreq->linger_id, 4558922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_PING); 4559922dab61SIlya Dryomov linger_submit(lreq); 4560922dab61SIlya Dryomov up_write(&osdc->lock); 4561922dab61SIlya Dryomov 4562922dab61SIlya Dryomov ret = linger_reg_commit_wait(lreq); 4563922dab61SIlya Dryomov if (ret) { 4564922dab61SIlya Dryomov linger_cancel(lreq); 4565922dab61SIlya Dryomov goto err_put_lreq; 4566922dab61SIlya Dryomov } 4567922dab61SIlya Dryomov 4568922dab61SIlya Dryomov return lreq; 4569922dab61SIlya Dryomov 4570922dab61SIlya Dryomov err_put_lreq: 4571922dab61SIlya Dryomov linger_put(lreq); 4572922dab61SIlya Dryomov return ERR_PTR(ret); 4573922dab61SIlya Dryomov } 4574922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_watch); 4575922dab61SIlya Dryomov 4576922dab61SIlya Dryomov /* 4577922dab61SIlya Dryomov * Releases a ref. 4578922dab61SIlya Dryomov * 4579922dab61SIlya Dryomov * Times out after mount_timeout to preserve rbd unmap behaviour 4580922dab61SIlya Dryomov * introduced in 2894e1d76974 ("rbd: timeout watch teardown on unmap 4581922dab61SIlya Dryomov * with mount_timeout"). 4582922dab61SIlya Dryomov */ 4583922dab61SIlya Dryomov int ceph_osdc_unwatch(struct ceph_osd_client *osdc, 4584922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 4585922dab61SIlya Dryomov { 4586922dab61SIlya Dryomov struct ceph_options *opts = osdc->client->options; 4587922dab61SIlya Dryomov struct ceph_osd_request *req; 4588922dab61SIlya Dryomov int ret; 4589922dab61SIlya Dryomov 4590922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4591922dab61SIlya Dryomov if (!req) 4592922dab61SIlya Dryomov return -ENOMEM; 4593922dab61SIlya Dryomov 4594922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4595922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 459654ea0046SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_WRITE; 45971134e091SDeepa Dinamani ktime_get_real_ts(&req->r_mtime); 4598922dab61SIlya Dryomov osd_req_op_watch_init(req, 0, lreq->linger_id, 4599922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_UNWATCH); 4600922dab61SIlya Dryomov 4601922dab61SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4602922dab61SIlya Dryomov if (ret) 4603922dab61SIlya Dryomov goto out_put_req; 4604922dab61SIlya Dryomov 4605922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4606922dab61SIlya Dryomov linger_cancel(lreq); 4607922dab61SIlya Dryomov linger_put(lreq); 4608922dab61SIlya Dryomov ret = wait_request_timeout(req, opts->mount_timeout); 4609922dab61SIlya Dryomov 4610922dab61SIlya Dryomov out_put_req: 4611922dab61SIlya Dryomov ceph_osdc_put_request(req); 4612922dab61SIlya Dryomov return ret; 4613922dab61SIlya Dryomov } 4614922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_unwatch); 4615922dab61SIlya Dryomov 4616922dab61SIlya Dryomov static int osd_req_op_notify_ack_init(struct ceph_osd_request *req, int which, 4617922dab61SIlya Dryomov u64 notify_id, u64 cookie, void *payload, 4618922dab61SIlya Dryomov size_t payload_len) 4619922dab61SIlya Dryomov { 4620922dab61SIlya Dryomov struct ceph_osd_req_op *op; 4621922dab61SIlya Dryomov struct ceph_pagelist *pl; 4622922dab61SIlya Dryomov int ret; 4623922dab61SIlya Dryomov 4624922dab61SIlya Dryomov op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY_ACK, 0); 4625922dab61SIlya Dryomov 4626922dab61SIlya Dryomov pl = kmalloc(sizeof(*pl), GFP_NOIO); 4627922dab61SIlya Dryomov if (!pl) 4628922dab61SIlya Dryomov return -ENOMEM; 4629922dab61SIlya Dryomov 4630922dab61SIlya Dryomov ceph_pagelist_init(pl); 4631922dab61SIlya Dryomov ret = ceph_pagelist_encode_64(pl, notify_id); 4632922dab61SIlya Dryomov ret |= ceph_pagelist_encode_64(pl, cookie); 4633922dab61SIlya Dryomov if (payload) { 4634922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 4635922dab61SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 4636922dab61SIlya Dryomov } else { 4637922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, 0); 4638922dab61SIlya Dryomov } 4639922dab61SIlya Dryomov if (ret) { 4640922dab61SIlya Dryomov ceph_pagelist_release(pl); 4641922dab61SIlya Dryomov return -ENOMEM; 4642922dab61SIlya Dryomov } 4643922dab61SIlya Dryomov 4644922dab61SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify_ack.request_data, pl); 4645922dab61SIlya Dryomov op->indata_len = pl->length; 4646922dab61SIlya Dryomov return 0; 4647922dab61SIlya Dryomov } 4648922dab61SIlya Dryomov 4649922dab61SIlya Dryomov int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, 4650922dab61SIlya Dryomov struct ceph_object_id *oid, 4651922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4652922dab61SIlya Dryomov u64 notify_id, 4653922dab61SIlya Dryomov u64 cookie, 4654922dab61SIlya Dryomov void *payload, 4655922dab61SIlya Dryomov size_t payload_len) 4656922dab61SIlya Dryomov { 4657922dab61SIlya Dryomov struct ceph_osd_request *req; 4658922dab61SIlya Dryomov int ret; 4659922dab61SIlya Dryomov 4660922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4661922dab61SIlya Dryomov if (!req) 4662922dab61SIlya Dryomov return -ENOMEM; 4663922dab61SIlya Dryomov 4664922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, oid); 4665922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, oloc); 4666922dab61SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_READ; 4667922dab61SIlya Dryomov 4668922dab61SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4669922dab61SIlya Dryomov if (ret) 4670922dab61SIlya Dryomov goto out_put_req; 4671922dab61SIlya Dryomov 4672922dab61SIlya Dryomov ret = osd_req_op_notify_ack_init(req, 0, notify_id, cookie, payload, 4673922dab61SIlya Dryomov payload_len); 4674922dab61SIlya Dryomov if (ret) 4675922dab61SIlya Dryomov goto out_put_req; 4676922dab61SIlya Dryomov 4677922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4678922dab61SIlya Dryomov ret = ceph_osdc_wait_request(osdc, req); 4679922dab61SIlya Dryomov 4680922dab61SIlya Dryomov out_put_req: 4681922dab61SIlya Dryomov ceph_osdc_put_request(req); 4682922dab61SIlya Dryomov return ret; 4683922dab61SIlya Dryomov } 4684922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify_ack); 4685922dab61SIlya Dryomov 468619079203SIlya Dryomov static int osd_req_op_notify_init(struct ceph_osd_request *req, int which, 468719079203SIlya Dryomov u64 cookie, u32 prot_ver, u32 timeout, 468819079203SIlya Dryomov void *payload, size_t payload_len) 468919079203SIlya Dryomov { 469019079203SIlya Dryomov struct ceph_osd_req_op *op; 469119079203SIlya Dryomov struct ceph_pagelist *pl; 469219079203SIlya Dryomov int ret; 469319079203SIlya Dryomov 469419079203SIlya Dryomov op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY, 0); 469519079203SIlya Dryomov op->notify.cookie = cookie; 469619079203SIlya Dryomov 469719079203SIlya Dryomov pl = kmalloc(sizeof(*pl), GFP_NOIO); 469819079203SIlya Dryomov if (!pl) 469919079203SIlya Dryomov return -ENOMEM; 470019079203SIlya Dryomov 470119079203SIlya Dryomov ceph_pagelist_init(pl); 470219079203SIlya Dryomov ret = ceph_pagelist_encode_32(pl, 1); /* prot_ver */ 470319079203SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, timeout); 470419079203SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 470519079203SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 470619079203SIlya Dryomov if (ret) { 470719079203SIlya Dryomov ceph_pagelist_release(pl); 470819079203SIlya Dryomov return -ENOMEM; 470919079203SIlya Dryomov } 471019079203SIlya Dryomov 471119079203SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify.request_data, pl); 471219079203SIlya Dryomov op->indata_len = pl->length; 471319079203SIlya Dryomov return 0; 471419079203SIlya Dryomov } 471519079203SIlya Dryomov 471619079203SIlya Dryomov /* 471719079203SIlya Dryomov * @timeout: in seconds 471819079203SIlya Dryomov * 471919079203SIlya Dryomov * @preply_{pages,len} are initialized both on success and error. 472019079203SIlya Dryomov * The caller is responsible for: 472119079203SIlya Dryomov * 472219079203SIlya Dryomov * ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len)) 472319079203SIlya Dryomov */ 472419079203SIlya Dryomov int ceph_osdc_notify(struct ceph_osd_client *osdc, 472519079203SIlya Dryomov struct ceph_object_id *oid, 472619079203SIlya Dryomov struct ceph_object_locator *oloc, 472719079203SIlya Dryomov void *payload, 472819079203SIlya Dryomov size_t payload_len, 472919079203SIlya Dryomov u32 timeout, 473019079203SIlya Dryomov struct page ***preply_pages, 473119079203SIlya Dryomov size_t *preply_len) 473219079203SIlya Dryomov { 473319079203SIlya Dryomov struct ceph_osd_linger_request *lreq; 473419079203SIlya Dryomov struct page **pages; 473519079203SIlya Dryomov int ret; 473619079203SIlya Dryomov 473719079203SIlya Dryomov WARN_ON(!timeout); 473819079203SIlya Dryomov if (preply_pages) { 473919079203SIlya Dryomov *preply_pages = NULL; 474019079203SIlya Dryomov *preply_len = 0; 474119079203SIlya Dryomov } 474219079203SIlya Dryomov 474319079203SIlya Dryomov lreq = linger_alloc(osdc); 474419079203SIlya Dryomov if (!lreq) 474519079203SIlya Dryomov return -ENOMEM; 474619079203SIlya Dryomov 474719079203SIlya Dryomov lreq->preply_pages = preply_pages; 474819079203SIlya Dryomov lreq->preply_len = preply_len; 474919079203SIlya Dryomov 475019079203SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 475119079203SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 475219079203SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_READ; 475319079203SIlya Dryomov 475419079203SIlya Dryomov lreq->reg_req = alloc_linger_request(lreq); 475519079203SIlya Dryomov if (!lreq->reg_req) { 475619079203SIlya Dryomov ret = -ENOMEM; 475719079203SIlya Dryomov goto out_put_lreq; 475819079203SIlya Dryomov } 475919079203SIlya Dryomov 476019079203SIlya Dryomov /* for notify_id */ 476119079203SIlya Dryomov pages = ceph_alloc_page_vector(1, GFP_NOIO); 476219079203SIlya Dryomov if (IS_ERR(pages)) { 476319079203SIlya Dryomov ret = PTR_ERR(pages); 476419079203SIlya Dryomov goto out_put_lreq; 476519079203SIlya Dryomov } 476619079203SIlya Dryomov 476719079203SIlya Dryomov down_write(&osdc->lock); 476819079203SIlya Dryomov linger_register(lreq); /* before osd_req_op_* */ 476919079203SIlya Dryomov ret = osd_req_op_notify_init(lreq->reg_req, 0, lreq->linger_id, 1, 477019079203SIlya Dryomov timeout, payload, payload_len); 477119079203SIlya Dryomov if (ret) { 477219079203SIlya Dryomov linger_unregister(lreq); 477319079203SIlya Dryomov up_write(&osdc->lock); 477419079203SIlya Dryomov ceph_release_page_vector(pages, 1); 477519079203SIlya Dryomov goto out_put_lreq; 477619079203SIlya Dryomov } 477719079203SIlya Dryomov ceph_osd_data_pages_init(osd_req_op_data(lreq->reg_req, 0, notify, 477819079203SIlya Dryomov response_data), 477919079203SIlya Dryomov pages, PAGE_SIZE, 0, false, true); 478019079203SIlya Dryomov linger_submit(lreq); 478119079203SIlya Dryomov up_write(&osdc->lock); 478219079203SIlya Dryomov 478319079203SIlya Dryomov ret = linger_reg_commit_wait(lreq); 478419079203SIlya Dryomov if (!ret) 478519079203SIlya Dryomov ret = linger_notify_finish_wait(lreq); 478619079203SIlya Dryomov else 478719079203SIlya Dryomov dout("lreq %p failed to initiate notify %d\n", lreq, ret); 478819079203SIlya Dryomov 478919079203SIlya Dryomov linger_cancel(lreq); 479019079203SIlya Dryomov out_put_lreq: 479119079203SIlya Dryomov linger_put(lreq); 479219079203SIlya Dryomov return ret; 479319079203SIlya Dryomov } 479419079203SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify); 479519079203SIlya Dryomov 47963d14c5d2SYehuda Sadeh /* 4797b07d3c4bSIlya Dryomov * Return the number of milliseconds since the watch was last 4798b07d3c4bSIlya Dryomov * confirmed, or an error. If there is an error, the watch is no 4799b07d3c4bSIlya Dryomov * longer valid, and should be destroyed with ceph_osdc_unwatch(). 4800b07d3c4bSIlya Dryomov */ 4801b07d3c4bSIlya Dryomov int ceph_osdc_watch_check(struct ceph_osd_client *osdc, 4802b07d3c4bSIlya Dryomov struct ceph_osd_linger_request *lreq) 4803b07d3c4bSIlya Dryomov { 4804b07d3c4bSIlya Dryomov unsigned long stamp, age; 4805b07d3c4bSIlya Dryomov int ret; 4806b07d3c4bSIlya Dryomov 4807b07d3c4bSIlya Dryomov down_read(&osdc->lock); 4808b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 4809b07d3c4bSIlya Dryomov stamp = lreq->watch_valid_thru; 4810b07d3c4bSIlya Dryomov if (!list_empty(&lreq->pending_lworks)) { 4811b07d3c4bSIlya Dryomov struct linger_work *lwork = 4812b07d3c4bSIlya Dryomov list_first_entry(&lreq->pending_lworks, 4813b07d3c4bSIlya Dryomov struct linger_work, 4814b07d3c4bSIlya Dryomov pending_item); 4815b07d3c4bSIlya Dryomov 4816b07d3c4bSIlya Dryomov if (time_before(lwork->queued_stamp, stamp)) 4817b07d3c4bSIlya Dryomov stamp = lwork->queued_stamp; 4818b07d3c4bSIlya Dryomov } 4819b07d3c4bSIlya Dryomov age = jiffies - stamp; 4820b07d3c4bSIlya Dryomov dout("%s lreq %p linger_id %llu age %lu last_error %d\n", __func__, 4821b07d3c4bSIlya Dryomov lreq, lreq->linger_id, age, lreq->last_error); 4822b07d3c4bSIlya Dryomov /* we are truncating to msecs, so return a safe upper bound */ 4823b07d3c4bSIlya Dryomov ret = lreq->last_error ?: 1 + jiffies_to_msecs(age); 4824b07d3c4bSIlya Dryomov 4825b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 4826b07d3c4bSIlya Dryomov up_read(&osdc->lock); 4827b07d3c4bSIlya Dryomov return ret; 4828b07d3c4bSIlya Dryomov } 4829b07d3c4bSIlya Dryomov 4830a4ed38d7SDouglas Fuller static int decode_watcher(void **p, void *end, struct ceph_watch_item *item) 4831a4ed38d7SDouglas Fuller { 4832a4ed38d7SDouglas Fuller u8 struct_v; 4833a4ed38d7SDouglas Fuller u32 struct_len; 4834a4ed38d7SDouglas Fuller int ret; 4835a4ed38d7SDouglas Fuller 4836a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 2, "watch_item_t", 4837a4ed38d7SDouglas Fuller &struct_v, &struct_len); 4838a4ed38d7SDouglas Fuller if (ret) 4839a4ed38d7SDouglas Fuller return ret; 4840a4ed38d7SDouglas Fuller 4841a4ed38d7SDouglas Fuller ceph_decode_copy(p, &item->name, sizeof(item->name)); 4842a4ed38d7SDouglas Fuller item->cookie = ceph_decode_64(p); 4843a4ed38d7SDouglas Fuller *p += 4; /* skip timeout_seconds */ 4844a4ed38d7SDouglas Fuller if (struct_v >= 2) { 4845a4ed38d7SDouglas Fuller ceph_decode_copy(p, &item->addr, sizeof(item->addr)); 4846a4ed38d7SDouglas Fuller ceph_decode_addr(&item->addr); 4847a4ed38d7SDouglas Fuller } 4848a4ed38d7SDouglas Fuller 4849a4ed38d7SDouglas Fuller dout("%s %s%llu cookie %llu addr %s\n", __func__, 4850a4ed38d7SDouglas Fuller ENTITY_NAME(item->name), item->cookie, 4851a4ed38d7SDouglas Fuller ceph_pr_addr(&item->addr.in_addr)); 4852a4ed38d7SDouglas Fuller return 0; 4853a4ed38d7SDouglas Fuller } 4854a4ed38d7SDouglas Fuller 4855a4ed38d7SDouglas Fuller static int decode_watchers(void **p, void *end, 4856a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 4857a4ed38d7SDouglas Fuller u32 *num_watchers) 4858a4ed38d7SDouglas Fuller { 4859a4ed38d7SDouglas Fuller u8 struct_v; 4860a4ed38d7SDouglas Fuller u32 struct_len; 4861a4ed38d7SDouglas Fuller int i; 4862a4ed38d7SDouglas Fuller int ret; 4863a4ed38d7SDouglas Fuller 4864a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 1, "obj_list_watch_response_t", 4865a4ed38d7SDouglas Fuller &struct_v, &struct_len); 4866a4ed38d7SDouglas Fuller if (ret) 4867a4ed38d7SDouglas Fuller return ret; 4868a4ed38d7SDouglas Fuller 4869a4ed38d7SDouglas Fuller *num_watchers = ceph_decode_32(p); 4870a4ed38d7SDouglas Fuller *watchers = kcalloc(*num_watchers, sizeof(**watchers), GFP_NOIO); 4871a4ed38d7SDouglas Fuller if (!*watchers) 4872a4ed38d7SDouglas Fuller return -ENOMEM; 4873a4ed38d7SDouglas Fuller 4874a4ed38d7SDouglas Fuller for (i = 0; i < *num_watchers; i++) { 4875a4ed38d7SDouglas Fuller ret = decode_watcher(p, end, *watchers + i); 4876a4ed38d7SDouglas Fuller if (ret) { 4877a4ed38d7SDouglas Fuller kfree(*watchers); 4878a4ed38d7SDouglas Fuller return ret; 4879a4ed38d7SDouglas Fuller } 4880a4ed38d7SDouglas Fuller } 4881a4ed38d7SDouglas Fuller 4882a4ed38d7SDouglas Fuller return 0; 4883a4ed38d7SDouglas Fuller } 4884a4ed38d7SDouglas Fuller 4885a4ed38d7SDouglas Fuller /* 4886a4ed38d7SDouglas Fuller * On success, the caller is responsible for: 4887a4ed38d7SDouglas Fuller * 4888a4ed38d7SDouglas Fuller * kfree(watchers); 4889a4ed38d7SDouglas Fuller */ 4890a4ed38d7SDouglas Fuller int ceph_osdc_list_watchers(struct ceph_osd_client *osdc, 4891a4ed38d7SDouglas Fuller struct ceph_object_id *oid, 4892a4ed38d7SDouglas Fuller struct ceph_object_locator *oloc, 4893a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 4894a4ed38d7SDouglas Fuller u32 *num_watchers) 4895a4ed38d7SDouglas Fuller { 4896a4ed38d7SDouglas Fuller struct ceph_osd_request *req; 4897a4ed38d7SDouglas Fuller struct page **pages; 4898a4ed38d7SDouglas Fuller int ret; 4899a4ed38d7SDouglas Fuller 4900a4ed38d7SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4901a4ed38d7SDouglas Fuller if (!req) 4902a4ed38d7SDouglas Fuller return -ENOMEM; 4903a4ed38d7SDouglas Fuller 4904a4ed38d7SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 4905a4ed38d7SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 4906a4ed38d7SDouglas Fuller req->r_flags = CEPH_OSD_FLAG_READ; 4907a4ed38d7SDouglas Fuller 4908a4ed38d7SDouglas Fuller ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4909a4ed38d7SDouglas Fuller if (ret) 4910a4ed38d7SDouglas Fuller goto out_put_req; 4911a4ed38d7SDouglas Fuller 4912a4ed38d7SDouglas Fuller pages = ceph_alloc_page_vector(1, GFP_NOIO); 4913a4ed38d7SDouglas Fuller if (IS_ERR(pages)) { 4914a4ed38d7SDouglas Fuller ret = PTR_ERR(pages); 4915a4ed38d7SDouglas Fuller goto out_put_req; 4916a4ed38d7SDouglas Fuller } 4917a4ed38d7SDouglas Fuller 4918a4ed38d7SDouglas Fuller osd_req_op_init(req, 0, CEPH_OSD_OP_LIST_WATCHERS, 0); 4919a4ed38d7SDouglas Fuller ceph_osd_data_pages_init(osd_req_op_data(req, 0, list_watchers, 4920a4ed38d7SDouglas Fuller response_data), 4921a4ed38d7SDouglas Fuller pages, PAGE_SIZE, 0, false, true); 4922a4ed38d7SDouglas Fuller 4923a4ed38d7SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 4924a4ed38d7SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 4925a4ed38d7SDouglas Fuller if (ret >= 0) { 4926a4ed38d7SDouglas Fuller void *p = page_address(pages[0]); 4927a4ed38d7SDouglas Fuller void *const end = p + req->r_ops[0].outdata_len; 4928a4ed38d7SDouglas Fuller 4929a4ed38d7SDouglas Fuller ret = decode_watchers(&p, end, watchers, num_watchers); 4930a4ed38d7SDouglas Fuller } 4931a4ed38d7SDouglas Fuller 4932a4ed38d7SDouglas Fuller out_put_req: 4933a4ed38d7SDouglas Fuller ceph_osdc_put_request(req); 4934a4ed38d7SDouglas Fuller return ret; 4935a4ed38d7SDouglas Fuller } 4936a4ed38d7SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_list_watchers); 4937a4ed38d7SDouglas Fuller 4938b07d3c4bSIlya Dryomov /* 4939dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 4940dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 4941dd935f44SJosh Durgin */ 4942f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 4943dd935f44SJosh Durgin { 494499d16943SIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 4945dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 4946dd935f44SJosh Durgin } 4947dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 4948dd935f44SJosh Durgin 49497cca78c9SIlya Dryomov void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc) 49507cca78c9SIlya Dryomov { 49517cca78c9SIlya Dryomov down_read(&osdc->lock); 49527cca78c9SIlya Dryomov maybe_request_map(osdc); 49537cca78c9SIlya Dryomov up_read(&osdc->lock); 49547cca78c9SIlya Dryomov } 49557cca78c9SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_maybe_request_map); 4956dd935f44SJosh Durgin 4957dd935f44SJosh Durgin /* 4958428a7158SDouglas Fuller * Execute an OSD class method on an object. 4959428a7158SDouglas Fuller * 4960428a7158SDouglas Fuller * @flags: CEPH_OSD_FLAG_* 49612544a020SIlya Dryomov * @resp_len: in/out param for reply length 4962428a7158SDouglas Fuller */ 4963428a7158SDouglas Fuller int ceph_osdc_call(struct ceph_osd_client *osdc, 4964428a7158SDouglas Fuller struct ceph_object_id *oid, 4965428a7158SDouglas Fuller struct ceph_object_locator *oloc, 4966428a7158SDouglas Fuller const char *class, const char *method, 4967428a7158SDouglas Fuller unsigned int flags, 4968428a7158SDouglas Fuller struct page *req_page, size_t req_len, 4969428a7158SDouglas Fuller struct page *resp_page, size_t *resp_len) 4970428a7158SDouglas Fuller { 4971428a7158SDouglas Fuller struct ceph_osd_request *req; 4972428a7158SDouglas Fuller int ret; 4973428a7158SDouglas Fuller 49742544a020SIlya Dryomov if (req_len > PAGE_SIZE || (resp_page && *resp_len > PAGE_SIZE)) 49752544a020SIlya Dryomov return -E2BIG; 49762544a020SIlya Dryomov 4977428a7158SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4978428a7158SDouglas Fuller if (!req) 4979428a7158SDouglas Fuller return -ENOMEM; 4980428a7158SDouglas Fuller 4981428a7158SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 4982428a7158SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 4983428a7158SDouglas Fuller req->r_flags = flags; 4984428a7158SDouglas Fuller 4985428a7158SDouglas Fuller ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4986428a7158SDouglas Fuller if (ret) 4987428a7158SDouglas Fuller goto out_put_req; 4988428a7158SDouglas Fuller 4989fe943d50SChengguang Xu ret = osd_req_op_cls_init(req, 0, CEPH_OSD_OP_CALL, class, method); 4990fe943d50SChengguang Xu if (ret) 4991fe943d50SChengguang Xu goto out_put_req; 4992fe943d50SChengguang Xu 4993428a7158SDouglas Fuller if (req_page) 4994428a7158SDouglas Fuller osd_req_op_cls_request_data_pages(req, 0, &req_page, req_len, 4995428a7158SDouglas Fuller 0, false, false); 4996428a7158SDouglas Fuller if (resp_page) 4997428a7158SDouglas Fuller osd_req_op_cls_response_data_pages(req, 0, &resp_page, 49982544a020SIlya Dryomov *resp_len, 0, false, false); 4999428a7158SDouglas Fuller 5000428a7158SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 5001428a7158SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 5002428a7158SDouglas Fuller if (ret >= 0) { 5003428a7158SDouglas Fuller ret = req->r_ops[0].rval; 5004428a7158SDouglas Fuller if (resp_page) 5005428a7158SDouglas Fuller *resp_len = req->r_ops[0].outdata_len; 5006428a7158SDouglas Fuller } 5007428a7158SDouglas Fuller 5008428a7158SDouglas Fuller out_put_req: 5009428a7158SDouglas Fuller ceph_osdc_put_request(req); 5010428a7158SDouglas Fuller return ret; 5011428a7158SDouglas Fuller } 5012428a7158SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_call); 5013428a7158SDouglas Fuller 5014428a7158SDouglas Fuller /* 50153d14c5d2SYehuda Sadeh * init, shutdown 50163d14c5d2SYehuda Sadeh */ 50173d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 50183d14c5d2SYehuda Sadeh { 50193d14c5d2SYehuda Sadeh int err; 50203d14c5d2SYehuda Sadeh 50213d14c5d2SYehuda Sadeh dout("init\n"); 50223d14c5d2SYehuda Sadeh osdc->client = client; 50235aea3dcdSIlya Dryomov init_rwsem(&osdc->lock); 50243d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 50253d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 50269dd2845cSIlya Dryomov spin_lock_init(&osdc->osd_lru_lock); 50275aea3dcdSIlya Dryomov osd_init(&osdc->homeless_osd); 50285aea3dcdSIlya Dryomov osdc->homeless_osd.o_osdc = osdc; 50295aea3dcdSIlya Dryomov osdc->homeless_osd.o_osd = CEPH_HOMELESS_OSD; 5030264048afSIlya Dryomov osdc->last_linger_id = CEPH_LINGER_ID_START; 5031922dab61SIlya Dryomov osdc->linger_requests = RB_ROOT; 50324609245eSIlya Dryomov osdc->map_checks = RB_ROOT; 50334609245eSIlya Dryomov osdc->linger_map_checks = RB_ROOT; 50343d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 50353d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 50363d14c5d2SYehuda Sadeh 50373d14c5d2SYehuda Sadeh err = -ENOMEM; 5038e5253a7bSIlya Dryomov osdc->osdmap = ceph_osdmap_alloc(); 5039e5253a7bSIlya Dryomov if (!osdc->osdmap) 5040e5253a7bSIlya Dryomov goto out; 5041e5253a7bSIlya Dryomov 50429e767adbSIlya Dryomov osdc->req_mempool = mempool_create_slab_pool(10, 50439e767adbSIlya Dryomov ceph_osd_request_cache); 50443d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 5045e5253a7bSIlya Dryomov goto out_map; 50463d14c5d2SYehuda Sadeh 5047d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 5048711da55dSIlya Dryomov PAGE_SIZE, 10, true, "osd_op"); 50493d14c5d2SYehuda Sadeh if (err < 0) 50503d14c5d2SYehuda Sadeh goto out_mempool; 5051d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 5052711da55dSIlya Dryomov PAGE_SIZE, 10, true, "osd_op_reply"); 50533d14c5d2SYehuda Sadeh if (err < 0) 50543d14c5d2SYehuda Sadeh goto out_msgpool; 5055a40c4f10SYehuda Sadeh 5056dbcae088SDan Carpenter err = -ENOMEM; 5057a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 5058dbcae088SDan Carpenter if (!osdc->notify_wq) 5059c172ec5cSIlya Dryomov goto out_msgpool_reply; 5060c172ec5cSIlya Dryomov 5061fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 5062fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 5063b37ee1b9SIlya Dryomov schedule_delayed_work(&osdc->osds_timeout_work, 5064b37ee1b9SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 5065b37ee1b9SIlya Dryomov 50663d14c5d2SYehuda Sadeh return 0; 50673d14c5d2SYehuda Sadeh 5068c172ec5cSIlya Dryomov out_msgpool_reply: 5069c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 50703d14c5d2SYehuda Sadeh out_msgpool: 50713d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 50723d14c5d2SYehuda Sadeh out_mempool: 50733d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 5074e5253a7bSIlya Dryomov out_map: 5075e5253a7bSIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 50763d14c5d2SYehuda Sadeh out: 50773d14c5d2SYehuda Sadeh return err; 50783d14c5d2SYehuda Sadeh } 50793d14c5d2SYehuda Sadeh 50803d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 50813d14c5d2SYehuda Sadeh { 5082a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 50833d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 50843d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 508542a2c09fSIlya Dryomov 50865aea3dcdSIlya Dryomov down_write(&osdc->lock); 508742a2c09fSIlya Dryomov while (!RB_EMPTY_ROOT(&osdc->osds)) { 508842a2c09fSIlya Dryomov struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 508942a2c09fSIlya Dryomov struct ceph_osd, o_node); 50905aea3dcdSIlya Dryomov close_osd(osd); 509142a2c09fSIlya Dryomov } 50925aea3dcdSIlya Dryomov up_write(&osdc->lock); 509302113a0fSElena Reshetova WARN_ON(refcount_read(&osdc->homeless_osd.o_ref) != 1); 50945aea3dcdSIlya Dryomov osd_cleanup(&osdc->homeless_osd); 50955aea3dcdSIlya Dryomov 50965aea3dcdSIlya Dryomov WARN_ON(!list_empty(&osdc->osd_lru)); 5097922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_requests)); 50984609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->map_checks)); 50994609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_map_checks)); 51005aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_requests)); 51015aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_homeless)); 510242a2c09fSIlya Dryomov 51033d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 51043d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 51053d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 51063d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 51073d14c5d2SYehuda Sadeh } 51083d14c5d2SYehuda Sadeh 51093d14c5d2SYehuda Sadeh /* 51103d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 51113d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 51123d14c5d2SYehuda Sadeh */ 51133d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 51143d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 51153d14c5d2SYehuda Sadeh u64 off, u64 *plen, 51163d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 5117b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 51183d14c5d2SYehuda Sadeh { 51193d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 51203d14c5d2SYehuda Sadeh int rc = 0; 51213d14c5d2SYehuda Sadeh 51223d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 51233d14c5d2SYehuda Sadeh vino.snap, off, *plen); 5124715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1, 51253d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 5126acead002SAlex Elder NULL, truncate_seq, truncate_size, 5127153e5167SAlex Elder false); 51286816282dSSage Weil if (IS_ERR(req)) 51296816282dSSage Weil return PTR_ERR(req); 51303d14c5d2SYehuda Sadeh 51313d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 5132406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 5133a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 51343d14c5d2SYehuda Sadeh 5135e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 513643bfe5deSAlex Elder off, *plen, *plen, page_align); 51373d14c5d2SYehuda Sadeh 51383d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 51393d14c5d2SYehuda Sadeh if (!rc) 51403d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 51413d14c5d2SYehuda Sadeh 51423d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 51433d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 51443d14c5d2SYehuda Sadeh return rc; 51453d14c5d2SYehuda Sadeh } 51463d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 51473d14c5d2SYehuda Sadeh 51483d14c5d2SYehuda Sadeh /* 51493d14c5d2SYehuda Sadeh * do a synchronous write on N pages 51503d14c5d2SYehuda Sadeh */ 51513d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 51523d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 51533d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 51543d14c5d2SYehuda Sadeh u64 off, u64 len, 51553d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 51563d14c5d2SYehuda Sadeh struct timespec *mtime, 515724808826SAlex Elder struct page **pages, int num_pages) 51583d14c5d2SYehuda Sadeh { 51593d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 51603d14c5d2SYehuda Sadeh int rc = 0; 5161b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 51623d14c5d2SYehuda Sadeh 5163715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 516454ea0046SIlya Dryomov CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, 5165acead002SAlex Elder snapc, truncate_seq, truncate_size, 5166153e5167SAlex Elder true); 51676816282dSSage Weil if (IS_ERR(req)) 51686816282dSSage Weil return PTR_ERR(req); 51693d14c5d2SYehuda Sadeh 51703d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 5171406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 517243bfe5deSAlex Elder false, false); 517343bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 51743d14c5d2SYehuda Sadeh 5175bb873b53SIlya Dryomov req->r_mtime = *mtime; 517687f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 51773d14c5d2SYehuda Sadeh if (!rc) 51783d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 51793d14c5d2SYehuda Sadeh 51803d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 51813d14c5d2SYehuda Sadeh if (rc == 0) 51823d14c5d2SYehuda Sadeh rc = len; 51833d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 51843d14c5d2SYehuda Sadeh return rc; 51853d14c5d2SYehuda Sadeh } 51863d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 51873d14c5d2SYehuda Sadeh 518857a35dfbSChengguang Xu int __init ceph_osdc_setup(void) 51895522ae0bSAlex Elder { 51903f1af42aSIlya Dryomov size_t size = sizeof(struct ceph_osd_request) + 51913f1af42aSIlya Dryomov CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op); 51923f1af42aSIlya Dryomov 51935522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 51943f1af42aSIlya Dryomov ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size, 51953f1af42aSIlya Dryomov 0, 0, NULL); 51965522ae0bSAlex Elder 51975522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 51985522ae0bSAlex Elder } 51995522ae0bSAlex Elder 52005522ae0bSAlex Elder void ceph_osdc_cleanup(void) 52015522ae0bSAlex Elder { 52025522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 52035522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 52045522ae0bSAlex Elder ceph_osd_request_cache = NULL; 52055522ae0bSAlex Elder } 52065522ae0bSAlex Elder 52073d14c5d2SYehuda Sadeh /* 52083d14c5d2SYehuda Sadeh * handle incoming message 52093d14c5d2SYehuda Sadeh */ 52103d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 52113d14c5d2SYehuda Sadeh { 52123d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 52135aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 52143d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 52153d14c5d2SYehuda Sadeh 52163d14c5d2SYehuda Sadeh switch (type) { 52173d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 52183d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 52193d14c5d2SYehuda Sadeh break; 52203d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 52215aea3dcdSIlya Dryomov handle_reply(osd, msg); 52223d14c5d2SYehuda Sadeh break; 5223a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5224a02a946dSIlya Dryomov handle_backoff(osd, msg); 5225a02a946dSIlya Dryomov break; 5226a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 5227a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 5228a40c4f10SYehuda Sadeh break; 52293d14c5d2SYehuda Sadeh 52303d14c5d2SYehuda Sadeh default: 52313d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 52323d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 52333d14c5d2SYehuda Sadeh } 52345aea3dcdSIlya Dryomov 52353d14c5d2SYehuda Sadeh ceph_msg_put(msg); 52363d14c5d2SYehuda Sadeh } 52373d14c5d2SYehuda Sadeh 52383d14c5d2SYehuda Sadeh /* 5239d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 5240d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 5241d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 52423d14c5d2SYehuda Sadeh */ 52433d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 52443d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 52453d14c5d2SYehuda Sadeh int *skip) 52463d14c5d2SYehuda Sadeh { 52473d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 52483d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 52495aea3dcdSIlya Dryomov struct ceph_msg *m = NULL; 52503d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 52513f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 52523d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 52535aea3dcdSIlya Dryomov u64 tid = le64_to_cpu(hdr->tid); 52543d14c5d2SYehuda Sadeh 52555aea3dcdSIlya Dryomov down_read(&osdc->lock); 52565aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 52575aea3dcdSIlya Dryomov dout("%s osd%d unknown, skipping\n", __func__, osd->o_osd); 52585aea3dcdSIlya Dryomov *skip = 1; 52595aea3dcdSIlya Dryomov goto out_unlock_osdc; 52605aea3dcdSIlya Dryomov } 52615aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(hdr->src.num)); 52625aea3dcdSIlya Dryomov 52635aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 52645aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 52653d14c5d2SYehuda Sadeh if (!req) { 5266cd8140c6SIlya Dryomov dout("%s osd%d tid %llu unknown, skipping\n", __func__, 5267cd8140c6SIlya Dryomov osd->o_osd, tid); 5268d15f9d69SIlya Dryomov *skip = 1; 52695aea3dcdSIlya Dryomov goto out_unlock_session; 52703d14c5d2SYehuda Sadeh } 52713d14c5d2SYehuda Sadeh 52728921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 52733d14c5d2SYehuda Sadeh 5274f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 5275d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 5276d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 5277d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 52783f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 52793f0a4ac5SIlya Dryomov false); 52803d14c5d2SYehuda Sadeh if (!m) 52815aea3dcdSIlya Dryomov goto out_unlock_session; 52823d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 52833d14c5d2SYehuda Sadeh req->r_reply = m; 52843d14c5d2SYehuda Sadeh } 52853d14c5d2SYehuda Sadeh 5286d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 5287d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 5288d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 5289d15f9d69SIlya Dryomov req->r_reply->data_length); 52903d14c5d2SYehuda Sadeh m = NULL; 5291d15f9d69SIlya Dryomov *skip = 1; 52925aea3dcdSIlya Dryomov goto out_unlock_session; 52933d14c5d2SYehuda Sadeh } 5294d15f9d69SIlya Dryomov 5295d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 52963d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 52973d14c5d2SYehuda Sadeh 52985aea3dcdSIlya Dryomov out_unlock_session: 52995aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 53005aea3dcdSIlya Dryomov out_unlock_osdc: 53015aea3dcdSIlya Dryomov up_read(&osdc->lock); 53023d14c5d2SYehuda Sadeh return m; 53033d14c5d2SYehuda Sadeh } 53043d14c5d2SYehuda Sadeh 530519079203SIlya Dryomov /* 530619079203SIlya Dryomov * TODO: switch to a msg-owned pagelist 530719079203SIlya Dryomov */ 530819079203SIlya Dryomov static struct ceph_msg *alloc_msg_with_page_vector(struct ceph_msg_header *hdr) 530919079203SIlya Dryomov { 531019079203SIlya Dryomov struct ceph_msg *m; 531119079203SIlya Dryomov int type = le16_to_cpu(hdr->type); 531219079203SIlya Dryomov u32 front_len = le32_to_cpu(hdr->front_len); 531319079203SIlya Dryomov u32 data_len = le32_to_cpu(hdr->data_len); 531419079203SIlya Dryomov 531519079203SIlya Dryomov m = ceph_msg_new(type, front_len, GFP_NOIO, false); 531619079203SIlya Dryomov if (!m) 531719079203SIlya Dryomov return NULL; 531819079203SIlya Dryomov 531919079203SIlya Dryomov if (data_len) { 532019079203SIlya Dryomov struct page **pages; 532119079203SIlya Dryomov struct ceph_osd_data osd_data; 532219079203SIlya Dryomov 532319079203SIlya Dryomov pages = ceph_alloc_page_vector(calc_pages_for(0, data_len), 532419079203SIlya Dryomov GFP_NOIO); 5325c22e853aSWei Yongjun if (IS_ERR(pages)) { 532619079203SIlya Dryomov ceph_msg_put(m); 532719079203SIlya Dryomov return NULL; 532819079203SIlya Dryomov } 532919079203SIlya Dryomov 533019079203SIlya Dryomov ceph_osd_data_pages_init(&osd_data, pages, data_len, 0, false, 533119079203SIlya Dryomov false); 533219079203SIlya Dryomov ceph_osdc_msg_data_add(m, &osd_data); 533319079203SIlya Dryomov } 533419079203SIlya Dryomov 533519079203SIlya Dryomov return m; 533619079203SIlya Dryomov } 533719079203SIlya Dryomov 53383d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 53393d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 53403d14c5d2SYehuda Sadeh int *skip) 53413d14c5d2SYehuda Sadeh { 53423d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 53433d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 53443d14c5d2SYehuda Sadeh 53451c20f2d2SAlex Elder *skip = 0; 53463d14c5d2SYehuda Sadeh switch (type) { 53473d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 5348a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5349a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 535019079203SIlya Dryomov return alloc_msg_with_page_vector(hdr); 53513d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 53523d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 53533d14c5d2SYehuda Sadeh default: 53545aea3dcdSIlya Dryomov pr_warn("%s osd%d unknown msg type %d, skipping\n", __func__, 53555aea3dcdSIlya Dryomov osd->o_osd, type); 53563d14c5d2SYehuda Sadeh *skip = 1; 53573d14c5d2SYehuda Sadeh return NULL; 53583d14c5d2SYehuda Sadeh } 53593d14c5d2SYehuda Sadeh } 53603d14c5d2SYehuda Sadeh 53613d14c5d2SYehuda Sadeh /* 53623d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 53633d14c5d2SYehuda Sadeh */ 53643d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 53653d14c5d2SYehuda Sadeh { 53663d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 53673d14c5d2SYehuda Sadeh if (get_osd(osd)) 53683d14c5d2SYehuda Sadeh return con; 53693d14c5d2SYehuda Sadeh return NULL; 53703d14c5d2SYehuda Sadeh } 53713d14c5d2SYehuda Sadeh 53723d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 53733d14c5d2SYehuda Sadeh { 53743d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 53753d14c5d2SYehuda Sadeh put_osd(osd); 53763d14c5d2SYehuda Sadeh } 53773d14c5d2SYehuda Sadeh 53783d14c5d2SYehuda Sadeh /* 53793d14c5d2SYehuda Sadeh * authentication 53803d14c5d2SYehuda Sadeh */ 5381a3530df3SAlex Elder /* 5382a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 5383a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 5384a3530df3SAlex Elder */ 5385a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 53868f43fb53SAlex Elder int *proto, int force_new) 53873d14c5d2SYehuda Sadeh { 53883d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 53893d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 53903d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 539174f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 53923d14c5d2SYehuda Sadeh 539374f1869fSAlex Elder if (force_new && auth->authorizer) { 53946c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(auth->authorizer); 539574f1869fSAlex Elder auth->authorizer = NULL; 53963d14c5d2SYehuda Sadeh } 539727859f97SSage Weil if (!auth->authorizer) { 539827859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 5399a3530df3SAlex Elder auth); 54003d14c5d2SYehuda Sadeh if (ret) 5401a3530df3SAlex Elder return ERR_PTR(ret); 540227859f97SSage Weil } else { 540327859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 54040bed9b5cSSage Weil auth); 54050bed9b5cSSage Weil if (ret) 54060bed9b5cSSage Weil return ERR_PTR(ret); 54073d14c5d2SYehuda Sadeh } 54083d14c5d2SYehuda Sadeh *proto = ac->protocol; 540974f1869fSAlex Elder 5410a3530df3SAlex Elder return auth; 54113d14c5d2SYehuda Sadeh } 54123d14c5d2SYehuda Sadeh 54133d14c5d2SYehuda Sadeh 54140dde5848SIlya Dryomov static int verify_authorizer_reply(struct ceph_connection *con) 54153d14c5d2SYehuda Sadeh { 54163d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 54173d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 54183d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 54193d14c5d2SYehuda Sadeh 54200dde5848SIlya Dryomov return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer); 54213d14c5d2SYehuda Sadeh } 54223d14c5d2SYehuda Sadeh 54233d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 54243d14c5d2SYehuda Sadeh { 54253d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 54263d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 54273d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 54283d14c5d2SYehuda Sadeh 542927859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 54303d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 54313d14c5d2SYehuda Sadeh } 54323d14c5d2SYehuda Sadeh 54338cb441c0SIlya Dryomov static void osd_reencode_message(struct ceph_msg *msg) 54348cb441c0SIlya Dryomov { 5435914902afSIlya Dryomov int type = le16_to_cpu(msg->hdr.type); 5436914902afSIlya Dryomov 5437914902afSIlya Dryomov if (type == CEPH_MSG_OSD_OP) 54388cb441c0SIlya Dryomov encode_request_finish(msg); 54398cb441c0SIlya Dryomov } 54408cb441c0SIlya Dryomov 544179dbd1baSIlya Dryomov static int osd_sign_message(struct ceph_msg *msg) 544233d07337SYan, Zheng { 544379dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 544433d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 544579dbd1baSIlya Dryomov 544633d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 544733d07337SYan, Zheng } 544833d07337SYan, Zheng 544979dbd1baSIlya Dryomov static int osd_check_message_signature(struct ceph_msg *msg) 545033d07337SYan, Zheng { 545179dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 545233d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 545379dbd1baSIlya Dryomov 545433d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 545533d07337SYan, Zheng } 545633d07337SYan, Zheng 54573d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 54583d14c5d2SYehuda Sadeh .get = get_osd_con, 54593d14c5d2SYehuda Sadeh .put = put_osd_con, 54603d14c5d2SYehuda Sadeh .dispatch = dispatch, 54613d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 54623d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 54633d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 54643d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 54658cb441c0SIlya Dryomov .reencode_message = osd_reencode_message, 546679dbd1baSIlya Dryomov .sign_message = osd_sign_message, 546779dbd1baSIlya Dryomov .check_message_signature = osd_check_message_signature, 54685aea3dcdSIlya Dryomov .fault = osd_fault, 54693d14c5d2SYehuda Sadeh }; 5470