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, 160b9e281c2SIlya Dryomov struct ceph_bvec_iter *bvec_pos) 161b9e281c2SIlya Dryomov { 162b9e281c2SIlya Dryomov osd_data->type = CEPH_OSD_DATA_TYPE_BVECS; 163b9e281c2SIlya Dryomov osd_data->bvec_pos = *bvec_pos; 164b9e281c2SIlya Dryomov } 165b9e281c2SIlya Dryomov 166863c7eb5SAlex Elder #define osd_req_op_data(oreq, whch, typ, fld) \ 167863c7eb5SAlex Elder ({ \ 1688a703a38SIoana Ciornei struct ceph_osd_request *__oreq = (oreq); \ 1698a703a38SIoana Ciornei unsigned int __whch = (whch); \ 1708a703a38SIoana Ciornei BUG_ON(__whch >= __oreq->r_num_ops); \ 1718a703a38SIoana Ciornei &__oreq->r_ops[__whch].typ.fld; \ 172863c7eb5SAlex Elder }) 173863c7eb5SAlex Elder 17449719778SAlex Elder static struct ceph_osd_data * 17549719778SAlex Elder osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which) 17649719778SAlex Elder { 17749719778SAlex Elder BUG_ON(which >= osd_req->r_num_ops); 17849719778SAlex Elder 17949719778SAlex Elder return &osd_req->r_ops[which].raw_data_in; 18049719778SAlex Elder } 18149719778SAlex Elder 182a4ce40a9SAlex Elder struct ceph_osd_data * 183a4ce40a9SAlex Elder osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req, 184406e2c9fSAlex Elder unsigned int which) 185a4ce40a9SAlex Elder { 186863c7eb5SAlex Elder return osd_req_op_data(osd_req, which, extent, osd_data); 187a4ce40a9SAlex Elder } 188a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data); 189a4ce40a9SAlex Elder 19049719778SAlex Elder void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req, 19149719778SAlex Elder unsigned int which, struct page **pages, 19249719778SAlex Elder u64 length, u32 alignment, 19349719778SAlex Elder bool pages_from_pool, bool own_pages) 19449719778SAlex Elder { 19549719778SAlex Elder struct ceph_osd_data *osd_data; 19649719778SAlex Elder 19749719778SAlex Elder osd_data = osd_req_op_raw_data_in(osd_req, which); 19849719778SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 19949719778SAlex Elder pages_from_pool, own_pages); 20049719778SAlex Elder } 20149719778SAlex Elder EXPORT_SYMBOL(osd_req_op_raw_data_in_pages); 20249719778SAlex Elder 203a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req, 204406e2c9fSAlex Elder unsigned int which, struct page **pages, 205406e2c9fSAlex Elder u64 length, u32 alignment, 206a4ce40a9SAlex Elder bool pages_from_pool, bool own_pages) 207a4ce40a9SAlex Elder { 208a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 209a4ce40a9SAlex Elder 210863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 211a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 212a4ce40a9SAlex Elder pages_from_pool, own_pages); 213a4ce40a9SAlex Elder } 214a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages); 215a4ce40a9SAlex Elder 216a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req, 217406e2c9fSAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 218a4ce40a9SAlex Elder { 219a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 220a4ce40a9SAlex Elder 221863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 222a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 223a4ce40a9SAlex Elder } 224a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist); 225a4ce40a9SAlex Elder 226a4ce40a9SAlex Elder #ifdef CONFIG_BLOCK 227a4ce40a9SAlex Elder void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req, 2285359a17dSIlya Dryomov unsigned int which, 2295359a17dSIlya Dryomov struct ceph_bio_iter *bio_pos, 2305359a17dSIlya Dryomov u32 bio_length) 231a4ce40a9SAlex Elder { 232a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 233863c7eb5SAlex Elder 234863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 2355359a17dSIlya Dryomov ceph_osd_data_bio_init(osd_data, bio_pos, bio_length); 236a4ce40a9SAlex Elder } 237a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio); 238a4ce40a9SAlex Elder #endif /* CONFIG_BLOCK */ 239a4ce40a9SAlex Elder 240b9e281c2SIlya Dryomov void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req, 241b9e281c2SIlya Dryomov unsigned int which, 242b9e281c2SIlya Dryomov struct ceph_bvec_iter *bvec_pos) 243b9e281c2SIlya Dryomov { 244b9e281c2SIlya Dryomov struct ceph_osd_data *osd_data; 245b9e281c2SIlya Dryomov 246b9e281c2SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, extent, osd_data); 247b9e281c2SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, bvec_pos); 248b9e281c2SIlya Dryomov } 249b9e281c2SIlya Dryomov EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvec_pos); 250b9e281c2SIlya Dryomov 251a4ce40a9SAlex Elder static void osd_req_op_cls_request_info_pagelist( 252a4ce40a9SAlex Elder struct ceph_osd_request *osd_req, 253a4ce40a9SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 254a4ce40a9SAlex Elder { 255a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 256a4ce40a9SAlex Elder 257863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_info); 258a4ce40a9SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 259a4ce40a9SAlex Elder } 260a4ce40a9SAlex Elder 26104017e29SAlex Elder void osd_req_op_cls_request_data_pagelist( 26204017e29SAlex Elder struct ceph_osd_request *osd_req, 26304017e29SAlex Elder unsigned int which, struct ceph_pagelist *pagelist) 26404017e29SAlex Elder { 26504017e29SAlex Elder struct ceph_osd_data *osd_data; 26604017e29SAlex Elder 267863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 26804017e29SAlex Elder ceph_osd_data_pagelist_init(osd_data, pagelist); 269bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += pagelist->length; 270bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += pagelist->length; 27104017e29SAlex Elder } 27204017e29SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist); 27304017e29SAlex Elder 2746c57b554SAlex Elder void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req, 2756c57b554SAlex Elder unsigned int which, struct page **pages, u64 length, 2766c57b554SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 2776c57b554SAlex Elder { 2786c57b554SAlex Elder struct ceph_osd_data *osd_data; 2796c57b554SAlex Elder 2806c57b554SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, request_data); 2816c57b554SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 2826c57b554SAlex Elder pages_from_pool, own_pages); 283bb873b53SIlya Dryomov osd_req->r_ops[which].cls.indata_len += length; 284bb873b53SIlya Dryomov osd_req->r_ops[which].indata_len += length; 2856c57b554SAlex Elder } 2866c57b554SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_request_data_pages); 2876c57b554SAlex Elder 288b9e281c2SIlya Dryomov void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req, 289b9e281c2SIlya Dryomov unsigned int which, 290b9e281c2SIlya Dryomov struct bio_vec *bvecs, u32 bytes) 291b9e281c2SIlya Dryomov { 292b9e281c2SIlya Dryomov struct ceph_osd_data *osd_data; 293b9e281c2SIlya Dryomov struct ceph_bvec_iter it = { 294b9e281c2SIlya Dryomov .bvecs = bvecs, 295b9e281c2SIlya Dryomov .iter = { .bi_size = bytes }, 296b9e281c2SIlya Dryomov }; 297b9e281c2SIlya Dryomov 298b9e281c2SIlya Dryomov osd_data = osd_req_op_data(osd_req, which, cls, request_data); 299b9e281c2SIlya Dryomov ceph_osd_data_bvecs_init(osd_data, &it); 300b9e281c2SIlya Dryomov osd_req->r_ops[which].cls.indata_len += bytes; 301b9e281c2SIlya Dryomov osd_req->r_ops[which].indata_len += bytes; 302b9e281c2SIlya Dryomov } 303b9e281c2SIlya Dryomov EXPORT_SYMBOL(osd_req_op_cls_request_data_bvecs); 304b9e281c2SIlya Dryomov 305a4ce40a9SAlex Elder void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req, 306a4ce40a9SAlex Elder unsigned int which, struct page **pages, u64 length, 307a4ce40a9SAlex Elder u32 alignment, bool pages_from_pool, bool own_pages) 308a4ce40a9SAlex Elder { 309a4ce40a9SAlex Elder struct ceph_osd_data *osd_data; 310a4ce40a9SAlex Elder 311863c7eb5SAlex Elder osd_data = osd_req_op_data(osd_req, which, cls, response_data); 312a4ce40a9SAlex Elder ceph_osd_data_pages_init(osd_data, pages, length, alignment, 313a4ce40a9SAlex Elder pages_from_pool, own_pages); 314a4ce40a9SAlex Elder } 315a4ce40a9SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_response_data_pages); 316a4ce40a9SAlex Elder 31723c08a9cSAlex Elder static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data) 31823c08a9cSAlex Elder { 31923c08a9cSAlex Elder switch (osd_data->type) { 32023c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_NONE: 32123c08a9cSAlex Elder return 0; 32223c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGES: 32323c08a9cSAlex Elder return osd_data->length; 32423c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_PAGELIST: 32523c08a9cSAlex Elder return (u64)osd_data->pagelist->length; 32623c08a9cSAlex Elder #ifdef CONFIG_BLOCK 32723c08a9cSAlex Elder case CEPH_OSD_DATA_TYPE_BIO: 32823c08a9cSAlex Elder return (u64)osd_data->bio_length; 32923c08a9cSAlex Elder #endif /* CONFIG_BLOCK */ 330b9e281c2SIlya Dryomov case CEPH_OSD_DATA_TYPE_BVECS: 331b9e281c2SIlya Dryomov return osd_data->bvec_pos.iter.bi_size; 33223c08a9cSAlex Elder default: 33323c08a9cSAlex Elder WARN(true, "unrecognized data type %d\n", (int)osd_data->type); 33423c08a9cSAlex Elder return 0; 33523c08a9cSAlex Elder } 33623c08a9cSAlex Elder } 33723c08a9cSAlex Elder 338c54d47bfSAlex Elder static void ceph_osd_data_release(struct ceph_osd_data *osd_data) 339c54d47bfSAlex Elder { 3405476492fSAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) { 341c54d47bfSAlex Elder int num_pages; 342c54d47bfSAlex Elder 343c54d47bfSAlex Elder num_pages = calc_pages_for((u64)osd_data->alignment, 344c54d47bfSAlex Elder (u64)osd_data->length); 345c54d47bfSAlex Elder ceph_release_page_vector(osd_data->pages, num_pages); 346c54d47bfSAlex Elder } 3475476492fSAlex Elder ceph_osd_data_init(osd_data); 3485476492fSAlex Elder } 3495476492fSAlex Elder 3505476492fSAlex Elder static void osd_req_op_data_release(struct ceph_osd_request *osd_req, 3515476492fSAlex Elder unsigned int which) 3525476492fSAlex Elder { 3535476492fSAlex Elder struct ceph_osd_req_op *op; 3545476492fSAlex Elder 3555476492fSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 3565476492fSAlex Elder op = &osd_req->r_ops[which]; 3575476492fSAlex Elder 3585476492fSAlex Elder switch (op->op) { 3595476492fSAlex Elder case CEPH_OSD_OP_READ: 3605476492fSAlex Elder case CEPH_OSD_OP_WRITE: 361e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 3625476492fSAlex Elder ceph_osd_data_release(&op->extent.osd_data); 3635476492fSAlex Elder break; 3645476492fSAlex Elder case CEPH_OSD_OP_CALL: 3655476492fSAlex Elder ceph_osd_data_release(&op->cls.request_info); 36604017e29SAlex Elder ceph_osd_data_release(&op->cls.request_data); 3675476492fSAlex Elder ceph_osd_data_release(&op->cls.response_data); 3685476492fSAlex Elder break; 369d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 370d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 371d74b50beSYan, Zheng ceph_osd_data_release(&op->xattr.osd_data); 372d74b50beSYan, Zheng break; 37366ba609fSYan, Zheng case CEPH_OSD_OP_STAT: 37466ba609fSYan, Zheng ceph_osd_data_release(&op->raw_data_in); 37566ba609fSYan, Zheng break; 376922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 377922dab61SIlya Dryomov ceph_osd_data_release(&op->notify_ack.request_data); 378922dab61SIlya Dryomov break; 37919079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 38019079203SIlya Dryomov ceph_osd_data_release(&op->notify.request_data); 38119079203SIlya Dryomov ceph_osd_data_release(&op->notify.response_data); 38219079203SIlya Dryomov break; 383a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 384a4ed38d7SDouglas Fuller ceph_osd_data_release(&op->list_watchers.response_data); 385a4ed38d7SDouglas Fuller break; 3865476492fSAlex Elder default: 3875476492fSAlex Elder break; 3885476492fSAlex Elder } 389c54d47bfSAlex Elder } 390c54d47bfSAlex Elder 3913d14c5d2SYehuda Sadeh /* 39263244fa1SIlya Dryomov * Assumes @t is zero-initialized. 39363244fa1SIlya Dryomov */ 39463244fa1SIlya Dryomov static void target_init(struct ceph_osd_request_target *t) 39563244fa1SIlya Dryomov { 39663244fa1SIlya Dryomov ceph_oid_init(&t->base_oid); 39763244fa1SIlya Dryomov ceph_oloc_init(&t->base_oloc); 39863244fa1SIlya Dryomov ceph_oid_init(&t->target_oid); 39963244fa1SIlya Dryomov ceph_oloc_init(&t->target_oloc); 40063244fa1SIlya Dryomov 40163244fa1SIlya Dryomov ceph_osds_init(&t->acting); 40263244fa1SIlya Dryomov ceph_osds_init(&t->up); 40363244fa1SIlya Dryomov t->size = -1; 40463244fa1SIlya Dryomov t->min_size = -1; 40563244fa1SIlya Dryomov 40663244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 40763244fa1SIlya Dryomov } 40863244fa1SIlya Dryomov 409922dab61SIlya Dryomov static void target_copy(struct ceph_osd_request_target *dest, 410922dab61SIlya Dryomov const struct ceph_osd_request_target *src) 411922dab61SIlya Dryomov { 412922dab61SIlya Dryomov ceph_oid_copy(&dest->base_oid, &src->base_oid); 413922dab61SIlya Dryomov ceph_oloc_copy(&dest->base_oloc, &src->base_oloc); 414922dab61SIlya Dryomov ceph_oid_copy(&dest->target_oid, &src->target_oid); 415922dab61SIlya Dryomov ceph_oloc_copy(&dest->target_oloc, &src->target_oloc); 416922dab61SIlya Dryomov 417922dab61SIlya Dryomov dest->pgid = src->pgid; /* struct */ 418dc98ff72SIlya Dryomov dest->spgid = src->spgid; /* struct */ 419922dab61SIlya Dryomov dest->pg_num = src->pg_num; 420922dab61SIlya Dryomov dest->pg_num_mask = src->pg_num_mask; 421922dab61SIlya Dryomov ceph_osds_copy(&dest->acting, &src->acting); 422922dab61SIlya Dryomov ceph_osds_copy(&dest->up, &src->up); 423922dab61SIlya Dryomov dest->size = src->size; 424922dab61SIlya Dryomov dest->min_size = src->min_size; 425922dab61SIlya Dryomov dest->sort_bitwise = src->sort_bitwise; 426922dab61SIlya Dryomov 427922dab61SIlya Dryomov dest->flags = src->flags; 428922dab61SIlya Dryomov dest->paused = src->paused; 429922dab61SIlya Dryomov 43004c7d789SIlya Dryomov dest->epoch = src->epoch; 431dc93e0e2SIlya Dryomov dest->last_force_resend = src->last_force_resend; 432dc93e0e2SIlya Dryomov 433922dab61SIlya Dryomov dest->osd = src->osd; 434922dab61SIlya Dryomov } 435922dab61SIlya Dryomov 43663244fa1SIlya Dryomov static void target_destroy(struct ceph_osd_request_target *t) 43763244fa1SIlya Dryomov { 43863244fa1SIlya Dryomov ceph_oid_destroy(&t->base_oid); 43930c156d9SYan, Zheng ceph_oloc_destroy(&t->base_oloc); 44063244fa1SIlya Dryomov ceph_oid_destroy(&t->target_oid); 44130c156d9SYan, Zheng ceph_oloc_destroy(&t->target_oloc); 44263244fa1SIlya Dryomov } 44363244fa1SIlya Dryomov 44463244fa1SIlya Dryomov /* 4453d14c5d2SYehuda Sadeh * requests 4463d14c5d2SYehuda Sadeh */ 4473540bfdbSIlya Dryomov static void request_release_checks(struct ceph_osd_request *req) 4483540bfdbSIlya Dryomov { 4493540bfdbSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_node)); 4504609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&req->r_mc_node)); 4513540bfdbSIlya Dryomov WARN_ON(!list_empty(&req->r_unsafe_item)); 4523540bfdbSIlya Dryomov WARN_ON(req->r_osd); 4533540bfdbSIlya Dryomov } 4543540bfdbSIlya Dryomov 4559e94af20SIlya Dryomov static void ceph_osdc_release_request(struct kref *kref) 4563d14c5d2SYehuda Sadeh { 4579e94af20SIlya Dryomov struct ceph_osd_request *req = container_of(kref, 4589e94af20SIlya Dryomov struct ceph_osd_request, r_kref); 4595476492fSAlex Elder unsigned int which; 4603d14c5d2SYehuda Sadeh 4619e94af20SIlya Dryomov dout("%s %p (r_request %p r_reply %p)\n", __func__, req, 4629e94af20SIlya Dryomov req->r_request, req->r_reply); 4633540bfdbSIlya Dryomov request_release_checks(req); 4649e94af20SIlya Dryomov 4653d14c5d2SYehuda Sadeh if (req->r_request) 4663d14c5d2SYehuda Sadeh ceph_msg_put(req->r_request); 4675aea3dcdSIlya Dryomov if (req->r_reply) 468ab8cb34aSAlex Elder ceph_msg_put(req->r_reply); 4690fff87ecSAlex Elder 4705476492fSAlex Elder for (which = 0; which < req->r_num_ops; which++) 4715476492fSAlex Elder osd_req_op_data_release(req, which); 4720fff87ecSAlex Elder 473a66dd383SIlya Dryomov target_destroy(&req->r_t); 4743d14c5d2SYehuda Sadeh ceph_put_snap_context(req->r_snapc); 475d30291b9SIlya Dryomov 4763d14c5d2SYehuda Sadeh if (req->r_mempool) 4773d14c5d2SYehuda Sadeh mempool_free(req, req->r_osdc->req_mempool); 4783f1af42aSIlya Dryomov else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS) 4795522ae0bSAlex Elder kmem_cache_free(ceph_osd_request_cache, req); 4803f1af42aSIlya Dryomov else 4813f1af42aSIlya Dryomov kfree(req); 4823d14c5d2SYehuda Sadeh } 4839e94af20SIlya Dryomov 4849e94af20SIlya Dryomov void ceph_osdc_get_request(struct ceph_osd_request *req) 4859e94af20SIlya Dryomov { 4869e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 4872c935bc5SPeter Zijlstra kref_read(&req->r_kref)); 4889e94af20SIlya Dryomov kref_get(&req->r_kref); 4899e94af20SIlya Dryomov } 4909e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_get_request); 4919e94af20SIlya Dryomov 4929e94af20SIlya Dryomov void ceph_osdc_put_request(struct ceph_osd_request *req) 4939e94af20SIlya Dryomov { 4943ed97d63SIlya Dryomov if (req) { 4959e94af20SIlya Dryomov dout("%s %p (was %d)\n", __func__, req, 4962c935bc5SPeter Zijlstra kref_read(&req->r_kref)); 4979e94af20SIlya Dryomov kref_put(&req->r_kref, ceph_osdc_release_request); 4989e94af20SIlya Dryomov } 4993ed97d63SIlya Dryomov } 5009e94af20SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_put_request); 5013d14c5d2SYehuda Sadeh 5023540bfdbSIlya Dryomov static void request_init(struct ceph_osd_request *req) 5033540bfdbSIlya Dryomov { 5043540bfdbSIlya Dryomov /* req only, each op is zeroed in _osd_req_op_init() */ 5053540bfdbSIlya Dryomov memset(req, 0, sizeof(*req)); 5063540bfdbSIlya Dryomov 5073540bfdbSIlya Dryomov kref_init(&req->r_kref); 5083540bfdbSIlya Dryomov init_completion(&req->r_completion); 5093540bfdbSIlya Dryomov RB_CLEAR_NODE(&req->r_node); 5104609245eSIlya Dryomov RB_CLEAR_NODE(&req->r_mc_node); 5113540bfdbSIlya Dryomov INIT_LIST_HEAD(&req->r_unsafe_item); 5123540bfdbSIlya Dryomov 5133540bfdbSIlya Dryomov target_init(&req->r_t); 5143540bfdbSIlya Dryomov } 5153540bfdbSIlya Dryomov 516922dab61SIlya Dryomov /* 517922dab61SIlya Dryomov * This is ugly, but it allows us to reuse linger registration and ping 518922dab61SIlya Dryomov * requests, keeping the structure of the code around send_linger{_ping}() 519922dab61SIlya Dryomov * reasonable. Setting up a min_nr=2 mempool for each linger request 520922dab61SIlya Dryomov * and dealing with copying ops (this blasts req only, watch op remains 521922dab61SIlya Dryomov * intact) isn't any better. 522922dab61SIlya Dryomov */ 523922dab61SIlya Dryomov static void request_reinit(struct ceph_osd_request *req) 524922dab61SIlya Dryomov { 525922dab61SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 526922dab61SIlya Dryomov bool mempool = req->r_mempool; 527922dab61SIlya Dryomov unsigned int num_ops = req->r_num_ops; 528922dab61SIlya Dryomov u64 snapid = req->r_snapid; 529922dab61SIlya Dryomov struct ceph_snap_context *snapc = req->r_snapc; 530922dab61SIlya Dryomov bool linger = req->r_linger; 531922dab61SIlya Dryomov struct ceph_msg *request_msg = req->r_request; 532922dab61SIlya Dryomov struct ceph_msg *reply_msg = req->r_reply; 533922dab61SIlya Dryomov 534922dab61SIlya Dryomov dout("%s req %p\n", __func__, req); 5352c935bc5SPeter Zijlstra WARN_ON(kref_read(&req->r_kref) != 1); 536922dab61SIlya Dryomov request_release_checks(req); 537922dab61SIlya Dryomov 5382c935bc5SPeter Zijlstra WARN_ON(kref_read(&request_msg->kref) != 1); 5392c935bc5SPeter Zijlstra WARN_ON(kref_read(&reply_msg->kref) != 1); 540922dab61SIlya Dryomov target_destroy(&req->r_t); 541922dab61SIlya Dryomov 542922dab61SIlya Dryomov request_init(req); 543922dab61SIlya Dryomov req->r_osdc = osdc; 544922dab61SIlya Dryomov req->r_mempool = mempool; 545922dab61SIlya Dryomov req->r_num_ops = num_ops; 546922dab61SIlya Dryomov req->r_snapid = snapid; 547922dab61SIlya Dryomov req->r_snapc = snapc; 548922dab61SIlya Dryomov req->r_linger = linger; 549922dab61SIlya Dryomov req->r_request = request_msg; 550922dab61SIlya Dryomov req->r_reply = reply_msg; 551922dab61SIlya Dryomov } 552922dab61SIlya Dryomov 5533d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, 5543d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 5551b83bef2SSage Weil unsigned int num_ops, 5563d14c5d2SYehuda Sadeh bool use_mempool, 55754a54007SAlex Elder gfp_t gfp_flags) 5583d14c5d2SYehuda Sadeh { 5593d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 5603d14c5d2SYehuda Sadeh 5613d14c5d2SYehuda Sadeh if (use_mempool) { 5623f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_SLAB_OPS); 5633d14c5d2SYehuda Sadeh req = mempool_alloc(osdc->req_mempool, gfp_flags); 5643f1af42aSIlya Dryomov } else if (num_ops <= CEPH_OSD_SLAB_OPS) { 5653f1af42aSIlya Dryomov req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags); 5663d14c5d2SYehuda Sadeh } else { 5673f1af42aSIlya Dryomov BUG_ON(num_ops > CEPH_OSD_MAX_OPS); 568acafe7e3SKees Cook req = kmalloc(struct_size(req, r_ops, num_ops), gfp_flags); 5693d14c5d2SYehuda Sadeh } 5703f1af42aSIlya Dryomov if (unlikely(!req)) 5713d14c5d2SYehuda Sadeh return NULL; 5723d14c5d2SYehuda Sadeh 5733540bfdbSIlya Dryomov request_init(req); 5743d14c5d2SYehuda Sadeh req->r_osdc = osdc; 5753d14c5d2SYehuda Sadeh req->r_mempool = use_mempool; 57679528734SAlex Elder req->r_num_ops = num_ops; 57784127282SIlya Dryomov req->r_snapid = CEPH_NOSNAP; 57884127282SIlya Dryomov req->r_snapc = ceph_get_snap_context(snapc); 5793d14c5d2SYehuda Sadeh 58013d1ad16SIlya Dryomov dout("%s req %p\n", __func__, req); 58113d1ad16SIlya Dryomov return req; 5823f1af42aSIlya Dryomov } 58313d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_request); 5843f1af42aSIlya Dryomov 5852e59ffd1SIlya Dryomov static int ceph_oloc_encoding_size(const struct ceph_object_locator *oloc) 58630c156d9SYan, Zheng { 58730c156d9SYan, Zheng return 8 + 4 + 4 + 4 + (oloc->pool_ns ? oloc->pool_ns->len : 0); 58830c156d9SYan, Zheng } 58930c156d9SYan, Zheng 59013d1ad16SIlya Dryomov int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp) 59113d1ad16SIlya Dryomov { 59213d1ad16SIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 59313d1ad16SIlya Dryomov struct ceph_msg *msg; 59413d1ad16SIlya Dryomov int msg_size; 5953d14c5d2SYehuda Sadeh 596d30291b9SIlya Dryomov WARN_ON(ceph_oid_empty(&req->r_base_oid)); 59730c156d9SYan, Zheng WARN_ON(ceph_oloc_empty(&req->r_base_oloc)); 598d30291b9SIlya Dryomov 59913d1ad16SIlya Dryomov /* create request message */ 6008cb441c0SIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 6018cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 6028cb441c0SIlya Dryomov msg_size += 4 + 4 + 4; /* hash, osdmap_epoch, flags */ 6038cb441c0SIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 6048cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid); /* reqid */ 6058cb441c0SIlya Dryomov msg_size += sizeof(struct ceph_blkin_trace_info); /* trace */ 6068cb441c0SIlya Dryomov msg_size += 4 + sizeof(struct ceph_timespec); /* client_inc, mtime */ 60730c156d9SYan, Zheng msg_size += CEPH_ENCODING_START_BLK_LEN + 60830c156d9SYan, Zheng ceph_oloc_encoding_size(&req->r_base_oloc); /* oloc */ 60913d1ad16SIlya Dryomov msg_size += 4 + req->r_base_oid.name_len; /* oid */ 61013d1ad16SIlya Dryomov msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op); 611ae458f5aSIlya Dryomov msg_size += 8; /* snapid */ 612ae458f5aSIlya Dryomov msg_size += 8; /* snap_seq */ 61313d1ad16SIlya Dryomov msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0); 6148cb441c0SIlya Dryomov msg_size += 4 + 8; /* retry_attempt, features */ 615ae458f5aSIlya Dryomov 61613d1ad16SIlya Dryomov if (req->r_mempool) 6173d14c5d2SYehuda Sadeh msg = ceph_msgpool_get(&osdc->msgpool_op, 0); 6183d14c5d2SYehuda Sadeh else 61913d1ad16SIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp, true); 62013d1ad16SIlya Dryomov if (!msg) 62113d1ad16SIlya Dryomov return -ENOMEM; 6223d14c5d2SYehuda Sadeh 6233d14c5d2SYehuda Sadeh memset(msg->front.iov_base, 0, msg->front.iov_len); 6243d14c5d2SYehuda Sadeh req->r_request = msg; 6253d14c5d2SYehuda Sadeh 62613d1ad16SIlya Dryomov /* create reply message */ 62713d1ad16SIlya Dryomov msg_size = OSD_OPREPLY_FRONT_LEN; 628711da55dSIlya Dryomov msg_size += req->r_base_oid.name_len; 629711da55dSIlya Dryomov msg_size += req->r_num_ops * sizeof(struct ceph_osd_op); 63013d1ad16SIlya Dryomov 63113d1ad16SIlya Dryomov if (req->r_mempool) 63213d1ad16SIlya Dryomov msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0); 63313d1ad16SIlya Dryomov else 63413d1ad16SIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, msg_size, gfp, true); 63513d1ad16SIlya Dryomov if (!msg) 63613d1ad16SIlya Dryomov return -ENOMEM; 63713d1ad16SIlya Dryomov 63813d1ad16SIlya Dryomov req->r_reply = msg; 63913d1ad16SIlya Dryomov 64013d1ad16SIlya Dryomov return 0; 64113d1ad16SIlya Dryomov } 64213d1ad16SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_alloc_messages); 6433d14c5d2SYehuda Sadeh 644a8dd0a37SAlex Elder static bool osd_req_opcode_valid(u16 opcode) 645a8dd0a37SAlex Elder { 646a8dd0a37SAlex Elder switch (opcode) { 64770b5bfa3SIlya Dryomov #define GENERATE_CASE(op, opcode, str) case CEPH_OSD_OP_##op: return true; 64870b5bfa3SIlya Dryomov __CEPH_FORALL_OSD_OPS(GENERATE_CASE) 64970b5bfa3SIlya Dryomov #undef GENERATE_CASE 650a8dd0a37SAlex Elder default: 651a8dd0a37SAlex Elder return false; 652a8dd0a37SAlex Elder } 653a8dd0a37SAlex Elder } 654a8dd0a37SAlex Elder 65533803f33SAlex Elder /* 65633803f33SAlex Elder * This is an osd op init function for opcodes that have no data or 65733803f33SAlex Elder * other information associated with them. It also serves as a 65833803f33SAlex Elder * common init routine for all the other init functions, below. 65933803f33SAlex Elder */ 660c99d2d4aSAlex Elder static struct ceph_osd_req_op * 66149719778SAlex Elder _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which, 662144cba14SYan, Zheng u16 opcode, u32 flags) 66333803f33SAlex Elder { 664c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 665c99d2d4aSAlex Elder 666c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 66733803f33SAlex Elder BUG_ON(!osd_req_opcode_valid(opcode)); 66833803f33SAlex Elder 669c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 67033803f33SAlex Elder memset(op, 0, sizeof (*op)); 67133803f33SAlex Elder op->op = opcode; 672144cba14SYan, Zheng op->flags = flags; 673c99d2d4aSAlex Elder 674c99d2d4aSAlex Elder return op; 67533803f33SAlex Elder } 67633803f33SAlex Elder 67749719778SAlex Elder void osd_req_op_init(struct ceph_osd_request *osd_req, 678144cba14SYan, Zheng unsigned int which, u16 opcode, u32 flags) 67949719778SAlex Elder { 680144cba14SYan, Zheng (void)_osd_req_op_init(osd_req, which, opcode, flags); 68149719778SAlex Elder } 68249719778SAlex Elder EXPORT_SYMBOL(osd_req_op_init); 68349719778SAlex Elder 684c99d2d4aSAlex Elder void osd_req_op_extent_init(struct ceph_osd_request *osd_req, 685c99d2d4aSAlex Elder unsigned int which, u16 opcode, 68633803f33SAlex Elder u64 offset, u64 length, 68733803f33SAlex Elder u64 truncate_size, u32 truncate_seq) 68833803f33SAlex Elder { 689144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 690144cba14SYan, Zheng opcode, 0); 69133803f33SAlex Elder size_t payload_len = 0; 69233803f33SAlex Elder 693ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 694e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO && 695e30b7577SIlya Dryomov opcode != CEPH_OSD_OP_TRUNCATE); 69633803f33SAlex Elder 69733803f33SAlex Elder op->extent.offset = offset; 69833803f33SAlex Elder op->extent.length = length; 69933803f33SAlex Elder op->extent.truncate_size = truncate_size; 70033803f33SAlex Elder op->extent.truncate_seq = truncate_seq; 701e30b7577SIlya Dryomov if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL) 70233803f33SAlex Elder payload_len += length; 70333803f33SAlex Elder 704de2aa102SIlya Dryomov op->indata_len = payload_len; 70533803f33SAlex Elder } 70633803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_extent_init); 70733803f33SAlex Elder 708c99d2d4aSAlex Elder void osd_req_op_extent_update(struct ceph_osd_request *osd_req, 709c99d2d4aSAlex Elder unsigned int which, u64 length) 710e5975c7cSAlex Elder { 711c99d2d4aSAlex Elder struct ceph_osd_req_op *op; 712c99d2d4aSAlex Elder u64 previous; 713c99d2d4aSAlex Elder 714c99d2d4aSAlex Elder BUG_ON(which >= osd_req->r_num_ops); 715c99d2d4aSAlex Elder op = &osd_req->r_ops[which]; 716c99d2d4aSAlex Elder previous = op->extent.length; 717e5975c7cSAlex Elder 718e5975c7cSAlex Elder if (length == previous) 719e5975c7cSAlex Elder return; /* Nothing to do */ 720e5975c7cSAlex Elder BUG_ON(length > previous); 721e5975c7cSAlex Elder 722e5975c7cSAlex Elder op->extent.length = length; 723d641df81SYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 724de2aa102SIlya Dryomov op->indata_len -= previous - length; 725e5975c7cSAlex Elder } 726e5975c7cSAlex Elder EXPORT_SYMBOL(osd_req_op_extent_update); 727e5975c7cSAlex Elder 7282c63f49aSYan, Zheng void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, 7292c63f49aSYan, Zheng unsigned int which, u64 offset_inc) 7302c63f49aSYan, Zheng { 7312c63f49aSYan, Zheng struct ceph_osd_req_op *op, *prev_op; 7322c63f49aSYan, Zheng 7332c63f49aSYan, Zheng BUG_ON(which + 1 >= osd_req->r_num_ops); 7342c63f49aSYan, Zheng 7352c63f49aSYan, Zheng prev_op = &osd_req->r_ops[which]; 7362c63f49aSYan, Zheng op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags); 7372c63f49aSYan, Zheng /* dup previous one */ 7382c63f49aSYan, Zheng op->indata_len = prev_op->indata_len; 7392c63f49aSYan, Zheng op->outdata_len = prev_op->outdata_len; 7402c63f49aSYan, Zheng op->extent = prev_op->extent; 7412c63f49aSYan, Zheng /* adjust offset */ 7422c63f49aSYan, Zheng op->extent.offset += offset_inc; 7432c63f49aSYan, Zheng op->extent.length -= offset_inc; 7442c63f49aSYan, Zheng 7452c63f49aSYan, Zheng if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL) 7462c63f49aSYan, Zheng op->indata_len -= offset_inc; 7472c63f49aSYan, Zheng } 7482c63f49aSYan, Zheng EXPORT_SYMBOL(osd_req_op_extent_dup_last); 7492c63f49aSYan, Zheng 750c99d2d4aSAlex Elder void osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which, 75104017e29SAlex Elder u16 opcode, const char *class, const char *method) 75233803f33SAlex Elder { 753144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 754144cba14SYan, Zheng opcode, 0); 7555f562df5SAlex Elder struct ceph_pagelist *pagelist; 75633803f33SAlex Elder size_t payload_len = 0; 75733803f33SAlex Elder size_t size; 75833803f33SAlex Elder 75933803f33SAlex Elder BUG_ON(opcode != CEPH_OSD_OP_CALL); 76033803f33SAlex Elder 7615f562df5SAlex Elder pagelist = kmalloc(sizeof (*pagelist), GFP_NOFS); 7625f562df5SAlex Elder BUG_ON(!pagelist); 7635f562df5SAlex Elder ceph_pagelist_init(pagelist); 7645f562df5SAlex Elder 76533803f33SAlex Elder op->cls.class_name = class; 76633803f33SAlex Elder size = strlen(class); 76733803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 76833803f33SAlex Elder op->cls.class_len = size; 7695f562df5SAlex Elder ceph_pagelist_append(pagelist, class, size); 77033803f33SAlex Elder payload_len += size; 77133803f33SAlex Elder 77233803f33SAlex Elder op->cls.method_name = method; 77333803f33SAlex Elder size = strlen(method); 77433803f33SAlex Elder BUG_ON(size > (size_t) U8_MAX); 77533803f33SAlex Elder op->cls.method_len = size; 7765f562df5SAlex Elder ceph_pagelist_append(pagelist, method, size); 77733803f33SAlex Elder payload_len += size; 77833803f33SAlex Elder 779a4ce40a9SAlex Elder osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist); 7805f562df5SAlex Elder 781de2aa102SIlya Dryomov op->indata_len = payload_len; 78233803f33SAlex Elder } 78333803f33SAlex Elder EXPORT_SYMBOL(osd_req_op_cls_init); 7848c042b0dSAlex Elder 785d74b50beSYan, Zheng int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, 786d74b50beSYan, Zheng u16 opcode, const char *name, const void *value, 787d74b50beSYan, Zheng size_t size, u8 cmp_op, u8 cmp_mode) 788d74b50beSYan, Zheng { 789144cba14SYan, Zheng struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 790144cba14SYan, Zheng opcode, 0); 791d74b50beSYan, Zheng struct ceph_pagelist *pagelist; 792d74b50beSYan, Zheng size_t payload_len; 793d74b50beSYan, Zheng 794d74b50beSYan, Zheng BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR); 795d74b50beSYan, Zheng 796d74b50beSYan, Zheng pagelist = kmalloc(sizeof(*pagelist), GFP_NOFS); 797d74b50beSYan, Zheng if (!pagelist) 798d74b50beSYan, Zheng return -ENOMEM; 799d74b50beSYan, Zheng 800d74b50beSYan, Zheng ceph_pagelist_init(pagelist); 801d74b50beSYan, Zheng 802d74b50beSYan, Zheng payload_len = strlen(name); 803d74b50beSYan, Zheng op->xattr.name_len = payload_len; 804d74b50beSYan, Zheng ceph_pagelist_append(pagelist, name, payload_len); 805d74b50beSYan, Zheng 806d74b50beSYan, Zheng op->xattr.value_len = size; 807d74b50beSYan, Zheng ceph_pagelist_append(pagelist, value, size); 808d74b50beSYan, Zheng payload_len += size; 809d74b50beSYan, Zheng 810d74b50beSYan, Zheng op->xattr.cmp_op = cmp_op; 811d74b50beSYan, Zheng op->xattr.cmp_mode = cmp_mode; 812d74b50beSYan, Zheng 813d74b50beSYan, Zheng ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist); 814de2aa102SIlya Dryomov op->indata_len = payload_len; 815d74b50beSYan, Zheng return 0; 816d74b50beSYan, Zheng } 817d74b50beSYan, Zheng EXPORT_SYMBOL(osd_req_op_xattr_init); 818d74b50beSYan, Zheng 819922dab61SIlya Dryomov /* 820922dab61SIlya Dryomov * @watch_opcode: CEPH_OSD_WATCH_OP_* 821922dab61SIlya Dryomov */ 822922dab61SIlya Dryomov static void osd_req_op_watch_init(struct ceph_osd_request *req, int which, 823922dab61SIlya Dryomov u64 cookie, u8 watch_opcode) 82433803f33SAlex Elder { 825922dab61SIlya Dryomov struct ceph_osd_req_op *op; 82633803f33SAlex Elder 827922dab61SIlya Dryomov op = _osd_req_op_init(req, which, CEPH_OSD_OP_WATCH, 0); 82833803f33SAlex Elder op->watch.cookie = cookie; 829922dab61SIlya Dryomov op->watch.op = watch_opcode; 830922dab61SIlya Dryomov op->watch.gen = 0; 83133803f33SAlex Elder } 83233803f33SAlex Elder 833c647b8a8SIlya Dryomov void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, 834c647b8a8SIlya Dryomov unsigned int which, 835c647b8a8SIlya Dryomov u64 expected_object_size, 836c647b8a8SIlya Dryomov u64 expected_write_size) 837c647b8a8SIlya Dryomov { 838c647b8a8SIlya Dryomov struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which, 839144cba14SYan, Zheng CEPH_OSD_OP_SETALLOCHINT, 840144cba14SYan, Zheng 0); 841c647b8a8SIlya Dryomov 842c647b8a8SIlya Dryomov op->alloc_hint.expected_object_size = expected_object_size; 843c647b8a8SIlya Dryomov op->alloc_hint.expected_write_size = expected_write_size; 844c647b8a8SIlya Dryomov 845c647b8a8SIlya Dryomov /* 846c647b8a8SIlya Dryomov * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed 847c647b8a8SIlya Dryomov * not worth a feature bit. Set FAILOK per-op flag to make 848c647b8a8SIlya Dryomov * sure older osds don't trip over an unsupported opcode. 849c647b8a8SIlya Dryomov */ 850c647b8a8SIlya Dryomov op->flags |= CEPH_OSD_OP_FLAG_FAILOK; 851c647b8a8SIlya Dryomov } 852c647b8a8SIlya Dryomov EXPORT_SYMBOL(osd_req_op_alloc_hint_init); 853c647b8a8SIlya Dryomov 85490af3602SAlex Elder static void ceph_osdc_msg_data_add(struct ceph_msg *msg, 855ec9123c5SAlex Elder struct ceph_osd_data *osd_data) 856ec9123c5SAlex Elder { 857ec9123c5SAlex Elder u64 length = ceph_osd_data_length(osd_data); 858ec9123c5SAlex Elder 859ec9123c5SAlex Elder if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) { 860ec9123c5SAlex Elder BUG_ON(length > (u64) SIZE_MAX); 861ec9123c5SAlex Elder if (length) 86290af3602SAlex Elder ceph_msg_data_add_pages(msg, osd_data->pages, 863ec9123c5SAlex Elder length, osd_data->alignment); 864ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) { 865ec9123c5SAlex Elder BUG_ON(!length); 86690af3602SAlex Elder ceph_msg_data_add_pagelist(msg, osd_data->pagelist); 867ec9123c5SAlex Elder #ifdef CONFIG_BLOCK 868ec9123c5SAlex Elder } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) { 8695359a17dSIlya Dryomov ceph_msg_data_add_bio(msg, &osd_data->bio_pos, length); 870ec9123c5SAlex Elder #endif 871b9e281c2SIlya Dryomov } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BVECS) { 872b9e281c2SIlya Dryomov ceph_msg_data_add_bvecs(msg, &osd_data->bvec_pos); 873ec9123c5SAlex Elder } else { 874ec9123c5SAlex Elder BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE); 875ec9123c5SAlex Elder } 876ec9123c5SAlex Elder } 877ec9123c5SAlex Elder 878bb873b53SIlya Dryomov static u32 osd_req_encode_op(struct ceph_osd_op *dst, 879bb873b53SIlya Dryomov const struct ceph_osd_req_op *src) 8803d14c5d2SYehuda Sadeh { 881a8dd0a37SAlex Elder if (WARN_ON(!osd_req_opcode_valid(src->op))) { 882a8dd0a37SAlex Elder pr_err("unrecognized osd opcode %d\n", src->op); 883a8dd0a37SAlex Elder 884a8dd0a37SAlex Elder return 0; 885a8dd0a37SAlex Elder } 8863d14c5d2SYehuda Sadeh 887065a68f9SAlex Elder switch (src->op) { 888fbfab539SAlex Elder case CEPH_OSD_OP_STAT: 889fbfab539SAlex Elder break; 8903d14c5d2SYehuda Sadeh case CEPH_OSD_OP_READ: 8913d14c5d2SYehuda Sadeh case CEPH_OSD_OP_WRITE: 892e30b7577SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 893ad7a60deSLi Wang case CEPH_OSD_OP_ZERO: 894ad7a60deSLi Wang case CEPH_OSD_OP_TRUNCATE: 895175face2SAlex Elder dst->extent.offset = cpu_to_le64(src->extent.offset); 896175face2SAlex Elder dst->extent.length = cpu_to_le64(src->extent.length); 8973d14c5d2SYehuda Sadeh dst->extent.truncate_size = 8983d14c5d2SYehuda Sadeh cpu_to_le64(src->extent.truncate_size); 8993d14c5d2SYehuda Sadeh dst->extent.truncate_seq = 9003d14c5d2SYehuda Sadeh cpu_to_le32(src->extent.truncate_seq); 9013d14c5d2SYehuda Sadeh break; 9023d14c5d2SYehuda Sadeh case CEPH_OSD_OP_CALL: 9033d14c5d2SYehuda Sadeh dst->cls.class_len = src->cls.class_len; 9043d14c5d2SYehuda Sadeh dst->cls.method_len = src->cls.method_len; 905bb873b53SIlya Dryomov dst->cls.indata_len = cpu_to_le32(src->cls.indata_len); 9063d14c5d2SYehuda Sadeh break; 907a40c4f10SYehuda Sadeh case CEPH_OSD_OP_WATCH: 908a40c4f10SYehuda Sadeh dst->watch.cookie = cpu_to_le64(src->watch.cookie); 909922dab61SIlya Dryomov dst->watch.ver = cpu_to_le64(0); 910922dab61SIlya Dryomov dst->watch.op = src->watch.op; 911922dab61SIlya Dryomov dst->watch.gen = cpu_to_le32(src->watch.gen); 912922dab61SIlya Dryomov break; 913922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 914a40c4f10SYehuda Sadeh break; 91519079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 91619079203SIlya Dryomov dst->notify.cookie = cpu_to_le64(src->notify.cookie); 91719079203SIlya Dryomov break; 918a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 919a4ed38d7SDouglas Fuller break; 920c647b8a8SIlya Dryomov case CEPH_OSD_OP_SETALLOCHINT: 921c647b8a8SIlya Dryomov dst->alloc_hint.expected_object_size = 922c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_object_size); 923c647b8a8SIlya Dryomov dst->alloc_hint.expected_write_size = 924c647b8a8SIlya Dryomov cpu_to_le64(src->alloc_hint.expected_write_size); 925c647b8a8SIlya Dryomov break; 926d74b50beSYan, Zheng case CEPH_OSD_OP_SETXATTR: 927d74b50beSYan, Zheng case CEPH_OSD_OP_CMPXATTR: 928d74b50beSYan, Zheng dst->xattr.name_len = cpu_to_le32(src->xattr.name_len); 929d74b50beSYan, Zheng dst->xattr.value_len = cpu_to_le32(src->xattr.value_len); 930d74b50beSYan, Zheng dst->xattr.cmp_op = src->xattr.cmp_op; 931d74b50beSYan, Zheng dst->xattr.cmp_mode = src->xattr.cmp_mode; 932d74b50beSYan, Zheng break; 933864e9197SYan, Zheng case CEPH_OSD_OP_CREATE: 934864e9197SYan, Zheng case CEPH_OSD_OP_DELETE: 935864e9197SYan, Zheng break; 9363d14c5d2SYehuda Sadeh default: 9374c46459cSAlex Elder pr_err("unsupported osd opcode %s\n", 9388f63ca2dSAlex Elder ceph_osd_op_name(src->op)); 9394c46459cSAlex Elder WARN_ON(1); 940a8dd0a37SAlex Elder 941a8dd0a37SAlex Elder return 0; 9423d14c5d2SYehuda Sadeh } 9437b25bf5fSIlya Dryomov 944a8dd0a37SAlex Elder dst->op = cpu_to_le16(src->op); 9457b25bf5fSIlya Dryomov dst->flags = cpu_to_le32(src->flags); 946de2aa102SIlya Dryomov dst->payload_len = cpu_to_le32(src->indata_len); 947175face2SAlex Elder 948bb873b53SIlya Dryomov return src->indata_len; 9493d14c5d2SYehuda Sadeh } 9503d14c5d2SYehuda Sadeh 9513d14c5d2SYehuda Sadeh /* 9523d14c5d2SYehuda Sadeh * build new request AND message, calculate layout, and adjust file 9533d14c5d2SYehuda Sadeh * extent as needed. 9543d14c5d2SYehuda Sadeh * 9553d14c5d2SYehuda Sadeh * if the file was recently truncated, we include information about its 9563d14c5d2SYehuda Sadeh * old and new size so that the object can be updated appropriately. (we 9573d14c5d2SYehuda Sadeh * avoid synchronously deleting truncated objects because it's slow.) 9583d14c5d2SYehuda Sadeh */ 9593d14c5d2SYehuda Sadeh struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc, 9603d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 9613d14c5d2SYehuda Sadeh struct ceph_vino vino, 962715e4cd4SYan, Zheng u64 off, u64 *plen, 963715e4cd4SYan, Zheng unsigned int which, int num_ops, 9643d14c5d2SYehuda Sadeh int opcode, int flags, 9653d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 9663d14c5d2SYehuda Sadeh u32 truncate_seq, 9673d14c5d2SYehuda Sadeh u64 truncate_size, 968153e5167SAlex Elder bool use_mempool) 9693d14c5d2SYehuda Sadeh { 9703d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 97175d1c941SAlex Elder u64 objnum = 0; 97275d1c941SAlex Elder u64 objoff = 0; 97375d1c941SAlex Elder u64 objlen = 0; 9746816282dSSage Weil int r; 9753d14c5d2SYehuda Sadeh 976ad7a60deSLi Wang BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE && 977864e9197SYan, Zheng opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE && 978864e9197SYan, Zheng opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE); 9793d14c5d2SYehuda Sadeh 980acead002SAlex Elder req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool, 981ae7ca4a3SAlex Elder GFP_NOFS); 98213d1ad16SIlya Dryomov if (!req) { 98313d1ad16SIlya Dryomov r = -ENOMEM; 98413d1ad16SIlya Dryomov goto fail; 98513d1ad16SIlya Dryomov } 98679528734SAlex Elder 9873d14c5d2SYehuda Sadeh /* calculate max write size */ 988a19dadfbSAlex Elder r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen); 98913d1ad16SIlya Dryomov if (r) 99013d1ad16SIlya Dryomov goto fail; 991a19dadfbSAlex Elder 992864e9197SYan, Zheng if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) { 993144cba14SYan, Zheng osd_req_op_init(req, which, opcode, 0); 994864e9197SYan, Zheng } else { 9957627151eSYan, Zheng u32 object_size = layout->object_size; 996864e9197SYan, Zheng u32 object_base = off - objoff; 997ccca4e37SYan, Zheng if (!(truncate_seq == 1 && truncate_size == -1ULL)) { 998d18d1e28SAlex Elder if (truncate_size <= object_base) { 999d18d1e28SAlex Elder truncate_size = 0; 1000d18d1e28SAlex Elder } else { 1001d18d1e28SAlex Elder truncate_size -= object_base; 1002d18d1e28SAlex Elder if (truncate_size > object_size) 1003d18d1e28SAlex Elder truncate_size = object_size; 1004d18d1e28SAlex Elder } 1005ccca4e37SYan, Zheng } 1006715e4cd4SYan, Zheng osd_req_op_extent_init(req, which, opcode, objoff, objlen, 1007b0270324SAlex Elder truncate_size, truncate_seq); 1008864e9197SYan, Zheng } 1009d18d1e28SAlex Elder 1010a1f4020aSJeff Layton req->r_abort_on_full = true; 1011bb873b53SIlya Dryomov req->r_flags = flags; 10127627151eSYan, Zheng req->r_base_oloc.pool = layout->pool_id; 101330c156d9SYan, Zheng req->r_base_oloc.pool_ns = ceph_try_get_string(layout->pool_ns); 1014d30291b9SIlya Dryomov ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum); 1015dbe0fc41SAlex Elder 1016bb873b53SIlya Dryomov req->r_snapid = vino.snap; 1017bb873b53SIlya Dryomov if (flags & CEPH_OSD_FLAG_WRITE) 1018bb873b53SIlya Dryomov req->r_data_offset = off; 1019bb873b53SIlya Dryomov 102013d1ad16SIlya Dryomov r = ceph_osdc_alloc_messages(req, GFP_NOFS); 102113d1ad16SIlya Dryomov if (r) 102213d1ad16SIlya Dryomov goto fail; 102313d1ad16SIlya Dryomov 10243d14c5d2SYehuda Sadeh return req; 102513d1ad16SIlya Dryomov 102613d1ad16SIlya Dryomov fail: 102713d1ad16SIlya Dryomov ceph_osdc_put_request(req); 102813d1ad16SIlya Dryomov return ERR_PTR(r); 10293d14c5d2SYehuda Sadeh } 10303d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_new_request); 10313d14c5d2SYehuda Sadeh 10323d14c5d2SYehuda Sadeh /* 10333d14c5d2SYehuda Sadeh * We keep osd requests in an rbtree, sorted by ->r_tid. 10343d14c5d2SYehuda Sadeh */ 1035fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node) 10364609245eSIlya Dryomov DEFINE_RB_FUNCS(request_mc, struct ceph_osd_request, r_tid, r_mc_node) 10373d14c5d2SYehuda Sadeh 10380247a0cfSIlya Dryomov static bool osd_homeless(struct ceph_osd *osd) 10390247a0cfSIlya Dryomov { 10400247a0cfSIlya Dryomov return osd->o_osd == CEPH_HOMELESS_OSD; 10410247a0cfSIlya Dryomov } 10420247a0cfSIlya Dryomov 10435aea3dcdSIlya Dryomov static bool osd_registered(struct ceph_osd *osd) 10443d14c5d2SYehuda Sadeh { 10455aea3dcdSIlya Dryomov verify_osdc_locked(osd->o_osdc); 10463d14c5d2SYehuda Sadeh 10475aea3dcdSIlya Dryomov return !RB_EMPTY_NODE(&osd->o_node); 10483d14c5d2SYehuda Sadeh } 10493d14c5d2SYehuda Sadeh 10503d14c5d2SYehuda Sadeh /* 10510247a0cfSIlya Dryomov * Assumes @osd is zero-initialized. 10520247a0cfSIlya Dryomov */ 10530247a0cfSIlya Dryomov static void osd_init(struct ceph_osd *osd) 10540247a0cfSIlya Dryomov { 105502113a0fSElena Reshetova refcount_set(&osd->o_ref, 1); 10560247a0cfSIlya Dryomov RB_CLEAR_NODE(&osd->o_node); 10575aea3dcdSIlya Dryomov osd->o_requests = RB_ROOT; 1058922dab61SIlya Dryomov osd->o_linger_requests = RB_ROOT; 1059a02a946dSIlya Dryomov osd->o_backoff_mappings = RB_ROOT; 1060a02a946dSIlya Dryomov osd->o_backoffs_by_id = RB_ROOT; 10610247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_osd_lru); 10620247a0cfSIlya Dryomov INIT_LIST_HEAD(&osd->o_keepalive_item); 10630247a0cfSIlya Dryomov osd->o_incarnation = 1; 10645aea3dcdSIlya Dryomov mutex_init(&osd->lock); 10650247a0cfSIlya Dryomov } 10660247a0cfSIlya Dryomov 10670247a0cfSIlya Dryomov static void osd_cleanup(struct ceph_osd *osd) 10680247a0cfSIlya Dryomov { 10690247a0cfSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&osd->o_node)); 10705aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 1071922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 1072a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoff_mappings)); 1073a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoffs_by_id)); 10740247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_osd_lru)); 10750247a0cfSIlya Dryomov WARN_ON(!list_empty(&osd->o_keepalive_item)); 10760247a0cfSIlya Dryomov 10770247a0cfSIlya Dryomov if (osd->o_auth.authorizer) { 10780247a0cfSIlya Dryomov WARN_ON(osd_homeless(osd)); 10790247a0cfSIlya Dryomov ceph_auth_destroy_authorizer(osd->o_auth.authorizer); 10800247a0cfSIlya Dryomov } 10810247a0cfSIlya Dryomov } 10820247a0cfSIlya Dryomov 10830247a0cfSIlya Dryomov /* 10843d14c5d2SYehuda Sadeh * Track open sessions with osds. 10853d14c5d2SYehuda Sadeh */ 1086e10006f8SAlex Elder static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum) 10873d14c5d2SYehuda Sadeh { 10883d14c5d2SYehuda Sadeh struct ceph_osd *osd; 10893d14c5d2SYehuda Sadeh 10900247a0cfSIlya Dryomov WARN_ON(onum == CEPH_HOMELESS_OSD); 10910247a0cfSIlya Dryomov 10927a28f59bSIlya Dryomov osd = kzalloc(sizeof(*osd), GFP_NOIO | __GFP_NOFAIL); 10930247a0cfSIlya Dryomov osd_init(osd); 10943d14c5d2SYehuda Sadeh osd->o_osdc = osdc; 1095e10006f8SAlex Elder osd->o_osd = onum; 10963d14c5d2SYehuda Sadeh 1097b7a9e5ddSSage Weil ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr); 10983d14c5d2SYehuda Sadeh 10993d14c5d2SYehuda Sadeh return osd; 11003d14c5d2SYehuda Sadeh } 11013d14c5d2SYehuda Sadeh 11023d14c5d2SYehuda Sadeh static struct ceph_osd *get_osd(struct ceph_osd *osd) 11033d14c5d2SYehuda Sadeh { 110402113a0fSElena Reshetova if (refcount_inc_not_zero(&osd->o_ref)) { 110502113a0fSElena Reshetova dout("get_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref)-1, 110602113a0fSElena Reshetova refcount_read(&osd->o_ref)); 11073d14c5d2SYehuda Sadeh return osd; 11083d14c5d2SYehuda Sadeh } else { 11093d14c5d2SYehuda Sadeh dout("get_osd %p FAIL\n", osd); 11103d14c5d2SYehuda Sadeh return NULL; 11113d14c5d2SYehuda Sadeh } 11123d14c5d2SYehuda Sadeh } 11133d14c5d2SYehuda Sadeh 11143d14c5d2SYehuda Sadeh static void put_osd(struct ceph_osd *osd) 11153d14c5d2SYehuda Sadeh { 111602113a0fSElena Reshetova dout("put_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref), 111702113a0fSElena Reshetova refcount_read(&osd->o_ref) - 1); 111802113a0fSElena Reshetova if (refcount_dec_and_test(&osd->o_ref)) { 11190247a0cfSIlya Dryomov osd_cleanup(osd); 11203d14c5d2SYehuda Sadeh kfree(osd); 11213d14c5d2SYehuda Sadeh } 11223d14c5d2SYehuda Sadeh } 11233d14c5d2SYehuda Sadeh 1124fcd00b68SIlya Dryomov DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node) 1125fcd00b68SIlya Dryomov 11269dd2845cSIlya Dryomov static void __move_osd_to_lru(struct ceph_osd *osd) 11273d14c5d2SYehuda Sadeh { 11289dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 11299dd2845cSIlya Dryomov 11309dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 11313d14c5d2SYehuda Sadeh BUG_ON(!list_empty(&osd->o_osd_lru)); 1132bbf37ec3SIlya Dryomov 11339dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 11343d14c5d2SYehuda Sadeh list_add_tail(&osd->o_osd_lru, &osdc->osd_lru); 11359dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 11369dd2845cSIlya Dryomov 1137a319bf56SIlya Dryomov osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl; 11383d14c5d2SYehuda Sadeh } 11393d14c5d2SYehuda Sadeh 11409dd2845cSIlya Dryomov static void maybe_move_osd_to_lru(struct ceph_osd *osd) 1141bbf37ec3SIlya Dryomov { 11425aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1143922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) 11449dd2845cSIlya Dryomov __move_osd_to_lru(osd); 1145bbf37ec3SIlya Dryomov } 1146bbf37ec3SIlya Dryomov 11473d14c5d2SYehuda Sadeh static void __remove_osd_from_lru(struct ceph_osd *osd) 11483d14c5d2SYehuda Sadeh { 11499dd2845cSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 11509dd2845cSIlya Dryomov 11519dd2845cSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 11529dd2845cSIlya Dryomov 11539dd2845cSIlya Dryomov spin_lock(&osdc->osd_lru_lock); 11543d14c5d2SYehuda Sadeh if (!list_empty(&osd->o_osd_lru)) 11553d14c5d2SYehuda Sadeh list_del_init(&osd->o_osd_lru); 11569dd2845cSIlya Dryomov spin_unlock(&osdc->osd_lru_lock); 11573d14c5d2SYehuda Sadeh } 11583d14c5d2SYehuda Sadeh 11593d14c5d2SYehuda Sadeh /* 11605aea3dcdSIlya Dryomov * Close the connection and assign any leftover requests to the 11615aea3dcdSIlya Dryomov * homeless session. 11625aea3dcdSIlya Dryomov */ 11635aea3dcdSIlya Dryomov static void close_osd(struct ceph_osd *osd) 11645aea3dcdSIlya Dryomov { 11655aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 11665aea3dcdSIlya Dryomov struct rb_node *n; 11675aea3dcdSIlya Dryomov 11685aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 11695aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 11705aea3dcdSIlya Dryomov 11715aea3dcdSIlya Dryomov ceph_con_close(&osd->o_con); 11725aea3dcdSIlya Dryomov 11735aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 11745aea3dcdSIlya Dryomov struct ceph_osd_request *req = 11755aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 11765aea3dcdSIlya Dryomov 11775aea3dcdSIlya Dryomov n = rb_next(n); /* unlink_request() */ 11785aea3dcdSIlya Dryomov 11795aea3dcdSIlya Dryomov dout(" reassigning req %p tid %llu\n", req, req->r_tid); 11805aea3dcdSIlya Dryomov unlink_request(osd, req); 11815aea3dcdSIlya Dryomov link_request(&osdc->homeless_osd, req); 11825aea3dcdSIlya Dryomov } 1183922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 1184922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 1185922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 1186922dab61SIlya Dryomov 1187922dab61SIlya Dryomov n = rb_next(n); /* unlink_linger() */ 1188922dab61SIlya Dryomov 1189922dab61SIlya Dryomov dout(" reassigning lreq %p linger_id %llu\n", lreq, 1190922dab61SIlya Dryomov lreq->linger_id); 1191922dab61SIlya Dryomov unlink_linger(osd, lreq); 1192922dab61SIlya Dryomov link_linger(&osdc->homeless_osd, lreq); 1193922dab61SIlya Dryomov } 1194a02a946dSIlya Dryomov clear_backoffs(osd); 11955aea3dcdSIlya Dryomov 11965aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 11975aea3dcdSIlya Dryomov erase_osd(&osdc->osds, osd); 11985aea3dcdSIlya Dryomov put_osd(osd); 11995aea3dcdSIlya Dryomov } 12005aea3dcdSIlya Dryomov 12015aea3dcdSIlya Dryomov /* 12023d14c5d2SYehuda Sadeh * reset osd connect 12033d14c5d2SYehuda Sadeh */ 12045aea3dcdSIlya Dryomov static int reopen_osd(struct ceph_osd *osd) 12053d14c5d2SYehuda Sadeh { 1206c3acb181SAlex Elder struct ceph_entity_addr *peer_addr; 12073d14c5d2SYehuda Sadeh 12085aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 12095aea3dcdSIlya Dryomov 12105aea3dcdSIlya Dryomov if (RB_EMPTY_ROOT(&osd->o_requests) && 1211922dab61SIlya Dryomov RB_EMPTY_ROOT(&osd->o_linger_requests)) { 12125aea3dcdSIlya Dryomov close_osd(osd); 1213c3acb181SAlex Elder return -ENODEV; 1214c3acb181SAlex Elder } 1215c3acb181SAlex Elder 12165aea3dcdSIlya Dryomov peer_addr = &osd->o_osdc->osdmap->osd_addr[osd->o_osd]; 1217c3acb181SAlex Elder if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) && 12183d14c5d2SYehuda Sadeh !ceph_con_opened(&osd->o_con)) { 12195aea3dcdSIlya Dryomov struct rb_node *n; 1220c3acb181SAlex Elder 12213d14c5d2SYehuda Sadeh dout("osd addr hasn't changed and connection never opened, " 12220b4af2e8SIlya Dryomov "letting msgr retry\n"); 12233d14c5d2SYehuda Sadeh /* touch each r_stamp for handle_timeout()'s benfit */ 12245aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 12255aea3dcdSIlya Dryomov struct ceph_osd_request *req = 12265aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 12273d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 12285aea3dcdSIlya Dryomov } 1229c3acb181SAlex Elder 1230c3acb181SAlex Elder return -EAGAIN; 12313d14c5d2SYehuda Sadeh } 1232c3acb181SAlex Elder 1233c3acb181SAlex Elder ceph_con_close(&osd->o_con); 1234c3acb181SAlex Elder ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr); 1235c3acb181SAlex Elder osd->o_incarnation++; 1236c3acb181SAlex Elder 1237c3acb181SAlex Elder return 0; 12383d14c5d2SYehuda Sadeh } 12393d14c5d2SYehuda Sadeh 12405aea3dcdSIlya Dryomov static struct ceph_osd *lookup_create_osd(struct ceph_osd_client *osdc, int o, 12415aea3dcdSIlya Dryomov bool wrlocked) 12423d14c5d2SYehuda Sadeh { 12435aea3dcdSIlya Dryomov struct ceph_osd *osd; 12445aea3dcdSIlya Dryomov 12455aea3dcdSIlya Dryomov if (wrlocked) 12465aea3dcdSIlya Dryomov verify_osdc_wrlocked(osdc); 12475aea3dcdSIlya Dryomov else 12485aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 12495aea3dcdSIlya Dryomov 12505aea3dcdSIlya Dryomov if (o != CEPH_HOMELESS_OSD) 12515aea3dcdSIlya Dryomov osd = lookup_osd(&osdc->osds, o); 12525aea3dcdSIlya Dryomov else 12535aea3dcdSIlya Dryomov osd = &osdc->homeless_osd; 12545aea3dcdSIlya Dryomov if (!osd) { 12555aea3dcdSIlya Dryomov if (!wrlocked) 12565aea3dcdSIlya Dryomov return ERR_PTR(-EAGAIN); 12575aea3dcdSIlya Dryomov 12585aea3dcdSIlya Dryomov osd = create_osd(osdc, o); 12595aea3dcdSIlya Dryomov insert_osd(&osdc->osds, osd); 12605aea3dcdSIlya Dryomov ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, 12615aea3dcdSIlya Dryomov &osdc->osdmap->osd_addr[osd->o_osd]); 12625aea3dcdSIlya Dryomov } 12635aea3dcdSIlya Dryomov 12645aea3dcdSIlya Dryomov dout("%s osdc %p osd%d -> osd %p\n", __func__, osdc, o, osd); 12655aea3dcdSIlya Dryomov return osd; 1266a40c4f10SYehuda Sadeh } 1267a40c4f10SYehuda Sadeh 12683d14c5d2SYehuda Sadeh /* 12695aea3dcdSIlya Dryomov * Create request <-> OSD session relation. 12705aea3dcdSIlya Dryomov * 12715aea3dcdSIlya Dryomov * @req has to be assigned a tid, @osd may be homeless. 12723d14c5d2SYehuda Sadeh */ 12735aea3dcdSIlya Dryomov static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req) 12743d14c5d2SYehuda Sadeh { 12755aea3dcdSIlya Dryomov verify_osd_locked(osd); 12765aea3dcdSIlya Dryomov WARN_ON(!req->r_tid || req->r_osd); 12775aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 127835f9f8a0SSage Weil req, req->r_tid); 12795aea3dcdSIlya Dryomov 12805aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 12815aea3dcdSIlya Dryomov __remove_osd_from_lru(osd); 12825aea3dcdSIlya Dryomov else 12835aea3dcdSIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 12845aea3dcdSIlya Dryomov 12855aea3dcdSIlya Dryomov get_osd(osd); 12865aea3dcdSIlya Dryomov insert_request(&osd->o_requests, req); 12875aea3dcdSIlya Dryomov req->r_osd = osd; 128835f9f8a0SSage Weil } 128935f9f8a0SSage Weil 12905aea3dcdSIlya Dryomov static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req) 12913d14c5d2SYehuda Sadeh { 12925aea3dcdSIlya Dryomov verify_osd_locked(osd); 12935aea3dcdSIlya Dryomov WARN_ON(req->r_osd != osd); 12945aea3dcdSIlya Dryomov dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd, 12955aea3dcdSIlya Dryomov req, req->r_tid); 12965aea3dcdSIlya Dryomov 12975aea3dcdSIlya Dryomov req->r_osd = NULL; 12985aea3dcdSIlya Dryomov erase_request(&osd->o_requests, req); 12995aea3dcdSIlya Dryomov put_osd(osd); 13005aea3dcdSIlya Dryomov 13015aea3dcdSIlya Dryomov if (!osd_homeless(osd)) 13025aea3dcdSIlya Dryomov maybe_move_osd_to_lru(osd); 13035aea3dcdSIlya Dryomov else 13045aea3dcdSIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 13053d14c5d2SYehuda Sadeh } 13063d14c5d2SYehuda Sadeh 130763244fa1SIlya Dryomov static bool __pool_full(struct ceph_pg_pool_info *pi) 130863244fa1SIlya Dryomov { 130963244fa1SIlya Dryomov return pi->flags & CEPH_POOL_FLAG_FULL; 131063244fa1SIlya Dryomov } 131163244fa1SIlya Dryomov 131242c1b124SIlya Dryomov static bool have_pool_full(struct ceph_osd_client *osdc) 131342c1b124SIlya Dryomov { 131442c1b124SIlya Dryomov struct rb_node *n; 131542c1b124SIlya Dryomov 131642c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 131742c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 131842c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 131942c1b124SIlya Dryomov 132042c1b124SIlya Dryomov if (__pool_full(pi)) 132142c1b124SIlya Dryomov return true; 132242c1b124SIlya Dryomov } 132342c1b124SIlya Dryomov 132442c1b124SIlya Dryomov return false; 132542c1b124SIlya Dryomov } 132642c1b124SIlya Dryomov 13275aea3dcdSIlya Dryomov static bool pool_full(struct ceph_osd_client *osdc, s64 pool_id) 13285aea3dcdSIlya Dryomov { 13295aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 13305aea3dcdSIlya Dryomov 13315aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 13325aea3dcdSIlya Dryomov if (!pi) 13335aea3dcdSIlya Dryomov return false; 13345aea3dcdSIlya Dryomov 13355aea3dcdSIlya Dryomov return __pool_full(pi); 13365aea3dcdSIlya Dryomov } 13375aea3dcdSIlya Dryomov 13383d14c5d2SYehuda Sadeh /* 1339d29adb34SJosh Durgin * Returns whether a request should be blocked from being sent 1340d29adb34SJosh Durgin * based on the current osdmap and osd_client settings. 1341d29adb34SJosh Durgin */ 134263244fa1SIlya Dryomov static bool target_should_be_paused(struct ceph_osd_client *osdc, 134363244fa1SIlya Dryomov const struct ceph_osd_request_target *t, 134463244fa1SIlya Dryomov struct ceph_pg_pool_info *pi) 134563244fa1SIlya Dryomov { 1346b7ec35b3SIlya Dryomov bool pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 1347b7ec35b3SIlya Dryomov bool pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 1348b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 134963244fa1SIlya Dryomov __pool_full(pi); 135063244fa1SIlya Dryomov 13516d637a54SIlya Dryomov WARN_ON(pi->id != t->target_oloc.pool); 135258eb7932SJeff Layton return ((t->flags & CEPH_OSD_FLAG_READ) && pauserd) || 135358eb7932SJeff Layton ((t->flags & CEPH_OSD_FLAG_WRITE) && pausewr) || 135458eb7932SJeff Layton (osdc->osdmap->epoch < osdc->epoch_barrier); 135563244fa1SIlya Dryomov } 135663244fa1SIlya Dryomov 135763244fa1SIlya Dryomov enum calc_target_result { 135863244fa1SIlya Dryomov CALC_TARGET_NO_ACTION = 0, 135963244fa1SIlya Dryomov CALC_TARGET_NEED_RESEND, 136063244fa1SIlya Dryomov CALC_TARGET_POOL_DNE, 136163244fa1SIlya Dryomov }; 136263244fa1SIlya Dryomov 136363244fa1SIlya Dryomov static enum calc_target_result calc_target(struct ceph_osd_client *osdc, 136463244fa1SIlya Dryomov struct ceph_osd_request_target *t, 13657de030d6SIlya Dryomov struct ceph_connection *con, 136663244fa1SIlya Dryomov bool any_change) 136763244fa1SIlya Dryomov { 136863244fa1SIlya Dryomov struct ceph_pg_pool_info *pi; 136963244fa1SIlya Dryomov struct ceph_pg pgid, last_pgid; 137063244fa1SIlya Dryomov struct ceph_osds up, acting; 137163244fa1SIlya Dryomov bool force_resend = false; 137284ed45dfSIlya Dryomov bool unpaused = false; 137384ed45dfSIlya Dryomov bool legacy_change; 13747de030d6SIlya Dryomov bool split = false; 1375b7ec35b3SIlya Dryomov bool sort_bitwise = ceph_osdmap_flag(osdc, CEPH_OSDMAP_SORTBITWISE); 1376ae78dd81SIlya Dryomov bool recovery_deletes = ceph_osdmap_flag(osdc, 1377ae78dd81SIlya Dryomov CEPH_OSDMAP_RECOVERY_DELETES); 137863244fa1SIlya Dryomov enum calc_target_result ct_res; 137963244fa1SIlya Dryomov int ret; 138063244fa1SIlya Dryomov 138104c7d789SIlya Dryomov t->epoch = osdc->osdmap->epoch; 138263244fa1SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool); 138363244fa1SIlya Dryomov if (!pi) { 138463244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 138563244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 138663244fa1SIlya Dryomov goto out; 138763244fa1SIlya Dryomov } 138863244fa1SIlya Dryomov 138963244fa1SIlya Dryomov if (osdc->osdmap->epoch == pi->last_force_request_resend) { 1390dc93e0e2SIlya Dryomov if (t->last_force_resend < pi->last_force_request_resend) { 1391dc93e0e2SIlya Dryomov t->last_force_resend = pi->last_force_request_resend; 139263244fa1SIlya Dryomov force_resend = true; 1393dc93e0e2SIlya Dryomov } else if (t->last_force_resend == 0) { 139463244fa1SIlya Dryomov force_resend = true; 139563244fa1SIlya Dryomov } 139663244fa1SIlya Dryomov } 139763244fa1SIlya Dryomov 1398db098ec4SIlya Dryomov /* apply tiering */ 1399db098ec4SIlya Dryomov ceph_oid_copy(&t->target_oid, &t->base_oid); 1400db098ec4SIlya Dryomov ceph_oloc_copy(&t->target_oloc, &t->base_oloc); 1401db098ec4SIlya Dryomov if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) { 140263244fa1SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_READ && pi->read_tier >= 0) 140363244fa1SIlya Dryomov t->target_oloc.pool = pi->read_tier; 140463244fa1SIlya Dryomov if (t->flags & CEPH_OSD_FLAG_WRITE && pi->write_tier >= 0) 140563244fa1SIlya Dryomov t->target_oloc.pool = pi->write_tier; 14066d637a54SIlya Dryomov 14076d637a54SIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool); 14086d637a54SIlya Dryomov if (!pi) { 14096d637a54SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 14106d637a54SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 14116d637a54SIlya Dryomov goto out; 14126d637a54SIlya Dryomov } 141363244fa1SIlya Dryomov } 141463244fa1SIlya Dryomov 1415df28152dSIlya Dryomov ret = __ceph_object_locator_to_pg(pi, &t->target_oid, &t->target_oloc, 1416df28152dSIlya Dryomov &pgid); 141763244fa1SIlya Dryomov if (ret) { 141863244fa1SIlya Dryomov WARN_ON(ret != -ENOENT); 141963244fa1SIlya Dryomov t->osd = CEPH_HOMELESS_OSD; 142063244fa1SIlya Dryomov ct_res = CALC_TARGET_POOL_DNE; 142163244fa1SIlya Dryomov goto out; 142263244fa1SIlya Dryomov } 142363244fa1SIlya Dryomov last_pgid.pool = pgid.pool; 142463244fa1SIlya Dryomov last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask); 142563244fa1SIlya Dryomov 1426df28152dSIlya Dryomov ceph_pg_to_up_acting_osds(osdc->osdmap, pi, &pgid, &up, &acting); 142763244fa1SIlya Dryomov if (any_change && 142863244fa1SIlya Dryomov ceph_is_new_interval(&t->acting, 142963244fa1SIlya Dryomov &acting, 143063244fa1SIlya Dryomov &t->up, 143163244fa1SIlya Dryomov &up, 143263244fa1SIlya Dryomov t->size, 143363244fa1SIlya Dryomov pi->size, 143463244fa1SIlya Dryomov t->min_size, 143563244fa1SIlya Dryomov pi->min_size, 143663244fa1SIlya Dryomov t->pg_num, 143763244fa1SIlya Dryomov pi->pg_num, 143863244fa1SIlya Dryomov t->sort_bitwise, 143963244fa1SIlya Dryomov sort_bitwise, 1440ae78dd81SIlya Dryomov t->recovery_deletes, 1441ae78dd81SIlya Dryomov recovery_deletes, 144263244fa1SIlya Dryomov &last_pgid)) 144363244fa1SIlya Dryomov force_resend = true; 144463244fa1SIlya Dryomov 144563244fa1SIlya Dryomov if (t->paused && !target_should_be_paused(osdc, t, pi)) { 144663244fa1SIlya Dryomov t->paused = false; 144784ed45dfSIlya Dryomov unpaused = true; 144863244fa1SIlya Dryomov } 144984ed45dfSIlya Dryomov legacy_change = ceph_pg_compare(&t->pgid, &pgid) || 145084ed45dfSIlya Dryomov ceph_osds_changed(&t->acting, &acting, any_change); 14517de030d6SIlya Dryomov if (t->pg_num) 14527de030d6SIlya Dryomov split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num); 145363244fa1SIlya Dryomov 14547de030d6SIlya Dryomov if (legacy_change || force_resend || split) { 145563244fa1SIlya Dryomov t->pgid = pgid; /* struct */ 1456df28152dSIlya Dryomov ceph_pg_to_primary_shard(osdc->osdmap, pi, &pgid, &t->spgid); 145763244fa1SIlya Dryomov ceph_osds_copy(&t->acting, &acting); 145863244fa1SIlya Dryomov ceph_osds_copy(&t->up, &up); 145963244fa1SIlya Dryomov t->size = pi->size; 146063244fa1SIlya Dryomov t->min_size = pi->min_size; 146163244fa1SIlya Dryomov t->pg_num = pi->pg_num; 146263244fa1SIlya Dryomov t->pg_num_mask = pi->pg_num_mask; 146363244fa1SIlya Dryomov t->sort_bitwise = sort_bitwise; 1464ae78dd81SIlya Dryomov t->recovery_deletes = recovery_deletes; 146563244fa1SIlya Dryomov 146663244fa1SIlya Dryomov t->osd = acting.primary; 146763244fa1SIlya Dryomov } 146863244fa1SIlya Dryomov 14697de030d6SIlya Dryomov if (unpaused || legacy_change || force_resend || 14707de030d6SIlya Dryomov (split && con && CEPH_HAVE_FEATURE(con->peer_features, 14717de030d6SIlya Dryomov RESEND_ON_SPLIT))) 147284ed45dfSIlya Dryomov ct_res = CALC_TARGET_NEED_RESEND; 147384ed45dfSIlya Dryomov else 147484ed45dfSIlya Dryomov ct_res = CALC_TARGET_NO_ACTION; 147584ed45dfSIlya Dryomov 147663244fa1SIlya Dryomov out: 147763244fa1SIlya Dryomov dout("%s t %p -> ct_res %d osd %d\n", __func__, t, ct_res, t->osd); 147863244fa1SIlya Dryomov return ct_res; 147963244fa1SIlya Dryomov } 148063244fa1SIlya Dryomov 1481a02a946dSIlya Dryomov static struct ceph_spg_mapping *alloc_spg_mapping(void) 1482a02a946dSIlya Dryomov { 1483a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1484a02a946dSIlya Dryomov 1485a02a946dSIlya Dryomov spg = kmalloc(sizeof(*spg), GFP_NOIO); 1486a02a946dSIlya Dryomov if (!spg) 1487a02a946dSIlya Dryomov return NULL; 1488a02a946dSIlya Dryomov 1489a02a946dSIlya Dryomov RB_CLEAR_NODE(&spg->node); 1490a02a946dSIlya Dryomov spg->backoffs = RB_ROOT; 1491a02a946dSIlya Dryomov return spg; 1492a02a946dSIlya Dryomov } 1493a02a946dSIlya Dryomov 1494a02a946dSIlya Dryomov static void free_spg_mapping(struct ceph_spg_mapping *spg) 1495a02a946dSIlya Dryomov { 1496a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&spg->node)); 1497a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&spg->backoffs)); 1498a02a946dSIlya Dryomov 1499a02a946dSIlya Dryomov kfree(spg); 1500a02a946dSIlya Dryomov } 1501a02a946dSIlya Dryomov 1502a02a946dSIlya Dryomov /* 1503a02a946dSIlya Dryomov * rbtree of ceph_spg_mapping for handling map<spg_t, ...>, similar to 1504a02a946dSIlya Dryomov * ceph_pg_mapping. Used to track OSD backoffs -- a backoff [range] is 1505a02a946dSIlya Dryomov * defined only within a specific spgid; it does not pass anything to 1506a02a946dSIlya Dryomov * children on split, or to another primary. 1507a02a946dSIlya Dryomov */ 1508a02a946dSIlya Dryomov DEFINE_RB_FUNCS2(spg_mapping, struct ceph_spg_mapping, spgid, ceph_spg_compare, 1509a02a946dSIlya Dryomov RB_BYPTR, const struct ceph_spg *, node) 1510a02a946dSIlya Dryomov 1511a02a946dSIlya Dryomov static u64 hoid_get_bitwise_key(const struct ceph_hobject_id *hoid) 1512a02a946dSIlya Dryomov { 1513a02a946dSIlya Dryomov return hoid->is_max ? 0x100000000ull : hoid->hash_reverse_bits; 1514a02a946dSIlya Dryomov } 1515a02a946dSIlya Dryomov 1516a02a946dSIlya Dryomov static void hoid_get_effective_key(const struct ceph_hobject_id *hoid, 1517a02a946dSIlya Dryomov void **pkey, size_t *pkey_len) 1518a02a946dSIlya Dryomov { 1519a02a946dSIlya Dryomov if (hoid->key_len) { 1520a02a946dSIlya Dryomov *pkey = hoid->key; 1521a02a946dSIlya Dryomov *pkey_len = hoid->key_len; 1522a02a946dSIlya Dryomov } else { 1523a02a946dSIlya Dryomov *pkey = hoid->oid; 1524a02a946dSIlya Dryomov *pkey_len = hoid->oid_len; 1525a02a946dSIlya Dryomov } 1526a02a946dSIlya Dryomov } 1527a02a946dSIlya Dryomov 1528a02a946dSIlya Dryomov static int compare_names(const void *name1, size_t name1_len, 1529a02a946dSIlya Dryomov const void *name2, size_t name2_len) 1530a02a946dSIlya Dryomov { 1531a02a946dSIlya Dryomov int ret; 1532a02a946dSIlya Dryomov 1533a02a946dSIlya Dryomov ret = memcmp(name1, name2, min(name1_len, name2_len)); 1534a02a946dSIlya Dryomov if (!ret) { 1535a02a946dSIlya Dryomov if (name1_len < name2_len) 1536a02a946dSIlya Dryomov ret = -1; 1537a02a946dSIlya Dryomov else if (name1_len > name2_len) 1538a02a946dSIlya Dryomov ret = 1; 1539a02a946dSIlya Dryomov } 1540a02a946dSIlya Dryomov return ret; 1541a02a946dSIlya Dryomov } 1542a02a946dSIlya Dryomov 1543a02a946dSIlya Dryomov static int hoid_compare(const struct ceph_hobject_id *lhs, 1544a02a946dSIlya Dryomov const struct ceph_hobject_id *rhs) 1545a02a946dSIlya Dryomov { 1546a02a946dSIlya Dryomov void *effective_key1, *effective_key2; 1547a02a946dSIlya Dryomov size_t effective_key1_len, effective_key2_len; 1548a02a946dSIlya Dryomov int ret; 1549a02a946dSIlya Dryomov 1550a02a946dSIlya Dryomov if (lhs->is_max < rhs->is_max) 1551a02a946dSIlya Dryomov return -1; 1552a02a946dSIlya Dryomov if (lhs->is_max > rhs->is_max) 1553a02a946dSIlya Dryomov return 1; 1554a02a946dSIlya Dryomov 1555a02a946dSIlya Dryomov if (lhs->pool < rhs->pool) 1556a02a946dSIlya Dryomov return -1; 1557a02a946dSIlya Dryomov if (lhs->pool > rhs->pool) 1558a02a946dSIlya Dryomov return 1; 1559a02a946dSIlya Dryomov 1560a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) < hoid_get_bitwise_key(rhs)) 1561a02a946dSIlya Dryomov return -1; 1562a02a946dSIlya Dryomov if (hoid_get_bitwise_key(lhs) > hoid_get_bitwise_key(rhs)) 1563a02a946dSIlya Dryomov return 1; 1564a02a946dSIlya Dryomov 1565a02a946dSIlya Dryomov ret = compare_names(lhs->nspace, lhs->nspace_len, 1566a02a946dSIlya Dryomov rhs->nspace, rhs->nspace_len); 1567a02a946dSIlya Dryomov if (ret) 1568a02a946dSIlya Dryomov return ret; 1569a02a946dSIlya Dryomov 1570a02a946dSIlya Dryomov hoid_get_effective_key(lhs, &effective_key1, &effective_key1_len); 1571a02a946dSIlya Dryomov hoid_get_effective_key(rhs, &effective_key2, &effective_key2_len); 1572a02a946dSIlya Dryomov ret = compare_names(effective_key1, effective_key1_len, 1573a02a946dSIlya Dryomov effective_key2, effective_key2_len); 1574a02a946dSIlya Dryomov if (ret) 1575a02a946dSIlya Dryomov return ret; 1576a02a946dSIlya Dryomov 1577a02a946dSIlya Dryomov ret = compare_names(lhs->oid, lhs->oid_len, rhs->oid, rhs->oid_len); 1578a02a946dSIlya Dryomov if (ret) 1579a02a946dSIlya Dryomov return ret; 1580a02a946dSIlya Dryomov 1581a02a946dSIlya Dryomov if (lhs->snapid < rhs->snapid) 1582a02a946dSIlya Dryomov return -1; 1583a02a946dSIlya Dryomov if (lhs->snapid > rhs->snapid) 1584a02a946dSIlya Dryomov return 1; 1585a02a946dSIlya Dryomov 1586a02a946dSIlya Dryomov return 0; 1587a02a946dSIlya Dryomov } 1588a02a946dSIlya Dryomov 1589a02a946dSIlya Dryomov /* 1590a02a946dSIlya Dryomov * For decoding ->begin and ->end of MOSDBackoff only -- no MIN/MAX 1591a02a946dSIlya Dryomov * compat stuff here. 1592a02a946dSIlya Dryomov * 1593a02a946dSIlya Dryomov * Assumes @hoid is zero-initialized. 1594a02a946dSIlya Dryomov */ 1595a02a946dSIlya Dryomov static int decode_hoid(void **p, void *end, struct ceph_hobject_id *hoid) 1596a02a946dSIlya Dryomov { 1597a02a946dSIlya Dryomov u8 struct_v; 1598a02a946dSIlya Dryomov u32 struct_len; 1599a02a946dSIlya Dryomov int ret; 1600a02a946dSIlya Dryomov 1601a02a946dSIlya Dryomov ret = ceph_start_decoding(p, end, 4, "hobject_t", &struct_v, 1602a02a946dSIlya Dryomov &struct_len); 1603a02a946dSIlya Dryomov if (ret) 1604a02a946dSIlya Dryomov return ret; 1605a02a946dSIlya Dryomov 1606a02a946dSIlya Dryomov if (struct_v < 4) { 1607a02a946dSIlya Dryomov pr_err("got struct_v %d < 4 of hobject_t\n", struct_v); 1608a02a946dSIlya Dryomov goto e_inval; 1609a02a946dSIlya Dryomov } 1610a02a946dSIlya Dryomov 1611a02a946dSIlya Dryomov hoid->key = ceph_extract_encoded_string(p, end, &hoid->key_len, 1612a02a946dSIlya Dryomov GFP_NOIO); 1613a02a946dSIlya Dryomov if (IS_ERR(hoid->key)) { 1614a02a946dSIlya Dryomov ret = PTR_ERR(hoid->key); 1615a02a946dSIlya Dryomov hoid->key = NULL; 1616a02a946dSIlya Dryomov return ret; 1617a02a946dSIlya Dryomov } 1618a02a946dSIlya Dryomov 1619a02a946dSIlya Dryomov hoid->oid = ceph_extract_encoded_string(p, end, &hoid->oid_len, 1620a02a946dSIlya Dryomov GFP_NOIO); 1621a02a946dSIlya Dryomov if (IS_ERR(hoid->oid)) { 1622a02a946dSIlya Dryomov ret = PTR_ERR(hoid->oid); 1623a02a946dSIlya Dryomov hoid->oid = NULL; 1624a02a946dSIlya Dryomov return ret; 1625a02a946dSIlya Dryomov } 1626a02a946dSIlya Dryomov 1627a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->snapid, e_inval); 1628a02a946dSIlya Dryomov ceph_decode_32_safe(p, end, hoid->hash, e_inval); 1629a02a946dSIlya Dryomov ceph_decode_8_safe(p, end, hoid->is_max, e_inval); 1630a02a946dSIlya Dryomov 1631a02a946dSIlya Dryomov hoid->nspace = ceph_extract_encoded_string(p, end, &hoid->nspace_len, 1632a02a946dSIlya Dryomov GFP_NOIO); 1633a02a946dSIlya Dryomov if (IS_ERR(hoid->nspace)) { 1634a02a946dSIlya Dryomov ret = PTR_ERR(hoid->nspace); 1635a02a946dSIlya Dryomov hoid->nspace = NULL; 1636a02a946dSIlya Dryomov return ret; 1637a02a946dSIlya Dryomov } 1638a02a946dSIlya Dryomov 1639a02a946dSIlya Dryomov ceph_decode_64_safe(p, end, hoid->pool, e_inval); 1640a02a946dSIlya Dryomov 1641a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1642a02a946dSIlya Dryomov return 0; 1643a02a946dSIlya Dryomov 1644a02a946dSIlya Dryomov e_inval: 1645a02a946dSIlya Dryomov return -EINVAL; 1646a02a946dSIlya Dryomov } 1647a02a946dSIlya Dryomov 1648a02a946dSIlya Dryomov static int hoid_encoding_size(const struct ceph_hobject_id *hoid) 1649a02a946dSIlya Dryomov { 1650a02a946dSIlya Dryomov return 8 + 4 + 1 + 8 + /* snapid, hash, is_max, pool */ 1651a02a946dSIlya Dryomov 4 + hoid->key_len + 4 + hoid->oid_len + 4 + hoid->nspace_len; 1652a02a946dSIlya Dryomov } 1653a02a946dSIlya Dryomov 1654a02a946dSIlya Dryomov static void encode_hoid(void **p, void *end, const struct ceph_hobject_id *hoid) 1655a02a946dSIlya Dryomov { 1656a02a946dSIlya Dryomov ceph_start_encoding(p, 4, 3, hoid_encoding_size(hoid)); 1657a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->key, hoid->key_len); 1658a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->oid, hoid->oid_len); 1659a02a946dSIlya Dryomov ceph_encode_64(p, hoid->snapid); 1660a02a946dSIlya Dryomov ceph_encode_32(p, hoid->hash); 1661a02a946dSIlya Dryomov ceph_encode_8(p, hoid->is_max); 1662a02a946dSIlya Dryomov ceph_encode_string(p, end, hoid->nspace, hoid->nspace_len); 1663a02a946dSIlya Dryomov ceph_encode_64(p, hoid->pool); 1664a02a946dSIlya Dryomov } 1665a02a946dSIlya Dryomov 1666a02a946dSIlya Dryomov static void free_hoid(struct ceph_hobject_id *hoid) 1667a02a946dSIlya Dryomov { 1668a02a946dSIlya Dryomov if (hoid) { 1669a02a946dSIlya Dryomov kfree(hoid->key); 1670a02a946dSIlya Dryomov kfree(hoid->oid); 1671a02a946dSIlya Dryomov kfree(hoid->nspace); 1672a02a946dSIlya Dryomov kfree(hoid); 1673a02a946dSIlya Dryomov } 1674a02a946dSIlya Dryomov } 1675a02a946dSIlya Dryomov 1676a02a946dSIlya Dryomov static struct ceph_osd_backoff *alloc_backoff(void) 1677a02a946dSIlya Dryomov { 1678a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1679a02a946dSIlya Dryomov 1680a02a946dSIlya Dryomov backoff = kzalloc(sizeof(*backoff), GFP_NOIO); 1681a02a946dSIlya Dryomov if (!backoff) 1682a02a946dSIlya Dryomov return NULL; 1683a02a946dSIlya Dryomov 1684a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->spg_node); 1685a02a946dSIlya Dryomov RB_CLEAR_NODE(&backoff->id_node); 1686a02a946dSIlya Dryomov return backoff; 1687a02a946dSIlya Dryomov } 1688a02a946dSIlya Dryomov 1689a02a946dSIlya Dryomov static void free_backoff(struct ceph_osd_backoff *backoff) 1690a02a946dSIlya Dryomov { 1691a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->spg_node)); 1692a02a946dSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&backoff->id_node)); 1693a02a946dSIlya Dryomov 1694a02a946dSIlya Dryomov free_hoid(backoff->begin); 1695a02a946dSIlya Dryomov free_hoid(backoff->end); 1696a02a946dSIlya Dryomov kfree(backoff); 1697a02a946dSIlya Dryomov } 1698a02a946dSIlya Dryomov 1699a02a946dSIlya Dryomov /* 1700a02a946dSIlya Dryomov * Within a specific spgid, backoffs are managed by ->begin hoid. 1701a02a946dSIlya Dryomov */ 1702a02a946dSIlya Dryomov DEFINE_RB_INSDEL_FUNCS2(backoff, struct ceph_osd_backoff, begin, hoid_compare, 1703a02a946dSIlya Dryomov RB_BYVAL, spg_node); 1704a02a946dSIlya Dryomov 1705a02a946dSIlya Dryomov static struct ceph_osd_backoff *lookup_containing_backoff(struct rb_root *root, 1706a02a946dSIlya Dryomov const struct ceph_hobject_id *hoid) 1707a02a946dSIlya Dryomov { 1708a02a946dSIlya Dryomov struct rb_node *n = root->rb_node; 1709a02a946dSIlya Dryomov 1710a02a946dSIlya Dryomov while (n) { 1711a02a946dSIlya Dryomov struct ceph_osd_backoff *cur = 1712a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_backoff, spg_node); 1713a02a946dSIlya Dryomov int cmp; 1714a02a946dSIlya Dryomov 1715a02a946dSIlya Dryomov cmp = hoid_compare(hoid, cur->begin); 1716a02a946dSIlya Dryomov if (cmp < 0) { 1717a02a946dSIlya Dryomov n = n->rb_left; 1718a02a946dSIlya Dryomov } else if (cmp > 0) { 1719a02a946dSIlya Dryomov if (hoid_compare(hoid, cur->end) < 0) 1720a02a946dSIlya Dryomov return cur; 1721a02a946dSIlya Dryomov 1722a02a946dSIlya Dryomov n = n->rb_right; 1723a02a946dSIlya Dryomov } else { 1724a02a946dSIlya Dryomov return cur; 1725a02a946dSIlya Dryomov } 1726a02a946dSIlya Dryomov } 1727a02a946dSIlya Dryomov 1728a02a946dSIlya Dryomov return NULL; 1729a02a946dSIlya Dryomov } 1730a02a946dSIlya Dryomov 1731a02a946dSIlya Dryomov /* 1732a02a946dSIlya Dryomov * Each backoff has a unique id within its OSD session. 1733a02a946dSIlya Dryomov */ 1734a02a946dSIlya Dryomov DEFINE_RB_FUNCS(backoff_by_id, struct ceph_osd_backoff, id, id_node) 1735a02a946dSIlya Dryomov 1736a02a946dSIlya Dryomov static void clear_backoffs(struct ceph_osd *osd) 1737a02a946dSIlya Dryomov { 1738a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&osd->o_backoff_mappings)) { 1739a02a946dSIlya Dryomov struct ceph_spg_mapping *spg = 1740a02a946dSIlya Dryomov rb_entry(rb_first(&osd->o_backoff_mappings), 1741a02a946dSIlya Dryomov struct ceph_spg_mapping, node); 1742a02a946dSIlya Dryomov 1743a02a946dSIlya Dryomov while (!RB_EMPTY_ROOT(&spg->backoffs)) { 1744a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff = 1745a02a946dSIlya Dryomov rb_entry(rb_first(&spg->backoffs), 1746a02a946dSIlya Dryomov struct ceph_osd_backoff, spg_node); 1747a02a946dSIlya Dryomov 1748a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 1749a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 1750a02a946dSIlya Dryomov free_backoff(backoff); 1751a02a946dSIlya Dryomov } 1752a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 1753a02a946dSIlya Dryomov free_spg_mapping(spg); 1754a02a946dSIlya Dryomov } 1755a02a946dSIlya Dryomov } 1756a02a946dSIlya Dryomov 1757a02a946dSIlya Dryomov /* 1758a02a946dSIlya Dryomov * Set up a temporary, non-owning view into @t. 1759a02a946dSIlya Dryomov */ 1760a02a946dSIlya Dryomov static void hoid_fill_from_target(struct ceph_hobject_id *hoid, 1761a02a946dSIlya Dryomov const struct ceph_osd_request_target *t) 1762a02a946dSIlya Dryomov { 1763a02a946dSIlya Dryomov hoid->key = NULL; 1764a02a946dSIlya Dryomov hoid->key_len = 0; 1765a02a946dSIlya Dryomov hoid->oid = t->target_oid.name; 1766a02a946dSIlya Dryomov hoid->oid_len = t->target_oid.name_len; 1767a02a946dSIlya Dryomov hoid->snapid = CEPH_NOSNAP; 1768a02a946dSIlya Dryomov hoid->hash = t->pgid.seed; 1769a02a946dSIlya Dryomov hoid->is_max = false; 1770a02a946dSIlya Dryomov if (t->target_oloc.pool_ns) { 1771a02a946dSIlya Dryomov hoid->nspace = t->target_oloc.pool_ns->str; 1772a02a946dSIlya Dryomov hoid->nspace_len = t->target_oloc.pool_ns->len; 1773a02a946dSIlya Dryomov } else { 1774a02a946dSIlya Dryomov hoid->nspace = NULL; 1775a02a946dSIlya Dryomov hoid->nspace_len = 0; 1776a02a946dSIlya Dryomov } 1777a02a946dSIlya Dryomov hoid->pool = t->target_oloc.pool; 1778a02a946dSIlya Dryomov ceph_hoid_build_hash_cache(hoid); 1779a02a946dSIlya Dryomov } 1780a02a946dSIlya Dryomov 1781a02a946dSIlya Dryomov static bool should_plug_request(struct ceph_osd_request *req) 1782a02a946dSIlya Dryomov { 1783a02a946dSIlya Dryomov struct ceph_osd *osd = req->r_osd; 1784a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 1785a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 1786a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 1787a02a946dSIlya Dryomov 1788a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &req->r_t.spgid); 1789a02a946dSIlya Dryomov if (!spg) 1790a02a946dSIlya Dryomov return false; 1791a02a946dSIlya Dryomov 1792a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, &req->r_t); 1793a02a946dSIlya Dryomov backoff = lookup_containing_backoff(&spg->backoffs, &hoid); 1794a02a946dSIlya Dryomov if (!backoff) 1795a02a946dSIlya Dryomov return false; 1796a02a946dSIlya Dryomov 1797a02a946dSIlya Dryomov dout("%s req %p tid %llu backoff osd%d spgid %llu.%xs%d id %llu\n", 1798a02a946dSIlya Dryomov __func__, req, req->r_tid, osd->o_osd, backoff->spgid.pgid.pool, 1799a02a946dSIlya Dryomov backoff->spgid.pgid.seed, backoff->spgid.shard, backoff->id); 1800a02a946dSIlya Dryomov return true; 1801a02a946dSIlya Dryomov } 1802a02a946dSIlya Dryomov 1803bb873b53SIlya Dryomov static void setup_request_data(struct ceph_osd_request *req, 1804bb873b53SIlya Dryomov struct ceph_msg *msg) 18053d14c5d2SYehuda Sadeh { 1806bb873b53SIlya Dryomov u32 data_len = 0; 1807bb873b53SIlya Dryomov int i; 18083d14c5d2SYehuda Sadeh 1809bb873b53SIlya Dryomov if (!list_empty(&msg->data)) 1810bb873b53SIlya Dryomov return; 18113d14c5d2SYehuda Sadeh 1812bb873b53SIlya Dryomov WARN_ON(msg->data_length); 1813bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 1814bb873b53SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[i]; 1815bb873b53SIlya Dryomov 1816bb873b53SIlya Dryomov switch (op->op) { 1817bb873b53SIlya Dryomov /* request */ 1818bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITE: 1819bb873b53SIlya Dryomov case CEPH_OSD_OP_WRITEFULL: 1820bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->extent.length); 1821bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->extent.osd_data); 1822bb873b53SIlya Dryomov break; 1823bb873b53SIlya Dryomov case CEPH_OSD_OP_SETXATTR: 1824bb873b53SIlya Dryomov case CEPH_OSD_OP_CMPXATTR: 1825bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->xattr.name_len + 1826bb873b53SIlya Dryomov op->xattr.value_len); 1827bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->xattr.osd_data); 1828bb873b53SIlya Dryomov break; 1829922dab61SIlya Dryomov case CEPH_OSD_OP_NOTIFY_ACK: 1830922dab61SIlya Dryomov ceph_osdc_msg_data_add(msg, 1831922dab61SIlya Dryomov &op->notify_ack.request_data); 1832922dab61SIlya Dryomov break; 1833bb873b53SIlya Dryomov 1834bb873b53SIlya Dryomov /* reply */ 1835bb873b53SIlya Dryomov case CEPH_OSD_OP_STAT: 1836bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1837bb873b53SIlya Dryomov &op->raw_data_in); 1838bb873b53SIlya Dryomov break; 1839bb873b53SIlya Dryomov case CEPH_OSD_OP_READ: 1840bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1841bb873b53SIlya Dryomov &op->extent.osd_data); 1842bb873b53SIlya Dryomov break; 1843a4ed38d7SDouglas Fuller case CEPH_OSD_OP_LIST_WATCHERS: 1844a4ed38d7SDouglas Fuller ceph_osdc_msg_data_add(req->r_reply, 1845a4ed38d7SDouglas Fuller &op->list_watchers.response_data); 1846a4ed38d7SDouglas Fuller break; 1847bb873b53SIlya Dryomov 1848bb873b53SIlya Dryomov /* both */ 1849bb873b53SIlya Dryomov case CEPH_OSD_OP_CALL: 1850bb873b53SIlya Dryomov WARN_ON(op->indata_len != op->cls.class_len + 1851bb873b53SIlya Dryomov op->cls.method_len + 1852bb873b53SIlya Dryomov op->cls.indata_len); 1853bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->cls.request_info); 1854bb873b53SIlya Dryomov /* optional, can be NONE */ 1855bb873b53SIlya Dryomov ceph_osdc_msg_data_add(msg, &op->cls.request_data); 1856bb873b53SIlya Dryomov /* optional, can be NONE */ 1857bb873b53SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 1858bb873b53SIlya Dryomov &op->cls.response_data); 1859bb873b53SIlya Dryomov break; 186019079203SIlya Dryomov case CEPH_OSD_OP_NOTIFY: 186119079203SIlya Dryomov ceph_osdc_msg_data_add(msg, 186219079203SIlya Dryomov &op->notify.request_data); 186319079203SIlya Dryomov ceph_osdc_msg_data_add(req->r_reply, 186419079203SIlya Dryomov &op->notify.response_data); 186519079203SIlya Dryomov break; 1866bb873b53SIlya Dryomov } 1867bb873b53SIlya Dryomov 1868bb873b53SIlya Dryomov data_len += op->indata_len; 1869bb873b53SIlya Dryomov } 1870bb873b53SIlya Dryomov 1871bb873b53SIlya Dryomov WARN_ON(data_len != msg->data_length); 1872bb873b53SIlya Dryomov } 1873bb873b53SIlya Dryomov 18742e59ffd1SIlya Dryomov static void encode_pgid(void **p, const struct ceph_pg *pgid) 18752e59ffd1SIlya Dryomov { 18762e59ffd1SIlya Dryomov ceph_encode_8(p, 1); 18772e59ffd1SIlya Dryomov ceph_encode_64(p, pgid->pool); 18782e59ffd1SIlya Dryomov ceph_encode_32(p, pgid->seed); 18792e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 18802e59ffd1SIlya Dryomov } 18812e59ffd1SIlya Dryomov 18828cb441c0SIlya Dryomov static void encode_spgid(void **p, const struct ceph_spg *spgid) 18838cb441c0SIlya Dryomov { 18848cb441c0SIlya Dryomov ceph_start_encoding(p, 1, 1, CEPH_PGID_ENCODING_LEN + 1); 18858cb441c0SIlya Dryomov encode_pgid(p, &spgid->pgid); 18868cb441c0SIlya Dryomov ceph_encode_8(p, spgid->shard); 18878cb441c0SIlya Dryomov } 18888cb441c0SIlya Dryomov 18892e59ffd1SIlya Dryomov static void encode_oloc(void **p, void *end, 18902e59ffd1SIlya Dryomov const struct ceph_object_locator *oloc) 18912e59ffd1SIlya Dryomov { 18922e59ffd1SIlya Dryomov ceph_start_encoding(p, 5, 4, ceph_oloc_encoding_size(oloc)); 18932e59ffd1SIlya Dryomov ceph_encode_64(p, oloc->pool); 18942e59ffd1SIlya Dryomov ceph_encode_32(p, -1); /* preferred */ 18952e59ffd1SIlya Dryomov ceph_encode_32(p, 0); /* key len */ 18962e59ffd1SIlya Dryomov if (oloc->pool_ns) 18972e59ffd1SIlya Dryomov ceph_encode_string(p, end, oloc->pool_ns->str, 18982e59ffd1SIlya Dryomov oloc->pool_ns->len); 18992e59ffd1SIlya Dryomov else 19002e59ffd1SIlya Dryomov ceph_encode_32(p, 0); 19012e59ffd1SIlya Dryomov } 19022e59ffd1SIlya Dryomov 19038cb441c0SIlya Dryomov static void encode_request_partial(struct ceph_osd_request *req, 19048cb441c0SIlya Dryomov struct ceph_msg *msg) 1905bb873b53SIlya Dryomov { 1906bb873b53SIlya Dryomov void *p = msg->front.iov_base; 1907bb873b53SIlya Dryomov void *const end = p + msg->front_alloc_len; 1908bb873b53SIlya Dryomov u32 data_len = 0; 1909bb873b53SIlya Dryomov int i; 1910bb873b53SIlya Dryomov 1911bb873b53SIlya Dryomov if (req->r_flags & CEPH_OSD_FLAG_WRITE) { 1912bb873b53SIlya Dryomov /* snapshots aren't writeable */ 1913bb873b53SIlya Dryomov WARN_ON(req->r_snapid != CEPH_NOSNAP); 1914bb873b53SIlya Dryomov } else { 1915bb873b53SIlya Dryomov WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec || 1916bb873b53SIlya Dryomov req->r_data_offset || req->r_snapc); 1917bb873b53SIlya Dryomov } 1918bb873b53SIlya Dryomov 1919bb873b53SIlya Dryomov setup_request_data(req, msg); 1920bb873b53SIlya Dryomov 19218cb441c0SIlya Dryomov encode_spgid(&p, &req->r_t.spgid); /* actual spg */ 19228cb441c0SIlya Dryomov ceph_encode_32(&p, req->r_t.pgid.seed); /* raw hash */ 1923bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_osdc->osdmap->epoch); 1924bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_flags); 19258cb441c0SIlya Dryomov 19268cb441c0SIlya Dryomov /* reqid */ 19278cb441c0SIlya Dryomov ceph_start_encoding(&p, 2, 2, sizeof(struct ceph_osd_reqid)); 19288cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_osd_reqid)); 19298cb441c0SIlya Dryomov p += sizeof(struct ceph_osd_reqid); 19308cb441c0SIlya Dryomov 19318cb441c0SIlya Dryomov /* trace */ 19328cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_blkin_trace_info)); 19338cb441c0SIlya Dryomov p += sizeof(struct ceph_blkin_trace_info); 19348cb441c0SIlya Dryomov 19358cb441c0SIlya Dryomov ceph_encode_32(&p, 0); /* client_inc, always 0 */ 1936bb873b53SIlya Dryomov ceph_encode_timespec(p, &req->r_mtime); 1937bb873b53SIlya Dryomov p += sizeof(struct ceph_timespec); 1938aa26d662SJeff Layton 19392e59ffd1SIlya Dryomov encode_oloc(&p, end, &req->r_t.target_oloc); 19402e59ffd1SIlya Dryomov ceph_encode_string(&p, end, req->r_t.target_oid.name, 19412e59ffd1SIlya Dryomov req->r_t.target_oid.name_len); 1942bb873b53SIlya Dryomov 1943bb873b53SIlya Dryomov /* ops, can imply data */ 1944bb873b53SIlya Dryomov ceph_encode_16(&p, req->r_num_ops); 1945bb873b53SIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 1946bb873b53SIlya Dryomov data_len += osd_req_encode_op(p, &req->r_ops[i]); 1947bb873b53SIlya Dryomov p += sizeof(struct ceph_osd_op); 1948bb873b53SIlya Dryomov } 1949bb873b53SIlya Dryomov 1950bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapid); /* snapid */ 1951bb873b53SIlya Dryomov if (req->r_snapc) { 1952bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->seq); 1953bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_snapc->num_snaps); 1954bb873b53SIlya Dryomov for (i = 0; i < req->r_snapc->num_snaps; i++) 1955bb873b53SIlya Dryomov ceph_encode_64(&p, req->r_snapc->snaps[i]); 1956bb873b53SIlya Dryomov } else { 1957bb873b53SIlya Dryomov ceph_encode_64(&p, 0); /* snap_seq */ 1958bb873b53SIlya Dryomov ceph_encode_32(&p, 0); /* snaps len */ 1959bb873b53SIlya Dryomov } 1960bb873b53SIlya Dryomov 1961bb873b53SIlya Dryomov ceph_encode_32(&p, req->r_attempts); /* retry_attempt */ 1962986e8989SIlya Dryomov BUG_ON(p > end - 8); /* space for features */ 1963bb873b53SIlya Dryomov 19648cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(8); /* MOSDOp v8 */ 19658cb441c0SIlya Dryomov /* front_len is finalized in encode_request_finish() */ 1966986e8989SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 1967986e8989SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 1968bb873b53SIlya Dryomov msg->hdr.data_len = cpu_to_le32(data_len); 1969bb873b53SIlya Dryomov /* 1970bb873b53SIlya Dryomov * The header "data_off" is a hint to the receiver allowing it 1971bb873b53SIlya Dryomov * to align received data into its buffers such that there's no 1972bb873b53SIlya Dryomov * need to re-copy it before writing it to disk (direct I/O). 1973bb873b53SIlya Dryomov */ 1974bb873b53SIlya Dryomov msg->hdr.data_off = cpu_to_le16(req->r_data_offset); 1975bb873b53SIlya Dryomov 19768cb441c0SIlya Dryomov dout("%s req %p msg %p oid %s oid_len %d\n", __func__, req, msg, 19778cb441c0SIlya Dryomov req->r_t.target_oid.name, req->r_t.target_oid.name_len); 19788cb441c0SIlya Dryomov } 19798cb441c0SIlya Dryomov 19808cb441c0SIlya Dryomov static void encode_request_finish(struct ceph_msg *msg) 19818cb441c0SIlya Dryomov { 19828cb441c0SIlya Dryomov void *p = msg->front.iov_base; 1983986e8989SIlya Dryomov void *const partial_end = p + msg->front.iov_len; 19848cb441c0SIlya Dryomov void *const end = p + msg->front_alloc_len; 19858cb441c0SIlya Dryomov 19868cb441c0SIlya Dryomov if (CEPH_HAVE_FEATURE(msg->con->peer_features, RESEND_ON_SPLIT)) { 19878cb441c0SIlya Dryomov /* luminous OSD -- encode features and be done */ 1988986e8989SIlya Dryomov p = partial_end; 19898cb441c0SIlya Dryomov ceph_encode_64(&p, msg->con->peer_features); 19908cb441c0SIlya Dryomov } else { 19918cb441c0SIlya Dryomov struct { 19928cb441c0SIlya Dryomov char spgid[CEPH_ENCODING_START_BLK_LEN + 19938cb441c0SIlya Dryomov CEPH_PGID_ENCODING_LEN + 1]; 19948cb441c0SIlya Dryomov __le32 hash; 19958cb441c0SIlya Dryomov __le32 epoch; 19968cb441c0SIlya Dryomov __le32 flags; 19978cb441c0SIlya Dryomov char reqid[CEPH_ENCODING_START_BLK_LEN + 19988cb441c0SIlya Dryomov sizeof(struct ceph_osd_reqid)]; 19998cb441c0SIlya Dryomov char trace[sizeof(struct ceph_blkin_trace_info)]; 20008cb441c0SIlya Dryomov __le32 client_inc; 20018cb441c0SIlya Dryomov struct ceph_timespec mtime; 20028cb441c0SIlya Dryomov } __packed head; 20038cb441c0SIlya Dryomov struct ceph_pg pgid; 20048cb441c0SIlya Dryomov void *oloc, *oid, *tail; 20058cb441c0SIlya Dryomov int oloc_len, oid_len, tail_len; 20068cb441c0SIlya Dryomov int len; 20078cb441c0SIlya Dryomov 20088cb441c0SIlya Dryomov /* 20098cb441c0SIlya Dryomov * Pre-luminous OSD -- reencode v8 into v4 using @head 20108cb441c0SIlya Dryomov * as a temporary buffer. Encode the raw PG; the rest 20118cb441c0SIlya Dryomov * is just a matter of moving oloc, oid and tail blobs 20128cb441c0SIlya Dryomov * around. 20138cb441c0SIlya Dryomov */ 20148cb441c0SIlya Dryomov memcpy(&head, p, sizeof(head)); 20158cb441c0SIlya Dryomov p += sizeof(head); 20168cb441c0SIlya Dryomov 20178cb441c0SIlya Dryomov oloc = p; 20188cb441c0SIlya Dryomov p += CEPH_ENCODING_START_BLK_LEN; 20198cb441c0SIlya Dryomov pgid.pool = ceph_decode_64(&p); 20208cb441c0SIlya Dryomov p += 4 + 4; /* preferred, key len */ 20218cb441c0SIlya Dryomov len = ceph_decode_32(&p); 20228cb441c0SIlya Dryomov p += len; /* nspace */ 20238cb441c0SIlya Dryomov oloc_len = p - oloc; 20248cb441c0SIlya Dryomov 20258cb441c0SIlya Dryomov oid = p; 20268cb441c0SIlya Dryomov len = ceph_decode_32(&p); 20278cb441c0SIlya Dryomov p += len; 20288cb441c0SIlya Dryomov oid_len = p - oid; 20298cb441c0SIlya Dryomov 20308cb441c0SIlya Dryomov tail = p; 2031986e8989SIlya Dryomov tail_len = partial_end - p; 20328cb441c0SIlya Dryomov 20338cb441c0SIlya Dryomov p = msg->front.iov_base; 20348cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.client_inc, sizeof(head.client_inc)); 20358cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.epoch, sizeof(head.epoch)); 20368cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.flags, sizeof(head.flags)); 20378cb441c0SIlya Dryomov ceph_encode_copy(&p, &head.mtime, sizeof(head.mtime)); 20388cb441c0SIlya Dryomov 20398cb441c0SIlya Dryomov /* reassert_version */ 20408cb441c0SIlya Dryomov memset(p, 0, sizeof(struct ceph_eversion)); 20418cb441c0SIlya Dryomov p += sizeof(struct ceph_eversion); 20428cb441c0SIlya Dryomov 20438cb441c0SIlya Dryomov BUG_ON(p >= oloc); 20448cb441c0SIlya Dryomov memmove(p, oloc, oloc_len); 20458cb441c0SIlya Dryomov p += oloc_len; 20468cb441c0SIlya Dryomov 20478cb441c0SIlya Dryomov pgid.seed = le32_to_cpu(head.hash); 20488cb441c0SIlya Dryomov encode_pgid(&p, &pgid); /* raw pg */ 20498cb441c0SIlya Dryomov 20508cb441c0SIlya Dryomov BUG_ON(p >= oid); 20518cb441c0SIlya Dryomov memmove(p, oid, oid_len); 20528cb441c0SIlya Dryomov p += oid_len; 20538cb441c0SIlya Dryomov 20548cb441c0SIlya Dryomov /* tail -- ops, snapid, snapc, retry_attempt */ 20558cb441c0SIlya Dryomov BUG_ON(p >= tail); 20568cb441c0SIlya Dryomov memmove(p, tail, tail_len); 20578cb441c0SIlya Dryomov p += tail_len; 20588cb441c0SIlya Dryomov 20598cb441c0SIlya Dryomov msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */ 20608cb441c0SIlya Dryomov } 20618cb441c0SIlya Dryomov 20628cb441c0SIlya Dryomov BUG_ON(p > end); 20638cb441c0SIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 20648cb441c0SIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 20658cb441c0SIlya Dryomov 20668cb441c0SIlya Dryomov dout("%s msg %p tid %llu %u+%u+%u v%d\n", __func__, msg, 20678cb441c0SIlya Dryomov le64_to_cpu(msg->hdr.tid), le32_to_cpu(msg->hdr.front_len), 20688cb441c0SIlya Dryomov le32_to_cpu(msg->hdr.middle_len), le32_to_cpu(msg->hdr.data_len), 20698cb441c0SIlya Dryomov le16_to_cpu(msg->hdr.version)); 2070bb873b53SIlya Dryomov } 2071bb873b53SIlya Dryomov 2072bb873b53SIlya Dryomov /* 2073bb873b53SIlya Dryomov * @req has to be assigned a tid and registered. 2074bb873b53SIlya Dryomov */ 2075bb873b53SIlya Dryomov static void send_request(struct ceph_osd_request *req) 2076bb873b53SIlya Dryomov { 2077bb873b53SIlya Dryomov struct ceph_osd *osd = req->r_osd; 2078bb873b53SIlya Dryomov 20795aea3dcdSIlya Dryomov verify_osd_locked(osd); 2080bb873b53SIlya Dryomov WARN_ON(osd->o_osd != req->r_t.osd); 2081bb873b53SIlya Dryomov 2082a02a946dSIlya Dryomov /* backoff? */ 2083a02a946dSIlya Dryomov if (should_plug_request(req)) 2084a02a946dSIlya Dryomov return; 2085a02a946dSIlya Dryomov 20865aea3dcdSIlya Dryomov /* 20875aea3dcdSIlya Dryomov * We may have a previously queued request message hanging 20885aea3dcdSIlya Dryomov * around. Cancel it to avoid corrupting the msgr. 20895aea3dcdSIlya Dryomov */ 20905aea3dcdSIlya Dryomov if (req->r_sent) 20915aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 20925aea3dcdSIlya Dryomov 2093bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR; 2094bb873b53SIlya Dryomov if (req->r_attempts) 2095bb873b53SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_RETRY; 2096bb873b53SIlya Dryomov else 2097bb873b53SIlya Dryomov WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY); 2098bb873b53SIlya Dryomov 20998cb441c0SIlya Dryomov encode_request_partial(req, req->r_request); 2100bb873b53SIlya Dryomov 210104c7d789SIlya 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", 2102bb873b53SIlya Dryomov __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed, 2103dc98ff72SIlya Dryomov req->r_t.spgid.pgid.pool, req->r_t.spgid.pgid.seed, 210404c7d789SIlya Dryomov req->r_t.spgid.shard, osd->o_osd, req->r_t.epoch, req->r_flags, 210504c7d789SIlya Dryomov req->r_attempts); 2106bb873b53SIlya Dryomov 2107bb873b53SIlya Dryomov req->r_t.paused = false; 21083d14c5d2SYehuda Sadeh req->r_stamp = jiffies; 2109bb873b53SIlya Dryomov req->r_attempts++; 21103d14c5d2SYehuda Sadeh 2111bb873b53SIlya Dryomov req->r_sent = osd->o_incarnation; 2112bb873b53SIlya Dryomov req->r_request->hdr.tid = cpu_to_le64(req->r_tid); 2113bb873b53SIlya Dryomov ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request)); 21143d14c5d2SYehuda Sadeh } 21153d14c5d2SYehuda Sadeh 211642c1b124SIlya Dryomov static void maybe_request_map(struct ceph_osd_client *osdc) 211742c1b124SIlya Dryomov { 211842c1b124SIlya Dryomov bool continuous = false; 211942c1b124SIlya Dryomov 21205aea3dcdSIlya Dryomov verify_osdc_locked(osdc); 212142c1b124SIlya Dryomov WARN_ON(!osdc->osdmap->epoch); 212242c1b124SIlya Dryomov 2123b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2124b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD) || 2125b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 212642c1b124SIlya Dryomov dout("%s osdc %p continuous\n", __func__, osdc); 212742c1b124SIlya Dryomov continuous = true; 212842c1b124SIlya Dryomov } else { 212942c1b124SIlya Dryomov dout("%s osdc %p onetime\n", __func__, osdc); 213042c1b124SIlya Dryomov } 213142c1b124SIlya Dryomov 213242c1b124SIlya Dryomov if (ceph_monc_want_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 213342c1b124SIlya Dryomov osdc->osdmap->epoch + 1, continuous)) 213442c1b124SIlya Dryomov ceph_monc_renew_subs(&osdc->client->monc); 213542c1b124SIlya Dryomov } 213642c1b124SIlya Dryomov 2137a1f4020aSJeff Layton static void complete_request(struct ceph_osd_request *req, int err); 21384609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req); 21394609245eSIlya Dryomov 21405aea3dcdSIlya Dryomov static void __submit_request(struct ceph_osd_request *req, bool wrlocked) 21410bbfdfe8SIlya Dryomov { 21425aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 21435aea3dcdSIlya Dryomov struct ceph_osd *osd; 21444609245eSIlya Dryomov enum calc_target_result ct_res; 21455aea3dcdSIlya Dryomov bool need_send = false; 21465aea3dcdSIlya Dryomov bool promoted = false; 2147a1f4020aSJeff Layton bool need_abort = false; 21480bbfdfe8SIlya Dryomov 2149b18b9550SIlya Dryomov WARN_ON(req->r_tid); 21505aea3dcdSIlya Dryomov dout("%s req %p wrlocked %d\n", __func__, req, wrlocked); 21515aea3dcdSIlya Dryomov 21525aea3dcdSIlya Dryomov again: 21537de030d6SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, NULL, false); 21544609245eSIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE && !wrlocked) 21554609245eSIlya Dryomov goto promote; 21564609245eSIlya Dryomov 21575aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, wrlocked); 21585aea3dcdSIlya Dryomov if (IS_ERR(osd)) { 21595aea3dcdSIlya Dryomov WARN_ON(PTR_ERR(osd) != -EAGAIN || wrlocked); 21605aea3dcdSIlya Dryomov goto promote; 21615aea3dcdSIlya Dryomov } 21625aea3dcdSIlya Dryomov 216358eb7932SJeff Layton if (osdc->osdmap->epoch < osdc->epoch_barrier) { 216458eb7932SJeff Layton dout("req %p epoch %u barrier %u\n", req, osdc->osdmap->epoch, 216558eb7932SJeff Layton osdc->epoch_barrier); 216658eb7932SJeff Layton req->r_t.paused = true; 216758eb7932SJeff Layton maybe_request_map(osdc); 216858eb7932SJeff Layton } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 2169b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) { 21705aea3dcdSIlya Dryomov dout("req %p pausewr\n", req); 21715aea3dcdSIlya Dryomov req->r_t.paused = true; 21725aea3dcdSIlya Dryomov maybe_request_map(osdc); 21735aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_READ) && 2174b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 21755aea3dcdSIlya Dryomov dout("req %p pauserd\n", req); 21765aea3dcdSIlya Dryomov req->r_t.paused = true; 21775aea3dcdSIlya Dryomov maybe_request_map(osdc); 21785aea3dcdSIlya Dryomov } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) && 21795aea3dcdSIlya Dryomov !(req->r_flags & (CEPH_OSD_FLAG_FULL_TRY | 21805aea3dcdSIlya Dryomov CEPH_OSD_FLAG_FULL_FORCE)) && 2181b7ec35b3SIlya Dryomov (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 21825aea3dcdSIlya Dryomov pool_full(osdc, req->r_t.base_oloc.pool))) { 21835aea3dcdSIlya Dryomov dout("req %p full/pool_full\n", req); 21845aea3dcdSIlya Dryomov pr_warn_ratelimited("FULL or reached pool quota\n"); 21855aea3dcdSIlya Dryomov req->r_t.paused = true; 21865aea3dcdSIlya Dryomov maybe_request_map(osdc); 2187a1f4020aSJeff Layton if (req->r_abort_on_full) 2188a1f4020aSJeff Layton need_abort = true; 21895aea3dcdSIlya Dryomov } else if (!osd_homeless(osd)) { 21905aea3dcdSIlya Dryomov need_send = true; 21910bbfdfe8SIlya Dryomov } else { 21925aea3dcdSIlya Dryomov maybe_request_map(osdc); 21930bbfdfe8SIlya Dryomov } 21940bbfdfe8SIlya Dryomov 21955aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 21965aea3dcdSIlya Dryomov /* 21975aea3dcdSIlya Dryomov * Assign the tid atomically with send_request() to protect 21985aea3dcdSIlya Dryomov * multiple writes to the same object from racing with each 21995aea3dcdSIlya Dryomov * other, resulting in out of order ops on the OSDs. 22005aea3dcdSIlya Dryomov */ 22015aea3dcdSIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 22025aea3dcdSIlya Dryomov link_request(osd, req); 22035aea3dcdSIlya Dryomov if (need_send) 22045aea3dcdSIlya Dryomov send_request(req); 2205a1f4020aSJeff Layton else if (need_abort) 2206a1f4020aSJeff Layton complete_request(req, -ENOSPC); 22075aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 22085aea3dcdSIlya Dryomov 22094609245eSIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE) 22104609245eSIlya Dryomov send_map_check(req); 22114609245eSIlya Dryomov 22125aea3dcdSIlya Dryomov if (promoted) 22135aea3dcdSIlya Dryomov downgrade_write(&osdc->lock); 22145aea3dcdSIlya Dryomov return; 22155aea3dcdSIlya Dryomov 22165aea3dcdSIlya Dryomov promote: 22175aea3dcdSIlya Dryomov up_read(&osdc->lock); 22185aea3dcdSIlya Dryomov down_write(&osdc->lock); 22195aea3dcdSIlya Dryomov wrlocked = true; 22205aea3dcdSIlya Dryomov promoted = true; 22215aea3dcdSIlya Dryomov goto again; 22220bbfdfe8SIlya Dryomov } 22230bbfdfe8SIlya Dryomov 22245aea3dcdSIlya Dryomov static void account_request(struct ceph_osd_request *req) 22255aea3dcdSIlya Dryomov { 222654ea0046SIlya Dryomov WARN_ON(req->r_flags & (CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK)); 2227b18b9550SIlya Dryomov WARN_ON(!(req->r_flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE))); 22285aea3dcdSIlya Dryomov 2229b18b9550SIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_ONDISK; 22305aea3dcdSIlya Dryomov atomic_inc(&req->r_osdc->num_requests); 22317cc5e38fSIlya Dryomov 22327cc5e38fSIlya Dryomov req->r_start_stamp = jiffies; 22335aea3dcdSIlya Dryomov } 22345aea3dcdSIlya Dryomov 22355aea3dcdSIlya Dryomov static void submit_request(struct ceph_osd_request *req, bool wrlocked) 22365aea3dcdSIlya Dryomov { 22375aea3dcdSIlya Dryomov ceph_osdc_get_request(req); 22385aea3dcdSIlya Dryomov account_request(req); 22395aea3dcdSIlya Dryomov __submit_request(req, wrlocked); 22405aea3dcdSIlya Dryomov } 22415aea3dcdSIlya Dryomov 224245ee2c1dSIlya Dryomov static void finish_request(struct ceph_osd_request *req) 22435aea3dcdSIlya Dryomov { 22445aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 22455aea3dcdSIlya Dryomov 22464609245eSIlya Dryomov WARN_ON(lookup_request_mc(&osdc->map_checks, req->r_tid)); 224704c7d789SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 224804c7d789SIlya Dryomov 224904c7d789SIlya Dryomov if (req->r_osd) 225004c7d789SIlya Dryomov unlink_request(req->r_osd, req); 22515aea3dcdSIlya Dryomov atomic_dec(&osdc->num_requests); 22525aea3dcdSIlya Dryomov 22535aea3dcdSIlya Dryomov /* 22545aea3dcdSIlya Dryomov * If an OSD has failed or returned and a request has been sent 22555aea3dcdSIlya Dryomov * twice, it's possible to get a reply and end up here while the 22565aea3dcdSIlya Dryomov * request message is queued for delivery. We will ignore the 22575aea3dcdSIlya Dryomov * reply, so not a big deal, but better to try and catch it. 22585aea3dcdSIlya Dryomov */ 22595aea3dcdSIlya Dryomov ceph_msg_revoke(req->r_request); 22605aea3dcdSIlya Dryomov ceph_msg_revoke_incoming(req->r_reply); 22615aea3dcdSIlya Dryomov } 22625aea3dcdSIlya Dryomov 2263fe5da05eSIlya Dryomov static void __complete_request(struct ceph_osd_request *req) 2264fe5da05eSIlya Dryomov { 2265b18b9550SIlya Dryomov if (req->r_callback) { 2266b18b9550SIlya Dryomov dout("%s req %p tid %llu cb %pf result %d\n", __func__, req, 2267b18b9550SIlya Dryomov req->r_tid, req->r_callback, req->r_result); 2268fe5da05eSIlya Dryomov req->r_callback(req); 2269b18b9550SIlya Dryomov } 2270fe5da05eSIlya Dryomov } 2271fe5da05eSIlya Dryomov 22724609245eSIlya Dryomov /* 2273b18b9550SIlya Dryomov * This is open-coded in handle_reply(). 22744609245eSIlya Dryomov */ 22754609245eSIlya Dryomov static void complete_request(struct ceph_osd_request *req, int err) 22764609245eSIlya Dryomov { 22774609245eSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 22784609245eSIlya Dryomov 22794609245eSIlya Dryomov req->r_result = err; 228045ee2c1dSIlya Dryomov finish_request(req); 22814609245eSIlya Dryomov __complete_request(req); 2282b18b9550SIlya Dryomov complete_all(&req->r_completion); 22834609245eSIlya Dryomov ceph_osdc_put_request(req); 22844609245eSIlya Dryomov } 22854609245eSIlya Dryomov 22864609245eSIlya Dryomov static void cancel_map_check(struct ceph_osd_request *req) 22874609245eSIlya Dryomov { 22884609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 22894609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 22904609245eSIlya Dryomov 22914609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 22924609245eSIlya Dryomov 22934609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 22944609245eSIlya Dryomov if (!lookup_req) 22954609245eSIlya Dryomov return; 22964609245eSIlya Dryomov 22974609245eSIlya Dryomov WARN_ON(lookup_req != req); 22984609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 22994609245eSIlya Dryomov ceph_osdc_put_request(req); 23004609245eSIlya Dryomov } 23014609245eSIlya Dryomov 23025aea3dcdSIlya Dryomov static void cancel_request(struct ceph_osd_request *req) 23035aea3dcdSIlya Dryomov { 23045aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 23055aea3dcdSIlya Dryomov 23064609245eSIlya Dryomov cancel_map_check(req); 230745ee2c1dSIlya Dryomov finish_request(req); 2308b18b9550SIlya Dryomov complete_all(&req->r_completion); 2309c297eb42SIlya Dryomov ceph_osdc_put_request(req); 23105aea3dcdSIlya Dryomov } 23115aea3dcdSIlya Dryomov 23127cc5e38fSIlya Dryomov static void abort_request(struct ceph_osd_request *req, int err) 23137cc5e38fSIlya Dryomov { 23147cc5e38fSIlya Dryomov dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err); 23157cc5e38fSIlya Dryomov 23167cc5e38fSIlya Dryomov cancel_map_check(req); 23177cc5e38fSIlya Dryomov complete_request(req, err); 23187cc5e38fSIlya Dryomov } 23197cc5e38fSIlya Dryomov 232058eb7932SJeff Layton static void update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 232158eb7932SJeff Layton { 232258eb7932SJeff Layton if (likely(eb > osdc->epoch_barrier)) { 232358eb7932SJeff Layton dout("updating epoch_barrier from %u to %u\n", 232458eb7932SJeff Layton osdc->epoch_barrier, eb); 232558eb7932SJeff Layton osdc->epoch_barrier = eb; 232658eb7932SJeff Layton /* Request map if we're not to the barrier yet */ 232758eb7932SJeff Layton if (eb > osdc->osdmap->epoch) 232858eb7932SJeff Layton maybe_request_map(osdc); 232958eb7932SJeff Layton } 233058eb7932SJeff Layton } 233158eb7932SJeff Layton 233258eb7932SJeff Layton void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb) 233358eb7932SJeff Layton { 233458eb7932SJeff Layton down_read(&osdc->lock); 233558eb7932SJeff Layton if (unlikely(eb > osdc->epoch_barrier)) { 233658eb7932SJeff Layton up_read(&osdc->lock); 233758eb7932SJeff Layton down_write(&osdc->lock); 233858eb7932SJeff Layton update_epoch_barrier(osdc, eb); 233958eb7932SJeff Layton up_write(&osdc->lock); 234058eb7932SJeff Layton } else { 234158eb7932SJeff Layton up_read(&osdc->lock); 234258eb7932SJeff Layton } 234358eb7932SJeff Layton } 234458eb7932SJeff Layton EXPORT_SYMBOL(ceph_osdc_update_epoch_barrier); 234558eb7932SJeff Layton 2346fc36d0a4SJeff Layton /* 2347fc36d0a4SJeff Layton * Drop all pending requests that are stalled waiting on a full condition to 234858eb7932SJeff Layton * clear, and complete them with ENOSPC as the return code. Set the 234958eb7932SJeff Layton * osdc->epoch_barrier to the latest map epoch that we've seen if any were 235058eb7932SJeff Layton * cancelled. 2351fc36d0a4SJeff Layton */ 2352fc36d0a4SJeff Layton static void ceph_osdc_abort_on_full(struct ceph_osd_client *osdc) 2353fc36d0a4SJeff Layton { 2354fc36d0a4SJeff Layton struct rb_node *n; 235558eb7932SJeff Layton bool victims = false; 2356fc36d0a4SJeff Layton 2357fc36d0a4SJeff Layton dout("enter abort_on_full\n"); 2358fc36d0a4SJeff Layton 2359fc36d0a4SJeff Layton if (!ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) && !have_pool_full(osdc)) 2360fc36d0a4SJeff Layton goto out; 2361fc36d0a4SJeff Layton 236258eb7932SJeff Layton /* Scan list and see if there is anything to abort */ 236358eb7932SJeff Layton for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 236458eb7932SJeff Layton struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 236558eb7932SJeff Layton struct rb_node *m; 236658eb7932SJeff Layton 236758eb7932SJeff Layton m = rb_first(&osd->o_requests); 236858eb7932SJeff Layton while (m) { 236958eb7932SJeff Layton struct ceph_osd_request *req = rb_entry(m, 237058eb7932SJeff Layton struct ceph_osd_request, r_node); 237158eb7932SJeff Layton m = rb_next(m); 237258eb7932SJeff Layton 237358eb7932SJeff Layton if (req->r_abort_on_full) { 237458eb7932SJeff Layton victims = true; 237558eb7932SJeff Layton break; 237658eb7932SJeff Layton } 237758eb7932SJeff Layton } 237858eb7932SJeff Layton if (victims) 237958eb7932SJeff Layton break; 238058eb7932SJeff Layton } 238158eb7932SJeff Layton 238258eb7932SJeff Layton if (!victims) 238358eb7932SJeff Layton goto out; 238458eb7932SJeff Layton 238558eb7932SJeff Layton /* 238658eb7932SJeff Layton * Update the barrier to current epoch if it's behind that point, 238758eb7932SJeff Layton * since we know we have some calls to be aborted in the tree. 238858eb7932SJeff Layton */ 238958eb7932SJeff Layton update_epoch_barrier(osdc, osdc->osdmap->epoch); 239058eb7932SJeff Layton 2391fc36d0a4SJeff Layton for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 2392fc36d0a4SJeff Layton struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 2393fc36d0a4SJeff Layton struct rb_node *m; 2394fc36d0a4SJeff Layton 2395fc36d0a4SJeff Layton m = rb_first(&osd->o_requests); 2396fc36d0a4SJeff Layton while (m) { 2397fc36d0a4SJeff Layton struct ceph_osd_request *req = rb_entry(m, 2398fc36d0a4SJeff Layton struct ceph_osd_request, r_node); 2399fc36d0a4SJeff Layton m = rb_next(m); 2400fc36d0a4SJeff Layton 2401fc36d0a4SJeff Layton if (req->r_abort_on_full && 2402fc36d0a4SJeff Layton (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 2403fc36d0a4SJeff Layton pool_full(osdc, req->r_t.target_oloc.pool))) 2404fc36d0a4SJeff Layton abort_request(req, -ENOSPC); 2405fc36d0a4SJeff Layton } 2406fc36d0a4SJeff Layton } 2407fc36d0a4SJeff Layton out: 240858eb7932SJeff Layton dout("return abort_on_full barrier=%u\n", osdc->epoch_barrier); 2409fc36d0a4SJeff Layton } 2410fc36d0a4SJeff Layton 24114609245eSIlya Dryomov static void check_pool_dne(struct ceph_osd_request *req) 24124609245eSIlya Dryomov { 24134609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 24144609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 24154609245eSIlya Dryomov 24164609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 24174609245eSIlya Dryomov WARN_ON(!map->epoch); 24184609245eSIlya Dryomov 24194609245eSIlya Dryomov if (req->r_attempts) { 24204609245eSIlya Dryomov /* 24214609245eSIlya Dryomov * We sent a request earlier, which means that 24224609245eSIlya Dryomov * previously the pool existed, and now it does not 24234609245eSIlya Dryomov * (i.e., it was deleted). 24244609245eSIlya Dryomov */ 24254609245eSIlya Dryomov req->r_map_dne_bound = map->epoch; 24264609245eSIlya Dryomov dout("%s req %p tid %llu pool disappeared\n", __func__, req, 24274609245eSIlya Dryomov req->r_tid); 24284609245eSIlya Dryomov } else { 24294609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u have %u\n", __func__, 24304609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, map->epoch); 24314609245eSIlya Dryomov } 24324609245eSIlya Dryomov 24334609245eSIlya Dryomov if (req->r_map_dne_bound) { 24344609245eSIlya Dryomov if (map->epoch >= req->r_map_dne_bound) { 24354609245eSIlya Dryomov /* we had a new enough map */ 24364609245eSIlya Dryomov pr_info_ratelimited("tid %llu pool does not exist\n", 24374609245eSIlya Dryomov req->r_tid); 24384609245eSIlya Dryomov complete_request(req, -ENOENT); 24394609245eSIlya Dryomov } 24404609245eSIlya Dryomov } else { 24414609245eSIlya Dryomov send_map_check(req); 24424609245eSIlya Dryomov } 24434609245eSIlya Dryomov } 24444609245eSIlya Dryomov 24454609245eSIlya Dryomov static void map_check_cb(struct ceph_mon_generic_request *greq) 24464609245eSIlya Dryomov { 24474609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 24484609245eSIlya Dryomov struct ceph_osd_request *req; 24494609245eSIlya Dryomov u64 tid = greq->private_data; 24504609245eSIlya Dryomov 24514609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 24524609245eSIlya Dryomov 24534609245eSIlya Dryomov down_write(&osdc->lock); 24544609245eSIlya Dryomov req = lookup_request_mc(&osdc->map_checks, tid); 24554609245eSIlya Dryomov if (!req) { 24564609245eSIlya Dryomov dout("%s tid %llu dne\n", __func__, tid); 24574609245eSIlya Dryomov goto out_unlock; 24584609245eSIlya Dryomov } 24594609245eSIlya Dryomov 24604609245eSIlya Dryomov dout("%s req %p tid %llu map_dne_bound %u newest %llu\n", __func__, 24614609245eSIlya Dryomov req, req->r_tid, req->r_map_dne_bound, greq->u.newest); 24624609245eSIlya Dryomov if (!req->r_map_dne_bound) 24634609245eSIlya Dryomov req->r_map_dne_bound = greq->u.newest; 24644609245eSIlya Dryomov erase_request_mc(&osdc->map_checks, req); 24654609245eSIlya Dryomov check_pool_dne(req); 24664609245eSIlya Dryomov 24674609245eSIlya Dryomov ceph_osdc_put_request(req); 24684609245eSIlya Dryomov out_unlock: 24694609245eSIlya Dryomov up_write(&osdc->lock); 24704609245eSIlya Dryomov } 24714609245eSIlya Dryomov 24724609245eSIlya Dryomov static void send_map_check(struct ceph_osd_request *req) 24734609245eSIlya Dryomov { 24744609245eSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 24754609245eSIlya Dryomov struct ceph_osd_request *lookup_req; 24764609245eSIlya Dryomov int ret; 24774609245eSIlya Dryomov 24784609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 24794609245eSIlya Dryomov 24804609245eSIlya Dryomov lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid); 24814609245eSIlya Dryomov if (lookup_req) { 24824609245eSIlya Dryomov WARN_ON(lookup_req != req); 24834609245eSIlya Dryomov return; 24844609245eSIlya Dryomov } 24854609245eSIlya Dryomov 24864609245eSIlya Dryomov ceph_osdc_get_request(req); 24874609245eSIlya Dryomov insert_request_mc(&osdc->map_checks, req); 24884609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 24894609245eSIlya Dryomov map_check_cb, req->r_tid); 24904609245eSIlya Dryomov WARN_ON(ret); 24914609245eSIlya Dryomov } 24924609245eSIlya Dryomov 24930bbfdfe8SIlya Dryomov /* 2494922dab61SIlya Dryomov * lingering requests, watch/notify v2 infrastructure 2495922dab61SIlya Dryomov */ 2496922dab61SIlya Dryomov static void linger_release(struct kref *kref) 2497922dab61SIlya Dryomov { 2498922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 2499922dab61SIlya Dryomov container_of(kref, struct ceph_osd_linger_request, kref); 2500922dab61SIlya Dryomov 2501922dab61SIlya Dryomov dout("%s lreq %p reg_req %p ping_req %p\n", __func__, lreq, 2502922dab61SIlya Dryomov lreq->reg_req, lreq->ping_req); 2503922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->node)); 2504922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->osdc_node)); 25054609245eSIlya Dryomov WARN_ON(!RB_EMPTY_NODE(&lreq->mc_node)); 2506922dab61SIlya Dryomov WARN_ON(!list_empty(&lreq->scan_item)); 2507b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lreq->pending_lworks)); 2508922dab61SIlya Dryomov WARN_ON(lreq->osd); 2509922dab61SIlya Dryomov 2510922dab61SIlya Dryomov if (lreq->reg_req) 2511922dab61SIlya Dryomov ceph_osdc_put_request(lreq->reg_req); 2512922dab61SIlya Dryomov if (lreq->ping_req) 2513922dab61SIlya Dryomov ceph_osdc_put_request(lreq->ping_req); 2514922dab61SIlya Dryomov target_destroy(&lreq->t); 2515922dab61SIlya Dryomov kfree(lreq); 2516922dab61SIlya Dryomov } 2517922dab61SIlya Dryomov 2518922dab61SIlya Dryomov static void linger_put(struct ceph_osd_linger_request *lreq) 2519922dab61SIlya Dryomov { 2520922dab61SIlya Dryomov if (lreq) 2521922dab61SIlya Dryomov kref_put(&lreq->kref, linger_release); 2522922dab61SIlya Dryomov } 2523922dab61SIlya Dryomov 2524922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2525922dab61SIlya Dryomov linger_get(struct ceph_osd_linger_request *lreq) 2526922dab61SIlya Dryomov { 2527922dab61SIlya Dryomov kref_get(&lreq->kref); 2528922dab61SIlya Dryomov return lreq; 2529922dab61SIlya Dryomov } 2530922dab61SIlya Dryomov 2531922dab61SIlya Dryomov static struct ceph_osd_linger_request * 2532922dab61SIlya Dryomov linger_alloc(struct ceph_osd_client *osdc) 2533922dab61SIlya Dryomov { 2534922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2535922dab61SIlya Dryomov 2536922dab61SIlya Dryomov lreq = kzalloc(sizeof(*lreq), GFP_NOIO); 2537922dab61SIlya Dryomov if (!lreq) 2538922dab61SIlya Dryomov return NULL; 2539922dab61SIlya Dryomov 2540922dab61SIlya Dryomov kref_init(&lreq->kref); 2541922dab61SIlya Dryomov mutex_init(&lreq->lock); 2542922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->node); 2543922dab61SIlya Dryomov RB_CLEAR_NODE(&lreq->osdc_node); 25444609245eSIlya Dryomov RB_CLEAR_NODE(&lreq->mc_node); 2545922dab61SIlya Dryomov INIT_LIST_HEAD(&lreq->scan_item); 2546b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lreq->pending_lworks); 2547922dab61SIlya Dryomov init_completion(&lreq->reg_commit_wait); 254819079203SIlya Dryomov init_completion(&lreq->notify_finish_wait); 2549922dab61SIlya Dryomov 2550922dab61SIlya Dryomov lreq->osdc = osdc; 2551922dab61SIlya Dryomov target_init(&lreq->t); 2552922dab61SIlya Dryomov 2553922dab61SIlya Dryomov dout("%s lreq %p\n", __func__, lreq); 2554922dab61SIlya Dryomov return lreq; 2555922dab61SIlya Dryomov } 2556922dab61SIlya Dryomov 2557922dab61SIlya Dryomov DEFINE_RB_INSDEL_FUNCS(linger, struct ceph_osd_linger_request, linger_id, node) 2558922dab61SIlya Dryomov DEFINE_RB_FUNCS(linger_osdc, struct ceph_osd_linger_request, linger_id, osdc_node) 25594609245eSIlya Dryomov DEFINE_RB_FUNCS(linger_mc, struct ceph_osd_linger_request, linger_id, mc_node) 2560922dab61SIlya Dryomov 2561922dab61SIlya Dryomov /* 2562922dab61SIlya Dryomov * Create linger request <-> OSD session relation. 2563922dab61SIlya Dryomov * 2564922dab61SIlya Dryomov * @lreq has to be registered, @osd may be homeless. 2565922dab61SIlya Dryomov */ 2566922dab61SIlya Dryomov static void link_linger(struct ceph_osd *osd, 2567922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2568922dab61SIlya Dryomov { 2569922dab61SIlya Dryomov verify_osd_locked(osd); 2570922dab61SIlya Dryomov WARN_ON(!lreq->linger_id || lreq->osd); 2571922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2572922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2573922dab61SIlya Dryomov 2574922dab61SIlya Dryomov if (!osd_homeless(osd)) 2575922dab61SIlya Dryomov __remove_osd_from_lru(osd); 2576922dab61SIlya Dryomov else 2577922dab61SIlya Dryomov atomic_inc(&osd->o_osdc->num_homeless); 2578922dab61SIlya Dryomov 2579922dab61SIlya Dryomov get_osd(osd); 2580922dab61SIlya Dryomov insert_linger(&osd->o_linger_requests, lreq); 2581922dab61SIlya Dryomov lreq->osd = osd; 2582922dab61SIlya Dryomov } 2583922dab61SIlya Dryomov 2584922dab61SIlya Dryomov static void unlink_linger(struct ceph_osd *osd, 2585922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 2586922dab61SIlya Dryomov { 2587922dab61SIlya Dryomov verify_osd_locked(osd); 2588922dab61SIlya Dryomov WARN_ON(lreq->osd != osd); 2589922dab61SIlya Dryomov dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd, 2590922dab61SIlya Dryomov osd->o_osd, lreq, lreq->linger_id); 2591922dab61SIlya Dryomov 2592922dab61SIlya Dryomov lreq->osd = NULL; 2593922dab61SIlya Dryomov erase_linger(&osd->o_linger_requests, lreq); 2594922dab61SIlya Dryomov put_osd(osd); 2595922dab61SIlya Dryomov 2596922dab61SIlya Dryomov if (!osd_homeless(osd)) 2597922dab61SIlya Dryomov maybe_move_osd_to_lru(osd); 2598922dab61SIlya Dryomov else 2599922dab61SIlya Dryomov atomic_dec(&osd->o_osdc->num_homeless); 2600922dab61SIlya Dryomov } 2601922dab61SIlya Dryomov 2602922dab61SIlya Dryomov static bool __linger_registered(struct ceph_osd_linger_request *lreq) 2603922dab61SIlya Dryomov { 2604922dab61SIlya Dryomov verify_osdc_locked(lreq->osdc); 2605922dab61SIlya Dryomov 2606922dab61SIlya Dryomov return !RB_EMPTY_NODE(&lreq->osdc_node); 2607922dab61SIlya Dryomov } 2608922dab61SIlya Dryomov 2609922dab61SIlya Dryomov static bool linger_registered(struct ceph_osd_linger_request *lreq) 2610922dab61SIlya Dryomov { 2611922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2612922dab61SIlya Dryomov bool registered; 2613922dab61SIlya Dryomov 2614922dab61SIlya Dryomov down_read(&osdc->lock); 2615922dab61SIlya Dryomov registered = __linger_registered(lreq); 2616922dab61SIlya Dryomov up_read(&osdc->lock); 2617922dab61SIlya Dryomov 2618922dab61SIlya Dryomov return registered; 2619922dab61SIlya Dryomov } 2620922dab61SIlya Dryomov 2621922dab61SIlya Dryomov static void linger_register(struct ceph_osd_linger_request *lreq) 2622922dab61SIlya Dryomov { 2623922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2624922dab61SIlya Dryomov 2625922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2626922dab61SIlya Dryomov WARN_ON(lreq->linger_id); 2627922dab61SIlya Dryomov 2628922dab61SIlya Dryomov linger_get(lreq); 2629922dab61SIlya Dryomov lreq->linger_id = ++osdc->last_linger_id; 2630922dab61SIlya Dryomov insert_linger_osdc(&osdc->linger_requests, lreq); 2631922dab61SIlya Dryomov } 2632922dab61SIlya Dryomov 2633922dab61SIlya Dryomov static void linger_unregister(struct ceph_osd_linger_request *lreq) 2634922dab61SIlya Dryomov { 2635922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2636922dab61SIlya Dryomov 2637922dab61SIlya Dryomov verify_osdc_wrlocked(osdc); 2638922dab61SIlya Dryomov 2639922dab61SIlya Dryomov erase_linger_osdc(&osdc->linger_requests, lreq); 2640922dab61SIlya Dryomov linger_put(lreq); 2641922dab61SIlya Dryomov } 2642922dab61SIlya Dryomov 2643922dab61SIlya Dryomov static void cancel_linger_request(struct ceph_osd_request *req) 2644922dab61SIlya Dryomov { 2645922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2646922dab61SIlya Dryomov 2647922dab61SIlya Dryomov WARN_ON(!req->r_linger); 2648922dab61SIlya Dryomov cancel_request(req); 2649922dab61SIlya Dryomov linger_put(lreq); 2650922dab61SIlya Dryomov } 2651922dab61SIlya Dryomov 2652922dab61SIlya Dryomov struct linger_work { 2653922dab61SIlya Dryomov struct work_struct work; 2654922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 2655b07d3c4bSIlya Dryomov struct list_head pending_item; 2656b07d3c4bSIlya Dryomov unsigned long queued_stamp; 2657922dab61SIlya Dryomov 2658922dab61SIlya Dryomov union { 2659922dab61SIlya Dryomov struct { 2660922dab61SIlya Dryomov u64 notify_id; 2661922dab61SIlya Dryomov u64 notifier_id; 2662922dab61SIlya Dryomov void *payload; /* points into @msg front */ 2663922dab61SIlya Dryomov size_t payload_len; 2664922dab61SIlya Dryomov 2665922dab61SIlya Dryomov struct ceph_msg *msg; /* for ceph_msg_put() */ 2666922dab61SIlya Dryomov } notify; 2667922dab61SIlya Dryomov struct { 2668922dab61SIlya Dryomov int err; 2669922dab61SIlya Dryomov } error; 2670922dab61SIlya Dryomov }; 2671922dab61SIlya Dryomov }; 2672922dab61SIlya Dryomov 2673922dab61SIlya Dryomov static struct linger_work *lwork_alloc(struct ceph_osd_linger_request *lreq, 2674922dab61SIlya Dryomov work_func_t workfn) 2675922dab61SIlya Dryomov { 2676922dab61SIlya Dryomov struct linger_work *lwork; 2677922dab61SIlya Dryomov 2678922dab61SIlya Dryomov lwork = kzalloc(sizeof(*lwork), GFP_NOIO); 2679922dab61SIlya Dryomov if (!lwork) 2680922dab61SIlya Dryomov return NULL; 2681922dab61SIlya Dryomov 2682922dab61SIlya Dryomov INIT_WORK(&lwork->work, workfn); 2683b07d3c4bSIlya Dryomov INIT_LIST_HEAD(&lwork->pending_item); 2684922dab61SIlya Dryomov lwork->lreq = linger_get(lreq); 2685922dab61SIlya Dryomov 2686922dab61SIlya Dryomov return lwork; 2687922dab61SIlya Dryomov } 2688922dab61SIlya Dryomov 2689922dab61SIlya Dryomov static void lwork_free(struct linger_work *lwork) 2690922dab61SIlya Dryomov { 2691922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2692922dab61SIlya Dryomov 2693b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 2694b07d3c4bSIlya Dryomov list_del(&lwork->pending_item); 2695b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 2696b07d3c4bSIlya Dryomov 2697922dab61SIlya Dryomov linger_put(lreq); 2698922dab61SIlya Dryomov kfree(lwork); 2699922dab61SIlya Dryomov } 2700922dab61SIlya Dryomov 2701922dab61SIlya Dryomov static void lwork_queue(struct linger_work *lwork) 2702922dab61SIlya Dryomov { 2703922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2704922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2705922dab61SIlya Dryomov 2706922dab61SIlya Dryomov verify_lreq_locked(lreq); 2707b07d3c4bSIlya Dryomov WARN_ON(!list_empty(&lwork->pending_item)); 2708b07d3c4bSIlya Dryomov 2709b07d3c4bSIlya Dryomov lwork->queued_stamp = jiffies; 2710b07d3c4bSIlya Dryomov list_add_tail(&lwork->pending_item, &lreq->pending_lworks); 2711922dab61SIlya Dryomov queue_work(osdc->notify_wq, &lwork->work); 2712922dab61SIlya Dryomov } 2713922dab61SIlya Dryomov 2714922dab61SIlya Dryomov static void do_watch_notify(struct work_struct *w) 2715922dab61SIlya Dryomov { 2716922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2717922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2718922dab61SIlya Dryomov 2719922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2720922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2721922dab61SIlya Dryomov goto out; 2722922dab61SIlya Dryomov } 2723922dab61SIlya Dryomov 272419079203SIlya Dryomov WARN_ON(!lreq->is_watch); 2725922dab61SIlya Dryomov dout("%s lreq %p notify_id %llu notifier_id %llu payload_len %zu\n", 2726922dab61SIlya Dryomov __func__, lreq, lwork->notify.notify_id, lwork->notify.notifier_id, 2727922dab61SIlya Dryomov lwork->notify.payload_len); 2728922dab61SIlya Dryomov lreq->wcb(lreq->data, lwork->notify.notify_id, lreq->linger_id, 2729922dab61SIlya Dryomov lwork->notify.notifier_id, lwork->notify.payload, 2730922dab61SIlya Dryomov lwork->notify.payload_len); 2731922dab61SIlya Dryomov 2732922dab61SIlya Dryomov out: 2733922dab61SIlya Dryomov ceph_msg_put(lwork->notify.msg); 2734922dab61SIlya Dryomov lwork_free(lwork); 2735922dab61SIlya Dryomov } 2736922dab61SIlya Dryomov 2737922dab61SIlya Dryomov static void do_watch_error(struct work_struct *w) 2738922dab61SIlya Dryomov { 2739922dab61SIlya Dryomov struct linger_work *lwork = container_of(w, struct linger_work, work); 2740922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = lwork->lreq; 2741922dab61SIlya Dryomov 2742922dab61SIlya Dryomov if (!linger_registered(lreq)) { 2743922dab61SIlya Dryomov dout("%s lreq %p not registered\n", __func__, lreq); 2744922dab61SIlya Dryomov goto out; 2745922dab61SIlya Dryomov } 2746922dab61SIlya Dryomov 2747922dab61SIlya Dryomov dout("%s lreq %p err %d\n", __func__, lreq, lwork->error.err); 2748922dab61SIlya Dryomov lreq->errcb(lreq->data, lreq->linger_id, lwork->error.err); 2749922dab61SIlya Dryomov 2750922dab61SIlya Dryomov out: 2751922dab61SIlya Dryomov lwork_free(lwork); 2752922dab61SIlya Dryomov } 2753922dab61SIlya Dryomov 2754922dab61SIlya Dryomov static void queue_watch_error(struct ceph_osd_linger_request *lreq) 2755922dab61SIlya Dryomov { 2756922dab61SIlya Dryomov struct linger_work *lwork; 2757922dab61SIlya Dryomov 2758922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_error); 2759922dab61SIlya Dryomov if (!lwork) { 2760922dab61SIlya Dryomov pr_err("failed to allocate error-lwork\n"); 2761922dab61SIlya Dryomov return; 2762922dab61SIlya Dryomov } 2763922dab61SIlya Dryomov 2764922dab61SIlya Dryomov lwork->error.err = lreq->last_error; 2765922dab61SIlya Dryomov lwork_queue(lwork); 2766922dab61SIlya Dryomov } 2767922dab61SIlya Dryomov 2768922dab61SIlya Dryomov static void linger_reg_commit_complete(struct ceph_osd_linger_request *lreq, 2769922dab61SIlya Dryomov int result) 2770922dab61SIlya Dryomov { 2771922dab61SIlya Dryomov if (!completion_done(&lreq->reg_commit_wait)) { 2772922dab61SIlya Dryomov lreq->reg_commit_error = (result <= 0 ? result : 0); 2773922dab61SIlya Dryomov complete_all(&lreq->reg_commit_wait); 2774922dab61SIlya Dryomov } 2775922dab61SIlya Dryomov } 2776922dab61SIlya Dryomov 2777922dab61SIlya Dryomov static void linger_commit_cb(struct ceph_osd_request *req) 2778922dab61SIlya Dryomov { 2779922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2780922dab61SIlya Dryomov 2781922dab61SIlya Dryomov mutex_lock(&lreq->lock); 2782922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d\n", __func__, lreq, 2783922dab61SIlya Dryomov lreq->linger_id, req->r_result); 2784922dab61SIlya Dryomov linger_reg_commit_complete(lreq, req->r_result); 2785922dab61SIlya Dryomov lreq->committed = true; 2786922dab61SIlya Dryomov 278719079203SIlya Dryomov if (!lreq->is_watch) { 278819079203SIlya Dryomov struct ceph_osd_data *osd_data = 278919079203SIlya Dryomov osd_req_op_data(req, 0, notify, response_data); 279019079203SIlya Dryomov void *p = page_address(osd_data->pages[0]); 279119079203SIlya Dryomov 279219079203SIlya Dryomov WARN_ON(req->r_ops[0].op != CEPH_OSD_OP_NOTIFY || 279319079203SIlya Dryomov osd_data->type != CEPH_OSD_DATA_TYPE_PAGES); 279419079203SIlya Dryomov 279519079203SIlya Dryomov /* make note of the notify_id */ 279619079203SIlya Dryomov if (req->r_ops[0].outdata_len >= sizeof(u64)) { 279719079203SIlya Dryomov lreq->notify_id = ceph_decode_64(&p); 279819079203SIlya Dryomov dout("lreq %p notify_id %llu\n", lreq, 279919079203SIlya Dryomov lreq->notify_id); 280019079203SIlya Dryomov } else { 280119079203SIlya Dryomov dout("lreq %p no notify_id\n", lreq); 280219079203SIlya Dryomov } 280319079203SIlya Dryomov } 280419079203SIlya Dryomov 2805922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2806922dab61SIlya Dryomov linger_put(lreq); 2807922dab61SIlya Dryomov } 2808922dab61SIlya Dryomov 2809922dab61SIlya Dryomov static int normalize_watch_error(int err) 2810922dab61SIlya Dryomov { 2811922dab61SIlya Dryomov /* 2812922dab61SIlya Dryomov * Translate ENOENT -> ENOTCONN so that a delete->disconnection 2813922dab61SIlya Dryomov * notification and a failure to reconnect because we raced with 2814922dab61SIlya Dryomov * the delete appear the same to the user. 2815922dab61SIlya Dryomov */ 2816922dab61SIlya Dryomov if (err == -ENOENT) 2817922dab61SIlya Dryomov err = -ENOTCONN; 2818922dab61SIlya Dryomov 2819922dab61SIlya Dryomov return err; 2820922dab61SIlya Dryomov } 2821922dab61SIlya Dryomov 2822922dab61SIlya Dryomov static void linger_reconnect_cb(struct ceph_osd_request *req) 2823922dab61SIlya Dryomov { 2824922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2825922dab61SIlya Dryomov 2826922dab61SIlya Dryomov mutex_lock(&lreq->lock); 2827922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d last_error %d\n", __func__, 2828922dab61SIlya Dryomov lreq, lreq->linger_id, req->r_result, lreq->last_error); 2829922dab61SIlya Dryomov if (req->r_result < 0) { 2830922dab61SIlya Dryomov if (!lreq->last_error) { 2831922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 2832922dab61SIlya Dryomov queue_watch_error(lreq); 2833922dab61SIlya Dryomov } 2834922dab61SIlya Dryomov } 2835922dab61SIlya Dryomov 2836922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2837922dab61SIlya Dryomov linger_put(lreq); 2838922dab61SIlya Dryomov } 2839922dab61SIlya Dryomov 2840922dab61SIlya Dryomov static void send_linger(struct ceph_osd_linger_request *lreq) 2841922dab61SIlya Dryomov { 2842922dab61SIlya Dryomov struct ceph_osd_request *req = lreq->reg_req; 2843922dab61SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[0]; 2844922dab61SIlya Dryomov 2845922dab61SIlya Dryomov verify_osdc_wrlocked(req->r_osdc); 2846922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 2847922dab61SIlya Dryomov 2848922dab61SIlya Dryomov if (req->r_osd) 2849922dab61SIlya Dryomov cancel_linger_request(req); 2850922dab61SIlya Dryomov 2851922dab61SIlya Dryomov request_reinit(req); 2852922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 2853922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 2854922dab61SIlya Dryomov req->r_flags = lreq->t.flags; 2855922dab61SIlya Dryomov req->r_mtime = lreq->mtime; 2856922dab61SIlya Dryomov 2857922dab61SIlya Dryomov mutex_lock(&lreq->lock); 285819079203SIlya Dryomov if (lreq->is_watch && lreq->committed) { 2859922dab61SIlya Dryomov WARN_ON(op->op != CEPH_OSD_OP_WATCH || 2860922dab61SIlya Dryomov op->watch.cookie != lreq->linger_id); 2861922dab61SIlya Dryomov op->watch.op = CEPH_OSD_WATCH_OP_RECONNECT; 2862922dab61SIlya Dryomov op->watch.gen = ++lreq->register_gen; 2863922dab61SIlya Dryomov dout("lreq %p reconnect register_gen %u\n", lreq, 2864922dab61SIlya Dryomov op->watch.gen); 2865922dab61SIlya Dryomov req->r_callback = linger_reconnect_cb; 2866922dab61SIlya Dryomov } else { 286719079203SIlya Dryomov if (!lreq->is_watch) 286819079203SIlya Dryomov lreq->notify_id = 0; 286919079203SIlya Dryomov else 2870922dab61SIlya Dryomov WARN_ON(op->watch.op != CEPH_OSD_WATCH_OP_WATCH); 2871922dab61SIlya Dryomov dout("lreq %p register\n", lreq); 2872922dab61SIlya Dryomov req->r_callback = linger_commit_cb; 2873922dab61SIlya Dryomov } 2874922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2875922dab61SIlya Dryomov 2876922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 2877922dab61SIlya Dryomov req->r_linger = true; 2878922dab61SIlya Dryomov 2879922dab61SIlya Dryomov submit_request(req, true); 2880922dab61SIlya Dryomov } 2881922dab61SIlya Dryomov 2882922dab61SIlya Dryomov static void linger_ping_cb(struct ceph_osd_request *req) 2883922dab61SIlya Dryomov { 2884922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = req->r_priv; 2885922dab61SIlya Dryomov 2886922dab61SIlya Dryomov mutex_lock(&lreq->lock); 2887922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu result %d ping_sent %lu last_error %d\n", 2888922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, req->r_result, lreq->ping_sent, 2889922dab61SIlya Dryomov lreq->last_error); 2890922dab61SIlya Dryomov if (lreq->register_gen == req->r_ops[0].watch.gen) { 2891b07d3c4bSIlya Dryomov if (!req->r_result) { 2892b07d3c4bSIlya Dryomov lreq->watch_valid_thru = lreq->ping_sent; 2893b07d3c4bSIlya Dryomov } else if (!lreq->last_error) { 2894922dab61SIlya Dryomov lreq->last_error = normalize_watch_error(req->r_result); 2895922dab61SIlya Dryomov queue_watch_error(lreq); 2896922dab61SIlya Dryomov } 2897922dab61SIlya Dryomov } else { 2898922dab61SIlya Dryomov dout("lreq %p register_gen %u ignoring old pong %u\n", lreq, 2899922dab61SIlya Dryomov lreq->register_gen, req->r_ops[0].watch.gen); 2900922dab61SIlya Dryomov } 2901922dab61SIlya Dryomov 2902922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 2903922dab61SIlya Dryomov linger_put(lreq); 2904922dab61SIlya Dryomov } 2905922dab61SIlya Dryomov 2906922dab61SIlya Dryomov static void send_linger_ping(struct ceph_osd_linger_request *lreq) 2907922dab61SIlya Dryomov { 2908922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2909922dab61SIlya Dryomov struct ceph_osd_request *req = lreq->ping_req; 2910922dab61SIlya Dryomov struct ceph_osd_req_op *op = &req->r_ops[0]; 2911922dab61SIlya Dryomov 2912b7ec35b3SIlya Dryomov if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) { 2913922dab61SIlya Dryomov dout("%s PAUSERD\n", __func__); 2914922dab61SIlya Dryomov return; 2915922dab61SIlya Dryomov } 2916922dab61SIlya Dryomov 2917922dab61SIlya Dryomov lreq->ping_sent = jiffies; 2918922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu ping_sent %lu register_gen %u\n", 2919922dab61SIlya Dryomov __func__, lreq, lreq->linger_id, lreq->ping_sent, 2920922dab61SIlya Dryomov lreq->register_gen); 2921922dab61SIlya Dryomov 2922922dab61SIlya Dryomov if (req->r_osd) 2923922dab61SIlya Dryomov cancel_linger_request(req); 2924922dab61SIlya Dryomov 2925922dab61SIlya Dryomov request_reinit(req); 2926922dab61SIlya Dryomov target_copy(&req->r_t, &lreq->t); 2927922dab61SIlya Dryomov 2928922dab61SIlya Dryomov WARN_ON(op->op != CEPH_OSD_OP_WATCH || 2929922dab61SIlya Dryomov op->watch.cookie != lreq->linger_id || 2930922dab61SIlya Dryomov op->watch.op != CEPH_OSD_WATCH_OP_PING); 2931922dab61SIlya Dryomov op->watch.gen = lreq->register_gen; 2932922dab61SIlya Dryomov req->r_callback = linger_ping_cb; 2933922dab61SIlya Dryomov req->r_priv = linger_get(lreq); 2934922dab61SIlya Dryomov req->r_linger = true; 2935922dab61SIlya Dryomov 2936922dab61SIlya Dryomov ceph_osdc_get_request(req); 2937922dab61SIlya Dryomov account_request(req); 2938922dab61SIlya Dryomov req->r_tid = atomic64_inc_return(&osdc->last_tid); 2939922dab61SIlya Dryomov link_request(lreq->osd, req); 2940922dab61SIlya Dryomov send_request(req); 2941922dab61SIlya Dryomov } 2942922dab61SIlya Dryomov 2943922dab61SIlya Dryomov static void linger_submit(struct ceph_osd_linger_request *lreq) 2944922dab61SIlya Dryomov { 2945922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2946922dab61SIlya Dryomov struct ceph_osd *osd; 2947922dab61SIlya Dryomov 29487de030d6SIlya Dryomov calc_target(osdc, &lreq->t, NULL, false); 2949922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 2950922dab61SIlya Dryomov link_linger(osd, lreq); 2951922dab61SIlya Dryomov 2952922dab61SIlya Dryomov send_linger(lreq); 2953922dab61SIlya Dryomov } 2954922dab61SIlya Dryomov 29554609245eSIlya Dryomov static void cancel_linger_map_check(struct ceph_osd_linger_request *lreq) 29564609245eSIlya Dryomov { 29574609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 29584609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 29594609245eSIlya Dryomov 29604609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 29614609245eSIlya Dryomov 29624609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 29634609245eSIlya Dryomov lreq->linger_id); 29644609245eSIlya Dryomov if (!lookup_lreq) 29654609245eSIlya Dryomov return; 29664609245eSIlya Dryomov 29674609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 29684609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 29694609245eSIlya Dryomov linger_put(lreq); 29704609245eSIlya Dryomov } 29714609245eSIlya Dryomov 2972922dab61SIlya Dryomov /* 2973922dab61SIlya Dryomov * @lreq has to be both registered and linked. 2974922dab61SIlya Dryomov */ 2975922dab61SIlya Dryomov static void __linger_cancel(struct ceph_osd_linger_request *lreq) 2976922dab61SIlya Dryomov { 297719079203SIlya Dryomov if (lreq->is_watch && lreq->ping_req->r_osd) 2978922dab61SIlya Dryomov cancel_linger_request(lreq->ping_req); 2979922dab61SIlya Dryomov if (lreq->reg_req->r_osd) 2980922dab61SIlya Dryomov cancel_linger_request(lreq->reg_req); 29814609245eSIlya Dryomov cancel_linger_map_check(lreq); 2982922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 2983922dab61SIlya Dryomov linger_unregister(lreq); 2984922dab61SIlya Dryomov } 2985922dab61SIlya Dryomov 2986922dab61SIlya Dryomov static void linger_cancel(struct ceph_osd_linger_request *lreq) 2987922dab61SIlya Dryomov { 2988922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 2989922dab61SIlya Dryomov 2990922dab61SIlya Dryomov down_write(&osdc->lock); 2991922dab61SIlya Dryomov if (__linger_registered(lreq)) 2992922dab61SIlya Dryomov __linger_cancel(lreq); 2993922dab61SIlya Dryomov up_write(&osdc->lock); 2994922dab61SIlya Dryomov } 2995922dab61SIlya Dryomov 29964609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq); 29974609245eSIlya Dryomov 29984609245eSIlya Dryomov static void check_linger_pool_dne(struct ceph_osd_linger_request *lreq) 29994609245eSIlya Dryomov { 30004609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 30014609245eSIlya Dryomov struct ceph_osdmap *map = osdc->osdmap; 30024609245eSIlya Dryomov 30034609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 30044609245eSIlya Dryomov WARN_ON(!map->epoch); 30054609245eSIlya Dryomov 30064609245eSIlya Dryomov if (lreq->register_gen) { 30074609245eSIlya Dryomov lreq->map_dne_bound = map->epoch; 30084609245eSIlya Dryomov dout("%s lreq %p linger_id %llu pool disappeared\n", __func__, 30094609245eSIlya Dryomov lreq, lreq->linger_id); 30104609245eSIlya Dryomov } else { 30114609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u have %u\n", 30124609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 30134609245eSIlya Dryomov map->epoch); 30144609245eSIlya Dryomov } 30154609245eSIlya Dryomov 30164609245eSIlya Dryomov if (lreq->map_dne_bound) { 30174609245eSIlya Dryomov if (map->epoch >= lreq->map_dne_bound) { 30184609245eSIlya Dryomov /* we had a new enough map */ 30194609245eSIlya Dryomov pr_info("linger_id %llu pool does not exist\n", 30204609245eSIlya Dryomov lreq->linger_id); 30214609245eSIlya Dryomov linger_reg_commit_complete(lreq, -ENOENT); 30224609245eSIlya Dryomov __linger_cancel(lreq); 30234609245eSIlya Dryomov } 30244609245eSIlya Dryomov } else { 30254609245eSIlya Dryomov send_linger_map_check(lreq); 30264609245eSIlya Dryomov } 30274609245eSIlya Dryomov } 30284609245eSIlya Dryomov 30294609245eSIlya Dryomov static void linger_map_check_cb(struct ceph_mon_generic_request *greq) 30304609245eSIlya Dryomov { 30314609245eSIlya Dryomov struct ceph_osd_client *osdc = &greq->monc->client->osdc; 30324609245eSIlya Dryomov struct ceph_osd_linger_request *lreq; 30334609245eSIlya Dryomov u64 linger_id = greq->private_data; 30344609245eSIlya Dryomov 30354609245eSIlya Dryomov WARN_ON(greq->result || !greq->u.newest); 30364609245eSIlya Dryomov 30374609245eSIlya Dryomov down_write(&osdc->lock); 30384609245eSIlya Dryomov lreq = lookup_linger_mc(&osdc->linger_map_checks, linger_id); 30394609245eSIlya Dryomov if (!lreq) { 30404609245eSIlya Dryomov dout("%s linger_id %llu dne\n", __func__, linger_id); 30414609245eSIlya Dryomov goto out_unlock; 30424609245eSIlya Dryomov } 30434609245eSIlya Dryomov 30444609245eSIlya Dryomov dout("%s lreq %p linger_id %llu map_dne_bound %u newest %llu\n", 30454609245eSIlya Dryomov __func__, lreq, lreq->linger_id, lreq->map_dne_bound, 30464609245eSIlya Dryomov greq->u.newest); 30474609245eSIlya Dryomov if (!lreq->map_dne_bound) 30484609245eSIlya Dryomov lreq->map_dne_bound = greq->u.newest; 30494609245eSIlya Dryomov erase_linger_mc(&osdc->linger_map_checks, lreq); 30504609245eSIlya Dryomov check_linger_pool_dne(lreq); 30514609245eSIlya Dryomov 30524609245eSIlya Dryomov linger_put(lreq); 30534609245eSIlya Dryomov out_unlock: 30544609245eSIlya Dryomov up_write(&osdc->lock); 30554609245eSIlya Dryomov } 30564609245eSIlya Dryomov 30574609245eSIlya Dryomov static void send_linger_map_check(struct ceph_osd_linger_request *lreq) 30584609245eSIlya Dryomov { 30594609245eSIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 30604609245eSIlya Dryomov struct ceph_osd_linger_request *lookup_lreq; 30614609245eSIlya Dryomov int ret; 30624609245eSIlya Dryomov 30634609245eSIlya Dryomov verify_osdc_wrlocked(osdc); 30644609245eSIlya Dryomov 30654609245eSIlya Dryomov lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks, 30664609245eSIlya Dryomov lreq->linger_id); 30674609245eSIlya Dryomov if (lookup_lreq) { 30684609245eSIlya Dryomov WARN_ON(lookup_lreq != lreq); 30694609245eSIlya Dryomov return; 30704609245eSIlya Dryomov } 30714609245eSIlya Dryomov 30724609245eSIlya Dryomov linger_get(lreq); 30734609245eSIlya Dryomov insert_linger_mc(&osdc->linger_map_checks, lreq); 30744609245eSIlya Dryomov ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap", 30754609245eSIlya Dryomov linger_map_check_cb, lreq->linger_id); 30764609245eSIlya Dryomov WARN_ON(ret); 30774609245eSIlya Dryomov } 30784609245eSIlya Dryomov 3079922dab61SIlya Dryomov static int linger_reg_commit_wait(struct ceph_osd_linger_request *lreq) 3080922dab61SIlya Dryomov { 3081922dab61SIlya Dryomov int ret; 3082922dab61SIlya Dryomov 3083922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 3084922dab61SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->reg_commit_wait); 3085922dab61SIlya Dryomov return ret ?: lreq->reg_commit_error; 3086922dab61SIlya Dryomov } 3087922dab61SIlya Dryomov 308819079203SIlya Dryomov static int linger_notify_finish_wait(struct ceph_osd_linger_request *lreq) 308919079203SIlya Dryomov { 309019079203SIlya Dryomov int ret; 309119079203SIlya Dryomov 309219079203SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id); 309319079203SIlya Dryomov ret = wait_for_completion_interruptible(&lreq->notify_finish_wait); 309419079203SIlya Dryomov return ret ?: lreq->notify_finish_error; 309519079203SIlya Dryomov } 309619079203SIlya Dryomov 3097922dab61SIlya Dryomov /* 3098fbca9635SIlya Dryomov * Timeout callback, called every N seconds. When 1 or more OSD 3099fbca9635SIlya Dryomov * requests has been active for more than N seconds, we send a keepalive 3100fbca9635SIlya Dryomov * (tag + timestamp) to its OSD to ensure any communications channel 3101fbca9635SIlya Dryomov * reset is detected. 31023d14c5d2SYehuda Sadeh */ 31033d14c5d2SYehuda Sadeh static void handle_timeout(struct work_struct *work) 31043d14c5d2SYehuda Sadeh { 31053d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 31063d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, timeout_work.work); 3107a319bf56SIlya Dryomov struct ceph_options *opts = osdc->client->options; 31085aea3dcdSIlya Dryomov unsigned long cutoff = jiffies - opts->osd_keepalive_timeout; 31097cc5e38fSIlya Dryomov unsigned long expiry_cutoff = jiffies - opts->osd_request_timeout; 31105aea3dcdSIlya Dryomov LIST_HEAD(slow_osds); 31115aea3dcdSIlya Dryomov struct rb_node *n, *p; 31123d14c5d2SYehuda Sadeh 31135aea3dcdSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 31145aea3dcdSIlya Dryomov down_write(&osdc->lock); 31153d14c5d2SYehuda Sadeh 31163d14c5d2SYehuda Sadeh /* 31173d14c5d2SYehuda Sadeh * ping osds that are a bit slow. this ensures that if there 31183d14c5d2SYehuda Sadeh * is a break in the TCP connection we will notice, and reopen 31193d14c5d2SYehuda Sadeh * a connection with that osd (from the fault callback). 31203d14c5d2SYehuda Sadeh */ 31215aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 31225aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 31235aea3dcdSIlya Dryomov bool found = false; 31243d14c5d2SYehuda Sadeh 31257cc5e38fSIlya Dryomov for (p = rb_first(&osd->o_requests); p; ) { 31265aea3dcdSIlya Dryomov struct ceph_osd_request *req = 31275aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 31285aea3dcdSIlya Dryomov 31297cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 31307cc5e38fSIlya Dryomov 31315aea3dcdSIlya Dryomov if (time_before(req->r_stamp, cutoff)) { 31325aea3dcdSIlya Dryomov dout(" req %p tid %llu on osd%d is laggy\n", 31335aea3dcdSIlya Dryomov req, req->r_tid, osd->o_osd); 31345aea3dcdSIlya Dryomov found = true; 31355aea3dcdSIlya Dryomov } 31367cc5e38fSIlya Dryomov if (opts->osd_request_timeout && 31377cc5e38fSIlya Dryomov time_before(req->r_start_stamp, expiry_cutoff)) { 31387cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 31397cc5e38fSIlya Dryomov req->r_tid, osd->o_osd); 31407cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 31417cc5e38fSIlya Dryomov } 31425aea3dcdSIlya Dryomov } 3143922dab61SIlya Dryomov for (p = rb_first(&osd->o_linger_requests); p; p = rb_next(p)) { 3144922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3145922dab61SIlya Dryomov rb_entry(p, struct ceph_osd_linger_request, node); 3146922dab61SIlya Dryomov 3147922dab61SIlya Dryomov dout(" lreq %p linger_id %llu is served by osd%d\n", 3148922dab61SIlya Dryomov lreq, lreq->linger_id, osd->o_osd); 3149922dab61SIlya Dryomov found = true; 3150922dab61SIlya Dryomov 3151922dab61SIlya Dryomov mutex_lock(&lreq->lock); 315219079203SIlya Dryomov if (lreq->is_watch && lreq->committed && !lreq->last_error) 3153922dab61SIlya Dryomov send_linger_ping(lreq); 3154922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 3155922dab61SIlya Dryomov } 31565aea3dcdSIlya Dryomov 31575aea3dcdSIlya Dryomov if (found) 31583d14c5d2SYehuda Sadeh list_move_tail(&osd->o_keepalive_item, &slow_osds); 31593d14c5d2SYehuda Sadeh } 31605aea3dcdSIlya Dryomov 31617cc5e38fSIlya Dryomov if (opts->osd_request_timeout) { 31627cc5e38fSIlya Dryomov for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) { 31637cc5e38fSIlya Dryomov struct ceph_osd_request *req = 31647cc5e38fSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 31657cc5e38fSIlya Dryomov 31667cc5e38fSIlya Dryomov p = rb_next(p); /* abort_request() */ 31677cc5e38fSIlya Dryomov 31687cc5e38fSIlya Dryomov if (time_before(req->r_start_stamp, expiry_cutoff)) { 31697cc5e38fSIlya Dryomov pr_err_ratelimited("tid %llu on osd%d timeout\n", 31707cc5e38fSIlya Dryomov req->r_tid, osdc->homeless_osd.o_osd); 31717cc5e38fSIlya Dryomov abort_request(req, -ETIMEDOUT); 31727cc5e38fSIlya Dryomov } 31737cc5e38fSIlya Dryomov } 31747cc5e38fSIlya Dryomov } 31757cc5e38fSIlya Dryomov 31765aea3dcdSIlya Dryomov if (atomic_read(&osdc->num_homeless) || !list_empty(&slow_osds)) 31775aea3dcdSIlya Dryomov maybe_request_map(osdc); 31785aea3dcdSIlya Dryomov 31793d14c5d2SYehuda Sadeh while (!list_empty(&slow_osds)) { 31805aea3dcdSIlya Dryomov struct ceph_osd *osd = list_first_entry(&slow_osds, 31815aea3dcdSIlya Dryomov struct ceph_osd, 31823d14c5d2SYehuda Sadeh o_keepalive_item); 31833d14c5d2SYehuda Sadeh list_del_init(&osd->o_keepalive_item); 31843d14c5d2SYehuda Sadeh ceph_con_keepalive(&osd->o_con); 31853d14c5d2SYehuda Sadeh } 31863d14c5d2SYehuda Sadeh 31875aea3dcdSIlya Dryomov up_write(&osdc->lock); 3188fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 3189fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 31903d14c5d2SYehuda Sadeh } 31913d14c5d2SYehuda Sadeh 31923d14c5d2SYehuda Sadeh static void handle_osds_timeout(struct work_struct *work) 31933d14c5d2SYehuda Sadeh { 31943d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = 31953d14c5d2SYehuda Sadeh container_of(work, struct ceph_osd_client, 31963d14c5d2SYehuda Sadeh osds_timeout_work.work); 3197a319bf56SIlya Dryomov unsigned long delay = osdc->client->options->osd_idle_ttl / 4; 319842a2c09fSIlya Dryomov struct ceph_osd *osd, *nosd; 31993d14c5d2SYehuda Sadeh 320042a2c09fSIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 32015aea3dcdSIlya Dryomov down_write(&osdc->lock); 320242a2c09fSIlya Dryomov list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) { 320342a2c09fSIlya Dryomov if (time_before(jiffies, osd->lru_ttl)) 320442a2c09fSIlya Dryomov break; 320542a2c09fSIlya Dryomov 32065aea3dcdSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests)); 3207922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests)); 32085aea3dcdSIlya Dryomov close_osd(osd); 320942a2c09fSIlya Dryomov } 321042a2c09fSIlya Dryomov 32115aea3dcdSIlya Dryomov up_write(&osdc->lock); 32123d14c5d2SYehuda Sadeh schedule_delayed_work(&osdc->osds_timeout_work, 32133d14c5d2SYehuda Sadeh round_jiffies_relative(delay)); 32143d14c5d2SYehuda Sadeh } 32153d14c5d2SYehuda Sadeh 3216205ee118SIlya Dryomov static int ceph_oloc_decode(void **p, void *end, 3217205ee118SIlya Dryomov struct ceph_object_locator *oloc) 3218205ee118SIlya Dryomov { 3219205ee118SIlya Dryomov u8 struct_v, struct_cv; 3220205ee118SIlya Dryomov u32 len; 3221205ee118SIlya Dryomov void *struct_end; 3222205ee118SIlya Dryomov int ret = 0; 3223205ee118SIlya Dryomov 3224205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3225205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3226205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3227205ee118SIlya Dryomov if (struct_v < 3) { 3228205ee118SIlya Dryomov pr_warn("got v %d < 3 cv %d of ceph_object_locator\n", 3229205ee118SIlya Dryomov struct_v, struct_cv); 3230205ee118SIlya Dryomov goto e_inval; 3231205ee118SIlya Dryomov } 3232205ee118SIlya Dryomov if (struct_cv > 6) { 3233205ee118SIlya Dryomov pr_warn("got v %d cv %d > 6 of ceph_object_locator\n", 3234205ee118SIlya Dryomov struct_v, struct_cv); 3235205ee118SIlya Dryomov goto e_inval; 3236205ee118SIlya Dryomov } 3237205ee118SIlya Dryomov len = ceph_decode_32(p); 3238205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3239205ee118SIlya Dryomov struct_end = *p + len; 3240205ee118SIlya Dryomov 3241205ee118SIlya Dryomov oloc->pool = ceph_decode_64(p); 3242205ee118SIlya Dryomov *p += 4; /* skip preferred */ 3243205ee118SIlya Dryomov 3244205ee118SIlya Dryomov len = ceph_decode_32(p); 3245205ee118SIlya Dryomov if (len > 0) { 3246205ee118SIlya Dryomov pr_warn("ceph_object_locator::key is set\n"); 3247205ee118SIlya Dryomov goto e_inval; 3248205ee118SIlya Dryomov } 3249205ee118SIlya Dryomov 3250205ee118SIlya Dryomov if (struct_v >= 5) { 3251cd08e0a2SYan, Zheng bool changed = false; 3252cd08e0a2SYan, Zheng 3253205ee118SIlya Dryomov len = ceph_decode_32(p); 3254205ee118SIlya Dryomov if (len > 0) { 325530c156d9SYan, Zheng ceph_decode_need(p, end, len, e_inval); 3256cd08e0a2SYan, Zheng if (!oloc->pool_ns || 3257cd08e0a2SYan, Zheng ceph_compare_string(oloc->pool_ns, *p, len)) 3258cd08e0a2SYan, Zheng changed = true; 325930c156d9SYan, Zheng *p += len; 3260cd08e0a2SYan, Zheng } else { 3261cd08e0a2SYan, Zheng if (oloc->pool_ns) 3262cd08e0a2SYan, Zheng changed = true; 3263cd08e0a2SYan, Zheng } 3264cd08e0a2SYan, Zheng if (changed) { 3265cd08e0a2SYan, Zheng /* redirect changes namespace */ 3266cd08e0a2SYan, Zheng pr_warn("ceph_object_locator::nspace is changed\n"); 3267cd08e0a2SYan, Zheng goto e_inval; 3268205ee118SIlya Dryomov } 3269205ee118SIlya Dryomov } 3270205ee118SIlya Dryomov 3271205ee118SIlya Dryomov if (struct_v >= 6) { 3272205ee118SIlya Dryomov s64 hash = ceph_decode_64(p); 3273205ee118SIlya Dryomov if (hash != -1) { 3274205ee118SIlya Dryomov pr_warn("ceph_object_locator::hash is set\n"); 3275205ee118SIlya Dryomov goto e_inval; 3276205ee118SIlya Dryomov } 3277205ee118SIlya Dryomov } 3278205ee118SIlya Dryomov 3279205ee118SIlya Dryomov /* skip the rest */ 3280205ee118SIlya Dryomov *p = struct_end; 3281205ee118SIlya Dryomov out: 3282205ee118SIlya Dryomov return ret; 3283205ee118SIlya Dryomov 3284205ee118SIlya Dryomov e_inval: 3285205ee118SIlya Dryomov ret = -EINVAL; 3286205ee118SIlya Dryomov goto out; 3287205ee118SIlya Dryomov } 3288205ee118SIlya Dryomov 3289205ee118SIlya Dryomov static int ceph_redirect_decode(void **p, void *end, 3290205ee118SIlya Dryomov struct ceph_request_redirect *redir) 3291205ee118SIlya Dryomov { 3292205ee118SIlya Dryomov u8 struct_v, struct_cv; 3293205ee118SIlya Dryomov u32 len; 3294205ee118SIlya Dryomov void *struct_end; 3295205ee118SIlya Dryomov int ret; 3296205ee118SIlya Dryomov 3297205ee118SIlya Dryomov ceph_decode_need(p, end, 1 + 1 + 4, e_inval); 3298205ee118SIlya Dryomov struct_v = ceph_decode_8(p); 3299205ee118SIlya Dryomov struct_cv = ceph_decode_8(p); 3300205ee118SIlya Dryomov if (struct_cv > 1) { 3301205ee118SIlya Dryomov pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n", 3302205ee118SIlya Dryomov struct_v, struct_cv); 3303205ee118SIlya Dryomov goto e_inval; 3304205ee118SIlya Dryomov } 3305205ee118SIlya Dryomov len = ceph_decode_32(p); 3306205ee118SIlya Dryomov ceph_decode_need(p, end, len, e_inval); 3307205ee118SIlya Dryomov struct_end = *p + len; 3308205ee118SIlya Dryomov 3309205ee118SIlya Dryomov ret = ceph_oloc_decode(p, end, &redir->oloc); 3310205ee118SIlya Dryomov if (ret) 3311205ee118SIlya Dryomov goto out; 3312205ee118SIlya Dryomov 3313205ee118SIlya Dryomov len = ceph_decode_32(p); 3314205ee118SIlya Dryomov if (len > 0) { 3315205ee118SIlya Dryomov pr_warn("ceph_request_redirect::object_name is set\n"); 3316205ee118SIlya Dryomov goto e_inval; 3317205ee118SIlya Dryomov } 3318205ee118SIlya Dryomov 3319205ee118SIlya Dryomov len = ceph_decode_32(p); 3320205ee118SIlya Dryomov *p += len; /* skip osd_instructions */ 3321205ee118SIlya Dryomov 3322205ee118SIlya Dryomov /* skip the rest */ 3323205ee118SIlya Dryomov *p = struct_end; 3324205ee118SIlya Dryomov out: 3325205ee118SIlya Dryomov return ret; 3326205ee118SIlya Dryomov 3327205ee118SIlya Dryomov e_inval: 3328205ee118SIlya Dryomov ret = -EINVAL; 3329205ee118SIlya Dryomov goto out; 3330205ee118SIlya Dryomov } 3331205ee118SIlya Dryomov 3332fe5da05eSIlya Dryomov struct MOSDOpReply { 3333fe5da05eSIlya Dryomov struct ceph_pg pgid; 3334fe5da05eSIlya Dryomov u64 flags; 3335fe5da05eSIlya Dryomov int result; 3336fe5da05eSIlya Dryomov u32 epoch; 3337fe5da05eSIlya Dryomov int num_ops; 3338fe5da05eSIlya Dryomov u32 outdata_len[CEPH_OSD_MAX_OPS]; 3339fe5da05eSIlya Dryomov s32 rval[CEPH_OSD_MAX_OPS]; 3340fe5da05eSIlya Dryomov int retry_attempt; 3341fe5da05eSIlya Dryomov struct ceph_eversion replay_version; 3342fe5da05eSIlya Dryomov u64 user_version; 3343fe5da05eSIlya Dryomov struct ceph_request_redirect redirect; 3344fe5da05eSIlya Dryomov }; 334525845472SSage Weil 3346fe5da05eSIlya Dryomov static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m) 33473d14c5d2SYehuda Sadeh { 3348fe5da05eSIlya Dryomov void *p = msg->front.iov_base; 3349fe5da05eSIlya Dryomov void *const end = p + msg->front.iov_len; 3350fe5da05eSIlya Dryomov u16 version = le16_to_cpu(msg->hdr.version); 3351fe5da05eSIlya Dryomov struct ceph_eversion bad_replay_version; 3352b0b31a8fSIlya Dryomov u8 decode_redir; 3353fe5da05eSIlya Dryomov u32 len; 3354fe5da05eSIlya Dryomov int ret; 3355fe5da05eSIlya Dryomov int i; 33563d14c5d2SYehuda Sadeh 3357fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, len, e_inval); 3358fe5da05eSIlya Dryomov ceph_decode_need(&p, end, len, e_inval); 3359fe5da05eSIlya Dryomov p += len; /* skip oid */ 33601b83bef2SSage Weil 3361fe5da05eSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->pgid); 3362fe5da05eSIlya Dryomov if (ret) 3363fe5da05eSIlya Dryomov return ret; 33641b83bef2SSage Weil 3365fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->flags, e_inval); 3366fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->result, e_inval); 3367fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval); 3368fe5da05eSIlya Dryomov memcpy(&bad_replay_version, p, sizeof(bad_replay_version)); 3369fe5da05eSIlya Dryomov p += sizeof(bad_replay_version); 3370fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->epoch, e_inval); 33711b83bef2SSage Weil 3372fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->num_ops, e_inval); 3373fe5da05eSIlya Dryomov if (m->num_ops > ARRAY_SIZE(m->outdata_len)) 3374fe5da05eSIlya Dryomov goto e_inval; 33751b83bef2SSage Weil 3376fe5da05eSIlya Dryomov ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op), 3377fe5da05eSIlya Dryomov e_inval); 3378fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) { 33791b83bef2SSage Weil struct ceph_osd_op *op = p; 33801b83bef2SSage Weil 3381fe5da05eSIlya Dryomov m->outdata_len[i] = le32_to_cpu(op->payload_len); 33821b83bef2SSage Weil p += sizeof(*op); 33831b83bef2SSage Weil } 3384fe5da05eSIlya Dryomov 3385fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval); 3386fe5da05eSIlya Dryomov for (i = 0; i < m->num_ops; i++) 3387fe5da05eSIlya Dryomov ceph_decode_32_safe(&p, end, m->rval[i], e_inval); 3388fe5da05eSIlya Dryomov 3389fe5da05eSIlya Dryomov if (version >= 5) { 3390fe5da05eSIlya Dryomov ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval); 3391fe5da05eSIlya Dryomov memcpy(&m->replay_version, p, sizeof(m->replay_version)); 3392fe5da05eSIlya Dryomov p += sizeof(m->replay_version); 3393fe5da05eSIlya Dryomov ceph_decode_64_safe(&p, end, m->user_version, e_inval); 3394fe5da05eSIlya Dryomov } else { 3395fe5da05eSIlya Dryomov m->replay_version = bad_replay_version; /* struct */ 3396fe5da05eSIlya Dryomov m->user_version = le64_to_cpu(m->replay_version.version); 33971b83bef2SSage Weil } 33981b83bef2SSage Weil 3399fe5da05eSIlya Dryomov if (version >= 6) { 3400fe5da05eSIlya Dryomov if (version >= 7) 3401fe5da05eSIlya Dryomov ceph_decode_8_safe(&p, end, decode_redir, e_inval); 3402b0b31a8fSIlya Dryomov else 3403b0b31a8fSIlya Dryomov decode_redir = 1; 3404b0b31a8fSIlya Dryomov } else { 3405b0b31a8fSIlya Dryomov decode_redir = 0; 3406b0b31a8fSIlya Dryomov } 3407b0b31a8fSIlya Dryomov 3408b0b31a8fSIlya Dryomov if (decode_redir) { 3409fe5da05eSIlya Dryomov ret = ceph_redirect_decode(&p, end, &m->redirect); 3410fe5da05eSIlya Dryomov if (ret) 3411fe5da05eSIlya Dryomov return ret; 3412205ee118SIlya Dryomov } else { 3413fe5da05eSIlya Dryomov ceph_oloc_init(&m->redirect.oloc); 3414205ee118SIlya Dryomov } 3415205ee118SIlya Dryomov 3416fe5da05eSIlya Dryomov return 0; 3417205ee118SIlya Dryomov 3418fe5da05eSIlya Dryomov e_inval: 3419fe5da05eSIlya Dryomov return -EINVAL; 3420fe5da05eSIlya Dryomov } 3421fe5da05eSIlya Dryomov 3422fe5da05eSIlya Dryomov /* 3423b18b9550SIlya Dryomov * Handle MOSDOpReply. Set ->r_result and call the callback if it is 3424b18b9550SIlya Dryomov * specified. 3425fe5da05eSIlya Dryomov */ 34265aea3dcdSIlya Dryomov static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg) 3427fe5da05eSIlya Dryomov { 34285aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 3429fe5da05eSIlya Dryomov struct ceph_osd_request *req; 3430fe5da05eSIlya Dryomov struct MOSDOpReply m; 3431fe5da05eSIlya Dryomov u64 tid = le64_to_cpu(msg->hdr.tid); 3432fe5da05eSIlya Dryomov u32 data_len = 0; 3433fe5da05eSIlya Dryomov int ret; 3434fe5da05eSIlya Dryomov int i; 3435fe5da05eSIlya Dryomov 3436fe5da05eSIlya Dryomov dout("%s msg %p tid %llu\n", __func__, msg, tid); 3437fe5da05eSIlya Dryomov 34385aea3dcdSIlya Dryomov down_read(&osdc->lock); 34395aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 34405aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 34415aea3dcdSIlya Dryomov goto out_unlock_osdc; 3442fe5da05eSIlya Dryomov } 34435aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 34445aea3dcdSIlya Dryomov 34455aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 34465aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 34475aea3dcdSIlya Dryomov if (!req) { 34485aea3dcdSIlya Dryomov dout("%s osd%d tid %llu unknown\n", __func__, osd->o_osd, tid); 34495aea3dcdSIlya Dryomov goto out_unlock_session; 34505aea3dcdSIlya Dryomov } 3451fe5da05eSIlya Dryomov 3452cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = req->r_t.target_oloc.pool_ns; 3453fe5da05eSIlya Dryomov ret = decode_MOSDOpReply(msg, &m); 3454cd08e0a2SYan, Zheng m.redirect.oloc.pool_ns = NULL; 3455fe5da05eSIlya Dryomov if (ret) { 3456fe5da05eSIlya Dryomov pr_err("failed to decode MOSDOpReply for tid %llu: %d\n", 3457fe5da05eSIlya Dryomov req->r_tid, ret); 3458fe5da05eSIlya Dryomov ceph_msg_dump(msg); 3459fe5da05eSIlya Dryomov goto fail_request; 3460fe5da05eSIlya Dryomov } 3461fe5da05eSIlya Dryomov dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n", 3462fe5da05eSIlya Dryomov __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed, 3463fe5da05eSIlya Dryomov m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch), 3464fe5da05eSIlya Dryomov le64_to_cpu(m.replay_version.version), m.user_version); 3465fe5da05eSIlya Dryomov 3466fe5da05eSIlya Dryomov if (m.retry_attempt >= 0) { 3467fe5da05eSIlya Dryomov if (m.retry_attempt != req->r_attempts - 1) { 3468fe5da05eSIlya Dryomov dout("req %p tid %llu retry_attempt %d != %d, ignoring\n", 3469fe5da05eSIlya Dryomov req, req->r_tid, m.retry_attempt, 3470fe5da05eSIlya Dryomov req->r_attempts - 1); 34715aea3dcdSIlya Dryomov goto out_unlock_session; 3472fe5da05eSIlya Dryomov } 3473fe5da05eSIlya Dryomov } else { 3474fe5da05eSIlya Dryomov WARN_ON(1); /* MOSDOpReply v4 is assumed */ 3475fe5da05eSIlya Dryomov } 3476fe5da05eSIlya Dryomov 3477fe5da05eSIlya Dryomov if (!ceph_oloc_empty(&m.redirect.oloc)) { 3478fe5da05eSIlya Dryomov dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid, 3479fe5da05eSIlya Dryomov m.redirect.oloc.pool); 34805aea3dcdSIlya Dryomov unlink_request(osd, req); 34815aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 3482205ee118SIlya Dryomov 348330c156d9SYan, Zheng /* 348430c156d9SYan, Zheng * Not ceph_oloc_copy() - changing pool_ns is not 348530c156d9SYan, Zheng * supported. 348630c156d9SYan, Zheng */ 348730c156d9SYan, Zheng req->r_t.target_oloc.pool = m.redirect.oloc.pool; 34885aea3dcdSIlya Dryomov req->r_flags |= CEPH_OSD_FLAG_REDIRECTED; 34895aea3dcdSIlya Dryomov req->r_tid = 0; 34905aea3dcdSIlya Dryomov __submit_request(req, false); 34915aea3dcdSIlya Dryomov goto out_unlock_osdc; 3492205ee118SIlya Dryomov } 3493205ee118SIlya Dryomov 3494fe5da05eSIlya Dryomov if (m.num_ops != req->r_num_ops) { 3495fe5da05eSIlya Dryomov pr_err("num_ops %d != %d for tid %llu\n", m.num_ops, 3496fe5da05eSIlya Dryomov req->r_num_ops, req->r_tid); 3497fe5da05eSIlya Dryomov goto fail_request; 3498fe5da05eSIlya Dryomov } 3499fe5da05eSIlya Dryomov for (i = 0; i < req->r_num_ops; i++) { 3500fe5da05eSIlya Dryomov dout(" req %p tid %llu op %d rval %d len %u\n", req, 3501fe5da05eSIlya Dryomov req->r_tid, i, m.rval[i], m.outdata_len[i]); 3502fe5da05eSIlya Dryomov req->r_ops[i].rval = m.rval[i]; 3503fe5da05eSIlya Dryomov req->r_ops[i].outdata_len = m.outdata_len[i]; 3504fe5da05eSIlya Dryomov data_len += m.outdata_len[i]; 3505fe5da05eSIlya Dryomov } 3506fe5da05eSIlya Dryomov if (data_len != le32_to_cpu(msg->hdr.data_len)) { 3507fe5da05eSIlya Dryomov pr_err("sum of lens %u != %u for tid %llu\n", data_len, 3508fe5da05eSIlya Dryomov le32_to_cpu(msg->hdr.data_len), req->r_tid); 3509fe5da05eSIlya Dryomov goto fail_request; 3510fe5da05eSIlya Dryomov } 3511b18b9550SIlya Dryomov dout("%s req %p tid %llu result %d data_len %u\n", __func__, 3512b18b9550SIlya Dryomov req, req->r_tid, m.result, data_len); 35133d14c5d2SYehuda Sadeh 3514b18b9550SIlya Dryomov /* 3515b18b9550SIlya Dryomov * Since we only ever request ONDISK, we should only ever get 3516b18b9550SIlya Dryomov * one (type of) reply back. 3517b18b9550SIlya Dryomov */ 3518b18b9550SIlya Dryomov WARN_ON(!(m.flags & CEPH_OSD_FLAG_ONDISK)); 3519fe5da05eSIlya Dryomov req->r_result = m.result ?: data_len; 352045ee2c1dSIlya Dryomov finish_request(req); 35215aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 35225aea3dcdSIlya Dryomov up_read(&osdc->lock); 35233d14c5d2SYehuda Sadeh 3524fe5da05eSIlya Dryomov __complete_request(req); 3525b18b9550SIlya Dryomov complete_all(&req->r_completion); 3526dc045a91SIlya Dryomov ceph_osdc_put_request(req); 35273d14c5d2SYehuda Sadeh return; 3528fe5da05eSIlya Dryomov 3529fe5da05eSIlya Dryomov fail_request: 35304609245eSIlya Dryomov complete_request(req, -EIO); 35315aea3dcdSIlya Dryomov out_unlock_session: 35325aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 35335aea3dcdSIlya Dryomov out_unlock_osdc: 35345aea3dcdSIlya Dryomov up_read(&osdc->lock); 35353d14c5d2SYehuda Sadeh } 35363d14c5d2SYehuda Sadeh 353742c1b124SIlya Dryomov static void set_pool_was_full(struct ceph_osd_client *osdc) 353842c1b124SIlya Dryomov { 353942c1b124SIlya Dryomov struct rb_node *n; 354042c1b124SIlya Dryomov 354142c1b124SIlya Dryomov for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) { 354242c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 354342c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 354442c1b124SIlya Dryomov 354542c1b124SIlya Dryomov pi->was_full = __pool_full(pi); 354642c1b124SIlya Dryomov } 354742c1b124SIlya Dryomov } 354842c1b124SIlya Dryomov 35495aea3dcdSIlya Dryomov static bool pool_cleared_full(struct ceph_osd_client *osdc, s64 pool_id) 35503d14c5d2SYehuda Sadeh { 35515aea3dcdSIlya Dryomov struct ceph_pg_pool_info *pi; 35523d14c5d2SYehuda Sadeh 35535aea3dcdSIlya Dryomov pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id); 35545aea3dcdSIlya Dryomov if (!pi) 35555aea3dcdSIlya Dryomov return false; 35563d14c5d2SYehuda Sadeh 35575aea3dcdSIlya Dryomov return pi->was_full && !__pool_full(pi); 35583d14c5d2SYehuda Sadeh } 35593d14c5d2SYehuda Sadeh 3560922dab61SIlya Dryomov static enum calc_target_result 3561922dab61SIlya Dryomov recalc_linger_target(struct ceph_osd_linger_request *lreq) 3562922dab61SIlya Dryomov { 3563922dab61SIlya Dryomov struct ceph_osd_client *osdc = lreq->osdc; 3564922dab61SIlya Dryomov enum calc_target_result ct_res; 3565922dab61SIlya Dryomov 35667de030d6SIlya Dryomov ct_res = calc_target(osdc, &lreq->t, NULL, true); 3567922dab61SIlya Dryomov if (ct_res == CALC_TARGET_NEED_RESEND) { 3568922dab61SIlya Dryomov struct ceph_osd *osd; 3569922dab61SIlya Dryomov 3570922dab61SIlya Dryomov osd = lookup_create_osd(osdc, lreq->t.osd, true); 3571922dab61SIlya Dryomov if (osd != lreq->osd) { 3572922dab61SIlya Dryomov unlink_linger(lreq->osd, lreq); 3573922dab61SIlya Dryomov link_linger(osd, lreq); 3574922dab61SIlya Dryomov } 3575922dab61SIlya Dryomov } 3576922dab61SIlya Dryomov 3577922dab61SIlya Dryomov return ct_res; 3578922dab61SIlya Dryomov } 3579922dab61SIlya Dryomov 35803d14c5d2SYehuda Sadeh /* 35815aea3dcdSIlya Dryomov * Requeue requests whose mapping to an OSD has changed. 35823d14c5d2SYehuda Sadeh */ 35835aea3dcdSIlya Dryomov static void scan_requests(struct ceph_osd *osd, 35845aea3dcdSIlya Dryomov bool force_resend, 35855aea3dcdSIlya Dryomov bool cleared_full, 35865aea3dcdSIlya Dryomov bool check_pool_cleared_full, 35875aea3dcdSIlya Dryomov struct rb_root *need_resend, 35885aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 35893d14c5d2SYehuda Sadeh { 35905aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 35915aea3dcdSIlya Dryomov struct rb_node *n; 35925aea3dcdSIlya Dryomov bool force_resend_writes; 35933d14c5d2SYehuda Sadeh 3594922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; ) { 3595922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3596922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 3597922dab61SIlya Dryomov enum calc_target_result ct_res; 3598922dab61SIlya Dryomov 3599922dab61SIlya Dryomov n = rb_next(n); /* recalc_linger_target() */ 3600922dab61SIlya Dryomov 3601922dab61SIlya Dryomov dout("%s lreq %p linger_id %llu\n", __func__, lreq, 3602922dab61SIlya Dryomov lreq->linger_id); 3603922dab61SIlya Dryomov ct_res = recalc_linger_target(lreq); 3604922dab61SIlya Dryomov switch (ct_res) { 3605922dab61SIlya Dryomov case CALC_TARGET_NO_ACTION: 3606922dab61SIlya Dryomov force_resend_writes = cleared_full || 3607922dab61SIlya Dryomov (check_pool_cleared_full && 3608922dab61SIlya Dryomov pool_cleared_full(osdc, lreq->t.base_oloc.pool)); 3609922dab61SIlya Dryomov if (!force_resend && !force_resend_writes) 3610922dab61SIlya Dryomov break; 3611922dab61SIlya Dryomov 3612922dab61SIlya Dryomov /* fall through */ 3613922dab61SIlya Dryomov case CALC_TARGET_NEED_RESEND: 36144609245eSIlya Dryomov cancel_linger_map_check(lreq); 3615922dab61SIlya Dryomov /* 3616922dab61SIlya Dryomov * scan_requests() for the previous epoch(s) 3617922dab61SIlya Dryomov * may have already added it to the list, since 3618922dab61SIlya Dryomov * it's not unlinked here. 3619922dab61SIlya Dryomov */ 3620922dab61SIlya Dryomov if (list_empty(&lreq->scan_item)) 3621922dab61SIlya Dryomov list_add_tail(&lreq->scan_item, need_resend_linger); 3622922dab61SIlya Dryomov break; 3623922dab61SIlya Dryomov case CALC_TARGET_POOL_DNE: 3624a10bcb19SIlya Dryomov list_del_init(&lreq->scan_item); 36254609245eSIlya Dryomov check_linger_pool_dne(lreq); 3626922dab61SIlya Dryomov break; 3627922dab61SIlya Dryomov } 3628922dab61SIlya Dryomov } 3629922dab61SIlya Dryomov 36305aea3dcdSIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 36315aea3dcdSIlya Dryomov struct ceph_osd_request *req = 36325aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 36335aea3dcdSIlya Dryomov enum calc_target_result ct_res; 3634ab60b16dSAlex Elder 36354609245eSIlya Dryomov n = rb_next(n); /* unlink_request(), check_pool_dne() */ 3636ab60b16dSAlex Elder 36375aea3dcdSIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 36387de030d6SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, &req->r_osd->o_con, 36397de030d6SIlya Dryomov false); 36405aea3dcdSIlya Dryomov switch (ct_res) { 36415aea3dcdSIlya Dryomov case CALC_TARGET_NO_ACTION: 36425aea3dcdSIlya Dryomov force_resend_writes = cleared_full || 36435aea3dcdSIlya Dryomov (check_pool_cleared_full && 36445aea3dcdSIlya Dryomov pool_cleared_full(osdc, req->r_t.base_oloc.pool)); 36455aea3dcdSIlya Dryomov if (!force_resend && 36465aea3dcdSIlya Dryomov (!(req->r_flags & CEPH_OSD_FLAG_WRITE) || 36475aea3dcdSIlya Dryomov !force_resend_writes)) 36485aea3dcdSIlya Dryomov break; 3649a40c4f10SYehuda Sadeh 36505aea3dcdSIlya Dryomov /* fall through */ 36515aea3dcdSIlya Dryomov case CALC_TARGET_NEED_RESEND: 36524609245eSIlya Dryomov cancel_map_check(req); 36535aea3dcdSIlya Dryomov unlink_request(osd, req); 36545aea3dcdSIlya Dryomov insert_request(need_resend, req); 36555aea3dcdSIlya Dryomov break; 36565aea3dcdSIlya Dryomov case CALC_TARGET_POOL_DNE: 36574609245eSIlya Dryomov check_pool_dne(req); 36585aea3dcdSIlya Dryomov break; 3659a40c4f10SYehuda Sadeh } 36603d14c5d2SYehuda Sadeh } 36613d14c5d2SYehuda Sadeh } 36626f6c7006SSage Weil 366342c1b124SIlya Dryomov static int handle_one_map(struct ceph_osd_client *osdc, 36645aea3dcdSIlya Dryomov void *p, void *end, bool incremental, 36655aea3dcdSIlya Dryomov struct rb_root *need_resend, 36665aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 366742c1b124SIlya Dryomov { 366842c1b124SIlya Dryomov struct ceph_osdmap *newmap; 366942c1b124SIlya Dryomov struct rb_node *n; 367042c1b124SIlya Dryomov bool skipped_map = false; 367142c1b124SIlya Dryomov bool was_full; 367242c1b124SIlya Dryomov 3673b7ec35b3SIlya Dryomov was_full = ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 367442c1b124SIlya Dryomov set_pool_was_full(osdc); 367542c1b124SIlya Dryomov 367642c1b124SIlya Dryomov if (incremental) 367742c1b124SIlya Dryomov newmap = osdmap_apply_incremental(&p, end, osdc->osdmap); 367842c1b124SIlya Dryomov else 367942c1b124SIlya Dryomov newmap = ceph_osdmap_decode(&p, end); 368042c1b124SIlya Dryomov if (IS_ERR(newmap)) 368142c1b124SIlya Dryomov return PTR_ERR(newmap); 368242c1b124SIlya Dryomov 368342c1b124SIlya Dryomov if (newmap != osdc->osdmap) { 368442c1b124SIlya Dryomov /* 368542c1b124SIlya Dryomov * Preserve ->was_full before destroying the old map. 368642c1b124SIlya Dryomov * For pools that weren't in the old map, ->was_full 368742c1b124SIlya Dryomov * should be false. 368842c1b124SIlya Dryomov */ 368942c1b124SIlya Dryomov for (n = rb_first(&newmap->pg_pools); n; n = rb_next(n)) { 369042c1b124SIlya Dryomov struct ceph_pg_pool_info *pi = 369142c1b124SIlya Dryomov rb_entry(n, struct ceph_pg_pool_info, node); 369242c1b124SIlya Dryomov struct ceph_pg_pool_info *old_pi; 369342c1b124SIlya Dryomov 369442c1b124SIlya Dryomov old_pi = ceph_pg_pool_by_id(osdc->osdmap, pi->id); 369542c1b124SIlya Dryomov if (old_pi) 369642c1b124SIlya Dryomov pi->was_full = old_pi->was_full; 369742c1b124SIlya Dryomov else 369842c1b124SIlya Dryomov WARN_ON(pi->was_full); 369942c1b124SIlya Dryomov } 370042c1b124SIlya Dryomov 370142c1b124SIlya Dryomov if (osdc->osdmap->epoch && 370242c1b124SIlya Dryomov osdc->osdmap->epoch + 1 < newmap->epoch) { 370342c1b124SIlya Dryomov WARN_ON(incremental); 370442c1b124SIlya Dryomov skipped_map = true; 370542c1b124SIlya Dryomov } 370642c1b124SIlya Dryomov 370742c1b124SIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 370842c1b124SIlya Dryomov osdc->osdmap = newmap; 370942c1b124SIlya Dryomov } 371042c1b124SIlya Dryomov 3711b7ec35b3SIlya Dryomov was_full &= !ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL); 37125aea3dcdSIlya Dryomov scan_requests(&osdc->homeless_osd, skipped_map, was_full, true, 37135aea3dcdSIlya Dryomov need_resend, need_resend_linger); 37145aea3dcdSIlya Dryomov 37155aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; ) { 37165aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 37175aea3dcdSIlya Dryomov 37185aea3dcdSIlya Dryomov n = rb_next(n); /* close_osd() */ 37195aea3dcdSIlya Dryomov 37205aea3dcdSIlya Dryomov scan_requests(osd, skipped_map, was_full, true, need_resend, 37215aea3dcdSIlya Dryomov need_resend_linger); 37225aea3dcdSIlya Dryomov if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) || 37235aea3dcdSIlya Dryomov memcmp(&osd->o_con.peer_addr, 37245aea3dcdSIlya Dryomov ceph_osd_addr(osdc->osdmap, osd->o_osd), 37255aea3dcdSIlya Dryomov sizeof(struct ceph_entity_addr))) 37265aea3dcdSIlya Dryomov close_osd(osd); 37275aea3dcdSIlya Dryomov } 372842c1b124SIlya Dryomov 372942c1b124SIlya Dryomov return 0; 373042c1b124SIlya Dryomov } 37316f6c7006SSage Weil 37325aea3dcdSIlya Dryomov static void kick_requests(struct ceph_osd_client *osdc, 37335aea3dcdSIlya Dryomov struct rb_root *need_resend, 37345aea3dcdSIlya Dryomov struct list_head *need_resend_linger) 37355aea3dcdSIlya Dryomov { 3736922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq, *nlreq; 373704c7d789SIlya Dryomov enum calc_target_result ct_res; 37385aea3dcdSIlya Dryomov struct rb_node *n; 37395aea3dcdSIlya Dryomov 374004c7d789SIlya Dryomov /* make sure need_resend targets reflect latest map */ 374104c7d789SIlya Dryomov for (n = rb_first(need_resend); n; ) { 374204c7d789SIlya Dryomov struct ceph_osd_request *req = 374304c7d789SIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 374404c7d789SIlya Dryomov 374504c7d789SIlya Dryomov n = rb_next(n); 374604c7d789SIlya Dryomov 374704c7d789SIlya Dryomov if (req->r_t.epoch < osdc->osdmap->epoch) { 374804c7d789SIlya Dryomov ct_res = calc_target(osdc, &req->r_t, NULL, false); 374904c7d789SIlya Dryomov if (ct_res == CALC_TARGET_POOL_DNE) { 375004c7d789SIlya Dryomov erase_request(need_resend, req); 375104c7d789SIlya Dryomov check_pool_dne(req); 375204c7d789SIlya Dryomov } 375304c7d789SIlya Dryomov } 375404c7d789SIlya Dryomov } 375504c7d789SIlya Dryomov 37565aea3dcdSIlya Dryomov for (n = rb_first(need_resend); n; ) { 37575aea3dcdSIlya Dryomov struct ceph_osd_request *req = 37585aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 37595aea3dcdSIlya Dryomov struct ceph_osd *osd; 37605aea3dcdSIlya Dryomov 37615aea3dcdSIlya Dryomov n = rb_next(n); 37625aea3dcdSIlya Dryomov erase_request(need_resend, req); /* before link_request() */ 37635aea3dcdSIlya Dryomov 37645aea3dcdSIlya Dryomov osd = lookup_create_osd(osdc, req->r_t.osd, true); 37655aea3dcdSIlya Dryomov link_request(osd, req); 37665aea3dcdSIlya Dryomov if (!req->r_linger) { 37675aea3dcdSIlya Dryomov if (!osd_homeless(osd) && !req->r_t.paused) 37685aea3dcdSIlya Dryomov send_request(req); 3769922dab61SIlya Dryomov } else { 3770922dab61SIlya Dryomov cancel_linger_request(req); 37715aea3dcdSIlya Dryomov } 37725aea3dcdSIlya Dryomov } 3773922dab61SIlya Dryomov 3774922dab61SIlya Dryomov list_for_each_entry_safe(lreq, nlreq, need_resend_linger, scan_item) { 3775922dab61SIlya Dryomov if (!osd_homeless(lreq->osd)) 3776922dab61SIlya Dryomov send_linger(lreq); 3777922dab61SIlya Dryomov 3778922dab61SIlya Dryomov list_del_init(&lreq->scan_item); 3779922dab61SIlya Dryomov } 37805aea3dcdSIlya Dryomov } 37815aea3dcdSIlya Dryomov 37823d14c5d2SYehuda Sadeh /* 37833d14c5d2SYehuda Sadeh * Process updated osd map. 37843d14c5d2SYehuda Sadeh * 37853d14c5d2SYehuda Sadeh * The message contains any number of incremental and full maps, normally 37863d14c5d2SYehuda Sadeh * indicating some sort of topology change in the cluster. Kick requests 37873d14c5d2SYehuda Sadeh * off to different OSDs as needed. 37883d14c5d2SYehuda Sadeh */ 37893d14c5d2SYehuda Sadeh void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg) 37903d14c5d2SYehuda Sadeh { 379142c1b124SIlya Dryomov void *p = msg->front.iov_base; 379242c1b124SIlya Dryomov void *const end = p + msg->front.iov_len; 37933d14c5d2SYehuda Sadeh u32 nr_maps, maplen; 37943d14c5d2SYehuda Sadeh u32 epoch; 37953d14c5d2SYehuda Sadeh struct ceph_fsid fsid; 37965aea3dcdSIlya Dryomov struct rb_root need_resend = RB_ROOT; 37975aea3dcdSIlya Dryomov LIST_HEAD(need_resend_linger); 379842c1b124SIlya Dryomov bool handled_incremental = false; 379942c1b124SIlya Dryomov bool was_pauserd, was_pausewr; 380042c1b124SIlya Dryomov bool pauserd, pausewr; 380142c1b124SIlya Dryomov int err; 38023d14c5d2SYehuda Sadeh 380342c1b124SIlya Dryomov dout("%s have %u\n", __func__, osdc->osdmap->epoch); 38045aea3dcdSIlya Dryomov down_write(&osdc->lock); 38053d14c5d2SYehuda Sadeh 38063d14c5d2SYehuda Sadeh /* verify fsid */ 38073d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, sizeof(fsid), bad); 38083d14c5d2SYehuda Sadeh ceph_decode_copy(&p, &fsid, sizeof(fsid)); 38093d14c5d2SYehuda Sadeh if (ceph_check_fsid(osdc->client, &fsid) < 0) 381042c1b124SIlya Dryomov goto bad; 38113d14c5d2SYehuda Sadeh 3812b7ec35b3SIlya Dryomov was_pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 3813b7ec35b3SIlya Dryomov was_pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 3814b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 381542c1b124SIlya Dryomov have_pool_full(osdc); 38169a1ea2dbSJosh Durgin 38173d14c5d2SYehuda Sadeh /* incremental maps */ 38183d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 38193d14c5d2SYehuda Sadeh dout(" %d inc maps\n", nr_maps); 38203d14c5d2SYehuda Sadeh while (nr_maps > 0) { 38213d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 38223d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 38233d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 38243d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 382542c1b124SIlya Dryomov if (osdc->osdmap->epoch && 382642c1b124SIlya Dryomov osdc->osdmap->epoch + 1 == epoch) { 38273d14c5d2SYehuda Sadeh dout("applying incremental map %u len %d\n", 38283d14c5d2SYehuda Sadeh epoch, maplen); 38295aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, true, 38305aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 383142c1b124SIlya Dryomov if (err) 38323d14c5d2SYehuda Sadeh goto bad; 383342c1b124SIlya Dryomov handled_incremental = true; 38343d14c5d2SYehuda Sadeh } else { 38353d14c5d2SYehuda Sadeh dout("ignoring incremental map %u len %d\n", 38363d14c5d2SYehuda Sadeh epoch, maplen); 38373d14c5d2SYehuda Sadeh } 383842c1b124SIlya Dryomov p += maplen; 38393d14c5d2SYehuda Sadeh nr_maps--; 38403d14c5d2SYehuda Sadeh } 384142c1b124SIlya Dryomov if (handled_incremental) 38423d14c5d2SYehuda Sadeh goto done; 38433d14c5d2SYehuda Sadeh 38443d14c5d2SYehuda Sadeh /* full maps */ 38453d14c5d2SYehuda Sadeh ceph_decode_32_safe(&p, end, nr_maps, bad); 38463d14c5d2SYehuda Sadeh dout(" %d full maps\n", nr_maps); 38473d14c5d2SYehuda Sadeh while (nr_maps) { 38483d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, 2*sizeof(u32), bad); 38493d14c5d2SYehuda Sadeh epoch = ceph_decode_32(&p); 38503d14c5d2SYehuda Sadeh maplen = ceph_decode_32(&p); 38513d14c5d2SYehuda Sadeh ceph_decode_need(&p, end, maplen, bad); 38523d14c5d2SYehuda Sadeh if (nr_maps > 1) { 38533d14c5d2SYehuda Sadeh dout("skipping non-latest full map %u len %d\n", 38543d14c5d2SYehuda Sadeh epoch, maplen); 3855e5253a7bSIlya Dryomov } else if (osdc->osdmap->epoch >= epoch) { 38563d14c5d2SYehuda Sadeh dout("skipping full map %u len %d, " 38573d14c5d2SYehuda Sadeh "older than our %u\n", epoch, maplen, 38583d14c5d2SYehuda Sadeh osdc->osdmap->epoch); 38593d14c5d2SYehuda Sadeh } else { 38603d14c5d2SYehuda Sadeh dout("taking full map %u len %d\n", epoch, maplen); 38615aea3dcdSIlya Dryomov err = handle_one_map(osdc, p, p + maplen, false, 38625aea3dcdSIlya Dryomov &need_resend, &need_resend_linger); 386342c1b124SIlya Dryomov if (err) 38643d14c5d2SYehuda Sadeh goto bad; 38653d14c5d2SYehuda Sadeh } 38663d14c5d2SYehuda Sadeh p += maplen; 38673d14c5d2SYehuda Sadeh nr_maps--; 38683d14c5d2SYehuda Sadeh } 38693d14c5d2SYehuda Sadeh 38703d14c5d2SYehuda Sadeh done: 3871cd634fb6SSage Weil /* 3872cd634fb6SSage Weil * subscribe to subsequent osdmap updates if full to ensure 3873cd634fb6SSage Weil * we find out when we are no longer full and stop returning 3874cd634fb6SSage Weil * ENOSPC. 3875cd634fb6SSage Weil */ 3876b7ec35b3SIlya Dryomov pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD); 3877b7ec35b3SIlya Dryomov pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) || 3878b7ec35b3SIlya Dryomov ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || 387942c1b124SIlya Dryomov have_pool_full(osdc); 388058eb7932SJeff Layton if (was_pauserd || was_pausewr || pauserd || pausewr || 388158eb7932SJeff Layton osdc->osdmap->epoch < osdc->epoch_barrier) 388242c1b124SIlya Dryomov maybe_request_map(osdc); 3883cd634fb6SSage Weil 38845aea3dcdSIlya Dryomov kick_requests(osdc, &need_resend, &need_resend_linger); 388542c1b124SIlya Dryomov 3886fc36d0a4SJeff Layton ceph_osdc_abort_on_full(osdc); 388742c1b124SIlya Dryomov ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP, 388842c1b124SIlya Dryomov osdc->osdmap->epoch); 38895aea3dcdSIlya Dryomov up_write(&osdc->lock); 38903d14c5d2SYehuda Sadeh wake_up_all(&osdc->client->auth_wq); 38913d14c5d2SYehuda Sadeh return; 38923d14c5d2SYehuda Sadeh 38933d14c5d2SYehuda Sadeh bad: 38943d14c5d2SYehuda Sadeh pr_err("osdc handle_map corrupt msg\n"); 38953d14c5d2SYehuda Sadeh ceph_msg_dump(msg); 38965aea3dcdSIlya Dryomov up_write(&osdc->lock); 38975aea3dcdSIlya Dryomov } 38985aea3dcdSIlya Dryomov 38995aea3dcdSIlya Dryomov /* 39005aea3dcdSIlya Dryomov * Resubmit requests pending on the given osd. 39015aea3dcdSIlya Dryomov */ 39025aea3dcdSIlya Dryomov static void kick_osd_requests(struct ceph_osd *osd) 39035aea3dcdSIlya Dryomov { 39045aea3dcdSIlya Dryomov struct rb_node *n; 39055aea3dcdSIlya Dryomov 3906a02a946dSIlya Dryomov clear_backoffs(osd); 3907a02a946dSIlya Dryomov 3908922dab61SIlya Dryomov for (n = rb_first(&osd->o_requests); n; ) { 39095aea3dcdSIlya Dryomov struct ceph_osd_request *req = 39105aea3dcdSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 39115aea3dcdSIlya Dryomov 3912922dab61SIlya Dryomov n = rb_next(n); /* cancel_linger_request() */ 3913922dab61SIlya Dryomov 39145aea3dcdSIlya Dryomov if (!req->r_linger) { 39155aea3dcdSIlya Dryomov if (!req->r_t.paused) 39165aea3dcdSIlya Dryomov send_request(req); 3917922dab61SIlya Dryomov } else { 3918922dab61SIlya Dryomov cancel_linger_request(req); 39195aea3dcdSIlya Dryomov } 39205aea3dcdSIlya Dryomov } 3921922dab61SIlya Dryomov for (n = rb_first(&osd->o_linger_requests); n; n = rb_next(n)) { 3922922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq = 3923922dab61SIlya Dryomov rb_entry(n, struct ceph_osd_linger_request, node); 3924922dab61SIlya Dryomov 3925922dab61SIlya Dryomov send_linger(lreq); 3926922dab61SIlya Dryomov } 39275aea3dcdSIlya Dryomov } 39285aea3dcdSIlya Dryomov 39295aea3dcdSIlya Dryomov /* 39305aea3dcdSIlya Dryomov * If the osd connection drops, we need to resubmit all requests. 39315aea3dcdSIlya Dryomov */ 39325aea3dcdSIlya Dryomov static void osd_fault(struct ceph_connection *con) 39335aea3dcdSIlya Dryomov { 39345aea3dcdSIlya Dryomov struct ceph_osd *osd = con->private; 39355aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 39365aea3dcdSIlya Dryomov 39375aea3dcdSIlya Dryomov dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd); 39385aea3dcdSIlya Dryomov 39395aea3dcdSIlya Dryomov down_write(&osdc->lock); 39405aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 39415aea3dcdSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 39425aea3dcdSIlya Dryomov goto out_unlock; 39435aea3dcdSIlya Dryomov } 39445aea3dcdSIlya Dryomov 39455aea3dcdSIlya Dryomov if (!reopen_osd(osd)) 39465aea3dcdSIlya Dryomov kick_osd_requests(osd); 39475aea3dcdSIlya Dryomov maybe_request_map(osdc); 39485aea3dcdSIlya Dryomov 39495aea3dcdSIlya Dryomov out_unlock: 39505aea3dcdSIlya Dryomov up_write(&osdc->lock); 39513d14c5d2SYehuda Sadeh } 39523d14c5d2SYehuda Sadeh 3953a02a946dSIlya Dryomov struct MOSDBackoff { 3954a02a946dSIlya Dryomov struct ceph_spg spgid; 3955a02a946dSIlya Dryomov u32 map_epoch; 3956a02a946dSIlya Dryomov u8 op; 3957a02a946dSIlya Dryomov u64 id; 3958a02a946dSIlya Dryomov struct ceph_hobject_id *begin; 3959a02a946dSIlya Dryomov struct ceph_hobject_id *end; 3960a02a946dSIlya Dryomov }; 3961a02a946dSIlya Dryomov 3962a02a946dSIlya Dryomov static int decode_MOSDBackoff(const struct ceph_msg *msg, struct MOSDBackoff *m) 3963a02a946dSIlya Dryomov { 3964a02a946dSIlya Dryomov void *p = msg->front.iov_base; 3965a02a946dSIlya Dryomov void *const end = p + msg->front.iov_len; 3966a02a946dSIlya Dryomov u8 struct_v; 3967a02a946dSIlya Dryomov u32 struct_len; 3968a02a946dSIlya Dryomov int ret; 3969a02a946dSIlya Dryomov 3970a02a946dSIlya Dryomov ret = ceph_start_decoding(&p, end, 1, "spg_t", &struct_v, &struct_len); 3971a02a946dSIlya Dryomov if (ret) 3972a02a946dSIlya Dryomov return ret; 3973a02a946dSIlya Dryomov 3974a02a946dSIlya Dryomov ret = ceph_decode_pgid(&p, end, &m->spgid.pgid); 3975a02a946dSIlya Dryomov if (ret) 3976a02a946dSIlya Dryomov return ret; 3977a02a946dSIlya Dryomov 3978a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->spgid.shard, e_inval); 3979a02a946dSIlya Dryomov ceph_decode_32_safe(&p, end, m->map_epoch, e_inval); 3980a02a946dSIlya Dryomov ceph_decode_8_safe(&p, end, m->op, e_inval); 3981a02a946dSIlya Dryomov ceph_decode_64_safe(&p, end, m->id, e_inval); 3982a02a946dSIlya Dryomov 3983a02a946dSIlya Dryomov m->begin = kzalloc(sizeof(*m->begin), GFP_NOIO); 3984a02a946dSIlya Dryomov if (!m->begin) 3985a02a946dSIlya Dryomov return -ENOMEM; 3986a02a946dSIlya Dryomov 3987a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->begin); 3988a02a946dSIlya Dryomov if (ret) { 3989a02a946dSIlya Dryomov free_hoid(m->begin); 3990a02a946dSIlya Dryomov return ret; 3991a02a946dSIlya Dryomov } 3992a02a946dSIlya Dryomov 3993a02a946dSIlya Dryomov m->end = kzalloc(sizeof(*m->end), GFP_NOIO); 3994a02a946dSIlya Dryomov if (!m->end) { 3995a02a946dSIlya Dryomov free_hoid(m->begin); 3996a02a946dSIlya Dryomov return -ENOMEM; 3997a02a946dSIlya Dryomov } 3998a02a946dSIlya Dryomov 3999a02a946dSIlya Dryomov ret = decode_hoid(&p, end, m->end); 4000a02a946dSIlya Dryomov if (ret) { 4001a02a946dSIlya Dryomov free_hoid(m->begin); 4002a02a946dSIlya Dryomov free_hoid(m->end); 4003a02a946dSIlya Dryomov return ret; 4004a02a946dSIlya Dryomov } 4005a02a946dSIlya Dryomov 4006a02a946dSIlya Dryomov return 0; 4007a02a946dSIlya Dryomov 4008a02a946dSIlya Dryomov e_inval: 4009a02a946dSIlya Dryomov return -EINVAL; 4010a02a946dSIlya Dryomov } 4011a02a946dSIlya Dryomov 4012a02a946dSIlya Dryomov static struct ceph_msg *create_backoff_message( 4013a02a946dSIlya Dryomov const struct ceph_osd_backoff *backoff, 4014a02a946dSIlya Dryomov u32 map_epoch) 4015a02a946dSIlya Dryomov { 4016a02a946dSIlya Dryomov struct ceph_msg *msg; 4017a02a946dSIlya Dryomov void *p, *end; 4018a02a946dSIlya Dryomov int msg_size; 4019a02a946dSIlya Dryomov 4020a02a946dSIlya Dryomov msg_size = CEPH_ENCODING_START_BLK_LEN + 4021a02a946dSIlya Dryomov CEPH_PGID_ENCODING_LEN + 1; /* spgid */ 4022a02a946dSIlya Dryomov msg_size += 4 + 1 + 8; /* map_epoch, op, id */ 4023a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4024a02a946dSIlya Dryomov hoid_encoding_size(backoff->begin); 4025a02a946dSIlya Dryomov msg_size += CEPH_ENCODING_START_BLK_LEN + 4026a02a946dSIlya Dryomov hoid_encoding_size(backoff->end); 4027a02a946dSIlya Dryomov 4028a02a946dSIlya Dryomov msg = ceph_msg_new(CEPH_MSG_OSD_BACKOFF, msg_size, GFP_NOIO, true); 4029a02a946dSIlya Dryomov if (!msg) 4030a02a946dSIlya Dryomov return NULL; 4031a02a946dSIlya Dryomov 4032a02a946dSIlya Dryomov p = msg->front.iov_base; 4033a02a946dSIlya Dryomov end = p + msg->front_alloc_len; 4034a02a946dSIlya Dryomov 4035a02a946dSIlya Dryomov encode_spgid(&p, &backoff->spgid); 4036a02a946dSIlya Dryomov ceph_encode_32(&p, map_epoch); 4037a02a946dSIlya Dryomov ceph_encode_8(&p, CEPH_OSD_BACKOFF_OP_ACK_BLOCK); 4038a02a946dSIlya Dryomov ceph_encode_64(&p, backoff->id); 4039a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->begin); 4040a02a946dSIlya Dryomov encode_hoid(&p, end, backoff->end); 4041a02a946dSIlya Dryomov BUG_ON(p != end); 4042a02a946dSIlya Dryomov 4043a02a946dSIlya Dryomov msg->front.iov_len = p - msg->front.iov_base; 4044a02a946dSIlya Dryomov msg->hdr.version = cpu_to_le16(1); /* MOSDBackoff v1 */ 4045a02a946dSIlya Dryomov msg->hdr.front_len = cpu_to_le32(msg->front.iov_len); 4046a02a946dSIlya Dryomov 4047a02a946dSIlya Dryomov return msg; 4048a02a946dSIlya Dryomov } 4049a02a946dSIlya Dryomov 4050a02a946dSIlya Dryomov static void handle_backoff_block(struct ceph_osd *osd, struct MOSDBackoff *m) 4051a02a946dSIlya Dryomov { 4052a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4053a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4054a02a946dSIlya Dryomov struct ceph_msg *msg; 4055a02a946dSIlya Dryomov 4056a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4057a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4058a02a946dSIlya Dryomov 4059a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &m->spgid); 4060a02a946dSIlya Dryomov if (!spg) { 4061a02a946dSIlya Dryomov spg = alloc_spg_mapping(); 4062a02a946dSIlya Dryomov if (!spg) { 4063a02a946dSIlya Dryomov pr_err("%s failed to allocate spg\n", __func__); 4064a02a946dSIlya Dryomov return; 4065a02a946dSIlya Dryomov } 4066a02a946dSIlya Dryomov spg->spgid = m->spgid; /* struct */ 4067a02a946dSIlya Dryomov insert_spg_mapping(&osd->o_backoff_mappings, spg); 4068a02a946dSIlya Dryomov } 4069a02a946dSIlya Dryomov 4070a02a946dSIlya Dryomov backoff = alloc_backoff(); 4071a02a946dSIlya Dryomov if (!backoff) { 4072a02a946dSIlya Dryomov pr_err("%s failed to allocate backoff\n", __func__); 4073a02a946dSIlya Dryomov return; 4074a02a946dSIlya Dryomov } 4075a02a946dSIlya Dryomov backoff->spgid = m->spgid; /* struct */ 4076a02a946dSIlya Dryomov backoff->id = m->id; 4077a02a946dSIlya Dryomov backoff->begin = m->begin; 4078a02a946dSIlya Dryomov m->begin = NULL; /* backoff now owns this */ 4079a02a946dSIlya Dryomov backoff->end = m->end; 4080a02a946dSIlya Dryomov m->end = NULL; /* ditto */ 4081a02a946dSIlya Dryomov 4082a02a946dSIlya Dryomov insert_backoff(&spg->backoffs, backoff); 4083a02a946dSIlya Dryomov insert_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4084a02a946dSIlya Dryomov 4085a02a946dSIlya Dryomov /* 4086a02a946dSIlya Dryomov * Ack with original backoff's epoch so that the OSD can 4087a02a946dSIlya Dryomov * discard this if there was a PG split. 4088a02a946dSIlya Dryomov */ 4089a02a946dSIlya Dryomov msg = create_backoff_message(backoff, m->map_epoch); 4090a02a946dSIlya Dryomov if (!msg) { 4091a02a946dSIlya Dryomov pr_err("%s failed to allocate msg\n", __func__); 4092a02a946dSIlya Dryomov return; 4093a02a946dSIlya Dryomov } 4094a02a946dSIlya Dryomov ceph_con_send(&osd->o_con, msg); 4095a02a946dSIlya Dryomov } 4096a02a946dSIlya Dryomov 4097a02a946dSIlya Dryomov static bool target_contained_by(const struct ceph_osd_request_target *t, 4098a02a946dSIlya Dryomov const struct ceph_hobject_id *begin, 4099a02a946dSIlya Dryomov const struct ceph_hobject_id *end) 4100a02a946dSIlya Dryomov { 4101a02a946dSIlya Dryomov struct ceph_hobject_id hoid; 4102a02a946dSIlya Dryomov int cmp; 4103a02a946dSIlya Dryomov 4104a02a946dSIlya Dryomov hoid_fill_from_target(&hoid, t); 4105a02a946dSIlya Dryomov cmp = hoid_compare(&hoid, begin); 4106a02a946dSIlya Dryomov return !cmp || (cmp > 0 && hoid_compare(&hoid, end) < 0); 4107a02a946dSIlya Dryomov } 4108a02a946dSIlya Dryomov 4109a02a946dSIlya Dryomov static void handle_backoff_unblock(struct ceph_osd *osd, 4110a02a946dSIlya Dryomov const struct MOSDBackoff *m) 4111a02a946dSIlya Dryomov { 4112a02a946dSIlya Dryomov struct ceph_spg_mapping *spg; 4113a02a946dSIlya Dryomov struct ceph_osd_backoff *backoff; 4114a02a946dSIlya Dryomov struct rb_node *n; 4115a02a946dSIlya Dryomov 4116a02a946dSIlya Dryomov dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd, 4117a02a946dSIlya Dryomov m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id); 4118a02a946dSIlya Dryomov 4119a02a946dSIlya Dryomov backoff = lookup_backoff_by_id(&osd->o_backoffs_by_id, m->id); 4120a02a946dSIlya Dryomov if (!backoff) { 4121a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu backoff dne\n", 4122a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4123a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4124a02a946dSIlya Dryomov return; 4125a02a946dSIlya Dryomov } 4126a02a946dSIlya Dryomov 4127a02a946dSIlya Dryomov if (hoid_compare(backoff->begin, m->begin) && 4128a02a946dSIlya Dryomov hoid_compare(backoff->end, m->end)) { 4129a02a946dSIlya Dryomov pr_err("%s osd%d spgid %llu.%xs%d id %llu bad range?\n", 4130a02a946dSIlya Dryomov __func__, osd->o_osd, m->spgid.pgid.pool, 4131a02a946dSIlya Dryomov m->spgid.pgid.seed, m->spgid.shard, m->id); 4132a02a946dSIlya Dryomov /* unblock it anyway... */ 4133a02a946dSIlya Dryomov } 4134a02a946dSIlya Dryomov 4135a02a946dSIlya Dryomov spg = lookup_spg_mapping(&osd->o_backoff_mappings, &backoff->spgid); 4136a02a946dSIlya Dryomov BUG_ON(!spg); 4137a02a946dSIlya Dryomov 4138a02a946dSIlya Dryomov erase_backoff(&spg->backoffs, backoff); 4139a02a946dSIlya Dryomov erase_backoff_by_id(&osd->o_backoffs_by_id, backoff); 4140a02a946dSIlya Dryomov free_backoff(backoff); 4141a02a946dSIlya Dryomov 4142a02a946dSIlya Dryomov if (RB_EMPTY_ROOT(&spg->backoffs)) { 4143a02a946dSIlya Dryomov erase_spg_mapping(&osd->o_backoff_mappings, spg); 4144a02a946dSIlya Dryomov free_spg_mapping(spg); 4145a02a946dSIlya Dryomov } 4146a02a946dSIlya Dryomov 4147a02a946dSIlya Dryomov for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) { 4148a02a946dSIlya Dryomov struct ceph_osd_request *req = 4149a02a946dSIlya Dryomov rb_entry(n, struct ceph_osd_request, r_node); 4150a02a946dSIlya Dryomov 4151a02a946dSIlya Dryomov if (!ceph_spg_compare(&req->r_t.spgid, &m->spgid)) { 4152a02a946dSIlya Dryomov /* 4153a02a946dSIlya Dryomov * Match against @m, not @backoff -- the PG may 4154a02a946dSIlya Dryomov * have split on the OSD. 4155a02a946dSIlya Dryomov */ 4156a02a946dSIlya Dryomov if (target_contained_by(&req->r_t, m->begin, m->end)) { 4157a02a946dSIlya Dryomov /* 4158a02a946dSIlya Dryomov * If no other installed backoff applies, 4159a02a946dSIlya Dryomov * resend. 4160a02a946dSIlya Dryomov */ 4161a02a946dSIlya Dryomov send_request(req); 4162a02a946dSIlya Dryomov } 4163a02a946dSIlya Dryomov } 4164a02a946dSIlya Dryomov } 4165a02a946dSIlya Dryomov } 4166a02a946dSIlya Dryomov 4167a02a946dSIlya Dryomov static void handle_backoff(struct ceph_osd *osd, struct ceph_msg *msg) 4168a02a946dSIlya Dryomov { 4169a02a946dSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 4170a02a946dSIlya Dryomov struct MOSDBackoff m; 4171a02a946dSIlya Dryomov int ret; 4172a02a946dSIlya Dryomov 4173a02a946dSIlya Dryomov down_read(&osdc->lock); 4174a02a946dSIlya Dryomov if (!osd_registered(osd)) { 4175a02a946dSIlya Dryomov dout("%s osd%d unknown\n", __func__, osd->o_osd); 4176a02a946dSIlya Dryomov up_read(&osdc->lock); 4177a02a946dSIlya Dryomov return; 4178a02a946dSIlya Dryomov } 4179a02a946dSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num)); 4180a02a946dSIlya Dryomov 4181a02a946dSIlya Dryomov mutex_lock(&osd->lock); 4182a02a946dSIlya Dryomov ret = decode_MOSDBackoff(msg, &m); 4183a02a946dSIlya Dryomov if (ret) { 4184a02a946dSIlya Dryomov pr_err("failed to decode MOSDBackoff: %d\n", ret); 4185a02a946dSIlya Dryomov ceph_msg_dump(msg); 4186a02a946dSIlya Dryomov goto out_unlock; 4187a02a946dSIlya Dryomov } 4188a02a946dSIlya Dryomov 4189a02a946dSIlya Dryomov switch (m.op) { 4190a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_BLOCK: 4191a02a946dSIlya Dryomov handle_backoff_block(osd, &m); 4192a02a946dSIlya Dryomov break; 4193a02a946dSIlya Dryomov case CEPH_OSD_BACKOFF_OP_UNBLOCK: 4194a02a946dSIlya Dryomov handle_backoff_unblock(osd, &m); 4195a02a946dSIlya Dryomov break; 4196a02a946dSIlya Dryomov default: 4197a02a946dSIlya Dryomov pr_err("%s osd%d unknown op %d\n", __func__, osd->o_osd, m.op); 4198a02a946dSIlya Dryomov } 4199a02a946dSIlya Dryomov 4200a02a946dSIlya Dryomov free_hoid(m.begin); 4201a02a946dSIlya Dryomov free_hoid(m.end); 4202a02a946dSIlya Dryomov 4203a02a946dSIlya Dryomov out_unlock: 4204a02a946dSIlya Dryomov mutex_unlock(&osd->lock); 4205a02a946dSIlya Dryomov up_read(&osdc->lock); 4206a02a946dSIlya Dryomov } 4207a02a946dSIlya Dryomov 42083d14c5d2SYehuda Sadeh /* 4209a40c4f10SYehuda Sadeh * Process osd watch notifications 4210a40c4f10SYehuda Sadeh */ 42113c663bbdSAlex Elder static void handle_watch_notify(struct ceph_osd_client *osdc, 42123c663bbdSAlex Elder struct ceph_msg *msg) 4213a40c4f10SYehuda Sadeh { 4214922dab61SIlya Dryomov void *p = msg->front.iov_base; 4215922dab61SIlya Dryomov void *const end = p + msg->front.iov_len; 4216922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4217922dab61SIlya Dryomov struct linger_work *lwork; 4218922dab61SIlya Dryomov u8 proto_ver, opcode; 4219922dab61SIlya Dryomov u64 cookie, notify_id; 4220922dab61SIlya Dryomov u64 notifier_id = 0; 422119079203SIlya Dryomov s32 return_code = 0; 4222922dab61SIlya Dryomov void *payload = NULL; 4223922dab61SIlya Dryomov u32 payload_len = 0; 4224a40c4f10SYehuda Sadeh 4225a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, proto_ver, bad); 4226a40c4f10SYehuda Sadeh ceph_decode_8_safe(&p, end, opcode, bad); 4227a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, cookie, bad); 4228922dab61SIlya Dryomov p += 8; /* skip ver */ 4229a40c4f10SYehuda Sadeh ceph_decode_64_safe(&p, end, notify_id, bad); 4230a40c4f10SYehuda Sadeh 4231922dab61SIlya Dryomov if (proto_ver >= 1) { 4232922dab61SIlya Dryomov ceph_decode_32_safe(&p, end, payload_len, bad); 4233922dab61SIlya Dryomov ceph_decode_need(&p, end, payload_len, bad); 4234922dab61SIlya Dryomov payload = p; 4235922dab61SIlya Dryomov p += payload_len; 4236a40c4f10SYehuda Sadeh } 4237a40c4f10SYehuda Sadeh 4238922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 2) 423919079203SIlya Dryomov ceph_decode_32_safe(&p, end, return_code, bad); 4240922dab61SIlya Dryomov 4241922dab61SIlya Dryomov if (le16_to_cpu(msg->hdr.version) >= 3) 4242922dab61SIlya Dryomov ceph_decode_64_safe(&p, end, notifier_id, bad); 4243922dab61SIlya Dryomov 4244922dab61SIlya Dryomov down_read(&osdc->lock); 4245922dab61SIlya Dryomov lreq = lookup_linger_osdc(&osdc->linger_requests, cookie); 4246922dab61SIlya Dryomov if (!lreq) { 4247922dab61SIlya Dryomov dout("%s opcode %d cookie %llu dne\n", __func__, opcode, 4248922dab61SIlya Dryomov cookie); 4249922dab61SIlya Dryomov goto out_unlock_osdc; 4250922dab61SIlya Dryomov } 4251922dab61SIlya Dryomov 4252922dab61SIlya Dryomov mutex_lock(&lreq->lock); 425319079203SIlya Dryomov dout("%s opcode %d cookie %llu lreq %p is_watch %d\n", __func__, 425419079203SIlya Dryomov opcode, cookie, lreq, lreq->is_watch); 4255922dab61SIlya Dryomov if (opcode == CEPH_WATCH_EVENT_DISCONNECT) { 4256922dab61SIlya Dryomov if (!lreq->last_error) { 4257922dab61SIlya Dryomov lreq->last_error = -ENOTCONN; 4258922dab61SIlya Dryomov queue_watch_error(lreq); 4259922dab61SIlya Dryomov } 426019079203SIlya Dryomov } else if (!lreq->is_watch) { 426119079203SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY_COMPLETE */ 426219079203SIlya Dryomov if (lreq->notify_id && lreq->notify_id != notify_id) { 426319079203SIlya Dryomov dout("lreq %p notify_id %llu != %llu, ignoring\n", lreq, 426419079203SIlya Dryomov lreq->notify_id, notify_id); 426519079203SIlya Dryomov } else if (!completion_done(&lreq->notify_finish_wait)) { 426619079203SIlya Dryomov struct ceph_msg_data *data = 426719079203SIlya Dryomov list_first_entry_or_null(&msg->data, 426819079203SIlya Dryomov struct ceph_msg_data, 426919079203SIlya Dryomov links); 427019079203SIlya Dryomov 427119079203SIlya Dryomov if (data) { 427219079203SIlya Dryomov if (lreq->preply_pages) { 427319079203SIlya Dryomov WARN_ON(data->type != 427419079203SIlya Dryomov CEPH_MSG_DATA_PAGES); 427519079203SIlya Dryomov *lreq->preply_pages = data->pages; 427619079203SIlya Dryomov *lreq->preply_len = data->length; 427719079203SIlya Dryomov } else { 427819079203SIlya Dryomov ceph_release_page_vector(data->pages, 427919079203SIlya Dryomov calc_pages_for(0, data->length)); 428019079203SIlya Dryomov } 428119079203SIlya Dryomov } 428219079203SIlya Dryomov lreq->notify_finish_error = return_code; 428319079203SIlya Dryomov complete_all(&lreq->notify_finish_wait); 428419079203SIlya Dryomov } 4285922dab61SIlya Dryomov } else { 4286922dab61SIlya Dryomov /* CEPH_WATCH_EVENT_NOTIFY */ 4287922dab61SIlya Dryomov lwork = lwork_alloc(lreq, do_watch_notify); 4288922dab61SIlya Dryomov if (!lwork) { 4289922dab61SIlya Dryomov pr_err("failed to allocate notify-lwork\n"); 4290922dab61SIlya Dryomov goto out_unlock_lreq; 4291922dab61SIlya Dryomov } 4292922dab61SIlya Dryomov 4293922dab61SIlya Dryomov lwork->notify.notify_id = notify_id; 4294922dab61SIlya Dryomov lwork->notify.notifier_id = notifier_id; 4295922dab61SIlya Dryomov lwork->notify.payload = payload; 4296922dab61SIlya Dryomov lwork->notify.payload_len = payload_len; 4297922dab61SIlya Dryomov lwork->notify.msg = ceph_msg_get(msg); 4298922dab61SIlya Dryomov lwork_queue(lwork); 4299922dab61SIlya Dryomov } 4300922dab61SIlya Dryomov 4301922dab61SIlya Dryomov out_unlock_lreq: 4302922dab61SIlya Dryomov mutex_unlock(&lreq->lock); 4303922dab61SIlya Dryomov out_unlock_osdc: 4304922dab61SIlya Dryomov up_read(&osdc->lock); 4305a40c4f10SYehuda Sadeh return; 4306a40c4f10SYehuda Sadeh 4307a40c4f10SYehuda Sadeh bad: 4308a40c4f10SYehuda Sadeh pr_err("osdc handle_watch_notify corrupt msg\n"); 4309a40c4f10SYehuda Sadeh } 4310a40c4f10SYehuda Sadeh 4311a40c4f10SYehuda Sadeh /* 43123d14c5d2SYehuda Sadeh * Register request, send initial attempt. 43133d14c5d2SYehuda Sadeh */ 43143d14c5d2SYehuda Sadeh int ceph_osdc_start_request(struct ceph_osd_client *osdc, 43153d14c5d2SYehuda Sadeh struct ceph_osd_request *req, 43163d14c5d2SYehuda Sadeh bool nofail) 43173d14c5d2SYehuda Sadeh { 43185aea3dcdSIlya Dryomov down_read(&osdc->lock); 43195aea3dcdSIlya Dryomov submit_request(req, false); 43205aea3dcdSIlya Dryomov up_read(&osdc->lock); 43213d14c5d2SYehuda Sadeh 43225aea3dcdSIlya Dryomov return 0; 43233d14c5d2SYehuda Sadeh } 43243d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_start_request); 43253d14c5d2SYehuda Sadeh 43263d14c5d2SYehuda Sadeh /* 4327c297eb42SIlya Dryomov * Unregister a registered request. The request is not completed: 4328c297eb42SIlya Dryomov * ->r_result isn't set and __complete_request() isn't called. 4329c9f9b93dSIlya Dryomov */ 4330c9f9b93dSIlya Dryomov void ceph_osdc_cancel_request(struct ceph_osd_request *req) 4331c9f9b93dSIlya Dryomov { 4332c9f9b93dSIlya Dryomov struct ceph_osd_client *osdc = req->r_osdc; 4333c9f9b93dSIlya Dryomov 43345aea3dcdSIlya Dryomov down_write(&osdc->lock); 43355aea3dcdSIlya Dryomov if (req->r_osd) 43365aea3dcdSIlya Dryomov cancel_request(req); 43375aea3dcdSIlya Dryomov up_write(&osdc->lock); 4338c9f9b93dSIlya Dryomov } 4339c9f9b93dSIlya Dryomov EXPORT_SYMBOL(ceph_osdc_cancel_request); 4340c9f9b93dSIlya Dryomov 4341c9f9b93dSIlya Dryomov /* 434242b06965SIlya Dryomov * @timeout: in jiffies, 0 means "wait forever" 434342b06965SIlya Dryomov */ 434442b06965SIlya Dryomov static int wait_request_timeout(struct ceph_osd_request *req, 434542b06965SIlya Dryomov unsigned long timeout) 434642b06965SIlya Dryomov { 434742b06965SIlya Dryomov long left; 434842b06965SIlya Dryomov 434942b06965SIlya Dryomov dout("%s req %p tid %llu\n", __func__, req, req->r_tid); 43500e76abf2SYan, Zheng left = wait_for_completion_killable_timeout(&req->r_completion, 435142b06965SIlya Dryomov ceph_timeout_jiffies(timeout)); 435242b06965SIlya Dryomov if (left <= 0) { 435342b06965SIlya Dryomov left = left ?: -ETIMEDOUT; 435442b06965SIlya Dryomov ceph_osdc_cancel_request(req); 435542b06965SIlya Dryomov } else { 435642b06965SIlya Dryomov left = req->r_result; /* completed */ 435742b06965SIlya Dryomov } 435842b06965SIlya Dryomov 435942b06965SIlya Dryomov return left; 436042b06965SIlya Dryomov } 436142b06965SIlya Dryomov 436242b06965SIlya Dryomov /* 43633d14c5d2SYehuda Sadeh * wait for a request to complete 43643d14c5d2SYehuda Sadeh */ 43653d14c5d2SYehuda Sadeh int ceph_osdc_wait_request(struct ceph_osd_client *osdc, 43663d14c5d2SYehuda Sadeh struct ceph_osd_request *req) 43673d14c5d2SYehuda Sadeh { 436842b06965SIlya Dryomov return wait_request_timeout(req, 0); 43693d14c5d2SYehuda Sadeh } 43703d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_wait_request); 43713d14c5d2SYehuda Sadeh 43723d14c5d2SYehuda Sadeh /* 43733d14c5d2SYehuda Sadeh * sync - wait for all in-flight requests to flush. avoid starvation. 43743d14c5d2SYehuda Sadeh */ 43753d14c5d2SYehuda Sadeh void ceph_osdc_sync(struct ceph_osd_client *osdc) 43763d14c5d2SYehuda Sadeh { 43775aea3dcdSIlya Dryomov struct rb_node *n, *p; 43785aea3dcdSIlya Dryomov u64 last_tid = atomic64_read(&osdc->last_tid); 43793d14c5d2SYehuda Sadeh 43805aea3dcdSIlya Dryomov again: 43815aea3dcdSIlya Dryomov down_read(&osdc->lock); 43825aea3dcdSIlya Dryomov for (n = rb_first(&osdc->osds); n; n = rb_next(n)) { 43835aea3dcdSIlya Dryomov struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node); 43845aea3dcdSIlya Dryomov 43855aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 43865aea3dcdSIlya Dryomov for (p = rb_first(&osd->o_requests); p; p = rb_next(p)) { 43875aea3dcdSIlya Dryomov struct ceph_osd_request *req = 43885aea3dcdSIlya Dryomov rb_entry(p, struct ceph_osd_request, r_node); 43895aea3dcdSIlya Dryomov 43903d14c5d2SYehuda Sadeh if (req->r_tid > last_tid) 43913d14c5d2SYehuda Sadeh break; 43923d14c5d2SYehuda Sadeh 43935aea3dcdSIlya Dryomov if (!(req->r_flags & CEPH_OSD_FLAG_WRITE)) 43943d14c5d2SYehuda Sadeh continue; 43953d14c5d2SYehuda Sadeh 43963d14c5d2SYehuda Sadeh ceph_osdc_get_request(req); 43975aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 43985aea3dcdSIlya Dryomov up_read(&osdc->lock); 43995aea3dcdSIlya Dryomov dout("%s waiting on req %p tid %llu last_tid %llu\n", 44005aea3dcdSIlya Dryomov __func__, req, req->r_tid, last_tid); 4401b18b9550SIlya Dryomov wait_for_completion(&req->r_completion); 44023d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 44035aea3dcdSIlya Dryomov goto again; 44043d14c5d2SYehuda Sadeh } 44055aea3dcdSIlya Dryomov 44065aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 44075aea3dcdSIlya Dryomov } 44085aea3dcdSIlya Dryomov 44095aea3dcdSIlya Dryomov up_read(&osdc->lock); 44105aea3dcdSIlya Dryomov dout("%s done last_tid %llu\n", __func__, last_tid); 44113d14c5d2SYehuda Sadeh } 44123d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_sync); 44133d14c5d2SYehuda Sadeh 4414922dab61SIlya Dryomov static struct ceph_osd_request * 4415922dab61SIlya Dryomov alloc_linger_request(struct ceph_osd_linger_request *lreq) 4416922dab61SIlya Dryomov { 4417922dab61SIlya Dryomov struct ceph_osd_request *req; 4418922dab61SIlya Dryomov 4419922dab61SIlya Dryomov req = ceph_osdc_alloc_request(lreq->osdc, NULL, 1, false, GFP_NOIO); 4420922dab61SIlya Dryomov if (!req) 4421922dab61SIlya Dryomov return NULL; 4422922dab61SIlya Dryomov 4423922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4424922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 4425922dab61SIlya Dryomov 4426922dab61SIlya Dryomov if (ceph_osdc_alloc_messages(req, GFP_NOIO)) { 4427922dab61SIlya Dryomov ceph_osdc_put_request(req); 4428922dab61SIlya Dryomov return NULL; 4429922dab61SIlya Dryomov } 4430922dab61SIlya Dryomov 4431922dab61SIlya Dryomov return req; 4432922dab61SIlya Dryomov } 4433922dab61SIlya Dryomov 4434922dab61SIlya Dryomov /* 4435922dab61SIlya Dryomov * Returns a handle, caller owns a ref. 4436922dab61SIlya Dryomov */ 4437922dab61SIlya Dryomov struct ceph_osd_linger_request * 4438922dab61SIlya Dryomov ceph_osdc_watch(struct ceph_osd_client *osdc, 4439922dab61SIlya Dryomov struct ceph_object_id *oid, 4440922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4441922dab61SIlya Dryomov rados_watchcb2_t wcb, 4442922dab61SIlya Dryomov rados_watcherrcb_t errcb, 4443922dab61SIlya Dryomov void *data) 4444922dab61SIlya Dryomov { 4445922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq; 4446922dab61SIlya Dryomov int ret; 4447922dab61SIlya Dryomov 4448922dab61SIlya Dryomov lreq = linger_alloc(osdc); 4449922dab61SIlya Dryomov if (!lreq) 4450922dab61SIlya Dryomov return ERR_PTR(-ENOMEM); 4451922dab61SIlya Dryomov 445219079203SIlya Dryomov lreq->is_watch = true; 4453922dab61SIlya Dryomov lreq->wcb = wcb; 4454922dab61SIlya Dryomov lreq->errcb = errcb; 4455922dab61SIlya Dryomov lreq->data = data; 4456b07d3c4bSIlya Dryomov lreq->watch_valid_thru = jiffies; 4457922dab61SIlya Dryomov 4458922dab61SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 4459922dab61SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 446054ea0046SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_WRITE; 44611134e091SDeepa Dinamani ktime_get_real_ts(&lreq->mtime); 4462922dab61SIlya Dryomov 4463922dab61SIlya Dryomov lreq->reg_req = alloc_linger_request(lreq); 4464922dab61SIlya Dryomov if (!lreq->reg_req) { 4465922dab61SIlya Dryomov ret = -ENOMEM; 4466922dab61SIlya Dryomov goto err_put_lreq; 4467922dab61SIlya Dryomov } 4468922dab61SIlya Dryomov 4469922dab61SIlya Dryomov lreq->ping_req = alloc_linger_request(lreq); 4470922dab61SIlya Dryomov if (!lreq->ping_req) { 4471922dab61SIlya Dryomov ret = -ENOMEM; 4472922dab61SIlya Dryomov goto err_put_lreq; 4473922dab61SIlya Dryomov } 4474922dab61SIlya Dryomov 4475922dab61SIlya Dryomov down_write(&osdc->lock); 4476922dab61SIlya Dryomov linger_register(lreq); /* before osd_req_op_* */ 4477922dab61SIlya Dryomov osd_req_op_watch_init(lreq->reg_req, 0, lreq->linger_id, 4478922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_WATCH); 4479922dab61SIlya Dryomov osd_req_op_watch_init(lreq->ping_req, 0, lreq->linger_id, 4480922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_PING); 4481922dab61SIlya Dryomov linger_submit(lreq); 4482922dab61SIlya Dryomov up_write(&osdc->lock); 4483922dab61SIlya Dryomov 4484922dab61SIlya Dryomov ret = linger_reg_commit_wait(lreq); 4485922dab61SIlya Dryomov if (ret) { 4486922dab61SIlya Dryomov linger_cancel(lreq); 4487922dab61SIlya Dryomov goto err_put_lreq; 4488922dab61SIlya Dryomov } 4489922dab61SIlya Dryomov 4490922dab61SIlya Dryomov return lreq; 4491922dab61SIlya Dryomov 4492922dab61SIlya Dryomov err_put_lreq: 4493922dab61SIlya Dryomov linger_put(lreq); 4494922dab61SIlya Dryomov return ERR_PTR(ret); 4495922dab61SIlya Dryomov } 4496922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_watch); 4497922dab61SIlya Dryomov 4498922dab61SIlya Dryomov /* 4499922dab61SIlya Dryomov * Releases a ref. 4500922dab61SIlya Dryomov * 4501922dab61SIlya Dryomov * Times out after mount_timeout to preserve rbd unmap behaviour 4502922dab61SIlya Dryomov * introduced in 2894e1d76974 ("rbd: timeout watch teardown on unmap 4503922dab61SIlya Dryomov * with mount_timeout"). 4504922dab61SIlya Dryomov */ 4505922dab61SIlya Dryomov int ceph_osdc_unwatch(struct ceph_osd_client *osdc, 4506922dab61SIlya Dryomov struct ceph_osd_linger_request *lreq) 4507922dab61SIlya Dryomov { 4508922dab61SIlya Dryomov struct ceph_options *opts = osdc->client->options; 4509922dab61SIlya Dryomov struct ceph_osd_request *req; 4510922dab61SIlya Dryomov int ret; 4511922dab61SIlya Dryomov 4512922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4513922dab61SIlya Dryomov if (!req) 4514922dab61SIlya Dryomov return -ENOMEM; 4515922dab61SIlya Dryomov 4516922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid); 4517922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc); 451854ea0046SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_WRITE; 45191134e091SDeepa Dinamani ktime_get_real_ts(&req->r_mtime); 4520922dab61SIlya Dryomov osd_req_op_watch_init(req, 0, lreq->linger_id, 4521922dab61SIlya Dryomov CEPH_OSD_WATCH_OP_UNWATCH); 4522922dab61SIlya Dryomov 4523922dab61SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4524922dab61SIlya Dryomov if (ret) 4525922dab61SIlya Dryomov goto out_put_req; 4526922dab61SIlya Dryomov 4527922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4528922dab61SIlya Dryomov linger_cancel(lreq); 4529922dab61SIlya Dryomov linger_put(lreq); 4530922dab61SIlya Dryomov ret = wait_request_timeout(req, opts->mount_timeout); 4531922dab61SIlya Dryomov 4532922dab61SIlya Dryomov out_put_req: 4533922dab61SIlya Dryomov ceph_osdc_put_request(req); 4534922dab61SIlya Dryomov return ret; 4535922dab61SIlya Dryomov } 4536922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_unwatch); 4537922dab61SIlya Dryomov 4538922dab61SIlya Dryomov static int osd_req_op_notify_ack_init(struct ceph_osd_request *req, int which, 4539922dab61SIlya Dryomov u64 notify_id, u64 cookie, void *payload, 4540922dab61SIlya Dryomov size_t payload_len) 4541922dab61SIlya Dryomov { 4542922dab61SIlya Dryomov struct ceph_osd_req_op *op; 4543922dab61SIlya Dryomov struct ceph_pagelist *pl; 4544922dab61SIlya Dryomov int ret; 4545922dab61SIlya Dryomov 4546922dab61SIlya Dryomov op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY_ACK, 0); 4547922dab61SIlya Dryomov 4548922dab61SIlya Dryomov pl = kmalloc(sizeof(*pl), GFP_NOIO); 4549922dab61SIlya Dryomov if (!pl) 4550922dab61SIlya Dryomov return -ENOMEM; 4551922dab61SIlya Dryomov 4552922dab61SIlya Dryomov ceph_pagelist_init(pl); 4553922dab61SIlya Dryomov ret = ceph_pagelist_encode_64(pl, notify_id); 4554922dab61SIlya Dryomov ret |= ceph_pagelist_encode_64(pl, cookie); 4555922dab61SIlya Dryomov if (payload) { 4556922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 4557922dab61SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 4558922dab61SIlya Dryomov } else { 4559922dab61SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, 0); 4560922dab61SIlya Dryomov } 4561922dab61SIlya Dryomov if (ret) { 4562922dab61SIlya Dryomov ceph_pagelist_release(pl); 4563922dab61SIlya Dryomov return -ENOMEM; 4564922dab61SIlya Dryomov } 4565922dab61SIlya Dryomov 4566922dab61SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify_ack.request_data, pl); 4567922dab61SIlya Dryomov op->indata_len = pl->length; 4568922dab61SIlya Dryomov return 0; 4569922dab61SIlya Dryomov } 4570922dab61SIlya Dryomov 4571922dab61SIlya Dryomov int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, 4572922dab61SIlya Dryomov struct ceph_object_id *oid, 4573922dab61SIlya Dryomov struct ceph_object_locator *oloc, 4574922dab61SIlya Dryomov u64 notify_id, 4575922dab61SIlya Dryomov u64 cookie, 4576922dab61SIlya Dryomov void *payload, 4577922dab61SIlya Dryomov size_t payload_len) 4578922dab61SIlya Dryomov { 4579922dab61SIlya Dryomov struct ceph_osd_request *req; 4580922dab61SIlya Dryomov int ret; 4581922dab61SIlya Dryomov 4582922dab61SIlya Dryomov req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4583922dab61SIlya Dryomov if (!req) 4584922dab61SIlya Dryomov return -ENOMEM; 4585922dab61SIlya Dryomov 4586922dab61SIlya Dryomov ceph_oid_copy(&req->r_base_oid, oid); 4587922dab61SIlya Dryomov ceph_oloc_copy(&req->r_base_oloc, oloc); 4588922dab61SIlya Dryomov req->r_flags = CEPH_OSD_FLAG_READ; 4589922dab61SIlya Dryomov 4590922dab61SIlya Dryomov ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4591922dab61SIlya Dryomov if (ret) 4592922dab61SIlya Dryomov goto out_put_req; 4593922dab61SIlya Dryomov 4594922dab61SIlya Dryomov ret = osd_req_op_notify_ack_init(req, 0, notify_id, cookie, payload, 4595922dab61SIlya Dryomov payload_len); 4596922dab61SIlya Dryomov if (ret) 4597922dab61SIlya Dryomov goto out_put_req; 4598922dab61SIlya Dryomov 4599922dab61SIlya Dryomov ceph_osdc_start_request(osdc, req, false); 4600922dab61SIlya Dryomov ret = ceph_osdc_wait_request(osdc, req); 4601922dab61SIlya Dryomov 4602922dab61SIlya Dryomov out_put_req: 4603922dab61SIlya Dryomov ceph_osdc_put_request(req); 4604922dab61SIlya Dryomov return ret; 4605922dab61SIlya Dryomov } 4606922dab61SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify_ack); 4607922dab61SIlya Dryomov 460819079203SIlya Dryomov static int osd_req_op_notify_init(struct ceph_osd_request *req, int which, 460919079203SIlya Dryomov u64 cookie, u32 prot_ver, u32 timeout, 461019079203SIlya Dryomov void *payload, size_t payload_len) 461119079203SIlya Dryomov { 461219079203SIlya Dryomov struct ceph_osd_req_op *op; 461319079203SIlya Dryomov struct ceph_pagelist *pl; 461419079203SIlya Dryomov int ret; 461519079203SIlya Dryomov 461619079203SIlya Dryomov op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY, 0); 461719079203SIlya Dryomov op->notify.cookie = cookie; 461819079203SIlya Dryomov 461919079203SIlya Dryomov pl = kmalloc(sizeof(*pl), GFP_NOIO); 462019079203SIlya Dryomov if (!pl) 462119079203SIlya Dryomov return -ENOMEM; 462219079203SIlya Dryomov 462319079203SIlya Dryomov ceph_pagelist_init(pl); 462419079203SIlya Dryomov ret = ceph_pagelist_encode_32(pl, 1); /* prot_ver */ 462519079203SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, timeout); 462619079203SIlya Dryomov ret |= ceph_pagelist_encode_32(pl, payload_len); 462719079203SIlya Dryomov ret |= ceph_pagelist_append(pl, payload, payload_len); 462819079203SIlya Dryomov if (ret) { 462919079203SIlya Dryomov ceph_pagelist_release(pl); 463019079203SIlya Dryomov return -ENOMEM; 463119079203SIlya Dryomov } 463219079203SIlya Dryomov 463319079203SIlya Dryomov ceph_osd_data_pagelist_init(&op->notify.request_data, pl); 463419079203SIlya Dryomov op->indata_len = pl->length; 463519079203SIlya Dryomov return 0; 463619079203SIlya Dryomov } 463719079203SIlya Dryomov 463819079203SIlya Dryomov /* 463919079203SIlya Dryomov * @timeout: in seconds 464019079203SIlya Dryomov * 464119079203SIlya Dryomov * @preply_{pages,len} are initialized both on success and error. 464219079203SIlya Dryomov * The caller is responsible for: 464319079203SIlya Dryomov * 464419079203SIlya Dryomov * ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len)) 464519079203SIlya Dryomov */ 464619079203SIlya Dryomov int ceph_osdc_notify(struct ceph_osd_client *osdc, 464719079203SIlya Dryomov struct ceph_object_id *oid, 464819079203SIlya Dryomov struct ceph_object_locator *oloc, 464919079203SIlya Dryomov void *payload, 465019079203SIlya Dryomov size_t payload_len, 465119079203SIlya Dryomov u32 timeout, 465219079203SIlya Dryomov struct page ***preply_pages, 465319079203SIlya Dryomov size_t *preply_len) 465419079203SIlya Dryomov { 465519079203SIlya Dryomov struct ceph_osd_linger_request *lreq; 465619079203SIlya Dryomov struct page **pages; 465719079203SIlya Dryomov int ret; 465819079203SIlya Dryomov 465919079203SIlya Dryomov WARN_ON(!timeout); 466019079203SIlya Dryomov if (preply_pages) { 466119079203SIlya Dryomov *preply_pages = NULL; 466219079203SIlya Dryomov *preply_len = 0; 466319079203SIlya Dryomov } 466419079203SIlya Dryomov 466519079203SIlya Dryomov lreq = linger_alloc(osdc); 466619079203SIlya Dryomov if (!lreq) 466719079203SIlya Dryomov return -ENOMEM; 466819079203SIlya Dryomov 466919079203SIlya Dryomov lreq->preply_pages = preply_pages; 467019079203SIlya Dryomov lreq->preply_len = preply_len; 467119079203SIlya Dryomov 467219079203SIlya Dryomov ceph_oid_copy(&lreq->t.base_oid, oid); 467319079203SIlya Dryomov ceph_oloc_copy(&lreq->t.base_oloc, oloc); 467419079203SIlya Dryomov lreq->t.flags = CEPH_OSD_FLAG_READ; 467519079203SIlya Dryomov 467619079203SIlya Dryomov lreq->reg_req = alloc_linger_request(lreq); 467719079203SIlya Dryomov if (!lreq->reg_req) { 467819079203SIlya Dryomov ret = -ENOMEM; 467919079203SIlya Dryomov goto out_put_lreq; 468019079203SIlya Dryomov } 468119079203SIlya Dryomov 468219079203SIlya Dryomov /* for notify_id */ 468319079203SIlya Dryomov pages = ceph_alloc_page_vector(1, GFP_NOIO); 468419079203SIlya Dryomov if (IS_ERR(pages)) { 468519079203SIlya Dryomov ret = PTR_ERR(pages); 468619079203SIlya Dryomov goto out_put_lreq; 468719079203SIlya Dryomov } 468819079203SIlya Dryomov 468919079203SIlya Dryomov down_write(&osdc->lock); 469019079203SIlya Dryomov linger_register(lreq); /* before osd_req_op_* */ 469119079203SIlya Dryomov ret = osd_req_op_notify_init(lreq->reg_req, 0, lreq->linger_id, 1, 469219079203SIlya Dryomov timeout, payload, payload_len); 469319079203SIlya Dryomov if (ret) { 469419079203SIlya Dryomov linger_unregister(lreq); 469519079203SIlya Dryomov up_write(&osdc->lock); 469619079203SIlya Dryomov ceph_release_page_vector(pages, 1); 469719079203SIlya Dryomov goto out_put_lreq; 469819079203SIlya Dryomov } 469919079203SIlya Dryomov ceph_osd_data_pages_init(osd_req_op_data(lreq->reg_req, 0, notify, 470019079203SIlya Dryomov response_data), 470119079203SIlya Dryomov pages, PAGE_SIZE, 0, false, true); 470219079203SIlya Dryomov linger_submit(lreq); 470319079203SIlya Dryomov up_write(&osdc->lock); 470419079203SIlya Dryomov 470519079203SIlya Dryomov ret = linger_reg_commit_wait(lreq); 470619079203SIlya Dryomov if (!ret) 470719079203SIlya Dryomov ret = linger_notify_finish_wait(lreq); 470819079203SIlya Dryomov else 470919079203SIlya Dryomov dout("lreq %p failed to initiate notify %d\n", lreq, ret); 471019079203SIlya Dryomov 471119079203SIlya Dryomov linger_cancel(lreq); 471219079203SIlya Dryomov out_put_lreq: 471319079203SIlya Dryomov linger_put(lreq); 471419079203SIlya Dryomov return ret; 471519079203SIlya Dryomov } 471619079203SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_notify); 471719079203SIlya Dryomov 47183d14c5d2SYehuda Sadeh /* 4719b07d3c4bSIlya Dryomov * Return the number of milliseconds since the watch was last 4720b07d3c4bSIlya Dryomov * confirmed, or an error. If there is an error, the watch is no 4721b07d3c4bSIlya Dryomov * longer valid, and should be destroyed with ceph_osdc_unwatch(). 4722b07d3c4bSIlya Dryomov */ 4723b07d3c4bSIlya Dryomov int ceph_osdc_watch_check(struct ceph_osd_client *osdc, 4724b07d3c4bSIlya Dryomov struct ceph_osd_linger_request *lreq) 4725b07d3c4bSIlya Dryomov { 4726b07d3c4bSIlya Dryomov unsigned long stamp, age; 4727b07d3c4bSIlya Dryomov int ret; 4728b07d3c4bSIlya Dryomov 4729b07d3c4bSIlya Dryomov down_read(&osdc->lock); 4730b07d3c4bSIlya Dryomov mutex_lock(&lreq->lock); 4731b07d3c4bSIlya Dryomov stamp = lreq->watch_valid_thru; 4732b07d3c4bSIlya Dryomov if (!list_empty(&lreq->pending_lworks)) { 4733b07d3c4bSIlya Dryomov struct linger_work *lwork = 4734b07d3c4bSIlya Dryomov list_first_entry(&lreq->pending_lworks, 4735b07d3c4bSIlya Dryomov struct linger_work, 4736b07d3c4bSIlya Dryomov pending_item); 4737b07d3c4bSIlya Dryomov 4738b07d3c4bSIlya Dryomov if (time_before(lwork->queued_stamp, stamp)) 4739b07d3c4bSIlya Dryomov stamp = lwork->queued_stamp; 4740b07d3c4bSIlya Dryomov } 4741b07d3c4bSIlya Dryomov age = jiffies - stamp; 4742b07d3c4bSIlya Dryomov dout("%s lreq %p linger_id %llu age %lu last_error %d\n", __func__, 4743b07d3c4bSIlya Dryomov lreq, lreq->linger_id, age, lreq->last_error); 4744b07d3c4bSIlya Dryomov /* we are truncating to msecs, so return a safe upper bound */ 4745b07d3c4bSIlya Dryomov ret = lreq->last_error ?: 1 + jiffies_to_msecs(age); 4746b07d3c4bSIlya Dryomov 4747b07d3c4bSIlya Dryomov mutex_unlock(&lreq->lock); 4748b07d3c4bSIlya Dryomov up_read(&osdc->lock); 4749b07d3c4bSIlya Dryomov return ret; 4750b07d3c4bSIlya Dryomov } 4751b07d3c4bSIlya Dryomov 4752a4ed38d7SDouglas Fuller static int decode_watcher(void **p, void *end, struct ceph_watch_item *item) 4753a4ed38d7SDouglas Fuller { 4754a4ed38d7SDouglas Fuller u8 struct_v; 4755a4ed38d7SDouglas Fuller u32 struct_len; 4756a4ed38d7SDouglas Fuller int ret; 4757a4ed38d7SDouglas Fuller 4758a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 2, "watch_item_t", 4759a4ed38d7SDouglas Fuller &struct_v, &struct_len); 4760a4ed38d7SDouglas Fuller if (ret) 4761a4ed38d7SDouglas Fuller return ret; 4762a4ed38d7SDouglas Fuller 4763a4ed38d7SDouglas Fuller ceph_decode_copy(p, &item->name, sizeof(item->name)); 4764a4ed38d7SDouglas Fuller item->cookie = ceph_decode_64(p); 4765a4ed38d7SDouglas Fuller *p += 4; /* skip timeout_seconds */ 4766a4ed38d7SDouglas Fuller if (struct_v >= 2) { 4767a4ed38d7SDouglas Fuller ceph_decode_copy(p, &item->addr, sizeof(item->addr)); 4768a4ed38d7SDouglas Fuller ceph_decode_addr(&item->addr); 4769a4ed38d7SDouglas Fuller } 4770a4ed38d7SDouglas Fuller 4771a4ed38d7SDouglas Fuller dout("%s %s%llu cookie %llu addr %s\n", __func__, 4772a4ed38d7SDouglas Fuller ENTITY_NAME(item->name), item->cookie, 4773a4ed38d7SDouglas Fuller ceph_pr_addr(&item->addr.in_addr)); 4774a4ed38d7SDouglas Fuller return 0; 4775a4ed38d7SDouglas Fuller } 4776a4ed38d7SDouglas Fuller 4777a4ed38d7SDouglas Fuller static int decode_watchers(void **p, void *end, 4778a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 4779a4ed38d7SDouglas Fuller u32 *num_watchers) 4780a4ed38d7SDouglas Fuller { 4781a4ed38d7SDouglas Fuller u8 struct_v; 4782a4ed38d7SDouglas Fuller u32 struct_len; 4783a4ed38d7SDouglas Fuller int i; 4784a4ed38d7SDouglas Fuller int ret; 4785a4ed38d7SDouglas Fuller 4786a4ed38d7SDouglas Fuller ret = ceph_start_decoding(p, end, 1, "obj_list_watch_response_t", 4787a4ed38d7SDouglas Fuller &struct_v, &struct_len); 4788a4ed38d7SDouglas Fuller if (ret) 4789a4ed38d7SDouglas Fuller return ret; 4790a4ed38d7SDouglas Fuller 4791a4ed38d7SDouglas Fuller *num_watchers = ceph_decode_32(p); 4792a4ed38d7SDouglas Fuller *watchers = kcalloc(*num_watchers, sizeof(**watchers), GFP_NOIO); 4793a4ed38d7SDouglas Fuller if (!*watchers) 4794a4ed38d7SDouglas Fuller return -ENOMEM; 4795a4ed38d7SDouglas Fuller 4796a4ed38d7SDouglas Fuller for (i = 0; i < *num_watchers; i++) { 4797a4ed38d7SDouglas Fuller ret = decode_watcher(p, end, *watchers + i); 4798a4ed38d7SDouglas Fuller if (ret) { 4799a4ed38d7SDouglas Fuller kfree(*watchers); 4800a4ed38d7SDouglas Fuller return ret; 4801a4ed38d7SDouglas Fuller } 4802a4ed38d7SDouglas Fuller } 4803a4ed38d7SDouglas Fuller 4804a4ed38d7SDouglas Fuller return 0; 4805a4ed38d7SDouglas Fuller } 4806a4ed38d7SDouglas Fuller 4807a4ed38d7SDouglas Fuller /* 4808a4ed38d7SDouglas Fuller * On success, the caller is responsible for: 4809a4ed38d7SDouglas Fuller * 4810a4ed38d7SDouglas Fuller * kfree(watchers); 4811a4ed38d7SDouglas Fuller */ 4812a4ed38d7SDouglas Fuller int ceph_osdc_list_watchers(struct ceph_osd_client *osdc, 4813a4ed38d7SDouglas Fuller struct ceph_object_id *oid, 4814a4ed38d7SDouglas Fuller struct ceph_object_locator *oloc, 4815a4ed38d7SDouglas Fuller struct ceph_watch_item **watchers, 4816a4ed38d7SDouglas Fuller u32 *num_watchers) 4817a4ed38d7SDouglas Fuller { 4818a4ed38d7SDouglas Fuller struct ceph_osd_request *req; 4819a4ed38d7SDouglas Fuller struct page **pages; 4820a4ed38d7SDouglas Fuller int ret; 4821a4ed38d7SDouglas Fuller 4822a4ed38d7SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4823a4ed38d7SDouglas Fuller if (!req) 4824a4ed38d7SDouglas Fuller return -ENOMEM; 4825a4ed38d7SDouglas Fuller 4826a4ed38d7SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 4827a4ed38d7SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 4828a4ed38d7SDouglas Fuller req->r_flags = CEPH_OSD_FLAG_READ; 4829a4ed38d7SDouglas Fuller 4830a4ed38d7SDouglas Fuller ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4831a4ed38d7SDouglas Fuller if (ret) 4832a4ed38d7SDouglas Fuller goto out_put_req; 4833a4ed38d7SDouglas Fuller 4834a4ed38d7SDouglas Fuller pages = ceph_alloc_page_vector(1, GFP_NOIO); 4835a4ed38d7SDouglas Fuller if (IS_ERR(pages)) { 4836a4ed38d7SDouglas Fuller ret = PTR_ERR(pages); 4837a4ed38d7SDouglas Fuller goto out_put_req; 4838a4ed38d7SDouglas Fuller } 4839a4ed38d7SDouglas Fuller 4840a4ed38d7SDouglas Fuller osd_req_op_init(req, 0, CEPH_OSD_OP_LIST_WATCHERS, 0); 4841a4ed38d7SDouglas Fuller ceph_osd_data_pages_init(osd_req_op_data(req, 0, list_watchers, 4842a4ed38d7SDouglas Fuller response_data), 4843a4ed38d7SDouglas Fuller pages, PAGE_SIZE, 0, false, true); 4844a4ed38d7SDouglas Fuller 4845a4ed38d7SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 4846a4ed38d7SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 4847a4ed38d7SDouglas Fuller if (ret >= 0) { 4848a4ed38d7SDouglas Fuller void *p = page_address(pages[0]); 4849a4ed38d7SDouglas Fuller void *const end = p + req->r_ops[0].outdata_len; 4850a4ed38d7SDouglas Fuller 4851a4ed38d7SDouglas Fuller ret = decode_watchers(&p, end, watchers, num_watchers); 4852a4ed38d7SDouglas Fuller } 4853a4ed38d7SDouglas Fuller 4854a4ed38d7SDouglas Fuller out_put_req: 4855a4ed38d7SDouglas Fuller ceph_osdc_put_request(req); 4856a4ed38d7SDouglas Fuller return ret; 4857a4ed38d7SDouglas Fuller } 4858a4ed38d7SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_list_watchers); 4859a4ed38d7SDouglas Fuller 4860b07d3c4bSIlya Dryomov /* 4861dd935f44SJosh Durgin * Call all pending notify callbacks - for use after a watch is 4862dd935f44SJosh Durgin * unregistered, to make sure no more callbacks for it will be invoked 4863dd935f44SJosh Durgin */ 4864f6479449Sstephen hemminger void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc) 4865dd935f44SJosh Durgin { 486699d16943SIlya Dryomov dout("%s osdc %p\n", __func__, osdc); 4867dd935f44SJosh Durgin flush_workqueue(osdc->notify_wq); 4868dd935f44SJosh Durgin } 4869dd935f44SJosh Durgin EXPORT_SYMBOL(ceph_osdc_flush_notifies); 4870dd935f44SJosh Durgin 48717cca78c9SIlya Dryomov void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc) 48727cca78c9SIlya Dryomov { 48737cca78c9SIlya Dryomov down_read(&osdc->lock); 48747cca78c9SIlya Dryomov maybe_request_map(osdc); 48757cca78c9SIlya Dryomov up_read(&osdc->lock); 48767cca78c9SIlya Dryomov } 48777cca78c9SIlya Dryomov EXPORT_SYMBOL(ceph_osdc_maybe_request_map); 4878dd935f44SJosh Durgin 4879dd935f44SJosh Durgin /* 4880428a7158SDouglas Fuller * Execute an OSD class method on an object. 4881428a7158SDouglas Fuller * 4882428a7158SDouglas Fuller * @flags: CEPH_OSD_FLAG_* 48832544a020SIlya Dryomov * @resp_len: in/out param for reply length 4884428a7158SDouglas Fuller */ 4885428a7158SDouglas Fuller int ceph_osdc_call(struct ceph_osd_client *osdc, 4886428a7158SDouglas Fuller struct ceph_object_id *oid, 4887428a7158SDouglas Fuller struct ceph_object_locator *oloc, 4888428a7158SDouglas Fuller const char *class, const char *method, 4889428a7158SDouglas Fuller unsigned int flags, 4890428a7158SDouglas Fuller struct page *req_page, size_t req_len, 4891428a7158SDouglas Fuller struct page *resp_page, size_t *resp_len) 4892428a7158SDouglas Fuller { 4893428a7158SDouglas Fuller struct ceph_osd_request *req; 4894428a7158SDouglas Fuller int ret; 4895428a7158SDouglas Fuller 48962544a020SIlya Dryomov if (req_len > PAGE_SIZE || (resp_page && *resp_len > PAGE_SIZE)) 48972544a020SIlya Dryomov return -E2BIG; 48982544a020SIlya Dryomov 4899428a7158SDouglas Fuller req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO); 4900428a7158SDouglas Fuller if (!req) 4901428a7158SDouglas Fuller return -ENOMEM; 4902428a7158SDouglas Fuller 4903428a7158SDouglas Fuller ceph_oid_copy(&req->r_base_oid, oid); 4904428a7158SDouglas Fuller ceph_oloc_copy(&req->r_base_oloc, oloc); 4905428a7158SDouglas Fuller req->r_flags = flags; 4906428a7158SDouglas Fuller 4907428a7158SDouglas Fuller ret = ceph_osdc_alloc_messages(req, GFP_NOIO); 4908428a7158SDouglas Fuller if (ret) 4909428a7158SDouglas Fuller goto out_put_req; 4910428a7158SDouglas Fuller 4911428a7158SDouglas Fuller osd_req_op_cls_init(req, 0, CEPH_OSD_OP_CALL, class, method); 4912428a7158SDouglas Fuller if (req_page) 4913428a7158SDouglas Fuller osd_req_op_cls_request_data_pages(req, 0, &req_page, req_len, 4914428a7158SDouglas Fuller 0, false, false); 4915428a7158SDouglas Fuller if (resp_page) 4916428a7158SDouglas Fuller osd_req_op_cls_response_data_pages(req, 0, &resp_page, 49172544a020SIlya Dryomov *resp_len, 0, false, false); 4918428a7158SDouglas Fuller 4919428a7158SDouglas Fuller ceph_osdc_start_request(osdc, req, false); 4920428a7158SDouglas Fuller ret = ceph_osdc_wait_request(osdc, req); 4921428a7158SDouglas Fuller if (ret >= 0) { 4922428a7158SDouglas Fuller ret = req->r_ops[0].rval; 4923428a7158SDouglas Fuller if (resp_page) 4924428a7158SDouglas Fuller *resp_len = req->r_ops[0].outdata_len; 4925428a7158SDouglas Fuller } 4926428a7158SDouglas Fuller 4927428a7158SDouglas Fuller out_put_req: 4928428a7158SDouglas Fuller ceph_osdc_put_request(req); 4929428a7158SDouglas Fuller return ret; 4930428a7158SDouglas Fuller } 4931428a7158SDouglas Fuller EXPORT_SYMBOL(ceph_osdc_call); 4932428a7158SDouglas Fuller 4933428a7158SDouglas Fuller /* 49343d14c5d2SYehuda Sadeh * init, shutdown 49353d14c5d2SYehuda Sadeh */ 49363d14c5d2SYehuda Sadeh int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client) 49373d14c5d2SYehuda Sadeh { 49383d14c5d2SYehuda Sadeh int err; 49393d14c5d2SYehuda Sadeh 49403d14c5d2SYehuda Sadeh dout("init\n"); 49413d14c5d2SYehuda Sadeh osdc->client = client; 49425aea3dcdSIlya Dryomov init_rwsem(&osdc->lock); 49433d14c5d2SYehuda Sadeh osdc->osds = RB_ROOT; 49443d14c5d2SYehuda Sadeh INIT_LIST_HEAD(&osdc->osd_lru); 49459dd2845cSIlya Dryomov spin_lock_init(&osdc->osd_lru_lock); 49465aea3dcdSIlya Dryomov osd_init(&osdc->homeless_osd); 49475aea3dcdSIlya Dryomov osdc->homeless_osd.o_osdc = osdc; 49485aea3dcdSIlya Dryomov osdc->homeless_osd.o_osd = CEPH_HOMELESS_OSD; 4949264048afSIlya Dryomov osdc->last_linger_id = CEPH_LINGER_ID_START; 4950922dab61SIlya Dryomov osdc->linger_requests = RB_ROOT; 49514609245eSIlya Dryomov osdc->map_checks = RB_ROOT; 49524609245eSIlya Dryomov osdc->linger_map_checks = RB_ROOT; 49533d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout); 49543d14c5d2SYehuda Sadeh INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout); 49553d14c5d2SYehuda Sadeh 49563d14c5d2SYehuda Sadeh err = -ENOMEM; 4957e5253a7bSIlya Dryomov osdc->osdmap = ceph_osdmap_alloc(); 4958e5253a7bSIlya Dryomov if (!osdc->osdmap) 4959e5253a7bSIlya Dryomov goto out; 4960e5253a7bSIlya Dryomov 49619e767adbSIlya Dryomov osdc->req_mempool = mempool_create_slab_pool(10, 49629e767adbSIlya Dryomov ceph_osd_request_cache); 49633d14c5d2SYehuda Sadeh if (!osdc->req_mempool) 4964e5253a7bSIlya Dryomov goto out_map; 49653d14c5d2SYehuda Sadeh 4966d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP, 4967711da55dSIlya Dryomov PAGE_SIZE, 10, true, "osd_op"); 49683d14c5d2SYehuda Sadeh if (err < 0) 49693d14c5d2SYehuda Sadeh goto out_mempool; 4970d50b409fSSage Weil err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY, 4971711da55dSIlya Dryomov PAGE_SIZE, 10, true, "osd_op_reply"); 49723d14c5d2SYehuda Sadeh if (err < 0) 49733d14c5d2SYehuda Sadeh goto out_msgpool; 4974a40c4f10SYehuda Sadeh 4975dbcae088SDan Carpenter err = -ENOMEM; 4976a40c4f10SYehuda Sadeh osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify"); 4977dbcae088SDan Carpenter if (!osdc->notify_wq) 4978c172ec5cSIlya Dryomov goto out_msgpool_reply; 4979c172ec5cSIlya Dryomov 4980fbca9635SIlya Dryomov schedule_delayed_work(&osdc->timeout_work, 4981fbca9635SIlya Dryomov osdc->client->options->osd_keepalive_timeout); 4982b37ee1b9SIlya Dryomov schedule_delayed_work(&osdc->osds_timeout_work, 4983b37ee1b9SIlya Dryomov round_jiffies_relative(osdc->client->options->osd_idle_ttl)); 4984b37ee1b9SIlya Dryomov 49853d14c5d2SYehuda Sadeh return 0; 49863d14c5d2SYehuda Sadeh 4987c172ec5cSIlya Dryomov out_msgpool_reply: 4988c172ec5cSIlya Dryomov ceph_msgpool_destroy(&osdc->msgpool_op_reply); 49893d14c5d2SYehuda Sadeh out_msgpool: 49903d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 49913d14c5d2SYehuda Sadeh out_mempool: 49923d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 4993e5253a7bSIlya Dryomov out_map: 4994e5253a7bSIlya Dryomov ceph_osdmap_destroy(osdc->osdmap); 49953d14c5d2SYehuda Sadeh out: 49963d14c5d2SYehuda Sadeh return err; 49973d14c5d2SYehuda Sadeh } 49983d14c5d2SYehuda Sadeh 49993d14c5d2SYehuda Sadeh void ceph_osdc_stop(struct ceph_osd_client *osdc) 50003d14c5d2SYehuda Sadeh { 5001a40c4f10SYehuda Sadeh flush_workqueue(osdc->notify_wq); 5002a40c4f10SYehuda Sadeh destroy_workqueue(osdc->notify_wq); 50033d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->timeout_work); 50043d14c5d2SYehuda Sadeh cancel_delayed_work_sync(&osdc->osds_timeout_work); 500542a2c09fSIlya Dryomov 50065aea3dcdSIlya Dryomov down_write(&osdc->lock); 500742a2c09fSIlya Dryomov while (!RB_EMPTY_ROOT(&osdc->osds)) { 500842a2c09fSIlya Dryomov struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds), 500942a2c09fSIlya Dryomov struct ceph_osd, o_node); 50105aea3dcdSIlya Dryomov close_osd(osd); 501142a2c09fSIlya Dryomov } 50125aea3dcdSIlya Dryomov up_write(&osdc->lock); 501302113a0fSElena Reshetova WARN_ON(refcount_read(&osdc->homeless_osd.o_ref) != 1); 50145aea3dcdSIlya Dryomov osd_cleanup(&osdc->homeless_osd); 50155aea3dcdSIlya Dryomov 50165aea3dcdSIlya Dryomov WARN_ON(!list_empty(&osdc->osd_lru)); 5017922dab61SIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_requests)); 50184609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->map_checks)); 50194609245eSIlya Dryomov WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_map_checks)); 50205aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_requests)); 50215aea3dcdSIlya Dryomov WARN_ON(atomic_read(&osdc->num_homeless)); 502242a2c09fSIlya Dryomov 50233d14c5d2SYehuda Sadeh ceph_osdmap_destroy(osdc->osdmap); 50243d14c5d2SYehuda Sadeh mempool_destroy(osdc->req_mempool); 50253d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op); 50263d14c5d2SYehuda Sadeh ceph_msgpool_destroy(&osdc->msgpool_op_reply); 50273d14c5d2SYehuda Sadeh } 50283d14c5d2SYehuda Sadeh 50293d14c5d2SYehuda Sadeh /* 50303d14c5d2SYehuda Sadeh * Read some contiguous pages. If we cross a stripe boundary, shorten 50313d14c5d2SYehuda Sadeh * *plen. Return number of bytes read, or error. 50323d14c5d2SYehuda Sadeh */ 50333d14c5d2SYehuda Sadeh int ceph_osdc_readpages(struct ceph_osd_client *osdc, 50343d14c5d2SYehuda Sadeh struct ceph_vino vino, struct ceph_file_layout *layout, 50353d14c5d2SYehuda Sadeh u64 off, u64 *plen, 50363d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 5037b7495fc2SSage Weil struct page **pages, int num_pages, int page_align) 50383d14c5d2SYehuda Sadeh { 50393d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 50403d14c5d2SYehuda Sadeh int rc = 0; 50413d14c5d2SYehuda Sadeh 50423d14c5d2SYehuda Sadeh dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino, 50433d14c5d2SYehuda Sadeh vino.snap, off, *plen); 5044715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, plen, 0, 1, 50453d14c5d2SYehuda Sadeh CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ, 5046acead002SAlex Elder NULL, truncate_seq, truncate_size, 5047153e5167SAlex Elder false); 50486816282dSSage Weil if (IS_ERR(req)) 50496816282dSSage Weil return PTR_ERR(req); 50503d14c5d2SYehuda Sadeh 50513d14c5d2SYehuda Sadeh /* it may be a short read due to an object boundary */ 5052406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, 5053a4ce40a9SAlex Elder pages, *plen, page_align, false, false); 50543d14c5d2SYehuda Sadeh 5055e0c59487SAlex Elder dout("readpages final extent is %llu~%llu (%llu bytes align %d)\n", 505643bfe5deSAlex Elder off, *plen, *plen, page_align); 50573d14c5d2SYehuda Sadeh 50583d14c5d2SYehuda Sadeh rc = ceph_osdc_start_request(osdc, req, false); 50593d14c5d2SYehuda Sadeh if (!rc) 50603d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 50613d14c5d2SYehuda Sadeh 50623d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 50633d14c5d2SYehuda Sadeh dout("readpages result %d\n", rc); 50643d14c5d2SYehuda Sadeh return rc; 50653d14c5d2SYehuda Sadeh } 50663d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_readpages); 50673d14c5d2SYehuda Sadeh 50683d14c5d2SYehuda Sadeh /* 50693d14c5d2SYehuda Sadeh * do a synchronous write on N pages 50703d14c5d2SYehuda Sadeh */ 50713d14c5d2SYehuda Sadeh int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino, 50723d14c5d2SYehuda Sadeh struct ceph_file_layout *layout, 50733d14c5d2SYehuda Sadeh struct ceph_snap_context *snapc, 50743d14c5d2SYehuda Sadeh u64 off, u64 len, 50753d14c5d2SYehuda Sadeh u32 truncate_seq, u64 truncate_size, 50763d14c5d2SYehuda Sadeh struct timespec *mtime, 507724808826SAlex Elder struct page **pages, int num_pages) 50783d14c5d2SYehuda Sadeh { 50793d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 50803d14c5d2SYehuda Sadeh int rc = 0; 5081b7495fc2SSage Weil int page_align = off & ~PAGE_MASK; 50823d14c5d2SYehuda Sadeh 5083715e4cd4SYan, Zheng req = ceph_osdc_new_request(osdc, layout, vino, off, &len, 0, 1, 508454ea0046SIlya Dryomov CEPH_OSD_OP_WRITE, CEPH_OSD_FLAG_WRITE, 5085acead002SAlex Elder snapc, truncate_seq, truncate_size, 5086153e5167SAlex Elder true); 50876816282dSSage Weil if (IS_ERR(req)) 50886816282dSSage Weil return PTR_ERR(req); 50893d14c5d2SYehuda Sadeh 50903d14c5d2SYehuda Sadeh /* it may be a short write due to an object boundary */ 5091406e2c9fSAlex Elder osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align, 509243bfe5deSAlex Elder false, false); 509343bfe5deSAlex Elder dout("writepages %llu~%llu (%llu bytes)\n", off, len, len); 50943d14c5d2SYehuda Sadeh 5095bb873b53SIlya Dryomov req->r_mtime = *mtime; 509687f979d3SAlex Elder rc = ceph_osdc_start_request(osdc, req, true); 50973d14c5d2SYehuda Sadeh if (!rc) 50983d14c5d2SYehuda Sadeh rc = ceph_osdc_wait_request(osdc, req); 50993d14c5d2SYehuda Sadeh 51003d14c5d2SYehuda Sadeh ceph_osdc_put_request(req); 51013d14c5d2SYehuda Sadeh if (rc == 0) 51023d14c5d2SYehuda Sadeh rc = len; 51033d14c5d2SYehuda Sadeh dout("writepages result %d\n", rc); 51043d14c5d2SYehuda Sadeh return rc; 51053d14c5d2SYehuda Sadeh } 51063d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_osdc_writepages); 51073d14c5d2SYehuda Sadeh 510857a35dfbSChengguang Xu int __init ceph_osdc_setup(void) 51095522ae0bSAlex Elder { 51103f1af42aSIlya Dryomov size_t size = sizeof(struct ceph_osd_request) + 51113f1af42aSIlya Dryomov CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op); 51123f1af42aSIlya Dryomov 51135522ae0bSAlex Elder BUG_ON(ceph_osd_request_cache); 51143f1af42aSIlya Dryomov ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size, 51153f1af42aSIlya Dryomov 0, 0, NULL); 51165522ae0bSAlex Elder 51175522ae0bSAlex Elder return ceph_osd_request_cache ? 0 : -ENOMEM; 51185522ae0bSAlex Elder } 51195522ae0bSAlex Elder 51205522ae0bSAlex Elder void ceph_osdc_cleanup(void) 51215522ae0bSAlex Elder { 51225522ae0bSAlex Elder BUG_ON(!ceph_osd_request_cache); 51235522ae0bSAlex Elder kmem_cache_destroy(ceph_osd_request_cache); 51245522ae0bSAlex Elder ceph_osd_request_cache = NULL; 51255522ae0bSAlex Elder } 51265522ae0bSAlex Elder 51273d14c5d2SYehuda Sadeh /* 51283d14c5d2SYehuda Sadeh * handle incoming message 51293d14c5d2SYehuda Sadeh */ 51303d14c5d2SYehuda Sadeh static void dispatch(struct ceph_connection *con, struct ceph_msg *msg) 51313d14c5d2SYehuda Sadeh { 51323d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 51335aea3dcdSIlya Dryomov struct ceph_osd_client *osdc = osd->o_osdc; 51343d14c5d2SYehuda Sadeh int type = le16_to_cpu(msg->hdr.type); 51353d14c5d2SYehuda Sadeh 51363d14c5d2SYehuda Sadeh switch (type) { 51373d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 51383d14c5d2SYehuda Sadeh ceph_osdc_handle_map(osdc, msg); 51393d14c5d2SYehuda Sadeh break; 51403d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 51415aea3dcdSIlya Dryomov handle_reply(osd, msg); 51423d14c5d2SYehuda Sadeh break; 5143a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5144a02a946dSIlya Dryomov handle_backoff(osd, msg); 5145a02a946dSIlya Dryomov break; 5146a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 5147a40c4f10SYehuda Sadeh handle_watch_notify(osdc, msg); 5148a40c4f10SYehuda Sadeh break; 51493d14c5d2SYehuda Sadeh 51503d14c5d2SYehuda Sadeh default: 51513d14c5d2SYehuda Sadeh pr_err("received unknown message type %d %s\n", type, 51523d14c5d2SYehuda Sadeh ceph_msg_type_name(type)); 51533d14c5d2SYehuda Sadeh } 51545aea3dcdSIlya Dryomov 51553d14c5d2SYehuda Sadeh ceph_msg_put(msg); 51563d14c5d2SYehuda Sadeh } 51573d14c5d2SYehuda Sadeh 51583d14c5d2SYehuda Sadeh /* 5159d15f9d69SIlya Dryomov * Lookup and return message for incoming reply. Don't try to do 5160d15f9d69SIlya Dryomov * anything about a larger than preallocated data portion of the 5161d15f9d69SIlya Dryomov * message at the moment - for now, just skip the message. 51623d14c5d2SYehuda Sadeh */ 51633d14c5d2SYehuda Sadeh static struct ceph_msg *get_reply(struct ceph_connection *con, 51643d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 51653d14c5d2SYehuda Sadeh int *skip) 51663d14c5d2SYehuda Sadeh { 51673d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 51683d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = osd->o_osdc; 51695aea3dcdSIlya Dryomov struct ceph_msg *m = NULL; 51703d14c5d2SYehuda Sadeh struct ceph_osd_request *req; 51713f0a4ac5SIlya Dryomov int front_len = le32_to_cpu(hdr->front_len); 51723d14c5d2SYehuda Sadeh int data_len = le32_to_cpu(hdr->data_len); 51735aea3dcdSIlya Dryomov u64 tid = le64_to_cpu(hdr->tid); 51743d14c5d2SYehuda Sadeh 51755aea3dcdSIlya Dryomov down_read(&osdc->lock); 51765aea3dcdSIlya Dryomov if (!osd_registered(osd)) { 51775aea3dcdSIlya Dryomov dout("%s osd%d unknown, skipping\n", __func__, osd->o_osd); 51785aea3dcdSIlya Dryomov *skip = 1; 51795aea3dcdSIlya Dryomov goto out_unlock_osdc; 51805aea3dcdSIlya Dryomov } 51815aea3dcdSIlya Dryomov WARN_ON(osd->o_osd != le64_to_cpu(hdr->src.num)); 51825aea3dcdSIlya Dryomov 51835aea3dcdSIlya Dryomov mutex_lock(&osd->lock); 51845aea3dcdSIlya Dryomov req = lookup_request(&osd->o_requests, tid); 51853d14c5d2SYehuda Sadeh if (!req) { 5186cd8140c6SIlya Dryomov dout("%s osd%d tid %llu unknown, skipping\n", __func__, 5187cd8140c6SIlya Dryomov osd->o_osd, tid); 5188d15f9d69SIlya Dryomov *skip = 1; 51895aea3dcdSIlya Dryomov goto out_unlock_session; 51903d14c5d2SYehuda Sadeh } 51913d14c5d2SYehuda Sadeh 51928921d114SAlex Elder ceph_msg_revoke_incoming(req->r_reply); 51933d14c5d2SYehuda Sadeh 5194f2be82b0SIlya Dryomov if (front_len > req->r_reply->front_alloc_len) { 5195d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu front %d > preallocated %d\n", 5196d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, front_len, 5197d15f9d69SIlya Dryomov req->r_reply->front_alloc_len); 51983f0a4ac5SIlya Dryomov m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS, 51993f0a4ac5SIlya Dryomov false); 52003d14c5d2SYehuda Sadeh if (!m) 52015aea3dcdSIlya Dryomov goto out_unlock_session; 52023d14c5d2SYehuda Sadeh ceph_msg_put(req->r_reply); 52033d14c5d2SYehuda Sadeh req->r_reply = m; 52043d14c5d2SYehuda Sadeh } 52053d14c5d2SYehuda Sadeh 5206d15f9d69SIlya Dryomov if (data_len > req->r_reply->data_length) { 5207d15f9d69SIlya Dryomov pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n", 5208d15f9d69SIlya Dryomov __func__, osd->o_osd, req->r_tid, data_len, 5209d15f9d69SIlya Dryomov req->r_reply->data_length); 52103d14c5d2SYehuda Sadeh m = NULL; 5211d15f9d69SIlya Dryomov *skip = 1; 52125aea3dcdSIlya Dryomov goto out_unlock_session; 52133d14c5d2SYehuda Sadeh } 5214d15f9d69SIlya Dryomov 5215d15f9d69SIlya Dryomov m = ceph_msg_get(req->r_reply); 52163d14c5d2SYehuda Sadeh dout("get_reply tid %lld %p\n", tid, m); 52173d14c5d2SYehuda Sadeh 52185aea3dcdSIlya Dryomov out_unlock_session: 52195aea3dcdSIlya Dryomov mutex_unlock(&osd->lock); 52205aea3dcdSIlya Dryomov out_unlock_osdc: 52215aea3dcdSIlya Dryomov up_read(&osdc->lock); 52223d14c5d2SYehuda Sadeh return m; 52233d14c5d2SYehuda Sadeh } 52243d14c5d2SYehuda Sadeh 522519079203SIlya Dryomov /* 522619079203SIlya Dryomov * TODO: switch to a msg-owned pagelist 522719079203SIlya Dryomov */ 522819079203SIlya Dryomov static struct ceph_msg *alloc_msg_with_page_vector(struct ceph_msg_header *hdr) 522919079203SIlya Dryomov { 523019079203SIlya Dryomov struct ceph_msg *m; 523119079203SIlya Dryomov int type = le16_to_cpu(hdr->type); 523219079203SIlya Dryomov u32 front_len = le32_to_cpu(hdr->front_len); 523319079203SIlya Dryomov u32 data_len = le32_to_cpu(hdr->data_len); 523419079203SIlya Dryomov 523519079203SIlya Dryomov m = ceph_msg_new(type, front_len, GFP_NOIO, false); 523619079203SIlya Dryomov if (!m) 523719079203SIlya Dryomov return NULL; 523819079203SIlya Dryomov 523919079203SIlya Dryomov if (data_len) { 524019079203SIlya Dryomov struct page **pages; 524119079203SIlya Dryomov struct ceph_osd_data osd_data; 524219079203SIlya Dryomov 524319079203SIlya Dryomov pages = ceph_alloc_page_vector(calc_pages_for(0, data_len), 524419079203SIlya Dryomov GFP_NOIO); 5245c22e853aSWei Yongjun if (IS_ERR(pages)) { 524619079203SIlya Dryomov ceph_msg_put(m); 524719079203SIlya Dryomov return NULL; 524819079203SIlya Dryomov } 524919079203SIlya Dryomov 525019079203SIlya Dryomov ceph_osd_data_pages_init(&osd_data, pages, data_len, 0, false, 525119079203SIlya Dryomov false); 525219079203SIlya Dryomov ceph_osdc_msg_data_add(m, &osd_data); 525319079203SIlya Dryomov } 525419079203SIlya Dryomov 525519079203SIlya Dryomov return m; 525619079203SIlya Dryomov } 525719079203SIlya Dryomov 52583d14c5d2SYehuda Sadeh static struct ceph_msg *alloc_msg(struct ceph_connection *con, 52593d14c5d2SYehuda Sadeh struct ceph_msg_header *hdr, 52603d14c5d2SYehuda Sadeh int *skip) 52613d14c5d2SYehuda Sadeh { 52623d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 52633d14c5d2SYehuda Sadeh int type = le16_to_cpu(hdr->type); 52643d14c5d2SYehuda Sadeh 52651c20f2d2SAlex Elder *skip = 0; 52663d14c5d2SYehuda Sadeh switch (type) { 52673d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_MAP: 5268a02a946dSIlya Dryomov case CEPH_MSG_OSD_BACKOFF: 5269a40c4f10SYehuda Sadeh case CEPH_MSG_WATCH_NOTIFY: 527019079203SIlya Dryomov return alloc_msg_with_page_vector(hdr); 52713d14c5d2SYehuda Sadeh case CEPH_MSG_OSD_OPREPLY: 52723d14c5d2SYehuda Sadeh return get_reply(con, hdr, skip); 52733d14c5d2SYehuda Sadeh default: 52745aea3dcdSIlya Dryomov pr_warn("%s osd%d unknown msg type %d, skipping\n", __func__, 52755aea3dcdSIlya Dryomov osd->o_osd, type); 52763d14c5d2SYehuda Sadeh *skip = 1; 52773d14c5d2SYehuda Sadeh return NULL; 52783d14c5d2SYehuda Sadeh } 52793d14c5d2SYehuda Sadeh } 52803d14c5d2SYehuda Sadeh 52813d14c5d2SYehuda Sadeh /* 52823d14c5d2SYehuda Sadeh * Wrappers to refcount containing ceph_osd struct 52833d14c5d2SYehuda Sadeh */ 52843d14c5d2SYehuda Sadeh static struct ceph_connection *get_osd_con(struct ceph_connection *con) 52853d14c5d2SYehuda Sadeh { 52863d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 52873d14c5d2SYehuda Sadeh if (get_osd(osd)) 52883d14c5d2SYehuda Sadeh return con; 52893d14c5d2SYehuda Sadeh return NULL; 52903d14c5d2SYehuda Sadeh } 52913d14c5d2SYehuda Sadeh 52923d14c5d2SYehuda Sadeh static void put_osd_con(struct ceph_connection *con) 52933d14c5d2SYehuda Sadeh { 52943d14c5d2SYehuda Sadeh struct ceph_osd *osd = con->private; 52953d14c5d2SYehuda Sadeh put_osd(osd); 52963d14c5d2SYehuda Sadeh } 52973d14c5d2SYehuda Sadeh 52983d14c5d2SYehuda Sadeh /* 52993d14c5d2SYehuda Sadeh * authentication 53003d14c5d2SYehuda Sadeh */ 5301a3530df3SAlex Elder /* 5302a3530df3SAlex Elder * Note: returned pointer is the address of a structure that's 5303a3530df3SAlex Elder * managed separately. Caller must *not* attempt to free it. 5304a3530df3SAlex Elder */ 5305a3530df3SAlex Elder static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con, 53068f43fb53SAlex Elder int *proto, int force_new) 53073d14c5d2SYehuda Sadeh { 53083d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 53093d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 53103d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 531174f1869fSAlex Elder struct ceph_auth_handshake *auth = &o->o_auth; 53123d14c5d2SYehuda Sadeh 531374f1869fSAlex Elder if (force_new && auth->authorizer) { 53146c1ea260SIlya Dryomov ceph_auth_destroy_authorizer(auth->authorizer); 531574f1869fSAlex Elder auth->authorizer = NULL; 53163d14c5d2SYehuda Sadeh } 531727859f97SSage Weil if (!auth->authorizer) { 531827859f97SSage Weil int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 5319a3530df3SAlex Elder auth); 53203d14c5d2SYehuda Sadeh if (ret) 5321a3530df3SAlex Elder return ERR_PTR(ret); 532227859f97SSage Weil } else { 532327859f97SSage Weil int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD, 53240bed9b5cSSage Weil auth); 53250bed9b5cSSage Weil if (ret) 53260bed9b5cSSage Weil return ERR_PTR(ret); 53273d14c5d2SYehuda Sadeh } 53283d14c5d2SYehuda Sadeh *proto = ac->protocol; 532974f1869fSAlex Elder 5330a3530df3SAlex Elder return auth; 53313d14c5d2SYehuda Sadeh } 53323d14c5d2SYehuda Sadeh 53333d14c5d2SYehuda Sadeh 53340dde5848SIlya Dryomov static int verify_authorizer_reply(struct ceph_connection *con) 53353d14c5d2SYehuda Sadeh { 53363d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 53373d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 53383d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 53393d14c5d2SYehuda Sadeh 53400dde5848SIlya Dryomov return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer); 53413d14c5d2SYehuda Sadeh } 53423d14c5d2SYehuda Sadeh 53433d14c5d2SYehuda Sadeh static int invalidate_authorizer(struct ceph_connection *con) 53443d14c5d2SYehuda Sadeh { 53453d14c5d2SYehuda Sadeh struct ceph_osd *o = con->private; 53463d14c5d2SYehuda Sadeh struct ceph_osd_client *osdc = o->o_osdc; 53473d14c5d2SYehuda Sadeh struct ceph_auth_client *ac = osdc->client->monc.auth; 53483d14c5d2SYehuda Sadeh 534927859f97SSage Weil ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD); 53503d14c5d2SYehuda Sadeh return ceph_monc_validate_auth(&osdc->client->monc); 53513d14c5d2SYehuda Sadeh } 53523d14c5d2SYehuda Sadeh 53538cb441c0SIlya Dryomov static void osd_reencode_message(struct ceph_msg *msg) 53548cb441c0SIlya Dryomov { 5355914902afSIlya Dryomov int type = le16_to_cpu(msg->hdr.type); 5356914902afSIlya Dryomov 5357914902afSIlya Dryomov if (type == CEPH_MSG_OSD_OP) 53588cb441c0SIlya Dryomov encode_request_finish(msg); 53598cb441c0SIlya Dryomov } 53608cb441c0SIlya Dryomov 536179dbd1baSIlya Dryomov static int osd_sign_message(struct ceph_msg *msg) 536233d07337SYan, Zheng { 536379dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 536433d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 536579dbd1baSIlya Dryomov 536633d07337SYan, Zheng return ceph_auth_sign_message(auth, msg); 536733d07337SYan, Zheng } 536833d07337SYan, Zheng 536979dbd1baSIlya Dryomov static int osd_check_message_signature(struct ceph_msg *msg) 537033d07337SYan, Zheng { 537179dbd1baSIlya Dryomov struct ceph_osd *o = msg->con->private; 537233d07337SYan, Zheng struct ceph_auth_handshake *auth = &o->o_auth; 537379dbd1baSIlya Dryomov 537433d07337SYan, Zheng return ceph_auth_check_message_signature(auth, msg); 537533d07337SYan, Zheng } 537633d07337SYan, Zheng 53773d14c5d2SYehuda Sadeh static const struct ceph_connection_operations osd_con_ops = { 53783d14c5d2SYehuda Sadeh .get = get_osd_con, 53793d14c5d2SYehuda Sadeh .put = put_osd_con, 53803d14c5d2SYehuda Sadeh .dispatch = dispatch, 53813d14c5d2SYehuda Sadeh .get_authorizer = get_authorizer, 53823d14c5d2SYehuda Sadeh .verify_authorizer_reply = verify_authorizer_reply, 53833d14c5d2SYehuda Sadeh .invalidate_authorizer = invalidate_authorizer, 53843d14c5d2SYehuda Sadeh .alloc_msg = alloc_msg, 53858cb441c0SIlya Dryomov .reencode_message = osd_reencode_message, 538679dbd1baSIlya Dryomov .sign_message = osd_sign_message, 538779dbd1baSIlya Dryomov .check_message_signature = osd_check_message_signature, 53885aea3dcdSIlya Dryomov .fault = osd_fault, 53893d14c5d2SYehuda Sadeh }; 5390