net/ceph/osd_client.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/ceph/ceph_debug.h>
   4
   5 #include <linux/module.h>
   6 #include <linux/err.h>
   7 #include <linux/highmem.h>
   8 #include <linux/mm.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/slab.h>
  11 #include <linux/uaccess.h>
  12 #ifdef CONFIG_BLOCK
  13 #include <linux/bio.h>
  14 #endif
  15
  16 #include <linux/ceph/ceph_features.h>
  17 #include <linux/ceph/libceph.h>
  18 #include <linux/ceph/osd_client.h>
  19 #include <linux/ceph/messenger.h>
  20 #include <linux/ceph/decode.h>
  21 #include <linux/ceph/auth.h>
  22 #include <linux/ceph/pagelist.h>
  23 #include <linux/ceph/striper.h>
  24
  25 #define OSD_OPREPLY_FRONT_LEN   512
  26
  27 static struct kmem_cache        *ceph_osd_request_cache;
  28
  29 static const struct ceph_connection_operations osd_con_ops;
  30
  31 /*
  32  * Implement client access to distributed object storage cluster.
  33  *
  34  * All data objects are stored within a cluster/cloud of OSDs, or
  35  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  36  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  37  * remote daemons serving up and coordinating consistent and safe
  38  * access to storage.
  39  *
  40  * Cluster membership and the mapping of data objects onto storage devices
  41  * are described by the osd map.
  42  *
  43  * We keep track of pending OSD requests (read, write), resubmit
  44  * requests to different OSDs when the cluster topology/data layout
  45  * change, or retry the affected requests when the communications
  46  * channel with an OSD is reset.
  47  */
  48
  49 static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req);
  50 static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req);
  51 static void link_linger(struct ceph_osd *osd,
  52                         struct ceph_osd_linger_request *lreq);
  53 static void unlink_linger(struct ceph_osd *osd,
  54                           struct ceph_osd_linger_request *lreq);
  55 static void clear_backoffs(struct ceph_osd *osd);
  56
  57 #if 1
  58 static inline bool rwsem_is_wrlocked(struct rw_semaphore *sem)
  59 {
  60         bool wrlocked = true;
  61
  62         if (unlikely(down_read_trylock(sem))) {
  63                 wrlocked = false;
  64                 up_read(sem);
  65         }
  66
  67         return wrlocked;
  68 }
  69 static inline void verify_osdc_locked(struct ceph_osd_client *osdc)
  70 {
  71         WARN_ON(!rwsem_is_locked(&osdc->lock));
  72 }
  73 static inline void verify_osdc_wrlocked(struct ceph_osd_client *osdc)
  74 {
  75         WARN_ON(!rwsem_is_wrlocked(&osdc->lock));
  76 }
  77 static inline void verify_osd_locked(struct ceph_osd *osd)
  78 {
  79         struct ceph_osd_client *osdc = osd->o_osdc;
  80
  81         WARN_ON(!(mutex_is_locked(&osd->lock) &&
  82                   rwsem_is_locked(&osdc->lock)) &&
  83                 !rwsem_is_wrlocked(&osdc->lock));
  84 }
  85 static inline void verify_lreq_locked(struct ceph_osd_linger_request *lreq)
  86 {
  87         WARN_ON(!mutex_is_locked(&lreq->lock));
  88 }
  89 #else
  90 static inline void verify_osdc_locked(struct ceph_osd_client *osdc) { }
  91 static inline void verify_osdc_wrlocked(struct ceph_osd_client *osdc) { }
  92 static inline void verify_osd_locked(struct ceph_osd *osd) { }
  93 static inline void verify_lreq_locked(struct ceph_osd_linger_request *lreq) { }
  94 #endif
  95
  96 /*
  97  * calculate the mapping of a file extent onto an object, and fill out the
  98  * request accordingly.  shorten extent as necessary if it crosses an
  99  * object boundary.
 100  *
 101  * fill osd op in request message.
 102  */
 103 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
 104                         u64 *objnum, u64 *objoff, u64 *objlen)
 105 {
 106         u64 orig_len = *plen;
 107         u32 xlen;
 108
 109         /* object extent? */
 110         ceph_calc_file_object_mapping(layout, off, orig_len, objnum,
 111                                           objoff, &xlen);
 112         *objlen = xlen;
 113         if (*objlen < orig_len) {
 114                 *plen = *objlen;
 115                 dout(" skipping last %llu, final file extent %llu~%llu\n",
 116                      orig_len - *plen, off, *plen);
 117         }
 118
 119         dout("calc_layout objnum=%llx %llu~%llu\n", *objnum, *objoff, *objlen);
 120         return 0;
 121 }
 122
 123 static void ceph_osd_data_init(struct ceph_osd_data *osd_data)
 124 {
 125         memset(osd_data, 0, sizeof (*osd_data));
 126         osd_data->type = CEPH_OSD_DATA_TYPE_NONE;
 127 }
 128
 129 /*
 130  * Consumes @pages if @own_pages is true.
 131  */
 132 static void ceph_osd_data_pages_init(struct ceph_osd_data *osd_data,
 133                         struct page **pages, u64 length, u32 alignment,
 134                         bool pages_from_pool, bool own_pages)
 135 {
 136         osd_data->type = CEPH_OSD_DATA_TYPE_PAGES;
 137         osd_data->pages = pages;
 138         osd_data->length = length;
 139         osd_data->alignment = alignment;
 140         osd_data->pages_from_pool = pages_from_pool;
 141         osd_data->own_pages = own_pages;
 142 }
 143
 144 /*
 145  * Consumes a ref on @pagelist.
 146  */
 147 static void ceph_osd_data_pagelist_init(struct ceph_osd_data *osd_data,
 148                         struct ceph_pagelist *pagelist)
 149 {
 150         osd_data->type = CEPH_OSD_DATA_TYPE_PAGELIST;
 151         osd_data->pagelist = pagelist;
 152 }
 153
 154 #ifdef CONFIG_BLOCK
 155 static void ceph_osd_data_bio_init(struct ceph_osd_data *osd_data,
 156                                    struct ceph_bio_iter *bio_pos,
 157                                    u32 bio_length)
 158 {
 159         osd_data->type = CEPH_OSD_DATA_TYPE_BIO;
 160         osd_data->bio_pos = *bio_pos;
 161         osd_data->bio_length = bio_length;
 162 }
 163 #endif /* CONFIG_BLOCK */
 164
 165 static void ceph_osd_data_bvecs_init(struct ceph_osd_data *osd_data,
 166                                      struct ceph_bvec_iter *bvec_pos,
 167                                      u32 num_bvecs)
 168 {
 169         osd_data->type = CEPH_OSD_DATA_TYPE_BVECS;
 170         osd_data->bvec_pos = *bvec_pos;
 171         osd_data->num_bvecs = num_bvecs;
 172 }
 173
 174 static struct ceph_osd_data *
 175 osd_req_op_raw_data_in(struct ceph_osd_request *osd_req, unsigned int which)
 176 {
 177         BUG_ON(which >= osd_req->r_num_ops);
 178
 179         return &osd_req->r_ops[which].raw_data_in;
 180 }
 181
 182 struct ceph_osd_data *
 183 osd_req_op_extent_osd_data(struct ceph_osd_request *osd_req,
 184                         unsigned int which)
 185 {
 186         return osd_req_op_data(osd_req, which, extent, osd_data);
 187 }
 188 EXPORT_SYMBOL(osd_req_op_extent_osd_data);
 189
 190 void osd_req_op_raw_data_in_pages(struct ceph_osd_request *osd_req,
 191                         unsigned int which, struct page **pages,
 192                         u64 length, u32 alignment,
 193                         bool pages_from_pool, bool own_pages)
 194 {
 195         struct ceph_osd_data *osd_data;
 196
 197         osd_data = osd_req_op_raw_data_in(osd_req, which);
 198         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 199                                 pages_from_pool, own_pages);
 200 }
 201 EXPORT_SYMBOL(osd_req_op_raw_data_in_pages);
 202
 203 void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *osd_req,
 204                         unsigned int which, struct page **pages,
 205                         u64 length, u32 alignment,
 206                         bool pages_from_pool, bool own_pages)
 207 {
 208         struct ceph_osd_data *osd_data;
 209
 210         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 211         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 212                                 pages_from_pool, own_pages);
 213 }
 214 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pages);
 215
 216 void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *osd_req,
 217                         unsigned int which, struct ceph_pagelist *pagelist)
 218 {
 219         struct ceph_osd_data *osd_data;
 220
 221         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 222         ceph_osd_data_pagelist_init(osd_data, pagelist);
 223 }
 224 EXPORT_SYMBOL(osd_req_op_extent_osd_data_pagelist);
 225
 226 #ifdef CONFIG_BLOCK
 227 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
 228                                     unsigned int which,
 229                                     struct ceph_bio_iter *bio_pos,
 230                                     u32 bio_length)
 231 {
 232         struct ceph_osd_data *osd_data;
 233
 234         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 235         ceph_osd_data_bio_init(osd_data, bio_pos, bio_length);
 236 }
 237 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bio);
 238 #endif /* CONFIG_BLOCK */
 239
 240 void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request *osd_req,
 241                                       unsigned int which,
 242                                       struct bio_vec *bvecs, u32 num_bvecs,
 243                                       u32 bytes)
 244 {
 245         struct ceph_osd_data *osd_data;
 246         struct ceph_bvec_iter it = {
 247                 .bvecs = bvecs,
 248                 .iter = { .bi_size = bytes },
 249         };
 250
 251         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 252         ceph_osd_data_bvecs_init(osd_data, &it, num_bvecs);
 253 }
 254 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvecs);
 255
 256 void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req,
 257                                          unsigned int which,
 258                                          struct ceph_bvec_iter *bvec_pos)
 259 {
 260         struct ceph_osd_data *osd_data;
 261
 262         osd_data = osd_req_op_data(osd_req, which, extent, osd_data);
 263         ceph_osd_data_bvecs_init(osd_data, bvec_pos, 0);
 264 }
 265 EXPORT_SYMBOL(osd_req_op_extent_osd_data_bvec_pos);
 266
 267 static void osd_req_op_cls_request_info_pagelist(
 268                         struct ceph_osd_request *osd_req,
 269                         unsigned int which, struct ceph_pagelist *pagelist)
 270 {
 271         struct ceph_osd_data *osd_data;
 272
 273         osd_data = osd_req_op_data(osd_req, which, cls, request_info);
 274         ceph_osd_data_pagelist_init(osd_data, pagelist);
 275 }
 276
 277 void osd_req_op_cls_request_data_pagelist(
 278                         struct ceph_osd_request *osd_req,
 279                         unsigned int which, struct ceph_pagelist *pagelist)
 280 {
 281         struct ceph_osd_data *osd_data;
 282
 283         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 284         ceph_osd_data_pagelist_init(osd_data, pagelist);
 285         osd_req->r_ops[which].cls.indata_len += pagelist->length;
 286         osd_req->r_ops[which].indata_len += pagelist->length;
 287 }
 288 EXPORT_SYMBOL(osd_req_op_cls_request_data_pagelist);
 289
 290 void osd_req_op_cls_request_data_pages(struct ceph_osd_request *osd_req,
 291                         unsigned int which, struct page **pages, u64 length,
 292                         u32 alignment, bool pages_from_pool, bool own_pages)
 293 {
 294         struct ceph_osd_data *osd_data;
 295
 296         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 297         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 298                                 pages_from_pool, own_pages);
 299         osd_req->r_ops[which].cls.indata_len += length;
 300         osd_req->r_ops[which].indata_len += length;
 301 }
 302 EXPORT_SYMBOL(osd_req_op_cls_request_data_pages);
 303
 304 void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req,
 305                                        unsigned int which,
 306                                        struct bio_vec *bvecs, u32 num_bvecs,
 307                                        u32 bytes)
 308 {
 309         struct ceph_osd_data *osd_data;
 310         struct ceph_bvec_iter it = {
 311                 .bvecs = bvecs,
 312                 .iter = { .bi_size = bytes },
 313         };
 314
 315         osd_data = osd_req_op_data(osd_req, which, cls, request_data);
 316         ceph_osd_data_bvecs_init(osd_data, &it, num_bvecs);
 317         osd_req->r_ops[which].cls.indata_len += bytes;
 318         osd_req->r_ops[which].indata_len += bytes;
 319 }
 320 EXPORT_SYMBOL(osd_req_op_cls_request_data_bvecs);
 321
 322 void osd_req_op_cls_response_data_pages(struct ceph_osd_request *osd_req,
 323                         unsigned int which, struct page **pages, u64 length,
 324                         u32 alignment, bool pages_from_pool, bool own_pages)
 325 {
 326         struct ceph_osd_data *osd_data;
 327
 328         osd_data = osd_req_op_data(osd_req, which, cls, response_data);
 329         ceph_osd_data_pages_init(osd_data, pages, length, alignment,
 330                                 pages_from_pool, own_pages);
 331 }
 332 EXPORT_SYMBOL(osd_req_op_cls_response_data_pages);
 333
 334 static u64 ceph_osd_data_length(struct ceph_osd_data *osd_data)
 335 {
 336         switch (osd_data->type) {
 337         case CEPH_OSD_DATA_TYPE_NONE:
 338                 return 0;
 339         case CEPH_OSD_DATA_TYPE_PAGES:
 340                 return osd_data->length;
 341         case CEPH_OSD_DATA_TYPE_PAGELIST:
 342                 return (u64)osd_data->pagelist->length;
 343 #ifdef CONFIG_BLOCK
 344         case CEPH_OSD_DATA_TYPE_BIO:
 345                 return (u64)osd_data->bio_length;
 346 #endif /* CONFIG_BLOCK */
 347         case CEPH_OSD_DATA_TYPE_BVECS:
 348                 return osd_data->bvec_pos.iter.bi_size;
 349         default:
 350                 WARN(true, "unrecognized data type %d\n", (int)osd_data->type);
 351                 return 0;
 352         }
 353 }
 354
 355 static void ceph_osd_data_release(struct ceph_osd_data *osd_data)
 356 {
 357         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES && osd_data->own_pages) {
 358                 int num_pages;
 359
 360                 num_pages = calc_pages_for((u64)osd_data->alignment,
 361                                                 (u64)osd_data->length);
 362                 ceph_release_page_vector(osd_data->pages, num_pages);
 363         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 364                 ceph_pagelist_release(osd_data->pagelist);
 365         }
 366         ceph_osd_data_init(osd_data);
 367 }
 368
 369 static void osd_req_op_data_release(struct ceph_osd_request *osd_req,
 370                         unsigned int which)
 371 {
 372         struct ceph_osd_req_op *op;
 373
 374         BUG_ON(which >= osd_req->r_num_ops);
 375         op = &osd_req->r_ops[which];
 376
 377         switch (op->op) {
 378         case CEPH_OSD_OP_READ:
 379         case CEPH_OSD_OP_WRITE:
 380         case CEPH_OSD_OP_WRITEFULL:
 381                 ceph_osd_data_release(&op->extent.osd_data);
 382                 break;
 383         case CEPH_OSD_OP_CALL:
 384                 ceph_osd_data_release(&op->cls.request_info);
 385                 ceph_osd_data_release(&op->cls.request_data);
 386                 ceph_osd_data_release(&op->cls.response_data);
 387                 break;
 388         case CEPH_OSD_OP_SETXATTR:
 389         case CEPH_OSD_OP_CMPXATTR:
 390                 ceph_osd_data_release(&op->xattr.osd_data);
 391                 break;
 392         case CEPH_OSD_OP_STAT:
 393                 ceph_osd_data_release(&op->raw_data_in);
 394                 break;
 395         case CEPH_OSD_OP_NOTIFY_ACK:
 396                 ceph_osd_data_release(&op->notify_ack.request_data);
 397                 break;
 398         case CEPH_OSD_OP_NOTIFY:
 399                 ceph_osd_data_release(&op->notify.request_data);
 400                 ceph_osd_data_release(&op->notify.response_data);
 401                 break;
 402         case CEPH_OSD_OP_LIST_WATCHERS:
 403                 ceph_osd_data_release(&op->list_watchers.response_data);
 404                 break;
 405         case CEPH_OSD_OP_COPY_FROM2:
 406                 ceph_osd_data_release(&op->copy_from.osd_data);
 407                 break;
 408         default:
 409                 break;
 410         }
 411 }
 412
 413 /*
 414  * Assumes @t is zero-initialized.
 415  */
 416 static void target_init(struct ceph_osd_request_target *t)
 417 {
 418         ceph_oid_init(&t->base_oid);
 419         ceph_oloc_init(&t->base_oloc);
 420         ceph_oid_init(&t->target_oid);
 421         ceph_oloc_init(&t->target_oloc);
 422
 423         ceph_osds_init(&t->acting);
 424         ceph_osds_init(&t->up);
 425         t->size = -1;
 426         t->min_size = -1;
 427
 428         t->osd = CEPH_HOMELESS_OSD;
 429 }
 430
 431 static void target_copy(struct ceph_osd_request_target *dest,
 432                         const struct ceph_osd_request_target *src)
 433 {
 434         ceph_oid_copy(&dest->base_oid, &src->base_oid);
 435         ceph_oloc_copy(&dest->base_oloc, &src->base_oloc);
 436         ceph_oid_copy(&dest->target_oid, &src->target_oid);
 437         ceph_oloc_copy(&dest->target_oloc, &src->target_oloc);
 438
 439         dest->pgid = src->pgid; /* struct */
 440         dest->spgid = src->spgid; /* struct */
 441         dest->pg_num = src->pg_num;
 442         dest->pg_num_mask = src->pg_num_mask;
 443         ceph_osds_copy(&dest->acting, &src->acting);
 444         ceph_osds_copy(&dest->up, &src->up);
 445         dest->size = src->size;
 446         dest->min_size = src->min_size;
 447         dest->sort_bitwise = src->sort_bitwise;
 448
 449         dest->flags = src->flags;
 450         dest->paused = src->paused;
 451
 452         dest->epoch = src->epoch;
 453         dest->last_force_resend = src->last_force_resend;
 454
 455         dest->osd = src->osd;
 456 }
 457
 458 static void target_destroy(struct ceph_osd_request_target *t)
 459 {
 460         ceph_oid_destroy(&t->base_oid);
 461         ceph_oloc_destroy(&t->base_oloc);
 462         ceph_oid_destroy(&t->target_oid);
 463         ceph_oloc_destroy(&t->target_oloc);
 464 }
 465
 466 /*
 467  * requests
 468  */
 469 static void request_release_checks(struct ceph_osd_request *req)
 470 {
 471         WARN_ON(!RB_EMPTY_NODE(&req->r_node));
 472         WARN_ON(!RB_EMPTY_NODE(&req->r_mc_node));
 473         WARN_ON(!list_empty(&req->r_private_item));
 474         WARN_ON(req->r_osd);
 475 }
 476
 477 static void ceph_osdc_release_request(struct kref *kref)
 478 {
 479         struct ceph_osd_request *req = container_of(kref,
 480                                             struct ceph_osd_request, r_kref);
 481         unsigned int which;
 482
 483         dout("%s %p (r_request %p r_reply %p)\n", __func__, req,
 484              req->r_request, req->r_reply);
 485         request_release_checks(req);
 486
 487         if (req->r_request)
 488                 ceph_msg_put(req->r_request);
 489         if (req->r_reply)
 490                 ceph_msg_put(req->r_reply);
 491
 492         for (which = 0; which < req->r_num_ops; which++)
 493                 osd_req_op_data_release(req, which);
 494
 495         target_destroy(&req->r_t);
 496         ceph_put_snap_context(req->r_snapc);
 497
 498         if (req->r_mempool)
 499                 mempool_free(req, req->r_osdc->req_mempool);
 500         else if (req->r_num_ops <= CEPH_OSD_SLAB_OPS)
 501                 kmem_cache_free(ceph_osd_request_cache, req);
 502         else
 503                 kfree(req);
 504 }
 505
 506 void ceph_osdc_get_request(struct ceph_osd_request *req)
 507 {
 508         dout("%s %p (was %d)\n", __func__, req,
 509              kref_read(&req->r_kref));
 510         kref_get(&req->r_kref);
 511 }
 512 EXPORT_SYMBOL(ceph_osdc_get_request);
 513
 514 void ceph_osdc_put_request(struct ceph_osd_request *req)
 515 {
 516         if (req) {
 517                 dout("%s %p (was %d)\n", __func__, req,
 518                      kref_read(&req->r_kref));
 519                 kref_put(&req->r_kref, ceph_osdc_release_request);
 520         }
 521 }
 522 EXPORT_SYMBOL(ceph_osdc_put_request);
 523
 524 static void request_init(struct ceph_osd_request *req)
 525 {
 526         /* req only, each op is zeroed in _osd_req_op_init() */
 527         memset(req, 0, sizeof(*req));
 528
 529         kref_init(&req->r_kref);
 530         init_completion(&req->r_completion);
 531         RB_CLEAR_NODE(&req->r_node);
 532         RB_CLEAR_NODE(&req->r_mc_node);
 533         INIT_LIST_HEAD(&req->r_private_item);
 534
 535         target_init(&req->r_t);
 536 }
 537
 538 /*
 539  * This is ugly, but it allows us to reuse linger registration and ping
 540  * requests, keeping the structure of the code around send_linger{_ping}()
 541  * reasonable.  Setting up a min_nr=2 mempool for each linger request
 542  * and dealing with copying ops (this blasts req only, watch op remains
 543  * intact) isn't any better.
 544  */
 545 static void request_reinit(struct ceph_osd_request *req)
 546 {
 547         struct ceph_osd_client *osdc = req->r_osdc;
 548         bool mempool = req->r_mempool;
 549         unsigned int num_ops = req->r_num_ops;
 550         u64 snapid = req->r_snapid;
 551         struct ceph_snap_context *snapc = req->r_snapc;
 552         bool linger = req->r_linger;
 553         struct ceph_msg *request_msg = req->r_request;
 554         struct ceph_msg *reply_msg = req->r_reply;
 555
 556         dout("%s req %p\n", __func__, req);
 557         WARN_ON(kref_read(&req->r_kref) != 1);
 558         request_release_checks(req);
 559
 560         WARN_ON(kref_read(&request_msg->kref) != 1);
 561         WARN_ON(kref_read(&reply_msg->kref) != 1);
 562         target_destroy(&req->r_t);
 563
 564         request_init(req);
 565         req->r_osdc = osdc;
 566         req->r_mempool = mempool;
 567         req->r_num_ops = num_ops;
 568         req->r_snapid = snapid;
 569         req->r_snapc = snapc;
 570         req->r_linger = linger;
 571         req->r_request = request_msg;
 572         req->r_reply = reply_msg;
 573 }
 574
 575 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 576                                                struct ceph_snap_context *snapc,
 577                                                unsigned int num_ops,
 578                                                bool use_mempool,
 579                                                gfp_t gfp_flags)
 580 {
 581         struct ceph_osd_request *req;
 582
 583         if (use_mempool) {
 584                 BUG_ON(num_ops > CEPH_OSD_SLAB_OPS);
 585                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 586         } else if (num_ops <= CEPH_OSD_SLAB_OPS) {
 587                 req = kmem_cache_alloc(ceph_osd_request_cache, gfp_flags);
 588         } else {
 589                 BUG_ON(num_ops > CEPH_OSD_MAX_OPS);
 590                 req = kmalloc(struct_size(req, r_ops, num_ops), gfp_flags);
 591         }
 592         if (unlikely(!req))
 593                 return NULL;
 594
 595         request_init(req);
 596         req->r_osdc = osdc;
 597         req->r_mempool = use_mempool;
 598         req->r_num_ops = num_ops;
 599         req->r_snapid = CEPH_NOSNAP;
 600         req->r_snapc = ceph_get_snap_context(snapc);
 601
 602         dout("%s req %p\n", __func__, req);
 603         return req;
 604 }
 605 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 606
 607 static int ceph_oloc_encoding_size(const struct ceph_object_locator *oloc)
 608 {
 609         return 8 + 4 + 4 + 4 + (oloc->pool_ns ? oloc->pool_ns->len : 0);
 610 }
 611
 612 static int __ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp,
 613                                       int num_request_data_items,
 614                                       int num_reply_data_items)
 615 {
 616         struct ceph_osd_client *osdc = req->r_osdc;
 617         struct ceph_msg *msg;
 618         int msg_size;
 619
 620         WARN_ON(req->r_request || req->r_reply);
 621         WARN_ON(ceph_oid_empty(&req->r_base_oid));
 622         WARN_ON(ceph_oloc_empty(&req->r_base_oloc));
 623
 624         /* create request message */
 625         msg_size = CEPH_ENCODING_START_BLK_LEN +
 626                         CEPH_PGID_ENCODING_LEN + 1; /* spgid */
 627         msg_size += 4 + 4 + 4; /* hash, osdmap_epoch, flags */
 628         msg_size += CEPH_ENCODING_START_BLK_LEN +
 629                         sizeof(struct ceph_osd_reqid); /* reqid */
 630         msg_size += sizeof(struct ceph_blkin_trace_info); /* trace */
 631         msg_size += 4 + sizeof(struct ceph_timespec); /* client_inc, mtime */
 632         msg_size += CEPH_ENCODING_START_BLK_LEN +
 633                         ceph_oloc_encoding_size(&req->r_base_oloc); /* oloc */
 634         msg_size += 4 + req->r_base_oid.name_len; /* oid */
 635         msg_size += 2 + req->r_num_ops * sizeof(struct ceph_osd_op);
 636         msg_size += 8; /* snapid */
 637         msg_size += 8; /* snap_seq */
 638         msg_size += 4 + 8 * (req->r_snapc ? req->r_snapc->num_snaps : 0);
 639         msg_size += 4 + 8; /* retry_attempt, features */
 640
 641         if (req->r_mempool)
 642                 msg = ceph_msgpool_get(&osdc->msgpool_op, msg_size,
 643                                        num_request_data_items);
 644         else
 645                 msg = ceph_msg_new2(CEPH_MSG_OSD_OP, msg_size,
 646                                     num_request_data_items, gfp, true);
 647         if (!msg)
 648                 return -ENOMEM;
 649
 650         memset(msg->front.iov_base, 0, msg->front.iov_len);
 651         req->r_request = msg;
 652
 653         /* create reply message */
 654         msg_size = OSD_OPREPLY_FRONT_LEN;
 655         msg_size += req->r_base_oid.name_len;
 656         msg_size += req->r_num_ops * sizeof(struct ceph_osd_op);
 657
 658         if (req->r_mempool)
 659                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, msg_size,
 660                                        num_reply_data_items);
 661         else
 662                 msg = ceph_msg_new2(CEPH_MSG_OSD_OPREPLY, msg_size,
 663                                     num_reply_data_items, gfp, true);
 664         if (!msg)
 665                 return -ENOMEM;
 666
 667         req->r_reply = msg;
 668
 669         return 0;
 670 }
 671
 672 static bool osd_req_opcode_valid(u16 opcode)
 673 {
 674         switch (opcode) {
 675 #define GENERATE_CASE(op, opcode, str)  case CEPH_OSD_OP_##op: return true;
 676 __CEPH_FORALL_OSD_OPS(GENERATE_CASE)
 677 #undef GENERATE_CASE
 678         default:
 679                 return false;
 680         }
 681 }
 682
 683 static void get_num_data_items(struct ceph_osd_request *req,
 684                                int *num_request_data_items,
 685                                int *num_reply_data_items)
 686 {
 687         struct ceph_osd_req_op *op;
 688
 689         *num_request_data_items = 0;
 690         *num_reply_data_items = 0;
 691
 692         for (op = req->r_ops; op != &req->r_ops[req->r_num_ops]; op++) {
 693                 switch (op->op) {
 694                 /* request */
 695                 case CEPH_OSD_OP_WRITE:
 696                 case CEPH_OSD_OP_WRITEFULL:
 697                 case CEPH_OSD_OP_SETXATTR:
 698                 case CEPH_OSD_OP_CMPXATTR:
 699                 case CEPH_OSD_OP_NOTIFY_ACK:
 700                 case CEPH_OSD_OP_COPY_FROM2:
 701                         *num_request_data_items += 1;
 702                         break;
 703
 704                 /* reply */
 705                 case CEPH_OSD_OP_STAT:
 706                 case CEPH_OSD_OP_READ:
 707                 case CEPH_OSD_OP_LIST_WATCHERS:
 708                         *num_reply_data_items += 1;
 709                         break;
 710
 711                 /* both */
 712                 case CEPH_OSD_OP_NOTIFY:
 713                         *num_request_data_items += 1;
 714                         *num_reply_data_items += 1;
 715                         break;
 716                 case CEPH_OSD_OP_CALL:
 717                         *num_request_data_items += 2;
 718                         *num_reply_data_items += 1;
 719                         break;
 720
 721                 default:
 722                         WARN_ON(!osd_req_opcode_valid(op->op));
 723                         break;
 724                 }
 725         }
 726 }
 727
 728 /*
 729  * oid, oloc and OSD op opcode(s) must be filled in before this function
 730  * is called.
 731  */
 732 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp)
 733 {
 734         int num_request_data_items, num_reply_data_items;
 735
 736         get_num_data_items(req, &num_request_data_items, &num_reply_data_items);
 737         return __ceph_osdc_alloc_messages(req, gfp, num_request_data_items,
 738                                           num_reply_data_items);
 739 }
 740 EXPORT_SYMBOL(ceph_osdc_alloc_messages);
 741
 742 /*
 743  * This is an osd op init function for opcodes that have no data or
 744  * other information associated with them.  It also serves as a
 745  * common init routine for all the other init functions, below.
 746  */
 747 static struct ceph_osd_req_op *
 748 _osd_req_op_init(struct ceph_osd_request *osd_req, unsigned int which,
 749                  u16 opcode, u32 flags)
 750 {
 751         struct ceph_osd_req_op *op;
 752
 753         BUG_ON(which >= osd_req->r_num_ops);
 754         BUG_ON(!osd_req_opcode_valid(opcode));
 755
 756         op = &osd_req->r_ops[which];
 757         memset(op, 0, sizeof (*op));
 758         op->op = opcode;
 759         op->flags = flags;
 760
 761         return op;
 762 }
 763
 764 void osd_req_op_init(struct ceph_osd_request *osd_req,
 765                      unsigned int which, u16 opcode, u32 flags)
 766 {
 767         (void)_osd_req_op_init(osd_req, which, opcode, flags);
 768 }
 769 EXPORT_SYMBOL(osd_req_op_init);
 770
 771 void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
 772                                 unsigned int which, u16 opcode,
 773                                 u64 offset, u64 length,
 774                                 u64 truncate_size, u32 truncate_seq)
 775 {
 776         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 777                                                       opcode, 0);
 778         size_t payload_len = 0;
 779
 780         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
 781                opcode != CEPH_OSD_OP_WRITEFULL && opcode != CEPH_OSD_OP_ZERO &&
 782                opcode != CEPH_OSD_OP_TRUNCATE);
 783
 784         op->extent.offset = offset;
 785         op->extent.length = length;
 786         op->extent.truncate_size = truncate_size;
 787         op->extent.truncate_seq = truncate_seq;
 788         if (opcode == CEPH_OSD_OP_WRITE || opcode == CEPH_OSD_OP_WRITEFULL)
 789                 payload_len += length;
 790
 791         op->indata_len = payload_len;
 792 }
 793 EXPORT_SYMBOL(osd_req_op_extent_init);
 794
 795 void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
 796                                 unsigned int which, u64 length)
 797 {
 798         struct ceph_osd_req_op *op;
 799         u64 previous;
 800
 801         BUG_ON(which >= osd_req->r_num_ops);
 802         op = &osd_req->r_ops[which];
 803         previous = op->extent.length;
 804
 805         if (length == previous)
 806                 return;         /* Nothing to do */
 807         BUG_ON(length > previous);
 808
 809         op->extent.length = length;
 810         if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
 811                 op->indata_len -= previous - length;
 812 }
 813 EXPORT_SYMBOL(osd_req_op_extent_update);
 814
 815 void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
 816                                 unsigned int which, u64 offset_inc)
 817 {
 818         struct ceph_osd_req_op *op, *prev_op;
 819
 820         BUG_ON(which + 1 >= osd_req->r_num_ops);
 821
 822         prev_op = &osd_req->r_ops[which];
 823         op = _osd_req_op_init(osd_req, which + 1, prev_op->op, prev_op->flags);
 824         /* dup previous one */
 825         op->indata_len = prev_op->indata_len;
 826         op->outdata_len = prev_op->outdata_len;
 827         op->extent = prev_op->extent;
 828         /* adjust offset */
 829         op->extent.offset += offset_inc;
 830         op->extent.length -= offset_inc;
 831
 832         if (op->op == CEPH_OSD_OP_WRITE || op->op == CEPH_OSD_OP_WRITEFULL)
 833                 op->indata_len -= offset_inc;
 834 }
 835 EXPORT_SYMBOL(osd_req_op_extent_dup_last);
 836
 837 int osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
 838                         const char *class, const char *method)
 839 {
 840         struct ceph_osd_req_op *op;
 841         struct ceph_pagelist *pagelist;
 842         size_t payload_len = 0;
 843         size_t size;
 844         int ret;
 845
 846         op = _osd_req_op_init(osd_req, which, CEPH_OSD_OP_CALL, 0);
 847
 848         pagelist = ceph_pagelist_alloc(GFP_NOFS);
 849         if (!pagelist)
 850                 return -ENOMEM;
 851
 852         op->cls.class_name = class;
 853         size = strlen(class);
 854         BUG_ON(size > (size_t) U8_MAX);
 855         op->cls.class_len = size;
 856         ret = ceph_pagelist_append(pagelist, class, size);
 857         if (ret)
 858                 goto err_pagelist_free;
 859         payload_len += size;
 860
 861         op->cls.method_name = method;
 862         size = strlen(method);
 863         BUG_ON(size > (size_t) U8_MAX);
 864         op->cls.method_len = size;
 865         ret = ceph_pagelist_append(pagelist, method, size);
 866         if (ret)
 867                 goto err_pagelist_free;
 868         payload_len += size;
 869
 870         osd_req_op_cls_request_info_pagelist(osd_req, which, pagelist);
 871         op->indata_len = payload_len;
 872         return 0;
 873
 874 err_pagelist_free:
 875         ceph_pagelist_release(pagelist);
 876         return ret;
 877 }
 878 EXPORT_SYMBOL(osd_req_op_cls_init);
 879
 880 int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
 881                           u16 opcode, const char *name, const void *value,
 882                           size_t size, u8 cmp_op, u8 cmp_mode)
 883 {
 884         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 885                                                       opcode, 0);
 886         struct ceph_pagelist *pagelist;
 887         size_t payload_len;
 888         int ret;
 889
 890         BUG_ON(opcode != CEPH_OSD_OP_SETXATTR && opcode != CEPH_OSD_OP_CMPXATTR);
 891
 892         pagelist = ceph_pagelist_alloc(GFP_NOFS);
 893         if (!pagelist)
 894                 return -ENOMEM;
 895
 896         payload_len = strlen(name);
 897         op->xattr.name_len = payload_len;
 898         ret = ceph_pagelist_append(pagelist, name, payload_len);
 899         if (ret)
 900                 goto err_pagelist_free;
 901
 902         op->xattr.value_len = size;
 903         ret = ceph_pagelist_append(pagelist, value, size);
 904         if (ret)
 905                 goto err_pagelist_free;
 906         payload_len += size;
 907
 908         op->xattr.cmp_op = cmp_op;
 909         op->xattr.cmp_mode = cmp_mode;
 910
 911         ceph_osd_data_pagelist_init(&op->xattr.osd_data, pagelist);
 912         op->indata_len = payload_len;
 913         return 0;
 914
 915 err_pagelist_free:
 916         ceph_pagelist_release(pagelist);
 917         return ret;
 918 }
 919 EXPORT_SYMBOL(osd_req_op_xattr_init);
 920
 921 /*
 922  * @watch_opcode: CEPH_OSD_WATCH_OP_*
 923  */
 924 static void osd_req_op_watch_init(struct ceph_osd_request *req, int which,
 925                                   u64 cookie, u8 watch_opcode)
 926 {
 927         struct ceph_osd_req_op *op;
 928
 929         op = _osd_req_op_init(req, which, CEPH_OSD_OP_WATCH, 0);
 930         op->watch.cookie = cookie;
 931         op->watch.op = watch_opcode;
 932         op->watch.gen = 0;
 933 }
 934
 935 /*
 936  * @flags: CEPH_OSD_OP_ALLOC_HINT_FLAG_*
 937  */
 938 void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
 939                                 unsigned int which,
 940                                 u64 expected_object_size,
 941                                 u64 expected_write_size,
 942                                 u32 flags)
 943 {
 944         struct ceph_osd_req_op *op = _osd_req_op_init(osd_req, which,
 945                                                       CEPH_OSD_OP_SETALLOCHINT,
 946                                                       0);
 947
 948         op->alloc_hint.expected_object_size = expected_object_size;
 949         op->alloc_hint.expected_write_size = expected_write_size;
 950         op->alloc_hint.flags = flags;
 951
 952         /*
 953          * CEPH_OSD_OP_SETALLOCHINT op is advisory and therefore deemed
 954          * not worth a feature bit.  Set FAILOK per-op flag to make
 955          * sure older osds don't trip over an unsupported opcode.
 956          */
 957         op->flags |= CEPH_OSD_OP_FLAG_FAILOK;
 958 }
 959 EXPORT_SYMBOL(osd_req_op_alloc_hint_init);
 960
 961 static void ceph_osdc_msg_data_add(struct ceph_msg *msg,
 962                                 struct ceph_osd_data *osd_data)
 963 {
 964         u64 length = ceph_osd_data_length(osd_data);
 965
 966         if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGES) {
 967                 BUG_ON(length > (u64) SIZE_MAX);
 968                 if (length)
 969                         ceph_msg_data_add_pages(msg, osd_data->pages,
 970                                         length, osd_data->alignment, false);
 971         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_PAGELIST) {
 972                 BUG_ON(!length);
 973                 ceph_msg_data_add_pagelist(msg, osd_data->pagelist);
 974 #ifdef CONFIG_BLOCK
 975         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BIO) {
 976                 ceph_msg_data_add_bio(msg, &osd_data->bio_pos, length);
 977 #endif
 978         } else if (osd_data->type == CEPH_OSD_DATA_TYPE_BVECS) {
 979                 ceph_msg_data_add_bvecs(msg, &osd_data->bvec_pos);
 980         } else {
 981                 BUG_ON(osd_data->type != CEPH_OSD_DATA_TYPE_NONE);
 982         }
 983 }
 984
 985 static u32 osd_req_encode_op(struct ceph_osd_op *dst,
 986                              const struct ceph_osd_req_op *src)
 987 {
 988         switch (src->op) {
 989         case CEPH_OSD_OP_STAT:
 990                 break;
 991         case CEPH_OSD_OP_READ:
 992         case CEPH_OSD_OP_WRITE:
 993         case CEPH_OSD_OP_WRITEFULL:
 994         case CEPH_OSD_OP_ZERO:
 995         case CEPH_OSD_OP_TRUNCATE:
 996                 dst->extent.offset = cpu_to_le64(src->extent.offset);
 997                 dst->extent.length = cpu_to_le64(src->extent.length);
 998                 dst->extent.truncate_size =
 999                         cpu_to_le64(src->extent.truncate_size);
1000                 dst->extent.truncate_seq =
1001                         cpu_to_le32(src->extent.truncate_seq);
1002                 break;
1003         case CEPH_OSD_OP_CALL:
1004                 dst->cls.class_len = src->cls.class_len;
1005                 dst->cls.method_len = src->cls.method_len;
1006                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
1007                 break;
1008         case CEPH_OSD_OP_WATCH:
1009                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
1010                 dst->watch.ver = cpu_to_le64(0);
1011                 dst->watch.op = src->watch.op;
1012                 dst->watch.gen = cpu_to_le32(src->watch.gen);
1013                 break;
1014         case CEPH_OSD_OP_NOTIFY_ACK:
1015                 break;
1016         case CEPH_OSD_OP_NOTIFY:
1017                 dst->notify.cookie = cpu_to_le64(src->notify.cookie);
1018                 break;
1019         case CEPH_OSD_OP_LIST_WATCHERS:
1020                 break;
1021         case CEPH_OSD_OP_SETALLOCHINT:
1022                 dst->alloc_hint.expected_object_size =
1023                     cpu_to_le64(src->alloc_hint.expected_object_size);
1024                 dst->alloc_hint.expected_write_size =
1025                     cpu_to_le64(src->alloc_hint.expected_write_size);
1026                 dst->alloc_hint.flags = cpu_to_le32(src->alloc_hint.flags);
1027                 break;
1028         case CEPH_OSD_OP_SETXATTR:
1029         case CEPH_OSD_OP_CMPXATTR:
1030                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
1031                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
1032                 dst->xattr.cmp_op = src->xattr.cmp_op;
1033                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
1034                 break;
1035         case CEPH_OSD_OP_CREATE:
1036         case CEPH_OSD_OP_DELETE:
1037                 break;
1038         case CEPH_OSD_OP_COPY_FROM2:
1039                 dst->copy_from.snapid = cpu_to_le64(src->copy_from.snapid);
1040                 dst->copy_from.src_version =
1041                         cpu_to_le64(src->copy_from.src_version);
1042                 dst->copy_from.flags = src->copy_from.flags;
1043                 dst->copy_from.src_fadvise_flags =
1044                         cpu_to_le32(src->copy_from.src_fadvise_flags);
1045                 break;
1046         default:
1047                 pr_err("unsupported osd opcode %s\n",
1048                         ceph_osd_op_name(src->op));
1049                 WARN_ON(1);
1050
1051                 return 0;
1052         }
1053
1054         dst->op = cpu_to_le16(src->op);
1055         dst->flags = cpu_to_le32(src->flags);
1056         dst->payload_len = cpu_to_le32(src->indata_len);
1057
1058         return src->indata_len;
1059 }
1060
1061 /*
1062  * build new request AND message, calculate layout, and adjust file
1063  * extent as needed.
1064  *
1065  * if the file was recently truncated, we include information about its
1066  * old and new size so that the object can be updated appropriately.  (we
1067  * avoid synchronously deleting truncated objects because it's slow.)
1068  */
1069 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
1070                                                struct ceph_file_layout *layout,
1071                                                struct ceph_vino vino,
1072                                                u64 off, u64 *plen,
1073                                                unsigned int which, int num_ops,
1074                                                int opcode, int flags,
1075                                                struct ceph_snap_context *snapc,
1076                                                u32 truncate_seq,
1077                                                u64 truncate_size,
1078                                                bool use_mempool)
1079 {
1080         struct ceph_osd_request *req;
1081         u64 objnum = 0;
1082         u64 objoff = 0;
1083         u64 objlen = 0;
1084         int r;
1085
1086         BUG_ON(opcode != CEPH_OSD_OP_READ && opcode != CEPH_OSD_OP_WRITE &&
1087                opcode != CEPH_OSD_OP_ZERO && opcode != CEPH_OSD_OP_TRUNCATE &&
1088                opcode != CEPH_OSD_OP_CREATE && opcode != CEPH_OSD_OP_DELETE);
1089
1090         req = ceph_osdc_alloc_request(osdc, snapc, num_ops, use_mempool,
1091                                         GFP_NOFS);
1092         if (!req) {
1093                 r = -ENOMEM;
1094                 goto fail;
1095         }
1096
1097         /* calculate max write size */
1098         r = calc_layout(layout, off, plen, &objnum, &objoff, &objlen);
1099         if (r)
1100                 goto fail;
1101
1102         if (opcode == CEPH_OSD_OP_CREATE || opcode == CEPH_OSD_OP_DELETE) {
1103                 osd_req_op_init(req, which, opcode, 0);
1104         } else {
1105                 u32 object_size = layout->object_size;
1106                 u32 object_base = off - objoff;
1107                 if (!(truncate_seq == 1 && truncate_size == -1ULL)) {
1108                         if (truncate_size <= object_base) {
1109                                 truncate_size = 0;
1110                         } else {
1111                                 truncate_size -= object_base;
1112                                 if (truncate_size > object_size)
1113                                         truncate_size = object_size;
1114                         }
1115                 }
1116                 osd_req_op_extent_init(req, which, opcode, objoff, objlen,
1117                                        truncate_size, truncate_seq);
1118         }
1119
1120         req->r_flags = flags;
1121         req->r_base_oloc.pool = layout->pool_id;
1122         req->r_base_oloc.pool_ns = ceph_try_get_string(layout->pool_ns);
1123         ceph_oid_printf(&req->r_base_oid, "%llx.%08llx", vino.ino, objnum);
1124
1125         req->r_snapid = vino.snap;
1126         if (flags & CEPH_OSD_FLAG_WRITE)
1127                 req->r_data_offset = off;
1128
1129         if (num_ops > 1)
1130                 /*
1131                  * This is a special case for ceph_writepages_start(), but it
1132                  * also covers ceph_uninline_data().  If more multi-op request
1133                  * use cases emerge, we will need a separate helper.
1134                  */
1135                 r = __ceph_osdc_alloc_messages(req, GFP_NOFS, num_ops, 0);
1136         else
1137                 r = ceph_osdc_alloc_messages(req, GFP_NOFS);
1138         if (r)
1139                 goto fail;
1140
1141         return req;
1142
1143 fail:
1144         ceph_osdc_put_request(req);
1145         return ERR_PTR(r);
1146 }
1147 EXPORT_SYMBOL(ceph_osdc_new_request);
1148
1149 /*
1150  * We keep osd requests in an rbtree, sorted by ->r_tid.
1151  */
1152 DEFINE_RB_FUNCS(request, struct ceph_osd_request, r_tid, r_node)
1153 DEFINE_RB_FUNCS(request_mc, struct ceph_osd_request, r_tid, r_mc_node)
1154
1155 /*
1156  * Call @fn on each OSD request as long as @fn returns 0.
1157  */
1158 static void for_each_request(struct ceph_osd_client *osdc,
1159                         int (*fn)(struct ceph_osd_request *req, void *arg),
1160                         void *arg)
1161 {
1162         struct rb_node *n, *p;
1163
1164         for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
1165                 struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
1166
1167                 for (p = rb_first(&osd->o_requests); p; ) {
1168                         struct ceph_osd_request *req =
1169                             rb_entry(p, struct ceph_osd_request, r_node);
1170
1171                         p = rb_next(p);
1172                         if (fn(req, arg))
1173                                 return;
1174                 }
1175         }
1176
1177         for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) {
1178                 struct ceph_osd_request *req =
1179                     rb_entry(p, struct ceph_osd_request, r_node);
1180
1181                 p = rb_next(p);
1182                 if (fn(req, arg))
1183                         return;
1184         }
1185 }
1186
1187 static bool osd_homeless(struct ceph_osd *osd)
1188 {
1189         return osd->o_osd == CEPH_HOMELESS_OSD;
1190 }
1191
1192 static bool osd_registered(struct ceph_osd *osd)
1193 {
1194         verify_osdc_locked(osd->o_osdc);
1195
1196         return !RB_EMPTY_NODE(&osd->o_node);
1197 }
1198
1199 /*
1200  * Assumes @osd is zero-initialized.
1201  */
1202 static void osd_init(struct ceph_osd *osd)
1203 {
1204         refcount_set(&osd->o_ref, 1);
1205         RB_CLEAR_NODE(&osd->o_node);
1206         osd->o_requests = RB_ROOT;
1207         osd->o_linger_requests = RB_ROOT;
1208         osd->o_backoff_mappings = RB_ROOT;
1209         osd->o_backoffs_by_id = RB_ROOT;
1210         INIT_LIST_HEAD(&osd->o_osd_lru);
1211         INIT_LIST_HEAD(&osd->o_keepalive_item);
1212         osd->o_incarnation = 1;
1213         mutex_init(&osd->lock);
1214 }
1215
1216 static void osd_cleanup(struct ceph_osd *osd)
1217 {
1218         WARN_ON(!RB_EMPTY_NODE(&osd->o_node));
1219         WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests));
1220         WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests));
1221         WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoff_mappings));
1222         WARN_ON(!RB_EMPTY_ROOT(&osd->o_backoffs_by_id));
1223         WARN_ON(!list_empty(&osd->o_osd_lru));
1224         WARN_ON(!list_empty(&osd->o_keepalive_item));
1225
1226         if (osd->o_auth.authorizer) {
1227                 WARN_ON(osd_homeless(osd));
1228                 ceph_auth_destroy_authorizer(osd->o_auth.authorizer);
1229         }
1230 }
1231
1232 /*
1233  * Track open sessions with osds.
1234  */
1235 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
1236 {
1237         struct ceph_osd *osd;
1238
1239         WARN_ON(onum == CEPH_HOMELESS_OSD);
1240
1241         osd = kzalloc(sizeof(*osd), GFP_NOIO | __GFP_NOFAIL);
1242         osd_init(osd);
1243         osd->o_osdc = osdc;
1244         osd->o_osd = onum;
1245
1246         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
1247
1248         return osd;
1249 }
1250
1251 static struct ceph_osd *get_osd(struct ceph_osd *osd)
1252 {
1253         if (refcount_inc_not_zero(&osd->o_ref)) {
1254                 dout("get_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref)-1,
1255                      refcount_read(&osd->o_ref));
1256                 return osd;
1257         } else {
1258                 dout("get_osd %p FAIL\n", osd);
1259                 return NULL;
1260         }
1261 }
1262
1263 static void put_osd(struct ceph_osd *osd)
1264 {
1265         dout("put_osd %p %d -> %d\n", osd, refcount_read(&osd->o_ref),
1266              refcount_read(&osd->o_ref) - 1);
1267         if (refcount_dec_and_test(&osd->o_ref)) {
1268                 osd_cleanup(osd);
1269                 kfree(osd);
1270         }
1271 }
1272
1273 DEFINE_RB_FUNCS(osd, struct ceph_osd, o_osd, o_node)
1274
1275 static void __move_osd_to_lru(struct ceph_osd *osd)
1276 {
1277         struct ceph_osd_client *osdc = osd->o_osdc;
1278
1279         dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1280         BUG_ON(!list_empty(&osd->o_osd_lru));
1281
1282         spin_lock(&osdc->osd_lru_lock);
1283         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
1284         spin_unlock(&osdc->osd_lru_lock);
1285
1286         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl;
1287 }
1288
1289 static void maybe_move_osd_to_lru(struct ceph_osd *osd)
1290 {
1291         if (RB_EMPTY_ROOT(&osd->o_requests) &&
1292             RB_EMPTY_ROOT(&osd->o_linger_requests))
1293                 __move_osd_to_lru(osd);
1294 }
1295
1296 static void __remove_osd_from_lru(struct ceph_osd *osd)
1297 {
1298         struct ceph_osd_client *osdc = osd->o_osdc;
1299
1300         dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1301
1302         spin_lock(&osdc->osd_lru_lock);
1303         if (!list_empty(&osd->o_osd_lru))
1304                 list_del_init(&osd->o_osd_lru);
1305         spin_unlock(&osdc->osd_lru_lock);
1306 }
1307
1308 /*
1309  * Close the connection and assign any leftover requests to the
1310  * homeless session.
1311  */
1312 static void close_osd(struct ceph_osd *osd)
1313 {
1314         struct ceph_osd_client *osdc = osd->o_osdc;
1315         struct rb_node *n;
1316
1317         verify_osdc_wrlocked(osdc);
1318         dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1319
1320         ceph_con_close(&osd->o_con);
1321
1322         for (n = rb_first(&osd->o_requests); n; ) {
1323                 struct ceph_osd_request *req =
1324                     rb_entry(n, struct ceph_osd_request, r_node);
1325
1326                 n = rb_next(n); /* unlink_request() */
1327
1328                 dout(" reassigning req %p tid %llu\n", req, req->r_tid);
1329                 unlink_request(osd, req);
1330                 link_request(&osdc->homeless_osd, req);
1331         }
1332         for (n = rb_first(&osd->o_linger_requests); n; ) {
1333                 struct ceph_osd_linger_request *lreq =
1334                     rb_entry(n, struct ceph_osd_linger_request, node);
1335
1336                 n = rb_next(n); /* unlink_linger() */
1337
1338                 dout(" reassigning lreq %p linger_id %llu\n", lreq,
1339                      lreq->linger_id);
1340                 unlink_linger(osd, lreq);
1341                 link_linger(&osdc->homeless_osd, lreq);
1342         }
1343         clear_backoffs(osd);
1344
1345         __remove_osd_from_lru(osd);
1346         erase_osd(&osdc->osds, osd);
1347         put_osd(osd);
1348 }
1349
1350 /*
1351  * reset osd connect
1352  */
1353 static int reopen_osd(struct ceph_osd *osd)
1354 {
1355         struct ceph_entity_addr *peer_addr;
1356
1357         dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
1358
1359         if (RB_EMPTY_ROOT(&osd->o_requests) &&
1360             RB_EMPTY_ROOT(&osd->o_linger_requests)) {
1361                 close_osd(osd);
1362                 return -ENODEV;
1363         }
1364
1365         peer_addr = &osd->o_osdc->osdmap->osd_addr[osd->o_osd];
1366         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
1367                         !ceph_con_opened(&osd->o_con)) {
1368                 struct rb_node *n;
1369
1370                 dout("osd addr hasn't changed and connection never opened, "
1371                      "letting msgr retry\n");
1372                 /* touch each r_stamp for handle_timeout()'s benfit */
1373                 for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) {
1374                         struct ceph_osd_request *req =
1375                             rb_entry(n, struct ceph_osd_request, r_node);
1376                         req->r_stamp = jiffies;
1377                 }
1378
1379                 return -EAGAIN;
1380         }
1381
1382         ceph_con_close(&osd->o_con);
1383         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
1384         osd->o_incarnation++;
1385
1386         return 0;
1387 }
1388
1389 static struct ceph_osd *lookup_create_osd(struct ceph_osd_client *osdc, int o,
1390                                           bool wrlocked)
1391 {
1392         struct ceph_osd *osd;
1393
1394         if (wrlocked)
1395                 verify_osdc_wrlocked(osdc);
1396         else
1397                 verify_osdc_locked(osdc);
1398
1399         if (o != CEPH_HOMELESS_OSD)
1400                 osd = lookup_osd(&osdc->osds, o);
1401         else
1402                 osd = &osdc->homeless_osd;
1403         if (!osd) {
1404                 if (!wrlocked)
1405                         return ERR_PTR(-EAGAIN);
1406
1407                 osd = create_osd(osdc, o);
1408                 insert_osd(&osdc->osds, osd);
1409                 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
1410                               &osdc->osdmap->osd_addr[osd->o_osd]);
1411         }
1412
1413         dout("%s osdc %p osd%d -> osd %p\n", __func__, osdc, o, osd);
1414         return osd;
1415 }
1416
1417 /*
1418  * Create request <-> OSD session relation.
1419  *
1420  * @req has to be assigned a tid, @osd may be homeless.
1421  */
1422 static void link_request(struct ceph_osd *osd, struct ceph_osd_request *req)
1423 {
1424         verify_osd_locked(osd);
1425         WARN_ON(!req->r_tid || req->r_osd);
1426         dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd,
1427              req, req->r_tid);
1428
1429         if (!osd_homeless(osd))
1430                 __remove_osd_from_lru(osd);
1431         else
1432                 atomic_inc(&osd->o_osdc->num_homeless);
1433
1434         get_osd(osd);
1435         insert_request(&osd->o_requests, req);
1436         req->r_osd = osd;
1437 }
1438
1439 static void unlink_request(struct ceph_osd *osd, struct ceph_osd_request *req)
1440 {
1441         verify_osd_locked(osd);
1442         WARN_ON(req->r_osd != osd);
1443         dout("%s osd %p osd%d req %p tid %llu\n", __func__, osd, osd->o_osd,
1444              req, req->r_tid);
1445
1446         req->r_osd = NULL;
1447         erase_request(&osd->o_requests, req);
1448         put_osd(osd);
1449
1450         if (!osd_homeless(osd))
1451                 maybe_move_osd_to_lru(osd);
1452         else
1453                 atomic_dec(&osd->o_osdc->num_homeless);
1454 }
1455
1456 static bool __pool_full(struct ceph_pg_pool_info *pi)
1457 {
1458         return pi->flags & CEPH_POOL_FLAG_FULL;
1459 }
1460
1461 static bool have_pool_full(struct ceph_osd_client *osdc)
1462 {
1463         struct rb_node *n;
1464
1465         for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) {
1466                 struct ceph_pg_pool_info *pi =
1467                     rb_entry(n, struct ceph_pg_pool_info, node);
1468
1469                 if (__pool_full(pi))
1470                         return true;
1471         }
1472
1473         return false;
1474 }
1475
1476 static bool pool_full(struct ceph_osd_client *osdc, s64 pool_id)
1477 {
1478         struct ceph_pg_pool_info *pi;
1479
1480         pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id);
1481         if (!pi)
1482                 return false;
1483
1484         return __pool_full(pi);
1485 }
1486
1487 /*
1488  * Returns whether a request should be blocked from being sent
1489  * based on the current osdmap and osd_client settings.
1490  */
1491 static bool target_should_be_paused(struct ceph_osd_client *osdc,
1492                                     const struct ceph_osd_request_target *t,
1493                                     struct ceph_pg_pool_info *pi)
1494 {
1495         bool pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD);
1496         bool pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) ||
1497                        ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
1498                        __pool_full(pi);
1499
1500         WARN_ON(pi->id != t->target_oloc.pool);
1501         return ((t->flags & CEPH_OSD_FLAG_READ) && pauserd) ||
1502                ((t->flags & CEPH_OSD_FLAG_WRITE) && pausewr) ||
1503                (osdc->osdmap->epoch < osdc->epoch_barrier);
1504 }
1505
1506 static int pick_random_replica(const struct ceph_osds *acting)
1507 {
1508         int i = prandom_u32() % acting->size;
1509
1510         dout("%s picked osd%d, primary osd%d\n", __func__,
1511              acting->osds[i], acting->primary);
1512         return i;
1513 }
1514
1515 /*
1516  * Picks the closest replica based on client's location given by
1517  * crush_location option.  Prefers the primary if the locality is
1518  * the same.
1519  */
1520 static int pick_closest_replica(struct ceph_osd_client *osdc,
1521                                 const struct ceph_osds *acting)
1522 {
1523         struct ceph_options *opt = osdc->client->options;
1524         int best_i, best_locality;
1525         int i = 0, locality;
1526
1527         do {
1528                 locality = ceph_get_crush_locality(osdc->osdmap,
1529                                                    acting->osds[i],
1530                                                    &opt->crush_locs);
1531                 if (i == 0 ||
1532                     (locality >= 0 && best_locality < 0) ||
1533                     (locality >= 0 && best_locality >= 0 &&
1534                      locality < best_locality)) {
1535                         best_i = i;
1536                         best_locality = locality;
1537                 }
1538         } while (++i < acting->size);
1539
1540         dout("%s picked osd%d with locality %d, primary osd%d\n", __func__,
1541              acting->osds[best_i], best_locality, acting->primary);
1542         return best_i;
1543 }
1544
1545 enum calc_target_result {
1546         CALC_TARGET_NO_ACTION = 0,
1547         CALC_TARGET_NEED_RESEND,
1548         CALC_TARGET_POOL_DNE,
1549 };
1550
1551 static enum calc_target_result calc_target(struct ceph_osd_client *osdc,
1552                                            struct ceph_osd_request_target *t,
1553                                            bool any_change)
1554 {
1555         struct ceph_pg_pool_info *pi;
1556         struct ceph_pg pgid, last_pgid;
1557         struct ceph_osds up, acting;
1558         bool is_read = t->flags & CEPH_OSD_FLAG_READ;
1559         bool is_write = t->flags & CEPH_OSD_FLAG_WRITE;
1560         bool force_resend = false;
1561         bool unpaused = false;
1562         bool legacy_change = false;
1563         bool split = false;
1564         bool sort_bitwise = ceph_osdmap_flag(osdc, CEPH_OSDMAP_SORTBITWISE);
1565         bool recovery_deletes = ceph_osdmap_flag(osdc,
1566                                                  CEPH_OSDMAP_RECOVERY_DELETES);
1567         enum calc_target_result ct_res;
1568
1569         t->epoch = osdc->osdmap->epoch;
1570         pi = ceph_pg_pool_by_id(osdc->osdmap, t->base_oloc.pool);
1571         if (!pi) {
1572                 t->osd = CEPH_HOMELESS_OSD;
1573                 ct_res = CALC_TARGET_POOL_DNE;
1574                 goto out;
1575         }
1576
1577         if (osdc->osdmap->epoch == pi->last_force_request_resend) {
1578                 if (t->last_force_resend < pi->last_force_request_resend) {
1579                         t->last_force_resend = pi->last_force_request_resend;
1580                         force_resend = true;
1581                 } else if (t->last_force_resend == 0) {
1582                         force_resend = true;
1583                 }
1584         }
1585
1586         /* apply tiering */
1587         ceph_oid_copy(&t->target_oid, &t->base_oid);
1588         ceph_oloc_copy(&t->target_oloc, &t->base_oloc);
1589         if ((t->flags & CEPH_OSD_FLAG_IGNORE_OVERLAY) == 0) {
1590                 if (is_read && pi->read_tier >= 0)
1591                         t->target_oloc.pool = pi->read_tier;
1592                 if (is_write && pi->write_tier >= 0)
1593                         t->target_oloc.pool = pi->write_tier;
1594
1595                 pi = ceph_pg_pool_by_id(osdc->osdmap, t->target_oloc.pool);
1596                 if (!pi) {
1597                         t->osd = CEPH_HOMELESS_OSD;
1598                         ct_res = CALC_TARGET_POOL_DNE;
1599                         goto out;
1600                 }
1601         }
1602
1603         __ceph_object_locator_to_pg(pi, &t->target_oid, &t->target_oloc, &pgid);
1604         last_pgid.pool = pgid.pool;
1605         last_pgid.seed = ceph_stable_mod(pgid.seed, t->pg_num, t->pg_num_mask);
1606
1607         ceph_pg_to_up_acting_osds(osdc->osdmap, pi, &pgid, &up, &acting);
1608         if (any_change &&
1609             ceph_is_new_interval(&t->acting,
1610                                  &acting,
1611                                  &t->up,
1612                                  &up,
1613                                  t->size,
1614                                  pi->size,
1615                                  t->min_size,
1616                                  pi->min_size,
1617                                  t->pg_num,
1618                                  pi->pg_num,
1619                                  t->sort_bitwise,
1620                                  sort_bitwise,
1621                                  t->recovery_deletes,
1622                                  recovery_deletes,
1623                                  &last_pgid))
1624                 force_resend = true;
1625
1626         if (t->paused && !target_should_be_paused(osdc, t, pi)) {
1627                 t->paused = false;
1628                 unpaused = true;
1629         }
1630         legacy_change = ceph_pg_compare(&t->pgid, &pgid) ||
1631                         ceph_osds_changed(&t->acting, &acting,
1632                                           t->used_replica || any_change);
1633         if (t->pg_num)
1634                 split = ceph_pg_is_split(&last_pgid, t->pg_num, pi->pg_num);
1635
1636         if (legacy_change || force_resend || split) {
1637                 t->pgid = pgid; /* struct */
1638                 ceph_pg_to_primary_shard(osdc->osdmap, pi, &pgid, &t->spgid);
1639                 ceph_osds_copy(&t->acting, &acting);
1640                 ceph_osds_copy(&t->up, &up);
1641                 t->size = pi->size;
1642                 t->min_size = pi->min_size;
1643                 t->pg_num = pi->pg_num;
1644                 t->pg_num_mask = pi->pg_num_mask;
1645                 t->sort_bitwise = sort_bitwise;
1646                 t->recovery_deletes = recovery_deletes;
1647
1648                 if ((t->flags & (CEPH_OSD_FLAG_BALANCE_READS |
1649                                  CEPH_OSD_FLAG_LOCALIZE_READS)) &&
1650                     !is_write && pi->type == CEPH_POOL_TYPE_REP &&
1651                     acting.size > 1) {
1652                         int pos;
1653
1654                         WARN_ON(!is_read || acting.osds[0] != acting.primary);
1655                         if (t->flags & CEPH_OSD_FLAG_BALANCE_READS) {
1656                                 pos = pick_random_replica(&acting);
1657                         } else {
1658                                 pos = pick_closest_replica(osdc, &acting);
1659                         }
1660                         t->osd = acting.osds[pos];
1661                         t->used_replica = pos > 0;
1662                 } else {
1663                         t->osd = acting.primary;
1664                         t->used_replica = false;
1665                 }
1666         }
1667
1668         if (unpaused || legacy_change || force_resend || split)
1669                 ct_res = CALC_TARGET_NEED_RESEND;
1670         else
1671                 ct_res = CALC_TARGET_NO_ACTION;
1672
1673 out:
1674         dout("%s t %p -> %d%d%d%d ct_res %d osd%d\n", __func__, t, unpaused,
1675              legacy_change, force_resend, split, ct_res, t->osd);
1676         return ct_res;
1677 }
1678
1679 static struct ceph_spg_mapping *alloc_spg_mapping(void)
1680 {
1681         struct ceph_spg_mapping *spg;
1682
1683         spg = kmalloc(sizeof(*spg), GFP_NOIO);
1684         if (!spg)
1685                 return NULL;
1686
1687         RB_CLEAR_NODE(&spg->node);
1688         spg->backoffs = RB_ROOT;
1689         return spg;
1690 }
1691
1692 static void free_spg_mapping(struct ceph_spg_mapping *spg)
1693 {
1694         WARN_ON(!RB_EMPTY_NODE(&spg->node));
1695         WARN_ON(!RB_EMPTY_ROOT(&spg->backoffs));
1696
1697         kfree(spg);
1698 }
1699
1700 /*
1701  * rbtree of ceph_spg_mapping for handling map<spg_t, ...>, similar to
1702  * ceph_pg_mapping.  Used to track OSD backoffs -- a backoff [range] is
1703  * defined only within a specific spgid; it does not pass anything to
1704  * children on split, or to another primary.
1705  */
1706 DEFINE_RB_FUNCS2(spg_mapping, struct ceph_spg_mapping, spgid, ceph_spg_compare,
1707                  RB_BYPTR, const struct ceph_spg *, node)
1708
1709 static u64 hoid_get_bitwise_key(const struct ceph_hobject_id *hoid)
1710 {
1711         return hoid->is_max ? 0x100000000ull : hoid->hash_reverse_bits;
1712 }
1713
1714 static void hoid_get_effective_key(const struct ceph_hobject_id *hoid,
1715                                    void **pkey, size_t *pkey_len)
1716 {
1717         if (hoid->key_len) {
1718                 *pkey = hoid->key;
1719                 *pkey_len = hoid->key_len;
1720         } else {
1721                 *pkey = hoid->oid;
1722                 *pkey_len = hoid->oid_len;
1723         }
1724 }
1725
1726 static int compare_names(const void *name1, size_t name1_len,
1727                          const void *name2, size_t name2_len)
1728 {
1729         int ret;
1730
1731         ret = memcmp(name1, name2, min(name1_len, name2_len));
1732         if (!ret) {
1733                 if (name1_len < name2_len)
1734                         ret = -1;
1735                 else if (name1_len > name2_len)
1736                         ret = 1;
1737         }
1738         return ret;
1739 }
1740
1741 static int hoid_compare(const struct ceph_hobject_id *lhs,
1742                         const struct ceph_hobject_id *rhs)
1743 {
1744         void *effective_key1, *effective_key2;
1745         size_t effective_key1_len, effective_key2_len;
1746         int ret;
1747
1748         if (lhs->is_max < rhs->is_max)
1749                 return -1;
1750         if (lhs->is_max > rhs->is_max)
1751                 return 1;
1752
1753         if (lhs->pool < rhs->pool)
1754                 return -1;
1755         if (lhs->pool > rhs->pool)
1756                 return 1;
1757
1758         if (hoid_get_bitwise_key(lhs) < hoid_get_bitwise_key(rhs))
1759                 return -1;
1760         if (hoid_get_bitwise_key(lhs) > hoid_get_bitwise_key(rhs))
1761                 return 1;
1762
1763         ret = compare_names(lhs->nspace, lhs->nspace_len,
1764                             rhs->nspace, rhs->nspace_len);
1765         if (ret)
1766                 return ret;
1767
1768         hoid_get_effective_key(lhs, &effective_key1, &effective_key1_len);
1769         hoid_get_effective_key(rhs, &effective_key2, &effective_key2_len);
1770         ret = compare_names(effective_key1, effective_key1_len,
1771                             effective_key2, effective_key2_len);
1772         if (ret)
1773                 return ret;
1774
1775         ret = compare_names(lhs->oid, lhs->oid_len, rhs->oid, rhs->oid_len);
1776         if (ret)
1777                 return ret;
1778
1779         if (lhs->snapid < rhs->snapid)
1780                 return -1;
1781         if (lhs->snapid > rhs->snapid)
1782                 return 1;
1783
1784         return 0;
1785 }
1786
1787 /*
1788  * For decoding ->begin and ->end of MOSDBackoff only -- no MIN/MAX
1789  * compat stuff here.
1790  *
1791  * Assumes @hoid is zero-initialized.
1792  */
1793 static int decode_hoid(void **p, void *end, struct ceph_hobject_id *hoid)
1794 {
1795         u8 struct_v;
1796         u32 struct_len;
1797         int ret;
1798
1799         ret = ceph_start_decoding(p, end, 4, "hobject_t", &struct_v,
1800                                   &struct_len);
1801         if (ret)
1802                 return ret;
1803
1804         if (struct_v < 4) {
1805                 pr_err("got struct_v %d < 4 of hobject_t\n", struct_v);
1806                 goto e_inval;
1807         }
1808
1809         hoid->key = ceph_extract_encoded_string(p, end, &hoid->key_len,
1810                                                 GFP_NOIO);
1811         if (IS_ERR(hoid->key)) {
1812                 ret = PTR_ERR(hoid->key);
1813                 hoid->key = NULL;
1814                 return ret;
1815         }
1816
1817         hoid->oid = ceph_extract_encoded_string(p, end, &hoid->oid_len,
1818                                                 GFP_NOIO);
1819         if (IS_ERR(hoid->oid)) {
1820                 ret = PTR_ERR(hoid->oid);
1821                 hoid->oid = NULL;
1822                 return ret;
1823         }
1824
1825         ceph_decode_64_safe(p, end, hoid->snapid, e_inval);
1826         ceph_decode_32_safe(p, end, hoid->hash, e_inval);
1827         ceph_decode_8_safe(p, end, hoid->is_max, e_inval);
1828
1829         hoid->nspace = ceph_extract_encoded_string(p, end, &hoid->nspace_len,
1830                                                    GFP_NOIO);
1831         if (IS_ERR(hoid->nspace)) {
1832                 ret = PTR_ERR(hoid->nspace);
1833                 hoid->nspace = NULL;
1834                 return ret;
1835         }
1836
1837         ceph_decode_64_safe(p, end, hoid->pool, e_inval);
1838
1839         ceph_hoid_build_hash_cache(hoid);
1840         return 0;
1841
1842 e_inval:
1843         return -EINVAL;
1844 }
1845
1846 static int hoid_encoding_size(const struct ceph_hobject_id *hoid)
1847 {
1848         return 8 + 4 + 1 + 8 + /* snapid, hash, is_max, pool */
1849                4 + hoid->key_len + 4 + hoid->oid_len + 4 + hoid->nspace_len;
1850 }
1851
1852 static void encode_hoid(void **p, void *end, const struct ceph_hobject_id *hoid)
1853 {
1854         ceph_start_encoding(p, 4, 3, hoid_encoding_size(hoid));
1855         ceph_encode_string(p, end, hoid->key, hoid->key_len);
1856         ceph_encode_string(p, end, hoid->oid, hoid->oid_len);
1857         ceph_encode_64(p, hoid->snapid);
1858         ceph_encode_32(p, hoid->hash);
1859         ceph_encode_8(p, hoid->is_max);
1860         ceph_encode_string(p, end, hoid->nspace, hoid->nspace_len);
1861         ceph_encode_64(p, hoid->pool);
1862 }
1863
1864 static void free_hoid(struct ceph_hobject_id *hoid)
1865 {
1866         if (hoid) {
1867                 kfree(hoid->key);
1868                 kfree(hoid->oid);
1869                 kfree(hoid->nspace);
1870                 kfree(hoid);
1871         }
1872 }
1873
1874 static struct ceph_osd_backoff *alloc_backoff(void)
1875 {
1876         struct ceph_osd_backoff *backoff;
1877
1878         backoff = kzalloc(sizeof(*backoff), GFP_NOIO);
1879         if (!backoff)
1880                 return NULL;
1881
1882         RB_CLEAR_NODE(&backoff->spg_node);
1883         RB_CLEAR_NODE(&backoff->id_node);
1884         return backoff;
1885 }
1886
1887 static void free_backoff(struct ceph_osd_backoff *backoff)
1888 {
1889         WARN_ON(!RB_EMPTY_NODE(&backoff->spg_node));
1890         WARN_ON(!RB_EMPTY_NODE(&backoff->id_node));
1891
1892         free_hoid(backoff->begin);
1893         free_hoid(backoff->end);
1894         kfree(backoff);
1895 }
1896
1897 /*
1898  * Within a specific spgid, backoffs are managed by ->begin hoid.
1899  */
1900 DEFINE_RB_INSDEL_FUNCS2(backoff, struct ceph_osd_backoff, begin, hoid_compare,
1901                         RB_BYVAL, spg_node);
1902
1903 static struct ceph_osd_backoff *lookup_containing_backoff(struct rb_root *root,
1904                                             const struct ceph_hobject_id *hoid)
1905 {
1906         struct rb_node *n = root->rb_node;
1907
1908         while (n) {
1909                 struct ceph_osd_backoff *cur =
1910                     rb_entry(n, struct ceph_osd_backoff, spg_node);
1911                 int cmp;
1912
1913                 cmp = hoid_compare(hoid, cur->begin);
1914                 if (cmp < 0) {
1915                         n = n->rb_left;
1916                 } else if (cmp > 0) {
1917                         if (hoid_compare(hoid, cur->end) < 0)
1918                                 return cur;
1919
1920                         n = n->rb_right;
1921                 } else {
1922                         return cur;
1923                 }
1924         }
1925
1926         return NULL;
1927 }
1928
1929 /*
1930  * Each backoff has a unique id within its OSD session.
1931  */
1932 DEFINE_RB_FUNCS(backoff_by_id, struct ceph_osd_backoff, id, id_node)
1933
1934 static void clear_backoffs(struct ceph_osd *osd)
1935 {
1936         while (!RB_EMPTY_ROOT(&osd->o_backoff_mappings)) {
1937                 struct ceph_spg_mapping *spg =
1938                     rb_entry(rb_first(&osd->o_backoff_mappings),
1939                              struct ceph_spg_mapping, node);
1940
1941                 while (!RB_EMPTY_ROOT(&spg->backoffs)) {
1942                         struct ceph_osd_backoff *backoff =
1943                             rb_entry(rb_first(&spg->backoffs),
1944                                      struct ceph_osd_backoff, spg_node);
1945
1946                         erase_backoff(&spg->backoffs, backoff);
1947                         erase_backoff_by_id(&osd->o_backoffs_by_id, backoff);
1948                         free_backoff(backoff);
1949                 }
1950                 erase_spg_mapping(&osd->o_backoff_mappings, spg);
1951                 free_spg_mapping(spg);
1952         }
1953 }
1954
1955 /*
1956  * Set up a temporary, non-owning view into @t.
1957  */
1958 static void hoid_fill_from_target(struct ceph_hobject_id *hoid,
1959                                   const struct ceph_osd_request_target *t)
1960 {
1961         hoid->key = NULL;
1962         hoid->key_len = 0;
1963         hoid->oid = t->target_oid.name;
1964         hoid->oid_len = t->target_oid.name_len;
1965         hoid->snapid = CEPH_NOSNAP;
1966         hoid->hash = t->pgid.seed;
1967         hoid->is_max = false;
1968         if (t->target_oloc.pool_ns) {
1969                 hoid->nspace = t->target_oloc.pool_ns->str;
1970                 hoid->nspace_len = t->target_oloc.pool_ns->len;
1971         } else {
1972                 hoid->nspace = NULL;
1973                 hoid->nspace_len = 0;
1974         }
1975         hoid->pool = t->target_oloc.pool;
1976         ceph_hoid_build_hash_cache(hoid);
1977 }
1978
1979 static bool should_plug_request(struct ceph_osd_request *req)
1980 {
1981         struct ceph_osd *osd = req->r_osd;
1982         struct ceph_spg_mapping *spg;
1983         struct ceph_osd_backoff *backoff;
1984         struct ceph_hobject_id hoid;
1985
1986         spg = lookup_spg_mapping(&osd->o_backoff_mappings, &req->r_t.spgid);
1987         if (!spg)
1988                 return false;
1989
1990         hoid_fill_from_target(&hoid, &req->r_t);
1991         backoff = lookup_containing_backoff(&spg->backoffs, &hoid);
1992         if (!backoff)
1993                 return false;
1994
1995         dout("%s req %p tid %llu backoff osd%d spgid %llu.%xs%d id %llu\n",
1996              __func__, req, req->r_tid, osd->o_osd, backoff->spgid.pgid.pool,
1997              backoff->spgid.pgid.seed, backoff->spgid.shard, backoff->id);
1998         return true;
1999 }
2000
2001 /*
2002  * Keep get_num_data_items() in sync with this function.
2003  */
2004 static void setup_request_data(struct ceph_osd_request *req)
2005 {
2006         struct ceph_msg *request_msg = req->r_request;
2007         struct ceph_msg *reply_msg = req->r_reply;
2008         struct ceph_osd_req_op *op;
2009
2010         if (req->r_request->num_data_items || req->r_reply->num_data_items)
2011                 return;
2012
2013         WARN_ON(request_msg->data_length || reply_msg->data_length);
2014         for (op = req->r_ops; op != &req->r_ops[req->r_num_ops]; op++) {
2015                 switch (op->op) {
2016                 /* request */
2017                 case CEPH_OSD_OP_WRITE:
2018                 case CEPH_OSD_OP_WRITEFULL:
2019                         WARN_ON(op->indata_len != op->extent.length);
2020                         ceph_osdc_msg_data_add(request_msg,
2021                                                &op->extent.osd_data);
2022                         break;
2023                 case CEPH_OSD_OP_SETXATTR:
2024                 case CEPH_OSD_OP_CMPXATTR:
2025                         WARN_ON(op->indata_len != op->xattr.name_len +
2026                                                   op->xattr.value_len);
2027                         ceph_osdc_msg_data_add(request_msg,
2028                                                &op->xattr.osd_data);
2029                         break;
2030                 case CEPH_OSD_OP_NOTIFY_ACK:
2031                         ceph_osdc_msg_data_add(request_msg,
2032                                                &op->notify_ack.request_data);
2033                         break;
2034                 case CEPH_OSD_OP_COPY_FROM2:
2035                         ceph_osdc_msg_data_add(request_msg,
2036                                                &op->copy_from.osd_data);
2037                         break;
2038
2039                 /* reply */
2040                 case CEPH_OSD_OP_STAT:
2041                         ceph_osdc_msg_data_add(reply_msg,
2042                                                &op->raw_data_in);
2043                         break;
2044                 case CEPH_OSD_OP_READ:
2045                         ceph_osdc_msg_data_add(reply_msg,
2046                                                &op->extent.osd_data);
2047                         break;
2048                 case CEPH_OSD_OP_LIST_WATCHERS:
2049                         ceph_osdc_msg_data_add(reply_msg,
2050                                                &op->list_watchers.response_data);
2051                         break;
2052
2053                 /* both */
2054                 case CEPH_OSD_OP_CALL:
2055                         WARN_ON(op->indata_len != op->cls.class_len +
2056                                                   op->cls.method_len +
2057                                                   op->cls.indata_len);
2058                         ceph_osdc_msg_data_add(request_msg,
2059                                                &op->cls.request_info);
2060                         /* optional, can be NONE */
2061                         ceph_osdc_msg_data_add(request_msg,
2062                                                &op->cls.request_data);
2063                         /* optional, can be NONE */
2064                         ceph_osdc_msg_data_add(reply_msg,
2065                                                &op->cls.response_data);
2066                         break;
2067                 case CEPH_OSD_OP_NOTIFY:
2068                         ceph_osdc_msg_data_add(request_msg,
2069                                                &op->notify.request_data);
2070                         ceph_osdc_msg_data_add(reply_msg,
2071                                                &op->notify.response_data);
2072                         break;
2073                 }
2074         }
2075 }
2076
2077 static void encode_pgid(void **p, const struct ceph_pg *pgid)
2078 {
2079         ceph_encode_8(p, 1);
2080         ceph_encode_64(p, pgid->pool);
2081         ceph_encode_32(p, pgid->seed);
2082         ceph_encode_32(p, -1); /* preferred */
2083 }
2084
2085 static void encode_spgid(void **p, const struct ceph_spg *spgid)
2086 {
2087         ceph_start_encoding(p, 1, 1, CEPH_PGID_ENCODING_LEN + 1);
2088         encode_pgid(p, &spgid->pgid);
2089         ceph_encode_8(p, spgid->shard);
2090 }
2091
2092 static void encode_oloc(void **p, void *end,
2093                         const struct ceph_object_locator *oloc)
2094 {
2095         ceph_start_encoding(p, 5, 4, ceph_oloc_encoding_size(oloc));
2096         ceph_encode_64(p, oloc->pool);
2097         ceph_encode_32(p, -1); /* preferred */
2098         ceph_encode_32(p, 0);  /* key len */
2099         if (oloc->pool_ns)
2100                 ceph_encode_string(p, end, oloc->pool_ns->str,
2101                                    oloc->pool_ns->len);
2102         else
2103                 ceph_encode_32(p, 0);
2104 }
2105
2106 static void encode_request_partial(struct ceph_osd_request *req,
2107                                    struct ceph_msg *msg)
2108 {
2109         void *p = msg->front.iov_base;
2110         void *const end = p + msg->front_alloc_len;
2111         u32 data_len = 0;
2112         int i;
2113
2114         if (req->r_flags & CEPH_OSD_FLAG_WRITE) {
2115                 /* snapshots aren't writeable */
2116                 WARN_ON(req->r_snapid != CEPH_NOSNAP);
2117         } else {
2118                 WARN_ON(req->r_mtime.tv_sec || req->r_mtime.tv_nsec ||
2119                         req->r_data_offset || req->r_snapc);
2120         }
2121
2122         setup_request_data(req);
2123
2124         encode_spgid(&p, &req->r_t.spgid); /* actual spg */
2125         ceph_encode_32(&p, req->r_t.pgid.seed); /* raw hash */
2126         ceph_encode_32(&p, req->r_osdc->osdmap->epoch);
2127         ceph_encode_32(&p, req->r_flags);
2128
2129         /* reqid */
2130         ceph_start_encoding(&p, 2, 2, sizeof(struct ceph_osd_reqid));
2131         memset(p, 0, sizeof(struct ceph_osd_reqid));
2132         p += sizeof(struct ceph_osd_reqid);
2133
2134         /* trace */
2135         memset(p, 0, sizeof(struct ceph_blkin_trace_info));
2136         p += sizeof(struct ceph_blkin_trace_info);
2137
2138         ceph_encode_32(&p, 0); /* client_inc, always 0 */
2139         ceph_encode_timespec64(p, &req->r_mtime);
2140         p += sizeof(struct ceph_timespec);
2141
2142         encode_oloc(&p, end, &req->r_t.target_oloc);
2143         ceph_encode_string(&p, end, req->r_t.target_oid.name,
2144                            req->r_t.target_oid.name_len);
2145
2146         /* ops, can imply data */
2147         ceph_encode_16(&p, req->r_num_ops);
2148         for (i = 0; i < req->r_num_ops; i++) {
2149                 data_len += osd_req_encode_op(p, &req->r_ops[i]);
2150                 p += sizeof(struct ceph_osd_op);
2151         }
2152
2153         ceph_encode_64(&p, req->r_snapid); /* snapid */
2154         if (req->r_snapc) {
2155                 ceph_encode_64(&p, req->r_snapc->seq);
2156                 ceph_encode_32(&p, req->r_snapc->num_snaps);
2157                 for (i = 0; i < req->r_snapc->num_snaps; i++)
2158                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
2159         } else {
2160                 ceph_encode_64(&p, 0); /* snap_seq */
2161                 ceph_encode_32(&p, 0); /* snaps len */
2162         }
2163
2164         ceph_encode_32(&p, req->r_attempts); /* retry_attempt */
2165         BUG_ON(p > end - 8); /* space for features */
2166
2167         msg->hdr.version = cpu_to_le16(8); /* MOSDOp v8 */
2168         /* front_len is finalized in encode_request_finish() */
2169         msg->front.iov_len = p - msg->front.iov_base;
2170         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2171         msg->hdr.data_len = cpu_to_le32(data_len);
2172         /*
2173          * The header "data_off" is a hint to the receiver allowing it
2174          * to align received data into its buffers such that there's no
2175          * need to re-copy it before writing it to disk (direct I/O).
2176          */
2177         msg->hdr.data_off = cpu_to_le16(req->r_data_offset);
2178
2179         dout("%s req %p msg %p oid %s oid_len %d\n", __func__, req, msg,
2180              req->r_t.target_oid.name, req->r_t.target_oid.name_len);
2181 }
2182
2183 static void encode_request_finish(struct ceph_msg *msg)
2184 {
2185         void *p = msg->front.iov_base;
2186         void *const partial_end = p + msg->front.iov_len;
2187         void *const end = p + msg->front_alloc_len;
2188
2189         if (CEPH_HAVE_FEATURE(msg->con->peer_features, RESEND_ON_SPLIT)) {
2190                 /* luminous OSD -- encode features and be done */
2191                 p = partial_end;
2192                 ceph_encode_64(&p, msg->con->peer_features);
2193         } else {
2194                 struct {
2195                         char spgid[CEPH_ENCODING_START_BLK_LEN +
2196                                    CEPH_PGID_ENCODING_LEN + 1];
2197                         __le32 hash;
2198                         __le32 epoch;
2199                         __le32 flags;
2200                         char reqid[CEPH_ENCODING_START_BLK_LEN +
2201                                    sizeof(struct ceph_osd_reqid)];
2202                         char trace[sizeof(struct ceph_blkin_trace_info)];
2203                         __le32 client_inc;
2204                         struct ceph_timespec mtime;
2205                 } __packed head;
2206                 struct ceph_pg pgid;
2207                 void *oloc, *oid, *tail;
2208                 int oloc_len, oid_len, tail_len;
2209                 int len;
2210
2211                 /*
2212                  * Pre-luminous OSD -- reencode v8 into v4 using @head
2213                  * as a temporary buffer.  Encode the raw PG; the rest
2214                  * is just a matter of moving oloc, oid and tail blobs
2215                  * around.
2216                  */
2217                 memcpy(&head, p, sizeof(head));
2218                 p += sizeof(head);
2219
2220                 oloc = p;
2221                 p += CEPH_ENCODING_START_BLK_LEN;
2222                 pgid.pool = ceph_decode_64(&p);
2223                 p += 4 + 4; /* preferred, key len */
2224                 len = ceph_decode_32(&p);
2225                 p += len;   /* nspace */
2226                 oloc_len = p - oloc;
2227
2228                 oid = p;
2229                 len = ceph_decode_32(&p);
2230                 p += len;
2231                 oid_len = p - oid;
2232
2233                 tail = p;
2234                 tail_len = partial_end - p;
2235
2236                 p = msg->front.iov_base;
2237                 ceph_encode_copy(&p, &head.client_inc, sizeof(head.client_inc));
2238                 ceph_encode_copy(&p, &head.epoch, sizeof(head.epoch));
2239                 ceph_encode_copy(&p, &head.flags, sizeof(head.flags));
2240                 ceph_encode_copy(&p, &head.mtime, sizeof(head.mtime));
2241
2242                 /* reassert_version */
2243                 memset(p, 0, sizeof(struct ceph_eversion));
2244                 p += sizeof(struct ceph_eversion);
2245
2246                 BUG_ON(p >= oloc);
2247                 memmove(p, oloc, oloc_len);
2248                 p += oloc_len;
2249
2250                 pgid.seed = le32_to_cpu(head.hash);
2251                 encode_pgid(&p, &pgid); /* raw pg */
2252
2253                 BUG_ON(p >= oid);
2254                 memmove(p, oid, oid_len);
2255                 p += oid_len;
2256
2257                 /* tail -- ops, snapid, snapc, retry_attempt */
2258                 BUG_ON(p >= tail);
2259                 memmove(p, tail, tail_len);
2260                 p += tail_len;
2261
2262                 msg->hdr.version = cpu_to_le16(4); /* MOSDOp v4 */
2263         }
2264
2265         BUG_ON(p > end);
2266         msg->front.iov_len = p - msg->front.iov_base;
2267         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
2268
2269         dout("%s msg %p tid %llu %u+%u+%u v%d\n", __func__, msg,
2270              le64_to_cpu(msg->hdr.tid), le32_to_cpu(msg->hdr.front_len),
2271              le32_to_cpu(msg->hdr.middle_len), le32_to_cpu(msg->hdr.data_len),
2272              le16_to_cpu(msg->hdr.version));
2273 }
2274
2275 /*
2276  * @req has to be assigned a tid and registered.
2277  */
2278 static void send_request(struct ceph_osd_request *req)
2279 {
2280         struct ceph_osd *osd = req->r_osd;
2281
2282         verify_osd_locked(osd);
2283         WARN_ON(osd->o_osd != req->r_t.osd);
2284
2285         /* backoff? */
2286         if (should_plug_request(req))
2287                 return;
2288
2289         /*
2290          * We may have a previously queued request message hanging
2291          * around.  Cancel it to avoid corrupting the msgr.
2292          */
2293         if (req->r_sent)
2294                 ceph_msg_revoke(req->r_request);
2295
2296         req->r_flags |= CEPH_OSD_FLAG_KNOWN_REDIR;
2297         if (req->r_attempts)
2298                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
2299         else
2300                 WARN_ON(req->r_flags & CEPH_OSD_FLAG_RETRY);
2301
2302         encode_request_partial(req, req->r_request);
2303
2304         dout("%s req %p tid %llu to pgid %llu.%x spgid %llu.%xs%d osd%d e%u flags 0x%x attempt %d\n",
2305              __func__, req, req->r_tid, req->r_t.pgid.pool, req->r_t.pgid.seed,
2306              req->r_t.spgid.pgid.pool, req->r_t.spgid.pgid.seed,
2307              req->r_t.spgid.shard, osd->o_osd, req->r_t.epoch, req->r_flags,
2308              req->r_attempts);
2309
2310         req->r_t.paused = false;
2311         req->r_stamp = jiffies;
2312         req->r_attempts++;
2313
2314         req->r_sent = osd->o_incarnation;
2315         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
2316         ceph_con_send(&osd->o_con, ceph_msg_get(req->r_request));
2317 }
2318
2319 static void maybe_request_map(struct ceph_osd_client *osdc)
2320 {
2321         bool continuous = false;
2322
2323         verify_osdc_locked(osdc);
2324         WARN_ON(!osdc->osdmap->epoch);
2325
2326         if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
2327             ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD) ||
2328             ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) {
2329                 dout("%s osdc %p continuous\n", __func__, osdc);
2330                 continuous = true;
2331         } else {
2332                 dout("%s osdc %p onetime\n", __func__, osdc);
2333         }
2334
2335         if (ceph_monc_want_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
2336                                osdc->osdmap->epoch + 1, continuous))
2337                 ceph_monc_renew_subs(&osdc->client->monc);
2338 }
2339
2340 static void complete_request(struct ceph_osd_request *req, int err);
2341 static void send_map_check(struct ceph_osd_request *req);
2342
2343 static void __submit_request(struct ceph_osd_request *req, bool wrlocked)
2344 {
2345         struct ceph_osd_client *osdc = req->r_osdc;
2346         struct ceph_osd *osd;
2347         enum calc_target_result ct_res;
2348         int err = 0;
2349         bool need_send = false;
2350         bool promoted = false;
2351
2352         WARN_ON(req->r_tid);
2353         dout("%s req %p wrlocked %d\n", __func__, req, wrlocked);
2354
2355 again:
2356         ct_res = calc_target(osdc, &req->r_t, false);
2357         if (ct_res == CALC_TARGET_POOL_DNE && !wrlocked)
2358                 goto promote;
2359
2360         osd = lookup_create_osd(osdc, req->r_t.osd, wrlocked);
2361         if (IS_ERR(osd)) {
2362                 WARN_ON(PTR_ERR(osd) != -EAGAIN || wrlocked);
2363                 goto promote;
2364         }
2365
2366         if (osdc->abort_err) {
2367                 dout("req %p abort_err %d\n", req, osdc->abort_err);
2368                 err = osdc->abort_err;
2369         } else if (osdc->osdmap->epoch < osdc->epoch_barrier) {
2370                 dout("req %p epoch %u barrier %u\n", req, osdc->osdmap->epoch,
2371                      osdc->epoch_barrier);
2372                 req->r_t.paused = true;
2373                 maybe_request_map(osdc);
2374         } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
2375                    ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR)) {
2376                 dout("req %p pausewr\n", req);
2377                 req->r_t.paused = true;
2378                 maybe_request_map(osdc);
2379         } else if ((req->r_flags & CEPH_OSD_FLAG_READ) &&
2380                    ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) {
2381                 dout("req %p pauserd\n", req);
2382                 req->r_t.paused = true;
2383                 maybe_request_map(osdc);
2384         } else if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
2385                    !(req->r_flags & (CEPH_OSD_FLAG_FULL_TRY |
2386                                      CEPH_OSD_FLAG_FULL_FORCE)) &&
2387                    (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
2388                     pool_full(osdc, req->r_t.base_oloc.pool))) {
2389                 dout("req %p full/pool_full\n", req);
2390                 if (ceph_test_opt(osdc->client, ABORT_ON_FULL)) {
2391                         err = -ENOSPC;
2392                 } else {
2393                         pr_warn_ratelimited("FULL or reached pool quota\n");
2394                         req->r_t.paused = true;
2395                         maybe_request_map(osdc);
2396                 }
2397         } else if (!osd_homeless(osd)) {
2398                 need_send = true;
2399         } else {
2400                 maybe_request_map(osdc);
2401         }
2402
2403         mutex_lock(&osd->lock);
2404         /*
2405          * Assign the tid atomically with send_request() to protect
2406          * multiple writes to the same object from racing with each
2407          * other, resulting in out of order ops on the OSDs.
2408          */
2409         req->r_tid = atomic64_inc_return(&osdc->last_tid);
2410         link_request(osd, req);
2411         if (need_send)
2412                 send_request(req);
2413         else if (err)
2414                 complete_request(req, err);
2415         mutex_unlock(&osd->lock);
2416
2417         if (!err && ct_res == CALC_TARGET_POOL_DNE)
2418                 send_map_check(req);
2419
2420         if (promoted)
2421                 downgrade_write(&osdc->lock);
2422         return;
2423
2424 promote:
2425         up_read(&osdc->lock);
2426         down_write(&osdc->lock);
2427         wrlocked = true;
2428         promoted = true;
2429         goto again;
2430 }
2431
2432 static void account_request(struct ceph_osd_request *req)
2433 {
2434         struct ceph_osd_client *osdc = req->r_osdc;
2435
2436         WARN_ON(req->r_flags & (CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK));
2437         WARN_ON(!(req->r_flags & (CEPH_OSD_FLAG_READ | CEPH_OSD_FLAG_WRITE)));
2438
2439         req->r_flags |= CEPH_OSD_FLAG_ONDISK;
2440         req->r_flags |= osdc->client->options->osd_req_flags;
2441         atomic_inc(&osdc->num_requests);
2442
2443         req->r_start_stamp = jiffies;
2444         req->r_start_latency = ktime_get();
2445 }
2446
2447 static void submit_request(struct ceph_osd_request *req, bool wrlocked)
2448 {
2449         ceph_osdc_get_request(req);
2450         account_request(req);
2451         __submit_request(req, wrlocked);
2452 }
2453
2454 static void finish_request(struct ceph_osd_request *req)
2455 {
2456         struct ceph_osd_client *osdc = req->r_osdc;
2457
2458         WARN_ON(lookup_request_mc(&osdc->map_checks, req->r_tid));
2459         dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
2460
2461         req->r_end_latency = ktime_get();
2462
2463         if (req->r_osd)
2464                 unlink_request(req->r_osd, req);
2465         atomic_dec(&osdc->num_requests);
2466
2467         /*
2468          * If an OSD has failed or returned and a request has been sent
2469          * twice, it's possible to get a reply and end up here while the
2470          * request message is queued for delivery.  We will ignore the
2471          * reply, so not a big deal, but better to try and catch it.
2472          */
2473         ceph_msg_revoke(req->r_request);
2474         ceph_msg_revoke_incoming(req->r_reply);
2475 }
2476
2477 static void __complete_request(struct ceph_osd_request *req)
2478 {
2479         dout("%s req %p tid %llu cb %ps result %d\n", __func__, req,
2480              req->r_tid, req->r_callback, req->r_result);
2481
2482         if (req->r_callback)
2483                 req->r_callback(req);
2484         complete_all(&req->r_completion);
2485         ceph_osdc_put_request(req);
2486 }
2487
2488 static void complete_request_workfn(struct work_struct *work)
2489 {
2490         struct ceph_osd_request *req =
2491             container_of(work, struct ceph_osd_request, r_complete_work);
2492
2493         __complete_request(req);
2494 }
2495
2496 /*
2497  * This is open-coded in handle_reply().
2498  */
2499 static void complete_request(struct ceph_osd_request *req, int err)
2500 {
2501         dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err);
2502
2503         req->r_result = err;
2504         finish_request(req);
2505
2506         INIT_WORK(&req->r_complete_work, complete_request_workfn);
2507         queue_work(req->r_osdc->completion_wq, &req->r_complete_work);
2508 }
2509
2510 static void cancel_map_check(struct ceph_osd_request *req)
2511 {
2512         struct ceph_osd_client *osdc = req->r_osdc;
2513         struct ceph_osd_request *lookup_req;
2514
2515         verify_osdc_wrlocked(osdc);
2516
2517         lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid);
2518         if (!lookup_req)
2519                 return;
2520
2521         WARN_ON(lookup_req != req);
2522         erase_request_mc(&osdc->map_checks, req);
2523         ceph_osdc_put_request(req);
2524 }
2525
2526 static void cancel_request(struct ceph_osd_request *req)
2527 {
2528         dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
2529
2530         cancel_map_check(req);
2531         finish_request(req);
2532         complete_all(&req->r_completion);
2533         ceph_osdc_put_request(req);
2534 }
2535
2536 static void abort_request(struct ceph_osd_request *req, int err)
2537 {
2538         dout("%s req %p tid %llu err %d\n", __func__, req, req->r_tid, err);
2539
2540         cancel_map_check(req);
2541         complete_request(req, err);
2542 }
2543
2544 static int abort_fn(struct ceph_osd_request *req, void *arg)
2545 {
2546         int err = *(int *)arg;
2547
2548         abort_request(req, err);
2549         return 0; /* continue iteration */
2550 }
2551
2552 /*
2553  * Abort all in-flight requests with @err and arrange for all future
2554  * requests to be failed immediately.
2555  */
2556 void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err)
2557 {
2558         dout("%s osdc %p err %d\n", __func__, osdc, err);
2559         down_write(&osdc->lock);
2560         for_each_request(osdc, abort_fn, &err);
2561         osdc->abort_err = err;
2562         up_write(&osdc->lock);
2563 }
2564 EXPORT_SYMBOL(ceph_osdc_abort_requests);
2565
2566 void ceph_osdc_clear_abort_err(struct ceph_osd_client *osdc)
2567 {
2568         down_write(&osdc->lock);
2569         osdc->abort_err = 0;
2570         up_write(&osdc->lock);
2571 }
2572 EXPORT_SYMBOL(ceph_osdc_clear_abort_err);
2573
2574 static void update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb)
2575 {
2576         if (likely(eb > osdc->epoch_barrier)) {
2577                 dout("updating epoch_barrier from %u to %u\n",
2578                                 osdc->epoch_barrier, eb);
2579                 osdc->epoch_barrier = eb;
2580                 /* Request map if we're not to the barrier yet */
2581                 if (eb > osdc->osdmap->epoch)
2582                         maybe_request_map(osdc);
2583         }
2584 }
2585
2586 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb)
2587 {
2588         down_read(&osdc->lock);
2589         if (unlikely(eb > osdc->epoch_barrier)) {
2590                 up_read(&osdc->lock);
2591                 down_write(&osdc->lock);
2592                 update_epoch_barrier(osdc, eb);
2593                 up_write(&osdc->lock);
2594         } else {
2595                 up_read(&osdc->lock);
2596         }
2597 }
2598 EXPORT_SYMBOL(ceph_osdc_update_epoch_barrier);
2599
2600 /*
2601  * We can end up releasing caps as a result of abort_request().
2602  * In that case, we probably want to ensure that the cap release message
2603  * has an updated epoch barrier in it, so set the epoch barrier prior to
2604  * aborting the first request.
2605  */
2606 static int abort_on_full_fn(struct ceph_osd_request *req, void *arg)
2607 {
2608         struct ceph_osd_client *osdc = req->r_osdc;
2609         bool *victims = arg;
2610
2611         if ((req->r_flags & CEPH_OSD_FLAG_WRITE) &&
2612             (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
2613              pool_full(osdc, req->r_t.base_oloc.pool))) {
2614                 if (!*victims) {
2615                         update_epoch_barrier(osdc, osdc->osdmap->epoch);
2616                         *victims = true;
2617                 }
2618                 abort_request(req, -ENOSPC);
2619         }
2620
2621         return 0; /* continue iteration */
2622 }
2623
2624 /*
2625  * Drop all pending requests that are stalled waiting on a full condition to
2626  * clear, and complete them with ENOSPC as the return code. Set the
2627  * osdc->epoch_barrier to the latest map epoch that we've seen if any were
2628  * cancelled.
2629  */
2630 static void ceph_osdc_abort_on_full(struct ceph_osd_client *osdc)
2631 {
2632         bool victims = false;
2633
2634         if (ceph_test_opt(osdc->client, ABORT_ON_FULL) &&
2635             (ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) || have_pool_full(osdc)))
2636                 for_each_request(osdc, abort_on_full_fn, &victims);
2637 }
2638
2639 static void check_pool_dne(struct ceph_osd_request *req)
2640 {
2641         struct ceph_osd_client *osdc = req->r_osdc;
2642         struct ceph_osdmap *map = osdc->osdmap;
2643
2644         verify_osdc_wrlocked(osdc);
2645         WARN_ON(!map->epoch);
2646
2647         if (req->r_attempts) {
2648                 /*
2649                  * We sent a request earlier, which means that
2650                  * previously the pool existed, and now it does not
2651                  * (i.e., it was deleted).
2652                  */
2653                 req->r_map_dne_bound = map->epoch;
2654                 dout("%s req %p tid %llu pool disappeared\n", __func__, req,
2655                      req->r_tid);
2656         } else {
2657                 dout("%s req %p tid %llu map_dne_bound %u have %u\n", __func__,
2658                      req, req->r_tid, req->r_map_dne_bound, map->epoch);
2659         }
2660
2661         if (req->r_map_dne_bound) {
2662                 if (map->epoch >= req->r_map_dne_bound) {
2663                         /* we had a new enough map */
2664                         pr_info_ratelimited("tid %llu pool does not exist\n",
2665                                             req->r_tid);
2666                         complete_request(req, -ENOENT);
2667                 }
2668         } else {
2669                 send_map_check(req);
2670         }
2671 }
2672
2673 static void map_check_cb(struct ceph_mon_generic_request *greq)
2674 {
2675         struct ceph_osd_client *osdc = &greq->monc->client->osdc;
2676         struct ceph_osd_request *req;
2677         u64 tid = greq->private_data;
2678
2679         WARN_ON(greq->result || !greq->u.newest);
2680
2681         down_write(&osdc->lock);
2682         req = lookup_request_mc(&osdc->map_checks, tid);
2683         if (!req) {
2684                 dout("%s tid %llu dne\n", __func__, tid);
2685                 goto out_unlock;
2686         }
2687
2688         dout("%s req %p tid %llu map_dne_bound %u newest %llu\n", __func__,
2689              req, req->r_tid, req->r_map_dne_bound, greq->u.newest);
2690         if (!req->r_map_dne_bound)
2691                 req->r_map_dne_bound = greq->u.newest;
2692         erase_request_mc(&osdc->map_checks, req);
2693         check_pool_dne(req);
2694
2695         ceph_osdc_put_request(req);
2696 out_unlock:
2697         up_write(&osdc->lock);
2698 }
2699
2700 static void send_map_check(struct ceph_osd_request *req)
2701 {
2702         struct ceph_osd_client *osdc = req->r_osdc;
2703         struct ceph_osd_request *lookup_req;
2704         int ret;
2705
2706         verify_osdc_wrlocked(osdc);
2707
2708         lookup_req = lookup_request_mc(&osdc->map_checks, req->r_tid);
2709         if (lookup_req) {
2710                 WARN_ON(lookup_req != req);
2711                 return;
2712         }
2713
2714         ceph_osdc_get_request(req);
2715         insert_request_mc(&osdc->map_checks, req);
2716         ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap",
2717                                           map_check_cb, req->r_tid);
2718         WARN_ON(ret);
2719 }
2720
2721 /*
2722  * lingering requests, watch/notify v2 infrastructure
2723  */
2724 static void linger_release(struct kref *kref)
2725 {
2726         struct ceph_osd_linger_request *lreq =
2727             container_of(kref, struct ceph_osd_linger_request, kref);
2728
2729         dout("%s lreq %p reg_req %p ping_req %p\n", __func__, lreq,
2730              lreq->reg_req, lreq->ping_req);
2731         WARN_ON(!RB_EMPTY_NODE(&lreq->node));
2732         WARN_ON(!RB_EMPTY_NODE(&lreq->osdc_node));
2733         WARN_ON(!RB_EMPTY_NODE(&lreq->mc_node));
2734         WARN_ON(!list_empty(&lreq->scan_item));
2735         WARN_ON(!list_empty(&lreq->pending_lworks));
2736         WARN_ON(lreq->osd);
2737
2738         if (lreq->reg_req)
2739                 ceph_osdc_put_request(lreq->reg_req);
2740         if (lreq->ping_req)
2741                 ceph_osdc_put_request(lreq->ping_req);
2742         target_destroy(&lreq->t);
2743         kfree(lreq);
2744 }
2745
2746 static void linger_put(struct ceph_osd_linger_request *lreq)
2747 {
2748         if (lreq)
2749                 kref_put(&lreq->kref, linger_release);
2750 }
2751
2752 static struct ceph_osd_linger_request *
2753 linger_get(struct ceph_osd_linger_request *lreq)
2754 {
2755         kref_get(&lreq->kref);
2756         return lreq;
2757 }
2758
2759 static struct ceph_osd_linger_request *
2760 linger_alloc(struct ceph_osd_client *osdc)
2761 {
2762         struct ceph_osd_linger_request *lreq;
2763
2764         lreq = kzalloc(sizeof(*lreq), GFP_NOIO);
2765         if (!lreq)
2766                 return NULL;
2767
2768         kref_init(&lreq->kref);
2769         mutex_init(&lreq->lock);
2770         RB_CLEAR_NODE(&lreq->node);
2771         RB_CLEAR_NODE(&lreq->osdc_node);
2772         RB_CLEAR_NODE(&lreq->mc_node);
2773         INIT_LIST_HEAD(&lreq->scan_item);
2774         INIT_LIST_HEAD(&lreq->pending_lworks);
2775         init_completion(&lreq->reg_commit_wait);
2776         init_completion(&lreq->notify_finish_wait);
2777
2778         lreq->osdc = osdc;
2779         target_init(&lreq->t);
2780
2781         dout("%s lreq %p\n", __func__, lreq);
2782         return lreq;
2783 }
2784
2785 DEFINE_RB_INSDEL_FUNCS(linger, struct ceph_osd_linger_request, linger_id, node)
2786 DEFINE_RB_FUNCS(linger_osdc, struct ceph_osd_linger_request, linger_id, osdc_node)
2787 DEFINE_RB_FUNCS(linger_mc, struct ceph_osd_linger_request, linger_id, mc_node)
2788
2789 /*
2790  * Create linger request <-> OSD session relation.
2791  *
2792  * @lreq has to be registered, @osd may be homeless.
2793  */
2794 static void link_linger(struct ceph_osd *osd,
2795                         struct ceph_osd_linger_request *lreq)
2796 {
2797         verify_osd_locked(osd);
2798         WARN_ON(!lreq->linger_id || lreq->osd);
2799         dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd,
2800              osd->o_osd, lreq, lreq->linger_id);
2801
2802         if (!osd_homeless(osd))
2803                 __remove_osd_from_lru(osd);
2804         else
2805                 atomic_inc(&osd->o_osdc->num_homeless);
2806
2807         get_osd(osd);
2808         insert_linger(&osd->o_linger_requests, lreq);
2809         lreq->osd = osd;
2810 }
2811
2812 static void unlink_linger(struct ceph_osd *osd,
2813                           struct ceph_osd_linger_request *lreq)
2814 {
2815         verify_osd_locked(osd);
2816         WARN_ON(lreq->osd != osd);
2817         dout("%s osd %p osd%d lreq %p linger_id %llu\n", __func__, osd,
2818              osd->o_osd, lreq, lreq->linger_id);
2819
2820         lreq->osd = NULL;
2821         erase_linger(&osd->o_linger_requests, lreq);
2822         put_osd(osd);
2823
2824         if (!osd_homeless(osd))
2825                 maybe_move_osd_to_lru(osd);
2826         else
2827                 atomic_dec(&osd->o_osdc->num_homeless);
2828 }
2829
2830 static bool __linger_registered(struct ceph_osd_linger_request *lreq)
2831 {
2832         verify_osdc_locked(lreq->osdc);
2833
2834         return !RB_EMPTY_NODE(&lreq->osdc_node);
2835 }
2836
2837 static bool linger_registered(struct ceph_osd_linger_request *lreq)
2838 {
2839         struct ceph_osd_client *osdc = lreq->osdc;
2840         bool registered;
2841
2842         down_read(&osdc->lock);
2843         registered = __linger_registered(lreq);
2844         up_read(&osdc->lock);
2845
2846         return registered;
2847 }
2848
2849 static void linger_register(struct ceph_osd_linger_request *lreq)
2850 {
2851         struct ceph_osd_client *osdc = lreq->osdc;
2852
2853         verify_osdc_wrlocked(osdc);
2854         WARN_ON(lreq->linger_id);
2855
2856         linger_get(lreq);
2857         lreq->linger_id = ++osdc->last_linger_id;
2858         insert_linger_osdc(&osdc->linger_requests, lreq);
2859 }
2860
2861 static void linger_unregister(struct ceph_osd_linger_request *lreq)
2862 {
2863         struct ceph_osd_client *osdc = lreq->osdc;
2864
2865         verify_osdc_wrlocked(osdc);
2866
2867         erase_linger_osdc(&osdc->linger_requests, lreq);
2868         linger_put(lreq);
2869 }
2870
2871 static void cancel_linger_request(struct ceph_osd_request *req)
2872 {
2873         struct ceph_osd_linger_request *lreq = req->r_priv;
2874
2875         WARN_ON(!req->r_linger);
2876         cancel_request(req);
2877         linger_put(lreq);
2878 }
2879
2880 struct linger_work {
2881         struct work_struct work;
2882         struct ceph_osd_linger_request *lreq;
2883         struct list_head pending_item;
2884         unsigned long queued_stamp;
2885
2886         union {
2887                 struct {
2888                         u64 notify_id;
2889                         u64 notifier_id;
2890                         void *payload; /* points into @msg front */
2891                         size_t payload_len;
2892
2893                         struct ceph_msg *msg; /* for ceph_msg_put() */
2894                 } notify;
2895                 struct {
2896                         int err;
2897                 } error;
2898         };
2899 };
2900
2901 static struct linger_work *lwork_alloc(struct ceph_osd_linger_request *lreq,
2902                                        work_func_t workfn)
2903 {
2904         struct linger_work *lwork;
2905
2906         lwork = kzalloc(sizeof(*lwork), GFP_NOIO);
2907         if (!lwork)
2908                 return NULL;
2909
2910         INIT_WORK(&lwork->work, workfn);
2911         INIT_LIST_HEAD(&lwork->pending_item);
2912         lwork->lreq = linger_get(lreq);
2913
2914         return lwork;
2915 }
2916
2917 static void lwork_free(struct linger_work *lwork)
2918 {
2919         struct ceph_osd_linger_request *lreq = lwork->lreq;
2920
2921         mutex_lock(&lreq->lock);
2922         list_del(&lwork->pending_item);
2923         mutex_unlock(&lreq->lock);
2924
2925         linger_put(lreq);
2926         kfree(lwork);
2927 }
2928
2929 static void lwork_queue(struct linger_work *lwork)
2930 {
2931         struct ceph_osd_linger_request *lreq = lwork->lreq;
2932         struct ceph_osd_client *osdc = lreq->osdc;
2933
2934         verify_lreq_locked(lreq);
2935         WARN_ON(!list_empty(&lwork->pending_item));
2936
2937         lwork->queued_stamp = jiffies;
2938         list_add_tail(&lwork->pending_item, &lreq->pending_lworks);
2939         queue_work(osdc->notify_wq, &lwork->work);
2940 }
2941
2942 static void do_watch_notify(struct work_struct *w)
2943 {
2944         struct linger_work *lwork = container_of(w, struct linger_work, work);
2945         struct ceph_osd_linger_request *lreq = lwork->lreq;
2946
2947         if (!linger_registered(lreq)) {
2948                 dout("%s lreq %p not registered\n", __func__, lreq);
2949                 goto out;
2950         }
2951
2952         WARN_ON(!lreq->is_watch);
2953         dout("%s lreq %p notify_id %llu notifier_id %llu payload_len %zu\n",
2954              __func__, lreq, lwork->notify.notify_id, lwork->notify.notifier_id,
2955              lwork->notify.payload_len);
2956         lreq->wcb(lreq->data, lwork->notify.notify_id, lreq->linger_id,
2957                   lwork->notify.notifier_id, lwork->notify.payload,
2958                   lwork->notify.payload_len);
2959
2960 out:
2961         ceph_msg_put(lwork->notify.msg);
2962         lwork_free(lwork);
2963 }
2964
2965 static void do_watch_error(struct work_struct *w)
2966 {
2967         struct linger_work *lwork = container_of(w, struct linger_work, work);
2968         struct ceph_osd_linger_request *lreq = lwork->lreq;
2969
2970         if (!linger_registered(lreq)) {
2971                 dout("%s lreq %p not registered\n", __func__, lreq);
2972                 goto out;
2973         }
2974
2975         dout("%s lreq %p err %d\n", __func__, lreq, lwork->error.err);
2976         lreq->errcb(lreq->data, lreq->linger_id, lwork->error.err);
2977
2978 out:
2979         lwork_free(lwork);
2980 }
2981
2982 static void queue_watch_error(struct ceph_osd_linger_request *lreq)
2983 {
2984         struct linger_work *lwork;
2985
2986         lwork = lwork_alloc(lreq, do_watch_error);
2987         if (!lwork) {
2988                 pr_err("failed to allocate error-lwork\n");
2989                 return;
2990         }
2991
2992         lwork->error.err = lreq->last_error;
2993         lwork_queue(lwork);
2994 }
2995
2996 static void linger_reg_commit_complete(struct ceph_osd_linger_request *lreq,
2997                                        int result)
2998 {
2999         if (!completion_done(&lreq->reg_commit_wait)) {
3000                 lreq->reg_commit_error = (result <= 0 ? result : 0);
3001                 complete_all(&lreq->reg_commit_wait);
3002         }
3003 }
3004
3005 static void linger_commit_cb(struct ceph_osd_request *req)
3006 {
3007         struct ceph_osd_linger_request *lreq = req->r_priv;
3008
3009         mutex_lock(&lreq->lock);
3010         dout("%s lreq %p linger_id %llu result %d\n", __func__, lreq,
3011              lreq->linger_id, req->r_result);
3012         linger_reg_commit_complete(lreq, req->r_result);
3013         lreq->committed = true;
3014
3015         if (!lreq->is_watch) {
3016                 struct ceph_osd_data *osd_data =
3017                     osd_req_op_data(req, 0, notify, response_data);
3018                 void *p = page_address(osd_data->pages[0]);
3019
3020                 WARN_ON(req->r_ops[0].op != CEPH_OSD_OP_NOTIFY ||
3021                         osd_data->type != CEPH_OSD_DATA_TYPE_PAGES);
3022
3023                 /* make note of the notify_id */
3024                 if (req->r_ops[0].outdata_len >= sizeof(u64)) {
3025                         lreq->notify_id = ceph_decode_64(&p);
3026                         dout("lreq %p notify_id %llu\n", lreq,
3027                              lreq->notify_id);
3028                 } else {
3029                         dout("lreq %p no notify_id\n", lreq);
3030                 }
3031         }
3032
3033         mutex_unlock(&lreq->lock);
3034         linger_put(lreq);
3035 }
3036
3037 static int normalize_watch_error(int err)
3038 {
3039         /*
3040          * Translate ENOENT -> ENOTCONN so that a delete->disconnection
3041          * notification and a failure to reconnect because we raced with
3042          * the delete appear the same to the user.
3043          */
3044         if (err == -ENOENT)
3045                 err = -ENOTCONN;
3046
3047         return err;
3048 }
3049
3050 static void linger_reconnect_cb(struct ceph_osd_request *req)
3051 {
3052         struct ceph_osd_linger_request *lreq = req->r_priv;
3053
3054         mutex_lock(&lreq->lock);
3055         dout("%s lreq %p linger_id %llu result %d last_error %d\n", __func__,
3056              lreq, lreq->linger_id, req->r_result, lreq->last_error);
3057         if (req->r_result < 0) {
3058                 if (!lreq->last_error) {
3059                         lreq->last_error = normalize_watch_error(req->r_result);
3060                         queue_watch_error(lreq);
3061                 }
3062         }
3063
3064         mutex_unlock(&lreq->lock);
3065         linger_put(lreq);
3066 }
3067
3068 static void send_linger(struct ceph_osd_linger_request *lreq)
3069 {
3070         struct ceph_osd_request *req = lreq->reg_req;
3071         struct ceph_osd_req_op *op = &req->r_ops[0];
3072
3073         verify_osdc_wrlocked(req->r_osdc);
3074         dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id);
3075
3076         if (req->r_osd)
3077                 cancel_linger_request(req);
3078
3079         request_reinit(req);
3080         ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid);
3081         ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc);
3082         req->r_flags = lreq->t.flags;
3083         req->r_mtime = lreq->mtime;
3084
3085         mutex_lock(&lreq->lock);
3086         if (lreq->is_watch && lreq->committed) {
3087                 WARN_ON(op->op != CEPH_OSD_OP_WATCH ||
3088                         op->watch.cookie != lreq->linger_id);
3089                 op->watch.op = CEPH_OSD_WATCH_OP_RECONNECT;
3090                 op->watch.gen = ++lreq->register_gen;
3091                 dout("lreq %p reconnect register_gen %u\n", lreq,
3092                      op->watch.gen);
3093                 req->r_callback = linger_reconnect_cb;
3094         } else {
3095                 if (!lreq->is_watch)
3096                         lreq->notify_id = 0;
3097                 else
3098                         WARN_ON(op->watch.op != CEPH_OSD_WATCH_OP_WATCH);
3099                 dout("lreq %p register\n", lreq);
3100                 req->r_callback = linger_commit_cb;
3101         }
3102         mutex_unlock(&lreq->lock);
3103
3104         req->r_priv = linger_get(lreq);
3105         req->r_linger = true;
3106
3107         submit_request(req, true);
3108 }
3109
3110 static void linger_ping_cb(struct ceph_osd_request *req)
3111 {
3112         struct ceph_osd_linger_request *lreq = req->r_priv;
3113
3114         mutex_lock(&lreq->lock);
3115         dout("%s lreq %p linger_id %llu result %d ping_sent %lu last_error %d\n",
3116              __func__, lreq, lreq->linger_id, req->r_result, lreq->ping_sent,
3117              lreq->last_error);
3118         if (lreq->register_gen == req->r_ops[0].watch.gen) {
3119                 if (!req->r_result) {
3120                         lreq->watch_valid_thru = lreq->ping_sent;
3121                 } else if (!lreq->last_error) {
3122                         lreq->last_error = normalize_watch_error(req->r_result);
3123                         queue_watch_error(lreq);
3124                 }
3125         } else {
3126                 dout("lreq %p register_gen %u ignoring old pong %u\n", lreq,
3127                      lreq->register_gen, req->r_ops[0].watch.gen);
3128         }
3129
3130         mutex_unlock(&lreq->lock);
3131         linger_put(lreq);
3132 }
3133
3134 static void send_linger_ping(struct ceph_osd_linger_request *lreq)
3135 {
3136         struct ceph_osd_client *osdc = lreq->osdc;
3137         struct ceph_osd_request *req = lreq->ping_req;
3138         struct ceph_osd_req_op *op = &req->r_ops[0];
3139
3140         if (ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD)) {
3141                 dout("%s PAUSERD\n", __func__);
3142                 return;
3143         }
3144
3145         lreq->ping_sent = jiffies;
3146         dout("%s lreq %p linger_id %llu ping_sent %lu register_gen %u\n",
3147              __func__, lreq, lreq->linger_id, lreq->ping_sent,
3148              lreq->register_gen);
3149
3150         if (req->r_osd)
3151                 cancel_linger_request(req);
3152
3153         request_reinit(req);
3154         target_copy(&req->r_t, &lreq->t);
3155
3156         WARN_ON(op->op != CEPH_OSD_OP_WATCH ||
3157                 op->watch.cookie != lreq->linger_id ||
3158                 op->watch.op != CEPH_OSD_WATCH_OP_PING);
3159         op->watch.gen = lreq->register_gen;
3160         req->r_callback = linger_ping_cb;
3161         req->r_priv = linger_get(lreq);
3162         req->r_linger = true;
3163
3164         ceph_osdc_get_request(req);
3165         account_request(req);
3166         req->r_tid = atomic64_inc_return(&osdc->last_tid);
3167         link_request(lreq->osd, req);
3168         send_request(req);
3169 }
3170
3171 static void linger_submit(struct ceph_osd_linger_request *lreq)
3172 {
3173         struct ceph_osd_client *osdc = lreq->osdc;
3174         struct ceph_osd *osd;
3175
3176         down_write(&osdc->lock);
3177         linger_register(lreq);
3178         if (lreq->is_watch) {
3179                 lreq->reg_req->r_ops[0].watch.cookie = lreq->linger_id;
3180                 lreq->ping_req->r_ops[0].watch.cookie = lreq->linger_id;
3181         } else {
3182                 lreq->reg_req->r_ops[0].notify.cookie = lreq->linger_id;
3183         }
3184
3185         calc_target(osdc, &lreq->t, false);
3186         osd = lookup_create_osd(osdc, lreq->t.osd, true);
3187         link_linger(osd, lreq);
3188
3189         send_linger(lreq);
3190         up_write(&osdc->lock);
3191 }
3192
3193 static void cancel_linger_map_check(struct ceph_osd_linger_request *lreq)
3194 {
3195         struct ceph_osd_client *osdc = lreq->osdc;
3196         struct ceph_osd_linger_request *lookup_lreq;
3197
3198         verify_osdc_wrlocked(osdc);
3199
3200         lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks,
3201                                        lreq->linger_id);
3202         if (!lookup_lreq)
3203                 return;
3204
3205         WARN_ON(lookup_lreq != lreq);
3206         erase_linger_mc(&osdc->linger_map_checks, lreq);
3207         linger_put(lreq);
3208 }
3209
3210 /*
3211  * @lreq has to be both registered and linked.
3212  */
3213 static void __linger_cancel(struct ceph_osd_linger_request *lreq)
3214 {
3215         if (lreq->is_watch && lreq->ping_req->r_osd)
3216                 cancel_linger_request(lreq->ping_req);
3217         if (lreq->reg_req->r_osd)
3218                 cancel_linger_request(lreq->reg_req);
3219         cancel_linger_map_check(lreq);
3220         unlink_linger(lreq->osd, lreq);
3221         linger_unregister(lreq);
3222 }
3223
3224 static void linger_cancel(struct ceph_osd_linger_request *lreq)
3225 {
3226         struct ceph_osd_client *osdc = lreq->osdc;
3227
3228         down_write(&osdc->lock);
3229         if (__linger_registered(lreq))
3230                 __linger_cancel(lreq);
3231         up_write(&osdc->lock);
3232 }
3233
3234 static void send_linger_map_check(struct ceph_osd_linger_request *lreq);
3235
3236 static void check_linger_pool_dne(struct ceph_osd_linger_request *lreq)
3237 {
3238         struct ceph_osd_client *osdc = lreq->osdc;
3239         struct ceph_osdmap *map = osdc->osdmap;
3240
3241         verify_osdc_wrlocked(osdc);
3242         WARN_ON(!map->epoch);
3243
3244         if (lreq->register_gen) {
3245                 lreq->map_dne_bound = map->epoch;
3246                 dout("%s lreq %p linger_id %llu pool disappeared\n", __func__,
3247                      lreq, lreq->linger_id);
3248         } else {
3249                 dout("%s lreq %p linger_id %llu map_dne_bound %u have %u\n",
3250                      __func__, lreq, lreq->linger_id, lreq->map_dne_bound,
3251                      map->epoch);
3252         }
3253
3254         if (lreq->map_dne_bound) {
3255                 if (map->epoch >= lreq->map_dne_bound) {
3256                         /* we had a new enough map */
3257                         pr_info("linger_id %llu pool does not exist\n",
3258                                 lreq->linger_id);
3259                         linger_reg_commit_complete(lreq, -ENOENT);
3260                         __linger_cancel(lreq);
3261                 }
3262         } else {
3263                 send_linger_map_check(lreq);
3264         }
3265 }
3266
3267 static void linger_map_check_cb(struct ceph_mon_generic_request *greq)
3268 {
3269         struct ceph_osd_client *osdc = &greq->monc->client->osdc;
3270         struct ceph_osd_linger_request *lreq;
3271         u64 linger_id = greq->private_data;
3272
3273         WARN_ON(greq->result || !greq->u.newest);
3274
3275         down_write(&osdc->lock);
3276         lreq = lookup_linger_mc(&osdc->linger_map_checks, linger_id);
3277         if (!lreq) {
3278                 dout("%s linger_id %llu dne\n", __func__, linger_id);
3279                 goto out_unlock;
3280         }
3281
3282         dout("%s lreq %p linger_id %llu map_dne_bound %u newest %llu\n",
3283              __func__, lreq, lreq->linger_id, lreq->map_dne_bound,
3284              greq->u.newest);
3285         if (!lreq->map_dne_bound)
3286                 lreq->map_dne_bound = greq->u.newest;
3287         erase_linger_mc(&osdc->linger_map_checks, lreq);
3288         check_linger_pool_dne(lreq);
3289
3290         linger_put(lreq);
3291 out_unlock:
3292         up_write(&osdc->lock);
3293 }
3294
3295 static void send_linger_map_check(struct ceph_osd_linger_request *lreq)
3296 {
3297         struct ceph_osd_client *osdc = lreq->osdc;
3298         struct ceph_osd_linger_request *lookup_lreq;
3299         int ret;
3300
3301         verify_osdc_wrlocked(osdc);
3302
3303         lookup_lreq = lookup_linger_mc(&osdc->linger_map_checks,
3304                                        lreq->linger_id);
3305         if (lookup_lreq) {
3306                 WARN_ON(lookup_lreq != lreq);
3307                 return;
3308         }
3309
3310         linger_get(lreq);
3311         insert_linger_mc(&osdc->linger_map_checks, lreq);
3312         ret = ceph_monc_get_version_async(&osdc->client->monc, "osdmap",
3313                                           linger_map_check_cb, lreq->linger_id);
3314         WARN_ON(ret);
3315 }
3316
3317 static int linger_reg_commit_wait(struct ceph_osd_linger_request *lreq)
3318 {
3319         int ret;
3320
3321         dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id);
3322         ret = wait_for_completion_interruptible(&lreq->reg_commit_wait);
3323         return ret ?: lreq->reg_commit_error;
3324 }
3325
3326 static int linger_notify_finish_wait(struct ceph_osd_linger_request *lreq)
3327 {
3328         int ret;
3329
3330         dout("%s lreq %p linger_id %llu\n", __func__, lreq, lreq->linger_id);
3331         ret = wait_for_completion_interruptible(&lreq->notify_finish_wait);
3332         return ret ?: lreq->notify_finish_error;
3333 }
3334
3335 /*
3336  * Timeout callback, called every N seconds.  When 1 or more OSD
3337  * requests has been active for more than N seconds, we send a keepalive
3338  * (tag + timestamp) to its OSD to ensure any communications channel
3339  * reset is detected.
3340  */
3341 static void handle_timeout(struct work_struct *work)
3342 {
3343         struct ceph_osd_client *osdc =
3344                 container_of(work, struct ceph_osd_client, timeout_work.work);
3345         struct ceph_options *opts = osdc->client->options;
3346         unsigned long cutoff = jiffies - opts->osd_keepalive_timeout;
3347         unsigned long expiry_cutoff = jiffies - opts->osd_request_timeout;
3348         LIST_HEAD(slow_osds);
3349         struct rb_node *n, *p;
3350
3351         dout("%s osdc %p\n", __func__, osdc);
3352         down_write(&osdc->lock);
3353
3354         /*
3355          * ping osds that are a bit slow.  this ensures that if there
3356          * is a break in the TCP connection we will notice, and reopen
3357          * a connection with that osd (from the fault callback).
3358          */
3359         for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
3360                 struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
3361                 bool found = false;
3362
3363                 for (p = rb_first(&osd->o_requests); p; ) {
3364                         struct ceph_osd_request *req =
3365                             rb_entry(p, struct ceph_osd_request, r_node);
3366
3367                         p = rb_next(p); /* abort_request() */
3368
3369                         if (time_before(req->r_stamp, cutoff)) {
3370                                 dout(" req %p tid %llu on osd%d is laggy\n",
3371                                      req, req->r_tid, osd->o_osd);
3372                                 found = true;
3373                         }
3374                         if (opts->osd_request_timeout &&
3375                             time_before(req->r_start_stamp, expiry_cutoff)) {
3376                                 pr_err_ratelimited("tid %llu on osd%d timeout\n",
3377                                        req->r_tid, osd->o_osd);
3378                                 abort_request(req, -ETIMEDOUT);
3379                         }
3380                 }
3381                 for (p = rb_first(&osd->o_linger_requests); p; p = rb_next(p)) {
3382                         struct ceph_osd_linger_request *lreq =
3383                             rb_entry(p, struct ceph_osd_linger_request, node);
3384
3385                         dout(" lreq %p linger_id %llu is served by osd%d\n",
3386                              lreq, lreq->linger_id, osd->o_osd);
3387                         found = true;
3388
3389                         mutex_lock(&lreq->lock);
3390                         if (lreq->is_watch && lreq->committed && !lreq->last_error)
3391                                 send_linger_ping(lreq);
3392                         mutex_unlock(&lreq->lock);
3393                 }
3394
3395                 if (found)
3396                         list_move_tail(&osd->o_keepalive_item, &slow_osds);
3397         }
3398
3399         if (opts->osd_request_timeout) {
3400                 for (p = rb_first(&osdc->homeless_osd.o_requests); p; ) {
3401                         struct ceph_osd_request *req =
3402                             rb_entry(p, struct ceph_osd_request, r_node);
3403
3404                         p = rb_next(p); /* abort_request() */
3405
3406                         if (time_before(req->r_start_stamp, expiry_cutoff)) {
3407                                 pr_err_ratelimited("tid %llu on osd%d timeout\n",
3408                                        req->r_tid, osdc->homeless_osd.o_osd);
3409                                 abort_request(req, -ETIMEDOUT);
3410                         }
3411                 }
3412         }
3413
3414         if (atomic_read(&osdc->num_homeless) || !list_empty(&slow_osds))
3415                 maybe_request_map(osdc);
3416
3417         while (!list_empty(&slow_osds)) {
3418                 struct ceph_osd *osd = list_first_entry(&slow_osds,
3419                                                         struct ceph_osd,
3420                                                         o_keepalive_item);
3421                 list_del_init(&osd->o_keepalive_item);
3422                 ceph_con_keepalive(&osd->o_con);
3423         }
3424
3425         up_write(&osdc->lock);
3426         schedule_delayed_work(&osdc->timeout_work,
3427                               osdc->client->options->osd_keepalive_timeout);
3428 }
3429
3430 static void handle_osds_timeout(struct work_struct *work)
3431 {
3432         struct ceph_osd_client *osdc =
3433                 container_of(work, struct ceph_osd_client,
3434                              osds_timeout_work.work);
3435         unsigned long delay = osdc->client->options->osd_idle_ttl / 4;
3436         struct ceph_osd *osd, *nosd;
3437
3438         dout("%s osdc %p\n", __func__, osdc);
3439         down_write(&osdc->lock);
3440         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
3441                 if (time_before(jiffies, osd->lru_ttl))
3442                         break;
3443
3444                 WARN_ON(!RB_EMPTY_ROOT(&osd->o_requests));
3445                 WARN_ON(!RB_EMPTY_ROOT(&osd->o_linger_requests));
3446                 close_osd(osd);
3447         }
3448
3449         up_write(&osdc->lock);
3450         schedule_delayed_work(&osdc->osds_timeout_work,
3451                               round_jiffies_relative(delay));
3452 }
3453
3454 static int ceph_oloc_decode(void **p, void *end,
3455                             struct ceph_object_locator *oloc)
3456 {
3457         u8 struct_v, struct_cv;
3458         u32 len;
3459         void *struct_end;
3460         int ret = 0;
3461
3462         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
3463         struct_v = ceph_decode_8(p);
3464         struct_cv = ceph_decode_8(p);
3465         if (struct_v < 3) {
3466                 pr_warn("got v %d < 3 cv %d of ceph_object_locator\n",
3467                         struct_v, struct_cv);
3468                 goto e_inval;
3469         }
3470         if (struct_cv > 6) {
3471                 pr_warn("got v %d cv %d > 6 of ceph_object_locator\n",
3472                         struct_v, struct_cv);
3473                 goto e_inval;
3474         }
3475         len = ceph_decode_32(p);
3476         ceph_decode_need(p, end, len, e_inval);
3477         struct_end = *p + len;
3478
3479         oloc->pool = ceph_decode_64(p);
3480         *p += 4; /* skip preferred */
3481
3482         len = ceph_decode_32(p);
3483         if (len > 0) {
3484                 pr_warn("ceph_object_locator::key is set\n");
3485                 goto e_inval;
3486         }
3487
3488         if (struct_v >= 5) {
3489                 bool changed = false;
3490
3491                 len = ceph_decode_32(p);
3492                 if (len > 0) {
3493                         ceph_decode_need(p, end, len, e_inval);
3494                         if (!oloc->pool_ns ||
3495                             ceph_compare_string(oloc->pool_ns, *p, len))
3496                                 changed = true;
3497                         *p += len;
3498                 } else {
3499                         if (oloc->pool_ns)
3500                                 changed = true;
3501                 }
3502                 if (changed) {
3503                         /* redirect changes namespace */
3504                         pr_warn("ceph_object_locator::nspace is changed\n");
3505                         goto e_inval;
3506                 }
3507         }
3508
3509         if (struct_v >= 6) {
3510                 s64 hash = ceph_decode_64(p);
3511                 if (hash != -1) {
3512                         pr_warn("ceph_object_locator::hash is set\n");
3513                         goto e_inval;
3514                 }
3515         }
3516
3517         /* skip the rest */
3518         *p = struct_end;
3519 out:
3520         return ret;
3521
3522 e_inval:
3523         ret = -EINVAL;
3524         goto out;
3525 }
3526
3527 static int ceph_redirect_decode(void **p, void *end,
3528                                 struct ceph_request_redirect *redir)
3529 {
3530         u8 struct_v, struct_cv;
3531         u32 len;
3532         void *struct_end;
3533         int ret;
3534
3535         ceph_decode_need(p, end, 1 + 1 + 4, e_inval);
3536         struct_v = ceph_decode_8(p);
3537         struct_cv = ceph_decode_8(p);
3538         if (struct_cv > 1) {
3539                 pr_warn("got v %d cv %d > 1 of ceph_request_redirect\n",
3540                         struct_v, struct_cv);
3541                 goto e_inval;
3542         }
3543         len = ceph_decode_32(p);
3544         ceph_decode_need(p, end, len, e_inval);
3545         struct_end = *p + len;
3546
3547         ret = ceph_oloc_decode(p, end, &redir->oloc);
3548         if (ret)
3549                 goto out;
3550
3551         len = ceph_decode_32(p);
3552         if (len > 0) {
3553                 pr_warn("ceph_request_redirect::object_name is set\n");
3554                 goto e_inval;
3555         }
3556
3557         /* skip the rest */
3558         *p = struct_end;
3559 out:
3560         return ret;
3561
3562 e_inval:
3563         ret = -EINVAL;
3564         goto out;
3565 }
3566
3567 struct MOSDOpReply {
3568         struct ceph_pg pgid;
3569         u64 flags;
3570         int result;
3571         u32 epoch;
3572         int num_ops;
3573         u32 outdata_len[CEPH_OSD_MAX_OPS];
3574         s32 rval[CEPH_OSD_MAX_OPS];
3575         int retry_attempt;
3576         struct ceph_eversion replay_version;
3577         u64 user_version;
3578         struct ceph_request_redirect redirect;
3579 };
3580
3581 static int decode_MOSDOpReply(const struct ceph_msg *msg, struct MOSDOpReply *m)
3582 {
3583         void *p = msg->front.iov_base;
3584         void *const end = p + msg->front.iov_len;
3585         u16 version = le16_to_cpu(msg->hdr.version);
3586         struct ceph_eversion bad_replay_version;
3587         u8 decode_redir;
3588         u32 len;
3589         int ret;
3590         int i;
3591
3592         ceph_decode_32_safe(&p, end, len, e_inval);
3593         ceph_decode_need(&p, end, len, e_inval);
3594         p += len; /* skip oid */
3595
3596         ret = ceph_decode_pgid(&p, end, &m->pgid);
3597         if (ret)
3598                 return ret;
3599
3600         ceph_decode_64_safe(&p, end, m->flags, e_inval);
3601         ceph_decode_32_safe(&p, end, m->result, e_inval);
3602         ceph_decode_need(&p, end, sizeof(bad_replay_version), e_inval);
3603         memcpy(&bad_replay_version, p, sizeof(bad_replay_version));
3604         p += sizeof(bad_replay_version);
3605         ceph_decode_32_safe(&p, end, m->epoch, e_inval);
3606
3607         ceph_decode_32_safe(&p, end, m->num_ops, e_inval);
3608         if (m->num_ops > ARRAY_SIZE(m->outdata_len))
3609                 goto e_inval;
3610
3611         ceph_decode_need(&p, end, m->num_ops * sizeof(struct ceph_osd_op),
3612                          e_inval);
3613         for (i = 0; i < m->num_ops; i++) {
3614                 struct ceph_osd_op *op = p;
3615
3616                 m->outdata_len[i] = le32_to_cpu(op->payload_len);
3617                 p += sizeof(*op);
3618         }
3619
3620         ceph_decode_32_safe(&p, end, m->retry_attempt, e_inval);
3621         for (i = 0; i < m->num_ops; i++)
3622                 ceph_decode_32_safe(&p, end, m->rval[i], e_inval);
3623
3624         if (version >= 5) {
3625                 ceph_decode_need(&p, end, sizeof(m->replay_version), e_inval);
3626                 memcpy(&m->replay_version, p, sizeof(m->replay_version));
3627                 p += sizeof(m->replay_version);
3628                 ceph_decode_64_safe(&p, end, m->user_version, e_inval);
3629         } else {
3630                 m->replay_version = bad_replay_version; /* struct */
3631                 m->user_version = le64_to_cpu(m->replay_version.version);
3632         }
3633
3634         if (version >= 6) {
3635                 if (version >= 7)
3636                         ceph_decode_8_safe(&p, end, decode_redir, e_inval);
3637                 else
3638                         decode_redir = 1;
3639         } else {
3640                 decode_redir = 0;
3641         }
3642
3643         if (decode_redir) {
3644                 ret = ceph_redirect_decode(&p, end, &m->redirect);
3645                 if (ret)
3646                         return ret;
3647         } else {
3648                 ceph_oloc_init(&m->redirect.oloc);
3649         }
3650
3651         return 0;
3652
3653 e_inval:
3654         return -EINVAL;
3655 }
3656
3657 /*
3658  * Handle MOSDOpReply.  Set ->r_result and call the callback if it is
3659  * specified.
3660  */
3661 static void handle_reply(struct ceph_osd *osd, struct ceph_msg *msg)
3662 {
3663         struct ceph_osd_client *osdc = osd->o_osdc;
3664         struct ceph_osd_request *req;
3665         struct MOSDOpReply m;
3666         u64 tid = le64_to_cpu(msg->hdr.tid);
3667         u32 data_len = 0;
3668         int ret;
3669         int i;
3670
3671         dout("%s msg %p tid %llu\n", __func__, msg, tid);
3672
3673         down_read(&osdc->lock);
3674         if (!osd_registered(osd)) {
3675                 dout("%s osd%d unknown\n", __func__, osd->o_osd);
3676                 goto out_unlock_osdc;
3677         }
3678         WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num));
3679
3680         mutex_lock(&osd->lock);
3681         req = lookup_request(&osd->o_requests, tid);
3682         if (!req) {
3683                 dout("%s osd%d tid %llu unknown\n", __func__, osd->o_osd, tid);
3684                 goto out_unlock_session;
3685         }
3686
3687         m.redirect.oloc.pool_ns = req->r_t.target_oloc.pool_ns;
3688         ret = decode_MOSDOpReply(msg, &m);
3689         m.redirect.oloc.pool_ns = NULL;
3690         if (ret) {
3691                 pr_err("failed to decode MOSDOpReply for tid %llu: %d\n",
3692                        req->r_tid, ret);
3693                 ceph_msg_dump(msg);
3694                 goto fail_request;
3695         }
3696         dout("%s req %p tid %llu flags 0x%llx pgid %llu.%x epoch %u attempt %d v %u'%llu uv %llu\n",
3697              __func__, req, req->r_tid, m.flags, m.pgid.pool, m.pgid.seed,
3698              m.epoch, m.retry_attempt, le32_to_cpu(m.replay_version.epoch),
3699              le64_to_cpu(m.replay_version.version), m.user_version);
3700
3701         if (m.retry_attempt >= 0) {
3702                 if (m.retry_attempt != req->r_attempts - 1) {
3703                         dout("req %p tid %llu retry_attempt %d != %d, ignoring\n",
3704                              req, req->r_tid, m.retry_attempt,
3705                              req->r_attempts - 1);
3706                         goto out_unlock_session;
3707                 }
3708         } else {
3709                 WARN_ON(1); /* MOSDOpReply v4 is assumed */
3710         }
3711
3712         if (!ceph_oloc_empty(&m.redirect.oloc)) {
3713                 dout("req %p tid %llu redirect pool %lld\n", req, req->r_tid,
3714                      m.redirect.oloc.pool);
3715                 unlink_request(osd, req);
3716                 mutex_unlock(&osd->lock);
3717
3718                 /*
3719                  * Not ceph_oloc_copy() - changing pool_ns is not
3720                  * supported.
3721                  */
3722                 req->r_t.target_oloc.pool = m.redirect.oloc.pool;
3723                 req->r_flags |= CEPH_OSD_FLAG_REDIRECTED |
3724                                 CEPH_OSD_FLAG_IGNORE_OVERLAY |
3725                                 CEPH_OSD_FLAG_IGNORE_CACHE;
3726                 req->r_tid = 0;
3727                 __submit_request(req, false);
3728                 goto out_unlock_osdc;
3729         }
3730
3731         if (m.result == -EAGAIN) {
3732                 dout("req %p tid %llu EAGAIN\n", req, req->r_tid);
3733                 unlink_request(osd, req);
3734                 mutex_unlock(&osd->lock);
3735
3736                 /*
3737                  * The object is missing on the replica or not (yet)
3738                  * readable.  Clear pgid to force a resend to the primary
3739                  * via legacy_change.
3740                  */
3741                 req->r_t.pgid.pool = 0;
3742                 req->r_t.pgid.seed = 0;
3743                 WARN_ON(!req->r_t.used_replica);
3744                 req->r_flags &= ~(CEPH_OSD_FLAG_BALANCE_READS |
3745                                   CEPH_OSD_FLAG_LOCALIZE_READS);
3746                 req->r_tid = 0;
3747                 __submit_request(req, false);
3748                 goto out_unlock_osdc;
3749         }
3750
3751         if (m.num_ops != req->r_num_ops) {
3752                 pr_err("num_ops %d != %d for tid %llu\n", m.num_ops,
3753                        req->r_num_ops, req->r_tid);
3754                 goto fail_request;
3755         }
3756         for (i = 0; i < req->r_num_ops; i++) {
3757                 dout(" req %p tid %llu op %d rval %d len %u\n", req,
3758                      req->r_tid, i, m.rval[i], m.outdata_len[i]);
3759                 req->r_ops[i].rval = m.rval[i];
3760                 req->r_ops[i].outdata_len = m.outdata_len[i];
3761                 data_len += m.outdata_len[i];
3762         }
3763         if (data_len != le32_to_cpu(msg->hdr.data_len)) {
3764                 pr_err("sum of lens %u != %u for tid %llu\n", data_len,
3765                        le32_to_cpu(msg->hdr.data_len), req->r_tid);
3766                 goto fail_request;
3767         }
3768         dout("%s req %p tid %llu result %d data_len %u\n", __func__,
3769              req, req->r_tid, m.result, data_len);
3770
3771         /*
3772          * Since we only ever request ONDISK, we should only ever get
3773          * one (type of) reply back.
3774          */
3775         WARN_ON(!(m.flags & CEPH_OSD_FLAG_ONDISK));
3776         req->r_result = m.result ?: data_len;
3777         finish_request(req);
3778         mutex_unlock(&osd->lock);
3779         up_read(&osdc->lock);
3780
3781         __complete_request(req);
3782         return;
3783
3784 fail_request:
3785         complete_request(req, -EIO);
3786 out_unlock_session:
3787         mutex_unlock(&osd->lock);
3788 out_unlock_osdc:
3789         up_read(&osdc->lock);
3790 }
3791
3792 static void set_pool_was_full(struct ceph_osd_client *osdc)
3793 {
3794         struct rb_node *n;
3795
3796         for (n = rb_first(&osdc->osdmap->pg_pools); n; n = rb_next(n)) {
3797                 struct ceph_pg_pool_info *pi =
3798                     rb_entry(n, struct ceph_pg_pool_info, node);
3799
3800                 pi->was_full = __pool_full(pi);
3801         }
3802 }
3803
3804 static bool pool_cleared_full(struct ceph_osd_client *osdc, s64 pool_id)
3805 {
3806         struct ceph_pg_pool_info *pi;
3807
3808         pi = ceph_pg_pool_by_id(osdc->osdmap, pool_id);
3809         if (!pi)
3810                 return false;
3811
3812         return pi->was_full && !__pool_full(pi);
3813 }
3814
3815 static enum calc_target_result
3816 recalc_linger_target(struct ceph_osd_linger_request *lreq)
3817 {
3818         struct ceph_osd_client *osdc = lreq->osdc;
3819         enum calc_target_result ct_res;
3820
3821         ct_res = calc_target(osdc, &lreq->t, true);
3822         if (ct_res == CALC_TARGET_NEED_RESEND) {
3823                 struct ceph_osd *osd;
3824
3825                 osd = lookup_create_osd(osdc, lreq->t.osd, true);
3826                 if (osd != lreq->osd) {
3827                         unlink_linger(lreq->osd, lreq);
3828                         link_linger(osd, lreq);
3829                 }
3830         }
3831
3832         return ct_res;
3833 }
3834
3835 /*
3836  * Requeue requests whose mapping to an OSD has changed.
3837  */
3838 static void scan_requests(struct ceph_osd *osd,
3839                           bool force_resend,
3840                           bool cleared_full,
3841                           bool check_pool_cleared_full,
3842                           struct rb_root *need_resend,
3843                           struct list_head *need_resend_linger)
3844 {
3845         struct ceph_osd_client *osdc = osd->o_osdc;
3846         struct rb_node *n;
3847         bool force_resend_writes;
3848
3849         for (n = rb_first(&osd->o_linger_requests); n; ) {
3850                 struct ceph_osd_linger_request *lreq =
3851                     rb_entry(n, struct ceph_osd_linger_request, node);
3852                 enum calc_target_result ct_res;
3853
3854                 n = rb_next(n); /* recalc_linger_target() */
3855
3856                 dout("%s lreq %p linger_id %llu\n", __func__, lreq,
3857                      lreq->linger_id);
3858                 ct_res = recalc_linger_target(lreq);
3859                 switch (ct_res) {
3860                 case CALC_TARGET_NO_ACTION:
3861                         force_resend_writes = cleared_full ||
3862                             (check_pool_cleared_full &&
3863                              pool_cleared_full(osdc, lreq->t.base_oloc.pool));
3864                         if (!force_resend && !force_resend_writes)
3865                                 break;
3866
3867                         /* fall through */
3868                 case CALC_TARGET_NEED_RESEND:
3869                         cancel_linger_map_check(lreq);
3870                         /*
3871                          * scan_requests() for the previous epoch(s)
3872                          * may have already added it to the list, since
3873                          * it's not unlinked here.
3874                          */
3875                         if (list_empty(&lreq->scan_item))
3876                                 list_add_tail(&lreq->scan_item, need_resend_linger);
3877                         break;
3878                 case CALC_TARGET_POOL_DNE:
3879                         list_del_init(&lreq->scan_item);
3880                         check_linger_pool_dne(lreq);
3881                         break;
3882                 }
3883         }
3884
3885         for (n = rb_first(&osd->o_requests); n; ) {
3886                 struct ceph_osd_request *req =
3887                     rb_entry(n, struct ceph_osd_request, r_node);
3888                 enum calc_target_result ct_res;
3889
3890                 n = rb_next(n); /* unlink_request(), check_pool_dne() */
3891
3892                 dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
3893                 ct_res = calc_target(osdc, &req->r_t, false);
3894                 switch (ct_res) {
3895                 case CALC_TARGET_NO_ACTION:
3896                         force_resend_writes = cleared_full ||
3897                             (check_pool_cleared_full &&
3898                              pool_cleared_full(osdc, req->r_t.base_oloc.pool));
3899                         if (!force_resend &&
3900                             (!(req->r_flags & CEPH_OSD_FLAG_WRITE) ||
3901                              !force_resend_writes))
3902                                 break;
3903
3904                         /* fall through */
3905                 case CALC_TARGET_NEED_RESEND:
3906                         cancel_map_check(req);
3907                         unlink_request(osd, req);
3908                         insert_request(need_resend, req);
3909                         break;
3910                 case CALC_TARGET_POOL_DNE:
3911                         check_pool_dne(req);
3912                         break;
3913                 }
3914         }
3915 }
3916
3917 static int handle_one_map(struct ceph_osd_client *osdc,
3918                           void *p, void *end, bool incremental,
3919                           struct rb_root *need_resend,
3920                           struct list_head *need_resend_linger)
3921 {
3922         struct ceph_osdmap *newmap;
3923         struct rb_node *n;
3924         bool skipped_map = false;
3925         bool was_full;
3926
3927         was_full = ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL);
3928         set_pool_was_full(osdc);
3929
3930         if (incremental)
3931                 newmap = osdmap_apply_incremental(&p, end, osdc->osdmap);
3932         else
3933                 newmap = ceph_osdmap_decode(&p, end);
3934         if (IS_ERR(newmap))
3935                 return PTR_ERR(newmap);
3936
3937         if (newmap != osdc->osdmap) {
3938                 /*
3939                  * Preserve ->was_full before destroying the old map.
3940                  * For pools that weren't in the old map, ->was_full
3941                  * should be false.
3942                  */
3943                 for (n = rb_first(&newmap->pg_pools); n; n = rb_next(n)) {
3944                         struct ceph_pg_pool_info *pi =
3945                             rb_entry(n, struct ceph_pg_pool_info, node);
3946                         struct ceph_pg_pool_info *old_pi;
3947
3948                         old_pi = ceph_pg_pool_by_id(osdc->osdmap, pi->id);
3949                         if (old_pi)
3950                                 pi->was_full = old_pi->was_full;
3951                         else
3952                                 WARN_ON(pi->was_full);
3953                 }
3954
3955                 if (osdc->osdmap->epoch &&
3956                     osdc->osdmap->epoch + 1 < newmap->epoch) {
3957                         WARN_ON(incremental);
3958                         skipped_map = true;
3959                 }
3960
3961                 ceph_osdmap_destroy(osdc->osdmap);
3962                 osdc->osdmap = newmap;
3963         }
3964
3965         was_full &= !ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL);
3966         scan_requests(&osdc->homeless_osd, skipped_map, was_full, true,
3967                       need_resend, need_resend_linger);
3968
3969         for (n = rb_first(&osdc->osds); n; ) {
3970                 struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
3971
3972                 n = rb_next(n); /* close_osd() */
3973
3974                 scan_requests(osd, skipped_map, was_full, true, need_resend,
3975                               need_resend_linger);
3976                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
3977                     memcmp(&osd->o_con.peer_addr,
3978                            ceph_osd_addr(osdc->osdmap, osd->o_osd),
3979                            sizeof(struct ceph_entity_addr)))
3980                         close_osd(osd);
3981         }
3982
3983         return 0;
3984 }
3985
3986 static void kick_requests(struct ceph_osd_client *osdc,
3987                           struct rb_root *need_resend,
3988                           struct list_head *need_resend_linger)
3989 {
3990         struct ceph_osd_linger_request *lreq, *nlreq;
3991         enum calc_target_result ct_res;
3992         struct rb_node *n;
3993
3994         /* make sure need_resend targets reflect latest map */
3995         for (n = rb_first(need_resend); n; ) {
3996                 struct ceph_osd_request *req =
3997                     rb_entry(n, struct ceph_osd_request, r_node);
3998
3999                 n = rb_next(n);
4000
4001                 if (req->r_t.epoch < osdc->osdmap->epoch) {
4002                         ct_res = calc_target(osdc, &req->r_t, false);
4003                         if (ct_res == CALC_TARGET_POOL_DNE) {
4004                                 erase_request(need_resend, req);
4005                                 check_pool_dne(req);
4006                         }
4007                 }
4008         }
4009
4010         for (n = rb_first(need_resend); n; ) {
4011                 struct ceph_osd_request *req =
4012                     rb_entry(n, struct ceph_osd_request, r_node);
4013                 struct ceph_osd *osd;
4014
4015                 n = rb_next(n);
4016                 erase_request(need_resend, req); /* before link_request() */
4017
4018                 osd = lookup_create_osd(osdc, req->r_t.osd, true);
4019                 link_request(osd, req);
4020                 if (!req->r_linger) {
4021                         if (!osd_homeless(osd) && !req->r_t.paused)
4022                                 send_request(req);
4023                 } else {
4024                         cancel_linger_request(req);
4025                 }
4026         }
4027
4028         list_for_each_entry_safe(lreq, nlreq, need_resend_linger, scan_item) {
4029                 if (!osd_homeless(lreq->osd))
4030                         send_linger(lreq);
4031
4032                 list_del_init(&lreq->scan_item);
4033         }
4034 }
4035
4036 /*
4037  * Process updated osd map.
4038  *
4039  * The message contains any number of incremental and full maps, normally
4040  * indicating some sort of topology change in the cluster.  Kick requests
4041  * off to different OSDs as needed.
4042  */
4043 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
4044 {
4045         void *p = msg->front.iov_base;
4046         void *const end = p + msg->front.iov_len;
4047         u32 nr_maps, maplen;
4048         u32 epoch;
4049         struct ceph_fsid fsid;
4050         struct rb_root need_resend = RB_ROOT;
4051         LIST_HEAD(need_resend_linger);
4052         bool handled_incremental = false;
4053         bool was_pauserd, was_pausewr;
4054         bool pauserd, pausewr;
4055         int err;
4056
4057         dout("%s have %u\n", __func__, osdc->osdmap->epoch);
4058         down_write(&osdc->lock);
4059
4060         /* verify fsid */
4061         ceph_decode_need(&p, end, sizeof(fsid), bad);
4062         ceph_decode_copy(&p, &fsid, sizeof(fsid));
4063         if (ceph_check_fsid(osdc->client, &fsid) < 0)
4064                 goto bad;
4065
4066         was_pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD);
4067         was_pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) ||
4068                       ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
4069                       have_pool_full(osdc);
4070
4071         /* incremental maps */
4072         ceph_decode_32_safe(&p, end, nr_maps, bad);
4073         dout(" %d inc maps\n", nr_maps);
4074         while (nr_maps > 0) {
4075                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
4076                 epoch = ceph_decode_32(&p);
4077                 maplen = ceph_decode_32(&p);
4078                 ceph_decode_need(&p, end, maplen, bad);
4079                 if (osdc->osdmap->epoch &&
4080                     osdc->osdmap->epoch + 1 == epoch) {
4081                         dout("applying incremental map %u len %d\n",
4082                              epoch, maplen);
4083                         err = handle_one_map(osdc, p, p + maplen, true,
4084                                              &need_resend, &need_resend_linger);
4085                         if (err)
4086                                 goto bad;
4087                         handled_incremental = true;
4088                 } else {
4089                         dout("ignoring incremental map %u len %d\n",
4090                              epoch, maplen);
4091                 }
4092                 p += maplen;
4093                 nr_maps--;
4094         }
4095         if (handled_incremental)
4096                 goto done;
4097
4098         /* full maps */
4099         ceph_decode_32_safe(&p, end, nr_maps, bad);
4100         dout(" %d full maps\n", nr_maps);
4101         while (nr_maps) {
4102                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
4103                 epoch = ceph_decode_32(&p);
4104                 maplen = ceph_decode_32(&p);
4105                 ceph_decode_need(&p, end, maplen, bad);
4106                 if (nr_maps > 1) {
4107                         dout("skipping non-latest full map %u len %d\n",
4108                              epoch, maplen);
4109                 } else if (osdc->osdmap->epoch >= epoch) {
4110                         dout("skipping full map %u len %d, "
4111                              "older than our %u\n", epoch, maplen,
4112                              osdc->osdmap->epoch);
4113                 } else {
4114                         dout("taking full map %u len %d\n", epoch, maplen);
4115                         err = handle_one_map(osdc, p, p + maplen, false,
4116                                              &need_resend, &need_resend_linger);
4117                         if (err)
4118                                 goto bad;
4119                 }
4120                 p += maplen;
4121                 nr_maps--;
4122         }
4123
4124 done:
4125         /*
4126          * subscribe to subsequent osdmap updates if full to ensure
4127          * we find out when we are no longer full and stop returning
4128          * ENOSPC.
4129          */
4130         pauserd = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSERD);
4131         pausewr = ceph_osdmap_flag(osdc, CEPH_OSDMAP_PAUSEWR) ||
4132                   ceph_osdmap_flag(osdc, CEPH_OSDMAP_FULL) ||
4133                   have_pool_full(osdc);
4134         if (was_pauserd || was_pausewr || pauserd || pausewr ||
4135             osdc->osdmap->epoch < osdc->epoch_barrier)
4136                 maybe_request_map(osdc);
4137
4138         kick_requests(osdc, &need_resend, &need_resend_linger);
4139
4140         ceph_osdc_abort_on_full(osdc);
4141         ceph_monc_got_map(&osdc->client->monc, CEPH_SUB_OSDMAP,
4142                           osdc->osdmap->epoch);
4143         up_write(&osdc->lock);
4144         wake_up_all(&osdc->client->auth_wq);
4145         return;
4146
4147 bad:
4148         pr_err("osdc handle_map corrupt msg\n");
4149         ceph_msg_dump(msg);
4150         up_write(&osdc->lock);
4151 }
4152
4153 /*
4154  * Resubmit requests pending on the given osd.
4155  */
4156 static void kick_osd_requests(struct ceph_osd *osd)
4157 {
4158         struct rb_node *n;
4159
4160         clear_backoffs(osd);
4161
4162         for (n = rb_first(&osd->o_requests); n; ) {
4163                 struct ceph_osd_request *req =
4164                     rb_entry(n, struct ceph_osd_request, r_node);
4165
4166                 n = rb_next(n); /* cancel_linger_request() */
4167
4168                 if (!req->r_linger) {
4169                         if (!req->r_t.paused)
4170                                 send_request(req);
4171                 } else {
4172                         cancel_linger_request(req);
4173                 }
4174         }
4175         for (n = rb_first(&osd->o_linger_requests); n; n = rb_next(n)) {
4176                 struct ceph_osd_linger_request *lreq =
4177                     rb_entry(n, struct ceph_osd_linger_request, node);
4178
4179                 send_linger(lreq);
4180         }
4181 }
4182
4183 /*
4184  * If the osd connection drops, we need to resubmit all requests.
4185  */
4186 static void osd_fault(struct ceph_connection *con)
4187 {
4188         struct ceph_osd *osd = con->private;
4189         struct ceph_osd_client *osdc = osd->o_osdc;
4190
4191         dout("%s osd %p osd%d\n", __func__, osd, osd->o_osd);
4192
4193         down_write(&osdc->lock);
4194         if (!osd_registered(osd)) {
4195                 dout("%s osd%d unknown\n", __func__, osd->o_osd);
4196                 goto out_unlock;
4197         }
4198
4199         if (!reopen_osd(osd))
4200                 kick_osd_requests(osd);
4201         maybe_request_map(osdc);
4202
4203 out_unlock:
4204         up_write(&osdc->lock);
4205 }
4206
4207 struct MOSDBackoff {
4208         struct ceph_spg spgid;
4209         u32 map_epoch;
4210         u8 op;
4211         u64 id;
4212         struct ceph_hobject_id *begin;
4213         struct ceph_hobject_id *end;
4214 };
4215
4216 static int decode_MOSDBackoff(const struct ceph_msg *msg, struct MOSDBackoff *m)
4217 {
4218         void *p = msg->front.iov_base;
4219         void *const end = p + msg->front.iov_len;
4220         u8 struct_v;
4221         u32 struct_len;
4222         int ret;
4223
4224         ret = ceph_start_decoding(&p, end, 1, "spg_t", &struct_v, &struct_len);
4225         if (ret)
4226                 return ret;
4227
4228         ret = ceph_decode_pgid(&p, end, &m->spgid.pgid);
4229         if (ret)
4230                 return ret;
4231
4232         ceph_decode_8_safe(&p, end, m->spgid.shard, e_inval);
4233         ceph_decode_32_safe(&p, end, m->map_epoch, e_inval);
4234         ceph_decode_8_safe(&p, end, m->op, e_inval);
4235         ceph_decode_64_safe(&p, end, m->id, e_inval);
4236
4237         m->begin = kzalloc(sizeof(*m->begin), GFP_NOIO);
4238         if (!m->begin)
4239                 return -ENOMEM;
4240
4241         ret = decode_hoid(&p, end, m->begin);
4242         if (ret) {
4243                 free_hoid(m->begin);
4244                 return ret;
4245         }
4246
4247         m->end = kzalloc(sizeof(*m->end), GFP_NOIO);
4248         if (!m->end) {
4249                 free_hoid(m->begin);
4250                 return -ENOMEM;
4251         }
4252
4253         ret = decode_hoid(&p, end, m->end);
4254         if (ret) {
4255                 free_hoid(m->begin);
4256                 free_hoid(m->end);
4257                 return ret;
4258         }
4259
4260         return 0;
4261
4262 e_inval:
4263         return -EINVAL;
4264 }
4265
4266 static struct ceph_msg *create_backoff_message(
4267                                 const struct ceph_osd_backoff *backoff,
4268                                 u32 map_epoch)
4269 {
4270         struct ceph_msg *msg;
4271         void *p, *end;
4272         int msg_size;
4273
4274         msg_size = CEPH_ENCODING_START_BLK_LEN +
4275                         CEPH_PGID_ENCODING_LEN + 1; /* spgid */
4276         msg_size += 4 + 1 + 8; /* map_epoch, op, id */
4277         msg_size += CEPH_ENCODING_START_BLK_LEN +
4278                         hoid_encoding_size(backoff->begin);
4279         msg_size += CEPH_ENCODING_START_BLK_LEN +
4280                         hoid_encoding_size(backoff->end);
4281
4282         msg = ceph_msg_new(CEPH_MSG_OSD_BACKOFF, msg_size, GFP_NOIO, true);
4283         if (!msg)
4284                 return NULL;
4285
4286         p = msg->front.iov_base;
4287         end = p + msg->front_alloc_len;
4288
4289         encode_spgid(&p, &backoff->spgid);
4290         ceph_encode_32(&p, map_epoch);
4291         ceph_encode_8(&p, CEPH_OSD_BACKOFF_OP_ACK_BLOCK);
4292         ceph_encode_64(&p, backoff->id);
4293         encode_hoid(&p, end, backoff->begin);
4294         encode_hoid(&p, end, backoff->end);
4295         BUG_ON(p != end);
4296
4297         msg->front.iov_len = p - msg->front.iov_base;
4298         msg->hdr.version = cpu_to_le16(1); /* MOSDBackoff v1 */
4299         msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
4300
4301         return msg;
4302 }
4303
4304 static void handle_backoff_block(struct ceph_osd *osd, struct MOSDBackoff *m)
4305 {
4306         struct ceph_spg_mapping *spg;
4307         struct ceph_osd_backoff *backoff;
4308         struct ceph_msg *msg;
4309
4310         dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd,
4311              m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id);
4312
4313         spg = lookup_spg_mapping(&osd->o_backoff_mappings, &m->spgid);
4314         if (!spg) {
4315                 spg = alloc_spg_mapping();
4316                 if (!spg) {
4317                         pr_err("%s failed to allocate spg\n", __func__);
4318                         return;
4319                 }
4320                 spg->spgid = m->spgid; /* struct */
4321                 insert_spg_mapping(&osd->o_backoff_mappings, spg);
4322         }
4323
4324         backoff = alloc_backoff();
4325         if (!backoff) {
4326                 pr_err("%s failed to allocate backoff\n", __func__);
4327                 return;
4328         }
4329         backoff->spgid = m->spgid; /* struct */
4330         backoff->id = m->id;
4331         backoff->begin = m->begin;
4332         m->begin = NULL; /* backoff now owns this */
4333         backoff->end = m->end;
4334         m->end = NULL;   /* ditto */
4335
4336         insert_backoff(&spg->backoffs, backoff);
4337         insert_backoff_by_id(&osd->o_backoffs_by_id, backoff);
4338
4339         /*
4340          * Ack with original backoff's epoch so that the OSD can
4341          * discard this if there was a PG split.
4342          */
4343         msg = create_backoff_message(backoff, m->map_epoch);
4344         if (!msg) {
4345                 pr_err("%s failed to allocate msg\n", __func__);
4346                 return;
4347         }
4348         ceph_con_send(&osd->o_con, msg);
4349 }
4350
4351 static bool target_contained_by(const struct ceph_osd_request_target *t,
4352                                 const struct ceph_hobject_id *begin,
4353                                 const struct ceph_hobject_id *end)
4354 {
4355         struct ceph_hobject_id hoid;
4356         int cmp;
4357
4358         hoid_fill_from_target(&hoid, t);
4359         cmp = hoid_compare(&hoid, begin);
4360         return !cmp || (cmp > 0 && hoid_compare(&hoid, end) < 0);
4361 }
4362
4363 static void handle_backoff_unblock(struct ceph_osd *osd,
4364                                    const struct MOSDBackoff *m)
4365 {
4366         struct ceph_spg_mapping *spg;
4367         struct ceph_osd_backoff *backoff;
4368         struct rb_node *n;
4369
4370         dout("%s osd%d spgid %llu.%xs%d id %llu\n", __func__, osd->o_osd,
4371              m->spgid.pgid.pool, m->spgid.pgid.seed, m->spgid.shard, m->id);
4372
4373         backoff = lookup_backoff_by_id(&osd->o_backoffs_by_id, m->id);
4374         if (!backoff) {
4375                 pr_err("%s osd%d spgid %llu.%xs%d id %llu backoff dne\n",
4376                        __func__, osd->o_osd, m->spgid.pgid.pool,
4377                        m->spgid.pgid.seed, m->spgid.shard, m->id);
4378                 return;
4379         }
4380
4381         if (hoid_compare(backoff->begin, m->begin) &&
4382             hoid_compare(backoff->end, m->end)) {
4383                 pr_err("%s osd%d spgid %llu.%xs%d id %llu bad range?\n",
4384                        __func__, osd->o_osd, m->spgid.pgid.pool,
4385                        m->spgid.pgid.seed, m->spgid.shard, m->id);
4386                 /* unblock it anyway... */
4387         }
4388
4389         spg = lookup_spg_mapping(&osd->o_backoff_mappings, &backoff->spgid);
4390         BUG_ON(!spg);
4391
4392         erase_backoff(&spg->backoffs, backoff);
4393         erase_backoff_by_id(&osd->o_backoffs_by_id, backoff);
4394         free_backoff(backoff);
4395
4396         if (RB_EMPTY_ROOT(&spg->backoffs)) {
4397                 erase_spg_mapping(&osd->o_backoff_mappings, spg);
4398                 free_spg_mapping(spg);
4399         }
4400
4401         for (n = rb_first(&osd->o_requests); n; n = rb_next(n)) {
4402                 struct ceph_osd_request *req =
4403                     rb_entry(n, struct ceph_osd_request, r_node);
4404
4405                 if (!ceph_spg_compare(&req->r_t.spgid, &m->spgid)) {
4406                         /*
4407                          * Match against @m, not @backoff -- the PG may
4408                          * have split on the OSD.
4409                          */
4410                         if (target_contained_by(&req->r_t, m->begin, m->end)) {
4411                                 /*
4412                                  * If no other installed backoff applies,
4413                                  * resend.
4414                                  */
4415                                 send_request(req);
4416                         }
4417                 }
4418         }
4419 }
4420
4421 static void handle_backoff(struct ceph_osd *osd, struct ceph_msg *msg)
4422 {
4423         struct ceph_osd_client *osdc = osd->o_osdc;
4424         struct MOSDBackoff m;
4425         int ret;
4426
4427         down_read(&osdc->lock);
4428         if (!osd_registered(osd)) {
4429                 dout("%s osd%d unknown\n", __func__, osd->o_osd);
4430                 up_read(&osdc->lock);
4431                 return;
4432         }
4433         WARN_ON(osd->o_osd != le64_to_cpu(msg->hdr.src.num));
4434
4435         mutex_lock(&osd->lock);
4436         ret = decode_MOSDBackoff(msg, &m);
4437         if (ret) {
4438                 pr_err("failed to decode MOSDBackoff: %d\n", ret);
4439                 ceph_msg_dump(msg);
4440                 goto out_unlock;
4441         }
4442
4443         switch (m.op) {
4444         case CEPH_OSD_BACKOFF_OP_BLOCK:
4445                 handle_backoff_block(osd, &m);
4446                 break;
4447         case CEPH_OSD_BACKOFF_OP_UNBLOCK:
4448                 handle_backoff_unblock(osd, &m);
4449                 break;
4450         default:
4451                 pr_err("%s osd%d unknown op %d\n", __func__, osd->o_osd, m.op);
4452         }
4453
4454         free_hoid(m.begin);
4455         free_hoid(m.end);
4456
4457 out_unlock:
4458         mutex_unlock(&osd->lock);
4459         up_read(&osdc->lock);
4460 }
4461
4462 /*
4463  * Process osd watch notifications
4464  */
4465 static void handle_watch_notify(struct ceph_osd_client *osdc,
4466                                 struct ceph_msg *msg)
4467 {
4468         void *p = msg->front.iov_base;
4469         void *const end = p + msg->front.iov_len;
4470         struct ceph_osd_linger_request *lreq;
4471         struct linger_work *lwork;
4472         u8 proto_ver, opcode;
4473         u64 cookie, notify_id;
4474         u64 notifier_id = 0;
4475         s32 return_code = 0;
4476         void *payload = NULL;
4477         u32 payload_len = 0;
4478
4479         ceph_decode_8_safe(&p, end, proto_ver, bad);
4480         ceph_decode_8_safe(&p, end, opcode, bad);
4481         ceph_decode_64_safe(&p, end, cookie, bad);
4482         p += 8; /* skip ver */
4483         ceph_decode_64_safe(&p, end, notify_id, bad);
4484
4485         if (proto_ver >= 1) {
4486                 ceph_decode_32_safe(&p, end, payload_len, bad);
4487                 ceph_decode_need(&p, end, payload_len, bad);
4488                 payload = p;
4489                 p += payload_len;
4490         }
4491
4492         if (le16_to_cpu(msg->hdr.version) >= 2)
4493                 ceph_decode_32_safe(&p, end, return_code, bad);
4494
4495         if (le16_to_cpu(msg->hdr.version) >= 3)
4496                 ceph_decode_64_safe(&p, end, notifier_id, bad);
4497
4498         down_read(&osdc->lock);
4499         lreq = lookup_linger_osdc(&osdc->linger_requests, cookie);
4500         if (!lreq) {
4501                 dout("%s opcode %d cookie %llu dne\n", __func__, opcode,
4502                      cookie);
4503                 goto out_unlock_osdc;
4504         }
4505
4506         mutex_lock(&lreq->lock);
4507         dout("%s opcode %d cookie %llu lreq %p is_watch %d\n", __func__,
4508              opcode, cookie, lreq, lreq->is_watch);
4509         if (opcode == CEPH_WATCH_EVENT_DISCONNECT) {
4510                 if (!lreq->last_error) {
4511                         lreq->last_error = -ENOTCONN;
4512                         queue_watch_error(lreq);
4513                 }
4514         } else if (!lreq->is_watch) {
4515                 /* CEPH_WATCH_EVENT_NOTIFY_COMPLETE */
4516                 if (lreq->notify_id && lreq->notify_id != notify_id) {
4517                         dout("lreq %p notify_id %llu != %llu, ignoring\n", lreq,
4518                              lreq->notify_id, notify_id);
4519                 } else if (!completion_done(&lreq->notify_finish_wait)) {
4520                         struct ceph_msg_data *data =
4521                             msg->num_data_items ? &msg->data[0] : NULL;
4522
4523                         if (data) {
4524                                 if (lreq->preply_pages) {
4525                                         WARN_ON(data->type !=
4526                                                         CEPH_MSG_DATA_PAGES);
4527                                         *lreq->preply_pages = data->pages;
4528                                         *lreq->preply_len = data->length;
4529                                         data->own_pages = false;
4530                                 }
4531                         }
4532                         lreq->notify_finish_error = return_code;
4533                         complete_all(&lreq->notify_finish_wait);
4534                 }
4535         } else {
4536                 /* CEPH_WATCH_EVENT_NOTIFY */
4537                 lwork = lwork_alloc(lreq, do_watch_notify);
4538                 if (!lwork) {
4539                         pr_err("failed to allocate notify-lwork\n");
4540                         goto out_unlock_lreq;
4541                 }
4542
4543                 lwork->notify.notify_id = notify_id;
4544                 lwork->notify.notifier_id = notifier_id;
4545                 lwork->notify.payload = payload;
4546                 lwork->notify.payload_len = payload_len;
4547                 lwork->notify.msg = ceph_msg_get(msg);
4548                 lwork_queue(lwork);
4549         }
4550
4551 out_unlock_lreq:
4552         mutex_unlock(&lreq->lock);
4553 out_unlock_osdc:
4554         up_read(&osdc->lock);
4555         return;
4556
4557 bad:
4558         pr_err("osdc handle_watch_notify corrupt msg\n");
4559 }
4560
4561 /*
4562  * Register request, send initial attempt.
4563  */
4564 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
4565                             struct ceph_osd_request *req,
4566                             bool nofail)
4567 {
4568         down_read(&osdc->lock);
4569         submit_request(req, false);
4570         up_read(&osdc->lock);
4571
4572         return 0;
4573 }
4574 EXPORT_SYMBOL(ceph_osdc_start_request);
4575
4576 /*
4577  * Unregister a registered request.  The request is not completed:
4578  * ->r_result isn't set and __complete_request() isn't called.
4579  */
4580 void ceph_osdc_cancel_request(struct ceph_osd_request *req)
4581 {
4582         struct ceph_osd_client *osdc = req->r_osdc;
4583
4584         down_write(&osdc->lock);
4585         if (req->r_osd)
4586                 cancel_request(req);
4587         up_write(&osdc->lock);
4588 }
4589 EXPORT_SYMBOL(ceph_osdc_cancel_request);
4590
4591 /*
4592  * @timeout: in jiffies, 0 means "wait forever"
4593  */
4594 static int wait_request_timeout(struct ceph_osd_request *req,
4595                                 unsigned long timeout)
4596 {
4597         long left;
4598
4599         dout("%s req %p tid %llu\n", __func__, req, req->r_tid);
4600         left = wait_for_completion_killable_timeout(&req->r_completion,
4601                                                 ceph_timeout_jiffies(timeout));
4602         if (left <= 0) {
4603                 left = left ?: -ETIMEDOUT;
4604                 ceph_osdc_cancel_request(req);
4605         } else {
4606                 left = req->r_result; /* completed */
4607         }
4608
4609         return left;
4610 }
4611
4612 /*
4613  * wait for a request to complete
4614  */
4615 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
4616                            struct ceph_osd_request *req)
4617 {
4618         return wait_request_timeout(req, 0);
4619 }
4620 EXPORT_SYMBOL(ceph_osdc_wait_request);
4621
4622 /*
4623  * sync - wait for all in-flight requests to flush.  avoid starvation.
4624  */
4625 void ceph_osdc_sync(struct ceph_osd_client *osdc)
4626 {
4627         struct rb_node *n, *p;
4628         u64 last_tid = atomic64_read(&osdc->last_tid);
4629
4630 again:
4631         down_read(&osdc->lock);
4632         for (n = rb_first(&osdc->osds); n; n = rb_next(n)) {
4633                 struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
4634
4635                 mutex_lock(&osd->lock);
4636                 for (p = rb_first(&osd->o_requests); p; p = rb_next(p)) {
4637                         struct ceph_osd_request *req =
4638                             rb_entry(p, struct ceph_osd_request, r_node);
4639
4640                         if (req->r_tid > last_tid)
4641                                 break;
4642
4643                         if (!(req->r_flags & CEPH_OSD_FLAG_WRITE))
4644                                 continue;
4645
4646                         ceph_osdc_get_request(req);
4647                         mutex_unlock(&osd->lock);
4648                         up_read(&osdc->lock);
4649                         dout("%s waiting on req %p tid %llu last_tid %llu\n",
4650                              __func__, req, req->r_tid, last_tid);
4651                         wait_for_completion(&req->r_completion);
4652                         ceph_osdc_put_request(req);
4653                         goto again;
4654                 }
4655
4656                 mutex_unlock(&osd->lock);
4657         }
4658
4659         up_read(&osdc->lock);
4660         dout("%s done last_tid %llu\n", __func__, last_tid);
4661 }
4662 EXPORT_SYMBOL(ceph_osdc_sync);
4663
4664 static struct ceph_osd_request *
4665 alloc_linger_request(struct ceph_osd_linger_request *lreq)
4666 {
4667         struct ceph_osd_request *req;
4668
4669         req = ceph_osdc_alloc_request(lreq->osdc, NULL, 1, false, GFP_NOIO);
4670         if (!req)
4671                 return NULL;
4672
4673         ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid);
4674         ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc);
4675         return req;
4676 }
4677
4678 static struct ceph_osd_request *
4679 alloc_watch_request(struct ceph_osd_linger_request *lreq, u8 watch_opcode)
4680 {
4681         struct ceph_osd_request *req;
4682
4683         req = alloc_linger_request(lreq);
4684         if (!req)
4685                 return NULL;
4686
4687         /*
4688          * Pass 0 for cookie because we don't know it yet, it will be
4689          * filled in by linger_submit().
4690          */
4691         osd_req_op_watch_init(req, 0, 0, watch_opcode);
4692
4693         if (ceph_osdc_alloc_messages(req, GFP_NOIO)) {
4694                 ceph_osdc_put_request(req);
4695                 return NULL;
4696         }
4697
4698         return req;
4699 }
4700
4701 /*
4702  * Returns a handle, caller owns a ref.
4703  */
4704 struct ceph_osd_linger_request *
4705 ceph_osdc_watch(struct ceph_osd_client *osdc,
4706                 struct ceph_object_id *oid,
4707                 struct ceph_object_locator *oloc,
4708                 rados_watchcb2_t wcb,
4709                 rados_watcherrcb_t errcb,
4710                 void *data)
4711 {
4712         struct ceph_osd_linger_request *lreq;
4713         int ret;
4714
4715         lreq = linger_alloc(osdc);
4716         if (!lreq)
4717                 return ERR_PTR(-ENOMEM);
4718
4719         lreq->is_watch = true;
4720         lreq->wcb = wcb;
4721         lreq->errcb = errcb;
4722         lreq->data = data;
4723         lreq->watch_valid_thru = jiffies;
4724
4725         ceph_oid_copy(&lreq->t.base_oid, oid);
4726         ceph_oloc_copy(&lreq->t.base_oloc, oloc);
4727         lreq->t.flags = CEPH_OSD_FLAG_WRITE;
4728         ktime_get_real_ts64(&lreq->mtime);
4729
4730         lreq->reg_req = alloc_watch_request(lreq, CEPH_OSD_WATCH_OP_WATCH);
4731         if (!lreq->reg_req) {
4732                 ret = -ENOMEM;
4733                 goto err_put_lreq;
4734         }
4735
4736         lreq->ping_req = alloc_watch_request(lreq, CEPH_OSD_WATCH_OP_PING);
4737         if (!lreq->ping_req) {
4738                 ret = -ENOMEM;
4739                 goto err_put_lreq;
4740         }
4741
4742         linger_submit(lreq);
4743         ret = linger_reg_commit_wait(lreq);
4744         if (ret) {
4745                 linger_cancel(lreq);
4746                 goto err_put_lreq;
4747         }
4748
4749         return lreq;
4750
4751 err_put_lreq:
4752         linger_put(lreq);
4753         return ERR_PTR(ret);
4754 }
4755 EXPORT_SYMBOL(ceph_osdc_watch);
4756
4757 /*
4758  * Releases a ref.
4759  *
4760  * Times out after mount_timeout to preserve rbd unmap behaviour
4761  * introduced in 2894e1d76974 ("rbd: timeout watch teardown on unmap
4762  * with mount_timeout").
4763  */
4764 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
4765                       struct ceph_osd_linger_request *lreq)
4766 {
4767         struct ceph_options *opts = osdc->client->options;
4768         struct ceph_osd_request *req;
4769         int ret;
4770
4771         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
4772         if (!req)
4773                 return -ENOMEM;
4774
4775         ceph_oid_copy(&req->r_base_oid, &lreq->t.base_oid);
4776         ceph_oloc_copy(&req->r_base_oloc, &lreq->t.base_oloc);
4777         req->r_flags = CEPH_OSD_FLAG_WRITE;
4778         ktime_get_real_ts64(&req->r_mtime);
4779         osd_req_op_watch_init(req, 0, lreq->linger_id,
4780                               CEPH_OSD_WATCH_OP_UNWATCH);
4781
4782         ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
4783         if (ret)
4784                 goto out_put_req;
4785
4786         ceph_osdc_start_request(osdc, req, false);
4787         linger_cancel(lreq);
4788         linger_put(lreq);
4789         ret = wait_request_timeout(req, opts->mount_timeout);
4790
4791 out_put_req:
4792         ceph_osdc_put_request(req);
4793         return ret;
4794 }
4795 EXPORT_SYMBOL(ceph_osdc_unwatch);
4796
4797 static int osd_req_op_notify_ack_init(struct ceph_osd_request *req, int which,
4798                                       u64 notify_id, u64 cookie, void *payload,
4799                                       u32 payload_len)
4800 {
4801         struct ceph_osd_req_op *op;
4802         struct ceph_pagelist *pl;
4803         int ret;
4804
4805         op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY_ACK, 0);
4806
4807         pl = ceph_pagelist_alloc(GFP_NOIO);
4808         if (!pl)
4809                 return -ENOMEM;
4810
4811         ret = ceph_pagelist_encode_64(pl, notify_id);
4812         ret |= ceph_pagelist_encode_64(pl, cookie);
4813         if (payload) {
4814                 ret |= ceph_pagelist_encode_32(pl, payload_len);
4815                 ret |= ceph_pagelist_append(pl, payload, payload_len);
4816         } else {
4817                 ret |= ceph_pagelist_encode_32(pl, 0);
4818         }
4819         if (ret) {
4820                 ceph_pagelist_release(pl);
4821                 return -ENOMEM;
4822         }
4823
4824         ceph_osd_data_pagelist_init(&op->notify_ack.request_data, pl);
4825         op->indata_len = pl->length;
4826         return 0;
4827 }
4828
4829 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
4830                          struct ceph_object_id *oid,
4831                          struct ceph_object_locator *oloc,
4832                          u64 notify_id,
4833                          u64 cookie,
4834                          void *payload,
4835                          u32 payload_len)
4836 {
4837         struct ceph_osd_request *req;
4838         int ret;
4839
4840         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
4841         if (!req)
4842                 return -ENOMEM;
4843
4844         ceph_oid_copy(&req->r_base_oid, oid);
4845         ceph_oloc_copy(&req->r_base_oloc, oloc);
4846         req->r_flags = CEPH_OSD_FLAG_READ;
4847
4848         ret = osd_req_op_notify_ack_init(req, 0, notify_id, cookie, payload,
4849                                          payload_len);
4850         if (ret)
4851                 goto out_put_req;
4852
4853         ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
4854         if (ret)
4855                 goto out_put_req;
4856
4857         ceph_osdc_start_request(osdc, req, false);
4858         ret = ceph_osdc_wait_request(osdc, req);
4859
4860 out_put_req:
4861         ceph_osdc_put_request(req);
4862         return ret;
4863 }
4864 EXPORT_SYMBOL(ceph_osdc_notify_ack);
4865
4866 static int osd_req_op_notify_init(struct ceph_osd_request *req, int which,
4867                                   u64 cookie, u32 prot_ver, u32 timeout,
4868                                   void *payload, u32 payload_len)
4869 {
4870         struct ceph_osd_req_op *op;
4871         struct ceph_pagelist *pl;
4872         int ret;
4873
4874         op = _osd_req_op_init(req, which, CEPH_OSD_OP_NOTIFY, 0);
4875         op->notify.cookie = cookie;
4876
4877         pl = ceph_pagelist_alloc(GFP_NOIO);
4878         if (!pl)
4879                 return -ENOMEM;
4880
4881         ret = ceph_pagelist_encode_32(pl, 1); /* prot_ver */
4882         ret |= ceph_pagelist_encode_32(pl, timeout);
4883         ret |= ceph_pagelist_encode_32(pl, payload_len);
4884         ret |= ceph_pagelist_append(pl, payload, payload_len);
4885         if (ret) {
4886                 ceph_pagelist_release(pl);
4887                 return -ENOMEM;
4888         }
4889
4890         ceph_osd_data_pagelist_init(&op->notify.request_data, pl);
4891         op->indata_len = pl->length;
4892         return 0;
4893 }
4894
4895 /*
4896  * @timeout: in seconds
4897  *
4898  * @preply_{pages,len} are initialized both on success and error.
4899  * The caller is responsible for:
4900  *
4901  *     ceph_release_page_vector(reply_pages, calc_pages_for(0, reply_len))
4902  */
4903 int ceph_osdc_notify(struct ceph_osd_client *osdc,
4904                      struct ceph_object_id *oid,
4905                      struct ceph_object_locator *oloc,
4906                      void *payload,
4907                      u32 payload_len,
4908                      u32 timeout,
4909                      struct page ***preply_pages,
4910                      size_t *preply_len)
4911 {
4912         struct ceph_osd_linger_request *lreq;
4913         struct page **pages;
4914         int ret;
4915
4916         WARN_ON(!timeout);
4917         if (preply_pages) {
4918                 *preply_pages = NULL;
4919                 *preply_len = 0;
4920         }
4921
4922         lreq = linger_alloc(osdc);
4923         if (!lreq)
4924                 return -ENOMEM;
4925
4926         lreq->preply_pages = preply_pages;
4927         lreq->preply_len = preply_len;
4928
4929         ceph_oid_copy(&lreq->t.base_oid, oid);
4930         ceph_oloc_copy(&lreq->t.base_oloc, oloc);
4931         lreq->t.flags = CEPH_OSD_FLAG_READ;
4932
4933         lreq->reg_req = alloc_linger_request(lreq);
4934         if (!lreq->reg_req) {
4935                 ret = -ENOMEM;
4936                 goto out_put_lreq;
4937         }
4938
4939         /*
4940          * Pass 0 for cookie because we don't know it yet, it will be
4941          * filled in by linger_submit().
4942          */
4943         ret = osd_req_op_notify_init(lreq->reg_req, 0, 0, 1, timeout,
4944                                      payload, payload_len);
4945         if (ret)
4946                 goto out_put_lreq;
4947
4948         /* for notify_id */
4949         pages = ceph_alloc_page_vector(1, GFP_NOIO);
4950         if (IS_ERR(pages)) {
4951                 ret = PTR_ERR(pages);
4952                 goto out_put_lreq;
4953         }
4954         ceph_osd_data_pages_init(osd_req_op_data(lreq->reg_req, 0, notify,
4955                                                  response_data),
4956                                  pages, PAGE_SIZE, 0, false, true);
4957
4958         ret = ceph_osdc_alloc_messages(lreq->reg_req, GFP_NOIO);
4959         if (ret)
4960                 goto out_put_lreq;
4961
4962         linger_submit(lreq);
4963         ret = linger_reg_commit_wait(lreq);
4964         if (!ret)
4965                 ret = linger_notify_finish_wait(lreq);
4966         else
4967                 dout("lreq %p failed to initiate notify %d\n", lreq, ret);
4968
4969         linger_cancel(lreq);
4970 out_put_lreq:
4971         linger_put(lreq);
4972         return ret;
4973 }
4974 EXPORT_SYMBOL(ceph_osdc_notify);
4975
4976 /*
4977  * Return the number of milliseconds since the watch was last
4978  * confirmed, or an error.  If there is an error, the watch is no
4979  * longer valid, and should be destroyed with ceph_osdc_unwatch().
4980  */
4981 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
4982                           struct ceph_osd_linger_request *lreq)
4983 {
4984         unsigned long stamp, age;
4985         int ret;
4986
4987         down_read(&osdc->lock);
4988         mutex_lock(&lreq->lock);
4989         stamp = lreq->watch_valid_thru;
4990         if (!list_empty(&lreq->pending_lworks)) {
4991                 struct linger_work *lwork =
4992                     list_first_entry(&lreq->pending_lworks,
4993                                      struct linger_work,
4994                                      pending_item);
4995
4996                 if (time_before(lwork->queued_stamp, stamp))
4997                         stamp = lwork->queued_stamp;
4998         }
4999         age = jiffies - stamp;
5000         dout("%s lreq %p linger_id %llu age %lu last_error %d\n", __func__,
5001              lreq, lreq->linger_id, age, lreq->last_error);
5002         /* we are truncating to msecs, so return a safe upper bound */
5003         ret = lreq->last_error ?: 1 + jiffies_to_msecs(age);
5004
5005         mutex_unlock(&lreq->lock);
5006         up_read(&osdc->lock);
5007         return ret;
5008 }
5009
5010 static int decode_watcher(void **p, void *end, struct ceph_watch_item *item)
5011 {
5012         u8 struct_v;
5013         u32 struct_len;
5014         int ret;
5015
5016         ret = ceph_start_decoding(p, end, 2, "watch_item_t",
5017                                   &struct_v, &struct_len);
5018         if (ret)
5019                 goto bad;
5020
5021         ret = -EINVAL;
5022         ceph_decode_copy_safe(p, end, &item->name, sizeof(item->name), bad);
5023         ceph_decode_64_safe(p, end, item->cookie, bad);
5024         ceph_decode_skip_32(p, end, bad); /* skip timeout seconds */
5025
5026         if (struct_v >= 2) {
5027                 ret = ceph_decode_entity_addr(p, end, &item->addr);
5028                 if (ret)
5029                         goto bad;
5030         } else {
5031                 ret = 0;
5032         }
5033
5034         dout("%s %s%llu cookie %llu addr %s\n", __func__,
5035              ENTITY_NAME(item->name), item->cookie,
5036              ceph_pr_addr(&item->addr));
5037 bad:
5038         return ret;
5039 }
5040
5041 static int decode_watchers(void **p, void *end,
5042                            struct ceph_watch_item **watchers,
5043                            u32 *num_watchers)
5044 {
5045         u8 struct_v;
5046         u32 struct_len;
5047         int i;
5048         int ret;
5049
5050         ret = ceph_start_decoding(p, end, 1, "obj_list_watch_response_t",
5051                                   &struct_v, &struct_len);
5052         if (ret)
5053                 return ret;
5054
5055         *num_watchers = ceph_decode_32(p);
5056         *watchers = kcalloc(*num_watchers, sizeof(**watchers), GFP_NOIO);
5057         if (!*watchers)
5058                 return -ENOMEM;
5059
5060         for (i = 0; i < *num_watchers; i++) {
5061                 ret = decode_watcher(p, end, *watchers + i);
5062                 if (ret) {
5063                         kfree(*watchers);
5064                         return ret;
5065                 }
5066         }
5067
5068         return 0;
5069 }
5070
5071 /*
5072  * On success, the caller is responsible for:
5073  *
5074  *     kfree(watchers);
5075  */
5076 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
5077                             struct ceph_object_id *oid,
5078                             struct ceph_object_locator *oloc,
5079                             struct ceph_watch_item **watchers,
5080                             u32 *num_watchers)
5081 {
5082         struct ceph_osd_request *req;
5083         struct page **pages;
5084         int ret;
5085
5086         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
5087         if (!req)
5088                 return -ENOMEM;
5089
5090         ceph_oid_copy(&req->r_base_oid, oid);
5091         ceph_oloc_copy(&req->r_base_oloc, oloc);
5092         req->r_flags = CEPH_OSD_FLAG_READ;
5093
5094         pages = ceph_alloc_page_vector(1, GFP_NOIO);
5095         if (IS_ERR(pages)) {
5096                 ret = PTR_ERR(pages);
5097                 goto out_put_req;
5098         }
5099
5100         osd_req_op_init(req, 0, CEPH_OSD_OP_LIST_WATCHERS, 0);
5101         ceph_osd_data_pages_init(osd_req_op_data(req, 0, list_watchers,
5102                                                  response_data),
5103                                  pages, PAGE_SIZE, 0, false, true);
5104
5105         ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
5106         if (ret)
5107                 goto out_put_req;
5108
5109         ceph_osdc_start_request(osdc, req, false);
5110         ret = ceph_osdc_wait_request(osdc, req);
5111         if (ret >= 0) {
5112                 void *p = page_address(pages[0]);
5113                 void *const end = p + req->r_ops[0].outdata_len;
5114
5115                 ret = decode_watchers(&p, end, watchers, num_watchers);
5116         }
5117
5118 out_put_req:
5119         ceph_osdc_put_request(req);
5120         return ret;
5121 }
5122 EXPORT_SYMBOL(ceph_osdc_list_watchers);
5123
5124 /*
5125  * Call all pending notify callbacks - for use after a watch is
5126  * unregistered, to make sure no more callbacks for it will be invoked
5127  */
5128 void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc)
5129 {
5130         dout("%s osdc %p\n", __func__, osdc);
5131         flush_workqueue(osdc->notify_wq);
5132 }
5133 EXPORT_SYMBOL(ceph_osdc_flush_notifies);
5134
5135 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc)
5136 {
5137         down_read(&osdc->lock);
5138         maybe_request_map(osdc);
5139         up_read(&osdc->lock);
5140 }
5141 EXPORT_SYMBOL(ceph_osdc_maybe_request_map);
5142
5143 /*
5144  * Execute an OSD class method on an object.
5145  *
5146  * @flags: CEPH_OSD_FLAG_*
5147  * @resp_len: in/out param for reply length
5148  */
5149 int ceph_osdc_call(struct ceph_osd_client *osdc,
5150                    struct ceph_object_id *oid,
5151                    struct ceph_object_locator *oloc,
5152                    const char *class, const char *method,
5153                    unsigned int flags,
5154                    struct page *req_page, size_t req_len,
5155                    struct page **resp_pages, size_t *resp_len)
5156 {
5157         struct ceph_osd_request *req;
5158         int ret;
5159
5160         if (req_len > PAGE_SIZE)
5161                 return -E2BIG;
5162
5163         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_NOIO);
5164         if (!req)
5165                 return -ENOMEM;
5166
5167         ceph_oid_copy(&req->r_base_oid, oid);
5168         ceph_oloc_copy(&req->r_base_oloc, oloc);
5169         req->r_flags = flags;
5170
5171         ret = osd_req_op_cls_init(req, 0, class, method);
5172         if (ret)
5173                 goto out_put_req;
5174
5175         if (req_page)
5176                 osd_req_op_cls_request_data_pages(req, 0, &req_page, req_len,
5177                                                   0, false, false);
5178         if (resp_pages)
5179                 osd_req_op_cls_response_data_pages(req, 0, resp_pages,
5180                                                    *resp_len, 0, false, false);
5181
5182         ret = ceph_osdc_alloc_messages(req, GFP_NOIO);
5183         if (ret)
5184                 goto out_put_req;
5185
5186         ceph_osdc_start_request(osdc, req, false);
5187         ret = ceph_osdc_wait_request(osdc, req);
5188         if (ret >= 0) {
5189                 ret = req->r_ops[0].rval;
5190                 if (resp_pages)
5191                         *resp_len = req->r_ops[0].outdata_len;
5192         }
5193
5194 out_put_req:
5195         ceph_osdc_put_request(req);
5196         return ret;
5197 }
5198 EXPORT_SYMBOL(ceph_osdc_call);
5199
5200 /*
5201  * reset all osd connections
5202  */
5203 void ceph_osdc_reopen_osds(struct ceph_osd_client *osdc)
5204 {
5205         struct rb_node *n;
5206
5207         down_write(&osdc->lock);
5208         for (n = rb_first(&osdc->osds); n; ) {
5209                 struct ceph_osd *osd = rb_entry(n, struct ceph_osd, o_node);
5210
5211                 n = rb_next(n);
5212                 if (!reopen_osd(osd))
5213                         kick_osd_requests(osd);
5214         }
5215         up_write(&osdc->lock);
5216 }
5217
5218 /*
5219  * init, shutdown
5220  */
5221 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
5222 {
5223         int err;
5224
5225         dout("init\n");
5226         osdc->client = client;
5227         init_rwsem(&osdc->lock);
5228         osdc->osds = RB_ROOT;
5229         INIT_LIST_HEAD(&osdc->osd_lru);
5230         spin_lock_init(&osdc->osd_lru_lock);
5231         osd_init(&osdc->homeless_osd);
5232         osdc->homeless_osd.o_osdc = osdc;
5233         osdc->homeless_osd.o_osd = CEPH_HOMELESS_OSD;
5234         osdc->last_linger_id = CEPH_LINGER_ID_START;
5235         osdc->linger_requests = RB_ROOT;
5236         osdc->map_checks = RB_ROOT;
5237         osdc->linger_map_checks = RB_ROOT;
5238         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
5239         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
5240
5241         err = -ENOMEM;
5242         osdc->osdmap = ceph_osdmap_alloc();
5243         if (!osdc->osdmap)
5244                 goto out;
5245
5246         osdc->req_mempool = mempool_create_slab_pool(10,
5247                                                      ceph_osd_request_cache);
5248         if (!osdc->req_mempool)
5249                 goto out_map;
5250
5251         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
5252                                 PAGE_SIZE, CEPH_OSD_SLAB_OPS, 10, "osd_op");
5253         if (err < 0)
5254                 goto out_mempool;
5255         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
5256                                 PAGE_SIZE, CEPH_OSD_SLAB_OPS, 10,
5257                                 "osd_op_reply");
5258         if (err < 0)
5259                 goto out_msgpool;
5260
5261         err = -ENOMEM;
5262         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
5263         if (!osdc->notify_wq)
5264                 goto out_msgpool_reply;
5265
5266         osdc->completion_wq = create_singlethread_workqueue("ceph-completion");
5267         if (!osdc->completion_wq)
5268                 goto out_notify_wq;
5269
5270         schedule_delayed_work(&osdc->timeout_work,
5271                               osdc->client->options->osd_keepalive_timeout);
5272         schedule_delayed_work(&osdc->osds_timeout_work,
5273             round_jiffies_relative(osdc->client->options->osd_idle_ttl));
5274
5275         return 0;
5276
5277 out_notify_wq:
5278         destroy_workqueue(osdc->notify_wq);
5279 out_msgpool_reply:
5280         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
5281 out_msgpool:
5282         ceph_msgpool_destroy(&osdc->msgpool_op);
5283 out_mempool:
5284         mempool_destroy(osdc->req_mempool);
5285 out_map:
5286         ceph_osdmap_destroy(osdc->osdmap);
5287 out:
5288         return err;
5289 }
5290
5291 void ceph_osdc_stop(struct ceph_osd_client *osdc)
5292 {
5293         destroy_workqueue(osdc->completion_wq);
5294         destroy_workqueue(osdc->notify_wq);
5295         cancel_delayed_work_sync(&osdc->timeout_work);
5296         cancel_delayed_work_sync(&osdc->osds_timeout_work);
5297
5298         down_write(&osdc->lock);
5299         while (!RB_EMPTY_ROOT(&osdc->osds)) {
5300                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
5301                                                 struct ceph_osd, o_node);
5302                 close_osd(osd);
5303         }
5304         up_write(&osdc->lock);
5305         WARN_ON(refcount_read(&osdc->homeless_osd.o_ref) != 1);
5306         osd_cleanup(&osdc->homeless_osd);
5307
5308         WARN_ON(!list_empty(&osdc->osd_lru));
5309         WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_requests));
5310         WARN_ON(!RB_EMPTY_ROOT(&osdc->map_checks));
5311         WARN_ON(!RB_EMPTY_ROOT(&osdc->linger_map_checks));
5312         WARN_ON(atomic_read(&osdc->num_requests));
5313         WARN_ON(atomic_read(&osdc->num_homeless));
5314
5315         ceph_osdmap_destroy(osdc->osdmap);
5316         mempool_destroy(osdc->req_mempool);
5317         ceph_msgpool_destroy(&osdc->msgpool_op);
5318         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
5319 }
5320
5321 static int osd_req_op_copy_from_init(struct ceph_osd_request *req,
5322                                      u64 src_snapid, u64 src_version,
5323                                      struct ceph_object_id *src_oid,
5324                                      struct ceph_object_locator *src_oloc,
5325                                      u32 src_fadvise_flags,
5326                                      u32 dst_fadvise_flags,
5327                                      u32 truncate_seq, u64 truncate_size,
5328                                      u8 copy_from_flags)
5329 {
5330         struct ceph_osd_req_op *op;
5331         struct page **pages;
5332         void *p, *end;
5333
5334         pages = ceph_alloc_page_vector(1, GFP_KERNEL);
5335         if (IS_ERR(pages))
5336                 return PTR_ERR(pages);
5337
5338         op = _osd_req_op_init(req, 0, CEPH_OSD_OP_COPY_FROM2,
5339                               dst_fadvise_flags);
5340         op->copy_from.snapid = src_snapid;
5341         op->copy_from.src_version = src_version;
5342         op->copy_from.flags = copy_from_flags;
5343         op->copy_from.src_fadvise_flags = src_fadvise_flags;
5344
5345         p = page_address(pages[0]);
5346         end = p + PAGE_SIZE;
5347         ceph_encode_string(&p, end, src_oid->name, src_oid->name_len);
5348         encode_oloc(&p, end, src_oloc);
5349         ceph_encode_32(&p, truncate_seq);
5350         ceph_encode_64(&p, truncate_size);
5351         op->indata_len = PAGE_SIZE - (end - p);
5352
5353         ceph_osd_data_pages_init(&op->copy_from.osd_data, pages,
5354                                  op->indata_len, 0, false, true);
5355         return 0;
5356 }
5357
5358 int ceph_osdc_copy_from(struct ceph_osd_client *osdc,
5359                         u64 src_snapid, u64 src_version,
5360                         struct ceph_object_id *src_oid,
5361                         struct ceph_object_locator *src_oloc,
5362                         u32 src_fadvise_flags,
5363                         struct ceph_object_id *dst_oid,
5364                         struct ceph_object_locator *dst_oloc,
5365                         u32 dst_fadvise_flags,
5366                         u32 truncate_seq, u64 truncate_size,
5367                         u8 copy_from_flags)
5368 {
5369         struct ceph_osd_request *req;
5370         int ret;
5371
5372         req = ceph_osdc_alloc_request(osdc, NULL, 1, false, GFP_KERNEL);
5373         if (!req)
5374                 return -ENOMEM;
5375
5376         req->r_flags = CEPH_OSD_FLAG_WRITE;
5377
5378         ceph_oloc_copy(&req->r_t.base_oloc, dst_oloc);
5379         ceph_oid_copy(&req->r_t.base_oid, dst_oid);
5380
5381         ret = osd_req_op_copy_from_init(req, src_snapid, src_version, src_oid,
5382                                         src_oloc, src_fadvise_flags,
5383                                         dst_fadvise_flags, truncate_seq,
5384                                         truncate_size, copy_from_flags);
5385         if (ret)
5386                 goto out;
5387
5388         ret = ceph_osdc_alloc_messages(req, GFP_KERNEL);
5389         if (ret)
5390                 goto out;
5391
5392         ceph_osdc_start_request(osdc, req, false);
5393         ret = ceph_osdc_wait_request(osdc, req);
5394
5395 out:
5396         ceph_osdc_put_request(req);
5397         return ret;
5398 }
5399 EXPORT_SYMBOL(ceph_osdc_copy_from);
5400
5401 int __init ceph_osdc_setup(void)
5402 {
5403         size_t size = sizeof(struct ceph_osd_request) +
5404             CEPH_OSD_SLAB_OPS * sizeof(struct ceph_osd_req_op);
5405
5406         BUG_ON(ceph_osd_request_cache);
5407         ceph_osd_request_cache = kmem_cache_create("ceph_osd_request", size,
5408                                                    0, 0, NULL);
5409
5410         return ceph_osd_request_cache ? 0 : -ENOMEM;
5411 }
5412
5413 void ceph_osdc_cleanup(void)
5414 {
5415         BUG_ON(!ceph_osd_request_cache);
5416         kmem_cache_destroy(ceph_osd_request_cache);
5417         ceph_osd_request_cache = NULL;
5418 }
5419
5420 /*
5421  * handle incoming message
5422  */
5423 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
5424 {
5425         struct ceph_osd *osd = con->private;
5426         struct ceph_osd_client *osdc = osd->o_osdc;
5427         int type = le16_to_cpu(msg->hdr.type);
5428
5429         switch (type) {
5430         case CEPH_MSG_OSD_MAP:
5431                 ceph_osdc_handle_map(osdc, msg);
5432                 break;
5433         case CEPH_MSG_OSD_OPREPLY:
5434                 handle_reply(osd, msg);
5435                 break;
5436         case CEPH_MSG_OSD_BACKOFF:
5437                 handle_backoff(osd, msg);
5438                 break;
5439         case CEPH_MSG_WATCH_NOTIFY:
5440                 handle_watch_notify(osdc, msg);
5441                 break;
5442
5443         default:
5444                 pr_err("received unknown message type %d %s\n", type,
5445                        ceph_msg_type_name(type));
5446         }
5447
5448         ceph_msg_put(msg);
5449 }
5450
5451 /*
5452  * Lookup and return message for incoming reply.  Don't try to do
5453  * anything about a larger than preallocated data portion of the
5454  * message at the moment - for now, just skip the message.
5455  */
5456 static struct ceph_msg *get_reply(struct ceph_connection *con,
5457                                   struct ceph_msg_header *hdr,
5458                                   int *skip)
5459 {
5460         struct ceph_osd *osd = con->private;
5461         struct ceph_osd_client *osdc = osd->o_osdc;
5462         struct ceph_msg *m = NULL;
5463         struct ceph_osd_request *req;
5464         int front_len = le32_to_cpu(hdr->front_len);
5465         int data_len = le32_to_cpu(hdr->data_len);
5466         u64 tid = le64_to_cpu(hdr->tid);
5467
5468         down_read(&osdc->lock);
5469         if (!osd_registered(osd)) {
5470                 dout("%s osd%d unknown, skipping\n", __func__, osd->o_osd);
5471                 *skip = 1;
5472                 goto out_unlock_osdc;
5473         }
5474         WARN_ON(osd->o_osd != le64_to_cpu(hdr->src.num));
5475
5476         mutex_lock(&osd->lock);
5477         req = lookup_request(&osd->o_requests, tid);
5478         if (!req) {
5479                 dout("%s osd%d tid %llu unknown, skipping\n", __func__,
5480                      osd->o_osd, tid);
5481                 *skip = 1;
5482                 goto out_unlock_session;
5483         }
5484
5485         ceph_msg_revoke_incoming(req->r_reply);
5486
5487         if (front_len > req->r_reply->front_alloc_len) {
5488                 pr_warn("%s osd%d tid %llu front %d > preallocated %d\n",
5489                         __func__, osd->o_osd, req->r_tid, front_len,
5490                         req->r_reply->front_alloc_len);
5491                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front_len, GFP_NOFS,
5492                                  false);
5493                 if (!m)
5494                         goto out_unlock_session;
5495                 ceph_msg_put(req->r_reply);
5496                 req->r_reply = m;
5497         }
5498
5499         if (data_len > req->r_reply->data_length) {
5500                 pr_warn("%s osd%d tid %llu data %d > preallocated %zu, skipping\n",
5501                         __func__, osd->o_osd, req->r_tid, data_len,
5502                         req->r_reply->data_length);
5503                 m = NULL;
5504                 *skip = 1;
5505                 goto out_unlock_session;
5506         }
5507
5508         m = ceph_msg_get(req->r_reply);
5509         dout("get_reply tid %lld %p\n", tid, m);
5510
5511 out_unlock_session:
5512         mutex_unlock(&osd->lock);
5513 out_unlock_osdc:
5514         up_read(&osdc->lock);
5515         return m;
5516 }
5517
5518 static struct ceph_msg *alloc_msg_with_page_vector(struct ceph_msg_header *hdr)
5519 {
5520         struct ceph_msg *m;
5521         int type = le16_to_cpu(hdr->type);
5522         u32 front_len = le32_to_cpu(hdr->front_len);
5523         u32 data_len = le32_to_cpu(hdr->data_len);
5524
5525         m = ceph_msg_new2(type, front_len, 1, GFP_NOIO, false);
5526         if (!m)
5527                 return NULL;
5528
5529         if (data_len) {
5530                 struct page **pages;
5531
5532                 pages = ceph_alloc_page_vector(calc_pages_for(0, data_len),
5533                                                GFP_NOIO);
5534                 if (IS_ERR(pages)) {
5535                         ceph_msg_put(m);
5536                         return NULL;
5537                 }
5538
5539                 ceph_msg_data_add_pages(m, pages, data_len, 0, true);
5540         }
5541
5542         return m;
5543 }
5544
5545 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
5546                                   struct ceph_msg_header *hdr,
5547                                   int *skip)
5548 {
5549         struct ceph_osd *osd = con->private;
5550         int type = le16_to_cpu(hdr->type);
5551
5552         *skip = 0;
5553         switch (type) {
5554         case CEPH_MSG_OSD_MAP:
5555         case CEPH_MSG_OSD_BACKOFF:
5556         case CEPH_MSG_WATCH_NOTIFY:
5557                 return alloc_msg_with_page_vector(hdr);
5558         case CEPH_MSG_OSD_OPREPLY:
5559                 return get_reply(con, hdr, skip);
5560         default:
5561                 pr_warn("%s osd%d unknown msg type %d, skipping\n", __func__,
5562                         osd->o_osd, type);
5563                 *skip = 1;
5564                 return NULL;
5565         }
5566 }
5567
5568 /*
5569  * Wrappers to refcount containing ceph_osd struct
5570  */
5571 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
5572 {
5573         struct ceph_osd *osd = con->private;
5574         if (get_osd(osd))
5575                 return con;
5576         return NULL;
5577 }
5578
5579 static void put_osd_con(struct ceph_connection *con)
5580 {
5581         struct ceph_osd *osd = con->private;
5582         put_osd(osd);
5583 }
5584
5585 /*
5586  * authentication
5587  */
5588 /*
5589  * Note: returned pointer is the address of a structure that's
5590  * managed separately.  Caller must *not* attempt to free it.
5591  */
5592 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
5593                                         int *proto, int force_new)
5594 {
5595         struct ceph_osd *o = con->private;
5596         struct ceph_osd_client *osdc = o->o_osdc;
5597         struct ceph_auth_client *ac = osdc->client->monc.auth;
5598         struct ceph_auth_handshake *auth = &o->o_auth;
5599
5600         if (force_new && auth->authorizer) {
5601                 ceph_auth_destroy_authorizer(auth->authorizer);
5602                 auth->authorizer = NULL;
5603         }
5604         if (!auth->authorizer) {
5605                 int ret = ceph_auth_create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
5606                                                       auth);
5607                 if (ret)
5608                         return ERR_PTR(ret);
5609         } else {
5610                 int ret = ceph_auth_update_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
5611                                                      auth);
5612                 if (ret)
5613                         return ERR_PTR(ret);
5614         }
5615         *proto = ac->protocol;
5616
5617         return auth;
5618 }
5619
5620 static int add_authorizer_challenge(struct ceph_connection *con,
5621                                     void *challenge_buf, int challenge_buf_len)
5622 {
5623         struct ceph_osd *o = con->private;
5624         struct ceph_osd_client *osdc = o->o_osdc;
5625         struct ceph_auth_client *ac = osdc->client->monc.auth;
5626
5627         return ceph_auth_add_authorizer_challenge(ac, o->o_auth.authorizer,
5628                                             challenge_buf, challenge_buf_len);
5629 }
5630
5631 static int verify_authorizer_reply(struct ceph_connection *con)
5632 {
5633         struct ceph_osd *o = con->private;
5634         struct ceph_osd_client *osdc = o->o_osdc;
5635         struct ceph_auth_client *ac = osdc->client->monc.auth;
5636
5637         return ceph_auth_verify_authorizer_reply(ac, o->o_auth.authorizer);
5638 }
5639
5640 static int invalidate_authorizer(struct ceph_connection *con)
5641 {
5642         struct ceph_osd *o = con->private;
5643         struct ceph_osd_client *osdc = o->o_osdc;
5644         struct ceph_auth_client *ac = osdc->client->monc.auth;
5645
5646         ceph_auth_invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
5647         return ceph_monc_validate_auth(&osdc->client->monc);
5648 }
5649
5650 static void osd_reencode_message(struct ceph_msg *msg)
5651 {
5652         int type = le16_to_cpu(msg->hdr.type);
5653
5654         if (type == CEPH_MSG_OSD_OP)
5655                 encode_request_finish(msg);
5656 }
5657
5658 static int osd_sign_message(struct ceph_msg *msg)
5659 {
5660         struct ceph_osd *o = msg->con->private;
5661         struct ceph_auth_handshake *auth = &o->o_auth;
5662
5663         return ceph_auth_sign_message(auth, msg);
5664 }
5665
5666 static int osd_check_message_signature(struct ceph_msg *msg)
5667 {
5668         struct ceph_osd *o = msg->con->private;
5669         struct ceph_auth_handshake *auth = &o->o_auth;
5670
5671         return ceph_auth_check_message_signature(auth, msg);
5672 }
5673
5674 static const struct ceph_connection_operations osd_con_ops = {
5675         .get = get_osd_con,
5676         .put = put_osd_con,
5677         .dispatch = dispatch,
5678         .get_authorizer = get_authorizer,
5679         .add_authorizer_challenge = add_authorizer_challenge,
5680         .verify_authorizer_reply = verify_authorizer_reply,
5681         .invalidate_authorizer = invalidate_authorizer,
5682         .alloc_msg = alloc_msg,
5683         .reencode_message = osd_reencode_message,
5684         .sign_message = osd_sign_message,
5685         .check_message_signature = osd_check_message_signature,
5686         .fault = osd_fault,
5687 };