net/ceph/osd_client.c

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