SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/ceph/osd_client.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/err.h>
   4 #include <linux/highmem.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/slab.h>
   8 #include <linux/uaccess.h>
   9
  10 #include "super.h"
  11 #include "osd_client.h"
  12 #include "messenger.h"
  13 #include "decode.h"
  14 #include "auth.h"
  15
  16 #define OSD_REPLY_RESERVE_FRONT_LEN     512
  17
  18 const static struct ceph_connection_operations osd_con_ops;
  19
  20 static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  21
  22 /*
  23  * Implement client access to distributed object storage cluster.
  24  *
  25  * All data objects are stored within a cluster/cloud of OSDs, or
  26  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  27  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  28  * remote daemons serving up and coordinating consistent and safe
  29  * access to storage.
  30  *
  31  * Cluster membership and the mapping of data objects onto storage devices
  32  * are described by the osd map.
  33  *
  34  * We keep track of pending OSD requests (read, write), resubmit
  35  * requests to different OSDs when the cluster topology/data layout
  36  * change, or retry the affected requests when the communications
  37  * channel with an OSD is reset.
  38  */
  39
  40 /*
  41  * calculate the mapping of a file extent onto an object, and fill out the
  42  * request accordingly.  shorten extent as necessary if it crosses an
  43  * object boundary.
  44  *
  45  * fill osd op in request message.
  46  */
  47 static void calc_layout(struct ceph_osd_client *osdc,
  48                         struct ceph_vino vino, struct ceph_file_layout *layout,
  49                         u64 off, u64 *plen,
  50                         struct ceph_osd_request *req)
  51 {
  52         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
  53         struct ceph_osd_op *op = (void *)(reqhead + 1);
  54         u64 orig_len = *plen;
  55         u64 objoff, objlen;    /* extent in object */
  56         u64 bno;
  57
  58         reqhead->snapid = cpu_to_le64(vino.snap);
  59
  60         /* object extent? */
  61         ceph_calc_file_object_mapping(layout, off, plen, &bno,
  62                                       &objoff, &objlen);
  63         if (*plen < orig_len)
  64                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  65                      orig_len - *plen, off, *plen);
  66
  67         sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
  68         req->r_oid_len = strlen(req->r_oid);
  69
  70         op->extent.offset = cpu_to_le64(objoff);
  71         op->extent.length = cpu_to_le64(objlen);
  72         req->r_num_pages = calc_pages_for(off, *plen);
  73
  74         dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
  75              req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
  76 }
  77
  78 static void remove_replies(struct ceph_osd_request *req)
  79 {
  80         int i;
  81         int max = ARRAY_SIZE(req->replies);
  82
  83         for (i=0; i<max; i++) {
  84                 if (req->replies[i])
  85                         ceph_msg_put(req->replies[i]);
  86         }
  87 }
  88
  89 /*
  90  * requests
  91  */
  92 void ceph_osdc_release_request(struct kref *kref)
  93 {
  94         struct ceph_osd_request *req = container_of(kref,
  95                                                     struct ceph_osd_request,
  96                                                     r_kref);
  97
  98         if (req->r_request)
  99                 ceph_msg_put(req->r_request);
 100         if (req->r_reply)
 101                 ceph_msg_put(req->r_reply);
 102         remove_replies(req);
 103         if (req->r_con_filling_msg) {
 104                 dout("release_request revoking pages %p from con %p\n",
 105                      req->r_pages, req->r_con_filling_msg);
 106                 ceph_con_revoke_message(req->r_con_filling_msg,
 107                                       req->r_reply);
 108                 ceph_con_put(req->r_con_filling_msg);
 109         }
 110         if (req->r_own_pages)
 111                 ceph_release_page_vector(req->r_pages,
 112                                          req->r_num_pages);
 113         ceph_put_snap_context(req->r_snapc);
 114         if (req->r_mempool)
 115                 mempool_free(req, req->r_osdc->req_mempool);
 116         else
 117                 kfree(req);
 118 }
 119
 120 static int alloc_replies(struct ceph_osd_request *req, int num_reply)
 121 {
 122         int i;
 123         int max = ARRAY_SIZE(req->replies);
 124
 125         BUG_ON(num_reply > max);
 126
 127         for (i=0; i<num_reply; i++) {
 128                 req->replies[i] = ceph_msg_new(0, OSD_REPLY_RESERVE_FRONT_LEN, 0, 0, NULL);
 129                 if (IS_ERR(req->replies[i])) {
 130                         int j;
 131                         int err = PTR_ERR(req->replies[i]);
 132                         for (j = 0; j<=i; j++) {
 133                                 ceph_msg_put(req->replies[j]);
 134                         }
 135                         return err;
 136                 }
 137         }
 138
 139         for (; i<max; i++) {
 140                 req->replies[i] = NULL;
 141         }
 142
 143         req->cur_reply = 0;
 144
 145         return 0;
 146 }
 147
 148 static struct ceph_msg *__get_next_reply(struct ceph_connection *con,
 149                                        struct ceph_osd_request *req,
 150                                        int front_len)
 151 {
 152         struct ceph_msg *reply;
 153         if (req->r_con_filling_msg) {
 154                 dout("revoking reply msg %p from old con %p\n", req->r_reply,
 155                      req->r_con_filling_msg);
 156                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
 157                 ceph_con_put(req->r_con_filling_msg);
 158                 req->cur_reply = 0;
 159         }
 160         reply = req->replies[req->cur_reply];
 161         if (!reply || front_len > OSD_REPLY_RESERVE_FRONT_LEN) {
 162                 /* maybe we can allocate it now? */
 163                 reply = ceph_msg_new(0, front_len, 0, 0, NULL);
 164                 if (!reply || IS_ERR(reply)) {
 165                         pr_err(" reply alloc failed, front_len=%d\n", front_len);
 166                         return ERR_PTR(-ENOMEM);
 167                 }
 168         }
 169         req->r_con_filling_msg = ceph_con_get(con);
 170         req->r_reply = ceph_msg_get(reply); /* for duration of read over socket */
 171         return ceph_msg_get(reply);
 172 }
 173
 174 /*
 175  * build new request AND message, calculate layout, and adjust file
 176  * extent as needed.
 177  *
 178  * if the file was recently truncated, we include information about its
 179  * old and new size so that the object can be updated appropriately.  (we
 180  * avoid synchronously deleting truncated objects because it's slow.)
 181  *
 182  * if @do_sync, include a 'startsync' command so that the osd will flush
 183  * data quickly.
 184  */
 185 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 186                                                struct ceph_file_layout *layout,
 187                                                struct ceph_vino vino,
 188                                                u64 off, u64 *plen,
 189                                                int opcode, int flags,
 190                                                struct ceph_snap_context *snapc,
 191                                                int do_sync,
 192                                                u32 truncate_seq,
 193                                                u64 truncate_size,
 194                                                struct timespec *mtime,
 195                                                bool use_mempool, int num_reply)
 196 {
 197         struct ceph_osd_request *req;
 198         struct ceph_msg *msg;
 199         struct ceph_osd_request_head *head;
 200         struct ceph_osd_op *op;
 201         void *p;
 202         int num_op = 1 + do_sync;
 203         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 204         int err, i;
 205
 206         if (use_mempool) {
 207                 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
 208                 memset(req, 0, sizeof(*req));
 209         } else {
 210                 req = kzalloc(sizeof(*req), GFP_NOFS);
 211         }
 212         if (req == NULL)
 213                 return ERR_PTR(-ENOMEM);
 214
 215         err = alloc_replies(req, num_reply);
 216         if (err) {
 217                 ceph_osdc_put_request(req);
 218                 return ERR_PTR(-ENOMEM);
 219         }
 220         req->r_num_prealloc_reply = num_reply;
 221
 222         req->r_osdc = osdc;
 223         req->r_mempool = use_mempool;
 224         kref_init(&req->r_kref);
 225         init_completion(&req->r_completion);
 226         init_completion(&req->r_safe_completion);
 227         INIT_LIST_HEAD(&req->r_unsafe_item);
 228         req->r_flags = flags;
 229
 230         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 231
 232         /* create message; allow space for oid */
 233         msg_size += 40;
 234         if (snapc)
 235                 msg_size += sizeof(u64) * snapc->num_snaps;
 236         if (use_mempool)
 237                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 238         else
 239                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, 0, 0, NULL);
 240         if (IS_ERR(msg)) {
 241                 ceph_osdc_put_request(req);
 242                 return ERR_PTR(PTR_ERR(msg));
 243         }
 244         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
 245         memset(msg->front.iov_base, 0, msg->front.iov_len);
 246         head = msg->front.iov_base;
 247         op = (void *)(head + 1);
 248         p = (void *)(op + num_op);
 249
 250         req->r_request = msg;
 251         req->r_snapc = ceph_get_snap_context(snapc);
 252
 253         head->client_inc = cpu_to_le32(1); /* always, for now. */
 254         head->flags = cpu_to_le32(flags);
 255         if (flags & CEPH_OSD_FLAG_WRITE)
 256                 ceph_encode_timespec(&head->mtime, mtime);
 257         head->num_ops = cpu_to_le16(num_op);
 258         op->op = cpu_to_le16(opcode);
 259
 260         /* calculate max write size */
 261         calc_layout(osdc, vino, layout, off, plen, req);
 262         req->r_file_layout = *layout;  /* keep a copy */
 263
 264         if (flags & CEPH_OSD_FLAG_WRITE) {
 265                 req->r_request->hdr.data_off = cpu_to_le16(off);
 266                 req->r_request->hdr.data_len = cpu_to_le32(*plen);
 267                 op->payload_len = cpu_to_le32(*plen);
 268         }
 269         op->extent.truncate_size = cpu_to_le64(truncate_size);
 270         op->extent.truncate_seq = cpu_to_le32(truncate_seq);
 271
 272         /* fill in oid */
 273         head->object_len = cpu_to_le32(req->r_oid_len);
 274         memcpy(p, req->r_oid, req->r_oid_len);
 275         p += req->r_oid_len;
 276
 277         if (do_sync) {
 278                 op++;
 279                 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
 280         }
 281         if (snapc) {
 282                 head->snap_seq = cpu_to_le64(snapc->seq);
 283                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 284                 for (i = 0; i < snapc->num_snaps; i++) {
 285                         put_unaligned_le64(snapc->snaps[i], p);
 286                         p += sizeof(u64);
 287                 }
 288         }
 289
 290         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 291         return req;
 292 }
 293
 294 /*
 295  * We keep osd requests in an rbtree, sorted by ->r_tid.
 296  */
 297 static void __insert_request(struct ceph_osd_client *osdc,
 298                              struct ceph_osd_request *new)
 299 {
 300         struct rb_node **p = &osdc->requests.rb_node;
 301         struct rb_node *parent = NULL;
 302         struct ceph_osd_request *req = NULL;
 303
 304         while (*p) {
 305                 parent = *p;
 306                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 307                 if (new->r_tid < req->r_tid)
 308                         p = &(*p)->rb_left;
 309                 else if (new->r_tid > req->r_tid)
 310                         p = &(*p)->rb_right;
 311                 else
 312                         BUG();
 313         }
 314
 315         rb_link_node(&new->r_node, parent, p);
 316         rb_insert_color(&new->r_node, &osdc->requests);
 317 }
 318
 319 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 320                                                  u64 tid)
 321 {
 322         struct ceph_osd_request *req;
 323         struct rb_node *n = osdc->requests.rb_node;
 324
 325         while (n) {
 326                 req = rb_entry(n, struct ceph_osd_request, r_node);
 327                 if (tid < req->r_tid)
 328                         n = n->rb_left;
 329                 else if (tid > req->r_tid)
 330                         n = n->rb_right;
 331                 else
 332                         return req;
 333         }
 334         return NULL;
 335 }
 336
 337 static struct ceph_osd_request *
 338 __lookup_request_ge(struct ceph_osd_client *osdc,
 339                     u64 tid)
 340 {
 341         struct ceph_osd_request *req;
 342         struct rb_node *n = osdc->requests.rb_node;
 343
 344         while (n) {
 345                 req = rb_entry(n, struct ceph_osd_request, r_node);
 346                 if (tid < req->r_tid) {
 347                         if (!n->rb_left)
 348                                 return req;
 349                         n = n->rb_left;
 350                 } else if (tid > req->r_tid) {
 351                         n = n->rb_right;
 352                 } else {
 353                         return req;
 354                 }
 355         }
 356         return NULL;
 357 }
 358
 359
 360 /*
 361  * If the osd connection drops, we need to resubmit all requests.
 362  */
 363 static void osd_reset(struct ceph_connection *con)
 364 {
 365         struct ceph_osd *osd = con->private;
 366         struct ceph_osd_client *osdc;
 367
 368         if (!osd)
 369                 return;
 370         dout("osd_reset osd%d\n", osd->o_osd);
 371         osdc = osd->o_osdc;
 372         down_read(&osdc->map_sem);
 373         kick_requests(osdc, osd);
 374         up_read(&osdc->map_sem);
 375 }
 376
 377 /*
 378  * Track open sessions with osds.
 379  */
 380 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
 381 {
 382         struct ceph_osd *osd;
 383
 384         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 385         if (!osd)
 386                 return NULL;
 387
 388         atomic_set(&osd->o_ref, 1);
 389         osd->o_osdc = osdc;
 390         INIT_LIST_HEAD(&osd->o_requests);
 391         INIT_LIST_HEAD(&osd->o_osd_lru);
 392         osd->o_incarnation = 1;
 393
 394         ceph_con_init(osdc->client->msgr, &osd->o_con);
 395         osd->o_con.private = osd;
 396         osd->o_con.ops = &osd_con_ops;
 397         osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
 398
 399         return osd;
 400 }
 401
 402 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 403 {
 404         if (atomic_inc_not_zero(&osd->o_ref)) {
 405                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 406                      atomic_read(&osd->o_ref));
 407                 return osd;
 408         } else {
 409                 dout("get_osd %p FAIL\n", osd);
 410                 return NULL;
 411         }
 412 }
 413
 414 static void put_osd(struct ceph_osd *osd)
 415 {
 416         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 417              atomic_read(&osd->o_ref) - 1);
 418         if (atomic_dec_and_test(&osd->o_ref))
 419                 kfree(osd);
 420 }
 421
 422 /*
 423  * remove an osd from our map
 424  */
 425 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 426 {
 427         dout("__remove_osd %p\n", osd);
 428         BUG_ON(!list_empty(&osd->o_requests));
 429         rb_erase(&osd->o_node, &osdc->osds);
 430         list_del_init(&osd->o_osd_lru);
 431         ceph_con_close(&osd->o_con);
 432         put_osd(osd);
 433 }
 434
 435 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 436                               struct ceph_osd *osd)
 437 {
 438         dout("__move_osd_to_lru %p\n", osd);
 439         BUG_ON(!list_empty(&osd->o_osd_lru));
 440         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 441         osd->lru_ttl = jiffies + osdc->client->mount_args->osd_idle_ttl * HZ;
 442 }
 443
 444 static void __remove_osd_from_lru(struct ceph_osd *osd)
 445 {
 446         dout("__remove_osd_from_lru %p\n", osd);
 447         if (!list_empty(&osd->o_osd_lru))
 448                 list_del_init(&osd->o_osd_lru);
 449 }
 450
 451 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
 452 {
 453         struct ceph_osd *osd, *nosd;
 454
 455         dout("__remove_old_osds %p\n", osdc);
 456         mutex_lock(&osdc->request_mutex);
 457         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 458                 if (!remove_all && time_before(jiffies, osd->lru_ttl))
 459                         break;
 460                 __remove_osd(osdc, osd);
 461         }
 462         mutex_unlock(&osdc->request_mutex);
 463 }
 464
 465 /*
 466  * reset osd connect
 467  */
 468 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 469 {
 470         int ret = 0;
 471
 472         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 473         if (list_empty(&osd->o_requests)) {
 474                 __remove_osd(osdc, osd);
 475         } else {
 476                 ceph_con_close(&osd->o_con);
 477                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
 478                 osd->o_incarnation++;
 479         }
 480         return ret;
 481 }
 482
 483 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 484 {
 485         struct rb_node **p = &osdc->osds.rb_node;
 486         struct rb_node *parent = NULL;
 487         struct ceph_osd *osd = NULL;
 488
 489         while (*p) {
 490                 parent = *p;
 491                 osd = rb_entry(parent, struct ceph_osd, o_node);
 492                 if (new->o_osd < osd->o_osd)
 493                         p = &(*p)->rb_left;
 494                 else if (new->o_osd > osd->o_osd)
 495                         p = &(*p)->rb_right;
 496                 else
 497                         BUG();
 498         }
 499
 500         rb_link_node(&new->o_node, parent, p);
 501         rb_insert_color(&new->o_node, &osdc->osds);
 502 }
 503
 504 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 505 {
 506         struct ceph_osd *osd;
 507         struct rb_node *n = osdc->osds.rb_node;
 508
 509         while (n) {
 510                 osd = rb_entry(n, struct ceph_osd, o_node);
 511                 if (o < osd->o_osd)
 512                         n = n->rb_left;
 513                 else if (o > osd->o_osd)
 514                         n = n->rb_right;
 515                 else
 516                         return osd;
 517         }
 518         return NULL;
 519 }
 520
 521
 522 /*
 523  * Register request, assign tid.  If this is the first request, set up
 524  * the timeout event.
 525  */
 526 static void register_request(struct ceph_osd_client *osdc,
 527                              struct ceph_osd_request *req)
 528 {
 529         mutex_lock(&osdc->request_mutex);
 530         req->r_tid = ++osdc->last_tid;
 531         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 532
 533         dout("register_request %p tid %lld\n", req, req->r_tid);
 534         __insert_request(osdc, req);
 535         ceph_osdc_get_request(req);
 536         osdc->num_requests++;
 537
 538         req->r_timeout_stamp =
 539                 jiffies + osdc->client->mount_args->osd_timeout*HZ;
 540
 541         if (osdc->num_requests == 1) {
 542                 osdc->timeout_tid = req->r_tid;
 543                 dout("  timeout on tid %llu at %lu\n", req->r_tid,
 544                      req->r_timeout_stamp);
 545                 schedule_delayed_work(&osdc->timeout_work,
 546                       round_jiffies_relative(req->r_timeout_stamp - jiffies));
 547         }
 548         mutex_unlock(&osdc->request_mutex);
 549 }
 550
 551 /*
 552  * called under osdc->request_mutex
 553  */
 554 static void __unregister_request(struct ceph_osd_client *osdc,
 555                                  struct ceph_osd_request *req)
 556 {
 557         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 558         rb_erase(&req->r_node, &osdc->requests);
 559         osdc->num_requests--;
 560
 561         if (req->r_osd) {
 562                 /* make sure the original request isn't in flight. */
 563                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 564
 565                 list_del_init(&req->r_osd_item);
 566                 if (list_empty(&req->r_osd->o_requests))
 567                         __move_osd_to_lru(osdc, req->r_osd);
 568                 req->r_osd = NULL;
 569         }
 570
 571         ceph_osdc_put_request(req);
 572
 573         if (req->r_tid == osdc->timeout_tid) {
 574                 if (osdc->num_requests == 0) {
 575                         dout("no requests, canceling timeout\n");
 576                         osdc->timeout_tid = 0;
 577                         cancel_delayed_work(&osdc->timeout_work);
 578                 } else {
 579                         req = rb_entry(rb_first(&osdc->requests),
 580                                        struct ceph_osd_request, r_node);
 581                         osdc->timeout_tid = req->r_tid;
 582                         dout("rescheduled timeout on tid %llu at %lu\n",
 583                              req->r_tid, req->r_timeout_stamp);
 584                         schedule_delayed_work(&osdc->timeout_work,
 585                               round_jiffies_relative(req->r_timeout_stamp -
 586                                                      jiffies));
 587                 }
 588         }
 589 }
 590
 591 /*
 592  * Cancel a previously queued request message
 593  */
 594 static void __cancel_request(struct ceph_osd_request *req)
 595 {
 596         if (req->r_sent) {
 597                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 598                 req->r_sent = 0;
 599         }
 600 }
 601
 602 /*
 603  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 604  * (as needed), and set the request r_osd appropriately.  If there is
 605  * no up osd, set r_osd to NULL.
 606  *
 607  * Return 0 if unchanged, 1 if changed, or negative on error.
 608  *
 609  * Caller should hold map_sem for read and request_mutex.
 610  */
 611 static int __map_osds(struct ceph_osd_client *osdc,
 612                       struct ceph_osd_request *req)
 613 {
 614         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 615         struct ceph_pg pgid;
 616         int o = -1;
 617         int err;
 618         struct ceph_osd *newosd = NULL;
 619
 620         dout("map_osds %p tid %lld\n", req, req->r_tid);
 621         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 622                                       &req->r_file_layout, osdc->osdmap);
 623         if (err)
 624                 return err;
 625         pgid = reqhead->layout.ol_pgid;
 626         req->r_pgid = pgid;
 627
 628         o = ceph_calc_pg_primary(osdc->osdmap, pgid);
 629
 630         if ((req->r_osd && req->r_osd->o_osd == o &&
 631              req->r_sent >= req->r_osd->o_incarnation) ||
 632             (req->r_osd == NULL && o == -1))
 633                 return 0;  /* no change */
 634
 635         dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
 636              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 637              req->r_osd ? req->r_osd->o_osd : -1);
 638
 639         if (req->r_osd) {
 640                 __cancel_request(req);
 641                 list_del_init(&req->r_osd_item);
 642                 if (list_empty(&req->r_osd->o_requests)) {
 643                         /* try to re-use r_osd if possible */
 644                         newosd = get_osd(req->r_osd);
 645                         __remove_osd(osdc, newosd);
 646                 }
 647                 req->r_osd = NULL;
 648         }
 649
 650         req->r_osd = __lookup_osd(osdc, o);
 651         if (!req->r_osd && o >= 0) {
 652                 if (newosd) {
 653                         req->r_osd = newosd;
 654                         newosd = NULL;
 655                 } else {
 656                         err = -ENOMEM;
 657                         req->r_osd = create_osd(osdc);
 658                         if (!req->r_osd)
 659                                 goto out;
 660                 }
 661
 662                 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
 663                 req->r_osd->o_osd = o;
 664                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
 665                 __insert_osd(osdc, req->r_osd);
 666
 667                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
 668         }
 669
 670         if (req->r_osd) {
 671                 __remove_osd_from_lru(req->r_osd);
 672                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 673         }
 674         err = 1;   /* osd changed */
 675
 676 out:
 677         if (newosd)
 678                 put_osd(newosd);
 679         return err;
 680 }
 681
 682 /*
 683  * caller should hold map_sem (for read) and request_mutex
 684  */
 685 static int __send_request(struct ceph_osd_client *osdc,
 686                           struct ceph_osd_request *req)
 687 {
 688         struct ceph_osd_request_head *reqhead;
 689         int err;
 690
 691         err = __map_osds(osdc, req);
 692         if (err < 0)
 693                 return err;
 694         if (req->r_osd == NULL) {
 695                 dout("send_request %p no up osds in pg\n", req);
 696                 ceph_monc_request_next_osdmap(&osdc->client->monc);
 697                 return 0;
 698         }
 699
 700         dout("send_request %p tid %llu to osd%d flags %d\n",
 701              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
 702
 703         reqhead = req->r_request->front.iov_base;
 704         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
 705         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
 706         reqhead->reassert_version = req->r_reassert_version;
 707
 708         req->r_timeout_stamp = jiffies+osdc->client->mount_args->osd_timeout*HZ;
 709
 710         ceph_msg_get(req->r_request); /* send consumes a ref */
 711         ceph_con_send(&req->r_osd->o_con, req->r_request);
 712         req->r_sent = req->r_osd->o_incarnation;
 713         return 0;
 714 }
 715
 716 /*
 717  * Timeout callback, called every N seconds when 1 or more osd
 718  * requests has been active for more than N seconds.  When this
 719  * happens, we ping all OSDs with requests who have timed out to
 720  * ensure any communications channel reset is detected.  Reset the
 721  * request timeouts another N seconds in the future as we go.
 722  * Reschedule the timeout event another N seconds in future (unless
 723  * there are no open requests).
 724  */
 725 static void handle_timeout(struct work_struct *work)
 726 {
 727         struct ceph_osd_client *osdc =
 728                 container_of(work, struct ceph_osd_client, timeout_work.work);
 729         struct ceph_osd_request *req;
 730         struct ceph_osd *osd;
 731         unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
 732         unsigned long next_timeout = timeout + jiffies;
 733         struct rb_node *p;
 734
 735         dout("timeout\n");
 736         down_read(&osdc->map_sem);
 737
 738         ceph_monc_request_next_osdmap(&osdc->client->monc);
 739
 740         mutex_lock(&osdc->request_mutex);
 741         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
 742                 req = rb_entry(p, struct ceph_osd_request, r_node);
 743
 744                 if (req->r_resend) {
 745                         int err;
 746
 747                         dout("osdc resending prev failed %lld\n", req->r_tid);
 748                         err = __send_request(osdc, req);
 749                         if (err)
 750                                 dout("osdc failed again on %lld\n", req->r_tid);
 751                         else
 752                                 req->r_resend = false;
 753                         continue;
 754                 }
 755         }
 756         for (p = rb_first(&osdc->osds); p; p = rb_next(p)) {
 757                 osd = rb_entry(p, struct ceph_osd, o_node);
 758                 if (list_empty(&osd->o_requests))
 759                         continue;
 760                 req = list_first_entry(&osd->o_requests,
 761                                        struct ceph_osd_request, r_osd_item);
 762                 if (time_before(jiffies, req->r_timeout_stamp))
 763                         continue;
 764
 765                 dout(" tid %llu (at least) timed out on osd%d\n",
 766                      req->r_tid, osd->o_osd);
 767                 req->r_timeout_stamp = next_timeout;
 768                 ceph_con_keepalive(&osd->o_con);
 769         }
 770
 771         if (osdc->timeout_tid)
 772                 schedule_delayed_work(&osdc->timeout_work,
 773                                       round_jiffies_relative(timeout));
 774
 775         mutex_unlock(&osdc->request_mutex);
 776
 777         up_read(&osdc->map_sem);
 778 }
 779
 780 static void handle_osds_timeout(struct work_struct *work)
 781 {
 782         struct ceph_osd_client *osdc =
 783                 container_of(work, struct ceph_osd_client,
 784                              osds_timeout_work.work);
 785         unsigned long delay =
 786                 osdc->client->mount_args->osd_idle_ttl * HZ >> 2;
 787
 788         dout("osds timeout\n");
 789         down_read(&osdc->map_sem);
 790         remove_old_osds(osdc, 0);
 791         up_read(&osdc->map_sem);
 792
 793         schedule_delayed_work(&osdc->osds_timeout_work,
 794                               round_jiffies_relative(delay));
 795 }
 796
 797 /*
 798  * handle osd op reply.  either call the callback if it is specified,
 799  * or do the completion to wake up the waiting thread.
 800  */
 801 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
 802                          struct ceph_connection *con)
 803 {
 804         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
 805         struct ceph_osd_request *req;
 806         u64 tid;
 807         int numops, object_len, flags;
 808
 809         tid = le64_to_cpu(msg->hdr.tid);
 810         if (msg->front.iov_len < sizeof(*rhead))
 811                 goto bad;
 812         numops = le32_to_cpu(rhead->num_ops);
 813         object_len = le32_to_cpu(rhead->object_len);
 814         if (msg->front.iov_len != sizeof(*rhead) + object_len +
 815             numops * sizeof(struct ceph_osd_op))
 816                 goto bad;
 817         dout("handle_reply %p tid %llu\n", msg, tid);
 818
 819         /* lookup */
 820         mutex_lock(&osdc->request_mutex);
 821         req = __lookup_request(osdc, tid);
 822         if (req == NULL) {
 823                 dout("handle_reply tid %llu dne\n", tid);
 824                 mutex_unlock(&osdc->request_mutex);
 825                 return;
 826         }
 827         ceph_osdc_get_request(req);
 828         flags = le32_to_cpu(rhead->flags);
 829
 830         /*
 831          * if this connection filled our message, drop our reference now, to
 832          * avoid a (safe but slower) revoke later.
 833          */
 834         if (req->r_con_filling_msg == con && req->r_reply == msg) {
 835                 dout(" got pages, dropping con_filling_msg ref %p\n", con);
 836                 req->r_con_filling_msg = NULL;
 837                 ceph_con_put(con);
 838         }
 839
 840         if (req->r_reply) {
 841                 /*
 842                  * once we see the message has been received, we don't
 843                  * need a ref (which is only needed for revoking
 844                  * pages)
 845                  */
 846                 ceph_msg_put(req->r_reply);
 847                 req->r_reply = NULL;
 848         }
 849
 850         if (!req->r_got_reply) {
 851                 unsigned bytes;
 852
 853                 req->r_result = le32_to_cpu(rhead->result);
 854                 bytes = le32_to_cpu(msg->hdr.data_len);
 855                 dout("handle_reply result %d bytes %d\n", req->r_result,
 856                      bytes);
 857                 if (req->r_result == 0)
 858                         req->r_result = bytes;
 859
 860                 /* in case this is a write and we need to replay, */
 861                 req->r_reassert_version = rhead->reassert_version;
 862
 863                 req->r_got_reply = 1;
 864         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
 865                 dout("handle_reply tid %llu dup ack\n", tid);
 866                 mutex_unlock(&osdc->request_mutex);
 867                 goto done;
 868         }
 869
 870         dout("handle_reply tid %llu flags %d\n", tid, flags);
 871
 872         /* either this is a read, or we got the safe response */
 873         if ((flags & CEPH_OSD_FLAG_ONDISK) ||
 874             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
 875                 __unregister_request(osdc, req);
 876
 877         mutex_unlock(&osdc->request_mutex);
 878
 879         if (req->r_callback)
 880                 req->r_callback(req, msg);
 881         else
 882                 complete(&req->r_completion);
 883
 884         if (flags & CEPH_OSD_FLAG_ONDISK) {
 885                 if (req->r_safe_callback)
 886                         req->r_safe_callback(req, msg);
 887                 complete(&req->r_safe_completion);  /* fsync waiter */
 888         }
 889
 890 done:
 891         ceph_osdc_put_request(req);
 892         return;
 893
 894 bad:
 895         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
 896                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
 897                (int)sizeof(*rhead));
 898         ceph_msg_dump(msg);
 899 }
 900
 901
 902 /*
 903  * Resubmit osd requests whose osd or osd address has changed.  Request
 904  * a new osd map if osds are down, or we are otherwise unable to determine
 905  * how to direct a request.
 906  *
 907  * Close connections to down osds.
 908  *
 909  * If @who is specified, resubmit requests for that specific osd.
 910  *
 911  * Caller should hold map_sem for read and request_mutex.
 912  */
 913 static void kick_requests(struct ceph_osd_client *osdc,
 914                           struct ceph_osd *kickosd)
 915 {
 916         struct ceph_osd_request *req;
 917         struct rb_node *p, *n;
 918         int needmap = 0;
 919         int err;
 920
 921         dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
 922         mutex_lock(&osdc->request_mutex);
 923         if (kickosd) {
 924                 __reset_osd(osdc, kickosd);
 925         } else {
 926                 for (p = rb_first(&osdc->osds); p; p = n) {
 927                         struct ceph_osd *osd =
 928                                 rb_entry(p, struct ceph_osd, o_node);
 929
 930                         n = rb_next(p);
 931                         if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
 932                             memcmp(&osd->o_con.peer_addr,
 933                                    ceph_osd_addr(osdc->osdmap,
 934                                                  osd->o_osd),
 935                                    sizeof(struct ceph_entity_addr)) != 0)
 936                                 __reset_osd(osdc, osd);
 937                 }
 938         }
 939
 940         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
 941                 req = rb_entry(p, struct ceph_osd_request, r_node);
 942
 943                 if (req->r_resend) {
 944                         dout(" r_resend set on tid %llu\n", req->r_tid);
 945                         __cancel_request(req);
 946                         goto kick;
 947                 }
 948                 if (req->r_osd && kickosd == req->r_osd) {
 949                         __cancel_request(req);
 950                         goto kick;
 951                 }
 952
 953                 err = __map_osds(osdc, req);
 954                 if (err == 0)
 955                         continue;  /* no change */
 956                 if (err < 0) {
 957                         /*
 958                          * FIXME: really, we should set the request
 959                          * error and fail if this isn't a 'nofail'
 960                          * request, but that's a fair bit more
 961                          * complicated to do.  So retry!
 962                          */
 963                         dout(" setting r_resend on %llu\n", req->r_tid);
 964                         req->r_resend = true;
 965                         continue;
 966                 }
 967                 if (req->r_osd == NULL) {
 968                         dout("tid %llu maps to no valid osd\n", req->r_tid);
 969                         needmap++;  /* request a newer map */
 970                         continue;
 971                 }
 972
 973 kick:
 974                 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
 975                      req->r_osd->o_osd);
 976                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
 977                 err = __send_request(osdc, req);
 978                 if (err) {
 979                         dout(" setting r_resend on %llu\n", req->r_tid);
 980                         req->r_resend = true;
 981                 }
 982         }
 983         mutex_unlock(&osdc->request_mutex);
 984
 985         if (needmap) {
 986                 dout("%d requests for down osds, need new map\n", needmap);
 987                 ceph_monc_request_next_osdmap(&osdc->client->monc);
 988         }
 989 }
 990
 991 /*
 992  * Process updated osd map.
 993  *
 994  * The message contains any number of incremental and full maps, normally
 995  * indicating some sort of topology change in the cluster.  Kick requests
 996  * off to different OSDs as needed.
 997  */
 998 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
 999 {
1000         void *p, *end, *next;
1001         u32 nr_maps, maplen;
1002         u32 epoch;
1003         struct ceph_osdmap *newmap = NULL, *oldmap;
1004         int err;
1005         struct ceph_fsid fsid;
1006
1007         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1008         p = msg->front.iov_base;
1009         end = p + msg->front.iov_len;
1010
1011         /* verify fsid */
1012         ceph_decode_need(&p, end, sizeof(fsid), bad);
1013         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1014         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1015                 return;
1016
1017         down_write(&osdc->map_sem);
1018
1019         /* incremental maps */
1020         ceph_decode_32_safe(&p, end, nr_maps, bad);
1021         dout(" %d inc maps\n", nr_maps);
1022         while (nr_maps > 0) {
1023                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1024                 epoch = ceph_decode_32(&p);
1025                 maplen = ceph_decode_32(&p);
1026                 ceph_decode_need(&p, end, maplen, bad);
1027                 next = p + maplen;
1028                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1029                         dout("applying incremental map %u len %d\n",
1030                              epoch, maplen);
1031                         newmap = osdmap_apply_incremental(&p, next,
1032                                                           osdc->osdmap,
1033                                                           osdc->client->msgr);
1034                         if (IS_ERR(newmap)) {
1035                                 err = PTR_ERR(newmap);
1036                                 goto bad;
1037                         }
1038                         BUG_ON(!newmap);
1039                         if (newmap != osdc->osdmap) {
1040                                 ceph_osdmap_destroy(osdc->osdmap);
1041                                 osdc->osdmap = newmap;
1042                         }
1043                 } else {
1044                         dout("ignoring incremental map %u len %d\n",
1045                              epoch, maplen);
1046                 }
1047                 p = next;
1048                 nr_maps--;
1049         }
1050         if (newmap)
1051                 goto done;
1052
1053         /* full maps */
1054         ceph_decode_32_safe(&p, end, nr_maps, bad);
1055         dout(" %d full maps\n", nr_maps);
1056         while (nr_maps) {
1057                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1058                 epoch = ceph_decode_32(&p);
1059                 maplen = ceph_decode_32(&p);
1060                 ceph_decode_need(&p, end, maplen, bad);
1061                 if (nr_maps > 1) {
1062                         dout("skipping non-latest full map %u len %d\n",
1063                              epoch, maplen);
1064                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1065                         dout("skipping full map %u len %d, "
1066                              "older than our %u\n", epoch, maplen,
1067                              osdc->osdmap->epoch);
1068                 } else {
1069                         dout("taking full map %u len %d\n", epoch, maplen);
1070                         newmap = osdmap_decode(&p, p+maplen);
1071                         if (IS_ERR(newmap)) {
1072                                 err = PTR_ERR(newmap);
1073                                 goto bad;
1074                         }
1075                         BUG_ON(!newmap);
1076                         oldmap = osdc->osdmap;
1077                         osdc->osdmap = newmap;
1078                         if (oldmap)
1079                                 ceph_osdmap_destroy(oldmap);
1080                 }
1081                 p += maplen;
1082                 nr_maps--;
1083         }
1084
1085 done:
1086         downgrade_write(&osdc->map_sem);
1087         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1088         if (newmap)
1089                 kick_requests(osdc, NULL);
1090         up_read(&osdc->map_sem);
1091         return;
1092
1093 bad:
1094         pr_err("osdc handle_map corrupt msg\n");
1095         ceph_msg_dump(msg);
1096         up_write(&osdc->map_sem);
1097         return;
1098 }
1099
1100
1101 /*
1102  * A read request prepares specific pages that data is to be read into.
1103  * When a message is being read off the wire, we call prepare_pages to
1104  * find those pages.
1105  *  0 = success, -1 failure.
1106  */
1107 static int __prepare_pages(struct ceph_connection *con,
1108                          struct ceph_msg_header *hdr,
1109                          struct ceph_osd_request *req,
1110                          u64 tid,
1111                          struct ceph_msg *m)
1112 {
1113         struct ceph_osd *osd = con->private;
1114         struct ceph_osd_client *osdc;
1115         int ret = -1;
1116         int data_len = le32_to_cpu(hdr->data_len);
1117         unsigned data_off = le16_to_cpu(hdr->data_off);
1118
1119         int want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1120
1121         if (!osd)
1122                 return -1;
1123
1124         osdc = osd->o_osdc;
1125
1126         dout("__prepare_pages on msg %p tid %llu, has %d pages, want %d\n", m,
1127              tid, req->r_num_pages, want);
1128         if (unlikely(req->r_num_pages < want))
1129                 goto out;
1130         m->pages = req->r_pages;
1131         m->nr_pages = req->r_num_pages;
1132         ret = 0; /* success */
1133 out:
1134         BUG_ON(ret < 0 || m->nr_pages < want);
1135
1136         return ret;
1137 }
1138
1139 /*
1140  * Register request, send initial attempt.
1141  */
1142 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1143                             struct ceph_osd_request *req,
1144                             bool nofail)
1145 {
1146         int rc = 0;
1147
1148         req->r_request->pages = req->r_pages;
1149         req->r_request->nr_pages = req->r_num_pages;
1150
1151         register_request(osdc, req);
1152
1153         down_read(&osdc->map_sem);
1154         mutex_lock(&osdc->request_mutex);
1155         /*
1156          * a racing kick_requests() may have sent the message for us
1157          * while we dropped request_mutex above, so only send now if
1158          * the request still han't been touched yet.
1159          */
1160         if (req->r_sent == 0) {
1161                 rc = __send_request(osdc, req);
1162                 if (rc) {
1163                         if (nofail) {
1164                                 dout("osdc_start_request failed send, "
1165                                      " marking %lld\n", req->r_tid);
1166                                 req->r_resend = true;
1167                                 rc = 0;
1168                         } else {
1169                                 __unregister_request(osdc, req);
1170                         }
1171                 }
1172         }
1173         mutex_unlock(&osdc->request_mutex);
1174         up_read(&osdc->map_sem);
1175         return rc;
1176 }
1177
1178 /*
1179  * wait for a request to complete
1180  */
1181 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1182                            struct ceph_osd_request *req)
1183 {
1184         int rc;
1185
1186         rc = wait_for_completion_interruptible(&req->r_completion);
1187         if (rc < 0) {
1188                 mutex_lock(&osdc->request_mutex);
1189                 __cancel_request(req);
1190                 __unregister_request(osdc, req);
1191                 mutex_unlock(&osdc->request_mutex);
1192                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1193                 return rc;
1194         }
1195
1196         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1197         return req->r_result;
1198 }
1199
1200 /*
1201  * sync - wait for all in-flight requests to flush.  avoid starvation.
1202  */
1203 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1204 {
1205         struct ceph_osd_request *req;
1206         u64 last_tid, next_tid = 0;
1207
1208         mutex_lock(&osdc->request_mutex);
1209         last_tid = osdc->last_tid;
1210         while (1) {
1211                 req = __lookup_request_ge(osdc, next_tid);
1212                 if (!req)
1213                         break;
1214                 if (req->r_tid > last_tid)
1215                         break;
1216
1217                 next_tid = req->r_tid + 1;
1218                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1219                         continue;
1220
1221                 ceph_osdc_get_request(req);
1222                 mutex_unlock(&osdc->request_mutex);
1223                 dout("sync waiting on tid %llu (last is %llu)\n",
1224                      req->r_tid, last_tid);
1225                 wait_for_completion(&req->r_safe_completion);
1226                 mutex_lock(&osdc->request_mutex);
1227                 ceph_osdc_put_request(req);
1228         }
1229         mutex_unlock(&osdc->request_mutex);
1230         dout("sync done (thru tid %llu)\n", last_tid);
1231 }
1232
1233 /*
1234  * init, shutdown
1235  */
1236 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1237 {
1238         int err;
1239
1240         dout("init\n");
1241         osdc->client = client;
1242         osdc->osdmap = NULL;
1243         init_rwsem(&osdc->map_sem);
1244         init_completion(&osdc->map_waiters);
1245         osdc->last_requested_map = 0;
1246         mutex_init(&osdc->request_mutex);
1247         osdc->timeout_tid = 0;
1248         osdc->last_tid = 0;
1249         osdc->osds = RB_ROOT;
1250         INIT_LIST_HEAD(&osdc->osd_lru);
1251         osdc->requests = RB_ROOT;
1252         osdc->num_requests = 0;
1253         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1254         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1255
1256         schedule_delayed_work(&osdc->osds_timeout_work,
1257            round_jiffies_relative(osdc->client->mount_args->osd_idle_ttl * HZ));
1258
1259         err = -ENOMEM;
1260         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1261                                         sizeof(struct ceph_osd_request));
1262         if (!osdc->req_mempool)
1263                 goto out;
1264
1265         err = ceph_msgpool_init(&osdc->msgpool_op, 4096, 10, true);
1266         if (err < 0)
1267                 goto out_mempool;
1268         return 0;
1269
1270 out_mempool:
1271         mempool_destroy(osdc->req_mempool);
1272 out:
1273         return err;
1274 }
1275
1276 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1277 {
1278         cancel_delayed_work_sync(&osdc->timeout_work);
1279         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1280         if (osdc->osdmap) {
1281                 ceph_osdmap_destroy(osdc->osdmap);
1282                 osdc->osdmap = NULL;
1283         }
1284         remove_old_osds(osdc, 1);
1285         mempool_destroy(osdc->req_mempool);
1286         ceph_msgpool_destroy(&osdc->msgpool_op);
1287 }
1288
1289 /*
1290  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1291  * *plen.  Return number of bytes read, or error.
1292  */
1293 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1294                         struct ceph_vino vino, struct ceph_file_layout *layout,
1295                         u64 off, u64 *plen,
1296                         u32 truncate_seq, u64 truncate_size,
1297                         struct page **pages, int num_pages)
1298 {
1299         struct ceph_osd_request *req;
1300         int rc = 0;
1301
1302         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1303              vino.snap, off, *plen);
1304         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1305                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1306                                     NULL, 0, truncate_seq, truncate_size, NULL,
1307                                     false, 1);
1308         if (IS_ERR(req))
1309                 return PTR_ERR(req);
1310
1311         /* it may be a short read due to an object boundary */
1312         req->r_pages = pages;
1313         num_pages = calc_pages_for(off, *plen);
1314         req->r_num_pages = num_pages;
1315
1316         dout("readpages  final extent is %llu~%llu (%d pages)\n",
1317              off, *plen, req->r_num_pages);
1318
1319         rc = ceph_osdc_start_request(osdc, req, false);
1320         if (!rc)
1321                 rc = ceph_osdc_wait_request(osdc, req);
1322
1323         ceph_osdc_put_request(req);
1324         dout("readpages result %d\n", rc);
1325         return rc;
1326 }
1327
1328 /*
1329  * do a synchronous write on N pages
1330  */
1331 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1332                          struct ceph_file_layout *layout,
1333                          struct ceph_snap_context *snapc,
1334                          u64 off, u64 len,
1335                          u32 truncate_seq, u64 truncate_size,
1336                          struct timespec *mtime,
1337                          struct page **pages, int num_pages,
1338                          int flags, int do_sync, bool nofail)
1339 {
1340         struct ceph_osd_request *req;
1341         int rc = 0;
1342
1343         BUG_ON(vino.snap != CEPH_NOSNAP);
1344         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1345                                     CEPH_OSD_OP_WRITE,
1346                                     flags | CEPH_OSD_FLAG_ONDISK |
1347                                             CEPH_OSD_FLAG_WRITE,
1348                                     snapc, do_sync,
1349                                     truncate_seq, truncate_size, mtime,
1350                                     nofail, 1);
1351         if (IS_ERR(req))
1352                 return PTR_ERR(req);
1353
1354         /* it may be a short write due to an object boundary */
1355         req->r_pages = pages;
1356         req->r_num_pages = calc_pages_for(off, len);
1357         dout("writepages %llu~%llu (%d pages)\n", off, len,
1358              req->r_num_pages);
1359
1360         rc = ceph_osdc_start_request(osdc, req, nofail);
1361         if (!rc)
1362                 rc = ceph_osdc_wait_request(osdc, req);
1363
1364         ceph_osdc_put_request(req);
1365         if (rc == 0)
1366                 rc = len;
1367         dout("writepages result %d\n", rc);
1368         return rc;
1369 }
1370
1371 /*
1372  * handle incoming message
1373  */
1374 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1375 {
1376         struct ceph_osd *osd = con->private;
1377         struct ceph_osd_client *osdc;
1378         int type = le16_to_cpu(msg->hdr.type);
1379
1380         if (!osd)
1381                 return;
1382         osdc = osd->o_osdc;
1383
1384         switch (type) {
1385         case CEPH_MSG_OSD_MAP:
1386                 ceph_osdc_handle_map(osdc, msg);
1387                 break;
1388         case CEPH_MSG_OSD_OPREPLY:
1389                 handle_reply(osdc, msg, con);
1390                 break;
1391
1392         default:
1393                 pr_err("received unknown message type %d %s\n", type,
1394                        ceph_msg_type_name(type));
1395         }
1396         ceph_msg_put(msg);
1397 }
1398
1399 /*
1400  * lookup and return message for incoming reply
1401  */
1402 static struct ceph_msg *get_reply(struct ceph_connection *con,
1403                                   struct ceph_msg_header *hdr,
1404                                   int *skip)
1405 {
1406         struct ceph_osd *osd = con->private;
1407         struct ceph_osd_client *osdc = osd->o_osdc;
1408         struct ceph_msg *m;
1409         struct ceph_osd_request *req;
1410         int front = le32_to_cpu(hdr->front_len);
1411         int data_len = le32_to_cpu(hdr->data_len);
1412         u64 tid;
1413         int err;
1414
1415         tid = le64_to_cpu(hdr->tid);
1416         mutex_lock(&osdc->request_mutex);
1417         req = __lookup_request(osdc, tid);
1418         if (!req) {
1419                 *skip = 1;
1420                 m = NULL;
1421                 pr_info("alloc_msg unknown tid %llu from osd%d\n", tid,
1422                         osd->o_osd);
1423                 goto out;
1424         }
1425         m = __get_next_reply(con, req, front);
1426         if (!m || IS_ERR(m)) {
1427                 *skip = 1;
1428                 goto out;
1429         }
1430
1431         if (data_len > 0) {
1432                 err = __prepare_pages(con, hdr, req, tid, m);
1433                 if (err < 0) {
1434                         *skip = 1;
1435                         ceph_msg_put(m);
1436                         m = ERR_PTR(err);
1437                 }
1438         }
1439         *skip = 0;
1440
1441 out:
1442         mutex_unlock(&osdc->request_mutex);
1443         return m;
1444
1445 }
1446
1447 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1448                                   struct ceph_msg_header *hdr,
1449                                   int *skip)
1450 {
1451         struct ceph_osd *osd = con->private;
1452         int type = le16_to_cpu(hdr->type);
1453         int front = le32_to_cpu(hdr->front_len);
1454
1455         switch (type) {
1456         case CEPH_MSG_OSD_MAP:
1457                 return ceph_msg_new(type, front, 0, 0, NULL);
1458         case CEPH_MSG_OSD_OPREPLY:
1459                 return get_reply(con, hdr, skip);
1460         default:
1461                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
1462                         osd->o_osd);
1463                 *skip = 1;
1464                 return NULL;
1465         }
1466 }
1467
1468 /*
1469  * Wrappers to refcount containing ceph_osd struct
1470  */
1471 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1472 {
1473         struct ceph_osd *osd = con->private;
1474         if (get_osd(osd))
1475                 return con;
1476         return NULL;
1477 }
1478
1479 static void put_osd_con(struct ceph_connection *con)
1480 {
1481         struct ceph_osd *osd = con->private;
1482         put_osd(osd);
1483 }
1484
1485 /*
1486  * authentication
1487  */
1488 static int get_authorizer(struct ceph_connection *con,
1489                           void **buf, int *len, int *proto,
1490                           void **reply_buf, int *reply_len, int force_new)
1491 {
1492         struct ceph_osd *o = con->private;
1493         struct ceph_osd_client *osdc = o->o_osdc;
1494         struct ceph_auth_client *ac = osdc->client->monc.auth;
1495         int ret = 0;
1496
1497         if (force_new && o->o_authorizer) {
1498                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1499                 o->o_authorizer = NULL;
1500         }
1501         if (o->o_authorizer == NULL) {
1502                 ret = ac->ops->create_authorizer(
1503                         ac, CEPH_ENTITY_TYPE_OSD,
1504                         &o->o_authorizer,
1505                         &o->o_authorizer_buf,
1506                         &o->o_authorizer_buf_len,
1507                         &o->o_authorizer_reply_buf,
1508                         &o->o_authorizer_reply_buf_len);
1509                 if (ret)
1510                 return ret;
1511         }
1512
1513         *proto = ac->protocol;
1514         *buf = o->o_authorizer_buf;
1515         *len = o->o_authorizer_buf_len;
1516         *reply_buf = o->o_authorizer_reply_buf;
1517         *reply_len = o->o_authorizer_reply_buf_len;
1518         return 0;
1519 }
1520
1521
1522 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1523 {
1524         struct ceph_osd *o = con->private;
1525         struct ceph_osd_client *osdc = o->o_osdc;
1526         struct ceph_auth_client *ac = osdc->client->monc.auth;
1527
1528         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1529 }
1530
1531 static int invalidate_authorizer(struct ceph_connection *con)
1532 {
1533         struct ceph_osd *o = con->private;
1534         struct ceph_osd_client *osdc = o->o_osdc;
1535         struct ceph_auth_client *ac = osdc->client->monc.auth;
1536
1537         if (ac->ops->invalidate_authorizer)
1538                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
1539
1540         return ceph_monc_validate_auth(&osdc->client->monc);
1541 }
1542
1543 const static struct ceph_connection_operations osd_con_ops = {
1544         .get = get_osd_con,
1545         .put = put_osd_con,
1546         .dispatch = dispatch,
1547         .get_authorizer = get_authorizer,
1548         .verify_authorizer_reply = verify_authorizer_reply,
1549         .invalidate_authorizer = invalidate_authorizer,
1550         .alloc_msg = alloc_msg,
1551         .fault = osd_reset,
1552 };