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