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