fs/ceph/osd_client.c

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