net/ceph/osd_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/module.h>
   4 #include <linux/err.h>
   5 #include <linux/highmem.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/slab.h>
   9 #include <linux/uaccess.h>
  10 #ifdef CONFIG_BLOCK
  11 #include <linux/bio.h>
  12 #endif
  13
  14 #include <linux/ceph/libceph.h>
  15 #include <linux/ceph/osd_client.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/auth.h>
  19 #include <linux/ceph/pagelist.h>
  20
  21 #define OSD_OP_FRONT_LEN        4096
  22 #define OSD_OPREPLY_FRONT_LEN   512
  23
  24 static const struct ceph_connection_operations osd_con_ops;
  25
  26 static void send_queued(struct ceph_osd_client *osdc);
  27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  28 static void __register_request(struct ceph_osd_client *osdc,
  29                                struct ceph_osd_request *req);
  30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  31                                         struct ceph_osd_request *req);
  32 static int __send_request(struct ceph_osd_client *osdc,
  33                           struct ceph_osd_request *req);
  34
  35 static int op_needs_trail(int op)
  36 {
  37         switch (op) {
  38         case CEPH_OSD_OP_GETXATTR:
  39         case CEPH_OSD_OP_SETXATTR:
  40         case CEPH_OSD_OP_CMPXATTR:
  41         case CEPH_OSD_OP_CALL:
  42         case CEPH_OSD_OP_NOTIFY:
  43                 return 1;
  44         default:
  45                 return 0;
  46         }
  47 }
  48
  49 static int op_has_extent(int op)
  50 {
  51         return (op == CEPH_OSD_OP_READ ||
  52                 op == CEPH_OSD_OP_WRITE);
  53 }
  54
  55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
  56                         struct ceph_file_layout *layout,
  57                         u64 snapid,
  58                         u64 off, u64 *plen, u64 *bno,
  59                         struct ceph_osd_request *req,
  60                         struct ceph_osd_req_op *op)
  61 {
  62         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
  63         u64 orig_len = *plen;
  64         u64 objoff, objlen;    /* extent in object */
  65
  66         reqhead->snapid = cpu_to_le64(snapid);
  67
  68         /* object extent? */
  69         ceph_calc_file_object_mapping(layout, off, plen, bno,
  70                                       &objoff, &objlen);
  71         if (*plen < orig_len)
  72                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  73                      orig_len - *plen, off, *plen);
  74
  75         if (op_has_extent(op->op)) {
  76                 op->extent.offset = objoff;
  77                 op->extent.length = objlen;
  78         }
  79         req->r_num_pages = calc_pages_for(off, *plen);
  80         req->r_page_alignment = off & ~PAGE_MASK;
  81         if (op->op == CEPH_OSD_OP_WRITE)
  82                 op->payload_len = *plen;
  83
  84         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
  85              *bno, objoff, objlen, req->r_num_pages);
  86
  87 }
  88 EXPORT_SYMBOL(ceph_calc_raw_layout);
  89
  90 /*
  91  * Implement client access to distributed object storage cluster.
  92  *
  93  * All data objects are stored within a cluster/cloud of OSDs, or
  94  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  95  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  96  * remote daemons serving up and coordinating consistent and safe
  97  * access to storage.
  98  *
  99  * Cluster membership and the mapping of data objects onto storage devices
 100  * are described by the osd map.
 101  *
 102  * We keep track of pending OSD requests (read, write), resubmit
 103  * requests to different OSDs when the cluster topology/data layout
 104  * change, or retry the affected requests when the communications
 105  * channel with an OSD is reset.
 106  */
 107
 108 /*
 109  * calculate the mapping of a file extent onto an object, and fill out the
 110  * request accordingly.  shorten extent as necessary if it crosses an
 111  * object boundary.
 112  *
 113  * fill osd op in request message.
 114  */
 115 static void calc_layout(struct ceph_osd_client *osdc,
 116                         struct ceph_vino vino,
 117                         struct ceph_file_layout *layout,
 118                         u64 off, u64 *plen,
 119                         struct ceph_osd_request *req,
 120                         struct ceph_osd_req_op *op)
 121 {
 122         u64 bno;
 123
 124         ceph_calc_raw_layout(osdc, layout, vino.snap, off,
 125                              plen, &bno, req, op);
 126
 127         sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
 128         req->r_oid_len = strlen(req->r_oid);
 129 }
 130
 131 /*
 132  * requests
 133  */
 134 void ceph_osdc_release_request(struct kref *kref)
 135 {
 136         struct ceph_osd_request *req = container_of(kref,
 137                                                     struct ceph_osd_request,
 138                                                     r_kref);
 139
 140         if (req->r_request)
 141                 ceph_msg_put(req->r_request);
 142         if (req->r_reply)
 143                 ceph_msg_put(req->r_reply);
 144         if (req->r_con_filling_msg) {
 145                 dout("release_request revoking pages %p from con %p\n",
 146                      req->r_pages, req->r_con_filling_msg);
 147                 ceph_con_revoke_message(req->r_con_filling_msg,
 148                                       req->r_reply);
 149                 ceph_con_put(req->r_con_filling_msg);
 150         }
 151         if (req->r_own_pages)
 152                 ceph_release_page_vector(req->r_pages,
 153                                          req->r_num_pages);
 154 #ifdef CONFIG_BLOCK
 155         if (req->r_bio)
 156                 bio_put(req->r_bio);
 157 #endif
 158         ceph_put_snap_context(req->r_snapc);
 159         if (req->r_trail) {
 160                 ceph_pagelist_release(req->r_trail);
 161                 kfree(req->r_trail);
 162         }
 163         if (req->r_mempool)
 164                 mempool_free(req, req->r_osdc->req_mempool);
 165         else
 166                 kfree(req);
 167 }
 168 EXPORT_SYMBOL(ceph_osdc_release_request);
 169
 170 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
 171 {
 172         int i = 0;
 173
 174         if (needs_trail)
 175                 *needs_trail = 0;
 176         while (ops[i].op) {
 177                 if (needs_trail && op_needs_trail(ops[i].op))
 178                         *needs_trail = 1;
 179                 i++;
 180         }
 181
 182         return i;
 183 }
 184
 185 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 186                                                int flags,
 187                                                struct ceph_snap_context *snapc,
 188                                                struct ceph_osd_req_op *ops,
 189                                                bool use_mempool,
 190                                                gfp_t gfp_flags,
 191                                                struct page **pages,
 192                                                struct bio *bio)
 193 {
 194         struct ceph_osd_request *req;
 195         struct ceph_msg *msg;
 196         int needs_trail;
 197         int num_op = get_num_ops(ops, &needs_trail);
 198         size_t msg_size = sizeof(struct ceph_osd_request_head);
 199
 200         msg_size += num_op*sizeof(struct ceph_osd_op);
 201
 202         if (use_mempool) {
 203                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 204                 memset(req, 0, sizeof(*req));
 205         } else {
 206                 req = kzalloc(sizeof(*req), gfp_flags);
 207         }
 208         if (req == NULL)
 209                 return NULL;
 210
 211         req->r_osdc = osdc;
 212         req->r_mempool = use_mempool;
 213
 214         kref_init(&req->r_kref);
 215         init_completion(&req->r_completion);
 216         init_completion(&req->r_safe_completion);
 217         INIT_LIST_HEAD(&req->r_unsafe_item);
 218         INIT_LIST_HEAD(&req->r_linger_item);
 219         INIT_LIST_HEAD(&req->r_linger_osd);
 220         req->r_flags = flags;
 221
 222         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 223
 224         /* create reply message */
 225         if (use_mempool)
 226                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 227         else
 228                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 229                                    OSD_OPREPLY_FRONT_LEN, gfp_flags);
 230         if (!msg) {
 231                 ceph_osdc_put_request(req);
 232                 return NULL;
 233         }
 234         req->r_reply = msg;
 235
 236         /* allocate space for the trailing data */
 237         if (needs_trail) {
 238                 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
 239                 if (!req->r_trail) {
 240                         ceph_osdc_put_request(req);
 241                         return NULL;
 242                 }
 243                 ceph_pagelist_init(req->r_trail);
 244         }
 245         /* create request message; allow space for oid */
 246         msg_size += 40;
 247         if (snapc)
 248                 msg_size += sizeof(u64) * snapc->num_snaps;
 249         if (use_mempool)
 250                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 251         else
 252                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags);
 253         if (!msg) {
 254                 ceph_osdc_put_request(req);
 255                 return NULL;
 256         }
 257
 258         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
 259         memset(msg->front.iov_base, 0, msg->front.iov_len);
 260
 261         req->r_request = msg;
 262         req->r_pages = pages;
 263 #ifdef CONFIG_BLOCK
 264         if (bio) {
 265                 req->r_bio = bio;
 266                 bio_get(req->r_bio);
 267         }
 268 #endif
 269
 270         return req;
 271 }
 272 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 273
 274 static void osd_req_encode_op(struct ceph_osd_request *req,
 275                               struct ceph_osd_op *dst,
 276                               struct ceph_osd_req_op *src)
 277 {
 278         dst->op = cpu_to_le16(src->op);
 279
 280         switch (dst->op) {
 281         case CEPH_OSD_OP_READ:
 282         case CEPH_OSD_OP_WRITE:
 283                 dst->extent.offset =
 284                         cpu_to_le64(src->extent.offset);
 285                 dst->extent.length =
 286                         cpu_to_le64(src->extent.length);
 287                 dst->extent.truncate_size =
 288                         cpu_to_le64(src->extent.truncate_size);
 289                 dst->extent.truncate_seq =
 290                         cpu_to_le32(src->extent.truncate_seq);
 291                 break;
 292
 293         case CEPH_OSD_OP_GETXATTR:
 294         case CEPH_OSD_OP_SETXATTR:
 295         case CEPH_OSD_OP_CMPXATTR:
 296                 BUG_ON(!req->r_trail);
 297
 298                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 299                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 300                 dst->xattr.cmp_op = src->xattr.cmp_op;
 301                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 302                 ceph_pagelist_append(req->r_trail, src->xattr.name,
 303                                      src->xattr.name_len);
 304                 ceph_pagelist_append(req->r_trail, src->xattr.val,
 305                                      src->xattr.value_len);
 306                 break;
 307         case CEPH_OSD_OP_CALL:
 308                 BUG_ON(!req->r_trail);
 309
 310                 dst->cls.class_len = src->cls.class_len;
 311                 dst->cls.method_len = src->cls.method_len;
 312                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 313
 314                 ceph_pagelist_append(req->r_trail, src->cls.class_name,
 315                                      src->cls.class_len);
 316                 ceph_pagelist_append(req->r_trail, src->cls.method_name,
 317                                      src->cls.method_len);
 318                 ceph_pagelist_append(req->r_trail, src->cls.indata,
 319                                      src->cls.indata_len);
 320                 break;
 321         case CEPH_OSD_OP_ROLLBACK:
 322                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
 323                 break;
 324         case CEPH_OSD_OP_STARTSYNC:
 325                 break;
 326         case CEPH_OSD_OP_NOTIFY:
 327                 {
 328                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
 329                         __le32 timeout = cpu_to_le32(src->watch.timeout);
 330
 331                         BUG_ON(!req->r_trail);
 332
 333                         ceph_pagelist_append(req->r_trail,
 334                                                 &prot_ver, sizeof(prot_ver));
 335                         ceph_pagelist_append(req->r_trail,
 336                                                 &timeout, sizeof(timeout));
 337                 }
 338         case CEPH_OSD_OP_NOTIFY_ACK:
 339         case CEPH_OSD_OP_WATCH:
 340                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 341                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 342                 dst->watch.flag = src->watch.flag;
 343                 break;
 344         default:
 345                 pr_err("unrecognized osd opcode %d\n", dst->op);
 346                 WARN_ON(1);
 347                 break;
 348         }
 349         dst->payload_len = cpu_to_le32(src->payload_len);
 350 }
 351
 352 /*
 353  * build new request AND message
 354  *
 355  */
 356 void ceph_osdc_build_request(struct ceph_osd_request *req,
 357                              u64 off, u64 *plen,
 358                              struct ceph_osd_req_op *src_ops,
 359                              struct ceph_snap_context *snapc,
 360                              struct timespec *mtime,
 361                              const char *oid,
 362                              int oid_len)
 363 {
 364         struct ceph_msg *msg = req->r_request;
 365         struct ceph_osd_request_head *head;
 366         struct ceph_osd_req_op *src_op;
 367         struct ceph_osd_op *op;
 368         void *p;
 369         int num_op = get_num_ops(src_ops, NULL);
 370         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 371         int flags = req->r_flags;
 372         u64 data_len = 0;
 373         int i;
 374
 375         head = msg->front.iov_base;
 376         op = (void *)(head + 1);
 377         p = (void *)(op + num_op);
 378
 379         req->r_snapc = ceph_get_snap_context(snapc);
 380
 381         head->client_inc = cpu_to_le32(1); /* always, for now. */
 382         head->flags = cpu_to_le32(flags);
 383         if (flags & CEPH_OSD_FLAG_WRITE)
 384                 ceph_encode_timespec(&head->mtime, mtime);
 385         head->num_ops = cpu_to_le16(num_op);
 386
 387
 388         /* fill in oid */
 389         head->object_len = cpu_to_le32(oid_len);
 390         memcpy(p, oid, oid_len);
 391         p += oid_len;
 392
 393         src_op = src_ops;
 394         while (src_op->op) {
 395                 osd_req_encode_op(req, op, src_op);
 396                 src_op++;
 397                 op++;
 398         }
 399
 400         if (req->r_trail)
 401                 data_len += req->r_trail->length;
 402
 403         if (snapc) {
 404                 head->snap_seq = cpu_to_le64(snapc->seq);
 405                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 406                 for (i = 0; i < snapc->num_snaps; i++) {
 407                         put_unaligned_le64(snapc->snaps[i], p);
 408                         p += sizeof(u64);
 409                 }
 410         }
 411
 412         if (flags & CEPH_OSD_FLAG_WRITE) {
 413                 req->r_request->hdr.data_off = cpu_to_le16(off);
 414                 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
 415         } else if (data_len) {
 416                 req->r_request->hdr.data_off = 0;
 417                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 418         }
 419
 420         req->r_request->page_alignment = req->r_page_alignment;
 421
 422         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 423         msg_size = p - msg->front.iov_base;
 424         msg->front.iov_len = msg_size;
 425         msg->hdr.front_len = cpu_to_le32(msg_size);
 426         return;
 427 }
 428 EXPORT_SYMBOL(ceph_osdc_build_request);
 429
 430 /*
 431  * build new request AND message, calculate layout, and adjust file
 432  * extent as needed.
 433  *
 434  * if the file was recently truncated, we include information about its
 435  * old and new size so that the object can be updated appropriately.  (we
 436  * avoid synchronously deleting truncated objects because it's slow.)
 437  *
 438  * if @do_sync, include a 'startsync' command so that the osd will flush
 439  * data quickly.
 440  */
 441 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 442                                                struct ceph_file_layout *layout,
 443                                                struct ceph_vino vino,
 444                                                u64 off, u64 *plen,
 445                                                int opcode, int flags,
 446                                                struct ceph_snap_context *snapc,
 447                                                int do_sync,
 448                                                u32 truncate_seq,
 449                                                u64 truncate_size,
 450                                                struct timespec *mtime,
 451                                                bool use_mempool, int num_reply,
 452                                                int page_align)
 453 {
 454         struct ceph_osd_req_op ops[3];
 455         struct ceph_osd_request *req;
 456
 457         ops[0].op = opcode;
 458         ops[0].extent.truncate_seq = truncate_seq;
 459         ops[0].extent.truncate_size = truncate_size;
 460         ops[0].payload_len = 0;
 461
 462         if (do_sync) {
 463                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 464                 ops[1].payload_len = 0;
 465                 ops[2].op = 0;
 466         } else
 467                 ops[1].op = 0;
 468
 469         req = ceph_osdc_alloc_request(osdc, flags,
 470                                          snapc, ops,
 471                                          use_mempool,
 472                                          GFP_NOFS, NULL, NULL);
 473         if (!req)
 474                 return NULL;
 475
 476         /* calculate max write size */
 477         calc_layout(osdc, vino, layout, off, plen, req, ops);
 478         req->r_file_layout = *layout;  /* keep a copy */
 479
 480         /* in case it differs from natural alignment that calc_layout
 481            filled in for us */
 482         req->r_page_alignment = page_align;
 483
 484         ceph_osdc_build_request(req, off, plen, ops,
 485                                 snapc,
 486                                 mtime,
 487                                 req->r_oid, req->r_oid_len);
 488
 489         return req;
 490 }
 491 EXPORT_SYMBOL(ceph_osdc_new_request);
 492
 493 /*
 494  * We keep osd requests in an rbtree, sorted by ->r_tid.
 495  */
 496 static void __insert_request(struct ceph_osd_client *osdc,
 497                              struct ceph_osd_request *new)
 498 {
 499         struct rb_node **p = &osdc->requests.rb_node;
 500         struct rb_node *parent = NULL;
 501         struct ceph_osd_request *req = NULL;
 502
 503         while (*p) {
 504                 parent = *p;
 505                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 506                 if (new->r_tid < req->r_tid)
 507                         p = &(*p)->rb_left;
 508                 else if (new->r_tid > req->r_tid)
 509                         p = &(*p)->rb_right;
 510                 else
 511                         BUG();
 512         }
 513
 514         rb_link_node(&new->r_node, parent, p);
 515         rb_insert_color(&new->r_node, &osdc->requests);
 516 }
 517
 518 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 519                                                  u64 tid)
 520 {
 521         struct ceph_osd_request *req;
 522         struct rb_node *n = osdc->requests.rb_node;
 523
 524         while (n) {
 525                 req = rb_entry(n, struct ceph_osd_request, r_node);
 526                 if (tid < req->r_tid)
 527                         n = n->rb_left;
 528                 else if (tid > req->r_tid)
 529                         n = n->rb_right;
 530                 else
 531                         return req;
 532         }
 533         return NULL;
 534 }
 535
 536 static struct ceph_osd_request *
 537 __lookup_request_ge(struct ceph_osd_client *osdc,
 538                     u64 tid)
 539 {
 540         struct ceph_osd_request *req;
 541         struct rb_node *n = osdc->requests.rb_node;
 542
 543         while (n) {
 544                 req = rb_entry(n, struct ceph_osd_request, r_node);
 545                 if (tid < req->r_tid) {
 546                         if (!n->rb_left)
 547                                 return req;
 548                         n = n->rb_left;
 549                 } else if (tid > req->r_tid) {
 550                         n = n->rb_right;
 551                 } else {
 552                         return req;
 553                 }
 554         }
 555         return NULL;
 556 }
 557
 558 /*
 559  * Resubmit requests pending on the given osd.
 560  */
 561 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 562                                 struct ceph_osd *osd)
 563 {
 564         struct ceph_osd_request *req, *nreq;
 565         int err;
 566
 567         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 568         err = __reset_osd(osdc, osd);
 569         if (err == -EAGAIN)
 570                 return;
 571
 572         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 573                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 574                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 575                      osd->o_osd);
 576                 if (!req->r_linger)
 577                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 578         }
 579
 580         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 581                                  r_linger_osd) {
 582                 /*
 583                  * reregister request prior to unregistering linger so
 584                  * that r_osd is preserved.
 585                  */
 586                 BUG_ON(!list_empty(&req->r_req_lru_item));
 587                 __register_request(osdc, req);
 588                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 589                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 590                 __unregister_linger_request(osdc, req);
 591                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 592                      osd->o_osd);
 593         }
 594 }
 595
 596 static void kick_osd_requests(struct ceph_osd_client *osdc,
 597                               struct ceph_osd *kickosd)
 598 {
 599         mutex_lock(&osdc->request_mutex);
 600         __kick_osd_requests(osdc, kickosd);
 601         mutex_unlock(&osdc->request_mutex);
 602 }
 603
 604 /*
 605  * If the osd connection drops, we need to resubmit all requests.
 606  */
 607 static void osd_reset(struct ceph_connection *con)
 608 {
 609         struct ceph_osd *osd = con->private;
 610         struct ceph_osd_client *osdc;
 611
 612         if (!osd)
 613                 return;
 614         dout("osd_reset osd%d\n", osd->o_osd);
 615         osdc = osd->o_osdc;
 616         down_read(&osdc->map_sem);
 617         kick_osd_requests(osdc, osd);
 618         send_queued(osdc);
 619         up_read(&osdc->map_sem);
 620 }
 621
 622 /*
 623  * Track open sessions with osds.
 624  */
 625 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
 626 {
 627         struct ceph_osd *osd;
 628
 629         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 630         if (!osd)
 631                 return NULL;
 632
 633         atomic_set(&osd->o_ref, 1);
 634         osd->o_osdc = osdc;
 635         INIT_LIST_HEAD(&osd->o_requests);
 636         INIT_LIST_HEAD(&osd->o_linger_requests);
 637         INIT_LIST_HEAD(&osd->o_osd_lru);
 638         osd->o_incarnation = 1;
 639
 640         ceph_con_init(osdc->client->msgr, &osd->o_con);
 641         osd->o_con.private = osd;
 642         osd->o_con.ops = &osd_con_ops;
 643         osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
 644
 645         INIT_LIST_HEAD(&osd->o_keepalive_item);
 646         return osd;
 647 }
 648
 649 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 650 {
 651         if (atomic_inc_not_zero(&osd->o_ref)) {
 652                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 653                      atomic_read(&osd->o_ref));
 654                 return osd;
 655         } else {
 656                 dout("get_osd %p FAIL\n", osd);
 657                 return NULL;
 658         }
 659 }
 660
 661 static void put_osd(struct ceph_osd *osd)
 662 {
 663         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 664              atomic_read(&osd->o_ref) - 1);
 665         if (atomic_dec_and_test(&osd->o_ref)) {
 666                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 667
 668                 if (osd->o_authorizer)
 669                         ac->ops->destroy_authorizer(ac, osd->o_authorizer);
 670                 kfree(osd);
 671         }
 672 }
 673
 674 /*
 675  * remove an osd from our map
 676  */
 677 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 678 {
 679         dout("__remove_osd %p\n", osd);
 680         BUG_ON(!list_empty(&osd->o_requests));
 681         rb_erase(&osd->o_node, &osdc->osds);
 682         list_del_init(&osd->o_osd_lru);
 683         ceph_con_close(&osd->o_con);
 684         put_osd(osd);
 685 }
 686
 687 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 688                               struct ceph_osd *osd)
 689 {
 690         dout("__move_osd_to_lru %p\n", osd);
 691         BUG_ON(!list_empty(&osd->o_osd_lru));
 692         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 693         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 694 }
 695
 696 static void __remove_osd_from_lru(struct ceph_osd *osd)
 697 {
 698         dout("__remove_osd_from_lru %p\n", osd);
 699         if (!list_empty(&osd->o_osd_lru))
 700                 list_del_init(&osd->o_osd_lru);
 701 }
 702
 703 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
 704 {
 705         struct ceph_osd *osd, *nosd;
 706
 707         dout("__remove_old_osds %p\n", osdc);
 708         mutex_lock(&osdc->request_mutex);
 709         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 710                 if (!remove_all && time_before(jiffies, osd->lru_ttl))
 711                         break;
 712                 __remove_osd(osdc, osd);
 713         }
 714         mutex_unlock(&osdc->request_mutex);
 715 }
 716
 717 /*
 718  * reset osd connect
 719  */
 720 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 721 {
 722         struct ceph_osd_request *req;
 723         int ret = 0;
 724
 725         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 726         if (list_empty(&osd->o_requests) &&
 727             list_empty(&osd->o_linger_requests)) {
 728                 __remove_osd(osdc, osd);
 729         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 730                           &osd->o_con.peer_addr,
 731                           sizeof(osd->o_con.peer_addr)) == 0 &&
 732                    !ceph_con_opened(&osd->o_con)) {
 733                 dout(" osd addr hasn't changed and connection never opened,"
 734                      " letting msgr retry");
 735                 /* touch each r_stamp for handle_timeout()'s benfit */
 736                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 737                         req->r_stamp = jiffies;
 738                 ret = -EAGAIN;
 739         } else {
 740                 ceph_con_close(&osd->o_con);
 741                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
 742                 osd->o_incarnation++;
 743         }
 744         return ret;
 745 }
 746
 747 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 748 {
 749         struct rb_node **p = &osdc->osds.rb_node;
 750         struct rb_node *parent = NULL;
 751         struct ceph_osd *osd = NULL;
 752
 753         while (*p) {
 754                 parent = *p;
 755                 osd = rb_entry(parent, struct ceph_osd, o_node);
 756                 if (new->o_osd < osd->o_osd)
 757                         p = &(*p)->rb_left;
 758                 else if (new->o_osd > osd->o_osd)
 759                         p = &(*p)->rb_right;
 760                 else
 761                         BUG();
 762         }
 763
 764         rb_link_node(&new->o_node, parent, p);
 765         rb_insert_color(&new->o_node, &osdc->osds);
 766 }
 767
 768 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 769 {
 770         struct ceph_osd *osd;
 771         struct rb_node *n = osdc->osds.rb_node;
 772
 773         while (n) {
 774                 osd = rb_entry(n, struct ceph_osd, o_node);
 775                 if (o < osd->o_osd)
 776                         n = n->rb_left;
 777                 else if (o > osd->o_osd)
 778                         n = n->rb_right;
 779                 else
 780                         return osd;
 781         }
 782         return NULL;
 783 }
 784
 785 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 786 {
 787         schedule_delayed_work(&osdc->timeout_work,
 788                         osdc->client->options->osd_keepalive_timeout * HZ);
 789 }
 790
 791 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 792 {
 793         cancel_delayed_work(&osdc->timeout_work);
 794 }
 795
 796 /*
 797  * Register request, assign tid.  If this is the first request, set up
 798  * the timeout event.
 799  */
 800 static void __register_request(struct ceph_osd_client *osdc,
 801                                struct ceph_osd_request *req)
 802 {
 803         req->r_tid = ++osdc->last_tid;
 804         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 805         INIT_LIST_HEAD(&req->r_req_lru_item);
 806
 807         dout("__register_request %p tid %lld\n", req, req->r_tid);
 808         __insert_request(osdc, req);
 809         ceph_osdc_get_request(req);
 810         osdc->num_requests++;
 811
 812         if (osdc->num_requests == 1) {
 813                 dout(" first request, scheduling timeout\n");
 814                 __schedule_osd_timeout(osdc);
 815         }
 816 }
 817
 818 static void register_request(struct ceph_osd_client *osdc,
 819                              struct ceph_osd_request *req)
 820 {
 821         mutex_lock(&osdc->request_mutex);
 822         __register_request(osdc, req);
 823         mutex_unlock(&osdc->request_mutex);
 824 }
 825
 826 /*
 827  * called under osdc->request_mutex
 828  */
 829 static void __unregister_request(struct ceph_osd_client *osdc,
 830                                  struct ceph_osd_request *req)
 831 {
 832         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 833         rb_erase(&req->r_node, &osdc->requests);
 834         osdc->num_requests--;
 835
 836         if (req->r_osd) {
 837                 /* make sure the original request isn't in flight. */
 838                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 839
 840                 list_del_init(&req->r_osd_item);
 841                 if (list_empty(&req->r_osd->o_requests) &&
 842                     list_empty(&req->r_osd->o_linger_requests)) {
 843                         dout("moving osd to %p lru\n", req->r_osd);
 844                         __move_osd_to_lru(osdc, req->r_osd);
 845                 }
 846                 if (list_empty(&req->r_linger_item))
 847                         req->r_osd = NULL;
 848         }
 849
 850         ceph_osdc_put_request(req);
 851
 852         list_del_init(&req->r_req_lru_item);
 853         if (osdc->num_requests == 0) {
 854                 dout(" no requests, canceling timeout\n");
 855                 __cancel_osd_timeout(osdc);
 856         }
 857 }
 858
 859 /*
 860  * Cancel a previously queued request message
 861  */
 862 static void __cancel_request(struct ceph_osd_request *req)
 863 {
 864         if (req->r_sent && req->r_osd) {
 865                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 866                 req->r_sent = 0;
 867         }
 868 }
 869
 870 static void __register_linger_request(struct ceph_osd_client *osdc,
 871                                     struct ceph_osd_request *req)
 872 {
 873         dout("__register_linger_request %p\n", req);
 874         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 875         list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests);
 876 }
 877
 878 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 879                                         struct ceph_osd_request *req)
 880 {
 881         dout("__unregister_linger_request %p\n", req);
 882         if (req->r_osd) {
 883                 list_del_init(&req->r_linger_item);
 884                 list_del_init(&req->r_linger_osd);
 885
 886                 if (list_empty(&req->r_osd->o_requests) &&
 887                     list_empty(&req->r_osd->o_linger_requests)) {
 888                         dout("moving osd to %p lru\n", req->r_osd);
 889                         __move_osd_to_lru(osdc, req->r_osd);
 890                 }
 891                 if (list_empty(&req->r_osd_item))
 892                         req->r_osd = NULL;
 893         }
 894 }
 895
 896 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 897                                          struct ceph_osd_request *req)
 898 {
 899         mutex_lock(&osdc->request_mutex);
 900         if (req->r_linger) {
 901                 __unregister_linger_request(osdc, req);
 902                 ceph_osdc_put_request(req);
 903         }
 904         mutex_unlock(&osdc->request_mutex);
 905 }
 906 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 907
 908 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 909                                   struct ceph_osd_request *req)
 910 {
 911         if (!req->r_linger) {
 912                 dout("set_request_linger %p\n", req);
 913                 req->r_linger = 1;
 914                 /*
 915                  * caller is now responsible for calling
 916                  * unregister_linger_request
 917                  */
 918                 ceph_osdc_get_request(req);
 919         }
 920 }
 921 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 922
 923 /*
 924  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 925  * (as needed), and set the request r_osd appropriately.  If there is
 926  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 927  * (unsent, homeless) or leave on in-flight lru.
 928  *
 929  * Return 0 if unchanged, 1 if changed, or negative on error.
 930  *
 931  * Caller should hold map_sem for read and request_mutex.
 932  */
 933 static int __map_request(struct ceph_osd_client *osdc,
 934                          struct ceph_osd_request *req)
 935 {
 936         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 937         struct ceph_pg pgid;
 938         int acting[CEPH_PG_MAX_SIZE];
 939         int o = -1, num = 0;
 940         int err;
 941
 942         dout("map_request %p tid %lld\n", req, req->r_tid);
 943         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 944                                       &req->r_file_layout, osdc->osdmap);
 945         if (err) {
 946                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 947                 return err;
 948         }
 949         pgid = reqhead->layout.ol_pgid;
 950         req->r_pgid = pgid;
 951
 952         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 953         if (err > 0) {
 954                 o = acting[0];
 955                 num = err;
 956         }
 957
 958         if ((req->r_osd && req->r_osd->o_osd == o &&
 959              req->r_sent >= req->r_osd->o_incarnation &&
 960              req->r_num_pg_osds == num &&
 961              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 962             (req->r_osd == NULL && o == -1))
 963                 return 0;  /* no change */
 964
 965         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 966              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 967              req->r_osd ? req->r_osd->o_osd : -1);
 968
 969         /* record full pg acting set */
 970         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 971         req->r_num_pg_osds = num;
 972
 973         if (req->r_osd) {
 974                 __cancel_request(req);
 975                 list_del_init(&req->r_osd_item);
 976                 req->r_osd = NULL;
 977         }
 978
 979         req->r_osd = __lookup_osd(osdc, o);
 980         if (!req->r_osd && o >= 0) {
 981                 err = -ENOMEM;
 982                 req->r_osd = create_osd(osdc);
 983                 if (!req->r_osd) {
 984                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
 985                         goto out;
 986                 }
 987
 988                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
 989                 req->r_osd->o_osd = o;
 990                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
 991                 __insert_osd(osdc, req->r_osd);
 992
 993                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
 994         }
 995
 996         if (req->r_osd) {
 997                 __remove_osd_from_lru(req->r_osd);
 998                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 999                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1000         } else {
1001                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1002         }
1003         err = 1;   /* osd or pg changed */
1004
1005 out:
1006         return err;
1007 }
1008
1009 /*
1010  * caller should hold map_sem (for read) and request_mutex
1011  */
1012 static int __send_request(struct ceph_osd_client *osdc,
1013                           struct ceph_osd_request *req)
1014 {
1015         struct ceph_osd_request_head *reqhead;
1016
1017         dout("send_request %p tid %llu to osd%d flags %d\n",
1018              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1019
1020         reqhead = req->r_request->front.iov_base;
1021         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1022         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1023         reqhead->reassert_version = req->r_reassert_version;
1024
1025         req->r_stamp = jiffies;
1026         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1027
1028         ceph_msg_get(req->r_request); /* send consumes a ref */
1029         ceph_con_send(&req->r_osd->o_con, req->r_request);
1030         req->r_sent = req->r_osd->o_incarnation;
1031         return 0;
1032 }
1033
1034 /*
1035  * Send any requests in the queue (req_unsent).
1036  */
1037 static void send_queued(struct ceph_osd_client *osdc)
1038 {
1039         struct ceph_osd_request *req, *tmp;
1040
1041         dout("send_queued\n");
1042         mutex_lock(&osdc->request_mutex);
1043         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1044                 __send_request(osdc, req);
1045         }
1046         mutex_unlock(&osdc->request_mutex);
1047 }
1048
1049 /*
1050  * Timeout callback, called every N seconds when 1 or more osd
1051  * requests has been active for more than N seconds.  When this
1052  * happens, we ping all OSDs with requests who have timed out to
1053  * ensure any communications channel reset is detected.  Reset the
1054  * request timeouts another N seconds in the future as we go.
1055  * Reschedule the timeout event another N seconds in future (unless
1056  * there are no open requests).
1057  */
1058 static void handle_timeout(struct work_struct *work)
1059 {
1060         struct ceph_osd_client *osdc =
1061                 container_of(work, struct ceph_osd_client, timeout_work.work);
1062         struct ceph_osd_request *req, *last_req = NULL;
1063         struct ceph_osd *osd;
1064         unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1065         unsigned long keepalive =
1066                 osdc->client->options->osd_keepalive_timeout * HZ;
1067         unsigned long last_stamp = 0;
1068         struct list_head slow_osds;
1069         dout("timeout\n");
1070         down_read(&osdc->map_sem);
1071
1072         ceph_monc_request_next_osdmap(&osdc->client->monc);
1073
1074         mutex_lock(&osdc->request_mutex);
1075
1076         /*
1077          * reset osds that appear to be _really_ unresponsive.  this
1078          * is a failsafe measure.. we really shouldn't be getting to
1079          * this point if the system is working properly.  the monitors
1080          * should mark the osd as failed and we should find out about
1081          * it from an updated osd map.
1082          */
1083         while (timeout && !list_empty(&osdc->req_lru)) {
1084                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1085                                  r_req_lru_item);
1086
1087                 if (time_before(jiffies, req->r_stamp + timeout))
1088                         break;
1089
1090                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
1091                 last_req = req;
1092                 last_stamp = req->r_stamp;
1093
1094                 osd = req->r_osd;
1095                 BUG_ON(!osd);
1096                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1097                            req->r_tid, osd->o_osd);
1098                 __kick_osd_requests(osdc, osd);
1099         }
1100
1101         /*
1102          * ping osds that are a bit slow.  this ensures that if there
1103          * is a break in the TCP connection we will notice, and reopen
1104          * a connection with that osd (from the fault callback).
1105          */
1106         INIT_LIST_HEAD(&slow_osds);
1107         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1108                 if (time_before(jiffies, req->r_stamp + keepalive))
1109                         break;
1110
1111                 osd = req->r_osd;
1112                 BUG_ON(!osd);
1113                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1114                      req->r_tid, osd->o_osd);
1115                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1116         }
1117         while (!list_empty(&slow_osds)) {
1118                 osd = list_entry(slow_osds.next, struct ceph_osd,
1119                                  o_keepalive_item);
1120                 list_del_init(&osd->o_keepalive_item);
1121                 ceph_con_keepalive(&osd->o_con);
1122         }
1123
1124         __schedule_osd_timeout(osdc);
1125         mutex_unlock(&osdc->request_mutex);
1126         send_queued(osdc);
1127         up_read(&osdc->map_sem);
1128 }
1129
1130 static void handle_osds_timeout(struct work_struct *work)
1131 {
1132         struct ceph_osd_client *osdc =
1133                 container_of(work, struct ceph_osd_client,
1134                              osds_timeout_work.work);
1135         unsigned long delay =
1136                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1137
1138         dout("osds timeout\n");
1139         down_read(&osdc->map_sem);
1140         remove_old_osds(osdc, 0);
1141         up_read(&osdc->map_sem);
1142
1143         schedule_delayed_work(&osdc->osds_timeout_work,
1144                               round_jiffies_relative(delay));
1145 }
1146
1147 /*
1148  * handle osd op reply.  either call the callback if it is specified,
1149  * or do the completion to wake up the waiting thread.
1150  */
1151 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1152                          struct ceph_connection *con)
1153 {
1154         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1155         struct ceph_osd_request *req;
1156         u64 tid;
1157         int numops, object_len, flags;
1158         s32 result;
1159
1160         tid = le64_to_cpu(msg->hdr.tid);
1161         if (msg->front.iov_len < sizeof(*rhead))
1162                 goto bad;
1163         numops = le32_to_cpu(rhead->num_ops);
1164         object_len = le32_to_cpu(rhead->object_len);
1165         result = le32_to_cpu(rhead->result);
1166         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1167             numops * sizeof(struct ceph_osd_op))
1168                 goto bad;
1169         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1170         /* lookup */
1171         mutex_lock(&osdc->request_mutex);
1172         req = __lookup_request(osdc, tid);
1173         if (req == NULL) {
1174                 dout("handle_reply tid %llu dne\n", tid);
1175                 mutex_unlock(&osdc->request_mutex);
1176                 return;
1177         }
1178         ceph_osdc_get_request(req);
1179         flags = le32_to_cpu(rhead->flags);
1180
1181         /*
1182          * if this connection filled our message, drop our reference now, to
1183          * avoid a (safe but slower) revoke later.
1184          */
1185         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1186                 dout(" dropping con_filling_msg ref %p\n", con);
1187                 req->r_con_filling_msg = NULL;
1188                 ceph_con_put(con);
1189         }
1190
1191         if (!req->r_got_reply) {
1192                 unsigned bytes;
1193
1194                 req->r_result = le32_to_cpu(rhead->result);
1195                 bytes = le32_to_cpu(msg->hdr.data_len);
1196                 dout("handle_reply result %d bytes %d\n", req->r_result,
1197                      bytes);
1198                 if (req->r_result == 0)
1199                         req->r_result = bytes;
1200
1201                 /* in case this is a write and we need to replay, */
1202                 req->r_reassert_version = rhead->reassert_version;
1203
1204                 req->r_got_reply = 1;
1205         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1206                 dout("handle_reply tid %llu dup ack\n", tid);
1207                 mutex_unlock(&osdc->request_mutex);
1208                 goto done;
1209         }
1210
1211         dout("handle_reply tid %llu flags %d\n", tid, flags);
1212
1213         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1214                 __register_linger_request(osdc, req);
1215
1216         /* either this is a read, or we got the safe response */
1217         if (result < 0 ||
1218             (flags & CEPH_OSD_FLAG_ONDISK) ||
1219             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1220                 __unregister_request(osdc, req);
1221
1222         mutex_unlock(&osdc->request_mutex);
1223
1224         if (req->r_callback)
1225                 req->r_callback(req, msg);
1226         else
1227                 complete_all(&req->r_completion);
1228
1229         if (flags & CEPH_OSD_FLAG_ONDISK) {
1230                 if (req->r_safe_callback)
1231                         req->r_safe_callback(req, msg);
1232                 complete_all(&req->r_safe_completion);  /* fsync waiter */
1233         }
1234
1235 done:
1236         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1237         ceph_osdc_put_request(req);
1238         return;
1239
1240 bad:
1241         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1242                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1243                (int)sizeof(*rhead));
1244         ceph_msg_dump(msg);
1245 }
1246
1247 static void reset_changed_osds(struct ceph_osd_client *osdc)
1248 {
1249         struct rb_node *p, *n;
1250
1251         for (p = rb_first(&osdc->osds); p; p = n) {
1252                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1253
1254                 n = rb_next(p);
1255                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1256                     memcmp(&osd->o_con.peer_addr,
1257                            ceph_osd_addr(osdc->osdmap,
1258                                          osd->o_osd),
1259                            sizeof(struct ceph_entity_addr)) != 0)
1260                         __reset_osd(osdc, osd);
1261         }
1262 }
1263
1264 /*
1265  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1266  * no osd, request a new map.
1267  *
1268  * Caller should hold map_sem for read and request_mutex.
1269  */
1270 static void kick_requests(struct ceph_osd_client *osdc)
1271 {
1272         struct ceph_osd_request *req, *nreq;
1273         struct rb_node *p;
1274         int needmap = 0;
1275         int err;
1276
1277         dout("kick_requests\n");
1278         mutex_lock(&osdc->request_mutex);
1279         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1280                 req = rb_entry(p, struct ceph_osd_request, r_node);
1281                 err = __map_request(osdc, req);
1282                 if (err < 0)
1283                         continue;  /* error */
1284                 if (req->r_osd == NULL) {
1285                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1286                         needmap++;  /* request a newer map */
1287                 } else if (err > 0) {
1288                         dout("%p tid %llu requeued on osd%d\n", req, req->r_tid,
1289                              req->r_osd ? req->r_osd->o_osd : -1);
1290                         if (!req->r_linger)
1291                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1292                 }
1293         }
1294
1295         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1296                                  r_linger_item) {
1297                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1298
1299                 err = __map_request(osdc, req);
1300                 if (err == 0)
1301                         continue;  /* no change and no osd was specified */
1302                 if (err < 0)
1303                         continue;  /* hrm! */
1304                 if (req->r_osd == NULL) {
1305                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1306                         needmap++;  /* request a newer map */
1307                         continue;
1308                 }
1309
1310                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1311                      req->r_osd ? req->r_osd->o_osd : -1);
1312                 __unregister_linger_request(osdc, req);
1313                 __register_request(osdc, req);
1314         }
1315         mutex_unlock(&osdc->request_mutex);
1316
1317         if (needmap) {
1318                 dout("%d requests for down osds, need new map\n", needmap);
1319                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1320         }
1321 }
1322
1323
1324 /*
1325  * Process updated osd map.
1326  *
1327  * The message contains any number of incremental and full maps, normally
1328  * indicating some sort of topology change in the cluster.  Kick requests
1329  * off to different OSDs as needed.
1330  */
1331 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1332 {
1333         void *p, *end, *next;
1334         u32 nr_maps, maplen;
1335         u32 epoch;
1336         struct ceph_osdmap *newmap = NULL, *oldmap;
1337         int err;
1338         struct ceph_fsid fsid;
1339
1340         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1341         p = msg->front.iov_base;
1342         end = p + msg->front.iov_len;
1343
1344         /* verify fsid */
1345         ceph_decode_need(&p, end, sizeof(fsid), bad);
1346         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1347         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1348                 return;
1349
1350         down_write(&osdc->map_sem);
1351
1352         /* incremental maps */
1353         ceph_decode_32_safe(&p, end, nr_maps, bad);
1354         dout(" %d inc maps\n", nr_maps);
1355         while (nr_maps > 0) {
1356                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1357                 epoch = ceph_decode_32(&p);
1358                 maplen = ceph_decode_32(&p);
1359                 ceph_decode_need(&p, end, maplen, bad);
1360                 next = p + maplen;
1361                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1362                         dout("applying incremental map %u len %d\n",
1363                              epoch, maplen);
1364                         newmap = osdmap_apply_incremental(&p, next,
1365                                                           osdc->osdmap,
1366                                                           osdc->client->msgr);
1367                         if (IS_ERR(newmap)) {
1368                                 err = PTR_ERR(newmap);
1369                                 goto bad;
1370                         }
1371                         BUG_ON(!newmap);
1372                         if (newmap != osdc->osdmap) {
1373                                 ceph_osdmap_destroy(osdc->osdmap);
1374                                 osdc->osdmap = newmap;
1375                         }
1376                         kick_requests(osdc);
1377                         reset_changed_osds(osdc);
1378                 } else {
1379                         dout("ignoring incremental map %u len %d\n",
1380                              epoch, maplen);
1381                 }
1382                 p = next;
1383                 nr_maps--;
1384         }
1385         if (newmap)
1386                 goto done;
1387
1388         /* full maps */
1389         ceph_decode_32_safe(&p, end, nr_maps, bad);
1390         dout(" %d full maps\n", nr_maps);
1391         while (nr_maps) {
1392                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1393                 epoch = ceph_decode_32(&p);
1394                 maplen = ceph_decode_32(&p);
1395                 ceph_decode_need(&p, end, maplen, bad);
1396                 if (nr_maps > 1) {
1397                         dout("skipping non-latest full map %u len %d\n",
1398                              epoch, maplen);
1399                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1400                         dout("skipping full map %u len %d, "
1401                              "older than our %u\n", epoch, maplen,
1402                              osdc->osdmap->epoch);
1403                 } else {
1404                         dout("taking full map %u len %d\n", epoch, maplen);
1405                         newmap = osdmap_decode(&p, p+maplen);
1406                         if (IS_ERR(newmap)) {
1407                                 err = PTR_ERR(newmap);
1408                                 goto bad;
1409                         }
1410                         BUG_ON(!newmap);
1411                         oldmap = osdc->osdmap;
1412                         osdc->osdmap = newmap;
1413                         if (oldmap)
1414                                 ceph_osdmap_destroy(oldmap);
1415                         kick_requests(osdc);
1416                 }
1417                 p += maplen;
1418                 nr_maps--;
1419         }
1420
1421 done:
1422         downgrade_write(&osdc->map_sem);
1423         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1424         send_queued(osdc);
1425         up_read(&osdc->map_sem);
1426         wake_up_all(&osdc->client->auth_wq);
1427         return;
1428
1429 bad:
1430         pr_err("osdc handle_map corrupt msg\n");
1431         ceph_msg_dump(msg);
1432         up_write(&osdc->map_sem);
1433         return;
1434 }
1435
1436 /*
1437  * watch/notify callback event infrastructure
1438  *
1439  * These callbacks are used both for watch and notify operations.
1440  */
1441 static void __release_event(struct kref *kref)
1442 {
1443         struct ceph_osd_event *event =
1444                 container_of(kref, struct ceph_osd_event, kref);
1445
1446         dout("__release_event %p\n", event);
1447         kfree(event);
1448 }
1449
1450 static void get_event(struct ceph_osd_event *event)
1451 {
1452         kref_get(&event->kref);
1453 }
1454
1455 void ceph_osdc_put_event(struct ceph_osd_event *event)
1456 {
1457         kref_put(&event->kref, __release_event);
1458 }
1459 EXPORT_SYMBOL(ceph_osdc_put_event);
1460
1461 static void __insert_event(struct ceph_osd_client *osdc,
1462                              struct ceph_osd_event *new)
1463 {
1464         struct rb_node **p = &osdc->event_tree.rb_node;
1465         struct rb_node *parent = NULL;
1466         struct ceph_osd_event *event = NULL;
1467
1468         while (*p) {
1469                 parent = *p;
1470                 event = rb_entry(parent, struct ceph_osd_event, node);
1471                 if (new->cookie < event->cookie)
1472                         p = &(*p)->rb_left;
1473                 else if (new->cookie > event->cookie)
1474                         p = &(*p)->rb_right;
1475                 else
1476                         BUG();
1477         }
1478
1479         rb_link_node(&new->node, parent, p);
1480         rb_insert_color(&new->node, &osdc->event_tree);
1481 }
1482
1483 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1484                                                 u64 cookie)
1485 {
1486         struct rb_node **p = &osdc->event_tree.rb_node;
1487         struct rb_node *parent = NULL;
1488         struct ceph_osd_event *event = NULL;
1489
1490         while (*p) {
1491                 parent = *p;
1492                 event = rb_entry(parent, struct ceph_osd_event, node);
1493                 if (cookie < event->cookie)
1494                         p = &(*p)->rb_left;
1495                 else if (cookie > event->cookie)
1496                         p = &(*p)->rb_right;
1497                 else
1498                         return event;
1499         }
1500         return NULL;
1501 }
1502
1503 static void __remove_event(struct ceph_osd_event *event)
1504 {
1505         struct ceph_osd_client *osdc = event->osdc;
1506
1507         if (!RB_EMPTY_NODE(&event->node)) {
1508                 dout("__remove_event removed %p\n", event);
1509                 rb_erase(&event->node, &osdc->event_tree);
1510                 ceph_osdc_put_event(event);
1511         } else {
1512                 dout("__remove_event didn't remove %p\n", event);
1513         }
1514 }
1515
1516 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1517                            void (*event_cb)(u64, u64, u8, void *),
1518                            int one_shot, void *data,
1519                            struct ceph_osd_event **pevent)
1520 {
1521         struct ceph_osd_event *event;
1522
1523         event = kmalloc(sizeof(*event), GFP_NOIO);
1524         if (!event)
1525                 return -ENOMEM;
1526
1527         dout("create_event %p\n", event);
1528         event->cb = event_cb;
1529         event->one_shot = one_shot;
1530         event->data = data;
1531         event->osdc = osdc;
1532         INIT_LIST_HEAD(&event->osd_node);
1533         kref_init(&event->kref);   /* one ref for us */
1534         kref_get(&event->kref);    /* one ref for the caller */
1535         init_completion(&event->completion);
1536
1537         spin_lock(&osdc->event_lock);
1538         event->cookie = ++osdc->event_count;
1539         __insert_event(osdc, event);
1540         spin_unlock(&osdc->event_lock);
1541
1542         *pevent = event;
1543         return 0;
1544 }
1545 EXPORT_SYMBOL(ceph_osdc_create_event);
1546
1547 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1548 {
1549         struct ceph_osd_client *osdc = event->osdc;
1550
1551         dout("cancel_event %p\n", event);
1552         spin_lock(&osdc->event_lock);
1553         __remove_event(event);
1554         spin_unlock(&osdc->event_lock);
1555         ceph_osdc_put_event(event); /* caller's */
1556 }
1557 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1558
1559
1560 static void do_event_work(struct work_struct *work)
1561 {
1562         struct ceph_osd_event_work *event_work =
1563                 container_of(work, struct ceph_osd_event_work, work);
1564         struct ceph_osd_event *event = event_work->event;
1565         u64 ver = event_work->ver;
1566         u64 notify_id = event_work->notify_id;
1567         u8 opcode = event_work->opcode;
1568
1569         dout("do_event_work completing %p\n", event);
1570         event->cb(ver, notify_id, opcode, event->data);
1571         complete(&event->completion);
1572         dout("do_event_work completed %p\n", event);
1573         ceph_osdc_put_event(event);
1574         kfree(event_work);
1575 }
1576
1577
1578 /*
1579  * Process osd watch notifications
1580  */
1581 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1582 {
1583         void *p, *end;
1584         u8 proto_ver;
1585         u64 cookie, ver, notify_id;
1586         u8 opcode;
1587         struct ceph_osd_event *event;
1588         struct ceph_osd_event_work *event_work;
1589
1590         p = msg->front.iov_base;
1591         end = p + msg->front.iov_len;
1592
1593         ceph_decode_8_safe(&p, end, proto_ver, bad);
1594         ceph_decode_8_safe(&p, end, opcode, bad);
1595         ceph_decode_64_safe(&p, end, cookie, bad);
1596         ceph_decode_64_safe(&p, end, ver, bad);
1597         ceph_decode_64_safe(&p, end, notify_id, bad);
1598
1599         spin_lock(&osdc->event_lock);
1600         event = __find_event(osdc, cookie);
1601         if (event) {
1602                 get_event(event);
1603                 if (event->one_shot)
1604                         __remove_event(event);
1605         }
1606         spin_unlock(&osdc->event_lock);
1607         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1608              cookie, ver, event);
1609         if (event) {
1610                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1611                 if (!event_work) {
1612                         dout("ERROR: could not allocate event_work\n");
1613                         goto done_err;
1614                 }
1615                 INIT_WORK(&event_work->work, do_event_work);
1616                 event_work->event = event;
1617                 event_work->ver = ver;
1618                 event_work->notify_id = notify_id;
1619                 event_work->opcode = opcode;
1620                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1621                         dout("WARNING: failed to queue notify event work\n");
1622                         goto done_err;
1623                 }
1624         }
1625
1626         return;
1627
1628 done_err:
1629         complete(&event->completion);
1630         ceph_osdc_put_event(event);
1631         return;
1632
1633 bad:
1634         pr_err("osdc handle_watch_notify corrupt msg\n");
1635         return;
1636 }
1637
1638 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1639 {
1640         int err;
1641
1642         dout("wait_event %p\n", event);
1643         err = wait_for_completion_interruptible_timeout(&event->completion,
1644                                                         timeout * HZ);
1645         ceph_osdc_put_event(event);
1646         if (err > 0)
1647                 err = 0;
1648         dout("wait_event %p returns %d\n", event, err);
1649         return err;
1650 }
1651 EXPORT_SYMBOL(ceph_osdc_wait_event);
1652
1653 /*
1654  * Register request, send initial attempt.
1655  */
1656 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1657                             struct ceph_osd_request *req,
1658                             bool nofail)
1659 {
1660         int rc = 0;
1661
1662         req->r_request->pages = req->r_pages;
1663         req->r_request->nr_pages = req->r_num_pages;
1664 #ifdef CONFIG_BLOCK
1665         req->r_request->bio = req->r_bio;
1666 #endif
1667         req->r_request->trail = req->r_trail;
1668
1669         register_request(osdc, req);
1670
1671         down_read(&osdc->map_sem);
1672         mutex_lock(&osdc->request_mutex);
1673         /*
1674          * a racing kick_requests() may have sent the message for us
1675          * while we dropped request_mutex above, so only send now if
1676          * the request still han't been touched yet.
1677          */
1678         if (req->r_sent == 0) {
1679                 rc = __map_request(osdc, req);
1680                 if (rc < 0)
1681                         goto out_unlock;
1682                 if (req->r_osd == NULL) {
1683                         dout("send_request %p no up osds in pg\n", req);
1684                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1685                 } else {
1686                         rc = __send_request(osdc, req);
1687                         if (rc) {
1688                                 if (nofail) {
1689                                         dout("osdc_start_request failed send, "
1690                                              " will retry %lld\n", req->r_tid);
1691                                         rc = 0;
1692                                 } else {
1693                                         __unregister_request(osdc, req);
1694                                 }
1695                         }
1696                 }
1697         }
1698
1699 out_unlock:
1700         mutex_unlock(&osdc->request_mutex);
1701         up_read(&osdc->map_sem);
1702         return rc;
1703 }
1704 EXPORT_SYMBOL(ceph_osdc_start_request);
1705
1706 /*
1707  * wait for a request to complete
1708  */
1709 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1710                            struct ceph_osd_request *req)
1711 {
1712         int rc;
1713
1714         rc = wait_for_completion_interruptible(&req->r_completion);
1715         if (rc < 0) {
1716                 mutex_lock(&osdc->request_mutex);
1717                 __cancel_request(req);
1718                 __unregister_request(osdc, req);
1719                 mutex_unlock(&osdc->request_mutex);
1720                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1721                 return rc;
1722         }
1723
1724         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1725         return req->r_result;
1726 }
1727 EXPORT_SYMBOL(ceph_osdc_wait_request);
1728
1729 /*
1730  * sync - wait for all in-flight requests to flush.  avoid starvation.
1731  */
1732 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1733 {
1734         struct ceph_osd_request *req;
1735         u64 last_tid, next_tid = 0;
1736
1737         mutex_lock(&osdc->request_mutex);
1738         last_tid = osdc->last_tid;
1739         while (1) {
1740                 req = __lookup_request_ge(osdc, next_tid);
1741                 if (!req)
1742                         break;
1743                 if (req->r_tid > last_tid)
1744                         break;
1745
1746                 next_tid = req->r_tid + 1;
1747                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1748                         continue;
1749
1750                 ceph_osdc_get_request(req);
1751                 mutex_unlock(&osdc->request_mutex);
1752                 dout("sync waiting on tid %llu (last is %llu)\n",
1753                      req->r_tid, last_tid);
1754                 wait_for_completion(&req->r_safe_completion);
1755                 mutex_lock(&osdc->request_mutex);
1756                 ceph_osdc_put_request(req);
1757         }
1758         mutex_unlock(&osdc->request_mutex);
1759         dout("sync done (thru tid %llu)\n", last_tid);
1760 }
1761 EXPORT_SYMBOL(ceph_osdc_sync);
1762
1763 /*
1764  * init, shutdown
1765  */
1766 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1767 {
1768         int err;
1769
1770         dout("init\n");
1771         osdc->client = client;
1772         osdc->osdmap = NULL;
1773         init_rwsem(&osdc->map_sem);
1774         init_completion(&osdc->map_waiters);
1775         osdc->last_requested_map = 0;
1776         mutex_init(&osdc->request_mutex);
1777         osdc->last_tid = 0;
1778         osdc->osds = RB_ROOT;
1779         INIT_LIST_HEAD(&osdc->osd_lru);
1780         osdc->requests = RB_ROOT;
1781         INIT_LIST_HEAD(&osdc->req_lru);
1782         INIT_LIST_HEAD(&osdc->req_unsent);
1783         INIT_LIST_HEAD(&osdc->req_notarget);
1784         INIT_LIST_HEAD(&osdc->req_linger);
1785         osdc->num_requests = 0;
1786         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1787         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1788         spin_lock_init(&osdc->event_lock);
1789         osdc->event_tree = RB_ROOT;
1790         osdc->event_count = 0;
1791
1792         schedule_delayed_work(&osdc->osds_timeout_work,
1793            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1794
1795         err = -ENOMEM;
1796         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1797                                         sizeof(struct ceph_osd_request));
1798         if (!osdc->req_mempool)
1799                 goto out;
1800
1801         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1802                                 "osd_op");
1803         if (err < 0)
1804                 goto out_mempool;
1805         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1806                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1807                                 "osd_op_reply");
1808         if (err < 0)
1809                 goto out_msgpool;
1810
1811         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1812         if (IS_ERR(osdc->notify_wq)) {
1813                 err = PTR_ERR(osdc->notify_wq);
1814                 osdc->notify_wq = NULL;
1815                 goto out_msgpool;
1816         }
1817         return 0;
1818
1819 out_msgpool:
1820         ceph_msgpool_destroy(&osdc->msgpool_op);
1821 out_mempool:
1822         mempool_destroy(osdc->req_mempool);
1823 out:
1824         return err;
1825 }
1826 EXPORT_SYMBOL(ceph_osdc_init);
1827
1828 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1829 {
1830         flush_workqueue(osdc->notify_wq);
1831         destroy_workqueue(osdc->notify_wq);
1832         cancel_delayed_work_sync(&osdc->timeout_work);
1833         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1834         if (osdc->osdmap) {
1835                 ceph_osdmap_destroy(osdc->osdmap);
1836                 osdc->osdmap = NULL;
1837         }
1838         remove_old_osds(osdc, 1);
1839         WARN_ON(!RB_EMPTY_ROOT(&osdc->osds));
1840         mempool_destroy(osdc->req_mempool);
1841         ceph_msgpool_destroy(&osdc->msgpool_op);
1842         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1843 }
1844 EXPORT_SYMBOL(ceph_osdc_stop);
1845
1846 /*
1847  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1848  * *plen.  Return number of bytes read, or error.
1849  */
1850 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1851                         struct ceph_vino vino, struct ceph_file_layout *layout,
1852                         u64 off, u64 *plen,
1853                         u32 truncate_seq, u64 truncate_size,
1854                         struct page **pages, int num_pages, int page_align)
1855 {
1856         struct ceph_osd_request *req;
1857         int rc = 0;
1858
1859         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1860              vino.snap, off, *plen);
1861         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1862                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1863                                     NULL, 0, truncate_seq, truncate_size, NULL,
1864                                     false, 1, page_align);
1865         if (!req)
1866                 return -ENOMEM;
1867
1868         /* it may be a short read due to an object boundary */
1869         req->r_pages = pages;
1870
1871         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1872              off, *plen, req->r_num_pages, page_align);
1873
1874         rc = ceph_osdc_start_request(osdc, req, false);
1875         if (!rc)
1876                 rc = ceph_osdc_wait_request(osdc, req);
1877
1878         ceph_osdc_put_request(req);
1879         dout("readpages result %d\n", rc);
1880         return rc;
1881 }
1882 EXPORT_SYMBOL(ceph_osdc_readpages);
1883
1884 /*
1885  * do a synchronous write on N pages
1886  */
1887 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1888                          struct ceph_file_layout *layout,
1889                          struct ceph_snap_context *snapc,
1890                          u64 off, u64 len,
1891                          u32 truncate_seq, u64 truncate_size,
1892                          struct timespec *mtime,
1893                          struct page **pages, int num_pages,
1894                          int flags, int do_sync, bool nofail)
1895 {
1896         struct ceph_osd_request *req;
1897         int rc = 0;
1898         int page_align = off & ~PAGE_MASK;
1899
1900         BUG_ON(vino.snap != CEPH_NOSNAP);
1901         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1902                                     CEPH_OSD_OP_WRITE,
1903                                     flags | CEPH_OSD_FLAG_ONDISK |
1904                                             CEPH_OSD_FLAG_WRITE,
1905                                     snapc, do_sync,
1906                                     truncate_seq, truncate_size, mtime,
1907                                     nofail, 1, page_align);
1908         if (!req)
1909                 return -ENOMEM;
1910
1911         /* it may be a short write due to an object boundary */
1912         req->r_pages = pages;
1913         dout("writepages %llu~%llu (%d pages)\n", off, len,
1914              req->r_num_pages);
1915
1916         rc = ceph_osdc_start_request(osdc, req, nofail);
1917         if (!rc)
1918                 rc = ceph_osdc_wait_request(osdc, req);
1919
1920         ceph_osdc_put_request(req);
1921         if (rc == 0)
1922                 rc = len;
1923         dout("writepages result %d\n", rc);
1924         return rc;
1925 }
1926 EXPORT_SYMBOL(ceph_osdc_writepages);
1927
1928 /*
1929  * handle incoming message
1930  */
1931 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1932 {
1933         struct ceph_osd *osd = con->private;
1934         struct ceph_osd_client *osdc;
1935         int type = le16_to_cpu(msg->hdr.type);
1936
1937         if (!osd)
1938                 goto out;
1939         osdc = osd->o_osdc;
1940
1941         switch (type) {
1942         case CEPH_MSG_OSD_MAP:
1943                 ceph_osdc_handle_map(osdc, msg);
1944                 break;
1945         case CEPH_MSG_OSD_OPREPLY:
1946                 handle_reply(osdc, msg, con);
1947                 break;
1948         case CEPH_MSG_WATCH_NOTIFY:
1949                 handle_watch_notify(osdc, msg);
1950                 break;
1951
1952         default:
1953                 pr_err("received unknown message type %d %s\n", type,
1954                        ceph_msg_type_name(type));
1955         }
1956 out:
1957         ceph_msg_put(msg);
1958 }
1959
1960 /*
1961  * lookup and return message for incoming reply.  set up reply message
1962  * pages.
1963  */
1964 static struct ceph_msg *get_reply(struct ceph_connection *con,
1965                                   struct ceph_msg_header *hdr,
1966                                   int *skip)
1967 {
1968         struct ceph_osd *osd = con->private;
1969         struct ceph_osd_client *osdc = osd->o_osdc;
1970         struct ceph_msg *m;
1971         struct ceph_osd_request *req;
1972         int front = le32_to_cpu(hdr->front_len);
1973         int data_len = le32_to_cpu(hdr->data_len);
1974         u64 tid;
1975
1976         tid = le64_to_cpu(hdr->tid);
1977         mutex_lock(&osdc->request_mutex);
1978         req = __lookup_request(osdc, tid);
1979         if (!req) {
1980                 *skip = 1;
1981                 m = NULL;
1982                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
1983                         osd->o_osd);
1984                 goto out;
1985         }
1986
1987         if (req->r_con_filling_msg) {
1988                 dout("get_reply revoking msg %p from old con %p\n",
1989                      req->r_reply, req->r_con_filling_msg);
1990                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
1991                 ceph_con_put(req->r_con_filling_msg);
1992                 req->r_con_filling_msg = NULL;
1993         }
1994
1995         if (front > req->r_reply->front.iov_len) {
1996                 pr_warning("get_reply front %d > preallocated %d\n",
1997                            front, (int)req->r_reply->front.iov_len);
1998                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS);
1999                 if (!m)
2000                         goto out;
2001                 ceph_msg_put(req->r_reply);
2002                 req->r_reply = m;
2003         }
2004         m = ceph_msg_get(req->r_reply);
2005
2006         if (data_len > 0) {
2007                 int want = calc_pages_for(req->r_page_alignment, data_len);
2008
2009                 if (unlikely(req->r_num_pages < want)) {
2010                         pr_warning("tid %lld reply %d > expected %d pages\n",
2011                                    tid, want, m->nr_pages);
2012                         *skip = 1;
2013                         ceph_msg_put(m);
2014                         m = NULL;
2015                         goto out;
2016                 }
2017                 m->pages = req->r_pages;
2018                 m->nr_pages = req->r_num_pages;
2019                 m->page_alignment = req->r_page_alignment;
2020 #ifdef CONFIG_BLOCK
2021                 m->bio = req->r_bio;
2022 #endif
2023         }
2024         *skip = 0;
2025         req->r_con_filling_msg = ceph_con_get(con);
2026         dout("get_reply tid %lld %p\n", tid, m);
2027
2028 out:
2029         mutex_unlock(&osdc->request_mutex);
2030         return m;
2031
2032 }
2033
2034 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2035                                   struct ceph_msg_header *hdr,
2036                                   int *skip)
2037 {
2038         struct ceph_osd *osd = con->private;
2039         int type = le16_to_cpu(hdr->type);
2040         int front = le32_to_cpu(hdr->front_len);
2041
2042         switch (type) {
2043         case CEPH_MSG_OSD_MAP:
2044         case CEPH_MSG_WATCH_NOTIFY:
2045                 return ceph_msg_new(type, front, GFP_NOFS);
2046         case CEPH_MSG_OSD_OPREPLY:
2047                 return get_reply(con, hdr, skip);
2048         default:
2049                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2050                         osd->o_osd);
2051                 *skip = 1;
2052                 return NULL;
2053         }
2054 }
2055
2056 /*
2057  * Wrappers to refcount containing ceph_osd struct
2058  */
2059 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2060 {
2061         struct ceph_osd *osd = con->private;
2062         if (get_osd(osd))
2063                 return con;
2064         return NULL;
2065 }
2066
2067 static void put_osd_con(struct ceph_connection *con)
2068 {
2069         struct ceph_osd *osd = con->private;
2070         put_osd(osd);
2071 }
2072
2073 /*
2074  * authentication
2075  */
2076 static int get_authorizer(struct ceph_connection *con,
2077                           void **buf, int *len, int *proto,
2078                           void **reply_buf, int *reply_len, int force_new)
2079 {
2080         struct ceph_osd *o = con->private;
2081         struct ceph_osd_client *osdc = o->o_osdc;
2082         struct ceph_auth_client *ac = osdc->client->monc.auth;
2083         int ret = 0;
2084
2085         if (force_new && o->o_authorizer) {
2086                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
2087                 o->o_authorizer = NULL;
2088         }
2089         if (o->o_authorizer == NULL) {
2090                 ret = ac->ops->create_authorizer(
2091                         ac, CEPH_ENTITY_TYPE_OSD,
2092                         &o->o_authorizer,
2093                         &o->o_authorizer_buf,
2094                         &o->o_authorizer_buf_len,
2095                         &o->o_authorizer_reply_buf,
2096                         &o->o_authorizer_reply_buf_len);
2097                 if (ret)
2098                         return ret;
2099         }
2100
2101         *proto = ac->protocol;
2102         *buf = o->o_authorizer_buf;
2103         *len = o->o_authorizer_buf_len;
2104         *reply_buf = o->o_authorizer_reply_buf;
2105         *reply_len = o->o_authorizer_reply_buf_len;
2106         return 0;
2107 }
2108
2109
2110 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2111 {
2112         struct ceph_osd *o = con->private;
2113         struct ceph_osd_client *osdc = o->o_osdc;
2114         struct ceph_auth_client *ac = osdc->client->monc.auth;
2115
2116         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
2117 }
2118
2119 static int invalidate_authorizer(struct ceph_connection *con)
2120 {
2121         struct ceph_osd *o = con->private;
2122         struct ceph_osd_client *osdc = o->o_osdc;
2123         struct ceph_auth_client *ac = osdc->client->monc.auth;
2124
2125         if (ac->ops->invalidate_authorizer)
2126                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2127
2128         return ceph_monc_validate_auth(&osdc->client->monc);
2129 }
2130
2131 static const struct ceph_connection_operations osd_con_ops = {
2132         .get = get_osd_con,
2133         .put = put_osd_con,
2134         .dispatch = dispatch,
2135         .get_authorizer = get_authorizer,
2136         .verify_authorizer_reply = verify_authorizer_reply,
2137         .invalidate_authorizer = invalidate_authorizer,
2138         .alloc_msg = alloc_msg,
2139         .fault = osd_reset,
2140 };