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 void __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 int 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         int r;
  66
  67         reqhead->snapid = cpu_to_le64(snapid);
  68
  69         /* object extent? */
  70         r = ceph_calc_file_object_mapping(layout, off, plen, bno,
  71                                           &objoff, &objlen);
  72         if (r < 0)
  73                 return r;
  74         if (*plen < orig_len)
  75                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  76                      orig_len - *plen, off, *plen);
  77
  78         if (op_has_extent(op->op)) {
  79                 op->extent.offset = objoff;
  80                 op->extent.length = objlen;
  81         }
  82         req->r_num_pages = calc_pages_for(off, *plen);
  83         req->r_page_alignment = off & ~PAGE_MASK;
  84         if (op->op == CEPH_OSD_OP_WRITE)
  85                 op->payload_len = *plen;
  86
  87         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
  88              *bno, objoff, objlen, req->r_num_pages);
  89         return 0;
  90 }
  91 EXPORT_SYMBOL(ceph_calc_raw_layout);
  92
  93 /*
  94  * Implement client access to distributed object storage cluster.
  95  *
  96  * All data objects are stored within a cluster/cloud of OSDs, or
  97  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  98  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  99  * remote daemons serving up and coordinating consistent and safe
 100  * access to storage.
 101  *
 102  * Cluster membership and the mapping of data objects onto storage devices
 103  * are described by the osd map.
 104  *
 105  * We keep track of pending OSD requests (read, write), resubmit
 106  * requests to different OSDs when the cluster topology/data layout
 107  * change, or retry the affected requests when the communications
 108  * channel with an OSD is reset.
 109  */
 110
 111 /*
 112  * calculate the mapping of a file extent onto an object, and fill out the
 113  * request accordingly.  shorten extent as necessary if it crosses an
 114  * object boundary.
 115  *
 116  * fill osd op in request message.
 117  */
 118 static int calc_layout(struct ceph_osd_client *osdc,
 119                        struct ceph_vino vino,
 120                        struct ceph_file_layout *layout,
 121                        u64 off, u64 *plen,
 122                        struct ceph_osd_request *req,
 123                        struct ceph_osd_req_op *op)
 124 {
 125         u64 bno;
 126         int r;
 127
 128         r = ceph_calc_raw_layout(osdc, layout, vino.snap, off,
 129                                  plen, &bno, req, op);
 130         if (r < 0)
 131                 return r;
 132
 133         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 134         req->r_oid_len = strlen(req->r_oid);
 135
 136         return r;
 137 }
 138
 139 /*
 140  * requests
 141  */
 142 void ceph_osdc_release_request(struct kref *kref)
 143 {
 144         struct ceph_osd_request *req = container_of(kref,
 145                                                     struct ceph_osd_request,
 146                                                     r_kref);
 147
 148         if (req->r_request)
 149                 ceph_msg_put(req->r_request);
 150         if (req->r_con_filling_msg) {
 151                 dout("%s revoking pages %p from con %p\n", __func__,
 152                      req->r_pages, req->r_con_filling_msg);
 153                 ceph_msg_revoke_incoming(req->r_reply);
 154                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 155         }
 156         if (req->r_reply)
 157                 ceph_msg_put(req->r_reply);
 158         if (req->r_own_pages)
 159                 ceph_release_page_vector(req->r_pages,
 160                                          req->r_num_pages);
 161 #ifdef CONFIG_BLOCK
 162         if (req->r_bio)
 163                 bio_put(req->r_bio);
 164 #endif
 165         ceph_put_snap_context(req->r_snapc);
 166         if (req->r_trail) {
 167                 ceph_pagelist_release(req->r_trail);
 168                 kfree(req->r_trail);
 169         }
 170         if (req->r_mempool)
 171                 mempool_free(req, req->r_osdc->req_mempool);
 172         else
 173                 kfree(req);
 174 }
 175 EXPORT_SYMBOL(ceph_osdc_release_request);
 176
 177 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
 178 {
 179         int i = 0;
 180
 181         if (needs_trail)
 182                 *needs_trail = 0;
 183         while (ops[i].op) {
 184                 if (needs_trail && op_needs_trail(ops[i].op))
 185                         *needs_trail = 1;
 186                 i++;
 187         }
 188
 189         return i;
 190 }
 191
 192 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 193                                                int flags,
 194                                                struct ceph_snap_context *snapc,
 195                                                struct ceph_osd_req_op *ops,
 196                                                bool use_mempool,
 197                                                gfp_t gfp_flags,
 198                                                struct page **pages,
 199                                                struct bio *bio)
 200 {
 201         struct ceph_osd_request *req;
 202         struct ceph_msg *msg;
 203         int needs_trail;
 204         int num_op = get_num_ops(ops, &needs_trail);
 205         size_t msg_size = sizeof(struct ceph_osd_request_head);
 206
 207         msg_size += num_op*sizeof(struct ceph_osd_op);
 208
 209         if (use_mempool) {
 210                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 211                 memset(req, 0, sizeof(*req));
 212         } else {
 213                 req = kzalloc(sizeof(*req), gfp_flags);
 214         }
 215         if (req == NULL)
 216                 return NULL;
 217
 218         req->r_osdc = osdc;
 219         req->r_mempool = use_mempool;
 220
 221         kref_init(&req->r_kref);
 222         init_completion(&req->r_completion);
 223         init_completion(&req->r_safe_completion);
 224         rb_init_node(&req->r_node);
 225         INIT_LIST_HEAD(&req->r_unsafe_item);
 226         INIT_LIST_HEAD(&req->r_linger_item);
 227         INIT_LIST_HEAD(&req->r_linger_osd);
 228         INIT_LIST_HEAD(&req->r_req_lru_item);
 229         INIT_LIST_HEAD(&req->r_osd_item);
 230
 231         req->r_flags = flags;
 232
 233         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 234
 235         /* create reply message */
 236         if (use_mempool)
 237                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 238         else
 239                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 240                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 241         if (!msg) {
 242                 ceph_osdc_put_request(req);
 243                 return NULL;
 244         }
 245         req->r_reply = msg;
 246
 247         /* allocate space for the trailing data */
 248         if (needs_trail) {
 249                 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
 250                 if (!req->r_trail) {
 251                         ceph_osdc_put_request(req);
 252                         return NULL;
 253                 }
 254                 ceph_pagelist_init(req->r_trail);
 255         }
 256
 257         /* create request message; allow space for oid */
 258         msg_size += MAX_OBJ_NAME_SIZE;
 259         if (snapc)
 260                 msg_size += sizeof(u64) * snapc->num_snaps;
 261         if (use_mempool)
 262                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 263         else
 264                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 265         if (!msg) {
 266                 ceph_osdc_put_request(req);
 267                 return NULL;
 268         }
 269
 270         memset(msg->front.iov_base, 0, msg->front.iov_len);
 271
 272         req->r_request = msg;
 273         req->r_pages = pages;
 274 #ifdef CONFIG_BLOCK
 275         if (bio) {
 276                 req->r_bio = bio;
 277                 bio_get(req->r_bio);
 278         }
 279 #endif
 280
 281         return req;
 282 }
 283 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 284
 285 static void osd_req_encode_op(struct ceph_osd_request *req,
 286                               struct ceph_osd_op *dst,
 287                               struct ceph_osd_req_op *src)
 288 {
 289         dst->op = cpu_to_le16(src->op);
 290
 291         switch (src->op) {
 292         case CEPH_OSD_OP_READ:
 293         case CEPH_OSD_OP_WRITE:
 294                 dst->extent.offset =
 295                         cpu_to_le64(src->extent.offset);
 296                 dst->extent.length =
 297                         cpu_to_le64(src->extent.length);
 298                 dst->extent.truncate_size =
 299                         cpu_to_le64(src->extent.truncate_size);
 300                 dst->extent.truncate_seq =
 301                         cpu_to_le32(src->extent.truncate_seq);
 302                 break;
 303
 304         case CEPH_OSD_OP_GETXATTR:
 305         case CEPH_OSD_OP_SETXATTR:
 306         case CEPH_OSD_OP_CMPXATTR:
 307                 BUG_ON(!req->r_trail);
 308
 309                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 310                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 311                 dst->xattr.cmp_op = src->xattr.cmp_op;
 312                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 313                 ceph_pagelist_append(req->r_trail, src->xattr.name,
 314                                      src->xattr.name_len);
 315                 ceph_pagelist_append(req->r_trail, src->xattr.val,
 316                                      src->xattr.value_len);
 317                 break;
 318         case CEPH_OSD_OP_CALL:
 319                 BUG_ON(!req->r_trail);
 320
 321                 dst->cls.class_len = src->cls.class_len;
 322                 dst->cls.method_len = src->cls.method_len;
 323                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 324
 325                 ceph_pagelist_append(req->r_trail, src->cls.class_name,
 326                                      src->cls.class_len);
 327                 ceph_pagelist_append(req->r_trail, src->cls.method_name,
 328                                      src->cls.method_len);
 329                 ceph_pagelist_append(req->r_trail, src->cls.indata,
 330                                      src->cls.indata_len);
 331                 break;
 332         case CEPH_OSD_OP_ROLLBACK:
 333                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
 334                 break;
 335         case CEPH_OSD_OP_STARTSYNC:
 336                 break;
 337         case CEPH_OSD_OP_NOTIFY:
 338                 {
 339                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
 340                         __le32 timeout = cpu_to_le32(src->watch.timeout);
 341
 342                         BUG_ON(!req->r_trail);
 343
 344                         ceph_pagelist_append(req->r_trail,
 345                                                 &prot_ver, sizeof(prot_ver));
 346                         ceph_pagelist_append(req->r_trail,
 347                                                 &timeout, sizeof(timeout));
 348                 }
 349         case CEPH_OSD_OP_NOTIFY_ACK:
 350         case CEPH_OSD_OP_WATCH:
 351                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 352                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 353                 dst->watch.flag = src->watch.flag;
 354                 break;
 355         default:
 356                 pr_err("unrecognized osd opcode %d\n", dst->op);
 357                 WARN_ON(1);
 358                 break;
 359         }
 360         dst->payload_len = cpu_to_le32(src->payload_len);
 361 }
 362
 363 /*
 364  * build new request AND message
 365  *
 366  */
 367 void ceph_osdc_build_request(struct ceph_osd_request *req,
 368                              u64 off, u64 *plen,
 369                              struct ceph_osd_req_op *src_ops,
 370                              struct ceph_snap_context *snapc,
 371                              struct timespec *mtime,
 372                              const char *oid,
 373                              int oid_len)
 374 {
 375         struct ceph_msg *msg = req->r_request;
 376         struct ceph_osd_request_head *head;
 377         struct ceph_osd_req_op *src_op;
 378         struct ceph_osd_op *op;
 379         void *p;
 380         int num_op = get_num_ops(src_ops, NULL);
 381         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 382         int flags = req->r_flags;
 383         u64 data_len = 0;
 384         int i;
 385
 386         head = msg->front.iov_base;
 387         op = (void *)(head + 1);
 388         p = (void *)(op + num_op);
 389
 390         req->r_snapc = ceph_get_snap_context(snapc);
 391
 392         head->client_inc = cpu_to_le32(1); /* always, for now. */
 393         head->flags = cpu_to_le32(flags);
 394         if (flags & CEPH_OSD_FLAG_WRITE)
 395                 ceph_encode_timespec(&head->mtime, mtime);
 396         head->num_ops = cpu_to_le16(num_op);
 397
 398
 399         /* fill in oid */
 400         head->object_len = cpu_to_le32(oid_len);
 401         memcpy(p, oid, oid_len);
 402         p += oid_len;
 403
 404         src_op = src_ops;
 405         while (src_op->op) {
 406                 osd_req_encode_op(req, op, src_op);
 407                 src_op++;
 408                 op++;
 409         }
 410
 411         if (req->r_trail)
 412                 data_len += req->r_trail->length;
 413
 414         if (snapc) {
 415                 head->snap_seq = cpu_to_le64(snapc->seq);
 416                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 417                 for (i = 0; i < snapc->num_snaps; i++) {
 418                         put_unaligned_le64(snapc->snaps[i], p);
 419                         p += sizeof(u64);
 420                 }
 421         }
 422
 423         if (flags & CEPH_OSD_FLAG_WRITE) {
 424                 req->r_request->hdr.data_off = cpu_to_le16(off);
 425                 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
 426         } else if (data_len) {
 427                 req->r_request->hdr.data_off = 0;
 428                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 429         }
 430
 431         req->r_request->page_alignment = req->r_page_alignment;
 432
 433         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 434         msg_size = p - msg->front.iov_base;
 435         msg->front.iov_len = msg_size;
 436         msg->hdr.front_len = cpu_to_le32(msg_size);
 437         return;
 438 }
 439 EXPORT_SYMBOL(ceph_osdc_build_request);
 440
 441 /*
 442  * build new request AND message, calculate layout, and adjust file
 443  * extent as needed.
 444  *
 445  * if the file was recently truncated, we include information about its
 446  * old and new size so that the object can be updated appropriately.  (we
 447  * avoid synchronously deleting truncated objects because it's slow.)
 448  *
 449  * if @do_sync, include a 'startsync' command so that the osd will flush
 450  * data quickly.
 451  */
 452 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 453                                                struct ceph_file_layout *layout,
 454                                                struct ceph_vino vino,
 455                                                u64 off, u64 *plen,
 456                                                int opcode, int flags,
 457                                                struct ceph_snap_context *snapc,
 458                                                int do_sync,
 459                                                u32 truncate_seq,
 460                                                u64 truncate_size,
 461                                                struct timespec *mtime,
 462                                                bool use_mempool, int num_reply,
 463                                                int page_align)
 464 {
 465         struct ceph_osd_req_op ops[3];
 466         struct ceph_osd_request *req;
 467         int r;
 468
 469         ops[0].op = opcode;
 470         ops[0].extent.truncate_seq = truncate_seq;
 471         ops[0].extent.truncate_size = truncate_size;
 472         ops[0].payload_len = 0;
 473
 474         if (do_sync) {
 475                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 476                 ops[1].payload_len = 0;
 477                 ops[2].op = 0;
 478         } else
 479                 ops[1].op = 0;
 480
 481         req = ceph_osdc_alloc_request(osdc, flags,
 482                                          snapc, ops,
 483                                          use_mempool,
 484                                          GFP_NOFS, NULL, NULL);
 485         if (!req)
 486                 return ERR_PTR(-ENOMEM);
 487
 488         /* calculate max write size */
 489         r = calc_layout(osdc, vino, layout, off, plen, req, ops);
 490         if (r < 0)
 491                 return ERR_PTR(r);
 492         req->r_file_layout = *layout;  /* keep a copy */
 493
 494         /* in case it differs from natural (file) alignment that
 495            calc_layout filled in for us */
 496         req->r_num_pages = calc_pages_for(page_align, *plen);
 497         req->r_page_alignment = page_align;
 498
 499         ceph_osdc_build_request(req, off, plen, ops,
 500                                 snapc,
 501                                 mtime,
 502                                 req->r_oid, req->r_oid_len);
 503
 504         return req;
 505 }
 506 EXPORT_SYMBOL(ceph_osdc_new_request);
 507
 508 /*
 509  * We keep osd requests in an rbtree, sorted by ->r_tid.
 510  */
 511 static void __insert_request(struct ceph_osd_client *osdc,
 512                              struct ceph_osd_request *new)
 513 {
 514         struct rb_node **p = &osdc->requests.rb_node;
 515         struct rb_node *parent = NULL;
 516         struct ceph_osd_request *req = NULL;
 517
 518         while (*p) {
 519                 parent = *p;
 520                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 521                 if (new->r_tid < req->r_tid)
 522                         p = &(*p)->rb_left;
 523                 else if (new->r_tid > req->r_tid)
 524                         p = &(*p)->rb_right;
 525                 else
 526                         BUG();
 527         }
 528
 529         rb_link_node(&new->r_node, parent, p);
 530         rb_insert_color(&new->r_node, &osdc->requests);
 531 }
 532
 533 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 534                                                  u64 tid)
 535 {
 536         struct ceph_osd_request *req;
 537         struct rb_node *n = osdc->requests.rb_node;
 538
 539         while (n) {
 540                 req = rb_entry(n, struct ceph_osd_request, r_node);
 541                 if (tid < req->r_tid)
 542                         n = n->rb_left;
 543                 else if (tid > req->r_tid)
 544                         n = n->rb_right;
 545                 else
 546                         return req;
 547         }
 548         return NULL;
 549 }
 550
 551 static struct ceph_osd_request *
 552 __lookup_request_ge(struct ceph_osd_client *osdc,
 553                     u64 tid)
 554 {
 555         struct ceph_osd_request *req;
 556         struct rb_node *n = osdc->requests.rb_node;
 557
 558         while (n) {
 559                 req = rb_entry(n, struct ceph_osd_request, r_node);
 560                 if (tid < req->r_tid) {
 561                         if (!n->rb_left)
 562                                 return req;
 563                         n = n->rb_left;
 564                 } else if (tid > req->r_tid) {
 565                         n = n->rb_right;
 566                 } else {
 567                         return req;
 568                 }
 569         }
 570         return NULL;
 571 }
 572
 573 /*
 574  * Resubmit requests pending on the given osd.
 575  */
 576 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 577                                 struct ceph_osd *osd)
 578 {
 579         struct ceph_osd_request *req, *nreq;
 580         int err;
 581
 582         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 583         err = __reset_osd(osdc, osd);
 584         if (err)
 585                 return;
 586
 587         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 588                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 589                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 590                      osd->o_osd);
 591                 if (!req->r_linger)
 592                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 593         }
 594
 595         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 596                                  r_linger_osd) {
 597                 /*
 598                  * reregister request prior to unregistering linger so
 599                  * that r_osd is preserved.
 600                  */
 601                 BUG_ON(!list_empty(&req->r_req_lru_item));
 602                 __register_request(osdc, req);
 603                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 604                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 605                 __unregister_linger_request(osdc, req);
 606                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 607                      osd->o_osd);
 608         }
 609 }
 610
 611 /*
 612  * If the osd connection drops, we need to resubmit all requests.
 613  */
 614 static void osd_reset(struct ceph_connection *con)
 615 {
 616         struct ceph_osd *osd = con->private;
 617         struct ceph_osd_client *osdc;
 618
 619         if (!osd)
 620                 return;
 621         dout("osd_reset osd%d\n", osd->o_osd);
 622         osdc = osd->o_osdc;
 623         down_read(&osdc->map_sem);
 624         mutex_lock(&osdc->request_mutex);
 625         __kick_osd_requests(osdc, osd);
 626         mutex_unlock(&osdc->request_mutex);
 627         send_queued(osdc);
 628         up_read(&osdc->map_sem);
 629 }
 630
 631 /*
 632  * Track open sessions with osds.
 633  */
 634 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 635 {
 636         struct ceph_osd *osd;
 637
 638         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 639         if (!osd)
 640                 return NULL;
 641
 642         atomic_set(&osd->o_ref, 1);
 643         osd->o_osdc = osdc;
 644         osd->o_osd = onum;
 645         RB_CLEAR_NODE(&osd->o_node);
 646         INIT_LIST_HEAD(&osd->o_requests);
 647         INIT_LIST_HEAD(&osd->o_linger_requests);
 648         INIT_LIST_HEAD(&osd->o_osd_lru);
 649         osd->o_incarnation = 1;
 650
 651         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 652
 653         INIT_LIST_HEAD(&osd->o_keepalive_item);
 654         return osd;
 655 }
 656
 657 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 658 {
 659         if (atomic_inc_not_zero(&osd->o_ref)) {
 660                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 661                      atomic_read(&osd->o_ref));
 662                 return osd;
 663         } else {
 664                 dout("get_osd %p FAIL\n", osd);
 665                 return NULL;
 666         }
 667 }
 668
 669 static void put_osd(struct ceph_osd *osd)
 670 {
 671         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 672              atomic_read(&osd->o_ref) - 1);
 673         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 674                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 675
 676                 if (ac->ops && ac->ops->destroy_authorizer)
 677                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 678                 kfree(osd);
 679         }
 680 }
 681
 682 /*
 683  * remove an osd from our map
 684  */
 685 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 686 {
 687         dout("__remove_osd %p\n", osd);
 688         BUG_ON(!list_empty(&osd->o_requests));
 689         rb_erase(&osd->o_node, &osdc->osds);
 690         list_del_init(&osd->o_osd_lru);
 691         ceph_con_close(&osd->o_con);
 692         put_osd(osd);
 693 }
 694
 695 static void remove_all_osds(struct ceph_osd_client *osdc)
 696 {
 697         dout("%s %p\n", __func__, osdc);
 698         mutex_lock(&osdc->request_mutex);
 699         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 700                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 701                                                 struct ceph_osd, o_node);
 702                 __remove_osd(osdc, osd);
 703         }
 704         mutex_unlock(&osdc->request_mutex);
 705 }
 706
 707 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 708                               struct ceph_osd *osd)
 709 {
 710         dout("__move_osd_to_lru %p\n", osd);
 711         BUG_ON(!list_empty(&osd->o_osd_lru));
 712         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 713         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 714 }
 715
 716 static void __remove_osd_from_lru(struct ceph_osd *osd)
 717 {
 718         dout("__remove_osd_from_lru %p\n", osd);
 719         if (!list_empty(&osd->o_osd_lru))
 720                 list_del_init(&osd->o_osd_lru);
 721 }
 722
 723 static void remove_old_osds(struct ceph_osd_client *osdc)
 724 {
 725         struct ceph_osd *osd, *nosd;
 726
 727         dout("__remove_old_osds %p\n", osdc);
 728         mutex_lock(&osdc->request_mutex);
 729         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 730                 if (time_before(jiffies, osd->lru_ttl))
 731                         break;
 732                 __remove_osd(osdc, osd);
 733         }
 734         mutex_unlock(&osdc->request_mutex);
 735 }
 736
 737 /*
 738  * reset osd connect
 739  */
 740 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 741 {
 742         struct ceph_osd_request *req;
 743         int ret = 0;
 744
 745         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 746         if (list_empty(&osd->o_requests) &&
 747             list_empty(&osd->o_linger_requests)) {
 748                 __remove_osd(osdc, osd);
 749                 ret = -ENODEV;
 750         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 751                           &osd->o_con.peer_addr,
 752                           sizeof(osd->o_con.peer_addr)) == 0 &&
 753                    !ceph_con_opened(&osd->o_con)) {
 754                 dout(" osd addr hasn't changed and connection never opened,"
 755                      " letting msgr retry");
 756                 /* touch each r_stamp for handle_timeout()'s benfit */
 757                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 758                         req->r_stamp = jiffies;
 759                 ret = -EAGAIN;
 760         } else {
 761                 ceph_con_close(&osd->o_con);
 762                 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
 763                               &osdc->osdmap->osd_addr[osd->o_osd]);
 764                 osd->o_incarnation++;
 765         }
 766         return ret;
 767 }
 768
 769 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 770 {
 771         struct rb_node **p = &osdc->osds.rb_node;
 772         struct rb_node *parent = NULL;
 773         struct ceph_osd *osd = NULL;
 774
 775         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 776         while (*p) {
 777                 parent = *p;
 778                 osd = rb_entry(parent, struct ceph_osd, o_node);
 779                 if (new->o_osd < osd->o_osd)
 780                         p = &(*p)->rb_left;
 781                 else if (new->o_osd > osd->o_osd)
 782                         p = &(*p)->rb_right;
 783                 else
 784                         BUG();
 785         }
 786
 787         rb_link_node(&new->o_node, parent, p);
 788         rb_insert_color(&new->o_node, &osdc->osds);
 789 }
 790
 791 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 792 {
 793         struct ceph_osd *osd;
 794         struct rb_node *n = osdc->osds.rb_node;
 795
 796         while (n) {
 797                 osd = rb_entry(n, struct ceph_osd, o_node);
 798                 if (o < osd->o_osd)
 799                         n = n->rb_left;
 800                 else if (o > osd->o_osd)
 801                         n = n->rb_right;
 802                 else
 803                         return osd;
 804         }
 805         return NULL;
 806 }
 807
 808 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 809 {
 810         schedule_delayed_work(&osdc->timeout_work,
 811                         osdc->client->options->osd_keepalive_timeout * HZ);
 812 }
 813
 814 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 815 {
 816         cancel_delayed_work(&osdc->timeout_work);
 817 }
 818
 819 /*
 820  * Register request, assign tid.  If this is the first request, set up
 821  * the timeout event.
 822  */
 823 static void __register_request(struct ceph_osd_client *osdc,
 824                                struct ceph_osd_request *req)
 825 {
 826         req->r_tid = ++osdc->last_tid;
 827         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 828         dout("__register_request %p tid %lld\n", req, req->r_tid);
 829         __insert_request(osdc, req);
 830         ceph_osdc_get_request(req);
 831         osdc->num_requests++;
 832         if (osdc->num_requests == 1) {
 833                 dout(" first request, scheduling timeout\n");
 834                 __schedule_osd_timeout(osdc);
 835         }
 836 }
 837
 838 static void register_request(struct ceph_osd_client *osdc,
 839                              struct ceph_osd_request *req)
 840 {
 841         mutex_lock(&osdc->request_mutex);
 842         __register_request(osdc, req);
 843         mutex_unlock(&osdc->request_mutex);
 844 }
 845
 846 /*
 847  * called under osdc->request_mutex
 848  */
 849 static void __unregister_request(struct ceph_osd_client *osdc,
 850                                  struct ceph_osd_request *req)
 851 {
 852         if (RB_EMPTY_NODE(&req->r_node)) {
 853                 dout("__unregister_request %p tid %lld not registered\n",
 854                         req, req->r_tid);
 855                 return;
 856         }
 857
 858         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 859         rb_erase(&req->r_node, &osdc->requests);
 860         osdc->num_requests--;
 861
 862         if (req->r_osd) {
 863                 /* make sure the original request isn't in flight. */
 864                 ceph_msg_revoke(req->r_request);
 865
 866                 list_del_init(&req->r_osd_item);
 867                 if (list_empty(&req->r_osd->o_requests) &&
 868                     list_empty(&req->r_osd->o_linger_requests)) {
 869                         dout("moving osd to %p lru\n", req->r_osd);
 870                         __move_osd_to_lru(osdc, req->r_osd);
 871                 }
 872                 if (list_empty(&req->r_linger_item))
 873                         req->r_osd = NULL;
 874         }
 875
 876         list_del_init(&req->r_req_lru_item);
 877         ceph_osdc_put_request(req);
 878
 879         if (osdc->num_requests == 0) {
 880                 dout(" no requests, canceling timeout\n");
 881                 __cancel_osd_timeout(osdc);
 882         }
 883 }
 884
 885 /*
 886  * Cancel a previously queued request message
 887  */
 888 static void __cancel_request(struct ceph_osd_request *req)
 889 {
 890         if (req->r_sent && req->r_osd) {
 891                 ceph_msg_revoke(req->r_request);
 892                 req->r_sent = 0;
 893         }
 894 }
 895
 896 static void __register_linger_request(struct ceph_osd_client *osdc,
 897                                     struct ceph_osd_request *req)
 898 {
 899         dout("__register_linger_request %p\n", req);
 900         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 901         if (req->r_osd)
 902                 list_add_tail(&req->r_linger_osd,
 903                               &req->r_osd->o_linger_requests);
 904 }
 905
 906 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 907                                         struct ceph_osd_request *req)
 908 {
 909         dout("__unregister_linger_request %p\n", req);
 910         list_del_init(&req->r_linger_item);
 911         if (req->r_osd) {
 912                 list_del_init(&req->r_linger_osd);
 913
 914                 if (list_empty(&req->r_osd->o_requests) &&
 915                     list_empty(&req->r_osd->o_linger_requests)) {
 916                         dout("moving osd to %p lru\n", req->r_osd);
 917                         __move_osd_to_lru(osdc, req->r_osd);
 918                 }
 919                 if (list_empty(&req->r_osd_item))
 920                         req->r_osd = NULL;
 921         }
 922 }
 923
 924 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 925                                          struct ceph_osd_request *req)
 926 {
 927         mutex_lock(&osdc->request_mutex);
 928         if (req->r_linger) {
 929                 __unregister_linger_request(osdc, req);
 930                 ceph_osdc_put_request(req);
 931         }
 932         mutex_unlock(&osdc->request_mutex);
 933 }
 934 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 935
 936 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 937                                   struct ceph_osd_request *req)
 938 {
 939         if (!req->r_linger) {
 940                 dout("set_request_linger %p\n", req);
 941                 req->r_linger = 1;
 942                 /*
 943                  * caller is now responsible for calling
 944                  * unregister_linger_request
 945                  */
 946                 ceph_osdc_get_request(req);
 947         }
 948 }
 949 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 950
 951 /*
 952  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 953  * (as needed), and set the request r_osd appropriately.  If there is
 954  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 955  * (unsent, homeless) or leave on in-flight lru.
 956  *
 957  * Return 0 if unchanged, 1 if changed, or negative on error.
 958  *
 959  * Caller should hold map_sem for read and request_mutex.
 960  */
 961 static int __map_request(struct ceph_osd_client *osdc,
 962                          struct ceph_osd_request *req, int force_resend)
 963 {
 964         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 965         struct ceph_pg pgid;
 966         int acting[CEPH_PG_MAX_SIZE];
 967         int o = -1, num = 0;
 968         int err;
 969
 970         dout("map_request %p tid %lld\n", req, req->r_tid);
 971         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 972                                       &req->r_file_layout, osdc->osdmap);
 973         if (err) {
 974                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 975                 return err;
 976         }
 977         pgid = reqhead->layout.ol_pgid;
 978         req->r_pgid = pgid;
 979
 980         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 981         if (err > 0) {
 982                 o = acting[0];
 983                 num = err;
 984         }
 985
 986         if ((!force_resend &&
 987              req->r_osd && req->r_osd->o_osd == o &&
 988              req->r_sent >= req->r_osd->o_incarnation &&
 989              req->r_num_pg_osds == num &&
 990              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 991             (req->r_osd == NULL && o == -1))
 992                 return 0;  /* no change */
 993
 994         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 995              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 996              req->r_osd ? req->r_osd->o_osd : -1);
 997
 998         /* record full pg acting set */
 999         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1000         req->r_num_pg_osds = num;
1001
1002         if (req->r_osd) {
1003                 __cancel_request(req);
1004                 list_del_init(&req->r_osd_item);
1005                 req->r_osd = NULL;
1006         }
1007
1008         req->r_osd = __lookup_osd(osdc, o);
1009         if (!req->r_osd && o >= 0) {
1010                 err = -ENOMEM;
1011                 req->r_osd = create_osd(osdc, o);
1012                 if (!req->r_osd) {
1013                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1014                         goto out;
1015                 }
1016
1017                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1018                 __insert_osd(osdc, req->r_osd);
1019
1020                 ceph_con_open(&req->r_osd->o_con,
1021                               CEPH_ENTITY_TYPE_OSD, o,
1022                               &osdc->osdmap->osd_addr[o]);
1023         }
1024
1025         if (req->r_osd) {
1026                 __remove_osd_from_lru(req->r_osd);
1027                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1028                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1029         } else {
1030                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1031         }
1032         err = 1;   /* osd or pg changed */
1033
1034 out:
1035         return err;
1036 }
1037
1038 /*
1039  * caller should hold map_sem (for read) and request_mutex
1040  */
1041 static void __send_request(struct ceph_osd_client *osdc,
1042                            struct ceph_osd_request *req)
1043 {
1044         struct ceph_osd_request_head *reqhead;
1045
1046         dout("send_request %p tid %llu to osd%d flags %d\n",
1047              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1048
1049         reqhead = req->r_request->front.iov_base;
1050         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1051         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1052         reqhead->reassert_version = req->r_reassert_version;
1053
1054         req->r_stamp = jiffies;
1055         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1056
1057         ceph_msg_get(req->r_request); /* send consumes a ref */
1058         ceph_con_send(&req->r_osd->o_con, req->r_request);
1059         req->r_sent = req->r_osd->o_incarnation;
1060 }
1061
1062 /*
1063  * Send any requests in the queue (req_unsent).
1064  */
1065 static void send_queued(struct ceph_osd_client *osdc)
1066 {
1067         struct ceph_osd_request *req, *tmp;
1068
1069         dout("send_queued\n");
1070         mutex_lock(&osdc->request_mutex);
1071         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1072                 __send_request(osdc, req);
1073         }
1074         mutex_unlock(&osdc->request_mutex);
1075 }
1076
1077 /*
1078  * Timeout callback, called every N seconds when 1 or more osd
1079  * requests has been active for more than N seconds.  When this
1080  * happens, we ping all OSDs with requests who have timed out to
1081  * ensure any communications channel reset is detected.  Reset the
1082  * request timeouts another N seconds in the future as we go.
1083  * Reschedule the timeout event another N seconds in future (unless
1084  * there are no open requests).
1085  */
1086 static void handle_timeout(struct work_struct *work)
1087 {
1088         struct ceph_osd_client *osdc =
1089                 container_of(work, struct ceph_osd_client, timeout_work.work);
1090         struct ceph_osd_request *req;
1091         struct ceph_osd *osd;
1092         unsigned long keepalive =
1093                 osdc->client->options->osd_keepalive_timeout * HZ;
1094         struct list_head slow_osds;
1095         dout("timeout\n");
1096         down_read(&osdc->map_sem);
1097
1098         ceph_monc_request_next_osdmap(&osdc->client->monc);
1099
1100         mutex_lock(&osdc->request_mutex);
1101
1102         /*
1103          * ping osds that are a bit slow.  this ensures that if there
1104          * is a break in the TCP connection we will notice, and reopen
1105          * a connection with that osd (from the fault callback).
1106          */
1107         INIT_LIST_HEAD(&slow_osds);
1108         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1109                 if (time_before(jiffies, req->r_stamp + keepalive))
1110                         break;
1111
1112                 osd = req->r_osd;
1113                 BUG_ON(!osd);
1114                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1115                      req->r_tid, osd->o_osd);
1116                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1117         }
1118         while (!list_empty(&slow_osds)) {
1119                 osd = list_entry(slow_osds.next, struct ceph_osd,
1120                                  o_keepalive_item);
1121                 list_del_init(&osd->o_keepalive_item);
1122                 ceph_con_keepalive(&osd->o_con);
1123         }
1124
1125         __schedule_osd_timeout(osdc);
1126         mutex_unlock(&osdc->request_mutex);
1127         send_queued(osdc);
1128         up_read(&osdc->map_sem);
1129 }
1130
1131 static void handle_osds_timeout(struct work_struct *work)
1132 {
1133         struct ceph_osd_client *osdc =
1134                 container_of(work, struct ceph_osd_client,
1135                              osds_timeout_work.work);
1136         unsigned long delay =
1137                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1138
1139         dout("osds timeout\n");
1140         down_read(&osdc->map_sem);
1141         remove_old_osds(osdc);
1142         up_read(&osdc->map_sem);
1143
1144         schedule_delayed_work(&osdc->osds_timeout_work,
1145                               round_jiffies_relative(delay));
1146 }
1147
1148 static void complete_request(struct ceph_osd_request *req)
1149 {
1150         if (req->r_safe_callback)
1151                 req->r_safe_callback(req, NULL);
1152         complete_all(&req->r_safe_completion);  /* fsync waiter */
1153 }
1154
1155 /*
1156  * handle osd op reply.  either call the callback if it is specified,
1157  * or do the completion to wake up the waiting thread.
1158  */
1159 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1160                          struct ceph_connection *con)
1161 {
1162         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1163         struct ceph_osd_request *req;
1164         u64 tid;
1165         int numops, object_len, flags;
1166         s32 result;
1167
1168         tid = le64_to_cpu(msg->hdr.tid);
1169         if (msg->front.iov_len < sizeof(*rhead))
1170                 goto bad;
1171         numops = le32_to_cpu(rhead->num_ops);
1172         object_len = le32_to_cpu(rhead->object_len);
1173         result = le32_to_cpu(rhead->result);
1174         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1175             numops * sizeof(struct ceph_osd_op))
1176                 goto bad;
1177         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1178         /* lookup */
1179         mutex_lock(&osdc->request_mutex);
1180         req = __lookup_request(osdc, tid);
1181         if (req == NULL) {
1182                 dout("handle_reply tid %llu dne\n", tid);
1183                 mutex_unlock(&osdc->request_mutex);
1184                 return;
1185         }
1186         ceph_osdc_get_request(req);
1187         flags = le32_to_cpu(rhead->flags);
1188
1189         /*
1190          * if this connection filled our message, drop our reference now, to
1191          * avoid a (safe but slower) revoke later.
1192          */
1193         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1194                 dout(" dropping con_filling_msg ref %p\n", con);
1195                 req->r_con_filling_msg = NULL;
1196                 con->ops->put(con);
1197         }
1198
1199         if (!req->r_got_reply) {
1200                 unsigned int bytes;
1201
1202                 req->r_result = le32_to_cpu(rhead->result);
1203                 bytes = le32_to_cpu(msg->hdr.data_len);
1204                 dout("handle_reply result %d bytes %d\n", req->r_result,
1205                      bytes);
1206                 if (req->r_result == 0)
1207                         req->r_result = bytes;
1208
1209                 /* in case this is a write and we need to replay, */
1210                 req->r_reassert_version = rhead->reassert_version;
1211
1212                 req->r_got_reply = 1;
1213         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1214                 dout("handle_reply tid %llu dup ack\n", tid);
1215                 mutex_unlock(&osdc->request_mutex);
1216                 goto done;
1217         }
1218
1219         dout("handle_reply tid %llu flags %d\n", tid, flags);
1220
1221         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1222                 __register_linger_request(osdc, req);
1223
1224         /* either this is a read, or we got the safe response */
1225         if (result < 0 ||
1226             (flags & CEPH_OSD_FLAG_ONDISK) ||
1227             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1228                 __unregister_request(osdc, req);
1229
1230         mutex_unlock(&osdc->request_mutex);
1231
1232         if (req->r_callback)
1233                 req->r_callback(req, msg);
1234         else
1235                 complete_all(&req->r_completion);
1236
1237         if (flags & CEPH_OSD_FLAG_ONDISK)
1238                 complete_request(req);
1239
1240 done:
1241         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1242         ceph_osdc_put_request(req);
1243         return;
1244
1245 bad:
1246         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1247                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1248                (int)sizeof(*rhead));
1249         ceph_msg_dump(msg);
1250 }
1251
1252 static void reset_changed_osds(struct ceph_osd_client *osdc)
1253 {
1254         struct rb_node *p, *n;
1255
1256         for (p = rb_first(&osdc->osds); p; p = n) {
1257                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1258
1259                 n = rb_next(p);
1260                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1261                     memcmp(&osd->o_con.peer_addr,
1262                            ceph_osd_addr(osdc->osdmap,
1263                                          osd->o_osd),
1264                            sizeof(struct ceph_entity_addr)) != 0)
1265                         __reset_osd(osdc, osd);
1266         }
1267 }
1268
1269 /*
1270  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1271  * no osd, request a new map.
1272  *
1273  * Caller should hold map_sem for read and request_mutex.
1274  */
1275 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1276 {
1277         struct ceph_osd_request *req, *nreq;
1278         struct rb_node *p;
1279         int needmap = 0;
1280         int err;
1281
1282         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1283         mutex_lock(&osdc->request_mutex);
1284         for (p = rb_first(&osdc->requests); p; ) {
1285                 req = rb_entry(p, struct ceph_osd_request, r_node);
1286                 p = rb_next(p);
1287                 err = __map_request(osdc, req, force_resend);
1288                 if (err < 0)
1289                         continue;  /* error */
1290                 if (req->r_osd == NULL) {
1291                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1292                         needmap++;  /* request a newer map */
1293                 } else if (err > 0) {
1294                         if (!req->r_linger) {
1295                                 dout("%p tid %llu requeued on osd%d\n", req,
1296                                      req->r_tid,
1297                                      req->r_osd ? req->r_osd->o_osd : -1);
1298                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1299                         }
1300                 }
1301                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1302                         /*
1303                          * register as a linger so that we will
1304                          * re-submit below and get a new tid
1305                          */
1306                         dout("%p tid %llu restart on osd%d\n",
1307                              req, req->r_tid,
1308                              req->r_osd ? req->r_osd->o_osd : -1);
1309                         __register_linger_request(osdc, req);
1310                         __unregister_request(osdc, req);
1311                 }
1312         }
1313
1314         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1315                                  r_linger_item) {
1316                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1317
1318                 err = __map_request(osdc, req, force_resend);
1319                 if (err == 0)
1320                         continue;  /* no change and no osd was specified */
1321                 if (err < 0)
1322                         continue;  /* hrm! */
1323                 if (req->r_osd == NULL) {
1324                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1325                         needmap++;  /* request a newer map */
1326                         continue;
1327                 }
1328
1329                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1330                      req->r_osd ? req->r_osd->o_osd : -1);
1331                 __unregister_linger_request(osdc, req);
1332                 __register_request(osdc, req);
1333         }
1334         mutex_unlock(&osdc->request_mutex);
1335
1336         if (needmap) {
1337                 dout("%d requests for down osds, need new map\n", needmap);
1338                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1339         }
1340 }
1341
1342
1343 /*
1344  * Process updated osd map.
1345  *
1346  * The message contains any number of incremental and full maps, normally
1347  * indicating some sort of topology change in the cluster.  Kick requests
1348  * off to different OSDs as needed.
1349  */
1350 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1351 {
1352         void *p, *end, *next;
1353         u32 nr_maps, maplen;
1354         u32 epoch;
1355         struct ceph_osdmap *newmap = NULL, *oldmap;
1356         int err;
1357         struct ceph_fsid fsid;
1358
1359         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1360         p = msg->front.iov_base;
1361         end = p + msg->front.iov_len;
1362
1363         /* verify fsid */
1364         ceph_decode_need(&p, end, sizeof(fsid), bad);
1365         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1366         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1367                 return;
1368
1369         down_write(&osdc->map_sem);
1370
1371         /* incremental maps */
1372         ceph_decode_32_safe(&p, end, nr_maps, bad);
1373         dout(" %d inc maps\n", nr_maps);
1374         while (nr_maps > 0) {
1375                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1376                 epoch = ceph_decode_32(&p);
1377                 maplen = ceph_decode_32(&p);
1378                 ceph_decode_need(&p, end, maplen, bad);
1379                 next = p + maplen;
1380                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1381                         dout("applying incremental map %u len %d\n",
1382                              epoch, maplen);
1383                         newmap = osdmap_apply_incremental(&p, next,
1384                                                           osdc->osdmap,
1385                                                           &osdc->client->msgr);
1386                         if (IS_ERR(newmap)) {
1387                                 err = PTR_ERR(newmap);
1388                                 goto bad;
1389                         }
1390                         BUG_ON(!newmap);
1391                         if (newmap != osdc->osdmap) {
1392                                 ceph_osdmap_destroy(osdc->osdmap);
1393                                 osdc->osdmap = newmap;
1394                         }
1395                         kick_requests(osdc, 0);
1396                         reset_changed_osds(osdc);
1397                 } else {
1398                         dout("ignoring incremental map %u len %d\n",
1399                              epoch, maplen);
1400                 }
1401                 p = next;
1402                 nr_maps--;
1403         }
1404         if (newmap)
1405                 goto done;
1406
1407         /* full maps */
1408         ceph_decode_32_safe(&p, end, nr_maps, bad);
1409         dout(" %d full maps\n", nr_maps);
1410         while (nr_maps) {
1411                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1412                 epoch = ceph_decode_32(&p);
1413                 maplen = ceph_decode_32(&p);
1414                 ceph_decode_need(&p, end, maplen, bad);
1415                 if (nr_maps > 1) {
1416                         dout("skipping non-latest full map %u len %d\n",
1417                              epoch, maplen);
1418                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1419                         dout("skipping full map %u len %d, "
1420                              "older than our %u\n", epoch, maplen,
1421                              osdc->osdmap->epoch);
1422                 } else {
1423                         int skipped_map = 0;
1424
1425                         dout("taking full map %u len %d\n", epoch, maplen);
1426                         newmap = osdmap_decode(&p, p+maplen);
1427                         if (IS_ERR(newmap)) {
1428                                 err = PTR_ERR(newmap);
1429                                 goto bad;
1430                         }
1431                         BUG_ON(!newmap);
1432                         oldmap = osdc->osdmap;
1433                         osdc->osdmap = newmap;
1434                         if (oldmap) {
1435                                 if (oldmap->epoch + 1 < newmap->epoch)
1436                                         skipped_map = 1;
1437                                 ceph_osdmap_destroy(oldmap);
1438                         }
1439                         kick_requests(osdc, skipped_map);
1440                 }
1441                 p += maplen;
1442                 nr_maps--;
1443         }
1444
1445 done:
1446         downgrade_write(&osdc->map_sem);
1447         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1448
1449         /*
1450          * subscribe to subsequent osdmap updates if full to ensure
1451          * we find out when we are no longer full and stop returning
1452          * ENOSPC.
1453          */
1454         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1455                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1456
1457         send_queued(osdc);
1458         up_read(&osdc->map_sem);
1459         wake_up_all(&osdc->client->auth_wq);
1460         return;
1461
1462 bad:
1463         pr_err("osdc handle_map corrupt msg\n");
1464         ceph_msg_dump(msg);
1465         up_write(&osdc->map_sem);
1466         return;
1467 }
1468
1469 /*
1470  * watch/notify callback event infrastructure
1471  *
1472  * These callbacks are used both for watch and notify operations.
1473  */
1474 static void __release_event(struct kref *kref)
1475 {
1476         struct ceph_osd_event *event =
1477                 container_of(kref, struct ceph_osd_event, kref);
1478
1479         dout("__release_event %p\n", event);
1480         kfree(event);
1481 }
1482
1483 static void get_event(struct ceph_osd_event *event)
1484 {
1485         kref_get(&event->kref);
1486 }
1487
1488 void ceph_osdc_put_event(struct ceph_osd_event *event)
1489 {
1490         kref_put(&event->kref, __release_event);
1491 }
1492 EXPORT_SYMBOL(ceph_osdc_put_event);
1493
1494 static void __insert_event(struct ceph_osd_client *osdc,
1495                              struct ceph_osd_event *new)
1496 {
1497         struct rb_node **p = &osdc->event_tree.rb_node;
1498         struct rb_node *parent = NULL;
1499         struct ceph_osd_event *event = NULL;
1500
1501         while (*p) {
1502                 parent = *p;
1503                 event = rb_entry(parent, struct ceph_osd_event, node);
1504                 if (new->cookie < event->cookie)
1505                         p = &(*p)->rb_left;
1506                 else if (new->cookie > event->cookie)
1507                         p = &(*p)->rb_right;
1508                 else
1509                         BUG();
1510         }
1511
1512         rb_link_node(&new->node, parent, p);
1513         rb_insert_color(&new->node, &osdc->event_tree);
1514 }
1515
1516 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1517                                                 u64 cookie)
1518 {
1519         struct rb_node **p = &osdc->event_tree.rb_node;
1520         struct rb_node *parent = NULL;
1521         struct ceph_osd_event *event = NULL;
1522
1523         while (*p) {
1524                 parent = *p;
1525                 event = rb_entry(parent, struct ceph_osd_event, node);
1526                 if (cookie < event->cookie)
1527                         p = &(*p)->rb_left;
1528                 else if (cookie > event->cookie)
1529                         p = &(*p)->rb_right;
1530                 else
1531                         return event;
1532         }
1533         return NULL;
1534 }
1535
1536 static void __remove_event(struct ceph_osd_event *event)
1537 {
1538         struct ceph_osd_client *osdc = event->osdc;
1539
1540         if (!RB_EMPTY_NODE(&event->node)) {
1541                 dout("__remove_event removed %p\n", event);
1542                 rb_erase(&event->node, &osdc->event_tree);
1543                 ceph_osdc_put_event(event);
1544         } else {
1545                 dout("__remove_event didn't remove %p\n", event);
1546         }
1547 }
1548
1549 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1550                            void (*event_cb)(u64, u64, u8, void *),
1551                            int one_shot, void *data,
1552                            struct ceph_osd_event **pevent)
1553 {
1554         struct ceph_osd_event *event;
1555
1556         event = kmalloc(sizeof(*event), GFP_NOIO);
1557         if (!event)
1558                 return -ENOMEM;
1559
1560         dout("create_event %p\n", event);
1561         event->cb = event_cb;
1562         event->one_shot = one_shot;
1563         event->data = data;
1564         event->osdc = osdc;
1565         INIT_LIST_HEAD(&event->osd_node);
1566         kref_init(&event->kref);   /* one ref for us */
1567         kref_get(&event->kref);    /* one ref for the caller */
1568         init_completion(&event->completion);
1569
1570         spin_lock(&osdc->event_lock);
1571         event->cookie = ++osdc->event_count;
1572         __insert_event(osdc, event);
1573         spin_unlock(&osdc->event_lock);
1574
1575         *pevent = event;
1576         return 0;
1577 }
1578 EXPORT_SYMBOL(ceph_osdc_create_event);
1579
1580 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1581 {
1582         struct ceph_osd_client *osdc = event->osdc;
1583
1584         dout("cancel_event %p\n", event);
1585         spin_lock(&osdc->event_lock);
1586         __remove_event(event);
1587         spin_unlock(&osdc->event_lock);
1588         ceph_osdc_put_event(event); /* caller's */
1589 }
1590 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1591
1592
1593 static void do_event_work(struct work_struct *work)
1594 {
1595         struct ceph_osd_event_work *event_work =
1596                 container_of(work, struct ceph_osd_event_work, work);
1597         struct ceph_osd_event *event = event_work->event;
1598         u64 ver = event_work->ver;
1599         u64 notify_id = event_work->notify_id;
1600         u8 opcode = event_work->opcode;
1601
1602         dout("do_event_work completing %p\n", event);
1603         event->cb(ver, notify_id, opcode, event->data);
1604         complete(&event->completion);
1605         dout("do_event_work completed %p\n", event);
1606         ceph_osdc_put_event(event);
1607         kfree(event_work);
1608 }
1609
1610
1611 /*
1612  * Process osd watch notifications
1613  */
1614 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1615 {
1616         void *p, *end;
1617         u8 proto_ver;
1618         u64 cookie, ver, notify_id;
1619         u8 opcode;
1620         struct ceph_osd_event *event;
1621         struct ceph_osd_event_work *event_work;
1622
1623         p = msg->front.iov_base;
1624         end = p + msg->front.iov_len;
1625
1626         ceph_decode_8_safe(&p, end, proto_ver, bad);
1627         ceph_decode_8_safe(&p, end, opcode, bad);
1628         ceph_decode_64_safe(&p, end, cookie, bad);
1629         ceph_decode_64_safe(&p, end, ver, bad);
1630         ceph_decode_64_safe(&p, end, notify_id, bad);
1631
1632         spin_lock(&osdc->event_lock);
1633         event = __find_event(osdc, cookie);
1634         if (event) {
1635                 get_event(event);
1636                 if (event->one_shot)
1637                         __remove_event(event);
1638         }
1639         spin_unlock(&osdc->event_lock);
1640         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1641              cookie, ver, event);
1642         if (event) {
1643                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1644                 if (!event_work) {
1645                         dout("ERROR: could not allocate event_work\n");
1646                         goto done_err;
1647                 }
1648                 INIT_WORK(&event_work->work, do_event_work);
1649                 event_work->event = event;
1650                 event_work->ver = ver;
1651                 event_work->notify_id = notify_id;
1652                 event_work->opcode = opcode;
1653                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1654                         dout("WARNING: failed to queue notify event work\n");
1655                         goto done_err;
1656                 }
1657         }
1658
1659         return;
1660
1661 done_err:
1662         complete(&event->completion);
1663         ceph_osdc_put_event(event);
1664         return;
1665
1666 bad:
1667         pr_err("osdc handle_watch_notify corrupt msg\n");
1668         return;
1669 }
1670
1671 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1672 {
1673         int err;
1674
1675         dout("wait_event %p\n", event);
1676         err = wait_for_completion_interruptible_timeout(&event->completion,
1677                                                         timeout * HZ);
1678         ceph_osdc_put_event(event);
1679         if (err > 0)
1680                 err = 0;
1681         dout("wait_event %p returns %d\n", event, err);
1682         return err;
1683 }
1684 EXPORT_SYMBOL(ceph_osdc_wait_event);
1685
1686 /*
1687  * Register request, send initial attempt.
1688  */
1689 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1690                             struct ceph_osd_request *req,
1691                             bool nofail)
1692 {
1693         int rc = 0;
1694
1695         req->r_request->pages = req->r_pages;
1696         req->r_request->nr_pages = req->r_num_pages;
1697 #ifdef CONFIG_BLOCK
1698         req->r_request->bio = req->r_bio;
1699 #endif
1700         req->r_request->trail = req->r_trail;
1701
1702         register_request(osdc, req);
1703
1704         down_read(&osdc->map_sem);
1705         mutex_lock(&osdc->request_mutex);
1706         /*
1707          * a racing kick_requests() may have sent the message for us
1708          * while we dropped request_mutex above, so only send now if
1709          * the request still han't been touched yet.
1710          */
1711         if (req->r_sent == 0) {
1712                 rc = __map_request(osdc, req, 0);
1713                 if (rc < 0) {
1714                         if (nofail) {
1715                                 dout("osdc_start_request failed map, "
1716                                      " will retry %lld\n", req->r_tid);
1717                                 rc = 0;
1718                         }
1719                         goto out_unlock;
1720                 }
1721                 if (req->r_osd == NULL) {
1722                         dout("send_request %p no up osds in pg\n", req);
1723                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1724                 } else {
1725                         __send_request(osdc, req);
1726                 }
1727                 rc = 0;
1728         }
1729
1730 out_unlock:
1731         mutex_unlock(&osdc->request_mutex);
1732         up_read(&osdc->map_sem);
1733         return rc;
1734 }
1735 EXPORT_SYMBOL(ceph_osdc_start_request);
1736
1737 /*
1738  * wait for a request to complete
1739  */
1740 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1741                            struct ceph_osd_request *req)
1742 {
1743         int rc;
1744
1745         rc = wait_for_completion_interruptible(&req->r_completion);
1746         if (rc < 0) {
1747                 mutex_lock(&osdc->request_mutex);
1748                 __cancel_request(req);
1749                 __unregister_request(osdc, req);
1750                 mutex_unlock(&osdc->request_mutex);
1751                 complete_request(req);
1752                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1753                 return rc;
1754         }
1755
1756         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1757         return req->r_result;
1758 }
1759 EXPORT_SYMBOL(ceph_osdc_wait_request);
1760
1761 /*
1762  * sync - wait for all in-flight requests to flush.  avoid starvation.
1763  */
1764 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1765 {
1766         struct ceph_osd_request *req;
1767         u64 last_tid, next_tid = 0;
1768
1769         mutex_lock(&osdc->request_mutex);
1770         last_tid = osdc->last_tid;
1771         while (1) {
1772                 req = __lookup_request_ge(osdc, next_tid);
1773                 if (!req)
1774                         break;
1775                 if (req->r_tid > last_tid)
1776                         break;
1777
1778                 next_tid = req->r_tid + 1;
1779                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1780                         continue;
1781
1782                 ceph_osdc_get_request(req);
1783                 mutex_unlock(&osdc->request_mutex);
1784                 dout("sync waiting on tid %llu (last is %llu)\n",
1785                      req->r_tid, last_tid);
1786                 wait_for_completion(&req->r_safe_completion);
1787                 mutex_lock(&osdc->request_mutex);
1788                 ceph_osdc_put_request(req);
1789         }
1790         mutex_unlock(&osdc->request_mutex);
1791         dout("sync done (thru tid %llu)\n", last_tid);
1792 }
1793 EXPORT_SYMBOL(ceph_osdc_sync);
1794
1795 /*
1796  * init, shutdown
1797  */
1798 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1799 {
1800         int err;
1801
1802         dout("init\n");
1803         osdc->client = client;
1804         osdc->osdmap = NULL;
1805         init_rwsem(&osdc->map_sem);
1806         init_completion(&osdc->map_waiters);
1807         osdc->last_requested_map = 0;
1808         mutex_init(&osdc->request_mutex);
1809         osdc->last_tid = 0;
1810         osdc->osds = RB_ROOT;
1811         INIT_LIST_HEAD(&osdc->osd_lru);
1812         osdc->requests = RB_ROOT;
1813         INIT_LIST_HEAD(&osdc->req_lru);
1814         INIT_LIST_HEAD(&osdc->req_unsent);
1815         INIT_LIST_HEAD(&osdc->req_notarget);
1816         INIT_LIST_HEAD(&osdc->req_linger);
1817         osdc->num_requests = 0;
1818         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1819         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1820         spin_lock_init(&osdc->event_lock);
1821         osdc->event_tree = RB_ROOT;
1822         osdc->event_count = 0;
1823
1824         schedule_delayed_work(&osdc->osds_timeout_work,
1825            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1826
1827         err = -ENOMEM;
1828         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1829                                         sizeof(struct ceph_osd_request));
1830         if (!osdc->req_mempool)
1831                 goto out;
1832
1833         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1834                                 OSD_OP_FRONT_LEN, 10, true,
1835                                 "osd_op");
1836         if (err < 0)
1837                 goto out_mempool;
1838         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1839                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1840                                 "osd_op_reply");
1841         if (err < 0)
1842                 goto out_msgpool;
1843
1844         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1845         if (IS_ERR(osdc->notify_wq)) {
1846                 err = PTR_ERR(osdc->notify_wq);
1847                 osdc->notify_wq = NULL;
1848                 goto out_msgpool;
1849         }
1850         return 0;
1851
1852 out_msgpool:
1853         ceph_msgpool_destroy(&osdc->msgpool_op);
1854 out_mempool:
1855         mempool_destroy(osdc->req_mempool);
1856 out:
1857         return err;
1858 }
1859 EXPORT_SYMBOL(ceph_osdc_init);
1860
1861 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1862 {
1863         flush_workqueue(osdc->notify_wq);
1864         destroy_workqueue(osdc->notify_wq);
1865         cancel_delayed_work_sync(&osdc->timeout_work);
1866         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1867         if (osdc->osdmap) {
1868                 ceph_osdmap_destroy(osdc->osdmap);
1869                 osdc->osdmap = NULL;
1870         }
1871         remove_all_osds(osdc);
1872         mempool_destroy(osdc->req_mempool);
1873         ceph_msgpool_destroy(&osdc->msgpool_op);
1874         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1875 }
1876 EXPORT_SYMBOL(ceph_osdc_stop);
1877
1878 /*
1879  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1880  * *plen.  Return number of bytes read, or error.
1881  */
1882 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1883                         struct ceph_vino vino, struct ceph_file_layout *layout,
1884                         u64 off, u64 *plen,
1885                         u32 truncate_seq, u64 truncate_size,
1886                         struct page **pages, int num_pages, int page_align)
1887 {
1888         struct ceph_osd_request *req;
1889         int rc = 0;
1890
1891         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1892              vino.snap, off, *plen);
1893         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1894                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1895                                     NULL, 0, truncate_seq, truncate_size, NULL,
1896                                     false, 1, page_align);
1897         if (IS_ERR(req))
1898                 return PTR_ERR(req);
1899
1900         /* it may be a short read due to an object boundary */
1901         req->r_pages = pages;
1902
1903         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1904              off, *plen, req->r_num_pages, page_align);
1905
1906         rc = ceph_osdc_start_request(osdc, req, false);
1907         if (!rc)
1908                 rc = ceph_osdc_wait_request(osdc, req);
1909
1910         ceph_osdc_put_request(req);
1911         dout("readpages result %d\n", rc);
1912         return rc;
1913 }
1914 EXPORT_SYMBOL(ceph_osdc_readpages);
1915
1916 /*
1917  * do a synchronous write on N pages
1918  */
1919 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1920                          struct ceph_file_layout *layout,
1921                          struct ceph_snap_context *snapc,
1922                          u64 off, u64 len,
1923                          u32 truncate_seq, u64 truncate_size,
1924                          struct timespec *mtime,
1925                          struct page **pages, int num_pages,
1926                          int flags, int do_sync, bool nofail)
1927 {
1928         struct ceph_osd_request *req;
1929         int rc = 0;
1930         int page_align = off & ~PAGE_MASK;
1931
1932         BUG_ON(vino.snap != CEPH_NOSNAP);
1933         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1934                                     CEPH_OSD_OP_WRITE,
1935                                     flags | CEPH_OSD_FLAG_ONDISK |
1936                                             CEPH_OSD_FLAG_WRITE,
1937                                     snapc, do_sync,
1938                                     truncate_seq, truncate_size, mtime,
1939                                     nofail, 1, page_align);
1940         if (IS_ERR(req))
1941                 return PTR_ERR(req);
1942
1943         /* it may be a short write due to an object boundary */
1944         req->r_pages = pages;
1945         dout("writepages %llu~%llu (%d pages)\n", off, len,
1946              req->r_num_pages);
1947
1948         rc = ceph_osdc_start_request(osdc, req, nofail);
1949         if (!rc)
1950                 rc = ceph_osdc_wait_request(osdc, req);
1951
1952         ceph_osdc_put_request(req);
1953         if (rc == 0)
1954                 rc = len;
1955         dout("writepages result %d\n", rc);
1956         return rc;
1957 }
1958 EXPORT_SYMBOL(ceph_osdc_writepages);
1959
1960 /*
1961  * handle incoming message
1962  */
1963 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1964 {
1965         struct ceph_osd *osd = con->private;
1966         struct ceph_osd_client *osdc;
1967         int type = le16_to_cpu(msg->hdr.type);
1968
1969         if (!osd)
1970                 goto out;
1971         osdc = osd->o_osdc;
1972
1973         switch (type) {
1974         case CEPH_MSG_OSD_MAP:
1975                 ceph_osdc_handle_map(osdc, msg);
1976                 break;
1977         case CEPH_MSG_OSD_OPREPLY:
1978                 handle_reply(osdc, msg, con);
1979                 break;
1980         case CEPH_MSG_WATCH_NOTIFY:
1981                 handle_watch_notify(osdc, msg);
1982                 break;
1983
1984         default:
1985                 pr_err("received unknown message type %d %s\n", type,
1986                        ceph_msg_type_name(type));
1987         }
1988 out:
1989         ceph_msg_put(msg);
1990 }
1991
1992 /*
1993  * lookup and return message for incoming reply.  set up reply message
1994  * pages.
1995  */
1996 static struct ceph_msg *get_reply(struct ceph_connection *con,
1997                                   struct ceph_msg_header *hdr,
1998                                   int *skip)
1999 {
2000         struct ceph_osd *osd = con->private;
2001         struct ceph_osd_client *osdc = osd->o_osdc;
2002         struct ceph_msg *m;
2003         struct ceph_osd_request *req;
2004         int front = le32_to_cpu(hdr->front_len);
2005         int data_len = le32_to_cpu(hdr->data_len);
2006         u64 tid;
2007
2008         tid = le64_to_cpu(hdr->tid);
2009         mutex_lock(&osdc->request_mutex);
2010         req = __lookup_request(osdc, tid);
2011         if (!req) {
2012                 *skip = 1;
2013                 m = NULL;
2014                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2015                      osd->o_osd);
2016                 goto out;
2017         }
2018
2019         if (req->r_con_filling_msg) {
2020                 dout("%s revoking msg %p from old con %p\n", __func__,
2021                      req->r_reply, req->r_con_filling_msg);
2022                 ceph_msg_revoke_incoming(req->r_reply);
2023                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2024                 req->r_con_filling_msg = NULL;
2025         }
2026
2027         if (front > req->r_reply->front.iov_len) {
2028                 pr_warning("get_reply front %d > preallocated %d\n",
2029                            front, (int)req->r_reply->front.iov_len);
2030                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2031                 if (!m)
2032                         goto out;
2033                 ceph_msg_put(req->r_reply);
2034                 req->r_reply = m;
2035         }
2036         m = ceph_msg_get(req->r_reply);
2037
2038         if (data_len > 0) {
2039                 int want = calc_pages_for(req->r_page_alignment, data_len);
2040
2041                 if (unlikely(req->r_num_pages < want)) {
2042                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2043                                    " had only %d pages ready\n", tid, data_len,
2044                                    want, req->r_num_pages);
2045                         *skip = 1;
2046                         ceph_msg_put(m);
2047                         m = NULL;
2048                         goto out;
2049                 }
2050                 m->pages = req->r_pages;
2051                 m->nr_pages = req->r_num_pages;
2052                 m->page_alignment = req->r_page_alignment;
2053 #ifdef CONFIG_BLOCK
2054                 m->bio = req->r_bio;
2055 #endif
2056         }
2057         *skip = 0;
2058         req->r_con_filling_msg = con->ops->get(con);
2059         dout("get_reply tid %lld %p\n", tid, m);
2060
2061 out:
2062         mutex_unlock(&osdc->request_mutex);
2063         return m;
2064
2065 }
2066
2067 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2068                                   struct ceph_msg_header *hdr,
2069                                   int *skip)
2070 {
2071         struct ceph_osd *osd = con->private;
2072         int type = le16_to_cpu(hdr->type);
2073         int front = le32_to_cpu(hdr->front_len);
2074
2075         *skip = 0;
2076         switch (type) {
2077         case CEPH_MSG_OSD_MAP:
2078         case CEPH_MSG_WATCH_NOTIFY:
2079                 return ceph_msg_new(type, front, GFP_NOFS, false);
2080         case CEPH_MSG_OSD_OPREPLY:
2081                 return get_reply(con, hdr, skip);
2082         default:
2083                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2084                         osd->o_osd);
2085                 *skip = 1;
2086                 return NULL;
2087         }
2088 }
2089
2090 /*
2091  * Wrappers to refcount containing ceph_osd struct
2092  */
2093 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2094 {
2095         struct ceph_osd *osd = con->private;
2096         if (get_osd(osd))
2097                 return con;
2098         return NULL;
2099 }
2100
2101 static void put_osd_con(struct ceph_connection *con)
2102 {
2103         struct ceph_osd *osd = con->private;
2104         put_osd(osd);
2105 }
2106
2107 /*
2108  * authentication
2109  */
2110 /*
2111  * Note: returned pointer is the address of a structure that's
2112  * managed separately.  Caller must *not* attempt to free it.
2113  */
2114 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2115                                         int *proto, int force_new)
2116 {
2117         struct ceph_osd *o = con->private;
2118         struct ceph_osd_client *osdc = o->o_osdc;
2119         struct ceph_auth_client *ac = osdc->client->monc.auth;
2120         struct ceph_auth_handshake *auth = &o->o_auth;
2121
2122         if (force_new && auth->authorizer) {
2123                 if (ac->ops && ac->ops->destroy_authorizer)
2124                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2125                 auth->authorizer = NULL;
2126         }
2127         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2128                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2129                                                         auth);
2130                 if (ret)
2131                         return ERR_PTR(ret);
2132         }
2133         *proto = ac->protocol;
2134
2135         return auth;
2136 }
2137
2138
2139 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2140 {
2141         struct ceph_osd *o = con->private;
2142         struct ceph_osd_client *osdc = o->o_osdc;
2143         struct ceph_auth_client *ac = osdc->client->monc.auth;
2144
2145         /*
2146          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2147          * XXX which do we do:  succeed or fail?
2148          */
2149         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2150 }
2151
2152 static int invalidate_authorizer(struct ceph_connection *con)
2153 {
2154         struct ceph_osd *o = con->private;
2155         struct ceph_osd_client *osdc = o->o_osdc;
2156         struct ceph_auth_client *ac = osdc->client->monc.auth;
2157
2158         if (ac->ops && ac->ops->invalidate_authorizer)
2159                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2160
2161         return ceph_monc_validate_auth(&osdc->client->monc);
2162 }
2163
2164 static const struct ceph_connection_operations osd_con_ops = {
2165         .get = get_osd_con,
2166         .put = put_osd_con,
2167         .dispatch = dispatch,
2168         .get_authorizer = get_authorizer,
2169         .verify_authorizer_reply = verify_authorizer_reply,
2170         .invalidate_authorizer = invalidate_authorizer,
2171         .alloc_msg = alloc_msg,
2172         .fault = osd_reset,
2173 };