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 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         snprintf(req->r_oid, sizeof(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         INIT_LIST_HEAD(&req->r_req_lru_item);
 221         req->r_flags = flags;
 222
 223         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 224
 225         /* create reply message */
 226         if (use_mempool)
 227                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 228         else
 229                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 230                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 231         if (!msg) {
 232                 ceph_osdc_put_request(req);
 233                 return NULL;
 234         }
 235         req->r_reply = msg;
 236
 237         /* allocate space for the trailing data */
 238         if (needs_trail) {
 239                 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
 240                 if (!req->r_trail) {
 241                         ceph_osdc_put_request(req);
 242                         return NULL;
 243                 }
 244                 ceph_pagelist_init(req->r_trail);
 245         }
 246         /* create request message; allow space for oid */
 247         msg_size += MAX_OBJ_NAME_SIZE;
 248         if (snapc)
 249                 msg_size += sizeof(u64) * snapc->num_snaps;
 250         if (use_mempool)
 251                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 252         else
 253                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 254         if (!msg) {
 255                 ceph_osdc_put_request(req);
 256                 return NULL;
 257         }
 258
 259         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
 260         memset(msg->front.iov_base, 0, msg->front.iov_len);
 261
 262         req->r_request = msg;
 263         req->r_pages = pages;
 264 #ifdef CONFIG_BLOCK
 265         if (bio) {
 266                 req->r_bio = bio;
 267                 bio_get(req->r_bio);
 268         }
 269 #endif
 270
 271         return req;
 272 }
 273 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 274
 275 static void osd_req_encode_op(struct ceph_osd_request *req,
 276                               struct ceph_osd_op *dst,
 277                               struct ceph_osd_req_op *src)
 278 {
 279         dst->op = cpu_to_le16(src->op);
 280
 281         switch (dst->op) {
 282         case CEPH_OSD_OP_READ:
 283         case CEPH_OSD_OP_WRITE:
 284                 dst->extent.offset =
 285                         cpu_to_le64(src->extent.offset);
 286                 dst->extent.length =
 287                         cpu_to_le64(src->extent.length);
 288                 dst->extent.truncate_size =
 289                         cpu_to_le64(src->extent.truncate_size);
 290                 dst->extent.truncate_seq =
 291                         cpu_to_le32(src->extent.truncate_seq);
 292                 break;
 293
 294         case CEPH_OSD_OP_GETXATTR:
 295         case CEPH_OSD_OP_SETXATTR:
 296         case CEPH_OSD_OP_CMPXATTR:
 297                 BUG_ON(!req->r_trail);
 298
 299                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 300                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 301                 dst->xattr.cmp_op = src->xattr.cmp_op;
 302                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 303                 ceph_pagelist_append(req->r_trail, src->xattr.name,
 304                                      src->xattr.name_len);
 305                 ceph_pagelist_append(req->r_trail, src->xattr.val,
 306                                      src->xattr.value_len);
 307                 break;
 308         case CEPH_OSD_OP_CALL:
 309                 BUG_ON(!req->r_trail);
 310
 311                 dst->cls.class_len = src->cls.class_len;
 312                 dst->cls.method_len = src->cls.method_len;
 313                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 314
 315                 ceph_pagelist_append(req->r_trail, src->cls.class_name,
 316                                      src->cls.class_len);
 317                 ceph_pagelist_append(req->r_trail, src->cls.method_name,
 318                                      src->cls.method_len);
 319                 ceph_pagelist_append(req->r_trail, src->cls.indata,
 320                                      src->cls.indata_len);
 321                 break;
 322         case CEPH_OSD_OP_ROLLBACK:
 323                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
 324                 break;
 325         case CEPH_OSD_OP_STARTSYNC:
 326                 break;
 327         case CEPH_OSD_OP_NOTIFY:
 328                 {
 329                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
 330                         __le32 timeout = cpu_to_le32(src->watch.timeout);
 331
 332                         BUG_ON(!req->r_trail);
 333
 334                         ceph_pagelist_append(req->r_trail,
 335                                                 &prot_ver, sizeof(prot_ver));
 336                         ceph_pagelist_append(req->r_trail,
 337                                                 &timeout, sizeof(timeout));
 338                 }
 339         case CEPH_OSD_OP_NOTIFY_ACK:
 340         case CEPH_OSD_OP_WATCH:
 341                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 342                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 343                 dst->watch.flag = src->watch.flag;
 344                 break;
 345         default:
 346                 pr_err("unrecognized osd opcode %d\n", dst->op);
 347                 WARN_ON(1);
 348                 break;
 349         }
 350         dst->payload_len = cpu_to_le32(src->payload_len);
 351 }
 352
 353 /*
 354  * build new request AND message
 355  *
 356  */
 357 void ceph_osdc_build_request(struct ceph_osd_request *req,
 358                              u64 off, u64 *plen,
 359                              struct ceph_osd_req_op *src_ops,
 360                              struct ceph_snap_context *snapc,
 361                              struct timespec *mtime,
 362                              const char *oid,
 363                              int oid_len)
 364 {
 365         struct ceph_msg *msg = req->r_request;
 366         struct ceph_osd_request_head *head;
 367         struct ceph_osd_req_op *src_op;
 368         struct ceph_osd_op *op;
 369         void *p;
 370         int num_op = get_num_ops(src_ops, NULL);
 371         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 372         int flags = req->r_flags;
 373         u64 data_len = 0;
 374         int i;
 375
 376         head = msg->front.iov_base;
 377         op = (void *)(head + 1);
 378         p = (void *)(op + num_op);
 379
 380         req->r_snapc = ceph_get_snap_context(snapc);
 381
 382         head->client_inc = cpu_to_le32(1); /* always, for now. */
 383         head->flags = cpu_to_le32(flags);
 384         if (flags & CEPH_OSD_FLAG_WRITE)
 385                 ceph_encode_timespec(&head->mtime, mtime);
 386         head->num_ops = cpu_to_le16(num_op);
 387
 388
 389         /* fill in oid */
 390         head->object_len = cpu_to_le32(oid_len);
 391         memcpy(p, oid, oid_len);
 392         p += oid_len;
 393
 394         src_op = src_ops;
 395         while (src_op->op) {
 396                 osd_req_encode_op(req, op, src_op);
 397                 src_op++;
 398                 op++;
 399         }
 400
 401         if (req->r_trail)
 402                 data_len += req->r_trail->length;
 403
 404         if (snapc) {
 405                 head->snap_seq = cpu_to_le64(snapc->seq);
 406                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 407                 for (i = 0; i < snapc->num_snaps; i++) {
 408                         put_unaligned_le64(snapc->snaps[i], p);
 409                         p += sizeof(u64);
 410                 }
 411         }
 412
 413         if (flags & CEPH_OSD_FLAG_WRITE) {
 414                 req->r_request->hdr.data_off = cpu_to_le16(off);
 415                 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
 416         } else if (data_len) {
 417                 req->r_request->hdr.data_off = 0;
 418                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 419         }
 420
 421         req->r_request->page_alignment = req->r_page_alignment;
 422
 423         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 424         msg_size = p - msg->front.iov_base;
 425         msg->front.iov_len = msg_size;
 426         msg->hdr.front_len = cpu_to_le32(msg_size);
 427         return;
 428 }
 429 EXPORT_SYMBOL(ceph_osdc_build_request);
 430
 431 /*
 432  * build new request AND message, calculate layout, and adjust file
 433  * extent as needed.
 434  *
 435  * if the file was recently truncated, we include information about its
 436  * old and new size so that the object can be updated appropriately.  (we
 437  * avoid synchronously deleting truncated objects because it's slow.)
 438  *
 439  * if @do_sync, include a 'startsync' command so that the osd will flush
 440  * data quickly.
 441  */
 442 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 443                                                struct ceph_file_layout *layout,
 444                                                struct ceph_vino vino,
 445                                                u64 off, u64 *plen,
 446                                                int opcode, int flags,
 447                                                struct ceph_snap_context *snapc,
 448                                                int do_sync,
 449                                                u32 truncate_seq,
 450                                                u64 truncate_size,
 451                                                struct timespec *mtime,
 452                                                bool use_mempool, int num_reply,
 453                                                int page_align)
 454 {
 455         struct ceph_osd_req_op ops[3];
 456         struct ceph_osd_request *req;
 457
 458         ops[0].op = opcode;
 459         ops[0].extent.truncate_seq = truncate_seq;
 460         ops[0].extent.truncate_size = truncate_size;
 461         ops[0].payload_len = 0;
 462
 463         if (do_sync) {
 464                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 465                 ops[1].payload_len = 0;
 466                 ops[2].op = 0;
 467         } else
 468                 ops[1].op = 0;
 469
 470         req = ceph_osdc_alloc_request(osdc, flags,
 471                                          snapc, ops,
 472                                          use_mempool,
 473                                          GFP_NOFS, NULL, NULL);
 474         if (!req)
 475                 return NULL;
 476
 477         /* calculate max write size */
 478         calc_layout(osdc, vino, layout, off, plen, req, ops);
 479         req->r_file_layout = *layout;  /* keep a copy */
 480
 481         /* in case it differs from natural (file) alignment that
 482            calc_layout filled in for us */
 483         req->r_num_pages = calc_pages_for(page_align, *plen);
 484         req->r_page_alignment = page_align;
 485
 486         ceph_osdc_build_request(req, off, plen, ops,
 487                                 snapc,
 488                                 mtime,
 489                                 req->r_oid, req->r_oid_len);
 490
 491         return req;
 492 }
 493 EXPORT_SYMBOL(ceph_osdc_new_request);
 494
 495 /*
 496  * We keep osd requests in an rbtree, sorted by ->r_tid.
 497  */
 498 static void __insert_request(struct ceph_osd_client *osdc,
 499                              struct ceph_osd_request *new)
 500 {
 501         struct rb_node **p = &osdc->requests.rb_node;
 502         struct rb_node *parent = NULL;
 503         struct ceph_osd_request *req = NULL;
 504
 505         while (*p) {
 506                 parent = *p;
 507                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 508                 if (new->r_tid < req->r_tid)
 509                         p = &(*p)->rb_left;
 510                 else if (new->r_tid > req->r_tid)
 511                         p = &(*p)->rb_right;
 512                 else
 513                         BUG();
 514         }
 515
 516         rb_link_node(&new->r_node, parent, p);
 517         rb_insert_color(&new->r_node, &osdc->requests);
 518 }
 519
 520 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 521                                                  u64 tid)
 522 {
 523         struct ceph_osd_request *req;
 524         struct rb_node *n = osdc->requests.rb_node;
 525
 526         while (n) {
 527                 req = rb_entry(n, struct ceph_osd_request, r_node);
 528                 if (tid < req->r_tid)
 529                         n = n->rb_left;
 530                 else if (tid > req->r_tid)
 531                         n = n->rb_right;
 532                 else
 533                         return req;
 534         }
 535         return NULL;
 536 }
 537
 538 static struct ceph_osd_request *
 539 __lookup_request_ge(struct ceph_osd_client *osdc,
 540                     u64 tid)
 541 {
 542         struct ceph_osd_request *req;
 543         struct rb_node *n = osdc->requests.rb_node;
 544
 545         while (n) {
 546                 req = rb_entry(n, struct ceph_osd_request, r_node);
 547                 if (tid < req->r_tid) {
 548                         if (!n->rb_left)
 549                                 return req;
 550                         n = n->rb_left;
 551                 } else if (tid > req->r_tid) {
 552                         n = n->rb_right;
 553                 } else {
 554                         return req;
 555                 }
 556         }
 557         return NULL;
 558 }
 559
 560 /*
 561  * Resubmit requests pending on the given osd.
 562  */
 563 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 564                                 struct ceph_osd *osd)
 565 {
 566         struct ceph_osd_request *req, *nreq;
 567         int err;
 568
 569         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 570         err = __reset_osd(osdc, osd);
 571         if (err == -EAGAIN)
 572                 return;
 573
 574         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 575                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 576                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 577                      osd->o_osd);
 578                 if (!req->r_linger)
 579                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 580         }
 581
 582         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 583                                  r_linger_osd) {
 584                 /*
 585                  * reregister request prior to unregistering linger so
 586                  * that r_osd is preserved.
 587                  */
 588                 BUG_ON(!list_empty(&req->r_req_lru_item));
 589                 __register_request(osdc, req);
 590                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 591                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 592                 __unregister_linger_request(osdc, req);
 593                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 594                      osd->o_osd);
 595         }
 596 }
 597
 598 static void kick_osd_requests(struct ceph_osd_client *osdc,
 599                               struct ceph_osd *kickosd)
 600 {
 601         mutex_lock(&osdc->request_mutex);
 602         __kick_osd_requests(osdc, kickosd);
 603         mutex_unlock(&osdc->request_mutex);
 604 }
 605
 606 /*
 607  * If the osd connection drops, we need to resubmit all requests.
 608  */
 609 static void osd_reset(struct ceph_connection *con)
 610 {
 611         struct ceph_osd *osd = con->private;
 612         struct ceph_osd_client *osdc;
 613
 614         if (!osd)
 615                 return;
 616         dout("osd_reset osd%d\n", osd->o_osd);
 617         osdc = osd->o_osdc;
 618         down_read(&osdc->map_sem);
 619         kick_osd_requests(osdc, osd);
 620         send_queued(osdc);
 621         up_read(&osdc->map_sem);
 622 }
 623
 624 /*
 625  * Track open sessions with osds.
 626  */
 627 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
 628 {
 629         struct ceph_osd *osd;
 630
 631         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 632         if (!osd)
 633                 return NULL;
 634
 635         atomic_set(&osd->o_ref, 1);
 636         osd->o_osdc = osdc;
 637         INIT_LIST_HEAD(&osd->o_requests);
 638         INIT_LIST_HEAD(&osd->o_linger_requests);
 639         INIT_LIST_HEAD(&osd->o_osd_lru);
 640         osd->o_incarnation = 1;
 641
 642         ceph_con_init(osdc->client->msgr, &osd->o_con);
 643         osd->o_con.private = osd;
 644         osd->o_con.ops = &osd_con_ops;
 645         osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
 646
 647         INIT_LIST_HEAD(&osd->o_keepalive_item);
 648         return osd;
 649 }
 650
 651 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 652 {
 653         if (atomic_inc_not_zero(&osd->o_ref)) {
 654                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 655                      atomic_read(&osd->o_ref));
 656                 return osd;
 657         } else {
 658                 dout("get_osd %p FAIL\n", osd);
 659                 return NULL;
 660         }
 661 }
 662
 663 static void put_osd(struct ceph_osd *osd)
 664 {
 665         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 666              atomic_read(&osd->o_ref) - 1);
 667         if (atomic_dec_and_test(&osd->o_ref)) {
 668                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 669
 670                 if (osd->o_authorizer)
 671                         ac->ops->destroy_authorizer(ac, osd->o_authorizer);
 672                 kfree(osd);
 673         }
 674 }
 675
 676 /*
 677  * remove an osd from our map
 678  */
 679 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 680 {
 681         dout("__remove_osd %p\n", osd);
 682         BUG_ON(!list_empty(&osd->o_requests));
 683         rb_erase(&osd->o_node, &osdc->osds);
 684         list_del_init(&osd->o_osd_lru);
 685         ceph_con_close(&osd->o_con);
 686         put_osd(osd);
 687 }
 688
 689 static void remove_all_osds(struct ceph_osd_client *osdc)
 690 {
 691         dout("__remove_old_osds %p\n", osdc);
 692         mutex_lock(&osdc->request_mutex);
 693         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 694                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 695                                                 struct ceph_osd, o_node);
 696                 __remove_osd(osdc, osd);
 697         }
 698         mutex_unlock(&osdc->request_mutex);
 699 }
 700
 701 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 702                               struct ceph_osd *osd)
 703 {
 704         dout("__move_osd_to_lru %p\n", osd);
 705         BUG_ON(!list_empty(&osd->o_osd_lru));
 706         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 707         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 708 }
 709
 710 static void __remove_osd_from_lru(struct ceph_osd *osd)
 711 {
 712         dout("__remove_osd_from_lru %p\n", osd);
 713         if (!list_empty(&osd->o_osd_lru))
 714                 list_del_init(&osd->o_osd_lru);
 715 }
 716
 717 static void remove_old_osds(struct ceph_osd_client *osdc)
 718 {
 719         struct ceph_osd *osd, *nosd;
 720
 721         dout("__remove_old_osds %p\n", osdc);
 722         mutex_lock(&osdc->request_mutex);
 723         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 724                 if (time_before(jiffies, osd->lru_ttl))
 725                         break;
 726                 __remove_osd(osdc, osd);
 727         }
 728         mutex_unlock(&osdc->request_mutex);
 729 }
 730
 731 /*
 732  * reset osd connect
 733  */
 734 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 735 {
 736         struct ceph_osd_request *req;
 737         int ret = 0;
 738
 739         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 740         if (list_empty(&osd->o_requests) &&
 741             list_empty(&osd->o_linger_requests)) {
 742                 __remove_osd(osdc, osd);
 743         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 744                           &osd->o_con.peer_addr,
 745                           sizeof(osd->o_con.peer_addr)) == 0 &&
 746                    !ceph_con_opened(&osd->o_con)) {
 747                 dout(" osd addr hasn't changed and connection never opened,"
 748                      " letting msgr retry");
 749                 /* touch each r_stamp for handle_timeout()'s benfit */
 750                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 751                         req->r_stamp = jiffies;
 752                 ret = -EAGAIN;
 753         } else {
 754                 ceph_con_close(&osd->o_con);
 755                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
 756                 osd->o_incarnation++;
 757         }
 758         return ret;
 759 }
 760
 761 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 762 {
 763         struct rb_node **p = &osdc->osds.rb_node;
 764         struct rb_node *parent = NULL;
 765         struct ceph_osd *osd = NULL;
 766
 767         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 768         while (*p) {
 769                 parent = *p;
 770                 osd = rb_entry(parent, struct ceph_osd, o_node);
 771                 if (new->o_osd < osd->o_osd)
 772                         p = &(*p)->rb_left;
 773                 else if (new->o_osd > osd->o_osd)
 774                         p = &(*p)->rb_right;
 775                 else
 776                         BUG();
 777         }
 778
 779         rb_link_node(&new->o_node, parent, p);
 780         rb_insert_color(&new->o_node, &osdc->osds);
 781 }
 782
 783 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 784 {
 785         struct ceph_osd *osd;
 786         struct rb_node *n = osdc->osds.rb_node;
 787
 788         while (n) {
 789                 osd = rb_entry(n, struct ceph_osd, o_node);
 790                 if (o < osd->o_osd)
 791                         n = n->rb_left;
 792                 else if (o > osd->o_osd)
 793                         n = n->rb_right;
 794                 else
 795                         return osd;
 796         }
 797         return NULL;
 798 }
 799
 800 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 801 {
 802         schedule_delayed_work(&osdc->timeout_work,
 803                         osdc->client->options->osd_keepalive_timeout * HZ);
 804 }
 805
 806 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 807 {
 808         cancel_delayed_work(&osdc->timeout_work);
 809 }
 810
 811 /*
 812  * Register request, assign tid.  If this is the first request, set up
 813  * the timeout event.
 814  */
 815 static void __register_request(struct ceph_osd_client *osdc,
 816                                struct ceph_osd_request *req)
 817 {
 818         req->r_tid = ++osdc->last_tid;
 819         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 820         dout("__register_request %p tid %lld\n", req, req->r_tid);
 821         __insert_request(osdc, req);
 822         ceph_osdc_get_request(req);
 823         osdc->num_requests++;
 824         if (osdc->num_requests == 1) {
 825                 dout(" first request, scheduling timeout\n");
 826                 __schedule_osd_timeout(osdc);
 827         }
 828 }
 829
 830 static void register_request(struct ceph_osd_client *osdc,
 831                              struct ceph_osd_request *req)
 832 {
 833         mutex_lock(&osdc->request_mutex);
 834         __register_request(osdc, req);
 835         mutex_unlock(&osdc->request_mutex);
 836 }
 837
 838 /*
 839  * called under osdc->request_mutex
 840  */
 841 static void __unregister_request(struct ceph_osd_client *osdc,
 842                                  struct ceph_osd_request *req)
 843 {
 844         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 845         rb_erase(&req->r_node, &osdc->requests);
 846         osdc->num_requests--;
 847
 848         if (req->r_osd) {
 849                 /* make sure the original request isn't in flight. */
 850                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 851
 852                 list_del_init(&req->r_osd_item);
 853                 if (list_empty(&req->r_osd->o_requests) &&
 854                     list_empty(&req->r_osd->o_linger_requests)) {
 855                         dout("moving osd to %p lru\n", req->r_osd);
 856                         __move_osd_to_lru(osdc, req->r_osd);
 857                 }
 858                 if (list_empty(&req->r_linger_item))
 859                         req->r_osd = NULL;
 860         }
 861
 862         ceph_osdc_put_request(req);
 863
 864         list_del_init(&req->r_req_lru_item);
 865         if (osdc->num_requests == 0) {
 866                 dout(" no requests, canceling timeout\n");
 867                 __cancel_osd_timeout(osdc);
 868         }
 869 }
 870
 871 /*
 872  * Cancel a previously queued request message
 873  */
 874 static void __cancel_request(struct ceph_osd_request *req)
 875 {
 876         if (req->r_sent && req->r_osd) {
 877                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 878                 req->r_sent = 0;
 879         }
 880 }
 881
 882 static void __register_linger_request(struct ceph_osd_client *osdc,
 883                                     struct ceph_osd_request *req)
 884 {
 885         dout("__register_linger_request %p\n", req);
 886         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 887         list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests);
 888 }
 889
 890 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 891                                         struct ceph_osd_request *req)
 892 {
 893         dout("__unregister_linger_request %p\n", req);
 894         if (req->r_osd) {
 895                 list_del_init(&req->r_linger_item);
 896                 list_del_init(&req->r_linger_osd);
 897
 898                 if (list_empty(&req->r_osd->o_requests) &&
 899                     list_empty(&req->r_osd->o_linger_requests)) {
 900                         dout("moving osd to %p lru\n", req->r_osd);
 901                         __move_osd_to_lru(osdc, req->r_osd);
 902                 }
 903                 if (list_empty(&req->r_osd_item))
 904                         req->r_osd = NULL;
 905         }
 906 }
 907
 908 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 909                                          struct ceph_osd_request *req)
 910 {
 911         mutex_lock(&osdc->request_mutex);
 912         if (req->r_linger) {
 913                 __unregister_linger_request(osdc, req);
 914                 ceph_osdc_put_request(req);
 915         }
 916         mutex_unlock(&osdc->request_mutex);
 917 }
 918 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 919
 920 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 921                                   struct ceph_osd_request *req)
 922 {
 923         if (!req->r_linger) {
 924                 dout("set_request_linger %p\n", req);
 925                 req->r_linger = 1;
 926                 /*
 927                  * caller is now responsible for calling
 928                  * unregister_linger_request
 929                  */
 930                 ceph_osdc_get_request(req);
 931         }
 932 }
 933 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 934
 935 /*
 936  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 937  * (as needed), and set the request r_osd appropriately.  If there is
 938  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 939  * (unsent, homeless) or leave on in-flight lru.
 940  *
 941  * Return 0 if unchanged, 1 if changed, or negative on error.
 942  *
 943  * Caller should hold map_sem for read and request_mutex.
 944  */
 945 static int __map_request(struct ceph_osd_client *osdc,
 946                          struct ceph_osd_request *req, int force_resend)
 947 {
 948         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 949         struct ceph_pg pgid;
 950         int acting[CEPH_PG_MAX_SIZE];
 951         int o = -1, num = 0;
 952         int err;
 953
 954         dout("map_request %p tid %lld\n", req, req->r_tid);
 955         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 956                                       &req->r_file_layout, osdc->osdmap);
 957         if (err) {
 958                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 959                 return err;
 960         }
 961         pgid = reqhead->layout.ol_pgid;
 962         req->r_pgid = pgid;
 963
 964         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 965         if (err > 0) {
 966                 o = acting[0];
 967                 num = err;
 968         }
 969
 970         if ((!force_resend &&
 971              req->r_osd && req->r_osd->o_osd == o &&
 972              req->r_sent >= req->r_osd->o_incarnation &&
 973              req->r_num_pg_osds == num &&
 974              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 975             (req->r_osd == NULL && o == -1))
 976                 return 0;  /* no change */
 977
 978         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 979              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 980              req->r_osd ? req->r_osd->o_osd : -1);
 981
 982         /* record full pg acting set */
 983         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 984         req->r_num_pg_osds = num;
 985
 986         if (req->r_osd) {
 987                 __cancel_request(req);
 988                 list_del_init(&req->r_osd_item);
 989                 req->r_osd = NULL;
 990         }
 991
 992         req->r_osd = __lookup_osd(osdc, o);
 993         if (!req->r_osd && o >= 0) {
 994                 err = -ENOMEM;
 995                 req->r_osd = create_osd(osdc);
 996                 if (!req->r_osd) {
 997                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
 998                         goto out;
 999                 }
1000
1001                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1002                 req->r_osd->o_osd = o;
1003                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
1004                 __insert_osd(osdc, req->r_osd);
1005
1006                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
1007         }
1008
1009         if (req->r_osd) {
1010                 __remove_osd_from_lru(req->r_osd);
1011                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1012                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1013         } else {
1014                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1015         }
1016         err = 1;   /* osd or pg changed */
1017
1018 out:
1019         return err;
1020 }
1021
1022 /*
1023  * caller should hold map_sem (for read) and request_mutex
1024  */
1025 static void __send_request(struct ceph_osd_client *osdc,
1026                            struct ceph_osd_request *req)
1027 {
1028         struct ceph_osd_request_head *reqhead;
1029
1030         dout("send_request %p tid %llu to osd%d flags %d\n",
1031              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1032
1033         reqhead = req->r_request->front.iov_base;
1034         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1035         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1036         reqhead->reassert_version = req->r_reassert_version;
1037
1038         req->r_stamp = jiffies;
1039         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1040
1041         ceph_msg_get(req->r_request); /* send consumes a ref */
1042         ceph_con_send(&req->r_osd->o_con, req->r_request);
1043         req->r_sent = req->r_osd->o_incarnation;
1044 }
1045
1046 /*
1047  * Send any requests in the queue (req_unsent).
1048  */
1049 static void send_queued(struct ceph_osd_client *osdc)
1050 {
1051         struct ceph_osd_request *req, *tmp;
1052
1053         dout("send_queued\n");
1054         mutex_lock(&osdc->request_mutex);
1055         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1056                 __send_request(osdc, req);
1057         }
1058         mutex_unlock(&osdc->request_mutex);
1059 }
1060
1061 /*
1062  * Timeout callback, called every N seconds when 1 or more osd
1063  * requests has been active for more than N seconds.  When this
1064  * happens, we ping all OSDs with requests who have timed out to
1065  * ensure any communications channel reset is detected.  Reset the
1066  * request timeouts another N seconds in the future as we go.
1067  * Reschedule the timeout event another N seconds in future (unless
1068  * there are no open requests).
1069  */
1070 static void handle_timeout(struct work_struct *work)
1071 {
1072         struct ceph_osd_client *osdc =
1073                 container_of(work, struct ceph_osd_client, timeout_work.work);
1074         struct ceph_osd_request *req, *last_req = NULL;
1075         struct ceph_osd *osd;
1076         unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1077         unsigned long keepalive =
1078                 osdc->client->options->osd_keepalive_timeout * HZ;
1079         unsigned long last_stamp = 0;
1080         struct list_head slow_osds;
1081         dout("timeout\n");
1082         down_read(&osdc->map_sem);
1083
1084         ceph_monc_request_next_osdmap(&osdc->client->monc);
1085
1086         mutex_lock(&osdc->request_mutex);
1087
1088         /*
1089          * reset osds that appear to be _really_ unresponsive.  this
1090          * is a failsafe measure.. we really shouldn't be getting to
1091          * this point if the system is working properly.  the monitors
1092          * should mark the osd as failed and we should find out about
1093          * it from an updated osd map.
1094          */
1095         while (timeout && !list_empty(&osdc->req_lru)) {
1096                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1097                                  r_req_lru_item);
1098
1099                 /* hasn't been long enough since we sent it? */
1100                 if (time_before(jiffies, req->r_stamp + timeout))
1101                         break;
1102
1103                 /* hasn't been long enough since it was acked? */
1104                 if (req->r_request->ack_stamp == 0 ||
1105                     time_before(jiffies, req->r_request->ack_stamp + timeout))
1106                         break;
1107
1108                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
1109                 last_req = req;
1110                 last_stamp = req->r_stamp;
1111
1112                 osd = req->r_osd;
1113                 BUG_ON(!osd);
1114                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1115                            req->r_tid, osd->o_osd);
1116                 __kick_osd_requests(osdc, osd);
1117         }
1118
1119         /*
1120          * ping osds that are a bit slow.  this ensures that if there
1121          * is a break in the TCP connection we will notice, and reopen
1122          * a connection with that osd (from the fault callback).
1123          */
1124         INIT_LIST_HEAD(&slow_osds);
1125         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1126                 if (time_before(jiffies, req->r_stamp + keepalive))
1127                         break;
1128
1129                 osd = req->r_osd;
1130                 BUG_ON(!osd);
1131                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1132                      req->r_tid, osd->o_osd);
1133                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1134         }
1135         while (!list_empty(&slow_osds)) {
1136                 osd = list_entry(slow_osds.next, struct ceph_osd,
1137                                  o_keepalive_item);
1138                 list_del_init(&osd->o_keepalive_item);
1139                 ceph_con_keepalive(&osd->o_con);
1140         }
1141
1142         __schedule_osd_timeout(osdc);
1143         mutex_unlock(&osdc->request_mutex);
1144         send_queued(osdc);
1145         up_read(&osdc->map_sem);
1146 }
1147
1148 static void handle_osds_timeout(struct work_struct *work)
1149 {
1150         struct ceph_osd_client *osdc =
1151                 container_of(work, struct ceph_osd_client,
1152                              osds_timeout_work.work);
1153         unsigned long delay =
1154                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1155
1156         dout("osds timeout\n");
1157         down_read(&osdc->map_sem);
1158         remove_old_osds(osdc);
1159         up_read(&osdc->map_sem);
1160
1161         schedule_delayed_work(&osdc->osds_timeout_work,
1162                               round_jiffies_relative(delay));
1163 }
1164
1165 static void complete_request(struct ceph_osd_request *req)
1166 {
1167         if (req->r_safe_callback)
1168                 req->r_safe_callback(req, NULL);
1169         complete_all(&req->r_safe_completion);  /* fsync waiter */
1170 }
1171
1172 /*
1173  * handle osd op reply.  either call the callback if it is specified,
1174  * or do the completion to wake up the waiting thread.
1175  */
1176 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1177                          struct ceph_connection *con)
1178 {
1179         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1180         struct ceph_osd_request *req;
1181         u64 tid;
1182         int numops, object_len, flags;
1183         s32 result;
1184
1185         tid = le64_to_cpu(msg->hdr.tid);
1186         if (msg->front.iov_len < sizeof(*rhead))
1187                 goto bad;
1188         numops = le32_to_cpu(rhead->num_ops);
1189         object_len = le32_to_cpu(rhead->object_len);
1190         result = le32_to_cpu(rhead->result);
1191         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1192             numops * sizeof(struct ceph_osd_op))
1193                 goto bad;
1194         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1195         /* lookup */
1196         mutex_lock(&osdc->request_mutex);
1197         req = __lookup_request(osdc, tid);
1198         if (req == NULL) {
1199                 dout("handle_reply tid %llu dne\n", tid);
1200                 mutex_unlock(&osdc->request_mutex);
1201                 return;
1202         }
1203         ceph_osdc_get_request(req);
1204         flags = le32_to_cpu(rhead->flags);
1205
1206         /*
1207          * if this connection filled our message, drop our reference now, to
1208          * avoid a (safe but slower) revoke later.
1209          */
1210         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1211                 dout(" dropping con_filling_msg ref %p\n", con);
1212                 req->r_con_filling_msg = NULL;
1213                 ceph_con_put(con);
1214         }
1215
1216         if (!req->r_got_reply) {
1217                 unsigned bytes;
1218
1219                 req->r_result = le32_to_cpu(rhead->result);
1220                 bytes = le32_to_cpu(msg->hdr.data_len);
1221                 dout("handle_reply result %d bytes %d\n", req->r_result,
1222                      bytes);
1223                 if (req->r_result == 0)
1224                         req->r_result = bytes;
1225
1226                 /* in case this is a write and we need to replay, */
1227                 req->r_reassert_version = rhead->reassert_version;
1228
1229                 req->r_got_reply = 1;
1230         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1231                 dout("handle_reply tid %llu dup ack\n", tid);
1232                 mutex_unlock(&osdc->request_mutex);
1233                 goto done;
1234         }
1235
1236         dout("handle_reply tid %llu flags %d\n", tid, flags);
1237
1238         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1239                 __register_linger_request(osdc, req);
1240
1241         /* either this is a read, or we got the safe response */
1242         if (result < 0 ||
1243             (flags & CEPH_OSD_FLAG_ONDISK) ||
1244             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1245                 __unregister_request(osdc, req);
1246
1247         mutex_unlock(&osdc->request_mutex);
1248
1249         if (req->r_callback)
1250                 req->r_callback(req, msg);
1251         else
1252                 complete_all(&req->r_completion);
1253
1254         if (flags & CEPH_OSD_FLAG_ONDISK)
1255                 complete_request(req);
1256
1257 done:
1258         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1259         ceph_osdc_put_request(req);
1260         return;
1261
1262 bad:
1263         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1264                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1265                (int)sizeof(*rhead));
1266         ceph_msg_dump(msg);
1267 }
1268
1269 static void reset_changed_osds(struct ceph_osd_client *osdc)
1270 {
1271         struct rb_node *p, *n;
1272
1273         for (p = rb_first(&osdc->osds); p; p = n) {
1274                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1275
1276                 n = rb_next(p);
1277                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1278                     memcmp(&osd->o_con.peer_addr,
1279                            ceph_osd_addr(osdc->osdmap,
1280                                          osd->o_osd),
1281                            sizeof(struct ceph_entity_addr)) != 0)
1282                         __reset_osd(osdc, osd);
1283         }
1284 }
1285
1286 /*
1287  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1288  * no osd, request a new map.
1289  *
1290  * Caller should hold map_sem for read and request_mutex.
1291  */
1292 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1293 {
1294         struct ceph_osd_request *req, *nreq;
1295         struct rb_node *p;
1296         int needmap = 0;
1297         int err;
1298
1299         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1300         mutex_lock(&osdc->request_mutex);
1301         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1302                 req = rb_entry(p, struct ceph_osd_request, r_node);
1303                 err = __map_request(osdc, req, force_resend);
1304                 if (err < 0)
1305                         continue;  /* error */
1306                 if (req->r_osd == NULL) {
1307                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1308                         needmap++;  /* request a newer map */
1309                 } else if (err > 0) {
1310                         dout("%p tid %llu requeued on osd%d\n", req, req->r_tid,
1311                              req->r_osd ? req->r_osd->o_osd : -1);
1312                         if (!req->r_linger)
1313                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1314                 }
1315         }
1316
1317         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1318                                  r_linger_item) {
1319                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1320
1321                 err = __map_request(osdc, req, force_resend);
1322                 if (err == 0)
1323                         continue;  /* no change and no osd was specified */
1324                 if (err < 0)
1325                         continue;  /* hrm! */
1326                 if (req->r_osd == NULL) {
1327                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1328                         needmap++;  /* request a newer map */
1329                         continue;
1330                 }
1331
1332                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1333                      req->r_osd ? req->r_osd->o_osd : -1);
1334                 __unregister_linger_request(osdc, req);
1335                 __register_request(osdc, req);
1336         }
1337         mutex_unlock(&osdc->request_mutex);
1338
1339         if (needmap) {
1340                 dout("%d requests for down osds, need new map\n", needmap);
1341                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1342         }
1343 }
1344
1345
1346 /*
1347  * Process updated osd map.
1348  *
1349  * The message contains any number of incremental and full maps, normally
1350  * indicating some sort of topology change in the cluster.  Kick requests
1351  * off to different OSDs as needed.
1352  */
1353 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1354 {
1355         void *p, *end, *next;
1356         u32 nr_maps, maplen;
1357         u32 epoch;
1358         struct ceph_osdmap *newmap = NULL, *oldmap;
1359         int err;
1360         struct ceph_fsid fsid;
1361
1362         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1363         p = msg->front.iov_base;
1364         end = p + msg->front.iov_len;
1365
1366         /* verify fsid */
1367         ceph_decode_need(&p, end, sizeof(fsid), bad);
1368         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1369         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1370                 return;
1371
1372         down_write(&osdc->map_sem);
1373
1374         /* incremental maps */
1375         ceph_decode_32_safe(&p, end, nr_maps, bad);
1376         dout(" %d inc maps\n", nr_maps);
1377         while (nr_maps > 0) {
1378                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1379                 epoch = ceph_decode_32(&p);
1380                 maplen = ceph_decode_32(&p);
1381                 ceph_decode_need(&p, end, maplen, bad);
1382                 next = p + maplen;
1383                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1384                         dout("applying incremental map %u len %d\n",
1385                              epoch, maplen);
1386                         newmap = osdmap_apply_incremental(&p, next,
1387                                                           osdc->osdmap,
1388                                                           osdc->client->msgr);
1389                         if (IS_ERR(newmap)) {
1390                                 err = PTR_ERR(newmap);
1391                                 goto bad;
1392                         }
1393                         BUG_ON(!newmap);
1394                         if (newmap != osdc->osdmap) {
1395                                 ceph_osdmap_destroy(osdc->osdmap);
1396                                 osdc->osdmap = newmap;
1397                         }
1398                         kick_requests(osdc, 0);
1399                         reset_changed_osds(osdc);
1400                 } else {
1401                         dout("ignoring incremental map %u len %d\n",
1402                              epoch, maplen);
1403                 }
1404                 p = next;
1405                 nr_maps--;
1406         }
1407         if (newmap)
1408                 goto done;
1409
1410         /* full maps */
1411         ceph_decode_32_safe(&p, end, nr_maps, bad);
1412         dout(" %d full maps\n", nr_maps);
1413         while (nr_maps) {
1414                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1415                 epoch = ceph_decode_32(&p);
1416                 maplen = ceph_decode_32(&p);
1417                 ceph_decode_need(&p, end, maplen, bad);
1418                 if (nr_maps > 1) {
1419                         dout("skipping non-latest full map %u len %d\n",
1420                              epoch, maplen);
1421                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1422                         dout("skipping full map %u len %d, "
1423                              "older than our %u\n", epoch, maplen,
1424                              osdc->osdmap->epoch);
1425                 } else {
1426                         int skipped_map = 0;
1427
1428                         dout("taking full map %u len %d\n", epoch, maplen);
1429                         newmap = osdmap_decode(&p, p+maplen);
1430                         if (IS_ERR(newmap)) {
1431                                 err = PTR_ERR(newmap);
1432                                 goto bad;
1433                         }
1434                         BUG_ON(!newmap);
1435                         oldmap = osdc->osdmap;
1436                         osdc->osdmap = newmap;
1437                         if (oldmap) {
1438                                 if (oldmap->epoch + 1 < newmap->epoch)
1439                                         skipped_map = 1;
1440                                 ceph_osdmap_destroy(oldmap);
1441                         }
1442                         kick_requests(osdc, skipped_map);
1443                 }
1444                 p += maplen;
1445                 nr_maps--;
1446         }
1447
1448 done:
1449         downgrade_write(&osdc->map_sem);
1450         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1451
1452         /*
1453          * subscribe to subsequent osdmap updates if full to ensure
1454          * we find out when we are no longer full and stop returning
1455          * ENOSPC.
1456          */
1457         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1458                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1459
1460         send_queued(osdc);
1461         up_read(&osdc->map_sem);
1462         wake_up_all(&osdc->client->auth_wq);
1463         return;
1464
1465 bad:
1466         pr_err("osdc handle_map corrupt msg\n");
1467         ceph_msg_dump(msg);
1468         up_write(&osdc->map_sem);
1469         return;
1470 }
1471
1472 /*
1473  * watch/notify callback event infrastructure
1474  *
1475  * These callbacks are used both for watch and notify operations.
1476  */
1477 static void __release_event(struct kref *kref)
1478 {
1479         struct ceph_osd_event *event =
1480                 container_of(kref, struct ceph_osd_event, kref);
1481
1482         dout("__release_event %p\n", event);
1483         kfree(event);
1484 }
1485
1486 static void get_event(struct ceph_osd_event *event)
1487 {
1488         kref_get(&event->kref);
1489 }
1490
1491 void ceph_osdc_put_event(struct ceph_osd_event *event)
1492 {
1493         kref_put(&event->kref, __release_event);
1494 }
1495 EXPORT_SYMBOL(ceph_osdc_put_event);
1496
1497 static void __insert_event(struct ceph_osd_client *osdc,
1498                              struct ceph_osd_event *new)
1499 {
1500         struct rb_node **p = &osdc->event_tree.rb_node;
1501         struct rb_node *parent = NULL;
1502         struct ceph_osd_event *event = NULL;
1503
1504         while (*p) {
1505                 parent = *p;
1506                 event = rb_entry(parent, struct ceph_osd_event, node);
1507                 if (new->cookie < event->cookie)
1508                         p = &(*p)->rb_left;
1509                 else if (new->cookie > event->cookie)
1510                         p = &(*p)->rb_right;
1511                 else
1512                         BUG();
1513         }
1514
1515         rb_link_node(&new->node, parent, p);
1516         rb_insert_color(&new->node, &osdc->event_tree);
1517 }
1518
1519 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1520                                                 u64 cookie)
1521 {
1522         struct rb_node **p = &osdc->event_tree.rb_node;
1523         struct rb_node *parent = NULL;
1524         struct ceph_osd_event *event = NULL;
1525
1526         while (*p) {
1527                 parent = *p;
1528                 event = rb_entry(parent, struct ceph_osd_event, node);
1529                 if (cookie < event->cookie)
1530                         p = &(*p)->rb_left;
1531                 else if (cookie > event->cookie)
1532                         p = &(*p)->rb_right;
1533                 else
1534                         return event;
1535         }
1536         return NULL;
1537 }
1538
1539 static void __remove_event(struct ceph_osd_event *event)
1540 {
1541         struct ceph_osd_client *osdc = event->osdc;
1542
1543         if (!RB_EMPTY_NODE(&event->node)) {
1544                 dout("__remove_event removed %p\n", event);
1545                 rb_erase(&event->node, &osdc->event_tree);
1546                 ceph_osdc_put_event(event);
1547         } else {
1548                 dout("__remove_event didn't remove %p\n", event);
1549         }
1550 }
1551
1552 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1553                            void (*event_cb)(u64, u64, u8, void *),
1554                            int one_shot, void *data,
1555                            struct ceph_osd_event **pevent)
1556 {
1557         struct ceph_osd_event *event;
1558
1559         event = kmalloc(sizeof(*event), GFP_NOIO);
1560         if (!event)
1561                 return -ENOMEM;
1562
1563         dout("create_event %p\n", event);
1564         event->cb = event_cb;
1565         event->one_shot = one_shot;
1566         event->data = data;
1567         event->osdc = osdc;
1568         INIT_LIST_HEAD(&event->osd_node);
1569         kref_init(&event->kref);   /* one ref for us */
1570         kref_get(&event->kref);    /* one ref for the caller */
1571         init_completion(&event->completion);
1572
1573         spin_lock(&osdc->event_lock);
1574         event->cookie = ++osdc->event_count;
1575         __insert_event(osdc, event);
1576         spin_unlock(&osdc->event_lock);
1577
1578         *pevent = event;
1579         return 0;
1580 }
1581 EXPORT_SYMBOL(ceph_osdc_create_event);
1582
1583 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1584 {
1585         struct ceph_osd_client *osdc = event->osdc;
1586
1587         dout("cancel_event %p\n", event);
1588         spin_lock(&osdc->event_lock);
1589         __remove_event(event);
1590         spin_unlock(&osdc->event_lock);
1591         ceph_osdc_put_event(event); /* caller's */
1592 }
1593 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1594
1595
1596 static void do_event_work(struct work_struct *work)
1597 {
1598         struct ceph_osd_event_work *event_work =
1599                 container_of(work, struct ceph_osd_event_work, work);
1600         struct ceph_osd_event *event = event_work->event;
1601         u64 ver = event_work->ver;
1602         u64 notify_id = event_work->notify_id;
1603         u8 opcode = event_work->opcode;
1604
1605         dout("do_event_work completing %p\n", event);
1606         event->cb(ver, notify_id, opcode, event->data);
1607         complete(&event->completion);
1608         dout("do_event_work completed %p\n", event);
1609         ceph_osdc_put_event(event);
1610         kfree(event_work);
1611 }
1612
1613
1614 /*
1615  * Process osd watch notifications
1616  */
1617 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1618 {
1619         void *p, *end;
1620         u8 proto_ver;
1621         u64 cookie, ver, notify_id;
1622         u8 opcode;
1623         struct ceph_osd_event *event;
1624         struct ceph_osd_event_work *event_work;
1625
1626         p = msg->front.iov_base;
1627         end = p + msg->front.iov_len;
1628
1629         ceph_decode_8_safe(&p, end, proto_ver, bad);
1630         ceph_decode_8_safe(&p, end, opcode, bad);
1631         ceph_decode_64_safe(&p, end, cookie, bad);
1632         ceph_decode_64_safe(&p, end, ver, bad);
1633         ceph_decode_64_safe(&p, end, notify_id, bad);
1634
1635         spin_lock(&osdc->event_lock);
1636         event = __find_event(osdc, cookie);
1637         if (event) {
1638                 get_event(event);
1639                 if (event->one_shot)
1640                         __remove_event(event);
1641         }
1642         spin_unlock(&osdc->event_lock);
1643         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1644              cookie, ver, event);
1645         if (event) {
1646                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1647                 if (!event_work) {
1648                         dout("ERROR: could not allocate event_work\n");
1649                         goto done_err;
1650                 }
1651                 INIT_WORK(&event_work->work, do_event_work);
1652                 event_work->event = event;
1653                 event_work->ver = ver;
1654                 event_work->notify_id = notify_id;
1655                 event_work->opcode = opcode;
1656                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1657                         dout("WARNING: failed to queue notify event work\n");
1658                         goto done_err;
1659                 }
1660         }
1661
1662         return;
1663
1664 done_err:
1665         complete(&event->completion);
1666         ceph_osdc_put_event(event);
1667         return;
1668
1669 bad:
1670         pr_err("osdc handle_watch_notify corrupt msg\n");
1671         return;
1672 }
1673
1674 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1675 {
1676         int err;
1677
1678         dout("wait_event %p\n", event);
1679         err = wait_for_completion_interruptible_timeout(&event->completion,
1680                                                         timeout * HZ);
1681         ceph_osdc_put_event(event);
1682         if (err > 0)
1683                 err = 0;
1684         dout("wait_event %p returns %d\n", event, err);
1685         return err;
1686 }
1687 EXPORT_SYMBOL(ceph_osdc_wait_event);
1688
1689 /*
1690  * Register request, send initial attempt.
1691  */
1692 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1693                             struct ceph_osd_request *req,
1694                             bool nofail)
1695 {
1696         int rc = 0;
1697
1698         req->r_request->pages = req->r_pages;
1699         req->r_request->nr_pages = req->r_num_pages;
1700 #ifdef CONFIG_BLOCK
1701         req->r_request->bio = req->r_bio;
1702 #endif
1703         req->r_request->trail = req->r_trail;
1704
1705         register_request(osdc, req);
1706
1707         down_read(&osdc->map_sem);
1708         mutex_lock(&osdc->request_mutex);
1709         /*
1710          * a racing kick_requests() may have sent the message for us
1711          * while we dropped request_mutex above, so only send now if
1712          * the request still han't been touched yet.
1713          */
1714         if (req->r_sent == 0) {
1715                 rc = __map_request(osdc, req, 0);
1716                 if (rc < 0) {
1717                         if (nofail) {
1718                                 dout("osdc_start_request failed map, "
1719                                      " will retry %lld\n", req->r_tid);
1720                                 rc = 0;
1721                         }
1722                         goto out_unlock;
1723                 }
1724                 if (req->r_osd == NULL) {
1725                         dout("send_request %p no up osds in pg\n", req);
1726                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1727                 } else {
1728                         __send_request(osdc, req);
1729                 }
1730                 rc = 0;
1731         }
1732
1733 out_unlock:
1734         mutex_unlock(&osdc->request_mutex);
1735         up_read(&osdc->map_sem);
1736         return rc;
1737 }
1738 EXPORT_SYMBOL(ceph_osdc_start_request);
1739
1740 /*
1741  * wait for a request to complete
1742  */
1743 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1744                            struct ceph_osd_request *req)
1745 {
1746         int rc;
1747
1748         rc = wait_for_completion_interruptible(&req->r_completion);
1749         if (rc < 0) {
1750                 mutex_lock(&osdc->request_mutex);
1751                 __cancel_request(req);
1752                 __unregister_request(osdc, req);
1753                 mutex_unlock(&osdc->request_mutex);
1754                 complete_request(req);
1755                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1756                 return rc;
1757         }
1758
1759         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1760         return req->r_result;
1761 }
1762 EXPORT_SYMBOL(ceph_osdc_wait_request);
1763
1764 /*
1765  * sync - wait for all in-flight requests to flush.  avoid starvation.
1766  */
1767 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1768 {
1769         struct ceph_osd_request *req;
1770         u64 last_tid, next_tid = 0;
1771
1772         mutex_lock(&osdc->request_mutex);
1773         last_tid = osdc->last_tid;
1774         while (1) {
1775                 req = __lookup_request_ge(osdc, next_tid);
1776                 if (!req)
1777                         break;
1778                 if (req->r_tid > last_tid)
1779                         break;
1780
1781                 next_tid = req->r_tid + 1;
1782                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1783                         continue;
1784
1785                 ceph_osdc_get_request(req);
1786                 mutex_unlock(&osdc->request_mutex);
1787                 dout("sync waiting on tid %llu (last is %llu)\n",
1788                      req->r_tid, last_tid);
1789                 wait_for_completion(&req->r_safe_completion);
1790                 mutex_lock(&osdc->request_mutex);
1791                 ceph_osdc_put_request(req);
1792         }
1793         mutex_unlock(&osdc->request_mutex);
1794         dout("sync done (thru tid %llu)\n", last_tid);
1795 }
1796 EXPORT_SYMBOL(ceph_osdc_sync);
1797
1798 /*
1799  * init, shutdown
1800  */
1801 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1802 {
1803         int err;
1804
1805         dout("init\n");
1806         osdc->client = client;
1807         osdc->osdmap = NULL;
1808         init_rwsem(&osdc->map_sem);
1809         init_completion(&osdc->map_waiters);
1810         osdc->last_requested_map = 0;
1811         mutex_init(&osdc->request_mutex);
1812         osdc->last_tid = 0;
1813         osdc->osds = RB_ROOT;
1814         INIT_LIST_HEAD(&osdc->osd_lru);
1815         osdc->requests = RB_ROOT;
1816         INIT_LIST_HEAD(&osdc->req_lru);
1817         INIT_LIST_HEAD(&osdc->req_unsent);
1818         INIT_LIST_HEAD(&osdc->req_notarget);
1819         INIT_LIST_HEAD(&osdc->req_linger);
1820         osdc->num_requests = 0;
1821         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1822         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1823         spin_lock_init(&osdc->event_lock);
1824         osdc->event_tree = RB_ROOT;
1825         osdc->event_count = 0;
1826
1827         schedule_delayed_work(&osdc->osds_timeout_work,
1828            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1829
1830         err = -ENOMEM;
1831         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1832                                         sizeof(struct ceph_osd_request));
1833         if (!osdc->req_mempool)
1834                 goto out;
1835
1836         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1837                                 "osd_op");
1838         if (err < 0)
1839                 goto out_mempool;
1840         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1841                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1842                                 "osd_op_reply");
1843         if (err < 0)
1844                 goto out_msgpool;
1845
1846         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1847         if (IS_ERR(osdc->notify_wq)) {
1848                 err = PTR_ERR(osdc->notify_wq);
1849                 osdc->notify_wq = NULL;
1850                 goto out_msgpool;
1851         }
1852         return 0;
1853
1854 out_msgpool:
1855         ceph_msgpool_destroy(&osdc->msgpool_op);
1856 out_mempool:
1857         mempool_destroy(osdc->req_mempool);
1858 out:
1859         return err;
1860 }
1861 EXPORT_SYMBOL(ceph_osdc_init);
1862
1863 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1864 {
1865         flush_workqueue(osdc->notify_wq);
1866         destroy_workqueue(osdc->notify_wq);
1867         cancel_delayed_work_sync(&osdc->timeout_work);
1868         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1869         if (osdc->osdmap) {
1870                 ceph_osdmap_destroy(osdc->osdmap);
1871                 osdc->osdmap = NULL;
1872         }
1873         remove_all_osds(osdc);
1874         mempool_destroy(osdc->req_mempool);
1875         ceph_msgpool_destroy(&osdc->msgpool_op);
1876         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1877 }
1878 EXPORT_SYMBOL(ceph_osdc_stop);
1879
1880 /*
1881  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1882  * *plen.  Return number of bytes read, or error.
1883  */
1884 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1885                         struct ceph_vino vino, struct ceph_file_layout *layout,
1886                         u64 off, u64 *plen,
1887                         u32 truncate_seq, u64 truncate_size,
1888                         struct page **pages, int num_pages, int page_align)
1889 {
1890         struct ceph_osd_request *req;
1891         int rc = 0;
1892
1893         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1894              vino.snap, off, *plen);
1895         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1896                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1897                                     NULL, 0, truncate_seq, truncate_size, NULL,
1898                                     false, 1, page_align);
1899         if (!req)
1900                 return -ENOMEM;
1901
1902         /* it may be a short read due to an object boundary */
1903         req->r_pages = pages;
1904
1905         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1906              off, *plen, req->r_num_pages, page_align);
1907
1908         rc = ceph_osdc_start_request(osdc, req, false);
1909         if (!rc)
1910                 rc = ceph_osdc_wait_request(osdc, req);
1911
1912         ceph_osdc_put_request(req);
1913         dout("readpages result %d\n", rc);
1914         return rc;
1915 }
1916 EXPORT_SYMBOL(ceph_osdc_readpages);
1917
1918 /*
1919  * do a synchronous write on N pages
1920  */
1921 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1922                          struct ceph_file_layout *layout,
1923                          struct ceph_snap_context *snapc,
1924                          u64 off, u64 len,
1925                          u32 truncate_seq, u64 truncate_size,
1926                          struct timespec *mtime,
1927                          struct page **pages, int num_pages,
1928                          int flags, int do_sync, bool nofail)
1929 {
1930         struct ceph_osd_request *req;
1931         int rc = 0;
1932         int page_align = off & ~PAGE_MASK;
1933
1934         BUG_ON(vino.snap != CEPH_NOSNAP);
1935         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1936                                     CEPH_OSD_OP_WRITE,
1937                                     flags | CEPH_OSD_FLAG_ONDISK |
1938                                             CEPH_OSD_FLAG_WRITE,
1939                                     snapc, do_sync,
1940                                     truncate_seq, truncate_size, mtime,
1941                                     nofail, 1, page_align);
1942         if (!req)
1943                 return -ENOMEM;
1944
1945         /* it may be a short write due to an object boundary */
1946         req->r_pages = pages;
1947         dout("writepages %llu~%llu (%d pages)\n", off, len,
1948              req->r_num_pages);
1949
1950         rc = ceph_osdc_start_request(osdc, req, nofail);
1951         if (!rc)
1952                 rc = ceph_osdc_wait_request(osdc, req);
1953
1954         ceph_osdc_put_request(req);
1955         if (rc == 0)
1956                 rc = len;
1957         dout("writepages result %d\n", rc);
1958         return rc;
1959 }
1960 EXPORT_SYMBOL(ceph_osdc_writepages);
1961
1962 /*
1963  * handle incoming message
1964  */
1965 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1966 {
1967         struct ceph_osd *osd = con->private;
1968         struct ceph_osd_client *osdc;
1969         int type = le16_to_cpu(msg->hdr.type);
1970
1971         if (!osd)
1972                 goto out;
1973         osdc = osd->o_osdc;
1974
1975         switch (type) {
1976         case CEPH_MSG_OSD_MAP:
1977                 ceph_osdc_handle_map(osdc, msg);
1978                 break;
1979         case CEPH_MSG_OSD_OPREPLY:
1980                 handle_reply(osdc, msg, con);
1981                 break;
1982         case CEPH_MSG_WATCH_NOTIFY:
1983                 handle_watch_notify(osdc, msg);
1984                 break;
1985
1986         default:
1987                 pr_err("received unknown message type %d %s\n", type,
1988                        ceph_msg_type_name(type));
1989         }
1990 out:
1991         ceph_msg_put(msg);
1992 }
1993
1994 /*
1995  * lookup and return message for incoming reply.  set up reply message
1996  * pages.
1997  */
1998 static struct ceph_msg *get_reply(struct ceph_connection *con,
1999                                   struct ceph_msg_header *hdr,
2000                                   int *skip)
2001 {
2002         struct ceph_osd *osd = con->private;
2003         struct ceph_osd_client *osdc = osd->o_osdc;
2004         struct ceph_msg *m;
2005         struct ceph_osd_request *req;
2006         int front = le32_to_cpu(hdr->front_len);
2007         int data_len = le32_to_cpu(hdr->data_len);
2008         u64 tid;
2009
2010         tid = le64_to_cpu(hdr->tid);
2011         mutex_lock(&osdc->request_mutex);
2012         req = __lookup_request(osdc, tid);
2013         if (!req) {
2014                 *skip = 1;
2015                 m = NULL;
2016                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
2017                         osd->o_osd);
2018                 goto out;
2019         }
2020
2021         if (req->r_con_filling_msg) {
2022                 dout("get_reply revoking msg %p from old con %p\n",
2023                      req->r_reply, req->r_con_filling_msg);
2024                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
2025                 ceph_con_put(req->r_con_filling_msg);
2026                 req->r_con_filling_msg = NULL;
2027         }
2028
2029         if (front > req->r_reply->front.iov_len) {
2030                 pr_warning("get_reply front %d > preallocated %d\n",
2031                            front, (int)req->r_reply->front.iov_len);
2032                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2033                 if (!m)
2034                         goto out;
2035                 ceph_msg_put(req->r_reply);
2036                 req->r_reply = m;
2037         }
2038         m = ceph_msg_get(req->r_reply);
2039
2040         if (data_len > 0) {
2041                 int want = calc_pages_for(req->r_page_alignment, data_len);
2042
2043                 if (unlikely(req->r_num_pages < want)) {
2044                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2045                                    " had only %d pages ready\n", tid, data_len,
2046                                    want, req->r_num_pages);
2047                         *skip = 1;
2048                         ceph_msg_put(m);
2049                         m = NULL;
2050                         goto out;
2051                 }
2052                 m->pages = req->r_pages;
2053                 m->nr_pages = req->r_num_pages;
2054                 m->page_alignment = req->r_page_alignment;
2055 #ifdef CONFIG_BLOCK
2056                 m->bio = req->r_bio;
2057 #endif
2058         }
2059         *skip = 0;
2060         req->r_con_filling_msg = ceph_con_get(con);
2061         dout("get_reply tid %lld %p\n", tid, m);
2062
2063 out:
2064         mutex_unlock(&osdc->request_mutex);
2065         return m;
2066
2067 }
2068
2069 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2070                                   struct ceph_msg_header *hdr,
2071                                   int *skip)
2072 {
2073         struct ceph_osd *osd = con->private;
2074         int type = le16_to_cpu(hdr->type);
2075         int front = le32_to_cpu(hdr->front_len);
2076
2077         switch (type) {
2078         case CEPH_MSG_OSD_MAP:
2079         case CEPH_MSG_WATCH_NOTIFY:
2080                 return ceph_msg_new(type, front, GFP_NOFS, false);
2081         case CEPH_MSG_OSD_OPREPLY:
2082                 return get_reply(con, hdr, skip);
2083         default:
2084                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2085                         osd->o_osd);
2086                 *skip = 1;
2087                 return NULL;
2088         }
2089 }
2090
2091 /*
2092  * Wrappers to refcount containing ceph_osd struct
2093  */
2094 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2095 {
2096         struct ceph_osd *osd = con->private;
2097         if (get_osd(osd))
2098                 return con;
2099         return NULL;
2100 }
2101
2102 static void put_osd_con(struct ceph_connection *con)
2103 {
2104         struct ceph_osd *osd = con->private;
2105         put_osd(osd);
2106 }
2107
2108 /*
2109  * authentication
2110  */
2111 static int get_authorizer(struct ceph_connection *con,
2112                           void **buf, int *len, int *proto,
2113                           void **reply_buf, int *reply_len, int force_new)
2114 {
2115         struct ceph_osd *o = con->private;
2116         struct ceph_osd_client *osdc = o->o_osdc;
2117         struct ceph_auth_client *ac = osdc->client->monc.auth;
2118         int ret = 0;
2119
2120         if (force_new && o->o_authorizer) {
2121                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
2122                 o->o_authorizer = NULL;
2123         }
2124         if (o->o_authorizer == NULL) {
2125                 ret = ac->ops->create_authorizer(
2126                         ac, CEPH_ENTITY_TYPE_OSD,
2127                         &o->o_authorizer,
2128                         &o->o_authorizer_buf,
2129                         &o->o_authorizer_buf_len,
2130                         &o->o_authorizer_reply_buf,
2131                         &o->o_authorizer_reply_buf_len);
2132                 if (ret)
2133                         return ret;
2134         }
2135
2136         *proto = ac->protocol;
2137         *buf = o->o_authorizer_buf;
2138         *len = o->o_authorizer_buf_len;
2139         *reply_buf = o->o_authorizer_reply_buf;
2140         *reply_len = o->o_authorizer_reply_buf_len;
2141         return 0;
2142 }
2143
2144
2145 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2146 {
2147         struct ceph_osd *o = con->private;
2148         struct ceph_osd_client *osdc = o->o_osdc;
2149         struct ceph_auth_client *ac = osdc->client->monc.auth;
2150
2151         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
2152 }
2153
2154 static int invalidate_authorizer(struct ceph_connection *con)
2155 {
2156         struct ceph_osd *o = con->private;
2157         struct ceph_osd_client *osdc = o->o_osdc;
2158         struct ceph_auth_client *ac = osdc->client->monc.auth;
2159
2160         if (ac->ops->invalidate_authorizer)
2161                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2162
2163         return ceph_monc_validate_auth(&osdc->client->monc);
2164 }
2165
2166 static const struct ceph_connection_operations osd_con_ops = {
2167         .get = get_osd_con,
2168         .put = put_osd_con,
2169         .dispatch = dispatch,
2170         .get_authorizer = get_authorizer,
2171         .verify_authorizer_reply = verify_authorizer_reply,
2172         .invalidate_authorizer = invalidate_authorizer,
2173         .alloc_msg = alloc_msg,
2174         .fault = osd_reset,
2175 };