Pileus Git - ~andy/linux/blob - 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 == -EAGAIN)
 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         INIT_LIST_HEAD(&osd->o_requests);
 646         INIT_LIST_HEAD(&osd->o_linger_requests);
 647         INIT_LIST_HEAD(&osd->o_osd_lru);
 648         osd->o_incarnation = 1;
 649
 650         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 651
 652         INIT_LIST_HEAD(&osd->o_keepalive_item);
 653         return osd;
 654 }
 655
 656 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 657 {
 658         if (atomic_inc_not_zero(&osd->o_ref)) {
 659                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 660                      atomic_read(&osd->o_ref));
 661                 return osd;
 662         } else {
 663                 dout("get_osd %p FAIL\n", osd);
 664                 return NULL;
 665         }
 666 }
 667
 668 static void put_osd(struct ceph_osd *osd)
 669 {
 670         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 671              atomic_read(&osd->o_ref) - 1);
 672         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 673                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 674
 675                 if (ac->ops && ac->ops->destroy_authorizer)
 676                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 677                 kfree(osd);
 678         }
 679 }
 680
 681 /*
 682  * remove an osd from our map
 683  */
 684 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 685 {
 686         dout("__remove_osd %p\n", osd);
 687         BUG_ON(!list_empty(&osd->o_requests));
 688         rb_erase(&osd->o_node, &osdc->osds);
 689         list_del_init(&osd->o_osd_lru);
 690         ceph_con_close(&osd->o_con);
 691         put_osd(osd);
 692 }
 693
 694 static void remove_all_osds(struct ceph_osd_client *osdc)
 695 {
 696         dout("%s %p\n", __func__, osdc);
 697         mutex_lock(&osdc->request_mutex);
 698         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 699                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 700                                                 struct ceph_osd, o_node);
 701                 __remove_osd(osdc, osd);
 702         }
 703         mutex_unlock(&osdc->request_mutex);
 704 }
 705
 706 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 707                               struct ceph_osd *osd)
 708 {
 709         dout("__move_osd_to_lru %p\n", osd);
 710         BUG_ON(!list_empty(&osd->o_osd_lru));
 711         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 712         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 713 }
 714
 715 static void __remove_osd_from_lru(struct ceph_osd *osd)
 716 {
 717         dout("__remove_osd_from_lru %p\n", osd);
 718         if (!list_empty(&osd->o_osd_lru))
 719                 list_del_init(&osd->o_osd_lru);
 720 }
 721
 722 static void remove_old_osds(struct ceph_osd_client *osdc)
 723 {
 724         struct ceph_osd *osd, *nosd;
 725
 726         dout("__remove_old_osds %p\n", osdc);
 727         mutex_lock(&osdc->request_mutex);
 728         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 729                 if (time_before(jiffies, osd->lru_ttl))
 730                         break;
 731                 __remove_osd(osdc, osd);
 732         }
 733         mutex_unlock(&osdc->request_mutex);
 734 }
 735
 736 /*
 737  * reset osd connect
 738  */
 739 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 740 {
 741         struct ceph_osd_request *req;
 742         int ret = 0;
 743
 744         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 745         if (list_empty(&osd->o_requests) &&
 746             list_empty(&osd->o_linger_requests)) {
 747                 __remove_osd(osdc, osd);
 748         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 749                           &osd->o_con.peer_addr,
 750                           sizeof(osd->o_con.peer_addr)) == 0 &&
 751                    !ceph_con_opened(&osd->o_con)) {
 752                 dout(" osd addr hasn't changed and connection never opened,"
 753                      " letting msgr retry");
 754                 /* touch each r_stamp for handle_timeout()'s benfit */
 755                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 756                         req->r_stamp = jiffies;
 757                 ret = -EAGAIN;
 758         } else {
 759                 ceph_con_close(&osd->o_con);
 760                 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
 761                               &osdc->osdmap->osd_addr[osd->o_osd]);
 762                 osd->o_incarnation++;
 763         }
 764         return ret;
 765 }
 766
 767 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 768 {
 769         struct rb_node **p = &osdc->osds.rb_node;
 770         struct rb_node *parent = NULL;
 771         struct ceph_osd *osd = NULL;
 772
 773         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 774         while (*p) {
 775                 parent = *p;
 776                 osd = rb_entry(parent, struct ceph_osd, o_node);
 777                 if (new->o_osd < osd->o_osd)
 778                         p = &(*p)->rb_left;
 779                 else if (new->o_osd > osd->o_osd)
 780                         p = &(*p)->rb_right;
 781                 else
 782                         BUG();
 783         }
 784
 785         rb_link_node(&new->o_node, parent, p);
 786         rb_insert_color(&new->o_node, &osdc->osds);
 787 }
 788
 789 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 790 {
 791         struct ceph_osd *osd;
 792         struct rb_node *n = osdc->osds.rb_node;
 793
 794         while (n) {
 795                 osd = rb_entry(n, struct ceph_osd, o_node);
 796                 if (o < osd->o_osd)
 797                         n = n->rb_left;
 798                 else if (o > osd->o_osd)
 799                         n = n->rb_right;
 800                 else
 801                         return osd;
 802         }
 803         return NULL;
 804 }
 805
 806 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 807 {
 808         schedule_delayed_work(&osdc->timeout_work,
 809                         osdc->client->options->osd_keepalive_timeout * HZ);
 810 }
 811
 812 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 813 {
 814         cancel_delayed_work(&osdc->timeout_work);
 815 }
 816
 817 /*
 818  * Register request, assign tid.  If this is the first request, set up
 819  * the timeout event.
 820  */
 821 static void __register_request(struct ceph_osd_client *osdc,
 822                                struct ceph_osd_request *req)
 823 {
 824         req->r_tid = ++osdc->last_tid;
 825         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 826         dout("__register_request %p tid %lld\n", req, req->r_tid);
 827         __insert_request(osdc, req);
 828         ceph_osdc_get_request(req);
 829         osdc->num_requests++;
 830         if (osdc->num_requests == 1) {
 831                 dout(" first request, scheduling timeout\n");
 832                 __schedule_osd_timeout(osdc);
 833         }
 834 }
 835
 836 static void register_request(struct ceph_osd_client *osdc,
 837                              struct ceph_osd_request *req)
 838 {
 839         mutex_lock(&osdc->request_mutex);
 840         __register_request(osdc, req);
 841         mutex_unlock(&osdc->request_mutex);
 842 }
 843
 844 /*
 845  * called under osdc->request_mutex
 846  */
 847 static void __unregister_request(struct ceph_osd_client *osdc,
 848                                  struct ceph_osd_request *req)
 849 {
 850         if (RB_EMPTY_NODE(&req->r_node)) {
 851                 dout("__unregister_request %p tid %lld not registered\n",
 852                         req, req->r_tid);
 853                 return;
 854         }
 855
 856         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 857         rb_erase(&req->r_node, &osdc->requests);
 858         osdc->num_requests--;
 859
 860         if (req->r_osd) {
 861                 /* make sure the original request isn't in flight. */
 862                 ceph_msg_revoke(req->r_request);
 863
 864                 list_del_init(&req->r_osd_item);
 865                 if (list_empty(&req->r_osd->o_requests) &&
 866                     list_empty(&req->r_osd->o_linger_requests)) {
 867                         dout("moving osd to %p lru\n", req->r_osd);
 868                         __move_osd_to_lru(osdc, req->r_osd);
 869                 }
 870                 if (list_empty(&req->r_linger_item))
 871                         req->r_osd = NULL;
 872         }
 873
 874         ceph_osdc_put_request(req);
 875
 876         list_del_init(&req->r_req_lru_item);
 877         if (osdc->num_requests == 0) {
 878                 dout(" no requests, canceling timeout\n");
 879                 __cancel_osd_timeout(osdc);
 880         }
 881 }
 882
 883 /*
 884  * Cancel a previously queued request message
 885  */
 886 static void __cancel_request(struct ceph_osd_request *req)
 887 {
 888         if (req->r_sent && req->r_osd) {
 889                 ceph_msg_revoke(req->r_request);
 890                 req->r_sent = 0;
 891         }
 892 }
 893
 894 static void __register_linger_request(struct ceph_osd_client *osdc,
 895                                     struct ceph_osd_request *req)
 896 {
 897         dout("__register_linger_request %p\n", req);
 898         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 899         if (req->r_osd)
 900                 list_add_tail(&req->r_linger_osd,
 901                               &req->r_osd->o_linger_requests);
 902 }
 903
 904 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 905                                         struct ceph_osd_request *req)
 906 {
 907         dout("__unregister_linger_request %p\n", req);
 908         if (req->r_osd) {
 909                 list_del_init(&req->r_linger_item);
 910                 list_del_init(&req->r_linger_osd);
 911
 912                 if (list_empty(&req->r_osd->o_requests) &&
 913                     list_empty(&req->r_osd->o_linger_requests)) {
 914                         dout("moving osd to %p lru\n", req->r_osd);
 915                         __move_osd_to_lru(osdc, req->r_osd);
 916                 }
 917                 if (list_empty(&req->r_osd_item))
 918                         req->r_osd = NULL;
 919         }
 920 }
 921
 922 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 923                                          struct ceph_osd_request *req)
 924 {
 925         mutex_lock(&osdc->request_mutex);
 926         if (req->r_linger) {
 927                 __unregister_linger_request(osdc, req);
 928                 ceph_osdc_put_request(req);
 929         }
 930         mutex_unlock(&osdc->request_mutex);
 931 }
 932 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 933
 934 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 935                                   struct ceph_osd_request *req)
 936 {
 937         if (!req->r_linger) {
 938                 dout("set_request_linger %p\n", req);
 939                 req->r_linger = 1;
 940                 /*
 941                  * caller is now responsible for calling
 942                  * unregister_linger_request
 943                  */
 944                 ceph_osdc_get_request(req);
 945         }
 946 }
 947 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 948
 949 /*
 950  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 951  * (as needed), and set the request r_osd appropriately.  If there is
 952  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 953  * (unsent, homeless) or leave on in-flight lru.
 954  *
 955  * Return 0 if unchanged, 1 if changed, or negative on error.
 956  *
 957  * Caller should hold map_sem for read and request_mutex.
 958  */
 959 static int __map_request(struct ceph_osd_client *osdc,
 960                          struct ceph_osd_request *req, int force_resend)
 961 {
 962         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 963         struct ceph_pg pgid;
 964         int acting[CEPH_PG_MAX_SIZE];
 965         int o = -1, num = 0;
 966         int err;
 967
 968         dout("map_request %p tid %lld\n", req, req->r_tid);
 969         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 970                                       &req->r_file_layout, osdc->osdmap);
 971         if (err) {
 972                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 973                 return err;
 974         }
 975         pgid = reqhead->layout.ol_pgid;
 976         req->r_pgid = pgid;
 977
 978         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 979         if (err > 0) {
 980                 o = acting[0];
 981                 num = err;
 982         }
 983
 984         if ((!force_resend &&
 985              req->r_osd && req->r_osd->o_osd == o &&
 986              req->r_sent >= req->r_osd->o_incarnation &&
 987              req->r_num_pg_osds == num &&
 988              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 989             (req->r_osd == NULL && o == -1))
 990                 return 0;  /* no change */
 991
 992         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 993              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 994              req->r_osd ? req->r_osd->o_osd : -1);
 995
 996         /* record full pg acting set */
 997         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 998         req->r_num_pg_osds = num;
 999
1000         if (req->r_osd) {
1001                 __cancel_request(req);
1002                 list_del_init(&req->r_osd_item);
1003                 req->r_osd = NULL;
1004         }
1005
1006         req->r_osd = __lookup_osd(osdc, o);
1007         if (!req->r_osd && o >= 0) {
1008                 err = -ENOMEM;
1009                 req->r_osd = create_osd(osdc, o);
1010                 if (!req->r_osd) {
1011                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1012                         goto out;
1013                 }
1014
1015                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1016                 __insert_osd(osdc, req->r_osd);
1017
1018                 ceph_con_open(&req->r_osd->o_con,
1019                               CEPH_ENTITY_TYPE_OSD, o,
1020                               &osdc->osdmap->osd_addr[o]);
1021         }
1022
1023         if (req->r_osd) {
1024                 __remove_osd_from_lru(req->r_osd);
1025                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1026                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1027         } else {
1028                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1029         }
1030         err = 1;   /* osd or pg changed */
1031
1032 out:
1033         return err;
1034 }
1035
1036 /*
1037  * caller should hold map_sem (for read) and request_mutex
1038  */
1039 static void __send_request(struct ceph_osd_client *osdc,
1040                            struct ceph_osd_request *req)
1041 {
1042         struct ceph_osd_request_head *reqhead;
1043
1044         dout("send_request %p tid %llu to osd%d flags %d\n",
1045              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1046
1047         reqhead = req->r_request->front.iov_base;
1048         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1049         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1050         reqhead->reassert_version = req->r_reassert_version;
1051
1052         req->r_stamp = jiffies;
1053         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1054
1055         ceph_msg_get(req->r_request); /* send consumes a ref */
1056         ceph_con_send(&req->r_osd->o_con, req->r_request);
1057         req->r_sent = req->r_osd->o_incarnation;
1058 }
1059
1060 /*
1061  * Send any requests in the queue (req_unsent).
1062  */
1063 static void send_queued(struct ceph_osd_client *osdc)
1064 {
1065         struct ceph_osd_request *req, *tmp;
1066
1067         dout("send_queued\n");
1068         mutex_lock(&osdc->request_mutex);
1069         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1070                 __send_request(osdc, req);
1071         }
1072         mutex_unlock(&osdc->request_mutex);
1073 }
1074
1075 /*
1076  * Timeout callback, called every N seconds when 1 or more osd
1077  * requests has been active for more than N seconds.  When this
1078  * happens, we ping all OSDs with requests who have timed out to
1079  * ensure any communications channel reset is detected.  Reset the
1080  * request timeouts another N seconds in the future as we go.
1081  * Reschedule the timeout event another N seconds in future (unless
1082  * there are no open requests).
1083  */
1084 static void handle_timeout(struct work_struct *work)
1085 {
1086         struct ceph_osd_client *osdc =
1087                 container_of(work, struct ceph_osd_client, timeout_work.work);
1088         struct ceph_osd_request *req;
1089         struct ceph_osd *osd;
1090         unsigned long keepalive =
1091                 osdc->client->options->osd_keepalive_timeout * HZ;
1092         struct list_head slow_osds;
1093         dout("timeout\n");
1094         down_read(&osdc->map_sem);
1095
1096         ceph_monc_request_next_osdmap(&osdc->client->monc);
1097
1098         mutex_lock(&osdc->request_mutex);
1099
1100         /*
1101          * ping osds that are a bit slow.  this ensures that if there
1102          * is a break in the TCP connection we will notice, and reopen
1103          * a connection with that osd (from the fault callback).
1104          */
1105         INIT_LIST_HEAD(&slow_osds);
1106         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1107                 if (time_before(jiffies, req->r_stamp + keepalive))
1108                         break;
1109
1110                 osd = req->r_osd;
1111                 BUG_ON(!osd);
1112                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1113                      req->r_tid, osd->o_osd);
1114                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1115         }
1116         while (!list_empty(&slow_osds)) {
1117                 osd = list_entry(slow_osds.next, struct ceph_osd,
1118                                  o_keepalive_item);
1119                 list_del_init(&osd->o_keepalive_item);
1120                 ceph_con_keepalive(&osd->o_con);
1121         }
1122
1123         __schedule_osd_timeout(osdc);
1124         mutex_unlock(&osdc->request_mutex);
1125         send_queued(osdc);
1126         up_read(&osdc->map_sem);
1127 }
1128
1129 static void handle_osds_timeout(struct work_struct *work)
1130 {
1131         struct ceph_osd_client *osdc =
1132                 container_of(work, struct ceph_osd_client,
1133                              osds_timeout_work.work);
1134         unsigned long delay =
1135                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1136
1137         dout("osds timeout\n");
1138         down_read(&osdc->map_sem);
1139         remove_old_osds(osdc);
1140         up_read(&osdc->map_sem);
1141
1142         schedule_delayed_work(&osdc->osds_timeout_work,
1143                               round_jiffies_relative(delay));
1144 }
1145
1146 static void complete_request(struct ceph_osd_request *req)
1147 {
1148         if (req->r_safe_callback)
1149                 req->r_safe_callback(req, NULL);
1150         complete_all(&req->r_safe_completion);  /* fsync waiter */
1151 }
1152
1153 /*
1154  * handle osd op reply.  either call the callback if it is specified,
1155  * or do the completion to wake up the waiting thread.
1156  */
1157 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1158                          struct ceph_connection *con)
1159 {
1160         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1161         struct ceph_osd_request *req;
1162         u64 tid;
1163         int numops, object_len, flags;
1164         s32 result;
1165
1166         tid = le64_to_cpu(msg->hdr.tid);
1167         if (msg->front.iov_len < sizeof(*rhead))
1168                 goto bad;
1169         numops = le32_to_cpu(rhead->num_ops);
1170         object_len = le32_to_cpu(rhead->object_len);
1171         result = le32_to_cpu(rhead->result);
1172         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1173             numops * sizeof(struct ceph_osd_op))
1174                 goto bad;
1175         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1176         /* lookup */
1177         mutex_lock(&osdc->request_mutex);
1178         req = __lookup_request(osdc, tid);
1179         if (req == NULL) {
1180                 dout("handle_reply tid %llu dne\n", tid);
1181                 mutex_unlock(&osdc->request_mutex);
1182                 return;
1183         }
1184         ceph_osdc_get_request(req);
1185         flags = le32_to_cpu(rhead->flags);
1186
1187         /*
1188          * if this connection filled our message, drop our reference now, to
1189          * avoid a (safe but slower) revoke later.
1190          */
1191         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1192                 dout(" dropping con_filling_msg ref %p\n", con);
1193                 req->r_con_filling_msg = NULL;
1194                 con->ops->put(con);
1195         }
1196
1197         if (!req->r_got_reply) {
1198                 unsigned int bytes;
1199
1200                 req->r_result = le32_to_cpu(rhead->result);
1201                 bytes = le32_to_cpu(msg->hdr.data_len);
1202                 dout("handle_reply result %d bytes %d\n", req->r_result,
1203                      bytes);
1204                 if (req->r_result == 0)
1205                         req->r_result = bytes;
1206
1207                 /* in case this is a write and we need to replay, */
1208                 req->r_reassert_version = rhead->reassert_version;
1209
1210                 req->r_got_reply = 1;
1211         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1212                 dout("handle_reply tid %llu dup ack\n", tid);
1213                 mutex_unlock(&osdc->request_mutex);
1214                 goto done;
1215         }
1216
1217         dout("handle_reply tid %llu flags %d\n", tid, flags);
1218
1219         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1220                 __register_linger_request(osdc, req);
1221
1222         /* either this is a read, or we got the safe response */
1223         if (result < 0 ||
1224             (flags & CEPH_OSD_FLAG_ONDISK) ||
1225             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1226                 __unregister_request(osdc, req);
1227
1228         mutex_unlock(&osdc->request_mutex);
1229
1230         if (req->r_callback)
1231                 req->r_callback(req, msg);
1232         else
1233                 complete_all(&req->r_completion);
1234
1235         if (flags & CEPH_OSD_FLAG_ONDISK)
1236                 complete_request(req);
1237
1238 done:
1239         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1240         ceph_osdc_put_request(req);
1241         return;
1242
1243 bad:
1244         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1245                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1246                (int)sizeof(*rhead));
1247         ceph_msg_dump(msg);
1248 }
1249
1250 static void reset_changed_osds(struct ceph_osd_client *osdc)
1251 {
1252         struct rb_node *p, *n;
1253
1254         for (p = rb_first(&osdc->osds); p; p = n) {
1255                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1256
1257                 n = rb_next(p);
1258                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1259                     memcmp(&osd->o_con.peer_addr,
1260                            ceph_osd_addr(osdc->osdmap,
1261                                          osd->o_osd),
1262                            sizeof(struct ceph_entity_addr)) != 0)
1263                         __reset_osd(osdc, osd);
1264         }
1265 }
1266
1267 /*
1268  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1269  * no osd, request a new map.
1270  *
1271  * Caller should hold map_sem for read and request_mutex.
1272  */
1273 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1274 {
1275         struct ceph_osd_request *req, *nreq;
1276         struct rb_node *p;
1277         int needmap = 0;
1278         int err;
1279
1280         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1281         mutex_lock(&osdc->request_mutex);
1282         for (p = rb_first(&osdc->requests); p; ) {
1283                 req = rb_entry(p, struct ceph_osd_request, r_node);
1284                 p = rb_next(p);
1285                 err = __map_request(osdc, req, force_resend);
1286                 if (err < 0)
1287                         continue;  /* error */
1288                 if (req->r_osd == NULL) {
1289                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1290                         needmap++;  /* request a newer map */
1291                 } else if (err > 0) {
1292                         if (!req->r_linger) {
1293                                 dout("%p tid %llu requeued on osd%d\n", req,
1294                                      req->r_tid,
1295                                      req->r_osd ? req->r_osd->o_osd : -1);
1296                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1297                         }
1298                 }
1299                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1300                         /*
1301                          * register as a linger so that we will
1302                          * re-submit below and get a new tid
1303                          */
1304                         dout("%p tid %llu restart on osd%d\n",
1305                              req, req->r_tid,
1306                              req->r_osd ? req->r_osd->o_osd : -1);
1307                         __register_linger_request(osdc, req);
1308                         __unregister_request(osdc, req);
1309                 }
1310         }
1311
1312         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1313                                  r_linger_item) {
1314                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1315
1316                 err = __map_request(osdc, req, force_resend);
1317                 if (err == 0)
1318                         continue;  /* no change and no osd was specified */
1319                 if (err < 0)
1320                         continue;  /* hrm! */
1321                 if (req->r_osd == NULL) {
1322                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1323                         needmap++;  /* request a newer map */
1324                         continue;
1325                 }
1326
1327                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1328                      req->r_osd ? req->r_osd->o_osd : -1);
1329                 __unregister_linger_request(osdc, req);
1330                 __register_request(osdc, req);
1331         }
1332         mutex_unlock(&osdc->request_mutex);
1333
1334         if (needmap) {
1335                 dout("%d requests for down osds, need new map\n", needmap);
1336                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1337         }
1338 }
1339
1340
1341 /*
1342  * Process updated osd map.
1343  *
1344  * The message contains any number of incremental and full maps, normally
1345  * indicating some sort of topology change in the cluster.  Kick requests
1346  * off to different OSDs as needed.
1347  */
1348 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1349 {
1350         void *p, *end, *next;
1351         u32 nr_maps, maplen;
1352         u32 epoch;
1353         struct ceph_osdmap *newmap = NULL, *oldmap;
1354         int err;
1355         struct ceph_fsid fsid;
1356
1357         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1358         p = msg->front.iov_base;
1359         end = p + msg->front.iov_len;
1360
1361         /* verify fsid */
1362         ceph_decode_need(&p, end, sizeof(fsid), bad);
1363         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1364         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1365                 return;
1366
1367         down_write(&osdc->map_sem);
1368
1369         /* incremental maps */
1370         ceph_decode_32_safe(&p, end, nr_maps, bad);
1371         dout(" %d inc maps\n", nr_maps);
1372         while (nr_maps > 0) {
1373                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1374                 epoch = ceph_decode_32(&p);
1375                 maplen = ceph_decode_32(&p);
1376                 ceph_decode_need(&p, end, maplen, bad);
1377                 next = p + maplen;
1378                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1379                         dout("applying incremental map %u len %d\n",
1380                              epoch, maplen);
1381                         newmap = osdmap_apply_incremental(&p, next,
1382                                                           osdc->osdmap,
1383                                                           &osdc->client->msgr);
1384                         if (IS_ERR(newmap)) {
1385                                 err = PTR_ERR(newmap);
1386                                 goto bad;
1387                         }
1388                         BUG_ON(!newmap);
1389                         if (newmap != osdc->osdmap) {
1390                                 ceph_osdmap_destroy(osdc->osdmap);
1391                                 osdc->osdmap = newmap;
1392                         }
1393                         kick_requests(osdc, 0);
1394                         reset_changed_osds(osdc);
1395                 } else {
1396                         dout("ignoring incremental map %u len %d\n",
1397                              epoch, maplen);
1398                 }
1399                 p = next;
1400                 nr_maps--;
1401         }
1402         if (newmap)
1403                 goto done;
1404
1405         /* full maps */
1406         ceph_decode_32_safe(&p, end, nr_maps, bad);
1407         dout(" %d full maps\n", nr_maps);
1408         while (nr_maps) {
1409                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1410                 epoch = ceph_decode_32(&p);
1411                 maplen = ceph_decode_32(&p);
1412                 ceph_decode_need(&p, end, maplen, bad);
1413                 if (nr_maps > 1) {
1414                         dout("skipping non-latest full map %u len %d\n",
1415                              epoch, maplen);
1416                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1417                         dout("skipping full map %u len %d, "
1418                              "older than our %u\n", epoch, maplen,
1419                              osdc->osdmap->epoch);
1420                 } else {
1421                         int skipped_map = 0;
1422
1423                         dout("taking full map %u len %d\n", epoch, maplen);
1424                         newmap = osdmap_decode(&p, p+maplen);
1425                         if (IS_ERR(newmap)) {
1426                                 err = PTR_ERR(newmap);
1427                                 goto bad;
1428                         }
1429                         BUG_ON(!newmap);
1430                         oldmap = osdc->osdmap;
1431                         osdc->osdmap = newmap;
1432                         if (oldmap) {
1433                                 if (oldmap->epoch + 1 < newmap->epoch)
1434                                         skipped_map = 1;
1435                                 ceph_osdmap_destroy(oldmap);
1436                         }
1437                         kick_requests(osdc, skipped_map);
1438                 }
1439                 p += maplen;
1440                 nr_maps--;
1441         }
1442
1443 done:
1444         downgrade_write(&osdc->map_sem);
1445         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1446
1447         /*
1448          * subscribe to subsequent osdmap updates if full to ensure
1449          * we find out when we are no longer full and stop returning
1450          * ENOSPC.
1451          */
1452         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1453                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1454
1455         send_queued(osdc);
1456         up_read(&osdc->map_sem);
1457         wake_up_all(&osdc->client->auth_wq);
1458         return;
1459
1460 bad:
1461         pr_err("osdc handle_map corrupt msg\n");
1462         ceph_msg_dump(msg);
1463         up_write(&osdc->map_sem);
1464         return;
1465 }
1466
1467 /*
1468  * watch/notify callback event infrastructure
1469  *
1470  * These callbacks are used both for watch and notify operations.
1471  */
1472 static void __release_event(struct kref *kref)
1473 {
1474         struct ceph_osd_event *event =
1475                 container_of(kref, struct ceph_osd_event, kref);
1476
1477         dout("__release_event %p\n", event);
1478         kfree(event);
1479 }
1480
1481 static void get_event(struct ceph_osd_event *event)
1482 {
1483         kref_get(&event->kref);
1484 }
1485
1486 void ceph_osdc_put_event(struct ceph_osd_event *event)
1487 {
1488         kref_put(&event->kref, __release_event);
1489 }
1490 EXPORT_SYMBOL(ceph_osdc_put_event);
1491
1492 static void __insert_event(struct ceph_osd_client *osdc,
1493                              struct ceph_osd_event *new)
1494 {
1495         struct rb_node **p = &osdc->event_tree.rb_node;
1496         struct rb_node *parent = NULL;
1497         struct ceph_osd_event *event = NULL;
1498
1499         while (*p) {
1500                 parent = *p;
1501                 event = rb_entry(parent, struct ceph_osd_event, node);
1502                 if (new->cookie < event->cookie)
1503                         p = &(*p)->rb_left;
1504                 else if (new->cookie > event->cookie)
1505                         p = &(*p)->rb_right;
1506                 else
1507                         BUG();
1508         }
1509
1510         rb_link_node(&new->node, parent, p);
1511         rb_insert_color(&new->node, &osdc->event_tree);
1512 }
1513
1514 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1515                                                 u64 cookie)
1516 {
1517         struct rb_node **p = &osdc->event_tree.rb_node;
1518         struct rb_node *parent = NULL;
1519         struct ceph_osd_event *event = NULL;
1520
1521         while (*p) {
1522                 parent = *p;
1523                 event = rb_entry(parent, struct ceph_osd_event, node);
1524                 if (cookie < event->cookie)
1525                         p = &(*p)->rb_left;
1526                 else if (cookie > event->cookie)
1527                         p = &(*p)->rb_right;
1528                 else
1529                         return event;
1530         }
1531         return NULL;
1532 }
1533
1534 static void __remove_event(struct ceph_osd_event *event)
1535 {
1536         struct ceph_osd_client *osdc = event->osdc;
1537
1538         if (!RB_EMPTY_NODE(&event->node)) {
1539                 dout("__remove_event removed %p\n", event);
1540                 rb_erase(&event->node, &osdc->event_tree);
1541                 ceph_osdc_put_event(event);
1542         } else {
1543                 dout("__remove_event didn't remove %p\n", event);
1544         }
1545 }
1546
1547 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1548                            void (*event_cb)(u64, u64, u8, void *),
1549                            int one_shot, void *data,
1550                            struct ceph_osd_event **pevent)
1551 {
1552         struct ceph_osd_event *event;
1553
1554         event = kmalloc(sizeof(*event), GFP_NOIO);
1555         if (!event)
1556                 return -ENOMEM;
1557
1558         dout("create_event %p\n", event);
1559         event->cb = event_cb;
1560         event->one_shot = one_shot;
1561         event->data = data;
1562         event->osdc = osdc;
1563         INIT_LIST_HEAD(&event->osd_node);
1564         kref_init(&event->kref);   /* one ref for us */
1565         kref_get(&event->kref);    /* one ref for the caller */
1566         init_completion(&event->completion);
1567
1568         spin_lock(&osdc->event_lock);
1569         event->cookie = ++osdc->event_count;
1570         __insert_event(osdc, event);
1571         spin_unlock(&osdc->event_lock);
1572
1573         *pevent = event;
1574         return 0;
1575 }
1576 EXPORT_SYMBOL(ceph_osdc_create_event);
1577
1578 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1579 {
1580         struct ceph_osd_client *osdc = event->osdc;
1581
1582         dout("cancel_event %p\n", event);
1583         spin_lock(&osdc->event_lock);
1584         __remove_event(event);
1585         spin_unlock(&osdc->event_lock);
1586         ceph_osdc_put_event(event); /* caller's */
1587 }
1588 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1589
1590
1591 static void do_event_work(struct work_struct *work)
1592 {
1593         struct ceph_osd_event_work *event_work =
1594                 container_of(work, struct ceph_osd_event_work, work);
1595         struct ceph_osd_event *event = event_work->event;
1596         u64 ver = event_work->ver;
1597         u64 notify_id = event_work->notify_id;
1598         u8 opcode = event_work->opcode;
1599
1600         dout("do_event_work completing %p\n", event);
1601         event->cb(ver, notify_id, opcode, event->data);
1602         complete(&event->completion);
1603         dout("do_event_work completed %p\n", event);
1604         ceph_osdc_put_event(event);
1605         kfree(event_work);
1606 }
1607
1608
1609 /*
1610  * Process osd watch notifications
1611  */
1612 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1613 {
1614         void *p, *end;
1615         u8 proto_ver;
1616         u64 cookie, ver, notify_id;
1617         u8 opcode;
1618         struct ceph_osd_event *event;
1619         struct ceph_osd_event_work *event_work;
1620
1621         p = msg->front.iov_base;
1622         end = p + msg->front.iov_len;
1623
1624         ceph_decode_8_safe(&p, end, proto_ver, bad);
1625         ceph_decode_8_safe(&p, end, opcode, bad);
1626         ceph_decode_64_safe(&p, end, cookie, bad);
1627         ceph_decode_64_safe(&p, end, ver, bad);
1628         ceph_decode_64_safe(&p, end, notify_id, bad);
1629
1630         spin_lock(&osdc->event_lock);
1631         event = __find_event(osdc, cookie);
1632         if (event) {
1633                 get_event(event);
1634                 if (event->one_shot)
1635                         __remove_event(event);
1636         }
1637         spin_unlock(&osdc->event_lock);
1638         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1639              cookie, ver, event);
1640         if (event) {
1641                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1642                 if (!event_work) {
1643                         dout("ERROR: could not allocate event_work\n");
1644                         goto done_err;
1645                 }
1646                 INIT_WORK(&event_work->work, do_event_work);
1647                 event_work->event = event;
1648                 event_work->ver = ver;
1649                 event_work->notify_id = notify_id;
1650                 event_work->opcode = opcode;
1651                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1652                         dout("WARNING: failed to queue notify event work\n");
1653                         goto done_err;
1654                 }
1655         }
1656
1657         return;
1658
1659 done_err:
1660         complete(&event->completion);
1661         ceph_osdc_put_event(event);
1662         return;
1663
1664 bad:
1665         pr_err("osdc handle_watch_notify corrupt msg\n");
1666         return;
1667 }
1668
1669 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1670 {
1671         int err;
1672
1673         dout("wait_event %p\n", event);
1674         err = wait_for_completion_interruptible_timeout(&event->completion,
1675                                                         timeout * HZ);
1676         ceph_osdc_put_event(event);
1677         if (err > 0)
1678                 err = 0;
1679         dout("wait_event %p returns %d\n", event, err);
1680         return err;
1681 }
1682 EXPORT_SYMBOL(ceph_osdc_wait_event);
1683
1684 /*
1685  * Register request, send initial attempt.
1686  */
1687 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1688                             struct ceph_osd_request *req,
1689                             bool nofail)
1690 {
1691         int rc = 0;
1692
1693         req->r_request->pages = req->r_pages;
1694         req->r_request->nr_pages = req->r_num_pages;
1695 #ifdef CONFIG_BLOCK
1696         req->r_request->bio = req->r_bio;
1697 #endif
1698         req->r_request->trail = req->r_trail;
1699
1700         register_request(osdc, req);
1701
1702         down_read(&osdc->map_sem);
1703         mutex_lock(&osdc->request_mutex);
1704         /*
1705          * a racing kick_requests() may have sent the message for us
1706          * while we dropped request_mutex above, so only send now if
1707          * the request still han't been touched yet.
1708          */
1709         if (req->r_sent == 0) {
1710                 rc = __map_request(osdc, req, 0);
1711                 if (rc < 0) {
1712                         if (nofail) {
1713                                 dout("osdc_start_request failed map, "
1714                                      " will retry %lld\n", req->r_tid);
1715                                 rc = 0;
1716                         }
1717                         goto out_unlock;
1718                 }
1719                 if (req->r_osd == NULL) {
1720                         dout("send_request %p no up osds in pg\n", req);
1721                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1722                 } else {
1723                         __send_request(osdc, req);
1724                 }
1725                 rc = 0;
1726         }
1727
1728 out_unlock:
1729         mutex_unlock(&osdc->request_mutex);
1730         up_read(&osdc->map_sem);
1731         return rc;
1732 }
1733 EXPORT_SYMBOL(ceph_osdc_start_request);
1734
1735 /*
1736  * wait for a request to complete
1737  */
1738 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1739                            struct ceph_osd_request *req)
1740 {
1741         int rc;
1742
1743         rc = wait_for_completion_interruptible(&req->r_completion);
1744         if (rc < 0) {
1745                 mutex_lock(&osdc->request_mutex);
1746                 __cancel_request(req);
1747                 __unregister_request(osdc, req);
1748                 mutex_unlock(&osdc->request_mutex);
1749                 complete_request(req);
1750                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1751                 return rc;
1752         }
1753
1754         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1755         return req->r_result;
1756 }
1757 EXPORT_SYMBOL(ceph_osdc_wait_request);
1758
1759 /*
1760  * sync - wait for all in-flight requests to flush.  avoid starvation.
1761  */
1762 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1763 {
1764         struct ceph_osd_request *req;
1765         u64 last_tid, next_tid = 0;
1766
1767         mutex_lock(&osdc->request_mutex);
1768         last_tid = osdc->last_tid;
1769         while (1) {
1770                 req = __lookup_request_ge(osdc, next_tid);
1771                 if (!req)
1772                         break;
1773                 if (req->r_tid > last_tid)
1774                         break;
1775
1776                 next_tid = req->r_tid + 1;
1777                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1778                         continue;
1779
1780                 ceph_osdc_get_request(req);
1781                 mutex_unlock(&osdc->request_mutex);
1782                 dout("sync waiting on tid %llu (last is %llu)\n",
1783                      req->r_tid, last_tid);
1784                 wait_for_completion(&req->r_safe_completion);
1785                 mutex_lock(&osdc->request_mutex);
1786                 ceph_osdc_put_request(req);
1787         }
1788         mutex_unlock(&osdc->request_mutex);
1789         dout("sync done (thru tid %llu)\n", last_tid);
1790 }
1791 EXPORT_SYMBOL(ceph_osdc_sync);
1792
1793 /*
1794  * init, shutdown
1795  */
1796 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1797 {
1798         int err;
1799
1800         dout("init\n");
1801         osdc->client = client;
1802         osdc->osdmap = NULL;
1803         init_rwsem(&osdc->map_sem);
1804         init_completion(&osdc->map_waiters);
1805         osdc->last_requested_map = 0;
1806         mutex_init(&osdc->request_mutex);
1807         osdc->last_tid = 0;
1808         osdc->osds = RB_ROOT;
1809         INIT_LIST_HEAD(&osdc->osd_lru);
1810         osdc->requests = RB_ROOT;
1811         INIT_LIST_HEAD(&osdc->req_lru);
1812         INIT_LIST_HEAD(&osdc->req_unsent);
1813         INIT_LIST_HEAD(&osdc->req_notarget);
1814         INIT_LIST_HEAD(&osdc->req_linger);
1815         osdc->num_requests = 0;
1816         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1817         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1818         spin_lock_init(&osdc->event_lock);
1819         osdc->event_tree = RB_ROOT;
1820         osdc->event_count = 0;
1821
1822         schedule_delayed_work(&osdc->osds_timeout_work,
1823            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1824
1825         err = -ENOMEM;
1826         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1827                                         sizeof(struct ceph_osd_request));
1828         if (!osdc->req_mempool)
1829                 goto out;
1830
1831         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1832                                 OSD_OP_FRONT_LEN, 10, true,
1833                                 "osd_op");
1834         if (err < 0)
1835                 goto out_mempool;
1836         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1837                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1838                                 "osd_op_reply");
1839         if (err < 0)
1840                 goto out_msgpool;
1841
1842         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1843         if (IS_ERR(osdc->notify_wq)) {
1844                 err = PTR_ERR(osdc->notify_wq);
1845                 osdc->notify_wq = NULL;
1846                 goto out_msgpool;
1847         }
1848         return 0;
1849
1850 out_msgpool:
1851         ceph_msgpool_destroy(&osdc->msgpool_op);
1852 out_mempool:
1853         mempool_destroy(osdc->req_mempool);
1854 out:
1855         return err;
1856 }
1857 EXPORT_SYMBOL(ceph_osdc_init);
1858
1859 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1860 {
1861         flush_workqueue(osdc->notify_wq);
1862         destroy_workqueue(osdc->notify_wq);
1863         cancel_delayed_work_sync(&osdc->timeout_work);
1864         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1865         if (osdc->osdmap) {
1866                 ceph_osdmap_destroy(osdc->osdmap);
1867                 osdc->osdmap = NULL;
1868         }
1869         remove_all_osds(osdc);
1870         mempool_destroy(osdc->req_mempool);
1871         ceph_msgpool_destroy(&osdc->msgpool_op);
1872         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1873 }
1874 EXPORT_SYMBOL(ceph_osdc_stop);
1875
1876 /*
1877  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1878  * *plen.  Return number of bytes read, or error.
1879  */
1880 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1881                         struct ceph_vino vino, struct ceph_file_layout *layout,
1882                         u64 off, u64 *plen,
1883                         u32 truncate_seq, u64 truncate_size,
1884                         struct page **pages, int num_pages, int page_align)
1885 {
1886         struct ceph_osd_request *req;
1887         int rc = 0;
1888
1889         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1890              vino.snap, off, *plen);
1891         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1892                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1893                                     NULL, 0, truncate_seq, truncate_size, NULL,
1894                                     false, 1, page_align);
1895         if (IS_ERR(req))
1896                 return PTR_ERR(req);
1897
1898         /* it may be a short read due to an object boundary */
1899         req->r_pages = pages;
1900
1901         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1902              off, *plen, req->r_num_pages, page_align);
1903
1904         rc = ceph_osdc_start_request(osdc, req, false);
1905         if (!rc)
1906                 rc = ceph_osdc_wait_request(osdc, req);
1907
1908         ceph_osdc_put_request(req);
1909         dout("readpages result %d\n", rc);
1910         return rc;
1911 }
1912 EXPORT_SYMBOL(ceph_osdc_readpages);
1913
1914 /*
1915  * do a synchronous write on N pages
1916  */
1917 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1918                          struct ceph_file_layout *layout,
1919                          struct ceph_snap_context *snapc,
1920                          u64 off, u64 len,
1921                          u32 truncate_seq, u64 truncate_size,
1922                          struct timespec *mtime,
1923                          struct page **pages, int num_pages,
1924                          int flags, int do_sync, bool nofail)
1925 {
1926         struct ceph_osd_request *req;
1927         int rc = 0;
1928         int page_align = off & ~PAGE_MASK;
1929
1930         BUG_ON(vino.snap != CEPH_NOSNAP);
1931         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1932                                     CEPH_OSD_OP_WRITE,
1933                                     flags | CEPH_OSD_FLAG_ONDISK |
1934                                             CEPH_OSD_FLAG_WRITE,
1935                                     snapc, do_sync,
1936                                     truncate_seq, truncate_size, mtime,
1937                                     nofail, 1, page_align);
1938         if (IS_ERR(req))
1939                 return PTR_ERR(req);
1940
1941         /* it may be a short write due to an object boundary */
1942         req->r_pages = pages;
1943         dout("writepages %llu~%llu (%d pages)\n", off, len,
1944              req->r_num_pages);
1945
1946         rc = ceph_osdc_start_request(osdc, req, nofail);
1947         if (!rc)
1948                 rc = ceph_osdc_wait_request(osdc, req);
1949
1950         ceph_osdc_put_request(req);
1951         if (rc == 0)
1952                 rc = len;
1953         dout("writepages result %d\n", rc);
1954         return rc;
1955 }
1956 EXPORT_SYMBOL(ceph_osdc_writepages);
1957
1958 /*
1959  * handle incoming message
1960  */
1961 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1962 {
1963         struct ceph_osd *osd = con->private;
1964         struct ceph_osd_client *osdc;
1965         int type = le16_to_cpu(msg->hdr.type);
1966
1967         if (!osd)
1968                 goto out;
1969         osdc = osd->o_osdc;
1970
1971         switch (type) {
1972         case CEPH_MSG_OSD_MAP:
1973                 ceph_osdc_handle_map(osdc, msg);
1974                 break;
1975         case CEPH_MSG_OSD_OPREPLY:
1976                 handle_reply(osdc, msg, con);
1977                 break;
1978         case CEPH_MSG_WATCH_NOTIFY:
1979                 handle_watch_notify(osdc, msg);
1980                 break;
1981
1982         default:
1983                 pr_err("received unknown message type %d %s\n", type,
1984                        ceph_msg_type_name(type));
1985         }
1986 out:
1987         ceph_msg_put(msg);
1988 }
1989
1990 /*
1991  * lookup and return message for incoming reply.  set up reply message
1992  * pages.
1993  */
1994 static struct ceph_msg *get_reply(struct ceph_connection *con,
1995                                   struct ceph_msg_header *hdr,
1996                                   int *skip)
1997 {
1998         struct ceph_osd *osd = con->private;
1999         struct ceph_osd_client *osdc = osd->o_osdc;
2000         struct ceph_msg *m;
2001         struct ceph_osd_request *req;
2002         int front = le32_to_cpu(hdr->front_len);
2003         int data_len = le32_to_cpu(hdr->data_len);
2004         u64 tid;
2005
2006         tid = le64_to_cpu(hdr->tid);
2007         mutex_lock(&osdc->request_mutex);
2008         req = __lookup_request(osdc, tid);
2009         if (!req) {
2010                 *skip = 1;
2011                 m = NULL;
2012                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2013                      osd->o_osd);
2014                 goto out;
2015         }
2016
2017         if (req->r_con_filling_msg) {
2018                 dout("%s revoking msg %p from old con %p\n", __func__,
2019                      req->r_reply, req->r_con_filling_msg);
2020                 ceph_msg_revoke_incoming(req->r_reply);
2021                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2022                 req->r_con_filling_msg = NULL;
2023         }
2024
2025         if (front > req->r_reply->front.iov_len) {
2026                 pr_warning("get_reply front %d > preallocated %d\n",
2027                            front, (int)req->r_reply->front.iov_len);
2028                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2029                 if (!m)
2030                         goto out;
2031                 ceph_msg_put(req->r_reply);
2032                 req->r_reply = m;
2033         }
2034         m = ceph_msg_get(req->r_reply);
2035
2036         if (data_len > 0) {
2037                 int want = calc_pages_for(req->r_page_alignment, data_len);
2038
2039                 if (unlikely(req->r_num_pages < want)) {
2040                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2041                                    " had only %d pages ready\n", tid, data_len,
2042                                    want, req->r_num_pages);
2043                         *skip = 1;
2044                         ceph_msg_put(m);
2045                         m = NULL;
2046                         goto out;
2047                 }
2048                 m->pages = req->r_pages;
2049                 m->nr_pages = req->r_num_pages;
2050                 m->page_alignment = req->r_page_alignment;
2051 #ifdef CONFIG_BLOCK
2052                 m->bio = req->r_bio;
2053 #endif
2054         }
2055         *skip = 0;
2056         req->r_con_filling_msg = con->ops->get(con);
2057         dout("get_reply tid %lld %p\n", tid, m);
2058
2059 out:
2060         mutex_unlock(&osdc->request_mutex);
2061         return m;
2062
2063 }
2064
2065 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2066                                   struct ceph_msg_header *hdr,
2067                                   int *skip)
2068 {
2069         struct ceph_osd *osd = con->private;
2070         int type = le16_to_cpu(hdr->type);
2071         int front = le32_to_cpu(hdr->front_len);
2072
2073         *skip = 0;
2074         switch (type) {
2075         case CEPH_MSG_OSD_MAP:
2076         case CEPH_MSG_WATCH_NOTIFY:
2077                 return ceph_msg_new(type, front, GFP_NOFS, false);
2078         case CEPH_MSG_OSD_OPREPLY:
2079                 return get_reply(con, hdr, skip);
2080         default:
2081                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2082                         osd->o_osd);
2083                 *skip = 1;
2084                 return NULL;
2085         }
2086 }
2087
2088 /*
2089  * Wrappers to refcount containing ceph_osd struct
2090  */
2091 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2092 {
2093         struct ceph_osd *osd = con->private;
2094         if (get_osd(osd))
2095                 return con;
2096         return NULL;
2097 }
2098
2099 static void put_osd_con(struct ceph_connection *con)
2100 {
2101         struct ceph_osd *osd = con->private;
2102         put_osd(osd);
2103 }
2104
2105 /*
2106  * authentication
2107  */
2108 /*
2109  * Note: returned pointer is the address of a structure that's
2110  * managed separately.  Caller must *not* attempt to free it.
2111  */
2112 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2113                                         int *proto, 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         struct ceph_auth_handshake *auth = &o->o_auth;
2119
2120         if (force_new && auth->authorizer) {
2121                 if (ac->ops && ac->ops->destroy_authorizer)
2122                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2123                 auth->authorizer = NULL;
2124         }
2125         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2126                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2127                                                         auth);
2128                 if (ret)
2129                         return ERR_PTR(ret);
2130         }
2131         *proto = ac->protocol;
2132
2133         return auth;
2134 }
2135
2136
2137 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2138 {
2139         struct ceph_osd *o = con->private;
2140         struct ceph_osd_client *osdc = o->o_osdc;
2141         struct ceph_auth_client *ac = osdc->client->monc.auth;
2142
2143         /*
2144          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2145          * XXX which do we do:  succeed or fail?
2146          */
2147         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2148 }
2149
2150 static int invalidate_authorizer(struct ceph_connection *con)
2151 {
2152         struct ceph_osd *o = con->private;
2153         struct ceph_osd_client *osdc = o->o_osdc;
2154         struct ceph_auth_client *ac = osdc->client->monc.auth;
2155
2156         if (ac->ops && ac->ops->invalidate_authorizer)
2157                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2158
2159         return ceph_monc_validate_auth(&osdc->client->monc);
2160 }
2161
2162 static const struct ceph_connection_operations osd_con_ops = {
2163         .get = get_osd_con,
2164         .put = put_osd_con,
2165         .dispatch = dispatch,
2166         .get_authorizer = get_authorizer,
2167         .verify_authorizer_reply = verify_authorizer_reply,
2168         .invalidate_authorizer = invalidate_authorizer,
2169         .alloc_msg = alloc_msg,
2170         .fault = osd_reset,
2171 };