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_has_extent(int op)
  36 {
  37         return (op == CEPH_OSD_OP_READ ||
  38                 op == CEPH_OSD_OP_WRITE);
  39 }
  40
  41 /*
  42  * Implement client access to distributed object storage cluster.
  43  *
  44  * All data objects are stored within a cluster/cloud of OSDs, or
  45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  47  * remote daemons serving up and coordinating consistent and safe
  48  * access to storage.
  49  *
  50  * Cluster membership and the mapping of data objects onto storage devices
  51  * are described by the osd map.
  52  *
  53  * We keep track of pending OSD requests (read, write), resubmit
  54  * requests to different OSDs when the cluster topology/data layout
  55  * change, or retry the affected requests when the communications
  56  * channel with an OSD is reset.
  57  */
  58
  59 /*
  60  * calculate the mapping of a file extent onto an object, and fill out the
  61  * request accordingly.  shorten extent as necessary if it crosses an
  62  * object boundary.
  63  *
  64  * fill osd op in request message.
  65  */
  66 static int calc_layout(struct ceph_vino vino,
  67                        struct ceph_file_layout *layout,
  68                        u64 off, u64 *plen,
  69                        struct ceph_osd_request *req,
  70                        struct ceph_osd_req_op *op)
  71 {
  72         u64 orig_len = *plen;
  73         u64 bno = 0;
  74         u64 objoff = 0;
  75         u64 objlen = 0;
  76         int r;
  77
  78         /* object extent? */
  79         r = ceph_calc_file_object_mapping(layout, off, orig_len, &bno,
  80                                           &objoff, &objlen);
  81         if (r < 0)
  82                 return r;
  83         if (objlen < orig_len) {
  84                 *plen = objlen;
  85                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  86                      orig_len - *plen, off, *plen);
  87         }
  88
  89         if (op_has_extent(op->op)) {
  90                 u32 osize = le32_to_cpu(layout->fl_object_size);
  91                 op->extent.offset = objoff;
  92                 op->extent.length = objlen;
  93                 if (op->extent.truncate_size <= off - objoff) {
  94                         op->extent.truncate_size = 0;
  95                 } else {
  96                         op->extent.truncate_size -= off - objoff;
  97                         if (op->extent.truncate_size > osize)
  98                                 op->extent.truncate_size = osize;
  99                 }
 100         }
 101         req->r_num_pages = calc_pages_for(off, *plen);
 102         req->r_page_alignment = off & ~PAGE_MASK;
 103         if (op->op == CEPH_OSD_OP_WRITE)
 104                 op->payload_len = *plen;
 105
 106         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
 107              bno, objoff, objlen, req->r_num_pages);
 108
 109         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 110         req->r_oid_len = strlen(req->r_oid);
 111
 112         return r;
 113 }
 114
 115 /*
 116  * requests
 117  */
 118 void ceph_osdc_release_request(struct kref *kref)
 119 {
 120         struct ceph_osd_request *req = container_of(kref,
 121                                                     struct ceph_osd_request,
 122                                                     r_kref);
 123
 124         if (req->r_request)
 125                 ceph_msg_put(req->r_request);
 126         if (req->r_con_filling_msg) {
 127                 dout("%s revoking msg %p from con %p\n", __func__,
 128                      req->r_reply, req->r_con_filling_msg);
 129                 ceph_msg_revoke_incoming(req->r_reply);
 130                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 131                 req->r_con_filling_msg = NULL;
 132         }
 133         if (req->r_reply)
 134                 ceph_msg_put(req->r_reply);
 135         if (req->r_own_pages)
 136                 ceph_release_page_vector(req->r_pages,
 137                                          req->r_num_pages);
 138         ceph_put_snap_context(req->r_snapc);
 139         ceph_pagelist_release(&req->r_trail);
 140         if (req->r_mempool)
 141                 mempool_free(req, req->r_osdc->req_mempool);
 142         else
 143                 kfree(req);
 144 }
 145 EXPORT_SYMBOL(ceph_osdc_release_request);
 146
 147 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 148                                                struct ceph_snap_context *snapc,
 149                                                unsigned int num_op,
 150                                                bool use_mempool,
 151                                                gfp_t gfp_flags)
 152 {
 153         struct ceph_osd_request *req;
 154         struct ceph_msg *msg;
 155         size_t msg_size = sizeof(struct ceph_osd_request_head);
 156
 157         msg_size += num_op*sizeof(struct ceph_osd_op);
 158
 159         if (use_mempool) {
 160                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 161                 memset(req, 0, sizeof(*req));
 162         } else {
 163                 req = kzalloc(sizeof(*req), gfp_flags);
 164         }
 165         if (req == NULL)
 166                 return NULL;
 167
 168         req->r_osdc = osdc;
 169         req->r_mempool = use_mempool;
 170
 171         kref_init(&req->r_kref);
 172         init_completion(&req->r_completion);
 173         init_completion(&req->r_safe_completion);
 174         RB_CLEAR_NODE(&req->r_node);
 175         INIT_LIST_HEAD(&req->r_unsafe_item);
 176         INIT_LIST_HEAD(&req->r_linger_item);
 177         INIT_LIST_HEAD(&req->r_linger_osd);
 178         INIT_LIST_HEAD(&req->r_req_lru_item);
 179         INIT_LIST_HEAD(&req->r_osd_item);
 180
 181         /* create reply message */
 182         if (use_mempool)
 183                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 184         else
 185                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 186                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 187         if (!msg) {
 188                 ceph_osdc_put_request(req);
 189                 return NULL;
 190         }
 191         req->r_reply = msg;
 192
 193         ceph_pagelist_init(&req->r_trail);
 194
 195         /* create request message; allow space for oid */
 196         msg_size += MAX_OBJ_NAME_SIZE;
 197         if (snapc)
 198                 msg_size += sizeof(u64) * snapc->num_snaps;
 199         if (use_mempool)
 200                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 201         else
 202                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 203         if (!msg) {
 204                 ceph_osdc_put_request(req);
 205                 return NULL;
 206         }
 207
 208         memset(msg->front.iov_base, 0, msg->front.iov_len);
 209
 210         req->r_request = msg;
 211
 212         return req;
 213 }
 214 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 215
 216 static void osd_req_encode_op(struct ceph_osd_request *req,
 217                               struct ceph_osd_op *dst,
 218                               struct ceph_osd_req_op *src)
 219 {
 220         dst->op = cpu_to_le16(src->op);
 221
 222         switch (src->op) {
 223         case CEPH_OSD_OP_STAT:
 224                 break;
 225         case CEPH_OSD_OP_READ:
 226         case CEPH_OSD_OP_WRITE:
 227                 dst->extent.offset =
 228                         cpu_to_le64(src->extent.offset);
 229                 dst->extent.length =
 230                         cpu_to_le64(src->extent.length);
 231                 dst->extent.truncate_size =
 232                         cpu_to_le64(src->extent.truncate_size);
 233                 dst->extent.truncate_seq =
 234                         cpu_to_le32(src->extent.truncate_seq);
 235                 break;
 236         case CEPH_OSD_OP_CALL:
 237                 dst->cls.class_len = src->cls.class_len;
 238                 dst->cls.method_len = src->cls.method_len;
 239                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 240
 241                 ceph_pagelist_append(&req->r_trail, src->cls.class_name,
 242                                      src->cls.class_len);
 243                 ceph_pagelist_append(&req->r_trail, src->cls.method_name,
 244                                      src->cls.method_len);
 245                 ceph_pagelist_append(&req->r_trail, src->cls.indata,
 246                                      src->cls.indata_len);
 247                 break;
 248         case CEPH_OSD_OP_STARTSYNC:
 249                 break;
 250         case CEPH_OSD_OP_NOTIFY_ACK:
 251         case CEPH_OSD_OP_WATCH:
 252                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 253                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 254                 dst->watch.flag = src->watch.flag;
 255                 break;
 256         default:
 257                 pr_err("unrecognized osd opcode %d\n", dst->op);
 258                 WARN_ON(1);
 259                 break;
 260         case CEPH_OSD_OP_MAPEXT:
 261         case CEPH_OSD_OP_MASKTRUNC:
 262         case CEPH_OSD_OP_SPARSE_READ:
 263         case CEPH_OSD_OP_NOTIFY:
 264         case CEPH_OSD_OP_ASSERT_VER:
 265         case CEPH_OSD_OP_WRITEFULL:
 266         case CEPH_OSD_OP_TRUNCATE:
 267         case CEPH_OSD_OP_ZERO:
 268         case CEPH_OSD_OP_DELETE:
 269         case CEPH_OSD_OP_APPEND:
 270         case CEPH_OSD_OP_SETTRUNC:
 271         case CEPH_OSD_OP_TRIMTRUNC:
 272         case CEPH_OSD_OP_TMAPUP:
 273         case CEPH_OSD_OP_TMAPPUT:
 274         case CEPH_OSD_OP_TMAPGET:
 275         case CEPH_OSD_OP_CREATE:
 276         case CEPH_OSD_OP_ROLLBACK:
 277         case CEPH_OSD_OP_OMAPGETKEYS:
 278         case CEPH_OSD_OP_OMAPGETVALS:
 279         case CEPH_OSD_OP_OMAPGETHEADER:
 280         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
 281         case CEPH_OSD_OP_MODE_RD:
 282         case CEPH_OSD_OP_OMAPSETVALS:
 283         case CEPH_OSD_OP_OMAPSETHEADER:
 284         case CEPH_OSD_OP_OMAPCLEAR:
 285         case CEPH_OSD_OP_OMAPRMKEYS:
 286         case CEPH_OSD_OP_OMAP_CMP:
 287         case CEPH_OSD_OP_CLONERANGE:
 288         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
 289         case CEPH_OSD_OP_SRC_CMPXATTR:
 290         case CEPH_OSD_OP_GETXATTR:
 291         case CEPH_OSD_OP_GETXATTRS:
 292         case CEPH_OSD_OP_CMPXATTR:
 293         case CEPH_OSD_OP_SETXATTR:
 294         case CEPH_OSD_OP_SETXATTRS:
 295         case CEPH_OSD_OP_RESETXATTRS:
 296         case CEPH_OSD_OP_RMXATTR:
 297         case CEPH_OSD_OP_PULL:
 298         case CEPH_OSD_OP_PUSH:
 299         case CEPH_OSD_OP_BALANCEREADS:
 300         case CEPH_OSD_OP_UNBALANCEREADS:
 301         case CEPH_OSD_OP_SCRUB:
 302         case CEPH_OSD_OP_SCRUB_RESERVE:
 303         case CEPH_OSD_OP_SCRUB_UNRESERVE:
 304         case CEPH_OSD_OP_SCRUB_STOP:
 305         case CEPH_OSD_OP_SCRUB_MAP:
 306         case CEPH_OSD_OP_WRLOCK:
 307         case CEPH_OSD_OP_WRUNLOCK:
 308         case CEPH_OSD_OP_RDLOCK:
 309         case CEPH_OSD_OP_RDUNLOCK:
 310         case CEPH_OSD_OP_UPLOCK:
 311         case CEPH_OSD_OP_DNLOCK:
 312         case CEPH_OSD_OP_PGLS:
 313         case CEPH_OSD_OP_PGLS_FILTER:
 314                 pr_err("unsupported osd opcode %s\n",
 315                         ceph_osd_op_name(dst->op));
 316                 WARN_ON(1);
 317                 break;
 318         }
 319         dst->payload_len = cpu_to_le32(src->payload_len);
 320 }
 321
 322 /*
 323  * build new request AND message
 324  *
 325  */
 326 void ceph_osdc_build_request(struct ceph_osd_request *req,
 327                              u64 off, u64 len, unsigned int num_op,
 328                              struct ceph_osd_req_op *src_ops,
 329                              struct ceph_snap_context *snapc, u64 snap_id,
 330                              struct timespec *mtime)
 331 {
 332         struct ceph_msg *msg = req->r_request;
 333         struct ceph_osd_request_head *head;
 334         struct ceph_osd_req_op *src_op;
 335         struct ceph_osd_op *op;
 336         void *p;
 337         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 338         int flags = req->r_flags;
 339         u64 data_len;
 340         int i;
 341
 342         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 343
 344         head = msg->front.iov_base;
 345         head->snapid = cpu_to_le64(snap_id);
 346         op = (void *)(head + 1);
 347         p = (void *)(op + num_op);
 348
 349         req->r_snapc = ceph_get_snap_context(snapc);
 350
 351         head->client_inc = cpu_to_le32(1); /* always, for now. */
 352         head->flags = cpu_to_le32(flags);
 353         if (flags & CEPH_OSD_FLAG_WRITE)
 354                 ceph_encode_timespec(&head->mtime, mtime);
 355         BUG_ON(num_op > (unsigned int) ((u16) -1));
 356         head->num_ops = cpu_to_le16(num_op);
 357
 358         /* fill in oid */
 359         head->object_len = cpu_to_le32(req->r_oid_len);
 360         memcpy(p, req->r_oid, req->r_oid_len);
 361         p += req->r_oid_len;
 362
 363         src_op = src_ops;
 364         while (num_op--)
 365                 osd_req_encode_op(req, op++, src_op++);
 366
 367         if (snapc) {
 368                 head->snap_seq = cpu_to_le64(snapc->seq);
 369                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 370                 for (i = 0; i < snapc->num_snaps; i++) {
 371                         put_unaligned_le64(snapc->snaps[i], p);
 372                         p += sizeof(u64);
 373                 }
 374         }
 375
 376         data_len = req->r_trail.length;
 377         if (flags & CEPH_OSD_FLAG_WRITE) {
 378                 req->r_request->hdr.data_off = cpu_to_le16(off);
 379                 data_len += len;
 380         }
 381         req->r_request->hdr.data_len = cpu_to_le32(data_len);
 382         req->r_request->page_alignment = req->r_page_alignment;
 383
 384         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 385         msg_size = p - msg->front.iov_base;
 386         msg->front.iov_len = msg_size;
 387         msg->hdr.front_len = cpu_to_le32(msg_size);
 388         return;
 389 }
 390 EXPORT_SYMBOL(ceph_osdc_build_request);
 391
 392 /*
 393  * build new request AND message, calculate layout, and adjust file
 394  * extent as needed.
 395  *
 396  * if the file was recently truncated, we include information about its
 397  * old and new size so that the object can be updated appropriately.  (we
 398  * avoid synchronously deleting truncated objects because it's slow.)
 399  *
 400  * if @do_sync, include a 'startsync' command so that the osd will flush
 401  * data quickly.
 402  */
 403 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 404                                                struct ceph_file_layout *layout,
 405                                                struct ceph_vino vino,
 406                                                u64 off, u64 *plen,
 407                                                int opcode, int flags,
 408                                                struct ceph_snap_context *snapc,
 409                                                int do_sync,
 410                                                u32 truncate_seq,
 411                                                u64 truncate_size,
 412                                                struct timespec *mtime,
 413                                                bool use_mempool,
 414                                                int page_align)
 415 {
 416         struct ceph_osd_req_op ops[2];
 417         struct ceph_osd_request *req;
 418         unsigned int num_op = 1;
 419         int r;
 420
 421         memset(&ops, 0, sizeof ops);
 422
 423         ops[0].op = opcode;
 424         ops[0].extent.truncate_seq = truncate_seq;
 425         ops[0].extent.truncate_size = truncate_size;
 426
 427         if (do_sync) {
 428                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 429                 num_op++;
 430         }
 431
 432         req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
 433                                         GFP_NOFS);
 434         if (!req)
 435                 return ERR_PTR(-ENOMEM);
 436         req->r_flags = flags;
 437
 438         /* calculate max write size */
 439         r = calc_layout(vino, layout, off, plen, req, ops);
 440         if (r < 0)
 441                 return ERR_PTR(r);
 442         req->r_file_layout = *layout;  /* keep a copy */
 443
 444         /* in case it differs from natural (file) alignment that
 445            calc_layout filled in for us */
 446         req->r_num_pages = calc_pages_for(page_align, *plen);
 447         req->r_page_alignment = page_align;
 448
 449         ceph_osdc_build_request(req, off, *plen, num_op, ops,
 450                                 snapc, vino.snap, mtime);
 451
 452         return req;
 453 }
 454 EXPORT_SYMBOL(ceph_osdc_new_request);
 455
 456 /*
 457  * We keep osd requests in an rbtree, sorted by ->r_tid.
 458  */
 459 static void __insert_request(struct ceph_osd_client *osdc,
 460                              struct ceph_osd_request *new)
 461 {
 462         struct rb_node **p = &osdc->requests.rb_node;
 463         struct rb_node *parent = NULL;
 464         struct ceph_osd_request *req = NULL;
 465
 466         while (*p) {
 467                 parent = *p;
 468                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 469                 if (new->r_tid < req->r_tid)
 470                         p = &(*p)->rb_left;
 471                 else if (new->r_tid > req->r_tid)
 472                         p = &(*p)->rb_right;
 473                 else
 474                         BUG();
 475         }
 476
 477         rb_link_node(&new->r_node, parent, p);
 478         rb_insert_color(&new->r_node, &osdc->requests);
 479 }
 480
 481 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 482                                                  u64 tid)
 483 {
 484         struct ceph_osd_request *req;
 485         struct rb_node *n = osdc->requests.rb_node;
 486
 487         while (n) {
 488                 req = rb_entry(n, struct ceph_osd_request, r_node);
 489                 if (tid < req->r_tid)
 490                         n = n->rb_left;
 491                 else if (tid > req->r_tid)
 492                         n = n->rb_right;
 493                 else
 494                         return req;
 495         }
 496         return NULL;
 497 }
 498
 499 static struct ceph_osd_request *
 500 __lookup_request_ge(struct ceph_osd_client *osdc,
 501                     u64 tid)
 502 {
 503         struct ceph_osd_request *req;
 504         struct rb_node *n = osdc->requests.rb_node;
 505
 506         while (n) {
 507                 req = rb_entry(n, struct ceph_osd_request, r_node);
 508                 if (tid < req->r_tid) {
 509                         if (!n->rb_left)
 510                                 return req;
 511                         n = n->rb_left;
 512                 } else if (tid > req->r_tid) {
 513                         n = n->rb_right;
 514                 } else {
 515                         return req;
 516                 }
 517         }
 518         return NULL;
 519 }
 520
 521 /*
 522  * Resubmit requests pending on the given osd.
 523  */
 524 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 525                                 struct ceph_osd *osd)
 526 {
 527         struct ceph_osd_request *req, *nreq;
 528         int err;
 529
 530         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 531         err = __reset_osd(osdc, osd);
 532         if (err)
 533                 return;
 534
 535         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 536                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 537                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 538                      osd->o_osd);
 539                 if (!req->r_linger)
 540                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 541         }
 542
 543         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 544                                  r_linger_osd) {
 545                 /*
 546                  * reregister request prior to unregistering linger so
 547                  * that r_osd is preserved.
 548                  */
 549                 BUG_ON(!list_empty(&req->r_req_lru_item));
 550                 __register_request(osdc, req);
 551                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 552                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 553                 __unregister_linger_request(osdc, req);
 554                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 555                      osd->o_osd);
 556         }
 557 }
 558
 559 /*
 560  * If the osd connection drops, we need to resubmit all requests.
 561  */
 562 static void osd_reset(struct ceph_connection *con)
 563 {
 564         struct ceph_osd *osd = con->private;
 565         struct ceph_osd_client *osdc;
 566
 567         if (!osd)
 568                 return;
 569         dout("osd_reset osd%d\n", osd->o_osd);
 570         osdc = osd->o_osdc;
 571         down_read(&osdc->map_sem);
 572         mutex_lock(&osdc->request_mutex);
 573         __kick_osd_requests(osdc, osd);
 574         __send_queued(osdc);
 575         mutex_unlock(&osdc->request_mutex);
 576         up_read(&osdc->map_sem);
 577 }
 578
 579 /*
 580  * Track open sessions with osds.
 581  */
 582 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 583 {
 584         struct ceph_osd *osd;
 585
 586         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 587         if (!osd)
 588                 return NULL;
 589
 590         atomic_set(&osd->o_ref, 1);
 591         osd->o_osdc = osdc;
 592         osd->o_osd = onum;
 593         RB_CLEAR_NODE(&osd->o_node);
 594         INIT_LIST_HEAD(&osd->o_requests);
 595         INIT_LIST_HEAD(&osd->o_linger_requests);
 596         INIT_LIST_HEAD(&osd->o_osd_lru);
 597         osd->o_incarnation = 1;
 598
 599         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 600
 601         INIT_LIST_HEAD(&osd->o_keepalive_item);
 602         return osd;
 603 }
 604
 605 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 606 {
 607         if (atomic_inc_not_zero(&osd->o_ref)) {
 608                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 609                      atomic_read(&osd->o_ref));
 610                 return osd;
 611         } else {
 612                 dout("get_osd %p FAIL\n", osd);
 613                 return NULL;
 614         }
 615 }
 616
 617 static void put_osd(struct ceph_osd *osd)
 618 {
 619         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 620              atomic_read(&osd->o_ref) - 1);
 621         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 622                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 623
 624                 if (ac->ops && ac->ops->destroy_authorizer)
 625                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 626                 kfree(osd);
 627         }
 628 }
 629
 630 /*
 631  * remove an osd from our map
 632  */
 633 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 634 {
 635         dout("__remove_osd %p\n", osd);
 636         BUG_ON(!list_empty(&osd->o_requests));
 637         rb_erase(&osd->o_node, &osdc->osds);
 638         list_del_init(&osd->o_osd_lru);
 639         ceph_con_close(&osd->o_con);
 640         put_osd(osd);
 641 }
 642
 643 static void remove_all_osds(struct ceph_osd_client *osdc)
 644 {
 645         dout("%s %p\n", __func__, osdc);
 646         mutex_lock(&osdc->request_mutex);
 647         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 648                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 649                                                 struct ceph_osd, o_node);
 650                 __remove_osd(osdc, osd);
 651         }
 652         mutex_unlock(&osdc->request_mutex);
 653 }
 654
 655 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 656                               struct ceph_osd *osd)
 657 {
 658         dout("__move_osd_to_lru %p\n", osd);
 659         BUG_ON(!list_empty(&osd->o_osd_lru));
 660         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 661         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 662 }
 663
 664 static void __remove_osd_from_lru(struct ceph_osd *osd)
 665 {
 666         dout("__remove_osd_from_lru %p\n", osd);
 667         if (!list_empty(&osd->o_osd_lru))
 668                 list_del_init(&osd->o_osd_lru);
 669 }
 670
 671 static void remove_old_osds(struct ceph_osd_client *osdc)
 672 {
 673         struct ceph_osd *osd, *nosd;
 674
 675         dout("__remove_old_osds %p\n", osdc);
 676         mutex_lock(&osdc->request_mutex);
 677         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 678                 if (time_before(jiffies, osd->lru_ttl))
 679                         break;
 680                 __remove_osd(osdc, osd);
 681         }
 682         mutex_unlock(&osdc->request_mutex);
 683 }
 684
 685 /*
 686  * reset osd connect
 687  */
 688 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 689 {
 690         struct ceph_entity_addr *peer_addr;
 691
 692         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 693         if (list_empty(&osd->o_requests) &&
 694             list_empty(&osd->o_linger_requests)) {
 695                 __remove_osd(osdc, osd);
 696
 697                 return -ENODEV;
 698         }
 699
 700         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
 701         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
 702                         !ceph_con_opened(&osd->o_con)) {
 703                 struct ceph_osd_request *req;
 704
 705                 dout(" osd addr hasn't changed and connection never opened,"
 706                      " letting msgr retry");
 707                 /* touch each r_stamp for handle_timeout()'s benfit */
 708                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 709                         req->r_stamp = jiffies;
 710
 711                 return -EAGAIN;
 712         }
 713
 714         ceph_con_close(&osd->o_con);
 715         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
 716         osd->o_incarnation++;
 717
 718         return 0;
 719 }
 720
 721 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 722 {
 723         struct rb_node **p = &osdc->osds.rb_node;
 724         struct rb_node *parent = NULL;
 725         struct ceph_osd *osd = NULL;
 726
 727         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 728         while (*p) {
 729                 parent = *p;
 730                 osd = rb_entry(parent, struct ceph_osd, o_node);
 731                 if (new->o_osd < osd->o_osd)
 732                         p = &(*p)->rb_left;
 733                 else if (new->o_osd > osd->o_osd)
 734                         p = &(*p)->rb_right;
 735                 else
 736                         BUG();
 737         }
 738
 739         rb_link_node(&new->o_node, parent, p);
 740         rb_insert_color(&new->o_node, &osdc->osds);
 741 }
 742
 743 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 744 {
 745         struct ceph_osd *osd;
 746         struct rb_node *n = osdc->osds.rb_node;
 747
 748         while (n) {
 749                 osd = rb_entry(n, struct ceph_osd, o_node);
 750                 if (o < osd->o_osd)
 751                         n = n->rb_left;
 752                 else if (o > osd->o_osd)
 753                         n = n->rb_right;
 754                 else
 755                         return osd;
 756         }
 757         return NULL;
 758 }
 759
 760 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 761 {
 762         schedule_delayed_work(&osdc->timeout_work,
 763                         osdc->client->options->osd_keepalive_timeout * HZ);
 764 }
 765
 766 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 767 {
 768         cancel_delayed_work(&osdc->timeout_work);
 769 }
 770
 771 /*
 772  * Register request, assign tid.  If this is the first request, set up
 773  * the timeout event.
 774  */
 775 static void __register_request(struct ceph_osd_client *osdc,
 776                                struct ceph_osd_request *req)
 777 {
 778         req->r_tid = ++osdc->last_tid;
 779         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 780         dout("__register_request %p tid %lld\n", req, req->r_tid);
 781         __insert_request(osdc, req);
 782         ceph_osdc_get_request(req);
 783         osdc->num_requests++;
 784         if (osdc->num_requests == 1) {
 785                 dout(" first request, scheduling timeout\n");
 786                 __schedule_osd_timeout(osdc);
 787         }
 788 }
 789
 790 static void register_request(struct ceph_osd_client *osdc,
 791                              struct ceph_osd_request *req)
 792 {
 793         mutex_lock(&osdc->request_mutex);
 794         __register_request(osdc, req);
 795         mutex_unlock(&osdc->request_mutex);
 796 }
 797
 798 /*
 799  * called under osdc->request_mutex
 800  */
 801 static void __unregister_request(struct ceph_osd_client *osdc,
 802                                  struct ceph_osd_request *req)
 803 {
 804         if (RB_EMPTY_NODE(&req->r_node)) {
 805                 dout("__unregister_request %p tid %lld not registered\n",
 806                         req, req->r_tid);
 807                 return;
 808         }
 809
 810         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 811         rb_erase(&req->r_node, &osdc->requests);
 812         osdc->num_requests--;
 813
 814         if (req->r_osd) {
 815                 /* make sure the original request isn't in flight. */
 816                 ceph_msg_revoke(req->r_request);
 817
 818                 list_del_init(&req->r_osd_item);
 819                 if (list_empty(&req->r_osd->o_requests) &&
 820                     list_empty(&req->r_osd->o_linger_requests)) {
 821                         dout("moving osd to %p lru\n", req->r_osd);
 822                         __move_osd_to_lru(osdc, req->r_osd);
 823                 }
 824                 if (list_empty(&req->r_linger_item))
 825                         req->r_osd = NULL;
 826         }
 827
 828         list_del_init(&req->r_req_lru_item);
 829         ceph_osdc_put_request(req);
 830
 831         if (osdc->num_requests == 0) {
 832                 dout(" no requests, canceling timeout\n");
 833                 __cancel_osd_timeout(osdc);
 834         }
 835 }
 836
 837 /*
 838  * Cancel a previously queued request message
 839  */
 840 static void __cancel_request(struct ceph_osd_request *req)
 841 {
 842         if (req->r_sent && req->r_osd) {
 843                 ceph_msg_revoke(req->r_request);
 844                 req->r_sent = 0;
 845         }
 846 }
 847
 848 static void __register_linger_request(struct ceph_osd_client *osdc,
 849                                     struct ceph_osd_request *req)
 850 {
 851         dout("__register_linger_request %p\n", req);
 852         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 853         if (req->r_osd)
 854                 list_add_tail(&req->r_linger_osd,
 855                               &req->r_osd->o_linger_requests);
 856 }
 857
 858 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 859                                         struct ceph_osd_request *req)
 860 {
 861         dout("__unregister_linger_request %p\n", req);
 862         list_del_init(&req->r_linger_item);
 863         if (req->r_osd) {
 864                 list_del_init(&req->r_linger_osd);
 865
 866                 if (list_empty(&req->r_osd->o_requests) &&
 867                     list_empty(&req->r_osd->o_linger_requests)) {
 868                         dout("moving osd to %p lru\n", req->r_osd);
 869                         __move_osd_to_lru(osdc, req->r_osd);
 870                 }
 871                 if (list_empty(&req->r_osd_item))
 872                         req->r_osd = NULL;
 873         }
 874 }
 875
 876 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 877                                          struct ceph_osd_request *req)
 878 {
 879         mutex_lock(&osdc->request_mutex);
 880         if (req->r_linger) {
 881                 __unregister_linger_request(osdc, req);
 882                 ceph_osdc_put_request(req);
 883         }
 884         mutex_unlock(&osdc->request_mutex);
 885 }
 886 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 887
 888 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 889                                   struct ceph_osd_request *req)
 890 {
 891         if (!req->r_linger) {
 892                 dout("set_request_linger %p\n", req);
 893                 req->r_linger = 1;
 894                 /*
 895                  * caller is now responsible for calling
 896                  * unregister_linger_request
 897                  */
 898                 ceph_osdc_get_request(req);
 899         }
 900 }
 901 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 902
 903 /*
 904  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 905  * (as needed), and set the request r_osd appropriately.  If there is
 906  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 907  * (unsent, homeless) or leave on in-flight lru.
 908  *
 909  * Return 0 if unchanged, 1 if changed, or negative on error.
 910  *
 911  * Caller should hold map_sem for read and request_mutex.
 912  */
 913 static int __map_request(struct ceph_osd_client *osdc,
 914                          struct ceph_osd_request *req, int force_resend)
 915 {
 916         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 917         struct ceph_pg_v1 pgid;
 918         int acting[CEPH_PG_MAX_SIZE];
 919         int o = -1, num = 0;
 920         int err;
 921
 922         dout("map_request %p tid %lld\n", req, req->r_tid);
 923         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 924                                       &req->r_file_layout, osdc->osdmap);
 925         if (err) {
 926                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 927                 return err;
 928         }
 929         pgid = reqhead->layout.ol_pgid;
 930         req->r_pgid = pgid;
 931
 932         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 933         if (err > 0) {
 934                 o = acting[0];
 935                 num = err;
 936         }
 937
 938         if ((!force_resend &&
 939              req->r_osd && req->r_osd->o_osd == o &&
 940              req->r_sent >= req->r_osd->o_incarnation &&
 941              req->r_num_pg_osds == num &&
 942              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 943             (req->r_osd == NULL && o == -1))
 944                 return 0;  /* no change */
 945
 946         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 947              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 948              req->r_osd ? req->r_osd->o_osd : -1);
 949
 950         /* record full pg acting set */
 951         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 952         req->r_num_pg_osds = num;
 953
 954         if (req->r_osd) {
 955                 __cancel_request(req);
 956                 list_del_init(&req->r_osd_item);
 957                 req->r_osd = NULL;
 958         }
 959
 960         req->r_osd = __lookup_osd(osdc, o);
 961         if (!req->r_osd && o >= 0) {
 962                 err = -ENOMEM;
 963                 req->r_osd = create_osd(osdc, o);
 964                 if (!req->r_osd) {
 965                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
 966                         goto out;
 967                 }
 968
 969                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
 970                 __insert_osd(osdc, req->r_osd);
 971
 972                 ceph_con_open(&req->r_osd->o_con,
 973                               CEPH_ENTITY_TYPE_OSD, o,
 974                               &osdc->osdmap->osd_addr[o]);
 975         }
 976
 977         if (req->r_osd) {
 978                 __remove_osd_from_lru(req->r_osd);
 979                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 980                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 981         } else {
 982                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 983         }
 984         err = 1;   /* osd or pg changed */
 985
 986 out:
 987         return err;
 988 }
 989
 990 /*
 991  * caller should hold map_sem (for read) and request_mutex
 992  */
 993 static void __send_request(struct ceph_osd_client *osdc,
 994                            struct ceph_osd_request *req)
 995 {
 996         struct ceph_osd_request_head *reqhead;
 997
 998         dout("send_request %p tid %llu to osd%d flags %d\n",
 999              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1000
1001         reqhead = req->r_request->front.iov_base;
1002         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1003         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1004         reqhead->reassert_version = req->r_reassert_version;
1005
1006         req->r_stamp = jiffies;
1007         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1008
1009         ceph_msg_get(req->r_request); /* send consumes a ref */
1010         ceph_con_send(&req->r_osd->o_con, req->r_request);
1011         req->r_sent = req->r_osd->o_incarnation;
1012 }
1013
1014 /*
1015  * Send any requests in the queue (req_unsent).
1016  */
1017 static void __send_queued(struct ceph_osd_client *osdc)
1018 {
1019         struct ceph_osd_request *req, *tmp;
1020
1021         dout("__send_queued\n");
1022         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1023                 __send_request(osdc, req);
1024 }
1025
1026 /*
1027  * Timeout callback, called every N seconds when 1 or more osd
1028  * requests has been active for more than N seconds.  When this
1029  * happens, we ping all OSDs with requests who have timed out to
1030  * ensure any communications channel reset is detected.  Reset the
1031  * request timeouts another N seconds in the future as we go.
1032  * Reschedule the timeout event another N seconds in future (unless
1033  * there are no open requests).
1034  */
1035 static void handle_timeout(struct work_struct *work)
1036 {
1037         struct ceph_osd_client *osdc =
1038                 container_of(work, struct ceph_osd_client, timeout_work.work);
1039         struct ceph_osd_request *req;
1040         struct ceph_osd *osd;
1041         unsigned long keepalive =
1042                 osdc->client->options->osd_keepalive_timeout * HZ;
1043         struct list_head slow_osds;
1044         dout("timeout\n");
1045         down_read(&osdc->map_sem);
1046
1047         ceph_monc_request_next_osdmap(&osdc->client->monc);
1048
1049         mutex_lock(&osdc->request_mutex);
1050
1051         /*
1052          * ping osds that are a bit slow.  this ensures that if there
1053          * is a break in the TCP connection we will notice, and reopen
1054          * a connection with that osd (from the fault callback).
1055          */
1056         INIT_LIST_HEAD(&slow_osds);
1057         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1058                 if (time_before(jiffies, req->r_stamp + keepalive))
1059                         break;
1060
1061                 osd = req->r_osd;
1062                 BUG_ON(!osd);
1063                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1064                      req->r_tid, osd->o_osd);
1065                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1066         }
1067         while (!list_empty(&slow_osds)) {
1068                 osd = list_entry(slow_osds.next, struct ceph_osd,
1069                                  o_keepalive_item);
1070                 list_del_init(&osd->o_keepalive_item);
1071                 ceph_con_keepalive(&osd->o_con);
1072         }
1073
1074         __schedule_osd_timeout(osdc);
1075         __send_queued(osdc);
1076         mutex_unlock(&osdc->request_mutex);
1077         up_read(&osdc->map_sem);
1078 }
1079
1080 static void handle_osds_timeout(struct work_struct *work)
1081 {
1082         struct ceph_osd_client *osdc =
1083                 container_of(work, struct ceph_osd_client,
1084                              osds_timeout_work.work);
1085         unsigned long delay =
1086                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1087
1088         dout("osds timeout\n");
1089         down_read(&osdc->map_sem);
1090         remove_old_osds(osdc);
1091         up_read(&osdc->map_sem);
1092
1093         schedule_delayed_work(&osdc->osds_timeout_work,
1094                               round_jiffies_relative(delay));
1095 }
1096
1097 static void complete_request(struct ceph_osd_request *req)
1098 {
1099         if (req->r_safe_callback)
1100                 req->r_safe_callback(req, NULL);
1101         complete_all(&req->r_safe_completion);  /* fsync waiter */
1102 }
1103
1104 /*
1105  * handle osd op reply.  either call the callback if it is specified,
1106  * or do the completion to wake up the waiting thread.
1107  */
1108 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1109                          struct ceph_connection *con)
1110 {
1111         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1112         struct ceph_osd_request *req;
1113         u64 tid;
1114         int numops, object_len, flags;
1115         s32 result;
1116
1117         tid = le64_to_cpu(msg->hdr.tid);
1118         if (msg->front.iov_len < sizeof(*rhead))
1119                 goto bad;
1120         numops = le32_to_cpu(rhead->num_ops);
1121         object_len = le32_to_cpu(rhead->object_len);
1122         result = le32_to_cpu(rhead->result);
1123         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1124             numops * sizeof(struct ceph_osd_op))
1125                 goto bad;
1126         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1127         /* lookup */
1128         mutex_lock(&osdc->request_mutex);
1129         req = __lookup_request(osdc, tid);
1130         if (req == NULL) {
1131                 dout("handle_reply tid %llu dne\n", tid);
1132                 mutex_unlock(&osdc->request_mutex);
1133                 return;
1134         }
1135         ceph_osdc_get_request(req);
1136         flags = le32_to_cpu(rhead->flags);
1137
1138         /*
1139          * if this connection filled our message, drop our reference now, to
1140          * avoid a (safe but slower) revoke later.
1141          */
1142         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1143                 dout(" dropping con_filling_msg ref %p\n", con);
1144                 req->r_con_filling_msg = NULL;
1145                 con->ops->put(con);
1146         }
1147
1148         if (!req->r_got_reply) {
1149                 unsigned int bytes;
1150
1151                 req->r_result = le32_to_cpu(rhead->result);
1152                 bytes = le32_to_cpu(msg->hdr.data_len);
1153                 dout("handle_reply result %d bytes %d\n", req->r_result,
1154                      bytes);
1155                 if (req->r_result == 0)
1156                         req->r_result = bytes;
1157
1158                 /* in case this is a write and we need to replay, */
1159                 req->r_reassert_version = rhead->reassert_version;
1160
1161                 req->r_got_reply = 1;
1162         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1163                 dout("handle_reply tid %llu dup ack\n", tid);
1164                 mutex_unlock(&osdc->request_mutex);
1165                 goto done;
1166         }
1167
1168         dout("handle_reply tid %llu flags %d\n", tid, flags);
1169
1170         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1171                 __register_linger_request(osdc, req);
1172
1173         /* either this is a read, or we got the safe response */
1174         if (result < 0 ||
1175             (flags & CEPH_OSD_FLAG_ONDISK) ||
1176             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1177                 __unregister_request(osdc, req);
1178
1179         mutex_unlock(&osdc->request_mutex);
1180
1181         if (req->r_callback)
1182                 req->r_callback(req, msg);
1183         else
1184                 complete_all(&req->r_completion);
1185
1186         if (flags & CEPH_OSD_FLAG_ONDISK)
1187                 complete_request(req);
1188
1189 done:
1190         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1191         ceph_osdc_put_request(req);
1192         return;
1193
1194 bad:
1195         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1196                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1197                (int)sizeof(*rhead));
1198         ceph_msg_dump(msg);
1199 }
1200
1201 static void reset_changed_osds(struct ceph_osd_client *osdc)
1202 {
1203         struct rb_node *p, *n;
1204
1205         for (p = rb_first(&osdc->osds); p; p = n) {
1206                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1207
1208                 n = rb_next(p);
1209                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1210                     memcmp(&osd->o_con.peer_addr,
1211                            ceph_osd_addr(osdc->osdmap,
1212                                          osd->o_osd),
1213                            sizeof(struct ceph_entity_addr)) != 0)
1214                         __reset_osd(osdc, osd);
1215         }
1216 }
1217
1218 /*
1219  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1220  * no osd, request a new map.
1221  *
1222  * Caller should hold map_sem for read.
1223  */
1224 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1225 {
1226         struct ceph_osd_request *req, *nreq;
1227         struct rb_node *p;
1228         int needmap = 0;
1229         int err;
1230
1231         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1232         mutex_lock(&osdc->request_mutex);
1233         for (p = rb_first(&osdc->requests); p; ) {
1234                 req = rb_entry(p, struct ceph_osd_request, r_node);
1235                 p = rb_next(p);
1236
1237                 /*
1238                  * For linger requests that have not yet been
1239                  * registered, move them to the linger list; they'll
1240                  * be sent to the osd in the loop below.  Unregister
1241                  * the request before re-registering it as a linger
1242                  * request to ensure the __map_request() below
1243                  * will decide it needs to be sent.
1244                  */
1245                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1246                         dout("%p tid %llu restart on osd%d\n",
1247                              req, req->r_tid,
1248                              req->r_osd ? req->r_osd->o_osd : -1);
1249                         __unregister_request(osdc, req);
1250                         __register_linger_request(osdc, req);
1251                         continue;
1252                 }
1253
1254                 err = __map_request(osdc, req, force_resend);
1255                 if (err < 0)
1256                         continue;  /* error */
1257                 if (req->r_osd == NULL) {
1258                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1259                         needmap++;  /* request a newer map */
1260                 } else if (err > 0) {
1261                         if (!req->r_linger) {
1262                                 dout("%p tid %llu requeued on osd%d\n", req,
1263                                      req->r_tid,
1264                                      req->r_osd ? req->r_osd->o_osd : -1);
1265                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1266                         }
1267                 }
1268         }
1269
1270         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1271                                  r_linger_item) {
1272                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1273
1274                 err = __map_request(osdc, req, force_resend);
1275                 dout("__map_request returned %d\n", err);
1276                 if (err == 0)
1277                         continue;  /* no change and no osd was specified */
1278                 if (err < 0)
1279                         continue;  /* hrm! */
1280                 if (req->r_osd == NULL) {
1281                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1282                         needmap++;  /* request a newer map */
1283                         continue;
1284                 }
1285
1286                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1287                      req->r_osd ? req->r_osd->o_osd : -1);
1288                 __register_request(osdc, req);
1289                 __unregister_linger_request(osdc, req);
1290         }
1291         mutex_unlock(&osdc->request_mutex);
1292
1293         if (needmap) {
1294                 dout("%d requests for down osds, need new map\n", needmap);
1295                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1296         }
1297         reset_changed_osds(osdc);
1298 }
1299
1300
1301 /*
1302  * Process updated osd map.
1303  *
1304  * The message contains any number of incremental and full maps, normally
1305  * indicating some sort of topology change in the cluster.  Kick requests
1306  * off to different OSDs as needed.
1307  */
1308 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1309 {
1310         void *p, *end, *next;
1311         u32 nr_maps, maplen;
1312         u32 epoch;
1313         struct ceph_osdmap *newmap = NULL, *oldmap;
1314         int err;
1315         struct ceph_fsid fsid;
1316
1317         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1318         p = msg->front.iov_base;
1319         end = p + msg->front.iov_len;
1320
1321         /* verify fsid */
1322         ceph_decode_need(&p, end, sizeof(fsid), bad);
1323         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1324         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1325                 return;
1326
1327         down_write(&osdc->map_sem);
1328
1329         /* incremental maps */
1330         ceph_decode_32_safe(&p, end, nr_maps, bad);
1331         dout(" %d inc maps\n", nr_maps);
1332         while (nr_maps > 0) {
1333                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1334                 epoch = ceph_decode_32(&p);
1335                 maplen = ceph_decode_32(&p);
1336                 ceph_decode_need(&p, end, maplen, bad);
1337                 next = p + maplen;
1338                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1339                         dout("applying incremental map %u len %d\n",
1340                              epoch, maplen);
1341                         newmap = osdmap_apply_incremental(&p, next,
1342                                                           osdc->osdmap,
1343                                                           &osdc->client->msgr);
1344                         if (IS_ERR(newmap)) {
1345                                 err = PTR_ERR(newmap);
1346                                 goto bad;
1347                         }
1348                         BUG_ON(!newmap);
1349                         if (newmap != osdc->osdmap) {
1350                                 ceph_osdmap_destroy(osdc->osdmap);
1351                                 osdc->osdmap = newmap;
1352                         }
1353                         kick_requests(osdc, 0);
1354                 } else {
1355                         dout("ignoring incremental map %u len %d\n",
1356                              epoch, maplen);
1357                 }
1358                 p = next;
1359                 nr_maps--;
1360         }
1361         if (newmap)
1362                 goto done;
1363
1364         /* full maps */
1365         ceph_decode_32_safe(&p, end, nr_maps, bad);
1366         dout(" %d full maps\n", nr_maps);
1367         while (nr_maps) {
1368                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1369                 epoch = ceph_decode_32(&p);
1370                 maplen = ceph_decode_32(&p);
1371                 ceph_decode_need(&p, end, maplen, bad);
1372                 if (nr_maps > 1) {
1373                         dout("skipping non-latest full map %u len %d\n",
1374                              epoch, maplen);
1375                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1376                         dout("skipping full map %u len %d, "
1377                              "older than our %u\n", epoch, maplen,
1378                              osdc->osdmap->epoch);
1379                 } else {
1380                         int skipped_map = 0;
1381
1382                         dout("taking full map %u len %d\n", epoch, maplen);
1383                         newmap = osdmap_decode(&p, p+maplen);
1384                         if (IS_ERR(newmap)) {
1385                                 err = PTR_ERR(newmap);
1386                                 goto bad;
1387                         }
1388                         BUG_ON(!newmap);
1389                         oldmap = osdc->osdmap;
1390                         osdc->osdmap = newmap;
1391                         if (oldmap) {
1392                                 if (oldmap->epoch + 1 < newmap->epoch)
1393                                         skipped_map = 1;
1394                                 ceph_osdmap_destroy(oldmap);
1395                         }
1396                         kick_requests(osdc, skipped_map);
1397                 }
1398                 p += maplen;
1399                 nr_maps--;
1400         }
1401
1402 done:
1403         downgrade_write(&osdc->map_sem);
1404         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1405
1406         /*
1407          * subscribe to subsequent osdmap updates if full to ensure
1408          * we find out when we are no longer full and stop returning
1409          * ENOSPC.
1410          */
1411         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1412                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1413
1414         mutex_lock(&osdc->request_mutex);
1415         __send_queued(osdc);
1416         mutex_unlock(&osdc->request_mutex);
1417         up_read(&osdc->map_sem);
1418         wake_up_all(&osdc->client->auth_wq);
1419         return;
1420
1421 bad:
1422         pr_err("osdc handle_map corrupt msg\n");
1423         ceph_msg_dump(msg);
1424         up_write(&osdc->map_sem);
1425         return;
1426 }
1427
1428 /*
1429  * watch/notify callback event infrastructure
1430  *
1431  * These callbacks are used both for watch and notify operations.
1432  */
1433 static void __release_event(struct kref *kref)
1434 {
1435         struct ceph_osd_event *event =
1436                 container_of(kref, struct ceph_osd_event, kref);
1437
1438         dout("__release_event %p\n", event);
1439         kfree(event);
1440 }
1441
1442 static void get_event(struct ceph_osd_event *event)
1443 {
1444         kref_get(&event->kref);
1445 }
1446
1447 void ceph_osdc_put_event(struct ceph_osd_event *event)
1448 {
1449         kref_put(&event->kref, __release_event);
1450 }
1451 EXPORT_SYMBOL(ceph_osdc_put_event);
1452
1453 static void __insert_event(struct ceph_osd_client *osdc,
1454                              struct ceph_osd_event *new)
1455 {
1456         struct rb_node **p = &osdc->event_tree.rb_node;
1457         struct rb_node *parent = NULL;
1458         struct ceph_osd_event *event = NULL;
1459
1460         while (*p) {
1461                 parent = *p;
1462                 event = rb_entry(parent, struct ceph_osd_event, node);
1463                 if (new->cookie < event->cookie)
1464                         p = &(*p)->rb_left;
1465                 else if (new->cookie > event->cookie)
1466                         p = &(*p)->rb_right;
1467                 else
1468                         BUG();
1469         }
1470
1471         rb_link_node(&new->node, parent, p);
1472         rb_insert_color(&new->node, &osdc->event_tree);
1473 }
1474
1475 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1476                                                 u64 cookie)
1477 {
1478         struct rb_node **p = &osdc->event_tree.rb_node;
1479         struct rb_node *parent = NULL;
1480         struct ceph_osd_event *event = NULL;
1481
1482         while (*p) {
1483                 parent = *p;
1484                 event = rb_entry(parent, struct ceph_osd_event, node);
1485                 if (cookie < event->cookie)
1486                         p = &(*p)->rb_left;
1487                 else if (cookie > event->cookie)
1488                         p = &(*p)->rb_right;
1489                 else
1490                         return event;
1491         }
1492         return NULL;
1493 }
1494
1495 static void __remove_event(struct ceph_osd_event *event)
1496 {
1497         struct ceph_osd_client *osdc = event->osdc;
1498
1499         if (!RB_EMPTY_NODE(&event->node)) {
1500                 dout("__remove_event removed %p\n", event);
1501                 rb_erase(&event->node, &osdc->event_tree);
1502                 ceph_osdc_put_event(event);
1503         } else {
1504                 dout("__remove_event didn't remove %p\n", event);
1505         }
1506 }
1507
1508 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1509                            void (*event_cb)(u64, u64, u8, void *),
1510                            void *data, struct ceph_osd_event **pevent)
1511 {
1512         struct ceph_osd_event *event;
1513
1514         event = kmalloc(sizeof(*event), GFP_NOIO);
1515         if (!event)
1516                 return -ENOMEM;
1517
1518         dout("create_event %p\n", event);
1519         event->cb = event_cb;
1520         event->one_shot = 0;
1521         event->data = data;
1522         event->osdc = osdc;
1523         INIT_LIST_HEAD(&event->osd_node);
1524         RB_CLEAR_NODE(&event->node);
1525         kref_init(&event->kref);   /* one ref for us */
1526         kref_get(&event->kref);    /* one ref for the caller */
1527
1528         spin_lock(&osdc->event_lock);
1529         event->cookie = ++osdc->event_count;
1530         __insert_event(osdc, event);
1531         spin_unlock(&osdc->event_lock);
1532
1533         *pevent = event;
1534         return 0;
1535 }
1536 EXPORT_SYMBOL(ceph_osdc_create_event);
1537
1538 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1539 {
1540         struct ceph_osd_client *osdc = event->osdc;
1541
1542         dout("cancel_event %p\n", event);
1543         spin_lock(&osdc->event_lock);
1544         __remove_event(event);
1545         spin_unlock(&osdc->event_lock);
1546         ceph_osdc_put_event(event); /* caller's */
1547 }
1548 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1549
1550
1551 static void do_event_work(struct work_struct *work)
1552 {
1553         struct ceph_osd_event_work *event_work =
1554                 container_of(work, struct ceph_osd_event_work, work);
1555         struct ceph_osd_event *event = event_work->event;
1556         u64 ver = event_work->ver;
1557         u64 notify_id = event_work->notify_id;
1558         u8 opcode = event_work->opcode;
1559
1560         dout("do_event_work completing %p\n", event);
1561         event->cb(ver, notify_id, opcode, event->data);
1562         dout("do_event_work completed %p\n", event);
1563         ceph_osdc_put_event(event);
1564         kfree(event_work);
1565 }
1566
1567
1568 /*
1569  * Process osd watch notifications
1570  */
1571 static void handle_watch_notify(struct ceph_osd_client *osdc,
1572                                 struct ceph_msg *msg)
1573 {
1574         void *p, *end;
1575         u8 proto_ver;
1576         u64 cookie, ver, notify_id;
1577         u8 opcode;
1578         struct ceph_osd_event *event;
1579         struct ceph_osd_event_work *event_work;
1580
1581         p = msg->front.iov_base;
1582         end = p + msg->front.iov_len;
1583
1584         ceph_decode_8_safe(&p, end, proto_ver, bad);
1585         ceph_decode_8_safe(&p, end, opcode, bad);
1586         ceph_decode_64_safe(&p, end, cookie, bad);
1587         ceph_decode_64_safe(&p, end, ver, bad);
1588         ceph_decode_64_safe(&p, end, notify_id, bad);
1589
1590         spin_lock(&osdc->event_lock);
1591         event = __find_event(osdc, cookie);
1592         if (event) {
1593                 BUG_ON(event->one_shot);
1594                 get_event(event);
1595         }
1596         spin_unlock(&osdc->event_lock);
1597         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1598              cookie, ver, event);
1599         if (event) {
1600                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1601                 if (!event_work) {
1602                         dout("ERROR: could not allocate event_work\n");
1603                         goto done_err;
1604                 }
1605                 INIT_WORK(&event_work->work, do_event_work);
1606                 event_work->event = event;
1607                 event_work->ver = ver;
1608                 event_work->notify_id = notify_id;
1609                 event_work->opcode = opcode;
1610                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1611                         dout("WARNING: failed to queue notify event work\n");
1612                         goto done_err;
1613                 }
1614         }
1615
1616         return;
1617
1618 done_err:
1619         ceph_osdc_put_event(event);
1620         return;
1621
1622 bad:
1623         pr_err("osdc handle_watch_notify corrupt msg\n");
1624         return;
1625 }
1626
1627 /*
1628  * Register request, send initial attempt.
1629  */
1630 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1631                             struct ceph_osd_request *req,
1632                             bool nofail)
1633 {
1634         int rc = 0;
1635
1636         req->r_request->pages = req->r_pages;
1637         req->r_request->nr_pages = req->r_num_pages;
1638 #ifdef CONFIG_BLOCK
1639         req->r_request->bio = req->r_bio;
1640 #endif
1641         req->r_request->trail = &req->r_trail;
1642
1643         register_request(osdc, req);
1644
1645         down_read(&osdc->map_sem);
1646         mutex_lock(&osdc->request_mutex);
1647         /*
1648          * a racing kick_requests() may have sent the message for us
1649          * while we dropped request_mutex above, so only send now if
1650          * the request still han't been touched yet.
1651          */
1652         if (req->r_sent == 0) {
1653                 rc = __map_request(osdc, req, 0);
1654                 if (rc < 0) {
1655                         if (nofail) {
1656                                 dout("osdc_start_request failed map, "
1657                                      " will retry %lld\n", req->r_tid);
1658                                 rc = 0;
1659                         }
1660                         goto out_unlock;
1661                 }
1662                 if (req->r_osd == NULL) {
1663                         dout("send_request %p no up osds in pg\n", req);
1664                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1665                 } else {
1666                         __send_request(osdc, req);
1667                 }
1668                 rc = 0;
1669         }
1670
1671 out_unlock:
1672         mutex_unlock(&osdc->request_mutex);
1673         up_read(&osdc->map_sem);
1674         return rc;
1675 }
1676 EXPORT_SYMBOL(ceph_osdc_start_request);
1677
1678 /*
1679  * wait for a request to complete
1680  */
1681 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1682                            struct ceph_osd_request *req)
1683 {
1684         int rc;
1685
1686         rc = wait_for_completion_interruptible(&req->r_completion);
1687         if (rc < 0) {
1688                 mutex_lock(&osdc->request_mutex);
1689                 __cancel_request(req);
1690                 __unregister_request(osdc, req);
1691                 mutex_unlock(&osdc->request_mutex);
1692                 complete_request(req);
1693                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1694                 return rc;
1695         }
1696
1697         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1698         return req->r_result;
1699 }
1700 EXPORT_SYMBOL(ceph_osdc_wait_request);
1701
1702 /*
1703  * sync - wait for all in-flight requests to flush.  avoid starvation.
1704  */
1705 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1706 {
1707         struct ceph_osd_request *req;
1708         u64 last_tid, next_tid = 0;
1709
1710         mutex_lock(&osdc->request_mutex);
1711         last_tid = osdc->last_tid;
1712         while (1) {
1713                 req = __lookup_request_ge(osdc, next_tid);
1714                 if (!req)
1715                         break;
1716                 if (req->r_tid > last_tid)
1717                         break;
1718
1719                 next_tid = req->r_tid + 1;
1720                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1721                         continue;
1722
1723                 ceph_osdc_get_request(req);
1724                 mutex_unlock(&osdc->request_mutex);
1725                 dout("sync waiting on tid %llu (last is %llu)\n",
1726                      req->r_tid, last_tid);
1727                 wait_for_completion(&req->r_safe_completion);
1728                 mutex_lock(&osdc->request_mutex);
1729                 ceph_osdc_put_request(req);
1730         }
1731         mutex_unlock(&osdc->request_mutex);
1732         dout("sync done (thru tid %llu)\n", last_tid);
1733 }
1734 EXPORT_SYMBOL(ceph_osdc_sync);
1735
1736 /*
1737  * init, shutdown
1738  */
1739 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1740 {
1741         int err;
1742
1743         dout("init\n");
1744         osdc->client = client;
1745         osdc->osdmap = NULL;
1746         init_rwsem(&osdc->map_sem);
1747         init_completion(&osdc->map_waiters);
1748         osdc->last_requested_map = 0;
1749         mutex_init(&osdc->request_mutex);
1750         osdc->last_tid = 0;
1751         osdc->osds = RB_ROOT;
1752         INIT_LIST_HEAD(&osdc->osd_lru);
1753         osdc->requests = RB_ROOT;
1754         INIT_LIST_HEAD(&osdc->req_lru);
1755         INIT_LIST_HEAD(&osdc->req_unsent);
1756         INIT_LIST_HEAD(&osdc->req_notarget);
1757         INIT_LIST_HEAD(&osdc->req_linger);
1758         osdc->num_requests = 0;
1759         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1760         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1761         spin_lock_init(&osdc->event_lock);
1762         osdc->event_tree = RB_ROOT;
1763         osdc->event_count = 0;
1764
1765         schedule_delayed_work(&osdc->osds_timeout_work,
1766            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1767
1768         err = -ENOMEM;
1769         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1770                                         sizeof(struct ceph_osd_request));
1771         if (!osdc->req_mempool)
1772                 goto out;
1773
1774         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1775                                 OSD_OP_FRONT_LEN, 10, true,
1776                                 "osd_op");
1777         if (err < 0)
1778                 goto out_mempool;
1779         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1780                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1781                                 "osd_op_reply");
1782         if (err < 0)
1783                 goto out_msgpool;
1784
1785         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1786         if (IS_ERR(osdc->notify_wq)) {
1787                 err = PTR_ERR(osdc->notify_wq);
1788                 osdc->notify_wq = NULL;
1789                 goto out_msgpool;
1790         }
1791         return 0;
1792
1793 out_msgpool:
1794         ceph_msgpool_destroy(&osdc->msgpool_op);
1795 out_mempool:
1796         mempool_destroy(osdc->req_mempool);
1797 out:
1798         return err;
1799 }
1800
1801 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1802 {
1803         flush_workqueue(osdc->notify_wq);
1804         destroy_workqueue(osdc->notify_wq);
1805         cancel_delayed_work_sync(&osdc->timeout_work);
1806         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1807         if (osdc->osdmap) {
1808                 ceph_osdmap_destroy(osdc->osdmap);
1809                 osdc->osdmap = NULL;
1810         }
1811         remove_all_osds(osdc);
1812         mempool_destroy(osdc->req_mempool);
1813         ceph_msgpool_destroy(&osdc->msgpool_op);
1814         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1815 }
1816
1817 /*
1818  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1819  * *plen.  Return number of bytes read, or error.
1820  */
1821 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1822                         struct ceph_vino vino, struct ceph_file_layout *layout,
1823                         u64 off, u64 *plen,
1824                         u32 truncate_seq, u64 truncate_size,
1825                         struct page **pages, int num_pages, int page_align)
1826 {
1827         struct ceph_osd_request *req;
1828         int rc = 0;
1829
1830         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1831              vino.snap, off, *plen);
1832         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1833                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1834                                     NULL, 0, truncate_seq, truncate_size, NULL,
1835                                     false, page_align);
1836         if (IS_ERR(req))
1837                 return PTR_ERR(req);
1838
1839         /* it may be a short read due to an object boundary */
1840         req->r_pages = pages;
1841
1842         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1843              off, *plen, req->r_num_pages, page_align);
1844
1845         rc = ceph_osdc_start_request(osdc, req, false);
1846         if (!rc)
1847                 rc = ceph_osdc_wait_request(osdc, req);
1848
1849         ceph_osdc_put_request(req);
1850         dout("readpages result %d\n", rc);
1851         return rc;
1852 }
1853 EXPORT_SYMBOL(ceph_osdc_readpages);
1854
1855 /*
1856  * do a synchronous write on N pages
1857  */
1858 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1859                          struct ceph_file_layout *layout,
1860                          struct ceph_snap_context *snapc,
1861                          u64 off, u64 len,
1862                          u32 truncate_seq, u64 truncate_size,
1863                          struct timespec *mtime,
1864                          struct page **pages, int num_pages)
1865 {
1866         struct ceph_osd_request *req;
1867         int rc = 0;
1868         int page_align = off & ~PAGE_MASK;
1869
1870         BUG_ON(vino.snap != CEPH_NOSNAP);
1871         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1872                                     CEPH_OSD_OP_WRITE,
1873                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1874                                     snapc, 0,
1875                                     truncate_seq, truncate_size, mtime,
1876                                     true, page_align);
1877         if (IS_ERR(req))
1878                 return PTR_ERR(req);
1879
1880         /* it may be a short write due to an object boundary */
1881         req->r_pages = pages;
1882         dout("writepages %llu~%llu (%d pages)\n", off, len,
1883              req->r_num_pages);
1884
1885         rc = ceph_osdc_start_request(osdc, req, true);
1886         if (!rc)
1887                 rc = ceph_osdc_wait_request(osdc, req);
1888
1889         ceph_osdc_put_request(req);
1890         if (rc == 0)
1891                 rc = len;
1892         dout("writepages result %d\n", rc);
1893         return rc;
1894 }
1895 EXPORT_SYMBOL(ceph_osdc_writepages);
1896
1897 /*
1898  * handle incoming message
1899  */
1900 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1901 {
1902         struct ceph_osd *osd = con->private;
1903         struct ceph_osd_client *osdc;
1904         int type = le16_to_cpu(msg->hdr.type);
1905
1906         if (!osd)
1907                 goto out;
1908         osdc = osd->o_osdc;
1909
1910         switch (type) {
1911         case CEPH_MSG_OSD_MAP:
1912                 ceph_osdc_handle_map(osdc, msg);
1913                 break;
1914         case CEPH_MSG_OSD_OPREPLY:
1915                 handle_reply(osdc, msg, con);
1916                 break;
1917         case CEPH_MSG_WATCH_NOTIFY:
1918                 handle_watch_notify(osdc, msg);
1919                 break;
1920
1921         default:
1922                 pr_err("received unknown message type %d %s\n", type,
1923                        ceph_msg_type_name(type));
1924         }
1925 out:
1926         ceph_msg_put(msg);
1927 }
1928
1929 /*
1930  * lookup and return message for incoming reply.  set up reply message
1931  * pages.
1932  */
1933 static struct ceph_msg *get_reply(struct ceph_connection *con,
1934                                   struct ceph_msg_header *hdr,
1935                                   int *skip)
1936 {
1937         struct ceph_osd *osd = con->private;
1938         struct ceph_osd_client *osdc = osd->o_osdc;
1939         struct ceph_msg *m;
1940         struct ceph_osd_request *req;
1941         int front = le32_to_cpu(hdr->front_len);
1942         int data_len = le32_to_cpu(hdr->data_len);
1943         u64 tid;
1944
1945         tid = le64_to_cpu(hdr->tid);
1946         mutex_lock(&osdc->request_mutex);
1947         req = __lookup_request(osdc, tid);
1948         if (!req) {
1949                 *skip = 1;
1950                 m = NULL;
1951                 dout("get_reply unknown tid %llu from osd%d\n", tid,
1952                      osd->o_osd);
1953                 goto out;
1954         }
1955
1956         if (req->r_con_filling_msg) {
1957                 dout("%s revoking msg %p from old con %p\n", __func__,
1958                      req->r_reply, req->r_con_filling_msg);
1959                 ceph_msg_revoke_incoming(req->r_reply);
1960                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
1961                 req->r_con_filling_msg = NULL;
1962         }
1963
1964         if (front > req->r_reply->front.iov_len) {
1965                 pr_warning("get_reply front %d > preallocated %d\n",
1966                            front, (int)req->r_reply->front.iov_len);
1967                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
1968                 if (!m)
1969                         goto out;
1970                 ceph_msg_put(req->r_reply);
1971                 req->r_reply = m;
1972         }
1973         m = ceph_msg_get(req->r_reply);
1974
1975         if (data_len > 0) {
1976                 int want = calc_pages_for(req->r_page_alignment, data_len);
1977
1978                 if (req->r_pages && unlikely(req->r_num_pages < want)) {
1979                         pr_warning("tid %lld reply has %d bytes %d pages, we"
1980                                    " had only %d pages ready\n", tid, data_len,
1981                                    want, req->r_num_pages);
1982                         *skip = 1;
1983                         ceph_msg_put(m);
1984                         m = NULL;
1985                         goto out;
1986                 }
1987                 m->pages = req->r_pages;
1988                 m->nr_pages = req->r_num_pages;
1989                 m->page_alignment = req->r_page_alignment;
1990 #ifdef CONFIG_BLOCK
1991                 m->bio = req->r_bio;
1992 #endif
1993         }
1994         *skip = 0;
1995         req->r_con_filling_msg = con->ops->get(con);
1996         dout("get_reply tid %lld %p\n", tid, m);
1997
1998 out:
1999         mutex_unlock(&osdc->request_mutex);
2000         return m;
2001
2002 }
2003
2004 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2005                                   struct ceph_msg_header *hdr,
2006                                   int *skip)
2007 {
2008         struct ceph_osd *osd = con->private;
2009         int type = le16_to_cpu(hdr->type);
2010         int front = le32_to_cpu(hdr->front_len);
2011
2012         *skip = 0;
2013         switch (type) {
2014         case CEPH_MSG_OSD_MAP:
2015         case CEPH_MSG_WATCH_NOTIFY:
2016                 return ceph_msg_new(type, front, GFP_NOFS, false);
2017         case CEPH_MSG_OSD_OPREPLY:
2018                 return get_reply(con, hdr, skip);
2019         default:
2020                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2021                         osd->o_osd);
2022                 *skip = 1;
2023                 return NULL;
2024         }
2025 }
2026
2027 /*
2028  * Wrappers to refcount containing ceph_osd struct
2029  */
2030 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2031 {
2032         struct ceph_osd *osd = con->private;
2033         if (get_osd(osd))
2034                 return con;
2035         return NULL;
2036 }
2037
2038 static void put_osd_con(struct ceph_connection *con)
2039 {
2040         struct ceph_osd *osd = con->private;
2041         put_osd(osd);
2042 }
2043
2044 /*
2045  * authentication
2046  */
2047 /*
2048  * Note: returned pointer is the address of a structure that's
2049  * managed separately.  Caller must *not* attempt to free it.
2050  */
2051 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2052                                         int *proto, int force_new)
2053 {
2054         struct ceph_osd *o = con->private;
2055         struct ceph_osd_client *osdc = o->o_osdc;
2056         struct ceph_auth_client *ac = osdc->client->monc.auth;
2057         struct ceph_auth_handshake *auth = &o->o_auth;
2058
2059         if (force_new && auth->authorizer) {
2060                 if (ac->ops && ac->ops->destroy_authorizer)
2061                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2062                 auth->authorizer = NULL;
2063         }
2064         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2065                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2066                                                         auth);
2067                 if (ret)
2068                         return ERR_PTR(ret);
2069         }
2070         *proto = ac->protocol;
2071
2072         return auth;
2073 }
2074
2075
2076 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2077 {
2078         struct ceph_osd *o = con->private;
2079         struct ceph_osd_client *osdc = o->o_osdc;
2080         struct ceph_auth_client *ac = osdc->client->monc.auth;
2081
2082         /*
2083          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2084          * XXX which do we do:  succeed or fail?
2085          */
2086         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2087 }
2088
2089 static int invalidate_authorizer(struct ceph_connection *con)
2090 {
2091         struct ceph_osd *o = con->private;
2092         struct ceph_osd_client *osdc = o->o_osdc;
2093         struct ceph_auth_client *ac = osdc->client->monc.auth;
2094
2095         if (ac->ops && ac->ops->invalidate_authorizer)
2096                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2097
2098         return ceph_monc_validate_auth(&osdc->client->monc);
2099 }
2100
2101 static const struct ceph_connection_operations osd_con_ops = {
2102         .get = get_osd_con,
2103         .put = put_osd_con,
2104         .dispatch = dispatch,
2105         .get_authorizer = get_authorizer,
2106         .verify_authorizer_reply = verify_authorizer_reply,
2107         .invalidate_authorizer = invalidate_authorizer,
2108         .alloc_msg = alloc_msg,
2109         .fault = osd_reset,
2110 };