net/ceph/osd_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/module.h>
   4 #include <linux/err.h>
   5 #include <linux/highmem.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/slab.h>
   9 #include <linux/uaccess.h>
  10 #ifdef CONFIG_BLOCK
  11 #include <linux/bio.h>
  12 #endif
  13
  14 #include <linux/ceph/libceph.h>
  15 #include <linux/ceph/osd_client.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/auth.h>
  19 #include <linux/ceph/pagelist.h>
  20
  21 #define OSD_OP_FRONT_LEN        4096
  22 #define OSD_OPREPLY_FRONT_LEN   512
  23
  24 static const struct ceph_connection_operations osd_con_ops;
  25
  26 static void send_queued(struct ceph_osd_client *osdc);
  27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  28 static void __register_request(struct ceph_osd_client *osdc,
  29                                struct ceph_osd_request *req);
  30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  31                                         struct ceph_osd_request *req);
  32 static void __send_request(struct ceph_osd_client *osdc,
  33                            struct ceph_osd_request *req);
  34
  35 static int op_needs_trail(int op)
  36 {
  37         switch (op) {
  38         case CEPH_OSD_OP_GETXATTR:
  39         case CEPH_OSD_OP_SETXATTR:
  40         case CEPH_OSD_OP_CMPXATTR:
  41         case CEPH_OSD_OP_CALL:
  42         case CEPH_OSD_OP_NOTIFY:
  43                 return 1;
  44         default:
  45                 return 0;
  46         }
  47 }
  48
  49 static int op_has_extent(int op)
  50 {
  51         return (op == CEPH_OSD_OP_READ ||
  52                 op == CEPH_OSD_OP_WRITE);
  53 }
  54
  55 void ceph_calc_raw_layout(struct ceph_osd_client *osdc,
  56                         struct ceph_file_layout *layout,
  57                         u64 snapid,
  58                         u64 off, u64 *plen, u64 *bno,
  59                         struct ceph_osd_request *req,
  60                         struct ceph_osd_req_op *op)
  61 {
  62         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
  63         u64 orig_len = *plen;
  64         u64 objoff, objlen;    /* extent in object */
  65
  66         reqhead->snapid = cpu_to_le64(snapid);
  67
  68         /* object extent? */
  69         ceph_calc_file_object_mapping(layout, off, plen, bno,
  70                                       &objoff, &objlen);
  71         if (*plen < orig_len)
  72                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  73                      orig_len - *plen, off, *plen);
  74
  75         if (op_has_extent(op->op)) {
  76                 op->extent.offset = objoff;
  77                 op->extent.length = objlen;
  78         }
  79         req->r_num_pages = calc_pages_for(off, *plen);
  80         req->r_page_alignment = off & ~PAGE_MASK;
  81         if (op->op == CEPH_OSD_OP_WRITE)
  82                 op->payload_len = *plen;
  83
  84         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
  85              *bno, objoff, objlen, req->r_num_pages);
  86
  87 }
  88 EXPORT_SYMBOL(ceph_calc_raw_layout);
  89
  90 /*
  91  * Implement client access to distributed object storage cluster.
  92  *
  93  * All data objects are stored within a cluster/cloud of OSDs, or
  94  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  95  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  96  * remote daemons serving up and coordinating consistent and safe
  97  * access to storage.
  98  *
  99  * Cluster membership and the mapping of data objects onto storage devices
 100  * are described by the osd map.
 101  *
 102  * We keep track of pending OSD requests (read, write), resubmit
 103  * requests to different OSDs when the cluster topology/data layout
 104  * change, or retry the affected requests when the communications
 105  * channel with an OSD is reset.
 106  */
 107
 108 /*
 109  * calculate the mapping of a file extent onto an object, and fill out the
 110  * request accordingly.  shorten extent as necessary if it crosses an
 111  * object boundary.
 112  *
 113  * fill osd op in request message.
 114  */
 115 static void calc_layout(struct ceph_osd_client *osdc,
 116                         struct ceph_vino vino,
 117                         struct ceph_file_layout *layout,
 118                         u64 off, u64 *plen,
 119                         struct ceph_osd_request *req,
 120                         struct ceph_osd_req_op *op)
 121 {
 122         u64 bno;
 123
 124         ceph_calc_raw_layout(osdc, layout, vino.snap, off,
 125                              plen, &bno, req, op);
 126
 127         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 128         req->r_oid_len = strlen(req->r_oid);
 129 }
 130
 131 /*
 132  * requests
 133  */
 134 void ceph_osdc_release_request(struct kref *kref)
 135 {
 136         struct ceph_osd_request *req = container_of(kref,
 137                                                     struct ceph_osd_request,
 138                                                     r_kref);
 139
 140         if (req->r_request)
 141                 ceph_msg_put(req->r_request);
 142         if (req->r_con_filling_msg) {
 143                 dout("%s revoking pages %p from con %p\n", __func__,
 144                      req->r_pages, req->r_con_filling_msg);
 145                 ceph_msg_revoke_incoming(req->r_reply);
 146                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 147         }
 148         if (req->r_reply)
 149                 ceph_msg_put(req->r_reply);
 150         if (req->r_own_pages)
 151                 ceph_release_page_vector(req->r_pages,
 152                                          req->r_num_pages);
 153 #ifdef CONFIG_BLOCK
 154         if (req->r_bio)
 155                 bio_put(req->r_bio);
 156 #endif
 157         ceph_put_snap_context(req->r_snapc);
 158         if (req->r_trail) {
 159                 ceph_pagelist_release(req->r_trail);
 160                 kfree(req->r_trail);
 161         }
 162         if (req->r_mempool)
 163                 mempool_free(req, req->r_osdc->req_mempool);
 164         else
 165                 kfree(req);
 166 }
 167 EXPORT_SYMBOL(ceph_osdc_release_request);
 168
 169 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
 170 {
 171         int i = 0;
 172
 173         if (needs_trail)
 174                 *needs_trail = 0;
 175         while (ops[i].op) {
 176                 if (needs_trail && op_needs_trail(ops[i].op))
 177                         *needs_trail = 1;
 178                 i++;
 179         }
 180
 181         return i;
 182 }
 183
 184 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 185                                                int flags,
 186                                                struct ceph_snap_context *snapc,
 187                                                struct ceph_osd_req_op *ops,
 188                                                bool use_mempool,
 189                                                gfp_t gfp_flags,
 190                                                struct page **pages,
 191                                                struct bio *bio)
 192 {
 193         struct ceph_osd_request *req;
 194         struct ceph_msg *msg;
 195         int needs_trail;
 196         int num_op = get_num_ops(ops, &needs_trail);
 197         size_t msg_size = sizeof(struct ceph_osd_request_head);
 198
 199         msg_size += num_op*sizeof(struct ceph_osd_op);
 200
 201         if (use_mempool) {
 202                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 203                 memset(req, 0, sizeof(*req));
 204         } else {
 205                 req = kzalloc(sizeof(*req), gfp_flags);
 206         }
 207         if (req == NULL)
 208                 return NULL;
 209
 210         req->r_osdc = osdc;
 211         req->r_mempool = use_mempool;
 212
 213         kref_init(&req->r_kref);
 214         init_completion(&req->r_completion);
 215         init_completion(&req->r_safe_completion);
 216         INIT_LIST_HEAD(&req->r_unsafe_item);
 217         INIT_LIST_HEAD(&req->r_linger_item);
 218         INIT_LIST_HEAD(&req->r_linger_osd);
 219         INIT_LIST_HEAD(&req->r_req_lru_item);
 220         req->r_flags = flags;
 221
 222         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 223
 224         /* create reply message */
 225         if (use_mempool)
 226                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 227         else
 228                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 229                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 230         if (!msg) {
 231                 ceph_osdc_put_request(req);
 232                 return NULL;
 233         }
 234         req->r_reply = msg;
 235
 236         /* allocate space for the trailing data */
 237         if (needs_trail) {
 238                 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
 239                 if (!req->r_trail) {
 240                         ceph_osdc_put_request(req);
 241                         return NULL;
 242                 }
 243                 ceph_pagelist_init(req->r_trail);
 244         }
 245         /* create request message; allow space for oid */
 246         msg_size += MAX_OBJ_NAME_SIZE;
 247         if (snapc)
 248                 msg_size += sizeof(u64) * snapc->num_snaps;
 249         if (use_mempool)
 250                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 251         else
 252                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 253         if (!msg) {
 254                 ceph_osdc_put_request(req);
 255                 return NULL;
 256         }
 257
 258         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
 259         memset(msg->front.iov_base, 0, msg->front.iov_len);
 260
 261         req->r_request = msg;
 262         req->r_pages = pages;
 263 #ifdef CONFIG_BLOCK
 264         if (bio) {
 265                 req->r_bio = bio;
 266                 bio_get(req->r_bio);
 267         }
 268 #endif
 269
 270         return req;
 271 }
 272 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 273
 274 static void osd_req_encode_op(struct ceph_osd_request *req,
 275                               struct ceph_osd_op *dst,
 276                               struct ceph_osd_req_op *src)
 277 {
 278         dst->op = cpu_to_le16(src->op);
 279
 280         switch (src->op) {
 281         case CEPH_OSD_OP_READ:
 282         case CEPH_OSD_OP_WRITE:
 283                 dst->extent.offset =
 284                         cpu_to_le64(src->extent.offset);
 285                 dst->extent.length =
 286                         cpu_to_le64(src->extent.length);
 287                 dst->extent.truncate_size =
 288                         cpu_to_le64(src->extent.truncate_size);
 289                 dst->extent.truncate_seq =
 290                         cpu_to_le32(src->extent.truncate_seq);
 291                 break;
 292
 293         case CEPH_OSD_OP_GETXATTR:
 294         case CEPH_OSD_OP_SETXATTR:
 295         case CEPH_OSD_OP_CMPXATTR:
 296                 BUG_ON(!req->r_trail);
 297
 298                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 299                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 300                 dst->xattr.cmp_op = src->xattr.cmp_op;
 301                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 302                 ceph_pagelist_append(req->r_trail, src->xattr.name,
 303                                      src->xattr.name_len);
 304                 ceph_pagelist_append(req->r_trail, src->xattr.val,
 305                                      src->xattr.value_len);
 306                 break;
 307         case CEPH_OSD_OP_CALL:
 308                 BUG_ON(!req->r_trail);
 309
 310                 dst->cls.class_len = src->cls.class_len;
 311                 dst->cls.method_len = src->cls.method_len;
 312                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 313
 314                 ceph_pagelist_append(req->r_trail, src->cls.class_name,
 315                                      src->cls.class_len);
 316                 ceph_pagelist_append(req->r_trail, src->cls.method_name,
 317                                      src->cls.method_len);
 318                 ceph_pagelist_append(req->r_trail, src->cls.indata,
 319                                      src->cls.indata_len);
 320                 break;
 321         case CEPH_OSD_OP_ROLLBACK:
 322                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
 323                 break;
 324         case CEPH_OSD_OP_STARTSYNC:
 325                 break;
 326         case CEPH_OSD_OP_NOTIFY:
 327                 {
 328                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
 329                         __le32 timeout = cpu_to_le32(src->watch.timeout);
 330
 331                         BUG_ON(!req->r_trail);
 332
 333                         ceph_pagelist_append(req->r_trail,
 334                                                 &prot_ver, sizeof(prot_ver));
 335                         ceph_pagelist_append(req->r_trail,
 336                                                 &timeout, sizeof(timeout));
 337                 }
 338         case CEPH_OSD_OP_NOTIFY_ACK:
 339         case CEPH_OSD_OP_WATCH:
 340                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 341                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 342                 dst->watch.flag = src->watch.flag;
 343                 break;
 344         default:
 345                 pr_err("unrecognized osd opcode %d\n", dst->op);
 346                 WARN_ON(1);
 347                 break;
 348         }
 349         dst->payload_len = cpu_to_le32(src->payload_len);
 350 }
 351
 352 /*
 353  * build new request AND message
 354  *
 355  */
 356 void ceph_osdc_build_request(struct ceph_osd_request *req,
 357                              u64 off, u64 *plen,
 358                              struct ceph_osd_req_op *src_ops,
 359                              struct ceph_snap_context *snapc,
 360                              struct timespec *mtime,
 361                              const char *oid,
 362                              int oid_len)
 363 {
 364         struct ceph_msg *msg = req->r_request;
 365         struct ceph_osd_request_head *head;
 366         struct ceph_osd_req_op *src_op;
 367         struct ceph_osd_op *op;
 368         void *p;
 369         int num_op = get_num_ops(src_ops, NULL);
 370         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 371         int flags = req->r_flags;
 372         u64 data_len = 0;
 373         int i;
 374
 375         head = msg->front.iov_base;
 376         op = (void *)(head + 1);
 377         p = (void *)(op + num_op);
 378
 379         req->r_snapc = ceph_get_snap_context(snapc);
 380
 381         head->client_inc = cpu_to_le32(1); /* always, for now. */
 382         head->flags = cpu_to_le32(flags);
 383         if (flags & CEPH_OSD_FLAG_WRITE)
 384                 ceph_encode_timespec(&head->mtime, mtime);
 385         head->num_ops = cpu_to_le16(num_op);
 386
 387
 388         /* fill in oid */
 389         head->object_len = cpu_to_le32(oid_len);
 390         memcpy(p, oid, oid_len);
 391         p += oid_len;
 392
 393         src_op = src_ops;
 394         while (src_op->op) {
 395                 osd_req_encode_op(req, op, src_op);
 396                 src_op++;
 397                 op++;
 398         }
 399
 400         if (req->r_trail)
 401                 data_len += req->r_trail->length;
 402
 403         if (snapc) {
 404                 head->snap_seq = cpu_to_le64(snapc->seq);
 405                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 406                 for (i = 0; i < snapc->num_snaps; i++) {
 407                         put_unaligned_le64(snapc->snaps[i], p);
 408                         p += sizeof(u64);
 409                 }
 410         }
 411
 412         if (flags & CEPH_OSD_FLAG_WRITE) {
 413                 req->r_request->hdr.data_off = cpu_to_le16(off);
 414                 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
 415         } else if (data_len) {
 416                 req->r_request->hdr.data_off = 0;
 417                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 418         }
 419
 420         req->r_request->page_alignment = req->r_page_alignment;
 421
 422         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 423         msg_size = p - msg->front.iov_base;
 424         msg->front.iov_len = msg_size;
 425         msg->hdr.front_len = cpu_to_le32(msg_size);
 426         return;
 427 }
 428 EXPORT_SYMBOL(ceph_osdc_build_request);
 429
 430 /*
 431  * build new request AND message, calculate layout, and adjust file
 432  * extent as needed.
 433  *
 434  * if the file was recently truncated, we include information about its
 435  * old and new size so that the object can be updated appropriately.  (we
 436  * avoid synchronously deleting truncated objects because it's slow.)
 437  *
 438  * if @do_sync, include a 'startsync' command so that the osd will flush
 439  * data quickly.
 440  */
 441 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 442                                                struct ceph_file_layout *layout,
 443                                                struct ceph_vino vino,
 444                                                u64 off, u64 *plen,
 445                                                int opcode, int flags,
 446                                                struct ceph_snap_context *snapc,
 447                                                int do_sync,
 448                                                u32 truncate_seq,
 449                                                u64 truncate_size,
 450                                                struct timespec *mtime,
 451                                                bool use_mempool, int num_reply,
 452                                                int page_align)
 453 {
 454         struct ceph_osd_req_op ops[3];
 455         struct ceph_osd_request *req;
 456
 457         ops[0].op = opcode;
 458         ops[0].extent.truncate_seq = truncate_seq;
 459         ops[0].extent.truncate_size = truncate_size;
 460         ops[0].payload_len = 0;
 461
 462         if (do_sync) {
 463                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 464                 ops[1].payload_len = 0;
 465                 ops[2].op = 0;
 466         } else
 467                 ops[1].op = 0;
 468
 469         req = ceph_osdc_alloc_request(osdc, flags,
 470                                          snapc, ops,
 471                                          use_mempool,
 472                                          GFP_NOFS, NULL, NULL);
 473         if (!req)
 474                 return NULL;
 475
 476         /* calculate max write size */
 477         calc_layout(osdc, vino, layout, off, plen, req, ops);
 478         req->r_file_layout = *layout;  /* keep a copy */
 479
 480         /* in case it differs from natural (file) alignment that
 481            calc_layout filled in for us */
 482         req->r_num_pages = calc_pages_for(page_align, *plen);
 483         req->r_page_alignment = page_align;
 484
 485         ceph_osdc_build_request(req, off, plen, ops,
 486                                 snapc,
 487                                 mtime,
 488                                 req->r_oid, req->r_oid_len);
 489
 490         return req;
 491 }
 492 EXPORT_SYMBOL(ceph_osdc_new_request);
 493
 494 /*
 495  * We keep osd requests in an rbtree, sorted by ->r_tid.
 496  */
 497 static void __insert_request(struct ceph_osd_client *osdc,
 498                              struct ceph_osd_request *new)
 499 {
 500         struct rb_node **p = &osdc->requests.rb_node;
 501         struct rb_node *parent = NULL;
 502         struct ceph_osd_request *req = NULL;
 503
 504         while (*p) {
 505                 parent = *p;
 506                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 507                 if (new->r_tid < req->r_tid)
 508                         p = &(*p)->rb_left;
 509                 else if (new->r_tid > req->r_tid)
 510                         p = &(*p)->rb_right;
 511                 else
 512                         BUG();
 513         }
 514
 515         rb_link_node(&new->r_node, parent, p);
 516         rb_insert_color(&new->r_node, &osdc->requests);
 517 }
 518
 519 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 520                                                  u64 tid)
 521 {
 522         struct ceph_osd_request *req;
 523         struct rb_node *n = osdc->requests.rb_node;
 524
 525         while (n) {
 526                 req = rb_entry(n, struct ceph_osd_request, r_node);
 527                 if (tid < req->r_tid)
 528                         n = n->rb_left;
 529                 else if (tid > req->r_tid)
 530                         n = n->rb_right;
 531                 else
 532                         return req;
 533         }
 534         return NULL;
 535 }
 536
 537 static struct ceph_osd_request *
 538 __lookup_request_ge(struct ceph_osd_client *osdc,
 539                     u64 tid)
 540 {
 541         struct ceph_osd_request *req;
 542         struct rb_node *n = osdc->requests.rb_node;
 543
 544         while (n) {
 545                 req = rb_entry(n, struct ceph_osd_request, r_node);
 546                 if (tid < req->r_tid) {
 547                         if (!n->rb_left)
 548                                 return req;
 549                         n = n->rb_left;
 550                 } else if (tid > req->r_tid) {
 551                         n = n->rb_right;
 552                 } else {
 553                         return req;
 554                 }
 555         }
 556         return NULL;
 557 }
 558
 559 /*
 560  * Resubmit requests pending on the given osd.
 561  */
 562 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 563                                 struct ceph_osd *osd)
 564 {
 565         struct ceph_osd_request *req, *nreq;
 566         int err;
 567
 568         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 569         err = __reset_osd(osdc, osd);
 570         if (err == -EAGAIN)
 571                 return;
 572
 573         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 574                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 575                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 576                      osd->o_osd);
 577                 if (!req->r_linger)
 578                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 579         }
 580
 581         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 582                                  r_linger_osd) {
 583                 /*
 584                  * reregister request prior to unregistering linger so
 585                  * that r_osd is preserved.
 586                  */
 587                 BUG_ON(!list_empty(&req->r_req_lru_item));
 588                 __register_request(osdc, req);
 589                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 590                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 591                 __unregister_linger_request(osdc, req);
 592                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 593                      osd->o_osd);
 594         }
 595 }
 596
 597 static void kick_osd_requests(struct ceph_osd_client *osdc,
 598                               struct ceph_osd *kickosd)
 599 {
 600         mutex_lock(&osdc->request_mutex);
 601         __kick_osd_requests(osdc, kickosd);
 602         mutex_unlock(&osdc->request_mutex);
 603 }
 604
 605 /*
 606  * If the osd connection drops, we need to resubmit all requests.
 607  */
 608 static void osd_reset(struct ceph_connection *con)
 609 {
 610         struct ceph_osd *osd = con->private;
 611         struct ceph_osd_client *osdc;
 612
 613         if (!osd)
 614                 return;
 615         dout("osd_reset osd%d\n", osd->o_osd);
 616         osdc = osd->o_osdc;
 617         down_read(&osdc->map_sem);
 618         kick_osd_requests(osdc, osd);
 619         send_queued(osdc);
 620         up_read(&osdc->map_sem);
 621 }
 622
 623 /*
 624  * Track open sessions with osds.
 625  */
 626 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 627 {
 628         struct ceph_osd *osd;
 629
 630         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 631         if (!osd)
 632                 return NULL;
 633
 634         atomic_set(&osd->o_ref, 1);
 635         osd->o_osdc = osdc;
 636         osd->o_osd = onum;
 637         INIT_LIST_HEAD(&osd->o_requests);
 638         INIT_LIST_HEAD(&osd->o_linger_requests);
 639         INIT_LIST_HEAD(&osd->o_osd_lru);
 640         osd->o_incarnation = 1;
 641
 642         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 643
 644         INIT_LIST_HEAD(&osd->o_keepalive_item);
 645         return osd;
 646 }
 647
 648 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 649 {
 650         if (atomic_inc_not_zero(&osd->o_ref)) {
 651                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 652                      atomic_read(&osd->o_ref));
 653                 return osd;
 654         } else {
 655                 dout("get_osd %p FAIL\n", osd);
 656                 return NULL;
 657         }
 658 }
 659
 660 static void put_osd(struct ceph_osd *osd)
 661 {
 662         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 663              atomic_read(&osd->o_ref) - 1);
 664         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 665                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 666
 667                 if (ac->ops && ac->ops->destroy_authorizer)
 668                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 669                 kfree(osd);
 670         }
 671 }
 672
 673 /*
 674  * remove an osd from our map
 675  */
 676 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 677 {
 678         dout("__remove_osd %p\n", osd);
 679         BUG_ON(!list_empty(&osd->o_requests));
 680         rb_erase(&osd->o_node, &osdc->osds);
 681         list_del_init(&osd->o_osd_lru);
 682         ceph_con_close(&osd->o_con);
 683         put_osd(osd);
 684 }
 685
 686 static void remove_all_osds(struct ceph_osd_client *osdc)
 687 {
 688         dout("__remove_old_osds %p\n", osdc);
 689         mutex_lock(&osdc->request_mutex);
 690         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 691                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 692                                                 struct ceph_osd, o_node);
 693                 __remove_osd(osdc, osd);
 694         }
 695         mutex_unlock(&osdc->request_mutex);
 696 }
 697
 698 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 699                               struct ceph_osd *osd)
 700 {
 701         dout("__move_osd_to_lru %p\n", osd);
 702         BUG_ON(!list_empty(&osd->o_osd_lru));
 703         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 704         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 705 }
 706
 707 static void __remove_osd_from_lru(struct ceph_osd *osd)
 708 {
 709         dout("__remove_osd_from_lru %p\n", osd);
 710         if (!list_empty(&osd->o_osd_lru))
 711                 list_del_init(&osd->o_osd_lru);
 712 }
 713
 714 static void remove_old_osds(struct ceph_osd_client *osdc)
 715 {
 716         struct ceph_osd *osd, *nosd;
 717
 718         dout("__remove_old_osds %p\n", osdc);
 719         mutex_lock(&osdc->request_mutex);
 720         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 721                 if (time_before(jiffies, osd->lru_ttl))
 722                         break;
 723                 __remove_osd(osdc, osd);
 724         }
 725         mutex_unlock(&osdc->request_mutex);
 726 }
 727
 728 /*
 729  * reset osd connect
 730  */
 731 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 732 {
 733         struct ceph_osd_request *req;
 734         int ret = 0;
 735
 736         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 737         if (list_empty(&osd->o_requests) &&
 738             list_empty(&osd->o_linger_requests)) {
 739                 __remove_osd(osdc, osd);
 740         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 741                           &osd->o_con.peer_addr,
 742                           sizeof(osd->o_con.peer_addr)) == 0 &&
 743                    !ceph_con_opened(&osd->o_con)) {
 744                 dout(" osd addr hasn't changed and connection never opened,"
 745                      " letting msgr retry");
 746                 /* touch each r_stamp for handle_timeout()'s benfit */
 747                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 748                         req->r_stamp = jiffies;
 749                 ret = -EAGAIN;
 750         } else {
 751                 ceph_con_close(&osd->o_con);
 752                 ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd,
 753                               &osdc->osdmap->osd_addr[osd->o_osd]);
 754                 osd->o_incarnation++;
 755         }
 756         return ret;
 757 }
 758
 759 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 760 {
 761         struct rb_node **p = &osdc->osds.rb_node;
 762         struct rb_node *parent = NULL;
 763         struct ceph_osd *osd = NULL;
 764
 765         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 766         while (*p) {
 767                 parent = *p;
 768                 osd = rb_entry(parent, struct ceph_osd, o_node);
 769                 if (new->o_osd < osd->o_osd)
 770                         p = &(*p)->rb_left;
 771                 else if (new->o_osd > osd->o_osd)
 772                         p = &(*p)->rb_right;
 773                 else
 774                         BUG();
 775         }
 776
 777         rb_link_node(&new->o_node, parent, p);
 778         rb_insert_color(&new->o_node, &osdc->osds);
 779 }
 780
 781 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 782 {
 783         struct ceph_osd *osd;
 784         struct rb_node *n = osdc->osds.rb_node;
 785
 786         while (n) {
 787                 osd = rb_entry(n, struct ceph_osd, o_node);
 788                 if (o < osd->o_osd)
 789                         n = n->rb_left;
 790                 else if (o > osd->o_osd)
 791                         n = n->rb_right;
 792                 else
 793                         return osd;
 794         }
 795         return NULL;
 796 }
 797
 798 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 799 {
 800         schedule_delayed_work(&osdc->timeout_work,
 801                         osdc->client->options->osd_keepalive_timeout * HZ);
 802 }
 803
 804 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 805 {
 806         cancel_delayed_work(&osdc->timeout_work);
 807 }
 808
 809 /*
 810  * Register request, assign tid.  If this is the first request, set up
 811  * the timeout event.
 812  */
 813 static void __register_request(struct ceph_osd_client *osdc,
 814                                struct ceph_osd_request *req)
 815 {
 816         req->r_tid = ++osdc->last_tid;
 817         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 818         dout("__register_request %p tid %lld\n", req, req->r_tid);
 819         __insert_request(osdc, req);
 820         ceph_osdc_get_request(req);
 821         osdc->num_requests++;
 822         if (osdc->num_requests == 1) {
 823                 dout(" first request, scheduling timeout\n");
 824                 __schedule_osd_timeout(osdc);
 825         }
 826 }
 827
 828 static void register_request(struct ceph_osd_client *osdc,
 829                              struct ceph_osd_request *req)
 830 {
 831         mutex_lock(&osdc->request_mutex);
 832         __register_request(osdc, req);
 833         mutex_unlock(&osdc->request_mutex);
 834 }
 835
 836 /*
 837  * called under osdc->request_mutex
 838  */
 839 static void __unregister_request(struct ceph_osd_client *osdc,
 840                                  struct ceph_osd_request *req)
 841 {
 842         if (RB_EMPTY_NODE(&req->r_node)) {
 843                 dout("__unregister_request %p tid %lld not registered\n",
 844                         req, req->r_tid);
 845                 return;
 846         }
 847
 848         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 849         rb_erase(&req->r_node, &osdc->requests);
 850         osdc->num_requests--;
 851
 852         if (req->r_osd) {
 853                 /* make sure the original request isn't in flight. */
 854                 ceph_msg_revoke(req->r_request);
 855
 856                 list_del_init(&req->r_osd_item);
 857                 if (list_empty(&req->r_osd->o_requests) &&
 858                     list_empty(&req->r_osd->o_linger_requests)) {
 859                         dout("moving osd to %p lru\n", req->r_osd);
 860                         __move_osd_to_lru(osdc, req->r_osd);
 861                 }
 862                 if (list_empty(&req->r_linger_item))
 863                         req->r_osd = NULL;
 864         }
 865
 866         ceph_osdc_put_request(req);
 867
 868         list_del_init(&req->r_req_lru_item);
 869         if (osdc->num_requests == 0) {
 870                 dout(" no requests, canceling timeout\n");
 871                 __cancel_osd_timeout(osdc);
 872         }
 873 }
 874
 875 /*
 876  * Cancel a previously queued request message
 877  */
 878 static void __cancel_request(struct ceph_osd_request *req)
 879 {
 880         if (req->r_sent && req->r_osd) {
 881                 ceph_msg_revoke(req->r_request);
 882                 req->r_sent = 0;
 883         }
 884 }
 885
 886 static void __register_linger_request(struct ceph_osd_client *osdc,
 887                                     struct ceph_osd_request *req)
 888 {
 889         dout("__register_linger_request %p\n", req);
 890         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 891         list_add_tail(&req->r_linger_osd, &req->r_osd->o_linger_requests);
 892 }
 893
 894 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 895                                         struct ceph_osd_request *req)
 896 {
 897         dout("__unregister_linger_request %p\n", req);
 898         if (req->r_osd) {
 899                 list_del_init(&req->r_linger_item);
 900                 list_del_init(&req->r_linger_osd);
 901
 902                 if (list_empty(&req->r_osd->o_requests) &&
 903                     list_empty(&req->r_osd->o_linger_requests)) {
 904                         dout("moving osd to %p lru\n", req->r_osd);
 905                         __move_osd_to_lru(osdc, req->r_osd);
 906                 }
 907                 if (list_empty(&req->r_osd_item))
 908                         req->r_osd = NULL;
 909         }
 910 }
 911
 912 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 913                                          struct ceph_osd_request *req)
 914 {
 915         mutex_lock(&osdc->request_mutex);
 916         if (req->r_linger) {
 917                 __unregister_linger_request(osdc, req);
 918                 ceph_osdc_put_request(req);
 919         }
 920         mutex_unlock(&osdc->request_mutex);
 921 }
 922 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 923
 924 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 925                                   struct ceph_osd_request *req)
 926 {
 927         if (!req->r_linger) {
 928                 dout("set_request_linger %p\n", req);
 929                 req->r_linger = 1;
 930                 /*
 931                  * caller is now responsible for calling
 932                  * unregister_linger_request
 933                  */
 934                 ceph_osdc_get_request(req);
 935         }
 936 }
 937 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 938
 939 /*
 940  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 941  * (as needed), and set the request r_osd appropriately.  If there is
 942  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 943  * (unsent, homeless) or leave on in-flight lru.
 944  *
 945  * Return 0 if unchanged, 1 if changed, or negative on error.
 946  *
 947  * Caller should hold map_sem for read and request_mutex.
 948  */
 949 static int __map_request(struct ceph_osd_client *osdc,
 950                          struct ceph_osd_request *req, int force_resend)
 951 {
 952         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 953         struct ceph_pg pgid;
 954         int acting[CEPH_PG_MAX_SIZE];
 955         int o = -1, num = 0;
 956         int err;
 957
 958         dout("map_request %p tid %lld\n", req, req->r_tid);
 959         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 960                                       &req->r_file_layout, osdc->osdmap);
 961         if (err) {
 962                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 963                 return err;
 964         }
 965         pgid = reqhead->layout.ol_pgid;
 966         req->r_pgid = pgid;
 967
 968         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 969         if (err > 0) {
 970                 o = acting[0];
 971                 num = err;
 972         }
 973
 974         if ((!force_resend &&
 975              req->r_osd && req->r_osd->o_osd == o &&
 976              req->r_sent >= req->r_osd->o_incarnation &&
 977              req->r_num_pg_osds == num &&
 978              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 979             (req->r_osd == NULL && o == -1))
 980                 return 0;  /* no change */
 981
 982         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 983              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 984              req->r_osd ? req->r_osd->o_osd : -1);
 985
 986         /* record full pg acting set */
 987         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 988         req->r_num_pg_osds = num;
 989
 990         if (req->r_osd) {
 991                 __cancel_request(req);
 992                 list_del_init(&req->r_osd_item);
 993                 req->r_osd = NULL;
 994         }
 995
 996         req->r_osd = __lookup_osd(osdc, o);
 997         if (!req->r_osd && o >= 0) {
 998                 err = -ENOMEM;
 999                 req->r_osd = create_osd(osdc, o);
1000                 if (!req->r_osd) {
1001                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1002                         goto out;
1003                 }
1004
1005                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1006                 __insert_osd(osdc, req->r_osd);
1007
1008                 ceph_con_open(&req->r_osd->o_con,
1009                               CEPH_ENTITY_TYPE_OSD, o,
1010                               &osdc->osdmap->osd_addr[o]);
1011         }
1012
1013         if (req->r_osd) {
1014                 __remove_osd_from_lru(req->r_osd);
1015                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1016                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1017         } else {
1018                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1019         }
1020         err = 1;   /* osd or pg changed */
1021
1022 out:
1023         return err;
1024 }
1025
1026 /*
1027  * caller should hold map_sem (for read) and request_mutex
1028  */
1029 static void __send_request(struct ceph_osd_client *osdc,
1030                            struct ceph_osd_request *req)
1031 {
1032         struct ceph_osd_request_head *reqhead;
1033
1034         dout("send_request %p tid %llu to osd%d flags %d\n",
1035              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1036
1037         reqhead = req->r_request->front.iov_base;
1038         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1039         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1040         reqhead->reassert_version = req->r_reassert_version;
1041
1042         req->r_stamp = jiffies;
1043         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1044
1045         ceph_msg_get(req->r_request); /* send consumes a ref */
1046         ceph_con_send(&req->r_osd->o_con, req->r_request);
1047         req->r_sent = req->r_osd->o_incarnation;
1048 }
1049
1050 /*
1051  * Send any requests in the queue (req_unsent).
1052  */
1053 static void send_queued(struct ceph_osd_client *osdc)
1054 {
1055         struct ceph_osd_request *req, *tmp;
1056
1057         dout("send_queued\n");
1058         mutex_lock(&osdc->request_mutex);
1059         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1060                 __send_request(osdc, req);
1061         }
1062         mutex_unlock(&osdc->request_mutex);
1063 }
1064
1065 /*
1066  * Timeout callback, called every N seconds when 1 or more osd
1067  * requests has been active for more than N seconds.  When this
1068  * happens, we ping all OSDs with requests who have timed out to
1069  * ensure any communications channel reset is detected.  Reset the
1070  * request timeouts another N seconds in the future as we go.
1071  * Reschedule the timeout event another N seconds in future (unless
1072  * there are no open requests).
1073  */
1074 static void handle_timeout(struct work_struct *work)
1075 {
1076         struct ceph_osd_client *osdc =
1077                 container_of(work, struct ceph_osd_client, timeout_work.work);
1078         struct ceph_osd_request *req, *last_req = NULL;
1079         struct ceph_osd *osd;
1080         unsigned long timeout = osdc->client->options->osd_timeout * HZ;
1081         unsigned long keepalive =
1082                 osdc->client->options->osd_keepalive_timeout * HZ;
1083         unsigned long last_stamp = 0;
1084         struct list_head slow_osds;
1085         dout("timeout\n");
1086         down_read(&osdc->map_sem);
1087
1088         ceph_monc_request_next_osdmap(&osdc->client->monc);
1089
1090         mutex_lock(&osdc->request_mutex);
1091
1092         /*
1093          * reset osds that appear to be _really_ unresponsive.  this
1094          * is a failsafe measure.. we really shouldn't be getting to
1095          * this point if the system is working properly.  the monitors
1096          * should mark the osd as failed and we should find out about
1097          * it from an updated osd map.
1098          */
1099         while (timeout && !list_empty(&osdc->req_lru)) {
1100                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
1101                                  r_req_lru_item);
1102
1103                 /* hasn't been long enough since we sent it? */
1104                 if (time_before(jiffies, req->r_stamp + timeout))
1105                         break;
1106
1107                 /* hasn't been long enough since it was acked? */
1108                 if (req->r_request->ack_stamp == 0 ||
1109                     time_before(jiffies, req->r_request->ack_stamp + timeout))
1110                         break;
1111
1112                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
1113                 last_req = req;
1114                 last_stamp = req->r_stamp;
1115
1116                 osd = req->r_osd;
1117                 BUG_ON(!osd);
1118                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
1119                            req->r_tid, osd->o_osd);
1120                 __kick_osd_requests(osdc, osd);
1121         }
1122
1123         /*
1124          * ping osds that are a bit slow.  this ensures that if there
1125          * is a break in the TCP connection we will notice, and reopen
1126          * a connection with that osd (from the fault callback).
1127          */
1128         INIT_LIST_HEAD(&slow_osds);
1129         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1130                 if (time_before(jiffies, req->r_stamp + keepalive))
1131                         break;
1132
1133                 osd = req->r_osd;
1134                 BUG_ON(!osd);
1135                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1136                      req->r_tid, osd->o_osd);
1137                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1138         }
1139         while (!list_empty(&slow_osds)) {
1140                 osd = list_entry(slow_osds.next, struct ceph_osd,
1141                                  o_keepalive_item);
1142                 list_del_init(&osd->o_keepalive_item);
1143                 ceph_con_keepalive(&osd->o_con);
1144         }
1145
1146         __schedule_osd_timeout(osdc);
1147         mutex_unlock(&osdc->request_mutex);
1148         send_queued(osdc);
1149         up_read(&osdc->map_sem);
1150 }
1151
1152 static void handle_osds_timeout(struct work_struct *work)
1153 {
1154         struct ceph_osd_client *osdc =
1155                 container_of(work, struct ceph_osd_client,
1156                              osds_timeout_work.work);
1157         unsigned long delay =
1158                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1159
1160         dout("osds timeout\n");
1161         down_read(&osdc->map_sem);
1162         remove_old_osds(osdc);
1163         up_read(&osdc->map_sem);
1164
1165         schedule_delayed_work(&osdc->osds_timeout_work,
1166                               round_jiffies_relative(delay));
1167 }
1168
1169 static void complete_request(struct ceph_osd_request *req)
1170 {
1171         if (req->r_safe_callback)
1172                 req->r_safe_callback(req, NULL);
1173         complete_all(&req->r_safe_completion);  /* fsync waiter */
1174 }
1175
1176 /*
1177  * handle osd op reply.  either call the callback if it is specified,
1178  * or do the completion to wake up the waiting thread.
1179  */
1180 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1181                          struct ceph_connection *con)
1182 {
1183         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1184         struct ceph_osd_request *req;
1185         u64 tid;
1186         int numops, object_len, flags;
1187         s32 result;
1188
1189         tid = le64_to_cpu(msg->hdr.tid);
1190         if (msg->front.iov_len < sizeof(*rhead))
1191                 goto bad;
1192         numops = le32_to_cpu(rhead->num_ops);
1193         object_len = le32_to_cpu(rhead->object_len);
1194         result = le32_to_cpu(rhead->result);
1195         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1196             numops * sizeof(struct ceph_osd_op))
1197                 goto bad;
1198         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1199         /* lookup */
1200         mutex_lock(&osdc->request_mutex);
1201         req = __lookup_request(osdc, tid);
1202         if (req == NULL) {
1203                 dout("handle_reply tid %llu dne\n", tid);
1204                 mutex_unlock(&osdc->request_mutex);
1205                 return;
1206         }
1207         ceph_osdc_get_request(req);
1208         flags = le32_to_cpu(rhead->flags);
1209
1210         /*
1211          * if this connection filled our message, drop our reference now, to
1212          * avoid a (safe but slower) revoke later.
1213          */
1214         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1215                 dout(" dropping con_filling_msg ref %p\n", con);
1216                 req->r_con_filling_msg = NULL;
1217                 con->ops->put(con);
1218         }
1219
1220         if (!req->r_got_reply) {
1221                 unsigned int bytes;
1222
1223                 req->r_result = le32_to_cpu(rhead->result);
1224                 bytes = le32_to_cpu(msg->hdr.data_len);
1225                 dout("handle_reply result %d bytes %d\n", req->r_result,
1226                      bytes);
1227                 if (req->r_result == 0)
1228                         req->r_result = bytes;
1229
1230                 /* in case this is a write and we need to replay, */
1231                 req->r_reassert_version = rhead->reassert_version;
1232
1233                 req->r_got_reply = 1;
1234         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1235                 dout("handle_reply tid %llu dup ack\n", tid);
1236                 mutex_unlock(&osdc->request_mutex);
1237                 goto done;
1238         }
1239
1240         dout("handle_reply tid %llu flags %d\n", tid, flags);
1241
1242         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1243                 __register_linger_request(osdc, req);
1244
1245         /* either this is a read, or we got the safe response */
1246         if (result < 0 ||
1247             (flags & CEPH_OSD_FLAG_ONDISK) ||
1248             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1249                 __unregister_request(osdc, req);
1250
1251         mutex_unlock(&osdc->request_mutex);
1252
1253         if (req->r_callback)
1254                 req->r_callback(req, msg);
1255         else
1256                 complete_all(&req->r_completion);
1257
1258         if (flags & CEPH_OSD_FLAG_ONDISK)
1259                 complete_request(req);
1260
1261 done:
1262         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1263         ceph_osdc_put_request(req);
1264         return;
1265
1266 bad:
1267         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1268                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1269                (int)sizeof(*rhead));
1270         ceph_msg_dump(msg);
1271 }
1272
1273 static void reset_changed_osds(struct ceph_osd_client *osdc)
1274 {
1275         struct rb_node *p, *n;
1276
1277         for (p = rb_first(&osdc->osds); p; p = n) {
1278                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1279
1280                 n = rb_next(p);
1281                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1282                     memcmp(&osd->o_con.peer_addr,
1283                            ceph_osd_addr(osdc->osdmap,
1284                                          osd->o_osd),
1285                            sizeof(struct ceph_entity_addr)) != 0)
1286                         __reset_osd(osdc, osd);
1287         }
1288 }
1289
1290 /*
1291  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1292  * no osd, request a new map.
1293  *
1294  * Caller should hold map_sem for read and request_mutex.
1295  */
1296 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1297 {
1298         struct ceph_osd_request *req, *nreq;
1299         struct rb_node *p;
1300         int needmap = 0;
1301         int err;
1302
1303         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1304         mutex_lock(&osdc->request_mutex);
1305         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
1306                 req = rb_entry(p, struct ceph_osd_request, r_node);
1307                 err = __map_request(osdc, req, force_resend);
1308                 if (err < 0)
1309                         continue;  /* error */
1310                 if (req->r_osd == NULL) {
1311                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1312                         needmap++;  /* request a newer map */
1313                 } else if (err > 0) {
1314                         dout("%p tid %llu requeued on osd%d\n", req, req->r_tid,
1315                              req->r_osd ? req->r_osd->o_osd : -1);
1316                         if (!req->r_linger)
1317                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1318                 }
1319         }
1320
1321         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1322                                  r_linger_item) {
1323                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1324
1325                 err = __map_request(osdc, req, force_resend);
1326                 if (err == 0)
1327                         continue;  /* no change and no osd was specified */
1328                 if (err < 0)
1329                         continue;  /* hrm! */
1330                 if (req->r_osd == NULL) {
1331                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1332                         needmap++;  /* request a newer map */
1333                         continue;
1334                 }
1335
1336                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1337                      req->r_osd ? req->r_osd->o_osd : -1);
1338                 __unregister_linger_request(osdc, req);
1339                 __register_request(osdc, req);
1340         }
1341         mutex_unlock(&osdc->request_mutex);
1342
1343         if (needmap) {
1344                 dout("%d requests for down osds, need new map\n", needmap);
1345                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1346         }
1347 }
1348
1349
1350 /*
1351  * Process updated osd map.
1352  *
1353  * The message contains any number of incremental and full maps, normally
1354  * indicating some sort of topology change in the cluster.  Kick requests
1355  * off to different OSDs as needed.
1356  */
1357 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1358 {
1359         void *p, *end, *next;
1360         u32 nr_maps, maplen;
1361         u32 epoch;
1362         struct ceph_osdmap *newmap = NULL, *oldmap;
1363         int err;
1364         struct ceph_fsid fsid;
1365
1366         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1367         p = msg->front.iov_base;
1368         end = p + msg->front.iov_len;
1369
1370         /* verify fsid */
1371         ceph_decode_need(&p, end, sizeof(fsid), bad);
1372         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1373         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1374                 return;
1375
1376         down_write(&osdc->map_sem);
1377
1378         /* incremental maps */
1379         ceph_decode_32_safe(&p, end, nr_maps, bad);
1380         dout(" %d inc maps\n", nr_maps);
1381         while (nr_maps > 0) {
1382                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1383                 epoch = ceph_decode_32(&p);
1384                 maplen = ceph_decode_32(&p);
1385                 ceph_decode_need(&p, end, maplen, bad);
1386                 next = p + maplen;
1387                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1388                         dout("applying incremental map %u len %d\n",
1389                              epoch, maplen);
1390                         newmap = osdmap_apply_incremental(&p, next,
1391                                                           osdc->osdmap,
1392                                                           &osdc->client->msgr);
1393                         if (IS_ERR(newmap)) {
1394                                 err = PTR_ERR(newmap);
1395                                 goto bad;
1396                         }
1397                         BUG_ON(!newmap);
1398                         if (newmap != osdc->osdmap) {
1399                                 ceph_osdmap_destroy(osdc->osdmap);
1400                                 osdc->osdmap = newmap;
1401                         }
1402                         kick_requests(osdc, 0);
1403                         reset_changed_osds(osdc);
1404                 } else {
1405                         dout("ignoring incremental map %u len %d\n",
1406                              epoch, maplen);
1407                 }
1408                 p = next;
1409                 nr_maps--;
1410         }
1411         if (newmap)
1412                 goto done;
1413
1414         /* full maps */
1415         ceph_decode_32_safe(&p, end, nr_maps, bad);
1416         dout(" %d full maps\n", nr_maps);
1417         while (nr_maps) {
1418                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1419                 epoch = ceph_decode_32(&p);
1420                 maplen = ceph_decode_32(&p);
1421                 ceph_decode_need(&p, end, maplen, bad);
1422                 if (nr_maps > 1) {
1423                         dout("skipping non-latest full map %u len %d\n",
1424                              epoch, maplen);
1425                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1426                         dout("skipping full map %u len %d, "
1427                              "older than our %u\n", epoch, maplen,
1428                              osdc->osdmap->epoch);
1429                 } else {
1430                         int skipped_map = 0;
1431
1432                         dout("taking full map %u len %d\n", epoch, maplen);
1433                         newmap = osdmap_decode(&p, p+maplen);
1434                         if (IS_ERR(newmap)) {
1435                                 err = PTR_ERR(newmap);
1436                                 goto bad;
1437                         }
1438                         BUG_ON(!newmap);
1439                         oldmap = osdc->osdmap;
1440                         osdc->osdmap = newmap;
1441                         if (oldmap) {
1442                                 if (oldmap->epoch + 1 < newmap->epoch)
1443                                         skipped_map = 1;
1444                                 ceph_osdmap_destroy(oldmap);
1445                         }
1446                         kick_requests(osdc, skipped_map);
1447                 }
1448                 p += maplen;
1449                 nr_maps--;
1450         }
1451
1452 done:
1453         downgrade_write(&osdc->map_sem);
1454         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1455
1456         /*
1457          * subscribe to subsequent osdmap updates if full to ensure
1458          * we find out when we are no longer full and stop returning
1459          * ENOSPC.
1460          */
1461         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1462                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1463
1464         send_queued(osdc);
1465         up_read(&osdc->map_sem);
1466         wake_up_all(&osdc->client->auth_wq);
1467         return;
1468
1469 bad:
1470         pr_err("osdc handle_map corrupt msg\n");
1471         ceph_msg_dump(msg);
1472         up_write(&osdc->map_sem);
1473         return;
1474 }
1475
1476 /*
1477  * watch/notify callback event infrastructure
1478  *
1479  * These callbacks are used both for watch and notify operations.
1480  */
1481 static void __release_event(struct kref *kref)
1482 {
1483         struct ceph_osd_event *event =
1484                 container_of(kref, struct ceph_osd_event, kref);
1485
1486         dout("__release_event %p\n", event);
1487         kfree(event);
1488 }
1489
1490 static void get_event(struct ceph_osd_event *event)
1491 {
1492         kref_get(&event->kref);
1493 }
1494
1495 void ceph_osdc_put_event(struct ceph_osd_event *event)
1496 {
1497         kref_put(&event->kref, __release_event);
1498 }
1499 EXPORT_SYMBOL(ceph_osdc_put_event);
1500
1501 static void __insert_event(struct ceph_osd_client *osdc,
1502                              struct ceph_osd_event *new)
1503 {
1504         struct rb_node **p = &osdc->event_tree.rb_node;
1505         struct rb_node *parent = NULL;
1506         struct ceph_osd_event *event = NULL;
1507
1508         while (*p) {
1509                 parent = *p;
1510                 event = rb_entry(parent, struct ceph_osd_event, node);
1511                 if (new->cookie < event->cookie)
1512                         p = &(*p)->rb_left;
1513                 else if (new->cookie > event->cookie)
1514                         p = &(*p)->rb_right;
1515                 else
1516                         BUG();
1517         }
1518
1519         rb_link_node(&new->node, parent, p);
1520         rb_insert_color(&new->node, &osdc->event_tree);
1521 }
1522
1523 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1524                                                 u64 cookie)
1525 {
1526         struct rb_node **p = &osdc->event_tree.rb_node;
1527         struct rb_node *parent = NULL;
1528         struct ceph_osd_event *event = NULL;
1529
1530         while (*p) {
1531                 parent = *p;
1532                 event = rb_entry(parent, struct ceph_osd_event, node);
1533                 if (cookie < event->cookie)
1534                         p = &(*p)->rb_left;
1535                 else if (cookie > event->cookie)
1536                         p = &(*p)->rb_right;
1537                 else
1538                         return event;
1539         }
1540         return NULL;
1541 }
1542
1543 static void __remove_event(struct ceph_osd_event *event)
1544 {
1545         struct ceph_osd_client *osdc = event->osdc;
1546
1547         if (!RB_EMPTY_NODE(&event->node)) {
1548                 dout("__remove_event removed %p\n", event);
1549                 rb_erase(&event->node, &osdc->event_tree);
1550                 ceph_osdc_put_event(event);
1551         } else {
1552                 dout("__remove_event didn't remove %p\n", event);
1553         }
1554 }
1555
1556 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1557                            void (*event_cb)(u64, u64, u8, void *),
1558                            int one_shot, void *data,
1559                            struct ceph_osd_event **pevent)
1560 {
1561         struct ceph_osd_event *event;
1562
1563         event = kmalloc(sizeof(*event), GFP_NOIO);
1564         if (!event)
1565                 return -ENOMEM;
1566
1567         dout("create_event %p\n", event);
1568         event->cb = event_cb;
1569         event->one_shot = one_shot;
1570         event->data = data;
1571         event->osdc = osdc;
1572         INIT_LIST_HEAD(&event->osd_node);
1573         kref_init(&event->kref);   /* one ref for us */
1574         kref_get(&event->kref);    /* one ref for the caller */
1575         init_completion(&event->completion);
1576
1577         spin_lock(&osdc->event_lock);
1578         event->cookie = ++osdc->event_count;
1579         __insert_event(osdc, event);
1580         spin_unlock(&osdc->event_lock);
1581
1582         *pevent = event;
1583         return 0;
1584 }
1585 EXPORT_SYMBOL(ceph_osdc_create_event);
1586
1587 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1588 {
1589         struct ceph_osd_client *osdc = event->osdc;
1590
1591         dout("cancel_event %p\n", event);
1592         spin_lock(&osdc->event_lock);
1593         __remove_event(event);
1594         spin_unlock(&osdc->event_lock);
1595         ceph_osdc_put_event(event); /* caller's */
1596 }
1597 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1598
1599
1600 static void do_event_work(struct work_struct *work)
1601 {
1602         struct ceph_osd_event_work *event_work =
1603                 container_of(work, struct ceph_osd_event_work, work);
1604         struct ceph_osd_event *event = event_work->event;
1605         u64 ver = event_work->ver;
1606         u64 notify_id = event_work->notify_id;
1607         u8 opcode = event_work->opcode;
1608
1609         dout("do_event_work completing %p\n", event);
1610         event->cb(ver, notify_id, opcode, event->data);
1611         complete(&event->completion);
1612         dout("do_event_work completed %p\n", event);
1613         ceph_osdc_put_event(event);
1614         kfree(event_work);
1615 }
1616
1617
1618 /*
1619  * Process osd watch notifications
1620  */
1621 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1622 {
1623         void *p, *end;
1624         u8 proto_ver;
1625         u64 cookie, ver, notify_id;
1626         u8 opcode;
1627         struct ceph_osd_event *event;
1628         struct ceph_osd_event_work *event_work;
1629
1630         p = msg->front.iov_base;
1631         end = p + msg->front.iov_len;
1632
1633         ceph_decode_8_safe(&p, end, proto_ver, bad);
1634         ceph_decode_8_safe(&p, end, opcode, bad);
1635         ceph_decode_64_safe(&p, end, cookie, bad);
1636         ceph_decode_64_safe(&p, end, ver, bad);
1637         ceph_decode_64_safe(&p, end, notify_id, bad);
1638
1639         spin_lock(&osdc->event_lock);
1640         event = __find_event(osdc, cookie);
1641         if (event) {
1642                 get_event(event);
1643                 if (event->one_shot)
1644                         __remove_event(event);
1645         }
1646         spin_unlock(&osdc->event_lock);
1647         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1648              cookie, ver, event);
1649         if (event) {
1650                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1651                 if (!event_work) {
1652                         dout("ERROR: could not allocate event_work\n");
1653                         goto done_err;
1654                 }
1655                 INIT_WORK(&event_work->work, do_event_work);
1656                 event_work->event = event;
1657                 event_work->ver = ver;
1658                 event_work->notify_id = notify_id;
1659                 event_work->opcode = opcode;
1660                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1661                         dout("WARNING: failed to queue notify event work\n");
1662                         goto done_err;
1663                 }
1664         }
1665
1666         return;
1667
1668 done_err:
1669         complete(&event->completion);
1670         ceph_osdc_put_event(event);
1671         return;
1672
1673 bad:
1674         pr_err("osdc handle_watch_notify corrupt msg\n");
1675         return;
1676 }
1677
1678 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1679 {
1680         int err;
1681
1682         dout("wait_event %p\n", event);
1683         err = wait_for_completion_interruptible_timeout(&event->completion,
1684                                                         timeout * HZ);
1685         ceph_osdc_put_event(event);
1686         if (err > 0)
1687                 err = 0;
1688         dout("wait_event %p returns %d\n", event, err);
1689         return err;
1690 }
1691 EXPORT_SYMBOL(ceph_osdc_wait_event);
1692
1693 /*
1694  * Register request, send initial attempt.
1695  */
1696 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1697                             struct ceph_osd_request *req,
1698                             bool nofail)
1699 {
1700         int rc = 0;
1701
1702         req->r_request->pages = req->r_pages;
1703         req->r_request->nr_pages = req->r_num_pages;
1704 #ifdef CONFIG_BLOCK
1705         req->r_request->bio = req->r_bio;
1706 #endif
1707         req->r_request->trail = req->r_trail;
1708
1709         register_request(osdc, req);
1710
1711         down_read(&osdc->map_sem);
1712         mutex_lock(&osdc->request_mutex);
1713         /*
1714          * a racing kick_requests() may have sent the message for us
1715          * while we dropped request_mutex above, so only send now if
1716          * the request still han't been touched yet.
1717          */
1718         if (req->r_sent == 0) {
1719                 rc = __map_request(osdc, req, 0);
1720                 if (rc < 0) {
1721                         if (nofail) {
1722                                 dout("osdc_start_request failed map, "
1723                                      " will retry %lld\n", req->r_tid);
1724                                 rc = 0;
1725                         }
1726                         goto out_unlock;
1727                 }
1728                 if (req->r_osd == NULL) {
1729                         dout("send_request %p no up osds in pg\n", req);
1730                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1731                 } else {
1732                         __send_request(osdc, req);
1733                 }
1734                 rc = 0;
1735         }
1736
1737 out_unlock:
1738         mutex_unlock(&osdc->request_mutex);
1739         up_read(&osdc->map_sem);
1740         return rc;
1741 }
1742 EXPORT_SYMBOL(ceph_osdc_start_request);
1743
1744 /*
1745  * wait for a request to complete
1746  */
1747 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1748                            struct ceph_osd_request *req)
1749 {
1750         int rc;
1751
1752         rc = wait_for_completion_interruptible(&req->r_completion);
1753         if (rc < 0) {
1754                 mutex_lock(&osdc->request_mutex);
1755                 __cancel_request(req);
1756                 __unregister_request(osdc, req);
1757                 mutex_unlock(&osdc->request_mutex);
1758                 complete_request(req);
1759                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1760                 return rc;
1761         }
1762
1763         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1764         return req->r_result;
1765 }
1766 EXPORT_SYMBOL(ceph_osdc_wait_request);
1767
1768 /*
1769  * sync - wait for all in-flight requests to flush.  avoid starvation.
1770  */
1771 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1772 {
1773         struct ceph_osd_request *req;
1774         u64 last_tid, next_tid = 0;
1775
1776         mutex_lock(&osdc->request_mutex);
1777         last_tid = osdc->last_tid;
1778         while (1) {
1779                 req = __lookup_request_ge(osdc, next_tid);
1780                 if (!req)
1781                         break;
1782                 if (req->r_tid > last_tid)
1783                         break;
1784
1785                 next_tid = req->r_tid + 1;
1786                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1787                         continue;
1788
1789                 ceph_osdc_get_request(req);
1790                 mutex_unlock(&osdc->request_mutex);
1791                 dout("sync waiting on tid %llu (last is %llu)\n",
1792                      req->r_tid, last_tid);
1793                 wait_for_completion(&req->r_safe_completion);
1794                 mutex_lock(&osdc->request_mutex);
1795                 ceph_osdc_put_request(req);
1796         }
1797         mutex_unlock(&osdc->request_mutex);
1798         dout("sync done (thru tid %llu)\n", last_tid);
1799 }
1800 EXPORT_SYMBOL(ceph_osdc_sync);
1801
1802 /*
1803  * init, shutdown
1804  */
1805 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1806 {
1807         int err;
1808
1809         dout("init\n");
1810         osdc->client = client;
1811         osdc->osdmap = NULL;
1812         init_rwsem(&osdc->map_sem);
1813         init_completion(&osdc->map_waiters);
1814         osdc->last_requested_map = 0;
1815         mutex_init(&osdc->request_mutex);
1816         osdc->last_tid = 0;
1817         osdc->osds = RB_ROOT;
1818         INIT_LIST_HEAD(&osdc->osd_lru);
1819         osdc->requests = RB_ROOT;
1820         INIT_LIST_HEAD(&osdc->req_lru);
1821         INIT_LIST_HEAD(&osdc->req_unsent);
1822         INIT_LIST_HEAD(&osdc->req_notarget);
1823         INIT_LIST_HEAD(&osdc->req_linger);
1824         osdc->num_requests = 0;
1825         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1826         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1827         spin_lock_init(&osdc->event_lock);
1828         osdc->event_tree = RB_ROOT;
1829         osdc->event_count = 0;
1830
1831         schedule_delayed_work(&osdc->osds_timeout_work,
1832            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1833
1834         err = -ENOMEM;
1835         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1836                                         sizeof(struct ceph_osd_request));
1837         if (!osdc->req_mempool)
1838                 goto out;
1839
1840         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1841                                 "osd_op");
1842         if (err < 0)
1843                 goto out_mempool;
1844         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1845                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1846                                 "osd_op_reply");
1847         if (err < 0)
1848                 goto out_msgpool;
1849
1850         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1851         if (IS_ERR(osdc->notify_wq)) {
1852                 err = PTR_ERR(osdc->notify_wq);
1853                 osdc->notify_wq = NULL;
1854                 goto out_msgpool;
1855         }
1856         return 0;
1857
1858 out_msgpool:
1859         ceph_msgpool_destroy(&osdc->msgpool_op);
1860 out_mempool:
1861         mempool_destroy(osdc->req_mempool);
1862 out:
1863         return err;
1864 }
1865 EXPORT_SYMBOL(ceph_osdc_init);
1866
1867 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1868 {
1869         flush_workqueue(osdc->notify_wq);
1870         destroy_workqueue(osdc->notify_wq);
1871         cancel_delayed_work_sync(&osdc->timeout_work);
1872         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1873         if (osdc->osdmap) {
1874                 ceph_osdmap_destroy(osdc->osdmap);
1875                 osdc->osdmap = NULL;
1876         }
1877         remove_all_osds(osdc);
1878         mempool_destroy(osdc->req_mempool);
1879         ceph_msgpool_destroy(&osdc->msgpool_op);
1880         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1881 }
1882 EXPORT_SYMBOL(ceph_osdc_stop);
1883
1884 /*
1885  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1886  * *plen.  Return number of bytes read, or error.
1887  */
1888 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1889                         struct ceph_vino vino, struct ceph_file_layout *layout,
1890                         u64 off, u64 *plen,
1891                         u32 truncate_seq, u64 truncate_size,
1892                         struct page **pages, int num_pages, int page_align)
1893 {
1894         struct ceph_osd_request *req;
1895         int rc = 0;
1896
1897         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1898              vino.snap, off, *plen);
1899         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1900                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1901                                     NULL, 0, truncate_seq, truncate_size, NULL,
1902                                     false, 1, page_align);
1903         if (!req)
1904                 return -ENOMEM;
1905
1906         /* it may be a short read due to an object boundary */
1907         req->r_pages = pages;
1908
1909         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1910              off, *plen, req->r_num_pages, page_align);
1911
1912         rc = ceph_osdc_start_request(osdc, req, false);
1913         if (!rc)
1914                 rc = ceph_osdc_wait_request(osdc, req);
1915
1916         ceph_osdc_put_request(req);
1917         dout("readpages result %d\n", rc);
1918         return rc;
1919 }
1920 EXPORT_SYMBOL(ceph_osdc_readpages);
1921
1922 /*
1923  * do a synchronous write on N pages
1924  */
1925 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1926                          struct ceph_file_layout *layout,
1927                          struct ceph_snap_context *snapc,
1928                          u64 off, u64 len,
1929                          u32 truncate_seq, u64 truncate_size,
1930                          struct timespec *mtime,
1931                          struct page **pages, int num_pages,
1932                          int flags, int do_sync, bool nofail)
1933 {
1934         struct ceph_osd_request *req;
1935         int rc = 0;
1936         int page_align = off & ~PAGE_MASK;
1937
1938         BUG_ON(vino.snap != CEPH_NOSNAP);
1939         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1940                                     CEPH_OSD_OP_WRITE,
1941                                     flags | CEPH_OSD_FLAG_ONDISK |
1942                                             CEPH_OSD_FLAG_WRITE,
1943                                     snapc, do_sync,
1944                                     truncate_seq, truncate_size, mtime,
1945                                     nofail, 1, page_align);
1946         if (!req)
1947                 return -ENOMEM;
1948
1949         /* it may be a short write due to an object boundary */
1950         req->r_pages = pages;
1951         dout("writepages %llu~%llu (%d pages)\n", off, len,
1952              req->r_num_pages);
1953
1954         rc = ceph_osdc_start_request(osdc, req, nofail);
1955         if (!rc)
1956                 rc = ceph_osdc_wait_request(osdc, req);
1957
1958         ceph_osdc_put_request(req);
1959         if (rc == 0)
1960                 rc = len;
1961         dout("writepages result %d\n", rc);
1962         return rc;
1963 }
1964 EXPORT_SYMBOL(ceph_osdc_writepages);
1965
1966 /*
1967  * handle incoming message
1968  */
1969 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1970 {
1971         struct ceph_osd *osd = con->private;
1972         struct ceph_osd_client *osdc;
1973         int type = le16_to_cpu(msg->hdr.type);
1974
1975         if (!osd)
1976                 goto out;
1977         osdc = osd->o_osdc;
1978
1979         switch (type) {
1980         case CEPH_MSG_OSD_MAP:
1981                 ceph_osdc_handle_map(osdc, msg);
1982                 break;
1983         case CEPH_MSG_OSD_OPREPLY:
1984                 handle_reply(osdc, msg, con);
1985                 break;
1986         case CEPH_MSG_WATCH_NOTIFY:
1987                 handle_watch_notify(osdc, msg);
1988                 break;
1989
1990         default:
1991                 pr_err("received unknown message type %d %s\n", type,
1992                        ceph_msg_type_name(type));
1993         }
1994 out:
1995         ceph_msg_put(msg);
1996 }
1997
1998 /*
1999  * lookup and return message for incoming reply.  set up reply message
2000  * pages.
2001  */
2002 static struct ceph_msg *get_reply(struct ceph_connection *con,
2003                                   struct ceph_msg_header *hdr,
2004                                   int *skip)
2005 {
2006         struct ceph_osd *osd = con->private;
2007         struct ceph_osd_client *osdc = osd->o_osdc;
2008         struct ceph_msg *m;
2009         struct ceph_osd_request *req;
2010         int front = le32_to_cpu(hdr->front_len);
2011         int data_len = le32_to_cpu(hdr->data_len);
2012         u64 tid;
2013
2014         tid = le64_to_cpu(hdr->tid);
2015         mutex_lock(&osdc->request_mutex);
2016         req = __lookup_request(osdc, tid);
2017         if (!req) {
2018                 *skip = 1;
2019                 m = NULL;
2020                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
2021                         osd->o_osd);
2022                 goto out;
2023         }
2024
2025         if (req->r_con_filling_msg) {
2026                 dout("%s revoking msg %p from old con %p\n", __func__,
2027                      req->r_reply, req->r_con_filling_msg);
2028                 ceph_msg_revoke_incoming(req->r_reply);
2029                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2030                 req->r_con_filling_msg = NULL;
2031         }
2032
2033         if (front > req->r_reply->front.iov_len) {
2034                 pr_warning("get_reply front %d > preallocated %d\n",
2035                            front, (int)req->r_reply->front.iov_len);
2036                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2037                 if (!m)
2038                         goto out;
2039                 ceph_msg_put(req->r_reply);
2040                 req->r_reply = m;
2041         }
2042         m = ceph_msg_get(req->r_reply);
2043
2044         if (data_len > 0) {
2045                 int want = calc_pages_for(req->r_page_alignment, data_len);
2046
2047                 if (unlikely(req->r_num_pages < want)) {
2048                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2049                                    " had only %d pages ready\n", tid, data_len,
2050                                    want, req->r_num_pages);
2051                         *skip = 1;
2052                         ceph_msg_put(m);
2053                         m = NULL;
2054                         goto out;
2055                 }
2056                 m->pages = req->r_pages;
2057                 m->nr_pages = req->r_num_pages;
2058                 m->page_alignment = req->r_page_alignment;
2059 #ifdef CONFIG_BLOCK
2060                 m->bio = req->r_bio;
2061 #endif
2062         }
2063         *skip = 0;
2064         req->r_con_filling_msg = con->ops->get(con);
2065         dout("get_reply tid %lld %p\n", tid, m);
2066
2067 out:
2068         mutex_unlock(&osdc->request_mutex);
2069         return m;
2070
2071 }
2072
2073 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2074                                   struct ceph_msg_header *hdr,
2075                                   int *skip)
2076 {
2077         struct ceph_osd *osd = con->private;
2078         int type = le16_to_cpu(hdr->type);
2079         int front = le32_to_cpu(hdr->front_len);
2080
2081         *skip = 0;
2082         switch (type) {
2083         case CEPH_MSG_OSD_MAP:
2084         case CEPH_MSG_WATCH_NOTIFY:
2085                 return ceph_msg_new(type, front, GFP_NOFS, false);
2086         case CEPH_MSG_OSD_OPREPLY:
2087                 return get_reply(con, hdr, skip);
2088         default:
2089                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2090                         osd->o_osd);
2091                 *skip = 1;
2092                 return NULL;
2093         }
2094 }
2095
2096 /*
2097  * Wrappers to refcount containing ceph_osd struct
2098  */
2099 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2100 {
2101         struct ceph_osd *osd = con->private;
2102         if (get_osd(osd))
2103                 return con;
2104         return NULL;
2105 }
2106
2107 static void put_osd_con(struct ceph_connection *con)
2108 {
2109         struct ceph_osd *osd = con->private;
2110         put_osd(osd);
2111 }
2112
2113 /*
2114  * authentication
2115  */
2116 /*
2117  * Note: returned pointer is the address of a structure that's
2118  * managed separately.  Caller must *not* attempt to free it.
2119  */
2120 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2121                                         int *proto, int force_new)
2122 {
2123         struct ceph_osd *o = con->private;
2124         struct ceph_osd_client *osdc = o->o_osdc;
2125         struct ceph_auth_client *ac = osdc->client->monc.auth;
2126         struct ceph_auth_handshake *auth = &o->o_auth;
2127
2128         if (force_new && auth->authorizer) {
2129                 if (ac->ops && ac->ops->destroy_authorizer)
2130                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2131                 auth->authorizer = NULL;
2132         }
2133         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2134                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2135                                                         auth);
2136                 if (ret)
2137                         return ERR_PTR(ret);
2138         }
2139         *proto = ac->protocol;
2140
2141         return auth;
2142 }
2143
2144
2145 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2146 {
2147         struct ceph_osd *o = con->private;
2148         struct ceph_osd_client *osdc = o->o_osdc;
2149         struct ceph_auth_client *ac = osdc->client->monc.auth;
2150
2151         /*
2152          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2153          * XXX which do we do:  succeed or fail?
2154          */
2155         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2156 }
2157
2158 static int invalidate_authorizer(struct ceph_connection *con)
2159 {
2160         struct ceph_osd *o = con->private;
2161         struct ceph_osd_client *osdc = o->o_osdc;
2162         struct ceph_auth_client *ac = osdc->client->monc.auth;
2163
2164         if (ac->ops && ac->ops->invalidate_authorizer)
2165                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2166
2167         return ceph_monc_validate_auth(&osdc->client->monc);
2168 }
2169
2170 static const struct ceph_connection_operations osd_con_ops = {
2171         .get = get_osd_con,
2172         .put = put_osd_con,
2173         .dispatch = dispatch,
2174         .get_authorizer = get_authorizer,
2175         .verify_authorizer_reply = verify_authorizer_reply,
2176         .invalidate_authorizer = invalidate_authorizer,
2177         .alloc_msg = alloc_msg,
2178         .fault = osd_reset,
2179 };