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 int ceph_calc_raw_layout(struct ceph_file_layout *layout,
  42                         u64 off, u64 *plen, u64 *bno,
  43                         struct ceph_osd_request *req,
  44                         struct ceph_osd_req_op *op)
  45 {
  46         u64 orig_len = *plen;
  47         u64 objoff, objlen;    /* extent in object */
  48         int r;
  49
  50         /* object extent? */
  51         r = ceph_calc_file_object_mapping(layout, off, orig_len, bno,
  52                                           &objoff, &objlen);
  53         if (r < 0)
  54                 return r;
  55         if (objlen < orig_len) {
  56                 *plen = objlen;
  57                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  58                      orig_len - *plen, off, *plen);
  59         }
  60
  61         if (op_has_extent(op->op)) {
  62                 u32 osize = le32_to_cpu(layout->fl_object_size);
  63                 op->extent.offset = objoff;
  64                 op->extent.length = objlen;
  65                 if (op->extent.truncate_size <= off - objoff) {
  66                         op->extent.truncate_size = 0;
  67                 } else {
  68                         op->extent.truncate_size -= off - objoff;
  69                         if (op->extent.truncate_size > osize)
  70                                 op->extent.truncate_size = osize;
  71                 }
  72         }
  73         req->r_num_pages = calc_pages_for(off, *plen);
  74         req->r_page_alignment = off & ~PAGE_MASK;
  75         if (op->op == CEPH_OSD_OP_WRITE)
  76                 op->payload_len = *plen;
  77
  78         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
  79              *bno, objoff, objlen, req->r_num_pages);
  80         return 0;
  81 }
  82 EXPORT_SYMBOL(ceph_calc_raw_layout);
  83
  84 /*
  85  * Implement client access to distributed object storage cluster.
  86  *
  87  * All data objects are stored within a cluster/cloud of OSDs, or
  88  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  89  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  90  * remote daemons serving up and coordinating consistent and safe
  91  * access to storage.
  92  *
  93  * Cluster membership and the mapping of data objects onto storage devices
  94  * are described by the osd map.
  95  *
  96  * We keep track of pending OSD requests (read, write), resubmit
  97  * requests to different OSDs when the cluster topology/data layout
  98  * change, or retry the affected requests when the communications
  99  * channel with an OSD is reset.
 100  */
 101
 102 /*
 103  * calculate the mapping of a file extent onto an object, and fill out the
 104  * request accordingly.  shorten extent as necessary if it crosses an
 105  * object boundary.
 106  *
 107  * fill osd op in request message.
 108  */
 109 static int calc_layout(struct ceph_vino vino,
 110                        struct ceph_file_layout *layout,
 111                        u64 off, u64 *plen,
 112                        struct ceph_osd_request *req,
 113                        struct ceph_osd_req_op *op)
 114 {
 115         u64 bno;
 116         int r;
 117
 118         r = ceph_calc_raw_layout(layout, off, plen, &bno, req, op);
 119         if (r < 0)
 120                 return r;
 121
 122         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 123         req->r_oid_len = strlen(req->r_oid);
 124
 125         return r;
 126 }
 127
 128 /*
 129  * requests
 130  */
 131 void ceph_osdc_release_request(struct kref *kref)
 132 {
 133         struct ceph_osd_request *req = container_of(kref,
 134                                                     struct ceph_osd_request,
 135                                                     r_kref);
 136
 137         if (req->r_request)
 138                 ceph_msg_put(req->r_request);
 139         if (req->r_con_filling_msg) {
 140                 dout("%s revoking msg %p from con %p\n", __func__,
 141                      req->r_reply, req->r_con_filling_msg);
 142                 ceph_msg_revoke_incoming(req->r_reply);
 143                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 144                 req->r_con_filling_msg = NULL;
 145         }
 146         if (req->r_reply)
 147                 ceph_msg_put(req->r_reply);
 148         if (req->r_own_pages)
 149                 ceph_release_page_vector(req->r_pages,
 150                                          req->r_num_pages);
 151         ceph_put_snap_context(req->r_snapc);
 152         ceph_pagelist_release(&req->r_trail);
 153         if (req->r_mempool)
 154                 mempool_free(req, req->r_osdc->req_mempool);
 155         else
 156                 kfree(req);
 157 }
 158 EXPORT_SYMBOL(ceph_osdc_release_request);
 159
 160 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 161                                                struct ceph_snap_context *snapc,
 162                                                unsigned int num_op,
 163                                                bool use_mempool,
 164                                                gfp_t gfp_flags)
 165 {
 166         struct ceph_osd_request *req;
 167         struct ceph_msg *msg;
 168         size_t msg_size = sizeof(struct ceph_osd_request_head);
 169
 170         msg_size += num_op*sizeof(struct ceph_osd_op);
 171
 172         if (use_mempool) {
 173                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 174                 memset(req, 0, sizeof(*req));
 175         } else {
 176                 req = kzalloc(sizeof(*req), gfp_flags);
 177         }
 178         if (req == NULL)
 179                 return NULL;
 180
 181         req->r_osdc = osdc;
 182         req->r_mempool = use_mempool;
 183
 184         kref_init(&req->r_kref);
 185         init_completion(&req->r_completion);
 186         init_completion(&req->r_safe_completion);
 187         RB_CLEAR_NODE(&req->r_node);
 188         INIT_LIST_HEAD(&req->r_unsafe_item);
 189         INIT_LIST_HEAD(&req->r_linger_item);
 190         INIT_LIST_HEAD(&req->r_linger_osd);
 191         INIT_LIST_HEAD(&req->r_req_lru_item);
 192         INIT_LIST_HEAD(&req->r_osd_item);
 193
 194         /* create reply message */
 195         if (use_mempool)
 196                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 197         else
 198                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 199                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 200         if (!msg) {
 201                 ceph_osdc_put_request(req);
 202                 return NULL;
 203         }
 204         req->r_reply = msg;
 205
 206         ceph_pagelist_init(&req->r_trail);
 207
 208         /* create request message; allow space for oid */
 209         msg_size += MAX_OBJ_NAME_SIZE;
 210         if (snapc)
 211                 msg_size += sizeof(u64) * snapc->num_snaps;
 212         if (use_mempool)
 213                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 214         else
 215                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 216         if (!msg) {
 217                 ceph_osdc_put_request(req);
 218                 return NULL;
 219         }
 220
 221         memset(msg->front.iov_base, 0, msg->front.iov_len);
 222
 223         req->r_request = msg;
 224
 225         return req;
 226 }
 227 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 228
 229 static void osd_req_encode_op(struct ceph_osd_request *req,
 230                               struct ceph_osd_op *dst,
 231                               struct ceph_osd_req_op *src)
 232 {
 233         dst->op = cpu_to_le16(src->op);
 234
 235         switch (src->op) {
 236         case CEPH_OSD_OP_READ:
 237         case CEPH_OSD_OP_WRITE:
 238                 dst->extent.offset =
 239                         cpu_to_le64(src->extent.offset);
 240                 dst->extent.length =
 241                         cpu_to_le64(src->extent.length);
 242                 dst->extent.truncate_size =
 243                         cpu_to_le64(src->extent.truncate_size);
 244                 dst->extent.truncate_seq =
 245                         cpu_to_le32(src->extent.truncate_seq);
 246                 break;
 247
 248         case CEPH_OSD_OP_GETXATTR:
 249         case CEPH_OSD_OP_SETXATTR:
 250         case CEPH_OSD_OP_CMPXATTR:
 251                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 252                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 253                 dst->xattr.cmp_op = src->xattr.cmp_op;
 254                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 255                 ceph_pagelist_append(&req->r_trail, src->xattr.name,
 256                                      src->xattr.name_len);
 257                 ceph_pagelist_append(&req->r_trail, src->xattr.val,
 258                                      src->xattr.value_len);
 259                 break;
 260         case CEPH_OSD_OP_CALL:
 261                 dst->cls.class_len = src->cls.class_len;
 262                 dst->cls.method_len = src->cls.method_len;
 263                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 264
 265                 ceph_pagelist_append(&req->r_trail, src->cls.class_name,
 266                                      src->cls.class_len);
 267                 ceph_pagelist_append(&req->r_trail, src->cls.method_name,
 268                                      src->cls.method_len);
 269                 ceph_pagelist_append(&req->r_trail, src->cls.indata,
 270                                      src->cls.indata_len);
 271                 break;
 272         case CEPH_OSD_OP_ROLLBACK:
 273                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
 274                 break;
 275         case CEPH_OSD_OP_STARTSYNC:
 276                 break;
 277         case CEPH_OSD_OP_NOTIFY:
 278                 {
 279                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
 280                         __le32 timeout = cpu_to_le32(src->watch.timeout);
 281
 282                         ceph_pagelist_append(&req->r_trail,
 283                                                 &prot_ver, sizeof(prot_ver));
 284                         ceph_pagelist_append(&req->r_trail,
 285                                                 &timeout, sizeof(timeout));
 286                 }
 287         case CEPH_OSD_OP_NOTIFY_ACK:
 288         case CEPH_OSD_OP_WATCH:
 289                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 290                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 291                 dst->watch.flag = src->watch.flag;
 292                 break;
 293         default:
 294                 pr_err("unrecognized osd opcode %d\n", dst->op);
 295                 WARN_ON(1);
 296                 break;
 297         }
 298         dst->payload_len = cpu_to_le32(src->payload_len);
 299 }
 300
 301 /*
 302  * build new request AND message
 303  *
 304  */
 305 void ceph_osdc_build_request(struct ceph_osd_request *req,
 306                              u64 off, u64 len, unsigned int num_op,
 307                              struct ceph_osd_req_op *src_ops,
 308                              struct ceph_snap_context *snapc, u64 snap_id,
 309                              struct timespec *mtime)
 310 {
 311         struct ceph_msg *msg = req->r_request;
 312         struct ceph_osd_request_head *head;
 313         struct ceph_osd_req_op *src_op;
 314         struct ceph_osd_op *op;
 315         void *p;
 316         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 317         int flags = req->r_flags;
 318         u64 data_len = 0;
 319         int i;
 320
 321         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 322
 323         head = msg->front.iov_base;
 324         head->snapid = cpu_to_le64(snap_id);
 325         op = (void *)(head + 1);
 326         p = (void *)(op + num_op);
 327
 328         req->r_snapc = ceph_get_snap_context(snapc);
 329
 330         head->client_inc = cpu_to_le32(1); /* always, for now. */
 331         head->flags = cpu_to_le32(flags);
 332         if (flags & CEPH_OSD_FLAG_WRITE)
 333                 ceph_encode_timespec(&head->mtime, mtime);
 334         BUG_ON(num_op > (unsigned int) ((u16) -1));
 335         head->num_ops = cpu_to_le16(num_op);
 336
 337         /* fill in oid */
 338         head->object_len = cpu_to_le32(req->r_oid_len);
 339         memcpy(p, req->r_oid, req->r_oid_len);
 340         p += req->r_oid_len;
 341
 342         src_op = src_ops;
 343         while (num_op--)
 344                 osd_req_encode_op(req, op++, src_op++);
 345
 346         data_len += req->r_trail.length;
 347
 348         if (snapc) {
 349                 head->snap_seq = cpu_to_le64(snapc->seq);
 350                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 351                 for (i = 0; i < snapc->num_snaps; i++) {
 352                         put_unaligned_le64(snapc->snaps[i], p);
 353                         p += sizeof(u64);
 354                 }
 355         }
 356
 357         if (flags & CEPH_OSD_FLAG_WRITE) {
 358                 req->r_request->hdr.data_off = cpu_to_le16(off);
 359                 req->r_request->hdr.data_len = cpu_to_le32(len + data_len);
 360         } else if (data_len) {
 361                 req->r_request->hdr.data_off = 0;
 362                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 363         }
 364
 365         req->r_request->page_alignment = req->r_page_alignment;
 366
 367         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 368         msg_size = p - msg->front.iov_base;
 369         msg->front.iov_len = msg_size;
 370         msg->hdr.front_len = cpu_to_le32(msg_size);
 371         return;
 372 }
 373 EXPORT_SYMBOL(ceph_osdc_build_request);
 374
 375 /*
 376  * build new request AND message, calculate layout, and adjust file
 377  * extent as needed.
 378  *
 379  * if the file was recently truncated, we include information about its
 380  * old and new size so that the object can be updated appropriately.  (we
 381  * avoid synchronously deleting truncated objects because it's slow.)
 382  *
 383  * if @do_sync, include a 'startsync' command so that the osd will flush
 384  * data quickly.
 385  */
 386 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 387                                                struct ceph_file_layout *layout,
 388                                                struct ceph_vino vino,
 389                                                u64 off, u64 *plen,
 390                                                int opcode, int flags,
 391                                                struct ceph_snap_context *snapc,
 392                                                int do_sync,
 393                                                u32 truncate_seq,
 394                                                u64 truncate_size,
 395                                                struct timespec *mtime,
 396                                                bool use_mempool,
 397                                                int page_align)
 398 {
 399         struct ceph_osd_req_op ops[2];
 400         struct ceph_osd_request *req;
 401         unsigned int num_op = 1;
 402         int r;
 403
 404         memset(&ops, 0, sizeof ops);
 405
 406         ops[0].op = opcode;
 407         ops[0].extent.truncate_seq = truncate_seq;
 408         ops[0].extent.truncate_size = truncate_size;
 409
 410         if (do_sync) {
 411                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 412                 num_op++;
 413         }
 414
 415         req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
 416                                         GFP_NOFS);
 417         if (!req)
 418                 return ERR_PTR(-ENOMEM);
 419         req->r_flags = flags;
 420
 421         /* calculate max write size */
 422         r = calc_layout(vino, layout, off, plen, req, ops);
 423         if (r < 0)
 424                 return ERR_PTR(r);
 425         req->r_file_layout = *layout;  /* keep a copy */
 426
 427         /* in case it differs from natural (file) alignment that
 428            calc_layout filled in for us */
 429         req->r_num_pages = calc_pages_for(page_align, *plen);
 430         req->r_page_alignment = page_align;
 431
 432         ceph_osdc_build_request(req, off, *plen, num_op, ops,
 433                                 snapc, vino.snap, mtime);
 434
 435         return req;
 436 }
 437 EXPORT_SYMBOL(ceph_osdc_new_request);
 438
 439 /*
 440  * We keep osd requests in an rbtree, sorted by ->r_tid.
 441  */
 442 static void __insert_request(struct ceph_osd_client *osdc,
 443                              struct ceph_osd_request *new)
 444 {
 445         struct rb_node **p = &osdc->requests.rb_node;
 446         struct rb_node *parent = NULL;
 447         struct ceph_osd_request *req = NULL;
 448
 449         while (*p) {
 450                 parent = *p;
 451                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 452                 if (new->r_tid < req->r_tid)
 453                         p = &(*p)->rb_left;
 454                 else if (new->r_tid > req->r_tid)
 455                         p = &(*p)->rb_right;
 456                 else
 457                         BUG();
 458         }
 459
 460         rb_link_node(&new->r_node, parent, p);
 461         rb_insert_color(&new->r_node, &osdc->requests);
 462 }
 463
 464 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 465                                                  u64 tid)
 466 {
 467         struct ceph_osd_request *req;
 468         struct rb_node *n = osdc->requests.rb_node;
 469
 470         while (n) {
 471                 req = rb_entry(n, struct ceph_osd_request, r_node);
 472                 if (tid < req->r_tid)
 473                         n = n->rb_left;
 474                 else if (tid > req->r_tid)
 475                         n = n->rb_right;
 476                 else
 477                         return req;
 478         }
 479         return NULL;
 480 }
 481
 482 static struct ceph_osd_request *
 483 __lookup_request_ge(struct ceph_osd_client *osdc,
 484                     u64 tid)
 485 {
 486         struct ceph_osd_request *req;
 487         struct rb_node *n = osdc->requests.rb_node;
 488
 489         while (n) {
 490                 req = rb_entry(n, struct ceph_osd_request, r_node);
 491                 if (tid < req->r_tid) {
 492                         if (!n->rb_left)
 493                                 return req;
 494                         n = n->rb_left;
 495                 } else if (tid > req->r_tid) {
 496                         n = n->rb_right;
 497                 } else {
 498                         return req;
 499                 }
 500         }
 501         return NULL;
 502 }
 503
 504 /*
 505  * Resubmit requests pending on the given osd.
 506  */
 507 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 508                                 struct ceph_osd *osd)
 509 {
 510         struct ceph_osd_request *req, *nreq;
 511         int err;
 512
 513         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 514         err = __reset_osd(osdc, osd);
 515         if (err)
 516                 return;
 517
 518         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 519                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 520                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 521                      osd->o_osd);
 522                 if (!req->r_linger)
 523                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 524         }
 525
 526         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 527                                  r_linger_osd) {
 528                 /*
 529                  * reregister request prior to unregistering linger so
 530                  * that r_osd is preserved.
 531                  */
 532                 BUG_ON(!list_empty(&req->r_req_lru_item));
 533                 __register_request(osdc, req);
 534                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 535                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 536                 __unregister_linger_request(osdc, req);
 537                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 538                      osd->o_osd);
 539         }
 540 }
 541
 542 /*
 543  * If the osd connection drops, we need to resubmit all requests.
 544  */
 545 static void osd_reset(struct ceph_connection *con)
 546 {
 547         struct ceph_osd *osd = con->private;
 548         struct ceph_osd_client *osdc;
 549
 550         if (!osd)
 551                 return;
 552         dout("osd_reset osd%d\n", osd->o_osd);
 553         osdc = osd->o_osdc;
 554         down_read(&osdc->map_sem);
 555         mutex_lock(&osdc->request_mutex);
 556         __kick_osd_requests(osdc, osd);
 557         mutex_unlock(&osdc->request_mutex);
 558         send_queued(osdc);
 559         up_read(&osdc->map_sem);
 560 }
 561
 562 /*
 563  * Track open sessions with osds.
 564  */
 565 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 566 {
 567         struct ceph_osd *osd;
 568
 569         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 570         if (!osd)
 571                 return NULL;
 572
 573         atomic_set(&osd->o_ref, 1);
 574         osd->o_osdc = osdc;
 575         osd->o_osd = onum;
 576         RB_CLEAR_NODE(&osd->o_node);
 577         INIT_LIST_HEAD(&osd->o_requests);
 578         INIT_LIST_HEAD(&osd->o_linger_requests);
 579         INIT_LIST_HEAD(&osd->o_osd_lru);
 580         osd->o_incarnation = 1;
 581
 582         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 583
 584         INIT_LIST_HEAD(&osd->o_keepalive_item);
 585         return osd;
 586 }
 587
 588 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 589 {
 590         if (atomic_inc_not_zero(&osd->o_ref)) {
 591                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 592                      atomic_read(&osd->o_ref));
 593                 return osd;
 594         } else {
 595                 dout("get_osd %p FAIL\n", osd);
 596                 return NULL;
 597         }
 598 }
 599
 600 static void put_osd(struct ceph_osd *osd)
 601 {
 602         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 603              atomic_read(&osd->o_ref) - 1);
 604         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 605                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 606
 607                 if (ac->ops && ac->ops->destroy_authorizer)
 608                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 609                 kfree(osd);
 610         }
 611 }
 612
 613 /*
 614  * remove an osd from our map
 615  */
 616 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 617 {
 618         dout("__remove_osd %p\n", osd);
 619         BUG_ON(!list_empty(&osd->o_requests));
 620         rb_erase(&osd->o_node, &osdc->osds);
 621         list_del_init(&osd->o_osd_lru);
 622         ceph_con_close(&osd->o_con);
 623         put_osd(osd);
 624 }
 625
 626 static void remove_all_osds(struct ceph_osd_client *osdc)
 627 {
 628         dout("%s %p\n", __func__, osdc);
 629         mutex_lock(&osdc->request_mutex);
 630         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 631                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 632                                                 struct ceph_osd, o_node);
 633                 __remove_osd(osdc, osd);
 634         }
 635         mutex_unlock(&osdc->request_mutex);
 636 }
 637
 638 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 639                               struct ceph_osd *osd)
 640 {
 641         dout("__move_osd_to_lru %p\n", osd);
 642         BUG_ON(!list_empty(&osd->o_osd_lru));
 643         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 644         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 645 }
 646
 647 static void __remove_osd_from_lru(struct ceph_osd *osd)
 648 {
 649         dout("__remove_osd_from_lru %p\n", osd);
 650         if (!list_empty(&osd->o_osd_lru))
 651                 list_del_init(&osd->o_osd_lru);
 652 }
 653
 654 static void remove_old_osds(struct ceph_osd_client *osdc)
 655 {
 656         struct ceph_osd *osd, *nosd;
 657
 658         dout("__remove_old_osds %p\n", osdc);
 659         mutex_lock(&osdc->request_mutex);
 660         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 661                 if (time_before(jiffies, osd->lru_ttl))
 662                         break;
 663                 __remove_osd(osdc, osd);
 664         }
 665         mutex_unlock(&osdc->request_mutex);
 666 }
 667
 668 /*
 669  * reset osd connect
 670  */
 671 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 672 {
 673         struct ceph_entity_addr *peer_addr;
 674
 675         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 676         if (list_empty(&osd->o_requests) &&
 677             list_empty(&osd->o_linger_requests)) {
 678                 __remove_osd(osdc, osd);
 679
 680                 return -ENODEV;
 681         }
 682
 683         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
 684         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
 685                         !ceph_con_opened(&osd->o_con)) {
 686                 struct ceph_osd_request *req;
 687
 688                 dout(" osd addr hasn't changed and connection never opened,"
 689                      " letting msgr retry");
 690                 /* touch each r_stamp for handle_timeout()'s benfit */
 691                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 692                         req->r_stamp = jiffies;
 693
 694                 return -EAGAIN;
 695         }
 696
 697         ceph_con_close(&osd->o_con);
 698         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
 699         osd->o_incarnation++;
 700
 701         return 0;
 702 }
 703
 704 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 705 {
 706         struct rb_node **p = &osdc->osds.rb_node;
 707         struct rb_node *parent = NULL;
 708         struct ceph_osd *osd = NULL;
 709
 710         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 711         while (*p) {
 712                 parent = *p;
 713                 osd = rb_entry(parent, struct ceph_osd, o_node);
 714                 if (new->o_osd < osd->o_osd)
 715                         p = &(*p)->rb_left;
 716                 else if (new->o_osd > osd->o_osd)
 717                         p = &(*p)->rb_right;
 718                 else
 719                         BUG();
 720         }
 721
 722         rb_link_node(&new->o_node, parent, p);
 723         rb_insert_color(&new->o_node, &osdc->osds);
 724 }
 725
 726 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 727 {
 728         struct ceph_osd *osd;
 729         struct rb_node *n = osdc->osds.rb_node;
 730
 731         while (n) {
 732                 osd = rb_entry(n, struct ceph_osd, o_node);
 733                 if (o < osd->o_osd)
 734                         n = n->rb_left;
 735                 else if (o > osd->o_osd)
 736                         n = n->rb_right;
 737                 else
 738                         return osd;
 739         }
 740         return NULL;
 741 }
 742
 743 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 744 {
 745         schedule_delayed_work(&osdc->timeout_work,
 746                         osdc->client->options->osd_keepalive_timeout * HZ);
 747 }
 748
 749 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 750 {
 751         cancel_delayed_work(&osdc->timeout_work);
 752 }
 753
 754 /*
 755  * Register request, assign tid.  If this is the first request, set up
 756  * the timeout event.
 757  */
 758 static void __register_request(struct ceph_osd_client *osdc,
 759                                struct ceph_osd_request *req)
 760 {
 761         req->r_tid = ++osdc->last_tid;
 762         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 763         dout("__register_request %p tid %lld\n", req, req->r_tid);
 764         __insert_request(osdc, req);
 765         ceph_osdc_get_request(req);
 766         osdc->num_requests++;
 767         if (osdc->num_requests == 1) {
 768                 dout(" first request, scheduling timeout\n");
 769                 __schedule_osd_timeout(osdc);
 770         }
 771 }
 772
 773 static void register_request(struct ceph_osd_client *osdc,
 774                              struct ceph_osd_request *req)
 775 {
 776         mutex_lock(&osdc->request_mutex);
 777         __register_request(osdc, req);
 778         mutex_unlock(&osdc->request_mutex);
 779 }
 780
 781 /*
 782  * called under osdc->request_mutex
 783  */
 784 static void __unregister_request(struct ceph_osd_client *osdc,
 785                                  struct ceph_osd_request *req)
 786 {
 787         if (RB_EMPTY_NODE(&req->r_node)) {
 788                 dout("__unregister_request %p tid %lld not registered\n",
 789                         req, req->r_tid);
 790                 return;
 791         }
 792
 793         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 794         rb_erase(&req->r_node, &osdc->requests);
 795         osdc->num_requests--;
 796
 797         if (req->r_osd) {
 798                 /* make sure the original request isn't in flight. */
 799                 ceph_msg_revoke(req->r_request);
 800
 801                 list_del_init(&req->r_osd_item);
 802                 if (list_empty(&req->r_osd->o_requests) &&
 803                     list_empty(&req->r_osd->o_linger_requests)) {
 804                         dout("moving osd to %p lru\n", req->r_osd);
 805                         __move_osd_to_lru(osdc, req->r_osd);
 806                 }
 807                 if (list_empty(&req->r_linger_item))
 808                         req->r_osd = NULL;
 809         }
 810
 811         list_del_init(&req->r_req_lru_item);
 812         ceph_osdc_put_request(req);
 813
 814         if (osdc->num_requests == 0) {
 815                 dout(" no requests, canceling timeout\n");
 816                 __cancel_osd_timeout(osdc);
 817         }
 818 }
 819
 820 /*
 821  * Cancel a previously queued request message
 822  */
 823 static void __cancel_request(struct ceph_osd_request *req)
 824 {
 825         if (req->r_sent && req->r_osd) {
 826                 ceph_msg_revoke(req->r_request);
 827                 req->r_sent = 0;
 828         }
 829 }
 830
 831 static void __register_linger_request(struct ceph_osd_client *osdc,
 832                                     struct ceph_osd_request *req)
 833 {
 834         dout("__register_linger_request %p\n", req);
 835         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 836         if (req->r_osd)
 837                 list_add_tail(&req->r_linger_osd,
 838                               &req->r_osd->o_linger_requests);
 839 }
 840
 841 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 842                                         struct ceph_osd_request *req)
 843 {
 844         dout("__unregister_linger_request %p\n", req);
 845         list_del_init(&req->r_linger_item);
 846         if (req->r_osd) {
 847                 list_del_init(&req->r_linger_osd);
 848
 849                 if (list_empty(&req->r_osd->o_requests) &&
 850                     list_empty(&req->r_osd->o_linger_requests)) {
 851                         dout("moving osd to %p lru\n", req->r_osd);
 852                         __move_osd_to_lru(osdc, req->r_osd);
 853                 }
 854                 if (list_empty(&req->r_osd_item))
 855                         req->r_osd = NULL;
 856         }
 857 }
 858
 859 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 860                                          struct ceph_osd_request *req)
 861 {
 862         mutex_lock(&osdc->request_mutex);
 863         if (req->r_linger) {
 864                 __unregister_linger_request(osdc, req);
 865                 ceph_osdc_put_request(req);
 866         }
 867         mutex_unlock(&osdc->request_mutex);
 868 }
 869 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 870
 871 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 872                                   struct ceph_osd_request *req)
 873 {
 874         if (!req->r_linger) {
 875                 dout("set_request_linger %p\n", req);
 876                 req->r_linger = 1;
 877                 /*
 878                  * caller is now responsible for calling
 879                  * unregister_linger_request
 880                  */
 881                 ceph_osdc_get_request(req);
 882         }
 883 }
 884 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 885
 886 /*
 887  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 888  * (as needed), and set the request r_osd appropriately.  If there is
 889  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 890  * (unsent, homeless) or leave on in-flight lru.
 891  *
 892  * Return 0 if unchanged, 1 if changed, or negative on error.
 893  *
 894  * Caller should hold map_sem for read and request_mutex.
 895  */
 896 static int __map_request(struct ceph_osd_client *osdc,
 897                          struct ceph_osd_request *req, int force_resend)
 898 {
 899         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 900         struct ceph_pg pgid;
 901         int acting[CEPH_PG_MAX_SIZE];
 902         int o = -1, num = 0;
 903         int err;
 904
 905         dout("map_request %p tid %lld\n", req, req->r_tid);
 906         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 907                                       &req->r_file_layout, osdc->osdmap);
 908         if (err) {
 909                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 910                 return err;
 911         }
 912         pgid = reqhead->layout.ol_pgid;
 913         req->r_pgid = pgid;
 914
 915         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 916         if (err > 0) {
 917                 o = acting[0];
 918                 num = err;
 919         }
 920
 921         if ((!force_resend &&
 922              req->r_osd && req->r_osd->o_osd == o &&
 923              req->r_sent >= req->r_osd->o_incarnation &&
 924              req->r_num_pg_osds == num &&
 925              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 926             (req->r_osd == NULL && o == -1))
 927                 return 0;  /* no change */
 928
 929         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 930              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 931              req->r_osd ? req->r_osd->o_osd : -1);
 932
 933         /* record full pg acting set */
 934         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 935         req->r_num_pg_osds = num;
 936
 937         if (req->r_osd) {
 938                 __cancel_request(req);
 939                 list_del_init(&req->r_osd_item);
 940                 req->r_osd = NULL;
 941         }
 942
 943         req->r_osd = __lookup_osd(osdc, o);
 944         if (!req->r_osd && o >= 0) {
 945                 err = -ENOMEM;
 946                 req->r_osd = create_osd(osdc, o);
 947                 if (!req->r_osd) {
 948                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
 949                         goto out;
 950                 }
 951
 952                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
 953                 __insert_osd(osdc, req->r_osd);
 954
 955                 ceph_con_open(&req->r_osd->o_con,
 956                               CEPH_ENTITY_TYPE_OSD, o,
 957                               &osdc->osdmap->osd_addr[o]);
 958         }
 959
 960         if (req->r_osd) {
 961                 __remove_osd_from_lru(req->r_osd);
 962                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 963                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 964         } else {
 965                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 966         }
 967         err = 1;   /* osd or pg changed */
 968
 969 out:
 970         return err;
 971 }
 972
 973 /*
 974  * caller should hold map_sem (for read) and request_mutex
 975  */
 976 static void __send_request(struct ceph_osd_client *osdc,
 977                            struct ceph_osd_request *req)
 978 {
 979         struct ceph_osd_request_head *reqhead;
 980
 981         dout("send_request %p tid %llu to osd%d flags %d\n",
 982              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
 983
 984         reqhead = req->r_request->front.iov_base;
 985         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
 986         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
 987         reqhead->reassert_version = req->r_reassert_version;
 988
 989         req->r_stamp = jiffies;
 990         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
 991
 992         ceph_msg_get(req->r_request); /* send consumes a ref */
 993         ceph_con_send(&req->r_osd->o_con, req->r_request);
 994         req->r_sent = req->r_osd->o_incarnation;
 995 }
 996
 997 /*
 998  * Send any requests in the queue (req_unsent).
 999  */
1000 static void send_queued(struct ceph_osd_client *osdc)
1001 {
1002         struct ceph_osd_request *req, *tmp;
1003
1004         dout("send_queued\n");
1005         mutex_lock(&osdc->request_mutex);
1006         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1007                 __send_request(osdc, req);
1008         }
1009         mutex_unlock(&osdc->request_mutex);
1010 }
1011
1012 /*
1013  * Timeout callback, called every N seconds when 1 or more osd
1014  * requests has been active for more than N seconds.  When this
1015  * happens, we ping all OSDs with requests who have timed out to
1016  * ensure any communications channel reset is detected.  Reset the
1017  * request timeouts another N seconds in the future as we go.
1018  * Reschedule the timeout event another N seconds in future (unless
1019  * there are no open requests).
1020  */
1021 static void handle_timeout(struct work_struct *work)
1022 {
1023         struct ceph_osd_client *osdc =
1024                 container_of(work, struct ceph_osd_client, timeout_work.work);
1025         struct ceph_osd_request *req;
1026         struct ceph_osd *osd;
1027         unsigned long keepalive =
1028                 osdc->client->options->osd_keepalive_timeout * HZ;
1029         struct list_head slow_osds;
1030         dout("timeout\n");
1031         down_read(&osdc->map_sem);
1032
1033         ceph_monc_request_next_osdmap(&osdc->client->monc);
1034
1035         mutex_lock(&osdc->request_mutex);
1036
1037         /*
1038          * ping osds that are a bit slow.  this ensures that if there
1039          * is a break in the TCP connection we will notice, and reopen
1040          * a connection with that osd (from the fault callback).
1041          */
1042         INIT_LIST_HEAD(&slow_osds);
1043         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1044                 if (time_before(jiffies, req->r_stamp + keepalive))
1045                         break;
1046
1047                 osd = req->r_osd;
1048                 BUG_ON(!osd);
1049                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1050                      req->r_tid, osd->o_osd);
1051                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1052         }
1053         while (!list_empty(&slow_osds)) {
1054                 osd = list_entry(slow_osds.next, struct ceph_osd,
1055                                  o_keepalive_item);
1056                 list_del_init(&osd->o_keepalive_item);
1057                 ceph_con_keepalive(&osd->o_con);
1058         }
1059
1060         __schedule_osd_timeout(osdc);
1061         mutex_unlock(&osdc->request_mutex);
1062         send_queued(osdc);
1063         up_read(&osdc->map_sem);
1064 }
1065
1066 static void handle_osds_timeout(struct work_struct *work)
1067 {
1068         struct ceph_osd_client *osdc =
1069                 container_of(work, struct ceph_osd_client,
1070                              osds_timeout_work.work);
1071         unsigned long delay =
1072                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1073
1074         dout("osds timeout\n");
1075         down_read(&osdc->map_sem);
1076         remove_old_osds(osdc);
1077         up_read(&osdc->map_sem);
1078
1079         schedule_delayed_work(&osdc->osds_timeout_work,
1080                               round_jiffies_relative(delay));
1081 }
1082
1083 static void complete_request(struct ceph_osd_request *req)
1084 {
1085         if (req->r_safe_callback)
1086                 req->r_safe_callback(req, NULL);
1087         complete_all(&req->r_safe_completion);  /* fsync waiter */
1088 }
1089
1090 /*
1091  * handle osd op reply.  either call the callback if it is specified,
1092  * or do the completion to wake up the waiting thread.
1093  */
1094 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1095                          struct ceph_connection *con)
1096 {
1097         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1098         struct ceph_osd_request *req;
1099         u64 tid;
1100         int numops, object_len, flags;
1101         s32 result;
1102
1103         tid = le64_to_cpu(msg->hdr.tid);
1104         if (msg->front.iov_len < sizeof(*rhead))
1105                 goto bad;
1106         numops = le32_to_cpu(rhead->num_ops);
1107         object_len = le32_to_cpu(rhead->object_len);
1108         result = le32_to_cpu(rhead->result);
1109         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1110             numops * sizeof(struct ceph_osd_op))
1111                 goto bad;
1112         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1113         /* lookup */
1114         mutex_lock(&osdc->request_mutex);
1115         req = __lookup_request(osdc, tid);
1116         if (req == NULL) {
1117                 dout("handle_reply tid %llu dne\n", tid);
1118                 mutex_unlock(&osdc->request_mutex);
1119                 return;
1120         }
1121         ceph_osdc_get_request(req);
1122         flags = le32_to_cpu(rhead->flags);
1123
1124         /*
1125          * if this connection filled our message, drop our reference now, to
1126          * avoid a (safe but slower) revoke later.
1127          */
1128         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1129                 dout(" dropping con_filling_msg ref %p\n", con);
1130                 req->r_con_filling_msg = NULL;
1131                 con->ops->put(con);
1132         }
1133
1134         if (!req->r_got_reply) {
1135                 unsigned int bytes;
1136
1137                 req->r_result = le32_to_cpu(rhead->result);
1138                 bytes = le32_to_cpu(msg->hdr.data_len);
1139                 dout("handle_reply result %d bytes %d\n", req->r_result,
1140                      bytes);
1141                 if (req->r_result == 0)
1142                         req->r_result = bytes;
1143
1144                 /* in case this is a write and we need to replay, */
1145                 req->r_reassert_version = rhead->reassert_version;
1146
1147                 req->r_got_reply = 1;
1148         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1149                 dout("handle_reply tid %llu dup ack\n", tid);
1150                 mutex_unlock(&osdc->request_mutex);
1151                 goto done;
1152         }
1153
1154         dout("handle_reply tid %llu flags %d\n", tid, flags);
1155
1156         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1157                 __register_linger_request(osdc, req);
1158
1159         /* either this is a read, or we got the safe response */
1160         if (result < 0 ||
1161             (flags & CEPH_OSD_FLAG_ONDISK) ||
1162             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1163                 __unregister_request(osdc, req);
1164
1165         mutex_unlock(&osdc->request_mutex);
1166
1167         if (req->r_callback)
1168                 req->r_callback(req, msg);
1169         else
1170                 complete_all(&req->r_completion);
1171
1172         if (flags & CEPH_OSD_FLAG_ONDISK)
1173                 complete_request(req);
1174
1175 done:
1176         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1177         ceph_osdc_put_request(req);
1178         return;
1179
1180 bad:
1181         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1182                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1183                (int)sizeof(*rhead));
1184         ceph_msg_dump(msg);
1185 }
1186
1187 static void reset_changed_osds(struct ceph_osd_client *osdc)
1188 {
1189         struct rb_node *p, *n;
1190
1191         for (p = rb_first(&osdc->osds); p; p = n) {
1192                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1193
1194                 n = rb_next(p);
1195                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1196                     memcmp(&osd->o_con.peer_addr,
1197                            ceph_osd_addr(osdc->osdmap,
1198                                          osd->o_osd),
1199                            sizeof(struct ceph_entity_addr)) != 0)
1200                         __reset_osd(osdc, osd);
1201         }
1202 }
1203
1204 /*
1205  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1206  * no osd, request a new map.
1207  *
1208  * Caller should hold map_sem for read.
1209  */
1210 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1211 {
1212         struct ceph_osd_request *req, *nreq;
1213         struct rb_node *p;
1214         int needmap = 0;
1215         int err;
1216
1217         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1218         mutex_lock(&osdc->request_mutex);
1219         for (p = rb_first(&osdc->requests); p; ) {
1220                 req = rb_entry(p, struct ceph_osd_request, r_node);
1221                 p = rb_next(p);
1222
1223                 /*
1224                  * For linger requests that have not yet been
1225                  * registered, move them to the linger list; they'll
1226                  * be sent to the osd in the loop below.  Unregister
1227                  * the request before re-registering it as a linger
1228                  * request to ensure the __map_request() below
1229                  * will decide it needs to be sent.
1230                  */
1231                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1232                         dout("%p tid %llu restart on osd%d\n",
1233                              req, req->r_tid,
1234                              req->r_osd ? req->r_osd->o_osd : -1);
1235                         __unregister_request(osdc, req);
1236                         __register_linger_request(osdc, req);
1237                         continue;
1238                 }
1239
1240                 err = __map_request(osdc, req, force_resend);
1241                 if (err < 0)
1242                         continue;  /* error */
1243                 if (req->r_osd == NULL) {
1244                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1245                         needmap++;  /* request a newer map */
1246                 } else if (err > 0) {
1247                         if (!req->r_linger) {
1248                                 dout("%p tid %llu requeued on osd%d\n", req,
1249                                      req->r_tid,
1250                                      req->r_osd ? req->r_osd->o_osd : -1);
1251                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1252                         }
1253                 }
1254         }
1255
1256         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1257                                  r_linger_item) {
1258                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1259
1260                 err = __map_request(osdc, req, force_resend);
1261                 dout("__map_request returned %d\n", err);
1262                 if (err == 0)
1263                         continue;  /* no change and no osd was specified */
1264                 if (err < 0)
1265                         continue;  /* hrm! */
1266                 if (req->r_osd == NULL) {
1267                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1268                         needmap++;  /* request a newer map */
1269                         continue;
1270                 }
1271
1272                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1273                      req->r_osd ? req->r_osd->o_osd : -1);
1274                 __register_request(osdc, req);
1275                 __unregister_linger_request(osdc, req);
1276         }
1277         mutex_unlock(&osdc->request_mutex);
1278
1279         if (needmap) {
1280                 dout("%d requests for down osds, need new map\n", needmap);
1281                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1282         }
1283         reset_changed_osds(osdc);
1284 }
1285
1286
1287 /*
1288  * Process updated osd map.
1289  *
1290  * The message contains any number of incremental and full maps, normally
1291  * indicating some sort of topology change in the cluster.  Kick requests
1292  * off to different OSDs as needed.
1293  */
1294 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1295 {
1296         void *p, *end, *next;
1297         u32 nr_maps, maplen;
1298         u32 epoch;
1299         struct ceph_osdmap *newmap = NULL, *oldmap;
1300         int err;
1301         struct ceph_fsid fsid;
1302
1303         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1304         p = msg->front.iov_base;
1305         end = p + msg->front.iov_len;
1306
1307         /* verify fsid */
1308         ceph_decode_need(&p, end, sizeof(fsid), bad);
1309         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1310         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1311                 return;
1312
1313         down_write(&osdc->map_sem);
1314
1315         /* incremental maps */
1316         ceph_decode_32_safe(&p, end, nr_maps, bad);
1317         dout(" %d inc maps\n", nr_maps);
1318         while (nr_maps > 0) {
1319                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1320                 epoch = ceph_decode_32(&p);
1321                 maplen = ceph_decode_32(&p);
1322                 ceph_decode_need(&p, end, maplen, bad);
1323                 next = p + maplen;
1324                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1325                         dout("applying incremental map %u len %d\n",
1326                              epoch, maplen);
1327                         newmap = osdmap_apply_incremental(&p, next,
1328                                                           osdc->osdmap,
1329                                                           &osdc->client->msgr);
1330                         if (IS_ERR(newmap)) {
1331                                 err = PTR_ERR(newmap);
1332                                 goto bad;
1333                         }
1334                         BUG_ON(!newmap);
1335                         if (newmap != osdc->osdmap) {
1336                                 ceph_osdmap_destroy(osdc->osdmap);
1337                                 osdc->osdmap = newmap;
1338                         }
1339                         kick_requests(osdc, 0);
1340                 } else {
1341                         dout("ignoring incremental map %u len %d\n",
1342                              epoch, maplen);
1343                 }
1344                 p = next;
1345                 nr_maps--;
1346         }
1347         if (newmap)
1348                 goto done;
1349
1350         /* full maps */
1351         ceph_decode_32_safe(&p, end, nr_maps, bad);
1352         dout(" %d full maps\n", nr_maps);
1353         while (nr_maps) {
1354                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1355                 epoch = ceph_decode_32(&p);
1356                 maplen = ceph_decode_32(&p);
1357                 ceph_decode_need(&p, end, maplen, bad);
1358                 if (nr_maps > 1) {
1359                         dout("skipping non-latest full map %u len %d\n",
1360                              epoch, maplen);
1361                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1362                         dout("skipping full map %u len %d, "
1363                              "older than our %u\n", epoch, maplen,
1364                              osdc->osdmap->epoch);
1365                 } else {
1366                         int skipped_map = 0;
1367
1368                         dout("taking full map %u len %d\n", epoch, maplen);
1369                         newmap = osdmap_decode(&p, p+maplen);
1370                         if (IS_ERR(newmap)) {
1371                                 err = PTR_ERR(newmap);
1372                                 goto bad;
1373                         }
1374                         BUG_ON(!newmap);
1375                         oldmap = osdc->osdmap;
1376                         osdc->osdmap = newmap;
1377                         if (oldmap) {
1378                                 if (oldmap->epoch + 1 < newmap->epoch)
1379                                         skipped_map = 1;
1380                                 ceph_osdmap_destroy(oldmap);
1381                         }
1382                         kick_requests(osdc, skipped_map);
1383                 }
1384                 p += maplen;
1385                 nr_maps--;
1386         }
1387
1388 done:
1389         downgrade_write(&osdc->map_sem);
1390         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1391
1392         /*
1393          * subscribe to subsequent osdmap updates if full to ensure
1394          * we find out when we are no longer full and stop returning
1395          * ENOSPC.
1396          */
1397         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1398                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1399
1400         send_queued(osdc);
1401         up_read(&osdc->map_sem);
1402         wake_up_all(&osdc->client->auth_wq);
1403         return;
1404
1405 bad:
1406         pr_err("osdc handle_map corrupt msg\n");
1407         ceph_msg_dump(msg);
1408         up_write(&osdc->map_sem);
1409         return;
1410 }
1411
1412 /*
1413  * watch/notify callback event infrastructure
1414  *
1415  * These callbacks are used both for watch and notify operations.
1416  */
1417 static void __release_event(struct kref *kref)
1418 {
1419         struct ceph_osd_event *event =
1420                 container_of(kref, struct ceph_osd_event, kref);
1421
1422         dout("__release_event %p\n", event);
1423         kfree(event);
1424 }
1425
1426 static void get_event(struct ceph_osd_event *event)
1427 {
1428         kref_get(&event->kref);
1429 }
1430
1431 void ceph_osdc_put_event(struct ceph_osd_event *event)
1432 {
1433         kref_put(&event->kref, __release_event);
1434 }
1435 EXPORT_SYMBOL(ceph_osdc_put_event);
1436
1437 static void __insert_event(struct ceph_osd_client *osdc,
1438                              struct ceph_osd_event *new)
1439 {
1440         struct rb_node **p = &osdc->event_tree.rb_node;
1441         struct rb_node *parent = NULL;
1442         struct ceph_osd_event *event = NULL;
1443
1444         while (*p) {
1445                 parent = *p;
1446                 event = rb_entry(parent, struct ceph_osd_event, node);
1447                 if (new->cookie < event->cookie)
1448                         p = &(*p)->rb_left;
1449                 else if (new->cookie > event->cookie)
1450                         p = &(*p)->rb_right;
1451                 else
1452                         BUG();
1453         }
1454
1455         rb_link_node(&new->node, parent, p);
1456         rb_insert_color(&new->node, &osdc->event_tree);
1457 }
1458
1459 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1460                                                 u64 cookie)
1461 {
1462         struct rb_node **p = &osdc->event_tree.rb_node;
1463         struct rb_node *parent = NULL;
1464         struct ceph_osd_event *event = NULL;
1465
1466         while (*p) {
1467                 parent = *p;
1468                 event = rb_entry(parent, struct ceph_osd_event, node);
1469                 if (cookie < event->cookie)
1470                         p = &(*p)->rb_left;
1471                 else if (cookie > event->cookie)
1472                         p = &(*p)->rb_right;
1473                 else
1474                         return event;
1475         }
1476         return NULL;
1477 }
1478
1479 static void __remove_event(struct ceph_osd_event *event)
1480 {
1481         struct ceph_osd_client *osdc = event->osdc;
1482
1483         if (!RB_EMPTY_NODE(&event->node)) {
1484                 dout("__remove_event removed %p\n", event);
1485                 rb_erase(&event->node, &osdc->event_tree);
1486                 ceph_osdc_put_event(event);
1487         } else {
1488                 dout("__remove_event didn't remove %p\n", event);
1489         }
1490 }
1491
1492 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1493                            void (*event_cb)(u64, u64, u8, void *),
1494                            int one_shot, void *data,
1495                            struct ceph_osd_event **pevent)
1496 {
1497         struct ceph_osd_event *event;
1498
1499         event = kmalloc(sizeof(*event), GFP_NOIO);
1500         if (!event)
1501                 return -ENOMEM;
1502
1503         dout("create_event %p\n", event);
1504         event->cb = event_cb;
1505         event->one_shot = one_shot;
1506         event->data = data;
1507         event->osdc = osdc;
1508         INIT_LIST_HEAD(&event->osd_node);
1509         RB_CLEAR_NODE(&event->node);
1510         kref_init(&event->kref);   /* one ref for us */
1511         kref_get(&event->kref);    /* one ref for the caller */
1512         init_completion(&event->completion);
1513
1514         spin_lock(&osdc->event_lock);
1515         event->cookie = ++osdc->event_count;
1516         __insert_event(osdc, event);
1517         spin_unlock(&osdc->event_lock);
1518
1519         *pevent = event;
1520         return 0;
1521 }
1522 EXPORT_SYMBOL(ceph_osdc_create_event);
1523
1524 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1525 {
1526         struct ceph_osd_client *osdc = event->osdc;
1527
1528         dout("cancel_event %p\n", event);
1529         spin_lock(&osdc->event_lock);
1530         __remove_event(event);
1531         spin_unlock(&osdc->event_lock);
1532         ceph_osdc_put_event(event); /* caller's */
1533 }
1534 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1535
1536
1537 static void do_event_work(struct work_struct *work)
1538 {
1539         struct ceph_osd_event_work *event_work =
1540                 container_of(work, struct ceph_osd_event_work, work);
1541         struct ceph_osd_event *event = event_work->event;
1542         u64 ver = event_work->ver;
1543         u64 notify_id = event_work->notify_id;
1544         u8 opcode = event_work->opcode;
1545
1546         dout("do_event_work completing %p\n", event);
1547         event->cb(ver, notify_id, opcode, event->data);
1548         complete(&event->completion);
1549         dout("do_event_work completed %p\n", event);
1550         ceph_osdc_put_event(event);
1551         kfree(event_work);
1552 }
1553
1554
1555 /*
1556  * Process osd watch notifications
1557  */
1558 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1559 {
1560         void *p, *end;
1561         u8 proto_ver;
1562         u64 cookie, ver, notify_id;
1563         u8 opcode;
1564         struct ceph_osd_event *event;
1565         struct ceph_osd_event_work *event_work;
1566
1567         p = msg->front.iov_base;
1568         end = p + msg->front.iov_len;
1569
1570         ceph_decode_8_safe(&p, end, proto_ver, bad);
1571         ceph_decode_8_safe(&p, end, opcode, bad);
1572         ceph_decode_64_safe(&p, end, cookie, bad);
1573         ceph_decode_64_safe(&p, end, ver, bad);
1574         ceph_decode_64_safe(&p, end, notify_id, bad);
1575
1576         spin_lock(&osdc->event_lock);
1577         event = __find_event(osdc, cookie);
1578         if (event) {
1579                 get_event(event);
1580                 if (event->one_shot)
1581                         __remove_event(event);
1582         }
1583         spin_unlock(&osdc->event_lock);
1584         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1585              cookie, ver, event);
1586         if (event) {
1587                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1588                 if (!event_work) {
1589                         dout("ERROR: could not allocate event_work\n");
1590                         goto done_err;
1591                 }
1592                 INIT_WORK(&event_work->work, do_event_work);
1593                 event_work->event = event;
1594                 event_work->ver = ver;
1595                 event_work->notify_id = notify_id;
1596                 event_work->opcode = opcode;
1597                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1598                         dout("WARNING: failed to queue notify event work\n");
1599                         goto done_err;
1600                 }
1601         }
1602
1603         return;
1604
1605 done_err:
1606         complete(&event->completion);
1607         ceph_osdc_put_event(event);
1608         return;
1609
1610 bad:
1611         pr_err("osdc handle_watch_notify corrupt msg\n");
1612         return;
1613 }
1614
1615 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1616 {
1617         int err;
1618
1619         dout("wait_event %p\n", event);
1620         err = wait_for_completion_interruptible_timeout(&event->completion,
1621                                                         timeout * HZ);
1622         ceph_osdc_put_event(event);
1623         if (err > 0)
1624                 err = 0;
1625         dout("wait_event %p returns %d\n", event, err);
1626         return err;
1627 }
1628 EXPORT_SYMBOL(ceph_osdc_wait_event);
1629
1630 /*
1631  * Register request, send initial attempt.
1632  */
1633 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1634                             struct ceph_osd_request *req,
1635                             bool nofail)
1636 {
1637         int rc = 0;
1638
1639         req->r_request->pages = req->r_pages;
1640         req->r_request->nr_pages = req->r_num_pages;
1641 #ifdef CONFIG_BLOCK
1642         req->r_request->bio = req->r_bio;
1643 #endif
1644         req->r_request->trail = &req->r_trail;
1645
1646         register_request(osdc, req);
1647
1648         down_read(&osdc->map_sem);
1649         mutex_lock(&osdc->request_mutex);
1650         /*
1651          * a racing kick_requests() may have sent the message for us
1652          * while we dropped request_mutex above, so only send now if
1653          * the request still han't been touched yet.
1654          */
1655         if (req->r_sent == 0) {
1656                 rc = __map_request(osdc, req, 0);
1657                 if (rc < 0) {
1658                         if (nofail) {
1659                                 dout("osdc_start_request failed map, "
1660                                      " will retry %lld\n", req->r_tid);
1661                                 rc = 0;
1662                         }
1663                         goto out_unlock;
1664                 }
1665                 if (req->r_osd == NULL) {
1666                         dout("send_request %p no up osds in pg\n", req);
1667                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1668                 } else {
1669                         __send_request(osdc, req);
1670                 }
1671                 rc = 0;
1672         }
1673
1674 out_unlock:
1675         mutex_unlock(&osdc->request_mutex);
1676         up_read(&osdc->map_sem);
1677         return rc;
1678 }
1679 EXPORT_SYMBOL(ceph_osdc_start_request);
1680
1681 /*
1682  * wait for a request to complete
1683  */
1684 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1685                            struct ceph_osd_request *req)
1686 {
1687         int rc;
1688
1689         rc = wait_for_completion_interruptible(&req->r_completion);
1690         if (rc < 0) {
1691                 mutex_lock(&osdc->request_mutex);
1692                 __cancel_request(req);
1693                 __unregister_request(osdc, req);
1694                 mutex_unlock(&osdc->request_mutex);
1695                 complete_request(req);
1696                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1697                 return rc;
1698         }
1699
1700         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1701         return req->r_result;
1702 }
1703 EXPORT_SYMBOL(ceph_osdc_wait_request);
1704
1705 /*
1706  * sync - wait for all in-flight requests to flush.  avoid starvation.
1707  */
1708 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1709 {
1710         struct ceph_osd_request *req;
1711         u64 last_tid, next_tid = 0;
1712
1713         mutex_lock(&osdc->request_mutex);
1714         last_tid = osdc->last_tid;
1715         while (1) {
1716                 req = __lookup_request_ge(osdc, next_tid);
1717                 if (!req)
1718                         break;
1719                 if (req->r_tid > last_tid)
1720                         break;
1721
1722                 next_tid = req->r_tid + 1;
1723                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1724                         continue;
1725
1726                 ceph_osdc_get_request(req);
1727                 mutex_unlock(&osdc->request_mutex);
1728                 dout("sync waiting on tid %llu (last is %llu)\n",
1729                      req->r_tid, last_tid);
1730                 wait_for_completion(&req->r_safe_completion);
1731                 mutex_lock(&osdc->request_mutex);
1732                 ceph_osdc_put_request(req);
1733         }
1734         mutex_unlock(&osdc->request_mutex);
1735         dout("sync done (thru tid %llu)\n", last_tid);
1736 }
1737 EXPORT_SYMBOL(ceph_osdc_sync);
1738
1739 /*
1740  * init, shutdown
1741  */
1742 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1743 {
1744         int err;
1745
1746         dout("init\n");
1747         osdc->client = client;
1748         osdc->osdmap = NULL;
1749         init_rwsem(&osdc->map_sem);
1750         init_completion(&osdc->map_waiters);
1751         osdc->last_requested_map = 0;
1752         mutex_init(&osdc->request_mutex);
1753         osdc->last_tid = 0;
1754         osdc->osds = RB_ROOT;
1755         INIT_LIST_HEAD(&osdc->osd_lru);
1756         osdc->requests = RB_ROOT;
1757         INIT_LIST_HEAD(&osdc->req_lru);
1758         INIT_LIST_HEAD(&osdc->req_unsent);
1759         INIT_LIST_HEAD(&osdc->req_notarget);
1760         INIT_LIST_HEAD(&osdc->req_linger);
1761         osdc->num_requests = 0;
1762         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1763         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1764         spin_lock_init(&osdc->event_lock);
1765         osdc->event_tree = RB_ROOT;
1766         osdc->event_count = 0;
1767
1768         schedule_delayed_work(&osdc->osds_timeout_work,
1769            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1770
1771         err = -ENOMEM;
1772         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1773                                         sizeof(struct ceph_osd_request));
1774         if (!osdc->req_mempool)
1775                 goto out;
1776
1777         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1778                                 OSD_OP_FRONT_LEN, 10, true,
1779                                 "osd_op");
1780         if (err < 0)
1781                 goto out_mempool;
1782         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1783                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1784                                 "osd_op_reply");
1785         if (err < 0)
1786                 goto out_msgpool;
1787
1788         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1789         if (IS_ERR(osdc->notify_wq)) {
1790                 err = PTR_ERR(osdc->notify_wq);
1791                 osdc->notify_wq = NULL;
1792                 goto out_msgpool;
1793         }
1794         return 0;
1795
1796 out_msgpool:
1797         ceph_msgpool_destroy(&osdc->msgpool_op);
1798 out_mempool:
1799         mempool_destroy(osdc->req_mempool);
1800 out:
1801         return err;
1802 }
1803 EXPORT_SYMBOL(ceph_osdc_init);
1804
1805 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1806 {
1807         flush_workqueue(osdc->notify_wq);
1808         destroy_workqueue(osdc->notify_wq);
1809         cancel_delayed_work_sync(&osdc->timeout_work);
1810         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1811         if (osdc->osdmap) {
1812                 ceph_osdmap_destroy(osdc->osdmap);
1813                 osdc->osdmap = NULL;
1814         }
1815         remove_all_osds(osdc);
1816         mempool_destroy(osdc->req_mempool);
1817         ceph_msgpool_destroy(&osdc->msgpool_op);
1818         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1819 }
1820 EXPORT_SYMBOL(ceph_osdc_stop);
1821
1822 /*
1823  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1824  * *plen.  Return number of bytes read, or error.
1825  */
1826 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1827                         struct ceph_vino vino, struct ceph_file_layout *layout,
1828                         u64 off, u64 *plen,
1829                         u32 truncate_seq, u64 truncate_size,
1830                         struct page **pages, int num_pages, int page_align)
1831 {
1832         struct ceph_osd_request *req;
1833         int rc = 0;
1834
1835         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1836              vino.snap, off, *plen);
1837         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1838                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1839                                     NULL, 0, truncate_seq, truncate_size, NULL,
1840                                     false, page_align);
1841         if (IS_ERR(req))
1842                 return PTR_ERR(req);
1843
1844         /* it may be a short read due to an object boundary */
1845         req->r_pages = pages;
1846
1847         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1848              off, *plen, req->r_num_pages, page_align);
1849
1850         rc = ceph_osdc_start_request(osdc, req, false);
1851         if (!rc)
1852                 rc = ceph_osdc_wait_request(osdc, req);
1853
1854         ceph_osdc_put_request(req);
1855         dout("readpages result %d\n", rc);
1856         return rc;
1857 }
1858 EXPORT_SYMBOL(ceph_osdc_readpages);
1859
1860 /*
1861  * do a synchronous write on N pages
1862  */
1863 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1864                          struct ceph_file_layout *layout,
1865                          struct ceph_snap_context *snapc,
1866                          u64 off, u64 len,
1867                          u32 truncate_seq, u64 truncate_size,
1868                          struct timespec *mtime,
1869                          struct page **pages, int num_pages)
1870 {
1871         struct ceph_osd_request *req;
1872         int rc = 0;
1873         int page_align = off & ~PAGE_MASK;
1874
1875         BUG_ON(vino.snap != CEPH_NOSNAP);
1876         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1877                                     CEPH_OSD_OP_WRITE,
1878                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1879                                     snapc, 0,
1880                                     truncate_seq, truncate_size, mtime,
1881                                     true, page_align);
1882         if (IS_ERR(req))
1883                 return PTR_ERR(req);
1884
1885         /* it may be a short write due to an object boundary */
1886         req->r_pages = pages;
1887         dout("writepages %llu~%llu (%d pages)\n", off, len,
1888              req->r_num_pages);
1889
1890         rc = ceph_osdc_start_request(osdc, req, true);
1891         if (!rc)
1892                 rc = ceph_osdc_wait_request(osdc, req);
1893
1894         ceph_osdc_put_request(req);
1895         if (rc == 0)
1896                 rc = len;
1897         dout("writepages result %d\n", rc);
1898         return rc;
1899 }
1900 EXPORT_SYMBOL(ceph_osdc_writepages);
1901
1902 /*
1903  * handle incoming message
1904  */
1905 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1906 {
1907         struct ceph_osd *osd = con->private;
1908         struct ceph_osd_client *osdc;
1909         int type = le16_to_cpu(msg->hdr.type);
1910
1911         if (!osd)
1912                 goto out;
1913         osdc = osd->o_osdc;
1914
1915         switch (type) {
1916         case CEPH_MSG_OSD_MAP:
1917                 ceph_osdc_handle_map(osdc, msg);
1918                 break;
1919         case CEPH_MSG_OSD_OPREPLY:
1920                 handle_reply(osdc, msg, con);
1921                 break;
1922         case CEPH_MSG_WATCH_NOTIFY:
1923                 handle_watch_notify(osdc, msg);
1924                 break;
1925
1926         default:
1927                 pr_err("received unknown message type %d %s\n", type,
1928                        ceph_msg_type_name(type));
1929         }
1930 out:
1931         ceph_msg_put(msg);
1932 }
1933
1934 /*
1935  * lookup and return message for incoming reply.  set up reply message
1936  * pages.
1937  */
1938 static struct ceph_msg *get_reply(struct ceph_connection *con,
1939                                   struct ceph_msg_header *hdr,
1940                                   int *skip)
1941 {
1942         struct ceph_osd *osd = con->private;
1943         struct ceph_osd_client *osdc = osd->o_osdc;
1944         struct ceph_msg *m;
1945         struct ceph_osd_request *req;
1946         int front = le32_to_cpu(hdr->front_len);
1947         int data_len = le32_to_cpu(hdr->data_len);
1948         u64 tid;
1949
1950         tid = le64_to_cpu(hdr->tid);
1951         mutex_lock(&osdc->request_mutex);
1952         req = __lookup_request(osdc, tid);
1953         if (!req) {
1954                 *skip = 1;
1955                 m = NULL;
1956                 dout("get_reply unknown tid %llu from osd%d\n", tid,
1957                      osd->o_osd);
1958                 goto out;
1959         }
1960
1961         if (req->r_con_filling_msg) {
1962                 dout("%s revoking msg %p from old con %p\n", __func__,
1963                      req->r_reply, req->r_con_filling_msg);
1964                 ceph_msg_revoke_incoming(req->r_reply);
1965                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
1966                 req->r_con_filling_msg = NULL;
1967         }
1968
1969         if (front > req->r_reply->front.iov_len) {
1970                 pr_warning("get_reply front %d > preallocated %d\n",
1971                            front, (int)req->r_reply->front.iov_len);
1972                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
1973                 if (!m)
1974                         goto out;
1975                 ceph_msg_put(req->r_reply);
1976                 req->r_reply = m;
1977         }
1978         m = ceph_msg_get(req->r_reply);
1979
1980         if (data_len > 0) {
1981                 int want = calc_pages_for(req->r_page_alignment, data_len);
1982
1983                 if (req->r_pages && unlikely(req->r_num_pages < want)) {
1984                         pr_warning("tid %lld reply has %d bytes %d pages, we"
1985                                    " had only %d pages ready\n", tid, data_len,
1986                                    want, req->r_num_pages);
1987                         *skip = 1;
1988                         ceph_msg_put(m);
1989                         m = NULL;
1990                         goto out;
1991                 }
1992                 m->pages = req->r_pages;
1993                 m->nr_pages = req->r_num_pages;
1994                 m->page_alignment = req->r_page_alignment;
1995 #ifdef CONFIG_BLOCK
1996                 m->bio = req->r_bio;
1997 #endif
1998         }
1999         *skip = 0;
2000         req->r_con_filling_msg = con->ops->get(con);
2001         dout("get_reply tid %lld %p\n", tid, m);
2002
2003 out:
2004         mutex_unlock(&osdc->request_mutex);
2005         return m;
2006
2007 }
2008
2009 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2010                                   struct ceph_msg_header *hdr,
2011                                   int *skip)
2012 {
2013         struct ceph_osd *osd = con->private;
2014         int type = le16_to_cpu(hdr->type);
2015         int front = le32_to_cpu(hdr->front_len);
2016
2017         *skip = 0;
2018         switch (type) {
2019         case CEPH_MSG_OSD_MAP:
2020         case CEPH_MSG_WATCH_NOTIFY:
2021                 return ceph_msg_new(type, front, GFP_NOFS, false);
2022         case CEPH_MSG_OSD_OPREPLY:
2023                 return get_reply(con, hdr, skip);
2024         default:
2025                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2026                         osd->o_osd);
2027                 *skip = 1;
2028                 return NULL;
2029         }
2030 }
2031
2032 /*
2033  * Wrappers to refcount containing ceph_osd struct
2034  */
2035 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2036 {
2037         struct ceph_osd *osd = con->private;
2038         if (get_osd(osd))
2039                 return con;
2040         return NULL;
2041 }
2042
2043 static void put_osd_con(struct ceph_connection *con)
2044 {
2045         struct ceph_osd *osd = con->private;
2046         put_osd(osd);
2047 }
2048
2049 /*
2050  * authentication
2051  */
2052 /*
2053  * Note: returned pointer is the address of a structure that's
2054  * managed separately.  Caller must *not* attempt to free it.
2055  */
2056 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2057                                         int *proto, int force_new)
2058 {
2059         struct ceph_osd *o = con->private;
2060         struct ceph_osd_client *osdc = o->o_osdc;
2061         struct ceph_auth_client *ac = osdc->client->monc.auth;
2062         struct ceph_auth_handshake *auth = &o->o_auth;
2063
2064         if (force_new && auth->authorizer) {
2065                 if (ac->ops && ac->ops->destroy_authorizer)
2066                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2067                 auth->authorizer = NULL;
2068         }
2069         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2070                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2071                                                         auth);
2072                 if (ret)
2073                         return ERR_PTR(ret);
2074         }
2075         *proto = ac->protocol;
2076
2077         return auth;
2078 }
2079
2080
2081 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2082 {
2083         struct ceph_osd *o = con->private;
2084         struct ceph_osd_client *osdc = o->o_osdc;
2085         struct ceph_auth_client *ac = osdc->client->monc.auth;
2086
2087         /*
2088          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2089          * XXX which do we do:  succeed or fail?
2090          */
2091         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2092 }
2093
2094 static int invalidate_authorizer(struct ceph_connection *con)
2095 {
2096         struct ceph_osd *o = con->private;
2097         struct ceph_osd_client *osdc = o->o_osdc;
2098         struct ceph_auth_client *ac = osdc->client->monc.auth;
2099
2100         if (ac->ops && ac->ops->invalidate_authorizer)
2101                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2102
2103         return ceph_monc_validate_auth(&osdc->client->monc);
2104 }
2105
2106 static const struct ceph_connection_operations osd_con_ops = {
2107         .get = get_osd_con,
2108         .put = put_osd_con,
2109         .dispatch = dispatch,
2110         .get_authorizer = get_authorizer,
2111         .verify_authorizer_reply = verify_authorizer_reply,
2112         .invalidate_authorizer = invalidate_authorizer,
2113         .alloc_msg = alloc_msg,
2114         .fault = osd_reset,
2115 };