net/ceph/osd_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/module.h>
   4 #include <linux/err.h>
   5 #include <linux/highmem.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/slab.h>
   9 #include <linux/uaccess.h>
  10 #ifdef CONFIG_BLOCK
  11 #include <linux/bio.h>
  12 #endif
  13
  14 #include <linux/ceph/libceph.h>
  15 #include <linux/ceph/osd_client.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/auth.h>
  19 #include <linux/ceph/pagelist.h>
  20
  21 #define OSD_OP_FRONT_LEN        4096
  22 #define OSD_OPREPLY_FRONT_LEN   512
  23
  24 static const struct ceph_connection_operations osd_con_ops;
  25
  26 static void __send_queued(struct ceph_osd_client *osdc);
  27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  28 static void __register_request(struct ceph_osd_client *osdc,
  29                                struct ceph_osd_request *req);
  30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  31                                         struct ceph_osd_request *req);
  32 static void __send_request(struct ceph_osd_client *osdc,
  33                            struct ceph_osd_request *req);
  34
  35 static int op_has_extent(int op)
  36 {
  37         return (op == CEPH_OSD_OP_READ ||
  38                 op == CEPH_OSD_OP_WRITE);
  39 }
  40
  41 /*
  42  * Implement client access to distributed object storage cluster.
  43  *
  44  * All data objects are stored within a cluster/cloud of OSDs, or
  45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  47  * remote daemons serving up and coordinating consistent and safe
  48  * access to storage.
  49  *
  50  * Cluster membership and the mapping of data objects onto storage devices
  51  * are described by the osd map.
  52  *
  53  * We keep track of pending OSD requests (read, write), resubmit
  54  * requests to different OSDs when the cluster topology/data layout
  55  * change, or retry the affected requests when the communications
  56  * channel with an OSD is reset.
  57  */
  58
  59 /*
  60  * calculate the mapping of a file extent onto an object, and fill out the
  61  * request accordingly.  shorten extent as necessary if it crosses an
  62  * object boundary.
  63  *
  64  * fill osd op in request message.
  65  */
  66 static int calc_layout(struct ceph_vino vino,
  67                        struct ceph_file_layout *layout,
  68                        u64 off, u64 *plen,
  69                        struct ceph_osd_request *req,
  70                        struct ceph_osd_req_op *op)
  71 {
  72         u64 orig_len = *plen;
  73         u64 bno = 0;
  74         u64 objoff = 0;
  75         u64 objlen = 0;
  76         int r;
  77
  78         /* object extent? */
  79         r = ceph_calc_file_object_mapping(layout, off, orig_len, &bno,
  80                                           &objoff, &objlen);
  81         if (r < 0)
  82                 return r;
  83         if (objlen < orig_len) {
  84                 *plen = objlen;
  85                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  86                      orig_len - *plen, off, *plen);
  87         }
  88
  89         if (op_has_extent(op->op)) {
  90                 u32 osize = le32_to_cpu(layout->fl_object_size);
  91                 op->extent.offset = objoff;
  92                 op->extent.length = objlen;
  93                 if (op->extent.truncate_size <= off - objoff) {
  94                         op->extent.truncate_size = 0;
  95                 } else {
  96                         op->extent.truncate_size -= off - objoff;
  97                         if (op->extent.truncate_size > osize)
  98                                 op->extent.truncate_size = osize;
  99                 }
 100         }
 101         req->r_num_pages = calc_pages_for(off, *plen);
 102         req->r_page_alignment = off & ~PAGE_MASK;
 103         if (op->op == CEPH_OSD_OP_WRITE)
 104                 op->payload_len = *plen;
 105
 106         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
 107              bno, objoff, objlen, req->r_num_pages);
 108
 109         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 110         req->r_oid_len = strlen(req->r_oid);
 111
 112         return r;
 113 }
 114
 115 /*
 116  * requests
 117  */
 118 void ceph_osdc_release_request(struct kref *kref)
 119 {
 120         struct ceph_osd_request *req = container_of(kref,
 121                                                     struct ceph_osd_request,
 122                                                     r_kref);
 123
 124         if (req->r_request)
 125                 ceph_msg_put(req->r_request);
 126         if (req->r_con_filling_msg) {
 127                 dout("%s revoking msg %p from con %p\n", __func__,
 128                      req->r_reply, req->r_con_filling_msg);
 129                 ceph_msg_revoke_incoming(req->r_reply);
 130                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 131                 req->r_con_filling_msg = NULL;
 132         }
 133         if (req->r_reply)
 134                 ceph_msg_put(req->r_reply);
 135         if (req->r_own_pages)
 136                 ceph_release_page_vector(req->r_pages,
 137                                          req->r_num_pages);
 138         ceph_put_snap_context(req->r_snapc);
 139         ceph_pagelist_release(&req->r_trail);
 140         if (req->r_mempool)
 141                 mempool_free(req, req->r_osdc->req_mempool);
 142         else
 143                 kfree(req);
 144 }
 145 EXPORT_SYMBOL(ceph_osdc_release_request);
 146
 147 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 148                                                struct ceph_snap_context *snapc,
 149                                                unsigned int num_op,
 150                                                bool use_mempool,
 151                                                gfp_t gfp_flags)
 152 {
 153         struct ceph_osd_request *req;
 154         struct ceph_msg *msg;
 155         size_t msg_size = sizeof(struct ceph_osd_request_head);
 156
 157         msg_size += num_op*sizeof(struct ceph_osd_op);
 158
 159         if (use_mempool) {
 160                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 161                 memset(req, 0, sizeof(*req));
 162         } else {
 163                 req = kzalloc(sizeof(*req), gfp_flags);
 164         }
 165         if (req == NULL)
 166                 return NULL;
 167
 168         req->r_osdc = osdc;
 169         req->r_mempool = use_mempool;
 170
 171         kref_init(&req->r_kref);
 172         init_completion(&req->r_completion);
 173         init_completion(&req->r_safe_completion);
 174         RB_CLEAR_NODE(&req->r_node);
 175         INIT_LIST_HEAD(&req->r_unsafe_item);
 176         INIT_LIST_HEAD(&req->r_linger_item);
 177         INIT_LIST_HEAD(&req->r_linger_osd);
 178         INIT_LIST_HEAD(&req->r_req_lru_item);
 179         INIT_LIST_HEAD(&req->r_osd_item);
 180
 181         /* create reply message */
 182         if (use_mempool)
 183                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 184         else
 185                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 186                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 187         if (!msg) {
 188                 ceph_osdc_put_request(req);
 189                 return NULL;
 190         }
 191         req->r_reply = msg;
 192
 193         ceph_pagelist_init(&req->r_trail);
 194
 195         /* create request message; allow space for oid */
 196         msg_size += MAX_OBJ_NAME_SIZE;
 197         if (snapc)
 198                 msg_size += sizeof(u64) * snapc->num_snaps;
 199         if (use_mempool)
 200                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 201         else
 202                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 203         if (!msg) {
 204                 ceph_osdc_put_request(req);
 205                 return NULL;
 206         }
 207
 208         memset(msg->front.iov_base, 0, msg->front.iov_len);
 209
 210         req->r_request = msg;
 211
 212         return req;
 213 }
 214 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 215
 216 static void osd_req_encode_op(struct ceph_osd_request *req,
 217                               struct ceph_osd_op *dst,
 218                               struct ceph_osd_req_op *src)
 219 {
 220         dst->op = cpu_to_le16(src->op);
 221
 222         switch (src->op) {
 223         case CEPH_OSD_OP_READ:
 224         case CEPH_OSD_OP_WRITE:
 225                 dst->extent.offset =
 226                         cpu_to_le64(src->extent.offset);
 227                 dst->extent.length =
 228                         cpu_to_le64(src->extent.length);
 229                 dst->extent.truncate_size =
 230                         cpu_to_le64(src->extent.truncate_size);
 231                 dst->extent.truncate_seq =
 232                         cpu_to_le32(src->extent.truncate_seq);
 233                 break;
 234         case CEPH_OSD_OP_CALL:
 235                 dst->cls.class_len = src->cls.class_len;
 236                 dst->cls.method_len = src->cls.method_len;
 237                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 238
 239                 ceph_pagelist_append(&req->r_trail, src->cls.class_name,
 240                                      src->cls.class_len);
 241                 ceph_pagelist_append(&req->r_trail, src->cls.method_name,
 242                                      src->cls.method_len);
 243                 ceph_pagelist_append(&req->r_trail, src->cls.indata,
 244                                      src->cls.indata_len);
 245                 break;
 246         case CEPH_OSD_OP_STARTSYNC:
 247                 break;
 248         case CEPH_OSD_OP_NOTIFY_ACK:
 249         case CEPH_OSD_OP_WATCH:
 250                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 251                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 252                 dst->watch.flag = src->watch.flag;
 253                 break;
 254         default:
 255                 pr_err("unrecognized osd opcode %d\n", dst->op);
 256                 WARN_ON(1);
 257                 break;
 258         case CEPH_OSD_OP_STAT:
 259         case CEPH_OSD_OP_MAPEXT:
 260         case CEPH_OSD_OP_MASKTRUNC:
 261         case CEPH_OSD_OP_SPARSE_READ:
 262         case CEPH_OSD_OP_NOTIFY:
 263         case CEPH_OSD_OP_ASSERT_VER:
 264         case CEPH_OSD_OP_WRITEFULL:
 265         case CEPH_OSD_OP_TRUNCATE:
 266         case CEPH_OSD_OP_ZERO:
 267         case CEPH_OSD_OP_DELETE:
 268         case CEPH_OSD_OP_APPEND:
 269         case CEPH_OSD_OP_SETTRUNC:
 270         case CEPH_OSD_OP_TRIMTRUNC:
 271         case CEPH_OSD_OP_TMAPUP:
 272         case CEPH_OSD_OP_TMAPPUT:
 273         case CEPH_OSD_OP_TMAPGET:
 274         case CEPH_OSD_OP_CREATE:
 275         case CEPH_OSD_OP_ROLLBACK:
 276         case CEPH_OSD_OP_OMAPGETKEYS:
 277         case CEPH_OSD_OP_OMAPGETVALS:
 278         case CEPH_OSD_OP_OMAPGETHEADER:
 279         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
 280         case CEPH_OSD_OP_MODE_RD:
 281         case CEPH_OSD_OP_OMAPSETVALS:
 282         case CEPH_OSD_OP_OMAPSETHEADER:
 283         case CEPH_OSD_OP_OMAPCLEAR:
 284         case CEPH_OSD_OP_OMAPRMKEYS:
 285         case CEPH_OSD_OP_OMAP_CMP:
 286         case CEPH_OSD_OP_CLONERANGE:
 287         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
 288         case CEPH_OSD_OP_SRC_CMPXATTR:
 289         case CEPH_OSD_OP_GETXATTR:
 290         case CEPH_OSD_OP_GETXATTRS:
 291         case CEPH_OSD_OP_CMPXATTR:
 292         case CEPH_OSD_OP_SETXATTR:
 293         case CEPH_OSD_OP_SETXATTRS:
 294         case CEPH_OSD_OP_RESETXATTRS:
 295         case CEPH_OSD_OP_RMXATTR:
 296         case CEPH_OSD_OP_PULL:
 297         case CEPH_OSD_OP_PUSH:
 298         case CEPH_OSD_OP_BALANCEREADS:
 299         case CEPH_OSD_OP_UNBALANCEREADS:
 300         case CEPH_OSD_OP_SCRUB:
 301         case CEPH_OSD_OP_SCRUB_RESERVE:
 302         case CEPH_OSD_OP_SCRUB_UNRESERVE:
 303         case CEPH_OSD_OP_SCRUB_STOP:
 304         case CEPH_OSD_OP_SCRUB_MAP:
 305         case CEPH_OSD_OP_WRLOCK:
 306         case CEPH_OSD_OP_WRUNLOCK:
 307         case CEPH_OSD_OP_RDLOCK:
 308         case CEPH_OSD_OP_RDUNLOCK:
 309         case CEPH_OSD_OP_UPLOCK:
 310         case CEPH_OSD_OP_DNLOCK:
 311         case CEPH_OSD_OP_PGLS:
 312         case CEPH_OSD_OP_PGLS_FILTER:
 313                 pr_err("unsupported osd opcode %s\n",
 314                         ceph_osd_op_name(dst->op));
 315                 WARN_ON(1);
 316                 break;
 317         }
 318         dst->payload_len = cpu_to_le32(src->payload_len);
 319 }
 320
 321 /*
 322  * build new request AND message
 323  *
 324  */
 325 void ceph_osdc_build_request(struct ceph_osd_request *req,
 326                              u64 off, u64 len, unsigned int num_op,
 327                              struct ceph_osd_req_op *src_ops,
 328                              struct ceph_snap_context *snapc, u64 snap_id,
 329                              struct timespec *mtime)
 330 {
 331         struct ceph_msg *msg = req->r_request;
 332         struct ceph_osd_request_head *head;
 333         struct ceph_osd_req_op *src_op;
 334         struct ceph_osd_op *op;
 335         void *p;
 336         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 337         int flags = req->r_flags;
 338         u64 data_len = 0;
 339         int i;
 340
 341         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 342
 343         head = msg->front.iov_base;
 344         head->snapid = cpu_to_le64(snap_id);
 345         op = (void *)(head + 1);
 346         p = (void *)(op + num_op);
 347
 348         req->r_snapc = ceph_get_snap_context(snapc);
 349
 350         head->client_inc = cpu_to_le32(1); /* always, for now. */
 351         head->flags = cpu_to_le32(flags);
 352         if (flags & CEPH_OSD_FLAG_WRITE)
 353                 ceph_encode_timespec(&head->mtime, mtime);
 354         BUG_ON(num_op > (unsigned int) ((u16) -1));
 355         head->num_ops = cpu_to_le16(num_op);
 356
 357         /* fill in oid */
 358         head->object_len = cpu_to_le32(req->r_oid_len);
 359         memcpy(p, req->r_oid, req->r_oid_len);
 360         p += req->r_oid_len;
 361
 362         src_op = src_ops;
 363         while (num_op--)
 364                 osd_req_encode_op(req, op++, src_op++);
 365
 366         data_len += req->r_trail.length;
 367
 368         if (snapc) {
 369                 head->snap_seq = cpu_to_le64(snapc->seq);
 370                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 371                 for (i = 0; i < snapc->num_snaps; i++) {
 372                         put_unaligned_le64(snapc->snaps[i], p);
 373                         p += sizeof(u64);
 374                 }
 375         }
 376
 377         if (flags & CEPH_OSD_FLAG_WRITE) {
 378                 req->r_request->hdr.data_off = cpu_to_le16(off);
 379                 req->r_request->hdr.data_len = cpu_to_le32(len + data_len);
 380         } else if (data_len) {
 381                 req->r_request->hdr.data_off = 0;
 382                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 383         }
 384
 385         req->r_request->page_alignment = req->r_page_alignment;
 386
 387         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 388         msg_size = p - msg->front.iov_base;
 389         msg->front.iov_len = msg_size;
 390         msg->hdr.front_len = cpu_to_le32(msg_size);
 391         return;
 392 }
 393 EXPORT_SYMBOL(ceph_osdc_build_request);
 394
 395 /*
 396  * build new request AND message, calculate layout, and adjust file
 397  * extent as needed.
 398  *
 399  * if the file was recently truncated, we include information about its
 400  * old and new size so that the object can be updated appropriately.  (we
 401  * avoid synchronously deleting truncated objects because it's slow.)
 402  *
 403  * if @do_sync, include a 'startsync' command so that the osd will flush
 404  * data quickly.
 405  */
 406 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 407                                                struct ceph_file_layout *layout,
 408                                                struct ceph_vino vino,
 409                                                u64 off, u64 *plen,
 410                                                int opcode, int flags,
 411                                                struct ceph_snap_context *snapc,
 412                                                int do_sync,
 413                                                u32 truncate_seq,
 414                                                u64 truncate_size,
 415                                                struct timespec *mtime,
 416                                                bool use_mempool,
 417                                                int page_align)
 418 {
 419         struct ceph_osd_req_op ops[2];
 420         struct ceph_osd_request *req;
 421         unsigned int num_op = 1;
 422         int r;
 423
 424         memset(&ops, 0, sizeof ops);
 425
 426         ops[0].op = opcode;
 427         ops[0].extent.truncate_seq = truncate_seq;
 428         ops[0].extent.truncate_size = truncate_size;
 429
 430         if (do_sync) {
 431                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 432                 num_op++;
 433         }
 434
 435         req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
 436                                         GFP_NOFS);
 437         if (!req)
 438                 return ERR_PTR(-ENOMEM);
 439         req->r_flags = flags;
 440
 441         /* calculate max write size */
 442         r = calc_layout(vino, layout, off, plen, req, ops);
 443         if (r < 0)
 444                 return ERR_PTR(r);
 445         req->r_file_layout = *layout;  /* keep a copy */
 446
 447         /* in case it differs from natural (file) alignment that
 448            calc_layout filled in for us */
 449         req->r_num_pages = calc_pages_for(page_align, *plen);
 450         req->r_page_alignment = page_align;
 451
 452         ceph_osdc_build_request(req, off, *plen, num_op, ops,
 453                                 snapc, vino.snap, mtime);
 454
 455         return req;
 456 }
 457 EXPORT_SYMBOL(ceph_osdc_new_request);
 458
 459 /*
 460  * We keep osd requests in an rbtree, sorted by ->r_tid.
 461  */
 462 static void __insert_request(struct ceph_osd_client *osdc,
 463                              struct ceph_osd_request *new)
 464 {
 465         struct rb_node **p = &osdc->requests.rb_node;
 466         struct rb_node *parent = NULL;
 467         struct ceph_osd_request *req = NULL;
 468
 469         while (*p) {
 470                 parent = *p;
 471                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 472                 if (new->r_tid < req->r_tid)
 473                         p = &(*p)->rb_left;
 474                 else if (new->r_tid > req->r_tid)
 475                         p = &(*p)->rb_right;
 476                 else
 477                         BUG();
 478         }
 479
 480         rb_link_node(&new->r_node, parent, p);
 481         rb_insert_color(&new->r_node, &osdc->requests);
 482 }
 483
 484 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 485                                                  u64 tid)
 486 {
 487         struct ceph_osd_request *req;
 488         struct rb_node *n = osdc->requests.rb_node;
 489
 490         while (n) {
 491                 req = rb_entry(n, struct ceph_osd_request, r_node);
 492                 if (tid < req->r_tid)
 493                         n = n->rb_left;
 494                 else if (tid > req->r_tid)
 495                         n = n->rb_right;
 496                 else
 497                         return req;
 498         }
 499         return NULL;
 500 }
 501
 502 static struct ceph_osd_request *
 503 __lookup_request_ge(struct ceph_osd_client *osdc,
 504                     u64 tid)
 505 {
 506         struct ceph_osd_request *req;
 507         struct rb_node *n = osdc->requests.rb_node;
 508
 509         while (n) {
 510                 req = rb_entry(n, struct ceph_osd_request, r_node);
 511                 if (tid < req->r_tid) {
 512                         if (!n->rb_left)
 513                                 return req;
 514                         n = n->rb_left;
 515                 } else if (tid > req->r_tid) {
 516                         n = n->rb_right;
 517                 } else {
 518                         return req;
 519                 }
 520         }
 521         return NULL;
 522 }
 523
 524 /*
 525  * Resubmit requests pending on the given osd.
 526  */
 527 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 528                                 struct ceph_osd *osd)
 529 {
 530         struct ceph_osd_request *req, *nreq;
 531         int err;
 532
 533         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 534         err = __reset_osd(osdc, osd);
 535         if (err)
 536                 return;
 537
 538         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 539                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 540                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 541                      osd->o_osd);
 542                 if (!req->r_linger)
 543                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 544         }
 545
 546         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 547                                  r_linger_osd) {
 548                 /*
 549                  * reregister request prior to unregistering linger so
 550                  * that r_osd is preserved.
 551                  */
 552                 BUG_ON(!list_empty(&req->r_req_lru_item));
 553                 __register_request(osdc, req);
 554                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 555                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 556                 __unregister_linger_request(osdc, req);
 557                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 558                      osd->o_osd);
 559         }
 560 }
 561
 562 /*
 563  * If the osd connection drops, we need to resubmit all requests.
 564  */
 565 static void osd_reset(struct ceph_connection *con)
 566 {
 567         struct ceph_osd *osd = con->private;
 568         struct ceph_osd_client *osdc;
 569
 570         if (!osd)
 571                 return;
 572         dout("osd_reset osd%d\n", osd->o_osd);
 573         osdc = osd->o_osdc;
 574         down_read(&osdc->map_sem);
 575         mutex_lock(&osdc->request_mutex);
 576         __kick_osd_requests(osdc, osd);
 577         __send_queued(osdc);
 578         mutex_unlock(&osdc->request_mutex);
 579         up_read(&osdc->map_sem);
 580 }
 581
 582 /*
 583  * Track open sessions with osds.
 584  */
 585 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 586 {
 587         struct ceph_osd *osd;
 588
 589         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 590         if (!osd)
 591                 return NULL;
 592
 593         atomic_set(&osd->o_ref, 1);
 594         osd->o_osdc = osdc;
 595         osd->o_osd = onum;
 596         RB_CLEAR_NODE(&osd->o_node);
 597         INIT_LIST_HEAD(&osd->o_requests);
 598         INIT_LIST_HEAD(&osd->o_linger_requests);
 599         INIT_LIST_HEAD(&osd->o_osd_lru);
 600         osd->o_incarnation = 1;
 601
 602         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 603
 604         INIT_LIST_HEAD(&osd->o_keepalive_item);
 605         return osd;
 606 }
 607
 608 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 609 {
 610         if (atomic_inc_not_zero(&osd->o_ref)) {
 611                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 612                      atomic_read(&osd->o_ref));
 613                 return osd;
 614         } else {
 615                 dout("get_osd %p FAIL\n", osd);
 616                 return NULL;
 617         }
 618 }
 619
 620 static void put_osd(struct ceph_osd *osd)
 621 {
 622         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 623              atomic_read(&osd->o_ref) - 1);
 624         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 625                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 626
 627                 if (ac->ops && ac->ops->destroy_authorizer)
 628                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 629                 kfree(osd);
 630         }
 631 }
 632
 633 /*
 634  * remove an osd from our map
 635  */
 636 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 637 {
 638         dout("__remove_osd %p\n", osd);
 639         BUG_ON(!list_empty(&osd->o_requests));
 640         rb_erase(&osd->o_node, &osdc->osds);
 641         list_del_init(&osd->o_osd_lru);
 642         ceph_con_close(&osd->o_con);
 643         put_osd(osd);
 644 }
 645
 646 static void remove_all_osds(struct ceph_osd_client *osdc)
 647 {
 648         dout("%s %p\n", __func__, osdc);
 649         mutex_lock(&osdc->request_mutex);
 650         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 651                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 652                                                 struct ceph_osd, o_node);
 653                 __remove_osd(osdc, osd);
 654         }
 655         mutex_unlock(&osdc->request_mutex);
 656 }
 657
 658 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 659                               struct ceph_osd *osd)
 660 {
 661         dout("__move_osd_to_lru %p\n", osd);
 662         BUG_ON(!list_empty(&osd->o_osd_lru));
 663         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 664         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 665 }
 666
 667 static void __remove_osd_from_lru(struct ceph_osd *osd)
 668 {
 669         dout("__remove_osd_from_lru %p\n", osd);
 670         if (!list_empty(&osd->o_osd_lru))
 671                 list_del_init(&osd->o_osd_lru);
 672 }
 673
 674 static void remove_old_osds(struct ceph_osd_client *osdc)
 675 {
 676         struct ceph_osd *osd, *nosd;
 677
 678         dout("__remove_old_osds %p\n", osdc);
 679         mutex_lock(&osdc->request_mutex);
 680         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 681                 if (time_before(jiffies, osd->lru_ttl))
 682                         break;
 683                 __remove_osd(osdc, osd);
 684         }
 685         mutex_unlock(&osdc->request_mutex);
 686 }
 687
 688 /*
 689  * reset osd connect
 690  */
 691 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 692 {
 693         struct ceph_entity_addr *peer_addr;
 694
 695         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 696         if (list_empty(&osd->o_requests) &&
 697             list_empty(&osd->o_linger_requests)) {
 698                 __remove_osd(osdc, osd);
 699
 700                 return -ENODEV;
 701         }
 702
 703         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
 704         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
 705                         !ceph_con_opened(&osd->o_con)) {
 706                 struct ceph_osd_request *req;
 707
 708                 dout(" osd addr hasn't changed and connection never opened,"
 709                      " letting msgr retry");
 710                 /* touch each r_stamp for handle_timeout()'s benfit */
 711                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 712                         req->r_stamp = jiffies;
 713
 714                 return -EAGAIN;
 715         }
 716
 717         ceph_con_close(&osd->o_con);
 718         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
 719         osd->o_incarnation++;
 720
 721         return 0;
 722 }
 723
 724 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 725 {
 726         struct rb_node **p = &osdc->osds.rb_node;
 727         struct rb_node *parent = NULL;
 728         struct ceph_osd *osd = NULL;
 729
 730         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 731         while (*p) {
 732                 parent = *p;
 733                 osd = rb_entry(parent, struct ceph_osd, o_node);
 734                 if (new->o_osd < osd->o_osd)
 735                         p = &(*p)->rb_left;
 736                 else if (new->o_osd > osd->o_osd)
 737                         p = &(*p)->rb_right;
 738                 else
 739                         BUG();
 740         }
 741
 742         rb_link_node(&new->o_node, parent, p);
 743         rb_insert_color(&new->o_node, &osdc->osds);
 744 }
 745
 746 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 747 {
 748         struct ceph_osd *osd;
 749         struct rb_node *n = osdc->osds.rb_node;
 750
 751         while (n) {
 752                 osd = rb_entry(n, struct ceph_osd, o_node);
 753                 if (o < osd->o_osd)
 754                         n = n->rb_left;
 755                 else if (o > osd->o_osd)
 756                         n = n->rb_right;
 757                 else
 758                         return osd;
 759         }
 760         return NULL;
 761 }
 762
 763 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 764 {
 765         schedule_delayed_work(&osdc->timeout_work,
 766                         osdc->client->options->osd_keepalive_timeout * HZ);
 767 }
 768
 769 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 770 {
 771         cancel_delayed_work(&osdc->timeout_work);
 772 }
 773
 774 /*
 775  * Register request, assign tid.  If this is the first request, set up
 776  * the timeout event.
 777  */
 778 static void __register_request(struct ceph_osd_client *osdc,
 779                                struct ceph_osd_request *req)
 780 {
 781         req->r_tid = ++osdc->last_tid;
 782         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 783         dout("__register_request %p tid %lld\n", req, req->r_tid);
 784         __insert_request(osdc, req);
 785         ceph_osdc_get_request(req);
 786         osdc->num_requests++;
 787         if (osdc->num_requests == 1) {
 788                 dout(" first request, scheduling timeout\n");
 789                 __schedule_osd_timeout(osdc);
 790         }
 791 }
 792
 793 static void register_request(struct ceph_osd_client *osdc,
 794                              struct ceph_osd_request *req)
 795 {
 796         mutex_lock(&osdc->request_mutex);
 797         __register_request(osdc, req);
 798         mutex_unlock(&osdc->request_mutex);
 799 }
 800
 801 /*
 802  * called under osdc->request_mutex
 803  */
 804 static void __unregister_request(struct ceph_osd_client *osdc,
 805                                  struct ceph_osd_request *req)
 806 {
 807         if (RB_EMPTY_NODE(&req->r_node)) {
 808                 dout("__unregister_request %p tid %lld not registered\n",
 809                         req, req->r_tid);
 810                 return;
 811         }
 812
 813         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 814         rb_erase(&req->r_node, &osdc->requests);
 815         osdc->num_requests--;
 816
 817         if (req->r_osd) {
 818                 /* make sure the original request isn't in flight. */
 819                 ceph_msg_revoke(req->r_request);
 820
 821                 list_del_init(&req->r_osd_item);
 822                 if (list_empty(&req->r_osd->o_requests) &&
 823                     list_empty(&req->r_osd->o_linger_requests)) {
 824                         dout("moving osd to %p lru\n", req->r_osd);
 825                         __move_osd_to_lru(osdc, req->r_osd);
 826                 }
 827                 if (list_empty(&req->r_linger_item))
 828                         req->r_osd = NULL;
 829         }
 830
 831         list_del_init(&req->r_req_lru_item);
 832         ceph_osdc_put_request(req);
 833
 834         if (osdc->num_requests == 0) {
 835                 dout(" no requests, canceling timeout\n");
 836                 __cancel_osd_timeout(osdc);
 837         }
 838 }
 839
 840 /*
 841  * Cancel a previously queued request message
 842  */
 843 static void __cancel_request(struct ceph_osd_request *req)
 844 {
 845         if (req->r_sent && req->r_osd) {
 846                 ceph_msg_revoke(req->r_request);
 847                 req->r_sent = 0;
 848         }
 849 }
 850
 851 static void __register_linger_request(struct ceph_osd_client *osdc,
 852                                     struct ceph_osd_request *req)
 853 {
 854         dout("__register_linger_request %p\n", req);
 855         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 856         if (req->r_osd)
 857                 list_add_tail(&req->r_linger_osd,
 858                               &req->r_osd->o_linger_requests);
 859 }
 860
 861 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 862                                         struct ceph_osd_request *req)
 863 {
 864         dout("__unregister_linger_request %p\n", req);
 865         list_del_init(&req->r_linger_item);
 866         if (req->r_osd) {
 867                 list_del_init(&req->r_linger_osd);
 868
 869                 if (list_empty(&req->r_osd->o_requests) &&
 870                     list_empty(&req->r_osd->o_linger_requests)) {
 871                         dout("moving osd to %p lru\n", req->r_osd);
 872                         __move_osd_to_lru(osdc, req->r_osd);
 873                 }
 874                 if (list_empty(&req->r_osd_item))
 875                         req->r_osd = NULL;
 876         }
 877 }
 878
 879 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 880                                          struct ceph_osd_request *req)
 881 {
 882         mutex_lock(&osdc->request_mutex);
 883         if (req->r_linger) {
 884                 __unregister_linger_request(osdc, req);
 885                 ceph_osdc_put_request(req);
 886         }
 887         mutex_unlock(&osdc->request_mutex);
 888 }
 889 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 890
 891 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 892                                   struct ceph_osd_request *req)
 893 {
 894         if (!req->r_linger) {
 895                 dout("set_request_linger %p\n", req);
 896                 req->r_linger = 1;
 897                 /*
 898                  * caller is now responsible for calling
 899                  * unregister_linger_request
 900                  */
 901                 ceph_osdc_get_request(req);
 902         }
 903 }
 904 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 905
 906 /*
 907  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 908  * (as needed), and set the request r_osd appropriately.  If there is
 909  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 910  * (unsent, homeless) or leave on in-flight lru.
 911  *
 912  * Return 0 if unchanged, 1 if changed, or negative on error.
 913  *
 914  * Caller should hold map_sem for read and request_mutex.
 915  */
 916 static int __map_request(struct ceph_osd_client *osdc,
 917                          struct ceph_osd_request *req, int force_resend)
 918 {
 919         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 920         struct ceph_pg pgid;
 921         int acting[CEPH_PG_MAX_SIZE];
 922         int o = -1, num = 0;
 923         int err;
 924
 925         dout("map_request %p tid %lld\n", req, req->r_tid);
 926         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 927                                       &req->r_file_layout, osdc->osdmap);
 928         if (err) {
 929                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 930                 return err;
 931         }
 932         pgid = reqhead->layout.ol_pgid;
 933         req->r_pgid = pgid;
 934
 935         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 936         if (err > 0) {
 937                 o = acting[0];
 938                 num = err;
 939         }
 940
 941         if ((!force_resend &&
 942              req->r_osd && req->r_osd->o_osd == o &&
 943              req->r_sent >= req->r_osd->o_incarnation &&
 944              req->r_num_pg_osds == num &&
 945              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 946             (req->r_osd == NULL && o == -1))
 947                 return 0;  /* no change */
 948
 949         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
 950              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 951              req->r_osd ? req->r_osd->o_osd : -1);
 952
 953         /* record full pg acting set */
 954         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 955         req->r_num_pg_osds = num;
 956
 957         if (req->r_osd) {
 958                 __cancel_request(req);
 959                 list_del_init(&req->r_osd_item);
 960                 req->r_osd = NULL;
 961         }
 962
 963         req->r_osd = __lookup_osd(osdc, o);
 964         if (!req->r_osd && o >= 0) {
 965                 err = -ENOMEM;
 966                 req->r_osd = create_osd(osdc, o);
 967                 if (!req->r_osd) {
 968                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
 969                         goto out;
 970                 }
 971
 972                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
 973                 __insert_osd(osdc, req->r_osd);
 974
 975                 ceph_con_open(&req->r_osd->o_con,
 976                               CEPH_ENTITY_TYPE_OSD, o,
 977                               &osdc->osdmap->osd_addr[o]);
 978         }
 979
 980         if (req->r_osd) {
 981                 __remove_osd_from_lru(req->r_osd);
 982                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 983                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 984         } else {
 985                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 986         }
 987         err = 1;   /* osd or pg changed */
 988
 989 out:
 990         return err;
 991 }
 992
 993 /*
 994  * caller should hold map_sem (for read) and request_mutex
 995  */
 996 static void __send_request(struct ceph_osd_client *osdc,
 997                            struct ceph_osd_request *req)
 998 {
 999         struct ceph_osd_request_head *reqhead;
1000
1001         dout("send_request %p tid %llu to osd%d flags %d\n",
1002              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1003
1004         reqhead = req->r_request->front.iov_base;
1005         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1006         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1007         reqhead->reassert_version = req->r_reassert_version;
1008
1009         req->r_stamp = jiffies;
1010         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1011
1012         ceph_msg_get(req->r_request); /* send consumes a ref */
1013         ceph_con_send(&req->r_osd->o_con, req->r_request);
1014         req->r_sent = req->r_osd->o_incarnation;
1015 }
1016
1017 /*
1018  * Send any requests in the queue (req_unsent).
1019  */
1020 static void __send_queued(struct ceph_osd_client *osdc)
1021 {
1022         struct ceph_osd_request *req, *tmp;
1023
1024         dout("__send_queued\n");
1025         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1026                 __send_request(osdc, req);
1027 }
1028
1029 /*
1030  * Timeout callback, called every N seconds when 1 or more osd
1031  * requests has been active for more than N seconds.  When this
1032  * happens, we ping all OSDs with requests who have timed out to
1033  * ensure any communications channel reset is detected.  Reset the
1034  * request timeouts another N seconds in the future as we go.
1035  * Reschedule the timeout event another N seconds in future (unless
1036  * there are no open requests).
1037  */
1038 static void handle_timeout(struct work_struct *work)
1039 {
1040         struct ceph_osd_client *osdc =
1041                 container_of(work, struct ceph_osd_client, timeout_work.work);
1042         struct ceph_osd_request *req;
1043         struct ceph_osd *osd;
1044         unsigned long keepalive =
1045                 osdc->client->options->osd_keepalive_timeout * HZ;
1046         struct list_head slow_osds;
1047         dout("timeout\n");
1048         down_read(&osdc->map_sem);
1049
1050         ceph_monc_request_next_osdmap(&osdc->client->monc);
1051
1052         mutex_lock(&osdc->request_mutex);
1053
1054         /*
1055          * ping osds that are a bit slow.  this ensures that if there
1056          * is a break in the TCP connection we will notice, and reopen
1057          * a connection with that osd (from the fault callback).
1058          */
1059         INIT_LIST_HEAD(&slow_osds);
1060         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1061                 if (time_before(jiffies, req->r_stamp + keepalive))
1062                         break;
1063
1064                 osd = req->r_osd;
1065                 BUG_ON(!osd);
1066                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1067                      req->r_tid, osd->o_osd);
1068                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1069         }
1070         while (!list_empty(&slow_osds)) {
1071                 osd = list_entry(slow_osds.next, struct ceph_osd,
1072                                  o_keepalive_item);
1073                 list_del_init(&osd->o_keepalive_item);
1074                 ceph_con_keepalive(&osd->o_con);
1075         }
1076
1077         __schedule_osd_timeout(osdc);
1078         __send_queued(osdc);
1079         mutex_unlock(&osdc->request_mutex);
1080         up_read(&osdc->map_sem);
1081 }
1082
1083 static void handle_osds_timeout(struct work_struct *work)
1084 {
1085         struct ceph_osd_client *osdc =
1086                 container_of(work, struct ceph_osd_client,
1087                              osds_timeout_work.work);
1088         unsigned long delay =
1089                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1090
1091         dout("osds timeout\n");
1092         down_read(&osdc->map_sem);
1093         remove_old_osds(osdc);
1094         up_read(&osdc->map_sem);
1095
1096         schedule_delayed_work(&osdc->osds_timeout_work,
1097                               round_jiffies_relative(delay));
1098 }
1099
1100 static void complete_request(struct ceph_osd_request *req)
1101 {
1102         if (req->r_safe_callback)
1103                 req->r_safe_callback(req, NULL);
1104         complete_all(&req->r_safe_completion);  /* fsync waiter */
1105 }
1106
1107 /*
1108  * handle osd op reply.  either call the callback if it is specified,
1109  * or do the completion to wake up the waiting thread.
1110  */
1111 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1112                          struct ceph_connection *con)
1113 {
1114         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1115         struct ceph_osd_request *req;
1116         u64 tid;
1117         int numops, object_len, flags;
1118         s32 result;
1119
1120         tid = le64_to_cpu(msg->hdr.tid);
1121         if (msg->front.iov_len < sizeof(*rhead))
1122                 goto bad;
1123         numops = le32_to_cpu(rhead->num_ops);
1124         object_len = le32_to_cpu(rhead->object_len);
1125         result = le32_to_cpu(rhead->result);
1126         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1127             numops * sizeof(struct ceph_osd_op))
1128                 goto bad;
1129         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1130         /* lookup */
1131         mutex_lock(&osdc->request_mutex);
1132         req = __lookup_request(osdc, tid);
1133         if (req == NULL) {
1134                 dout("handle_reply tid %llu dne\n", tid);
1135                 mutex_unlock(&osdc->request_mutex);
1136                 return;
1137         }
1138         ceph_osdc_get_request(req);
1139         flags = le32_to_cpu(rhead->flags);
1140
1141         /*
1142          * if this connection filled our message, drop our reference now, to
1143          * avoid a (safe but slower) revoke later.
1144          */
1145         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1146                 dout(" dropping con_filling_msg ref %p\n", con);
1147                 req->r_con_filling_msg = NULL;
1148                 con->ops->put(con);
1149         }
1150
1151         if (!req->r_got_reply) {
1152                 unsigned int bytes;
1153
1154                 req->r_result = le32_to_cpu(rhead->result);
1155                 bytes = le32_to_cpu(msg->hdr.data_len);
1156                 dout("handle_reply result %d bytes %d\n", req->r_result,
1157                      bytes);
1158                 if (req->r_result == 0)
1159                         req->r_result = bytes;
1160
1161                 /* in case this is a write and we need to replay, */
1162                 req->r_reassert_version = rhead->reassert_version;
1163
1164                 req->r_got_reply = 1;
1165         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1166                 dout("handle_reply tid %llu dup ack\n", tid);
1167                 mutex_unlock(&osdc->request_mutex);
1168                 goto done;
1169         }
1170
1171         dout("handle_reply tid %llu flags %d\n", tid, flags);
1172
1173         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1174                 __register_linger_request(osdc, req);
1175
1176         /* either this is a read, or we got the safe response */
1177         if (result < 0 ||
1178             (flags & CEPH_OSD_FLAG_ONDISK) ||
1179             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1180                 __unregister_request(osdc, req);
1181
1182         mutex_unlock(&osdc->request_mutex);
1183
1184         if (req->r_callback)
1185                 req->r_callback(req, msg);
1186         else
1187                 complete_all(&req->r_completion);
1188
1189         if (flags & CEPH_OSD_FLAG_ONDISK)
1190                 complete_request(req);
1191
1192 done:
1193         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1194         ceph_osdc_put_request(req);
1195         return;
1196
1197 bad:
1198         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1199                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1200                (int)sizeof(*rhead));
1201         ceph_msg_dump(msg);
1202 }
1203
1204 static void reset_changed_osds(struct ceph_osd_client *osdc)
1205 {
1206         struct rb_node *p, *n;
1207
1208         for (p = rb_first(&osdc->osds); p; p = n) {
1209                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1210
1211                 n = rb_next(p);
1212                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1213                     memcmp(&osd->o_con.peer_addr,
1214                            ceph_osd_addr(osdc->osdmap,
1215                                          osd->o_osd),
1216                            sizeof(struct ceph_entity_addr)) != 0)
1217                         __reset_osd(osdc, osd);
1218         }
1219 }
1220
1221 /*
1222  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1223  * no osd, request a new map.
1224  *
1225  * Caller should hold map_sem for read.
1226  */
1227 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1228 {
1229         struct ceph_osd_request *req, *nreq;
1230         struct rb_node *p;
1231         int needmap = 0;
1232         int err;
1233
1234         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1235         mutex_lock(&osdc->request_mutex);
1236         for (p = rb_first(&osdc->requests); p; ) {
1237                 req = rb_entry(p, struct ceph_osd_request, r_node);
1238                 p = rb_next(p);
1239
1240                 /*
1241                  * For linger requests that have not yet been
1242                  * registered, move them to the linger list; they'll
1243                  * be sent to the osd in the loop below.  Unregister
1244                  * the request before re-registering it as a linger
1245                  * request to ensure the __map_request() below
1246                  * will decide it needs to be sent.
1247                  */
1248                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1249                         dout("%p tid %llu restart on osd%d\n",
1250                              req, req->r_tid,
1251                              req->r_osd ? req->r_osd->o_osd : -1);
1252                         __unregister_request(osdc, req);
1253                         __register_linger_request(osdc, req);
1254                         continue;
1255                 }
1256
1257                 err = __map_request(osdc, req, force_resend);
1258                 if (err < 0)
1259                         continue;  /* error */
1260                 if (req->r_osd == NULL) {
1261                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1262                         needmap++;  /* request a newer map */
1263                 } else if (err > 0) {
1264                         if (!req->r_linger) {
1265                                 dout("%p tid %llu requeued on osd%d\n", req,
1266                                      req->r_tid,
1267                                      req->r_osd ? req->r_osd->o_osd : -1);
1268                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1269                         }
1270                 }
1271         }
1272
1273         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1274                                  r_linger_item) {
1275                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1276
1277                 err = __map_request(osdc, req, force_resend);
1278                 dout("__map_request returned %d\n", err);
1279                 if (err == 0)
1280                         continue;  /* no change and no osd was specified */
1281                 if (err < 0)
1282                         continue;  /* hrm! */
1283                 if (req->r_osd == NULL) {
1284                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1285                         needmap++;  /* request a newer map */
1286                         continue;
1287                 }
1288
1289                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1290                      req->r_osd ? req->r_osd->o_osd : -1);
1291                 __register_request(osdc, req);
1292                 __unregister_linger_request(osdc, req);
1293         }
1294         mutex_unlock(&osdc->request_mutex);
1295
1296         if (needmap) {
1297                 dout("%d requests for down osds, need new map\n", needmap);
1298                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1299         }
1300         reset_changed_osds(osdc);
1301 }
1302
1303
1304 /*
1305  * Process updated osd map.
1306  *
1307  * The message contains any number of incremental and full maps, normally
1308  * indicating some sort of topology change in the cluster.  Kick requests
1309  * off to different OSDs as needed.
1310  */
1311 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1312 {
1313         void *p, *end, *next;
1314         u32 nr_maps, maplen;
1315         u32 epoch;
1316         struct ceph_osdmap *newmap = NULL, *oldmap;
1317         int err;
1318         struct ceph_fsid fsid;
1319
1320         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1321         p = msg->front.iov_base;
1322         end = p + msg->front.iov_len;
1323
1324         /* verify fsid */
1325         ceph_decode_need(&p, end, sizeof(fsid), bad);
1326         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1327         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1328                 return;
1329
1330         down_write(&osdc->map_sem);
1331
1332         /* incremental maps */
1333         ceph_decode_32_safe(&p, end, nr_maps, bad);
1334         dout(" %d inc maps\n", nr_maps);
1335         while (nr_maps > 0) {
1336                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1337                 epoch = ceph_decode_32(&p);
1338                 maplen = ceph_decode_32(&p);
1339                 ceph_decode_need(&p, end, maplen, bad);
1340                 next = p + maplen;
1341                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1342                         dout("applying incremental map %u len %d\n",
1343                              epoch, maplen);
1344                         newmap = osdmap_apply_incremental(&p, next,
1345                                                           osdc->osdmap,
1346                                                           &osdc->client->msgr);
1347                         if (IS_ERR(newmap)) {
1348                                 err = PTR_ERR(newmap);
1349                                 goto bad;
1350                         }
1351                         BUG_ON(!newmap);
1352                         if (newmap != osdc->osdmap) {
1353                                 ceph_osdmap_destroy(osdc->osdmap);
1354                                 osdc->osdmap = newmap;
1355                         }
1356                         kick_requests(osdc, 0);
1357                 } else {
1358                         dout("ignoring incremental map %u len %d\n",
1359                              epoch, maplen);
1360                 }
1361                 p = next;
1362                 nr_maps--;
1363         }
1364         if (newmap)
1365                 goto done;
1366
1367         /* full maps */
1368         ceph_decode_32_safe(&p, end, nr_maps, bad);
1369         dout(" %d full maps\n", nr_maps);
1370         while (nr_maps) {
1371                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1372                 epoch = ceph_decode_32(&p);
1373                 maplen = ceph_decode_32(&p);
1374                 ceph_decode_need(&p, end, maplen, bad);
1375                 if (nr_maps > 1) {
1376                         dout("skipping non-latest full map %u len %d\n",
1377                              epoch, maplen);
1378                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1379                         dout("skipping full map %u len %d, "
1380                              "older than our %u\n", epoch, maplen,
1381                              osdc->osdmap->epoch);
1382                 } else {
1383                         int skipped_map = 0;
1384
1385                         dout("taking full map %u len %d\n", epoch, maplen);
1386                         newmap = osdmap_decode(&p, p+maplen);
1387                         if (IS_ERR(newmap)) {
1388                                 err = PTR_ERR(newmap);
1389                                 goto bad;
1390                         }
1391                         BUG_ON(!newmap);
1392                         oldmap = osdc->osdmap;
1393                         osdc->osdmap = newmap;
1394                         if (oldmap) {
1395                                 if (oldmap->epoch + 1 < newmap->epoch)
1396                                         skipped_map = 1;
1397                                 ceph_osdmap_destroy(oldmap);
1398                         }
1399                         kick_requests(osdc, skipped_map);
1400                 }
1401                 p += maplen;
1402                 nr_maps--;
1403         }
1404
1405 done:
1406         downgrade_write(&osdc->map_sem);
1407         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1408
1409         /*
1410          * subscribe to subsequent osdmap updates if full to ensure
1411          * we find out when we are no longer full and stop returning
1412          * ENOSPC.
1413          */
1414         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1415                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1416
1417         mutex_lock(&osdc->request_mutex);
1418         __send_queued(osdc);
1419         mutex_unlock(&osdc->request_mutex);
1420         up_read(&osdc->map_sem);
1421         wake_up_all(&osdc->client->auth_wq);
1422         return;
1423
1424 bad:
1425         pr_err("osdc handle_map corrupt msg\n");
1426         ceph_msg_dump(msg);
1427         up_write(&osdc->map_sem);
1428         return;
1429 }
1430
1431 /*
1432  * watch/notify callback event infrastructure
1433  *
1434  * These callbacks are used both for watch and notify operations.
1435  */
1436 static void __release_event(struct kref *kref)
1437 {
1438         struct ceph_osd_event *event =
1439                 container_of(kref, struct ceph_osd_event, kref);
1440
1441         dout("__release_event %p\n", event);
1442         kfree(event);
1443 }
1444
1445 static void get_event(struct ceph_osd_event *event)
1446 {
1447         kref_get(&event->kref);
1448 }
1449
1450 void ceph_osdc_put_event(struct ceph_osd_event *event)
1451 {
1452         kref_put(&event->kref, __release_event);
1453 }
1454 EXPORT_SYMBOL(ceph_osdc_put_event);
1455
1456 static void __insert_event(struct ceph_osd_client *osdc,
1457                              struct ceph_osd_event *new)
1458 {
1459         struct rb_node **p = &osdc->event_tree.rb_node;
1460         struct rb_node *parent = NULL;
1461         struct ceph_osd_event *event = NULL;
1462
1463         while (*p) {
1464                 parent = *p;
1465                 event = rb_entry(parent, struct ceph_osd_event, node);
1466                 if (new->cookie < event->cookie)
1467                         p = &(*p)->rb_left;
1468                 else if (new->cookie > event->cookie)
1469                         p = &(*p)->rb_right;
1470                 else
1471                         BUG();
1472         }
1473
1474         rb_link_node(&new->node, parent, p);
1475         rb_insert_color(&new->node, &osdc->event_tree);
1476 }
1477
1478 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1479                                                 u64 cookie)
1480 {
1481         struct rb_node **p = &osdc->event_tree.rb_node;
1482         struct rb_node *parent = NULL;
1483         struct ceph_osd_event *event = NULL;
1484
1485         while (*p) {
1486                 parent = *p;
1487                 event = rb_entry(parent, struct ceph_osd_event, node);
1488                 if (cookie < event->cookie)
1489                         p = &(*p)->rb_left;
1490                 else if (cookie > event->cookie)
1491                         p = &(*p)->rb_right;
1492                 else
1493                         return event;
1494         }
1495         return NULL;
1496 }
1497
1498 static void __remove_event(struct ceph_osd_event *event)
1499 {
1500         struct ceph_osd_client *osdc = event->osdc;
1501
1502         if (!RB_EMPTY_NODE(&event->node)) {
1503                 dout("__remove_event removed %p\n", event);
1504                 rb_erase(&event->node, &osdc->event_tree);
1505                 ceph_osdc_put_event(event);
1506         } else {
1507                 dout("__remove_event didn't remove %p\n", event);
1508         }
1509 }
1510
1511 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1512                            void (*event_cb)(u64, u64, u8, void *),
1513                            void *data, struct ceph_osd_event **pevent)
1514 {
1515         struct ceph_osd_event *event;
1516
1517         event = kmalloc(sizeof(*event), GFP_NOIO);
1518         if (!event)
1519                 return -ENOMEM;
1520
1521         dout("create_event %p\n", event);
1522         event->cb = event_cb;
1523         event->one_shot = 0;
1524         event->data = data;
1525         event->osdc = osdc;
1526         INIT_LIST_HEAD(&event->osd_node);
1527         RB_CLEAR_NODE(&event->node);
1528         kref_init(&event->kref);   /* one ref for us */
1529         kref_get(&event->kref);    /* one ref for the caller */
1530
1531         spin_lock(&osdc->event_lock);
1532         event->cookie = ++osdc->event_count;
1533         __insert_event(osdc, event);
1534         spin_unlock(&osdc->event_lock);
1535
1536         *pevent = event;
1537         return 0;
1538 }
1539 EXPORT_SYMBOL(ceph_osdc_create_event);
1540
1541 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1542 {
1543         struct ceph_osd_client *osdc = event->osdc;
1544
1545         dout("cancel_event %p\n", event);
1546         spin_lock(&osdc->event_lock);
1547         __remove_event(event);
1548         spin_unlock(&osdc->event_lock);
1549         ceph_osdc_put_event(event); /* caller's */
1550 }
1551 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1552
1553
1554 static void do_event_work(struct work_struct *work)
1555 {
1556         struct ceph_osd_event_work *event_work =
1557                 container_of(work, struct ceph_osd_event_work, work);
1558         struct ceph_osd_event *event = event_work->event;
1559         u64 ver = event_work->ver;
1560         u64 notify_id = event_work->notify_id;
1561         u8 opcode = event_work->opcode;
1562
1563         dout("do_event_work completing %p\n", event);
1564         event->cb(ver, notify_id, opcode, event->data);
1565         dout("do_event_work completed %p\n", event);
1566         ceph_osdc_put_event(event);
1567         kfree(event_work);
1568 }
1569
1570
1571 /*
1572  * Process osd watch notifications
1573  */
1574 static void handle_watch_notify(struct ceph_osd_client *osdc,
1575                                 struct ceph_msg *msg)
1576 {
1577         void *p, *end;
1578         u8 proto_ver;
1579         u64 cookie, ver, notify_id;
1580         u8 opcode;
1581         struct ceph_osd_event *event;
1582         struct ceph_osd_event_work *event_work;
1583
1584         p = msg->front.iov_base;
1585         end = p + msg->front.iov_len;
1586
1587         ceph_decode_8_safe(&p, end, proto_ver, bad);
1588         ceph_decode_8_safe(&p, end, opcode, bad);
1589         ceph_decode_64_safe(&p, end, cookie, bad);
1590         ceph_decode_64_safe(&p, end, ver, bad);
1591         ceph_decode_64_safe(&p, end, notify_id, bad);
1592
1593         spin_lock(&osdc->event_lock);
1594         event = __find_event(osdc, cookie);
1595         if (event) {
1596                 BUG_ON(event->one_shot);
1597                 get_event(event);
1598         }
1599         spin_unlock(&osdc->event_lock);
1600         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1601              cookie, ver, event);
1602         if (event) {
1603                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1604                 if (!event_work) {
1605                         dout("ERROR: could not allocate event_work\n");
1606                         goto done_err;
1607                 }
1608                 INIT_WORK(&event_work->work, do_event_work);
1609                 event_work->event = event;
1610                 event_work->ver = ver;
1611                 event_work->notify_id = notify_id;
1612                 event_work->opcode = opcode;
1613                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1614                         dout("WARNING: failed to queue notify event work\n");
1615                         goto done_err;
1616                 }
1617         }
1618
1619         return;
1620
1621 done_err:
1622         ceph_osdc_put_event(event);
1623         return;
1624
1625 bad:
1626         pr_err("osdc handle_watch_notify corrupt msg\n");
1627         return;
1628 }
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
1804 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1805 {
1806         flush_workqueue(osdc->notify_wq);
1807         destroy_workqueue(osdc->notify_wq);
1808         cancel_delayed_work_sync(&osdc->timeout_work);
1809         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1810         if (osdc->osdmap) {
1811                 ceph_osdmap_destroy(osdc->osdmap);
1812                 osdc->osdmap = NULL;
1813         }
1814         remove_all_osds(osdc);
1815         mempool_destroy(osdc->req_mempool);
1816         ceph_msgpool_destroy(&osdc->msgpool_op);
1817         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1818 }
1819
1820 /*
1821  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1822  * *plen.  Return number of bytes read, or error.
1823  */
1824 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1825                         struct ceph_vino vino, struct ceph_file_layout *layout,
1826                         u64 off, u64 *plen,
1827                         u32 truncate_seq, u64 truncate_size,
1828                         struct page **pages, int num_pages, int page_align)
1829 {
1830         struct ceph_osd_request *req;
1831         int rc = 0;
1832
1833         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1834              vino.snap, off, *plen);
1835         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1836                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1837                                     NULL, 0, truncate_seq, truncate_size, NULL,
1838                                     false, page_align);
1839         if (IS_ERR(req))
1840                 return PTR_ERR(req);
1841
1842         /* it may be a short read due to an object boundary */
1843         req->r_pages = pages;
1844
1845         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1846              off, *plen, req->r_num_pages, page_align);
1847
1848         rc = ceph_osdc_start_request(osdc, req, false);
1849         if (!rc)
1850                 rc = ceph_osdc_wait_request(osdc, req);
1851
1852         ceph_osdc_put_request(req);
1853         dout("readpages result %d\n", rc);
1854         return rc;
1855 }
1856 EXPORT_SYMBOL(ceph_osdc_readpages);
1857
1858 /*
1859  * do a synchronous write on N pages
1860  */
1861 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1862                          struct ceph_file_layout *layout,
1863                          struct ceph_snap_context *snapc,
1864                          u64 off, u64 len,
1865                          u32 truncate_seq, u64 truncate_size,
1866                          struct timespec *mtime,
1867                          struct page **pages, int num_pages)
1868 {
1869         struct ceph_osd_request *req;
1870         int rc = 0;
1871         int page_align = off & ~PAGE_MASK;
1872
1873         BUG_ON(vino.snap != CEPH_NOSNAP);
1874         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1875                                     CEPH_OSD_OP_WRITE,
1876                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1877                                     snapc, 0,
1878                                     truncate_seq, truncate_size, mtime,
1879                                     true, page_align);
1880         if (IS_ERR(req))
1881                 return PTR_ERR(req);
1882
1883         /* it may be a short write due to an object boundary */
1884         req->r_pages = pages;
1885         dout("writepages %llu~%llu (%d pages)\n", off, len,
1886              req->r_num_pages);
1887
1888         rc = ceph_osdc_start_request(osdc, req, true);
1889         if (!rc)
1890                 rc = ceph_osdc_wait_request(osdc, req);
1891
1892         ceph_osdc_put_request(req);
1893         if (rc == 0)
1894                 rc = len;
1895         dout("writepages result %d\n", rc);
1896         return rc;
1897 }
1898 EXPORT_SYMBOL(ceph_osdc_writepages);
1899
1900 /*
1901  * handle incoming message
1902  */
1903 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1904 {
1905         struct ceph_osd *osd = con->private;
1906         struct ceph_osd_client *osdc;
1907         int type = le16_to_cpu(msg->hdr.type);
1908
1909         if (!osd)
1910                 goto out;
1911         osdc = osd->o_osdc;
1912
1913         switch (type) {
1914         case CEPH_MSG_OSD_MAP:
1915                 ceph_osdc_handle_map(osdc, msg);
1916                 break;
1917         case CEPH_MSG_OSD_OPREPLY:
1918                 handle_reply(osdc, msg, con);
1919                 break;
1920         case CEPH_MSG_WATCH_NOTIFY:
1921                 handle_watch_notify(osdc, msg);
1922                 break;
1923
1924         default:
1925                 pr_err("received unknown message type %d %s\n", type,
1926                        ceph_msg_type_name(type));
1927         }
1928 out:
1929         ceph_msg_put(msg);
1930 }
1931
1932 /*
1933  * lookup and return message for incoming reply.  set up reply message
1934  * pages.
1935  */
1936 static struct ceph_msg *get_reply(struct ceph_connection *con,
1937                                   struct ceph_msg_header *hdr,
1938                                   int *skip)
1939 {
1940         struct ceph_osd *osd = con->private;
1941         struct ceph_osd_client *osdc = osd->o_osdc;
1942         struct ceph_msg *m;
1943         struct ceph_osd_request *req;
1944         int front = le32_to_cpu(hdr->front_len);
1945         int data_len = le32_to_cpu(hdr->data_len);
1946         u64 tid;
1947
1948         tid = le64_to_cpu(hdr->tid);
1949         mutex_lock(&osdc->request_mutex);
1950         req = __lookup_request(osdc, tid);
1951         if (!req) {
1952                 *skip = 1;
1953                 m = NULL;
1954                 dout("get_reply unknown tid %llu from osd%d\n", tid,
1955                      osd->o_osd);
1956                 goto out;
1957         }
1958
1959         if (req->r_con_filling_msg) {
1960                 dout("%s revoking msg %p from old con %p\n", __func__,
1961                      req->r_reply, req->r_con_filling_msg);
1962                 ceph_msg_revoke_incoming(req->r_reply);
1963                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
1964                 req->r_con_filling_msg = NULL;
1965         }
1966
1967         if (front > req->r_reply->front.iov_len) {
1968                 pr_warning("get_reply front %d > preallocated %d\n",
1969                            front, (int)req->r_reply->front.iov_len);
1970                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
1971                 if (!m)
1972                         goto out;
1973                 ceph_msg_put(req->r_reply);
1974                 req->r_reply = m;
1975         }
1976         m = ceph_msg_get(req->r_reply);
1977
1978         if (data_len > 0) {
1979                 int want = calc_pages_for(req->r_page_alignment, data_len);
1980
1981                 if (req->r_pages && unlikely(req->r_num_pages < want)) {
1982                         pr_warning("tid %lld reply has %d bytes %d pages, we"
1983                                    " had only %d pages ready\n", tid, data_len,
1984                                    want, req->r_num_pages);
1985                         *skip = 1;
1986                         ceph_msg_put(m);
1987                         m = NULL;
1988                         goto out;
1989                 }
1990                 m->pages = req->r_pages;
1991                 m->nr_pages = req->r_num_pages;
1992                 m->page_alignment = req->r_page_alignment;
1993 #ifdef CONFIG_BLOCK
1994                 m->bio = req->r_bio;
1995 #endif
1996         }
1997         *skip = 0;
1998         req->r_con_filling_msg = con->ops->get(con);
1999         dout("get_reply tid %lld %p\n", tid, m);
2000
2001 out:
2002         mutex_unlock(&osdc->request_mutex);
2003         return m;
2004
2005 }
2006
2007 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2008                                   struct ceph_msg_header *hdr,
2009                                   int *skip)
2010 {
2011         struct ceph_osd *osd = con->private;
2012         int type = le16_to_cpu(hdr->type);
2013         int front = le32_to_cpu(hdr->front_len);
2014
2015         *skip = 0;
2016         switch (type) {
2017         case CEPH_MSG_OSD_MAP:
2018         case CEPH_MSG_WATCH_NOTIFY:
2019                 return ceph_msg_new(type, front, GFP_NOFS, false);
2020         case CEPH_MSG_OSD_OPREPLY:
2021                 return get_reply(con, hdr, skip);
2022         default:
2023                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2024                         osd->o_osd);
2025                 *skip = 1;
2026                 return NULL;
2027         }
2028 }
2029
2030 /*
2031  * Wrappers to refcount containing ceph_osd struct
2032  */
2033 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2034 {
2035         struct ceph_osd *osd = con->private;
2036         if (get_osd(osd))
2037                 return con;
2038         return NULL;
2039 }
2040
2041 static void put_osd_con(struct ceph_connection *con)
2042 {
2043         struct ceph_osd *osd = con->private;
2044         put_osd(osd);
2045 }
2046
2047 /*
2048  * authentication
2049  */
2050 /*
2051  * Note: returned pointer is the address of a structure that's
2052  * managed separately.  Caller must *not* attempt to free it.
2053  */
2054 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2055                                         int *proto, int force_new)
2056 {
2057         struct ceph_osd *o = con->private;
2058         struct ceph_osd_client *osdc = o->o_osdc;
2059         struct ceph_auth_client *ac = osdc->client->monc.auth;
2060         struct ceph_auth_handshake *auth = &o->o_auth;
2061
2062         if (force_new && auth->authorizer) {
2063                 if (ac->ops && ac->ops->destroy_authorizer)
2064                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2065                 auth->authorizer = NULL;
2066         }
2067         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2068                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2069                                                         auth);
2070                 if (ret)
2071                         return ERR_PTR(ret);
2072         }
2073         *proto = ac->protocol;
2074
2075         return auth;
2076 }
2077
2078
2079 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2080 {
2081         struct ceph_osd *o = con->private;
2082         struct ceph_osd_client *osdc = o->o_osdc;
2083         struct ceph_auth_client *ac = osdc->client->monc.auth;
2084
2085         /*
2086          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2087          * XXX which do we do:  succeed or fail?
2088          */
2089         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2090 }
2091
2092 static int invalidate_authorizer(struct ceph_connection *con)
2093 {
2094         struct ceph_osd *o = con->private;
2095         struct ceph_osd_client *osdc = o->o_osdc;
2096         struct ceph_auth_client *ac = osdc->client->monc.auth;
2097
2098         if (ac->ops && ac->ops->invalidate_authorizer)
2099                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2100
2101         return ceph_monc_validate_auth(&osdc->client->monc);
2102 }
2103
2104 static const struct ceph_connection_operations osd_con_ops = {
2105         .get = get_osd_con,
2106         .put = put_osd_con,
2107         .dispatch = dispatch,
2108         .get_authorizer = get_authorizer,
2109         .verify_authorizer_reply = verify_authorizer_reply,
2110         .invalidate_authorizer = invalidate_authorizer,
2111         .alloc_msg = alloc_msg,
2112         .fault = osd_reset,
2113 };