net/ceph/osd_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/module.h>
   4 #include <linux/err.h>
   5 #include <linux/highmem.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/slab.h>
   9 #include <linux/uaccess.h>
  10 #ifdef CONFIG_BLOCK
  11 #include <linux/bio.h>
  12 #endif
  13
  14 #include <linux/ceph/libceph.h>
  15 #include <linux/ceph/osd_client.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/auth.h>
  19 #include <linux/ceph/pagelist.h>
  20
  21 #define OSD_OP_FRONT_LEN        4096
  22 #define OSD_OPREPLY_FRONT_LEN   512
  23
  24 static const struct ceph_connection_operations osd_con_ops;
  25
  26 static void send_queued(struct ceph_osd_client *osdc);
  27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  28 static void __register_request(struct ceph_osd_client *osdc,
  29                                struct ceph_osd_request *req);
  30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  31                                         struct ceph_osd_request *req);
  32 static void __send_request(struct ceph_osd_client *osdc,
  33                            struct ceph_osd_request *req);
  34
  35 static int op_needs_trail(int op)
  36 {
  37         switch (op) {
  38         case CEPH_OSD_OP_GETXATTR:
  39         case CEPH_OSD_OP_SETXATTR:
  40         case CEPH_OSD_OP_CMPXATTR:
  41         case CEPH_OSD_OP_CALL:
  42         case CEPH_OSD_OP_NOTIFY:
  43                 return 1;
  44         default:
  45                 return 0;
  46         }
  47 }
  48
  49 static int op_has_extent(int op)
  50 {
  51         return (op == CEPH_OSD_OP_READ ||
  52                 op == CEPH_OSD_OP_WRITE);
  53 }
  54
  55 int ceph_calc_raw_layout(struct ceph_osd_client *osdc,
  56                         struct ceph_file_layout *layout,
  57                         u64 snapid,
  58                         u64 off, u64 *plen, u64 *bno,
  59                         struct ceph_osd_request *req,
  60                         struct ceph_osd_req_op *op)
  61 {
  62         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
  63         u64 orig_len = *plen;
  64         u64 objoff, objlen;    /* extent in object */
  65         int r;
  66
  67         reqhead->snapid = cpu_to_le64(snapid);
  68
  69         /* object extent? */
  70         r = ceph_calc_file_object_mapping(layout, off, plen, bno,
  71                                           &objoff, &objlen);
  72         if (r < 0)
  73                 return r;
  74         if (*plen < orig_len)
  75                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  76                      orig_len - *plen, off, *plen);
  77
  78         if (op_has_extent(op->op)) {
  79                 u32 osize = le32_to_cpu(layout->fl_object_size);
  80                 op->extent.offset = objoff;
  81                 op->extent.length = objlen;
  82                 if (op->extent.truncate_size <= off - objoff) {
  83                         op->extent.truncate_size = 0;
  84                 } else {
  85                         op->extent.truncate_size -= off - objoff;
  86                         if (op->extent.truncate_size > osize)
  87                                 op->extent.truncate_size = osize;
  88                 }
  89         }
  90         req->r_num_pages = calc_pages_for(off, *plen);
  91         req->r_page_alignment = off & ~PAGE_MASK;
  92         if (op->op == CEPH_OSD_OP_WRITE)
  93                 op->payload_len = *plen;
  94
  95         dout("calc_layout bno=%llx %llu~%llu (%d pages)\n",
  96              *bno, objoff, objlen, req->r_num_pages);
  97         return 0;
  98 }
  99 EXPORT_SYMBOL(ceph_calc_raw_layout);
 100
 101 /*
 102  * Implement client access to distributed object storage cluster.
 103  *
 104  * All data objects are stored within a cluster/cloud of OSDs, or
 105  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
 106  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
 107  * remote daemons serving up and coordinating consistent and safe
 108  * access to storage.
 109  *
 110  * Cluster membership and the mapping of data objects onto storage devices
 111  * are described by the osd map.
 112  *
 113  * We keep track of pending OSD requests (read, write), resubmit
 114  * requests to different OSDs when the cluster topology/data layout
 115  * change, or retry the affected requests when the communications
 116  * channel with an OSD is reset.
 117  */
 118
 119 /*
 120  * calculate the mapping of a file extent onto an object, and fill out the
 121  * request accordingly.  shorten extent as necessary if it crosses an
 122  * object boundary.
 123  *
 124  * fill osd op in request message.
 125  */
 126 static int calc_layout(struct ceph_osd_client *osdc,
 127                        struct ceph_vino vino,
 128                        struct ceph_file_layout *layout,
 129                        u64 off, u64 *plen,
 130                        struct ceph_osd_request *req,
 131                        struct ceph_osd_req_op *op)
 132 {
 133         u64 bno;
 134         int r;
 135
 136         r = ceph_calc_raw_layout(osdc, layout, vino.snap, off,
 137                                  plen, &bno, req, op);
 138         if (r < 0)
 139                 return r;
 140
 141         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 142         req->r_oid_len = strlen(req->r_oid);
 143
 144         return r;
 145 }
 146
 147 /*
 148  * requests
 149  */
 150 void ceph_osdc_release_request(struct kref *kref)
 151 {
 152         struct ceph_osd_request *req = container_of(kref,
 153                                                     struct ceph_osd_request,
 154                                                     r_kref);
 155
 156         if (req->r_request)
 157                 ceph_msg_put(req->r_request);
 158         if (req->r_con_filling_msg) {
 159                 dout("%s revoking pages %p from con %p\n", __func__,
 160                      req->r_pages, req->r_con_filling_msg);
 161                 ceph_msg_revoke_incoming(req->r_reply);
 162                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 163         }
 164         if (req->r_reply)
 165                 ceph_msg_put(req->r_reply);
 166         if (req->r_own_pages)
 167                 ceph_release_page_vector(req->r_pages,
 168                                          req->r_num_pages);
 169 #ifdef CONFIG_BLOCK
 170         if (req->r_bio)
 171                 bio_put(req->r_bio);
 172 #endif
 173         ceph_put_snap_context(req->r_snapc);
 174         if (req->r_trail) {
 175                 ceph_pagelist_release(req->r_trail);
 176                 kfree(req->r_trail);
 177         }
 178         if (req->r_mempool)
 179                 mempool_free(req, req->r_osdc->req_mempool);
 180         else
 181                 kfree(req);
 182 }
 183 EXPORT_SYMBOL(ceph_osdc_release_request);
 184
 185 static int get_num_ops(struct ceph_osd_req_op *ops, int *needs_trail)
 186 {
 187         int i = 0;
 188
 189         if (needs_trail)
 190                 *needs_trail = 0;
 191         while (ops[i].op) {
 192                 if (needs_trail && op_needs_trail(ops[i].op))
 193                         *needs_trail = 1;
 194                 i++;
 195         }
 196
 197         return i;
 198 }
 199
 200 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 201                                                int flags,
 202                                                struct ceph_snap_context *snapc,
 203                                                struct ceph_osd_req_op *ops,
 204                                                bool use_mempool,
 205                                                gfp_t gfp_flags,
 206                                                struct page **pages,
 207                                                struct bio *bio)
 208 {
 209         struct ceph_osd_request *req;
 210         struct ceph_msg *msg;
 211         int needs_trail;
 212         int num_op = get_num_ops(ops, &needs_trail);
 213         size_t msg_size = sizeof(struct ceph_osd_request_head);
 214
 215         msg_size += num_op*sizeof(struct ceph_osd_op);
 216
 217         if (use_mempool) {
 218                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 219                 memset(req, 0, sizeof(*req));
 220         } else {
 221                 req = kzalloc(sizeof(*req), gfp_flags);
 222         }
 223         if (req == NULL)
 224                 return NULL;
 225
 226         req->r_osdc = osdc;
 227         req->r_mempool = use_mempool;
 228
 229         kref_init(&req->r_kref);
 230         init_completion(&req->r_completion);
 231         init_completion(&req->r_safe_completion);
 232         RB_CLEAR_NODE(&req->r_node);
 233         INIT_LIST_HEAD(&req->r_unsafe_item);
 234         INIT_LIST_HEAD(&req->r_linger_item);
 235         INIT_LIST_HEAD(&req->r_linger_osd);
 236         INIT_LIST_HEAD(&req->r_req_lru_item);
 237         INIT_LIST_HEAD(&req->r_osd_item);
 238
 239         req->r_flags = flags;
 240
 241         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 242
 243         /* create reply message */
 244         if (use_mempool)
 245                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 246         else
 247                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 248                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 249         if (!msg) {
 250                 ceph_osdc_put_request(req);
 251                 return NULL;
 252         }
 253         req->r_reply = msg;
 254
 255         /* allocate space for the trailing data */
 256         if (needs_trail) {
 257                 req->r_trail = kmalloc(sizeof(struct ceph_pagelist), gfp_flags);
 258                 if (!req->r_trail) {
 259                         ceph_osdc_put_request(req);
 260                         return NULL;
 261                 }
 262                 ceph_pagelist_init(req->r_trail);
 263         }
 264
 265         /* create request message; allow space for oid */
 266         msg_size += MAX_OBJ_NAME_SIZE;
 267         if (snapc)
 268                 msg_size += sizeof(u64) * snapc->num_snaps;
 269         if (use_mempool)
 270                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 271         else
 272                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 273         if (!msg) {
 274                 ceph_osdc_put_request(req);
 275                 return NULL;
 276         }
 277
 278         memset(msg->front.iov_base, 0, msg->front.iov_len);
 279
 280         req->r_request = msg;
 281         req->r_pages = pages;
 282 #ifdef CONFIG_BLOCK
 283         if (bio) {
 284                 req->r_bio = bio;
 285                 bio_get(req->r_bio);
 286         }
 287 #endif
 288
 289         return req;
 290 }
 291 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 292
 293 static void osd_req_encode_op(struct ceph_osd_request *req,
 294                               struct ceph_osd_op *dst,
 295                               struct ceph_osd_req_op *src)
 296 {
 297         dst->op = cpu_to_le16(src->op);
 298
 299         switch (src->op) {
 300         case CEPH_OSD_OP_READ:
 301         case CEPH_OSD_OP_WRITE:
 302                 dst->extent.offset =
 303                         cpu_to_le64(src->extent.offset);
 304                 dst->extent.length =
 305                         cpu_to_le64(src->extent.length);
 306                 dst->extent.truncate_size =
 307                         cpu_to_le64(src->extent.truncate_size);
 308                 dst->extent.truncate_seq =
 309                         cpu_to_le32(src->extent.truncate_seq);
 310                 break;
 311
 312         case CEPH_OSD_OP_GETXATTR:
 313         case CEPH_OSD_OP_SETXATTR:
 314         case CEPH_OSD_OP_CMPXATTR:
 315                 BUG_ON(!req->r_trail);
 316
 317                 dst->xattr.name_len = cpu_to_le32(src->xattr.name_len);
 318                 dst->xattr.value_len = cpu_to_le32(src->xattr.value_len);
 319                 dst->xattr.cmp_op = src->xattr.cmp_op;
 320                 dst->xattr.cmp_mode = src->xattr.cmp_mode;
 321                 ceph_pagelist_append(req->r_trail, src->xattr.name,
 322                                      src->xattr.name_len);
 323                 ceph_pagelist_append(req->r_trail, src->xattr.val,
 324                                      src->xattr.value_len);
 325                 break;
 326         case CEPH_OSD_OP_CALL:
 327                 BUG_ON(!req->r_trail);
 328
 329                 dst->cls.class_len = src->cls.class_len;
 330                 dst->cls.method_len = src->cls.method_len;
 331                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 332
 333                 ceph_pagelist_append(req->r_trail, src->cls.class_name,
 334                                      src->cls.class_len);
 335                 ceph_pagelist_append(req->r_trail, src->cls.method_name,
 336                                      src->cls.method_len);
 337                 ceph_pagelist_append(req->r_trail, src->cls.indata,
 338                                      src->cls.indata_len);
 339                 break;
 340         case CEPH_OSD_OP_ROLLBACK:
 341                 dst->snap.snapid = cpu_to_le64(src->snap.snapid);
 342                 break;
 343         case CEPH_OSD_OP_STARTSYNC:
 344                 break;
 345         case CEPH_OSD_OP_NOTIFY:
 346                 {
 347                         __le32 prot_ver = cpu_to_le32(src->watch.prot_ver);
 348                         __le32 timeout = cpu_to_le32(src->watch.timeout);
 349
 350                         BUG_ON(!req->r_trail);
 351
 352                         ceph_pagelist_append(req->r_trail,
 353                                                 &prot_ver, sizeof(prot_ver));
 354                         ceph_pagelist_append(req->r_trail,
 355                                                 &timeout, sizeof(timeout));
 356                 }
 357         case CEPH_OSD_OP_NOTIFY_ACK:
 358         case CEPH_OSD_OP_WATCH:
 359                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 360                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 361                 dst->watch.flag = src->watch.flag;
 362                 break;
 363         default:
 364                 pr_err("unrecognized osd opcode %d\n", dst->op);
 365                 WARN_ON(1);
 366                 break;
 367         }
 368         dst->payload_len = cpu_to_le32(src->payload_len);
 369 }
 370
 371 /*
 372  * build new request AND message
 373  *
 374  */
 375 void ceph_osdc_build_request(struct ceph_osd_request *req,
 376                              u64 off, u64 *plen,
 377                              struct ceph_osd_req_op *src_ops,
 378                              struct ceph_snap_context *snapc,
 379                              struct timespec *mtime,
 380                              const char *oid,
 381                              int oid_len)
 382 {
 383         struct ceph_msg *msg = req->r_request;
 384         struct ceph_osd_request_head *head;
 385         struct ceph_osd_req_op *src_op;
 386         struct ceph_osd_op *op;
 387         void *p;
 388         int num_op = get_num_ops(src_ops, NULL);
 389         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 390         int flags = req->r_flags;
 391         u64 data_len = 0;
 392         int i;
 393
 394         head = msg->front.iov_base;
 395         op = (void *)(head + 1);
 396         p = (void *)(op + num_op);
 397
 398         req->r_snapc = ceph_get_snap_context(snapc);
 399
 400         head->client_inc = cpu_to_le32(1); /* always, for now. */
 401         head->flags = cpu_to_le32(flags);
 402         if (flags & CEPH_OSD_FLAG_WRITE)
 403                 ceph_encode_timespec(&head->mtime, mtime);
 404         head->num_ops = cpu_to_le16(num_op);
 405
 406
 407         /* fill in oid */
 408         head->object_len = cpu_to_le32(oid_len);
 409         memcpy(p, oid, oid_len);
 410         p += oid_len;
 411
 412         src_op = src_ops;
 413         while (src_op->op) {
 414                 osd_req_encode_op(req, op, src_op);
 415                 src_op++;
 416                 op++;
 417         }
 418
 419         if (req->r_trail)
 420                 data_len += req->r_trail->length;
 421
 422         if (snapc) {
 423                 head->snap_seq = cpu_to_le64(snapc->seq);
 424                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 425                 for (i = 0; i < snapc->num_snaps; i++) {
 426                         put_unaligned_le64(snapc->snaps[i], p);
 427                         p += sizeof(u64);
 428                 }
 429         }
 430
 431         if (flags & CEPH_OSD_FLAG_WRITE) {
 432                 req->r_request->hdr.data_off = cpu_to_le16(off);
 433                 req->r_request->hdr.data_len = cpu_to_le32(*plen + data_len);
 434         } else if (data_len) {
 435                 req->r_request->hdr.data_off = 0;
 436                 req->r_request->hdr.data_len = cpu_to_le32(data_len);
 437         }
 438
 439         req->r_request->page_alignment = req->r_page_alignment;
 440
 441         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 442         msg_size = p - msg->front.iov_base;
 443         msg->front.iov_len = msg_size;
 444         msg->hdr.front_len = cpu_to_le32(msg_size);
 445         return;
 446 }
 447 EXPORT_SYMBOL(ceph_osdc_build_request);
 448
 449 /*
 450  * build new request AND message, calculate layout, and adjust file
 451  * extent as needed.
 452  *
 453  * if the file was recently truncated, we include information about its
 454  * old and new size so that the object can be updated appropriately.  (we
 455  * avoid synchronously deleting truncated objects because it's slow.)
 456  *
 457  * if @do_sync, include a 'startsync' command so that the osd will flush
 458  * data quickly.
 459  */
 460 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 461                                                struct ceph_file_layout *layout,
 462                                                struct ceph_vino vino,
 463                                                u64 off, u64 *plen,
 464                                                int opcode, int flags,
 465                                                struct ceph_snap_context *snapc,
 466                                                int do_sync,
 467                                                u32 truncate_seq,
 468                                                u64 truncate_size,
 469                                                struct timespec *mtime,
 470                                                bool use_mempool, int num_reply,
 471                                                int page_align)
 472 {
 473         struct ceph_osd_req_op ops[3];
 474         struct ceph_osd_request *req;
 475         int r;
 476
 477         ops[0].op = opcode;
 478         ops[0].extent.truncate_seq = truncate_seq;
 479         ops[0].extent.truncate_size = truncate_size;
 480         ops[0].payload_len = 0;
 481
 482         if (do_sync) {
 483                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 484                 ops[1].payload_len = 0;
 485                 ops[2].op = 0;
 486         } else
 487                 ops[1].op = 0;
 488
 489         req = ceph_osdc_alloc_request(osdc, flags,
 490                                          snapc, ops,
 491                                          use_mempool,
 492                                          GFP_NOFS, NULL, NULL);
 493         if (!req)
 494                 return ERR_PTR(-ENOMEM);
 495
 496         /* calculate max write size */
 497         r = calc_layout(osdc, vino, layout, off, plen, req, ops);
 498         if (r < 0)
 499                 return ERR_PTR(r);
 500         req->r_file_layout = *layout;  /* keep a copy */
 501
 502         /* in case it differs from natural (file) alignment that
 503            calc_layout filled in for us */
 504         req->r_num_pages = calc_pages_for(page_align, *plen);
 505         req->r_page_alignment = page_align;
 506
 507         ceph_osdc_build_request(req, off, plen, ops,
 508                                 snapc,
 509                                 mtime,
 510                                 req->r_oid, req->r_oid_len);
 511
 512         return req;
 513 }
 514 EXPORT_SYMBOL(ceph_osdc_new_request);
 515
 516 /*
 517  * We keep osd requests in an rbtree, sorted by ->r_tid.
 518  */
 519 static void __insert_request(struct ceph_osd_client *osdc,
 520                              struct ceph_osd_request *new)
 521 {
 522         struct rb_node **p = &osdc->requests.rb_node;
 523         struct rb_node *parent = NULL;
 524         struct ceph_osd_request *req = NULL;
 525
 526         while (*p) {
 527                 parent = *p;
 528                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 529                 if (new->r_tid < req->r_tid)
 530                         p = &(*p)->rb_left;
 531                 else if (new->r_tid > req->r_tid)
 532                         p = &(*p)->rb_right;
 533                 else
 534                         BUG();
 535         }
 536
 537         rb_link_node(&new->r_node, parent, p);
 538         rb_insert_color(&new->r_node, &osdc->requests);
 539 }
 540
 541 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 542                                                  u64 tid)
 543 {
 544         struct ceph_osd_request *req;
 545         struct rb_node *n = osdc->requests.rb_node;
 546
 547         while (n) {
 548                 req = rb_entry(n, struct ceph_osd_request, r_node);
 549                 if (tid < req->r_tid)
 550                         n = n->rb_left;
 551                 else if (tid > req->r_tid)
 552                         n = n->rb_right;
 553                 else
 554                         return req;
 555         }
 556         return NULL;
 557 }
 558
 559 static struct ceph_osd_request *
 560 __lookup_request_ge(struct ceph_osd_client *osdc,
 561                     u64 tid)
 562 {
 563         struct ceph_osd_request *req;
 564         struct rb_node *n = osdc->requests.rb_node;
 565
 566         while (n) {
 567                 req = rb_entry(n, struct ceph_osd_request, r_node);
 568                 if (tid < req->r_tid) {
 569                         if (!n->rb_left)
 570                                 return req;
 571                         n = n->rb_left;
 572                 } else if (tid > req->r_tid) {
 573                         n = n->rb_right;
 574                 } else {
 575                         return req;
 576                 }
 577         }
 578         return NULL;
 579 }
 580
 581 /*
 582  * Resubmit requests pending on the given osd.
 583  */
 584 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 585                                 struct ceph_osd *osd)
 586 {
 587         struct ceph_osd_request *req, *nreq;
 588         int err;
 589
 590         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 591         err = __reset_osd(osdc, osd);
 592         if (err)
 593                 return;
 594
 595         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 596                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 597                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 598                      osd->o_osd);
 599                 if (!req->r_linger)
 600                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 601         }
 602
 603         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 604                                  r_linger_osd) {
 605                 /*
 606                  * reregister request prior to unregistering linger so
 607                  * that r_osd is preserved.
 608                  */
 609                 BUG_ON(!list_empty(&req->r_req_lru_item));
 610                 __register_request(osdc, req);
 611                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 612                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 613                 __unregister_linger_request(osdc, req);
 614                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 615                      osd->o_osd);
 616         }
 617 }
 618
 619 /*
 620  * If the osd connection drops, we need to resubmit all requests.
 621  */
 622 static void osd_reset(struct ceph_connection *con)
 623 {
 624         struct ceph_osd *osd = con->private;
 625         struct ceph_osd_client *osdc;
 626
 627         if (!osd)
 628                 return;
 629         dout("osd_reset osd%d\n", osd->o_osd);
 630         osdc = osd->o_osdc;
 631         down_read(&osdc->map_sem);
 632         mutex_lock(&osdc->request_mutex);
 633         __kick_osd_requests(osdc, osd);
 634         mutex_unlock(&osdc->request_mutex);
 635         send_queued(osdc);
 636         up_read(&osdc->map_sem);
 637 }
 638
 639 /*
 640  * Track open sessions with osds.
 641  */
 642 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 643 {
 644         struct ceph_osd *osd;
 645
 646         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 647         if (!osd)
 648                 return NULL;
 649
 650         atomic_set(&osd->o_ref, 1);
 651         osd->o_osdc = osdc;
 652         osd->o_osd = onum;
 653         RB_CLEAR_NODE(&osd->o_node);
 654         INIT_LIST_HEAD(&osd->o_requests);
 655         INIT_LIST_HEAD(&osd->o_linger_requests);
 656         INIT_LIST_HEAD(&osd->o_osd_lru);
 657         osd->o_incarnation = 1;
 658
 659         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 660
 661         INIT_LIST_HEAD(&osd->o_keepalive_item);
 662         return osd;
 663 }
 664
 665 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 666 {
 667         if (atomic_inc_not_zero(&osd->o_ref)) {
 668                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 669                      atomic_read(&osd->o_ref));
 670                 return osd;
 671         } else {
 672                 dout("get_osd %p FAIL\n", osd);
 673                 return NULL;
 674         }
 675 }
 676
 677 static void put_osd(struct ceph_osd *osd)
 678 {
 679         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 680              atomic_read(&osd->o_ref) - 1);
 681         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 682                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 683
 684                 if (ac->ops && ac->ops->destroy_authorizer)
 685                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 686                 kfree(osd);
 687         }
 688 }
 689
 690 /*
 691  * remove an osd from our map
 692  */
 693 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 694 {
 695         dout("__remove_osd %p\n", osd);
 696         BUG_ON(!list_empty(&osd->o_requests));
 697         rb_erase(&osd->o_node, &osdc->osds);
 698         list_del_init(&osd->o_osd_lru);
 699         ceph_con_close(&osd->o_con);
 700         put_osd(osd);
 701 }
 702
 703 static void remove_all_osds(struct ceph_osd_client *osdc)
 704 {
 705         dout("%s %p\n", __func__, osdc);
 706         mutex_lock(&osdc->request_mutex);
 707         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 708                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 709                                                 struct ceph_osd, o_node);
 710                 __remove_osd(osdc, osd);
 711         }
 712         mutex_unlock(&osdc->request_mutex);
 713 }
 714
 715 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 716                               struct ceph_osd *osd)
 717 {
 718         dout("__move_osd_to_lru %p\n", osd);
 719         BUG_ON(!list_empty(&osd->o_osd_lru));
 720         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 721         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 722 }
 723
 724 static void __remove_osd_from_lru(struct ceph_osd *osd)
 725 {
 726         dout("__remove_osd_from_lru %p\n", osd);
 727         if (!list_empty(&osd->o_osd_lru))
 728                 list_del_init(&osd->o_osd_lru);
 729 }
 730
 731 static void remove_old_osds(struct ceph_osd_client *osdc)
 732 {
 733         struct ceph_osd *osd, *nosd;
 734
 735         dout("__remove_old_osds %p\n", osdc);
 736         mutex_lock(&osdc->request_mutex);
 737         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 738                 if (time_before(jiffies, osd->lru_ttl))
 739                         break;
 740                 __remove_osd(osdc, osd);
 741         }
 742         mutex_unlock(&osdc->request_mutex);
 743 }
 744
 745 /*
 746  * reset osd connect
 747  */
 748 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 749 {
 750         struct ceph_entity_addr *peer_addr;
 751
 752         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 753         if (list_empty(&osd->o_requests) &&
 754             list_empty(&osd->o_linger_requests)) {
 755                 __remove_osd(osdc, osd);
 756
 757                 return -ENODEV;
 758         }
 759
 760         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
 761         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
 762                         !ceph_con_opened(&osd->o_con)) {
 763                 struct ceph_osd_request *req;
 764
 765                 dout(" osd addr hasn't changed and connection never opened,"
 766                      " letting msgr retry");
 767                 /* touch each r_stamp for handle_timeout()'s benfit */
 768                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 769                         req->r_stamp = jiffies;
 770
 771                 return -EAGAIN;
 772         }
 773
 774         ceph_con_close(&osd->o_con);
 775         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
 776         osd->o_incarnation++;
 777
 778         return 0;
 779 }
 780
 781 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 782 {
 783         struct rb_node **p = &osdc->osds.rb_node;
 784         struct rb_node *parent = NULL;
 785         struct ceph_osd *osd = NULL;
 786
 787         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 788         while (*p) {
 789                 parent = *p;
 790                 osd = rb_entry(parent, struct ceph_osd, o_node);
 791                 if (new->o_osd < osd->o_osd)
 792                         p = &(*p)->rb_left;
 793                 else if (new->o_osd > osd->o_osd)
 794                         p = &(*p)->rb_right;
 795                 else
 796                         BUG();
 797         }
 798
 799         rb_link_node(&new->o_node, parent, p);
 800         rb_insert_color(&new->o_node, &osdc->osds);
 801 }
 802
 803 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 804 {
 805         struct ceph_osd *osd;
 806         struct rb_node *n = osdc->osds.rb_node;
 807
 808         while (n) {
 809                 osd = rb_entry(n, struct ceph_osd, o_node);
 810                 if (o < osd->o_osd)
 811                         n = n->rb_left;
 812                 else if (o > osd->o_osd)
 813                         n = n->rb_right;
 814                 else
 815                         return osd;
 816         }
 817         return NULL;
 818 }
 819
 820 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 821 {
 822         schedule_delayed_work(&osdc->timeout_work,
 823                         osdc->client->options->osd_keepalive_timeout * HZ);
 824 }
 825
 826 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 827 {
 828         cancel_delayed_work(&osdc->timeout_work);
 829 }
 830
 831 /*
 832  * Register request, assign tid.  If this is the first request, set up
 833  * the timeout event.
 834  */
 835 static void __register_request(struct ceph_osd_client *osdc,
 836                                struct ceph_osd_request *req)
 837 {
 838         req->r_tid = ++osdc->last_tid;
 839         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 840         dout("__register_request %p tid %lld\n", req, req->r_tid);
 841         __insert_request(osdc, req);
 842         ceph_osdc_get_request(req);
 843         osdc->num_requests++;
 844         if (osdc->num_requests == 1) {
 845                 dout(" first request, scheduling timeout\n");
 846                 __schedule_osd_timeout(osdc);
 847         }
 848 }
 849
 850 static void register_request(struct ceph_osd_client *osdc,
 851                              struct ceph_osd_request *req)
 852 {
 853         mutex_lock(&osdc->request_mutex);
 854         __register_request(osdc, req);
 855         mutex_unlock(&osdc->request_mutex);
 856 }
 857
 858 /*
 859  * called under osdc->request_mutex
 860  */
 861 static void __unregister_request(struct ceph_osd_client *osdc,
 862                                  struct ceph_osd_request *req)
 863 {
 864         if (RB_EMPTY_NODE(&req->r_node)) {
 865                 dout("__unregister_request %p tid %lld not registered\n",
 866                         req, req->r_tid);
 867                 return;
 868         }
 869
 870         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 871         rb_erase(&req->r_node, &osdc->requests);
 872         osdc->num_requests--;
 873
 874         if (req->r_osd) {
 875                 /* make sure the original request isn't in flight. */
 876                 ceph_msg_revoke(req->r_request);
 877
 878                 list_del_init(&req->r_osd_item);
 879                 if (list_empty(&req->r_osd->o_requests) &&
 880                     list_empty(&req->r_osd->o_linger_requests)) {
 881                         dout("moving osd to %p lru\n", req->r_osd);
 882                         __move_osd_to_lru(osdc, req->r_osd);
 883                 }
 884                 if (list_empty(&req->r_linger_item))
 885                         req->r_osd = NULL;
 886         }
 887
 888         list_del_init(&req->r_req_lru_item);
 889         ceph_osdc_put_request(req);
 890
 891         if (osdc->num_requests == 0) {
 892                 dout(" no requests, canceling timeout\n");
 893                 __cancel_osd_timeout(osdc);
 894         }
 895 }
 896
 897 /*
 898  * Cancel a previously queued request message
 899  */
 900 static void __cancel_request(struct ceph_osd_request *req)
 901 {
 902         if (req->r_sent && req->r_osd) {
 903                 ceph_msg_revoke(req->r_request);
 904                 req->r_sent = 0;
 905         }
 906 }
 907
 908 static void __register_linger_request(struct ceph_osd_client *osdc,
 909                                     struct ceph_osd_request *req)
 910 {
 911         dout("__register_linger_request %p\n", req);
 912         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 913         if (req->r_osd)
 914                 list_add_tail(&req->r_linger_osd,
 915                               &req->r_osd->o_linger_requests);
 916 }
 917
 918 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 919                                         struct ceph_osd_request *req)
 920 {
 921         dout("__unregister_linger_request %p\n", req);
 922         list_del_init(&req->r_linger_item);
 923         if (req->r_osd) {
 924                 list_del_init(&req->r_linger_osd);
 925
 926                 if (list_empty(&req->r_osd->o_requests) &&
 927                     list_empty(&req->r_osd->o_linger_requests)) {
 928                         dout("moving osd to %p lru\n", req->r_osd);
 929                         __move_osd_to_lru(osdc, req->r_osd);
 930                 }
 931                 if (list_empty(&req->r_osd_item))
 932                         req->r_osd = NULL;
 933         }
 934 }
 935
 936 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 937                                          struct ceph_osd_request *req)
 938 {
 939         mutex_lock(&osdc->request_mutex);
 940         if (req->r_linger) {
 941                 __unregister_linger_request(osdc, req);
 942                 ceph_osdc_put_request(req);
 943         }
 944         mutex_unlock(&osdc->request_mutex);
 945 }
 946 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 947
 948 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 949                                   struct ceph_osd_request *req)
 950 {
 951         if (!req->r_linger) {
 952                 dout("set_request_linger %p\n", req);
 953                 req->r_linger = 1;
 954                 /*
 955                  * caller is now responsible for calling
 956                  * unregister_linger_request
 957                  */
 958                 ceph_osdc_get_request(req);
 959         }
 960 }
 961 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 962
 963 /*
 964  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 965  * (as needed), and set the request r_osd appropriately.  If there is
 966  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 967  * (unsent, homeless) or leave on in-flight lru.
 968  *
 969  * Return 0 if unchanged, 1 if changed, or negative on error.
 970  *
 971  * Caller should hold map_sem for read and request_mutex.
 972  */
 973 static int __map_request(struct ceph_osd_client *osdc,
 974                          struct ceph_osd_request *req, int force_resend)
 975 {
 976         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 977         struct ceph_pg pgid;
 978         int acting[CEPH_PG_MAX_SIZE];
 979         int o = -1, num = 0;
 980         int err;
 981
 982         dout("map_request %p tid %lld\n", req, req->r_tid);
 983         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 984                                       &req->r_file_layout, osdc->osdmap);
 985         if (err) {
 986                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 987                 return err;
 988         }
 989         pgid = reqhead->layout.ol_pgid;
 990         req->r_pgid = pgid;
 991
 992         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 993         if (err > 0) {
 994                 o = acting[0];
 995                 num = err;
 996         }
 997
 998         if ((!force_resend &&
 999              req->r_osd && req->r_osd->o_osd == o &&
1000              req->r_sent >= req->r_osd->o_incarnation &&
1001              req->r_num_pg_osds == num &&
1002              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
1003             (req->r_osd == NULL && o == -1))
1004                 return 0;  /* no change */
1005
1006         dout("map_request tid %llu pgid %d.%x osd%d (was osd%d)\n",
1007              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
1008              req->r_osd ? req->r_osd->o_osd : -1);
1009
1010         /* record full pg acting set */
1011         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
1012         req->r_num_pg_osds = num;
1013
1014         if (req->r_osd) {
1015                 __cancel_request(req);
1016                 list_del_init(&req->r_osd_item);
1017                 req->r_osd = NULL;
1018         }
1019
1020         req->r_osd = __lookup_osd(osdc, o);
1021         if (!req->r_osd && o >= 0) {
1022                 err = -ENOMEM;
1023                 req->r_osd = create_osd(osdc, o);
1024                 if (!req->r_osd) {
1025                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
1026                         goto out;
1027                 }
1028
1029                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
1030                 __insert_osd(osdc, req->r_osd);
1031
1032                 ceph_con_open(&req->r_osd->o_con,
1033                               CEPH_ENTITY_TYPE_OSD, o,
1034                               &osdc->osdmap->osd_addr[o]);
1035         }
1036
1037         if (req->r_osd) {
1038                 __remove_osd_from_lru(req->r_osd);
1039                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1040                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1041         } else {
1042                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1043         }
1044         err = 1;   /* osd or pg changed */
1045
1046 out:
1047         return err;
1048 }
1049
1050 /*
1051  * caller should hold map_sem (for read) and request_mutex
1052  */
1053 static void __send_request(struct ceph_osd_client *osdc,
1054                            struct ceph_osd_request *req)
1055 {
1056         struct ceph_osd_request_head *reqhead;
1057
1058         dout("send_request %p tid %llu to osd%d flags %d\n",
1059              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
1060
1061         reqhead = req->r_request->front.iov_base;
1062         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
1063         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
1064         reqhead->reassert_version = req->r_reassert_version;
1065
1066         req->r_stamp = jiffies;
1067         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1068
1069         ceph_msg_get(req->r_request); /* send consumes a ref */
1070         ceph_con_send(&req->r_osd->o_con, req->r_request);
1071         req->r_sent = req->r_osd->o_incarnation;
1072 }
1073
1074 /*
1075  * Send any requests in the queue (req_unsent).
1076  */
1077 static void send_queued(struct ceph_osd_client *osdc)
1078 {
1079         struct ceph_osd_request *req, *tmp;
1080
1081         dout("send_queued\n");
1082         mutex_lock(&osdc->request_mutex);
1083         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item) {
1084                 __send_request(osdc, req);
1085         }
1086         mutex_unlock(&osdc->request_mutex);
1087 }
1088
1089 /*
1090  * Timeout callback, called every N seconds when 1 or more osd
1091  * requests has been active for more than N seconds.  When this
1092  * happens, we ping all OSDs with requests who have timed out to
1093  * ensure any communications channel reset is detected.  Reset the
1094  * request timeouts another N seconds in the future as we go.
1095  * Reschedule the timeout event another N seconds in future (unless
1096  * there are no open requests).
1097  */
1098 static void handle_timeout(struct work_struct *work)
1099 {
1100         struct ceph_osd_client *osdc =
1101                 container_of(work, struct ceph_osd_client, timeout_work.work);
1102         struct ceph_osd_request *req;
1103         struct ceph_osd *osd;
1104         unsigned long keepalive =
1105                 osdc->client->options->osd_keepalive_timeout * HZ;
1106         struct list_head slow_osds;
1107         dout("timeout\n");
1108         down_read(&osdc->map_sem);
1109
1110         ceph_monc_request_next_osdmap(&osdc->client->monc);
1111
1112         mutex_lock(&osdc->request_mutex);
1113
1114         /*
1115          * ping osds that are a bit slow.  this ensures that if there
1116          * is a break in the TCP connection we will notice, and reopen
1117          * a connection with that osd (from the fault callback).
1118          */
1119         INIT_LIST_HEAD(&slow_osds);
1120         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1121                 if (time_before(jiffies, req->r_stamp + keepalive))
1122                         break;
1123
1124                 osd = req->r_osd;
1125                 BUG_ON(!osd);
1126                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1127                      req->r_tid, osd->o_osd);
1128                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1129         }
1130         while (!list_empty(&slow_osds)) {
1131                 osd = list_entry(slow_osds.next, struct ceph_osd,
1132                                  o_keepalive_item);
1133                 list_del_init(&osd->o_keepalive_item);
1134                 ceph_con_keepalive(&osd->o_con);
1135         }
1136
1137         __schedule_osd_timeout(osdc);
1138         mutex_unlock(&osdc->request_mutex);
1139         send_queued(osdc);
1140         up_read(&osdc->map_sem);
1141 }
1142
1143 static void handle_osds_timeout(struct work_struct *work)
1144 {
1145         struct ceph_osd_client *osdc =
1146                 container_of(work, struct ceph_osd_client,
1147                              osds_timeout_work.work);
1148         unsigned long delay =
1149                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1150
1151         dout("osds timeout\n");
1152         down_read(&osdc->map_sem);
1153         remove_old_osds(osdc);
1154         up_read(&osdc->map_sem);
1155
1156         schedule_delayed_work(&osdc->osds_timeout_work,
1157                               round_jiffies_relative(delay));
1158 }
1159
1160 static void complete_request(struct ceph_osd_request *req)
1161 {
1162         if (req->r_safe_callback)
1163                 req->r_safe_callback(req, NULL);
1164         complete_all(&req->r_safe_completion);  /* fsync waiter */
1165 }
1166
1167 /*
1168  * handle osd op reply.  either call the callback if it is specified,
1169  * or do the completion to wake up the waiting thread.
1170  */
1171 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1172                          struct ceph_connection *con)
1173 {
1174         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
1175         struct ceph_osd_request *req;
1176         u64 tid;
1177         int numops, object_len, flags;
1178         s32 result;
1179
1180         tid = le64_to_cpu(msg->hdr.tid);
1181         if (msg->front.iov_len < sizeof(*rhead))
1182                 goto bad;
1183         numops = le32_to_cpu(rhead->num_ops);
1184         object_len = le32_to_cpu(rhead->object_len);
1185         result = le32_to_cpu(rhead->result);
1186         if (msg->front.iov_len != sizeof(*rhead) + object_len +
1187             numops * sizeof(struct ceph_osd_op))
1188                 goto bad;
1189         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
1190         /* lookup */
1191         mutex_lock(&osdc->request_mutex);
1192         req = __lookup_request(osdc, tid);
1193         if (req == NULL) {
1194                 dout("handle_reply tid %llu dne\n", tid);
1195                 mutex_unlock(&osdc->request_mutex);
1196                 return;
1197         }
1198         ceph_osdc_get_request(req);
1199         flags = le32_to_cpu(rhead->flags);
1200
1201         /*
1202          * if this connection filled our message, drop our reference now, to
1203          * avoid a (safe but slower) revoke later.
1204          */
1205         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1206                 dout(" dropping con_filling_msg ref %p\n", con);
1207                 req->r_con_filling_msg = NULL;
1208                 con->ops->put(con);
1209         }
1210
1211         if (!req->r_got_reply) {
1212                 unsigned int bytes;
1213
1214                 req->r_result = le32_to_cpu(rhead->result);
1215                 bytes = le32_to_cpu(msg->hdr.data_len);
1216                 dout("handle_reply result %d bytes %d\n", req->r_result,
1217                      bytes);
1218                 if (req->r_result == 0)
1219                         req->r_result = bytes;
1220
1221                 /* in case this is a write and we need to replay, */
1222                 req->r_reassert_version = rhead->reassert_version;
1223
1224                 req->r_got_reply = 1;
1225         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1226                 dout("handle_reply tid %llu dup ack\n", tid);
1227                 mutex_unlock(&osdc->request_mutex);
1228                 goto done;
1229         }
1230
1231         dout("handle_reply tid %llu flags %d\n", tid, flags);
1232
1233         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1234                 __register_linger_request(osdc, req);
1235
1236         /* either this is a read, or we got the safe response */
1237         if (result < 0 ||
1238             (flags & CEPH_OSD_FLAG_ONDISK) ||
1239             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1240                 __unregister_request(osdc, req);
1241
1242         mutex_unlock(&osdc->request_mutex);
1243
1244         if (req->r_callback)
1245                 req->r_callback(req, msg);
1246         else
1247                 complete_all(&req->r_completion);
1248
1249         if (flags & CEPH_OSD_FLAG_ONDISK)
1250                 complete_request(req);
1251
1252 done:
1253         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1254         ceph_osdc_put_request(req);
1255         return;
1256
1257 bad:
1258         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
1259                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
1260                (int)sizeof(*rhead));
1261         ceph_msg_dump(msg);
1262 }
1263
1264 static void reset_changed_osds(struct ceph_osd_client *osdc)
1265 {
1266         struct rb_node *p, *n;
1267
1268         for (p = rb_first(&osdc->osds); p; p = n) {
1269                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1270
1271                 n = rb_next(p);
1272                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1273                     memcmp(&osd->o_con.peer_addr,
1274                            ceph_osd_addr(osdc->osdmap,
1275                                          osd->o_osd),
1276                            sizeof(struct ceph_entity_addr)) != 0)
1277                         __reset_osd(osdc, osd);
1278         }
1279 }
1280
1281 /*
1282  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1283  * no osd, request a new map.
1284  *
1285  * Caller should hold map_sem for read.
1286  */
1287 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1288 {
1289         struct ceph_osd_request *req, *nreq;
1290         struct rb_node *p;
1291         int needmap = 0;
1292         int err;
1293
1294         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1295         mutex_lock(&osdc->request_mutex);
1296         for (p = rb_first(&osdc->requests); p; ) {
1297                 req = rb_entry(p, struct ceph_osd_request, r_node);
1298                 p = rb_next(p);
1299
1300                 /*
1301                  * For linger requests that have not yet been
1302                  * registered, move them to the linger list; they'll
1303                  * be sent to the osd in the loop below.  Unregister
1304                  * the request before re-registering it as a linger
1305                  * request to ensure the __map_request() below
1306                  * will decide it needs to be sent.
1307                  */
1308                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1309                         dout("%p tid %llu restart on osd%d\n",
1310                              req, req->r_tid,
1311                              req->r_osd ? req->r_osd->o_osd : -1);
1312                         __unregister_request(osdc, req);
1313                         __register_linger_request(osdc, req);
1314                         continue;
1315                 }
1316
1317                 err = __map_request(osdc, req, force_resend);
1318                 if (err < 0)
1319                         continue;  /* error */
1320                 if (req->r_osd == NULL) {
1321                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1322                         needmap++;  /* request a newer map */
1323                 } else if (err > 0) {
1324                         if (!req->r_linger) {
1325                                 dout("%p tid %llu requeued on osd%d\n", req,
1326                                      req->r_tid,
1327                                      req->r_osd ? req->r_osd->o_osd : -1);
1328                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1329                         }
1330                 }
1331         }
1332
1333         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1334                                  r_linger_item) {
1335                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1336
1337                 err = __map_request(osdc, req, force_resend);
1338                 dout("__map_request returned %d\n", err);
1339                 if (err == 0)
1340                         continue;  /* no change and no osd was specified */
1341                 if (err < 0)
1342                         continue;  /* hrm! */
1343                 if (req->r_osd == NULL) {
1344                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1345                         needmap++;  /* request a newer map */
1346                         continue;
1347                 }
1348
1349                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1350                      req->r_osd ? req->r_osd->o_osd : -1);
1351                 __register_request(osdc, req);
1352                 __unregister_linger_request(osdc, req);
1353         }
1354         mutex_unlock(&osdc->request_mutex);
1355
1356         if (needmap) {
1357                 dout("%d requests for down osds, need new map\n", needmap);
1358                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1359         }
1360         reset_changed_osds(osdc);
1361 }
1362
1363
1364 /*
1365  * Process updated osd map.
1366  *
1367  * The message contains any number of incremental and full maps, normally
1368  * indicating some sort of topology change in the cluster.  Kick requests
1369  * off to different OSDs as needed.
1370  */
1371 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1372 {
1373         void *p, *end, *next;
1374         u32 nr_maps, maplen;
1375         u32 epoch;
1376         struct ceph_osdmap *newmap = NULL, *oldmap;
1377         int err;
1378         struct ceph_fsid fsid;
1379
1380         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1381         p = msg->front.iov_base;
1382         end = p + msg->front.iov_len;
1383
1384         /* verify fsid */
1385         ceph_decode_need(&p, end, sizeof(fsid), bad);
1386         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1387         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1388                 return;
1389
1390         down_write(&osdc->map_sem);
1391
1392         /* incremental maps */
1393         ceph_decode_32_safe(&p, end, nr_maps, bad);
1394         dout(" %d inc maps\n", nr_maps);
1395         while (nr_maps > 0) {
1396                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1397                 epoch = ceph_decode_32(&p);
1398                 maplen = ceph_decode_32(&p);
1399                 ceph_decode_need(&p, end, maplen, bad);
1400                 next = p + maplen;
1401                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1402                         dout("applying incremental map %u len %d\n",
1403                              epoch, maplen);
1404                         newmap = osdmap_apply_incremental(&p, next,
1405                                                           osdc->osdmap,
1406                                                           &osdc->client->msgr);
1407                         if (IS_ERR(newmap)) {
1408                                 err = PTR_ERR(newmap);
1409                                 goto bad;
1410                         }
1411                         BUG_ON(!newmap);
1412                         if (newmap != osdc->osdmap) {
1413                                 ceph_osdmap_destroy(osdc->osdmap);
1414                                 osdc->osdmap = newmap;
1415                         }
1416                         kick_requests(osdc, 0);
1417                 } else {
1418                         dout("ignoring incremental map %u len %d\n",
1419                              epoch, maplen);
1420                 }
1421                 p = next;
1422                 nr_maps--;
1423         }
1424         if (newmap)
1425                 goto done;
1426
1427         /* full maps */
1428         ceph_decode_32_safe(&p, end, nr_maps, bad);
1429         dout(" %d full maps\n", nr_maps);
1430         while (nr_maps) {
1431                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1432                 epoch = ceph_decode_32(&p);
1433                 maplen = ceph_decode_32(&p);
1434                 ceph_decode_need(&p, end, maplen, bad);
1435                 if (nr_maps > 1) {
1436                         dout("skipping non-latest full map %u len %d\n",
1437                              epoch, maplen);
1438                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1439                         dout("skipping full map %u len %d, "
1440                              "older than our %u\n", epoch, maplen,
1441                              osdc->osdmap->epoch);
1442                 } else {
1443                         int skipped_map = 0;
1444
1445                         dout("taking full map %u len %d\n", epoch, maplen);
1446                         newmap = osdmap_decode(&p, p+maplen);
1447                         if (IS_ERR(newmap)) {
1448                                 err = PTR_ERR(newmap);
1449                                 goto bad;
1450                         }
1451                         BUG_ON(!newmap);
1452                         oldmap = osdc->osdmap;
1453                         osdc->osdmap = newmap;
1454                         if (oldmap) {
1455                                 if (oldmap->epoch + 1 < newmap->epoch)
1456                                         skipped_map = 1;
1457                                 ceph_osdmap_destroy(oldmap);
1458                         }
1459                         kick_requests(osdc, skipped_map);
1460                 }
1461                 p += maplen;
1462                 nr_maps--;
1463         }
1464
1465 done:
1466         downgrade_write(&osdc->map_sem);
1467         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1468
1469         /*
1470          * subscribe to subsequent osdmap updates if full to ensure
1471          * we find out when we are no longer full and stop returning
1472          * ENOSPC.
1473          */
1474         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1475                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1476
1477         send_queued(osdc);
1478         up_read(&osdc->map_sem);
1479         wake_up_all(&osdc->client->auth_wq);
1480         return;
1481
1482 bad:
1483         pr_err("osdc handle_map corrupt msg\n");
1484         ceph_msg_dump(msg);
1485         up_write(&osdc->map_sem);
1486         return;
1487 }
1488
1489 /*
1490  * watch/notify callback event infrastructure
1491  *
1492  * These callbacks are used both for watch and notify operations.
1493  */
1494 static void __release_event(struct kref *kref)
1495 {
1496         struct ceph_osd_event *event =
1497                 container_of(kref, struct ceph_osd_event, kref);
1498
1499         dout("__release_event %p\n", event);
1500         kfree(event);
1501 }
1502
1503 static void get_event(struct ceph_osd_event *event)
1504 {
1505         kref_get(&event->kref);
1506 }
1507
1508 void ceph_osdc_put_event(struct ceph_osd_event *event)
1509 {
1510         kref_put(&event->kref, __release_event);
1511 }
1512 EXPORT_SYMBOL(ceph_osdc_put_event);
1513
1514 static void __insert_event(struct ceph_osd_client *osdc,
1515                              struct ceph_osd_event *new)
1516 {
1517         struct rb_node **p = &osdc->event_tree.rb_node;
1518         struct rb_node *parent = NULL;
1519         struct ceph_osd_event *event = NULL;
1520
1521         while (*p) {
1522                 parent = *p;
1523                 event = rb_entry(parent, struct ceph_osd_event, node);
1524                 if (new->cookie < event->cookie)
1525                         p = &(*p)->rb_left;
1526                 else if (new->cookie > event->cookie)
1527                         p = &(*p)->rb_right;
1528                 else
1529                         BUG();
1530         }
1531
1532         rb_link_node(&new->node, parent, p);
1533         rb_insert_color(&new->node, &osdc->event_tree);
1534 }
1535
1536 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1537                                                 u64 cookie)
1538 {
1539         struct rb_node **p = &osdc->event_tree.rb_node;
1540         struct rb_node *parent = NULL;
1541         struct ceph_osd_event *event = NULL;
1542
1543         while (*p) {
1544                 parent = *p;
1545                 event = rb_entry(parent, struct ceph_osd_event, node);
1546                 if (cookie < event->cookie)
1547                         p = &(*p)->rb_left;
1548                 else if (cookie > event->cookie)
1549                         p = &(*p)->rb_right;
1550                 else
1551                         return event;
1552         }
1553         return NULL;
1554 }
1555
1556 static void __remove_event(struct ceph_osd_event *event)
1557 {
1558         struct ceph_osd_client *osdc = event->osdc;
1559
1560         if (!RB_EMPTY_NODE(&event->node)) {
1561                 dout("__remove_event removed %p\n", event);
1562                 rb_erase(&event->node, &osdc->event_tree);
1563                 ceph_osdc_put_event(event);
1564         } else {
1565                 dout("__remove_event didn't remove %p\n", event);
1566         }
1567 }
1568
1569 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1570                            void (*event_cb)(u64, u64, u8, void *),
1571                            int one_shot, void *data,
1572                            struct ceph_osd_event **pevent)
1573 {
1574         struct ceph_osd_event *event;
1575
1576         event = kmalloc(sizeof(*event), GFP_NOIO);
1577         if (!event)
1578                 return -ENOMEM;
1579
1580         dout("create_event %p\n", event);
1581         event->cb = event_cb;
1582         event->one_shot = one_shot;
1583         event->data = data;
1584         event->osdc = osdc;
1585         INIT_LIST_HEAD(&event->osd_node);
1586         RB_CLEAR_NODE(&event->node);
1587         kref_init(&event->kref);   /* one ref for us */
1588         kref_get(&event->kref);    /* one ref for the caller */
1589         init_completion(&event->completion);
1590
1591         spin_lock(&osdc->event_lock);
1592         event->cookie = ++osdc->event_count;
1593         __insert_event(osdc, event);
1594         spin_unlock(&osdc->event_lock);
1595
1596         *pevent = event;
1597         return 0;
1598 }
1599 EXPORT_SYMBOL(ceph_osdc_create_event);
1600
1601 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1602 {
1603         struct ceph_osd_client *osdc = event->osdc;
1604
1605         dout("cancel_event %p\n", event);
1606         spin_lock(&osdc->event_lock);
1607         __remove_event(event);
1608         spin_unlock(&osdc->event_lock);
1609         ceph_osdc_put_event(event); /* caller's */
1610 }
1611 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1612
1613
1614 static void do_event_work(struct work_struct *work)
1615 {
1616         struct ceph_osd_event_work *event_work =
1617                 container_of(work, struct ceph_osd_event_work, work);
1618         struct ceph_osd_event *event = event_work->event;
1619         u64 ver = event_work->ver;
1620         u64 notify_id = event_work->notify_id;
1621         u8 opcode = event_work->opcode;
1622
1623         dout("do_event_work completing %p\n", event);
1624         event->cb(ver, notify_id, opcode, event->data);
1625         complete(&event->completion);
1626         dout("do_event_work completed %p\n", event);
1627         ceph_osdc_put_event(event);
1628         kfree(event_work);
1629 }
1630
1631
1632 /*
1633  * Process osd watch notifications
1634  */
1635 void handle_watch_notify(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1636 {
1637         void *p, *end;
1638         u8 proto_ver;
1639         u64 cookie, ver, notify_id;
1640         u8 opcode;
1641         struct ceph_osd_event *event;
1642         struct ceph_osd_event_work *event_work;
1643
1644         p = msg->front.iov_base;
1645         end = p + msg->front.iov_len;
1646
1647         ceph_decode_8_safe(&p, end, proto_ver, bad);
1648         ceph_decode_8_safe(&p, end, opcode, bad);
1649         ceph_decode_64_safe(&p, end, cookie, bad);
1650         ceph_decode_64_safe(&p, end, ver, bad);
1651         ceph_decode_64_safe(&p, end, notify_id, bad);
1652
1653         spin_lock(&osdc->event_lock);
1654         event = __find_event(osdc, cookie);
1655         if (event) {
1656                 get_event(event);
1657                 if (event->one_shot)
1658                         __remove_event(event);
1659         }
1660         spin_unlock(&osdc->event_lock);
1661         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1662              cookie, ver, event);
1663         if (event) {
1664                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1665                 if (!event_work) {
1666                         dout("ERROR: could not allocate event_work\n");
1667                         goto done_err;
1668                 }
1669                 INIT_WORK(&event_work->work, do_event_work);
1670                 event_work->event = event;
1671                 event_work->ver = ver;
1672                 event_work->notify_id = notify_id;
1673                 event_work->opcode = opcode;
1674                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1675                         dout("WARNING: failed to queue notify event work\n");
1676                         goto done_err;
1677                 }
1678         }
1679
1680         return;
1681
1682 done_err:
1683         complete(&event->completion);
1684         ceph_osdc_put_event(event);
1685         return;
1686
1687 bad:
1688         pr_err("osdc handle_watch_notify corrupt msg\n");
1689         return;
1690 }
1691
1692 int ceph_osdc_wait_event(struct ceph_osd_event *event, unsigned long timeout)
1693 {
1694         int err;
1695
1696         dout("wait_event %p\n", event);
1697         err = wait_for_completion_interruptible_timeout(&event->completion,
1698                                                         timeout * HZ);
1699         ceph_osdc_put_event(event);
1700         if (err > 0)
1701                 err = 0;
1702         dout("wait_event %p returns %d\n", event, err);
1703         return err;
1704 }
1705 EXPORT_SYMBOL(ceph_osdc_wait_event);
1706
1707 /*
1708  * Register request, send initial attempt.
1709  */
1710 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1711                             struct ceph_osd_request *req,
1712                             bool nofail)
1713 {
1714         int rc = 0;
1715
1716         req->r_request->pages = req->r_pages;
1717         req->r_request->nr_pages = req->r_num_pages;
1718 #ifdef CONFIG_BLOCK
1719         req->r_request->bio = req->r_bio;
1720 #endif
1721         req->r_request->trail = req->r_trail;
1722
1723         register_request(osdc, req);
1724
1725         down_read(&osdc->map_sem);
1726         mutex_lock(&osdc->request_mutex);
1727         /*
1728          * a racing kick_requests() may have sent the message for us
1729          * while we dropped request_mutex above, so only send now if
1730          * the request still han't been touched yet.
1731          */
1732         if (req->r_sent == 0) {
1733                 rc = __map_request(osdc, req, 0);
1734                 if (rc < 0) {
1735                         if (nofail) {
1736                                 dout("osdc_start_request failed map, "
1737                                      " will retry %lld\n", req->r_tid);
1738                                 rc = 0;
1739                         }
1740                         goto out_unlock;
1741                 }
1742                 if (req->r_osd == NULL) {
1743                         dout("send_request %p no up osds in pg\n", req);
1744                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1745                 } else {
1746                         __send_request(osdc, req);
1747                 }
1748                 rc = 0;
1749         }
1750
1751 out_unlock:
1752         mutex_unlock(&osdc->request_mutex);
1753         up_read(&osdc->map_sem);
1754         return rc;
1755 }
1756 EXPORT_SYMBOL(ceph_osdc_start_request);
1757
1758 /*
1759  * wait for a request to complete
1760  */
1761 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1762                            struct ceph_osd_request *req)
1763 {
1764         int rc;
1765
1766         rc = wait_for_completion_interruptible(&req->r_completion);
1767         if (rc < 0) {
1768                 mutex_lock(&osdc->request_mutex);
1769                 __cancel_request(req);
1770                 __unregister_request(osdc, req);
1771                 mutex_unlock(&osdc->request_mutex);
1772                 complete_request(req);
1773                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1774                 return rc;
1775         }
1776
1777         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1778         return req->r_result;
1779 }
1780 EXPORT_SYMBOL(ceph_osdc_wait_request);
1781
1782 /*
1783  * sync - wait for all in-flight requests to flush.  avoid starvation.
1784  */
1785 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1786 {
1787         struct ceph_osd_request *req;
1788         u64 last_tid, next_tid = 0;
1789
1790         mutex_lock(&osdc->request_mutex);
1791         last_tid = osdc->last_tid;
1792         while (1) {
1793                 req = __lookup_request_ge(osdc, next_tid);
1794                 if (!req)
1795                         break;
1796                 if (req->r_tid > last_tid)
1797                         break;
1798
1799                 next_tid = req->r_tid + 1;
1800                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1801                         continue;
1802
1803                 ceph_osdc_get_request(req);
1804                 mutex_unlock(&osdc->request_mutex);
1805                 dout("sync waiting on tid %llu (last is %llu)\n",
1806                      req->r_tid, last_tid);
1807                 wait_for_completion(&req->r_safe_completion);
1808                 mutex_lock(&osdc->request_mutex);
1809                 ceph_osdc_put_request(req);
1810         }
1811         mutex_unlock(&osdc->request_mutex);
1812         dout("sync done (thru tid %llu)\n", last_tid);
1813 }
1814 EXPORT_SYMBOL(ceph_osdc_sync);
1815
1816 /*
1817  * init, shutdown
1818  */
1819 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1820 {
1821         int err;
1822
1823         dout("init\n");
1824         osdc->client = client;
1825         osdc->osdmap = NULL;
1826         init_rwsem(&osdc->map_sem);
1827         init_completion(&osdc->map_waiters);
1828         osdc->last_requested_map = 0;
1829         mutex_init(&osdc->request_mutex);
1830         osdc->last_tid = 0;
1831         osdc->osds = RB_ROOT;
1832         INIT_LIST_HEAD(&osdc->osd_lru);
1833         osdc->requests = RB_ROOT;
1834         INIT_LIST_HEAD(&osdc->req_lru);
1835         INIT_LIST_HEAD(&osdc->req_unsent);
1836         INIT_LIST_HEAD(&osdc->req_notarget);
1837         INIT_LIST_HEAD(&osdc->req_linger);
1838         osdc->num_requests = 0;
1839         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1840         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1841         spin_lock_init(&osdc->event_lock);
1842         osdc->event_tree = RB_ROOT;
1843         osdc->event_count = 0;
1844
1845         schedule_delayed_work(&osdc->osds_timeout_work,
1846            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1847
1848         err = -ENOMEM;
1849         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1850                                         sizeof(struct ceph_osd_request));
1851         if (!osdc->req_mempool)
1852                 goto out;
1853
1854         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1855                                 OSD_OP_FRONT_LEN, 10, true,
1856                                 "osd_op");
1857         if (err < 0)
1858                 goto out_mempool;
1859         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1860                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1861                                 "osd_op_reply");
1862         if (err < 0)
1863                 goto out_msgpool;
1864
1865         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1866         if (IS_ERR(osdc->notify_wq)) {
1867                 err = PTR_ERR(osdc->notify_wq);
1868                 osdc->notify_wq = NULL;
1869                 goto out_msgpool;
1870         }
1871         return 0;
1872
1873 out_msgpool:
1874         ceph_msgpool_destroy(&osdc->msgpool_op);
1875 out_mempool:
1876         mempool_destroy(osdc->req_mempool);
1877 out:
1878         return err;
1879 }
1880 EXPORT_SYMBOL(ceph_osdc_init);
1881
1882 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1883 {
1884         flush_workqueue(osdc->notify_wq);
1885         destroy_workqueue(osdc->notify_wq);
1886         cancel_delayed_work_sync(&osdc->timeout_work);
1887         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1888         if (osdc->osdmap) {
1889                 ceph_osdmap_destroy(osdc->osdmap);
1890                 osdc->osdmap = NULL;
1891         }
1892         remove_all_osds(osdc);
1893         mempool_destroy(osdc->req_mempool);
1894         ceph_msgpool_destroy(&osdc->msgpool_op);
1895         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1896 }
1897 EXPORT_SYMBOL(ceph_osdc_stop);
1898
1899 /*
1900  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1901  * *plen.  Return number of bytes read, or error.
1902  */
1903 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1904                         struct ceph_vino vino, struct ceph_file_layout *layout,
1905                         u64 off, u64 *plen,
1906                         u32 truncate_seq, u64 truncate_size,
1907                         struct page **pages, int num_pages, int page_align)
1908 {
1909         struct ceph_osd_request *req;
1910         int rc = 0;
1911
1912         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1913              vino.snap, off, *plen);
1914         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1915                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1916                                     NULL, 0, truncate_seq, truncate_size, NULL,
1917                                     false, 1, page_align);
1918         if (IS_ERR(req))
1919                 return PTR_ERR(req);
1920
1921         /* it may be a short read due to an object boundary */
1922         req->r_pages = pages;
1923
1924         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1925              off, *plen, req->r_num_pages, page_align);
1926
1927         rc = ceph_osdc_start_request(osdc, req, false);
1928         if (!rc)
1929                 rc = ceph_osdc_wait_request(osdc, req);
1930
1931         ceph_osdc_put_request(req);
1932         dout("readpages result %d\n", rc);
1933         return rc;
1934 }
1935 EXPORT_SYMBOL(ceph_osdc_readpages);
1936
1937 /*
1938  * do a synchronous write on N pages
1939  */
1940 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1941                          struct ceph_file_layout *layout,
1942                          struct ceph_snap_context *snapc,
1943                          u64 off, u64 len,
1944                          u32 truncate_seq, u64 truncate_size,
1945                          struct timespec *mtime,
1946                          struct page **pages, int num_pages,
1947                          int flags, int do_sync, bool nofail)
1948 {
1949         struct ceph_osd_request *req;
1950         int rc = 0;
1951         int page_align = off & ~PAGE_MASK;
1952
1953         BUG_ON(vino.snap != CEPH_NOSNAP);
1954         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1955                                     CEPH_OSD_OP_WRITE,
1956                                     flags | CEPH_OSD_FLAG_ONDISK |
1957                                             CEPH_OSD_FLAG_WRITE,
1958                                     snapc, do_sync,
1959                                     truncate_seq, truncate_size, mtime,
1960                                     nofail, 1, page_align);
1961         if (IS_ERR(req))
1962                 return PTR_ERR(req);
1963
1964         /* it may be a short write due to an object boundary */
1965         req->r_pages = pages;
1966         dout("writepages %llu~%llu (%d pages)\n", off, len,
1967              req->r_num_pages);
1968
1969         rc = ceph_osdc_start_request(osdc, req, nofail);
1970         if (!rc)
1971                 rc = ceph_osdc_wait_request(osdc, req);
1972
1973         ceph_osdc_put_request(req);
1974         if (rc == 0)
1975                 rc = len;
1976         dout("writepages result %d\n", rc);
1977         return rc;
1978 }
1979 EXPORT_SYMBOL(ceph_osdc_writepages);
1980
1981 /*
1982  * handle incoming message
1983  */
1984 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1985 {
1986         struct ceph_osd *osd = con->private;
1987         struct ceph_osd_client *osdc;
1988         int type = le16_to_cpu(msg->hdr.type);
1989
1990         if (!osd)
1991                 goto out;
1992         osdc = osd->o_osdc;
1993
1994         switch (type) {
1995         case CEPH_MSG_OSD_MAP:
1996                 ceph_osdc_handle_map(osdc, msg);
1997                 break;
1998         case CEPH_MSG_OSD_OPREPLY:
1999                 handle_reply(osdc, msg, con);
2000                 break;
2001         case CEPH_MSG_WATCH_NOTIFY:
2002                 handle_watch_notify(osdc, msg);
2003                 break;
2004
2005         default:
2006                 pr_err("received unknown message type %d %s\n", type,
2007                        ceph_msg_type_name(type));
2008         }
2009 out:
2010         ceph_msg_put(msg);
2011 }
2012
2013 /*
2014  * lookup and return message for incoming reply.  set up reply message
2015  * pages.
2016  */
2017 static struct ceph_msg *get_reply(struct ceph_connection *con,
2018                                   struct ceph_msg_header *hdr,
2019                                   int *skip)
2020 {
2021         struct ceph_osd *osd = con->private;
2022         struct ceph_osd_client *osdc = osd->o_osdc;
2023         struct ceph_msg *m;
2024         struct ceph_osd_request *req;
2025         int front = le32_to_cpu(hdr->front_len);
2026         int data_len = le32_to_cpu(hdr->data_len);
2027         u64 tid;
2028
2029         tid = le64_to_cpu(hdr->tid);
2030         mutex_lock(&osdc->request_mutex);
2031         req = __lookup_request(osdc, tid);
2032         if (!req) {
2033                 *skip = 1;
2034                 m = NULL;
2035                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2036                      osd->o_osd);
2037                 goto out;
2038         }
2039
2040         if (req->r_con_filling_msg) {
2041                 dout("%s revoking msg %p from old con %p\n", __func__,
2042                      req->r_reply, req->r_con_filling_msg);
2043                 ceph_msg_revoke_incoming(req->r_reply);
2044                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2045                 req->r_con_filling_msg = NULL;
2046         }
2047
2048         if (front > req->r_reply->front.iov_len) {
2049                 pr_warning("get_reply front %d > preallocated %d\n",
2050                            front, (int)req->r_reply->front.iov_len);
2051                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2052                 if (!m)
2053                         goto out;
2054                 ceph_msg_put(req->r_reply);
2055                 req->r_reply = m;
2056         }
2057         m = ceph_msg_get(req->r_reply);
2058
2059         if (data_len > 0) {
2060                 int want = calc_pages_for(req->r_page_alignment, data_len);
2061
2062                 if (unlikely(req->r_num_pages < want)) {
2063                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2064                                    " had only %d pages ready\n", tid, data_len,
2065                                    want, req->r_num_pages);
2066                         *skip = 1;
2067                         ceph_msg_put(m);
2068                         m = NULL;
2069                         goto out;
2070                 }
2071                 m->pages = req->r_pages;
2072                 m->nr_pages = req->r_num_pages;
2073                 m->page_alignment = req->r_page_alignment;
2074 #ifdef CONFIG_BLOCK
2075                 m->bio = req->r_bio;
2076 #endif
2077         }
2078         *skip = 0;
2079         req->r_con_filling_msg = con->ops->get(con);
2080         dout("get_reply tid %lld %p\n", tid, m);
2081
2082 out:
2083         mutex_unlock(&osdc->request_mutex);
2084         return m;
2085
2086 }
2087
2088 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2089                                   struct ceph_msg_header *hdr,
2090                                   int *skip)
2091 {
2092         struct ceph_osd *osd = con->private;
2093         int type = le16_to_cpu(hdr->type);
2094         int front = le32_to_cpu(hdr->front_len);
2095
2096         *skip = 0;
2097         switch (type) {
2098         case CEPH_MSG_OSD_MAP:
2099         case CEPH_MSG_WATCH_NOTIFY:
2100                 return ceph_msg_new(type, front, GFP_NOFS, false);
2101         case CEPH_MSG_OSD_OPREPLY:
2102                 return get_reply(con, hdr, skip);
2103         default:
2104                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2105                         osd->o_osd);
2106                 *skip = 1;
2107                 return NULL;
2108         }
2109 }
2110
2111 /*
2112  * Wrappers to refcount containing ceph_osd struct
2113  */
2114 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2115 {
2116         struct ceph_osd *osd = con->private;
2117         if (get_osd(osd))
2118                 return con;
2119         return NULL;
2120 }
2121
2122 static void put_osd_con(struct ceph_connection *con)
2123 {
2124         struct ceph_osd *osd = con->private;
2125         put_osd(osd);
2126 }
2127
2128 /*
2129  * authentication
2130  */
2131 /*
2132  * Note: returned pointer is the address of a structure that's
2133  * managed separately.  Caller must *not* attempt to free it.
2134  */
2135 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2136                                         int *proto, int force_new)
2137 {
2138         struct ceph_osd *o = con->private;
2139         struct ceph_osd_client *osdc = o->o_osdc;
2140         struct ceph_auth_client *ac = osdc->client->monc.auth;
2141         struct ceph_auth_handshake *auth = &o->o_auth;
2142
2143         if (force_new && auth->authorizer) {
2144                 if (ac->ops && ac->ops->destroy_authorizer)
2145                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2146                 auth->authorizer = NULL;
2147         }
2148         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2149                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2150                                                         auth);
2151                 if (ret)
2152                         return ERR_PTR(ret);
2153         }
2154         *proto = ac->protocol;
2155
2156         return auth;
2157 }
2158
2159
2160 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2161 {
2162         struct ceph_osd *o = con->private;
2163         struct ceph_osd_client *osdc = o->o_osdc;
2164         struct ceph_auth_client *ac = osdc->client->monc.auth;
2165
2166         /*
2167          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2168          * XXX which do we do:  succeed or fail?
2169          */
2170         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2171 }
2172
2173 static int invalidate_authorizer(struct ceph_connection *con)
2174 {
2175         struct ceph_osd *o = con->private;
2176         struct ceph_osd_client *osdc = o->o_osdc;
2177         struct ceph_auth_client *ac = osdc->client->monc.auth;
2178
2179         if (ac->ops && ac->ops->invalidate_authorizer)
2180                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2181
2182         return ceph_monc_validate_auth(&osdc->client->monc);
2183 }
2184
2185 static const struct ceph_connection_operations osd_con_ops = {
2186         .get = get_osd_con,
2187         .put = put_osd_con,
2188         .dispatch = dispatch,
2189         .get_authorizer = get_authorizer,
2190         .verify_authorizer_reply = verify_authorizer_reply,
2191         .invalidate_authorizer = invalidate_authorizer,
2192         .alloc_msg = alloc_msg,
2193         .fault = osd_reset,
2194 };