net/ceph/osd_client.c

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