net/ceph/osd_client.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/module.h>
   4 #include <linux/err.h>
   5 #include <linux/highmem.h>
   6 #include <linux/mm.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/slab.h>
   9 #include <linux/uaccess.h>
  10 #ifdef CONFIG_BLOCK
  11 #include <linux/bio.h>
  12 #endif
  13
  14 #include <linux/ceph/libceph.h>
  15 #include <linux/ceph/osd_client.h>
  16 #include <linux/ceph/messenger.h>
  17 #include <linux/ceph/decode.h>
  18 #include <linux/ceph/auth.h>
  19 #include <linux/ceph/pagelist.h>
  20
  21 #define OSD_OP_FRONT_LEN        4096
  22 #define OSD_OPREPLY_FRONT_LEN   512
  23
  24 static const struct ceph_connection_operations osd_con_ops;
  25
  26 static void __send_queued(struct ceph_osd_client *osdc);
  27 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  28 static void __register_request(struct ceph_osd_client *osdc,
  29                                struct ceph_osd_request *req);
  30 static void __unregister_linger_request(struct ceph_osd_client *osdc,
  31                                         struct ceph_osd_request *req);
  32 static void __send_request(struct ceph_osd_client *osdc,
  33                            struct ceph_osd_request *req);
  34
  35 static int op_has_extent(int op)
  36 {
  37         return (op == CEPH_OSD_OP_READ ||
  38                 op == CEPH_OSD_OP_WRITE);
  39 }
  40
  41 /*
  42  * Implement client access to distributed object storage cluster.
  43  *
  44  * All data objects are stored within a cluster/cloud of OSDs, or
  45  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  46  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  47  * remote daemons serving up and coordinating consistent and safe
  48  * access to storage.
  49  *
  50  * Cluster membership and the mapping of data objects onto storage devices
  51  * are described by the osd map.
  52  *
  53  * We keep track of pending OSD requests (read, write), resubmit
  54  * requests to different OSDs when the cluster topology/data layout
  55  * change, or retry the affected requests when the communications
  56  * channel with an OSD is reset.
  57  */
  58
  59 /*
  60  * calculate the mapping of a file extent onto an object, and fill out the
  61  * request accordingly.  shorten extent as necessary if it crosses an
  62  * object boundary.
  63  *
  64  * fill osd op in request message.
  65  */
  66 static int calc_layout(struct ceph_file_layout *layout, u64 off, u64 *plen,
  67                        struct ceph_osd_req_op *op, u64 *bno)
  68 {
  69         u64 orig_len = *plen;
  70         u64 objoff = 0;
  71         u64 objlen = 0;
  72         int r;
  73
  74         /* object extent? */
  75         r = ceph_calc_file_object_mapping(layout, off, orig_len, bno,
  76                                           &objoff, &objlen);
  77         if (r < 0)
  78                 return r;
  79         if (objlen < orig_len) {
  80                 *plen = objlen;
  81                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  82                      orig_len - *plen, off, *plen);
  83         }
  84
  85         if (op_has_extent(op->op)) {
  86                 u32 osize = le32_to_cpu(layout->fl_object_size);
  87                 op->extent.offset = objoff;
  88                 op->extent.length = objlen;
  89                 if (op->extent.truncate_size <= off - objoff) {
  90                         op->extent.truncate_size = 0;
  91                 } else {
  92                         op->extent.truncate_size -= off - objoff;
  93                         if (op->extent.truncate_size > osize)
  94                                 op->extent.truncate_size = osize;
  95                 }
  96         }
  97         if (op->op == CEPH_OSD_OP_WRITE)
  98                 op->payload_len = *plen;
  99
 100         dout("calc_layout bno=%llx %llu~%llu\n", *bno, objoff, objlen);
 101
 102         return 0;
 103 }
 104
 105 /*
 106  * requests
 107  */
 108 void ceph_osdc_release_request(struct kref *kref)
 109 {
 110         struct ceph_osd_request *req = container_of(kref,
 111                                                     struct ceph_osd_request,
 112                                                     r_kref);
 113
 114         if (req->r_request)
 115                 ceph_msg_put(req->r_request);
 116         if (req->r_con_filling_msg) {
 117                 dout("%s revoking msg %p from con %p\n", __func__,
 118                      req->r_reply, req->r_con_filling_msg);
 119                 ceph_msg_revoke_incoming(req->r_reply);
 120                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
 121                 req->r_con_filling_msg = NULL;
 122         }
 123         if (req->r_reply)
 124                 ceph_msg_put(req->r_reply);
 125         if (req->r_own_pages)
 126                 ceph_release_page_vector(req->r_pages,
 127                                          req->r_num_pages);
 128         ceph_put_snap_context(req->r_snapc);
 129         ceph_pagelist_release(&req->r_trail);
 130         if (req->r_mempool)
 131                 mempool_free(req, req->r_osdc->req_mempool);
 132         else
 133                 kfree(req);
 134 }
 135 EXPORT_SYMBOL(ceph_osdc_release_request);
 136
 137 struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
 138                                                struct ceph_snap_context *snapc,
 139                                                unsigned int num_ops,
 140                                                bool use_mempool,
 141                                                gfp_t gfp_flags)
 142 {
 143         struct ceph_osd_request *req;
 144         struct ceph_msg *msg;
 145         size_t msg_size;
 146
 147         msg_size = 4 + 4 + 8 + 8 + 4+8;
 148         msg_size += 2 + 4 + 8 + 4 + 4; /* oloc */
 149         msg_size += 1 + 8 + 4 + 4;     /* pg_t */
 150         msg_size += 4 + MAX_OBJ_NAME_SIZE;
 151         msg_size += 2 + num_ops*sizeof(struct ceph_osd_op);
 152         msg_size += 8;  /* snapid */
 153         msg_size += 8;  /* snap_seq */
 154         msg_size += 8 * (snapc ? snapc->num_snaps : 0);  /* snaps */
 155         msg_size += 4;
 156
 157         if (use_mempool) {
 158                 req = mempool_alloc(osdc->req_mempool, gfp_flags);
 159                 memset(req, 0, sizeof(*req));
 160         } else {
 161                 req = kzalloc(sizeof(*req), gfp_flags);
 162         }
 163         if (req == NULL)
 164                 return NULL;
 165
 166         req->r_osdc = osdc;
 167         req->r_mempool = use_mempool;
 168
 169         kref_init(&req->r_kref);
 170         init_completion(&req->r_completion);
 171         init_completion(&req->r_safe_completion);
 172         RB_CLEAR_NODE(&req->r_node);
 173         INIT_LIST_HEAD(&req->r_unsafe_item);
 174         INIT_LIST_HEAD(&req->r_linger_item);
 175         INIT_LIST_HEAD(&req->r_linger_osd);
 176         INIT_LIST_HEAD(&req->r_req_lru_item);
 177         INIT_LIST_HEAD(&req->r_osd_item);
 178
 179         /* create reply message */
 180         if (use_mempool)
 181                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 182         else
 183                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 184                                    OSD_OPREPLY_FRONT_LEN, gfp_flags, true);
 185         if (!msg) {
 186                 ceph_osdc_put_request(req);
 187                 return NULL;
 188         }
 189         req->r_reply = msg;
 190
 191         ceph_pagelist_init(&req->r_trail);
 192
 193         /* create request message; allow space for oid */
 194         if (use_mempool)
 195                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 196         else
 197                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, gfp_flags, true);
 198         if (!msg) {
 199                 ceph_osdc_put_request(req);
 200                 return NULL;
 201         }
 202
 203         memset(msg->front.iov_base, 0, msg->front.iov_len);
 204
 205         req->r_request = msg;
 206
 207         return req;
 208 }
 209 EXPORT_SYMBOL(ceph_osdc_alloc_request);
 210
 211 static void osd_req_encode_op(struct ceph_osd_request *req,
 212                               struct ceph_osd_op *dst,
 213                               struct ceph_osd_req_op *src)
 214 {
 215         dst->op = cpu_to_le16(src->op);
 216
 217         switch (src->op) {
 218         case CEPH_OSD_OP_STAT:
 219                 break;
 220         case CEPH_OSD_OP_READ:
 221         case CEPH_OSD_OP_WRITE:
 222                 dst->extent.offset =
 223                         cpu_to_le64(src->extent.offset);
 224                 dst->extent.length =
 225                         cpu_to_le64(src->extent.length);
 226                 dst->extent.truncate_size =
 227                         cpu_to_le64(src->extent.truncate_size);
 228                 dst->extent.truncate_seq =
 229                         cpu_to_le32(src->extent.truncate_seq);
 230                 break;
 231         case CEPH_OSD_OP_CALL:
 232                 dst->cls.class_len = src->cls.class_len;
 233                 dst->cls.method_len = src->cls.method_len;
 234                 dst->cls.indata_len = cpu_to_le32(src->cls.indata_len);
 235
 236                 ceph_pagelist_append(&req->r_trail, src->cls.class_name,
 237                                      src->cls.class_len);
 238                 ceph_pagelist_append(&req->r_trail, src->cls.method_name,
 239                                      src->cls.method_len);
 240                 ceph_pagelist_append(&req->r_trail, src->cls.indata,
 241                                      src->cls.indata_len);
 242                 break;
 243         case CEPH_OSD_OP_STARTSYNC:
 244                 break;
 245         case CEPH_OSD_OP_NOTIFY_ACK:
 246         case CEPH_OSD_OP_WATCH:
 247                 dst->watch.cookie = cpu_to_le64(src->watch.cookie);
 248                 dst->watch.ver = cpu_to_le64(src->watch.ver);
 249                 dst->watch.flag = src->watch.flag;
 250                 break;
 251         default:
 252                 pr_err("unrecognized osd opcode %d\n", dst->op);
 253                 WARN_ON(1);
 254                 break;
 255         case CEPH_OSD_OP_MAPEXT:
 256         case CEPH_OSD_OP_MASKTRUNC:
 257         case CEPH_OSD_OP_SPARSE_READ:
 258         case CEPH_OSD_OP_NOTIFY:
 259         case CEPH_OSD_OP_ASSERT_VER:
 260         case CEPH_OSD_OP_WRITEFULL:
 261         case CEPH_OSD_OP_TRUNCATE:
 262         case CEPH_OSD_OP_ZERO:
 263         case CEPH_OSD_OP_DELETE:
 264         case CEPH_OSD_OP_APPEND:
 265         case CEPH_OSD_OP_SETTRUNC:
 266         case CEPH_OSD_OP_TRIMTRUNC:
 267         case CEPH_OSD_OP_TMAPUP:
 268         case CEPH_OSD_OP_TMAPPUT:
 269         case CEPH_OSD_OP_TMAPGET:
 270         case CEPH_OSD_OP_CREATE:
 271         case CEPH_OSD_OP_ROLLBACK:
 272         case CEPH_OSD_OP_OMAPGETKEYS:
 273         case CEPH_OSD_OP_OMAPGETVALS:
 274         case CEPH_OSD_OP_OMAPGETHEADER:
 275         case CEPH_OSD_OP_OMAPGETVALSBYKEYS:
 276         case CEPH_OSD_OP_MODE_RD:
 277         case CEPH_OSD_OP_OMAPSETVALS:
 278         case CEPH_OSD_OP_OMAPSETHEADER:
 279         case CEPH_OSD_OP_OMAPCLEAR:
 280         case CEPH_OSD_OP_OMAPRMKEYS:
 281         case CEPH_OSD_OP_OMAP_CMP:
 282         case CEPH_OSD_OP_CLONERANGE:
 283         case CEPH_OSD_OP_ASSERT_SRC_VERSION:
 284         case CEPH_OSD_OP_SRC_CMPXATTR:
 285         case CEPH_OSD_OP_GETXATTR:
 286         case CEPH_OSD_OP_GETXATTRS:
 287         case CEPH_OSD_OP_CMPXATTR:
 288         case CEPH_OSD_OP_SETXATTR:
 289         case CEPH_OSD_OP_SETXATTRS:
 290         case CEPH_OSD_OP_RESETXATTRS:
 291         case CEPH_OSD_OP_RMXATTR:
 292         case CEPH_OSD_OP_PULL:
 293         case CEPH_OSD_OP_PUSH:
 294         case CEPH_OSD_OP_BALANCEREADS:
 295         case CEPH_OSD_OP_UNBALANCEREADS:
 296         case CEPH_OSD_OP_SCRUB:
 297         case CEPH_OSD_OP_SCRUB_RESERVE:
 298         case CEPH_OSD_OP_SCRUB_UNRESERVE:
 299         case CEPH_OSD_OP_SCRUB_STOP:
 300         case CEPH_OSD_OP_SCRUB_MAP:
 301         case CEPH_OSD_OP_WRLOCK:
 302         case CEPH_OSD_OP_WRUNLOCK:
 303         case CEPH_OSD_OP_RDLOCK:
 304         case CEPH_OSD_OP_RDUNLOCK:
 305         case CEPH_OSD_OP_UPLOCK:
 306         case CEPH_OSD_OP_DNLOCK:
 307         case CEPH_OSD_OP_PGLS:
 308         case CEPH_OSD_OP_PGLS_FILTER:
 309                 pr_err("unsupported osd opcode %s\n",
 310                         ceph_osd_op_name(dst->op));
 311                 WARN_ON(1);
 312                 break;
 313         }
 314         dst->payload_len = cpu_to_le32(src->payload_len);
 315 }
 316
 317 /*
 318  * build new request AND message
 319  *
 320  */
 321 void ceph_osdc_build_request(struct ceph_osd_request *req,
 322                              u64 off, u64 len, unsigned int num_ops,
 323                              struct ceph_osd_req_op *src_ops,
 324                              struct ceph_snap_context *snapc, u64 snap_id,
 325                              struct timespec *mtime)
 326 {
 327         struct ceph_msg *msg = req->r_request;
 328         struct ceph_osd_req_op *src_op;
 329         void *p;
 330         size_t msg_size;
 331         int flags = req->r_flags;
 332         u64 data_len;
 333         int i;
 334
 335         req->r_num_ops = num_ops;
 336         req->r_snapid = snap_id;
 337         req->r_snapc = ceph_get_snap_context(snapc);
 338
 339         /* encode request */
 340         msg->hdr.version = cpu_to_le16(4);
 341
 342         p = msg->front.iov_base;
 343         ceph_encode_32(&p, 1);   /* client_inc  is always 1 */
 344         req->r_request_osdmap_epoch = p;
 345         p += 4;
 346         req->r_request_flags = p;
 347         p += 4;
 348         if (req->r_flags & CEPH_OSD_FLAG_WRITE)
 349                 ceph_encode_timespec(p, mtime);
 350         p += sizeof(struct ceph_timespec);
 351         req->r_request_reassert_version = p;
 352         p += sizeof(struct ceph_eversion); /* will get filled in */
 353
 354         /* oloc */
 355         ceph_encode_8(&p, 4);
 356         ceph_encode_8(&p, 4);
 357         ceph_encode_32(&p, 8 + 4 + 4);
 358         req->r_request_pool = p;
 359         p += 8;
 360         ceph_encode_32(&p, -1);  /* preferred */
 361         ceph_encode_32(&p, 0);   /* key len */
 362
 363         ceph_encode_8(&p, 1);
 364         req->r_request_pgid = p;
 365         p += 8 + 4;
 366         ceph_encode_32(&p, -1);  /* preferred */
 367
 368         /* oid */
 369         ceph_encode_32(&p, req->r_oid_len);
 370         memcpy(p, req->r_oid, req->r_oid_len);
 371         dout("oid '%.*s' len %d\n", req->r_oid_len, req->r_oid, req->r_oid_len);
 372         p += req->r_oid_len;
 373
 374         /* ops */
 375         ceph_encode_16(&p, num_ops);
 376         src_op = src_ops;
 377         req->r_request_ops = p;
 378         for (i = 0; i < num_ops; i++, src_op++) {
 379                 osd_req_encode_op(req, p, src_op);
 380                 p += sizeof(struct ceph_osd_op);
 381         }
 382
 383         /* snaps */
 384         ceph_encode_64(&p, req->r_snapid);
 385         ceph_encode_64(&p, req->r_snapc ? req->r_snapc->seq : 0);
 386         ceph_encode_32(&p, req->r_snapc ? req->r_snapc->num_snaps : 0);
 387         if (req->r_snapc) {
 388                 for (i = 0; i < snapc->num_snaps; i++) {
 389                         ceph_encode_64(&p, req->r_snapc->snaps[i]);
 390                 }
 391         }
 392
 393         req->r_request_attempts = p;
 394         p += 4;
 395
 396         data_len = req->r_trail.length;
 397         if (flags & CEPH_OSD_FLAG_WRITE) {
 398                 req->r_request->hdr.data_off = cpu_to_le16(off);
 399                 data_len += len;
 400         }
 401         req->r_request->hdr.data_len = cpu_to_le32(data_len);
 402
 403         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 404         msg_size = p - msg->front.iov_base;
 405         msg->front.iov_len = msg_size;
 406         msg->hdr.front_len = cpu_to_le32(msg_size);
 407
 408         dout("build_request msg_size was %d num_ops %d\n", (int)msg_size,
 409              num_ops);
 410         return;
 411 }
 412 EXPORT_SYMBOL(ceph_osdc_build_request);
 413
 414 /*
 415  * build new request AND message, calculate layout, and adjust file
 416  * extent as needed.
 417  *
 418  * if the file was recently truncated, we include information about its
 419  * old and new size so that the object can be updated appropriately.  (we
 420  * avoid synchronously deleting truncated objects because it's slow.)
 421  *
 422  * if @do_sync, include a 'startsync' command so that the osd will flush
 423  * data quickly.
 424  */
 425 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 426                                                struct ceph_file_layout *layout,
 427                                                struct ceph_vino vino,
 428                                                u64 off, u64 *plen,
 429                                                int opcode, int flags,
 430                                                struct ceph_snap_context *snapc,
 431                                                int do_sync,
 432                                                u32 truncate_seq,
 433                                                u64 truncate_size,
 434                                                struct timespec *mtime,
 435                                                bool use_mempool,
 436                                                int page_align)
 437 {
 438         struct ceph_osd_req_op ops[2];
 439         struct ceph_osd_request *req;
 440         unsigned int num_op = 1;
 441         u64 bno = 0;
 442         int r;
 443
 444         memset(&ops, 0, sizeof ops);
 445
 446         ops[0].op = opcode;
 447         ops[0].extent.truncate_seq = truncate_seq;
 448         ops[0].extent.truncate_size = truncate_size;
 449
 450         if (do_sync) {
 451                 ops[1].op = CEPH_OSD_OP_STARTSYNC;
 452                 num_op++;
 453         }
 454
 455         req = ceph_osdc_alloc_request(osdc, snapc, num_op, use_mempool,
 456                                         GFP_NOFS);
 457         if (!req)
 458                 return ERR_PTR(-ENOMEM);
 459         req->r_flags = flags;
 460
 461         /* calculate max write size */
 462         r = calc_layout(layout, off, plen, ops, &bno);
 463         if (r < 0) {
 464                 ceph_osdc_put_request(req);
 465                 return ERR_PTR(r);
 466         }
 467
 468         req->r_file_layout = *layout;  /* keep a copy */
 469
 470         snprintf(req->r_oid, sizeof(req->r_oid), "%llx.%08llx", vino.ino, bno);
 471         req->r_oid_len = strlen(req->r_oid);
 472
 473         /* The alignment may differ from the natural (file) alignment */
 474
 475         req->r_num_pages = calc_pages_for(page_align, *plen);
 476         req->r_page_alignment = page_align;
 477
 478         ceph_osdc_build_request(req, off, *plen, num_op, ops,
 479                                 snapc, vino.snap, mtime);
 480
 481         return req;
 482 }
 483 EXPORT_SYMBOL(ceph_osdc_new_request);
 484
 485 /*
 486  * We keep osd requests in an rbtree, sorted by ->r_tid.
 487  */
 488 static void __insert_request(struct ceph_osd_client *osdc,
 489                              struct ceph_osd_request *new)
 490 {
 491         struct rb_node **p = &osdc->requests.rb_node;
 492         struct rb_node *parent = NULL;
 493         struct ceph_osd_request *req = NULL;
 494
 495         while (*p) {
 496                 parent = *p;
 497                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 498                 if (new->r_tid < req->r_tid)
 499                         p = &(*p)->rb_left;
 500                 else if (new->r_tid > req->r_tid)
 501                         p = &(*p)->rb_right;
 502                 else
 503                         BUG();
 504         }
 505
 506         rb_link_node(&new->r_node, parent, p);
 507         rb_insert_color(&new->r_node, &osdc->requests);
 508 }
 509
 510 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 511                                                  u64 tid)
 512 {
 513         struct ceph_osd_request *req;
 514         struct rb_node *n = osdc->requests.rb_node;
 515
 516         while (n) {
 517                 req = rb_entry(n, struct ceph_osd_request, r_node);
 518                 if (tid < req->r_tid)
 519                         n = n->rb_left;
 520                 else if (tid > req->r_tid)
 521                         n = n->rb_right;
 522                 else
 523                         return req;
 524         }
 525         return NULL;
 526 }
 527
 528 static struct ceph_osd_request *
 529 __lookup_request_ge(struct ceph_osd_client *osdc,
 530                     u64 tid)
 531 {
 532         struct ceph_osd_request *req;
 533         struct rb_node *n = osdc->requests.rb_node;
 534
 535         while (n) {
 536                 req = rb_entry(n, struct ceph_osd_request, r_node);
 537                 if (tid < req->r_tid) {
 538                         if (!n->rb_left)
 539                                 return req;
 540                         n = n->rb_left;
 541                 } else if (tid > req->r_tid) {
 542                         n = n->rb_right;
 543                 } else {
 544                         return req;
 545                 }
 546         }
 547         return NULL;
 548 }
 549
 550 /*
 551  * Resubmit requests pending on the given osd.
 552  */
 553 static void __kick_osd_requests(struct ceph_osd_client *osdc,
 554                                 struct ceph_osd *osd)
 555 {
 556         struct ceph_osd_request *req, *nreq;
 557         int err;
 558
 559         dout("__kick_osd_requests osd%d\n", osd->o_osd);
 560         err = __reset_osd(osdc, osd);
 561         if (err)
 562                 return;
 563
 564         list_for_each_entry(req, &osd->o_requests, r_osd_item) {
 565                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
 566                 dout("requeued %p tid %llu osd%d\n", req, req->r_tid,
 567                      osd->o_osd);
 568                 if (!req->r_linger)
 569                         req->r_flags |= CEPH_OSD_FLAG_RETRY;
 570         }
 571
 572         list_for_each_entry_safe(req, nreq, &osd->o_linger_requests,
 573                                  r_linger_osd) {
 574                 /*
 575                  * reregister request prior to unregistering linger so
 576                  * that r_osd is preserved.
 577                  */
 578                 BUG_ON(!list_empty(&req->r_req_lru_item));
 579                 __register_request(osdc, req);
 580                 list_add(&req->r_req_lru_item, &osdc->req_unsent);
 581                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 582                 __unregister_linger_request(osdc, req);
 583                 dout("requeued lingering %p tid %llu osd%d\n", req, req->r_tid,
 584                      osd->o_osd);
 585         }
 586 }
 587
 588 /*
 589  * If the osd connection drops, we need to resubmit all requests.
 590  */
 591 static void osd_reset(struct ceph_connection *con)
 592 {
 593         struct ceph_osd *osd = con->private;
 594         struct ceph_osd_client *osdc;
 595
 596         if (!osd)
 597                 return;
 598         dout("osd_reset osd%d\n", osd->o_osd);
 599         osdc = osd->o_osdc;
 600         down_read(&osdc->map_sem);
 601         mutex_lock(&osdc->request_mutex);
 602         __kick_osd_requests(osdc, osd);
 603         __send_queued(osdc);
 604         mutex_unlock(&osdc->request_mutex);
 605         up_read(&osdc->map_sem);
 606 }
 607
 608 /*
 609  * Track open sessions with osds.
 610  */
 611 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc, int onum)
 612 {
 613         struct ceph_osd *osd;
 614
 615         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 616         if (!osd)
 617                 return NULL;
 618
 619         atomic_set(&osd->o_ref, 1);
 620         osd->o_osdc = osdc;
 621         osd->o_osd = onum;
 622         RB_CLEAR_NODE(&osd->o_node);
 623         INIT_LIST_HEAD(&osd->o_requests);
 624         INIT_LIST_HEAD(&osd->o_linger_requests);
 625         INIT_LIST_HEAD(&osd->o_osd_lru);
 626         osd->o_incarnation = 1;
 627
 628         ceph_con_init(&osd->o_con, osd, &osd_con_ops, &osdc->client->msgr);
 629
 630         INIT_LIST_HEAD(&osd->o_keepalive_item);
 631         return osd;
 632 }
 633
 634 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 635 {
 636         if (atomic_inc_not_zero(&osd->o_ref)) {
 637                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 638                      atomic_read(&osd->o_ref));
 639                 return osd;
 640         } else {
 641                 dout("get_osd %p FAIL\n", osd);
 642                 return NULL;
 643         }
 644 }
 645
 646 static void put_osd(struct ceph_osd *osd)
 647 {
 648         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 649              atomic_read(&osd->o_ref) - 1);
 650         if (atomic_dec_and_test(&osd->o_ref) && osd->o_auth.authorizer) {
 651                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 652
 653                 if (ac->ops && ac->ops->destroy_authorizer)
 654                         ac->ops->destroy_authorizer(ac, osd->o_auth.authorizer);
 655                 kfree(osd);
 656         }
 657 }
 658
 659 /*
 660  * remove an osd from our map
 661  */
 662 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 663 {
 664         dout("__remove_osd %p\n", osd);
 665         BUG_ON(!list_empty(&osd->o_requests));
 666         rb_erase(&osd->o_node, &osdc->osds);
 667         list_del_init(&osd->o_osd_lru);
 668         ceph_con_close(&osd->o_con);
 669         put_osd(osd);
 670 }
 671
 672 static void remove_all_osds(struct ceph_osd_client *osdc)
 673 {
 674         dout("%s %p\n", __func__, osdc);
 675         mutex_lock(&osdc->request_mutex);
 676         while (!RB_EMPTY_ROOT(&osdc->osds)) {
 677                 struct ceph_osd *osd = rb_entry(rb_first(&osdc->osds),
 678                                                 struct ceph_osd, o_node);
 679                 __remove_osd(osdc, osd);
 680         }
 681         mutex_unlock(&osdc->request_mutex);
 682 }
 683
 684 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 685                               struct ceph_osd *osd)
 686 {
 687         dout("__move_osd_to_lru %p\n", osd);
 688         BUG_ON(!list_empty(&osd->o_osd_lru));
 689         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 690         osd->lru_ttl = jiffies + osdc->client->options->osd_idle_ttl * HZ;
 691 }
 692
 693 static void __remove_osd_from_lru(struct ceph_osd *osd)
 694 {
 695         dout("__remove_osd_from_lru %p\n", osd);
 696         if (!list_empty(&osd->o_osd_lru))
 697                 list_del_init(&osd->o_osd_lru);
 698 }
 699
 700 static void remove_old_osds(struct ceph_osd_client *osdc)
 701 {
 702         struct ceph_osd *osd, *nosd;
 703
 704         dout("__remove_old_osds %p\n", osdc);
 705         mutex_lock(&osdc->request_mutex);
 706         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 707                 if (time_before(jiffies, osd->lru_ttl))
 708                         break;
 709                 __remove_osd(osdc, osd);
 710         }
 711         mutex_unlock(&osdc->request_mutex);
 712 }
 713
 714 /*
 715  * reset osd connect
 716  */
 717 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 718 {
 719         struct ceph_entity_addr *peer_addr;
 720
 721         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 722         if (list_empty(&osd->o_requests) &&
 723             list_empty(&osd->o_linger_requests)) {
 724                 __remove_osd(osdc, osd);
 725
 726                 return -ENODEV;
 727         }
 728
 729         peer_addr = &osdc->osdmap->osd_addr[osd->o_osd];
 730         if (!memcmp(peer_addr, &osd->o_con.peer_addr, sizeof (*peer_addr)) &&
 731                         !ceph_con_opened(&osd->o_con)) {
 732                 struct ceph_osd_request *req;
 733
 734                 dout(" osd addr hasn't changed and connection never opened,"
 735                      " letting msgr retry");
 736                 /* touch each r_stamp for handle_timeout()'s benfit */
 737                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 738                         req->r_stamp = jiffies;
 739
 740                 return -EAGAIN;
 741         }
 742
 743         ceph_con_close(&osd->o_con);
 744         ceph_con_open(&osd->o_con, CEPH_ENTITY_TYPE_OSD, osd->o_osd, peer_addr);
 745         osd->o_incarnation++;
 746
 747         return 0;
 748 }
 749
 750 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 751 {
 752         struct rb_node **p = &osdc->osds.rb_node;
 753         struct rb_node *parent = NULL;
 754         struct ceph_osd *osd = NULL;
 755
 756         dout("__insert_osd %p osd%d\n", new, new->o_osd);
 757         while (*p) {
 758                 parent = *p;
 759                 osd = rb_entry(parent, struct ceph_osd, o_node);
 760                 if (new->o_osd < osd->o_osd)
 761                         p = &(*p)->rb_left;
 762                 else if (new->o_osd > osd->o_osd)
 763                         p = &(*p)->rb_right;
 764                 else
 765                         BUG();
 766         }
 767
 768         rb_link_node(&new->o_node, parent, p);
 769         rb_insert_color(&new->o_node, &osdc->osds);
 770 }
 771
 772 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 773 {
 774         struct ceph_osd *osd;
 775         struct rb_node *n = osdc->osds.rb_node;
 776
 777         while (n) {
 778                 osd = rb_entry(n, struct ceph_osd, o_node);
 779                 if (o < osd->o_osd)
 780                         n = n->rb_left;
 781                 else if (o > osd->o_osd)
 782                         n = n->rb_right;
 783                 else
 784                         return osd;
 785         }
 786         return NULL;
 787 }
 788
 789 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 790 {
 791         schedule_delayed_work(&osdc->timeout_work,
 792                         osdc->client->options->osd_keepalive_timeout * HZ);
 793 }
 794
 795 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 796 {
 797         cancel_delayed_work(&osdc->timeout_work);
 798 }
 799
 800 /*
 801  * Register request, assign tid.  If this is the first request, set up
 802  * the timeout event.
 803  */
 804 static void __register_request(struct ceph_osd_client *osdc,
 805                                struct ceph_osd_request *req)
 806 {
 807         req->r_tid = ++osdc->last_tid;
 808         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 809         dout("__register_request %p tid %lld\n", req, req->r_tid);
 810         __insert_request(osdc, req);
 811         ceph_osdc_get_request(req);
 812         osdc->num_requests++;
 813         if (osdc->num_requests == 1) {
 814                 dout(" first request, scheduling timeout\n");
 815                 __schedule_osd_timeout(osdc);
 816         }
 817 }
 818
 819 static void register_request(struct ceph_osd_client *osdc,
 820                              struct ceph_osd_request *req)
 821 {
 822         mutex_lock(&osdc->request_mutex);
 823         __register_request(osdc, req);
 824         mutex_unlock(&osdc->request_mutex);
 825 }
 826
 827 /*
 828  * called under osdc->request_mutex
 829  */
 830 static void __unregister_request(struct ceph_osd_client *osdc,
 831                                  struct ceph_osd_request *req)
 832 {
 833         if (RB_EMPTY_NODE(&req->r_node)) {
 834                 dout("__unregister_request %p tid %lld not registered\n",
 835                         req, req->r_tid);
 836                 return;
 837         }
 838
 839         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 840         rb_erase(&req->r_node, &osdc->requests);
 841         osdc->num_requests--;
 842
 843         if (req->r_osd) {
 844                 /* make sure the original request isn't in flight. */
 845                 ceph_msg_revoke(req->r_request);
 846
 847                 list_del_init(&req->r_osd_item);
 848                 if (list_empty(&req->r_osd->o_requests) &&
 849                     list_empty(&req->r_osd->o_linger_requests)) {
 850                         dout("moving osd to %p lru\n", req->r_osd);
 851                         __move_osd_to_lru(osdc, req->r_osd);
 852                 }
 853                 if (list_empty(&req->r_linger_item))
 854                         req->r_osd = NULL;
 855         }
 856
 857         list_del_init(&req->r_req_lru_item);
 858         ceph_osdc_put_request(req);
 859
 860         if (osdc->num_requests == 0) {
 861                 dout(" no requests, canceling timeout\n");
 862                 __cancel_osd_timeout(osdc);
 863         }
 864 }
 865
 866 /*
 867  * Cancel a previously queued request message
 868  */
 869 static void __cancel_request(struct ceph_osd_request *req)
 870 {
 871         if (req->r_sent && req->r_osd) {
 872                 ceph_msg_revoke(req->r_request);
 873                 req->r_sent = 0;
 874         }
 875 }
 876
 877 static void __register_linger_request(struct ceph_osd_client *osdc,
 878                                     struct ceph_osd_request *req)
 879 {
 880         dout("__register_linger_request %p\n", req);
 881         list_add_tail(&req->r_linger_item, &osdc->req_linger);
 882         if (req->r_osd)
 883                 list_add_tail(&req->r_linger_osd,
 884                               &req->r_osd->o_linger_requests);
 885 }
 886
 887 static void __unregister_linger_request(struct ceph_osd_client *osdc,
 888                                         struct ceph_osd_request *req)
 889 {
 890         dout("__unregister_linger_request %p\n", req);
 891         list_del_init(&req->r_linger_item);
 892         if (req->r_osd) {
 893                 list_del_init(&req->r_linger_osd);
 894
 895                 if (list_empty(&req->r_osd->o_requests) &&
 896                     list_empty(&req->r_osd->o_linger_requests)) {
 897                         dout("moving osd to %p lru\n", req->r_osd);
 898                         __move_osd_to_lru(osdc, req->r_osd);
 899                 }
 900                 if (list_empty(&req->r_osd_item))
 901                         req->r_osd = NULL;
 902         }
 903 }
 904
 905 void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc,
 906                                          struct ceph_osd_request *req)
 907 {
 908         mutex_lock(&osdc->request_mutex);
 909         if (req->r_linger) {
 910                 __unregister_linger_request(osdc, req);
 911                 ceph_osdc_put_request(req);
 912         }
 913         mutex_unlock(&osdc->request_mutex);
 914 }
 915 EXPORT_SYMBOL(ceph_osdc_unregister_linger_request);
 916
 917 void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc,
 918                                   struct ceph_osd_request *req)
 919 {
 920         if (!req->r_linger) {
 921                 dout("set_request_linger %p\n", req);
 922                 req->r_linger = 1;
 923                 /*
 924                  * caller is now responsible for calling
 925                  * unregister_linger_request
 926                  */
 927                 ceph_osdc_get_request(req);
 928         }
 929 }
 930 EXPORT_SYMBOL(ceph_osdc_set_request_linger);
 931
 932 /*
 933  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 934  * (as needed), and set the request r_osd appropriately.  If there is
 935  * no up osd, set r_osd to NULL.  Move the request to the appropriate list
 936  * (unsent, homeless) or leave on in-flight lru.
 937  *
 938  * Return 0 if unchanged, 1 if changed, or negative on error.
 939  *
 940  * Caller should hold map_sem for read and request_mutex.
 941  */
 942 static int __map_request(struct ceph_osd_client *osdc,
 943                          struct ceph_osd_request *req, int force_resend)
 944 {
 945         struct ceph_pg pgid;
 946         int acting[CEPH_PG_MAX_SIZE];
 947         int o = -1, num = 0;
 948         int err;
 949
 950         dout("map_request %p tid %lld\n", req, req->r_tid);
 951         err = ceph_calc_object_layout(&pgid, req->r_oid,
 952                                       &req->r_file_layout, osdc->osdmap);
 953         if (err) {
 954                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
 955                 return err;
 956         }
 957         req->r_pgid = pgid;
 958
 959         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 960         if (err > 0) {
 961                 o = acting[0];
 962                 num = err;
 963         }
 964
 965         if ((!force_resend &&
 966              req->r_osd && req->r_osd->o_osd == o &&
 967              req->r_sent >= req->r_osd->o_incarnation &&
 968              req->r_num_pg_osds == num &&
 969              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 970             (req->r_osd == NULL && o == -1))
 971                 return 0;  /* no change */
 972
 973         dout("map_request tid %llu pgid %lld.%x osd%d (was osd%d)\n",
 974              req->r_tid, pgid.pool, pgid.seed, o,
 975              req->r_osd ? req->r_osd->o_osd : -1);
 976
 977         /* record full pg acting set */
 978         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 979         req->r_num_pg_osds = num;
 980
 981         if (req->r_osd) {
 982                 __cancel_request(req);
 983                 list_del_init(&req->r_osd_item);
 984                 req->r_osd = NULL;
 985         }
 986
 987         req->r_osd = __lookup_osd(osdc, o);
 988         if (!req->r_osd && o >= 0) {
 989                 err = -ENOMEM;
 990                 req->r_osd = create_osd(osdc, o);
 991                 if (!req->r_osd) {
 992                         list_move(&req->r_req_lru_item, &osdc->req_notarget);
 993                         goto out;
 994                 }
 995
 996                 dout("map_request osd %p is osd%d\n", req->r_osd, o);
 997                 __insert_osd(osdc, req->r_osd);
 998
 999                 ceph_con_open(&req->r_osd->o_con,
1000                               CEPH_ENTITY_TYPE_OSD, o,
1001                               &osdc->osdmap->osd_addr[o]);
1002         }
1003
1004         if (req->r_osd) {
1005                 __remove_osd_from_lru(req->r_osd);
1006                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
1007                 list_move(&req->r_req_lru_item, &osdc->req_unsent);
1008         } else {
1009                 list_move(&req->r_req_lru_item, &osdc->req_notarget);
1010         }
1011         err = 1;   /* osd or pg changed */
1012
1013 out:
1014         return err;
1015 }
1016
1017 /*
1018  * caller should hold map_sem (for read) and request_mutex
1019  */
1020 static void __send_request(struct ceph_osd_client *osdc,
1021                            struct ceph_osd_request *req)
1022 {
1023         void *p;
1024
1025         dout("send_request %p tid %llu to osd%d flags %d pg %lld.%x\n",
1026              req, req->r_tid, req->r_osd->o_osd, req->r_flags,
1027              (unsigned long long)req->r_pgid.pool, req->r_pgid.seed);
1028
1029         /* fill in message content that changes each time we send it */
1030         put_unaligned_le32(osdc->osdmap->epoch, req->r_request_osdmap_epoch);
1031         put_unaligned_le32(req->r_flags, req->r_request_flags);
1032         put_unaligned_le64(req->r_pgid.pool, req->r_request_pool);
1033         p = req->r_request_pgid;
1034         ceph_encode_64(&p, req->r_pgid.pool);
1035         ceph_encode_32(&p, req->r_pgid.seed);
1036         put_unaligned_le64(1, req->r_request_attempts);  /* FIXME */
1037         memcpy(req->r_request_reassert_version, &req->r_reassert_version,
1038                sizeof(req->r_reassert_version));
1039
1040         req->r_stamp = jiffies;
1041         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
1042
1043         ceph_msg_get(req->r_request); /* send consumes a ref */
1044         ceph_con_send(&req->r_osd->o_con, req->r_request);
1045         req->r_sent = req->r_osd->o_incarnation;
1046 }
1047
1048 /*
1049  * Send any requests in the queue (req_unsent).
1050  */
1051 static void __send_queued(struct ceph_osd_client *osdc)
1052 {
1053         struct ceph_osd_request *req, *tmp;
1054
1055         dout("__send_queued\n");
1056         list_for_each_entry_safe(req, tmp, &osdc->req_unsent, r_req_lru_item)
1057                 __send_request(osdc, req);
1058 }
1059
1060 /*
1061  * Timeout callback, called every N seconds when 1 or more osd
1062  * requests has been active for more than N seconds.  When this
1063  * happens, we ping all OSDs with requests who have timed out to
1064  * ensure any communications channel reset is detected.  Reset the
1065  * request timeouts another N seconds in the future as we go.
1066  * Reschedule the timeout event another N seconds in future (unless
1067  * there are no open requests).
1068  */
1069 static void handle_timeout(struct work_struct *work)
1070 {
1071         struct ceph_osd_client *osdc =
1072                 container_of(work, struct ceph_osd_client, timeout_work.work);
1073         struct ceph_osd_request *req;
1074         struct ceph_osd *osd;
1075         unsigned long keepalive =
1076                 osdc->client->options->osd_keepalive_timeout * HZ;
1077         struct list_head slow_osds;
1078         dout("timeout\n");
1079         down_read(&osdc->map_sem);
1080
1081         ceph_monc_request_next_osdmap(&osdc->client->monc);
1082
1083         mutex_lock(&osdc->request_mutex);
1084
1085         /*
1086          * ping osds that are a bit slow.  this ensures that if there
1087          * is a break in the TCP connection we will notice, and reopen
1088          * a connection with that osd (from the fault callback).
1089          */
1090         INIT_LIST_HEAD(&slow_osds);
1091         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
1092                 if (time_before(jiffies, req->r_stamp + keepalive))
1093                         break;
1094
1095                 osd = req->r_osd;
1096                 BUG_ON(!osd);
1097                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
1098                      req->r_tid, osd->o_osd);
1099                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
1100         }
1101         while (!list_empty(&slow_osds)) {
1102                 osd = list_entry(slow_osds.next, struct ceph_osd,
1103                                  o_keepalive_item);
1104                 list_del_init(&osd->o_keepalive_item);
1105                 ceph_con_keepalive(&osd->o_con);
1106         }
1107
1108         __schedule_osd_timeout(osdc);
1109         __send_queued(osdc);
1110         mutex_unlock(&osdc->request_mutex);
1111         up_read(&osdc->map_sem);
1112 }
1113
1114 static void handle_osds_timeout(struct work_struct *work)
1115 {
1116         struct ceph_osd_client *osdc =
1117                 container_of(work, struct ceph_osd_client,
1118                              osds_timeout_work.work);
1119         unsigned long delay =
1120                 osdc->client->options->osd_idle_ttl * HZ >> 2;
1121
1122         dout("osds timeout\n");
1123         down_read(&osdc->map_sem);
1124         remove_old_osds(osdc);
1125         up_read(&osdc->map_sem);
1126
1127         schedule_delayed_work(&osdc->osds_timeout_work,
1128                               round_jiffies_relative(delay));
1129 }
1130
1131 static void complete_request(struct ceph_osd_request *req)
1132 {
1133         if (req->r_safe_callback)
1134                 req->r_safe_callback(req, NULL);
1135         complete_all(&req->r_safe_completion);  /* fsync waiter */
1136 }
1137
1138 static int __decode_pgid(void **p, void *end, struct ceph_pg *pgid)
1139 {
1140         __u8 v;
1141
1142         ceph_decode_need(p, end, 1 + 8 + 4 + 4, bad);
1143         v = ceph_decode_8(p);
1144         if (v > 1) {
1145                 pr_warning("do not understand pg encoding %d > 1", v);
1146                 return -EINVAL;
1147         }
1148         pgid->pool = ceph_decode_64(p);
1149         pgid->seed = ceph_decode_32(p);
1150         *p += 4;
1151         return 0;
1152
1153 bad:
1154         pr_warning("incomplete pg encoding");
1155         return -EINVAL;
1156 }
1157
1158 /*
1159  * handle osd op reply.  either call the callback if it is specified,
1160  * or do the completion to wake up the waiting thread.
1161  */
1162 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
1163                          struct ceph_connection *con)
1164 {
1165         void *p, *end;
1166         struct ceph_osd_request *req;
1167         u64 tid;
1168         int object_len;
1169         int numops, payload_len, flags;
1170         s32 result;
1171         s32 retry_attempt;
1172         struct ceph_pg pg;
1173         int err;
1174         u32 reassert_epoch;
1175         u64 reassert_version;
1176         u32 osdmap_epoch;
1177         int i;
1178
1179         tid = le64_to_cpu(msg->hdr.tid);
1180         dout("handle_reply %p tid %llu\n", msg, tid);
1181
1182         p = msg->front.iov_base;
1183         end = p + msg->front.iov_len;
1184
1185         ceph_decode_need(&p, end, 4, bad);
1186         object_len = ceph_decode_32(&p);
1187         ceph_decode_need(&p, end, object_len, bad);
1188         p += object_len;
1189
1190         err = __decode_pgid(&p, end, &pg);
1191         if (err)
1192                 goto bad;
1193
1194         ceph_decode_need(&p, end, 8 + 4 + 4 + 8 + 4, bad);
1195         flags = ceph_decode_64(&p);
1196         result = ceph_decode_32(&p);
1197         reassert_epoch = ceph_decode_32(&p);
1198         reassert_version = ceph_decode_64(&p);
1199         osdmap_epoch = ceph_decode_32(&p);
1200
1201         /* lookup */
1202         mutex_lock(&osdc->request_mutex);
1203         req = __lookup_request(osdc, tid);
1204         if (req == NULL) {
1205                 dout("handle_reply tid %llu dne\n", tid);
1206                 mutex_unlock(&osdc->request_mutex);
1207                 return;
1208         }
1209         ceph_osdc_get_request(req);
1210
1211         dout("handle_reply %p tid %llu req %p result %d\n", msg, tid,
1212              req, result);
1213
1214         ceph_decode_need(&p, end, 4, bad);
1215         numops = ceph_decode_32(&p);
1216         if (numops > CEPH_OSD_MAX_OP)
1217                 goto bad_put;
1218         if (numops != req->r_num_ops)
1219                 goto bad_put;
1220         payload_len = 0;
1221         ceph_decode_need(&p, end, numops * sizeof(struct ceph_osd_op), bad);
1222         for (i = 0; i < numops; i++) {
1223                 struct ceph_osd_op *op = p;
1224                 int len;
1225
1226                 len = le32_to_cpu(op->payload_len);
1227                 req->r_reply_op_len[i] = len;
1228                 dout(" op %d has %d bytes\n", i, len);
1229                 payload_len += len;
1230                 p += sizeof(*op);
1231         }
1232         if (payload_len != le32_to_cpu(msg->hdr.data_len)) {
1233                 pr_warning("sum of op payload lens %d != data_len %d",
1234                            payload_len, le32_to_cpu(msg->hdr.data_len));
1235                 goto bad_put;
1236         }
1237
1238         ceph_decode_need(&p, end, 4 + numops * 4, bad);
1239         retry_attempt = ceph_decode_32(&p);
1240         for (i = 0; i < numops; i++)
1241                 req->r_reply_op_result[i] = ceph_decode_32(&p);
1242
1243         /*
1244          * if this connection filled our message, drop our reference now, to
1245          * avoid a (safe but slower) revoke later.
1246          */
1247         if (req->r_con_filling_msg == con && req->r_reply == msg) {
1248                 dout(" dropping con_filling_msg ref %p\n", con);
1249                 req->r_con_filling_msg = NULL;
1250                 con->ops->put(con);
1251         }
1252
1253         if (!req->r_got_reply) {
1254                 unsigned int bytes;
1255
1256                 req->r_result = result;
1257                 bytes = le32_to_cpu(msg->hdr.data_len);
1258                 dout("handle_reply result %d bytes %d\n", req->r_result,
1259                      bytes);
1260                 if (req->r_result == 0)
1261                         req->r_result = bytes;
1262
1263                 /* in case this is a write and we need to replay, */
1264                 req->r_reassert_version.epoch = cpu_to_le32(reassert_epoch);
1265                 req->r_reassert_version.version = cpu_to_le64(reassert_version);
1266
1267                 req->r_got_reply = 1;
1268         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
1269                 dout("handle_reply tid %llu dup ack\n", tid);
1270                 mutex_unlock(&osdc->request_mutex);
1271                 goto done;
1272         }
1273
1274         dout("handle_reply tid %llu flags %d\n", tid, flags);
1275
1276         if (req->r_linger && (flags & CEPH_OSD_FLAG_ONDISK))
1277                 __register_linger_request(osdc, req);
1278
1279         /* either this is a read, or we got the safe response */
1280         if (result < 0 ||
1281             (flags & CEPH_OSD_FLAG_ONDISK) ||
1282             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
1283                 __unregister_request(osdc, req);
1284
1285         mutex_unlock(&osdc->request_mutex);
1286
1287         if (req->r_callback)
1288                 req->r_callback(req, msg);
1289         else
1290                 complete_all(&req->r_completion);
1291
1292         if (flags & CEPH_OSD_FLAG_ONDISK)
1293                 complete_request(req);
1294
1295 done:
1296         dout("req=%p req->r_linger=%d\n", req, req->r_linger);
1297         ceph_osdc_put_request(req);
1298         return;
1299
1300 bad_put:
1301         ceph_osdc_put_request(req);
1302 bad:
1303         pr_err("corrupt osd_op_reply got %d %d\n",
1304                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len));
1305         ceph_msg_dump(msg);
1306 }
1307
1308 static void reset_changed_osds(struct ceph_osd_client *osdc)
1309 {
1310         struct rb_node *p, *n;
1311
1312         for (p = rb_first(&osdc->osds); p; p = n) {
1313                 struct ceph_osd *osd = rb_entry(p, struct ceph_osd, o_node);
1314
1315                 n = rb_next(p);
1316                 if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
1317                     memcmp(&osd->o_con.peer_addr,
1318                            ceph_osd_addr(osdc->osdmap,
1319                                          osd->o_osd),
1320                            sizeof(struct ceph_entity_addr)) != 0)
1321                         __reset_osd(osdc, osd);
1322         }
1323 }
1324
1325 /*
1326  * Requeue requests whose mapping to an OSD has changed.  If requests map to
1327  * no osd, request a new map.
1328  *
1329  * Caller should hold map_sem for read.
1330  */
1331 static void kick_requests(struct ceph_osd_client *osdc, int force_resend)
1332 {
1333         struct ceph_osd_request *req, *nreq;
1334         struct rb_node *p;
1335         int needmap = 0;
1336         int err;
1337
1338         dout("kick_requests %s\n", force_resend ? " (force resend)" : "");
1339         mutex_lock(&osdc->request_mutex);
1340         for (p = rb_first(&osdc->requests); p; ) {
1341                 req = rb_entry(p, struct ceph_osd_request, r_node);
1342                 p = rb_next(p);
1343
1344                 /*
1345                  * For linger requests that have not yet been
1346                  * registered, move them to the linger list; they'll
1347                  * be sent to the osd in the loop below.  Unregister
1348                  * the request before re-registering it as a linger
1349                  * request to ensure the __map_request() below
1350                  * will decide it needs to be sent.
1351                  */
1352                 if (req->r_linger && list_empty(&req->r_linger_item)) {
1353                         dout("%p tid %llu restart on osd%d\n",
1354                              req, req->r_tid,
1355                              req->r_osd ? req->r_osd->o_osd : -1);
1356                         __unregister_request(osdc, req);
1357                         __register_linger_request(osdc, req);
1358                         continue;
1359                 }
1360
1361                 err = __map_request(osdc, req, force_resend);
1362                 if (err < 0)
1363                         continue;  /* error */
1364                 if (req->r_osd == NULL) {
1365                         dout("%p tid %llu maps to no osd\n", req, req->r_tid);
1366                         needmap++;  /* request a newer map */
1367                 } else if (err > 0) {
1368                         if (!req->r_linger) {
1369                                 dout("%p tid %llu requeued on osd%d\n", req,
1370                                      req->r_tid,
1371                                      req->r_osd ? req->r_osd->o_osd : -1);
1372                                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
1373                         }
1374                 }
1375         }
1376
1377         list_for_each_entry_safe(req, nreq, &osdc->req_linger,
1378                                  r_linger_item) {
1379                 dout("linger req=%p req->r_osd=%p\n", req, req->r_osd);
1380
1381                 err = __map_request(osdc, req, force_resend);
1382                 dout("__map_request returned %d\n", err);
1383                 if (err == 0)
1384                         continue;  /* no change and no osd was specified */
1385                 if (err < 0)
1386                         continue;  /* hrm! */
1387                 if (req->r_osd == NULL) {
1388                         dout("tid %llu maps to no valid osd\n", req->r_tid);
1389                         needmap++;  /* request a newer map */
1390                         continue;
1391                 }
1392
1393                 dout("kicking lingering %p tid %llu osd%d\n", req, req->r_tid,
1394                      req->r_osd ? req->r_osd->o_osd : -1);
1395                 __register_request(osdc, req);
1396                 __unregister_linger_request(osdc, req);
1397         }
1398         mutex_unlock(&osdc->request_mutex);
1399
1400         if (needmap) {
1401                 dout("%d requests for down osds, need new map\n", needmap);
1402                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1403         }
1404         reset_changed_osds(osdc);
1405 }
1406
1407
1408 /*
1409  * Process updated osd map.
1410  *
1411  * The message contains any number of incremental and full maps, normally
1412  * indicating some sort of topology change in the cluster.  Kick requests
1413  * off to different OSDs as needed.
1414  */
1415 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
1416 {
1417         void *p, *end, *next;
1418         u32 nr_maps, maplen;
1419         u32 epoch;
1420         struct ceph_osdmap *newmap = NULL, *oldmap;
1421         int err;
1422         struct ceph_fsid fsid;
1423
1424         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1425         p = msg->front.iov_base;
1426         end = p + msg->front.iov_len;
1427
1428         /* verify fsid */
1429         ceph_decode_need(&p, end, sizeof(fsid), bad);
1430         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1431         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1432                 return;
1433
1434         down_write(&osdc->map_sem);
1435
1436         /* incremental maps */
1437         ceph_decode_32_safe(&p, end, nr_maps, bad);
1438         dout(" %d inc maps\n", nr_maps);
1439         while (nr_maps > 0) {
1440                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1441                 epoch = ceph_decode_32(&p);
1442                 maplen = ceph_decode_32(&p);
1443                 ceph_decode_need(&p, end, maplen, bad);
1444                 next = p + maplen;
1445                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1446                         dout("applying incremental map %u len %d\n",
1447                              epoch, maplen);
1448                         newmap = osdmap_apply_incremental(&p, next,
1449                                                           osdc->osdmap,
1450                                                           &osdc->client->msgr);
1451                         if (IS_ERR(newmap)) {
1452                                 err = PTR_ERR(newmap);
1453                                 goto bad;
1454                         }
1455                         BUG_ON(!newmap);
1456                         if (newmap != osdc->osdmap) {
1457                                 ceph_osdmap_destroy(osdc->osdmap);
1458                                 osdc->osdmap = newmap;
1459                         }
1460                         kick_requests(osdc, 0);
1461                 } else {
1462                         dout("ignoring incremental map %u len %d\n",
1463                              epoch, maplen);
1464                 }
1465                 p = next;
1466                 nr_maps--;
1467         }
1468         if (newmap)
1469                 goto done;
1470
1471         /* full maps */
1472         ceph_decode_32_safe(&p, end, nr_maps, bad);
1473         dout(" %d full maps\n", nr_maps);
1474         while (nr_maps) {
1475                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1476                 epoch = ceph_decode_32(&p);
1477                 maplen = ceph_decode_32(&p);
1478                 ceph_decode_need(&p, end, maplen, bad);
1479                 if (nr_maps > 1) {
1480                         dout("skipping non-latest full map %u len %d\n",
1481                              epoch, maplen);
1482                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1483                         dout("skipping full map %u len %d, "
1484                              "older than our %u\n", epoch, maplen,
1485                              osdc->osdmap->epoch);
1486                 } else {
1487                         int skipped_map = 0;
1488
1489                         dout("taking full map %u len %d\n", epoch, maplen);
1490                         newmap = osdmap_decode(&p, p+maplen);
1491                         if (IS_ERR(newmap)) {
1492                                 err = PTR_ERR(newmap);
1493                                 goto bad;
1494                         }
1495                         BUG_ON(!newmap);
1496                         oldmap = osdc->osdmap;
1497                         osdc->osdmap = newmap;
1498                         if (oldmap) {
1499                                 if (oldmap->epoch + 1 < newmap->epoch)
1500                                         skipped_map = 1;
1501                                 ceph_osdmap_destroy(oldmap);
1502                         }
1503                         kick_requests(osdc, skipped_map);
1504                 }
1505                 p += maplen;
1506                 nr_maps--;
1507         }
1508
1509 done:
1510         downgrade_write(&osdc->map_sem);
1511         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1512
1513         /*
1514          * subscribe to subsequent osdmap updates if full to ensure
1515          * we find out when we are no longer full and stop returning
1516          * ENOSPC.
1517          */
1518         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
1519                 ceph_monc_request_next_osdmap(&osdc->client->monc);
1520
1521         mutex_lock(&osdc->request_mutex);
1522         __send_queued(osdc);
1523         mutex_unlock(&osdc->request_mutex);
1524         up_read(&osdc->map_sem);
1525         wake_up_all(&osdc->client->auth_wq);
1526         return;
1527
1528 bad:
1529         pr_err("osdc handle_map corrupt msg\n");
1530         ceph_msg_dump(msg);
1531         up_write(&osdc->map_sem);
1532         return;
1533 }
1534
1535 /*
1536  * watch/notify callback event infrastructure
1537  *
1538  * These callbacks are used both for watch and notify operations.
1539  */
1540 static void __release_event(struct kref *kref)
1541 {
1542         struct ceph_osd_event *event =
1543                 container_of(kref, struct ceph_osd_event, kref);
1544
1545         dout("__release_event %p\n", event);
1546         kfree(event);
1547 }
1548
1549 static void get_event(struct ceph_osd_event *event)
1550 {
1551         kref_get(&event->kref);
1552 }
1553
1554 void ceph_osdc_put_event(struct ceph_osd_event *event)
1555 {
1556         kref_put(&event->kref, __release_event);
1557 }
1558 EXPORT_SYMBOL(ceph_osdc_put_event);
1559
1560 static void __insert_event(struct ceph_osd_client *osdc,
1561                              struct ceph_osd_event *new)
1562 {
1563         struct rb_node **p = &osdc->event_tree.rb_node;
1564         struct rb_node *parent = NULL;
1565         struct ceph_osd_event *event = NULL;
1566
1567         while (*p) {
1568                 parent = *p;
1569                 event = rb_entry(parent, struct ceph_osd_event, node);
1570                 if (new->cookie < event->cookie)
1571                         p = &(*p)->rb_left;
1572                 else if (new->cookie > event->cookie)
1573                         p = &(*p)->rb_right;
1574                 else
1575                         BUG();
1576         }
1577
1578         rb_link_node(&new->node, parent, p);
1579         rb_insert_color(&new->node, &osdc->event_tree);
1580 }
1581
1582 static struct ceph_osd_event *__find_event(struct ceph_osd_client *osdc,
1583                                                 u64 cookie)
1584 {
1585         struct rb_node **p = &osdc->event_tree.rb_node;
1586         struct rb_node *parent = NULL;
1587         struct ceph_osd_event *event = NULL;
1588
1589         while (*p) {
1590                 parent = *p;
1591                 event = rb_entry(parent, struct ceph_osd_event, node);
1592                 if (cookie < event->cookie)
1593                         p = &(*p)->rb_left;
1594                 else if (cookie > event->cookie)
1595                         p = &(*p)->rb_right;
1596                 else
1597                         return event;
1598         }
1599         return NULL;
1600 }
1601
1602 static void __remove_event(struct ceph_osd_event *event)
1603 {
1604         struct ceph_osd_client *osdc = event->osdc;
1605
1606         if (!RB_EMPTY_NODE(&event->node)) {
1607                 dout("__remove_event removed %p\n", event);
1608                 rb_erase(&event->node, &osdc->event_tree);
1609                 ceph_osdc_put_event(event);
1610         } else {
1611                 dout("__remove_event didn't remove %p\n", event);
1612         }
1613 }
1614
1615 int ceph_osdc_create_event(struct ceph_osd_client *osdc,
1616                            void (*event_cb)(u64, u64, u8, void *),
1617                            void *data, struct ceph_osd_event **pevent)
1618 {
1619         struct ceph_osd_event *event;
1620
1621         event = kmalloc(sizeof(*event), GFP_NOIO);
1622         if (!event)
1623                 return -ENOMEM;
1624
1625         dout("create_event %p\n", event);
1626         event->cb = event_cb;
1627         event->one_shot = 0;
1628         event->data = data;
1629         event->osdc = osdc;
1630         INIT_LIST_HEAD(&event->osd_node);
1631         RB_CLEAR_NODE(&event->node);
1632         kref_init(&event->kref);   /* one ref for us */
1633         kref_get(&event->kref);    /* one ref for the caller */
1634
1635         spin_lock(&osdc->event_lock);
1636         event->cookie = ++osdc->event_count;
1637         __insert_event(osdc, event);
1638         spin_unlock(&osdc->event_lock);
1639
1640         *pevent = event;
1641         return 0;
1642 }
1643 EXPORT_SYMBOL(ceph_osdc_create_event);
1644
1645 void ceph_osdc_cancel_event(struct ceph_osd_event *event)
1646 {
1647         struct ceph_osd_client *osdc = event->osdc;
1648
1649         dout("cancel_event %p\n", event);
1650         spin_lock(&osdc->event_lock);
1651         __remove_event(event);
1652         spin_unlock(&osdc->event_lock);
1653         ceph_osdc_put_event(event); /* caller's */
1654 }
1655 EXPORT_SYMBOL(ceph_osdc_cancel_event);
1656
1657
1658 static void do_event_work(struct work_struct *work)
1659 {
1660         struct ceph_osd_event_work *event_work =
1661                 container_of(work, struct ceph_osd_event_work, work);
1662         struct ceph_osd_event *event = event_work->event;
1663         u64 ver = event_work->ver;
1664         u64 notify_id = event_work->notify_id;
1665         u8 opcode = event_work->opcode;
1666
1667         dout("do_event_work completing %p\n", event);
1668         event->cb(ver, notify_id, opcode, event->data);
1669         dout("do_event_work completed %p\n", event);
1670         ceph_osdc_put_event(event);
1671         kfree(event_work);
1672 }
1673
1674
1675 /*
1676  * Process osd watch notifications
1677  */
1678 static void handle_watch_notify(struct ceph_osd_client *osdc,
1679                                 struct ceph_msg *msg)
1680 {
1681         void *p, *end;
1682         u8 proto_ver;
1683         u64 cookie, ver, notify_id;
1684         u8 opcode;
1685         struct ceph_osd_event *event;
1686         struct ceph_osd_event_work *event_work;
1687
1688         p = msg->front.iov_base;
1689         end = p + msg->front.iov_len;
1690
1691         ceph_decode_8_safe(&p, end, proto_ver, bad);
1692         ceph_decode_8_safe(&p, end, opcode, bad);
1693         ceph_decode_64_safe(&p, end, cookie, bad);
1694         ceph_decode_64_safe(&p, end, ver, bad);
1695         ceph_decode_64_safe(&p, end, notify_id, bad);
1696
1697         spin_lock(&osdc->event_lock);
1698         event = __find_event(osdc, cookie);
1699         if (event) {
1700                 BUG_ON(event->one_shot);
1701                 get_event(event);
1702         }
1703         spin_unlock(&osdc->event_lock);
1704         dout("handle_watch_notify cookie %lld ver %lld event %p\n",
1705              cookie, ver, event);
1706         if (event) {
1707                 event_work = kmalloc(sizeof(*event_work), GFP_NOIO);
1708                 if (!event_work) {
1709                         dout("ERROR: could not allocate event_work\n");
1710                         goto done_err;
1711                 }
1712                 INIT_WORK(&event_work->work, do_event_work);
1713                 event_work->event = event;
1714                 event_work->ver = ver;
1715                 event_work->notify_id = notify_id;
1716                 event_work->opcode = opcode;
1717                 if (!queue_work(osdc->notify_wq, &event_work->work)) {
1718                         dout("WARNING: failed to queue notify event work\n");
1719                         goto done_err;
1720                 }
1721         }
1722
1723         return;
1724
1725 done_err:
1726         ceph_osdc_put_event(event);
1727         return;
1728
1729 bad:
1730         pr_err("osdc handle_watch_notify corrupt msg\n");
1731         return;
1732 }
1733
1734 /*
1735  * Register request, send initial attempt.
1736  */
1737 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1738                             struct ceph_osd_request *req,
1739                             bool nofail)
1740 {
1741         int rc = 0;
1742
1743         req->r_request->pages = req->r_pages;
1744         req->r_request->page_count = req->r_num_pages;
1745         req->r_request->page_alignment = req->r_page_alignment;
1746 #ifdef CONFIG_BLOCK
1747         req->r_request->bio = req->r_bio;
1748 #endif
1749         req->r_request->trail = &req->r_trail;
1750
1751         register_request(osdc, req);
1752
1753         down_read(&osdc->map_sem);
1754         mutex_lock(&osdc->request_mutex);
1755         /*
1756          * a racing kick_requests() may have sent the message for us
1757          * while we dropped request_mutex above, so only send now if
1758          * the request still han't been touched yet.
1759          */
1760         if (req->r_sent == 0) {
1761                 rc = __map_request(osdc, req, 0);
1762                 if (rc < 0) {
1763                         if (nofail) {
1764                                 dout("osdc_start_request failed map, "
1765                                      " will retry %lld\n", req->r_tid);
1766                                 rc = 0;
1767                         }
1768                         goto out_unlock;
1769                 }
1770                 if (req->r_osd == NULL) {
1771                         dout("send_request %p no up osds in pg\n", req);
1772                         ceph_monc_request_next_osdmap(&osdc->client->monc);
1773                 } else {
1774                         __send_request(osdc, req);
1775                 }
1776                 rc = 0;
1777         }
1778
1779 out_unlock:
1780         mutex_unlock(&osdc->request_mutex);
1781         up_read(&osdc->map_sem);
1782         return rc;
1783 }
1784 EXPORT_SYMBOL(ceph_osdc_start_request);
1785
1786 /*
1787  * wait for a request to complete
1788  */
1789 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1790                            struct ceph_osd_request *req)
1791 {
1792         int rc;
1793
1794         rc = wait_for_completion_interruptible(&req->r_completion);
1795         if (rc < 0) {
1796                 mutex_lock(&osdc->request_mutex);
1797                 __cancel_request(req);
1798                 __unregister_request(osdc, req);
1799                 mutex_unlock(&osdc->request_mutex);
1800                 complete_request(req);
1801                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1802                 return rc;
1803         }
1804
1805         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1806         return req->r_result;
1807 }
1808 EXPORT_SYMBOL(ceph_osdc_wait_request);
1809
1810 /*
1811  * sync - wait for all in-flight requests to flush.  avoid starvation.
1812  */
1813 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1814 {
1815         struct ceph_osd_request *req;
1816         u64 last_tid, next_tid = 0;
1817
1818         mutex_lock(&osdc->request_mutex);
1819         last_tid = osdc->last_tid;
1820         while (1) {
1821                 req = __lookup_request_ge(osdc, next_tid);
1822                 if (!req)
1823                         break;
1824                 if (req->r_tid > last_tid)
1825                         break;
1826
1827                 next_tid = req->r_tid + 1;
1828                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1829                         continue;
1830
1831                 ceph_osdc_get_request(req);
1832                 mutex_unlock(&osdc->request_mutex);
1833                 dout("sync waiting on tid %llu (last is %llu)\n",
1834                      req->r_tid, last_tid);
1835                 wait_for_completion(&req->r_safe_completion);
1836                 mutex_lock(&osdc->request_mutex);
1837                 ceph_osdc_put_request(req);
1838         }
1839         mutex_unlock(&osdc->request_mutex);
1840         dout("sync done (thru tid %llu)\n", last_tid);
1841 }
1842 EXPORT_SYMBOL(ceph_osdc_sync);
1843
1844 /*
1845  * init, shutdown
1846  */
1847 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1848 {
1849         int err;
1850
1851         dout("init\n");
1852         osdc->client = client;
1853         osdc->osdmap = NULL;
1854         init_rwsem(&osdc->map_sem);
1855         init_completion(&osdc->map_waiters);
1856         osdc->last_requested_map = 0;
1857         mutex_init(&osdc->request_mutex);
1858         osdc->last_tid = 0;
1859         osdc->osds = RB_ROOT;
1860         INIT_LIST_HEAD(&osdc->osd_lru);
1861         osdc->requests = RB_ROOT;
1862         INIT_LIST_HEAD(&osdc->req_lru);
1863         INIT_LIST_HEAD(&osdc->req_unsent);
1864         INIT_LIST_HEAD(&osdc->req_notarget);
1865         INIT_LIST_HEAD(&osdc->req_linger);
1866         osdc->num_requests = 0;
1867         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1868         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1869         spin_lock_init(&osdc->event_lock);
1870         osdc->event_tree = RB_ROOT;
1871         osdc->event_count = 0;
1872
1873         schedule_delayed_work(&osdc->osds_timeout_work,
1874            round_jiffies_relative(osdc->client->options->osd_idle_ttl * HZ));
1875
1876         err = -ENOMEM;
1877         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1878                                         sizeof(struct ceph_osd_request));
1879         if (!osdc->req_mempool)
1880                 goto out;
1881
1882         err = ceph_msgpool_init(&osdc->msgpool_op, CEPH_MSG_OSD_OP,
1883                                 OSD_OP_FRONT_LEN, 10, true,
1884                                 "osd_op");
1885         if (err < 0)
1886                 goto out_mempool;
1887         err = ceph_msgpool_init(&osdc->msgpool_op_reply, CEPH_MSG_OSD_OPREPLY,
1888                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1889                                 "osd_op_reply");
1890         if (err < 0)
1891                 goto out_msgpool;
1892
1893         osdc->notify_wq = create_singlethread_workqueue("ceph-watch-notify");
1894         if (IS_ERR(osdc->notify_wq)) {
1895                 err = PTR_ERR(osdc->notify_wq);
1896                 osdc->notify_wq = NULL;
1897                 goto out_msgpool;
1898         }
1899         return 0;
1900
1901 out_msgpool:
1902         ceph_msgpool_destroy(&osdc->msgpool_op);
1903 out_mempool:
1904         mempool_destroy(osdc->req_mempool);
1905 out:
1906         return err;
1907 }
1908
1909 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1910 {
1911         flush_workqueue(osdc->notify_wq);
1912         destroy_workqueue(osdc->notify_wq);
1913         cancel_delayed_work_sync(&osdc->timeout_work);
1914         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1915         if (osdc->osdmap) {
1916                 ceph_osdmap_destroy(osdc->osdmap);
1917                 osdc->osdmap = NULL;
1918         }
1919         remove_all_osds(osdc);
1920         mempool_destroy(osdc->req_mempool);
1921         ceph_msgpool_destroy(&osdc->msgpool_op);
1922         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1923 }
1924
1925 /*
1926  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1927  * *plen.  Return number of bytes read, or error.
1928  */
1929 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1930                         struct ceph_vino vino, struct ceph_file_layout *layout,
1931                         u64 off, u64 *plen,
1932                         u32 truncate_seq, u64 truncate_size,
1933                         struct page **pages, int num_pages, int page_align)
1934 {
1935         struct ceph_osd_request *req;
1936         int rc = 0;
1937
1938         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1939              vino.snap, off, *plen);
1940         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1941                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1942                                     NULL, 0, truncate_seq, truncate_size, NULL,
1943                                     false, page_align);
1944         if (IS_ERR(req))
1945                 return PTR_ERR(req);
1946
1947         /* it may be a short read due to an object boundary */
1948         req->r_pages = pages;
1949
1950         dout("readpages  final extent is %llu~%llu (%d pages align %d)\n",
1951              off, *plen, req->r_num_pages, page_align);
1952
1953         rc = ceph_osdc_start_request(osdc, req, false);
1954         if (!rc)
1955                 rc = ceph_osdc_wait_request(osdc, req);
1956
1957         ceph_osdc_put_request(req);
1958         dout("readpages result %d\n", rc);
1959         return rc;
1960 }
1961 EXPORT_SYMBOL(ceph_osdc_readpages);
1962
1963 /*
1964  * do a synchronous write on N pages
1965  */
1966 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1967                          struct ceph_file_layout *layout,
1968                          struct ceph_snap_context *snapc,
1969                          u64 off, u64 len,
1970                          u32 truncate_seq, u64 truncate_size,
1971                          struct timespec *mtime,
1972                          struct page **pages, int num_pages)
1973 {
1974         struct ceph_osd_request *req;
1975         int rc = 0;
1976         int page_align = off & ~PAGE_MASK;
1977
1978         BUG_ON(vino.snap != CEPH_NOSNAP);
1979         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1980                                     CEPH_OSD_OP_WRITE,
1981                                     CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1982                                     snapc, 0,
1983                                     truncate_seq, truncate_size, mtime,
1984                                     true, page_align);
1985         if (IS_ERR(req))
1986                 return PTR_ERR(req);
1987
1988         /* it may be a short write due to an object boundary */
1989         req->r_pages = pages;
1990         dout("writepages %llu~%llu (%d pages)\n", off, len,
1991              req->r_num_pages);
1992
1993         rc = ceph_osdc_start_request(osdc, req, true);
1994         if (!rc)
1995                 rc = ceph_osdc_wait_request(osdc, req);
1996
1997         ceph_osdc_put_request(req);
1998         if (rc == 0)
1999                 rc = len;
2000         dout("writepages result %d\n", rc);
2001         return rc;
2002 }
2003 EXPORT_SYMBOL(ceph_osdc_writepages);
2004
2005 /*
2006  * handle incoming message
2007  */
2008 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
2009 {
2010         struct ceph_osd *osd = con->private;
2011         struct ceph_osd_client *osdc;
2012         int type = le16_to_cpu(msg->hdr.type);
2013
2014         if (!osd)
2015                 goto out;
2016         osdc = osd->o_osdc;
2017
2018         switch (type) {
2019         case CEPH_MSG_OSD_MAP:
2020                 ceph_osdc_handle_map(osdc, msg);
2021                 break;
2022         case CEPH_MSG_OSD_OPREPLY:
2023                 handle_reply(osdc, msg, con);
2024                 break;
2025         case CEPH_MSG_WATCH_NOTIFY:
2026                 handle_watch_notify(osdc, msg);
2027                 break;
2028
2029         default:
2030                 pr_err("received unknown message type %d %s\n", type,
2031                        ceph_msg_type_name(type));
2032         }
2033 out:
2034         ceph_msg_put(msg);
2035 }
2036
2037 /*
2038  * lookup and return message for incoming reply.  set up reply message
2039  * pages.
2040  */
2041 static struct ceph_msg *get_reply(struct ceph_connection *con,
2042                                   struct ceph_msg_header *hdr,
2043                                   int *skip)
2044 {
2045         struct ceph_osd *osd = con->private;
2046         struct ceph_osd_client *osdc = osd->o_osdc;
2047         struct ceph_msg *m;
2048         struct ceph_osd_request *req;
2049         int front = le32_to_cpu(hdr->front_len);
2050         int data_len = le32_to_cpu(hdr->data_len);
2051         u64 tid;
2052
2053         tid = le64_to_cpu(hdr->tid);
2054         mutex_lock(&osdc->request_mutex);
2055         req = __lookup_request(osdc, tid);
2056         if (!req) {
2057                 *skip = 1;
2058                 m = NULL;
2059                 dout("get_reply unknown tid %llu from osd%d\n", tid,
2060                      osd->o_osd);
2061                 goto out;
2062         }
2063
2064         if (req->r_con_filling_msg) {
2065                 dout("%s revoking msg %p from old con %p\n", __func__,
2066                      req->r_reply, req->r_con_filling_msg);
2067                 ceph_msg_revoke_incoming(req->r_reply);
2068                 req->r_con_filling_msg->ops->put(req->r_con_filling_msg);
2069                 req->r_con_filling_msg = NULL;
2070         }
2071
2072         if (front > req->r_reply->front.iov_len) {
2073                 pr_warning("get_reply front %d > preallocated %d\n",
2074                            front, (int)req->r_reply->front.iov_len);
2075                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS, false);
2076                 if (!m)
2077                         goto out;
2078                 ceph_msg_put(req->r_reply);
2079                 req->r_reply = m;
2080         }
2081         m = ceph_msg_get(req->r_reply);
2082
2083         if (data_len > 0) {
2084                 int want = calc_pages_for(req->r_page_alignment, data_len);
2085
2086                 if (req->r_pages && unlikely(req->r_num_pages < want)) {
2087                         pr_warning("tid %lld reply has %d bytes %d pages, we"
2088                                    " had only %d pages ready\n", tid, data_len,
2089                                    want, req->r_num_pages);
2090                         *skip = 1;
2091                         ceph_msg_put(m);
2092                         m = NULL;
2093                         goto out;
2094                 }
2095                 m->pages = req->r_pages;
2096                 m->page_count = req->r_num_pages;
2097                 m->page_alignment = req->r_page_alignment;
2098 #ifdef CONFIG_BLOCK
2099                 m->bio = req->r_bio;
2100 #endif
2101         }
2102         *skip = 0;
2103         req->r_con_filling_msg = con->ops->get(con);
2104         dout("get_reply tid %lld %p\n", tid, m);
2105
2106 out:
2107         mutex_unlock(&osdc->request_mutex);
2108         return m;
2109
2110 }
2111
2112 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
2113                                   struct ceph_msg_header *hdr,
2114                                   int *skip)
2115 {
2116         struct ceph_osd *osd = con->private;
2117         int type = le16_to_cpu(hdr->type);
2118         int front = le32_to_cpu(hdr->front_len);
2119
2120         *skip = 0;
2121         switch (type) {
2122         case CEPH_MSG_OSD_MAP:
2123         case CEPH_MSG_WATCH_NOTIFY:
2124                 return ceph_msg_new(type, front, GFP_NOFS, false);
2125         case CEPH_MSG_OSD_OPREPLY:
2126                 return get_reply(con, hdr, skip);
2127         default:
2128                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
2129                         osd->o_osd);
2130                 *skip = 1;
2131                 return NULL;
2132         }
2133 }
2134
2135 /*
2136  * Wrappers to refcount containing ceph_osd struct
2137  */
2138 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
2139 {
2140         struct ceph_osd *osd = con->private;
2141         if (get_osd(osd))
2142                 return con;
2143         return NULL;
2144 }
2145
2146 static void put_osd_con(struct ceph_connection *con)
2147 {
2148         struct ceph_osd *osd = con->private;
2149         put_osd(osd);
2150 }
2151
2152 /*
2153  * authentication
2154  */
2155 /*
2156  * Note: returned pointer is the address of a structure that's
2157  * managed separately.  Caller must *not* attempt to free it.
2158  */
2159 static struct ceph_auth_handshake *get_authorizer(struct ceph_connection *con,
2160                                         int *proto, int force_new)
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         struct ceph_auth_handshake *auth = &o->o_auth;
2166
2167         if (force_new && auth->authorizer) {
2168                 if (ac->ops && ac->ops->destroy_authorizer)
2169                         ac->ops->destroy_authorizer(ac, auth->authorizer);
2170                 auth->authorizer = NULL;
2171         }
2172         if (!auth->authorizer && ac->ops && ac->ops->create_authorizer) {
2173                 int ret = ac->ops->create_authorizer(ac, CEPH_ENTITY_TYPE_OSD,
2174                                                         auth);
2175                 if (ret)
2176                         return ERR_PTR(ret);
2177         }
2178         *proto = ac->protocol;
2179
2180         return auth;
2181 }
2182
2183
2184 static int verify_authorizer_reply(struct ceph_connection *con, int len)
2185 {
2186         struct ceph_osd *o = con->private;
2187         struct ceph_osd_client *osdc = o->o_osdc;
2188         struct ceph_auth_client *ac = osdc->client->monc.auth;
2189
2190         /*
2191          * XXX If ac->ops or ac->ops->verify_authorizer_reply is null,
2192          * XXX which do we do:  succeed or fail?
2193          */
2194         return ac->ops->verify_authorizer_reply(ac, o->o_auth.authorizer, len);
2195 }
2196
2197 static int invalidate_authorizer(struct ceph_connection *con)
2198 {
2199         struct ceph_osd *o = con->private;
2200         struct ceph_osd_client *osdc = o->o_osdc;
2201         struct ceph_auth_client *ac = osdc->client->monc.auth;
2202
2203         if (ac->ops && ac->ops->invalidate_authorizer)
2204                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
2205
2206         return ceph_monc_validate_auth(&osdc->client->monc);
2207 }
2208
2209 static const struct ceph_connection_operations osd_con_ops = {
2210         .get = get_osd_con,
2211         .put = put_osd_con,
2212         .dispatch = dispatch,
2213         .get_authorizer = get_authorizer,
2214         .verify_authorizer_reply = verify_authorizer_reply,
2215         .invalidate_authorizer = invalidate_authorizer,
2216         .alloc_msg = alloc_msg,
2217         .fault = osd_reset,
2218 };