fs/ceph/osd_client.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/err.h>
   4 #include <linux/highmem.h>
   5 #include <linux/mm.h>
   6 #include <linux/pagemap.h>
   7 #include <linux/slab.h>
   8 #include <linux/uaccess.h>
   9
  10 #include "super.h"
  11 #include "osd_client.h"
  12 #include "messenger.h"
  13 #include "decode.h"
  14 #include "auth.h"
  15
  16 #define OSD_OP_FRONT_LEN        4096
  17 #define OSD_OPREPLY_FRONT_LEN   512
  18
  19 static const struct ceph_connection_operations osd_con_ops;
  20 static int __kick_requests(struct ceph_osd_client *osdc,
  21                           struct ceph_osd *kickosd);
  22
  23 static void kick_requests(struct ceph_osd_client *osdc, struct ceph_osd *osd);
  24
  25 /*
  26  * Implement client access to distributed object storage cluster.
  27  *
  28  * All data objects are stored within a cluster/cloud of OSDs, or
  29  * "object storage devices."  (Note that Ceph OSDs have _nothing_ to
  30  * do with the T10 OSD extensions to SCSI.)  Ceph OSDs are simply
  31  * remote daemons serving up and coordinating consistent and safe
  32  * access to storage.
  33  *
  34  * Cluster membership and the mapping of data objects onto storage devices
  35  * are described by the osd map.
  36  *
  37  * We keep track of pending OSD requests (read, write), resubmit
  38  * requests to different OSDs when the cluster topology/data layout
  39  * change, or retry the affected requests when the communications
  40  * channel with an OSD is reset.
  41  */
  42
  43 /*
  44  * calculate the mapping of a file extent onto an object, and fill out the
  45  * request accordingly.  shorten extent as necessary if it crosses an
  46  * object boundary.
  47  *
  48  * fill osd op in request message.
  49  */
  50 static void calc_layout(struct ceph_osd_client *osdc,
  51                         struct ceph_vino vino, struct ceph_file_layout *layout,
  52                         u64 off, u64 *plen,
  53                         struct ceph_osd_request *req)
  54 {
  55         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
  56         struct ceph_osd_op *op = (void *)(reqhead + 1);
  57         u64 orig_len = *plen;
  58         u64 objoff, objlen;    /* extent in object */
  59         u64 bno;
  60
  61         reqhead->snapid = cpu_to_le64(vino.snap);
  62
  63         /* object extent? */
  64         ceph_calc_file_object_mapping(layout, off, plen, &bno,
  65                                       &objoff, &objlen);
  66         if (*plen < orig_len)
  67                 dout(" skipping last %llu, final file extent %llu~%llu\n",
  68                      orig_len - *plen, off, *plen);
  69
  70         sprintf(req->r_oid, "%llx.%08llx", vino.ino, bno);
  71         req->r_oid_len = strlen(req->r_oid);
  72
  73         op->extent.offset = cpu_to_le64(objoff);
  74         op->extent.length = cpu_to_le64(objlen);
  75         req->r_num_pages = calc_pages_for(off, *plen);
  76
  77         dout("calc_layout %s (%d) %llu~%llu (%d pages)\n",
  78              req->r_oid, req->r_oid_len, objoff, objlen, req->r_num_pages);
  79 }
  80
  81 /*
  82  * requests
  83  */
  84 void ceph_osdc_release_request(struct kref *kref)
  85 {
  86         struct ceph_osd_request *req = container_of(kref,
  87                                                     struct ceph_osd_request,
  88                                                     r_kref);
  89
  90         if (req->r_request)
  91                 ceph_msg_put(req->r_request);
  92         if (req->r_reply)
  93                 ceph_msg_put(req->r_reply);
  94         if (req->r_con_filling_msg) {
  95                 dout("release_request revoking pages %p from con %p\n",
  96                      req->r_pages, req->r_con_filling_msg);
  97                 ceph_con_revoke_message(req->r_con_filling_msg,
  98                                       req->r_reply);
  99                 ceph_con_put(req->r_con_filling_msg);
 100         }
 101         if (req->r_own_pages)
 102                 ceph_release_page_vector(req->r_pages,
 103                                          req->r_num_pages);
 104         ceph_put_snap_context(req->r_snapc);
 105         if (req->r_mempool)
 106                 mempool_free(req, req->r_osdc->req_mempool);
 107         else
 108                 kfree(req);
 109 }
 110
 111 /*
 112  * build new request AND message, calculate layout, and adjust file
 113  * extent as needed.
 114  *
 115  * if the file was recently truncated, we include information about its
 116  * old and new size so that the object can be updated appropriately.  (we
 117  * avoid synchronously deleting truncated objects because it's slow.)
 118  *
 119  * if @do_sync, include a 'startsync' command so that the osd will flush
 120  * data quickly.
 121  */
 122 struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *osdc,
 123                                                struct ceph_file_layout *layout,
 124                                                struct ceph_vino vino,
 125                                                u64 off, u64 *plen,
 126                                                int opcode, int flags,
 127                                                struct ceph_snap_context *snapc,
 128                                                int do_sync,
 129                                                u32 truncate_seq,
 130                                                u64 truncate_size,
 131                                                struct timespec *mtime,
 132                                                bool use_mempool, int num_reply)
 133 {
 134         struct ceph_osd_request *req;
 135         struct ceph_msg *msg;
 136         struct ceph_osd_request_head *head;
 137         struct ceph_osd_op *op;
 138         void *p;
 139         int num_op = 1 + do_sync;
 140         size_t msg_size = sizeof(*head) + num_op*sizeof(*op);
 141         int i;
 142
 143         if (use_mempool) {
 144                 req = mempool_alloc(osdc->req_mempool, GFP_NOFS);
 145                 memset(req, 0, sizeof(*req));
 146         } else {
 147                 req = kzalloc(sizeof(*req), GFP_NOFS);
 148         }
 149         if (req == NULL)
 150                 return NULL;
 151
 152         req->r_osdc = osdc;
 153         req->r_mempool = use_mempool;
 154         kref_init(&req->r_kref);
 155         init_completion(&req->r_completion);
 156         init_completion(&req->r_safe_completion);
 157         INIT_LIST_HEAD(&req->r_unsafe_item);
 158         req->r_flags = flags;
 159
 160         WARN_ON((flags & (CEPH_OSD_FLAG_READ|CEPH_OSD_FLAG_WRITE)) == 0);
 161
 162         /* create reply message */
 163         if (use_mempool)
 164                 msg = ceph_msgpool_get(&osdc->msgpool_op_reply, 0);
 165         else
 166                 msg = ceph_msg_new(CEPH_MSG_OSD_OPREPLY,
 167                                    OSD_OPREPLY_FRONT_LEN, GFP_NOFS);
 168         if (!msg) {
 169                 ceph_osdc_put_request(req);
 170                 return NULL;
 171         }
 172         req->r_reply = msg;
 173
 174         /* create request message; allow space for oid */
 175         msg_size += 40;
 176         if (snapc)
 177                 msg_size += sizeof(u64) * snapc->num_snaps;
 178         if (use_mempool)
 179                 msg = ceph_msgpool_get(&osdc->msgpool_op, 0);
 180         else
 181                 msg = ceph_msg_new(CEPH_MSG_OSD_OP, msg_size, GFP_NOFS);
 182         if (!msg) {
 183                 ceph_osdc_put_request(req);
 184                 return NULL;
 185         }
 186         msg->hdr.type = cpu_to_le16(CEPH_MSG_OSD_OP);
 187         memset(msg->front.iov_base, 0, msg->front.iov_len);
 188         head = msg->front.iov_base;
 189         op = (void *)(head + 1);
 190         p = (void *)(op + num_op);
 191
 192         req->r_request = msg;
 193         req->r_snapc = ceph_get_snap_context(snapc);
 194
 195         head->client_inc = cpu_to_le32(1); /* always, for now. */
 196         head->flags = cpu_to_le32(flags);
 197         if (flags & CEPH_OSD_FLAG_WRITE)
 198                 ceph_encode_timespec(&head->mtime, mtime);
 199         head->num_ops = cpu_to_le16(num_op);
 200         op->op = cpu_to_le16(opcode);
 201
 202         /* calculate max write size */
 203         calc_layout(osdc, vino, layout, off, plen, req);
 204         req->r_file_layout = *layout;  /* keep a copy */
 205
 206         if (flags & CEPH_OSD_FLAG_WRITE) {
 207                 req->r_request->hdr.data_off = cpu_to_le16(off);
 208                 req->r_request->hdr.data_len = cpu_to_le32(*plen);
 209                 op->payload_len = cpu_to_le32(*plen);
 210         }
 211         op->extent.truncate_size = cpu_to_le64(truncate_size);
 212         op->extent.truncate_seq = cpu_to_le32(truncate_seq);
 213
 214         /* fill in oid */
 215         head->object_len = cpu_to_le32(req->r_oid_len);
 216         memcpy(p, req->r_oid, req->r_oid_len);
 217         p += req->r_oid_len;
 218
 219         if (do_sync) {
 220                 op++;
 221                 op->op = cpu_to_le16(CEPH_OSD_OP_STARTSYNC);
 222         }
 223         if (snapc) {
 224                 head->snap_seq = cpu_to_le64(snapc->seq);
 225                 head->num_snaps = cpu_to_le32(snapc->num_snaps);
 226                 for (i = 0; i < snapc->num_snaps; i++) {
 227                         put_unaligned_le64(snapc->snaps[i], p);
 228                         p += sizeof(u64);
 229                 }
 230         }
 231
 232         BUG_ON(p > msg->front.iov_base + msg->front.iov_len);
 233         msg_size = p - msg->front.iov_base;
 234         msg->front.iov_len = msg_size;
 235         msg->hdr.front_len = cpu_to_le32(msg_size);
 236         return req;
 237 }
 238
 239 /*
 240  * We keep osd requests in an rbtree, sorted by ->r_tid.
 241  */
 242 static void __insert_request(struct ceph_osd_client *osdc,
 243                              struct ceph_osd_request *new)
 244 {
 245         struct rb_node **p = &osdc->requests.rb_node;
 246         struct rb_node *parent = NULL;
 247         struct ceph_osd_request *req = NULL;
 248
 249         while (*p) {
 250                 parent = *p;
 251                 req = rb_entry(parent, struct ceph_osd_request, r_node);
 252                 if (new->r_tid < req->r_tid)
 253                         p = &(*p)->rb_left;
 254                 else if (new->r_tid > req->r_tid)
 255                         p = &(*p)->rb_right;
 256                 else
 257                         BUG();
 258         }
 259
 260         rb_link_node(&new->r_node, parent, p);
 261         rb_insert_color(&new->r_node, &osdc->requests);
 262 }
 263
 264 static struct ceph_osd_request *__lookup_request(struct ceph_osd_client *osdc,
 265                                                  u64 tid)
 266 {
 267         struct ceph_osd_request *req;
 268         struct rb_node *n = osdc->requests.rb_node;
 269
 270         while (n) {
 271                 req = rb_entry(n, struct ceph_osd_request, r_node);
 272                 if (tid < req->r_tid)
 273                         n = n->rb_left;
 274                 else if (tid > req->r_tid)
 275                         n = n->rb_right;
 276                 else
 277                         return req;
 278         }
 279         return NULL;
 280 }
 281
 282 static struct ceph_osd_request *
 283 __lookup_request_ge(struct ceph_osd_client *osdc,
 284                     u64 tid)
 285 {
 286         struct ceph_osd_request *req;
 287         struct rb_node *n = osdc->requests.rb_node;
 288
 289         while (n) {
 290                 req = rb_entry(n, struct ceph_osd_request, r_node);
 291                 if (tid < req->r_tid) {
 292                         if (!n->rb_left)
 293                                 return req;
 294                         n = n->rb_left;
 295                 } else if (tid > req->r_tid) {
 296                         n = n->rb_right;
 297                 } else {
 298                         return req;
 299                 }
 300         }
 301         return NULL;
 302 }
 303
 304
 305 /*
 306  * If the osd connection drops, we need to resubmit all requests.
 307  */
 308 static void osd_reset(struct ceph_connection *con)
 309 {
 310         struct ceph_osd *osd = con->private;
 311         struct ceph_osd_client *osdc;
 312
 313         if (!osd)
 314                 return;
 315         dout("osd_reset osd%d\n", osd->o_osd);
 316         osdc = osd->o_osdc;
 317         down_read(&osdc->map_sem);
 318         kick_requests(osdc, osd);
 319         up_read(&osdc->map_sem);
 320 }
 321
 322 /*
 323  * Track open sessions with osds.
 324  */
 325 static struct ceph_osd *create_osd(struct ceph_osd_client *osdc)
 326 {
 327         struct ceph_osd *osd;
 328
 329         osd = kzalloc(sizeof(*osd), GFP_NOFS);
 330         if (!osd)
 331                 return NULL;
 332
 333         atomic_set(&osd->o_ref, 1);
 334         osd->o_osdc = osdc;
 335         INIT_LIST_HEAD(&osd->o_requests);
 336         INIT_LIST_HEAD(&osd->o_osd_lru);
 337         osd->o_incarnation = 1;
 338
 339         ceph_con_init(osdc->client->msgr, &osd->o_con);
 340         osd->o_con.private = osd;
 341         osd->o_con.ops = &osd_con_ops;
 342         osd->o_con.peer_name.type = CEPH_ENTITY_TYPE_OSD;
 343
 344         INIT_LIST_HEAD(&osd->o_keepalive_item);
 345         return osd;
 346 }
 347
 348 static struct ceph_osd *get_osd(struct ceph_osd *osd)
 349 {
 350         if (atomic_inc_not_zero(&osd->o_ref)) {
 351                 dout("get_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref)-1,
 352                      atomic_read(&osd->o_ref));
 353                 return osd;
 354         } else {
 355                 dout("get_osd %p FAIL\n", osd);
 356                 return NULL;
 357         }
 358 }
 359
 360 static void put_osd(struct ceph_osd *osd)
 361 {
 362         dout("put_osd %p %d -> %d\n", osd, atomic_read(&osd->o_ref),
 363              atomic_read(&osd->o_ref) - 1);
 364         if (atomic_dec_and_test(&osd->o_ref)) {
 365                 struct ceph_auth_client *ac = osd->o_osdc->client->monc.auth;
 366
 367                 if (osd->o_authorizer)
 368                         ac->ops->destroy_authorizer(ac, osd->o_authorizer);
 369                 kfree(osd);
 370         }
 371 }
 372
 373 /*
 374  * remove an osd from our map
 375  */
 376 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 377 {
 378         dout("__remove_osd %p\n", osd);
 379         BUG_ON(!list_empty(&osd->o_requests));
 380         rb_erase(&osd->o_node, &osdc->osds);
 381         list_del_init(&osd->o_osd_lru);
 382         ceph_con_close(&osd->o_con);
 383         put_osd(osd);
 384 }
 385
 386 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 387                               struct ceph_osd *osd)
 388 {
 389         dout("__move_osd_to_lru %p\n", osd);
 390         BUG_ON(!list_empty(&osd->o_osd_lru));
 391         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 392         osd->lru_ttl = jiffies + osdc->client->mount_args->osd_idle_ttl * HZ;
 393 }
 394
 395 static void __remove_osd_from_lru(struct ceph_osd *osd)
 396 {
 397         dout("__remove_osd_from_lru %p\n", osd);
 398         if (!list_empty(&osd->o_osd_lru))
 399                 list_del_init(&osd->o_osd_lru);
 400 }
 401
 402 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
 403 {
 404         struct ceph_osd *osd, *nosd;
 405
 406         dout("__remove_old_osds %p\n", osdc);
 407         mutex_lock(&osdc->request_mutex);
 408         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 409                 if (!remove_all && time_before(jiffies, osd->lru_ttl))
 410                         break;
 411                 __remove_osd(osdc, osd);
 412         }
 413         mutex_unlock(&osdc->request_mutex);
 414 }
 415
 416 /*
 417  * reset osd connect
 418  */
 419 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 420 {
 421         struct ceph_osd_request *req;
 422         int ret = 0;
 423
 424         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 425         if (list_empty(&osd->o_requests)) {
 426                 __remove_osd(osdc, osd);
 427         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 428                           &osd->o_con.peer_addr,
 429                           sizeof(osd->o_con.peer_addr)) == 0 &&
 430                    !ceph_con_opened(&osd->o_con)) {
 431                 dout(" osd addr hasn't changed and connection never opened,"
 432                      " letting msgr retry");
 433                 /* touch each r_stamp for handle_timeout()'s benfit */
 434                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 435                         req->r_stamp = jiffies;
 436                 ret = -EAGAIN;
 437         } else {
 438                 ceph_con_close(&osd->o_con);
 439                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
 440                 osd->o_incarnation++;
 441         }
 442         return ret;
 443 }
 444
 445 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 446 {
 447         struct rb_node **p = &osdc->osds.rb_node;
 448         struct rb_node *parent = NULL;
 449         struct ceph_osd *osd = NULL;
 450
 451         while (*p) {
 452                 parent = *p;
 453                 osd = rb_entry(parent, struct ceph_osd, o_node);
 454                 if (new->o_osd < osd->o_osd)
 455                         p = &(*p)->rb_left;
 456                 else if (new->o_osd > osd->o_osd)
 457                         p = &(*p)->rb_right;
 458                 else
 459                         BUG();
 460         }
 461
 462         rb_link_node(&new->o_node, parent, p);
 463         rb_insert_color(&new->o_node, &osdc->osds);
 464 }
 465
 466 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 467 {
 468         struct ceph_osd *osd;
 469         struct rb_node *n = osdc->osds.rb_node;
 470
 471         while (n) {
 472                 osd = rb_entry(n, struct ceph_osd, o_node);
 473                 if (o < osd->o_osd)
 474                         n = n->rb_left;
 475                 else if (o > osd->o_osd)
 476                         n = n->rb_right;
 477                 else
 478                         return osd;
 479         }
 480         return NULL;
 481 }
 482
 483 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 484 {
 485         schedule_delayed_work(&osdc->timeout_work,
 486                         osdc->client->mount_args->osd_keepalive_timeout * HZ);
 487 }
 488
 489 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 490 {
 491         cancel_delayed_work(&osdc->timeout_work);
 492 }
 493
 494 /*
 495  * Register request, assign tid.  If this is the first request, set up
 496  * the timeout event.
 497  */
 498 static void register_request(struct ceph_osd_client *osdc,
 499                              struct ceph_osd_request *req)
 500 {
 501         mutex_lock(&osdc->request_mutex);
 502         req->r_tid = ++osdc->last_tid;
 503         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 504         INIT_LIST_HEAD(&req->r_req_lru_item);
 505
 506         dout("register_request %p tid %lld\n", req, req->r_tid);
 507         __insert_request(osdc, req);
 508         ceph_osdc_get_request(req);
 509         osdc->num_requests++;
 510
 511         if (osdc->num_requests == 1) {
 512                 dout(" first request, scheduling timeout\n");
 513                 __schedule_osd_timeout(osdc);
 514         }
 515         mutex_unlock(&osdc->request_mutex);
 516 }
 517
 518 /*
 519  * called under osdc->request_mutex
 520  */
 521 static void __unregister_request(struct ceph_osd_client *osdc,
 522                                  struct ceph_osd_request *req)
 523 {
 524         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 525         rb_erase(&req->r_node, &osdc->requests);
 526         osdc->num_requests--;
 527
 528         if (req->r_osd) {
 529                 /* make sure the original request isn't in flight. */
 530                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 531
 532                 list_del_init(&req->r_osd_item);
 533                 if (list_empty(&req->r_osd->o_requests))
 534                         __move_osd_to_lru(osdc, req->r_osd);
 535                 req->r_osd = NULL;
 536         }
 537
 538         ceph_osdc_put_request(req);
 539
 540         list_del_init(&req->r_req_lru_item);
 541         if (osdc->num_requests == 0) {
 542                 dout(" no requests, canceling timeout\n");
 543                 __cancel_osd_timeout(osdc);
 544         }
 545 }
 546
 547 /*
 548  * Cancel a previously queued request message
 549  */
 550 static void __cancel_request(struct ceph_osd_request *req)
 551 {
 552         if (req->r_sent) {
 553                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 554                 req->r_sent = 0;
 555         }
 556         list_del_init(&req->r_req_lru_item);
 557 }
 558
 559 /*
 560  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 561  * (as needed), and set the request r_osd appropriately.  If there is
 562  * no up osd, set r_osd to NULL.
 563  *
 564  * Return 0 if unchanged, 1 if changed, or negative on error.
 565  *
 566  * Caller should hold map_sem for read and request_mutex.
 567  */
 568 static int __map_osds(struct ceph_osd_client *osdc,
 569                       struct ceph_osd_request *req)
 570 {
 571         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 572         struct ceph_pg pgid;
 573         int acting[CEPH_PG_MAX_SIZE];
 574         int o = -1, num = 0;
 575         int err;
 576
 577         dout("map_osds %p tid %lld\n", req, req->r_tid);
 578         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 579                                       &req->r_file_layout, osdc->osdmap);
 580         if (err)
 581                 return err;
 582         pgid = reqhead->layout.ol_pgid;
 583         req->r_pgid = pgid;
 584
 585         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 586         if (err > 0) {
 587                 o = acting[0];
 588                 num = err;
 589         }
 590
 591         if ((req->r_osd && req->r_osd->o_osd == o &&
 592              req->r_sent >= req->r_osd->o_incarnation &&
 593              req->r_num_pg_osds == num &&
 594              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 595             (req->r_osd == NULL && o == -1))
 596                 return 0;  /* no change */
 597
 598         dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
 599              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 600              req->r_osd ? req->r_osd->o_osd : -1);
 601
 602         /* record full pg acting set */
 603         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 604         req->r_num_pg_osds = num;
 605
 606         if (req->r_osd) {
 607                 __cancel_request(req);
 608                 list_del_init(&req->r_osd_item);
 609                 req->r_osd = NULL;
 610         }
 611
 612         req->r_osd = __lookup_osd(osdc, o);
 613         if (!req->r_osd && o >= 0) {
 614                 err = -ENOMEM;
 615                 req->r_osd = create_osd(osdc);
 616                 if (!req->r_osd)
 617                         goto out;
 618
 619                 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
 620                 req->r_osd->o_osd = o;
 621                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
 622                 __insert_osd(osdc, req->r_osd);
 623
 624                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
 625         }
 626
 627         if (req->r_osd) {
 628                 __remove_osd_from_lru(req->r_osd);
 629                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 630         }
 631         err = 1;   /* osd or pg changed */
 632
 633 out:
 634         return err;
 635 }
 636
 637 /*
 638  * caller should hold map_sem (for read) and request_mutex
 639  */
 640 static int __send_request(struct ceph_osd_client *osdc,
 641                           struct ceph_osd_request *req)
 642 {
 643         struct ceph_osd_request_head *reqhead;
 644         int err;
 645
 646         err = __map_osds(osdc, req);
 647         if (err < 0)
 648                 return err;
 649         if (req->r_osd == NULL) {
 650                 dout("send_request %p no up osds in pg\n", req);
 651                 ceph_monc_request_next_osdmap(&osdc->client->monc);
 652                 return 0;
 653         }
 654
 655         dout("send_request %p tid %llu to osd%d flags %d\n",
 656              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
 657
 658         reqhead = req->r_request->front.iov_base;
 659         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
 660         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
 661         reqhead->reassert_version = req->r_reassert_version;
 662
 663         req->r_stamp = jiffies;
 664         list_move_tail(&req->r_req_lru_item, &osdc->req_lru);
 665
 666         ceph_msg_get(req->r_request); /* send consumes a ref */
 667         ceph_con_send(&req->r_osd->o_con, req->r_request);
 668         req->r_sent = req->r_osd->o_incarnation;
 669         return 0;
 670 }
 671
 672 /*
 673  * Timeout callback, called every N seconds when 1 or more osd
 674  * requests has been active for more than N seconds.  When this
 675  * happens, we ping all OSDs with requests who have timed out to
 676  * ensure any communications channel reset is detected.  Reset the
 677  * request timeouts another N seconds in the future as we go.
 678  * Reschedule the timeout event another N seconds in future (unless
 679  * there are no open requests).
 680  */
 681 static void handle_timeout(struct work_struct *work)
 682 {
 683         struct ceph_osd_client *osdc =
 684                 container_of(work, struct ceph_osd_client, timeout_work.work);
 685         struct ceph_osd_request *req, *last_req = NULL;
 686         struct ceph_osd *osd;
 687         unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
 688         unsigned long keepalive =
 689                 osdc->client->mount_args->osd_keepalive_timeout * HZ;
 690         unsigned long last_stamp = 0;
 691         struct rb_node *p;
 692         struct list_head slow_osds;
 693
 694         dout("timeout\n");
 695         down_read(&osdc->map_sem);
 696
 697         ceph_monc_request_next_osdmap(&osdc->client->monc);
 698
 699         mutex_lock(&osdc->request_mutex);
 700         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
 701                 req = rb_entry(p, struct ceph_osd_request, r_node);
 702
 703                 if (req->r_resend) {
 704                         int err;
 705
 706                         dout("osdc resending prev failed %lld\n", req->r_tid);
 707                         err = __send_request(osdc, req);
 708                         if (err)
 709                                 dout("osdc failed again on %lld\n", req->r_tid);
 710                         else
 711                                 req->r_resend = false;
 712                         continue;
 713                 }
 714         }
 715
 716         /*
 717          * reset osds that appear to be _really_ unresponsive.  this
 718          * is a failsafe measure.. we really shouldn't be getting to
 719          * this point if the system is working properly.  the monitors
 720          * should mark the osd as failed and we should find out about
 721          * it from an updated osd map.
 722          */
 723         while (timeout && !list_empty(&osdc->req_lru)) {
 724                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
 725                                  r_req_lru_item);
 726
 727                 if (time_before(jiffies, req->r_stamp + timeout))
 728                         break;
 729
 730                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
 731                 last_req = req;
 732                 last_stamp = req->r_stamp;
 733
 734                 osd = req->r_osd;
 735                 BUG_ON(!osd);
 736                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
 737                            req->r_tid, osd->o_osd);
 738                 __kick_requests(osdc, osd);
 739         }
 740
 741         /*
 742          * ping osds that are a bit slow.  this ensures that if there
 743          * is a break in the TCP connection we will notice, and reopen
 744          * a connection with that osd (from the fault callback).
 745          */
 746         INIT_LIST_HEAD(&slow_osds);
 747         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
 748                 if (time_before(jiffies, req->r_stamp + keepalive))
 749                         break;
 750
 751                 osd = req->r_osd;
 752                 BUG_ON(!osd);
 753                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
 754                      req->r_tid, osd->o_osd);
 755                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
 756         }
 757         while (!list_empty(&slow_osds)) {
 758                 osd = list_entry(slow_osds.next, struct ceph_osd,
 759                                  o_keepalive_item);
 760                 list_del_init(&osd->o_keepalive_item);
 761                 ceph_con_keepalive(&osd->o_con);
 762         }
 763
 764         __schedule_osd_timeout(osdc);
 765         mutex_unlock(&osdc->request_mutex);
 766
 767         up_read(&osdc->map_sem);
 768 }
 769
 770 static void handle_osds_timeout(struct work_struct *work)
 771 {
 772         struct ceph_osd_client *osdc =
 773                 container_of(work, struct ceph_osd_client,
 774                              osds_timeout_work.work);
 775         unsigned long delay =
 776                 osdc->client->mount_args->osd_idle_ttl * HZ >> 2;
 777
 778         dout("osds timeout\n");
 779         down_read(&osdc->map_sem);
 780         remove_old_osds(osdc, 0);
 781         up_read(&osdc->map_sem);
 782
 783         schedule_delayed_work(&osdc->osds_timeout_work,
 784                               round_jiffies_relative(delay));
 785 }
 786
 787 /*
 788  * handle osd op reply.  either call the callback if it is specified,
 789  * or do the completion to wake up the waiting thread.
 790  */
 791 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
 792                          struct ceph_connection *con)
 793 {
 794         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
 795         struct ceph_osd_request *req;
 796         u64 tid;
 797         int numops, object_len, flags;
 798         s32 result;
 799
 800         tid = le64_to_cpu(msg->hdr.tid);
 801         if (msg->front.iov_len < sizeof(*rhead))
 802                 goto bad;
 803         numops = le32_to_cpu(rhead->num_ops);
 804         object_len = le32_to_cpu(rhead->object_len);
 805         result = le32_to_cpu(rhead->result);
 806         if (msg->front.iov_len != sizeof(*rhead) + object_len +
 807             numops * sizeof(struct ceph_osd_op))
 808                 goto bad;
 809         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
 810
 811         /* lookup */
 812         mutex_lock(&osdc->request_mutex);
 813         req = __lookup_request(osdc, tid);
 814         if (req == NULL) {
 815                 dout("handle_reply tid %llu dne\n", tid);
 816                 mutex_unlock(&osdc->request_mutex);
 817                 return;
 818         }
 819         ceph_osdc_get_request(req);
 820         flags = le32_to_cpu(rhead->flags);
 821
 822         /*
 823          * if this connection filled our message, drop our reference now, to
 824          * avoid a (safe but slower) revoke later.
 825          */
 826         if (req->r_con_filling_msg == con && req->r_reply == msg) {
 827                 dout(" dropping con_filling_msg ref %p\n", con);
 828                 req->r_con_filling_msg = NULL;
 829                 ceph_con_put(con);
 830         }
 831
 832         if (!req->r_got_reply) {
 833                 unsigned bytes;
 834
 835                 req->r_result = le32_to_cpu(rhead->result);
 836                 bytes = le32_to_cpu(msg->hdr.data_len);
 837                 dout("handle_reply result %d bytes %d\n", req->r_result,
 838                      bytes);
 839                 if (req->r_result == 0)
 840                         req->r_result = bytes;
 841
 842                 /* in case this is a write and we need to replay, */
 843                 req->r_reassert_version = rhead->reassert_version;
 844
 845                 req->r_got_reply = 1;
 846         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
 847                 dout("handle_reply tid %llu dup ack\n", tid);
 848                 mutex_unlock(&osdc->request_mutex);
 849                 goto done;
 850         }
 851
 852         dout("handle_reply tid %llu flags %d\n", tid, flags);
 853
 854         /* either this is a read, or we got the safe response */
 855         if (result < 0 ||
 856             (flags & CEPH_OSD_FLAG_ONDISK) ||
 857             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
 858                 __unregister_request(osdc, req);
 859
 860         mutex_unlock(&osdc->request_mutex);
 861
 862         if (req->r_callback)
 863                 req->r_callback(req, msg);
 864         else
 865                 complete_all(&req->r_completion);
 866
 867         if (flags & CEPH_OSD_FLAG_ONDISK) {
 868                 if (req->r_safe_callback)
 869                         req->r_safe_callback(req, msg);
 870                 complete_all(&req->r_safe_completion);  /* fsync waiter */
 871         }
 872
 873 done:
 874         ceph_osdc_put_request(req);
 875         return;
 876
 877 bad:
 878         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
 879                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
 880                (int)sizeof(*rhead));
 881         ceph_msg_dump(msg);
 882 }
 883
 884
 885 static int __kick_requests(struct ceph_osd_client *osdc,
 886                           struct ceph_osd *kickosd)
 887 {
 888         struct ceph_osd_request *req;
 889         struct rb_node *p, *n;
 890         int needmap = 0;
 891         int err;
 892
 893         dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
 894         if (kickosd) {
 895                 err = __reset_osd(osdc, kickosd);
 896                 if (err == -EAGAIN)
 897                         return 1;
 898         } else {
 899                 for (p = rb_first(&osdc->osds); p; p = n) {
 900                         struct ceph_osd *osd =
 901                                 rb_entry(p, struct ceph_osd, o_node);
 902
 903                         n = rb_next(p);
 904                         if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
 905                             memcmp(&osd->o_con.peer_addr,
 906                                    ceph_osd_addr(osdc->osdmap,
 907                                                  osd->o_osd),
 908                                    sizeof(struct ceph_entity_addr)) != 0)
 909                                 __reset_osd(osdc, osd);
 910                 }
 911         }
 912
 913         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
 914                 req = rb_entry(p, struct ceph_osd_request, r_node);
 915
 916                 if (req->r_resend) {
 917                         dout(" r_resend set on tid %llu\n", req->r_tid);
 918                         __cancel_request(req);
 919                         goto kick;
 920                 }
 921                 if (req->r_osd && kickosd == req->r_osd) {
 922                         __cancel_request(req);
 923                         goto kick;
 924                 }
 925
 926                 err = __map_osds(osdc, req);
 927                 if (err == 0)
 928                         continue;  /* no change */
 929                 if (err < 0) {
 930                         /*
 931                          * FIXME: really, we should set the request
 932                          * error and fail if this isn't a 'nofail'
 933                          * request, but that's a fair bit more
 934                          * complicated to do.  So retry!
 935                          */
 936                         dout(" setting r_resend on %llu\n", req->r_tid);
 937                         req->r_resend = true;
 938                         continue;
 939                 }
 940                 if (req->r_osd == NULL) {
 941                         dout("tid %llu maps to no valid osd\n", req->r_tid);
 942                         needmap++;  /* request a newer map */
 943                         continue;
 944                 }
 945
 946 kick:
 947                 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
 948                      req->r_osd ? req->r_osd->o_osd : -1);
 949                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
 950                 err = __send_request(osdc, req);
 951                 if (err) {
 952                         dout(" setting r_resend on %llu\n", req->r_tid);
 953                         req->r_resend = true;
 954                 }
 955         }
 956
 957         return needmap;
 958 }
 959
 960 /*
 961  * Resubmit osd requests whose osd or osd address has changed.  Request
 962  * a new osd map if osds are down, or we are otherwise unable to determine
 963  * how to direct a request.
 964  *
 965  * Close connections to down osds.
 966  *
 967  * If @who is specified, resubmit requests for that specific osd.
 968  *
 969  * Caller should hold map_sem for read and request_mutex.
 970  */
 971 static void kick_requests(struct ceph_osd_client *osdc,
 972                           struct ceph_osd *kickosd)
 973 {
 974         int needmap;
 975
 976         mutex_lock(&osdc->request_mutex);
 977         needmap = __kick_requests(osdc, kickosd);
 978         mutex_unlock(&osdc->request_mutex);
 979
 980         if (needmap) {
 981                 dout("%d requests for down osds, need new map\n", needmap);
 982                 ceph_monc_request_next_osdmap(&osdc->client->monc);
 983         }
 984
 985 }
 986 /*
 987  * Process updated osd map.
 988  *
 989  * The message contains any number of incremental and full maps, normally
 990  * indicating some sort of topology change in the cluster.  Kick requests
 991  * off to different OSDs as needed.
 992  */
 993 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
 994 {
 995         void *p, *end, *next;
 996         u32 nr_maps, maplen;
 997         u32 epoch;
 998         struct ceph_osdmap *newmap = NULL, *oldmap;
 999         int err;
1000         struct ceph_fsid fsid;
1001
1002         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
1003         p = msg->front.iov_base;
1004         end = p + msg->front.iov_len;
1005
1006         /* verify fsid */
1007         ceph_decode_need(&p, end, sizeof(fsid), bad);
1008         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1009         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1010                 return;
1011
1012         down_write(&osdc->map_sem);
1013
1014         /* incremental maps */
1015         ceph_decode_32_safe(&p, end, nr_maps, bad);
1016         dout(" %d inc maps\n", nr_maps);
1017         while (nr_maps > 0) {
1018                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1019                 epoch = ceph_decode_32(&p);
1020                 maplen = ceph_decode_32(&p);
1021                 ceph_decode_need(&p, end, maplen, bad);
1022                 next = p + maplen;
1023                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1024                         dout("applying incremental map %u len %d\n",
1025                              epoch, maplen);
1026                         newmap = osdmap_apply_incremental(&p, next,
1027                                                           osdc->osdmap,
1028                                                           osdc->client->msgr);
1029                         if (IS_ERR(newmap)) {
1030                                 err = PTR_ERR(newmap);
1031                                 goto bad;
1032                         }
1033                         BUG_ON(!newmap);
1034                         if (newmap != osdc->osdmap) {
1035                                 ceph_osdmap_destroy(osdc->osdmap);
1036                                 osdc->osdmap = newmap;
1037                         }
1038                 } else {
1039                         dout("ignoring incremental map %u len %d\n",
1040                              epoch, maplen);
1041                 }
1042                 p = next;
1043                 nr_maps--;
1044         }
1045         if (newmap)
1046                 goto done;
1047
1048         /* full maps */
1049         ceph_decode_32_safe(&p, end, nr_maps, bad);
1050         dout(" %d full maps\n", nr_maps);
1051         while (nr_maps) {
1052                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1053                 epoch = ceph_decode_32(&p);
1054                 maplen = ceph_decode_32(&p);
1055                 ceph_decode_need(&p, end, maplen, bad);
1056                 if (nr_maps > 1) {
1057                         dout("skipping non-latest full map %u len %d\n",
1058                              epoch, maplen);
1059                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1060                         dout("skipping full map %u len %d, "
1061                              "older than our %u\n", epoch, maplen,
1062                              osdc->osdmap->epoch);
1063                 } else {
1064                         dout("taking full map %u len %d\n", epoch, maplen);
1065                         newmap = osdmap_decode(&p, p+maplen);
1066                         if (IS_ERR(newmap)) {
1067                                 err = PTR_ERR(newmap);
1068                                 goto bad;
1069                         }
1070                         BUG_ON(!newmap);
1071                         oldmap = osdc->osdmap;
1072                         osdc->osdmap = newmap;
1073                         if (oldmap)
1074                                 ceph_osdmap_destroy(oldmap);
1075                 }
1076                 p += maplen;
1077                 nr_maps--;
1078         }
1079
1080 done:
1081         downgrade_write(&osdc->map_sem);
1082         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1083         if (newmap)
1084                 kick_requests(osdc, NULL);
1085         up_read(&osdc->map_sem);
1086         wake_up_all(&osdc->client->auth_wq);
1087         return;
1088
1089 bad:
1090         pr_err("osdc handle_map corrupt msg\n");
1091         ceph_msg_dump(msg);
1092         up_write(&osdc->map_sem);
1093         return;
1094 }
1095
1096 /*
1097  * Register request, send initial attempt.
1098  */
1099 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1100                             struct ceph_osd_request *req,
1101                             bool nofail)
1102 {
1103         int rc = 0;
1104
1105         req->r_request->pages = req->r_pages;
1106         req->r_request->nr_pages = req->r_num_pages;
1107
1108         register_request(osdc, req);
1109
1110         down_read(&osdc->map_sem);
1111         mutex_lock(&osdc->request_mutex);
1112         /*
1113          * a racing kick_requests() may have sent the message for us
1114          * while we dropped request_mutex above, so only send now if
1115          * the request still han't been touched yet.
1116          */
1117         if (req->r_sent == 0) {
1118                 rc = __send_request(osdc, req);
1119                 if (rc) {
1120                         if (nofail) {
1121                                 dout("osdc_start_request failed send, "
1122                                      " marking %lld\n", req->r_tid);
1123                                 req->r_resend = true;
1124                                 rc = 0;
1125                         } else {
1126                                 __unregister_request(osdc, req);
1127                         }
1128                 }
1129         }
1130         mutex_unlock(&osdc->request_mutex);
1131         up_read(&osdc->map_sem);
1132         return rc;
1133 }
1134
1135 /*
1136  * wait for a request to complete
1137  */
1138 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1139                            struct ceph_osd_request *req)
1140 {
1141         int rc;
1142
1143         rc = wait_for_completion_interruptible(&req->r_completion);
1144         if (rc < 0) {
1145                 mutex_lock(&osdc->request_mutex);
1146                 __cancel_request(req);
1147                 __unregister_request(osdc, req);
1148                 mutex_unlock(&osdc->request_mutex);
1149                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1150                 return rc;
1151         }
1152
1153         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1154         return req->r_result;
1155 }
1156
1157 /*
1158  * sync - wait for all in-flight requests to flush.  avoid starvation.
1159  */
1160 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1161 {
1162         struct ceph_osd_request *req;
1163         u64 last_tid, next_tid = 0;
1164
1165         mutex_lock(&osdc->request_mutex);
1166         last_tid = osdc->last_tid;
1167         while (1) {
1168                 req = __lookup_request_ge(osdc, next_tid);
1169                 if (!req)
1170                         break;
1171                 if (req->r_tid > last_tid)
1172                         break;
1173
1174                 next_tid = req->r_tid + 1;
1175                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1176                         continue;
1177
1178                 ceph_osdc_get_request(req);
1179                 mutex_unlock(&osdc->request_mutex);
1180                 dout("sync waiting on tid %llu (last is %llu)\n",
1181                      req->r_tid, last_tid);
1182                 wait_for_completion(&req->r_safe_completion);
1183                 mutex_lock(&osdc->request_mutex);
1184                 ceph_osdc_put_request(req);
1185         }
1186         mutex_unlock(&osdc->request_mutex);
1187         dout("sync done (thru tid %llu)\n", last_tid);
1188 }
1189
1190 /*
1191  * init, shutdown
1192  */
1193 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1194 {
1195         int err;
1196
1197         dout("init\n");
1198         osdc->client = client;
1199         osdc->osdmap = NULL;
1200         init_rwsem(&osdc->map_sem);
1201         init_completion(&osdc->map_waiters);
1202         osdc->last_requested_map = 0;
1203         mutex_init(&osdc->request_mutex);
1204         osdc->last_tid = 0;
1205         osdc->osds = RB_ROOT;
1206         INIT_LIST_HEAD(&osdc->osd_lru);
1207         osdc->requests = RB_ROOT;
1208         INIT_LIST_HEAD(&osdc->req_lru);
1209         osdc->num_requests = 0;
1210         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1211         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1212
1213         schedule_delayed_work(&osdc->osds_timeout_work,
1214            round_jiffies_relative(osdc->client->mount_args->osd_idle_ttl * HZ));
1215
1216         err = -ENOMEM;
1217         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1218                                         sizeof(struct ceph_osd_request));
1219         if (!osdc->req_mempool)
1220                 goto out;
1221
1222         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1223                                 "osd_op");
1224         if (err < 0)
1225                 goto out_mempool;
1226         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1227                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1228                                 "osd_op_reply");
1229         if (err < 0)
1230                 goto out_msgpool;
1231         return 0;
1232
1233 out_msgpool:
1234         ceph_msgpool_destroy(&osdc->msgpool_op);
1235 out_mempool:
1236         mempool_destroy(osdc->req_mempool);
1237 out:
1238         return err;
1239 }
1240
1241 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1242 {
1243         cancel_delayed_work_sync(&osdc->timeout_work);
1244         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1245         if (osdc->osdmap) {
1246                 ceph_osdmap_destroy(osdc->osdmap);
1247                 osdc->osdmap = NULL;
1248         }
1249         remove_old_osds(osdc, 1);
1250         mempool_destroy(osdc->req_mempool);
1251         ceph_msgpool_destroy(&osdc->msgpool_op);
1252         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1253 }
1254
1255 /*
1256  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1257  * *plen.  Return number of bytes read, or error.
1258  */
1259 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1260                         struct ceph_vino vino, struct ceph_file_layout *layout,
1261                         u64 off, u64 *plen,
1262                         u32 truncate_seq, u64 truncate_size,
1263                         struct page **pages, int num_pages)
1264 {
1265         struct ceph_osd_request *req;
1266         int rc = 0;
1267
1268         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1269              vino.snap, off, *plen);
1270         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1271                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1272                                     NULL, 0, truncate_seq, truncate_size, NULL,
1273                                     false, 1);
1274         if (!req)
1275                 return -ENOMEM;
1276
1277         /* it may be a short read due to an object boundary */
1278         req->r_pages = pages;
1279
1280         dout("readpages  final extent is %llu~%llu (%d pages)\n",
1281              off, *plen, req->r_num_pages);
1282
1283         rc = ceph_osdc_start_request(osdc, req, false);
1284         if (!rc)
1285                 rc = ceph_osdc_wait_request(osdc, req);
1286
1287         ceph_osdc_put_request(req);
1288         dout("readpages result %d\n", rc);
1289         return rc;
1290 }
1291
1292 /*
1293  * do a synchronous write on N pages
1294  */
1295 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1296                          struct ceph_file_layout *layout,
1297                          struct ceph_snap_context *snapc,
1298                          u64 off, u64 len,
1299                          u32 truncate_seq, u64 truncate_size,
1300                          struct timespec *mtime,
1301                          struct page **pages, int num_pages,
1302                          int flags, int do_sync, bool nofail)
1303 {
1304         struct ceph_osd_request *req;
1305         int rc = 0;
1306
1307         BUG_ON(vino.snap != CEPH_NOSNAP);
1308         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1309                                     CEPH_OSD_OP_WRITE,
1310                                     flags | CEPH_OSD_FLAG_ONDISK |
1311                                             CEPH_OSD_FLAG_WRITE,
1312                                     snapc, do_sync,
1313                                     truncate_seq, truncate_size, mtime,
1314                                     nofail, 1);
1315         if (!req)
1316                 return -ENOMEM;
1317
1318         /* it may be a short write due to an object boundary */
1319         req->r_pages = pages;
1320         dout("writepages %llu~%llu (%d pages)\n", off, len,
1321              req->r_num_pages);
1322
1323         rc = ceph_osdc_start_request(osdc, req, nofail);
1324         if (!rc)
1325                 rc = ceph_osdc_wait_request(osdc, req);
1326
1327         ceph_osdc_put_request(req);
1328         if (rc == 0)
1329                 rc = len;
1330         dout("writepages result %d\n", rc);
1331         return rc;
1332 }
1333
1334 /*
1335  * handle incoming message
1336  */
1337 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1338 {
1339         struct ceph_osd *osd = con->private;
1340         struct ceph_osd_client *osdc;
1341         int type = le16_to_cpu(msg->hdr.type);
1342
1343         if (!osd)
1344                 goto out;
1345         osdc = osd->o_osdc;
1346
1347         switch (type) {
1348         case CEPH_MSG_OSD_MAP:
1349                 ceph_osdc_handle_map(osdc, msg);
1350                 break;
1351         case CEPH_MSG_OSD_OPREPLY:
1352                 handle_reply(osdc, msg, con);
1353                 break;
1354
1355         default:
1356                 pr_err("received unknown message type %d %s\n", type,
1357                        ceph_msg_type_name(type));
1358         }
1359 out:
1360         ceph_msg_put(msg);
1361 }
1362
1363 /*
1364  * lookup and return message for incoming reply.  set up reply message
1365  * pages.
1366  */
1367 static struct ceph_msg *get_reply(struct ceph_connection *con,
1368                                   struct ceph_msg_header *hdr,
1369                                   int *skip)
1370 {
1371         struct ceph_osd *osd = con->private;
1372         struct ceph_osd_client *osdc = osd->o_osdc;
1373         struct ceph_msg *m;
1374         struct ceph_osd_request *req;
1375         int front = le32_to_cpu(hdr->front_len);
1376         int data_len = le32_to_cpu(hdr->data_len);
1377         u64 tid;
1378
1379         tid = le64_to_cpu(hdr->tid);
1380         mutex_lock(&osdc->request_mutex);
1381         req = __lookup_request(osdc, tid);
1382         if (!req) {
1383                 *skip = 1;
1384                 m = NULL;
1385                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
1386                         osd->o_osd);
1387                 goto out;
1388         }
1389
1390         if (req->r_con_filling_msg) {
1391                 dout("get_reply revoking msg %p from old con %p\n",
1392                      req->r_reply, req->r_con_filling_msg);
1393                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
1394                 ceph_con_put(req->r_con_filling_msg);
1395                 req->r_con_filling_msg = NULL;
1396         }
1397
1398         if (front > req->r_reply->front.iov_len) {
1399                 pr_warning("get_reply front %d > preallocated %d\n",
1400                            front, (int)req->r_reply->front.iov_len);
1401                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS);
1402                 if (!m)
1403                         goto out;
1404                 ceph_msg_put(req->r_reply);
1405                 req->r_reply = m;
1406         }
1407         m = ceph_msg_get(req->r_reply);
1408
1409         if (data_len > 0) {
1410                 unsigned data_off = le16_to_cpu(hdr->data_off);
1411                 int want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1412
1413                 if (unlikely(req->r_num_pages < want)) {
1414                         pr_warning("tid %lld reply %d > expected %d pages\n",
1415                                    tid, want, m->nr_pages);
1416                         *skip = 1;
1417                         ceph_msg_put(m);
1418                         m = NULL;
1419                         goto out;
1420                 }
1421                 m->pages = req->r_pages;
1422                 m->nr_pages = req->r_num_pages;
1423         }
1424         *skip = 0;
1425         req->r_con_filling_msg = ceph_con_get(con);
1426         dout("get_reply tid %lld %p\n", tid, m);
1427
1428 out:
1429         mutex_unlock(&osdc->request_mutex);
1430         return m;
1431
1432 }
1433
1434 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1435                                   struct ceph_msg_header *hdr,
1436                                   int *skip)
1437 {
1438         struct ceph_osd *osd = con->private;
1439         int type = le16_to_cpu(hdr->type);
1440         int front = le32_to_cpu(hdr->front_len);
1441
1442         switch (type) {
1443         case CEPH_MSG_OSD_MAP:
1444                 return ceph_msg_new(type, front, GFP_NOFS);
1445         case CEPH_MSG_OSD_OPREPLY:
1446                 return get_reply(con, hdr, skip);
1447         default:
1448                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
1449                         osd->o_osd);
1450                 *skip = 1;
1451                 return NULL;
1452         }
1453 }
1454
1455 /*
1456  * Wrappers to refcount containing ceph_osd struct
1457  */
1458 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1459 {
1460         struct ceph_osd *osd = con->private;
1461         if (get_osd(osd))
1462                 return con;
1463         return NULL;
1464 }
1465
1466 static void put_osd_con(struct ceph_connection *con)
1467 {
1468         struct ceph_osd *osd = con->private;
1469         put_osd(osd);
1470 }
1471
1472 /*
1473  * authentication
1474  */
1475 static int get_authorizer(struct ceph_connection *con,
1476                           void **buf, int *len, int *proto,
1477                           void **reply_buf, int *reply_len, int force_new)
1478 {
1479         struct ceph_osd *o = con->private;
1480         struct ceph_osd_client *osdc = o->o_osdc;
1481         struct ceph_auth_client *ac = osdc->client->monc.auth;
1482         int ret = 0;
1483
1484         if (force_new && o->o_authorizer) {
1485                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1486                 o->o_authorizer = NULL;
1487         }
1488         if (o->o_authorizer == NULL) {
1489                 ret = ac->ops->create_authorizer(
1490                         ac, CEPH_ENTITY_TYPE_OSD,
1491                         &o->o_authorizer,
1492                         &o->o_authorizer_buf,
1493                         &o->o_authorizer_buf_len,
1494                         &o->o_authorizer_reply_buf,
1495                         &o->o_authorizer_reply_buf_len);
1496                 if (ret)
1497                         return ret;
1498         }
1499
1500         *proto = ac->protocol;
1501         *buf = o->o_authorizer_buf;
1502         *len = o->o_authorizer_buf_len;
1503         *reply_buf = o->o_authorizer_reply_buf;
1504         *reply_len = o->o_authorizer_reply_buf_len;
1505         return 0;
1506 }
1507
1508
1509 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1510 {
1511         struct ceph_osd *o = con->private;
1512         struct ceph_osd_client *osdc = o->o_osdc;
1513         struct ceph_auth_client *ac = osdc->client->monc.auth;
1514
1515         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1516 }
1517
1518 static int invalidate_authorizer(struct ceph_connection *con)
1519 {
1520         struct ceph_osd *o = con->private;
1521         struct ceph_osd_client *osdc = o->o_osdc;
1522         struct ceph_auth_client *ac = osdc->client->monc.auth;
1523
1524         if (ac->ops->invalidate_authorizer)
1525                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
1526
1527         return ceph_monc_validate_auth(&osdc->client->monc);
1528 }
1529
1530 static const struct ceph_connection_operations osd_con_ops = {
1531         .get = get_osd_con,
1532         .put = put_osd_con,
1533         .dispatch = dispatch,
1534         .get_authorizer = get_authorizer,
1535         .verify_authorizer_reply = verify_authorizer_reply,
1536         .invalidate_authorizer = invalidate_authorizer,
1537         .alloc_msg = alloc_msg,
1538         .fault = osd_reset,
1539 };