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                 kfree(osd);
 366 }
 367
 368 /*
 369  * remove an osd from our map
 370  */
 371 static void __remove_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 372 {
 373         dout("__remove_osd %p\n", osd);
 374         BUG_ON(!list_empty(&osd->o_requests));
 375         rb_erase(&osd->o_node, &osdc->osds);
 376         list_del_init(&osd->o_osd_lru);
 377         ceph_con_close(&osd->o_con);
 378         put_osd(osd);
 379 }
 380
 381 static void __move_osd_to_lru(struct ceph_osd_client *osdc,
 382                               struct ceph_osd *osd)
 383 {
 384         dout("__move_osd_to_lru %p\n", osd);
 385         BUG_ON(!list_empty(&osd->o_osd_lru));
 386         list_add_tail(&osd->o_osd_lru, &osdc->osd_lru);
 387         osd->lru_ttl = jiffies + osdc->client->mount_args->osd_idle_ttl * HZ;
 388 }
 389
 390 static void __remove_osd_from_lru(struct ceph_osd *osd)
 391 {
 392         dout("__remove_osd_from_lru %p\n", osd);
 393         if (!list_empty(&osd->o_osd_lru))
 394                 list_del_init(&osd->o_osd_lru);
 395 }
 396
 397 static void remove_old_osds(struct ceph_osd_client *osdc, int remove_all)
 398 {
 399         struct ceph_osd *osd, *nosd;
 400
 401         dout("__remove_old_osds %p\n", osdc);
 402         mutex_lock(&osdc->request_mutex);
 403         list_for_each_entry_safe(osd, nosd, &osdc->osd_lru, o_osd_lru) {
 404                 if (!remove_all && time_before(jiffies, osd->lru_ttl))
 405                         break;
 406                 __remove_osd(osdc, osd);
 407         }
 408         mutex_unlock(&osdc->request_mutex);
 409 }
 410
 411 /*
 412  * reset osd connect
 413  */
 414 static int __reset_osd(struct ceph_osd_client *osdc, struct ceph_osd *osd)
 415 {
 416         struct ceph_osd_request *req;
 417         int ret = 0;
 418
 419         dout("__reset_osd %p osd%d\n", osd, osd->o_osd);
 420         if (list_empty(&osd->o_requests)) {
 421                 __remove_osd(osdc, osd);
 422         } else if (memcmp(&osdc->osdmap->osd_addr[osd->o_osd],
 423                           &osd->o_con.peer_addr,
 424                           sizeof(osd->o_con.peer_addr)) == 0 &&
 425                    !ceph_con_opened(&osd->o_con)) {
 426                 dout(" osd addr hasn't changed and connection never opened,"
 427                      " letting msgr retry");
 428                 /* touch each r_stamp for handle_timeout()'s benfit */
 429                 list_for_each_entry(req, &osd->o_requests, r_osd_item)
 430                         req->r_stamp = jiffies;
 431                 ret = -EAGAIN;
 432         } else {
 433                 ceph_con_close(&osd->o_con);
 434                 ceph_con_open(&osd->o_con, &osdc->osdmap->osd_addr[osd->o_osd]);
 435                 osd->o_incarnation++;
 436         }
 437         return ret;
 438 }
 439
 440 static void __insert_osd(struct ceph_osd_client *osdc, struct ceph_osd *new)
 441 {
 442         struct rb_node **p = &osdc->osds.rb_node;
 443         struct rb_node *parent = NULL;
 444         struct ceph_osd *osd = NULL;
 445
 446         while (*p) {
 447                 parent = *p;
 448                 osd = rb_entry(parent, struct ceph_osd, o_node);
 449                 if (new->o_osd < osd->o_osd)
 450                         p = &(*p)->rb_left;
 451                 else if (new->o_osd > osd->o_osd)
 452                         p = &(*p)->rb_right;
 453                 else
 454                         BUG();
 455         }
 456
 457         rb_link_node(&new->o_node, parent, p);
 458         rb_insert_color(&new->o_node, &osdc->osds);
 459 }
 460
 461 static struct ceph_osd *__lookup_osd(struct ceph_osd_client *osdc, int o)
 462 {
 463         struct ceph_osd *osd;
 464         struct rb_node *n = osdc->osds.rb_node;
 465
 466         while (n) {
 467                 osd = rb_entry(n, struct ceph_osd, o_node);
 468                 if (o < osd->o_osd)
 469                         n = n->rb_left;
 470                 else if (o > osd->o_osd)
 471                         n = n->rb_right;
 472                 else
 473                         return osd;
 474         }
 475         return NULL;
 476 }
 477
 478 static void __schedule_osd_timeout(struct ceph_osd_client *osdc)
 479 {
 480         schedule_delayed_work(&osdc->timeout_work,
 481                         osdc->client->mount_args->osd_keepalive_timeout * HZ);
 482 }
 483
 484 static void __cancel_osd_timeout(struct ceph_osd_client *osdc)
 485 {
 486         cancel_delayed_work(&osdc->timeout_work);
 487 }
 488
 489 /*
 490  * Register request, assign tid.  If this is the first request, set up
 491  * the timeout event.
 492  */
 493 static void register_request(struct ceph_osd_client *osdc,
 494                              struct ceph_osd_request *req)
 495 {
 496         mutex_lock(&osdc->request_mutex);
 497         req->r_tid = ++osdc->last_tid;
 498         req->r_request->hdr.tid = cpu_to_le64(req->r_tid);
 499         INIT_LIST_HEAD(&req->r_req_lru_item);
 500
 501         dout("register_request %p tid %lld\n", req, req->r_tid);
 502         __insert_request(osdc, req);
 503         ceph_osdc_get_request(req);
 504         osdc->num_requests++;
 505
 506         if (osdc->num_requests == 1) {
 507                 dout(" first request, scheduling timeout\n");
 508                 __schedule_osd_timeout(osdc);
 509         }
 510         mutex_unlock(&osdc->request_mutex);
 511 }
 512
 513 /*
 514  * called under osdc->request_mutex
 515  */
 516 static void __unregister_request(struct ceph_osd_client *osdc,
 517                                  struct ceph_osd_request *req)
 518 {
 519         dout("__unregister_request %p tid %lld\n", req, req->r_tid);
 520         rb_erase(&req->r_node, &osdc->requests);
 521         osdc->num_requests--;
 522
 523         if (req->r_osd) {
 524                 /* make sure the original request isn't in flight. */
 525                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 526
 527                 list_del_init(&req->r_osd_item);
 528                 if (list_empty(&req->r_osd->o_requests))
 529                         __move_osd_to_lru(osdc, req->r_osd);
 530                 req->r_osd = NULL;
 531         }
 532
 533         ceph_osdc_put_request(req);
 534
 535         list_del_init(&req->r_req_lru_item);
 536         if (osdc->num_requests == 0) {
 537                 dout(" no requests, canceling timeout\n");
 538                 __cancel_osd_timeout(osdc);
 539         }
 540 }
 541
 542 /*
 543  * Cancel a previously queued request message
 544  */
 545 static void __cancel_request(struct ceph_osd_request *req)
 546 {
 547         if (req->r_sent) {
 548                 ceph_con_revoke(&req->r_osd->o_con, req->r_request);
 549                 req->r_sent = 0;
 550         }
 551         list_del_init(&req->r_req_lru_item);
 552 }
 553
 554 /*
 555  * Pick an osd (the first 'up' osd in the pg), allocate the osd struct
 556  * (as needed), and set the request r_osd appropriately.  If there is
 557  * no up osd, set r_osd to NULL.
 558  *
 559  * Return 0 if unchanged, 1 if changed, or negative on error.
 560  *
 561  * Caller should hold map_sem for read and request_mutex.
 562  */
 563 static int __map_osds(struct ceph_osd_client *osdc,
 564                       struct ceph_osd_request *req)
 565 {
 566         struct ceph_osd_request_head *reqhead = req->r_request->front.iov_base;
 567         struct ceph_pg pgid;
 568         int acting[CEPH_PG_MAX_SIZE];
 569         int o = -1, num = 0;
 570         int err;
 571
 572         dout("map_osds %p tid %lld\n", req, req->r_tid);
 573         err = ceph_calc_object_layout(&reqhead->layout, req->r_oid,
 574                                       &req->r_file_layout, osdc->osdmap);
 575         if (err)
 576                 return err;
 577         pgid = reqhead->layout.ol_pgid;
 578         req->r_pgid = pgid;
 579
 580         err = ceph_calc_pg_acting(osdc->osdmap, pgid, acting);
 581         if (err > 0) {
 582                 o = acting[0];
 583                 num = err;
 584         }
 585
 586         if ((req->r_osd && req->r_osd->o_osd == o &&
 587              req->r_sent >= req->r_osd->o_incarnation &&
 588              req->r_num_pg_osds == num &&
 589              memcmp(req->r_pg_osds, acting, sizeof(acting[0])*num) == 0) ||
 590             (req->r_osd == NULL && o == -1))
 591                 return 0;  /* no change */
 592
 593         dout("map_osds tid %llu pgid %d.%x osd%d (was osd%d)\n",
 594              req->r_tid, le32_to_cpu(pgid.pool), le16_to_cpu(pgid.ps), o,
 595              req->r_osd ? req->r_osd->o_osd : -1);
 596
 597         /* record full pg acting set */
 598         memcpy(req->r_pg_osds, acting, sizeof(acting[0]) * num);
 599         req->r_num_pg_osds = num;
 600
 601         if (req->r_osd) {
 602                 __cancel_request(req);
 603                 list_del_init(&req->r_osd_item);
 604                 req->r_osd = NULL;
 605         }
 606
 607         req->r_osd = __lookup_osd(osdc, o);
 608         if (!req->r_osd && o >= 0) {
 609                 err = -ENOMEM;
 610                 req->r_osd = create_osd(osdc);
 611                 if (!req->r_osd)
 612                         goto out;
 613
 614                 dout("map_osds osd %p is osd%d\n", req->r_osd, o);
 615                 req->r_osd->o_osd = o;
 616                 req->r_osd->o_con.peer_name.num = cpu_to_le64(o);
 617                 __insert_osd(osdc, req->r_osd);
 618
 619                 ceph_con_open(&req->r_osd->o_con, &osdc->osdmap->osd_addr[o]);
 620         }
 621
 622         if (req->r_osd) {
 623                 __remove_osd_from_lru(req->r_osd);
 624                 list_add(&req->r_osd_item, &req->r_osd->o_requests);
 625         }
 626         err = 1;   /* osd or pg changed */
 627
 628 out:
 629         return err;
 630 }
 631
 632 /*
 633  * caller should hold map_sem (for read) and request_mutex
 634  */
 635 static int __send_request(struct ceph_osd_client *osdc,
 636                           struct ceph_osd_request *req)
 637 {
 638         struct ceph_osd_request_head *reqhead;
 639         int err;
 640
 641         err = __map_osds(osdc, req);
 642         if (err < 0)
 643                 return err;
 644         if (req->r_osd == NULL) {
 645                 dout("send_request %p no up osds in pg\n", req);
 646                 ceph_monc_request_next_osdmap(&osdc->client->monc);
 647                 return 0;
 648         }
 649
 650         dout("send_request %p tid %llu to osd%d flags %d\n",
 651              req, req->r_tid, req->r_osd->o_osd, req->r_flags);
 652
 653         reqhead = req->r_request->front.iov_base;
 654         reqhead->osdmap_epoch = cpu_to_le32(osdc->osdmap->epoch);
 655         reqhead->flags |= cpu_to_le32(req->r_flags);  /* e.g., RETRY */
 656         reqhead->reassert_version = req->r_reassert_version;
 657
 658         req->r_stamp = jiffies;
 659         list_move_tail(&osdc->req_lru, &req->r_req_lru_item);
 660
 661         ceph_msg_get(req->r_request); /* send consumes a ref */
 662         ceph_con_send(&req->r_osd->o_con, req->r_request);
 663         req->r_sent = req->r_osd->o_incarnation;
 664         return 0;
 665 }
 666
 667 /*
 668  * Timeout callback, called every N seconds when 1 or more osd
 669  * requests has been active for more than N seconds.  When this
 670  * happens, we ping all OSDs with requests who have timed out to
 671  * ensure any communications channel reset is detected.  Reset the
 672  * request timeouts another N seconds in the future as we go.
 673  * Reschedule the timeout event another N seconds in future (unless
 674  * there are no open requests).
 675  */
 676 static void handle_timeout(struct work_struct *work)
 677 {
 678         struct ceph_osd_client *osdc =
 679                 container_of(work, struct ceph_osd_client, timeout_work.work);
 680         struct ceph_osd_request *req, *last_req = NULL;
 681         struct ceph_osd *osd;
 682         unsigned long timeout = osdc->client->mount_args->osd_timeout * HZ;
 683         unsigned long keepalive =
 684                 osdc->client->mount_args->osd_keepalive_timeout * HZ;
 685         unsigned long last_stamp = 0;
 686         struct rb_node *p;
 687         struct list_head slow_osds;
 688
 689         dout("timeout\n");
 690         down_read(&osdc->map_sem);
 691
 692         ceph_monc_request_next_osdmap(&osdc->client->monc);
 693
 694         mutex_lock(&osdc->request_mutex);
 695         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
 696                 req = rb_entry(p, struct ceph_osd_request, r_node);
 697
 698                 if (req->r_resend) {
 699                         int err;
 700
 701                         dout("osdc resending prev failed %lld\n", req->r_tid);
 702                         err = __send_request(osdc, req);
 703                         if (err)
 704                                 dout("osdc failed again on %lld\n", req->r_tid);
 705                         else
 706                                 req->r_resend = false;
 707                         continue;
 708                 }
 709         }
 710
 711         /*
 712          * reset osds that appear to be _really_ unresponsive.  this
 713          * is a failsafe measure.. we really shouldn't be getting to
 714          * this point if the system is working properly.  the monitors
 715          * should mark the osd as failed and we should find out about
 716          * it from an updated osd map.
 717          */
 718         while (timeout && !list_empty(&osdc->req_lru)) {
 719                 req = list_entry(osdc->req_lru.next, struct ceph_osd_request,
 720                                  r_req_lru_item);
 721
 722                 if (time_before(jiffies, req->r_stamp + timeout))
 723                         break;
 724
 725                 BUG_ON(req == last_req && req->r_stamp == last_stamp);
 726                 last_req = req;
 727                 last_stamp = req->r_stamp;
 728
 729                 osd = req->r_osd;
 730                 BUG_ON(!osd);
 731                 pr_warning(" tid %llu timed out on osd%d, will reset osd\n",
 732                            req->r_tid, osd->o_osd);
 733                 __kick_requests(osdc, osd);
 734         }
 735
 736         /*
 737          * ping osds that are a bit slow.  this ensures that if there
 738          * is a break in the TCP connection we will notice, and reopen
 739          * a connection with that osd (from the fault callback).
 740          */
 741         INIT_LIST_HEAD(&slow_osds);
 742         list_for_each_entry(req, &osdc->req_lru, r_req_lru_item) {
 743                 if (time_before(jiffies, req->r_stamp + keepalive))
 744                         break;
 745
 746                 osd = req->r_osd;
 747                 BUG_ON(!osd);
 748                 dout(" tid %llu is slow, will send keepalive on osd%d\n",
 749                      req->r_tid, osd->o_osd);
 750                 list_move_tail(&osd->o_keepalive_item, &slow_osds);
 751         }
 752         while (!list_empty(&slow_osds)) {
 753                 osd = list_entry(slow_osds.next, struct ceph_osd,
 754                                  o_keepalive_item);
 755                 list_del_init(&osd->o_keepalive_item);
 756                 ceph_con_keepalive(&osd->o_con);
 757         }
 758
 759         __schedule_osd_timeout(osdc);
 760         mutex_unlock(&osdc->request_mutex);
 761
 762         up_read(&osdc->map_sem);
 763 }
 764
 765 static void handle_osds_timeout(struct work_struct *work)
 766 {
 767         struct ceph_osd_client *osdc =
 768                 container_of(work, struct ceph_osd_client,
 769                              osds_timeout_work.work);
 770         unsigned long delay =
 771                 osdc->client->mount_args->osd_idle_ttl * HZ >> 2;
 772
 773         dout("osds timeout\n");
 774         down_read(&osdc->map_sem);
 775         remove_old_osds(osdc, 0);
 776         up_read(&osdc->map_sem);
 777
 778         schedule_delayed_work(&osdc->osds_timeout_work,
 779                               round_jiffies_relative(delay));
 780 }
 781
 782 /*
 783  * handle osd op reply.  either call the callback if it is specified,
 784  * or do the completion to wake up the waiting thread.
 785  */
 786 static void handle_reply(struct ceph_osd_client *osdc, struct ceph_msg *msg,
 787                          struct ceph_connection *con)
 788 {
 789         struct ceph_osd_reply_head *rhead = msg->front.iov_base;
 790         struct ceph_osd_request *req;
 791         u64 tid;
 792         int numops, object_len, flags;
 793         s32 result;
 794
 795         tid = le64_to_cpu(msg->hdr.tid);
 796         if (msg->front.iov_len < sizeof(*rhead))
 797                 goto bad;
 798         numops = le32_to_cpu(rhead->num_ops);
 799         object_len = le32_to_cpu(rhead->object_len);
 800         result = le32_to_cpu(rhead->result);
 801         if (msg->front.iov_len != sizeof(*rhead) + object_len +
 802             numops * sizeof(struct ceph_osd_op))
 803                 goto bad;
 804         dout("handle_reply %p tid %llu result %d\n", msg, tid, (int)result);
 805
 806         /* lookup */
 807         mutex_lock(&osdc->request_mutex);
 808         req = __lookup_request(osdc, tid);
 809         if (req == NULL) {
 810                 dout("handle_reply tid %llu dne\n", tid);
 811                 mutex_unlock(&osdc->request_mutex);
 812                 return;
 813         }
 814         ceph_osdc_get_request(req);
 815         flags = le32_to_cpu(rhead->flags);
 816
 817         /*
 818          * if this connection filled our message, drop our reference now, to
 819          * avoid a (safe but slower) revoke later.
 820          */
 821         if (req->r_con_filling_msg == con && req->r_reply == msg) {
 822                 dout(" dropping con_filling_msg ref %p\n", con);
 823                 req->r_con_filling_msg = NULL;
 824                 ceph_con_put(con);
 825         }
 826
 827         if (!req->r_got_reply) {
 828                 unsigned bytes;
 829
 830                 req->r_result = le32_to_cpu(rhead->result);
 831                 bytes = le32_to_cpu(msg->hdr.data_len);
 832                 dout("handle_reply result %d bytes %d\n", req->r_result,
 833                      bytes);
 834                 if (req->r_result == 0)
 835                         req->r_result = bytes;
 836
 837                 /* in case this is a write and we need to replay, */
 838                 req->r_reassert_version = rhead->reassert_version;
 839
 840                 req->r_got_reply = 1;
 841         } else if ((flags & CEPH_OSD_FLAG_ONDISK) == 0) {
 842                 dout("handle_reply tid %llu dup ack\n", tid);
 843                 mutex_unlock(&osdc->request_mutex);
 844                 goto done;
 845         }
 846
 847         dout("handle_reply tid %llu flags %d\n", tid, flags);
 848
 849         /* either this is a read, or we got the safe response */
 850         if (result < 0 ||
 851             (flags & CEPH_OSD_FLAG_ONDISK) ||
 852             ((flags & CEPH_OSD_FLAG_WRITE) == 0))
 853                 __unregister_request(osdc, req);
 854
 855         mutex_unlock(&osdc->request_mutex);
 856
 857         if (req->r_callback)
 858                 req->r_callback(req, msg);
 859         else
 860                 complete(&req->r_completion);
 861
 862         if (flags & CEPH_OSD_FLAG_ONDISK) {
 863                 if (req->r_safe_callback)
 864                         req->r_safe_callback(req, msg);
 865                 complete(&req->r_safe_completion);  /* fsync waiter */
 866         }
 867
 868 done:
 869         ceph_osdc_put_request(req);
 870         return;
 871
 872 bad:
 873         pr_err("corrupt osd_op_reply got %d %d expected %d\n",
 874                (int)msg->front.iov_len, le32_to_cpu(msg->hdr.front_len),
 875                (int)sizeof(*rhead));
 876         ceph_msg_dump(msg);
 877 }
 878
 879
 880 static int __kick_requests(struct ceph_osd_client *osdc,
 881                           struct ceph_osd *kickosd)
 882 {
 883         struct ceph_osd_request *req;
 884         struct rb_node *p, *n;
 885         int needmap = 0;
 886         int err;
 887
 888         dout("kick_requests osd%d\n", kickosd ? kickosd->o_osd : -1);
 889         if (kickosd) {
 890                 err = __reset_osd(osdc, kickosd);
 891                 if (err == -EAGAIN)
 892                         return 1;
 893         } else {
 894                 for (p = rb_first(&osdc->osds); p; p = n) {
 895                         struct ceph_osd *osd =
 896                                 rb_entry(p, struct ceph_osd, o_node);
 897
 898                         n = rb_next(p);
 899                         if (!ceph_osd_is_up(osdc->osdmap, osd->o_osd) ||
 900                             memcmp(&osd->o_con.peer_addr,
 901                                    ceph_osd_addr(osdc->osdmap,
 902                                                  osd->o_osd),
 903                                    sizeof(struct ceph_entity_addr)) != 0)
 904                                 __reset_osd(osdc, osd);
 905                 }
 906         }
 907
 908         for (p = rb_first(&osdc->requests); p; p = rb_next(p)) {
 909                 req = rb_entry(p, struct ceph_osd_request, r_node);
 910
 911                 if (req->r_resend) {
 912                         dout(" r_resend set on tid %llu\n", req->r_tid);
 913                         __cancel_request(req);
 914                         goto kick;
 915                 }
 916                 if (req->r_osd && kickosd == req->r_osd) {
 917                         __cancel_request(req);
 918                         goto kick;
 919                 }
 920
 921                 err = __map_osds(osdc, req);
 922                 if (err == 0)
 923                         continue;  /* no change */
 924                 if (err < 0) {
 925                         /*
 926                          * FIXME: really, we should set the request
 927                          * error and fail if this isn't a 'nofail'
 928                          * request, but that's a fair bit more
 929                          * complicated to do.  So retry!
 930                          */
 931                         dout(" setting r_resend on %llu\n", req->r_tid);
 932                         req->r_resend = true;
 933                         continue;
 934                 }
 935                 if (req->r_osd == NULL) {
 936                         dout("tid %llu maps to no valid osd\n", req->r_tid);
 937                         needmap++;  /* request a newer map */
 938                         continue;
 939                 }
 940
 941 kick:
 942                 dout("kicking %p tid %llu osd%d\n", req, req->r_tid,
 943                      req->r_osd ? req->r_osd->o_osd : -1);
 944                 req->r_flags |= CEPH_OSD_FLAG_RETRY;
 945                 err = __send_request(osdc, req);
 946                 if (err) {
 947                         dout(" setting r_resend on %llu\n", req->r_tid);
 948                         req->r_resend = true;
 949                 }
 950         }
 951
 952         return needmap;
 953 }
 954
 955 /*
 956  * Resubmit osd requests whose osd or osd address has changed.  Request
 957  * a new osd map if osds are down, or we are otherwise unable to determine
 958  * how to direct a request.
 959  *
 960  * Close connections to down osds.
 961  *
 962  * If @who is specified, resubmit requests for that specific osd.
 963  *
 964  * Caller should hold map_sem for read and request_mutex.
 965  */
 966 static void kick_requests(struct ceph_osd_client *osdc,
 967                           struct ceph_osd *kickosd)
 968 {
 969         int needmap;
 970
 971         mutex_lock(&osdc->request_mutex);
 972         needmap = __kick_requests(osdc, kickosd);
 973         mutex_unlock(&osdc->request_mutex);
 974
 975         if (needmap) {
 976                 dout("%d requests for down osds, need new map\n", needmap);
 977                 ceph_monc_request_next_osdmap(&osdc->client->monc);
 978         }
 979
 980 }
 981 /*
 982  * Process updated osd map.
 983  *
 984  * The message contains any number of incremental and full maps, normally
 985  * indicating some sort of topology change in the cluster.  Kick requests
 986  * off to different OSDs as needed.
 987  */
 988 void ceph_osdc_handle_map(struct ceph_osd_client *osdc, struct ceph_msg *msg)
 989 {
 990         void *p, *end, *next;
 991         u32 nr_maps, maplen;
 992         u32 epoch;
 993         struct ceph_osdmap *newmap = NULL, *oldmap;
 994         int err;
 995         struct ceph_fsid fsid;
 996
 997         dout("handle_map have %u\n", osdc->osdmap ? osdc->osdmap->epoch : 0);
 998         p = msg->front.iov_base;
 999         end = p + msg->front.iov_len;
1000
1001         /* verify fsid */
1002         ceph_decode_need(&p, end, sizeof(fsid), bad);
1003         ceph_decode_copy(&p, &fsid, sizeof(fsid));
1004         if (ceph_check_fsid(osdc->client, &fsid) < 0)
1005                 return;
1006
1007         down_write(&osdc->map_sem);
1008
1009         /* incremental maps */
1010         ceph_decode_32_safe(&p, end, nr_maps, bad);
1011         dout(" %d inc maps\n", nr_maps);
1012         while (nr_maps > 0) {
1013                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1014                 epoch = ceph_decode_32(&p);
1015                 maplen = ceph_decode_32(&p);
1016                 ceph_decode_need(&p, end, maplen, bad);
1017                 next = p + maplen;
1018                 if (osdc->osdmap && osdc->osdmap->epoch+1 == epoch) {
1019                         dout("applying incremental map %u len %d\n",
1020                              epoch, maplen);
1021                         newmap = osdmap_apply_incremental(&p, next,
1022                                                           osdc->osdmap,
1023                                                           osdc->client->msgr);
1024                         if (IS_ERR(newmap)) {
1025                                 err = PTR_ERR(newmap);
1026                                 goto bad;
1027                         }
1028                         BUG_ON(!newmap);
1029                         if (newmap != osdc->osdmap) {
1030                                 ceph_osdmap_destroy(osdc->osdmap);
1031                                 osdc->osdmap = newmap;
1032                         }
1033                 } else {
1034                         dout("ignoring incremental map %u len %d\n",
1035                              epoch, maplen);
1036                 }
1037                 p = next;
1038                 nr_maps--;
1039         }
1040         if (newmap)
1041                 goto done;
1042
1043         /* full maps */
1044         ceph_decode_32_safe(&p, end, nr_maps, bad);
1045         dout(" %d full maps\n", nr_maps);
1046         while (nr_maps) {
1047                 ceph_decode_need(&p, end, 2*sizeof(u32), bad);
1048                 epoch = ceph_decode_32(&p);
1049                 maplen = ceph_decode_32(&p);
1050                 ceph_decode_need(&p, end, maplen, bad);
1051                 if (nr_maps > 1) {
1052                         dout("skipping non-latest full map %u len %d\n",
1053                              epoch, maplen);
1054                 } else if (osdc->osdmap && osdc->osdmap->epoch >= epoch) {
1055                         dout("skipping full map %u len %d, "
1056                              "older than our %u\n", epoch, maplen,
1057                              osdc->osdmap->epoch);
1058                 } else {
1059                         dout("taking full map %u len %d\n", epoch, maplen);
1060                         newmap = osdmap_decode(&p, p+maplen);
1061                         if (IS_ERR(newmap)) {
1062                                 err = PTR_ERR(newmap);
1063                                 goto bad;
1064                         }
1065                         BUG_ON(!newmap);
1066                         oldmap = osdc->osdmap;
1067                         osdc->osdmap = newmap;
1068                         if (oldmap)
1069                                 ceph_osdmap_destroy(oldmap);
1070                 }
1071                 p += maplen;
1072                 nr_maps--;
1073         }
1074
1075 done:
1076         downgrade_write(&osdc->map_sem);
1077         ceph_monc_got_osdmap(&osdc->client->monc, osdc->osdmap->epoch);
1078         if (newmap)
1079                 kick_requests(osdc, NULL);
1080         up_read(&osdc->map_sem);
1081         wake_up(&osdc->client->auth_wq);
1082         return;
1083
1084 bad:
1085         pr_err("osdc handle_map corrupt msg\n");
1086         ceph_msg_dump(msg);
1087         up_write(&osdc->map_sem);
1088         return;
1089 }
1090
1091 /*
1092  * Register request, send initial attempt.
1093  */
1094 int ceph_osdc_start_request(struct ceph_osd_client *osdc,
1095                             struct ceph_osd_request *req,
1096                             bool nofail)
1097 {
1098         int rc = 0;
1099
1100         req->r_request->pages = req->r_pages;
1101         req->r_request->nr_pages = req->r_num_pages;
1102
1103         register_request(osdc, req);
1104
1105         down_read(&osdc->map_sem);
1106         mutex_lock(&osdc->request_mutex);
1107         /*
1108          * a racing kick_requests() may have sent the message for us
1109          * while we dropped request_mutex above, so only send now if
1110          * the request still han't been touched yet.
1111          */
1112         if (req->r_sent == 0) {
1113                 rc = __send_request(osdc, req);
1114                 if (rc) {
1115                         if (nofail) {
1116                                 dout("osdc_start_request failed send, "
1117                                      " marking %lld\n", req->r_tid);
1118                                 req->r_resend = true;
1119                                 rc = 0;
1120                         } else {
1121                                 __unregister_request(osdc, req);
1122                         }
1123                 }
1124         }
1125         mutex_unlock(&osdc->request_mutex);
1126         up_read(&osdc->map_sem);
1127         return rc;
1128 }
1129
1130 /*
1131  * wait for a request to complete
1132  */
1133 int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
1134                            struct ceph_osd_request *req)
1135 {
1136         int rc;
1137
1138         rc = wait_for_completion_interruptible(&req->r_completion);
1139         if (rc < 0) {
1140                 mutex_lock(&osdc->request_mutex);
1141                 __cancel_request(req);
1142                 __unregister_request(osdc, req);
1143                 mutex_unlock(&osdc->request_mutex);
1144                 dout("wait_request tid %llu canceled/timed out\n", req->r_tid);
1145                 return rc;
1146         }
1147
1148         dout("wait_request tid %llu result %d\n", req->r_tid, req->r_result);
1149         return req->r_result;
1150 }
1151
1152 /*
1153  * sync - wait for all in-flight requests to flush.  avoid starvation.
1154  */
1155 void ceph_osdc_sync(struct ceph_osd_client *osdc)
1156 {
1157         struct ceph_osd_request *req;
1158         u64 last_tid, next_tid = 0;
1159
1160         mutex_lock(&osdc->request_mutex);
1161         last_tid = osdc->last_tid;
1162         while (1) {
1163                 req = __lookup_request_ge(osdc, next_tid);
1164                 if (!req)
1165                         break;
1166                 if (req->r_tid > last_tid)
1167                         break;
1168
1169                 next_tid = req->r_tid + 1;
1170                 if ((req->r_flags & CEPH_OSD_FLAG_WRITE) == 0)
1171                         continue;
1172
1173                 ceph_osdc_get_request(req);
1174                 mutex_unlock(&osdc->request_mutex);
1175                 dout("sync waiting on tid %llu (last is %llu)\n",
1176                      req->r_tid, last_tid);
1177                 wait_for_completion(&req->r_safe_completion);
1178                 mutex_lock(&osdc->request_mutex);
1179                 ceph_osdc_put_request(req);
1180         }
1181         mutex_unlock(&osdc->request_mutex);
1182         dout("sync done (thru tid %llu)\n", last_tid);
1183 }
1184
1185 /*
1186  * init, shutdown
1187  */
1188 int ceph_osdc_init(struct ceph_osd_client *osdc, struct ceph_client *client)
1189 {
1190         int err;
1191
1192         dout("init\n");
1193         osdc->client = client;
1194         osdc->osdmap = NULL;
1195         init_rwsem(&osdc->map_sem);
1196         init_completion(&osdc->map_waiters);
1197         osdc->last_requested_map = 0;
1198         mutex_init(&osdc->request_mutex);
1199         osdc->last_tid = 0;
1200         osdc->osds = RB_ROOT;
1201         INIT_LIST_HEAD(&osdc->osd_lru);
1202         osdc->requests = RB_ROOT;
1203         INIT_LIST_HEAD(&osdc->req_lru);
1204         osdc->num_requests = 0;
1205         INIT_DELAYED_WORK(&osdc->timeout_work, handle_timeout);
1206         INIT_DELAYED_WORK(&osdc->osds_timeout_work, handle_osds_timeout);
1207
1208         schedule_delayed_work(&osdc->osds_timeout_work,
1209            round_jiffies_relative(osdc->client->mount_args->osd_idle_ttl * HZ));
1210
1211         err = -ENOMEM;
1212         osdc->req_mempool = mempool_create_kmalloc_pool(10,
1213                                         sizeof(struct ceph_osd_request));
1214         if (!osdc->req_mempool)
1215                 goto out;
1216
1217         err = ceph_msgpool_init(&osdc->msgpool_op, OSD_OP_FRONT_LEN, 10, true,
1218                                 "osd_op");
1219         if (err < 0)
1220                 goto out_mempool;
1221         err = ceph_msgpool_init(&osdc->msgpool_op_reply,
1222                                 OSD_OPREPLY_FRONT_LEN, 10, true,
1223                                 "osd_op_reply");
1224         if (err < 0)
1225                 goto out_msgpool;
1226         return 0;
1227
1228 out_msgpool:
1229         ceph_msgpool_destroy(&osdc->msgpool_op);
1230 out_mempool:
1231         mempool_destroy(osdc->req_mempool);
1232 out:
1233         return err;
1234 }
1235
1236 void ceph_osdc_stop(struct ceph_osd_client *osdc)
1237 {
1238         cancel_delayed_work_sync(&osdc->timeout_work);
1239         cancel_delayed_work_sync(&osdc->osds_timeout_work);
1240         if (osdc->osdmap) {
1241                 ceph_osdmap_destroy(osdc->osdmap);
1242                 osdc->osdmap = NULL;
1243         }
1244         remove_old_osds(osdc, 1);
1245         mempool_destroy(osdc->req_mempool);
1246         ceph_msgpool_destroy(&osdc->msgpool_op);
1247         ceph_msgpool_destroy(&osdc->msgpool_op_reply);
1248 }
1249
1250 /*
1251  * Read some contiguous pages.  If we cross a stripe boundary, shorten
1252  * *plen.  Return number of bytes read, or error.
1253  */
1254 int ceph_osdc_readpages(struct ceph_osd_client *osdc,
1255                         struct ceph_vino vino, struct ceph_file_layout *layout,
1256                         u64 off, u64 *plen,
1257                         u32 truncate_seq, u64 truncate_size,
1258                         struct page **pages, int num_pages)
1259 {
1260         struct ceph_osd_request *req;
1261         int rc = 0;
1262
1263         dout("readpages on ino %llx.%llx on %llu~%llu\n", vino.ino,
1264              vino.snap, off, *plen);
1265         req = ceph_osdc_new_request(osdc, layout, vino, off, plen,
1266                                     CEPH_OSD_OP_READ, CEPH_OSD_FLAG_READ,
1267                                     NULL, 0, truncate_seq, truncate_size, NULL,
1268                                     false, 1);
1269         if (!req)
1270                 return -ENOMEM;
1271
1272         /* it may be a short read due to an object boundary */
1273         req->r_pages = pages;
1274         num_pages = calc_pages_for(off, *plen);
1275         req->r_num_pages = num_pages;
1276
1277         dout("readpages  final extent is %llu~%llu (%d pages)\n",
1278              off, *plen, req->r_num_pages);
1279
1280         rc = ceph_osdc_start_request(osdc, req, false);
1281         if (!rc)
1282                 rc = ceph_osdc_wait_request(osdc, req);
1283
1284         ceph_osdc_put_request(req);
1285         dout("readpages result %d\n", rc);
1286         return rc;
1287 }
1288
1289 /*
1290  * do a synchronous write on N pages
1291  */
1292 int ceph_osdc_writepages(struct ceph_osd_client *osdc, struct ceph_vino vino,
1293                          struct ceph_file_layout *layout,
1294                          struct ceph_snap_context *snapc,
1295                          u64 off, u64 len,
1296                          u32 truncate_seq, u64 truncate_size,
1297                          struct timespec *mtime,
1298                          struct page **pages, int num_pages,
1299                          int flags, int do_sync, bool nofail)
1300 {
1301         struct ceph_osd_request *req;
1302         int rc = 0;
1303
1304         BUG_ON(vino.snap != CEPH_NOSNAP);
1305         req = ceph_osdc_new_request(osdc, layout, vino, off, &len,
1306                                     CEPH_OSD_OP_WRITE,
1307                                     flags | CEPH_OSD_FLAG_ONDISK |
1308                                             CEPH_OSD_FLAG_WRITE,
1309                                     snapc, do_sync,
1310                                     truncate_seq, truncate_size, mtime,
1311                                     nofail, 1);
1312         if (!req)
1313                 return -ENOMEM;
1314
1315         /* it may be a short write due to an object boundary */
1316         req->r_pages = pages;
1317         req->r_num_pages = calc_pages_for(off, len);
1318         dout("writepages %llu~%llu (%d pages)\n", off, len,
1319              req->r_num_pages);
1320
1321         rc = ceph_osdc_start_request(osdc, req, nofail);
1322         if (!rc)
1323                 rc = ceph_osdc_wait_request(osdc, req);
1324
1325         ceph_osdc_put_request(req);
1326         if (rc == 0)
1327                 rc = len;
1328         dout("writepages result %d\n", rc);
1329         return rc;
1330 }
1331
1332 /*
1333  * handle incoming message
1334  */
1335 static void dispatch(struct ceph_connection *con, struct ceph_msg *msg)
1336 {
1337         struct ceph_osd *osd = con->private;
1338         struct ceph_osd_client *osdc;
1339         int type = le16_to_cpu(msg->hdr.type);
1340
1341         if (!osd)
1342                 return;
1343         osdc = osd->o_osdc;
1344
1345         switch (type) {
1346         case CEPH_MSG_OSD_MAP:
1347                 ceph_osdc_handle_map(osdc, msg);
1348                 break;
1349         case CEPH_MSG_OSD_OPREPLY:
1350                 handle_reply(osdc, msg, con);
1351                 break;
1352
1353         default:
1354                 pr_err("received unknown message type %d %s\n", type,
1355                        ceph_msg_type_name(type));
1356         }
1357         ceph_msg_put(msg);
1358 }
1359
1360 /*
1361  * lookup and return message for incoming reply.  set up reply message
1362  * pages.
1363  */
1364 static struct ceph_msg *get_reply(struct ceph_connection *con,
1365                                   struct ceph_msg_header *hdr,
1366                                   int *skip)
1367 {
1368         struct ceph_osd *osd = con->private;
1369         struct ceph_osd_client *osdc = osd->o_osdc;
1370         struct ceph_msg *m;
1371         struct ceph_osd_request *req;
1372         int front = le32_to_cpu(hdr->front_len);
1373         int data_len = le32_to_cpu(hdr->data_len);
1374         u64 tid;
1375
1376         tid = le64_to_cpu(hdr->tid);
1377         mutex_lock(&osdc->request_mutex);
1378         req = __lookup_request(osdc, tid);
1379         if (!req) {
1380                 *skip = 1;
1381                 m = NULL;
1382                 pr_info("get_reply unknown tid %llu from osd%d\n", tid,
1383                         osd->o_osd);
1384                 goto out;
1385         }
1386
1387         if (req->r_con_filling_msg) {
1388                 dout("get_reply revoking msg %p from old con %p\n",
1389                      req->r_reply, req->r_con_filling_msg);
1390                 ceph_con_revoke_message(req->r_con_filling_msg, req->r_reply);
1391                 ceph_con_put(req->r_con_filling_msg);
1392                 req->r_con_filling_msg = NULL;
1393         }
1394
1395         if (front > req->r_reply->front.iov_len) {
1396                 pr_warning("get_reply front %d > preallocated %d\n",
1397                            front, (int)req->r_reply->front.iov_len);
1398                 m = ceph_msg_new(CEPH_MSG_OSD_OPREPLY, front, GFP_NOFS);
1399                 if (!m)
1400                         goto out;
1401                 ceph_msg_put(req->r_reply);
1402                 req->r_reply = m;
1403         }
1404         m = ceph_msg_get(req->r_reply);
1405
1406         if (data_len > 0) {
1407                 unsigned data_off = le16_to_cpu(hdr->data_off);
1408                 int want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1409
1410                 if (unlikely(req->r_num_pages < want)) {
1411                         pr_warning("tid %lld reply %d > expected %d pages\n",
1412                                    tid, want, m->nr_pages);
1413                         *skip = 1;
1414                         ceph_msg_put(m);
1415                         m = NULL;
1416                         goto out;
1417                 }
1418                 m->pages = req->r_pages;
1419                 m->nr_pages = req->r_num_pages;
1420         }
1421         *skip = 0;
1422         req->r_con_filling_msg = ceph_con_get(con);
1423         dout("get_reply tid %lld %p\n", tid, m);
1424
1425 out:
1426         mutex_unlock(&osdc->request_mutex);
1427         return m;
1428
1429 }
1430
1431 static struct ceph_msg *alloc_msg(struct ceph_connection *con,
1432                                   struct ceph_msg_header *hdr,
1433                                   int *skip)
1434 {
1435         struct ceph_osd *osd = con->private;
1436         int type = le16_to_cpu(hdr->type);
1437         int front = le32_to_cpu(hdr->front_len);
1438
1439         switch (type) {
1440         case CEPH_MSG_OSD_MAP:
1441                 return ceph_msg_new(type, front, GFP_NOFS);
1442         case CEPH_MSG_OSD_OPREPLY:
1443                 return get_reply(con, hdr, skip);
1444         default:
1445                 pr_info("alloc_msg unexpected msg type %d from osd%d\n", type,
1446                         osd->o_osd);
1447                 *skip = 1;
1448                 return NULL;
1449         }
1450 }
1451
1452 /*
1453  * Wrappers to refcount containing ceph_osd struct
1454  */
1455 static struct ceph_connection *get_osd_con(struct ceph_connection *con)
1456 {
1457         struct ceph_osd *osd = con->private;
1458         if (get_osd(osd))
1459                 return con;
1460         return NULL;
1461 }
1462
1463 static void put_osd_con(struct ceph_connection *con)
1464 {
1465         struct ceph_osd *osd = con->private;
1466         put_osd(osd);
1467 }
1468
1469 /*
1470  * authentication
1471  */
1472 static int get_authorizer(struct ceph_connection *con,
1473                           void **buf, int *len, int *proto,
1474                           void **reply_buf, int *reply_len, int force_new)
1475 {
1476         struct ceph_osd *o = con->private;
1477         struct ceph_osd_client *osdc = o->o_osdc;
1478         struct ceph_auth_client *ac = osdc->client->monc.auth;
1479         int ret = 0;
1480
1481         if (force_new && o->o_authorizer) {
1482                 ac->ops->destroy_authorizer(ac, o->o_authorizer);
1483                 o->o_authorizer = NULL;
1484         }
1485         if (o->o_authorizer == NULL) {
1486                 ret = ac->ops->create_authorizer(
1487                         ac, CEPH_ENTITY_TYPE_OSD,
1488                         &o->o_authorizer,
1489                         &o->o_authorizer_buf,
1490                         &o->o_authorizer_buf_len,
1491                         &o->o_authorizer_reply_buf,
1492                         &o->o_authorizer_reply_buf_len);
1493                 if (ret)
1494                 return ret;
1495         }
1496
1497         *proto = ac->protocol;
1498         *buf = o->o_authorizer_buf;
1499         *len = o->o_authorizer_buf_len;
1500         *reply_buf = o->o_authorizer_reply_buf;
1501         *reply_len = o->o_authorizer_reply_buf_len;
1502         return 0;
1503 }
1504
1505
1506 static int verify_authorizer_reply(struct ceph_connection *con, int len)
1507 {
1508         struct ceph_osd *o = con->private;
1509         struct ceph_osd_client *osdc = o->o_osdc;
1510         struct ceph_auth_client *ac = osdc->client->monc.auth;
1511
1512         return ac->ops->verify_authorizer_reply(ac, o->o_authorizer, len);
1513 }
1514
1515 static int invalidate_authorizer(struct ceph_connection *con)
1516 {
1517         struct ceph_osd *o = con->private;
1518         struct ceph_osd_client *osdc = o->o_osdc;
1519         struct ceph_auth_client *ac = osdc->client->monc.auth;
1520
1521         if (ac->ops->invalidate_authorizer)
1522                 ac->ops->invalidate_authorizer(ac, CEPH_ENTITY_TYPE_OSD);
1523
1524         return ceph_monc_validate_auth(&osdc->client->monc);
1525 }
1526
1527 static const struct ceph_connection_operations osd_con_ops = {
1528         .get = get_osd_con,
1529         .put = put_osd_con,
1530         .dispatch = dispatch,
1531         .get_authorizer = get_authorizer,
1532         .verify_authorizer_reply = verify_authorizer_reply,
1533         .invalidate_authorizer = invalidate_authorizer,
1534         .alloc_msg = alloc_msg,
1535         .fault = osd_reset,
1536 };