drivers/block/drbd/drbd_req.c

   1 /*
   2    drbd_req.c
   3
   4    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   5
   6    Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
   7    Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   8    Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
   9
  10    drbd is free software; you can redistribute it and/or modify
  11    it under the terms of the GNU General Public License as published by
  12    the Free Software Foundation; either version 2, or (at your option)
  13    any later version.
  14
  15    drbd is distributed in the hope that it will be useful,
  16    but WITHOUT ANY WARRANTY; without even the implied warranty of
  17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18    GNU General Public License for more details.
  19
  20    You should have received a copy of the GNU General Public License
  21    along with drbd; see the file COPYING.  If not, write to
  22    the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
  23
  24  */
  25
  26 #include <linux/module.h>
  27
  28 #include <linux/slab.h>
  29 #include <linux/drbd.h>
  30 #include "drbd_int.h"
  31 #include "drbd_req.h"
  32
  33
  34 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size);
  35
  36 /* Update disk stats at start of I/O request */
  37 static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
  38 {
  39         const int rw = bio_data_dir(bio);
  40         int cpu;
  41         cpu = part_stat_lock();
  42         part_round_stats(cpu, &mdev->vdisk->part0);
  43         part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
  44         part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
  45         (void) cpu; /* The macro invocations above want the cpu argument, I do not like
  46                        the compiler warning about cpu only assigned but never used... */
  47         part_inc_in_flight(&mdev->vdisk->part0, rw);
  48         part_stat_unlock();
  49 }
  50
  51 /* Update disk stats when completing request upwards */
  52 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
  53 {
  54         int rw = bio_data_dir(req->master_bio);
  55         unsigned long duration = jiffies - req->start_time;
  56         int cpu;
  57         cpu = part_stat_lock();
  58         part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
  59         part_round_stats(cpu, &mdev->vdisk->part0);
  60         part_dec_in_flight(&mdev->vdisk->part0, rw);
  61         part_stat_unlock();
  62 }
  63
  64 static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
  65                                                struct bio *bio_src)
  66 {
  67         struct drbd_request *req;
  68
  69         req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
  70         if (!req)
  71                 return NULL;
  72
  73         drbd_req_make_private_bio(req, bio_src);
  74         req->rq_state    = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
  75         req->w.mdev      = mdev;
  76         req->master_bio  = bio_src;
  77         req->epoch       = 0;
  78
  79         drbd_clear_interval(&req->i);
  80         req->i.sector     = bio_src->bi_sector;
  81         req->i.size      = bio_src->bi_size;
  82         req->i.local = true;
  83         req->i.waiting = false;
  84
  85         INIT_LIST_HEAD(&req->tl_requests);
  86         INIT_LIST_HEAD(&req->w.list);
  87
  88         return req;
  89 }
  90
  91 static void drbd_req_free(struct drbd_request *req)
  92 {
  93         mempool_free(req, drbd_request_mempool);
  94 }
  95
  96 /* rw is bio_data_dir(), only READ or WRITE */
  97 static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
  98 {
  99         const unsigned long s = req->rq_state;
 100
 101         /* remove it from the transfer log.
 102          * well, only if it had been there in the first
 103          * place... if it had not (local only or conflicting
 104          * and never sent), it should still be "empty" as
 105          * initialized in drbd_req_new(), so we can list_del() it
 106          * here unconditionally */
 107         list_del_init(&req->tl_requests);
 108
 109         /* if it was a write, we may have to set the corresponding
 110          * bit(s) out-of-sync first. If it had a local part, we need to
 111          * release the reference to the activity log. */
 112         if (rw == WRITE) {
 113                 /* Set out-of-sync unless both OK flags are set
 114                  * (local only or remote failed).
 115                  * Other places where we set out-of-sync:
 116                  * READ with local io-error */
 117                 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
 118                         drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
 119
 120                 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
 121                         drbd_set_in_sync(mdev, req->i.sector, req->i.size);
 122
 123                 /* one might be tempted to move the drbd_al_complete_io
 124                  * to the local io completion callback drbd_request_endio.
 125                  * but, if this was a mirror write, we may only
 126                  * drbd_al_complete_io after this is RQ_NET_DONE,
 127                  * otherwise the extent could be dropped from the al
 128                  * before it has actually been written on the peer.
 129                  * if we crash before our peer knows about the request,
 130                  * but after the extent has been dropped from the al,
 131                  * we would forget to resync the corresponding extent.
 132                  */
 133                 if (s & RQ_LOCAL_MASK) {
 134                         if (get_ldev_if_state(mdev, D_FAILED)) {
 135                                 if (s & RQ_IN_ACT_LOG)
 136                                         drbd_al_complete_io(mdev, &req->i);
 137                                 put_ldev(mdev);
 138                         } else if (__ratelimit(&drbd_ratelimit_state)) {
 139                                 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
 140                                          "but my Disk seems to have failed :(\n",
 141                                          (unsigned long long) req->i.sector, req->i.size);
 142                         }
 143                 }
 144         }
 145
 146         if (s & RQ_POSTPONED)
 147                 drbd_restart_write(req);
 148         else
 149                 drbd_req_free(req);
 150 }
 151
 152 static void queue_barrier(struct drbd_conf *mdev)
 153 {
 154         struct drbd_tl_epoch *b;
 155         struct drbd_tconn *tconn = mdev->tconn;
 156
 157         /* We are within the req_lock. Once we queued the barrier for sending,
 158          * we set the CREATE_BARRIER bit. It is cleared as soon as a new
 159          * barrier/epoch object is added. This is the only place this bit is
 160          * set. It indicates that the barrier for this epoch is already queued,
 161          * and no new epoch has been created yet. */
 162         if (test_bit(CREATE_BARRIER, &tconn->flags))
 163                 return;
 164
 165         b = tconn->newest_tle;
 166         b->w.cb = w_send_barrier;
 167         b->w.mdev = mdev;
 168         /* inc_ap_pending done here, so we won't
 169          * get imbalanced on connection loss.
 170          * dec_ap_pending will be done in got_BarrierAck
 171          * or (on connection loss) in tl_clear.  */
 172         inc_ap_pending(mdev);
 173         drbd_queue_work(&tconn->data.work, &b->w);
 174         set_bit(CREATE_BARRIER, &tconn->flags);
 175 }
 176
 177 static void _about_to_complete_local_write(struct drbd_conf *mdev,
 178         struct drbd_request *req)
 179 {
 180         const unsigned long s = req->rq_state;
 181
 182         /* Before we can signal completion to the upper layers,
 183          * we may need to close the current epoch.
 184          * We can skip this, if this request has not even been sent, because we
 185          * did not have a fully established connection yet/anymore, during
 186          * bitmap exchange, or while we are C_AHEAD due to congestion policy.
 187          */
 188         if (mdev->state.conn >= C_CONNECTED &&
 189             (s & RQ_NET_SENT) != 0 &&
 190             req->epoch == mdev->tconn->newest_tle->br_number)
 191                 queue_barrier(mdev);
 192 }
 193
 194 void complete_master_bio(struct drbd_conf *mdev,
 195                 struct bio_and_error *m)
 196 {
 197         bio_endio(m->bio, m->error);
 198         dec_ap_bio(mdev);
 199 }
 200
 201
 202 static void drbd_remove_request_interval(struct rb_root *root,
 203                                          struct drbd_request *req)
 204 {
 205         struct drbd_conf *mdev = req->w.mdev;
 206         struct drbd_interval *i = &req->i;
 207
 208         drbd_remove_interval(root, i);
 209
 210         /* Wake up any processes waiting for this request to complete.  */
 211         if (i->waiting)
 212                 wake_up(&mdev->misc_wait);
 213 }
 214
 215 static void maybe_wakeup_conflicting_requests(struct drbd_request *req)
 216 {
 217         const unsigned long s = req->rq_state;
 218         if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
 219                 return;
 220         if (req->i.waiting)
 221                 /* Retry all conflicting peer requests.  */
 222                 wake_up(&req->w.mdev->misc_wait);
 223 }
 224
 225 static
 226 void req_may_be_done(struct drbd_request *req)
 227 {
 228         const unsigned long s = req->rq_state;
 229         struct drbd_conf *mdev = req->w.mdev;
 230         int rw = req->rq_state & RQ_WRITE ? WRITE : READ;
 231
 232         /* req->master_bio still present means: Not yet completed.
 233          *
 234          * Unless this is RQ_POSTPONED, which will cause _req_is_done() to
 235          * queue it on the retry workqueue instead of destroying it.
 236          */
 237         if (req->master_bio && !(s & RQ_POSTPONED))
 238                 return;
 239
 240         /* Local still pending, even though master_bio is already completed?
 241          * may happen for RQ_LOCAL_ABORTED requests. */
 242         if (s & RQ_LOCAL_PENDING)
 243                 return;
 244
 245         if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
 246                 /* this is disconnected (local only) operation,
 247                  * or protocol A, B, or C P_BARRIER_ACK,
 248                  * or killed from the transfer log due to connection loss. */
 249                 _req_is_done(mdev, req, rw);
 250         }
 251         /* else: network part and not DONE yet. that is
 252          * protocol A, B, or C, barrier ack still pending... */
 253 }
 254
 255 /* Helper for __req_mod().
 256  * Set m->bio to the master bio, if it is fit to be completed,
 257  * or leave it alone (it is initialized to NULL in __req_mod),
 258  * if it has already been completed, or cannot be completed yet.
 259  * If m->bio is set, the error status to be returned is placed in m->error.
 260  */
 261 static
 262 void req_may_be_completed(struct drbd_request *req, struct bio_and_error *m)
 263 {
 264         const unsigned long s = req->rq_state;
 265         struct drbd_conf *mdev = req->w.mdev;
 266
 267         /* we must not complete the master bio, while it is
 268          *      still being processed by _drbd_send_zc_bio (drbd_send_dblock)
 269          *      not yet acknowledged by the peer
 270          *      not yet completed by the local io subsystem
 271          * these flags may get cleared in any order by
 272          *      the worker,
 273          *      the receiver,
 274          *      the bio_endio completion callbacks.
 275          */
 276         if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
 277                 return;
 278         if (s & RQ_NET_QUEUED)
 279                 return;
 280         if (s & RQ_NET_PENDING)
 281                 return;
 282
 283         if (req->master_bio) {
 284                 int rw = bio_rw(req->master_bio);
 285
 286                 /* this is DATA_RECEIVED (remote read)
 287                  * or protocol C P_WRITE_ACK
 288                  * or protocol B P_RECV_ACK
 289                  * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
 290                  * or canceled or failed,
 291                  * or killed from the transfer log due to connection loss.
 292                  */
 293
 294                 /*
 295                  * figure out whether to report success or failure.
 296                  *
 297                  * report success when at least one of the operations succeeded.
 298                  * or, to put the other way,
 299                  * only report failure, when both operations failed.
 300                  *
 301                  * what to do about the failures is handled elsewhere.
 302                  * what we need to do here is just: complete the master_bio.
 303                  *
 304                  * local completion error, if any, has been stored as ERR_PTR
 305                  * in private_bio within drbd_request_endio.
 306                  */
 307                 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
 308                 int error = PTR_ERR(req->private_bio);
 309
 310                 /* remove the request from the conflict detection
 311                  * respective block_id verification hash */
 312                 if (!drbd_interval_empty(&req->i)) {
 313                         struct rb_root *root;
 314
 315                         if (rw == WRITE)
 316                                 root = &mdev->write_requests;
 317                         else
 318                                 root = &mdev->read_requests;
 319                         drbd_remove_request_interval(root, req);
 320                 } else if (!(s & RQ_POSTPONED))
 321                         D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
 322
 323                 /* for writes we need to do some extra housekeeping */
 324                 if (rw == WRITE)
 325                         _about_to_complete_local_write(mdev, req);
 326
 327                 /* Update disk stats */
 328                 _drbd_end_io_acct(mdev, req);
 329
 330                 /* if READ failed,
 331                  * have it be pushed back to the retry work queue,
 332                  * so it will re-enter __drbd_make_request,
 333                  * and be re-assigned to a suitable local or remote path,
 334                  * or failed if we do not have access to good data anymore.
 335                  * READA may fail.
 336                  * WRITE should have used all available paths already.
 337                  */
 338                 if (!ok && rw == READ)
 339                         req->rq_state |= RQ_POSTPONED;
 340
 341                 if (!(req->rq_state & RQ_POSTPONED)) {
 342                         m->error = ok ? 0 : (error ?: -EIO);
 343                         m->bio = req->master_bio;
 344                         req->master_bio = NULL;
 345                 } else {
 346                         /* Assert that this will be _req_is_done()
 347                          * with this very invokation. */
 348                         /* FIXME:
 349                          * what about (RQ_LOCAL_PENDING | RQ_LOCAL_ABORTED)?
 350                          */
 351                         D_ASSERT(!(s & RQ_LOCAL_PENDING));
 352                         D_ASSERT((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE));
 353                 }
 354         }
 355         req_may_be_done(req);
 356 }
 357
 358 static void req_may_be_completed_not_susp(struct drbd_request *req, struct bio_and_error *m)
 359 {
 360         struct drbd_conf *mdev = req->w.mdev;
 361
 362         if (!drbd_suspended(mdev))
 363                 req_may_be_completed(req, m);
 364 }
 365
 366 /* obviously this could be coded as many single functions
 367  * instead of one huge switch,
 368  * or by putting the code directly in the respective locations
 369  * (as it has been before).
 370  *
 371  * but having it this way
 372  *  enforces that it is all in this one place, where it is easier to audit,
 373  *  it makes it obvious that whatever "event" "happens" to a request should
 374  *  happen "atomically" within the req_lock,
 375  *  and it enforces that we have to think in a very structured manner
 376  *  about the "events" that may happen to a request during its life time ...
 377  */
 378 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
 379                 struct bio_and_error *m)
 380 {
 381         struct drbd_conf *mdev = req->w.mdev;
 382         struct net_conf *nc;
 383         int p, rv = 0;
 384
 385         if (m)
 386                 m->bio = NULL;
 387
 388         switch (what) {
 389         default:
 390                 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
 391                 break;
 392
 393         /* does not happen...
 394          * initialization done in drbd_req_new
 395         case CREATED:
 396                 break;
 397                 */
 398
 399         case TO_BE_SENT: /* via network */
 400                 /* reached via __drbd_make_request
 401                  * and from w_read_retry_remote */
 402                 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
 403                 req->rq_state |= RQ_NET_PENDING;
 404                 rcu_read_lock();
 405                 nc = rcu_dereference(mdev->tconn->net_conf);
 406                 p = nc->wire_protocol;
 407                 rcu_read_unlock();
 408                 req->rq_state |=
 409                         p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
 410                         p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
 411                 inc_ap_pending(mdev);
 412                 break;
 413
 414         case TO_BE_SUBMITTED: /* locally */
 415                 /* reached via __drbd_make_request */
 416                 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
 417                 req->rq_state |= RQ_LOCAL_PENDING;
 418                 break;
 419
 420         case COMPLETED_OK:
 421                 if (req->rq_state & RQ_WRITE)
 422                         mdev->writ_cnt += req->i.size >> 9;
 423                 else
 424                         mdev->read_cnt += req->i.size >> 9;
 425
 426                 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
 427                 req->rq_state &= ~RQ_LOCAL_PENDING;
 428
 429                 maybe_wakeup_conflicting_requests(req);
 430                 req_may_be_completed_not_susp(req, m);
 431                 break;
 432
 433         case ABORT_DISK_IO:
 434                 req->rq_state |= RQ_LOCAL_ABORTED;
 435                 req_may_be_completed_not_susp(req, m);
 436                 break;
 437
 438         case WRITE_COMPLETED_WITH_ERROR:
 439                 req->rq_state |= RQ_LOCAL_COMPLETED;
 440                 req->rq_state &= ~RQ_LOCAL_PENDING;
 441
 442                 __drbd_chk_io_error(mdev, false);
 443                 maybe_wakeup_conflicting_requests(req);
 444                 req_may_be_completed_not_susp(req, m);
 445                 break;
 446
 447         case READ_AHEAD_COMPLETED_WITH_ERROR:
 448                 /* it is legal to fail READA */
 449                 req->rq_state |= RQ_LOCAL_COMPLETED;
 450                 req->rq_state &= ~RQ_LOCAL_PENDING;
 451                 req_may_be_completed_not_susp(req, m);
 452                 break;
 453
 454         case READ_COMPLETED_WITH_ERROR:
 455                 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
 456
 457                 req->rq_state |= RQ_LOCAL_COMPLETED;
 458                 req->rq_state &= ~RQ_LOCAL_PENDING;
 459
 460                 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
 461
 462                 __drbd_chk_io_error(mdev, false);
 463                 req_may_be_completed_not_susp(req, m);
 464                 break;
 465
 466         case QUEUE_FOR_NET_READ:
 467                 /* READ or READA, and
 468                  * no local disk,
 469                  * or target area marked as invalid,
 470                  * or just got an io-error. */
 471                 /* from __drbd_make_request
 472                  * or from bio_endio during read io-error recovery */
 473
 474                 /* So we can verify the handle in the answer packet.
 475                  * Corresponding drbd_remove_request_interval is in
 476                  * req_may_be_completed() */
 477                 D_ASSERT(drbd_interval_empty(&req->i));
 478                 drbd_insert_interval(&mdev->read_requests, &req->i);
 479
 480                 set_bit(UNPLUG_REMOTE, &mdev->flags);
 481
 482                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 483                 D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0);
 484                 req->rq_state |= RQ_NET_QUEUED;
 485                 req->w.cb = w_send_read_req;
 486                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 487                 break;
 488
 489         case QUEUE_FOR_NET_WRITE:
 490                 /* assert something? */
 491                 /* from __drbd_make_request only */
 492
 493                 /* Corresponding drbd_remove_request_interval is in
 494                  * req_may_be_completed() */
 495                 D_ASSERT(drbd_interval_empty(&req->i));
 496                 drbd_insert_interval(&mdev->write_requests, &req->i);
 497
 498                 /* NOTE
 499                  * In case the req ended up on the transfer log before being
 500                  * queued on the worker, it could lead to this request being
 501                  * missed during cleanup after connection loss.
 502                  * So we have to do both operations here,
 503                  * within the same lock that protects the transfer log.
 504                  *
 505                  * _req_add_to_epoch(req); this has to be after the
 506                  * _maybe_start_new_epoch(req); which happened in
 507                  * __drbd_make_request, because we now may set the bit
 508                  * again ourselves to close the current epoch.
 509                  *
 510                  * Add req to the (now) current epoch (barrier). */
 511
 512                 /* otherwise we may lose an unplug, which may cause some remote
 513                  * io-scheduler timeout to expire, increasing maximum latency,
 514                  * hurting performance. */
 515                 set_bit(UNPLUG_REMOTE, &mdev->flags);
 516
 517                 /* see __drbd_make_request,
 518                  * just after it grabs the req_lock */
 519                 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->tconn->flags) == 0);
 520
 521                 req->epoch = mdev->tconn->newest_tle->br_number;
 522
 523                 /* increment size of current epoch */
 524                 mdev->tconn->newest_tle->n_writes++;
 525
 526                 /* queue work item to send data */
 527                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 528                 req->rq_state |= RQ_NET_QUEUED;
 529                 req->w.cb =  w_send_dblock;
 530                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 531
 532                 /* close the epoch, in case it outgrew the limit */
 533                 rcu_read_lock();
 534                 nc = rcu_dereference(mdev->tconn->net_conf);
 535                 p = nc->max_epoch_size;
 536                 rcu_read_unlock();
 537                 if (mdev->tconn->newest_tle->n_writes >= p)
 538                         queue_barrier(mdev);
 539
 540                 break;
 541
 542         case QUEUE_FOR_SEND_OOS:
 543                 req->rq_state |= RQ_NET_QUEUED;
 544                 req->w.cb =  w_send_out_of_sync;
 545                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 546                 break;
 547
 548         case READ_RETRY_REMOTE_CANCELED:
 549         case SEND_CANCELED:
 550         case SEND_FAILED:
 551                 /* real cleanup will be done from tl_clear.  just update flags
 552                  * so it is no longer marked as on the worker queue */
 553                 req->rq_state &= ~RQ_NET_QUEUED;
 554                 /* if we did it right, tl_clear should be scheduled only after
 555                  * this, so this should not be necessary! */
 556                 req_may_be_completed_not_susp(req, m);
 557                 break;
 558
 559         case HANDED_OVER_TO_NETWORK:
 560                 /* assert something? */
 561                 if (bio_data_dir(req->master_bio) == WRITE)
 562                         atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
 563
 564                 if (bio_data_dir(req->master_bio) == WRITE &&
 565                     !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
 566                         /* this is what is dangerous about protocol A:
 567                          * pretend it was successfully written on the peer. */
 568                         if (req->rq_state & RQ_NET_PENDING) {
 569                                 dec_ap_pending(mdev);
 570                                 req->rq_state &= ~RQ_NET_PENDING;
 571                                 req->rq_state |= RQ_NET_OK;
 572                         } /* else: neg-ack was faster... */
 573                         /* it is still not yet RQ_NET_DONE until the
 574                          * corresponding epoch barrier got acked as well,
 575                          * so we know what to dirty on connection loss */
 576                 }
 577                 req->rq_state &= ~RQ_NET_QUEUED;
 578                 req->rq_state |= RQ_NET_SENT;
 579                 req_may_be_completed_not_susp(req, m);
 580                 break;
 581
 582         case OOS_HANDED_TO_NETWORK:
 583                 /* Was not set PENDING, no longer QUEUED, so is now DONE
 584                  * as far as this connection is concerned. */
 585                 req->rq_state &= ~RQ_NET_QUEUED;
 586                 req->rq_state |= RQ_NET_DONE;
 587                 req_may_be_completed_not_susp(req, m);
 588                 break;
 589
 590         case CONNECTION_LOST_WHILE_PENDING:
 591                 /* transfer log cleanup after connection loss */
 592                 /* assert something? */
 593                 if (req->rq_state & RQ_NET_PENDING)
 594                         dec_ap_pending(mdev);
 595
 596                 p = !(req->rq_state & RQ_WRITE) && req->rq_state & RQ_NET_PENDING;
 597
 598                 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
 599                 req->rq_state |= RQ_NET_DONE;
 600                 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
 601                         atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 602
 603                 req_may_be_completed(req, m); /* Allowed while state.susp */
 604                 break;
 605
 606         case DISCARD_WRITE:
 607                 /* for discarded conflicting writes of multiple primaries,
 608                  * there is no need to keep anything in the tl, potential
 609                  * node crashes are covered by the activity log. */
 610                 req->rq_state |= RQ_NET_DONE;
 611                 /* fall through */
 612         case WRITE_ACKED_BY_PEER_AND_SIS:
 613         case WRITE_ACKED_BY_PEER:
 614                 if (what == WRITE_ACKED_BY_PEER_AND_SIS)
 615                         req->rq_state |= RQ_NET_SIS;
 616                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 617                 /* protocol C; successfully written on peer.
 618                  * Nothing more to do here.
 619                  * We want to keep the tl in place for all protocols, to cater
 620                  * for volatile write-back caches on lower level devices. */
 621
 622                 goto ack_common;
 623         case RECV_ACKED_BY_PEER:
 624                 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
 625                 /* protocol B; pretends to be successfully written on peer.
 626                  * see also notes above in HANDED_OVER_TO_NETWORK about
 627                  * protocol != C */
 628         ack_common:
 629                 req->rq_state |= RQ_NET_OK;
 630                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 631                 dec_ap_pending(mdev);
 632                 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 633                 req->rq_state &= ~RQ_NET_PENDING;
 634                 maybe_wakeup_conflicting_requests(req);
 635                 req_may_be_completed_not_susp(req, m);
 636                 break;
 637
 638         case POSTPONE_WRITE:
 639                 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
 640                 /* If this node has already detected the write conflict, the
 641                  * worker will be waiting on misc_wait.  Wake it up once this
 642                  * request has completed locally.
 643                  */
 644                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 645                 req->rq_state |= RQ_POSTPONED;
 646                 maybe_wakeup_conflicting_requests(req);
 647                 req_may_be_completed_not_susp(req, m);
 648                 break;
 649
 650         case NEG_ACKED:
 651                 /* assert something? */
 652                 if (req->rq_state & RQ_NET_PENDING) {
 653                         dec_ap_pending(mdev);
 654                         if (req->rq_state & RQ_WRITE)
 655                                 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
 656                 }
 657                 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
 658
 659                 req->rq_state |= RQ_NET_DONE;
 660
 661                 maybe_wakeup_conflicting_requests(req);
 662                 req_may_be_completed_not_susp(req, m);
 663                 /* else: done by HANDED_OVER_TO_NETWORK */
 664                 break;
 665
 666         case FAIL_FROZEN_DISK_IO:
 667                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 668                         break;
 669
 670                 req_may_be_completed(req, m); /* Allowed while state.susp */
 671                 break;
 672
 673         case RESTART_FROZEN_DISK_IO:
 674                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 675                         break;
 676
 677                 req->rq_state &= ~RQ_LOCAL_COMPLETED;
 678
 679                 rv = MR_READ;
 680                 if (bio_data_dir(req->master_bio) == WRITE)
 681                         rv = MR_WRITE;
 682
 683                 get_ldev(mdev);
 684                 req->w.cb = w_restart_disk_io;
 685                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 686                 break;
 687
 688         case RESEND:
 689                 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 690                    before the connection loss (B&C only); only P_BARRIER_ACK was missing.
 691                    Throwing them out of the TL here by pretending we got a BARRIER_ACK.
 692                    During connection handshake, we ensure that the peer was not rebooted. */
 693                 if (!(req->rq_state & RQ_NET_OK)) {
 694                         if (req->w.cb) {
 695                                 drbd_queue_work(&mdev->tconn->data.work, &req->w);
 696                                 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 697                         }
 698                         break;
 699                 }
 700                 /* else, fall through to BARRIER_ACKED */
 701
 702         case BARRIER_ACKED:
 703                 if (!(req->rq_state & RQ_WRITE))
 704                         break;
 705
 706                 if (req->rq_state & RQ_NET_PENDING) {
 707                         /* barrier came in before all requests were acked.
 708                          * this is bad, because if the connection is lost now,
 709                          * we won't be able to clean them up... */
 710                         dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
 711                         list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests);
 712                 }
 713                 if ((req->rq_state & RQ_NET_MASK) != 0) {
 714                         req->rq_state |= RQ_NET_DONE;
 715                         if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK)))
 716                                 atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
 717                 }
 718                 req_may_be_done(req); /* Allowed while state.susp */
 719                 break;
 720
 721         case DATA_RECEIVED:
 722                 D_ASSERT(req->rq_state & RQ_NET_PENDING);
 723                 dec_ap_pending(mdev);
 724                 req->rq_state &= ~RQ_NET_PENDING;
 725                 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
 726                 req_may_be_completed_not_susp(req, m);
 727                 break;
 728         };
 729
 730         return rv;
 731 }
 732
 733 /* we may do a local read if:
 734  * - we are consistent (of course),
 735  * - or we are generally inconsistent,
 736  *   BUT we are still/already IN SYNC for this area.
 737  *   since size may be bigger than BM_BLOCK_SIZE,
 738  *   we may need to check several bits.
 739  */
 740 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
 741 {
 742         unsigned long sbnr, ebnr;
 743         sector_t esector, nr_sectors;
 744
 745         if (mdev->state.disk == D_UP_TO_DATE)
 746                 return true;
 747         if (mdev->state.disk != D_INCONSISTENT)
 748                 return false;
 749         esector = sector + (size >> 9) - 1;
 750         nr_sectors = drbd_get_capacity(mdev->this_bdev);
 751         D_ASSERT(sector  < nr_sectors);
 752         D_ASSERT(esector < nr_sectors);
 753
 754         sbnr = BM_SECT_TO_BIT(sector);
 755         ebnr = BM_SECT_TO_BIT(esector);
 756
 757         return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
 758 }
 759
 760 static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector)
 761 {
 762         enum drbd_read_balancing rbm;
 763         struct backing_dev_info *bdi;
 764         int stripe_shift;
 765
 766         if (mdev->state.pdsk < D_UP_TO_DATE)
 767                 return false;
 768
 769         rcu_read_lock();
 770         rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
 771         rcu_read_unlock();
 772
 773         switch (rbm) {
 774         case RB_CONGESTED_REMOTE:
 775                 bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
 776                 return bdi_read_congested(bdi);
 777         case RB_LEAST_PENDING:
 778                 return atomic_read(&mdev->local_cnt) >
 779                         atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
 780         case RB_32K_STRIPING:  /* stripe_shift = 15 */
 781         case RB_64K_STRIPING:
 782         case RB_128K_STRIPING:
 783         case RB_256K_STRIPING:
 784         case RB_512K_STRIPING:
 785         case RB_1M_STRIPING:   /* stripe_shift = 20 */
 786                 stripe_shift = (rbm - RB_32K_STRIPING + 15);
 787                 return (sector >> (stripe_shift - 9)) & 1;
 788         case RB_ROUND_ROBIN:
 789                 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
 790         case RB_PREFER_REMOTE:
 791                 return true;
 792         case RB_PREFER_LOCAL:
 793         default:
 794                 return false;
 795         }
 796 }
 797
 798 /*
 799  * complete_conflicting_writes  -  wait for any conflicting write requests
 800  *
 801  * The write_requests tree contains all active write requests which we
 802  * currently know about.  Wait for any requests to complete which conflict with
 803  * the new one.
 804  *
 805  * Only way out: remove the conflicting intervals from the tree.
 806  */
 807 static void complete_conflicting_writes(struct drbd_request *req)
 808 {
 809         DEFINE_WAIT(wait);
 810         struct drbd_conf *mdev = req->w.mdev;
 811         struct drbd_interval *i;
 812         sector_t sector = req->i.sector;
 813         int size = req->i.size;
 814
 815         i = drbd_find_overlap(&mdev->write_requests, sector, size);
 816         if (!i)
 817                 return;
 818
 819         for (;;) {
 820                 prepare_to_wait(&mdev->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
 821                 i = drbd_find_overlap(&mdev->write_requests, sector, size);
 822                 if (!i)
 823                         break;
 824                 /* Indicate to wake up device->misc_wait on progress.  */
 825                 i->waiting = true;
 826                 spin_unlock_irq(&mdev->tconn->req_lock);
 827                 schedule();
 828                 spin_lock_irq(&mdev->tconn->req_lock);
 829         }
 830         finish_wait(&mdev->misc_wait, &wait);
 831 }
 832
 833 int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
 834 {
 835         const int rw = bio_rw(bio);
 836         const int size = bio->bi_size;
 837         const sector_t sector = bio->bi_sector;
 838         struct drbd_tl_epoch *b = NULL;
 839         struct drbd_request *req;
 840         struct net_conf *nc;
 841         int local, remote, send_oos = 0;
 842         int err = 0;
 843         int ret = 0;
 844         union drbd_dev_state s;
 845
 846         /* allocate outside of all locks; */
 847         req = drbd_req_new(mdev, bio);
 848         if (!req) {
 849                 dec_ap_bio(mdev);
 850                 /* only pass the error to the upper layers.
 851                  * if user cannot handle io errors, that's not our business. */
 852                 dev_err(DEV, "could not kmalloc() req\n");
 853                 bio_endio(bio, -ENOMEM);
 854                 return 0;
 855         }
 856         req->start_time = start_time;
 857
 858         local = get_ldev(mdev);
 859         if (!local) {
 860                 bio_put(req->private_bio); /* or we get a bio leak */
 861                 req->private_bio = NULL;
 862         }
 863         if (rw == WRITE) {
 864                 remote = 1;
 865         } else {
 866                 /* READ || READA */
 867                 if (local) {
 868                         if (!drbd_may_do_local_read(mdev, sector, size) ||
 869                             remote_due_to_read_balancing(mdev, sector)) {
 870                                 /* we could kick the syncer to
 871                                  * sync this extent asap, wait for
 872                                  * it, then continue locally.
 873                                  * Or just issue the request remotely.
 874                                  */
 875                                 local = 0;
 876                                 bio_put(req->private_bio);
 877                                 req->private_bio = NULL;
 878                                 put_ldev(mdev);
 879                         }
 880                 }
 881                 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
 882         }
 883
 884         /* If we have a disk, but a READA request is mapped to remote,
 885          * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
 886          * Just fail that READA request right here.
 887          *
 888          * THINK: maybe fail all READA when not local?
 889          *        or make this configurable...
 890          *        if network is slow, READA won't do any good.
 891          */
 892         if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
 893                 err = -EWOULDBLOCK;
 894                 goto fail_and_free_req;
 895         }
 896
 897         /* For WRITES going to the local disk, grab a reference on the target
 898          * extent.  This waits for any resync activity in the corresponding
 899          * resync extent to finish, and, if necessary, pulls in the target
 900          * extent into the activity log, which involves further disk io because
 901          * of transactional on-disk meta data updates. */
 902         if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
 903                 req->rq_state |= RQ_IN_ACT_LOG;
 904                 drbd_al_begin_io(mdev, &req->i);
 905         }
 906
 907         s = mdev->state;
 908         remote = remote && drbd_should_do_remote(s);
 909         send_oos = rw == WRITE && drbd_should_send_out_of_sync(s);
 910         D_ASSERT(!(remote && send_oos));
 911
 912         if (!(local || remote) && !drbd_suspended(mdev)) {
 913                 if (__ratelimit(&drbd_ratelimit_state))
 914                         dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
 915                 err = -EIO;
 916                 goto fail_free_complete;
 917         }
 918
 919         /* For WRITE request, we have to make sure that we have an
 920          * unused_spare_tle, in case we need to start a new epoch.
 921          * I try to be smart and avoid to pre-allocate always "just in case",
 922          * but there is a race between testing the bit and pointer outside the
 923          * spinlock, and grabbing the spinlock.
 924          * if we lost that race, we retry.  */
 925         if (rw == WRITE && (remote || send_oos) &&
 926             mdev->tconn->unused_spare_tle == NULL &&
 927             test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
 928 allocate_barrier:
 929                 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
 930                 if (!b) {
 931                         dev_err(DEV, "Failed to alloc barrier.\n");
 932                         err = -ENOMEM;
 933                         goto fail_free_complete;
 934                 }
 935         }
 936
 937         /* GOOD, everything prepared, grab the spin_lock */
 938         spin_lock_irq(&mdev->tconn->req_lock);
 939
 940         if (rw == WRITE) {
 941                 /* This may temporarily give up the req_lock,
 942                  * but will re-aquire it before it returns here.
 943                  * Needs to be before the check on drbd_suspended() */
 944                 complete_conflicting_writes(req);
 945         }
 946
 947         if (drbd_suspended(mdev)) {
 948                 /* If we got suspended, use the retry mechanism in
 949                    drbd_make_request() to restart processing of this
 950                    bio. In the next call to drbd_make_request
 951                    we sleep in inc_ap_bio() */
 952                 ret = 1;
 953                 spin_unlock_irq(&mdev->tconn->req_lock);
 954                 goto fail_free_complete;
 955         }
 956
 957         if (remote || send_oos) {
 958                 remote = drbd_should_do_remote(mdev->state);
 959                 send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
 960                 D_ASSERT(!(remote && send_oos));
 961
 962                 if (!(remote || send_oos))
 963                         dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
 964                 if (!(local || remote)) {
 965                         dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
 966                         spin_unlock_irq(&mdev->tconn->req_lock);
 967                         err = -EIO;
 968                         goto fail_free_complete;
 969                 }
 970         }
 971
 972         if (b && mdev->tconn->unused_spare_tle == NULL) {
 973                 mdev->tconn->unused_spare_tle = b;
 974                 b = NULL;
 975         }
 976         if (rw == WRITE && (remote || send_oos) &&
 977             mdev->tconn->unused_spare_tle == NULL &&
 978             test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
 979                 /* someone closed the current epoch
 980                  * while we were grabbing the spinlock */
 981                 spin_unlock_irq(&mdev->tconn->req_lock);
 982                 goto allocate_barrier;
 983         }
 984
 985
 986         /* Update disk stats */
 987         _drbd_start_io_acct(mdev, req, bio);
 988
 989         /* _maybe_start_new_epoch(mdev);
 990          * If we need to generate a write barrier packet, we have to add the
 991          * new epoch (barrier) object, and queue the barrier packet for sending,
 992          * and queue the req's data after it _within the same lock_, otherwise
 993          * we have race conditions were the reorder domains could be mixed up.
 994          *
 995          * Even read requests may start a new epoch and queue the corresponding
 996          * barrier packet.  To get the write ordering right, we only have to
 997          * make sure that, if this is a write request and it triggered a
 998          * barrier packet, this request is queued within the same spinlock. */
 999         if ((remote || send_oos) && mdev->tconn->unused_spare_tle &&
1000             test_and_clear_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
1001                 _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle);
1002                 mdev->tconn->unused_spare_tle = NULL;
1003         } else {
1004                 D_ASSERT(!(remote && rw == WRITE &&
1005                            test_bit(CREATE_BARRIER, &mdev->tconn->flags)));
1006         }
1007
1008         /* NOTE
1009          * Actually, 'local' may be wrong here already, since we may have failed
1010          * to write to the meta data, and may become wrong anytime because of
1011          * local io-error for some other request, which would lead to us
1012          * "detaching" the local disk.
1013          *
1014          * 'remote' may become wrong any time because the network could fail.
1015          *
1016          * This is a harmless race condition, though, since it is handled
1017          * correctly at the appropriate places; so it just defers the failure
1018          * of the respective operation.
1019          */
1020
1021         /* mark them early for readability.
1022          * this just sets some state flags. */
1023         if (remote)
1024                 _req_mod(req, TO_BE_SENT);
1025         if (local)
1026                 _req_mod(req, TO_BE_SUBMITTED);
1027
1028         list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests);
1029
1030         /* NOTE remote first: to get the concurrent write detection right,
1031          * we must register the request before start of local IO.  */
1032         if (remote) {
1033                 /* either WRITE and C_CONNECTED,
1034                  * or READ, and no local disk,
1035                  * or READ, but not in sync.
1036                  */
1037                 _req_mod(req, (rw == WRITE)
1038                                 ? QUEUE_FOR_NET_WRITE
1039                                 : QUEUE_FOR_NET_READ);
1040         }
1041         if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
1042                 _req_mod(req, QUEUE_FOR_SEND_OOS);
1043
1044         rcu_read_lock();
1045         nc = rcu_dereference(mdev->tconn->net_conf);
1046         if (remote &&
1047             nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) {
1048                 int congested = 0;
1049
1050                 if (nc->cong_fill &&
1051                     atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
1052                         dev_info(DEV, "Congestion-fill threshold reached\n");
1053                         congested = 1;
1054                 }
1055
1056                 if (mdev->act_log->used >= nc->cong_extents) {
1057                         dev_info(DEV, "Congestion-extents threshold reached\n");
1058                         congested = 1;
1059                 }
1060
1061                 if (congested) {
1062                         queue_barrier(mdev); /* last barrier, after mirrored writes */
1063
1064                         if (nc->on_congestion == OC_PULL_AHEAD)
1065                                 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
1066                         else  /*nc->on_congestion == OC_DISCONNECT */
1067                                 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
1068                 }
1069         }
1070         rcu_read_unlock();
1071
1072         spin_unlock_irq(&mdev->tconn->req_lock);
1073         kfree(b); /* if someone else has beaten us to it... */
1074
1075         if (local) {
1076                 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
1077
1078                 /* State may have changed since we grabbed our reference on the
1079                  * mdev->ldev member. Double check, and short-circuit to endio.
1080                  * In case the last activity log transaction failed to get on
1081                  * stable storage, and this is a WRITE, we may not even submit
1082                  * this bio. */
1083                 if (get_ldev(mdev)) {
1084                         if (drbd_insert_fault(mdev,   rw == WRITE ? DRBD_FAULT_DT_WR
1085                                                     : rw == READ  ? DRBD_FAULT_DT_RD
1086                                                     :               DRBD_FAULT_DT_RA))
1087                                 bio_endio(req->private_bio, -EIO);
1088                         else
1089                                 generic_make_request(req->private_bio);
1090                         put_ldev(mdev);
1091                 } else
1092                         bio_endio(req->private_bio, -EIO);
1093         }
1094
1095         return 0;
1096
1097 fail_free_complete:
1098         if (req->rq_state & RQ_IN_ACT_LOG)
1099                 drbd_al_complete_io(mdev, &req->i);
1100 fail_and_free_req:
1101         if (local) {
1102                 bio_put(req->private_bio);
1103                 req->private_bio = NULL;
1104                 put_ldev(mdev);
1105         }
1106         if (!ret)
1107                 bio_endio(bio, err);
1108
1109         drbd_req_free(req);
1110         dec_ap_bio(mdev);
1111         kfree(b);
1112
1113         return ret;
1114 }
1115
1116 int drbd_make_request(struct request_queue *q, struct bio *bio)
1117 {
1118         struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1119         unsigned long start_time;
1120
1121         start_time = jiffies;
1122
1123         /*
1124          * what we "blindly" assume:
1125          */
1126         D_ASSERT(bio->bi_size > 0);
1127         D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1128
1129         do {
1130                 inc_ap_bio(mdev);
1131         } while (__drbd_make_request(mdev, bio, start_time));
1132
1133         return 0;
1134 }
1135
1136 /* This is called by bio_add_page().
1137  *
1138  * q->max_hw_sectors and other global limits are already enforced there.
1139  *
1140  * We need to call down to our lower level device,
1141  * in case it has special restrictions.
1142  *
1143  * We also may need to enforce configured max-bio-bvecs limits.
1144  *
1145  * As long as the BIO is empty we have to allow at least one bvec,
1146  * regardless of size and offset, so no need to ask lower levels.
1147  */
1148 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1149 {
1150         struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1151         unsigned int bio_size = bvm->bi_size;
1152         int limit = DRBD_MAX_BIO_SIZE;
1153         int backing_limit;
1154
1155         if (bio_size && get_ldev(mdev)) {
1156                 struct request_queue * const b =
1157                         mdev->ldev->backing_bdev->bd_disk->queue;
1158                 if (b->merge_bvec_fn) {
1159                         backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1160                         limit = min(limit, backing_limit);
1161                 }
1162                 put_ldev(mdev);
1163         }
1164         return limit;
1165 }
1166
1167 void request_timer_fn(unsigned long data)
1168 {
1169         struct drbd_conf *mdev = (struct drbd_conf *) data;
1170         struct drbd_tconn *tconn = mdev->tconn;
1171         struct drbd_request *req; /* oldest request */
1172         struct list_head *le;
1173         struct net_conf *nc;
1174         unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1175         unsigned long now;
1176
1177         rcu_read_lock();
1178         nc = rcu_dereference(tconn->net_conf);
1179         if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS)
1180                 ent = nc->timeout * HZ/10 * nc->ko_count;
1181
1182         if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */
1183                 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1184                 put_ldev(mdev);
1185         }
1186         rcu_read_unlock();
1187
1188         et = min_not_zero(dt, ent);
1189
1190         if (!et)
1191                 return; /* Recurring timer stopped */
1192
1193         now = jiffies;
1194
1195         spin_lock_irq(&tconn->req_lock);
1196         le = &tconn->oldest_tle->requests;
1197         if (list_empty(le)) {
1198                 spin_unlock_irq(&tconn->req_lock);
1199                 mod_timer(&mdev->request_timer, now + et);
1200                 return;
1201         }
1202
1203         le = le->prev;
1204         req = list_entry(le, struct drbd_request, tl_requests);
1205
1206         /* The request is considered timed out, if
1207          * - we have some effective timeout from the configuration,
1208          *   with above state restrictions applied,
1209          * - the oldest request is waiting for a response from the network
1210          *   resp. the local disk,
1211          * - the oldest request is in fact older than the effective timeout,
1212          * - the connection was established (resp. disk was attached)
1213          *   for longer than the timeout already.
1214          * Note that for 32bit jiffies and very stable connections/disks,
1215          * we may have a wrap around, which is catched by
1216          *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1217          *
1218          * Side effect: once per 32bit wrap-around interval, which means every
1219          * ~198 days with 250 HZ, we have a window where the timeout would need
1220          * to expire twice (worst case) to become effective. Good enough.
1221          */
1222         if (ent && req->rq_state & RQ_NET_PENDING &&
1223                  time_after(now, req->start_time + ent) &&
1224                 !time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) {
1225                 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1226                 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1227         }
1228         if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev &&
1229                  time_after(now, req->start_time + dt) &&
1230                 !time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) {
1231                 dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1232                 __drbd_chk_io_error(mdev, 1);
1233         }
1234         nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
1235         spin_unlock_irq(&tconn->req_lock);
1236         mod_timer(&mdev->request_timer, nt);
1237 }