drivers/block/drbd/drbd_req.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3    drbd_req.c
   4
   5    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   6
   7    Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
   8    Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   9    Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
  10
  11
  12  */
  13
  14 #include <linux/module.h>
  15
  16 #include <linux/slab.h>
  17 #include <linux/drbd.h>
  18 #include "drbd_int.h"
  19 #include "drbd_req.h"
  20
  21
  22 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
  23
  24 static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src)
  25 {
  26         struct drbd_request *req;
  27
  28         req = mempool_alloc(&drbd_request_mempool, GFP_NOIO);
  29         if (!req)
  30                 return NULL;
  31         memset(req, 0, sizeof(*req));
  32
  33         req->private_bio = bio_alloc_clone(device->ldev->backing_bdev, bio_src,
  34                                            GFP_NOIO, &drbd_io_bio_set);
  35         req->private_bio->bi_private = req;
  36         req->private_bio->bi_end_io = drbd_request_endio;
  37
  38         req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
  39                       | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0)
  40                       | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
  41         req->device = device;
  42         req->master_bio = bio_src;
  43         req->epoch = 0;
  44
  45         drbd_clear_interval(&req->i);
  46         req->i.sector     = bio_src->bi_iter.bi_sector;
  47         req->i.size      = bio_src->bi_iter.bi_size;
  48         req->i.local = true;
  49         req->i.waiting = false;
  50
  51         INIT_LIST_HEAD(&req->tl_requests);
  52         INIT_LIST_HEAD(&req->w.list);
  53         INIT_LIST_HEAD(&req->req_pending_master_completion);
  54         INIT_LIST_HEAD(&req->req_pending_local);
  55
  56         /* one reference to be put by __drbd_make_request */
  57         atomic_set(&req->completion_ref, 1);
  58         /* one kref as long as completion_ref > 0 */
  59         kref_init(&req->kref);
  60         return req;
  61 }
  62
  63 static void drbd_remove_request_interval(struct rb_root *root,
  64                                          struct drbd_request *req)
  65 {
  66         struct drbd_device *device = req->device;
  67         struct drbd_interval *i = &req->i;
  68
  69         drbd_remove_interval(root, i);
  70
  71         /* Wake up any processes waiting for this request to complete.  */
  72         if (i->waiting)
  73                 wake_up(&device->misc_wait);
  74 }
  75
  76 void drbd_req_destroy(struct kref *kref)
  77 {
  78         struct drbd_request *req = container_of(kref, struct drbd_request, kref);
  79         struct drbd_device *device = req->device;
  80         const unsigned s = req->rq_state;
  81
  82         if ((req->master_bio && !(s & RQ_POSTPONED)) ||
  83                 atomic_read(&req->completion_ref) ||
  84                 (s & RQ_LOCAL_PENDING) ||
  85                 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
  86                 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
  87                                 s, atomic_read(&req->completion_ref));
  88                 return;
  89         }
  90
  91         /* If called from mod_rq_state (expected normal case) or
  92          * drbd_send_and_submit (the less likely normal path), this holds the
  93          * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
  94          * though it may be still empty (never added to the transfer log).
  95          *
  96          * If called from do_retry(), we do NOT hold the req_lock, but we are
  97          * still allowed to unconditionally list_del(&req->tl_requests),
  98          * because it will be on a local on-stack list only. */
  99         list_del_init(&req->tl_requests);
 100
 101         /* finally remove the request from the conflict detection
 102          * respective block_id verification interval tree. */
 103         if (!drbd_interval_empty(&req->i)) {
 104                 struct rb_root *root;
 105
 106                 if (s & RQ_WRITE)
 107                         root = &device->write_requests;
 108                 else
 109                         root = &device->read_requests;
 110                 drbd_remove_request_interval(root, req);
 111         } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
 112                 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
 113                         s, (unsigned long long)req->i.sector, req->i.size);
 114
 115         /* if it was a write, we may have to set the corresponding
 116          * bit(s) out-of-sync first. If it had a local part, we need to
 117          * release the reference to the activity log. */
 118         if (s & RQ_WRITE) {
 119                 /* Set out-of-sync unless both OK flags are set
 120                  * (local only or remote failed).
 121                  * Other places where we set out-of-sync:
 122                  * READ with local io-error */
 123
 124                 /* There is a special case:
 125                  * we may notice late that IO was suspended,
 126                  * and postpone, or schedule for retry, a write,
 127                  * before it even was submitted or sent.
 128                  * In that case we do not want to touch the bitmap at all.
 129                  */
 130                 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
 131                         if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
 132                                 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
 133
 134                         if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
 135                                 drbd_set_in_sync(device, req->i.sector, req->i.size);
 136                 }
 137
 138                 /* one might be tempted to move the drbd_al_complete_io
 139                  * to the local io completion callback drbd_request_endio.
 140                  * but, if this was a mirror write, we may only
 141                  * drbd_al_complete_io after this is RQ_NET_DONE,
 142                  * otherwise the extent could be dropped from the al
 143                  * before it has actually been written on the peer.
 144                  * if we crash before our peer knows about the request,
 145                  * but after the extent has been dropped from the al,
 146                  * we would forget to resync the corresponding extent.
 147                  */
 148                 if (s & RQ_IN_ACT_LOG) {
 149                         if (get_ldev_if_state(device, D_FAILED)) {
 150                                 drbd_al_complete_io(device, &req->i);
 151                                 put_ldev(device);
 152                         } else if (__ratelimit(&drbd_ratelimit_state)) {
 153                                 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
 154                                          "but my Disk seems to have failed :(\n",
 155                                          (unsigned long long) req->i.sector, req->i.size);
 156                         }
 157                 }
 158         }
 159
 160         mempool_free(req, &drbd_request_mempool);
 161 }
 162
 163 static void wake_all_senders(struct drbd_connection *connection)
 164 {
 165         wake_up(&connection->sender_work.q_wait);
 166 }
 167
 168 /* must hold resource->req_lock */
 169 void start_new_tl_epoch(struct drbd_connection *connection)
 170 {
 171         /* no point closing an epoch, if it is empty, anyways. */
 172         if (connection->current_tle_writes == 0)
 173                 return;
 174
 175         connection->current_tle_writes = 0;
 176         atomic_inc(&connection->current_tle_nr);
 177         wake_all_senders(connection);
 178 }
 179
 180 void complete_master_bio(struct drbd_device *device,
 181                 struct bio_and_error *m)
 182 {
 183         if (unlikely(m->error))
 184                 m->bio->bi_status = errno_to_blk_status(m->error);
 185         bio_endio(m->bio);
 186         dec_ap_bio(device);
 187 }
 188
 189
 190 /* Helper for __req_mod().
 191  * Set m->bio to the master bio, if it is fit to be completed,
 192  * or leave it alone (it is initialized to NULL in __req_mod),
 193  * if it has already been completed, or cannot be completed yet.
 194  * If m->bio is set, the error status to be returned is placed in m->error.
 195  */
 196 static
 197 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
 198 {
 199         const unsigned s = req->rq_state;
 200         struct drbd_device *device = req->device;
 201         int error, ok;
 202
 203         /* we must not complete the master bio, while it is
 204          *      still being processed by _drbd_send_zc_bio (drbd_send_dblock)
 205          *      not yet acknowledged by the peer
 206          *      not yet completed by the local io subsystem
 207          * these flags may get cleared in any order by
 208          *      the worker,
 209          *      the receiver,
 210          *      the bio_endio completion callbacks.
 211          */
 212         if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
 213             (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
 214             (s & RQ_COMPLETION_SUSP)) {
 215                 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
 216                 return;
 217         }
 218
 219         if (!req->master_bio) {
 220                 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
 221                 return;
 222         }
 223
 224         /*
 225          * figure out whether to report success or failure.
 226          *
 227          * report success when at least one of the operations succeeded.
 228          * or, to put the other way,
 229          * only report failure, when both operations failed.
 230          *
 231          * what to do about the failures is handled elsewhere.
 232          * what we need to do here is just: complete the master_bio.
 233          *
 234          * local completion error, if any, has been stored as ERR_PTR
 235          * in private_bio within drbd_request_endio.
 236          */
 237         ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
 238         error = PTR_ERR(req->private_bio);
 239
 240         /* Before we can signal completion to the upper layers,
 241          * we may need to close the current transfer log epoch.
 242          * We are within the request lock, so we can simply compare
 243          * the request epoch number with the current transfer log
 244          * epoch number.  If they match, increase the current_tle_nr,
 245          * and reset the transfer log epoch write_cnt.
 246          */
 247         if (op_is_write(bio_op(req->master_bio)) &&
 248             req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
 249                 start_new_tl_epoch(first_peer_device(device)->connection);
 250
 251         /* Update disk stats */
 252         bio_end_io_acct(req->master_bio, req->start_jif);
 253
 254         /* If READ failed,
 255          * have it be pushed back to the retry work queue,
 256          * so it will re-enter __drbd_make_request(),
 257          * and be re-assigned to a suitable local or remote path,
 258          * or failed if we do not have access to good data anymore.
 259          *
 260          * Unless it was failed early by __drbd_make_request(),
 261          * because no path was available, in which case
 262          * it was not even added to the transfer_log.
 263          *
 264          * read-ahead may fail, and will not be retried.
 265          *
 266          * WRITE should have used all available paths already.
 267          */
 268         if (!ok &&
 269             bio_op(req->master_bio) == REQ_OP_READ &&
 270             !(req->master_bio->bi_opf & REQ_RAHEAD) &&
 271             !list_empty(&req->tl_requests))
 272                 req->rq_state |= RQ_POSTPONED;
 273
 274         if (!(req->rq_state & RQ_POSTPONED)) {
 275                 m->error = ok ? 0 : (error ?: -EIO);
 276                 m->bio = req->master_bio;
 277                 req->master_bio = NULL;
 278                 /* We leave it in the tree, to be able to verify later
 279                  * write-acks in protocol != C during resync.
 280                  * But we mark it as "complete", so it won't be counted as
 281                  * conflict in a multi-primary setup. */
 282                 req->i.completed = true;
 283         }
 284
 285         if (req->i.waiting)
 286                 wake_up(&device->misc_wait);
 287
 288         /* Either we are about to complete to upper layers,
 289          * or we will restart this request.
 290          * In either case, the request object will be destroyed soon,
 291          * so better remove it from all lists. */
 292         list_del_init(&req->req_pending_master_completion);
 293 }
 294
 295 /* still holds resource->req_lock */
 296 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
 297 {
 298         struct drbd_device *device = req->device;
 299         D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
 300
 301         if (!put)
 302                 return;
 303
 304         if (!atomic_sub_and_test(put, &req->completion_ref))
 305                 return;
 306
 307         drbd_req_complete(req, m);
 308
 309         /* local completion may still come in later,
 310          * we need to keep the req object around. */
 311         if (req->rq_state & RQ_LOCAL_ABORTED)
 312                 return;
 313
 314         if (req->rq_state & RQ_POSTPONED) {
 315                 /* don't destroy the req object just yet,
 316                  * but queue it for retry */
 317                 drbd_restart_request(req);
 318                 return;
 319         }
 320
 321         kref_put(&req->kref, drbd_req_destroy);
 322 }
 323
 324 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
 325 {
 326         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 327         if (!connection)
 328                 return;
 329         if (connection->req_next == NULL)
 330                 connection->req_next = req;
 331 }
 332
 333 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
 334 {
 335         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 336         struct drbd_request *iter = req;
 337         if (!connection)
 338                 return;
 339         if (connection->req_next != req)
 340                 return;
 341
 342         req = NULL;
 343         list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) {
 344                 const unsigned int s = iter->rq_state;
 345
 346                 if (s & RQ_NET_QUEUED) {
 347                         req = iter;
 348                         break;
 349                 }
 350         }
 351         connection->req_next = req;
 352 }
 353
 354 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
 355 {
 356         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 357         if (!connection)
 358                 return;
 359         if (connection->req_ack_pending == NULL)
 360                 connection->req_ack_pending = req;
 361 }
 362
 363 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
 364 {
 365         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 366         struct drbd_request *iter = req;
 367         if (!connection)
 368                 return;
 369         if (connection->req_ack_pending != req)
 370                 return;
 371
 372         req = NULL;
 373         list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) {
 374                 const unsigned int s = iter->rq_state;
 375
 376                 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING)) {
 377                         req = iter;
 378                         break;
 379                 }
 380         }
 381         connection->req_ack_pending = req;
 382 }
 383
 384 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
 385 {
 386         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 387         if (!connection)
 388                 return;
 389         if (connection->req_not_net_done == NULL)
 390                 connection->req_not_net_done = req;
 391 }
 392
 393 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
 394 {
 395         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 396         struct drbd_request *iter = req;
 397         if (!connection)
 398                 return;
 399         if (connection->req_not_net_done != req)
 400                 return;
 401
 402         req = NULL;
 403         list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) {
 404                 const unsigned int s = iter->rq_state;
 405
 406                 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE)) {
 407                         req = iter;
 408                         break;
 409                 }
 410         }
 411         connection->req_not_net_done = req;
 412 }
 413
 414 /* I'd like this to be the only place that manipulates
 415  * req->completion_ref and req->kref. */
 416 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
 417                 int clear, int set)
 418 {
 419         struct drbd_device *device = req->device;
 420         struct drbd_peer_device *peer_device = first_peer_device(device);
 421         unsigned s = req->rq_state;
 422         int c_put = 0;
 423
 424         if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
 425                 set |= RQ_COMPLETION_SUSP;
 426
 427         /* apply */
 428
 429         req->rq_state &= ~clear;
 430         req->rq_state |= set;
 431
 432         /* no change? */
 433         if (req->rq_state == s)
 434                 return;
 435
 436         /* intent: get references */
 437
 438         kref_get(&req->kref);
 439
 440         if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
 441                 atomic_inc(&req->completion_ref);
 442
 443         if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
 444                 inc_ap_pending(device);
 445                 atomic_inc(&req->completion_ref);
 446         }
 447
 448         if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
 449                 atomic_inc(&req->completion_ref);
 450                 set_if_null_req_next(peer_device, req);
 451         }
 452
 453         if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
 454                 kref_get(&req->kref); /* wait for the DONE */
 455
 456         if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
 457                 /* potentially already completed in the ack_receiver thread */
 458                 if (!(s & RQ_NET_DONE)) {
 459                         atomic_add(req->i.size >> 9, &device->ap_in_flight);
 460                         set_if_null_req_not_net_done(peer_device, req);
 461                 }
 462                 if (req->rq_state & RQ_NET_PENDING)
 463                         set_if_null_req_ack_pending(peer_device, req);
 464         }
 465
 466         if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
 467                 atomic_inc(&req->completion_ref);
 468
 469         /* progress: put references */
 470
 471         if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
 472                 ++c_put;
 473
 474         if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
 475                 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
 476                 ++c_put;
 477         }
 478
 479         if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
 480                 if (req->rq_state & RQ_LOCAL_ABORTED)
 481                         kref_put(&req->kref, drbd_req_destroy);
 482                 else
 483                         ++c_put;
 484                 list_del_init(&req->req_pending_local);
 485         }
 486
 487         if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
 488                 dec_ap_pending(device);
 489                 ++c_put;
 490                 req->acked_jif = jiffies;
 491                 advance_conn_req_ack_pending(peer_device, req);
 492         }
 493
 494         if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
 495                 ++c_put;
 496                 advance_conn_req_next(peer_device, req);
 497         }
 498
 499         if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
 500                 if (s & RQ_NET_SENT)
 501                         atomic_sub(req->i.size >> 9, &device->ap_in_flight);
 502                 if (s & RQ_EXP_BARR_ACK)
 503                         kref_put(&req->kref, drbd_req_destroy);
 504                 req->net_done_jif = jiffies;
 505
 506                 /* in ahead/behind mode, or just in case,
 507                  * before we finally destroy this request,
 508                  * the caching pointers must not reference it anymore */
 509                 advance_conn_req_next(peer_device, req);
 510                 advance_conn_req_ack_pending(peer_device, req);
 511                 advance_conn_req_not_net_done(peer_device, req);
 512         }
 513
 514         /* potentially complete and destroy */
 515
 516         /* If we made progress, retry conflicting peer requests, if any. */
 517         if (req->i.waiting)
 518                 wake_up(&device->misc_wait);
 519
 520         drbd_req_put_completion_ref(req, m, c_put);
 521         kref_put(&req->kref, drbd_req_destroy);
 522 }
 523
 524 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
 525 {
 526         char b[BDEVNAME_SIZE];
 527
 528         if (!__ratelimit(&drbd_ratelimit_state))
 529                 return;
 530
 531         drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
 532                         (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
 533                         (unsigned long long)req->i.sector,
 534                         req->i.size >> 9,
 535                         bdevname(device->ldev->backing_bdev, b));
 536 }
 537
 538 /* Helper for HANDED_OVER_TO_NETWORK.
 539  * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
 540  * Is it also still "PENDING"?
 541  * --> If so, clear PENDING and set NET_OK below.
 542  * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
 543  * (and we must not set RQ_NET_OK) */
 544 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
 545 {
 546         return (req->rq_state &
 547                    (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
 548                 == (RQ_WRITE|RQ_NET_PENDING);
 549 }
 550
 551 /* obviously this could be coded as many single functions
 552  * instead of one huge switch,
 553  * or by putting the code directly in the respective locations
 554  * (as it has been before).
 555  *
 556  * but having it this way
 557  *  enforces that it is all in this one place, where it is easier to audit,
 558  *  it makes it obvious that whatever "event" "happens" to a request should
 559  *  happen "atomically" within the req_lock,
 560  *  and it enforces that we have to think in a very structured manner
 561  *  about the "events" that may happen to a request during its life time ...
 562  */
 563 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
 564                 struct bio_and_error *m)
 565 {
 566         struct drbd_device *const device = req->device;
 567         struct drbd_peer_device *const peer_device = first_peer_device(device);
 568         struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
 569         struct net_conf *nc;
 570         int p, rv = 0;
 571
 572         if (m)
 573                 m->bio = NULL;
 574
 575         switch (what) {
 576         default:
 577                 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
 578                 break;
 579
 580         /* does not happen...
 581          * initialization done in drbd_req_new
 582         case CREATED:
 583                 break;
 584                 */
 585
 586         case TO_BE_SENT: /* via network */
 587                 /* reached via __drbd_make_request
 588                  * and from w_read_retry_remote */
 589                 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
 590                 rcu_read_lock();
 591                 nc = rcu_dereference(connection->net_conf);
 592                 p = nc->wire_protocol;
 593                 rcu_read_unlock();
 594                 req->rq_state |=
 595                         p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
 596                         p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
 597                 mod_rq_state(req, m, 0, RQ_NET_PENDING);
 598                 break;
 599
 600         case TO_BE_SUBMITTED: /* locally */
 601                 /* reached via __drbd_make_request */
 602                 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
 603                 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
 604                 break;
 605
 606         case COMPLETED_OK:
 607                 if (req->rq_state & RQ_WRITE)
 608                         device->writ_cnt += req->i.size >> 9;
 609                 else
 610                         device->read_cnt += req->i.size >> 9;
 611
 612                 mod_rq_state(req, m, RQ_LOCAL_PENDING,
 613                                 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
 614                 break;
 615
 616         case ABORT_DISK_IO:
 617                 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
 618                 break;
 619
 620         case WRITE_COMPLETED_WITH_ERROR:
 621                 drbd_report_io_error(device, req);
 622                 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
 623                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 624                 break;
 625
 626         case READ_COMPLETED_WITH_ERROR:
 627                 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
 628                 drbd_report_io_error(device, req);
 629                 __drbd_chk_io_error(device, DRBD_READ_ERROR);
 630                 fallthrough;
 631         case READ_AHEAD_COMPLETED_WITH_ERROR:
 632                 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
 633                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 634                 break;
 635
 636         case DISCARD_COMPLETED_NOTSUPP:
 637         case DISCARD_COMPLETED_WITH_ERROR:
 638                 /* I'd rather not detach from local disk just because it
 639                  * failed a REQ_OP_DISCARD. */
 640                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 641                 break;
 642
 643         case QUEUE_FOR_NET_READ:
 644                 /* READ, and
 645                  * no local disk,
 646                  * or target area marked as invalid,
 647                  * or just got an io-error. */
 648                 /* from __drbd_make_request
 649                  * or from bio_endio during read io-error recovery */
 650
 651                 /* So we can verify the handle in the answer packet.
 652                  * Corresponding drbd_remove_request_interval is in
 653                  * drbd_req_complete() */
 654                 D_ASSERT(device, drbd_interval_empty(&req->i));
 655                 drbd_insert_interval(&device->read_requests, &req->i);
 656
 657                 set_bit(UNPLUG_REMOTE, &device->flags);
 658
 659                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 660                 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
 661                 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 662                 req->w.cb = w_send_read_req;
 663                 drbd_queue_work(&connection->sender_work,
 664                                 &req->w);
 665                 break;
 666
 667         case QUEUE_FOR_NET_WRITE:
 668                 /* assert something? */
 669                 /* from __drbd_make_request only */
 670
 671                 /* Corresponding drbd_remove_request_interval is in
 672                  * drbd_req_complete() */
 673                 D_ASSERT(device, drbd_interval_empty(&req->i));
 674                 drbd_insert_interval(&device->write_requests, &req->i);
 675
 676                 /* NOTE
 677                  * In case the req ended up on the transfer log before being
 678                  * queued on the worker, it could lead to this request being
 679                  * missed during cleanup after connection loss.
 680                  * So we have to do both operations here,
 681                  * within the same lock that protects the transfer log.
 682                  *
 683                  * _req_add_to_epoch(req); this has to be after the
 684                  * _maybe_start_new_epoch(req); which happened in
 685                  * __drbd_make_request, because we now may set the bit
 686                  * again ourselves to close the current epoch.
 687                  *
 688                  * Add req to the (now) current epoch (barrier). */
 689
 690                 /* otherwise we may lose an unplug, which may cause some remote
 691                  * io-scheduler timeout to expire, increasing maximum latency,
 692                  * hurting performance. */
 693                 set_bit(UNPLUG_REMOTE, &device->flags);
 694
 695                 /* queue work item to send data */
 696                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 697                 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
 698                 req->w.cb =  w_send_dblock;
 699                 drbd_queue_work(&connection->sender_work,
 700                                 &req->w);
 701
 702                 /* close the epoch, in case it outgrew the limit */
 703                 rcu_read_lock();
 704                 nc = rcu_dereference(connection->net_conf);
 705                 p = nc->max_epoch_size;
 706                 rcu_read_unlock();
 707                 if (connection->current_tle_writes >= p)
 708                         start_new_tl_epoch(connection);
 709
 710                 break;
 711
 712         case QUEUE_FOR_SEND_OOS:
 713                 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 714                 req->w.cb =  w_send_out_of_sync;
 715                 drbd_queue_work(&connection->sender_work,
 716                                 &req->w);
 717                 break;
 718
 719         case READ_RETRY_REMOTE_CANCELED:
 720         case SEND_CANCELED:
 721         case SEND_FAILED:
 722                 /* real cleanup will be done from tl_clear.  just update flags
 723                  * so it is no longer marked as on the worker queue */
 724                 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
 725                 break;
 726
 727         case HANDED_OVER_TO_NETWORK:
 728                 /* assert something? */
 729                 if (is_pending_write_protocol_A(req))
 730                         /* this is what is dangerous about protocol A:
 731                          * pretend it was successfully written on the peer. */
 732                         mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
 733                                                 RQ_NET_SENT|RQ_NET_OK);
 734                 else
 735                         mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
 736                 /* It is still not yet RQ_NET_DONE until the
 737                  * corresponding epoch barrier got acked as well,
 738                  * so we know what to dirty on connection loss. */
 739                 break;
 740
 741         case OOS_HANDED_TO_NETWORK:
 742                 /* Was not set PENDING, no longer QUEUED, so is now DONE
 743                  * as far as this connection is concerned. */
 744                 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
 745                 break;
 746
 747         case CONNECTION_LOST_WHILE_PENDING:
 748                 /* transfer log cleanup after connection loss */
 749                 mod_rq_state(req, m,
 750                                 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
 751                                 RQ_NET_DONE);
 752                 break;
 753
 754         case CONFLICT_RESOLVED:
 755                 /* for superseded conflicting writes of multiple primaries,
 756                  * there is no need to keep anything in the tl, potential
 757                  * node crashes are covered by the activity log.
 758                  *
 759                  * If this request had been marked as RQ_POSTPONED before,
 760                  * it will actually not be completed, but "restarted",
 761                  * resubmitted from the retry worker context. */
 762                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 763                 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 764                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
 765                 break;
 766
 767         case WRITE_ACKED_BY_PEER_AND_SIS:
 768                 req->rq_state |= RQ_NET_SIS;
 769                 fallthrough;
 770         case WRITE_ACKED_BY_PEER:
 771                 /* Normal operation protocol C: successfully written on peer.
 772                  * During resync, even in protocol != C,
 773                  * we requested an explicit write ack anyways.
 774                  * Which means we cannot even assert anything here.
 775                  * Nothing more to do here.
 776                  * We want to keep the tl in place for all protocols, to cater
 777                  * for volatile write-back caches on lower level devices. */
 778                 goto ack_common;
 779         case RECV_ACKED_BY_PEER:
 780                 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
 781                 /* protocol B; pretends to be successfully written on peer.
 782                  * see also notes above in HANDED_OVER_TO_NETWORK about
 783                  * protocol != C */
 784         ack_common:
 785                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
 786                 break;
 787
 788         case POSTPONE_WRITE:
 789                 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 790                 /* If this node has already detected the write conflict, the
 791                  * worker will be waiting on misc_wait.  Wake it up once this
 792                  * request has completed locally.
 793                  */
 794                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 795                 req->rq_state |= RQ_POSTPONED;
 796                 if (req->i.waiting)
 797                         wake_up(&device->misc_wait);
 798                 /* Do not clear RQ_NET_PENDING. This request will make further
 799                  * progress via restart_conflicting_writes() or
 800                  * fail_postponed_requests(). Hopefully. */
 801                 break;
 802
 803         case NEG_ACKED:
 804                 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
 805                 break;
 806
 807         case FAIL_FROZEN_DISK_IO:
 808                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 809                         break;
 810                 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 811                 break;
 812
 813         case RESTART_FROZEN_DISK_IO:
 814                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 815                         break;
 816
 817                 mod_rq_state(req, m,
 818                                 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
 819                                 RQ_LOCAL_PENDING);
 820
 821                 rv = MR_READ;
 822                 if (bio_data_dir(req->master_bio) == WRITE)
 823                         rv = MR_WRITE;
 824
 825                 get_ldev(device); /* always succeeds in this call path */
 826                 req->w.cb = w_restart_disk_io;
 827                 drbd_queue_work(&connection->sender_work,
 828                                 &req->w);
 829                 break;
 830
 831         case RESEND:
 832                 /* Simply complete (local only) READs. */
 833                 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
 834                         mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 835                         break;
 836                 }
 837
 838                 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 839                    before the connection loss (B&C only); only P_BARRIER_ACK
 840                    (or the local completion?) was missing when we suspended.
 841                    Throwing them out of the TL here by pretending we got a BARRIER_ACK.
 842                    During connection handshake, we ensure that the peer was not rebooted. */
 843                 if (!(req->rq_state & RQ_NET_OK)) {
 844                         /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
 845                          * in that case we must not set RQ_NET_PENDING. */
 846
 847                         mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
 848                         if (req->w.cb) {
 849                                 /* w.cb expected to be w_send_dblock, or w_send_read_req */
 850                                 drbd_queue_work(&connection->sender_work,
 851                                                 &req->w);
 852                                 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 853                         } /* else: FIXME can this happen? */
 854                         break;
 855                 }
 856                 fallthrough;    /* to BARRIER_ACKED */
 857
 858         case BARRIER_ACKED:
 859                 /* barrier ack for READ requests does not make sense */
 860                 if (!(req->rq_state & RQ_WRITE))
 861                         break;
 862
 863                 if (req->rq_state & RQ_NET_PENDING) {
 864                         /* barrier came in before all requests were acked.
 865                          * this is bad, because if the connection is lost now,
 866                          * we won't be able to clean them up... */
 867                         drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
 868                 }
 869                 /* Allowed to complete requests, even while suspended.
 870                  * As this is called for all requests within a matching epoch,
 871                  * we need to filter, and only set RQ_NET_DONE for those that
 872                  * have actually been on the wire. */
 873                 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
 874                                 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
 875                 break;
 876
 877         case DATA_RECEIVED:
 878                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 879                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
 880                 break;
 881
 882         case QUEUE_AS_DRBD_BARRIER:
 883                 start_new_tl_epoch(connection);
 884                 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
 885                 break;
 886         }
 887
 888         return rv;
 889 }
 890
 891 /* we may do a local read if:
 892  * - we are consistent (of course),
 893  * - or we are generally inconsistent,
 894  *   BUT we are still/already IN SYNC for this area.
 895  *   since size may be bigger than BM_BLOCK_SIZE,
 896  *   we may need to check several bits.
 897  */
 898 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
 899 {
 900         unsigned long sbnr, ebnr;
 901         sector_t esector, nr_sectors;
 902
 903         if (device->state.disk == D_UP_TO_DATE)
 904                 return true;
 905         if (device->state.disk != D_INCONSISTENT)
 906                 return false;
 907         esector = sector + (size >> 9) - 1;
 908         nr_sectors = get_capacity(device->vdisk);
 909         D_ASSERT(device, sector  < nr_sectors);
 910         D_ASSERT(device, esector < nr_sectors);
 911
 912         sbnr = BM_SECT_TO_BIT(sector);
 913         ebnr = BM_SECT_TO_BIT(esector);
 914
 915         return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
 916 }
 917
 918 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
 919                 enum drbd_read_balancing rbm)
 920 {
 921         int stripe_shift;
 922
 923         switch (rbm) {
 924         case RB_CONGESTED_REMOTE:
 925                 return 0;
 926         case RB_LEAST_PENDING:
 927                 return atomic_read(&device->local_cnt) >
 928                         atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
 929         case RB_32K_STRIPING:  /* stripe_shift = 15 */
 930         case RB_64K_STRIPING:
 931         case RB_128K_STRIPING:
 932         case RB_256K_STRIPING:
 933         case RB_512K_STRIPING:
 934         case RB_1M_STRIPING:   /* stripe_shift = 20 */
 935                 stripe_shift = (rbm - RB_32K_STRIPING + 15);
 936                 return (sector >> (stripe_shift - 9)) & 1;
 937         case RB_ROUND_ROBIN:
 938                 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
 939         case RB_PREFER_REMOTE:
 940                 return true;
 941         case RB_PREFER_LOCAL:
 942         default:
 943                 return false;
 944         }
 945 }
 946
 947 /*
 948  * complete_conflicting_writes  -  wait for any conflicting write requests
 949  *
 950  * The write_requests tree contains all active write requests which we
 951  * currently know about.  Wait for any requests to complete which conflict with
 952  * the new one.
 953  *
 954  * Only way out: remove the conflicting intervals from the tree.
 955  */
 956 static void complete_conflicting_writes(struct drbd_request *req)
 957 {
 958         DEFINE_WAIT(wait);
 959         struct drbd_device *device = req->device;
 960         struct drbd_interval *i;
 961         sector_t sector = req->i.sector;
 962         int size = req->i.size;
 963
 964         for (;;) {
 965                 drbd_for_each_overlap(i, &device->write_requests, sector, size) {
 966                         /* Ignore, if already completed to upper layers. */
 967                         if (i->completed)
 968                                 continue;
 969                         /* Handle the first found overlap.  After the schedule
 970                          * we have to restart the tree walk. */
 971                         break;
 972                 }
 973                 if (!i) /* if any */
 974                         break;
 975
 976                 /* Indicate to wake up device->misc_wait on progress.  */
 977                 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
 978                 i->waiting = true;
 979                 spin_unlock_irq(&device->resource->req_lock);
 980                 schedule();
 981                 spin_lock_irq(&device->resource->req_lock);
 982         }
 983         finish_wait(&device->misc_wait, &wait);
 984 }
 985
 986 /* called within req_lock */
 987 static void maybe_pull_ahead(struct drbd_device *device)
 988 {
 989         struct drbd_connection *connection = first_peer_device(device)->connection;
 990         struct net_conf *nc;
 991         bool congested = false;
 992         enum drbd_on_congestion on_congestion;
 993
 994         rcu_read_lock();
 995         nc = rcu_dereference(connection->net_conf);
 996         on_congestion = nc ? nc->on_congestion : OC_BLOCK;
 997         rcu_read_unlock();
 998         if (on_congestion == OC_BLOCK ||
 999             connection->agreed_pro_version < 96)
1000                 return;
1001
1002         if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1003                 return; /* nothing to do ... */
1004
1005         /* If I don't even have good local storage, we can not reasonably try
1006          * to pull ahead of the peer. We also need the local reference to make
1007          * sure device->act_log is there.
1008          */
1009         if (!get_ldev_if_state(device, D_UP_TO_DATE))
1010                 return;
1011
1012         if (nc->cong_fill &&
1013             atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1014                 drbd_info(device, "Congestion-fill threshold reached\n");
1015                 congested = true;
1016         }
1017
1018         if (device->act_log->used >= nc->cong_extents) {
1019                 drbd_info(device, "Congestion-extents threshold reached\n");
1020                 congested = true;
1021         }
1022
1023         if (congested) {
1024                 /* start a new epoch for non-mirrored writes */
1025                 start_new_tl_epoch(first_peer_device(device)->connection);
1026
1027                 if (on_congestion == OC_PULL_AHEAD)
1028                         _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1029                 else  /*nc->on_congestion == OC_DISCONNECT */
1030                         _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1031         }
1032         put_ldev(device);
1033 }
1034
1035 /* If this returns false, and req->private_bio is still set,
1036  * this should be submitted locally.
1037  *
1038  * If it returns false, but req->private_bio is not set,
1039  * we do not have access to good data :(
1040  *
1041  * Otherwise, this destroys req->private_bio, if any,
1042  * and returns true.
1043  */
1044 static bool do_remote_read(struct drbd_request *req)
1045 {
1046         struct drbd_device *device = req->device;
1047         enum drbd_read_balancing rbm;
1048
1049         if (req->private_bio) {
1050                 if (!drbd_may_do_local_read(device,
1051                                         req->i.sector, req->i.size)) {
1052                         bio_put(req->private_bio);
1053                         req->private_bio = NULL;
1054                         put_ldev(device);
1055                 }
1056         }
1057
1058         if (device->state.pdsk != D_UP_TO_DATE)
1059                 return false;
1060
1061         if (req->private_bio == NULL)
1062                 return true;
1063
1064         /* TODO: improve read balancing decisions, take into account drbd
1065          * protocol, pending requests etc. */
1066
1067         rcu_read_lock();
1068         rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1069         rcu_read_unlock();
1070
1071         if (rbm == RB_PREFER_LOCAL && req->private_bio)
1072                 return false; /* submit locally */
1073
1074         if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1075                 if (req->private_bio) {
1076                         bio_put(req->private_bio);
1077                         req->private_bio = NULL;
1078                         put_ldev(device);
1079                 }
1080                 return true;
1081         }
1082
1083         return false;
1084 }
1085
1086 bool drbd_should_do_remote(union drbd_dev_state s)
1087 {
1088         return s.pdsk == D_UP_TO_DATE ||
1089                 (s.pdsk >= D_INCONSISTENT &&
1090                  s.conn >= C_WF_BITMAP_T &&
1091                  s.conn < C_AHEAD);
1092         /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1093            That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1094            states. */
1095 }
1096
1097 static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1098 {
1099         return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1100         /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1101            since we enter state C_AHEAD only if proto >= 96 */
1102 }
1103
1104 /* returns number of connections (== 1, for drbd 8.4)
1105  * expected to actually write this data,
1106  * which does NOT include those that we are L_AHEAD for. */
1107 static int drbd_process_write_request(struct drbd_request *req)
1108 {
1109         struct drbd_device *device = req->device;
1110         int remote, send_oos;
1111
1112         remote = drbd_should_do_remote(device->state);
1113         send_oos = drbd_should_send_out_of_sync(device->state);
1114
1115         /* Need to replicate writes.  Unless it is an empty flush,
1116          * which is better mapped to a DRBD P_BARRIER packet,
1117          * also for drbd wire protocol compatibility reasons.
1118          * If this was a flush, just start a new epoch.
1119          * Unless the current epoch was empty anyways, or we are not currently
1120          * replicating, in which case there is no point. */
1121         if (unlikely(req->i.size == 0)) {
1122                 /* The only size==0 bios we expect are empty flushes. */
1123                 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1124                 if (remote)
1125                         _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1126                 return remote;
1127         }
1128
1129         if (!remote && !send_oos)
1130                 return 0;
1131
1132         D_ASSERT(device, !(remote && send_oos));
1133
1134         if (remote) {
1135                 _req_mod(req, TO_BE_SENT);
1136                 _req_mod(req, QUEUE_FOR_NET_WRITE);
1137         } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1138                 _req_mod(req, QUEUE_FOR_SEND_OOS);
1139
1140         return remote;
1141 }
1142
1143 static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1144 {
1145         int err = drbd_issue_discard_or_zero_out(req->device,
1146                                 req->i.sector, req->i.size >> 9, flags);
1147         if (err)
1148                 req->private_bio->bi_status = BLK_STS_IOERR;
1149         bio_endio(req->private_bio);
1150 }
1151
1152 static void
1153 drbd_submit_req_private_bio(struct drbd_request *req)
1154 {
1155         struct drbd_device *device = req->device;
1156         struct bio *bio = req->private_bio;
1157         unsigned int type;
1158
1159         if (bio_op(bio) != REQ_OP_READ)
1160                 type = DRBD_FAULT_DT_WR;
1161         else if (bio->bi_opf & REQ_RAHEAD)
1162                 type = DRBD_FAULT_DT_RA;
1163         else
1164                 type = DRBD_FAULT_DT_RD;
1165
1166         /* State may have changed since we grabbed our reference on the
1167          * ->ldev member. Double check, and short-circuit to endio.
1168          * In case the last activity log transaction failed to get on
1169          * stable storage, and this is a WRITE, we may not even submit
1170          * this bio. */
1171         if (get_ldev(device)) {
1172                 if (drbd_insert_fault(device, type))
1173                         bio_io_error(bio);
1174                 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1175                         drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1176                             ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1177                 else if (bio_op(bio) == REQ_OP_DISCARD)
1178                         drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1179                 else
1180                         submit_bio_noacct(bio);
1181                 put_ldev(device);
1182         } else
1183                 bio_io_error(bio);
1184 }
1185
1186 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1187 {
1188         spin_lock_irq(&device->resource->req_lock);
1189         list_add_tail(&req->tl_requests, &device->submit.writes);
1190         list_add_tail(&req->req_pending_master_completion,
1191                         &device->pending_master_completion[1 /* WRITE */]);
1192         spin_unlock_irq(&device->resource->req_lock);
1193         queue_work(device->submit.wq, &device->submit.worker);
1194         /* do_submit() may sleep internally on al_wait, too */
1195         wake_up(&device->al_wait);
1196 }
1197
1198 /* returns the new drbd_request pointer, if the caller is expected to
1199  * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1200  * request on the submitter thread.
1201  * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1202  */
1203 static struct drbd_request *
1204 drbd_request_prepare(struct drbd_device *device, struct bio *bio)
1205 {
1206         const int rw = bio_data_dir(bio);
1207         struct drbd_request *req;
1208
1209         /* allocate outside of all locks; */
1210         req = drbd_req_new(device, bio);
1211         if (!req) {
1212                 dec_ap_bio(device);
1213                 /* only pass the error to the upper layers.
1214                  * if user cannot handle io errors, that's not our business. */
1215                 drbd_err(device, "could not kmalloc() req\n");
1216                 bio->bi_status = BLK_STS_RESOURCE;
1217                 bio_endio(bio);
1218                 return ERR_PTR(-ENOMEM);
1219         }
1220
1221         /* Update disk stats */
1222         req->start_jif = bio_start_io_acct(req->master_bio);
1223
1224         if (!get_ldev(device)) {
1225                 bio_put(req->private_bio);
1226                 req->private_bio = NULL;
1227         }
1228
1229         /* process discards always from our submitter thread */
1230         if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1231             bio_op(bio) == REQ_OP_DISCARD)
1232                 goto queue_for_submitter_thread;
1233
1234         if (rw == WRITE && req->private_bio && req->i.size
1235         && !test_bit(AL_SUSPENDED, &device->flags)) {
1236                 if (!drbd_al_begin_io_fastpath(device, &req->i))
1237                         goto queue_for_submitter_thread;
1238                 req->rq_state |= RQ_IN_ACT_LOG;
1239                 req->in_actlog_jif = jiffies;
1240         }
1241         return req;
1242
1243  queue_for_submitter_thread:
1244         atomic_inc(&device->ap_actlog_cnt);
1245         drbd_queue_write(device, req);
1246         return NULL;
1247 }
1248
1249 /* Require at least one path to current data.
1250  * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1251  * We would not allow to read what was written,
1252  * we would not have bumped the data generation uuids,
1253  * we would cause data divergence for all the wrong reasons.
1254  *
1255  * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1256  * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1257  * and queues for retry later.
1258  */
1259 static bool may_do_writes(struct drbd_device *device)
1260 {
1261         const union drbd_dev_state s = device->state;
1262         return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1263 }
1264
1265 struct drbd_plug_cb {
1266         struct blk_plug_cb cb;
1267         struct drbd_request *most_recent_req;
1268         /* do we need more? */
1269 };
1270
1271 static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1272 {
1273         struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1274         struct drbd_resource *resource = plug->cb.data;
1275         struct drbd_request *req = plug->most_recent_req;
1276
1277         kfree(cb);
1278         if (!req)
1279                 return;
1280
1281         spin_lock_irq(&resource->req_lock);
1282         /* In case the sender did not process it yet, raise the flag to
1283          * have it followed with P_UNPLUG_REMOTE just after. */
1284         req->rq_state |= RQ_UNPLUG;
1285         /* but also queue a generic unplug */
1286         drbd_queue_unplug(req->device);
1287         kref_put(&req->kref, drbd_req_destroy);
1288         spin_unlock_irq(&resource->req_lock);
1289 }
1290
1291 static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1292 {
1293         /* A lot of text to say
1294          * return (struct drbd_plug_cb*)blk_check_plugged(); */
1295         struct drbd_plug_cb *plug;
1296         struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1297
1298         if (cb)
1299                 plug = container_of(cb, struct drbd_plug_cb, cb);
1300         else
1301                 plug = NULL;
1302         return plug;
1303 }
1304
1305 static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1306 {
1307         struct drbd_request *tmp = plug->most_recent_req;
1308         /* Will be sent to some peer.
1309          * Remember to tag it with UNPLUG_REMOTE on unplug */
1310         kref_get(&req->kref);
1311         plug->most_recent_req = req;
1312         if (tmp)
1313                 kref_put(&tmp->kref, drbd_req_destroy);
1314 }
1315
1316 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1317 {
1318         struct drbd_resource *resource = device->resource;
1319         const int rw = bio_data_dir(req->master_bio);
1320         struct bio_and_error m = { NULL, };
1321         bool no_remote = false;
1322         bool submit_private_bio = false;
1323
1324         spin_lock_irq(&resource->req_lock);
1325         if (rw == WRITE) {
1326                 /* This may temporarily give up the req_lock,
1327                  * but will re-aquire it before it returns here.
1328                  * Needs to be before the check on drbd_suspended() */
1329                 complete_conflicting_writes(req);
1330                 /* no more giving up req_lock from now on! */
1331
1332                 /* check for congestion, and potentially stop sending
1333                  * full data updates, but start sending "dirty bits" only. */
1334                 maybe_pull_ahead(device);
1335         }
1336
1337
1338         if (drbd_suspended(device)) {
1339                 /* push back and retry: */
1340                 req->rq_state |= RQ_POSTPONED;
1341                 if (req->private_bio) {
1342                         bio_put(req->private_bio);
1343                         req->private_bio = NULL;
1344                         put_ldev(device);
1345                 }
1346                 goto out;
1347         }
1348
1349         /* We fail READ early, if we can not serve it.
1350          * We must do this before req is registered on any lists.
1351          * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1352         if (rw != WRITE) {
1353                 if (!do_remote_read(req) && !req->private_bio)
1354                         goto nodata;
1355         }
1356
1357         /* which transfer log epoch does this belong to? */
1358         req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1359
1360         /* no point in adding empty flushes to the transfer log,
1361          * they are mapped to drbd barriers already. */
1362         if (likely(req->i.size!=0)) {
1363                 if (rw == WRITE)
1364                         first_peer_device(device)->connection->current_tle_writes++;
1365
1366                 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1367         }
1368
1369         if (rw == WRITE) {
1370                 if (req->private_bio && !may_do_writes(device)) {
1371                         bio_put(req->private_bio);
1372                         req->private_bio = NULL;
1373                         put_ldev(device);
1374                         goto nodata;
1375                 }
1376                 if (!drbd_process_write_request(req))
1377                         no_remote = true;
1378         } else {
1379                 /* We either have a private_bio, or we can read from remote.
1380                  * Otherwise we had done the goto nodata above. */
1381                 if (req->private_bio == NULL) {
1382                         _req_mod(req, TO_BE_SENT);
1383                         _req_mod(req, QUEUE_FOR_NET_READ);
1384                 } else
1385                         no_remote = true;
1386         }
1387
1388         if (no_remote == false) {
1389                 struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1390                 if (plug)
1391                         drbd_update_plug(plug, req);
1392         }
1393
1394         /* If it took the fast path in drbd_request_prepare, add it here.
1395          * The slow path has added it already. */
1396         if (list_empty(&req->req_pending_master_completion))
1397                 list_add_tail(&req->req_pending_master_completion,
1398                         &device->pending_master_completion[rw == WRITE]);
1399         if (req->private_bio) {
1400                 /* needs to be marked within the same spinlock */
1401                 req->pre_submit_jif = jiffies;
1402                 list_add_tail(&req->req_pending_local,
1403                         &device->pending_completion[rw == WRITE]);
1404                 _req_mod(req, TO_BE_SUBMITTED);
1405                 /* but we need to give up the spinlock to submit */
1406                 submit_private_bio = true;
1407         } else if (no_remote) {
1408 nodata:
1409                 if (__ratelimit(&drbd_ratelimit_state))
1410                         drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1411                                         (unsigned long long)req->i.sector, req->i.size >> 9);
1412                 /* A write may have been queued for send_oos, however.
1413                  * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1414         }
1415
1416 out:
1417         drbd_req_put_completion_ref(req, &m, 1);
1418         spin_unlock_irq(&resource->req_lock);
1419
1420         /* Even though above is a kref_put(), this is safe.
1421          * As long as we still need to submit our private bio,
1422          * we hold a completion ref, and the request cannot disappear.
1423          * If however this request did not even have a private bio to submit
1424          * (e.g. remote read), req may already be invalid now.
1425          * That's why we cannot check on req->private_bio. */
1426         if (submit_private_bio)
1427                 drbd_submit_req_private_bio(req);
1428         if (m.bio)
1429                 complete_master_bio(device, &m);
1430 }
1431
1432 void __drbd_make_request(struct drbd_device *device, struct bio *bio)
1433 {
1434         struct drbd_request *req = drbd_request_prepare(device, bio);
1435         if (IS_ERR_OR_NULL(req))
1436                 return;
1437         drbd_send_and_submit(device, req);
1438 }
1439
1440 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1441 {
1442         struct blk_plug plug;
1443         struct drbd_request *req, *tmp;
1444
1445         blk_start_plug(&plug);
1446         list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1447                 const int rw = bio_data_dir(req->master_bio);
1448
1449                 if (rw == WRITE /* rw != WRITE should not even end up here! */
1450                 && req->private_bio && req->i.size
1451                 && !test_bit(AL_SUSPENDED, &device->flags)) {
1452                         if (!drbd_al_begin_io_fastpath(device, &req->i))
1453                                 continue;
1454
1455                         req->rq_state |= RQ_IN_ACT_LOG;
1456                         req->in_actlog_jif = jiffies;
1457                         atomic_dec(&device->ap_actlog_cnt);
1458                 }
1459
1460                 list_del_init(&req->tl_requests);
1461                 drbd_send_and_submit(device, req);
1462         }
1463         blk_finish_plug(&plug);
1464 }
1465
1466 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1467                                             struct list_head *incoming,
1468                                             struct list_head *pending,
1469                                             struct list_head *later)
1470 {
1471         struct drbd_request *req;
1472         int wake = 0;
1473         int err;
1474
1475         spin_lock_irq(&device->al_lock);
1476         while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1477                 err = drbd_al_begin_io_nonblock(device, &req->i);
1478                 if (err == -ENOBUFS)
1479                         break;
1480                 if (err == -EBUSY)
1481                         wake = 1;
1482                 if (err)
1483                         list_move_tail(&req->tl_requests, later);
1484                 else
1485                         list_move_tail(&req->tl_requests, pending);
1486         }
1487         spin_unlock_irq(&device->al_lock);
1488         if (wake)
1489                 wake_up(&device->al_wait);
1490         return !list_empty(pending);
1491 }
1492
1493 static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1494 {
1495         struct blk_plug plug;
1496         struct drbd_request *req;
1497
1498         blk_start_plug(&plug);
1499         while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1500                 req->rq_state |= RQ_IN_ACT_LOG;
1501                 req->in_actlog_jif = jiffies;
1502                 atomic_dec(&device->ap_actlog_cnt);
1503                 list_del_init(&req->tl_requests);
1504                 drbd_send_and_submit(device, req);
1505         }
1506         blk_finish_plug(&plug);
1507 }
1508
1509 void do_submit(struct work_struct *ws)
1510 {
1511         struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1512         LIST_HEAD(incoming);    /* from drbd_make_request() */
1513         LIST_HEAD(pending);     /* to be submitted after next AL-transaction commit */
1514         LIST_HEAD(busy);        /* blocked by resync requests */
1515
1516         /* grab new incoming requests */
1517         spin_lock_irq(&device->resource->req_lock);
1518         list_splice_tail_init(&device->submit.writes, &incoming);
1519         spin_unlock_irq(&device->resource->req_lock);
1520
1521         for (;;) {
1522                 DEFINE_WAIT(wait);
1523
1524                 /* move used-to-be-busy back to front of incoming */
1525                 list_splice_init(&busy, &incoming);
1526                 submit_fast_path(device, &incoming);
1527                 if (list_empty(&incoming))
1528                         break;
1529
1530                 for (;;) {
1531                         prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1532
1533                         list_splice_init(&busy, &incoming);
1534                         prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1535                         if (!list_empty(&pending))
1536                                 break;
1537
1538                         schedule();
1539
1540                         /* If all currently "hot" activity log extents are kept busy by
1541                          * incoming requests, we still must not totally starve new
1542                          * requests to "cold" extents.
1543                          * Something left on &incoming means there had not been
1544                          * enough update slots available, and the activity log
1545                          * has been marked as "starving".
1546                          *
1547                          * Try again now, without looking for new requests,
1548                          * effectively blocking all new requests until we made
1549                          * at least _some_ progress with what we currently have.
1550                          */
1551                         if (!list_empty(&incoming))
1552                                 continue;
1553
1554                         /* Nothing moved to pending, but nothing left
1555                          * on incoming: all moved to busy!
1556                          * Grab new and iterate. */
1557                         spin_lock_irq(&device->resource->req_lock);
1558                         list_splice_tail_init(&device->submit.writes, &incoming);
1559                         spin_unlock_irq(&device->resource->req_lock);
1560                 }
1561                 finish_wait(&device->al_wait, &wait);
1562
1563                 /* If the transaction was full, before all incoming requests
1564                  * had been processed, skip ahead to commit, and iterate
1565                  * without splicing in more incoming requests from upper layers.
1566                  *
1567                  * Else, if all incoming have been processed,
1568                  * they have become either "pending" (to be submitted after
1569                  * next transaction commit) or "busy" (blocked by resync).
1570                  *
1571                  * Maybe more was queued, while we prepared the transaction?
1572                  * Try to stuff those into this transaction as well.
1573                  * Be strictly non-blocking here,
1574                  * we already have something to commit.
1575                  *
1576                  * Commit if we don't make any more progres.
1577                  */
1578
1579                 while (list_empty(&incoming)) {
1580                         LIST_HEAD(more_pending);
1581                         LIST_HEAD(more_incoming);
1582                         bool made_progress;
1583
1584                         /* It is ok to look outside the lock,
1585                          * it's only an optimization anyways */
1586                         if (list_empty(&device->submit.writes))
1587                                 break;
1588
1589                         spin_lock_irq(&device->resource->req_lock);
1590                         list_splice_tail_init(&device->submit.writes, &more_incoming);
1591                         spin_unlock_irq(&device->resource->req_lock);
1592
1593                         if (list_empty(&more_incoming))
1594                                 break;
1595
1596                         made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1597
1598                         list_splice_tail_init(&more_pending, &pending);
1599                         list_splice_tail_init(&more_incoming, &incoming);
1600                         if (!made_progress)
1601                                 break;
1602                 }
1603
1604                 drbd_al_begin_io_commit(device);
1605                 send_and_submit_pending(device, &pending);
1606         }
1607 }
1608
1609 void drbd_submit_bio(struct bio *bio)
1610 {
1611         struct drbd_device *device = bio->bi_bdev->bd_disk->private_data;
1612
1613         blk_queue_split(&bio);
1614
1615         /*
1616          * what we "blindly" assume:
1617          */
1618         D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1619
1620         inc_ap_bio(device);
1621         __drbd_make_request(device, bio);
1622 }
1623
1624 static bool net_timeout_reached(struct drbd_request *net_req,
1625                 struct drbd_connection *connection,
1626                 unsigned long now, unsigned long ent,
1627                 unsigned int ko_count, unsigned int timeout)
1628 {
1629         struct drbd_device *device = net_req->device;
1630
1631         if (!time_after(now, net_req->pre_send_jif + ent))
1632                 return false;
1633
1634         if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1635                 return false;
1636
1637         if (net_req->rq_state & RQ_NET_PENDING) {
1638                 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1639                         jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1640                 return true;
1641         }
1642
1643         /* We received an ACK already (or are using protocol A),
1644          * but are waiting for the epoch closing barrier ack.
1645          * Check if we sent the barrier already.  We should not blame the peer
1646          * for being unresponsive, if we did not even ask it yet. */
1647         if (net_req->epoch == connection->send.current_epoch_nr) {
1648                 drbd_warn(device,
1649                         "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1650                         jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1651                 return false;
1652         }
1653
1654         /* Worst case: we may have been blocked for whatever reason, then
1655          * suddenly are able to send a lot of requests (and epoch separating
1656          * barriers) in quick succession.
1657          * The timestamp of the net_req may be much too old and not correspond
1658          * to the sending time of the relevant unack'ed barrier packet, so
1659          * would trigger a spurious timeout.  The latest barrier packet may
1660          * have a too recent timestamp to trigger the timeout, potentially miss
1661          * a timeout.  Right now we don't have a place to conveniently store
1662          * these timestamps.
1663          * But in this particular situation, the application requests are still
1664          * completed to upper layers, DRBD should still "feel" responsive.
1665          * No need yet to kill this connection, it may still recover.
1666          * If not, eventually we will have queued enough into the network for
1667          * us to block. From that point of view, the timestamp of the last sent
1668          * barrier packet is relevant enough.
1669          */
1670         if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1671                 drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1672                         connection->send.last_sent_barrier_jif, now,
1673                         jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1674                 return true;
1675         }
1676         return false;
1677 }
1678
1679 /* A request is considered timed out, if
1680  * - we have some effective timeout from the configuration,
1681  *   with some state restrictions applied,
1682  * - the oldest request is waiting for a response from the network
1683  *   resp. the local disk,
1684  * - the oldest request is in fact older than the effective timeout,
1685  * - the connection was established (resp. disk was attached)
1686  *   for longer than the timeout already.
1687  * Note that for 32bit jiffies and very stable connections/disks,
1688  * we may have a wrap around, which is catched by
1689  *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1690  *
1691  * Side effect: once per 32bit wrap-around interval, which means every
1692  * ~198 days with 250 HZ, we have a window where the timeout would need
1693  * to expire twice (worst case) to become effective. Good enough.
1694  */
1695
1696 void request_timer_fn(struct timer_list *t)
1697 {
1698         struct drbd_device *device = from_timer(device, t, request_timer);
1699         struct drbd_connection *connection = first_peer_device(device)->connection;
1700         struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1701         struct net_conf *nc;
1702         unsigned long oldest_submit_jif;
1703         unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1704         unsigned long now;
1705         unsigned int ko_count = 0, timeout = 0;
1706
1707         rcu_read_lock();
1708         nc = rcu_dereference(connection->net_conf);
1709         if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1710                 ko_count = nc->ko_count;
1711                 timeout = nc->timeout;
1712         }
1713
1714         if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1715                 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1716                 put_ldev(device);
1717         }
1718         rcu_read_unlock();
1719
1720
1721         ent = timeout * HZ/10 * ko_count;
1722         et = min_not_zero(dt, ent);
1723
1724         if (!et)
1725                 return; /* Recurring timer stopped */
1726
1727         now = jiffies;
1728         nt = now + et;
1729
1730         spin_lock_irq(&device->resource->req_lock);
1731         req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1732         req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1733
1734         /* maybe the oldest request waiting for the peer is in fact still
1735          * blocking in tcp sendmsg.  That's ok, though, that's handled via the
1736          * socket send timeout, requesting a ping, and bumping ko-count in
1737          * we_should_drop_the_connection().
1738          */
1739
1740         /* check the oldest request we did successfully sent,
1741          * but which is still waiting for an ACK. */
1742         req_peer = connection->req_ack_pending;
1743
1744         /* if we don't have such request (e.g. protocoll A)
1745          * check the oldest requests which is still waiting on its epoch
1746          * closing barrier ack. */
1747         if (!req_peer)
1748                 req_peer = connection->req_not_net_done;
1749
1750         /* evaluate the oldest peer request only in one timer! */
1751         if (req_peer && req_peer->device != device)
1752                 req_peer = NULL;
1753
1754         /* do we have something to evaluate? */
1755         if (req_peer == NULL && req_write == NULL && req_read == NULL)
1756                 goto out;
1757
1758         oldest_submit_jif =
1759                 (req_write && req_read)
1760                 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1761                   ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1762                 : req_write ? req_write->pre_submit_jif
1763                 : req_read ? req_read->pre_submit_jif : now;
1764
1765         if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1766                 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1767
1768         if (dt && oldest_submit_jif != now &&
1769                  time_after(now, oldest_submit_jif + dt) &&
1770                 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1771                 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1772                 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1773         }
1774
1775         /* Reschedule timer for the nearest not already expired timeout.
1776          * Fallback to now + min(effective network timeout, disk timeout). */
1777         ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1778                 ? req_peer->pre_send_jif + ent : now + et;
1779         dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1780                 ? oldest_submit_jif + dt : now + et;
1781         nt = time_before(ent, dt) ? ent : dt;
1782 out:
1783         spin_unlock_irq(&device->resource->req_lock);
1784         mod_timer(&device->request_timer, nt);
1785 }