root/drivers/block/drbd/drbd_req.c

DEFINITIONS

1. _drbd_start_io_acct
2. _drbd_end_io_acct
3. drbd_req_new
4. drbd_remove_request_interval
5. drbd_req_destroy
6. wake_all_senders
7. start_new_tl_epoch
8. complete_master_bio
9. drbd_req_complete
10. drbd_req_put_completion_ref
11. set_if_null_req_next
12. advance_conn_req_next
13. set_if_null_req_ack_pending
14. advance_conn_req_ack_pending
15. set_if_null_req_not_net_done
16. advance_conn_req_not_net_done
17. mod_rq_state
18. drbd_report_io_error
19. is_pending_write_protocol_A
20. __req_mod
21. drbd_may_do_local_read
22. remote_due_to_read_balancing
23. complete_conflicting_writes
24. maybe_pull_ahead
25. do_remote_read
26. drbd_should_do_remote
27. drbd_should_send_out_of_sync
28. drbd_process_write_request
29. drbd_process_discard_or_zeroes_req
30. drbd_submit_req_private_bio
31. drbd_queue_write
32. drbd_request_prepare
33. may_do_writes
34. drbd_unplug
35. drbd_check_plugged
36. drbd_update_plug
37. drbd_send_and_submit
38. __drbd_make_request
39. submit_fast_path
40. prepare_al_transaction_nonblock
41. send_and_submit_pending
42. do_submit
43. drbd_make_request
44. net_timeout_reached
45. request_timer_fn

   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 /* Update disk stats at start of I/O request */
  25 static void _drbd_start_io_acct(struct drbd_device *device, struct drbd_request *req)
  26 {
  27         struct request_queue *q = device->rq_queue;
  28 
  29         generic_start_io_acct(q, bio_op(req->master_bio),
  30                                 req->i.size >> 9, &device->vdisk->part0);
  31 }
  32 
  33 /* Update disk stats when completing request upwards */
  34 static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
  35 {
  36         struct request_queue *q = device->rq_queue;
  37 
  38         generic_end_io_acct(q, bio_op(req->master_bio),
  39                             &device->vdisk->part0, req->start_jif);
  40 }
  41 
  42 static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src)
  43 {
  44         struct drbd_request *req;
  45 
  46         req = mempool_alloc(&drbd_request_mempool, GFP_NOIO);
  47         if (!req)
  48                 return NULL;
  49         memset(req, 0, sizeof(*req));
  50 
  51         drbd_req_make_private_bio(req, bio_src);
  52         req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
  53                       | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0)
  54                       | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0)
  55                       | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
  56         req->device = device;
  57         req->master_bio = bio_src;
  58         req->epoch = 0;
  59 
  60         drbd_clear_interval(&req->i);
  61         req->i.sector     = bio_src->bi_iter.bi_sector;
  62         req->i.size      = bio_src->bi_iter.bi_size;
  63         req->i.local = true;
  64         req->i.waiting = false;
  65 
  66         INIT_LIST_HEAD(&req->tl_requests);
  67         INIT_LIST_HEAD(&req->w.list);
  68         INIT_LIST_HEAD(&req->req_pending_master_completion);
  69         INIT_LIST_HEAD(&req->req_pending_local);
  70 
  71         /* one reference to be put by __drbd_make_request */
  72         atomic_set(&req->completion_ref, 1);
  73         /* one kref as long as completion_ref > 0 */
  74         kref_init(&req->kref);
  75         return req;
  76 }
  77 
  78 static void drbd_remove_request_interval(struct rb_root *root,
  79                                          struct drbd_request *req)
  80 {
  81         struct drbd_device *device = req->device;
  82         struct drbd_interval *i = &req->i;
  83 
  84         drbd_remove_interval(root, i);
  85 
  86         /* Wake up any processes waiting for this request to complete.  */
  87         if (i->waiting)
  88                 wake_up(&device->misc_wait);
  89 }
  90 
  91 void drbd_req_destroy(struct kref *kref)
  92 {
  93         struct drbd_request *req = container_of(kref, struct drbd_request, kref);
  94         struct drbd_device *device = req->device;
  95         const unsigned s = req->rq_state;
  96 
  97         if ((req->master_bio && !(s & RQ_POSTPONED)) ||
  98                 atomic_read(&req->completion_ref) ||
  99                 (s & RQ_LOCAL_PENDING) ||
 100                 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
 101                 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
 102                                 s, atomic_read(&req->completion_ref));
 103                 return;
 104         }
 105 
 106         /* If called from mod_rq_state (expected normal case) or
 107          * drbd_send_and_submit (the less likely normal path), this holds the
 108          * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
 109          * though it may be still empty (never added to the transfer log).
 110          *
 111          * If called from do_retry(), we do NOT hold the req_lock, but we are
 112          * still allowed to unconditionally list_del(&req->tl_requests),
 113          * because it will be on a local on-stack list only. */
 114         list_del_init(&req->tl_requests);
 115 
 116         /* finally remove the request from the conflict detection
 117          * respective block_id verification interval tree. */
 118         if (!drbd_interval_empty(&req->i)) {
 119                 struct rb_root *root;
 120 
 121                 if (s & RQ_WRITE)
 122                         root = &device->write_requests;
 123                 else
 124                         root = &device->read_requests;
 125                 drbd_remove_request_interval(root, req);
 126         } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
 127                 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
 128                         s, (unsigned long long)req->i.sector, req->i.size);
 129 
 130         /* if it was a write, we may have to set the corresponding
 131          * bit(s) out-of-sync first. If it had a local part, we need to
 132          * release the reference to the activity log. */
 133         if (s & RQ_WRITE) {
 134                 /* Set out-of-sync unless both OK flags are set
 135                  * (local only or remote failed).
 136                  * Other places where we set out-of-sync:
 137                  * READ with local io-error */
 138 
 139                 /* There is a special case:
 140                  * we may notice late that IO was suspended,
 141                  * and postpone, or schedule for retry, a write,
 142                  * before it even was submitted or sent.
 143                  * In that case we do not want to touch the bitmap at all.
 144                  */
 145                 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
 146                         if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
 147                                 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
 148 
 149                         if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
 150                                 drbd_set_in_sync(device, req->i.sector, req->i.size);
 151                 }
 152 
 153                 /* one might be tempted to move the drbd_al_complete_io
 154                  * to the local io completion callback drbd_request_endio.
 155                  * but, if this was a mirror write, we may only
 156                  * drbd_al_complete_io after this is RQ_NET_DONE,
 157                  * otherwise the extent could be dropped from the al
 158                  * before it has actually been written on the peer.
 159                  * if we crash before our peer knows about the request,
 160                  * but after the extent has been dropped from the al,
 161                  * we would forget to resync the corresponding extent.
 162                  */
 163                 if (s & RQ_IN_ACT_LOG) {
 164                         if (get_ldev_if_state(device, D_FAILED)) {
 165                                 drbd_al_complete_io(device, &req->i);
 166                                 put_ldev(device);
 167                         } else if (__ratelimit(&drbd_ratelimit_state)) {
 168                                 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
 169                                          "but my Disk seems to have failed :(\n",
 170                                          (unsigned long long) req->i.sector, req->i.size);
 171                         }
 172                 }
 173         }
 174 
 175         mempool_free(req, &drbd_request_mempool);
 176 }
 177 
 178 static void wake_all_senders(struct drbd_connection *connection)
 179 {
 180         wake_up(&connection->sender_work.q_wait);
 181 }
 182 
 183 /* must hold resource->req_lock */
 184 void start_new_tl_epoch(struct drbd_connection *connection)
 185 {
 186         /* no point closing an epoch, if it is empty, anyways. */
 187         if (connection->current_tle_writes == 0)
 188                 return;
 189 
 190         connection->current_tle_writes = 0;
 191         atomic_inc(&connection->current_tle_nr);
 192         wake_all_senders(connection);
 193 }
 194 
 195 void complete_master_bio(struct drbd_device *device,
 196                 struct bio_and_error *m)
 197 {
 198         m->bio->bi_status = errno_to_blk_status(m->error);
 199         bio_endio(m->bio);
 200         dec_ap_bio(device);
 201 }
 202 
 203 
 204 /* Helper for __req_mod().
 205  * Set m->bio to the master bio, if it is fit to be completed,
 206  * or leave it alone (it is initialized to NULL in __req_mod),
 207  * if it has already been completed, or cannot be completed yet.
 208  * If m->bio is set, the error status to be returned is placed in m->error.
 209  */
 210 static
 211 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
 212 {
 213         const unsigned s = req->rq_state;
 214         struct drbd_device *device = req->device;
 215         int error, ok;
 216 
 217         /* we must not complete the master bio, while it is
 218          *      still being processed by _drbd_send_zc_bio (drbd_send_dblock)
 219          *      not yet acknowledged by the peer
 220          *      not yet completed by the local io subsystem
 221          * these flags may get cleared in any order by
 222          *      the worker,
 223          *      the receiver,
 224          *      the bio_endio completion callbacks.
 225          */
 226         if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
 227             (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
 228             (s & RQ_COMPLETION_SUSP)) {
 229                 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
 230                 return;
 231         }
 232 
 233         if (!req->master_bio) {
 234                 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
 235                 return;
 236         }
 237 
 238         /*
 239          * figure out whether to report success or failure.
 240          *
 241          * report success when at least one of the operations succeeded.
 242          * or, to put the other way,
 243          * only report failure, when both operations failed.
 244          *
 245          * what to do about the failures is handled elsewhere.
 246          * what we need to do here is just: complete the master_bio.
 247          *
 248          * local completion error, if any, has been stored as ERR_PTR
 249          * in private_bio within drbd_request_endio.
 250          */
 251         ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
 252         error = PTR_ERR(req->private_bio);
 253 
 254         /* Before we can signal completion to the upper layers,
 255          * we may need to close the current transfer log epoch.
 256          * We are within the request lock, so we can simply compare
 257          * the request epoch number with the current transfer log
 258          * epoch number.  If they match, increase the current_tle_nr,
 259          * and reset the transfer log epoch write_cnt.
 260          */
 261         if (op_is_write(bio_op(req->master_bio)) &&
 262             req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
 263                 start_new_tl_epoch(first_peer_device(device)->connection);
 264 
 265         /* Update disk stats */
 266         _drbd_end_io_acct(device, req);
 267 
 268         /* If READ failed,
 269          * have it be pushed back to the retry work queue,
 270          * so it will re-enter __drbd_make_request(),
 271          * and be re-assigned to a suitable local or remote path,
 272          * or failed if we do not have access to good data anymore.
 273          *
 274          * Unless it was failed early by __drbd_make_request(),
 275          * because no path was available, in which case
 276          * it was not even added to the transfer_log.
 277          *
 278          * read-ahead may fail, and will not be retried.
 279          *
 280          * WRITE should have used all available paths already.
 281          */
 282         if (!ok &&
 283             bio_op(req->master_bio) == REQ_OP_READ &&
 284             !(req->master_bio->bi_opf & REQ_RAHEAD) &&
 285             !list_empty(&req->tl_requests))
 286                 req->rq_state |= RQ_POSTPONED;
 287 
 288         if (!(req->rq_state & RQ_POSTPONED)) {
 289                 m->error = ok ? 0 : (error ?: -EIO);
 290                 m->bio = req->master_bio;
 291                 req->master_bio = NULL;
 292                 /* We leave it in the tree, to be able to verify later
 293                  * write-acks in protocol != C during resync.
 294                  * But we mark it as "complete", so it won't be counted as
 295                  * conflict in a multi-primary setup. */
 296                 req->i.completed = true;
 297         }
 298 
 299         if (req->i.waiting)
 300                 wake_up(&device->misc_wait);
 301 
 302         /* Either we are about to complete to upper layers,
 303          * or we will restart this request.
 304          * In either case, the request object will be destroyed soon,
 305          * so better remove it from all lists. */
 306         list_del_init(&req->req_pending_master_completion);
 307 }
 308 
 309 /* still holds resource->req_lock */
 310 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
 311 {
 312         struct drbd_device *device = req->device;
 313         D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
 314 
 315         if (!put)
 316                 return;
 317 
 318         if (!atomic_sub_and_test(put, &req->completion_ref))
 319                 return;
 320 
 321         drbd_req_complete(req, m);
 322 
 323         /* local completion may still come in later,
 324          * we need to keep the req object around. */
 325         if (req->rq_state & RQ_LOCAL_ABORTED)
 326                 return;
 327 
 328         if (req->rq_state & RQ_POSTPONED) {
 329                 /* don't destroy the req object just yet,
 330                  * but queue it for retry */
 331                 drbd_restart_request(req);
 332                 return;
 333         }
 334 
 335         kref_put(&req->kref, drbd_req_destroy);
 336 }
 337 
 338 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
 339 {
 340         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 341         if (!connection)
 342                 return;
 343         if (connection->req_next == NULL)
 344                 connection->req_next = req;
 345 }
 346 
 347 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
 348 {
 349         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 350         if (!connection)
 351                 return;
 352         if (connection->req_next != req)
 353                 return;
 354         list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
 355                 const unsigned s = req->rq_state;
 356                 if (s & RQ_NET_QUEUED)
 357                         break;
 358         }
 359         if (&req->tl_requests == &connection->transfer_log)
 360                 req = NULL;
 361         connection->req_next = req;
 362 }
 363 
 364 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
 365 {
 366         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 367         if (!connection)
 368                 return;
 369         if (connection->req_ack_pending == NULL)
 370                 connection->req_ack_pending = req;
 371 }
 372 
 373 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
 374 {
 375         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 376         if (!connection)
 377                 return;
 378         if (connection->req_ack_pending != req)
 379                 return;
 380         list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
 381                 const unsigned s = req->rq_state;
 382                 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
 383                         break;
 384         }
 385         if (&req->tl_requests == &connection->transfer_log)
 386                 req = NULL;
 387         connection->req_ack_pending = req;
 388 }
 389 
 390 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
 391 {
 392         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 393         if (!connection)
 394                 return;
 395         if (connection->req_not_net_done == NULL)
 396                 connection->req_not_net_done = req;
 397 }
 398 
 399 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
 400 {
 401         struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
 402         if (!connection)
 403                 return;
 404         if (connection->req_not_net_done != req)
 405                 return;
 406         list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
 407                 const unsigned s = req->rq_state;
 408                 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
 409                         break;
 410         }
 411         if (&req->tl_requests == &connection->transfer_log)
 412                 req = NULL;
 413         connection->req_not_net_done = req;
 414 }
 415 
 416 /* I'd like this to be the only place that manipulates
 417  * req->completion_ref and req->kref. */
 418 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
 419                 int clear, int set)
 420 {
 421         struct drbd_device *device = req->device;
 422         struct drbd_peer_device *peer_device = first_peer_device(device);
 423         unsigned s = req->rq_state;
 424         int c_put = 0;
 425 
 426         if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
 427                 set |= RQ_COMPLETION_SUSP;
 428 
 429         /* apply */
 430 
 431         req->rq_state &= ~clear;
 432         req->rq_state |= set;
 433 
 434         /* no change? */
 435         if (req->rq_state == s)
 436                 return;
 437 
 438         /* intent: get references */
 439 
 440         kref_get(&req->kref);
 441 
 442         if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
 443                 atomic_inc(&req->completion_ref);
 444 
 445         if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
 446                 inc_ap_pending(device);
 447                 atomic_inc(&req->completion_ref);
 448         }
 449 
 450         if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
 451                 atomic_inc(&req->completion_ref);
 452                 set_if_null_req_next(peer_device, req);
 453         }
 454 
 455         if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
 456                 kref_get(&req->kref); /* wait for the DONE */
 457 
 458         if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
 459                 /* potentially already completed in the ack_receiver thread */
 460                 if (!(s & RQ_NET_DONE)) {
 461                         atomic_add(req->i.size >> 9, &device->ap_in_flight);
 462                         set_if_null_req_not_net_done(peer_device, req);
 463                 }
 464                 if (req->rq_state & RQ_NET_PENDING)
 465                         set_if_null_req_ack_pending(peer_device, req);
 466         }
 467 
 468         if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
 469                 atomic_inc(&req->completion_ref);
 470 
 471         /* progress: put references */
 472 
 473         if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
 474                 ++c_put;
 475 
 476         if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
 477                 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
 478                 ++c_put;
 479         }
 480 
 481         if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
 482                 if (req->rq_state & RQ_LOCAL_ABORTED)
 483                         kref_put(&req->kref, drbd_req_destroy);
 484                 else
 485                         ++c_put;
 486                 list_del_init(&req->req_pending_local);
 487         }
 488 
 489         if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
 490                 dec_ap_pending(device);
 491                 ++c_put;
 492                 req->acked_jif = jiffies;
 493                 advance_conn_req_ack_pending(peer_device, req);
 494         }
 495 
 496         if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
 497                 ++c_put;
 498                 advance_conn_req_next(peer_device, req);
 499         }
 500 
 501         if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
 502                 if (s & RQ_NET_SENT)
 503                         atomic_sub(req->i.size >> 9, &device->ap_in_flight);
 504                 if (s & RQ_EXP_BARR_ACK)
 505                         kref_put(&req->kref, drbd_req_destroy);
 506                 req->net_done_jif = jiffies;
 507 
 508                 /* in ahead/behind mode, or just in case,
 509                  * before we finally destroy this request,
 510                  * the caching pointers must not reference it anymore */
 511                 advance_conn_req_next(peer_device, req);
 512                 advance_conn_req_ack_pending(peer_device, req);
 513                 advance_conn_req_not_net_done(peer_device, req);
 514         }
 515 
 516         /* potentially complete and destroy */
 517 
 518         /* If we made progress, retry conflicting peer requests, if any. */
 519         if (req->i.waiting)
 520                 wake_up(&device->misc_wait);
 521 
 522         drbd_req_put_completion_ref(req, m, c_put);
 523         kref_put(&req->kref, drbd_req_destroy);
 524 }
 525 
 526 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
 527 {
 528         char b[BDEVNAME_SIZE];
 529 
 530         if (!__ratelimit(&drbd_ratelimit_state))
 531                 return;
 532 
 533         drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
 534                         (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
 535                         (unsigned long long)req->i.sector,
 536                         req->i.size >> 9,
 537                         bdevname(device->ldev->backing_bdev, b));
 538 }
 539 
 540 /* Helper for HANDED_OVER_TO_NETWORK.
 541  * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
 542  * Is it also still "PENDING"?
 543  * --> If so, clear PENDING and set NET_OK below.
 544  * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
 545  * (and we must not set RQ_NET_OK) */
 546 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
 547 {
 548         return (req->rq_state &
 549                    (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
 550                 == (RQ_WRITE|RQ_NET_PENDING);
 551 }
 552 
 553 /* obviously this could be coded as many single functions
 554  * instead of one huge switch,
 555  * or by putting the code directly in the respective locations
 556  * (as it has been before).
 557  *
 558  * but having it this way
 559  *  enforces that it is all in this one place, where it is easier to audit,
 560  *  it makes it obvious that whatever "event" "happens" to a request should
 561  *  happen "atomically" within the req_lock,
 562  *  and it enforces that we have to think in a very structured manner
 563  *  about the "events" that may happen to a request during its life time ...
 564  */
 565 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
 566                 struct bio_and_error *m)
 567 {
 568         struct drbd_device *const device = req->device;
 569         struct drbd_peer_device *const peer_device = first_peer_device(device);
 570         struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
 571         struct net_conf *nc;
 572         int p, rv = 0;
 573 
 574         if (m)
 575                 m->bio = NULL;
 576 
 577         switch (what) {
 578         default:
 579                 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
 580                 break;
 581 
 582         /* does not happen...
 583          * initialization done in drbd_req_new
 584         case CREATED:
 585                 break;
 586                 */
 587 
 588         case TO_BE_SENT: /* via network */
 589                 /* reached via __drbd_make_request
 590                  * and from w_read_retry_remote */
 591                 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
 592                 rcu_read_lock();
 593                 nc = rcu_dereference(connection->net_conf);
 594                 p = nc->wire_protocol;
 595                 rcu_read_unlock();
 596                 req->rq_state |=
 597                         p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
 598                         p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
 599                 mod_rq_state(req, m, 0, RQ_NET_PENDING);
 600                 break;
 601 
 602         case TO_BE_SUBMITTED: /* locally */
 603                 /* reached via __drbd_make_request */
 604                 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
 605                 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
 606                 break;
 607 
 608         case COMPLETED_OK:
 609                 if (req->rq_state & RQ_WRITE)
 610                         device->writ_cnt += req->i.size >> 9;
 611                 else
 612                         device->read_cnt += req->i.size >> 9;
 613 
 614                 mod_rq_state(req, m, RQ_LOCAL_PENDING,
 615                                 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
 616                 break;
 617 
 618         case ABORT_DISK_IO:
 619                 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
 620                 break;
 621 
 622         case WRITE_COMPLETED_WITH_ERROR:
 623                 drbd_report_io_error(device, req);
 624                 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
 625                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 626                 break;
 627 
 628         case READ_COMPLETED_WITH_ERROR:
 629                 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
 630                 drbd_report_io_error(device, req);
 631                 __drbd_chk_io_error(device, DRBD_READ_ERROR);
 632                 /* fall through. */
 633         case READ_AHEAD_COMPLETED_WITH_ERROR:
 634                 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
 635                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 636                 break;
 637 
 638         case DISCARD_COMPLETED_NOTSUPP:
 639         case DISCARD_COMPLETED_WITH_ERROR:
 640                 /* I'd rather not detach from local disk just because it
 641                  * failed a REQ_OP_DISCARD. */
 642                 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 643                 break;
 644 
 645         case QUEUE_FOR_NET_READ:
 646                 /* READ, and
 647                  * no local disk,
 648                  * or target area marked as invalid,
 649                  * or just got an io-error. */
 650                 /* from __drbd_make_request
 651                  * or from bio_endio during read io-error recovery */
 652 
 653                 /* So we can verify the handle in the answer packet.
 654                  * Corresponding drbd_remove_request_interval is in
 655                  * drbd_req_complete() */
 656                 D_ASSERT(device, drbd_interval_empty(&req->i));
 657                 drbd_insert_interval(&device->read_requests, &req->i);
 658 
 659                 set_bit(UNPLUG_REMOTE, &device->flags);
 660 
 661                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 662                 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
 663                 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 664                 req->w.cb = w_send_read_req;
 665                 drbd_queue_work(&connection->sender_work,
 666                                 &req->w);
 667                 break;
 668 
 669         case QUEUE_FOR_NET_WRITE:
 670                 /* assert something? */
 671                 /* from __drbd_make_request only */
 672 
 673                 /* Corresponding drbd_remove_request_interval is in
 674                  * drbd_req_complete() */
 675                 D_ASSERT(device, drbd_interval_empty(&req->i));
 676                 drbd_insert_interval(&device->write_requests, &req->i);
 677 
 678                 /* NOTE
 679                  * In case the req ended up on the transfer log before being
 680                  * queued on the worker, it could lead to this request being
 681                  * missed during cleanup after connection loss.
 682                  * So we have to do both operations here,
 683                  * within the same lock that protects the transfer log.
 684                  *
 685                  * _req_add_to_epoch(req); this has to be after the
 686                  * _maybe_start_new_epoch(req); which happened in
 687                  * __drbd_make_request, because we now may set the bit
 688                  * again ourselves to close the current epoch.
 689                  *
 690                  * Add req to the (now) current epoch (barrier). */
 691 
 692                 /* otherwise we may lose an unplug, which may cause some remote
 693                  * io-scheduler timeout to expire, increasing maximum latency,
 694                  * hurting performance. */
 695                 set_bit(UNPLUG_REMOTE, &device->flags);
 696 
 697                 /* queue work item to send data */
 698                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 699                 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
 700                 req->w.cb =  w_send_dblock;
 701                 drbd_queue_work(&connection->sender_work,
 702                                 &req->w);
 703 
 704                 /* close the epoch, in case it outgrew the limit */
 705                 rcu_read_lock();
 706                 nc = rcu_dereference(connection->net_conf);
 707                 p = nc->max_epoch_size;
 708                 rcu_read_unlock();
 709                 if (connection->current_tle_writes >= p)
 710                         start_new_tl_epoch(connection);
 711 
 712                 break;
 713 
 714         case QUEUE_FOR_SEND_OOS:
 715                 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 716                 req->w.cb =  w_send_out_of_sync;
 717                 drbd_queue_work(&connection->sender_work,
 718                                 &req->w);
 719                 break;
 720 
 721         case READ_RETRY_REMOTE_CANCELED:
 722         case SEND_CANCELED:
 723         case SEND_FAILED:
 724                 /* real cleanup will be done from tl_clear.  just update flags
 725                  * so it is no longer marked as on the worker queue */
 726                 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
 727                 break;
 728 
 729         case HANDED_OVER_TO_NETWORK:
 730                 /* assert something? */
 731                 if (is_pending_write_protocol_A(req))
 732                         /* this is what is dangerous about protocol A:
 733                          * pretend it was successfully written on the peer. */
 734                         mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
 735                                                 RQ_NET_SENT|RQ_NET_OK);
 736                 else
 737                         mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
 738                 /* It is still not yet RQ_NET_DONE until the
 739                  * corresponding epoch barrier got acked as well,
 740                  * so we know what to dirty on connection loss. */
 741                 break;
 742 
 743         case OOS_HANDED_TO_NETWORK:
 744                 /* Was not set PENDING, no longer QUEUED, so is now DONE
 745                  * as far as this connection is concerned. */
 746                 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
 747                 break;
 748 
 749         case CONNECTION_LOST_WHILE_PENDING:
 750                 /* transfer log cleanup after connection loss */
 751                 mod_rq_state(req, m,
 752                                 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
 753                                 RQ_NET_DONE);
 754                 break;
 755 
 756         case CONFLICT_RESOLVED:
 757                 /* for superseded conflicting writes of multiple primaries,
 758                  * there is no need to keep anything in the tl, potential
 759                  * node crashes are covered by the activity log.
 760                  *
 761                  * If this request had been marked as RQ_POSTPONED before,
 762                  * it will actually not be completed, but "restarted",
 763                  * resubmitted from the retry worker context. */
 764                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 765                 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 766                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
 767                 break;
 768 
 769         case WRITE_ACKED_BY_PEER_AND_SIS:
 770                 req->rq_state |= RQ_NET_SIS;
 771         case WRITE_ACKED_BY_PEER:
 772                 /* Normal operation protocol C: successfully written on peer.
 773                  * During resync, even in protocol != C,
 774                  * we requested an explicit write ack anyways.
 775                  * Which means we cannot even assert anything here.
 776                  * Nothing more to do here.
 777                  * We want to keep the tl in place for all protocols, to cater
 778                  * for volatile write-back caches on lower level devices. */
 779                 goto ack_common;
 780         case RECV_ACKED_BY_PEER:
 781                 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
 782                 /* protocol B; pretends to be successfully written on peer.
 783                  * see also notes above in HANDED_OVER_TO_NETWORK about
 784                  * protocol != C */
 785         ack_common:
 786                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
 787                 break;
 788 
 789         case POSTPONE_WRITE:
 790                 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 791                 /* If this node has already detected the write conflict, the
 792                  * worker will be waiting on misc_wait.  Wake it up once this
 793                  * request has completed locally.
 794                  */
 795                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 796                 req->rq_state |= RQ_POSTPONED;
 797                 if (req->i.waiting)
 798                         wake_up(&device->misc_wait);
 799                 /* Do not clear RQ_NET_PENDING. This request will make further
 800                  * progress via restart_conflicting_writes() or
 801                  * fail_postponed_requests(). Hopefully. */
 802                 break;
 803 
 804         case NEG_ACKED:
 805                 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
 806                 break;
 807 
 808         case FAIL_FROZEN_DISK_IO:
 809                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 810                         break;
 811                 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 812                 break;
 813 
 814         case RESTART_FROZEN_DISK_IO:
 815                 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 816                         break;
 817 
 818                 mod_rq_state(req, m,
 819                                 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
 820                                 RQ_LOCAL_PENDING);
 821 
 822                 rv = MR_READ;
 823                 if (bio_data_dir(req->master_bio) == WRITE)
 824                         rv = MR_WRITE;
 825 
 826                 get_ldev(device); /* always succeeds in this call path */
 827                 req->w.cb = w_restart_disk_io;
 828                 drbd_queue_work(&connection->sender_work,
 829                                 &req->w);
 830                 break;
 831 
 832         case RESEND:
 833                 /* Simply complete (local only) READs. */
 834                 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
 835                         mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 836                         break;
 837                 }
 838 
 839                 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 840                    before the connection loss (B&C only); only P_BARRIER_ACK
 841                    (or the local completion?) was missing when we suspended.
 842                    Throwing them out of the TL here by pretending we got a BARRIER_ACK.
 843                    During connection handshake, we ensure that the peer was not rebooted. */
 844                 if (!(req->rq_state & RQ_NET_OK)) {
 845                         /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
 846                          * in that case we must not set RQ_NET_PENDING. */
 847 
 848                         mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
 849                         if (req->w.cb) {
 850                                 /* w.cb expected to be w_send_dblock, or w_send_read_req */
 851                                 drbd_queue_work(&connection->sender_work,
 852                                                 &req->w);
 853                                 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 854                         } /* else: FIXME can this happen? */
 855                         break;
 856                 }
 857                 /* else, fall through - to BARRIER_ACKED */
 858 
 859         case BARRIER_ACKED:
 860                 /* barrier ack for READ requests does not make sense */
 861                 if (!(req->rq_state & RQ_WRITE))
 862                         break;
 863 
 864                 if (req->rq_state & RQ_NET_PENDING) {
 865                         /* barrier came in before all requests were acked.
 866                          * this is bad, because if the connection is lost now,
 867                          * we won't be able to clean them up... */
 868                         drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
 869                 }
 870                 /* Allowed to complete requests, even while suspended.
 871                  * As this is called for all requests within a matching epoch,
 872                  * we need to filter, and only set RQ_NET_DONE for those that
 873                  * have actually been on the wire. */
 874                 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
 875                                 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
 876                 break;
 877 
 878         case DATA_RECEIVED:
 879                 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 880                 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
 881                 break;
 882 
 883         case QUEUE_AS_DRBD_BARRIER:
 884                 start_new_tl_epoch(connection);
 885                 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
 886                 break;
 887         };
 888 
 889         return rv;
 890 }
 891 
 892 /* we may do a local read if:
 893  * - we are consistent (of course),
 894  * - or we are generally inconsistent,
 895  *   BUT we are still/already IN SYNC for this area.
 896  *   since size may be bigger than BM_BLOCK_SIZE,
 897  *   we may need to check several bits.
 898  */
 899 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
 900 {
 901         unsigned long sbnr, ebnr;
 902         sector_t esector, nr_sectors;
 903 
 904         if (device->state.disk == D_UP_TO_DATE)
 905                 return true;
 906         if (device->state.disk != D_INCONSISTENT)
 907                 return false;
 908         esector = sector + (size >> 9) - 1;
 909         nr_sectors = drbd_get_capacity(device->this_bdev);
 910         D_ASSERT(device, sector  < nr_sectors);
 911         D_ASSERT(device, esector < nr_sectors);
 912 
 913         sbnr = BM_SECT_TO_BIT(sector);
 914         ebnr = BM_SECT_TO_BIT(esector);
 915 
 916         return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
 917 }
 918 
 919 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
 920                 enum drbd_read_balancing rbm)
 921 {
 922         struct backing_dev_info *bdi;
 923         int stripe_shift;
 924 
 925         switch (rbm) {
 926         case RB_CONGESTED_REMOTE:
 927                 bdi = device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
 928                 return bdi_read_congested(bdi);
 929         case RB_LEAST_PENDING:
 930                 return atomic_read(&device->local_cnt) >
 931                         atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
 932         case RB_32K_STRIPING:  /* stripe_shift = 15 */
 933         case RB_64K_STRIPING:
 934         case RB_128K_STRIPING:
 935         case RB_256K_STRIPING:
 936         case RB_512K_STRIPING:
 937         case RB_1M_STRIPING:   /* stripe_shift = 20 */
 938                 stripe_shift = (rbm - RB_32K_STRIPING + 15);
 939                 return (sector >> (stripe_shift - 9)) & 1;
 940         case RB_ROUND_ROBIN:
 941                 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
 942         case RB_PREFER_REMOTE:
 943                 return true;
 944         case RB_PREFER_LOCAL:
 945         default:
 946                 return false;
 947         }
 948 }
 949 
 950 /*
 951  * complete_conflicting_writes  -  wait for any conflicting write requests
 952  *
 953  * The write_requests tree contains all active write requests which we
 954  * currently know about.  Wait for any requests to complete which conflict with
 955  * the new one.
 956  *
 957  * Only way out: remove the conflicting intervals from the tree.
 958  */
 959 static void complete_conflicting_writes(struct drbd_request *req)
 960 {
 961         DEFINE_WAIT(wait);
 962         struct drbd_device *device = req->device;
 963         struct drbd_interval *i;
 964         sector_t sector = req->i.sector;
 965         int size = req->i.size;
 966 
 967         for (;;) {
 968                 drbd_for_each_overlap(i, &device->write_requests, sector, size) {
 969                         /* Ignore, if already completed to upper layers. */
 970                         if (i->completed)
 971                                 continue;
 972                         /* Handle the first found overlap.  After the schedule
 973                          * we have to restart the tree walk. */
 974                         break;
 975                 }
 976                 if (!i) /* if any */
 977                         break;
 978 
 979                 /* Indicate to wake up device->misc_wait on progress.  */
 980                 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
 981                 i->waiting = true;
 982                 spin_unlock_irq(&device->resource->req_lock);
 983                 schedule();
 984                 spin_lock_irq(&device->resource->req_lock);
 985         }
 986         finish_wait(&device->misc_wait, &wait);
 987 }
 988 
 989 /* called within req_lock */
 990 static void maybe_pull_ahead(struct drbd_device *device)
 991 {
 992         struct drbd_connection *connection = first_peer_device(device)->connection;
 993         struct net_conf *nc;
 994         bool congested = false;
 995         enum drbd_on_congestion on_congestion;
 996 
 997         rcu_read_lock();
 998         nc = rcu_dereference(connection->net_conf);
 999         on_congestion = nc ? nc->on_congestion : OC_BLOCK;
1000         rcu_read_unlock();
1001         if (on_congestion == OC_BLOCK ||
1002             connection->agreed_pro_version < 96)
1003                 return;
1004 
1005         if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1006                 return; /* nothing to do ... */
1007 
1008         /* If I don't even have good local storage, we can not reasonably try
1009          * to pull ahead of the peer. We also need the local reference to make
1010          * sure device->act_log is there.
1011          */
1012         if (!get_ldev_if_state(device, D_UP_TO_DATE))
1013                 return;
1014 
1015         if (nc->cong_fill &&
1016             atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1017                 drbd_info(device, "Congestion-fill threshold reached\n");
1018                 congested = true;
1019         }
1020 
1021         if (device->act_log->used >= nc->cong_extents) {
1022                 drbd_info(device, "Congestion-extents threshold reached\n");
1023                 congested = true;
1024         }
1025 
1026         if (congested) {
1027                 /* start a new epoch for non-mirrored writes */
1028                 start_new_tl_epoch(first_peer_device(device)->connection);
1029 
1030                 if (on_congestion == OC_PULL_AHEAD)
1031                         _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1032                 else  /*nc->on_congestion == OC_DISCONNECT */
1033                         _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1034         }
1035         put_ldev(device);
1036 }
1037 
1038 /* If this returns false, and req->private_bio is still set,
1039  * this should be submitted locally.
1040  *
1041  * If it returns false, but req->private_bio is not set,
1042  * we do not have access to good data :(
1043  *
1044  * Otherwise, this destroys req->private_bio, if any,
1045  * and returns true.
1046  */
1047 static bool do_remote_read(struct drbd_request *req)
1048 {
1049         struct drbd_device *device = req->device;
1050         enum drbd_read_balancing rbm;
1051 
1052         if (req->private_bio) {
1053                 if (!drbd_may_do_local_read(device,
1054                                         req->i.sector, req->i.size)) {
1055                         bio_put(req->private_bio);
1056                         req->private_bio = NULL;
1057                         put_ldev(device);
1058                 }
1059         }
1060 
1061         if (device->state.pdsk != D_UP_TO_DATE)
1062                 return false;
1063 
1064         if (req->private_bio == NULL)
1065                 return true;
1066 
1067         /* TODO: improve read balancing decisions, take into account drbd
1068          * protocol, pending requests etc. */
1069 
1070         rcu_read_lock();
1071         rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1072         rcu_read_unlock();
1073 
1074         if (rbm == RB_PREFER_LOCAL && req->private_bio)
1075                 return false; /* submit locally */
1076 
1077         if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1078                 if (req->private_bio) {
1079                         bio_put(req->private_bio);
1080                         req->private_bio = NULL;
1081                         put_ldev(device);
1082                 }
1083                 return true;
1084         }
1085 
1086         return false;
1087 }
1088 
1089 bool drbd_should_do_remote(union drbd_dev_state s)
1090 {
1091         return s.pdsk == D_UP_TO_DATE ||
1092                 (s.pdsk >= D_INCONSISTENT &&
1093                  s.conn >= C_WF_BITMAP_T &&
1094                  s.conn < C_AHEAD);
1095         /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1096            That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1097            states. */
1098 }
1099 
1100 static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1101 {
1102         return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1103         /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1104            since we enter state C_AHEAD only if proto >= 96 */
1105 }
1106 
1107 /* returns number of connections (== 1, for drbd 8.4)
1108  * expected to actually write this data,
1109  * which does NOT include those that we are L_AHEAD for. */
1110 static int drbd_process_write_request(struct drbd_request *req)
1111 {
1112         struct drbd_device *device = req->device;
1113         int remote, send_oos;
1114 
1115         remote = drbd_should_do_remote(device->state);
1116         send_oos = drbd_should_send_out_of_sync(device->state);
1117 
1118         /* Need to replicate writes.  Unless it is an empty flush,
1119          * which is better mapped to a DRBD P_BARRIER packet,
1120          * also for drbd wire protocol compatibility reasons.
1121          * If this was a flush, just start a new epoch.
1122          * Unless the current epoch was empty anyways, or we are not currently
1123          * replicating, in which case there is no point. */
1124         if (unlikely(req->i.size == 0)) {
1125                 /* The only size==0 bios we expect are empty flushes. */
1126                 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1127                 if (remote)
1128                         _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1129                 return remote;
1130         }
1131 
1132         if (!remote && !send_oos)
1133                 return 0;
1134 
1135         D_ASSERT(device, !(remote && send_oos));
1136 
1137         if (remote) {
1138                 _req_mod(req, TO_BE_SENT);
1139                 _req_mod(req, QUEUE_FOR_NET_WRITE);
1140         } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1141                 _req_mod(req, QUEUE_FOR_SEND_OOS);
1142 
1143         return remote;
1144 }
1145 
1146 static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1147 {
1148         int err = drbd_issue_discard_or_zero_out(req->device,
1149                                 req->i.sector, req->i.size >> 9, flags);
1150         if (err)
1151                 req->private_bio->bi_status = BLK_STS_IOERR;
1152         bio_endio(req->private_bio);
1153 }
1154 
1155 static void
1156 drbd_submit_req_private_bio(struct drbd_request *req)
1157 {
1158         struct drbd_device *device = req->device;
1159         struct bio *bio = req->private_bio;
1160         unsigned int type;
1161 
1162         if (bio_op(bio) != REQ_OP_READ)
1163                 type = DRBD_FAULT_DT_WR;
1164         else if (bio->bi_opf & REQ_RAHEAD)
1165                 type = DRBD_FAULT_DT_RA;
1166         else
1167                 type = DRBD_FAULT_DT_RD;
1168 
1169         bio_set_dev(bio, device->ldev->backing_bdev);
1170 
1171         /* State may have changed since we grabbed our reference on the
1172          * ->ldev member. Double check, and short-circuit to endio.
1173          * In case the last activity log transaction failed to get on
1174          * stable storage, and this is a WRITE, we may not even submit
1175          * this bio. */
1176         if (get_ldev(device)) {
1177                 if (drbd_insert_fault(device, type))
1178                         bio_io_error(bio);
1179                 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1180                         drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1181                             ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1182                 else if (bio_op(bio) == REQ_OP_DISCARD)
1183                         drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1184                 else
1185                         generic_make_request(bio);
1186                 put_ldev(device);
1187         } else
1188                 bio_io_error(bio);
1189 }
1190 
1191 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1192 {
1193         spin_lock_irq(&device->resource->req_lock);
1194         list_add_tail(&req->tl_requests, &device->submit.writes);
1195         list_add_tail(&req->req_pending_master_completion,
1196                         &device->pending_master_completion[1 /* WRITE */]);
1197         spin_unlock_irq(&device->resource->req_lock);
1198         queue_work(device->submit.wq, &device->submit.worker);
1199         /* do_submit() may sleep internally on al_wait, too */
1200         wake_up(&device->al_wait);
1201 }
1202 
1203 /* returns the new drbd_request pointer, if the caller is expected to
1204  * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1205  * request on the submitter thread.
1206  * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1207  */
1208 static struct drbd_request *
1209 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1210 {
1211         const int rw = bio_data_dir(bio);
1212         struct drbd_request *req;
1213 
1214         /* allocate outside of all locks; */
1215         req = drbd_req_new(device, bio);
1216         if (!req) {
1217                 dec_ap_bio(device);
1218                 /* only pass the error to the upper layers.
1219                  * if user cannot handle io errors, that's not our business. */
1220                 drbd_err(device, "could not kmalloc() req\n");
1221                 bio->bi_status = BLK_STS_RESOURCE;
1222                 bio_endio(bio);
1223                 return ERR_PTR(-ENOMEM);
1224         }
1225         req->start_jif = start_jif;
1226 
1227         if (!get_ldev(device)) {
1228                 bio_put(req->private_bio);
1229                 req->private_bio = NULL;
1230         }
1231 
1232         /* Update disk stats */
1233         _drbd_start_io_acct(device, req);
1234 
1235         /* process discards always from our submitter thread */
1236         if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1237             bio_op(bio) == REQ_OP_DISCARD)
1238                 goto queue_for_submitter_thread;
1239 
1240         if (rw == WRITE && req->private_bio && req->i.size
1241         && !test_bit(AL_SUSPENDED, &device->flags)) {
1242                 if (!drbd_al_begin_io_fastpath(device, &req->i))
1243                         goto queue_for_submitter_thread;
1244                 req->rq_state |= RQ_IN_ACT_LOG;
1245                 req->in_actlog_jif = jiffies;
1246         }
1247         return req;
1248 
1249  queue_for_submitter_thread:
1250         atomic_inc(&device->ap_actlog_cnt);
1251         drbd_queue_write(device, req);
1252         return NULL;
1253 }
1254 
1255 /* Require at least one path to current data.
1256  * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1257  * We would not allow to read what was written,
1258  * we would not have bumped the data generation uuids,
1259  * we would cause data divergence for all the wrong reasons.
1260  *
1261  * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1262  * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1263  * and queues for retry later.
1264  */
1265 static bool may_do_writes(struct drbd_device *device)
1266 {
1267         const union drbd_dev_state s = device->state;
1268         return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1269 }
1270 
1271 struct drbd_plug_cb {
1272         struct blk_plug_cb cb;
1273         struct drbd_request *most_recent_req;
1274         /* do we need more? */
1275 };
1276 
1277 static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1278 {
1279         struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1280         struct drbd_resource *resource = plug->cb.data;
1281         struct drbd_request *req = plug->most_recent_req;
1282 
1283         kfree(cb);
1284         if (!req)
1285                 return;
1286 
1287         spin_lock_irq(&resource->req_lock);
1288         /* In case the sender did not process it yet, raise the flag to
1289          * have it followed with P_UNPLUG_REMOTE just after. */
1290         req->rq_state |= RQ_UNPLUG;
1291         /* but also queue a generic unplug */
1292         drbd_queue_unplug(req->device);
1293         kref_put(&req->kref, drbd_req_destroy);
1294         spin_unlock_irq(&resource->req_lock);
1295 }
1296 
1297 static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1298 {
1299         /* A lot of text to say
1300          * return (struct drbd_plug_cb*)blk_check_plugged(); */
1301         struct drbd_plug_cb *plug;
1302         struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1303 
1304         if (cb)
1305                 plug = container_of(cb, struct drbd_plug_cb, cb);
1306         else
1307                 plug = NULL;
1308         return plug;
1309 }
1310 
1311 static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1312 {
1313         struct drbd_request *tmp = plug->most_recent_req;
1314         /* Will be sent to some peer.
1315          * Remember to tag it with UNPLUG_REMOTE on unplug */
1316         kref_get(&req->kref);
1317         plug->most_recent_req = req;
1318         if (tmp)
1319                 kref_put(&tmp->kref, drbd_req_destroy);
1320 }
1321 
1322 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1323 {
1324         struct drbd_resource *resource = device->resource;
1325         const int rw = bio_data_dir(req->master_bio);
1326         struct bio_and_error m = { NULL, };
1327         bool no_remote = false;
1328         bool submit_private_bio = false;
1329 
1330         spin_lock_irq(&resource->req_lock);
1331         if (rw == WRITE) {
1332                 /* This may temporarily give up the req_lock,
1333                  * but will re-aquire it before it returns here.
1334                  * Needs to be before the check on drbd_suspended() */
1335                 complete_conflicting_writes(req);
1336                 /* no more giving up req_lock from now on! */
1337 
1338                 /* check for congestion, and potentially stop sending
1339                  * full data updates, but start sending "dirty bits" only. */
1340                 maybe_pull_ahead(device);
1341         }
1342 
1343 
1344         if (drbd_suspended(device)) {
1345                 /* push back and retry: */
1346                 req->rq_state |= RQ_POSTPONED;
1347                 if (req->private_bio) {
1348                         bio_put(req->private_bio);
1349                         req->private_bio = NULL;
1350                         put_ldev(device);
1351                 }
1352                 goto out;
1353         }
1354 
1355         /* We fail READ early, if we can not serve it.
1356          * We must do this before req is registered on any lists.
1357          * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1358         if (rw != WRITE) {
1359                 if (!do_remote_read(req) && !req->private_bio)
1360                         goto nodata;
1361         }
1362 
1363         /* which transfer log epoch does this belong to? */
1364         req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1365 
1366         /* no point in adding empty flushes to the transfer log,
1367          * they are mapped to drbd barriers already. */
1368         if (likely(req->i.size!=0)) {
1369                 if (rw == WRITE)
1370                         first_peer_device(device)->connection->current_tle_writes++;
1371 
1372                 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1373         }
1374 
1375         if (rw == WRITE) {
1376                 if (req->private_bio && !may_do_writes(device)) {
1377                         bio_put(req->private_bio);
1378                         req->private_bio = NULL;
1379                         put_ldev(device);
1380                         goto nodata;
1381                 }
1382                 if (!drbd_process_write_request(req))
1383                         no_remote = true;
1384         } else {
1385                 /* We either have a private_bio, or we can read from remote.
1386                  * Otherwise we had done the goto nodata above. */
1387                 if (req->private_bio == NULL) {
1388                         _req_mod(req, TO_BE_SENT);
1389                         _req_mod(req, QUEUE_FOR_NET_READ);
1390                 } else
1391                         no_remote = true;
1392         }
1393 
1394         if (no_remote == false) {
1395                 struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1396                 if (plug)
1397                         drbd_update_plug(plug, req);
1398         }
1399 
1400         /* If it took the fast path in drbd_request_prepare, add it here.
1401          * The slow path has added it already. */
1402         if (list_empty(&req->req_pending_master_completion))
1403                 list_add_tail(&req->req_pending_master_completion,
1404                         &device->pending_master_completion[rw == WRITE]);
1405         if (req->private_bio) {
1406                 /* needs to be marked within the same spinlock */
1407                 req->pre_submit_jif = jiffies;
1408                 list_add_tail(&req->req_pending_local,
1409                         &device->pending_completion[rw == WRITE]);
1410                 _req_mod(req, TO_BE_SUBMITTED);
1411                 /* but we need to give up the spinlock to submit */
1412                 submit_private_bio = true;
1413         } else if (no_remote) {
1414 nodata:
1415                 if (__ratelimit(&drbd_ratelimit_state))
1416                         drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1417                                         (unsigned long long)req->i.sector, req->i.size >> 9);
1418                 /* A write may have been queued for send_oos, however.
1419                  * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1420         }
1421 
1422 out:
1423         drbd_req_put_completion_ref(req, &m, 1);
1424         spin_unlock_irq(&resource->req_lock);
1425 
1426         /* Even though above is a kref_put(), this is safe.
1427          * As long as we still need to submit our private bio,
1428          * we hold a completion ref, and the request cannot disappear.
1429          * If however this request did not even have a private bio to submit
1430          * (e.g. remote read), req may already be invalid now.
1431          * That's why we cannot check on req->private_bio. */
1432         if (submit_private_bio)
1433                 drbd_submit_req_private_bio(req);
1434         if (m.bio)
1435                 complete_master_bio(device, &m);
1436 }
1437 
1438 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1439 {
1440         struct drbd_request *req = drbd_request_prepare(device, bio, start_jif);
1441         if (IS_ERR_OR_NULL(req))
1442                 return;
1443         drbd_send_and_submit(device, req);
1444 }
1445 
1446 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1447 {
1448         struct blk_plug plug;
1449         struct drbd_request *req, *tmp;
1450 
1451         blk_start_plug(&plug);
1452         list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1453                 const int rw = bio_data_dir(req->master_bio);
1454 
1455                 if (rw == WRITE /* rw != WRITE should not even end up here! */
1456                 && req->private_bio && req->i.size
1457                 && !test_bit(AL_SUSPENDED, &device->flags)) {
1458                         if (!drbd_al_begin_io_fastpath(device, &req->i))
1459                                 continue;
1460 
1461                         req->rq_state |= RQ_IN_ACT_LOG;
1462                         req->in_actlog_jif = jiffies;
1463                         atomic_dec(&device->ap_actlog_cnt);
1464                 }
1465 
1466                 list_del_init(&req->tl_requests);
1467                 drbd_send_and_submit(device, req);
1468         }
1469         blk_finish_plug(&plug);
1470 }
1471 
1472 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1473                                             struct list_head *incoming,
1474                                             struct list_head *pending,
1475                                             struct list_head *later)
1476 {
1477         struct drbd_request *req;
1478         int wake = 0;
1479         int err;
1480 
1481         spin_lock_irq(&device->al_lock);
1482         while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1483                 err = drbd_al_begin_io_nonblock(device, &req->i);
1484                 if (err == -ENOBUFS)
1485                         break;
1486                 if (err == -EBUSY)
1487                         wake = 1;
1488                 if (err)
1489                         list_move_tail(&req->tl_requests, later);
1490                 else
1491                         list_move_tail(&req->tl_requests, pending);
1492         }
1493         spin_unlock_irq(&device->al_lock);
1494         if (wake)
1495                 wake_up(&device->al_wait);
1496         return !list_empty(pending);
1497 }
1498 
1499 static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1500 {
1501         struct blk_plug plug;
1502         struct drbd_request *req;
1503 
1504         blk_start_plug(&plug);
1505         while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1506                 req->rq_state |= RQ_IN_ACT_LOG;
1507                 req->in_actlog_jif = jiffies;
1508                 atomic_dec(&device->ap_actlog_cnt);
1509                 list_del_init(&req->tl_requests);
1510                 drbd_send_and_submit(device, req);
1511         }
1512         blk_finish_plug(&plug);
1513 }
1514 
1515 void do_submit(struct work_struct *ws)
1516 {
1517         struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1518         LIST_HEAD(incoming);    /* from drbd_make_request() */
1519         LIST_HEAD(pending);     /* to be submitted after next AL-transaction commit */
1520         LIST_HEAD(busy);        /* blocked by resync requests */
1521 
1522         /* grab new incoming requests */
1523         spin_lock_irq(&device->resource->req_lock);
1524         list_splice_tail_init(&device->submit.writes, &incoming);
1525         spin_unlock_irq(&device->resource->req_lock);
1526 
1527         for (;;) {
1528                 DEFINE_WAIT(wait);
1529 
1530                 /* move used-to-be-busy back to front of incoming */
1531                 list_splice_init(&busy, &incoming);
1532                 submit_fast_path(device, &incoming);
1533                 if (list_empty(&incoming))
1534                         break;
1535 
1536                 for (;;) {
1537                         prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1538 
1539                         list_splice_init(&busy, &incoming);
1540                         prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1541                         if (!list_empty(&pending))
1542                                 break;
1543 
1544                         schedule();
1545 
1546                         /* If all currently "hot" activity log extents are kept busy by
1547                          * incoming requests, we still must not totally starve new
1548                          * requests to "cold" extents.
1549                          * Something left on &incoming means there had not been
1550                          * enough update slots available, and the activity log
1551                          * has been marked as "starving".
1552                          *
1553                          * Try again now, without looking for new requests,
1554                          * effectively blocking all new requests until we made
1555                          * at least _some_ progress with what we currently have.
1556                          */
1557                         if (!list_empty(&incoming))
1558                                 continue;
1559 
1560                         /* Nothing moved to pending, but nothing left
1561                          * on incoming: all moved to busy!
1562                          * Grab new and iterate. */
1563                         spin_lock_irq(&device->resource->req_lock);
1564                         list_splice_tail_init(&device->submit.writes, &incoming);
1565                         spin_unlock_irq(&device->resource->req_lock);
1566                 }
1567                 finish_wait(&device->al_wait, &wait);
1568 
1569                 /* If the transaction was full, before all incoming requests
1570                  * had been processed, skip ahead to commit, and iterate
1571                  * without splicing in more incoming requests from upper layers.
1572                  *
1573                  * Else, if all incoming have been processed,
1574                  * they have become either "pending" (to be submitted after
1575                  * next transaction commit) or "busy" (blocked by resync).
1576                  *
1577                  * Maybe more was queued, while we prepared the transaction?
1578                  * Try to stuff those into this transaction as well.
1579                  * Be strictly non-blocking here,
1580                  * we already have something to commit.
1581                  *
1582                  * Commit if we don't make any more progres.
1583                  */
1584 
1585                 while (list_empty(&incoming)) {
1586                         LIST_HEAD(more_pending);
1587                         LIST_HEAD(more_incoming);
1588                         bool made_progress;
1589 
1590                         /* It is ok to look outside the lock,
1591                          * it's only an optimization anyways */
1592                         if (list_empty(&device->submit.writes))
1593                                 break;
1594 
1595                         spin_lock_irq(&device->resource->req_lock);
1596                         list_splice_tail_init(&device->submit.writes, &more_incoming);
1597                         spin_unlock_irq(&device->resource->req_lock);
1598 
1599                         if (list_empty(&more_incoming))
1600                                 break;
1601 
1602                         made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1603 
1604                         list_splice_tail_init(&more_pending, &pending);
1605                         list_splice_tail_init(&more_incoming, &incoming);
1606                         if (!made_progress)
1607                                 break;
1608                 }
1609 
1610                 drbd_al_begin_io_commit(device);
1611                 send_and_submit_pending(device, &pending);
1612         }
1613 }
1614 
1615 blk_qc_t drbd_make_request(struct request_queue *q, struct bio *bio)
1616 {
1617         struct drbd_device *device = (struct drbd_device *) q->queuedata;
1618         unsigned long start_jif;
1619 
1620         blk_queue_split(q, &bio);
1621 
1622         start_jif = jiffies;
1623 
1624         /*
1625          * what we "blindly" assume:
1626          */
1627         D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1628 
1629         inc_ap_bio(device);
1630         __drbd_make_request(device, bio, start_jif);
1631         return BLK_QC_T_NONE;
1632 }
1633 
1634 static bool net_timeout_reached(struct drbd_request *net_req,
1635                 struct drbd_connection *connection,
1636                 unsigned long now, unsigned long ent,
1637                 unsigned int ko_count, unsigned int timeout)
1638 {
1639         struct drbd_device *device = net_req->device;
1640 
1641         if (!time_after(now, net_req->pre_send_jif + ent))
1642                 return false;
1643 
1644         if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1645                 return false;
1646 
1647         if (net_req->rq_state & RQ_NET_PENDING) {
1648                 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1649                         jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1650                 return true;
1651         }
1652 
1653         /* We received an ACK already (or are using protocol A),
1654          * but are waiting for the epoch closing barrier ack.
1655          * Check if we sent the barrier already.  We should not blame the peer
1656          * for being unresponsive, if we did not even ask it yet. */
1657         if (net_req->epoch == connection->send.current_epoch_nr) {
1658                 drbd_warn(device,
1659                         "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1660                         jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1661                 return false;
1662         }
1663 
1664         /* Worst case: we may have been blocked for whatever reason, then
1665          * suddenly are able to send a lot of requests (and epoch separating
1666          * barriers) in quick succession.
1667          * The timestamp of the net_req may be much too old and not correspond
1668          * to the sending time of the relevant unack'ed barrier packet, so
1669          * would trigger a spurious timeout.  The latest barrier packet may
1670          * have a too recent timestamp to trigger the timeout, potentially miss
1671          * a timeout.  Right now we don't have a place to conveniently store
1672          * these timestamps.
1673          * But in this particular situation, the application requests are still
1674          * completed to upper layers, DRBD should still "feel" responsive.
1675          * No need yet to kill this connection, it may still recover.
1676          * If not, eventually we will have queued enough into the network for
1677          * us to block. From that point of view, the timestamp of the last sent
1678          * barrier packet is relevant enough.
1679          */
1680         if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1681                 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",
1682                         connection->send.last_sent_barrier_jif, now,
1683                         jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1684                 return true;
1685         }
1686         return false;
1687 }
1688 
1689 /* A request is considered timed out, if
1690  * - we have some effective timeout from the configuration,
1691  *   with some state restrictions applied,
1692  * - the oldest request is waiting for a response from the network
1693  *   resp. the local disk,
1694  * - the oldest request is in fact older than the effective timeout,
1695  * - the connection was established (resp. disk was attached)
1696  *   for longer than the timeout already.
1697  * Note that for 32bit jiffies and very stable connections/disks,
1698  * we may have a wrap around, which is catched by
1699  *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1700  *
1701  * Side effect: once per 32bit wrap-around interval, which means every
1702  * ~198 days with 250 HZ, we have a window where the timeout would need
1703  * to expire twice (worst case) to become effective. Good enough.
1704  */
1705 
1706 void request_timer_fn(struct timer_list *t)
1707 {
1708         struct drbd_device *device = from_timer(device, t, request_timer);
1709         struct drbd_connection *connection = first_peer_device(device)->connection;
1710         struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1711         struct net_conf *nc;
1712         unsigned long oldest_submit_jif;
1713         unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1714         unsigned long now;
1715         unsigned int ko_count = 0, timeout = 0;
1716 
1717         rcu_read_lock();
1718         nc = rcu_dereference(connection->net_conf);
1719         if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1720                 ko_count = nc->ko_count;
1721                 timeout = nc->timeout;
1722         }
1723 
1724         if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1725                 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1726                 put_ldev(device);
1727         }
1728         rcu_read_unlock();
1729 
1730 
1731         ent = timeout * HZ/10 * ko_count;
1732         et = min_not_zero(dt, ent);
1733 
1734         if (!et)
1735                 return; /* Recurring timer stopped */
1736 
1737         now = jiffies;
1738         nt = now + et;
1739 
1740         spin_lock_irq(&device->resource->req_lock);
1741         req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1742         req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1743 
1744         /* maybe the oldest request waiting for the peer is in fact still
1745          * blocking in tcp sendmsg.  That's ok, though, that's handled via the
1746          * socket send timeout, requesting a ping, and bumping ko-count in
1747          * we_should_drop_the_connection().
1748          */
1749 
1750         /* check the oldest request we did successfully sent,
1751          * but which is still waiting for an ACK. */
1752         req_peer = connection->req_ack_pending;
1753 
1754         /* if we don't have such request (e.g. protocoll A)
1755          * check the oldest requests which is still waiting on its epoch
1756          * closing barrier ack. */
1757         if (!req_peer)
1758                 req_peer = connection->req_not_net_done;
1759 
1760         /* evaluate the oldest peer request only in one timer! */
1761         if (req_peer && req_peer->device != device)
1762                 req_peer = NULL;
1763 
1764         /* do we have something to evaluate? */
1765         if (req_peer == NULL && req_write == NULL && req_read == NULL)
1766                 goto out;
1767 
1768         oldest_submit_jif =
1769                 (req_write && req_read)
1770                 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1771                   ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1772                 : req_write ? req_write->pre_submit_jif
1773                 : req_read ? req_read->pre_submit_jif : now;
1774 
1775         if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1776                 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1777 
1778         if (dt && oldest_submit_jif != now &&
1779                  time_after(now, oldest_submit_jif + dt) &&
1780                 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1781                 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1782                 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1783         }
1784 
1785         /* Reschedule timer for the nearest not already expired timeout.
1786          * Fallback to now + min(effective network timeout, disk timeout). */
1787         ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1788                 ? req_peer->pre_send_jif + ent : now + et;
1789         dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1790                 ? oldest_submit_jif + dt : now + et;
1791         nt = time_before(ent, dt) ? ent : dt;
1792 out:
1793         spin_unlock_irq(&device->resource->req_lock);
1794         mod_timer(&device->request_timer, nt);
1795 }