4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 drbd is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
15 drbd is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with drbd; see the file COPYING. If not, write to
22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
26 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/drbd.h>
34 /* Update disk stats at start of I/O request */
35 static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
37 const int rw = bio_data_dir(bio);
39 cpu = part_stat_lock();
40 part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
41 part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
42 part_inc_in_flight(&mdev->vdisk->part0, rw);
46 /* Update disk stats when completing request upwards */
47 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
49 int rw = bio_data_dir(req->master_bio);
50 unsigned long duration = jiffies - req->start_time;
52 cpu = part_stat_lock();
53 part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
54 part_round_stats(cpu, &mdev->vdisk->part0);
55 part_dec_in_flight(&mdev->vdisk->part0, rw);
59 static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
61 const unsigned long s = req->rq_state;
63 /* remove it from the transfer log.
64 * well, only if it had been there in the first
65 * place... if it had not (local only or conflicting
66 * and never sent), it should still be "empty" as
67 * initialized in drbd_req_new(), so we can list_del() it
68 * here unconditionally */
69 list_del(&req->tl_requests);
71 /* if it was a write, we may have to set the corresponding
72 * bit(s) out-of-sync first. If it had a local part, we need to
73 * release the reference to the activity log. */
75 /* Set out-of-sync unless both OK flags are set
76 * (local only or remote failed).
77 * Other places where we set out-of-sync:
78 * READ with local io-error */
79 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
80 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
82 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
83 drbd_set_in_sync(mdev, req->i.sector, req->i.size);
85 /* one might be tempted to move the drbd_al_complete_io
86 * to the local io completion callback drbd_endio_pri.
87 * but, if this was a mirror write, we may only
88 * drbd_al_complete_io after this is RQ_NET_DONE,
89 * otherwise the extent could be dropped from the al
90 * before it has actually been written on the peer.
91 * if we crash before our peer knows about the request,
92 * but after the extent has been dropped from the al,
93 * we would forget to resync the corresponding extent.
95 if (s & RQ_LOCAL_MASK) {
96 if (get_ldev_if_state(mdev, D_FAILED)) {
97 if (s & RQ_IN_ACT_LOG)
98 drbd_al_complete_io(mdev, req->i.sector);
100 } else if (__ratelimit(&drbd_ratelimit_state)) {
101 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu), "
102 "but my Disk seems to have failed :(\n",
103 (unsigned long long) req->i.sector);
111 static void queue_barrier(struct drbd_conf *mdev)
113 struct drbd_tl_epoch *b;
115 /* We are within the req_lock. Once we queued the barrier for sending,
116 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
117 * barrier/epoch object is added. This is the only place this bit is
118 * set. It indicates that the barrier for this epoch is already queued,
119 * and no new epoch has been created yet. */
120 if (test_bit(CREATE_BARRIER, &mdev->flags))
123 b = mdev->newest_tle;
124 b->w.cb = w_send_barrier;
125 /* inc_ap_pending done here, so we won't
126 * get imbalanced on connection loss.
127 * dec_ap_pending will be done in got_BarrierAck
128 * or (on connection loss) in tl_clear. */
129 inc_ap_pending(mdev);
130 drbd_queue_work(&mdev->data.work, &b->w);
131 set_bit(CREATE_BARRIER, &mdev->flags);
134 static void _about_to_complete_local_write(struct drbd_conf *mdev,
135 struct drbd_request *req)
137 const unsigned long s = req->rq_state;
139 /* Before we can signal completion to the upper layers,
140 * we may need to close the current epoch.
141 * We can skip this, if this request has not even been sent, because we
142 * did not have a fully established connection yet/anymore, during
143 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
145 if (mdev->state.conn >= C_CONNECTED &&
146 (s & RQ_NET_SENT) != 0 &&
147 req->epoch == mdev->newest_tle->br_number)
150 /* we need to do the conflict detection stuff,
151 * if the epoch_entries tree is non-empty and
152 * this request has completed on the network */
153 if ((s & RQ_NET_DONE) && !RB_EMPTY_ROOT(&mdev->epoch_entries)) {
154 const sector_t sector = req->i.sector;
155 const int size = req->i.size;
156 struct drbd_interval *i;
159 * there must be no conflicting requests, since
160 * they must have been failed on the spot */
162 i = drbd_find_overlap(&mdev->write_requests, sector, size);
164 struct drbd_request *req2 =
165 container_of(i, struct drbd_request, i);
167 dev_alert(DEV, "LOGIC BUG: completed: %p %llus +%u; "
168 "other: %p %llus +%u\n",
169 req, (unsigned long long)sector, size,
170 i, (unsigned long long)req2->i.sector, req2->i.size);
173 /* maybe "wake" those conflicting epoch entries
174 * that wait for this request to finish.
176 * currently, there can be only _one_ such ee
177 * (well, or some more, which would be pending
178 * P_DISCARD_ACK not yet sent by the asender...),
179 * since we block the receiver thread upon the
180 * first conflict detection, which will wait on
181 * misc_wait. maybe we want to assert that?
183 * anyways, if we found one,
184 * we just have to do a wake_up. */
185 i = drbd_find_overlap(&mdev->epoch_entries, sector, size);
187 wake_up(&mdev->misc_wait);
191 void complete_master_bio(struct drbd_conf *mdev,
192 struct bio_and_error *m)
194 bio_endio(m->bio, m->error);
198 /* Helper for __req_mod().
199 * Set m->bio to the master bio, if it is fit to be completed,
200 * or leave it alone (it is initialized to NULL in __req_mod),
201 * if it has already been completed, or cannot be completed yet.
202 * If m->bio is set, the error status to be returned is placed in m->error.
204 void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
206 const unsigned long s = req->rq_state;
207 struct drbd_conf *mdev = req->mdev;
208 /* only WRITES may end up here without a master bio (on barrier ack) */
209 int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE;
211 /* we must not complete the master bio, while it is
212 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
213 * not yet acknowledged by the peer
214 * not yet completed by the local io subsystem
215 * these flags may get cleared in any order by
218 * the bio_endio completion callbacks.
220 if (s & RQ_NET_QUEUED)
222 if (s & RQ_NET_PENDING)
224 if (s & RQ_LOCAL_PENDING)
227 if (req->master_bio) {
228 /* this is DATA_RECEIVED (remote read)
229 * or protocol C P_WRITE_ACK
230 * or protocol B P_RECV_ACK
231 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
232 * or canceled or failed,
233 * or killed from the transfer log due to connection loss.
237 * figure out whether to report success or failure.
239 * report success when at least one of the operations succeeded.
240 * or, to put the other way,
241 * only report failure, when both operations failed.
243 * what to do about the failures is handled elsewhere.
244 * what we need to do here is just: complete the master_bio.
246 * local completion error, if any, has been stored as ERR_PTR
247 * in private_bio within drbd_endio_pri.
249 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
250 int error = PTR_ERR(req->private_bio);
252 /* remove the request from the conflict detection
253 * respective block_id verification hash */
254 if (!drbd_interval_empty(&req->i)) {
255 struct rb_root *root;
258 root = &mdev->write_requests;
260 root = &mdev->read_requests;
261 drbd_remove_interval(root, &req->i);
263 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
265 /* for writes we need to do some extra housekeeping */
267 _about_to_complete_local_write(mdev, req);
269 /* Update disk stats */
270 _drbd_end_io_acct(mdev, req);
272 m->error = ok ? 0 : (error ?: -EIO);
273 m->bio = req->master_bio;
274 req->master_bio = NULL;
277 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
278 /* this is disconnected (local only) operation,
279 * or protocol C P_WRITE_ACK,
280 * or protocol A or B P_BARRIER_ACK,
281 * or killed from the transfer log due to connection loss. */
282 _req_is_done(mdev, req, rw);
284 /* else: network part and not DONE yet. that is
285 * protocol A or B, barrier ack still pending... */
288 static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
290 struct drbd_conf *mdev = req->mdev;
292 if (!is_susp(mdev->state))
293 _req_may_be_done(req, m);
297 * checks whether there was an overlapping request
298 * or ee already registered.
300 * if so, return 1, in which case this request is completed on the spot,
301 * without ever being submitted or send.
303 * return 0 if it is ok to submit this request.
306 * paranoia: assume something above us is broken, and issues different write
307 * requests for the same block simultaneously...
309 * To ensure these won't be reordered differently on both nodes, resulting in
310 * diverging data sets, we discard the later one(s). Not that this is supposed
311 * to happen, but this is the rationale why we also have to check for
312 * conflicting requests with local origin, and why we have to do so regardless
313 * of whether we allowed multiple primaries.
315 * In case we only have one primary, the epoch_entries tree is empty.
317 static int _req_conflicts(struct drbd_request *req)
319 struct drbd_conf *mdev = req->mdev;
320 const sector_t sector = req->i.sector;
321 const int size = req->i.size;
322 struct drbd_interval *i;
324 D_ASSERT(drbd_interval_empty(&req->i));
326 if (!get_net_conf(mdev))
329 i = drbd_find_overlap(&mdev->write_requests, sector, size);
331 struct drbd_request *req2 =
332 container_of(i, struct drbd_request, i);
334 dev_alert(DEV, "%s[%u] Concurrent local write detected! "
335 "[DISCARD L] new: %llus +%u; "
336 "pending: %llus +%u\n",
337 current->comm, current->pid,
338 (unsigned long long)sector, size,
339 (unsigned long long)req2->i.sector, req2->i.size);
343 if (!RB_EMPTY_ROOT(&mdev->epoch_entries)) {
344 /* check for overlapping requests with remote origin */
345 i = drbd_find_overlap(&mdev->epoch_entries, sector, size);
347 struct drbd_epoch_entry *e =
348 container_of(i, struct drbd_epoch_entry, i);
350 dev_alert(DEV, "%s[%u] Concurrent remote write detected!"
351 " [DISCARD L] new: %llus +%u; "
352 "pending: %llus +%u\n",
353 current->comm, current->pid,
354 (unsigned long long)sector, size,
355 (unsigned long long)e->i.sector, e->i.size);
360 /* this is like it should be, and what we expected.
361 * our users do behave after all... */
370 /* obviously this could be coded as many single functions
371 * instead of one huge switch,
372 * or by putting the code directly in the respective locations
373 * (as it has been before).
375 * but having it this way
376 * enforces that it is all in this one place, where it is easier to audit,
377 * it makes it obvious that whatever "event" "happens" to a request should
378 * happen "atomically" within the req_lock,
379 * and it enforces that we have to think in a very structured manner
380 * about the "events" that may happen to a request during its life time ...
382 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
383 struct bio_and_error *m)
385 struct drbd_conf *mdev = req->mdev;
391 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
394 /* does not happen...
395 * initialization done in drbd_req_new
400 case TO_BE_SENT: /* via network */
401 /* reached via drbd_make_request_common
402 * and from w_read_retry_remote */
403 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
404 req->rq_state |= RQ_NET_PENDING;
405 inc_ap_pending(mdev);
408 case TO_BE_SUBMITTED: /* locally */
409 /* reached via drbd_make_request_common */
410 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
411 req->rq_state |= RQ_LOCAL_PENDING;
415 if (bio_data_dir(req->master_bio) == WRITE)
416 mdev->writ_cnt += req->i.size >> 9;
418 mdev->read_cnt += req->i.size >> 9;
420 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
421 req->rq_state &= ~RQ_LOCAL_PENDING;
423 _req_may_be_done_not_susp(req, m);
427 case WRITE_COMPLETED_WITH_ERROR:
428 req->rq_state |= RQ_LOCAL_COMPLETED;
429 req->rq_state &= ~RQ_LOCAL_PENDING;
431 __drbd_chk_io_error(mdev, false);
432 _req_may_be_done_not_susp(req, m);
436 case READ_AHEAD_COMPLETED_WITH_ERROR:
437 /* it is legal to fail READA */
438 req->rq_state |= RQ_LOCAL_COMPLETED;
439 req->rq_state &= ~RQ_LOCAL_PENDING;
440 _req_may_be_done_not_susp(req, m);
444 case READ_COMPLETED_WITH_ERROR:
445 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
447 req->rq_state |= RQ_LOCAL_COMPLETED;
448 req->rq_state &= ~RQ_LOCAL_PENDING;
450 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
452 __drbd_chk_io_error(mdev, false);
455 /* no point in retrying if there is no good remote data,
456 * or we have no connection. */
457 if (mdev->state.pdsk != D_UP_TO_DATE) {
458 _req_may_be_done_not_susp(req, m);
462 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
463 req->rq_state |= RQ_NET_PENDING;
464 inc_ap_pending(mdev);
465 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
467 case QUEUE_FOR_NET_READ:
468 /* READ or READA, and
470 * or target area marked as invalid,
471 * or just got an io-error. */
472 /* from drbd_make_request_common
473 * or from bio_endio during read io-error recovery */
475 /* so we can verify the handle in the answer packet
476 * corresponding hlist_del is in _req_may_be_done() */
477 drbd_insert_interval(&mdev->read_requests, &req->i);
479 set_bit(UNPLUG_REMOTE, &mdev->flags);
481 D_ASSERT(req->rq_state & RQ_NET_PENDING);
482 req->rq_state |= RQ_NET_QUEUED;
483 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
484 ? w_read_retry_remote
486 drbd_queue_work(&mdev->data.work, &req->w);
489 case QUEUE_FOR_NET_WRITE:
490 /* assert something? */
491 /* from drbd_make_request_common only */
493 /* corresponding hlist_del is in _req_may_be_done() */
494 drbd_insert_interval(&mdev->write_requests, &req->i);
497 * In case the req ended up on the transfer log before being
498 * queued on the worker, it could lead to this request being
499 * missed during cleanup after connection loss.
500 * So we have to do both operations here,
501 * within the same lock that protects the transfer log.
503 * _req_add_to_epoch(req); this has to be after the
504 * _maybe_start_new_epoch(req); which happened in
505 * drbd_make_request_common, because we now may set the bit
506 * again ourselves to close the current epoch.
508 * Add req to the (now) current epoch (barrier). */
510 /* otherwise we may lose an unplug, which may cause some remote
511 * io-scheduler timeout to expire, increasing maximum latency,
512 * hurting performance. */
513 set_bit(UNPLUG_REMOTE, &mdev->flags);
515 /* see drbd_make_request_common,
516 * just after it grabs the req_lock */
517 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
519 req->epoch = mdev->newest_tle->br_number;
521 /* increment size of current epoch */
522 mdev->newest_tle->n_writes++;
524 /* queue work item to send data */
525 D_ASSERT(req->rq_state & RQ_NET_PENDING);
526 req->rq_state |= RQ_NET_QUEUED;
527 req->w.cb = w_send_dblock;
528 drbd_queue_work(&mdev->data.work, &req->w);
530 /* close the epoch, in case it outgrew the limit */
531 if (mdev->newest_tle->n_writes >= mdev->tconn->net_conf->max_epoch_size)
536 case QUEUE_FOR_SEND_OOS:
537 req->rq_state |= RQ_NET_QUEUED;
538 req->w.cb = w_send_oos;
539 drbd_queue_work(&mdev->data.work, &req->w);
542 case OOS_HANDED_TO_NETWORK:
543 /* actually the same */
545 /* treat it the same */
547 /* real cleanup will be done from tl_clear. just update flags
548 * so it is no longer marked as on the worker queue */
549 req->rq_state &= ~RQ_NET_QUEUED;
550 /* if we did it right, tl_clear should be scheduled only after
551 * this, so this should not be necessary! */
552 _req_may_be_done_not_susp(req, m);
555 case HANDED_OVER_TO_NETWORK:
556 /* assert something? */
557 if (bio_data_dir(req->master_bio) == WRITE)
558 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
560 if (bio_data_dir(req->master_bio) == WRITE &&
561 mdev->tconn->net_conf->wire_protocol == DRBD_PROT_A) {
562 /* this is what is dangerous about protocol A:
563 * pretend it was successfully written on the peer. */
564 if (req->rq_state & RQ_NET_PENDING) {
565 dec_ap_pending(mdev);
566 req->rq_state &= ~RQ_NET_PENDING;
567 req->rq_state |= RQ_NET_OK;
568 } /* else: neg-ack was faster... */
569 /* it is still not yet RQ_NET_DONE until the
570 * corresponding epoch barrier got acked as well,
571 * so we know what to dirty on connection loss */
573 req->rq_state &= ~RQ_NET_QUEUED;
574 req->rq_state |= RQ_NET_SENT;
575 /* because _drbd_send_zc_bio could sleep, and may want to
576 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
577 * "COMPLETED_OK" events came in, once we return from
578 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
579 * whether it is done already, and end it. */
580 _req_may_be_done_not_susp(req, m);
583 case READ_RETRY_REMOTE_CANCELED:
584 req->rq_state &= ~RQ_NET_QUEUED;
585 /* fall through, in case we raced with drbd_disconnect */
586 case CONNECTION_LOST_WHILE_PENDING:
587 /* transfer log cleanup after connection loss */
588 /* assert something? */
589 if (req->rq_state & RQ_NET_PENDING)
590 dec_ap_pending(mdev);
591 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
592 req->rq_state |= RQ_NET_DONE;
593 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
594 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
596 /* if it is still queued, we may not complete it here.
597 * it will be canceled soon. */
598 if (!(req->rq_state & RQ_NET_QUEUED))
599 _req_may_be_done(req, m); /* Allowed while state.susp */
602 case WRITE_ACKED_BY_PEER_AND_SIS:
603 req->rq_state |= RQ_NET_SIS;
604 case CONFLICT_DISCARDED_BY_PEER:
605 /* for discarded conflicting writes of multiple primaries,
606 * there is no need to keep anything in the tl, potential
607 * node crashes are covered by the activity log. */
608 if (what == CONFLICT_DISCARDED_BY_PEER)
609 dev_alert(DEV, "Got DiscardAck packet %llus +%u!"
610 " DRBD is not a random data generator!\n",
611 (unsigned long long)req->i.sector, req->i.size);
612 req->rq_state |= RQ_NET_DONE;
614 case WRITE_ACKED_BY_PEER:
615 /* protocol C; successfully written on peer.
616 * Nothing to do here.
617 * We want to keep the tl in place for all protocols, to cater
618 * for volatile write-back caches on lower level devices.
620 * A barrier request is expected to have forced all prior
621 * requests onto stable storage, so completion of a barrier
622 * request could set NET_DONE right here, and not wait for the
623 * P_BARRIER_ACK, but that is an unnecessary optimization. */
625 /* this makes it effectively the same as for: */
626 case RECV_ACKED_BY_PEER:
627 /* protocol B; pretends to be successfully written on peer.
628 * see also notes above in HANDED_OVER_TO_NETWORK about
630 req->rq_state |= RQ_NET_OK;
631 D_ASSERT(req->rq_state & RQ_NET_PENDING);
632 dec_ap_pending(mdev);
633 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
634 req->rq_state &= ~RQ_NET_PENDING;
635 _req_may_be_done_not_susp(req, m);
639 /* assert something? */
640 if (req->rq_state & RQ_NET_PENDING) {
641 dec_ap_pending(mdev);
642 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
644 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
646 req->rq_state |= RQ_NET_DONE;
647 _req_may_be_done_not_susp(req, m);
648 /* else: done by HANDED_OVER_TO_NETWORK */
651 case FAIL_FROZEN_DISK_IO:
652 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
655 _req_may_be_done(req, m); /* Allowed while state.susp */
658 case RESTART_FROZEN_DISK_IO:
659 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
662 req->rq_state &= ~RQ_LOCAL_COMPLETED;
665 if (bio_data_dir(req->master_bio) == WRITE)
669 req->w.cb = w_restart_disk_io;
670 drbd_queue_work(&mdev->data.work, &req->w);
674 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
675 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
676 Trowing them out of the TL here by pretending we got a BARRIER_ACK
677 We ensure that the peer was not rebooted */
678 if (!(req->rq_state & RQ_NET_OK)) {
680 drbd_queue_work(&mdev->data.work, &req->w);
681 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
685 /* else, fall through to BARRIER_ACKED */
688 if (!(req->rq_state & RQ_WRITE))
691 if (req->rq_state & RQ_NET_PENDING) {
692 /* barrier came in before all requests have been acked.
693 * this is bad, because if the connection is lost now,
694 * we won't be able to clean them up... */
695 dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
696 list_move(&req->tl_requests, &mdev->out_of_sequence_requests);
698 if ((req->rq_state & RQ_NET_MASK) != 0) {
699 req->rq_state |= RQ_NET_DONE;
700 if (mdev->tconn->net_conf->wire_protocol == DRBD_PROT_A)
701 atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
703 _req_may_be_done(req, m); /* Allowed while state.susp */
707 D_ASSERT(req->rq_state & RQ_NET_PENDING);
708 dec_ap_pending(mdev);
709 req->rq_state &= ~RQ_NET_PENDING;
710 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
711 _req_may_be_done_not_susp(req, m);
718 /* we may do a local read if:
719 * - we are consistent (of course),
720 * - or we are generally inconsistent,
721 * BUT we are still/already IN SYNC for this area.
722 * since size may be bigger than BM_BLOCK_SIZE,
723 * we may need to check several bits.
725 static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
727 unsigned long sbnr, ebnr;
728 sector_t esector, nr_sectors;
730 if (mdev->state.disk == D_UP_TO_DATE)
732 if (mdev->state.disk >= D_OUTDATED)
734 if (mdev->state.disk < D_INCONSISTENT)
736 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
737 nr_sectors = drbd_get_capacity(mdev->this_bdev);
738 esector = sector + (size >> 9) - 1;
740 D_ASSERT(sector < nr_sectors);
741 D_ASSERT(esector < nr_sectors);
743 sbnr = BM_SECT_TO_BIT(sector);
744 ebnr = BM_SECT_TO_BIT(esector);
746 return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr);
749 static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
751 const int rw = bio_rw(bio);
752 const int size = bio->bi_size;
753 const sector_t sector = bio->bi_sector;
754 struct drbd_tl_epoch *b = NULL;
755 struct drbd_request *req;
756 int local, remote, send_oos = 0;
760 /* allocate outside of all locks; */
761 req = drbd_req_new(mdev, bio);
764 /* only pass the error to the upper layers.
765 * if user cannot handle io errors, that's not our business. */
766 dev_err(DEV, "could not kmalloc() req\n");
767 bio_endio(bio, -ENOMEM);
770 req->start_time = start_time;
772 local = get_ldev(mdev);
774 bio_put(req->private_bio); /* or we get a bio leak */
775 req->private_bio = NULL;
782 if (!drbd_may_do_local_read(mdev, sector, size)) {
783 /* we could kick the syncer to
784 * sync this extent asap, wait for
785 * it, then continue locally.
786 * Or just issue the request remotely.
789 bio_put(req->private_bio);
790 req->private_bio = NULL;
794 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
797 /* If we have a disk, but a READA request is mapped to remote,
798 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
799 * Just fail that READA request right here.
801 * THINK: maybe fail all READA when not local?
802 * or make this configurable...
803 * if network is slow, READA won't do any good.
805 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
807 goto fail_and_free_req;
810 /* For WRITES going to the local disk, grab a reference on the target
811 * extent. This waits for any resync activity in the corresponding
812 * resync extent to finish, and, if necessary, pulls in the target
813 * extent into the activity log, which involves further disk io because
814 * of transactional on-disk meta data updates. */
815 if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
816 req->rq_state |= RQ_IN_ACT_LOG;
817 drbd_al_begin_io(mdev, sector);
820 remote = remote && drbd_should_do_remote(mdev->state);
821 send_oos = rw == WRITE && drbd_should_send_oos(mdev->state);
822 D_ASSERT(!(remote && send_oos));
824 if (!(local || remote) && !is_susp(mdev->state)) {
825 if (__ratelimit(&drbd_ratelimit_state))
826 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
827 goto fail_free_complete;
830 /* For WRITE request, we have to make sure that we have an
831 * unused_spare_tle, in case we need to start a new epoch.
832 * I try to be smart and avoid to pre-allocate always "just in case",
833 * but there is a race between testing the bit and pointer outside the
834 * spinlock, and grabbing the spinlock.
835 * if we lost that race, we retry. */
836 if (rw == WRITE && (remote || send_oos) &&
837 mdev->unused_spare_tle == NULL &&
838 test_bit(CREATE_BARRIER, &mdev->flags)) {
840 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
842 dev_err(DEV, "Failed to alloc barrier.\n");
844 goto fail_free_complete;
848 /* GOOD, everything prepared, grab the spin_lock */
849 spin_lock_irq(&mdev->req_lock);
851 if (is_susp(mdev->state)) {
852 /* If we got suspended, use the retry mechanism of
853 generic_make_request() to restart processing of this
854 bio. In the next call to drbd_make_request
855 we sleep in inc_ap_bio() */
857 spin_unlock_irq(&mdev->req_lock);
858 goto fail_free_complete;
861 if (remote || send_oos) {
862 remote = drbd_should_do_remote(mdev->state);
863 send_oos = rw == WRITE && drbd_should_send_oos(mdev->state);
864 D_ASSERT(!(remote && send_oos));
866 if (!(remote || send_oos))
867 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
868 if (!(local || remote)) {
869 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
870 spin_unlock_irq(&mdev->req_lock);
871 goto fail_free_complete;
875 if (b && mdev->unused_spare_tle == NULL) {
876 mdev->unused_spare_tle = b;
879 if (rw == WRITE && (remote || send_oos) &&
880 mdev->unused_spare_tle == NULL &&
881 test_bit(CREATE_BARRIER, &mdev->flags)) {
882 /* someone closed the current epoch
883 * while we were grabbing the spinlock */
884 spin_unlock_irq(&mdev->req_lock);
885 goto allocate_barrier;
889 /* Update disk stats */
890 _drbd_start_io_acct(mdev, req, bio);
892 /* _maybe_start_new_epoch(mdev);
893 * If we need to generate a write barrier packet, we have to add the
894 * new epoch (barrier) object, and queue the barrier packet for sending,
895 * and queue the req's data after it _within the same lock_, otherwise
896 * we have race conditions were the reorder domains could be mixed up.
898 * Even read requests may start a new epoch and queue the corresponding
899 * barrier packet. To get the write ordering right, we only have to
900 * make sure that, if this is a write request and it triggered a
901 * barrier packet, this request is queued within the same spinlock. */
902 if ((remote || send_oos) && mdev->unused_spare_tle &&
903 test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
904 _tl_add_barrier(mdev, mdev->unused_spare_tle);
905 mdev->unused_spare_tle = NULL;
907 D_ASSERT(!(remote && rw == WRITE &&
908 test_bit(CREATE_BARRIER, &mdev->flags)));
912 * Actually, 'local' may be wrong here already, since we may have failed
913 * to write to the meta data, and may become wrong anytime because of
914 * local io-error for some other request, which would lead to us
915 * "detaching" the local disk.
917 * 'remote' may become wrong any time because the network could fail.
919 * This is a harmless race condition, though, since it is handled
920 * correctly at the appropriate places; so it just defers the failure
921 * of the respective operation.
924 /* mark them early for readability.
925 * this just sets some state flags. */
927 _req_mod(req, TO_BE_SENT);
929 _req_mod(req, TO_BE_SUBMITTED);
931 /* check this request on the collision detection hash tables.
932 * if we have a conflict, just complete it here.
933 * THINK do we want to check reads, too? (I don't think so...) */
934 if (rw == WRITE && _req_conflicts(req))
935 goto fail_conflicting;
937 list_add_tail(&req->tl_requests, &mdev->newest_tle->requests);
939 /* NOTE remote first: to get the concurrent write detection right,
940 * we must register the request before start of local IO. */
942 /* either WRITE and C_CONNECTED,
943 * or READ, and no local disk,
944 * or READ, but not in sync.
946 _req_mod(req, (rw == WRITE)
947 ? QUEUE_FOR_NET_WRITE
948 : QUEUE_FOR_NET_READ);
950 if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
951 _req_mod(req, QUEUE_FOR_SEND_OOS);
954 mdev->tconn->net_conf->on_congestion != OC_BLOCK && mdev->agreed_pro_version >= 96) {
957 if (mdev->tconn->net_conf->cong_fill &&
958 atomic_read(&mdev->ap_in_flight) >= mdev->tconn->net_conf->cong_fill) {
959 dev_info(DEV, "Congestion-fill threshold reached\n");
963 if (mdev->act_log->used >= mdev->tconn->net_conf->cong_extents) {
964 dev_info(DEV, "Congestion-extents threshold reached\n");
969 queue_barrier(mdev); /* last barrier, after mirrored writes */
971 if (mdev->tconn->net_conf->on_congestion == OC_PULL_AHEAD)
972 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
973 else /*mdev->tconn->net_conf->on_congestion == OC_DISCONNECT */
974 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
978 spin_unlock_irq(&mdev->req_lock);
979 kfree(b); /* if someone else has beaten us to it... */
982 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
984 /* State may have changed since we grabbed our reference on the
985 * mdev->ldev member. Double check, and short-circuit to endio.
986 * In case the last activity log transaction failed to get on
987 * stable storage, and this is a WRITE, we may not even submit
989 if (get_ldev(mdev)) {
990 if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
991 : rw == READ ? DRBD_FAULT_DT_RD
993 bio_endio(req->private_bio, -EIO);
995 generic_make_request(req->private_bio);
998 bio_endio(req->private_bio, -EIO);
1004 /* this is a conflicting request.
1005 * even though it may have been only _partially_
1006 * overlapping with one of the currently pending requests,
1007 * without even submitting or sending it, we will
1008 * pretend that it was successfully served right now.
1010 _drbd_end_io_acct(mdev, req);
1011 spin_unlock_irq(&mdev->req_lock);
1013 dec_ap_pending(mdev);
1014 /* THINK: do we want to fail it (-EIO), or pretend success?
1015 * this pretends success. */
1019 if (req->rq_state & RQ_IN_ACT_LOG)
1020 drbd_al_complete_io(mdev, sector);
1023 bio_put(req->private_bio);
1024 req->private_bio = NULL;
1028 bio_endio(bio, err);
1037 /* helper function for drbd_make_request
1038 * if we can determine just by the mdev (state) that this request will fail,
1040 * otherwise return 0
1042 static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write)
1044 if (mdev->state.role != R_PRIMARY &&
1045 (!allow_oos || is_write)) {
1046 if (__ratelimit(&drbd_ratelimit_state)) {
1047 dev_err(DEV, "Process %s[%u] tried to %s; "
1048 "since we are not in Primary state, "
1049 "we cannot allow this\n",
1050 current->comm, current->pid,
1051 is_write ? "WRITE" : "READ");
1059 int drbd_make_request(struct request_queue *q, struct bio *bio)
1061 unsigned int s_enr, e_enr;
1062 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1063 unsigned long start_time;
1065 if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) {
1066 bio_endio(bio, -EPERM);
1070 start_time = jiffies;
1073 * what we "blindly" assume:
1075 D_ASSERT(bio->bi_size > 0);
1076 D_ASSERT((bio->bi_size & 0x1ff) == 0);
1077 D_ASSERT(bio->bi_idx == 0);
1079 /* to make some things easier, force alignment of requests within the
1080 * granularity of our hash tables */
1081 s_enr = bio->bi_sector >> HT_SHIFT;
1082 e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT;
1084 if (likely(s_enr == e_enr)) {
1085 inc_ap_bio(mdev, 1);
1086 return drbd_make_request_common(mdev, bio, start_time);
1089 /* can this bio be split generically?
1090 * Maybe add our own split-arbitrary-bios function. */
1091 if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_BIO_SIZE) {
1092 /* rather error out here than BUG in bio_split */
1093 dev_err(DEV, "bio would need to, but cannot, be split: "
1094 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1095 bio->bi_vcnt, bio->bi_idx, bio->bi_size,
1096 (unsigned long long)bio->bi_sector);
1097 bio_endio(bio, -EINVAL);
1099 /* This bio crosses some boundary, so we have to split it. */
1100 struct bio_pair *bp;
1101 /* works for the "do not cross hash slot boundaries" case
1102 * e.g. sector 262269, size 4096
1103 * s_enr = 262269 >> 6 = 4097
1104 * e_enr = (262269+8-1) >> 6 = 4098
1106 * sps = 64, mask = 63
1107 * first_sectors = 64 - (262269 & 63) = 3
1109 const sector_t sect = bio->bi_sector;
1110 const int sps = 1 << HT_SHIFT; /* sectors per slot */
1111 const int mask = sps - 1;
1112 const sector_t first_sectors = sps - (sect & mask);
1113 bp = bio_split(bio, first_sectors);
1115 /* we need to get a "reference count" (ap_bio_cnt)
1116 * to avoid races with the disconnect/reconnect/suspend code.
1117 * In case we need to split the bio here, we need to get three references
1118 * atomically, otherwise we might deadlock when trying to submit the
1120 inc_ap_bio(mdev, 3);
1122 D_ASSERT(e_enr == s_enr + 1);
1124 while (drbd_make_request_common(mdev, &bp->bio1, start_time))
1125 inc_ap_bio(mdev, 1);
1127 while (drbd_make_request_common(mdev, &bp->bio2, start_time))
1128 inc_ap_bio(mdev, 1);
1132 bio_pair_release(bp);
1137 /* This is called by bio_add_page(). With this function we reduce
1138 * the number of BIOs that span over multiple DRBD_MAX_BIO_SIZEs
1139 * units (was AL_EXTENTs).
1141 * we do the calculation within the lower 32bit of the byte offsets,
1142 * since we don't care for actual offset, but only check whether it
1143 * would cross "activity log extent" boundaries.
1145 * As long as the BIO is empty we have to allow at least one bvec,
1146 * regardless of size and offset. so the resulting bio may still
1147 * cross extent boundaries. those are dealt with (bio_split) in
1148 * drbd_make_request.
1150 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1152 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1153 unsigned int bio_offset =
1154 (unsigned int)bvm->bi_sector << 9; /* 32 bit */
1155 unsigned int bio_size = bvm->bi_size;
1156 int limit, backing_limit;
1158 limit = DRBD_MAX_BIO_SIZE
1159 - ((bio_offset & (DRBD_MAX_BIO_SIZE-1)) + bio_size);
1162 if (bio_size == 0) {
1163 if (limit <= bvec->bv_len)
1164 limit = bvec->bv_len;
1165 } else if (limit && get_ldev(mdev)) {
1166 struct request_queue * const b =
1167 mdev->ldev->backing_bdev->bd_disk->queue;
1168 if (b->merge_bvec_fn) {
1169 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1170 limit = min(limit, backing_limit);
1177 void request_timer_fn(unsigned long data)
1179 struct drbd_conf *mdev = (struct drbd_conf *) data;
1180 struct drbd_request *req; /* oldest request */
1181 struct list_head *le;
1182 unsigned long et = 0; /* effective timeout = ko_count * timeout */
1184 if (get_net_conf(mdev)) {
1185 et = mdev->tconn->net_conf->timeout*HZ/10 * mdev->tconn->net_conf->ko_count;
1188 if (!et || mdev->state.conn < C_WF_REPORT_PARAMS)
1189 return; /* Recurring timer stopped */
1191 spin_lock_irq(&mdev->req_lock);
1192 le = &mdev->oldest_tle->requests;
1193 if (list_empty(le)) {
1194 spin_unlock_irq(&mdev->req_lock);
1195 mod_timer(&mdev->request_timer, jiffies + et);
1200 req = list_entry(le, struct drbd_request, tl_requests);
1201 if (time_is_before_eq_jiffies(req->start_time + et)) {
1202 if (req->rq_state & RQ_NET_PENDING) {
1203 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1204 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE, NULL);
1206 dev_warn(DEV, "Local backing block device frozen?\n");
1207 mod_timer(&mdev->request_timer, jiffies + et);
1210 mod_timer(&mdev->request_timer, req->start_time + et);
1213 spin_unlock_irq(&mdev->req_lock);