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;
138 struct drbd_epoch_entry *e;
139 struct hlist_node *n;
140 struct hlist_head *slot;
142 /* Before we can signal completion to the upper layers,
143 * we may need to close the current epoch.
144 * We can skip this, if this request has not even been sent, because we
145 * did not have a fully established connection yet/anymore, during
146 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
148 if (mdev->state.conn >= C_CONNECTED &&
149 (s & RQ_NET_SENT) != 0 &&
150 req->epoch == mdev->newest_tle->br_number)
153 /* we need to do the conflict detection stuff,
154 * if we have the ee_hash (two_primaries) and
155 * this has been on the network */
156 if ((s & RQ_NET_DONE) && mdev->ee_hash != NULL) {
157 const sector_t sector = req->i.sector;
158 const int size = req->i.size;
159 struct drbd_interval *i;
162 * there must be no conflicting requests, since
163 * they must have been failed on the spot */
165 i = drbd_find_overlap(&mdev->write_requests, sector, size);
167 struct drbd_request *req2 =
168 container_of(i, struct drbd_request, i);
170 dev_alert(DEV, "LOGIC BUG: completed: %p %llus +%u; "
171 "other: %p %llus +%u\n",
172 req, (unsigned long long)sector, size,
173 i, (unsigned long long)req2->i.sector, req2->i.size);
176 /* maybe "wake" those conflicting epoch entries
177 * that wait for this request to finish.
179 * currently, there can be only _one_ such ee
180 * (well, or some more, which would be pending
181 * P_DISCARD_ACK not yet sent by the asender...),
182 * since we block the receiver thread upon the
183 * first conflict detection, which will wait on
184 * misc_wait. maybe we want to assert that?
186 * anyways, if we found one,
187 * we just have to do a wake_up. */
188 #define OVERLAPS overlaps(sector, size, e->sector, e->size)
189 slot = ee_hash_slot(mdev, req->i.sector);
190 hlist_for_each_entry(e, n, slot, collision) {
192 wake_up(&mdev->misc_wait);
200 void complete_master_bio(struct drbd_conf *mdev,
201 struct bio_and_error *m)
203 bio_endio(m->bio, m->error);
207 /* Helper for __req_mod().
208 * Set m->bio to the master bio, if it is fit to be completed,
209 * or leave it alone (it is initialized to NULL in __req_mod),
210 * if it has already been completed, or cannot be completed yet.
211 * If m->bio is set, the error status to be returned is placed in m->error.
213 void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
215 const unsigned long s = req->rq_state;
216 struct drbd_conf *mdev = req->mdev;
217 /* only WRITES may end up here without a master bio (on barrier ack) */
218 int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE;
220 /* we must not complete the master bio, while it is
221 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
222 * not yet acknowledged by the peer
223 * not yet completed by the local io subsystem
224 * these flags may get cleared in any order by
227 * the bio_endio completion callbacks.
229 if (s & RQ_NET_QUEUED)
231 if (s & RQ_NET_PENDING)
233 if (s & RQ_LOCAL_PENDING)
236 if (req->master_bio) {
237 /* this is data_received (remote read)
238 * or protocol C P_WRITE_ACK
239 * or protocol B P_RECV_ACK
240 * or protocol A "handed_over_to_network" (SendAck)
241 * or canceled or failed,
242 * or killed from the transfer log due to connection loss.
246 * figure out whether to report success or failure.
248 * report success when at least one of the operations succeeded.
249 * or, to put the other way,
250 * only report failure, when both operations failed.
252 * what to do about the failures is handled elsewhere.
253 * what we need to do here is just: complete the master_bio.
255 * local completion error, if any, has been stored as ERR_PTR
256 * in private_bio within drbd_endio_pri.
258 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
259 int error = PTR_ERR(req->private_bio);
261 /* remove the request from the conflict detection
262 * respective block_id verification hash */
263 if (!drbd_interval_empty(&req->i)) {
264 struct rb_root *root;
266 hlist_del(&req->collision);
268 root = &mdev->write_requests;
270 root = &mdev->read_requests;
271 drbd_remove_interval(root, &req->i);
273 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
275 /* for writes we need to do some extra housekeeping */
277 _about_to_complete_local_write(mdev, req);
279 /* Update disk stats */
280 _drbd_end_io_acct(mdev, req);
282 m->error = ok ? 0 : (error ?: -EIO);
283 m->bio = req->master_bio;
284 req->master_bio = NULL;
287 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
288 /* this is disconnected (local only) operation,
289 * or protocol C P_WRITE_ACK,
290 * or protocol A or B P_BARRIER_ACK,
291 * or killed from the transfer log due to connection loss. */
292 _req_is_done(mdev, req, rw);
294 /* else: network part and not DONE yet. that is
295 * protocol A or B, barrier ack still pending... */
298 static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
300 struct drbd_conf *mdev = req->mdev;
302 if (!is_susp(mdev->state))
303 _req_may_be_done(req, m);
307 * checks whether there was an overlapping request
308 * or ee already registered.
310 * if so, return 1, in which case this request is completed on the spot,
311 * without ever being submitted or send.
313 * return 0 if it is ok to submit this request.
316 * paranoia: assume something above us is broken, and issues different write
317 * requests for the same block simultaneously...
319 * To ensure these won't be reordered differently on both nodes, resulting in
320 * diverging data sets, we discard the later one(s). Not that this is supposed
321 * to happen, but this is the rationale why we also have to check for
322 * conflicting requests with local origin, and why we have to do so regardless
323 * of whether we allowed multiple primaries.
325 * BTW, in case we only have one primary, the ee_hash is empty anyways, and the
326 * second hlist_for_each_entry becomes a noop. This is even simpler than to
327 * grab a reference on the net_conf, and check for the two_primaries flag...
329 static int _req_conflicts(struct drbd_request *req)
331 struct drbd_conf *mdev = req->mdev;
332 const sector_t sector = req->i.sector;
333 const int size = req->i.size;
334 struct drbd_interval *i;
335 struct drbd_epoch_entry *e;
336 struct hlist_node *n;
337 struct hlist_head *slot;
339 D_ASSERT(hlist_unhashed(&req->collision));
340 D_ASSERT(drbd_interval_empty(&req->i));
342 if (!get_net_conf(mdev))
346 ERR_IF (mdev->tl_hash_s == 0)
347 goto out_no_conflict;
348 BUG_ON(mdev->tl_hash == NULL);
350 i = drbd_find_overlap(&mdev->write_requests, sector, size);
352 struct drbd_request *req2 =
353 container_of(i, struct drbd_request, i);
355 dev_alert(DEV, "%s[%u] Concurrent local write detected! "
356 "[DISCARD L] new: %llus +%u; "
357 "pending: %llus +%u\n",
358 current->comm, current->pid,
359 (unsigned long long)sector, size,
360 (unsigned long long)req2->i.sector, req2->i.size);
364 if (mdev->ee_hash_s) {
365 /* now, check for overlapping requests with remote origin */
366 BUG_ON(mdev->ee_hash == NULL);
367 #define OVERLAPS overlaps(e->sector, e->size, sector, size)
368 slot = ee_hash_slot(mdev, sector);
369 hlist_for_each_entry(e, n, slot, collision) {
371 dev_alert(DEV, "%s[%u] Concurrent remote write detected!"
372 " [DISCARD L] new: %llus +%u; "
373 "pending: %llus +%u\n",
374 current->comm, current->pid,
375 (unsigned long long)sector, size,
376 (unsigned long long)e->sector, e->size);
384 /* this is like it should be, and what we expected.
385 * our users do behave after all... */
394 /* obviously this could be coded as many single functions
395 * instead of one huge switch,
396 * or by putting the code directly in the respective locations
397 * (as it has been before).
399 * but having it this way
400 * enforces that it is all in this one place, where it is easier to audit,
401 * it makes it obvious that whatever "event" "happens" to a request should
402 * happen "atomically" within the req_lock,
403 * and it enforces that we have to think in a very structured manner
404 * about the "events" that may happen to a request during its life time ...
406 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
407 struct bio_and_error *m)
409 struct drbd_conf *mdev = req->mdev;
415 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
418 /* does not happen...
419 * initialization done in drbd_req_new
424 case to_be_send: /* via network */
425 /* reached via drbd_make_request_common
426 * and from w_read_retry_remote */
427 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
428 req->rq_state |= RQ_NET_PENDING;
429 inc_ap_pending(mdev);
432 case to_be_submitted: /* locally */
433 /* reached via drbd_make_request_common */
434 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
435 req->rq_state |= RQ_LOCAL_PENDING;
439 if (bio_data_dir(req->master_bio) == WRITE)
440 mdev->writ_cnt += req->i.size >> 9;
442 mdev->read_cnt += req->i.size >> 9;
444 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
445 req->rq_state &= ~RQ_LOCAL_PENDING;
447 _req_may_be_done_not_susp(req, m);
451 case write_completed_with_error:
452 req->rq_state |= RQ_LOCAL_COMPLETED;
453 req->rq_state &= ~RQ_LOCAL_PENDING;
455 __drbd_chk_io_error(mdev, false);
456 _req_may_be_done_not_susp(req, m);
460 case read_ahead_completed_with_error:
461 /* it is legal to fail READA */
462 req->rq_state |= RQ_LOCAL_COMPLETED;
463 req->rq_state &= ~RQ_LOCAL_PENDING;
464 _req_may_be_done_not_susp(req, m);
468 case read_completed_with_error:
469 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
471 req->rq_state |= RQ_LOCAL_COMPLETED;
472 req->rq_state &= ~RQ_LOCAL_PENDING;
474 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
476 __drbd_chk_io_error(mdev, false);
479 /* no point in retrying if there is no good remote data,
480 * or we have no connection. */
481 if (mdev->state.pdsk != D_UP_TO_DATE) {
482 _req_may_be_done_not_susp(req, m);
486 /* _req_mod(req,to_be_send); oops, recursion... */
487 req->rq_state |= RQ_NET_PENDING;
488 inc_ap_pending(mdev);
489 /* fall through: _req_mod(req,queue_for_net_read); */
491 case queue_for_net_read:
492 /* READ or READA, and
494 * or target area marked as invalid,
495 * or just got an io-error. */
496 /* from drbd_make_request_common
497 * or from bio_endio during read io-error recovery */
499 /* so we can verify the handle in the answer packet
500 * corresponding hlist_del is in _req_may_be_done() */
501 hlist_add_head(&req->collision, ar_hash_slot(mdev, req->i.sector));
502 drbd_insert_interval(&mdev->read_requests, &req->i);
504 set_bit(UNPLUG_REMOTE, &mdev->flags);
506 D_ASSERT(req->rq_state & RQ_NET_PENDING);
507 req->rq_state |= RQ_NET_QUEUED;
508 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
509 ? w_read_retry_remote
511 drbd_queue_work(&mdev->data.work, &req->w);
514 case queue_for_net_write:
515 /* assert something? */
516 /* from drbd_make_request_common only */
518 hlist_add_head(&req->collision, tl_hash_slot(mdev, req->i.sector));
519 /* corresponding hlist_del is in _req_may_be_done() */
520 drbd_insert_interval(&mdev->write_requests, &req->i);
523 * In case the req ended up on the transfer log before being
524 * queued on the worker, it could lead to this request being
525 * missed during cleanup after connection loss.
526 * So we have to do both operations here,
527 * within the same lock that protects the transfer log.
529 * _req_add_to_epoch(req); this has to be after the
530 * _maybe_start_new_epoch(req); which happened in
531 * drbd_make_request_common, because we now may set the bit
532 * again ourselves to close the current epoch.
534 * Add req to the (now) current epoch (barrier). */
536 /* otherwise we may lose an unplug, which may cause some remote
537 * io-scheduler timeout to expire, increasing maximum latency,
538 * hurting performance. */
539 set_bit(UNPLUG_REMOTE, &mdev->flags);
541 /* see drbd_make_request_common,
542 * just after it grabs the req_lock */
543 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
545 req->epoch = mdev->newest_tle->br_number;
547 /* increment size of current epoch */
548 mdev->newest_tle->n_writes++;
550 /* queue work item to send data */
551 D_ASSERT(req->rq_state & RQ_NET_PENDING);
552 req->rq_state |= RQ_NET_QUEUED;
553 req->w.cb = w_send_dblock;
554 drbd_queue_work(&mdev->data.work, &req->w);
556 /* close the epoch, in case it outgrew the limit */
557 if (mdev->newest_tle->n_writes >= mdev->net_conf->max_epoch_size)
562 case queue_for_send_oos:
563 req->rq_state |= RQ_NET_QUEUED;
564 req->w.cb = w_send_oos;
565 drbd_queue_work(&mdev->data.work, &req->w);
568 case oos_handed_to_network:
569 /* actually the same */
571 /* treat it the same */
573 /* real cleanup will be done from tl_clear. just update flags
574 * so it is no longer marked as on the worker queue */
575 req->rq_state &= ~RQ_NET_QUEUED;
576 /* if we did it right, tl_clear should be scheduled only after
577 * this, so this should not be necessary! */
578 _req_may_be_done_not_susp(req, m);
581 case handed_over_to_network:
582 /* assert something? */
583 if (bio_data_dir(req->master_bio) == WRITE)
584 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
586 if (bio_data_dir(req->master_bio) == WRITE &&
587 mdev->net_conf->wire_protocol == DRBD_PROT_A) {
588 /* this is what is dangerous about protocol A:
589 * pretend it was successfully written on the peer. */
590 if (req->rq_state & RQ_NET_PENDING) {
591 dec_ap_pending(mdev);
592 req->rq_state &= ~RQ_NET_PENDING;
593 req->rq_state |= RQ_NET_OK;
594 } /* else: neg-ack was faster... */
595 /* it is still not yet RQ_NET_DONE until the
596 * corresponding epoch barrier got acked as well,
597 * so we know what to dirty on connection loss */
599 req->rq_state &= ~RQ_NET_QUEUED;
600 req->rq_state |= RQ_NET_SENT;
601 /* because _drbd_send_zc_bio could sleep, and may want to
602 * dereference the bio even after the "write_acked_by_peer" and
603 * "completed_ok" events came in, once we return from
604 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
605 * whether it is done already, and end it. */
606 _req_may_be_done_not_susp(req, m);
609 case read_retry_remote_canceled:
610 req->rq_state &= ~RQ_NET_QUEUED;
611 /* fall through, in case we raced with drbd_disconnect */
612 case connection_lost_while_pending:
613 /* transfer log cleanup after connection loss */
614 /* assert something? */
615 if (req->rq_state & RQ_NET_PENDING)
616 dec_ap_pending(mdev);
617 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
618 req->rq_state |= RQ_NET_DONE;
619 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
620 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
622 /* if it is still queued, we may not complete it here.
623 * it will be canceled soon. */
624 if (!(req->rq_state & RQ_NET_QUEUED))
625 _req_may_be_done(req, m); /* Allowed while state.susp */
628 case write_acked_by_peer_and_sis:
629 req->rq_state |= RQ_NET_SIS;
630 case conflict_discarded_by_peer:
631 /* for discarded conflicting writes of multiple primaries,
632 * there is no need to keep anything in the tl, potential
633 * node crashes are covered by the activity log. */
634 if (what == conflict_discarded_by_peer)
635 dev_alert(DEV, "Got DiscardAck packet %llus +%u!"
636 " DRBD is not a random data generator!\n",
637 (unsigned long long)req->i.sector, req->i.size);
638 req->rq_state |= RQ_NET_DONE;
640 case write_acked_by_peer:
641 /* protocol C; successfully written on peer.
642 * Nothing to do here.
643 * We want to keep the tl in place for all protocols, to cater
644 * for volatile write-back caches on lower level devices.
646 * A barrier request is expected to have forced all prior
647 * requests onto stable storage, so completion of a barrier
648 * request could set NET_DONE right here, and not wait for the
649 * P_BARRIER_ACK, but that is an unnecessary optimization. */
651 /* this makes it effectively the same as for: */
652 case recv_acked_by_peer:
653 /* protocol B; pretends to be successfully written on peer.
654 * see also notes above in handed_over_to_network about
656 req->rq_state |= RQ_NET_OK;
657 D_ASSERT(req->rq_state & RQ_NET_PENDING);
658 dec_ap_pending(mdev);
659 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
660 req->rq_state &= ~RQ_NET_PENDING;
661 _req_may_be_done_not_susp(req, m);
665 /* assert something? */
666 if (req->rq_state & RQ_NET_PENDING) {
667 dec_ap_pending(mdev);
668 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
670 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
672 req->rq_state |= RQ_NET_DONE;
673 _req_may_be_done_not_susp(req, m);
674 /* else: done by handed_over_to_network */
677 case fail_frozen_disk_io:
678 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
681 _req_may_be_done(req, m); /* Allowed while state.susp */
684 case restart_frozen_disk_io:
685 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
688 req->rq_state &= ~RQ_LOCAL_COMPLETED;
691 if (bio_data_dir(req->master_bio) == WRITE)
695 req->w.cb = w_restart_disk_io;
696 drbd_queue_work(&mdev->data.work, &req->w);
700 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
701 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
702 Trowing them out of the TL here by pretending we got a BARRIER_ACK
703 We ensure that the peer was not rebooted */
704 if (!(req->rq_state & RQ_NET_OK)) {
706 drbd_queue_work(&mdev->data.work, &req->w);
707 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
711 /* else, fall through to barrier_acked */
714 if (!(req->rq_state & RQ_WRITE))
717 if (req->rq_state & RQ_NET_PENDING) {
718 /* barrier came in before all requests have been acked.
719 * this is bad, because if the connection is lost now,
720 * we won't be able to clean them up... */
721 dev_err(DEV, "FIXME (barrier_acked but pending)\n");
722 list_move(&req->tl_requests, &mdev->out_of_sequence_requests);
724 if ((req->rq_state & RQ_NET_MASK) != 0) {
725 req->rq_state |= RQ_NET_DONE;
726 if (mdev->net_conf->wire_protocol == DRBD_PROT_A)
727 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
729 _req_may_be_done(req, m); /* Allowed while state.susp */
733 D_ASSERT(req->rq_state & RQ_NET_PENDING);
734 dec_ap_pending(mdev);
735 req->rq_state &= ~RQ_NET_PENDING;
736 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
737 _req_may_be_done_not_susp(req, m);
744 /* we may do a local read if:
745 * - we are consistent (of course),
746 * - or we are generally inconsistent,
747 * BUT we are still/already IN SYNC for this area.
748 * since size may be bigger than BM_BLOCK_SIZE,
749 * we may need to check several bits.
751 static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
753 unsigned long sbnr, ebnr;
754 sector_t esector, nr_sectors;
756 if (mdev->state.disk == D_UP_TO_DATE)
758 if (mdev->state.disk >= D_OUTDATED)
760 if (mdev->state.disk < D_INCONSISTENT)
762 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
763 nr_sectors = drbd_get_capacity(mdev->this_bdev);
764 esector = sector + (size >> 9) - 1;
766 D_ASSERT(sector < nr_sectors);
767 D_ASSERT(esector < nr_sectors);
769 sbnr = BM_SECT_TO_BIT(sector);
770 ebnr = BM_SECT_TO_BIT(esector);
772 return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr);
775 static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
777 const int rw = bio_rw(bio);
778 const int size = bio->bi_size;
779 const sector_t sector = bio->bi_sector;
780 struct drbd_tl_epoch *b = NULL;
781 struct drbd_request *req;
782 int local, remote, send_oos = 0;
786 /* allocate outside of all locks; */
787 req = drbd_req_new(mdev, bio);
790 /* only pass the error to the upper layers.
791 * if user cannot handle io errors, that's not our business. */
792 dev_err(DEV, "could not kmalloc() req\n");
793 bio_endio(bio, -ENOMEM);
796 req->start_time = start_time;
798 local = get_ldev(mdev);
800 bio_put(req->private_bio); /* or we get a bio leak */
801 req->private_bio = NULL;
808 if (!drbd_may_do_local_read(mdev, sector, size)) {
809 /* we could kick the syncer to
810 * sync this extent asap, wait for
811 * it, then continue locally.
812 * Or just issue the request remotely.
815 bio_put(req->private_bio);
816 req->private_bio = NULL;
820 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
823 /* If we have a disk, but a READA request is mapped to remote,
824 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
825 * Just fail that READA request right here.
827 * THINK: maybe fail all READA when not local?
828 * or make this configurable...
829 * if network is slow, READA won't do any good.
831 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
833 goto fail_and_free_req;
836 /* For WRITES going to the local disk, grab a reference on the target
837 * extent. This waits for any resync activity in the corresponding
838 * resync extent to finish, and, if necessary, pulls in the target
839 * extent into the activity log, which involves further disk io because
840 * of transactional on-disk meta data updates. */
841 if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
842 req->rq_state |= RQ_IN_ACT_LOG;
843 drbd_al_begin_io(mdev, sector);
846 remote = remote && drbd_should_do_remote(mdev->state);
847 send_oos = rw == WRITE && drbd_should_send_oos(mdev->state);
848 D_ASSERT(!(remote && send_oos));
850 if (!(local || remote) && !is_susp(mdev->state)) {
851 if (__ratelimit(&drbd_ratelimit_state))
852 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
853 goto fail_free_complete;
856 /* For WRITE request, we have to make sure that we have an
857 * unused_spare_tle, in case we need to start a new epoch.
858 * I try to be smart and avoid to pre-allocate always "just in case",
859 * but there is a race between testing the bit and pointer outside the
860 * spinlock, and grabbing the spinlock.
861 * if we lost that race, we retry. */
862 if (rw == WRITE && (remote || send_oos) &&
863 mdev->unused_spare_tle == NULL &&
864 test_bit(CREATE_BARRIER, &mdev->flags)) {
866 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
868 dev_err(DEV, "Failed to alloc barrier.\n");
870 goto fail_free_complete;
874 /* GOOD, everything prepared, grab the spin_lock */
875 spin_lock_irq(&mdev->req_lock);
877 if (is_susp(mdev->state)) {
878 /* If we got suspended, use the retry mechanism of
879 generic_make_request() to restart processing of this
880 bio. In the next call to drbd_make_request
881 we sleep in inc_ap_bio() */
883 spin_unlock_irq(&mdev->req_lock);
884 goto fail_free_complete;
887 if (remote || send_oos) {
888 remote = drbd_should_do_remote(mdev->state);
889 send_oos = rw == WRITE && drbd_should_send_oos(mdev->state);
890 D_ASSERT(!(remote && send_oos));
892 if (!(remote || send_oos))
893 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
894 if (!(local || remote)) {
895 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
896 spin_unlock_irq(&mdev->req_lock);
897 goto fail_free_complete;
901 if (b && mdev->unused_spare_tle == NULL) {
902 mdev->unused_spare_tle = b;
905 if (rw == WRITE && (remote || send_oos) &&
906 mdev->unused_spare_tle == NULL &&
907 test_bit(CREATE_BARRIER, &mdev->flags)) {
908 /* someone closed the current epoch
909 * while we were grabbing the spinlock */
910 spin_unlock_irq(&mdev->req_lock);
911 goto allocate_barrier;
915 /* Update disk stats */
916 _drbd_start_io_acct(mdev, req, bio);
918 /* _maybe_start_new_epoch(mdev);
919 * If we need to generate a write barrier packet, we have to add the
920 * new epoch (barrier) object, and queue the barrier packet for sending,
921 * and queue the req's data after it _within the same lock_, otherwise
922 * we have race conditions were the reorder domains could be mixed up.
924 * Even read requests may start a new epoch and queue the corresponding
925 * barrier packet. To get the write ordering right, we only have to
926 * make sure that, if this is a write request and it triggered a
927 * barrier packet, this request is queued within the same spinlock. */
928 if ((remote || send_oos) && mdev->unused_spare_tle &&
929 test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
930 _tl_add_barrier(mdev, mdev->unused_spare_tle);
931 mdev->unused_spare_tle = NULL;
933 D_ASSERT(!(remote && rw == WRITE &&
934 test_bit(CREATE_BARRIER, &mdev->flags)));
938 * Actually, 'local' may be wrong here already, since we may have failed
939 * to write to the meta data, and may become wrong anytime because of
940 * local io-error for some other request, which would lead to us
941 * "detaching" the local disk.
943 * 'remote' may become wrong any time because the network could fail.
945 * This is a harmless race condition, though, since it is handled
946 * correctly at the appropriate places; so it just defers the failure
947 * of the respective operation.
950 /* mark them early for readability.
951 * this just sets some state flags. */
953 _req_mod(req, to_be_send);
955 _req_mod(req, to_be_submitted);
957 /* check this request on the collision detection hash tables.
958 * if we have a conflict, just complete it here.
959 * THINK do we want to check reads, too? (I don't think so...) */
960 if (rw == WRITE && _req_conflicts(req))
961 goto fail_conflicting;
963 list_add_tail(&req->tl_requests, &mdev->newest_tle->requests);
965 /* NOTE remote first: to get the concurrent write detection right,
966 * we must register the request before start of local IO. */
968 /* either WRITE and C_CONNECTED,
969 * or READ, and no local disk,
970 * or READ, but not in sync.
972 _req_mod(req, (rw == WRITE)
973 ? queue_for_net_write
974 : queue_for_net_read);
976 if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
977 _req_mod(req, queue_for_send_oos);
980 mdev->net_conf->on_congestion != OC_BLOCK && mdev->agreed_pro_version >= 96) {
983 if (mdev->net_conf->cong_fill &&
984 atomic_read(&mdev->ap_in_flight) >= mdev->net_conf->cong_fill) {
985 dev_info(DEV, "Congestion-fill threshold reached\n");
989 if (mdev->act_log->used >= mdev->net_conf->cong_extents) {
990 dev_info(DEV, "Congestion-extents threshold reached\n");
995 queue_barrier(mdev); /* last barrier, after mirrored writes */
997 if (mdev->net_conf->on_congestion == OC_PULL_AHEAD)
998 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
999 else /*mdev->net_conf->on_congestion == OC_DISCONNECT */
1000 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
1004 spin_unlock_irq(&mdev->req_lock);
1005 kfree(b); /* if someone else has beaten us to it... */
1008 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
1010 /* State may have changed since we grabbed our reference on the
1011 * mdev->ldev member. Double check, and short-circuit to endio.
1012 * In case the last activity log transaction failed to get on
1013 * stable storage, and this is a WRITE, we may not even submit
1015 if (get_ldev(mdev)) {
1016 if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
1017 : rw == READ ? DRBD_FAULT_DT_RD
1018 : DRBD_FAULT_DT_RA))
1019 bio_endio(req->private_bio, -EIO);
1021 generic_make_request(req->private_bio);
1024 bio_endio(req->private_bio, -EIO);
1030 /* this is a conflicting request.
1031 * even though it may have been only _partially_
1032 * overlapping with one of the currently pending requests,
1033 * without even submitting or sending it, we will
1034 * pretend that it was successfully served right now.
1036 _drbd_end_io_acct(mdev, req);
1037 spin_unlock_irq(&mdev->req_lock);
1039 dec_ap_pending(mdev);
1040 /* THINK: do we want to fail it (-EIO), or pretend success?
1041 * this pretends success. */
1045 if (req->rq_state & RQ_IN_ACT_LOG)
1046 drbd_al_complete_io(mdev, sector);
1049 bio_put(req->private_bio);
1050 req->private_bio = NULL;
1054 bio_endio(bio, err);
1063 /* helper function for drbd_make_request
1064 * if we can determine just by the mdev (state) that this request will fail,
1066 * otherwise return 0
1068 static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write)
1070 if (mdev->state.role != R_PRIMARY &&
1071 (!allow_oos || is_write)) {
1072 if (__ratelimit(&drbd_ratelimit_state)) {
1073 dev_err(DEV, "Process %s[%u] tried to %s; "
1074 "since we are not in Primary state, "
1075 "we cannot allow this\n",
1076 current->comm, current->pid,
1077 is_write ? "WRITE" : "READ");
1085 int drbd_make_request(struct request_queue *q, struct bio *bio)
1087 unsigned int s_enr, e_enr;
1088 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1089 unsigned long start_time;
1091 if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) {
1092 bio_endio(bio, -EPERM);
1096 start_time = jiffies;
1099 * what we "blindly" assume:
1101 D_ASSERT(bio->bi_size > 0);
1102 D_ASSERT((bio->bi_size & 0x1ff) == 0);
1103 D_ASSERT(bio->bi_idx == 0);
1105 /* to make some things easier, force alignment of requests within the
1106 * granularity of our hash tables */
1107 s_enr = bio->bi_sector >> HT_SHIFT;
1108 e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT;
1110 if (likely(s_enr == e_enr)) {
1111 inc_ap_bio(mdev, 1);
1112 return drbd_make_request_common(mdev, bio, start_time);
1115 /* can this bio be split generically?
1116 * Maybe add our own split-arbitrary-bios function. */
1117 if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_BIO_SIZE) {
1118 /* rather error out here than BUG in bio_split */
1119 dev_err(DEV, "bio would need to, but cannot, be split: "
1120 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1121 bio->bi_vcnt, bio->bi_idx, bio->bi_size,
1122 (unsigned long long)bio->bi_sector);
1123 bio_endio(bio, -EINVAL);
1125 /* This bio crosses some boundary, so we have to split it. */
1126 struct bio_pair *bp;
1127 /* works for the "do not cross hash slot boundaries" case
1128 * e.g. sector 262269, size 4096
1129 * s_enr = 262269 >> 6 = 4097
1130 * e_enr = (262269+8-1) >> 6 = 4098
1132 * sps = 64, mask = 63
1133 * first_sectors = 64 - (262269 & 63) = 3
1135 const sector_t sect = bio->bi_sector;
1136 const int sps = 1 << HT_SHIFT; /* sectors per slot */
1137 const int mask = sps - 1;
1138 const sector_t first_sectors = sps - (sect & mask);
1139 bp = bio_split(bio, first_sectors);
1141 /* we need to get a "reference count" (ap_bio_cnt)
1142 * to avoid races with the disconnect/reconnect/suspend code.
1143 * In case we need to split the bio here, we need to get three references
1144 * atomically, otherwise we might deadlock when trying to submit the
1146 inc_ap_bio(mdev, 3);
1148 D_ASSERT(e_enr == s_enr + 1);
1150 while (drbd_make_request_common(mdev, &bp->bio1, start_time))
1151 inc_ap_bio(mdev, 1);
1153 while (drbd_make_request_common(mdev, &bp->bio2, start_time))
1154 inc_ap_bio(mdev, 1);
1158 bio_pair_release(bp);
1163 /* This is called by bio_add_page(). With this function we reduce
1164 * the number of BIOs that span over multiple DRBD_MAX_BIO_SIZEs
1165 * units (was AL_EXTENTs).
1167 * we do the calculation within the lower 32bit of the byte offsets,
1168 * since we don't care for actual offset, but only check whether it
1169 * would cross "activity log extent" boundaries.
1171 * As long as the BIO is empty we have to allow at least one bvec,
1172 * regardless of size and offset. so the resulting bio may still
1173 * cross extent boundaries. those are dealt with (bio_split) in
1174 * drbd_make_request.
1176 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1178 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1179 unsigned int bio_offset =
1180 (unsigned int)bvm->bi_sector << 9; /* 32 bit */
1181 unsigned int bio_size = bvm->bi_size;
1182 int limit, backing_limit;
1184 limit = DRBD_MAX_BIO_SIZE
1185 - ((bio_offset & (DRBD_MAX_BIO_SIZE-1)) + bio_size);
1188 if (bio_size == 0) {
1189 if (limit <= bvec->bv_len)
1190 limit = bvec->bv_len;
1191 } else if (limit && get_ldev(mdev)) {
1192 struct request_queue * const b =
1193 mdev->ldev->backing_bdev->bd_disk->queue;
1194 if (b->merge_bvec_fn) {
1195 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1196 limit = min(limit, backing_limit);
1203 void request_timer_fn(unsigned long data)
1205 struct drbd_conf *mdev = (struct drbd_conf *) data;
1206 struct drbd_request *req; /* oldest request */
1207 struct list_head *le;
1208 unsigned long et = 0; /* effective timeout = ko_count * timeout */
1210 if (get_net_conf(mdev)) {
1211 et = mdev->net_conf->timeout*HZ/10 * mdev->net_conf->ko_count;
1214 if (!et || mdev->state.conn < C_WF_REPORT_PARAMS)
1215 return; /* Recurring timer stopped */
1217 spin_lock_irq(&mdev->req_lock);
1218 le = &mdev->oldest_tle->requests;
1219 if (list_empty(le)) {
1220 spin_unlock_irq(&mdev->req_lock);
1221 mod_timer(&mdev->request_timer, jiffies + et);
1226 req = list_entry(le, struct drbd_request, tl_requests);
1227 if (time_is_before_eq_jiffies(req->start_time + et)) {
1228 if (req->rq_state & RQ_NET_PENDING) {
1229 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1230 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE, NULL);
1232 dev_warn(DEV, "Local backing block device frozen?\n");
1233 mod_timer(&mdev->request_timer, jiffies + et);
1236 mod_timer(&mdev->request_timer, req->start_time + et);
1239 spin_unlock_irq(&mdev->req_lock);