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->sector, req->size);
82 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
83 drbd_set_in_sync(mdev, req->sector, req->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 drbd_al_complete_io(mdev, req->sector);
99 } else if (__ratelimit(&drbd_ratelimit_state)) {
100 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu), "
101 "but my Disk seems to have failed :(\n",
102 (unsigned long long) req->sector);
110 static void queue_barrier(struct drbd_conf *mdev)
112 struct drbd_tl_epoch *b;
114 /* We are within the req_lock. Once we queued the barrier for sending,
115 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
116 * barrier/epoch object is added. This is the only place this bit is
117 * set. It indicates that the barrier for this epoch is already queued,
118 * and no new epoch has been created yet. */
119 if (test_bit(CREATE_BARRIER, &mdev->flags))
122 b = mdev->newest_tle;
123 b->w.cb = w_send_barrier;
124 /* inc_ap_pending done here, so we won't
125 * get imbalanced on connection loss.
126 * dec_ap_pending will be done in got_BarrierAck
127 * or (on connection loss) in tl_clear. */
128 inc_ap_pending(mdev);
129 drbd_queue_work(&mdev->data.work, &b->w);
130 set_bit(CREATE_BARRIER, &mdev->flags);
133 static void _about_to_complete_local_write(struct drbd_conf *mdev,
134 struct drbd_request *req)
136 const unsigned long s = req->rq_state;
137 struct drbd_request *i;
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 if (mdev->state.conn >= C_CONNECTED &&
145 req->epoch == mdev->newest_tle->br_number)
148 /* we need to do the conflict detection stuff,
149 * if we have the ee_hash (two_primaries) and
150 * this has been on the network */
151 if ((s & RQ_NET_DONE) && mdev->ee_hash != NULL) {
152 const sector_t sector = req->sector;
153 const int size = req->size;
156 * there must be no conflicting requests, since
157 * they must have been failed on the spot */
158 #define OVERLAPS overlaps(sector, size, i->sector, i->size)
159 slot = tl_hash_slot(mdev, sector);
160 hlist_for_each_entry(i, n, slot, colision) {
162 dev_alert(DEV, "LOGIC BUG: completed: %p %llus +%u; "
163 "other: %p %llus +%u\n",
164 req, (unsigned long long)sector, size,
165 i, (unsigned long long)i->sector, i->size);
169 /* maybe "wake" those conflicting epoch entries
170 * that wait for this request to finish.
172 * currently, there can be only _one_ such ee
173 * (well, or some more, which would be pending
174 * P_DISCARD_ACK not yet sent by the asender...),
175 * since we block the receiver thread upon the
176 * first conflict detection, which will wait on
177 * misc_wait. maybe we want to assert that?
179 * anyways, if we found one,
180 * we just have to do a wake_up. */
182 #define OVERLAPS overlaps(sector, size, e->sector, e->size)
183 slot = ee_hash_slot(mdev, req->sector);
184 hlist_for_each_entry(e, n, slot, colision) {
186 wake_up(&mdev->misc_wait);
194 void complete_master_bio(struct drbd_conf *mdev,
195 struct bio_and_error *m)
197 bio_endio(m->bio, m->error);
201 /* Helper for __req_mod().
202 * Set m->bio to the master bio, if it is fit to be completed,
203 * or leave it alone (it is initialized to NULL in __req_mod),
204 * if it has already been completed, or cannot be completed yet.
205 * If m->bio is set, the error status to be returned is placed in m->error.
207 void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
209 const unsigned long s = req->rq_state;
210 struct drbd_conf *mdev = req->mdev;
211 /* only WRITES may end up here without a master bio (on barrier ack) */
212 int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE;
214 /* we must not complete the master bio, while it is
215 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
216 * not yet acknowledged by the peer
217 * not yet completed by the local io subsystem
218 * these flags may get cleared in any order by
221 * the bio_endio completion callbacks.
223 if (s & RQ_NET_QUEUED)
225 if (s & RQ_NET_PENDING)
227 if (s & RQ_LOCAL_PENDING)
230 if (req->master_bio) {
231 /* this is data_received (remote read)
232 * or protocol C P_WRITE_ACK
233 * or protocol B P_RECV_ACK
234 * or protocol A "handed_over_to_network" (SendAck)
235 * or canceled or failed,
236 * or killed from the transfer log due to connection loss.
240 * figure out whether to report success or failure.
242 * report success when at least one of the operations succeeded.
243 * or, to put the other way,
244 * only report failure, when both operations failed.
246 * what to do about the failures is handled elsewhere.
247 * what we need to do here is just: complete the master_bio.
249 * local completion error, if any, has been stored as ERR_PTR
250 * in private_bio within drbd_endio_pri.
252 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
253 int error = PTR_ERR(req->private_bio);
255 /* remove the request from the conflict detection
256 * respective block_id verification hash */
257 if (!hlist_unhashed(&req->colision))
258 hlist_del(&req->colision);
260 D_ASSERT((s & RQ_NET_MASK) == 0);
262 /* for writes we need to do some extra housekeeping */
264 _about_to_complete_local_write(mdev, req);
266 /* Update disk stats */
267 _drbd_end_io_acct(mdev, req);
269 m->error = ok ? 0 : (error ?: -EIO);
270 m->bio = req->master_bio;
271 req->master_bio = NULL;
274 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
275 /* this is disconnected (local only) operation,
276 * or protocol C P_WRITE_ACK,
277 * or protocol A or B P_BARRIER_ACK,
278 * or killed from the transfer log due to connection loss. */
279 _req_is_done(mdev, req, rw);
281 /* else: network part and not DONE yet. that is
282 * protocol A or B, barrier ack still pending... */
285 static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
287 struct drbd_conf *mdev = req->mdev;
289 if (!mdev->state.susp)
290 _req_may_be_done(req, m);
294 * checks whether there was an overlapping request
295 * or ee already registered.
297 * if so, return 1, in which case this request is completed on the spot,
298 * without ever being submitted or send.
300 * return 0 if it is ok to submit this request.
303 * paranoia: assume something above us is broken, and issues different write
304 * requests for the same block simultaneously...
306 * To ensure these won't be reordered differently on both nodes, resulting in
307 * diverging data sets, we discard the later one(s). Not that this is supposed
308 * to happen, but this is the rationale why we also have to check for
309 * conflicting requests with local origin, and why we have to do so regardless
310 * of whether we allowed multiple primaries.
312 * BTW, in case we only have one primary, the ee_hash is empty anyways, and the
313 * second hlist_for_each_entry becomes a noop. This is even simpler than to
314 * grab a reference on the net_conf, and check for the two_primaries flag...
316 static int _req_conflicts(struct drbd_request *req)
318 struct drbd_conf *mdev = req->mdev;
319 const sector_t sector = req->sector;
320 const int size = req->size;
321 struct drbd_request *i;
322 struct drbd_epoch_entry *e;
323 struct hlist_node *n;
324 struct hlist_head *slot;
326 D_ASSERT(hlist_unhashed(&req->colision));
328 if (!get_net_conf(mdev))
332 ERR_IF (mdev->tl_hash_s == 0)
333 goto out_no_conflict;
334 BUG_ON(mdev->tl_hash == NULL);
336 #define OVERLAPS overlaps(i->sector, i->size, sector, size)
337 slot = tl_hash_slot(mdev, sector);
338 hlist_for_each_entry(i, n, slot, colision) {
340 dev_alert(DEV, "%s[%u] Concurrent local write detected! "
341 "[DISCARD L] new: %llus +%u; "
342 "pending: %llus +%u\n",
343 current->comm, current->pid,
344 (unsigned long long)sector, size,
345 (unsigned long long)i->sector, i->size);
350 if (mdev->ee_hash_s) {
351 /* now, check for overlapping requests with remote origin */
352 BUG_ON(mdev->ee_hash == NULL);
354 #define OVERLAPS overlaps(e->sector, e->size, sector, size)
355 slot = ee_hash_slot(mdev, sector);
356 hlist_for_each_entry(e, n, slot, colision) {
358 dev_alert(DEV, "%s[%u] Concurrent remote write detected!"
359 " [DISCARD L] new: %llus +%u; "
360 "pending: %llus +%u\n",
361 current->comm, current->pid,
362 (unsigned long long)sector, size,
363 (unsigned long long)e->sector, e->size);
371 /* this is like it should be, and what we expected.
372 * our users do behave after all... */
381 /* obviously this could be coded as many single functions
382 * instead of one huge switch,
383 * or by putting the code directly in the respective locations
384 * (as it has been before).
386 * but having it this way
387 * enforces that it is all in this one place, where it is easier to audit,
388 * it makes it obvious that whatever "event" "happens" to a request should
389 * happen "atomically" within the req_lock,
390 * and it enforces that we have to think in a very structured manner
391 * about the "events" that may happen to a request during its life time ...
393 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
394 struct bio_and_error *m)
396 struct drbd_conf *mdev = req->mdev;
402 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
405 /* does not happen...
406 * initialization done in drbd_req_new
411 case to_be_send: /* via network */
412 /* reached via drbd_make_request_common
413 * and from w_read_retry_remote */
414 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
415 req->rq_state |= RQ_NET_PENDING;
416 inc_ap_pending(mdev);
419 case to_be_submitted: /* locally */
420 /* reached via drbd_make_request_common */
421 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
422 req->rq_state |= RQ_LOCAL_PENDING;
426 if (bio_data_dir(req->master_bio) == WRITE)
427 mdev->writ_cnt += req->size>>9;
429 mdev->read_cnt += req->size>>9;
431 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
432 req->rq_state &= ~RQ_LOCAL_PENDING;
434 _req_may_be_done_not_susp(req, m);
438 case write_completed_with_error:
439 req->rq_state |= RQ_LOCAL_COMPLETED;
440 req->rq_state &= ~RQ_LOCAL_PENDING;
442 __drbd_chk_io_error(mdev, FALSE);
443 _req_may_be_done_not_susp(req, m);
447 case read_ahead_completed_with_error:
448 /* it is legal to fail READA */
449 req->rq_state |= RQ_LOCAL_COMPLETED;
450 req->rq_state &= ~RQ_LOCAL_PENDING;
451 _req_may_be_done_not_susp(req, m);
455 case read_completed_with_error:
456 drbd_set_out_of_sync(mdev, req->sector, req->size);
458 req->rq_state |= RQ_LOCAL_COMPLETED;
459 req->rq_state &= ~RQ_LOCAL_PENDING;
461 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
463 __drbd_chk_io_error(mdev, FALSE);
466 /* no point in retrying if there is no good remote data,
467 * or we have no connection. */
468 if (mdev->state.pdsk != D_UP_TO_DATE) {
469 _req_may_be_done_not_susp(req, m);
473 /* _req_mod(req,to_be_send); oops, recursion... */
474 req->rq_state |= RQ_NET_PENDING;
475 inc_ap_pending(mdev);
476 /* fall through: _req_mod(req,queue_for_net_read); */
478 case queue_for_net_read:
479 /* READ or READA, and
481 * or target area marked as invalid,
482 * or just got an io-error. */
483 /* from drbd_make_request_common
484 * or from bio_endio during read io-error recovery */
486 /* so we can verify the handle in the answer packet
487 * corresponding hlist_del is in _req_may_be_done() */
488 hlist_add_head(&req->colision, ar_hash_slot(mdev, req->sector));
490 set_bit(UNPLUG_REMOTE, &mdev->flags);
492 D_ASSERT(req->rq_state & RQ_NET_PENDING);
493 req->rq_state |= RQ_NET_QUEUED;
494 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
495 ? w_read_retry_remote
497 drbd_queue_work(&mdev->data.work, &req->w);
500 case queue_for_net_write:
501 /* assert something? */
502 /* from drbd_make_request_common only */
504 hlist_add_head(&req->colision, tl_hash_slot(mdev, req->sector));
505 /* corresponding hlist_del is in _req_may_be_done() */
508 * In case the req ended up on the transfer log before being
509 * queued on the worker, it could lead to this request being
510 * missed during cleanup after connection loss.
511 * So we have to do both operations here,
512 * within the same lock that protects the transfer log.
514 * _req_add_to_epoch(req); this has to be after the
515 * _maybe_start_new_epoch(req); which happened in
516 * drbd_make_request_common, because we now may set the bit
517 * again ourselves to close the current epoch.
519 * Add req to the (now) current epoch (barrier). */
521 /* otherwise we may lose an unplug, which may cause some remote
522 * io-scheduler timeout to expire, increasing maximum latency,
523 * hurting performance. */
524 set_bit(UNPLUG_REMOTE, &mdev->flags);
526 /* see drbd_make_request_common,
527 * just after it grabs the req_lock */
528 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
530 req->epoch = mdev->newest_tle->br_number;
532 /* increment size of current epoch */
533 mdev->newest_tle->n_writes++;
535 /* queue work item to send data */
536 D_ASSERT(req->rq_state & RQ_NET_PENDING);
537 req->rq_state |= RQ_NET_QUEUED;
538 req->w.cb = w_send_dblock;
539 drbd_queue_work(&mdev->data.work, &req->w);
541 /* close the epoch, in case it outgrew the limit */
542 if (mdev->newest_tle->n_writes >= mdev->net_conf->max_epoch_size)
548 /* treat it the same */
550 /* real cleanup will be done from tl_clear. just update flags
551 * so it is no longer marked as on the worker queue */
552 req->rq_state &= ~RQ_NET_QUEUED;
553 /* if we did it right, tl_clear should be scheduled only after
554 * this, so this should not be necessary! */
555 _req_may_be_done_not_susp(req, m);
558 case handed_over_to_network:
559 /* assert something? */
560 if (bio_data_dir(req->master_bio) == WRITE &&
561 mdev->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 it is still queued, we may not complete it here.
594 * it will be canceled soon. */
595 if (!(req->rq_state & RQ_NET_QUEUED))
596 _req_may_be_done(req, m); /* Allowed while state.susp */
599 case write_acked_by_peer_and_sis:
600 req->rq_state |= RQ_NET_SIS;
601 case conflict_discarded_by_peer:
602 /* for discarded conflicting writes of multiple primaries,
603 * there is no need to keep anything in the tl, potential
604 * node crashes are covered by the activity log. */
605 if (what == conflict_discarded_by_peer)
606 dev_alert(DEV, "Got DiscardAck packet %llus +%u!"
607 " DRBD is not a random data generator!\n",
608 (unsigned long long)req->sector, req->size);
609 req->rq_state |= RQ_NET_DONE;
611 case write_acked_by_peer:
612 /* protocol C; successfully written on peer.
613 * Nothing to do here.
614 * We want to keep the tl in place for all protocols, to cater
615 * for volatile write-back caches on lower level devices.
617 * A barrier request is expected to have forced all prior
618 * requests onto stable storage, so completion of a barrier
619 * request could set NET_DONE right here, and not wait for the
620 * P_BARRIER_ACK, but that is an unnecessary optimization. */
622 /* this makes it effectively the same as for: */
623 case recv_acked_by_peer:
624 /* protocol B; pretends to be successfully written on peer.
625 * see also notes above in handed_over_to_network about
627 req->rq_state |= RQ_NET_OK;
628 D_ASSERT(req->rq_state & RQ_NET_PENDING);
629 dec_ap_pending(mdev);
630 req->rq_state &= ~RQ_NET_PENDING;
631 _req_may_be_done_not_susp(req, m);
635 /* assert something? */
636 if (req->rq_state & RQ_NET_PENDING)
637 dec_ap_pending(mdev);
638 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
640 req->rq_state |= RQ_NET_DONE;
641 _req_may_be_done_not_susp(req, m);
642 /* else: done by handed_over_to_network */
645 case fail_frozen_disk_io:
646 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
649 _req_may_be_done(req, m); /* Allowed while state.susp */
652 case restart_frozen_disk_io:
653 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
656 req->rq_state &= ~RQ_LOCAL_COMPLETED;
659 if (bio_data_dir(req->master_bio) == WRITE)
663 req->w.cb = w_restart_disk_io;
664 drbd_queue_work(&mdev->data.work, &req->w);
668 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
669 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
670 Trowing them out of the TL here by pretending we got a BARRIER_ACK
671 We ensure that the peer was not rebooted */
672 if (!(req->rq_state & RQ_NET_OK)) {
674 drbd_queue_work(&mdev->data.work, &req->w);
675 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
679 /* else, fall through to barrier_acked */
682 if (!(req->rq_state & RQ_WRITE))
685 if (req->rq_state & RQ_NET_PENDING) {
686 /* barrier came in before all requests have been acked.
687 * this is bad, because if the connection is lost now,
688 * we won't be able to clean them up... */
689 dev_err(DEV, "FIXME (barrier_acked but pending)\n");
690 list_move(&req->tl_requests, &mdev->out_of_sequence_requests);
692 D_ASSERT(req->rq_state & RQ_NET_SENT);
693 req->rq_state |= RQ_NET_DONE;
694 _req_may_be_done(req, m); /* Allowed while state.susp */
698 D_ASSERT(req->rq_state & RQ_NET_PENDING);
699 dec_ap_pending(mdev);
700 req->rq_state &= ~RQ_NET_PENDING;
701 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
702 _req_may_be_done_not_susp(req, m);
709 /* we may do a local read if:
710 * - we are consistent (of course),
711 * - or we are generally inconsistent,
712 * BUT we are still/already IN SYNC for this area.
713 * since size may be bigger than BM_BLOCK_SIZE,
714 * we may need to check several bits.
716 static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
718 unsigned long sbnr, ebnr;
719 sector_t esector, nr_sectors;
721 if (mdev->state.disk == D_UP_TO_DATE)
723 if (mdev->state.disk >= D_OUTDATED)
725 if (mdev->state.disk < D_INCONSISTENT)
727 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
728 nr_sectors = drbd_get_capacity(mdev->this_bdev);
729 esector = sector + (size >> 9) - 1;
731 D_ASSERT(sector < nr_sectors);
732 D_ASSERT(esector < nr_sectors);
734 sbnr = BM_SECT_TO_BIT(sector);
735 ebnr = BM_SECT_TO_BIT(esector);
737 return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr);
740 static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio)
742 const int rw = bio_rw(bio);
743 const int size = bio->bi_size;
744 const sector_t sector = bio->bi_sector;
745 struct drbd_tl_epoch *b = NULL;
746 struct drbd_request *req;
751 /* allocate outside of all locks; */
752 req = drbd_req_new(mdev, bio);
755 /* only pass the error to the upper layers.
756 * if user cannot handle io errors, that's not our business. */
757 dev_err(DEV, "could not kmalloc() req\n");
758 bio_endio(bio, -ENOMEM);
762 local = get_ldev(mdev);
764 bio_put(req->private_bio); /* or we get a bio leak */
765 req->private_bio = NULL;
772 if (!drbd_may_do_local_read(mdev, sector, size)) {
773 /* we could kick the syncer to
774 * sync this extent asap, wait for
775 * it, then continue locally.
776 * Or just issue the request remotely.
779 bio_put(req->private_bio);
780 req->private_bio = NULL;
784 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
787 /* If we have a disk, but a READA request is mapped to remote,
788 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
789 * Just fail that READA request right here.
791 * THINK: maybe fail all READA when not local?
792 * or make this configurable...
793 * if network is slow, READA won't do any good.
795 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
797 goto fail_and_free_req;
800 /* For WRITES going to the local disk, grab a reference on the target
801 * extent. This waits for any resync activity in the corresponding
802 * resync extent to finish, and, if necessary, pulls in the target
803 * extent into the activity log, which involves further disk io because
804 * of transactional on-disk meta data updates. */
805 if (rw == WRITE && local)
806 drbd_al_begin_io(mdev, sector);
808 remote = remote && (mdev->state.pdsk == D_UP_TO_DATE ||
809 (mdev->state.pdsk == D_INCONSISTENT &&
810 mdev->state.conn >= C_CONNECTED));
812 if (!(local || remote) && !mdev->state.susp) {
813 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
814 goto fail_free_complete;
817 /* For WRITE request, we have to make sure that we have an
818 * unused_spare_tle, in case we need to start a new epoch.
819 * I try to be smart and avoid to pre-allocate always "just in case",
820 * but there is a race between testing the bit and pointer outside the
821 * spinlock, and grabbing the spinlock.
822 * if we lost that race, we retry. */
823 if (rw == WRITE && remote &&
824 mdev->unused_spare_tle == NULL &&
825 test_bit(CREATE_BARRIER, &mdev->flags)) {
827 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
829 dev_err(DEV, "Failed to alloc barrier.\n");
831 goto fail_free_complete;
835 /* GOOD, everything prepared, grab the spin_lock */
836 spin_lock_irq(&mdev->req_lock);
838 if (mdev->state.susp) {
839 /* If we got suspended, use the retry mechanism of
840 generic_make_request() to restart processing of this
841 bio. In the next call to drbd_make_request_26
842 we sleep in inc_ap_bio() */
844 spin_unlock_irq(&mdev->req_lock);
845 goto fail_free_complete;
849 remote = (mdev->state.pdsk == D_UP_TO_DATE ||
850 (mdev->state.pdsk == D_INCONSISTENT &&
851 mdev->state.conn >= C_CONNECTED));
853 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
854 if (!(local || remote)) {
855 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
856 spin_unlock_irq(&mdev->req_lock);
857 goto fail_free_complete;
861 if (b && mdev->unused_spare_tle == NULL) {
862 mdev->unused_spare_tle = b;
865 if (rw == WRITE && remote &&
866 mdev->unused_spare_tle == NULL &&
867 test_bit(CREATE_BARRIER, &mdev->flags)) {
868 /* someone closed the current epoch
869 * while we were grabbing the spinlock */
870 spin_unlock_irq(&mdev->req_lock);
871 goto allocate_barrier;
875 /* Update disk stats */
876 _drbd_start_io_acct(mdev, req, bio);
878 /* _maybe_start_new_epoch(mdev);
879 * If we need to generate a write barrier packet, we have to add the
880 * new epoch (barrier) object, and queue the barrier packet for sending,
881 * and queue the req's data after it _within the same lock_, otherwise
882 * we have race conditions were the reorder domains could be mixed up.
884 * Even read requests may start a new epoch and queue the corresponding
885 * barrier packet. To get the write ordering right, we only have to
886 * make sure that, if this is a write request and it triggered a
887 * barrier packet, this request is queued within the same spinlock. */
888 if (remote && mdev->unused_spare_tle &&
889 test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
890 _tl_add_barrier(mdev, mdev->unused_spare_tle);
891 mdev->unused_spare_tle = NULL;
893 D_ASSERT(!(remote && rw == WRITE &&
894 test_bit(CREATE_BARRIER, &mdev->flags)));
898 * Actually, 'local' may be wrong here already, since we may have failed
899 * to write to the meta data, and may become wrong anytime because of
900 * local io-error for some other request, which would lead to us
901 * "detaching" the local disk.
903 * 'remote' may become wrong any time because the network could fail.
905 * This is a harmless race condition, though, since it is handled
906 * correctly at the appropriate places; so it just defers the failure
907 * of the respective operation.
910 /* mark them early for readability.
911 * this just sets some state flags. */
913 _req_mod(req, to_be_send);
915 _req_mod(req, to_be_submitted);
917 /* check this request on the collision detection hash tables.
918 * if we have a conflict, just complete it here.
919 * THINK do we want to check reads, too? (I don't think so...) */
920 if (rw == WRITE && _req_conflicts(req))
921 goto fail_conflicting;
923 list_add_tail(&req->tl_requests, &mdev->newest_tle->requests);
925 /* NOTE remote first: to get the concurrent write detection right,
926 * we must register the request before start of local IO. */
928 /* either WRITE and C_CONNECTED,
929 * or READ, and no local disk,
930 * or READ, but not in sync.
932 _req_mod(req, (rw == WRITE)
933 ? queue_for_net_write
934 : queue_for_net_read);
936 spin_unlock_irq(&mdev->req_lock);
937 kfree(b); /* if someone else has beaten us to it... */
940 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
942 if (FAULT_ACTIVE(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
943 : rw == READ ? DRBD_FAULT_DT_RD
945 bio_endio(req->private_bio, -EIO);
947 generic_make_request(req->private_bio);
950 /* we need to plug ALWAYS since we possibly need to kick lo_dev.
951 * we plug after submit, so we won't miss an unplug event */
952 drbd_plug_device(mdev);
957 /* this is a conflicting request.
958 * even though it may have been only _partially_
959 * overlapping with one of the currently pending requests,
960 * without even submitting or sending it, we will
961 * pretend that it was successfully served right now.
963 _drbd_end_io_acct(mdev, req);
964 spin_unlock_irq(&mdev->req_lock);
966 dec_ap_pending(mdev);
967 /* THINK: do we want to fail it (-EIO), or pretend success?
968 * this pretends success. */
972 if (rw == WRITE && local)
973 drbd_al_complete_io(mdev, sector);
976 bio_put(req->private_bio);
977 req->private_bio = NULL;
990 /* helper function for drbd_make_request
991 * if we can determine just by the mdev (state) that this request will fail,
995 static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write)
997 if (mdev->state.role != R_PRIMARY &&
998 (!allow_oos || is_write)) {
999 if (__ratelimit(&drbd_ratelimit_state)) {
1000 dev_err(DEV, "Process %s[%u] tried to %s; "
1001 "since we are not in Primary state, "
1002 "we cannot allow this\n",
1003 current->comm, current->pid,
1004 is_write ? "WRITE" : "READ");
1012 int drbd_make_request_26(struct request_queue *q, struct bio *bio)
1014 unsigned int s_enr, e_enr;
1015 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1017 if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) {
1018 bio_endio(bio, -EPERM);
1022 /* Reject barrier requests if we know the underlying device does
1024 * XXX: Need to get this info from peer as well some how so we
1025 * XXX: reject if EITHER side/data/metadata area does not support them.
1027 * because of those XXX, this is not yet enabled,
1028 * i.e. in drbd_init_set_defaults we set the NO_BARRIER_SUPP bit.
1030 if (unlikely(bio->bi_rw & REQ_HARDBARRIER) && test_bit(NO_BARRIER_SUPP, &mdev->flags)) {
1031 /* dev_warn(DEV, "Rejecting barrier request as underlying device does not support\n"); */
1032 bio_endio(bio, -EOPNOTSUPP);
1037 * what we "blindly" assume:
1039 D_ASSERT(bio->bi_size > 0);
1040 D_ASSERT((bio->bi_size & 0x1ff) == 0);
1041 D_ASSERT(bio->bi_idx == 0);
1043 /* to make some things easier, force alignment of requests within the
1044 * granularity of our hash tables */
1045 s_enr = bio->bi_sector >> HT_SHIFT;
1046 e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT;
1048 if (likely(s_enr == e_enr)) {
1049 inc_ap_bio(mdev, 1);
1050 return drbd_make_request_common(mdev, bio);
1053 /* can this bio be split generically?
1054 * Maybe add our own split-arbitrary-bios function. */
1055 if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_SEGMENT_SIZE) {
1056 /* rather error out here than BUG in bio_split */
1057 dev_err(DEV, "bio would need to, but cannot, be split: "
1058 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1059 bio->bi_vcnt, bio->bi_idx, bio->bi_size,
1060 (unsigned long long)bio->bi_sector);
1061 bio_endio(bio, -EINVAL);
1063 /* This bio crosses some boundary, so we have to split it. */
1064 struct bio_pair *bp;
1065 /* works for the "do not cross hash slot boundaries" case
1066 * e.g. sector 262269, size 4096
1067 * s_enr = 262269 >> 6 = 4097
1068 * e_enr = (262269+8-1) >> 6 = 4098
1070 * sps = 64, mask = 63
1071 * first_sectors = 64 - (262269 & 63) = 3
1073 const sector_t sect = bio->bi_sector;
1074 const int sps = 1 << HT_SHIFT; /* sectors per slot */
1075 const int mask = sps - 1;
1076 const sector_t first_sectors = sps - (sect & mask);
1078 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,28)
1083 /* we need to get a "reference count" (ap_bio_cnt)
1084 * to avoid races with the disconnect/reconnect/suspend code.
1085 * In case we need to split the bio here, we need to get three references
1086 * atomically, otherwise we might deadlock when trying to submit the
1088 inc_ap_bio(mdev, 3);
1090 D_ASSERT(e_enr == s_enr + 1);
1092 while (drbd_make_request_common(mdev, &bp->bio1))
1093 inc_ap_bio(mdev, 1);
1095 while (drbd_make_request_common(mdev, &bp->bio2))
1096 inc_ap_bio(mdev, 1);
1100 bio_pair_release(bp);
1105 /* This is called by bio_add_page(). With this function we reduce
1106 * the number of BIOs that span over multiple DRBD_MAX_SEGMENT_SIZEs
1107 * units (was AL_EXTENTs).
1109 * we do the calculation within the lower 32bit of the byte offsets,
1110 * since we don't care for actual offset, but only check whether it
1111 * would cross "activity log extent" boundaries.
1113 * As long as the BIO is empty we have to allow at least one bvec,
1114 * regardless of size and offset. so the resulting bio may still
1115 * cross extent boundaries. those are dealt with (bio_split) in
1116 * drbd_make_request_26.
1118 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1120 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1121 unsigned int bio_offset =
1122 (unsigned int)bvm->bi_sector << 9; /* 32 bit */
1123 unsigned int bio_size = bvm->bi_size;
1124 int limit, backing_limit;
1126 limit = DRBD_MAX_SEGMENT_SIZE
1127 - ((bio_offset & (DRBD_MAX_SEGMENT_SIZE-1)) + bio_size);
1130 if (bio_size == 0) {
1131 if (limit <= bvec->bv_len)
1132 limit = bvec->bv_len;
1133 } else if (limit && get_ldev(mdev)) {
1134 struct request_queue * const b =
1135 mdev->ldev->backing_bdev->bd_disk->queue;
1136 if (b->merge_bvec_fn) {
1137 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1138 limit = min(limit, backing_limit);