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 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size);
36 /* Update disk stats at start of I/O request */
37 static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
39 const int rw = bio_data_dir(bio);
41 cpu = part_stat_lock();
42 part_round_stats(cpu, &mdev->vdisk->part0);
43 part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
44 part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
45 (void) cpu; /* The macro invocations above want the cpu argument, I do not like
46 the compiler warning about cpu only assigned but never used... */
47 part_inc_in_flight(&mdev->vdisk->part0, rw);
51 /* Update disk stats when completing request upwards */
52 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
54 int rw = bio_data_dir(req->master_bio);
55 unsigned long duration = jiffies - req->start_time;
57 cpu = part_stat_lock();
58 part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
59 part_round_stats(cpu, &mdev->vdisk->part0);
60 part_dec_in_flight(&mdev->vdisk->part0, rw);
64 static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
67 struct drbd_request *req;
69 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
73 drbd_req_make_private_bio(req, bio_src);
74 req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
76 req->master_bio = bio_src;
79 drbd_clear_interval(&req->i);
80 req->i.sector = bio_src->bi_sector;
81 req->i.size = bio_src->bi_size;
83 req->i.waiting = false;
85 INIT_LIST_HEAD(&req->tl_requests);
86 INIT_LIST_HEAD(&req->w.list);
88 /* one reference to be put by __drbd_make_request */
89 atomic_set(&req->completion_ref, 1);
90 /* one kref as long as completion_ref > 0 */
91 kref_init(&req->kref);
95 void drbd_req_destroy(struct kref *kref)
97 struct drbd_request *req = container_of(kref, struct drbd_request, kref);
98 struct drbd_conf *mdev = req->w.mdev;
99 const unsigned s = req->rq_state;
101 if ((req->master_bio && !(s & RQ_POSTPONED)) ||
102 atomic_read(&req->completion_ref) ||
103 (s & RQ_LOCAL_PENDING) ||
104 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
105 dev_err(DEV, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
106 s, atomic_read(&req->completion_ref));
110 /* remove it from the transfer log.
111 * well, only if it had been there in the first
112 * place... if it had not (local only or conflicting
113 * and never sent), it should still be "empty" as
114 * initialized in drbd_req_new(), so we can list_del() it
115 * here unconditionally */
116 list_del_init(&req->tl_requests);
118 /* if it was a write, we may have to set the corresponding
119 * bit(s) out-of-sync first. If it had a local part, we need to
120 * release the reference to the activity log. */
122 /* Set out-of-sync unless both OK flags are set
123 * (local only or remote failed).
124 * Other places where we set out-of-sync:
125 * READ with local io-error */
126 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
127 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
129 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
130 drbd_set_in_sync(mdev, req->i.sector, req->i.size);
132 /* one might be tempted to move the drbd_al_complete_io
133 * to the local io completion callback drbd_request_endio.
134 * but, if this was a mirror write, we may only
135 * drbd_al_complete_io after this is RQ_NET_DONE,
136 * otherwise the extent could be dropped from the al
137 * before it has actually been written on the peer.
138 * if we crash before our peer knows about the request,
139 * but after the extent has been dropped from the al,
140 * we would forget to resync the corresponding extent.
142 if (s & RQ_LOCAL_MASK) {
143 if (get_ldev_if_state(mdev, D_FAILED)) {
144 if (s & RQ_IN_ACT_LOG)
145 drbd_al_complete_io(mdev, &req->i);
147 } else if (__ratelimit(&drbd_ratelimit_state)) {
148 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
149 "but my Disk seems to have failed :(\n",
150 (unsigned long long) req->i.sector, req->i.size);
155 mempool_free(req, drbd_request_mempool);
158 static void wake_all_senders(struct drbd_tconn *tconn) {
159 wake_up(&tconn->sender_work.q_wait);
162 /* must hold resource->req_lock */
163 static void start_new_tl_epoch(struct drbd_tconn *tconn)
165 /* no point closing an epoch, if it is empty, anyways. */
166 if (tconn->current_tle_writes == 0)
169 tconn->current_tle_writes = 0;
170 atomic_inc(&tconn->current_tle_nr);
171 wake_all_senders(tconn);
174 void complete_master_bio(struct drbd_conf *mdev,
175 struct bio_and_error *m)
177 bio_endio(m->bio, m->error);
182 static void drbd_remove_request_interval(struct rb_root *root,
183 struct drbd_request *req)
185 struct drbd_conf *mdev = req->w.mdev;
186 struct drbd_interval *i = &req->i;
188 drbd_remove_interval(root, i);
190 /* Wake up any processes waiting for this request to complete. */
192 wake_up(&mdev->misc_wait);
195 /* Helper for __req_mod().
196 * Set m->bio to the master bio, if it is fit to be completed,
197 * or leave it alone (it is initialized to NULL in __req_mod),
198 * if it has already been completed, or cannot be completed yet.
199 * If m->bio is set, the error status to be returned is placed in m->error.
202 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
204 const unsigned s = req->rq_state;
205 struct drbd_conf *mdev = req->w.mdev;
209 /* we must not complete the master bio, while it is
210 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
211 * not yet acknowledged by the peer
212 * not yet completed by the local io subsystem
213 * these flags may get cleared in any order by
216 * the bio_endio completion callbacks.
218 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
219 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
220 (s & RQ_COMPLETION_SUSP)) {
221 dev_err(DEV, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
225 if (!req->master_bio) {
226 dev_err(DEV, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
230 rw = bio_rw(req->master_bio);
233 * figure out whether to report success or failure.
235 * report success when at least one of the operations succeeded.
236 * or, to put the other way,
237 * only report failure, when both operations failed.
239 * what to do about the failures is handled elsewhere.
240 * what we need to do here is just: complete the master_bio.
242 * local completion error, if any, has been stored as ERR_PTR
243 * in private_bio within drbd_request_endio.
245 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
246 error = PTR_ERR(req->private_bio);
248 /* remove the request from the conflict detection
249 * respective block_id verification hash */
250 if (!drbd_interval_empty(&req->i)) {
251 struct rb_root *root;
254 root = &mdev->write_requests;
256 root = &mdev->read_requests;
257 drbd_remove_request_interval(root, req);
258 } else if (!(s & RQ_POSTPONED))
259 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
261 /* Before we can signal completion to the upper layers,
262 * we may need to close the current transfer log epoch.
263 * We are within the request lock, so we can simply compare
264 * the request epoch number with the current transfer log
265 * epoch number. If they match, increase the current_tle_nr,
266 * and reset the transfer log epoch write_cnt.
269 req->epoch == atomic_read(&mdev->tconn->current_tle_nr))
270 start_new_tl_epoch(mdev->tconn);
272 /* Update disk stats */
273 _drbd_end_io_acct(mdev, req);
276 * have it be pushed back to the retry work queue,
277 * so it will re-enter __drbd_make_request(),
278 * and be re-assigned to a suitable local or remote path,
279 * or failed if we do not have access to good data anymore.
281 * Unless it was failed early by __drbd_make_request(),
282 * because no path was available, in which case
283 * it was not even added to the transfer_log.
285 * READA may fail, and will not be retried.
287 * WRITE should have used all available paths already.
289 if (!ok && rw == READ && !list_empty(&req->tl_requests))
290 req->rq_state |= RQ_POSTPONED;
292 if (!(req->rq_state & RQ_POSTPONED)) {
293 m->error = ok ? 0 : (error ?: -EIO);
294 m->bio = req->master_bio;
295 req->master_bio = NULL;
299 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
301 struct drbd_conf *mdev = req->w.mdev;
302 D_ASSERT(m || (req->rq_state & RQ_POSTPONED));
304 if (!atomic_sub_and_test(put, &req->completion_ref))
307 if (drbd_suspended(mdev)) {
308 /* We do not allow completion while suspended. Re-get a
309 * reference, so whatever happens when this is resumed
310 * may put and complete. */
312 D_ASSERT(!(req->rq_state & RQ_COMPLETION_SUSP));
313 req->rq_state |= RQ_COMPLETION_SUSP;
314 atomic_inc(&req->completion_ref);
319 drbd_req_complete(req, m);
321 if (req->rq_state & RQ_POSTPONED) {
322 /* don't destroy the req object just yet,
323 * but queue it for retry */
324 drbd_restart_request(req);
331 /* I'd like this to be the only place that manipulates
332 * req->completion_ref and req->kref. */
333 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
336 struct drbd_conf *mdev = req->w.mdev;
337 unsigned s = req->rq_state;
343 req->rq_state &= ~clear;
344 req->rq_state |= set;
347 if (req->rq_state == s)
350 /* intent: get references */
352 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
353 atomic_inc(&req->completion_ref);
355 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
356 inc_ap_pending(mdev);
357 atomic_inc(&req->completion_ref);
360 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
361 atomic_inc(&req->completion_ref);
363 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
364 kref_get(&req->kref); /* wait for the DONE */
366 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
367 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
369 /* progress: put references */
371 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
374 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
375 D_ASSERT(req->rq_state & RQ_LOCAL_PENDING);
376 /* local completion may still come in later,
377 * we need to keep the req object around. */
378 kref_get(&req->kref);
382 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
383 if (req->rq_state & RQ_LOCAL_ABORTED)
389 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
390 dec_ap_pending(mdev);
394 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
397 if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
398 if (req->rq_state & RQ_NET_SENT)
399 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
403 /* potentially complete and destroy */
405 if (k_put || c_put) {
406 /* Completion does it's own kref_put. If we are going to
407 * kref_sub below, we need req to be still around then. */
408 int at_least = k_put + !!c_put;
409 int refcount = atomic_read(&req->kref.refcount);
410 if (refcount < at_least)
412 "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
413 s, req->rq_state, refcount, at_least);
416 /* If we made progress, retry conflicting peer requests, if any. */
418 wake_up(&mdev->misc_wait);
421 k_put += drbd_req_put_completion_ref(req, m, c_put);
423 kref_sub(&req->kref, k_put, drbd_req_destroy);
426 /* obviously this could be coded as many single functions
427 * instead of one huge switch,
428 * or by putting the code directly in the respective locations
429 * (as it has been before).
431 * but having it this way
432 * enforces that it is all in this one place, where it is easier to audit,
433 * it makes it obvious that whatever "event" "happens" to a request should
434 * happen "atomically" within the req_lock,
435 * and it enforces that we have to think in a very structured manner
436 * about the "events" that may happen to a request during its life time ...
438 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
439 struct bio_and_error *m)
441 struct drbd_conf *mdev = req->w.mdev;
450 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
453 /* does not happen...
454 * initialization done in drbd_req_new
459 case TO_BE_SENT: /* via network */
460 /* reached via __drbd_make_request
461 * and from w_read_retry_remote */
462 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
464 nc = rcu_dereference(mdev->tconn->net_conf);
465 p = nc->wire_protocol;
468 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
469 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
470 mod_rq_state(req, m, 0, RQ_NET_PENDING);
473 case TO_BE_SUBMITTED: /* locally */
474 /* reached via __drbd_make_request */
475 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
476 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
480 if (req->rq_state & RQ_WRITE)
481 mdev->writ_cnt += req->i.size >> 9;
483 mdev->read_cnt += req->i.size >> 9;
485 mod_rq_state(req, m, RQ_LOCAL_PENDING,
486 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
490 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
493 case READ_COMPLETED_WITH_ERROR:
494 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
496 case WRITE_COMPLETED_WITH_ERROR:
497 __drbd_chk_io_error(mdev, DRBD_IO_ERROR);
499 case READ_AHEAD_COMPLETED_WITH_ERROR:
500 /* it is legal to fail READA, no __drbd_chk_io_error in that case. */
501 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
504 case QUEUE_FOR_NET_READ:
505 /* READ or READA, and
507 * or target area marked as invalid,
508 * or just got an io-error. */
509 /* from __drbd_make_request
510 * or from bio_endio during read io-error recovery */
512 /* So we can verify the handle in the answer packet.
513 * Corresponding drbd_remove_request_interval is in
514 * drbd_req_complete() */
515 D_ASSERT(drbd_interval_empty(&req->i));
516 drbd_insert_interval(&mdev->read_requests, &req->i);
518 set_bit(UNPLUG_REMOTE, &mdev->flags);
520 D_ASSERT(req->rq_state & RQ_NET_PENDING);
521 D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0);
522 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
523 req->w.cb = w_send_read_req;
524 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
527 case QUEUE_FOR_NET_WRITE:
528 /* assert something? */
529 /* from __drbd_make_request only */
531 /* Corresponding drbd_remove_request_interval is in
532 * drbd_req_complete() */
533 D_ASSERT(drbd_interval_empty(&req->i));
534 drbd_insert_interval(&mdev->write_requests, &req->i);
537 * In case the req ended up on the transfer log before being
538 * queued on the worker, it could lead to this request being
539 * missed during cleanup after connection loss.
540 * So we have to do both operations here,
541 * within the same lock that protects the transfer log.
543 * _req_add_to_epoch(req); this has to be after the
544 * _maybe_start_new_epoch(req); which happened in
545 * __drbd_make_request, because we now may set the bit
546 * again ourselves to close the current epoch.
548 * Add req to the (now) current epoch (barrier). */
550 /* otherwise we may lose an unplug, which may cause some remote
551 * io-scheduler timeout to expire, increasing maximum latency,
552 * hurting performance. */
553 set_bit(UNPLUG_REMOTE, &mdev->flags);
555 /* queue work item to send data */
556 D_ASSERT(req->rq_state & RQ_NET_PENDING);
557 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
558 req->w.cb = w_send_dblock;
559 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
561 /* close the epoch, in case it outgrew the limit */
563 nc = rcu_dereference(mdev->tconn->net_conf);
564 p = nc->max_epoch_size;
566 if (mdev->tconn->current_tle_writes >= p)
567 start_new_tl_epoch(mdev->tconn);
571 case QUEUE_FOR_SEND_OOS:
572 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
573 req->w.cb = w_send_out_of_sync;
574 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
577 case READ_RETRY_REMOTE_CANCELED:
580 /* real cleanup will be done from tl_clear. just update flags
581 * so it is no longer marked as on the worker queue */
582 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
585 case HANDED_OVER_TO_NETWORK:
586 /* assert something? */
587 if (bio_data_dir(req->master_bio) == WRITE &&
588 !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
589 /* this is what is dangerous about protocol A:
590 * pretend it was successfully written on the peer. */
591 if (req->rq_state & RQ_NET_PENDING)
592 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
593 /* else: neg-ack was faster... */
594 /* it is still not yet RQ_NET_DONE until the
595 * corresponding epoch barrier got acked as well,
596 * so we know what to dirty on connection loss */
598 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
601 case OOS_HANDED_TO_NETWORK:
602 /* Was not set PENDING, no longer QUEUED, so is now DONE
603 * as far as this connection is concerned. */
604 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
607 case CONNECTION_LOST_WHILE_PENDING:
608 /* transfer log cleanup after connection loss */
610 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
615 /* for discarded conflicting writes of multiple primaries,
616 * there is no need to keep anything in the tl, potential
617 * node crashes are covered by the activity log.
619 * If this request had been marked as RQ_POSTPONED before,
620 * it will actually not be discarded, but "restarted",
621 * resubmitted from the retry worker context. */
622 D_ASSERT(req->rq_state & RQ_NET_PENDING);
623 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
624 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
627 case WRITE_ACKED_BY_PEER_AND_SIS:
628 req->rq_state |= RQ_NET_SIS;
629 case WRITE_ACKED_BY_PEER:
630 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
631 /* protocol C; successfully written on peer.
632 * Nothing more to do here.
633 * We want to keep the tl in place for all protocols, to cater
634 * for volatile write-back caches on lower level devices. */
637 case RECV_ACKED_BY_PEER:
638 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
639 /* protocol B; pretends to be successfully written on peer.
640 * see also notes above in HANDED_OVER_TO_NETWORK about
643 D_ASSERT(req->rq_state & RQ_NET_PENDING);
644 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
648 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
649 /* If this node has already detected the write conflict, the
650 * worker will be waiting on misc_wait. Wake it up once this
651 * request has completed locally.
653 D_ASSERT(req->rq_state & RQ_NET_PENDING);
654 req->rq_state |= RQ_POSTPONED;
656 wake_up(&mdev->misc_wait);
657 /* Do not clear RQ_NET_PENDING. This request will make further
658 * progress via restart_conflicting_writes() or
659 * fail_postponed_requests(). Hopefully. */
663 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, RQ_NET_DONE);
666 case FAIL_FROZEN_DISK_IO:
667 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
669 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
672 case RESTART_FROZEN_DISK_IO:
673 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
677 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
681 if (bio_data_dir(req->master_bio) == WRITE)
684 get_ldev(mdev); /* always succeeds in this call path */
685 req->w.cb = w_restart_disk_io;
686 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
690 /* Simply complete (local only) READs. */
691 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
692 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
696 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
697 before the connection loss (B&C only); only P_BARRIER_ACK
698 (or the local completion?) was missing when we suspended.
699 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
700 During connection handshake, we ensure that the peer was not rebooted. */
701 if (!(req->rq_state & RQ_NET_OK)) {
702 /* FIXME could this possibly be a req->w.cb == w_send_out_of_sync?
703 * in that case we must not set RQ_NET_PENDING. */
705 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
707 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
708 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
709 } /* else: FIXME can this happen? */
712 /* else, fall through to BARRIER_ACKED */
715 /* barrier ack for READ requests does not make sense */
716 if (!(req->rq_state & RQ_WRITE))
719 if (req->rq_state & RQ_NET_PENDING) {
720 /* barrier came in before all requests were acked.
721 * this is bad, because if the connection is lost now,
722 * we won't be able to clean them up... */
723 dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
725 /* Allowed to complete requests, even while suspended.
726 * As this is called for all requests within a matching epoch,
727 * we need to filter, and only set RQ_NET_DONE for those that
728 * have actually been on the wire. */
729 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
730 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
734 D_ASSERT(req->rq_state & RQ_NET_PENDING);
735 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
742 /* we may do a local read if:
743 * - we are consistent (of course),
744 * - or we are generally inconsistent,
745 * BUT we are still/already IN SYNC for this area.
746 * since size may be bigger than BM_BLOCK_SIZE,
747 * we may need to check several bits.
749 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
751 unsigned long sbnr, ebnr;
752 sector_t esector, nr_sectors;
754 if (mdev->state.disk == D_UP_TO_DATE)
756 if (mdev->state.disk != D_INCONSISTENT)
758 esector = sector + (size >> 9) - 1;
759 nr_sectors = drbd_get_capacity(mdev->this_bdev);
760 D_ASSERT(sector < nr_sectors);
761 D_ASSERT(esector < nr_sectors);
763 sbnr = BM_SECT_TO_BIT(sector);
764 ebnr = BM_SECT_TO_BIT(esector);
766 return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
769 static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector,
770 enum drbd_read_balancing rbm)
772 struct backing_dev_info *bdi;
776 case RB_CONGESTED_REMOTE:
777 bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
778 return bdi_read_congested(bdi);
779 case RB_LEAST_PENDING:
780 return atomic_read(&mdev->local_cnt) >
781 atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
782 case RB_32K_STRIPING: /* stripe_shift = 15 */
783 case RB_64K_STRIPING:
784 case RB_128K_STRIPING:
785 case RB_256K_STRIPING:
786 case RB_512K_STRIPING:
787 case RB_1M_STRIPING: /* stripe_shift = 20 */
788 stripe_shift = (rbm - RB_32K_STRIPING + 15);
789 return (sector >> (stripe_shift - 9)) & 1;
791 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
792 case RB_PREFER_REMOTE:
794 case RB_PREFER_LOCAL:
801 * complete_conflicting_writes - wait for any conflicting write requests
803 * The write_requests tree contains all active write requests which we
804 * currently know about. Wait for any requests to complete which conflict with
807 * Only way out: remove the conflicting intervals from the tree.
809 static void complete_conflicting_writes(struct drbd_request *req)
812 struct drbd_conf *mdev = req->w.mdev;
813 struct drbd_interval *i;
814 sector_t sector = req->i.sector;
815 int size = req->i.size;
817 i = drbd_find_overlap(&mdev->write_requests, sector, size);
822 prepare_to_wait(&mdev->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
823 i = drbd_find_overlap(&mdev->write_requests, sector, size);
826 /* Indicate to wake up device->misc_wait on progress. */
828 spin_unlock_irq(&mdev->tconn->req_lock);
830 spin_lock_irq(&mdev->tconn->req_lock);
832 finish_wait(&mdev->misc_wait, &wait);
835 /* called within req_lock and rcu_read_lock() */
836 static void maybe_pull_ahead(struct drbd_conf *mdev)
838 struct drbd_tconn *tconn = mdev->tconn;
840 bool congested = false;
841 enum drbd_on_congestion on_congestion;
843 nc = rcu_dereference(tconn->net_conf);
844 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
845 if (on_congestion == OC_BLOCK ||
846 tconn->agreed_pro_version < 96)
849 /* If I don't even have good local storage, we can not reasonably try
850 * to pull ahead of the peer. We also need the local reference to make
851 * sure mdev->act_log is there.
853 if (!get_ldev_if_state(mdev, D_UP_TO_DATE))
857 atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
858 dev_info(DEV, "Congestion-fill threshold reached\n");
862 if (mdev->act_log->used >= nc->cong_extents) {
863 dev_info(DEV, "Congestion-extents threshold reached\n");
868 /* start a new epoch for non-mirrored writes */
869 start_new_tl_epoch(mdev->tconn);
871 if (on_congestion == OC_PULL_AHEAD)
872 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
873 else /*nc->on_congestion == OC_DISCONNECT */
874 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
879 /* If this returns false, and req->private_bio is still set,
880 * this should be submitted locally.
882 * If it returns false, but req->private_bio is not set,
883 * we do not have access to good data :(
885 * Otherwise, this destroys req->private_bio, if any,
888 static bool do_remote_read(struct drbd_request *req)
890 struct drbd_conf *mdev = req->w.mdev;
891 enum drbd_read_balancing rbm;
893 if (req->private_bio) {
894 if (!drbd_may_do_local_read(mdev,
895 req->i.sector, req->i.size)) {
896 bio_put(req->private_bio);
897 req->private_bio = NULL;
902 if (mdev->state.pdsk != D_UP_TO_DATE)
905 if (req->private_bio == NULL)
908 /* TODO: improve read balancing decisions, take into account drbd
909 * protocol, pending requests etc. */
912 rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
915 if (rbm == RB_PREFER_LOCAL && req->private_bio)
916 return false; /* submit locally */
918 if (remote_due_to_read_balancing(mdev, req->i.sector, rbm)) {
919 if (req->private_bio) {
920 bio_put(req->private_bio);
921 req->private_bio = NULL;
930 /* returns number of connections (== 1, for drbd 8.4)
931 * expected to actually write this data,
932 * which does NOT include those that we are L_AHEAD for. */
933 static int drbd_process_write_request(struct drbd_request *req)
935 struct drbd_conf *mdev = req->w.mdev;
936 int remote, send_oos;
939 remote = drbd_should_do_remote(mdev->state);
941 maybe_pull_ahead(mdev);
942 remote = drbd_should_do_remote(mdev->state);
944 send_oos = drbd_should_send_out_of_sync(mdev->state);
947 /* Need to replicate writes. Unless it is an empty flush,
948 * which is better mapped to a DRBD P_BARRIER packet,
949 * also for drbd wire protocol compatibility reasons.
950 * If this was a flush, just start a new epoch.
951 * Unless the current epoch was empty anyways, or we are not currently
952 * replicating, in which case there is no point. */
953 if (unlikely(req->i.size == 0)) {
954 /* The only size==0 bios we expect are empty flushes. */
955 D_ASSERT(req->master_bio->bi_rw & REQ_FLUSH);
957 start_new_tl_epoch(mdev->tconn);
961 if (!remote && !send_oos)
964 D_ASSERT(!(remote && send_oos));
967 _req_mod(req, TO_BE_SENT);
968 _req_mod(req, QUEUE_FOR_NET_WRITE);
969 } else if (drbd_set_out_of_sync(mdev, req->i.sector, req->i.size))
970 _req_mod(req, QUEUE_FOR_SEND_OOS);
976 drbd_submit_req_private_bio(struct drbd_request *req)
978 struct drbd_conf *mdev = req->w.mdev;
979 struct bio *bio = req->private_bio;
980 const int rw = bio_rw(bio);
982 bio->bi_bdev = mdev->ldev->backing_bdev;
984 /* State may have changed since we grabbed our reference on the
985 * ->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,
991 rw == WRITE ? DRBD_FAULT_DT_WR
992 : rw == READ ? DRBD_FAULT_DT_RD
994 bio_endio(bio, -EIO);
996 generic_make_request(bio);
999 bio_endio(bio, -EIO);
1002 void __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
1004 const int rw = bio_rw(bio);
1005 struct bio_and_error m = { NULL, };
1006 struct drbd_request *req;
1007 bool no_remote = false;
1009 /* allocate outside of all locks; */
1010 req = drbd_req_new(mdev, bio);
1013 /* only pass the error to the upper layers.
1014 * if user cannot handle io errors, that's not our business. */
1015 dev_err(DEV, "could not kmalloc() req\n");
1016 bio_endio(bio, -ENOMEM);
1019 req->start_time = start_time;
1021 if (!get_ldev(mdev)) {
1022 bio_put(req->private_bio);
1023 req->private_bio = NULL;
1026 /* For WRITES going to the local disk, grab a reference on the target
1027 * extent. This waits for any resync activity in the corresponding
1028 * resync extent to finish, and, if necessary, pulls in the target
1029 * extent into the activity log, which involves further disk io because
1030 * of transactional on-disk meta data updates.
1031 * Empty flushes don't need to go into the activity log, they can only
1032 * flush data for pending writes which are already in there. */
1033 if (rw == WRITE && req->private_bio && req->i.size
1034 && !test_bit(AL_SUSPENDED, &mdev->flags)) {
1035 req->rq_state |= RQ_IN_ACT_LOG;
1036 drbd_al_begin_io(mdev, &req->i);
1039 spin_lock_irq(&mdev->tconn->req_lock);
1041 /* This may temporarily give up the req_lock,
1042 * but will re-aquire it before it returns here.
1043 * Needs to be before the check on drbd_suspended() */
1044 complete_conflicting_writes(req);
1047 /* no more giving up req_lock from now on! */
1049 if (drbd_suspended(mdev)) {
1050 /* push back and retry: */
1051 req->rq_state |= RQ_POSTPONED;
1052 if (req->private_bio) {
1053 bio_put(req->private_bio);
1054 req->private_bio = NULL;
1059 /* Update disk stats */
1060 _drbd_start_io_acct(mdev, req, bio);
1062 /* We fail READ/READA early, if we can not serve it.
1063 * We must do this before req is registered on any lists.
1064 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1066 if (!do_remote_read(req) && !req->private_bio)
1070 /* which transfer log epoch does this belong to? */
1071 req->epoch = atomic_read(&mdev->tconn->current_tle_nr);
1073 /* no point in adding empty flushes to the transfer log,
1074 * they are mapped to drbd barriers already. */
1075 if (likely(req->i.size!=0)) {
1077 mdev->tconn->current_tle_writes++;
1079 list_add_tail(&req->tl_requests, &mdev->tconn->transfer_log);
1083 if (!drbd_process_write_request(req))
1086 /* We either have a private_bio, or we can read from remote.
1087 * Otherwise we had done the goto nodata above. */
1088 if (req->private_bio == NULL) {
1089 _req_mod(req, TO_BE_SENT);
1090 _req_mod(req, QUEUE_FOR_NET_READ);
1095 if (req->private_bio) {
1096 /* needs to be marked within the same spinlock */
1097 _req_mod(req, TO_BE_SUBMITTED);
1098 /* but we need to give up the spinlock to submit */
1099 spin_unlock_irq(&mdev->tconn->req_lock);
1100 drbd_submit_req_private_bio(req);
1101 spin_lock_irq(&mdev->tconn->req_lock);
1102 } else if (no_remote) {
1104 if (__ratelimit(&drbd_ratelimit_state))
1105 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
1106 /* A write may have been queued for send_oos, however.
1107 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1111 if (drbd_req_put_completion_ref(req, &m, 1))
1112 kref_put(&req->kref, drbd_req_destroy);
1113 spin_unlock_irq(&mdev->tconn->req_lock);
1116 complete_master_bio(mdev, &m);
1120 int drbd_make_request(struct request_queue *q, struct bio *bio)
1122 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1123 unsigned long start_time;
1125 start_time = jiffies;
1128 * what we "blindly" assume:
1130 D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1133 __drbd_make_request(mdev, bio, start_time);
1138 /* This is called by bio_add_page().
1140 * q->max_hw_sectors and other global limits are already enforced there.
1142 * We need to call down to our lower level device,
1143 * in case it has special restrictions.
1145 * We also may need to enforce configured max-bio-bvecs limits.
1147 * As long as the BIO is empty we have to allow at least one bvec,
1148 * regardless of size and offset, so no need to ask lower levels.
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_size = bvm->bi_size;
1154 int limit = DRBD_MAX_BIO_SIZE;
1157 if (bio_size && get_ldev(mdev)) {
1158 struct request_queue * const b =
1159 mdev->ldev->backing_bdev->bd_disk->queue;
1160 if (b->merge_bvec_fn) {
1161 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1162 limit = min(limit, backing_limit);
1169 struct drbd_request *find_oldest_request(struct drbd_tconn *tconn)
1171 /* Walk the transfer log,
1172 * and find the oldest not yet completed request */
1173 struct drbd_request *r;
1174 list_for_each_entry(r, &tconn->transfer_log, tl_requests) {
1175 if (atomic_read(&r->completion_ref))
1181 void request_timer_fn(unsigned long data)
1183 struct drbd_conf *mdev = (struct drbd_conf *) data;
1184 struct drbd_tconn *tconn = mdev->tconn;
1185 struct drbd_request *req; /* oldest request */
1186 struct net_conf *nc;
1187 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1191 nc = rcu_dereference(tconn->net_conf);
1192 if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS)
1193 ent = nc->timeout * HZ/10 * nc->ko_count;
1195 if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */
1196 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1201 et = min_not_zero(dt, ent);
1204 return; /* Recurring timer stopped */
1208 spin_lock_irq(&tconn->req_lock);
1209 req = find_oldest_request(tconn);
1211 spin_unlock_irq(&tconn->req_lock);
1212 mod_timer(&mdev->request_timer, now + et);
1216 /* The request is considered timed out, if
1217 * - we have some effective timeout from the configuration,
1218 * with above state restrictions applied,
1219 * - the oldest request is waiting for a response from the network
1220 * resp. the local disk,
1221 * - the oldest request is in fact older than the effective timeout,
1222 * - the connection was established (resp. disk was attached)
1223 * for longer than the timeout already.
1224 * Note that for 32bit jiffies and very stable connections/disks,
1225 * we may have a wrap around, which is catched by
1226 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1228 * Side effect: once per 32bit wrap-around interval, which means every
1229 * ~198 days with 250 HZ, we have a window where the timeout would need
1230 * to expire twice (worst case) to become effective. Good enough.
1232 if (ent && req->rq_state & RQ_NET_PENDING &&
1233 time_after(now, req->start_time + ent) &&
1234 !time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) {
1235 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1236 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1238 if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev &&
1239 time_after(now, req->start_time + dt) &&
1240 !time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) {
1241 dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1242 __drbd_chk_io_error(mdev, DRBD_FORCE_DETACH);
1244 nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
1245 spin_unlock_irq(&tconn->req_lock);
1246 mod_timer(&mdev->request_timer, nt);