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 tconn->current_tle_writes = 0;
166 atomic_inc(&tconn->current_tle_nr);
167 wake_all_senders(tconn);
170 void complete_master_bio(struct drbd_conf *mdev,
171 struct bio_and_error *m)
173 bio_endio(m->bio, m->error);
178 static void drbd_remove_request_interval(struct rb_root *root,
179 struct drbd_request *req)
181 struct drbd_conf *mdev = req->w.mdev;
182 struct drbd_interval *i = &req->i;
184 drbd_remove_interval(root, i);
186 /* Wake up any processes waiting for this request to complete. */
188 wake_up(&mdev->misc_wait);
191 /* Helper for __req_mod().
192 * Set m->bio to the master bio, if it is fit to be completed,
193 * or leave it alone (it is initialized to NULL in __req_mod),
194 * if it has already been completed, or cannot be completed yet.
195 * If m->bio is set, the error status to be returned is placed in m->error.
198 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
200 const unsigned s = req->rq_state;
201 struct drbd_conf *mdev = req->w.mdev;
205 /* we must not complete the master bio, while it is
206 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
207 * not yet acknowledged by the peer
208 * not yet completed by the local io subsystem
209 * these flags may get cleared in any order by
212 * the bio_endio completion callbacks.
214 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
215 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
216 (s & RQ_COMPLETION_SUSP)) {
217 dev_err(DEV, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
221 if (!req->master_bio) {
222 dev_err(DEV, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
226 rw = bio_rw(req->master_bio);
229 * figure out whether to report success or failure.
231 * report success when at least one of the operations succeeded.
232 * or, to put the other way,
233 * only report failure, when both operations failed.
235 * what to do about the failures is handled elsewhere.
236 * what we need to do here is just: complete the master_bio.
238 * local completion error, if any, has been stored as ERR_PTR
239 * in private_bio within drbd_request_endio.
241 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
242 error = PTR_ERR(req->private_bio);
244 /* remove the request from the conflict detection
245 * respective block_id verification hash */
246 if (!drbd_interval_empty(&req->i)) {
247 struct rb_root *root;
250 root = &mdev->write_requests;
252 root = &mdev->read_requests;
253 drbd_remove_request_interval(root, req);
254 } else if (!(s & RQ_POSTPONED))
255 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
257 /* Before we can signal completion to the upper layers,
258 * we may need to close the current transfer log epoch.
259 * We are within the request lock, so we can simply compare
260 * the request epoch number with the current transfer log
261 * epoch number. If they match, increase the current_tle_nr,
262 * and reset the transfer log epoch write_cnt.
265 req->epoch == atomic_read(&mdev->tconn->current_tle_nr))
266 start_new_tl_epoch(mdev->tconn);
268 /* Update disk stats */
269 _drbd_end_io_acct(mdev, req);
272 * have it be pushed back to the retry work queue,
273 * so it will re-enter __drbd_make_request(),
274 * and be re-assigned to a suitable local or remote path,
275 * or failed if we do not have access to good data anymore.
277 * Unless it was failed early by __drbd_make_request(),
278 * because no path was available, in which case
279 * it was not even added to the transfer_log.
281 * READA may fail, and will not be retried.
283 * WRITE should have used all available paths already.
285 if (!ok && rw == READ && !list_empty(&req->tl_requests))
286 req->rq_state |= RQ_POSTPONED;
288 if (!(req->rq_state & RQ_POSTPONED)) {
289 m->error = ok ? 0 : (error ?: -EIO);
290 m->bio = req->master_bio;
291 req->master_bio = NULL;
295 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
297 struct drbd_conf *mdev = req->w.mdev;
298 D_ASSERT(m || (req->rq_state & RQ_POSTPONED));
300 if (!atomic_sub_and_test(put, &req->completion_ref))
303 if (drbd_suspended(mdev)) {
304 /* We do not allow completion while suspended. Re-get a
305 * reference, so whatever happens when this is resumed
306 * may put and complete. */
308 D_ASSERT(!(req->rq_state & RQ_COMPLETION_SUSP));
309 req->rq_state |= RQ_COMPLETION_SUSP;
310 atomic_inc(&req->completion_ref);
315 drbd_req_complete(req, m);
317 if (req->rq_state & RQ_POSTPONED) {
318 /* don't destroy the req object just yet,
319 * but queue it for retry */
320 drbd_restart_request(req);
327 /* I'd like this to be the only place that manipulates
328 * req->completion_ref and req->kref. */
329 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
332 struct drbd_conf *mdev = req->w.mdev;
333 unsigned s = req->rq_state;
339 req->rq_state &= ~clear;
340 req->rq_state |= set;
343 if (req->rq_state == s)
346 /* intent: get references */
348 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
349 atomic_inc(&req->completion_ref);
351 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
352 inc_ap_pending(mdev);
353 atomic_inc(&req->completion_ref);
356 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
357 atomic_inc(&req->completion_ref);
359 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
360 kref_get(&req->kref); /* wait for the DONE */
362 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
363 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
365 /* progress: put references */
367 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
370 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
371 D_ASSERT(req->rq_state & RQ_LOCAL_PENDING);
372 /* local completion may still come in later,
373 * we need to keep the req object around. */
374 kref_get(&req->kref);
378 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
379 if (req->rq_state & RQ_LOCAL_ABORTED)
385 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
386 dec_ap_pending(mdev);
390 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
393 if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
394 if (req->rq_state & RQ_NET_SENT)
395 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
399 /* potentially complete and destroy */
401 if (k_put || c_put) {
402 /* Completion does it's own kref_put. If we are going to
403 * kref_sub below, we need req to be still around then. */
404 int at_least = k_put + !!c_put;
405 int refcount = atomic_read(&req->kref.refcount);
406 if (refcount < at_least)
408 "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
409 s, req->rq_state, refcount, at_least);
412 /* If we made progress, retry conflicting peer requests, if any. */
414 wake_up(&mdev->misc_wait);
417 k_put += drbd_req_put_completion_ref(req, m, c_put);
419 kref_sub(&req->kref, k_put, drbd_req_destroy);
422 /* obviously this could be coded as many single functions
423 * instead of one huge switch,
424 * or by putting the code directly in the respective locations
425 * (as it has been before).
427 * but having it this way
428 * enforces that it is all in this one place, where it is easier to audit,
429 * it makes it obvious that whatever "event" "happens" to a request should
430 * happen "atomically" within the req_lock,
431 * and it enforces that we have to think in a very structured manner
432 * about the "events" that may happen to a request during its life time ...
434 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
435 struct bio_and_error *m)
437 struct drbd_conf *mdev = req->w.mdev;
446 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
449 /* does not happen...
450 * initialization done in drbd_req_new
455 case TO_BE_SENT: /* via network */
456 /* reached via __drbd_make_request
457 * and from w_read_retry_remote */
458 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
460 nc = rcu_dereference(mdev->tconn->net_conf);
461 p = nc->wire_protocol;
464 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
465 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
466 mod_rq_state(req, m, 0, RQ_NET_PENDING);
469 case TO_BE_SUBMITTED: /* locally */
470 /* reached via __drbd_make_request */
471 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
472 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
476 if (req->rq_state & RQ_WRITE)
477 mdev->writ_cnt += req->i.size >> 9;
479 mdev->read_cnt += req->i.size >> 9;
481 mod_rq_state(req, m, RQ_LOCAL_PENDING,
482 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
486 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
489 case READ_COMPLETED_WITH_ERROR:
490 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
492 case WRITE_COMPLETED_WITH_ERROR:
493 __drbd_chk_io_error(mdev, DRBD_IO_ERROR);
495 case READ_AHEAD_COMPLETED_WITH_ERROR:
496 /* it is legal to fail READA, no __drbd_chk_io_error in that case. */
497 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
500 case QUEUE_FOR_NET_READ:
501 /* READ or READA, and
503 * or target area marked as invalid,
504 * or just got an io-error. */
505 /* from __drbd_make_request
506 * or from bio_endio during read io-error recovery */
508 /* So we can verify the handle in the answer packet.
509 * Corresponding drbd_remove_request_interval is in
510 * drbd_req_complete() */
511 D_ASSERT(drbd_interval_empty(&req->i));
512 drbd_insert_interval(&mdev->read_requests, &req->i);
514 set_bit(UNPLUG_REMOTE, &mdev->flags);
516 D_ASSERT(req->rq_state & RQ_NET_PENDING);
517 D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0);
518 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
519 req->w.cb = w_send_read_req;
520 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
523 case QUEUE_FOR_NET_WRITE:
524 /* assert something? */
525 /* from __drbd_make_request only */
527 /* Corresponding drbd_remove_request_interval is in
528 * drbd_req_complete() */
529 D_ASSERT(drbd_interval_empty(&req->i));
530 drbd_insert_interval(&mdev->write_requests, &req->i);
533 * In case the req ended up on the transfer log before being
534 * queued on the worker, it could lead to this request being
535 * missed during cleanup after connection loss.
536 * So we have to do both operations here,
537 * within the same lock that protects the transfer log.
539 * _req_add_to_epoch(req); this has to be after the
540 * _maybe_start_new_epoch(req); which happened in
541 * __drbd_make_request, because we now may set the bit
542 * again ourselves to close the current epoch.
544 * Add req to the (now) current epoch (barrier). */
546 /* otherwise we may lose an unplug, which may cause some remote
547 * io-scheduler timeout to expire, increasing maximum latency,
548 * hurting performance. */
549 set_bit(UNPLUG_REMOTE, &mdev->flags);
551 /* queue work item to send data */
552 D_ASSERT(req->rq_state & RQ_NET_PENDING);
553 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
554 req->w.cb = w_send_dblock;
555 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
557 /* close the epoch, in case it outgrew the limit */
559 nc = rcu_dereference(mdev->tconn->net_conf);
560 p = nc->max_epoch_size;
562 if (mdev->tconn->current_tle_writes >= p)
563 start_new_tl_epoch(mdev->tconn);
567 case QUEUE_FOR_SEND_OOS:
568 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
569 req->w.cb = w_send_out_of_sync;
570 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
573 case READ_RETRY_REMOTE_CANCELED:
576 /* real cleanup will be done from tl_clear. just update flags
577 * so it is no longer marked as on the worker queue */
578 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
581 case HANDED_OVER_TO_NETWORK:
582 /* assert something? */
583 if (bio_data_dir(req->master_bio) == WRITE &&
584 !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
585 /* this is what is dangerous about protocol A:
586 * pretend it was successfully written on the peer. */
587 if (req->rq_state & RQ_NET_PENDING)
588 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
589 /* else: neg-ack was faster... */
590 /* it is still not yet RQ_NET_DONE until the
591 * corresponding epoch barrier got acked as well,
592 * so we know what to dirty on connection loss */
594 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
597 case OOS_HANDED_TO_NETWORK:
598 /* Was not set PENDING, no longer QUEUED, so is now DONE
599 * as far as this connection is concerned. */
600 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
603 case CONNECTION_LOST_WHILE_PENDING:
604 /* transfer log cleanup after connection loss */
606 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
611 /* for discarded conflicting writes of multiple primaries,
612 * there is no need to keep anything in the tl, potential
613 * node crashes are covered by the activity log.
615 * If this request had been marked as RQ_POSTPONED before,
616 * it will actually not be discarded, but "restarted",
617 * resubmitted from the retry worker context. */
618 D_ASSERT(req->rq_state & RQ_NET_PENDING);
619 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
620 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
623 case WRITE_ACKED_BY_PEER_AND_SIS:
624 req->rq_state |= RQ_NET_SIS;
625 case WRITE_ACKED_BY_PEER:
626 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
627 /* protocol C; successfully written on peer.
628 * Nothing more to do here.
629 * We want to keep the tl in place for all protocols, to cater
630 * for volatile write-back caches on lower level devices. */
633 case RECV_ACKED_BY_PEER:
634 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
635 /* protocol B; pretends to be successfully written on peer.
636 * see also notes above in HANDED_OVER_TO_NETWORK about
639 D_ASSERT(req->rq_state & RQ_NET_PENDING);
640 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
644 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
645 /* If this node has already detected the write conflict, the
646 * worker will be waiting on misc_wait. Wake it up once this
647 * request has completed locally.
649 D_ASSERT(req->rq_state & RQ_NET_PENDING);
650 req->rq_state |= RQ_POSTPONED;
652 wake_up(&mdev->misc_wait);
653 /* Do not clear RQ_NET_PENDING. This request will make further
654 * progress via restart_conflicting_writes() or
655 * fail_postponed_requests(). Hopefully. */
659 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, RQ_NET_DONE);
662 case FAIL_FROZEN_DISK_IO:
663 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
665 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
668 case RESTART_FROZEN_DISK_IO:
669 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
673 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
677 if (bio_data_dir(req->master_bio) == WRITE)
680 get_ldev(mdev); /* always succeeds in this call path */
681 req->w.cb = w_restart_disk_io;
682 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
686 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
687 before the connection loss (B&C only); only P_BARRIER_ACK
688 (or the local completion?) was missing when we suspended.
689 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
690 During connection handshake, we ensure that the peer was not rebooted. */
691 if (!(req->rq_state & RQ_NET_OK)) {
692 /* FIXME could this possibly be a req->w.cb == w_send_out_of_sync?
693 * in that case we must not set RQ_NET_PENDING. */
695 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
697 drbd_queue_work(&mdev->tconn->sender_work, &req->w);
698 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
699 } /* else: FIXME can this happen? */
702 /* else, fall through to BARRIER_ACKED */
705 /* barrier ack for READ requests does not make sense */
706 if (!(req->rq_state & RQ_WRITE))
709 if (req->rq_state & RQ_NET_PENDING) {
710 /* barrier came in before all requests were acked.
711 * this is bad, because if the connection is lost now,
712 * we won't be able to clean them up... */
713 dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
715 /* Allowed to complete requests, even while suspended.
716 * As this is called for all requests within a matching epoch,
717 * we need to filter, and only set RQ_NET_DONE for those that
718 * have actually been on the wire. */
719 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
720 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
724 D_ASSERT(req->rq_state & RQ_NET_PENDING);
725 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
732 /* we may do a local read if:
733 * - we are consistent (of course),
734 * - or we are generally inconsistent,
735 * BUT we are still/already IN SYNC for this area.
736 * since size may be bigger than BM_BLOCK_SIZE,
737 * we may need to check several bits.
739 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
741 unsigned long sbnr, ebnr;
742 sector_t esector, nr_sectors;
744 if (mdev->state.disk == D_UP_TO_DATE)
746 if (mdev->state.disk != D_INCONSISTENT)
748 esector = sector + (size >> 9) - 1;
749 nr_sectors = drbd_get_capacity(mdev->this_bdev);
750 D_ASSERT(sector < nr_sectors);
751 D_ASSERT(esector < nr_sectors);
753 sbnr = BM_SECT_TO_BIT(sector);
754 ebnr = BM_SECT_TO_BIT(esector);
756 return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
759 static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector,
760 enum drbd_read_balancing rbm)
762 struct backing_dev_info *bdi;
766 case RB_CONGESTED_REMOTE:
767 bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
768 return bdi_read_congested(bdi);
769 case RB_LEAST_PENDING:
770 return atomic_read(&mdev->local_cnt) >
771 atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
772 case RB_32K_STRIPING: /* stripe_shift = 15 */
773 case RB_64K_STRIPING:
774 case RB_128K_STRIPING:
775 case RB_256K_STRIPING:
776 case RB_512K_STRIPING:
777 case RB_1M_STRIPING: /* stripe_shift = 20 */
778 stripe_shift = (rbm - RB_32K_STRIPING + 15);
779 return (sector >> (stripe_shift - 9)) & 1;
781 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
782 case RB_PREFER_REMOTE:
784 case RB_PREFER_LOCAL:
791 * complete_conflicting_writes - wait for any conflicting write requests
793 * The write_requests tree contains all active write requests which we
794 * currently know about. Wait for any requests to complete which conflict with
797 * Only way out: remove the conflicting intervals from the tree.
799 static void complete_conflicting_writes(struct drbd_request *req)
802 struct drbd_conf *mdev = req->w.mdev;
803 struct drbd_interval *i;
804 sector_t sector = req->i.sector;
805 int size = req->i.size;
807 i = drbd_find_overlap(&mdev->write_requests, sector, size);
812 prepare_to_wait(&mdev->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
813 i = drbd_find_overlap(&mdev->write_requests, sector, size);
816 /* Indicate to wake up device->misc_wait on progress. */
818 spin_unlock_irq(&mdev->tconn->req_lock);
820 spin_lock_irq(&mdev->tconn->req_lock);
822 finish_wait(&mdev->misc_wait, &wait);
825 /* called within req_lock and rcu_read_lock() */
826 static void maybe_pull_ahead(struct drbd_conf *mdev)
828 struct drbd_tconn *tconn = mdev->tconn;
830 bool congested = false;
831 enum drbd_on_congestion on_congestion;
833 nc = rcu_dereference(tconn->net_conf);
834 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
835 if (on_congestion == OC_BLOCK ||
836 tconn->agreed_pro_version < 96)
839 /* If I don't even have good local storage, we can not reasonably try
840 * to pull ahead of the peer. We also need the local reference to make
841 * sure mdev->act_log is there.
843 if (!get_ldev_if_state(mdev, D_UP_TO_DATE))
847 atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
848 dev_info(DEV, "Congestion-fill threshold reached\n");
852 if (mdev->act_log->used >= nc->cong_extents) {
853 dev_info(DEV, "Congestion-extents threshold reached\n");
858 if (mdev->tconn->current_tle_writes)
859 /* start a new epoch for non-mirrored writes */
860 start_new_tl_epoch(mdev->tconn);
862 if (on_congestion == OC_PULL_AHEAD)
863 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
864 else /*nc->on_congestion == OC_DISCONNECT */
865 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
870 /* If this returns false, and req->private_bio is still set,
871 * this should be submitted locally.
873 * If it returns false, but req->private_bio is not set,
874 * we do not have access to good data :(
876 * Otherwise, this destroys req->private_bio, if any,
879 static bool do_remote_read(struct drbd_request *req)
881 struct drbd_conf *mdev = req->w.mdev;
882 enum drbd_read_balancing rbm;
884 if (req->private_bio) {
885 if (!drbd_may_do_local_read(mdev,
886 req->i.sector, req->i.size)) {
887 bio_put(req->private_bio);
888 req->private_bio = NULL;
893 if (mdev->state.pdsk != D_UP_TO_DATE)
896 if (req->private_bio == NULL)
899 /* TODO: improve read balancing decisions, take into account drbd
900 * protocol, pending requests etc. */
903 rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
906 if (rbm == RB_PREFER_LOCAL && req->private_bio)
907 return false; /* submit locally */
909 if (remote_due_to_read_balancing(mdev, req->i.sector, rbm)) {
910 if (req->private_bio) {
911 bio_put(req->private_bio);
912 req->private_bio = NULL;
921 /* returns number of connections (== 1, for drbd 8.4)
922 * expected to actually write this data,
923 * which does NOT include those that we are L_AHEAD for. */
924 static int drbd_process_write_request(struct drbd_request *req)
926 struct drbd_conf *mdev = req->w.mdev;
927 int remote, send_oos;
930 remote = drbd_should_do_remote(mdev->state);
932 maybe_pull_ahead(mdev);
933 remote = drbd_should_do_remote(mdev->state);
935 send_oos = drbd_should_send_out_of_sync(mdev->state);
938 /* Need to replicate writes. Unless it is an empty flush,
939 * which is better mapped to a DRBD P_BARRIER packet,
940 * also for drbd wire protocol compatibility reasons.
941 * If this was a flush, just start a new epoch.
942 * Unless the current epoch was empty anyways, or we are not currently
943 * replicating, in which case there is no point. */
944 if (unlikely(req->i.size == 0)) {
945 /* The only size==0 bios we expect are empty flushes. */
946 D_ASSERT(req->master_bio->bi_rw & REQ_FLUSH);
947 if (remote && mdev->tconn->current_tle_writes)
948 start_new_tl_epoch(mdev->tconn);
952 if (!remote && !send_oos)
955 D_ASSERT(!(remote && send_oos));
958 _req_mod(req, TO_BE_SENT);
959 _req_mod(req, QUEUE_FOR_NET_WRITE);
960 } else if (drbd_set_out_of_sync(mdev, req->i.sector, req->i.size))
961 _req_mod(req, QUEUE_FOR_SEND_OOS);
967 drbd_submit_req_private_bio(struct drbd_request *req)
969 struct drbd_conf *mdev = req->w.mdev;
970 struct bio *bio = req->private_bio;
971 const int rw = bio_rw(bio);
973 bio->bi_bdev = mdev->ldev->backing_bdev;
975 /* State may have changed since we grabbed our reference on the
976 * ->ldev member. Double check, and short-circuit to endio.
977 * In case the last activity log transaction failed to get on
978 * stable storage, and this is a WRITE, we may not even submit
980 if (get_ldev(mdev)) {
981 if (drbd_insert_fault(mdev,
982 rw == WRITE ? DRBD_FAULT_DT_WR
983 : rw == READ ? DRBD_FAULT_DT_RD
985 bio_endio(bio, -EIO);
987 generic_make_request(bio);
990 bio_endio(bio, -EIO);
993 void __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
995 const int rw = bio_rw(bio);
996 struct bio_and_error m = { NULL, };
997 struct drbd_request *req;
998 bool no_remote = false;
1000 /* allocate outside of all locks; */
1001 req = drbd_req_new(mdev, bio);
1004 /* only pass the error to the upper layers.
1005 * if user cannot handle io errors, that's not our business. */
1006 dev_err(DEV, "could not kmalloc() req\n");
1007 bio_endio(bio, -ENOMEM);
1010 req->start_time = start_time;
1012 if (!get_ldev(mdev)) {
1013 bio_put(req->private_bio);
1014 req->private_bio = NULL;
1017 /* For WRITES going to the local disk, grab a reference on the target
1018 * extent. This waits for any resync activity in the corresponding
1019 * resync extent to finish, and, if necessary, pulls in the target
1020 * extent into the activity log, which involves further disk io because
1021 * of transactional on-disk meta data updates.
1022 * Empty flushes don't need to go into the activity log, they can only
1023 * flush data for pending writes which are already in there. */
1024 if (rw == WRITE && req->private_bio && req->i.size
1025 && !test_bit(AL_SUSPENDED, &mdev->flags)) {
1026 req->rq_state |= RQ_IN_ACT_LOG;
1027 drbd_al_begin_io(mdev, &req->i);
1030 spin_lock_irq(&mdev->tconn->req_lock);
1032 /* This may temporarily give up the req_lock,
1033 * but will re-aquire it before it returns here.
1034 * Needs to be before the check on drbd_suspended() */
1035 complete_conflicting_writes(req);
1038 /* no more giving up req_lock from now on! */
1040 if (drbd_suspended(mdev)) {
1041 /* push back and retry: */
1042 req->rq_state |= RQ_POSTPONED;
1043 if (req->private_bio) {
1044 bio_put(req->private_bio);
1045 req->private_bio = NULL;
1050 /* Update disk stats */
1051 _drbd_start_io_acct(mdev, req, bio);
1053 /* We fail READ/READA early, if we can not serve it.
1054 * We must do this before req is registered on any lists.
1055 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1057 if (!do_remote_read(req) && !req->private_bio)
1061 /* which transfer log epoch does this belong to? */
1062 req->epoch = atomic_read(&mdev->tconn->current_tle_nr);
1064 mdev->tconn->current_tle_writes++;
1066 /* no point in adding empty flushes to the transfer log,
1067 * they are mapped to drbd barriers already. */
1068 if (likely(req->i.size!=0))
1069 list_add_tail(&req->tl_requests, &mdev->tconn->transfer_log);
1072 if (!drbd_process_write_request(req))
1075 /* We either have a private_bio, or we can read from remote.
1076 * Otherwise we had done the goto nodata above. */
1077 if (req->private_bio == NULL) {
1078 _req_mod(req, TO_BE_SENT);
1079 _req_mod(req, QUEUE_FOR_NET_READ);
1084 if (req->private_bio) {
1085 /* needs to be marked within the same spinlock */
1086 _req_mod(req, TO_BE_SUBMITTED);
1087 /* but we need to give up the spinlock to submit */
1088 spin_unlock_irq(&mdev->tconn->req_lock);
1089 drbd_submit_req_private_bio(req);
1090 spin_lock_irq(&mdev->tconn->req_lock);
1091 } else if (no_remote) {
1093 if (__ratelimit(&drbd_ratelimit_state))
1094 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
1095 /* A write may have been queued for send_oos, however.
1096 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1100 if (drbd_req_put_completion_ref(req, &m, 1))
1101 kref_put(&req->kref, drbd_req_destroy);
1102 spin_unlock_irq(&mdev->tconn->req_lock);
1105 complete_master_bio(mdev, &m);
1109 int drbd_make_request(struct request_queue *q, struct bio *bio)
1111 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1112 unsigned long start_time;
1114 start_time = jiffies;
1117 * what we "blindly" assume:
1119 D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1122 __drbd_make_request(mdev, bio, start_time);
1127 /* This is called by bio_add_page().
1129 * q->max_hw_sectors and other global limits are already enforced there.
1131 * We need to call down to our lower level device,
1132 * in case it has special restrictions.
1134 * We also may need to enforce configured max-bio-bvecs limits.
1136 * As long as the BIO is empty we have to allow at least one bvec,
1137 * regardless of size and offset, so no need to ask lower levels.
1139 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1141 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1142 unsigned int bio_size = bvm->bi_size;
1143 int limit = DRBD_MAX_BIO_SIZE;
1146 if (bio_size && get_ldev(mdev)) {
1147 struct request_queue * const b =
1148 mdev->ldev->backing_bdev->bd_disk->queue;
1149 if (b->merge_bvec_fn) {
1150 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1151 limit = min(limit, backing_limit);
1158 struct drbd_request *find_oldest_request(struct drbd_tconn *tconn)
1160 /* Walk the transfer log,
1161 * and find the oldest not yet completed request */
1162 struct drbd_request *r;
1163 list_for_each_entry(r, &tconn->transfer_log, tl_requests) {
1164 if (atomic_read(&r->completion_ref))
1170 void request_timer_fn(unsigned long data)
1172 struct drbd_conf *mdev = (struct drbd_conf *) data;
1173 struct drbd_tconn *tconn = mdev->tconn;
1174 struct drbd_request *req; /* oldest request */
1175 struct net_conf *nc;
1176 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1180 nc = rcu_dereference(tconn->net_conf);
1181 if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS)
1182 ent = nc->timeout * HZ/10 * nc->ko_count;
1184 if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */
1185 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1190 et = min_not_zero(dt, ent);
1193 return; /* Recurring timer stopped */
1197 spin_lock_irq(&tconn->req_lock);
1198 req = find_oldest_request(tconn);
1200 spin_unlock_irq(&tconn->req_lock);
1201 mod_timer(&mdev->request_timer, now + et);
1205 /* The request is considered timed out, if
1206 * - we have some effective timeout from the configuration,
1207 * with above state restrictions applied,
1208 * - the oldest request is waiting for a response from the network
1209 * resp. the local disk,
1210 * - the oldest request is in fact older than the effective timeout,
1211 * - the connection was established (resp. disk was attached)
1212 * for longer than the timeout already.
1213 * Note that for 32bit jiffies and very stable connections/disks,
1214 * we may have a wrap around, which is catched by
1215 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1217 * Side effect: once per 32bit wrap-around interval, which means every
1218 * ~198 days with 250 HZ, we have a window where the timeout would need
1219 * to expire twice (worst case) to become effective. Good enough.
1221 if (ent && req->rq_state & RQ_NET_PENDING &&
1222 time_after(now, req->start_time + ent) &&
1223 !time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) {
1224 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1225 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1227 if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev &&
1228 time_after(now, req->start_time + dt) &&
1229 !time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) {
1230 dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1231 __drbd_chk_io_error(mdev, DRBD_FORCE_DETACH);
1233 nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
1234 spin_unlock_irq(&tconn->req_lock);
1235 mod_timer(&mdev->request_timer, nt);