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_device *device, sector_t sector, int size);
36 /* Update disk stats at start of I/O request */
37 static void _drbd_start_io_acct(struct drbd_device *device, struct drbd_request *req)
39 const int rw = bio_data_dir(req->master_bio);
41 cpu = part_stat_lock();
42 part_round_stats(cpu, &device->vdisk->part0);
43 part_stat_inc(cpu, &device->vdisk->part0, ios[rw]);
44 part_stat_add(cpu, &device->vdisk->part0, sectors[rw], req->i.size >> 9);
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(&device->vdisk->part0, rw);
51 /* Update disk stats when completing request upwards */
52 static void _drbd_end_io_acct(struct drbd_device *device, 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, &device->vdisk->part0, ticks[rw], duration);
59 part_round_stats(cpu, &device->vdisk->part0);
60 part_dec_in_flight(&device->vdisk->part0, rw);
64 static struct drbd_request *drbd_req_new(struct drbd_device *device,
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_iter.bi_sector;
81 req->i.size = bio_src->bi_iter.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_device *device = req->device;
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 drbd_err(device, "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 */
127 /* There is a special case:
128 * we may notice late that IO was suspended,
129 * and postpone, or schedule for retry, a write,
130 * before it even was submitted or sent.
131 * In that case we do not want to touch the bitmap at all.
133 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
134 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
135 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
137 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
138 drbd_set_in_sync(device, req->i.sector, req->i.size);
141 /* one might be tempted to move the drbd_al_complete_io
142 * to the local io completion callback drbd_request_endio.
143 * but, if this was a mirror write, we may only
144 * drbd_al_complete_io after this is RQ_NET_DONE,
145 * otherwise the extent could be dropped from the al
146 * before it has actually been written on the peer.
147 * if we crash before our peer knows about the request,
148 * but after the extent has been dropped from the al,
149 * we would forget to resync the corresponding extent.
151 if (s & RQ_IN_ACT_LOG) {
152 if (get_ldev_if_state(device, D_FAILED)) {
153 drbd_al_complete_io(device, &req->i);
155 } else if (__ratelimit(&drbd_ratelimit_state)) {
156 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
157 "but my Disk seems to have failed :(\n",
158 (unsigned long long) req->i.sector, req->i.size);
163 mempool_free(req, drbd_request_mempool);
166 static void wake_all_senders(struct drbd_connection *connection)
168 wake_up(&connection->sender_work.q_wait);
171 /* must hold resource->req_lock */
172 void start_new_tl_epoch(struct drbd_connection *connection)
174 /* no point closing an epoch, if it is empty, anyways. */
175 if (connection->current_tle_writes == 0)
178 connection->current_tle_writes = 0;
179 atomic_inc(&connection->current_tle_nr);
180 wake_all_senders(connection);
183 void complete_master_bio(struct drbd_device *device,
184 struct bio_and_error *m)
186 bio_endio(m->bio, m->error);
191 static void drbd_remove_request_interval(struct rb_root *root,
192 struct drbd_request *req)
194 struct drbd_device *device = req->device;
195 struct drbd_interval *i = &req->i;
197 drbd_remove_interval(root, i);
199 /* Wake up any processes waiting for this request to complete. */
201 wake_up(&device->misc_wait);
204 /* Helper for __req_mod().
205 * Set m->bio to the master bio, if it is fit to be completed,
206 * or leave it alone (it is initialized to NULL in __req_mod),
207 * if it has already been completed, or cannot be completed yet.
208 * If m->bio is set, the error status to be returned is placed in m->error.
211 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
213 const unsigned s = req->rq_state;
214 struct drbd_device *device = req->device;
218 /* we must not complete the master bio, while it is
219 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
220 * not yet acknowledged by the peer
221 * not yet completed by the local io subsystem
222 * these flags may get cleared in any order by
225 * the bio_endio completion callbacks.
227 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
228 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
229 (s & RQ_COMPLETION_SUSP)) {
230 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
234 if (!req->master_bio) {
235 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
239 rw = bio_rw(req->master_bio);
242 * figure out whether to report success or failure.
244 * report success when at least one of the operations succeeded.
245 * or, to put the other way,
246 * only report failure, when both operations failed.
248 * what to do about the failures is handled elsewhere.
249 * what we need to do here is just: complete the master_bio.
251 * local completion error, if any, has been stored as ERR_PTR
252 * in private_bio within drbd_request_endio.
254 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
255 error = PTR_ERR(req->private_bio);
257 /* remove the request from the conflict detection
258 * respective block_id verification hash */
259 if (!drbd_interval_empty(&req->i)) {
260 struct rb_root *root;
263 root = &device->write_requests;
265 root = &device->read_requests;
266 drbd_remove_request_interval(root, req);
269 /* Before we can signal completion to the upper layers,
270 * we may need to close the current transfer log epoch.
271 * We are within the request lock, so we can simply compare
272 * the request epoch number with the current transfer log
273 * epoch number. If they match, increase the current_tle_nr,
274 * and reset the transfer log epoch write_cnt.
277 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
278 start_new_tl_epoch(first_peer_device(device)->connection);
280 /* Update disk stats */
281 _drbd_end_io_acct(device, req);
284 * have it be pushed back to the retry work queue,
285 * so it will re-enter __drbd_make_request(),
286 * and be re-assigned to a suitable local or remote path,
287 * or failed if we do not have access to good data anymore.
289 * Unless it was failed early by __drbd_make_request(),
290 * because no path was available, in which case
291 * it was not even added to the transfer_log.
293 * READA may fail, and will not be retried.
295 * WRITE should have used all available paths already.
297 if (!ok && rw == READ && !list_empty(&req->tl_requests))
298 req->rq_state |= RQ_POSTPONED;
300 if (!(req->rq_state & RQ_POSTPONED)) {
301 m->error = ok ? 0 : (error ?: -EIO);
302 m->bio = req->master_bio;
303 req->master_bio = NULL;
307 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
309 struct drbd_device *device = req->device;
310 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
312 if (!atomic_sub_and_test(put, &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_device *device = req->device;
333 unsigned s = req->rq_state;
337 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
338 set |= RQ_COMPLETION_SUSP;
342 req->rq_state &= ~clear;
343 req->rq_state |= set;
346 if (req->rq_state == s)
349 /* intent: get references */
351 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
352 atomic_inc(&req->completion_ref);
354 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
355 inc_ap_pending(device);
356 atomic_inc(&req->completion_ref);
359 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
360 atomic_inc(&req->completion_ref);
362 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
363 kref_get(&req->kref); /* wait for the DONE */
365 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
366 atomic_add(req->i.size >> 9, &device->ap_in_flight);
368 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
369 atomic_inc(&req->completion_ref);
371 /* progress: put references */
373 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
376 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
377 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
378 /* local completion may still come in later,
379 * we need to keep the req object around. */
380 kref_get(&req->kref);
384 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
385 if (req->rq_state & RQ_LOCAL_ABORTED)
391 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
392 dec_ap_pending(device);
396 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
399 if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
400 if (req->rq_state & RQ_NET_SENT)
401 atomic_sub(req->i.size >> 9, &device->ap_in_flight);
405 /* potentially complete and destroy */
407 if (k_put || c_put) {
408 /* Completion does it's own kref_put. If we are going to
409 * kref_sub below, we need req to be still around then. */
410 int at_least = k_put + !!c_put;
411 int refcount = atomic_read(&req->kref.refcount);
412 if (refcount < at_least)
414 "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
415 s, req->rq_state, refcount, at_least);
418 /* If we made progress, retry conflicting peer requests, if any. */
420 wake_up(&device->misc_wait);
423 k_put += drbd_req_put_completion_ref(req, m, c_put);
425 kref_sub(&req->kref, k_put, drbd_req_destroy);
428 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
430 char b[BDEVNAME_SIZE];
432 if (!__ratelimit(&drbd_ratelimit_state))
435 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
436 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
437 (unsigned long long)req->i.sector,
439 bdevname(device->ldev->backing_bdev, b));
442 /* obviously this could be coded as many single functions
443 * instead of one huge switch,
444 * or by putting the code directly in the respective locations
445 * (as it has been before).
447 * but having it this way
448 * enforces that it is all in this one place, where it is easier to audit,
449 * it makes it obvious that whatever "event" "happens" to a request should
450 * happen "atomically" within the req_lock,
451 * and it enforces that we have to think in a very structured manner
452 * about the "events" that may happen to a request during its life time ...
454 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
455 struct bio_and_error *m)
457 struct drbd_device *device = req->device;
466 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
469 /* does not happen...
470 * initialization done in drbd_req_new
475 case TO_BE_SENT: /* via network */
476 /* reached via __drbd_make_request
477 * and from w_read_retry_remote */
478 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
480 nc = rcu_dereference(first_peer_device(device)->connection->net_conf);
481 p = nc->wire_protocol;
484 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
485 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
486 mod_rq_state(req, m, 0, RQ_NET_PENDING);
489 case TO_BE_SUBMITTED: /* locally */
490 /* reached via __drbd_make_request */
491 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
492 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
496 if (req->rq_state & RQ_WRITE)
497 device->writ_cnt += req->i.size >> 9;
499 device->read_cnt += req->i.size >> 9;
501 mod_rq_state(req, m, RQ_LOCAL_PENDING,
502 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
506 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
509 case WRITE_COMPLETED_WITH_ERROR:
510 drbd_report_io_error(device, req);
511 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
512 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
515 case READ_COMPLETED_WITH_ERROR:
516 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
517 drbd_report_io_error(device, req);
518 __drbd_chk_io_error(device, DRBD_READ_ERROR);
520 case READ_AHEAD_COMPLETED_WITH_ERROR:
521 /* it is legal to fail READA, no __drbd_chk_io_error in that case. */
522 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
525 case DISCARD_COMPLETED_NOTSUPP:
526 case DISCARD_COMPLETED_WITH_ERROR:
527 /* I'd rather not detach from local disk just because it
528 * failed a REQ_DISCARD. */
529 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
532 case QUEUE_FOR_NET_READ:
533 /* READ or READA, and
535 * or target area marked as invalid,
536 * or just got an io-error. */
537 /* from __drbd_make_request
538 * or from bio_endio during read io-error recovery */
540 /* So we can verify the handle in the answer packet.
541 * Corresponding drbd_remove_request_interval is in
542 * drbd_req_complete() */
543 D_ASSERT(device, drbd_interval_empty(&req->i));
544 drbd_insert_interval(&device->read_requests, &req->i);
546 set_bit(UNPLUG_REMOTE, &device->flags);
548 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
549 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
550 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
551 req->w.cb = w_send_read_req;
552 drbd_queue_work(&first_peer_device(device)->connection->sender_work,
556 case QUEUE_FOR_NET_WRITE:
557 /* assert something? */
558 /* from __drbd_make_request only */
560 /* Corresponding drbd_remove_request_interval is in
561 * drbd_req_complete() */
562 D_ASSERT(device, drbd_interval_empty(&req->i));
563 drbd_insert_interval(&device->write_requests, &req->i);
566 * In case the req ended up on the transfer log before being
567 * queued on the worker, it could lead to this request being
568 * missed during cleanup after connection loss.
569 * So we have to do both operations here,
570 * within the same lock that protects the transfer log.
572 * _req_add_to_epoch(req); this has to be after the
573 * _maybe_start_new_epoch(req); which happened in
574 * __drbd_make_request, because we now may set the bit
575 * again ourselves to close the current epoch.
577 * Add req to the (now) current epoch (barrier). */
579 /* otherwise we may lose an unplug, which may cause some remote
580 * io-scheduler timeout to expire, increasing maximum latency,
581 * hurting performance. */
582 set_bit(UNPLUG_REMOTE, &device->flags);
584 /* queue work item to send data */
585 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
586 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
587 req->w.cb = w_send_dblock;
588 drbd_queue_work(&first_peer_device(device)->connection->sender_work,
591 /* close the epoch, in case it outgrew the limit */
593 nc = rcu_dereference(first_peer_device(device)->connection->net_conf);
594 p = nc->max_epoch_size;
596 if (first_peer_device(device)->connection->current_tle_writes >= p)
597 start_new_tl_epoch(first_peer_device(device)->connection);
601 case QUEUE_FOR_SEND_OOS:
602 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
603 req->w.cb = w_send_out_of_sync;
604 drbd_queue_work(&first_peer_device(device)->connection->sender_work,
608 case READ_RETRY_REMOTE_CANCELED:
611 /* real cleanup will be done from tl_clear. just update flags
612 * so it is no longer marked as on the worker queue */
613 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
616 case HANDED_OVER_TO_NETWORK:
617 /* assert something? */
618 if (bio_data_dir(req->master_bio) == WRITE &&
619 !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
620 /* this is what is dangerous about protocol A:
621 * pretend it was successfully written on the peer. */
622 if (req->rq_state & RQ_NET_PENDING)
623 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
624 /* else: neg-ack was faster... */
625 /* it is still not yet RQ_NET_DONE until the
626 * corresponding epoch barrier got acked as well,
627 * so we know what to dirty on connection loss */
629 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
632 case OOS_HANDED_TO_NETWORK:
633 /* Was not set PENDING, no longer QUEUED, so is now DONE
634 * as far as this connection is concerned. */
635 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
638 case CONNECTION_LOST_WHILE_PENDING:
639 /* transfer log cleanup after connection loss */
641 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
645 case CONFLICT_RESOLVED:
646 /* for superseded conflicting writes of multiple primaries,
647 * there is no need to keep anything in the tl, potential
648 * node crashes are covered by the activity log.
650 * If this request had been marked as RQ_POSTPONED before,
651 * it will actually not be completed, but "restarted",
652 * resubmitted from the retry worker context. */
653 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
654 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
655 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
658 case WRITE_ACKED_BY_PEER_AND_SIS:
659 req->rq_state |= RQ_NET_SIS;
660 case WRITE_ACKED_BY_PEER:
661 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
662 /* protocol C; successfully written on peer.
663 * Nothing more to do here.
664 * We want to keep the tl in place for all protocols, to cater
665 * for volatile write-back caches on lower level devices. */
668 case RECV_ACKED_BY_PEER:
669 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
670 /* protocol B; pretends to be successfully written on peer.
671 * see also notes above in HANDED_OVER_TO_NETWORK about
674 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
675 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
679 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
680 /* If this node has already detected the write conflict, the
681 * worker will be waiting on misc_wait. Wake it up once this
682 * request has completed locally.
684 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
685 req->rq_state |= RQ_POSTPONED;
687 wake_up(&device->misc_wait);
688 /* Do not clear RQ_NET_PENDING. This request will make further
689 * progress via restart_conflicting_writes() or
690 * fail_postponed_requests(). Hopefully. */
694 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
697 case FAIL_FROZEN_DISK_IO:
698 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
700 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
703 case RESTART_FROZEN_DISK_IO:
704 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
708 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
712 if (bio_data_dir(req->master_bio) == WRITE)
715 get_ldev(device); /* always succeeds in this call path */
716 req->w.cb = w_restart_disk_io;
717 drbd_queue_work(&first_peer_device(device)->connection->sender_work,
722 /* Simply complete (local only) READs. */
723 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
724 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
728 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
729 before the connection loss (B&C only); only P_BARRIER_ACK
730 (or the local completion?) was missing when we suspended.
731 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
732 During connection handshake, we ensure that the peer was not rebooted. */
733 if (!(req->rq_state & RQ_NET_OK)) {
734 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
735 * in that case we must not set RQ_NET_PENDING. */
737 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
739 drbd_queue_work(&first_peer_device(device)->connection->sender_work,
741 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
742 } /* else: FIXME can this happen? */
745 /* else, fall through to BARRIER_ACKED */
748 /* barrier ack for READ requests does not make sense */
749 if (!(req->rq_state & RQ_WRITE))
752 if (req->rq_state & RQ_NET_PENDING) {
753 /* barrier came in before all requests were acked.
754 * this is bad, because if the connection is lost now,
755 * we won't be able to clean them up... */
756 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
758 /* Allowed to complete requests, even while suspended.
759 * As this is called for all requests within a matching epoch,
760 * we need to filter, and only set RQ_NET_DONE for those that
761 * have actually been on the wire. */
762 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
763 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
767 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
768 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
771 case QUEUE_AS_DRBD_BARRIER:
772 start_new_tl_epoch(first_peer_device(device)->connection);
773 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
780 /* we may do a local read if:
781 * - we are consistent (of course),
782 * - or we are generally inconsistent,
783 * BUT we are still/already IN SYNC for this area.
784 * since size may be bigger than BM_BLOCK_SIZE,
785 * we may need to check several bits.
787 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
789 unsigned long sbnr, ebnr;
790 sector_t esector, nr_sectors;
792 if (device->state.disk == D_UP_TO_DATE)
794 if (device->state.disk != D_INCONSISTENT)
796 esector = sector + (size >> 9) - 1;
797 nr_sectors = drbd_get_capacity(device->this_bdev);
798 D_ASSERT(device, sector < nr_sectors);
799 D_ASSERT(device, esector < nr_sectors);
801 sbnr = BM_SECT_TO_BIT(sector);
802 ebnr = BM_SECT_TO_BIT(esector);
804 return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
807 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
808 enum drbd_read_balancing rbm)
810 struct backing_dev_info *bdi;
814 case RB_CONGESTED_REMOTE:
815 bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
816 return bdi_read_congested(bdi);
817 case RB_LEAST_PENDING:
818 return atomic_read(&device->local_cnt) >
819 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
820 case RB_32K_STRIPING: /* stripe_shift = 15 */
821 case RB_64K_STRIPING:
822 case RB_128K_STRIPING:
823 case RB_256K_STRIPING:
824 case RB_512K_STRIPING:
825 case RB_1M_STRIPING: /* stripe_shift = 20 */
826 stripe_shift = (rbm - RB_32K_STRIPING + 15);
827 return (sector >> (stripe_shift - 9)) & 1;
829 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
830 case RB_PREFER_REMOTE:
832 case RB_PREFER_LOCAL:
839 * complete_conflicting_writes - wait for any conflicting write requests
841 * The write_requests tree contains all active write requests which we
842 * currently know about. Wait for any requests to complete which conflict with
845 * Only way out: remove the conflicting intervals from the tree.
847 static void complete_conflicting_writes(struct drbd_request *req)
850 struct drbd_device *device = req->device;
851 struct drbd_interval *i;
852 sector_t sector = req->i.sector;
853 int size = req->i.size;
855 i = drbd_find_overlap(&device->write_requests, sector, size);
860 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
861 i = drbd_find_overlap(&device->write_requests, sector, size);
864 /* Indicate to wake up device->misc_wait on progress. */
866 spin_unlock_irq(&device->resource->req_lock);
868 spin_lock_irq(&device->resource->req_lock);
870 finish_wait(&device->misc_wait, &wait);
873 /* called within req_lock and rcu_read_lock() */
874 static void maybe_pull_ahead(struct drbd_device *device)
876 struct drbd_connection *connection = first_peer_device(device)->connection;
878 bool congested = false;
879 enum drbd_on_congestion on_congestion;
882 nc = rcu_dereference(connection->net_conf);
883 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
885 if (on_congestion == OC_BLOCK ||
886 connection->agreed_pro_version < 96)
889 /* If I don't even have good local storage, we can not reasonably try
890 * to pull ahead of the peer. We also need the local reference to make
891 * sure device->act_log is there.
893 if (!get_ldev_if_state(device, D_UP_TO_DATE))
897 atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
898 drbd_info(device, "Congestion-fill threshold reached\n");
902 if (device->act_log->used >= nc->cong_extents) {
903 drbd_info(device, "Congestion-extents threshold reached\n");
908 /* start a new epoch for non-mirrored writes */
909 start_new_tl_epoch(first_peer_device(device)->connection);
911 if (on_congestion == OC_PULL_AHEAD)
912 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
913 else /*nc->on_congestion == OC_DISCONNECT */
914 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
919 /* If this returns false, and req->private_bio is still set,
920 * this should be submitted locally.
922 * If it returns false, but req->private_bio is not set,
923 * we do not have access to good data :(
925 * Otherwise, this destroys req->private_bio, if any,
928 static bool do_remote_read(struct drbd_request *req)
930 struct drbd_device *device = req->device;
931 enum drbd_read_balancing rbm;
933 if (req->private_bio) {
934 if (!drbd_may_do_local_read(device,
935 req->i.sector, req->i.size)) {
936 bio_put(req->private_bio);
937 req->private_bio = NULL;
942 if (device->state.pdsk != D_UP_TO_DATE)
945 if (req->private_bio == NULL)
948 /* TODO: improve read balancing decisions, take into account drbd
949 * protocol, pending requests etc. */
952 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
955 if (rbm == RB_PREFER_LOCAL && req->private_bio)
956 return false; /* submit locally */
958 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
959 if (req->private_bio) {
960 bio_put(req->private_bio);
961 req->private_bio = NULL;
970 /* returns number of connections (== 1, for drbd 8.4)
971 * expected to actually write this data,
972 * which does NOT include those that we are L_AHEAD for. */
973 static int drbd_process_write_request(struct drbd_request *req)
975 struct drbd_device *device = req->device;
976 int remote, send_oos;
978 remote = drbd_should_do_remote(device->state);
979 send_oos = drbd_should_send_out_of_sync(device->state);
981 /* Need to replicate writes. Unless it is an empty flush,
982 * which is better mapped to a DRBD P_BARRIER packet,
983 * also for drbd wire protocol compatibility reasons.
984 * If this was a flush, just start a new epoch.
985 * Unless the current epoch was empty anyways, or we are not currently
986 * replicating, in which case there is no point. */
987 if (unlikely(req->i.size == 0)) {
988 /* The only size==0 bios we expect are empty flushes. */
989 D_ASSERT(device, req->master_bio->bi_rw & REQ_FLUSH);
991 _req_mod(req, QUEUE_AS_DRBD_BARRIER);
995 if (!remote && !send_oos)
998 D_ASSERT(device, !(remote && send_oos));
1001 _req_mod(req, TO_BE_SENT);
1002 _req_mod(req, QUEUE_FOR_NET_WRITE);
1003 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1004 _req_mod(req, QUEUE_FOR_SEND_OOS);
1010 drbd_submit_req_private_bio(struct drbd_request *req)
1012 struct drbd_device *device = req->device;
1013 struct bio *bio = req->private_bio;
1014 const int rw = bio_rw(bio);
1016 bio->bi_bdev = device->ldev->backing_bdev;
1018 /* State may have changed since we grabbed our reference on the
1019 * ->ldev member. Double check, and short-circuit to endio.
1020 * In case the last activity log transaction failed to get on
1021 * stable storage, and this is a WRITE, we may not even submit
1023 if (get_ldev(device)) {
1024 if (drbd_insert_fault(device,
1025 rw == WRITE ? DRBD_FAULT_DT_WR
1026 : rw == READ ? DRBD_FAULT_DT_RD
1027 : DRBD_FAULT_DT_RA))
1028 bio_endio(bio, -EIO);
1030 generic_make_request(bio);
1033 bio_endio(bio, -EIO);
1036 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1038 spin_lock(&device->submit.lock);
1039 list_add_tail(&req->tl_requests, &device->submit.writes);
1040 spin_unlock(&device->submit.lock);
1041 queue_work(device->submit.wq, &device->submit.worker);
1044 /* returns the new drbd_request pointer, if the caller is expected to
1045 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1046 * request on the submitter thread.
1047 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1049 static struct drbd_request *
1050 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_time)
1052 const int rw = bio_data_dir(bio);
1053 struct drbd_request *req;
1055 /* allocate outside of all locks; */
1056 req = drbd_req_new(device, bio);
1059 /* only pass the error to the upper layers.
1060 * if user cannot handle io errors, that's not our business. */
1061 drbd_err(device, "could not kmalloc() req\n");
1062 bio_endio(bio, -ENOMEM);
1063 return ERR_PTR(-ENOMEM);
1065 req->start_time = start_time;
1067 if (!get_ldev(device)) {
1068 bio_put(req->private_bio);
1069 req->private_bio = NULL;
1072 /* Update disk stats */
1073 _drbd_start_io_acct(device, req);
1075 if (rw == WRITE && req->private_bio && req->i.size
1076 && !test_bit(AL_SUSPENDED, &device->flags)) {
1077 if (!drbd_al_begin_io_fastpath(device, &req->i)) {
1078 drbd_queue_write(device, req);
1081 req->rq_state |= RQ_IN_ACT_LOG;
1087 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1089 const int rw = bio_rw(req->master_bio);
1090 struct bio_and_error m = { NULL, };
1091 bool no_remote = false;
1093 spin_lock_irq(&device->resource->req_lock);
1095 /* This may temporarily give up the req_lock,
1096 * but will re-aquire it before it returns here.
1097 * Needs to be before the check on drbd_suspended() */
1098 complete_conflicting_writes(req);
1099 /* no more giving up req_lock from now on! */
1101 /* check for congestion, and potentially stop sending
1102 * full data updates, but start sending "dirty bits" only. */
1103 maybe_pull_ahead(device);
1107 if (drbd_suspended(device)) {
1108 /* push back and retry: */
1109 req->rq_state |= RQ_POSTPONED;
1110 if (req->private_bio) {
1111 bio_put(req->private_bio);
1112 req->private_bio = NULL;
1118 /* We fail READ/READA early, if we can not serve it.
1119 * We must do this before req is registered on any lists.
1120 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1122 if (!do_remote_read(req) && !req->private_bio)
1126 /* which transfer log epoch does this belong to? */
1127 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1129 /* no point in adding empty flushes to the transfer log,
1130 * they are mapped to drbd barriers already. */
1131 if (likely(req->i.size!=0)) {
1133 first_peer_device(device)->connection->current_tle_writes++;
1135 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1139 if (!drbd_process_write_request(req))
1142 /* We either have a private_bio, or we can read from remote.
1143 * Otherwise we had done the goto nodata above. */
1144 if (req->private_bio == NULL) {
1145 _req_mod(req, TO_BE_SENT);
1146 _req_mod(req, QUEUE_FOR_NET_READ);
1151 if (req->private_bio) {
1152 /* needs to be marked within the same spinlock */
1153 _req_mod(req, TO_BE_SUBMITTED);
1154 /* but we need to give up the spinlock to submit */
1155 spin_unlock_irq(&device->resource->req_lock);
1156 drbd_submit_req_private_bio(req);
1157 spin_lock_irq(&device->resource->req_lock);
1158 } else if (no_remote) {
1160 if (__ratelimit(&drbd_ratelimit_state))
1161 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1162 (unsigned long long)req->i.sector, req->i.size >> 9);
1163 /* A write may have been queued for send_oos, however.
1164 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1168 if (drbd_req_put_completion_ref(req, &m, 1))
1169 kref_put(&req->kref, drbd_req_destroy);
1170 spin_unlock_irq(&device->resource->req_lock);
1173 complete_master_bio(device, &m);
1176 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_time)
1178 struct drbd_request *req = drbd_request_prepare(device, bio, start_time);
1179 if (IS_ERR_OR_NULL(req))
1181 drbd_send_and_submit(device, req);
1184 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1186 struct drbd_request *req, *tmp;
1187 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1188 const int rw = bio_data_dir(req->master_bio);
1190 if (rw == WRITE /* rw != WRITE should not even end up here! */
1191 && req->private_bio && req->i.size
1192 && !test_bit(AL_SUSPENDED, &device->flags)) {
1193 if (!drbd_al_begin_io_fastpath(device, &req->i))
1196 req->rq_state |= RQ_IN_ACT_LOG;
1199 list_del_init(&req->tl_requests);
1200 drbd_send_and_submit(device, req);
1204 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1205 struct list_head *incoming,
1206 struct list_head *pending)
1208 struct drbd_request *req, *tmp;
1212 spin_lock_irq(&device->al_lock);
1213 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1214 err = drbd_al_begin_io_nonblock(device, &req->i);
1219 req->rq_state |= RQ_IN_ACT_LOG;
1220 list_move_tail(&req->tl_requests, pending);
1222 spin_unlock_irq(&device->al_lock);
1224 wake_up(&device->al_wait);
1226 return !list_empty(pending);
1229 void do_submit(struct work_struct *ws)
1231 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1232 LIST_HEAD(incoming);
1234 struct drbd_request *req, *tmp;
1237 spin_lock(&device->submit.lock);
1238 list_splice_tail_init(&device->submit.writes, &incoming);
1239 spin_unlock(&device->submit.lock);
1241 submit_fast_path(device, &incoming);
1242 if (list_empty(&incoming))
1246 wait_event(device->al_wait, prepare_al_transaction_nonblock(device, &incoming, &pending));
1247 /* Maybe more was queued, while we prepared the transaction?
1248 * Try to stuff them into this transaction as well.
1249 * Be strictly non-blocking here, no wait_event, we already
1250 * have something to commit.
1251 * Stop if we don't make any more progres.
1254 LIST_HEAD(more_pending);
1255 LIST_HEAD(more_incoming);
1258 /* It is ok to look outside the lock,
1259 * it's only an optimization anyways */
1260 if (list_empty(&device->submit.writes))
1263 spin_lock(&device->submit.lock);
1264 list_splice_tail_init(&device->submit.writes, &more_incoming);
1265 spin_unlock(&device->submit.lock);
1267 if (list_empty(&more_incoming))
1270 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending);
1272 list_splice_tail_init(&more_pending, &pending);
1273 list_splice_tail_init(&more_incoming, &incoming);
1278 drbd_al_begin_io_commit(device, false);
1280 list_for_each_entry_safe(req, tmp, &pending, tl_requests) {
1281 list_del_init(&req->tl_requests);
1282 drbd_send_and_submit(device, req);
1285 /* If all currently hot activity log extents are kept busy by
1286 * incoming requests, we still must not totally starve new
1287 * requests to cold extents. In that case, prepare one request
1288 * in blocking mode. */
1289 list_for_each_entry_safe(req, tmp, &incoming, tl_requests) {
1290 list_del_init(&req->tl_requests);
1291 req->rq_state |= RQ_IN_ACT_LOG;
1292 if (!drbd_al_begin_io_prepare(device, &req->i)) {
1293 /* Corresponding extent was hot after all? */
1294 drbd_send_and_submit(device, req);
1296 /* Found a request to a cold extent.
1297 * Put on "pending" list,
1298 * and try to cumulate with more. */
1299 list_add(&req->tl_requests, &pending);
1300 goto skip_fast_path;
1306 void drbd_make_request(struct request_queue *q, struct bio *bio)
1308 struct drbd_device *device = (struct drbd_device *) q->queuedata;
1309 unsigned long start_time;
1311 start_time = jiffies;
1314 * what we "blindly" assume:
1316 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1319 __drbd_make_request(device, bio, start_time);
1322 /* This is called by bio_add_page().
1324 * q->max_hw_sectors and other global limits are already enforced there.
1326 * We need to call down to our lower level device,
1327 * in case it has special restrictions.
1329 * We also may need to enforce configured max-bio-bvecs limits.
1331 * As long as the BIO is empty we have to allow at least one bvec,
1332 * regardless of size and offset, so no need to ask lower levels.
1334 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1336 struct drbd_device *device = (struct drbd_device *) q->queuedata;
1337 unsigned int bio_size = bvm->bi_size;
1338 int limit = DRBD_MAX_BIO_SIZE;
1341 if (bio_size && get_ldev(device)) {
1342 unsigned int max_hw_sectors = queue_max_hw_sectors(q);
1343 struct request_queue * const b =
1344 device->ldev->backing_bdev->bd_disk->queue;
1345 if (b->merge_bvec_fn) {
1346 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1347 limit = min(limit, backing_limit);
1350 if ((limit >> 9) > max_hw_sectors)
1351 limit = max_hw_sectors << 9;
1356 static void find_oldest_requests(
1357 struct drbd_connection *connection,
1358 struct drbd_device *device,
1359 struct drbd_request **oldest_req_waiting_for_peer,
1360 struct drbd_request **oldest_req_waiting_for_disk)
1362 struct drbd_request *r;
1363 *oldest_req_waiting_for_peer = NULL;
1364 *oldest_req_waiting_for_disk = NULL;
1365 list_for_each_entry(r, &connection->transfer_log, tl_requests) {
1366 const unsigned s = r->rq_state;
1367 if (!*oldest_req_waiting_for_peer
1368 && ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE)))
1369 *oldest_req_waiting_for_peer = r;
1371 if (!*oldest_req_waiting_for_disk
1372 && (s & RQ_LOCAL_PENDING) && r->device == device)
1373 *oldest_req_waiting_for_disk = r;
1375 if (*oldest_req_waiting_for_peer && *oldest_req_waiting_for_disk)
1380 void request_timer_fn(unsigned long data)
1382 struct drbd_device *device = (struct drbd_device *) data;
1383 struct drbd_connection *connection = first_peer_device(device)->connection;
1384 struct drbd_request *req_disk, *req_peer; /* oldest request */
1385 struct net_conf *nc;
1386 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1390 nc = rcu_dereference(connection->net_conf);
1391 if (nc && device->state.conn >= C_WF_REPORT_PARAMS)
1392 ent = nc->timeout * HZ/10 * nc->ko_count;
1394 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1395 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1400 et = min_not_zero(dt, ent);
1403 return; /* Recurring timer stopped */
1407 spin_lock_irq(&device->resource->req_lock);
1408 find_oldest_requests(connection, device, &req_peer, &req_disk);
1409 if (req_peer == NULL && req_disk == NULL) {
1410 spin_unlock_irq(&device->resource->req_lock);
1411 mod_timer(&device->request_timer, now + et);
1415 /* The request is considered timed out, if
1416 * - we have some effective timeout from the configuration,
1417 * with above state restrictions applied,
1418 * - the oldest request is waiting for a response from the network
1419 * resp. the local disk,
1420 * - the oldest request is in fact older than the effective timeout,
1421 * - the connection was established (resp. disk was attached)
1422 * for longer than the timeout already.
1423 * Note that for 32bit jiffies and very stable connections/disks,
1424 * we may have a wrap around, which is catched by
1425 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1427 * Side effect: once per 32bit wrap-around interval, which means every
1428 * ~198 days with 250 HZ, we have a window where the timeout would need
1429 * to expire twice (worst case) to become effective. Good enough.
1431 if (ent && req_peer &&
1432 time_after(now, req_peer->start_time + ent) &&
1433 !time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) {
1434 drbd_warn(device, "Remote failed to finish a request within ko-count * timeout\n");
1435 _drbd_set_state(_NS(device, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1437 if (dt && req_disk &&
1438 time_after(now, req_disk->start_time + dt) &&
1439 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1440 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1441 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1444 /* Reschedule timer for the nearest not already expired timeout.
1445 * Fallback to now + min(effective network timeout, disk timeout). */
1446 ent = (ent && req_peer && time_before(now, req_peer->start_time + ent))
1447 ? req_peer->start_time + ent : now + et;
1448 dt = (dt && req_disk && time_before(now, req_disk->start_time + dt))
1449 ? req_disk->start_time + dt : now + et;
1450 nt = time_before(ent, dt) ? ent : dt;
1451 spin_unlock_irq(&connection->resource->req_lock);
1452 mod_timer(&device->request_timer, nt);