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 struct request_queue *q = device->rq_queue;
41 generic_start_io_acct(q, bio_data_dir(req->master_bio),
42 req->i.size >> 9, &device->vdisk->part0);
45 /* Update disk stats when completing request upwards */
46 static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
48 struct request_queue *q = device->rq_queue;
50 generic_end_io_acct(q, bio_data_dir(req->master_bio),
51 &device->vdisk->part0, req->start_jif);
54 static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src)
56 struct drbd_request *req;
58 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
61 memset(req, 0, sizeof(*req));
63 drbd_req_make_private_bio(req, bio_src);
64 req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
65 | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0)
66 | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_UNMAP : 0)
67 | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
69 req->master_bio = bio_src;
72 drbd_clear_interval(&req->i);
73 req->i.sector = bio_src->bi_iter.bi_sector;
74 req->i.size = bio_src->bi_iter.bi_size;
76 req->i.waiting = false;
78 INIT_LIST_HEAD(&req->tl_requests);
79 INIT_LIST_HEAD(&req->w.list);
80 INIT_LIST_HEAD(&req->req_pending_master_completion);
81 INIT_LIST_HEAD(&req->req_pending_local);
83 /* one reference to be put by __drbd_make_request */
84 atomic_set(&req->completion_ref, 1);
85 /* one kref as long as completion_ref > 0 */
86 kref_init(&req->kref);
90 static void drbd_remove_request_interval(struct rb_root *root,
91 struct drbd_request *req)
93 struct drbd_device *device = req->device;
94 struct drbd_interval *i = &req->i;
96 drbd_remove_interval(root, i);
98 /* Wake up any processes waiting for this request to complete. */
100 wake_up(&device->misc_wait);
103 void drbd_req_destroy(struct kref *kref)
105 struct drbd_request *req = container_of(kref, struct drbd_request, kref);
106 struct drbd_device *device = req->device;
107 const unsigned s = req->rq_state;
109 if ((req->master_bio && !(s & RQ_POSTPONED)) ||
110 atomic_read(&req->completion_ref) ||
111 (s & RQ_LOCAL_PENDING) ||
112 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
113 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
114 s, atomic_read(&req->completion_ref));
118 /* If called from mod_rq_state (expected normal case) or
119 * drbd_send_and_submit (the less likely normal path), this holds the
120 * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
121 * though it may be still empty (never added to the transfer log).
123 * If called from do_retry(), we do NOT hold the req_lock, but we are
124 * still allowed to unconditionally list_del(&req->tl_requests),
125 * because it will be on a local on-stack list only. */
126 list_del_init(&req->tl_requests);
128 /* finally remove the request from the conflict detection
129 * respective block_id verification interval tree. */
130 if (!drbd_interval_empty(&req->i)) {
131 struct rb_root *root;
134 root = &device->write_requests;
136 root = &device->read_requests;
137 drbd_remove_request_interval(root, req);
138 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
139 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
140 s, (unsigned long long)req->i.sector, req->i.size);
142 /* if it was a write, we may have to set the corresponding
143 * bit(s) out-of-sync first. If it had a local part, we need to
144 * release the reference to the activity log. */
146 /* Set out-of-sync unless both OK flags are set
147 * (local only or remote failed).
148 * Other places where we set out-of-sync:
149 * READ with local io-error */
151 /* There is a special case:
152 * we may notice late that IO was suspended,
153 * and postpone, or schedule for retry, a write,
154 * before it even was submitted or sent.
155 * In that case we do not want to touch the bitmap at all.
157 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
158 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
159 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
161 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
162 drbd_set_in_sync(device, req->i.sector, req->i.size);
165 /* one might be tempted to move the drbd_al_complete_io
166 * to the local io completion callback drbd_request_endio.
167 * but, if this was a mirror write, we may only
168 * drbd_al_complete_io after this is RQ_NET_DONE,
169 * otherwise the extent could be dropped from the al
170 * before it has actually been written on the peer.
171 * if we crash before our peer knows about the request,
172 * but after the extent has been dropped from the al,
173 * we would forget to resync the corresponding extent.
175 if (s & RQ_IN_ACT_LOG) {
176 if (get_ldev_if_state(device, D_FAILED)) {
177 drbd_al_complete_io(device, &req->i);
179 } else if (__ratelimit(&drbd_ratelimit_state)) {
180 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
181 "but my Disk seems to have failed :(\n",
182 (unsigned long long) req->i.sector, req->i.size);
187 mempool_free(req, drbd_request_mempool);
190 static void wake_all_senders(struct drbd_connection *connection)
192 wake_up(&connection->sender_work.q_wait);
195 /* must hold resource->req_lock */
196 void start_new_tl_epoch(struct drbd_connection *connection)
198 /* no point closing an epoch, if it is empty, anyways. */
199 if (connection->current_tle_writes == 0)
202 connection->current_tle_writes = 0;
203 atomic_inc(&connection->current_tle_nr);
204 wake_all_senders(connection);
207 void complete_master_bio(struct drbd_device *device,
208 struct bio_and_error *m)
210 m->bio->bi_status = errno_to_blk_status(m->error);
216 /* Helper for __req_mod().
217 * Set m->bio to the master bio, if it is fit to be completed,
218 * or leave it alone (it is initialized to NULL in __req_mod),
219 * if it has already been completed, or cannot be completed yet.
220 * If m->bio is set, the error status to be returned is placed in m->error.
223 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
225 const unsigned s = req->rq_state;
226 struct drbd_device *device = req->device;
229 /* we must not complete the master bio, while it is
230 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
231 * not yet acknowledged by the peer
232 * not yet completed by the local io subsystem
233 * these flags may get cleared in any order by
236 * the bio_endio completion callbacks.
238 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
239 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
240 (s & RQ_COMPLETION_SUSP)) {
241 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
245 if (!req->master_bio) {
246 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
251 * figure out whether to report success or failure.
253 * report success when at least one of the operations succeeded.
254 * or, to put the other way,
255 * only report failure, when both operations failed.
257 * what to do about the failures is handled elsewhere.
258 * what we need to do here is just: complete the master_bio.
260 * local completion error, if any, has been stored as ERR_PTR
261 * in private_bio within drbd_request_endio.
263 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
264 error = PTR_ERR(req->private_bio);
266 /* Before we can signal completion to the upper layers,
267 * we may need to close the current transfer log epoch.
268 * We are within the request lock, so we can simply compare
269 * the request epoch number with the current transfer log
270 * epoch number. If they match, increase the current_tle_nr,
271 * and reset the transfer log epoch write_cnt.
273 if (op_is_write(bio_op(req->master_bio)) &&
274 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
275 start_new_tl_epoch(first_peer_device(device)->connection);
277 /* Update disk stats */
278 _drbd_end_io_acct(device, req);
281 * have it be pushed back to the retry work queue,
282 * so it will re-enter __drbd_make_request(),
283 * and be re-assigned to a suitable local or remote path,
284 * or failed if we do not have access to good data anymore.
286 * Unless it was failed early by __drbd_make_request(),
287 * because no path was available, in which case
288 * it was not even added to the transfer_log.
290 * read-ahead may fail, and will not be retried.
292 * WRITE should have used all available paths already.
295 bio_op(req->master_bio) == REQ_OP_READ &&
296 !(req->master_bio->bi_opf & REQ_RAHEAD) &&
297 !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;
304 /* We leave it in the tree, to be able to verify later
305 * write-acks in protocol != C during resync.
306 * But we mark it as "complete", so it won't be counted as
307 * conflict in a multi-primary setup. */
308 req->i.completed = true;
312 wake_up(&device->misc_wait);
314 /* Either we are about to complete to upper layers,
315 * or we will restart this request.
316 * In either case, the request object will be destroyed soon,
317 * so better remove it from all lists. */
318 list_del_init(&req->req_pending_master_completion);
321 /* still holds resource->req_lock */
322 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
324 struct drbd_device *device = req->device;
325 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
330 if (!atomic_sub_and_test(put, &req->completion_ref))
333 drbd_req_complete(req, m);
335 /* local completion may still come in later,
336 * we need to keep the req object around. */
337 if (req->rq_state & RQ_LOCAL_ABORTED)
340 if (req->rq_state & RQ_POSTPONED) {
341 /* don't destroy the req object just yet,
342 * but queue it for retry */
343 drbd_restart_request(req);
347 kref_put(&req->kref, drbd_req_destroy);
350 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
352 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
355 if (connection->req_next == NULL)
356 connection->req_next = req;
359 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
361 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
364 if (connection->req_next != req)
366 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
367 const unsigned s = req->rq_state;
368 if (s & RQ_NET_QUEUED)
371 if (&req->tl_requests == &connection->transfer_log)
373 connection->req_next = req;
376 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
378 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
381 if (connection->req_ack_pending == NULL)
382 connection->req_ack_pending = req;
385 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
387 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
390 if (connection->req_ack_pending != req)
392 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
393 const unsigned s = req->rq_state;
394 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
397 if (&req->tl_requests == &connection->transfer_log)
399 connection->req_ack_pending = req;
402 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
404 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
407 if (connection->req_not_net_done == NULL)
408 connection->req_not_net_done = req;
411 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
413 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
416 if (connection->req_not_net_done != req)
418 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
419 const unsigned s = req->rq_state;
420 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
423 if (&req->tl_requests == &connection->transfer_log)
425 connection->req_not_net_done = req;
428 /* I'd like this to be the only place that manipulates
429 * req->completion_ref and req->kref. */
430 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
433 struct drbd_device *device = req->device;
434 struct drbd_peer_device *peer_device = first_peer_device(device);
435 unsigned s = req->rq_state;
438 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
439 set |= RQ_COMPLETION_SUSP;
443 req->rq_state &= ~clear;
444 req->rq_state |= set;
447 if (req->rq_state == s)
450 /* intent: get references */
452 kref_get(&req->kref);
454 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
455 atomic_inc(&req->completion_ref);
457 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
458 inc_ap_pending(device);
459 atomic_inc(&req->completion_ref);
462 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
463 atomic_inc(&req->completion_ref);
464 set_if_null_req_next(peer_device, req);
467 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
468 kref_get(&req->kref); /* wait for the DONE */
470 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
471 /* potentially already completed in the ack_receiver thread */
472 if (!(s & RQ_NET_DONE)) {
473 atomic_add(req->i.size >> 9, &device->ap_in_flight);
474 set_if_null_req_not_net_done(peer_device, req);
476 if (req->rq_state & RQ_NET_PENDING)
477 set_if_null_req_ack_pending(peer_device, req);
480 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
481 atomic_inc(&req->completion_ref);
483 /* progress: put references */
485 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
488 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
489 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
493 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
494 if (req->rq_state & RQ_LOCAL_ABORTED)
495 kref_put(&req->kref, drbd_req_destroy);
498 list_del_init(&req->req_pending_local);
501 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
502 dec_ap_pending(device);
504 req->acked_jif = jiffies;
505 advance_conn_req_ack_pending(peer_device, req);
508 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
510 advance_conn_req_next(peer_device, req);
513 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
515 atomic_sub(req->i.size >> 9, &device->ap_in_flight);
516 if (s & RQ_EXP_BARR_ACK)
517 kref_put(&req->kref, drbd_req_destroy);
518 req->net_done_jif = jiffies;
520 /* in ahead/behind mode, or just in case,
521 * before we finally destroy this request,
522 * the caching pointers must not reference it anymore */
523 advance_conn_req_next(peer_device, req);
524 advance_conn_req_ack_pending(peer_device, req);
525 advance_conn_req_not_net_done(peer_device, req);
528 /* potentially complete and destroy */
530 /* If we made progress, retry conflicting peer requests, if any. */
532 wake_up(&device->misc_wait);
534 drbd_req_put_completion_ref(req, m, c_put);
535 kref_put(&req->kref, drbd_req_destroy);
538 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
540 char b[BDEVNAME_SIZE];
542 if (!__ratelimit(&drbd_ratelimit_state))
545 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
546 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
547 (unsigned long long)req->i.sector,
549 bdevname(device->ldev->backing_bdev, b));
552 /* Helper for HANDED_OVER_TO_NETWORK.
553 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
554 * Is it also still "PENDING"?
555 * --> If so, clear PENDING and set NET_OK below.
556 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
557 * (and we must not set RQ_NET_OK) */
558 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
560 return (req->rq_state &
561 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
562 == (RQ_WRITE|RQ_NET_PENDING);
565 /* obviously this could be coded as many single functions
566 * instead of one huge switch,
567 * or by putting the code directly in the respective locations
568 * (as it has been before).
570 * but having it this way
571 * enforces that it is all in this one place, where it is easier to audit,
572 * it makes it obvious that whatever "event" "happens" to a request should
573 * happen "atomically" within the req_lock,
574 * and it enforces that we have to think in a very structured manner
575 * about the "events" that may happen to a request during its life time ...
577 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
578 struct bio_and_error *m)
580 struct drbd_device *const device = req->device;
581 struct drbd_peer_device *const peer_device = first_peer_device(device);
582 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
591 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
594 /* does not happen...
595 * initialization done in drbd_req_new
600 case TO_BE_SENT: /* via network */
601 /* reached via __drbd_make_request
602 * and from w_read_retry_remote */
603 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
605 nc = rcu_dereference(connection->net_conf);
606 p = nc->wire_protocol;
609 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
610 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
611 mod_rq_state(req, m, 0, RQ_NET_PENDING);
614 case TO_BE_SUBMITTED: /* locally */
615 /* reached via __drbd_make_request */
616 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
617 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
621 if (req->rq_state & RQ_WRITE)
622 device->writ_cnt += req->i.size >> 9;
624 device->read_cnt += req->i.size >> 9;
626 mod_rq_state(req, m, RQ_LOCAL_PENDING,
627 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
631 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
634 case WRITE_COMPLETED_WITH_ERROR:
635 drbd_report_io_error(device, req);
636 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
637 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
640 case READ_COMPLETED_WITH_ERROR:
641 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
642 drbd_report_io_error(device, req);
643 __drbd_chk_io_error(device, DRBD_READ_ERROR);
645 case READ_AHEAD_COMPLETED_WITH_ERROR:
646 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
647 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
650 case DISCARD_COMPLETED_NOTSUPP:
651 case DISCARD_COMPLETED_WITH_ERROR:
652 /* I'd rather not detach from local disk just because it
653 * failed a REQ_DISCARD. */
654 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
657 case QUEUE_FOR_NET_READ:
660 * or target area marked as invalid,
661 * or just got an io-error. */
662 /* from __drbd_make_request
663 * or from bio_endio during read io-error recovery */
665 /* So we can verify the handle in the answer packet.
666 * Corresponding drbd_remove_request_interval is in
667 * drbd_req_complete() */
668 D_ASSERT(device, drbd_interval_empty(&req->i));
669 drbd_insert_interval(&device->read_requests, &req->i);
671 set_bit(UNPLUG_REMOTE, &device->flags);
673 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
674 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
675 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
676 req->w.cb = w_send_read_req;
677 drbd_queue_work(&connection->sender_work,
681 case QUEUE_FOR_NET_WRITE:
682 /* assert something? */
683 /* from __drbd_make_request only */
685 /* Corresponding drbd_remove_request_interval is in
686 * drbd_req_complete() */
687 D_ASSERT(device, drbd_interval_empty(&req->i));
688 drbd_insert_interval(&device->write_requests, &req->i);
691 * In case the req ended up on the transfer log before being
692 * queued on the worker, it could lead to this request being
693 * missed during cleanup after connection loss.
694 * So we have to do both operations here,
695 * within the same lock that protects the transfer log.
697 * _req_add_to_epoch(req); this has to be after the
698 * _maybe_start_new_epoch(req); which happened in
699 * __drbd_make_request, because we now may set the bit
700 * again ourselves to close the current epoch.
702 * Add req to the (now) current epoch (barrier). */
704 /* otherwise we may lose an unplug, which may cause some remote
705 * io-scheduler timeout to expire, increasing maximum latency,
706 * hurting performance. */
707 set_bit(UNPLUG_REMOTE, &device->flags);
709 /* queue work item to send data */
710 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
711 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
712 req->w.cb = w_send_dblock;
713 drbd_queue_work(&connection->sender_work,
716 /* close the epoch, in case it outgrew the limit */
718 nc = rcu_dereference(connection->net_conf);
719 p = nc->max_epoch_size;
721 if (connection->current_tle_writes >= p)
722 start_new_tl_epoch(connection);
726 case QUEUE_FOR_SEND_OOS:
727 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
728 req->w.cb = w_send_out_of_sync;
729 drbd_queue_work(&connection->sender_work,
733 case READ_RETRY_REMOTE_CANCELED:
736 /* real cleanup will be done from tl_clear. just update flags
737 * so it is no longer marked as on the worker queue */
738 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
741 case HANDED_OVER_TO_NETWORK:
742 /* assert something? */
743 if (is_pending_write_protocol_A(req))
744 /* this is what is dangerous about protocol A:
745 * pretend it was successfully written on the peer. */
746 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
747 RQ_NET_SENT|RQ_NET_OK);
749 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
750 /* It is still not yet RQ_NET_DONE until the
751 * corresponding epoch barrier got acked as well,
752 * so we know what to dirty on connection loss. */
755 case OOS_HANDED_TO_NETWORK:
756 /* Was not set PENDING, no longer QUEUED, so is now DONE
757 * as far as this connection is concerned. */
758 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
761 case CONNECTION_LOST_WHILE_PENDING:
762 /* transfer log cleanup after connection loss */
764 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
768 case CONFLICT_RESOLVED:
769 /* for superseded conflicting writes of multiple primaries,
770 * there is no need to keep anything in the tl, potential
771 * node crashes are covered by the activity log.
773 * If this request had been marked as RQ_POSTPONED before,
774 * it will actually not be completed, but "restarted",
775 * resubmitted from the retry worker context. */
776 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
777 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
778 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
781 case WRITE_ACKED_BY_PEER_AND_SIS:
782 req->rq_state |= RQ_NET_SIS;
783 case WRITE_ACKED_BY_PEER:
784 /* Normal operation protocol C: successfully written on peer.
785 * During resync, even in protocol != C,
786 * we requested an explicit write ack anyways.
787 * Which means we cannot even assert anything here.
788 * Nothing more to do here.
789 * We want to keep the tl in place for all protocols, to cater
790 * for volatile write-back caches on lower level devices. */
792 case RECV_ACKED_BY_PEER:
793 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
794 /* protocol B; pretends to be successfully written on peer.
795 * see also notes above in HANDED_OVER_TO_NETWORK about
798 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
802 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
803 /* If this node has already detected the write conflict, the
804 * worker will be waiting on misc_wait. Wake it up once this
805 * request has completed locally.
807 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
808 req->rq_state |= RQ_POSTPONED;
810 wake_up(&device->misc_wait);
811 /* Do not clear RQ_NET_PENDING. This request will make further
812 * progress via restart_conflicting_writes() or
813 * fail_postponed_requests(). Hopefully. */
817 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
820 case FAIL_FROZEN_DISK_IO:
821 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
823 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
826 case RESTART_FROZEN_DISK_IO:
827 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
831 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
835 if (bio_data_dir(req->master_bio) == WRITE)
838 get_ldev(device); /* always succeeds in this call path */
839 req->w.cb = w_restart_disk_io;
840 drbd_queue_work(&connection->sender_work,
845 /* Simply complete (local only) READs. */
846 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
847 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
851 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
852 before the connection loss (B&C only); only P_BARRIER_ACK
853 (or the local completion?) was missing when we suspended.
854 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
855 During connection handshake, we ensure that the peer was not rebooted. */
856 if (!(req->rq_state & RQ_NET_OK)) {
857 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
858 * in that case we must not set RQ_NET_PENDING. */
860 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
862 /* w.cb expected to be w_send_dblock, or w_send_read_req */
863 drbd_queue_work(&connection->sender_work,
865 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
866 } /* else: FIXME can this happen? */
869 /* else, fall through to BARRIER_ACKED */
872 /* barrier ack for READ requests does not make sense */
873 if (!(req->rq_state & RQ_WRITE))
876 if (req->rq_state & RQ_NET_PENDING) {
877 /* barrier came in before all requests were acked.
878 * this is bad, because if the connection is lost now,
879 * we won't be able to clean them up... */
880 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
882 /* Allowed to complete requests, even while suspended.
883 * As this is called for all requests within a matching epoch,
884 * we need to filter, and only set RQ_NET_DONE for those that
885 * have actually been on the wire. */
886 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
887 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
891 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
892 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
895 case QUEUE_AS_DRBD_BARRIER:
896 start_new_tl_epoch(connection);
897 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
904 /* we may do a local read if:
905 * - we are consistent (of course),
906 * - or we are generally inconsistent,
907 * BUT we are still/already IN SYNC for this area.
908 * since size may be bigger than BM_BLOCK_SIZE,
909 * we may need to check several bits.
911 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
913 unsigned long sbnr, ebnr;
914 sector_t esector, nr_sectors;
916 if (device->state.disk == D_UP_TO_DATE)
918 if (device->state.disk != D_INCONSISTENT)
920 esector = sector + (size >> 9) - 1;
921 nr_sectors = drbd_get_capacity(device->this_bdev);
922 D_ASSERT(device, sector < nr_sectors);
923 D_ASSERT(device, esector < nr_sectors);
925 sbnr = BM_SECT_TO_BIT(sector);
926 ebnr = BM_SECT_TO_BIT(esector);
928 return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
931 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
932 enum drbd_read_balancing rbm)
934 struct backing_dev_info *bdi;
938 case RB_CONGESTED_REMOTE:
939 bdi = device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
940 return bdi_read_congested(bdi);
941 case RB_LEAST_PENDING:
942 return atomic_read(&device->local_cnt) >
943 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
944 case RB_32K_STRIPING: /* stripe_shift = 15 */
945 case RB_64K_STRIPING:
946 case RB_128K_STRIPING:
947 case RB_256K_STRIPING:
948 case RB_512K_STRIPING:
949 case RB_1M_STRIPING: /* stripe_shift = 20 */
950 stripe_shift = (rbm - RB_32K_STRIPING + 15);
951 return (sector >> (stripe_shift - 9)) & 1;
953 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
954 case RB_PREFER_REMOTE:
956 case RB_PREFER_LOCAL:
963 * complete_conflicting_writes - wait for any conflicting write requests
965 * The write_requests tree contains all active write requests which we
966 * currently know about. Wait for any requests to complete which conflict with
969 * Only way out: remove the conflicting intervals from the tree.
971 static void complete_conflicting_writes(struct drbd_request *req)
974 struct drbd_device *device = req->device;
975 struct drbd_interval *i;
976 sector_t sector = req->i.sector;
977 int size = req->i.size;
980 drbd_for_each_overlap(i, &device->write_requests, sector, size) {
981 /* Ignore, if already completed to upper layers. */
984 /* Handle the first found overlap. After the schedule
985 * we have to restart the tree walk. */
991 /* Indicate to wake up device->misc_wait on progress. */
992 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
994 spin_unlock_irq(&device->resource->req_lock);
996 spin_lock_irq(&device->resource->req_lock);
998 finish_wait(&device->misc_wait, &wait);
1001 /* called within req_lock */
1002 static void maybe_pull_ahead(struct drbd_device *device)
1004 struct drbd_connection *connection = first_peer_device(device)->connection;
1005 struct net_conf *nc;
1006 bool congested = false;
1007 enum drbd_on_congestion on_congestion;
1010 nc = rcu_dereference(connection->net_conf);
1011 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
1013 if (on_congestion == OC_BLOCK ||
1014 connection->agreed_pro_version < 96)
1017 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1018 return; /* nothing to do ... */
1020 /* If I don't even have good local storage, we can not reasonably try
1021 * to pull ahead of the peer. We also need the local reference to make
1022 * sure device->act_log is there.
1024 if (!get_ldev_if_state(device, D_UP_TO_DATE))
1027 if (nc->cong_fill &&
1028 atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1029 drbd_info(device, "Congestion-fill threshold reached\n");
1033 if (device->act_log->used >= nc->cong_extents) {
1034 drbd_info(device, "Congestion-extents threshold reached\n");
1039 /* start a new epoch for non-mirrored writes */
1040 start_new_tl_epoch(first_peer_device(device)->connection);
1042 if (on_congestion == OC_PULL_AHEAD)
1043 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1044 else /*nc->on_congestion == OC_DISCONNECT */
1045 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1050 /* If this returns false, and req->private_bio is still set,
1051 * this should be submitted locally.
1053 * If it returns false, but req->private_bio is not set,
1054 * we do not have access to good data :(
1056 * Otherwise, this destroys req->private_bio, if any,
1059 static bool do_remote_read(struct drbd_request *req)
1061 struct drbd_device *device = req->device;
1062 enum drbd_read_balancing rbm;
1064 if (req->private_bio) {
1065 if (!drbd_may_do_local_read(device,
1066 req->i.sector, req->i.size)) {
1067 bio_put(req->private_bio);
1068 req->private_bio = NULL;
1073 if (device->state.pdsk != D_UP_TO_DATE)
1076 if (req->private_bio == NULL)
1079 /* TODO: improve read balancing decisions, take into account drbd
1080 * protocol, pending requests etc. */
1083 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1086 if (rbm == RB_PREFER_LOCAL && req->private_bio)
1087 return false; /* submit locally */
1089 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1090 if (req->private_bio) {
1091 bio_put(req->private_bio);
1092 req->private_bio = NULL;
1101 bool drbd_should_do_remote(union drbd_dev_state s)
1103 return s.pdsk == D_UP_TO_DATE ||
1104 (s.pdsk >= D_INCONSISTENT &&
1105 s.conn >= C_WF_BITMAP_T &&
1107 /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1108 That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1112 static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1114 return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1115 /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1116 since we enter state C_AHEAD only if proto >= 96 */
1119 /* returns number of connections (== 1, for drbd 8.4)
1120 * expected to actually write this data,
1121 * which does NOT include those that we are L_AHEAD for. */
1122 static int drbd_process_write_request(struct drbd_request *req)
1124 struct drbd_device *device = req->device;
1125 int remote, send_oos;
1127 remote = drbd_should_do_remote(device->state);
1128 send_oos = drbd_should_send_out_of_sync(device->state);
1130 /* Need to replicate writes. Unless it is an empty flush,
1131 * which is better mapped to a DRBD P_BARRIER packet,
1132 * also for drbd wire protocol compatibility reasons.
1133 * If this was a flush, just start a new epoch.
1134 * Unless the current epoch was empty anyways, or we are not currently
1135 * replicating, in which case there is no point. */
1136 if (unlikely(req->i.size == 0)) {
1137 /* The only size==0 bios we expect are empty flushes. */
1138 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1140 _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1144 if (!remote && !send_oos)
1147 D_ASSERT(device, !(remote && send_oos));
1150 _req_mod(req, TO_BE_SENT);
1151 _req_mod(req, QUEUE_FOR_NET_WRITE);
1152 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1153 _req_mod(req, QUEUE_FOR_SEND_OOS);
1158 static void drbd_process_discard_req(struct drbd_request *req)
1160 struct block_device *bdev = req->device->ldev->backing_bdev;
1162 if (blkdev_issue_zeroout(bdev, req->i.sector, req->i.size >> 9,
1164 req->private_bio->bi_status = BLK_STS_IOERR;
1165 bio_endio(req->private_bio);
1169 drbd_submit_req_private_bio(struct drbd_request *req)
1171 struct drbd_device *device = req->device;
1172 struct bio *bio = req->private_bio;
1175 if (bio_op(bio) != REQ_OP_READ)
1176 type = DRBD_FAULT_DT_WR;
1177 else if (bio->bi_opf & REQ_RAHEAD)
1178 type = DRBD_FAULT_DT_RA;
1180 type = DRBD_FAULT_DT_RD;
1182 bio_set_dev(bio, device->ldev->backing_bdev);
1184 /* State may have changed since we grabbed our reference on the
1185 * ->ldev member. Double check, and short-circuit to endio.
1186 * In case the last activity log transaction failed to get on
1187 * stable storage, and this is a WRITE, we may not even submit
1189 if (get_ldev(device)) {
1190 if (drbd_insert_fault(device, type))
1192 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1193 bio_op(bio) == REQ_OP_DISCARD)
1194 drbd_process_discard_req(req);
1196 generic_make_request(bio);
1202 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1204 spin_lock_irq(&device->resource->req_lock);
1205 list_add_tail(&req->tl_requests, &device->submit.writes);
1206 list_add_tail(&req->req_pending_master_completion,
1207 &device->pending_master_completion[1 /* WRITE */]);
1208 spin_unlock_irq(&device->resource->req_lock);
1209 queue_work(device->submit.wq, &device->submit.worker);
1210 /* do_submit() may sleep internally on al_wait, too */
1211 wake_up(&device->al_wait);
1214 /* returns the new drbd_request pointer, if the caller is expected to
1215 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1216 * request on the submitter thread.
1217 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1219 static struct drbd_request *
1220 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1222 const int rw = bio_data_dir(bio);
1223 struct drbd_request *req;
1225 /* allocate outside of all locks; */
1226 req = drbd_req_new(device, bio);
1229 /* only pass the error to the upper layers.
1230 * if user cannot handle io errors, that's not our business. */
1231 drbd_err(device, "could not kmalloc() req\n");
1232 bio->bi_status = BLK_STS_RESOURCE;
1234 return ERR_PTR(-ENOMEM);
1236 req->start_jif = start_jif;
1238 if (!get_ldev(device)) {
1239 bio_put(req->private_bio);
1240 req->private_bio = NULL;
1243 /* Update disk stats */
1244 _drbd_start_io_acct(device, req);
1246 /* process discards always from our submitter thread */
1247 if ((bio_op(bio) & REQ_OP_WRITE_ZEROES) ||
1248 (bio_op(bio) & REQ_OP_DISCARD))
1249 goto queue_for_submitter_thread;
1251 if (rw == WRITE && req->private_bio && req->i.size
1252 && !test_bit(AL_SUSPENDED, &device->flags)) {
1253 if (!drbd_al_begin_io_fastpath(device, &req->i))
1254 goto queue_for_submitter_thread;
1255 req->rq_state |= RQ_IN_ACT_LOG;
1256 req->in_actlog_jif = jiffies;
1260 queue_for_submitter_thread:
1261 atomic_inc(&device->ap_actlog_cnt);
1262 drbd_queue_write(device, req);
1266 /* Require at least one path to current data.
1267 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1268 * We would not allow to read what was written,
1269 * we would not have bumped the data generation uuids,
1270 * we would cause data divergence for all the wrong reasons.
1272 * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1273 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1274 * and queues for retry later.
1276 static bool may_do_writes(struct drbd_device *device)
1278 const union drbd_dev_state s = device->state;
1279 return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1282 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1284 struct drbd_resource *resource = device->resource;
1285 const int rw = bio_data_dir(req->master_bio);
1286 struct bio_and_error m = { NULL, };
1287 bool no_remote = false;
1288 bool submit_private_bio = false;
1290 spin_lock_irq(&resource->req_lock);
1292 /* This may temporarily give up the req_lock,
1293 * but will re-aquire it before it returns here.
1294 * Needs to be before the check on drbd_suspended() */
1295 complete_conflicting_writes(req);
1296 /* no more giving up req_lock from now on! */
1298 /* check for congestion, and potentially stop sending
1299 * full data updates, but start sending "dirty bits" only. */
1300 maybe_pull_ahead(device);
1304 if (drbd_suspended(device)) {
1305 /* push back and retry: */
1306 req->rq_state |= RQ_POSTPONED;
1307 if (req->private_bio) {
1308 bio_put(req->private_bio);
1309 req->private_bio = NULL;
1315 /* We fail READ early, if we can not serve it.
1316 * We must do this before req is registered on any lists.
1317 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1319 if (!do_remote_read(req) && !req->private_bio)
1323 /* which transfer log epoch does this belong to? */
1324 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1326 /* no point in adding empty flushes to the transfer log,
1327 * they are mapped to drbd barriers already. */
1328 if (likely(req->i.size!=0)) {
1330 first_peer_device(device)->connection->current_tle_writes++;
1332 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1336 if (req->private_bio && !may_do_writes(device)) {
1337 bio_put(req->private_bio);
1338 req->private_bio = NULL;
1342 if (!drbd_process_write_request(req))
1345 /* We either have a private_bio, or we can read from remote.
1346 * Otherwise we had done the goto nodata above. */
1347 if (req->private_bio == NULL) {
1348 _req_mod(req, TO_BE_SENT);
1349 _req_mod(req, QUEUE_FOR_NET_READ);
1354 /* If it took the fast path in drbd_request_prepare, add it here.
1355 * The slow path has added it already. */
1356 if (list_empty(&req->req_pending_master_completion))
1357 list_add_tail(&req->req_pending_master_completion,
1358 &device->pending_master_completion[rw == WRITE]);
1359 if (req->private_bio) {
1360 /* needs to be marked within the same spinlock */
1361 req->pre_submit_jif = jiffies;
1362 list_add_tail(&req->req_pending_local,
1363 &device->pending_completion[rw == WRITE]);
1364 _req_mod(req, TO_BE_SUBMITTED);
1365 /* but we need to give up the spinlock to submit */
1366 submit_private_bio = true;
1367 } else if (no_remote) {
1369 if (__ratelimit(&drbd_ratelimit_state))
1370 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1371 (unsigned long long)req->i.sector, req->i.size >> 9);
1372 /* A write may have been queued for send_oos, however.
1373 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1377 drbd_req_put_completion_ref(req, &m, 1);
1378 spin_unlock_irq(&resource->req_lock);
1380 /* Even though above is a kref_put(), this is safe.
1381 * As long as we still need to submit our private bio,
1382 * we hold a completion ref, and the request cannot disappear.
1383 * If however this request did not even have a private bio to submit
1384 * (e.g. remote read), req may already be invalid now.
1385 * That's why we cannot check on req->private_bio. */
1386 if (submit_private_bio)
1387 drbd_submit_req_private_bio(req);
1389 complete_master_bio(device, &m);
1392 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1394 struct drbd_request *req = drbd_request_prepare(device, bio, start_jif);
1395 if (IS_ERR_OR_NULL(req))
1397 drbd_send_and_submit(device, req);
1400 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1402 struct drbd_request *req, *tmp;
1403 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1404 const int rw = bio_data_dir(req->master_bio);
1406 if (rw == WRITE /* rw != WRITE should not even end up here! */
1407 && req->private_bio && req->i.size
1408 && !test_bit(AL_SUSPENDED, &device->flags)) {
1409 if (!drbd_al_begin_io_fastpath(device, &req->i))
1412 req->rq_state |= RQ_IN_ACT_LOG;
1413 req->in_actlog_jif = jiffies;
1414 atomic_dec(&device->ap_actlog_cnt);
1417 list_del_init(&req->tl_requests);
1418 drbd_send_and_submit(device, req);
1422 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1423 struct list_head *incoming,
1424 struct list_head *pending,
1425 struct list_head *later)
1427 struct drbd_request *req, *tmp;
1431 spin_lock_irq(&device->al_lock);
1432 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1433 err = drbd_al_begin_io_nonblock(device, &req->i);
1434 if (err == -ENOBUFS)
1439 list_move_tail(&req->tl_requests, later);
1441 list_move_tail(&req->tl_requests, pending);
1443 spin_unlock_irq(&device->al_lock);
1445 wake_up(&device->al_wait);
1446 return !list_empty(pending);
1449 void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1451 struct drbd_request *req, *tmp;
1453 list_for_each_entry_safe(req, tmp, pending, tl_requests) {
1454 req->rq_state |= RQ_IN_ACT_LOG;
1455 req->in_actlog_jif = jiffies;
1456 atomic_dec(&device->ap_actlog_cnt);
1457 list_del_init(&req->tl_requests);
1458 drbd_send_and_submit(device, req);
1462 void do_submit(struct work_struct *ws)
1464 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1465 LIST_HEAD(incoming); /* from drbd_make_request() */
1466 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */
1467 LIST_HEAD(busy); /* blocked by resync requests */
1469 /* grab new incoming requests */
1470 spin_lock_irq(&device->resource->req_lock);
1471 list_splice_tail_init(&device->submit.writes, &incoming);
1472 spin_unlock_irq(&device->resource->req_lock);
1477 /* move used-to-be-busy back to front of incoming */
1478 list_splice_init(&busy, &incoming);
1479 submit_fast_path(device, &incoming);
1480 if (list_empty(&incoming))
1484 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1486 list_splice_init(&busy, &incoming);
1487 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1488 if (!list_empty(&pending))
1493 /* If all currently "hot" activity log extents are kept busy by
1494 * incoming requests, we still must not totally starve new
1495 * requests to "cold" extents.
1496 * Something left on &incoming means there had not been
1497 * enough update slots available, and the activity log
1498 * has been marked as "starving".
1500 * Try again now, without looking for new requests,
1501 * effectively blocking all new requests until we made
1502 * at least _some_ progress with what we currently have.
1504 if (!list_empty(&incoming))
1507 /* Nothing moved to pending, but nothing left
1508 * on incoming: all moved to busy!
1509 * Grab new and iterate. */
1510 spin_lock_irq(&device->resource->req_lock);
1511 list_splice_tail_init(&device->submit.writes, &incoming);
1512 spin_unlock_irq(&device->resource->req_lock);
1514 finish_wait(&device->al_wait, &wait);
1516 /* If the transaction was full, before all incoming requests
1517 * had been processed, skip ahead to commit, and iterate
1518 * without splicing in more incoming requests from upper layers.
1520 * Else, if all incoming have been processed,
1521 * they have become either "pending" (to be submitted after
1522 * next transaction commit) or "busy" (blocked by resync).
1524 * Maybe more was queued, while we prepared the transaction?
1525 * Try to stuff those into this transaction as well.
1526 * Be strictly non-blocking here,
1527 * we already have something to commit.
1529 * Commit if we don't make any more progres.
1532 while (list_empty(&incoming)) {
1533 LIST_HEAD(more_pending);
1534 LIST_HEAD(more_incoming);
1537 /* It is ok to look outside the lock,
1538 * it's only an optimization anyways */
1539 if (list_empty(&device->submit.writes))
1542 spin_lock_irq(&device->resource->req_lock);
1543 list_splice_tail_init(&device->submit.writes, &more_incoming);
1544 spin_unlock_irq(&device->resource->req_lock);
1546 if (list_empty(&more_incoming))
1549 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1551 list_splice_tail_init(&more_pending, &pending);
1552 list_splice_tail_init(&more_incoming, &incoming);
1557 drbd_al_begin_io_commit(device);
1558 send_and_submit_pending(device, &pending);
1562 blk_qc_t drbd_make_request(struct request_queue *q, struct bio *bio)
1564 struct drbd_device *device = (struct drbd_device *) q->queuedata;
1565 unsigned long start_jif;
1567 blk_queue_split(q, &bio);
1569 start_jif = jiffies;
1572 * what we "blindly" assume:
1574 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1577 __drbd_make_request(device, bio, start_jif);
1578 return BLK_QC_T_NONE;
1581 static bool net_timeout_reached(struct drbd_request *net_req,
1582 struct drbd_connection *connection,
1583 unsigned long now, unsigned long ent,
1584 unsigned int ko_count, unsigned int timeout)
1586 struct drbd_device *device = net_req->device;
1588 if (!time_after(now, net_req->pre_send_jif + ent))
1591 if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1594 if (net_req->rq_state & RQ_NET_PENDING) {
1595 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1596 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1600 /* We received an ACK already (or are using protocol A),
1601 * but are waiting for the epoch closing barrier ack.
1602 * Check if we sent the barrier already. We should not blame the peer
1603 * for being unresponsive, if we did not even ask it yet. */
1604 if (net_req->epoch == connection->send.current_epoch_nr) {
1606 "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1607 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1611 /* Worst case: we may have been blocked for whatever reason, then
1612 * suddenly are able to send a lot of requests (and epoch separating
1613 * barriers) in quick succession.
1614 * The timestamp of the net_req may be much too old and not correspond
1615 * to the sending time of the relevant unack'ed barrier packet, so
1616 * would trigger a spurious timeout. The latest barrier packet may
1617 * have a too recent timestamp to trigger the timeout, potentially miss
1618 * a timeout. Right now we don't have a place to conveniently store
1620 * But in this particular situation, the application requests are still
1621 * completed to upper layers, DRBD should still "feel" responsive.
1622 * No need yet to kill this connection, it may still recover.
1623 * If not, eventually we will have queued enough into the network for
1624 * us to block. From that point of view, the timestamp of the last sent
1625 * barrier packet is relevant enough.
1627 if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1628 drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1629 connection->send.last_sent_barrier_jif, now,
1630 jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1636 /* A request is considered timed out, if
1637 * - we have some effective timeout from the configuration,
1638 * with some state restrictions applied,
1639 * - the oldest request is waiting for a response from the network
1640 * resp. the local disk,
1641 * - the oldest request is in fact older than the effective timeout,
1642 * - the connection was established (resp. disk was attached)
1643 * for longer than the timeout already.
1644 * Note that for 32bit jiffies and very stable connections/disks,
1645 * we may have a wrap around, which is catched by
1646 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1648 * Side effect: once per 32bit wrap-around interval, which means every
1649 * ~198 days with 250 HZ, we have a window where the timeout would need
1650 * to expire twice (worst case) to become effective. Good enough.
1653 void request_timer_fn(unsigned long data)
1655 struct drbd_device *device = (struct drbd_device *) data;
1656 struct drbd_connection *connection = first_peer_device(device)->connection;
1657 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1658 struct net_conf *nc;
1659 unsigned long oldest_submit_jif;
1660 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1662 unsigned int ko_count = 0, timeout = 0;
1665 nc = rcu_dereference(connection->net_conf);
1666 if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1667 ko_count = nc->ko_count;
1668 timeout = nc->timeout;
1671 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1672 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1678 ent = timeout * HZ/10 * ko_count;
1679 et = min_not_zero(dt, ent);
1682 return; /* Recurring timer stopped */
1687 spin_lock_irq(&device->resource->req_lock);
1688 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1689 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1691 /* maybe the oldest request waiting for the peer is in fact still
1692 * blocking in tcp sendmsg. That's ok, though, that's handled via the
1693 * socket send timeout, requesting a ping, and bumping ko-count in
1694 * we_should_drop_the_connection().
1697 /* check the oldest request we did successfully sent,
1698 * but which is still waiting for an ACK. */
1699 req_peer = connection->req_ack_pending;
1701 /* if we don't have such request (e.g. protocoll A)
1702 * check the oldest requests which is still waiting on its epoch
1703 * closing barrier ack. */
1705 req_peer = connection->req_not_net_done;
1707 /* evaluate the oldest peer request only in one timer! */
1708 if (req_peer && req_peer->device != device)
1711 /* do we have something to evaluate? */
1712 if (req_peer == NULL && req_write == NULL && req_read == NULL)
1716 (req_write && req_read)
1717 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1718 ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1719 : req_write ? req_write->pre_submit_jif
1720 : req_read ? req_read->pre_submit_jif : now;
1722 if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1723 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1725 if (dt && oldest_submit_jif != now &&
1726 time_after(now, oldest_submit_jif + dt) &&
1727 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1728 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1729 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1732 /* Reschedule timer for the nearest not already expired timeout.
1733 * Fallback to now + min(effective network timeout, disk timeout). */
1734 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1735 ? req_peer->pre_send_jif + ent : now + et;
1736 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1737 ? oldest_submit_jif + dt : now + et;
1738 nt = time_before(ent, dt) ? ent : dt;
1740 spin_unlock_irq(&device->resource->req_lock);
1741 mod_timer(&device->request_timer, nt);