1 // SPDX-License-Identifier: GPL-2.0-or-later
5 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
7 Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
8 Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
9 Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
14 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/drbd.h>
22 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
24 static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src)
26 struct drbd_request *req;
28 req = mempool_alloc(&drbd_request_mempool, GFP_NOIO);
31 memset(req, 0, sizeof(*req));
33 req->private_bio = bio_clone_fast(bio_src, GFP_NOIO, &drbd_io_bio_set);
34 req->private_bio->bi_private = req;
35 req->private_bio->bi_end_io = drbd_request_endio;
37 req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
38 | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0)
39 | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0)
40 | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
42 req->master_bio = bio_src;
45 drbd_clear_interval(&req->i);
46 req->i.sector = bio_src->bi_iter.bi_sector;
47 req->i.size = bio_src->bi_iter.bi_size;
49 req->i.waiting = false;
51 INIT_LIST_HEAD(&req->tl_requests);
52 INIT_LIST_HEAD(&req->w.list);
53 INIT_LIST_HEAD(&req->req_pending_master_completion);
54 INIT_LIST_HEAD(&req->req_pending_local);
56 /* one reference to be put by __drbd_make_request */
57 atomic_set(&req->completion_ref, 1);
58 /* one kref as long as completion_ref > 0 */
59 kref_init(&req->kref);
63 static void drbd_remove_request_interval(struct rb_root *root,
64 struct drbd_request *req)
66 struct drbd_device *device = req->device;
67 struct drbd_interval *i = &req->i;
69 drbd_remove_interval(root, i);
71 /* Wake up any processes waiting for this request to complete. */
73 wake_up(&device->misc_wait);
76 void drbd_req_destroy(struct kref *kref)
78 struct drbd_request *req = container_of(kref, struct drbd_request, kref);
79 struct drbd_device *device = req->device;
80 const unsigned s = req->rq_state;
82 if ((req->master_bio && !(s & RQ_POSTPONED)) ||
83 atomic_read(&req->completion_ref) ||
84 (s & RQ_LOCAL_PENDING) ||
85 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
86 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
87 s, atomic_read(&req->completion_ref));
91 /* If called from mod_rq_state (expected normal case) or
92 * drbd_send_and_submit (the less likely normal path), this holds the
93 * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
94 * though it may be still empty (never added to the transfer log).
96 * If called from do_retry(), we do NOT hold the req_lock, but we are
97 * still allowed to unconditionally list_del(&req->tl_requests),
98 * because it will be on a local on-stack list only. */
99 list_del_init(&req->tl_requests);
101 /* finally remove the request from the conflict detection
102 * respective block_id verification interval tree. */
103 if (!drbd_interval_empty(&req->i)) {
104 struct rb_root *root;
107 root = &device->write_requests;
109 root = &device->read_requests;
110 drbd_remove_request_interval(root, req);
111 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
112 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
113 s, (unsigned long long)req->i.sector, req->i.size);
115 /* if it was a write, we may have to set the corresponding
116 * bit(s) out-of-sync first. If it had a local part, we need to
117 * release the reference to the activity log. */
119 /* Set out-of-sync unless both OK flags are set
120 * (local only or remote failed).
121 * Other places where we set out-of-sync:
122 * READ with local io-error */
124 /* There is a special case:
125 * we may notice late that IO was suspended,
126 * and postpone, or schedule for retry, a write,
127 * before it even was submitted or sent.
128 * In that case we do not want to touch the bitmap at all.
130 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
131 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
132 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
134 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
135 drbd_set_in_sync(device, req->i.sector, req->i.size);
138 /* one might be tempted to move the drbd_al_complete_io
139 * to the local io completion callback drbd_request_endio.
140 * but, if this was a mirror write, we may only
141 * drbd_al_complete_io after this is RQ_NET_DONE,
142 * otherwise the extent could be dropped from the al
143 * before it has actually been written on the peer.
144 * if we crash before our peer knows about the request,
145 * but after the extent has been dropped from the al,
146 * we would forget to resync the corresponding extent.
148 if (s & RQ_IN_ACT_LOG) {
149 if (get_ldev_if_state(device, D_FAILED)) {
150 drbd_al_complete_io(device, &req->i);
152 } else if (__ratelimit(&drbd_ratelimit_state)) {
153 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
154 "but my Disk seems to have failed :(\n",
155 (unsigned long long) req->i.sector, req->i.size);
160 mempool_free(req, &drbd_request_mempool);
163 static void wake_all_senders(struct drbd_connection *connection)
165 wake_up(&connection->sender_work.q_wait);
168 /* must hold resource->req_lock */
169 void start_new_tl_epoch(struct drbd_connection *connection)
171 /* no point closing an epoch, if it is empty, anyways. */
172 if (connection->current_tle_writes == 0)
175 connection->current_tle_writes = 0;
176 atomic_inc(&connection->current_tle_nr);
177 wake_all_senders(connection);
180 void complete_master_bio(struct drbd_device *device,
181 struct bio_and_error *m)
183 m->bio->bi_status = errno_to_blk_status(m->error);
189 /* Helper for __req_mod().
190 * Set m->bio to the master bio, if it is fit to be completed,
191 * or leave it alone (it is initialized to NULL in __req_mod),
192 * if it has already been completed, or cannot be completed yet.
193 * If m->bio is set, the error status to be returned is placed in m->error.
196 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
198 const unsigned s = req->rq_state;
199 struct drbd_device *device = req->device;
202 /* we must not complete the master bio, while it is
203 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
204 * not yet acknowledged by the peer
205 * not yet completed by the local io subsystem
206 * these flags may get cleared in any order by
209 * the bio_endio completion callbacks.
211 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
212 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
213 (s & RQ_COMPLETION_SUSP)) {
214 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
218 if (!req->master_bio) {
219 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
224 * figure out whether to report success or failure.
226 * report success when at least one of the operations succeeded.
227 * or, to put the other way,
228 * only report failure, when both operations failed.
230 * what to do about the failures is handled elsewhere.
231 * what we need to do here is just: complete the master_bio.
233 * local completion error, if any, has been stored as ERR_PTR
234 * in private_bio within drbd_request_endio.
236 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
237 error = PTR_ERR(req->private_bio);
239 /* Before we can signal completion to the upper layers,
240 * we may need to close the current transfer log epoch.
241 * We are within the request lock, so we can simply compare
242 * the request epoch number with the current transfer log
243 * epoch number. If they match, increase the current_tle_nr,
244 * and reset the transfer log epoch write_cnt.
246 if (op_is_write(bio_op(req->master_bio)) &&
247 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
248 start_new_tl_epoch(first_peer_device(device)->connection);
250 /* Update disk stats */
251 bio_end_io_acct(req->master_bio, req->start_jif);
254 * have it be pushed back to the retry work queue,
255 * so it will re-enter __drbd_make_request(),
256 * and be re-assigned to a suitable local or remote path,
257 * or failed if we do not have access to good data anymore.
259 * Unless it was failed early by __drbd_make_request(),
260 * because no path was available, in which case
261 * it was not even added to the transfer_log.
263 * read-ahead may fail, and will not be retried.
265 * WRITE should have used all available paths already.
268 bio_op(req->master_bio) == REQ_OP_READ &&
269 !(req->master_bio->bi_opf & REQ_RAHEAD) &&
270 !list_empty(&req->tl_requests))
271 req->rq_state |= RQ_POSTPONED;
273 if (!(req->rq_state & RQ_POSTPONED)) {
274 m->error = ok ? 0 : (error ?: -EIO);
275 m->bio = req->master_bio;
276 req->master_bio = NULL;
277 /* We leave it in the tree, to be able to verify later
278 * write-acks in protocol != C during resync.
279 * But we mark it as "complete", so it won't be counted as
280 * conflict in a multi-primary setup. */
281 req->i.completed = true;
285 wake_up(&device->misc_wait);
287 /* Either we are about to complete to upper layers,
288 * or we will restart this request.
289 * In either case, the request object will be destroyed soon,
290 * so better remove it from all lists. */
291 list_del_init(&req->req_pending_master_completion);
294 /* still holds resource->req_lock */
295 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
297 struct drbd_device *device = req->device;
298 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
303 if (!atomic_sub_and_test(put, &req->completion_ref))
306 drbd_req_complete(req, m);
308 /* local completion may still come in later,
309 * we need to keep the req object around. */
310 if (req->rq_state & RQ_LOCAL_ABORTED)
313 if (req->rq_state & RQ_POSTPONED) {
314 /* don't destroy the req object just yet,
315 * but queue it for retry */
316 drbd_restart_request(req);
320 kref_put(&req->kref, drbd_req_destroy);
323 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
325 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
328 if (connection->req_next == NULL)
329 connection->req_next = req;
332 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
334 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
337 if (connection->req_next != req)
339 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
340 const unsigned s = req->rq_state;
341 if (s & RQ_NET_QUEUED)
344 if (&req->tl_requests == &connection->transfer_log)
346 connection->req_next = req;
349 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
351 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
354 if (connection->req_ack_pending == NULL)
355 connection->req_ack_pending = req;
358 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
360 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
363 if (connection->req_ack_pending != req)
365 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
366 const unsigned s = req->rq_state;
367 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
370 if (&req->tl_requests == &connection->transfer_log)
372 connection->req_ack_pending = req;
375 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
377 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
380 if (connection->req_not_net_done == NULL)
381 connection->req_not_net_done = req;
384 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
386 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
389 if (connection->req_not_net_done != req)
391 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
392 const unsigned s = req->rq_state;
393 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
396 if (&req->tl_requests == &connection->transfer_log)
398 connection->req_not_net_done = req;
401 /* I'd like this to be the only place that manipulates
402 * req->completion_ref and req->kref. */
403 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
406 struct drbd_device *device = req->device;
407 struct drbd_peer_device *peer_device = first_peer_device(device);
408 unsigned s = req->rq_state;
411 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
412 set |= RQ_COMPLETION_SUSP;
416 req->rq_state &= ~clear;
417 req->rq_state |= set;
420 if (req->rq_state == s)
423 /* intent: get references */
425 kref_get(&req->kref);
427 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
428 atomic_inc(&req->completion_ref);
430 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
431 inc_ap_pending(device);
432 atomic_inc(&req->completion_ref);
435 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
436 atomic_inc(&req->completion_ref);
437 set_if_null_req_next(peer_device, req);
440 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
441 kref_get(&req->kref); /* wait for the DONE */
443 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
444 /* potentially already completed in the ack_receiver thread */
445 if (!(s & RQ_NET_DONE)) {
446 atomic_add(req->i.size >> 9, &device->ap_in_flight);
447 set_if_null_req_not_net_done(peer_device, req);
449 if (req->rq_state & RQ_NET_PENDING)
450 set_if_null_req_ack_pending(peer_device, req);
453 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
454 atomic_inc(&req->completion_ref);
456 /* progress: put references */
458 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
461 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
462 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
466 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
467 if (req->rq_state & RQ_LOCAL_ABORTED)
468 kref_put(&req->kref, drbd_req_destroy);
471 list_del_init(&req->req_pending_local);
474 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
475 dec_ap_pending(device);
477 req->acked_jif = jiffies;
478 advance_conn_req_ack_pending(peer_device, req);
481 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
483 advance_conn_req_next(peer_device, req);
486 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
488 atomic_sub(req->i.size >> 9, &device->ap_in_flight);
489 if (s & RQ_EXP_BARR_ACK)
490 kref_put(&req->kref, drbd_req_destroy);
491 req->net_done_jif = jiffies;
493 /* in ahead/behind mode, or just in case,
494 * before we finally destroy this request,
495 * the caching pointers must not reference it anymore */
496 advance_conn_req_next(peer_device, req);
497 advance_conn_req_ack_pending(peer_device, req);
498 advance_conn_req_not_net_done(peer_device, req);
501 /* potentially complete and destroy */
503 /* If we made progress, retry conflicting peer requests, if any. */
505 wake_up(&device->misc_wait);
507 drbd_req_put_completion_ref(req, m, c_put);
508 kref_put(&req->kref, drbd_req_destroy);
511 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
513 char b[BDEVNAME_SIZE];
515 if (!__ratelimit(&drbd_ratelimit_state))
518 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
519 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
520 (unsigned long long)req->i.sector,
522 bdevname(device->ldev->backing_bdev, b));
525 /* Helper for HANDED_OVER_TO_NETWORK.
526 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
527 * Is it also still "PENDING"?
528 * --> If so, clear PENDING and set NET_OK below.
529 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
530 * (and we must not set RQ_NET_OK) */
531 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
533 return (req->rq_state &
534 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
535 == (RQ_WRITE|RQ_NET_PENDING);
538 /* obviously this could be coded as many single functions
539 * instead of one huge switch,
540 * or by putting the code directly in the respective locations
541 * (as it has been before).
543 * but having it this way
544 * enforces that it is all in this one place, where it is easier to audit,
545 * it makes it obvious that whatever "event" "happens" to a request should
546 * happen "atomically" within the req_lock,
547 * and it enforces that we have to think in a very structured manner
548 * about the "events" that may happen to a request during its life time ...
550 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
551 struct bio_and_error *m)
553 struct drbd_device *const device = req->device;
554 struct drbd_peer_device *const peer_device = first_peer_device(device);
555 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
564 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
567 /* does not happen...
568 * initialization done in drbd_req_new
573 case TO_BE_SENT: /* via network */
574 /* reached via __drbd_make_request
575 * and from w_read_retry_remote */
576 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
578 nc = rcu_dereference(connection->net_conf);
579 p = nc->wire_protocol;
582 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
583 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
584 mod_rq_state(req, m, 0, RQ_NET_PENDING);
587 case TO_BE_SUBMITTED: /* locally */
588 /* reached via __drbd_make_request */
589 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
590 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
594 if (req->rq_state & RQ_WRITE)
595 device->writ_cnt += req->i.size >> 9;
597 device->read_cnt += req->i.size >> 9;
599 mod_rq_state(req, m, RQ_LOCAL_PENDING,
600 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
604 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
607 case WRITE_COMPLETED_WITH_ERROR:
608 drbd_report_io_error(device, req);
609 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
610 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
613 case READ_COMPLETED_WITH_ERROR:
614 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
615 drbd_report_io_error(device, req);
616 __drbd_chk_io_error(device, DRBD_READ_ERROR);
618 case READ_AHEAD_COMPLETED_WITH_ERROR:
619 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
620 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
623 case DISCARD_COMPLETED_NOTSUPP:
624 case DISCARD_COMPLETED_WITH_ERROR:
625 /* I'd rather not detach from local disk just because it
626 * failed a REQ_OP_DISCARD. */
627 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
630 case QUEUE_FOR_NET_READ:
633 * or target area marked as invalid,
634 * or just got an io-error. */
635 /* from __drbd_make_request
636 * or from bio_endio during read io-error recovery */
638 /* So we can verify the handle in the answer packet.
639 * Corresponding drbd_remove_request_interval is in
640 * drbd_req_complete() */
641 D_ASSERT(device, drbd_interval_empty(&req->i));
642 drbd_insert_interval(&device->read_requests, &req->i);
644 set_bit(UNPLUG_REMOTE, &device->flags);
646 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
647 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
648 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
649 req->w.cb = w_send_read_req;
650 drbd_queue_work(&connection->sender_work,
654 case QUEUE_FOR_NET_WRITE:
655 /* assert something? */
656 /* from __drbd_make_request only */
658 /* Corresponding drbd_remove_request_interval is in
659 * drbd_req_complete() */
660 D_ASSERT(device, drbd_interval_empty(&req->i));
661 drbd_insert_interval(&device->write_requests, &req->i);
664 * In case the req ended up on the transfer log before being
665 * queued on the worker, it could lead to this request being
666 * missed during cleanup after connection loss.
667 * So we have to do both operations here,
668 * within the same lock that protects the transfer log.
670 * _req_add_to_epoch(req); this has to be after the
671 * _maybe_start_new_epoch(req); which happened in
672 * __drbd_make_request, because we now may set the bit
673 * again ourselves to close the current epoch.
675 * Add req to the (now) current epoch (barrier). */
677 /* otherwise we may lose an unplug, which may cause some remote
678 * io-scheduler timeout to expire, increasing maximum latency,
679 * hurting performance. */
680 set_bit(UNPLUG_REMOTE, &device->flags);
682 /* queue work item to send data */
683 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
684 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
685 req->w.cb = w_send_dblock;
686 drbd_queue_work(&connection->sender_work,
689 /* close the epoch, in case it outgrew the limit */
691 nc = rcu_dereference(connection->net_conf);
692 p = nc->max_epoch_size;
694 if (connection->current_tle_writes >= p)
695 start_new_tl_epoch(connection);
699 case QUEUE_FOR_SEND_OOS:
700 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
701 req->w.cb = w_send_out_of_sync;
702 drbd_queue_work(&connection->sender_work,
706 case READ_RETRY_REMOTE_CANCELED:
709 /* real cleanup will be done from tl_clear. just update flags
710 * so it is no longer marked as on the worker queue */
711 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
714 case HANDED_OVER_TO_NETWORK:
715 /* assert something? */
716 if (is_pending_write_protocol_A(req))
717 /* this is what is dangerous about protocol A:
718 * pretend it was successfully written on the peer. */
719 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
720 RQ_NET_SENT|RQ_NET_OK);
722 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
723 /* It is still not yet RQ_NET_DONE until the
724 * corresponding epoch barrier got acked as well,
725 * so we know what to dirty on connection loss. */
728 case OOS_HANDED_TO_NETWORK:
729 /* Was not set PENDING, no longer QUEUED, so is now DONE
730 * as far as this connection is concerned. */
731 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
734 case CONNECTION_LOST_WHILE_PENDING:
735 /* transfer log cleanup after connection loss */
737 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
741 case CONFLICT_RESOLVED:
742 /* for superseded conflicting writes of multiple primaries,
743 * there is no need to keep anything in the tl, potential
744 * node crashes are covered by the activity log.
746 * If this request had been marked as RQ_POSTPONED before,
747 * it will actually not be completed, but "restarted",
748 * resubmitted from the retry worker context. */
749 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
750 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
751 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
754 case WRITE_ACKED_BY_PEER_AND_SIS:
755 req->rq_state |= RQ_NET_SIS;
757 case WRITE_ACKED_BY_PEER:
758 /* Normal operation protocol C: successfully written on peer.
759 * During resync, even in protocol != C,
760 * we requested an explicit write ack anyways.
761 * Which means we cannot even assert anything here.
762 * Nothing more to do here.
763 * We want to keep the tl in place for all protocols, to cater
764 * for volatile write-back caches on lower level devices. */
766 case RECV_ACKED_BY_PEER:
767 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
768 /* protocol B; pretends to be successfully written on peer.
769 * see also notes above in HANDED_OVER_TO_NETWORK about
772 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
776 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
777 /* If this node has already detected the write conflict, the
778 * worker will be waiting on misc_wait. Wake it up once this
779 * request has completed locally.
781 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
782 req->rq_state |= RQ_POSTPONED;
784 wake_up(&device->misc_wait);
785 /* Do not clear RQ_NET_PENDING. This request will make further
786 * progress via restart_conflicting_writes() or
787 * fail_postponed_requests(). Hopefully. */
791 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
794 case FAIL_FROZEN_DISK_IO:
795 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
797 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
800 case RESTART_FROZEN_DISK_IO:
801 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
805 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
809 if (bio_data_dir(req->master_bio) == WRITE)
812 get_ldev(device); /* always succeeds in this call path */
813 req->w.cb = w_restart_disk_io;
814 drbd_queue_work(&connection->sender_work,
819 /* Simply complete (local only) READs. */
820 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
821 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
825 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
826 before the connection loss (B&C only); only P_BARRIER_ACK
827 (or the local completion?) was missing when we suspended.
828 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
829 During connection handshake, we ensure that the peer was not rebooted. */
830 if (!(req->rq_state & RQ_NET_OK)) {
831 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
832 * in that case we must not set RQ_NET_PENDING. */
834 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
836 /* w.cb expected to be w_send_dblock, or w_send_read_req */
837 drbd_queue_work(&connection->sender_work,
839 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
840 } /* else: FIXME can this happen? */
843 fallthrough; /* to BARRIER_ACKED */
846 /* barrier ack for READ requests does not make sense */
847 if (!(req->rq_state & RQ_WRITE))
850 if (req->rq_state & RQ_NET_PENDING) {
851 /* barrier came in before all requests were acked.
852 * this is bad, because if the connection is lost now,
853 * we won't be able to clean them up... */
854 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
856 /* Allowed to complete requests, even while suspended.
857 * As this is called for all requests within a matching epoch,
858 * we need to filter, and only set RQ_NET_DONE for those that
859 * have actually been on the wire. */
860 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
861 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
865 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
866 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
869 case QUEUE_AS_DRBD_BARRIER:
870 start_new_tl_epoch(connection);
871 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
878 /* we may do a local read if:
879 * - we are consistent (of course),
880 * - or we are generally inconsistent,
881 * BUT we are still/already IN SYNC for this area.
882 * since size may be bigger than BM_BLOCK_SIZE,
883 * we may need to check several bits.
885 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
887 unsigned long sbnr, ebnr;
888 sector_t esector, nr_sectors;
890 if (device->state.disk == D_UP_TO_DATE)
892 if (device->state.disk != D_INCONSISTENT)
894 esector = sector + (size >> 9) - 1;
895 nr_sectors = get_capacity(device->vdisk);
896 D_ASSERT(device, sector < nr_sectors);
897 D_ASSERT(device, esector < nr_sectors);
899 sbnr = BM_SECT_TO_BIT(sector);
900 ebnr = BM_SECT_TO_BIT(esector);
902 return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
905 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
906 enum drbd_read_balancing rbm)
908 struct backing_dev_info *bdi;
912 case RB_CONGESTED_REMOTE:
913 bdi = device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
914 return bdi_read_congested(bdi);
915 case RB_LEAST_PENDING:
916 return atomic_read(&device->local_cnt) >
917 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
918 case RB_32K_STRIPING: /* stripe_shift = 15 */
919 case RB_64K_STRIPING:
920 case RB_128K_STRIPING:
921 case RB_256K_STRIPING:
922 case RB_512K_STRIPING:
923 case RB_1M_STRIPING: /* stripe_shift = 20 */
924 stripe_shift = (rbm - RB_32K_STRIPING + 15);
925 return (sector >> (stripe_shift - 9)) & 1;
927 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
928 case RB_PREFER_REMOTE:
930 case RB_PREFER_LOCAL:
937 * complete_conflicting_writes - wait for any conflicting write requests
939 * The write_requests tree contains all active write requests which we
940 * currently know about. Wait for any requests to complete which conflict with
943 * Only way out: remove the conflicting intervals from the tree.
945 static void complete_conflicting_writes(struct drbd_request *req)
948 struct drbd_device *device = req->device;
949 struct drbd_interval *i;
950 sector_t sector = req->i.sector;
951 int size = req->i.size;
954 drbd_for_each_overlap(i, &device->write_requests, sector, size) {
955 /* Ignore, if already completed to upper layers. */
958 /* Handle the first found overlap. After the schedule
959 * we have to restart the tree walk. */
965 /* Indicate to wake up device->misc_wait on progress. */
966 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
968 spin_unlock_irq(&device->resource->req_lock);
970 spin_lock_irq(&device->resource->req_lock);
972 finish_wait(&device->misc_wait, &wait);
975 /* called within req_lock */
976 static void maybe_pull_ahead(struct drbd_device *device)
978 struct drbd_connection *connection = first_peer_device(device)->connection;
980 bool congested = false;
981 enum drbd_on_congestion on_congestion;
984 nc = rcu_dereference(connection->net_conf);
985 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
987 if (on_congestion == OC_BLOCK ||
988 connection->agreed_pro_version < 96)
991 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
992 return; /* nothing to do ... */
994 /* If I don't even have good local storage, we can not reasonably try
995 * to pull ahead of the peer. We also need the local reference to make
996 * sure device->act_log is there.
998 if (!get_ldev_if_state(device, D_UP_TO_DATE))
1001 if (nc->cong_fill &&
1002 atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1003 drbd_info(device, "Congestion-fill threshold reached\n");
1007 if (device->act_log->used >= nc->cong_extents) {
1008 drbd_info(device, "Congestion-extents threshold reached\n");
1013 /* start a new epoch for non-mirrored writes */
1014 start_new_tl_epoch(first_peer_device(device)->connection);
1016 if (on_congestion == OC_PULL_AHEAD)
1017 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1018 else /*nc->on_congestion == OC_DISCONNECT */
1019 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1024 /* If this returns false, and req->private_bio is still set,
1025 * this should be submitted locally.
1027 * If it returns false, but req->private_bio is not set,
1028 * we do not have access to good data :(
1030 * Otherwise, this destroys req->private_bio, if any,
1033 static bool do_remote_read(struct drbd_request *req)
1035 struct drbd_device *device = req->device;
1036 enum drbd_read_balancing rbm;
1038 if (req->private_bio) {
1039 if (!drbd_may_do_local_read(device,
1040 req->i.sector, req->i.size)) {
1041 bio_put(req->private_bio);
1042 req->private_bio = NULL;
1047 if (device->state.pdsk != D_UP_TO_DATE)
1050 if (req->private_bio == NULL)
1053 /* TODO: improve read balancing decisions, take into account drbd
1054 * protocol, pending requests etc. */
1057 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1060 if (rbm == RB_PREFER_LOCAL && req->private_bio)
1061 return false; /* submit locally */
1063 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1064 if (req->private_bio) {
1065 bio_put(req->private_bio);
1066 req->private_bio = NULL;
1075 bool drbd_should_do_remote(union drbd_dev_state s)
1077 return s.pdsk == D_UP_TO_DATE ||
1078 (s.pdsk >= D_INCONSISTENT &&
1079 s.conn >= C_WF_BITMAP_T &&
1081 /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1082 That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1086 static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1088 return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1089 /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1090 since we enter state C_AHEAD only if proto >= 96 */
1093 /* returns number of connections (== 1, for drbd 8.4)
1094 * expected to actually write this data,
1095 * which does NOT include those that we are L_AHEAD for. */
1096 static int drbd_process_write_request(struct drbd_request *req)
1098 struct drbd_device *device = req->device;
1099 int remote, send_oos;
1101 remote = drbd_should_do_remote(device->state);
1102 send_oos = drbd_should_send_out_of_sync(device->state);
1104 /* Need to replicate writes. Unless it is an empty flush,
1105 * which is better mapped to a DRBD P_BARRIER packet,
1106 * also for drbd wire protocol compatibility reasons.
1107 * If this was a flush, just start a new epoch.
1108 * Unless the current epoch was empty anyways, or we are not currently
1109 * replicating, in which case there is no point. */
1110 if (unlikely(req->i.size == 0)) {
1111 /* The only size==0 bios we expect are empty flushes. */
1112 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1114 _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1118 if (!remote && !send_oos)
1121 D_ASSERT(device, !(remote && send_oos));
1124 _req_mod(req, TO_BE_SENT);
1125 _req_mod(req, QUEUE_FOR_NET_WRITE);
1126 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1127 _req_mod(req, QUEUE_FOR_SEND_OOS);
1132 static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1134 int err = drbd_issue_discard_or_zero_out(req->device,
1135 req->i.sector, req->i.size >> 9, flags);
1137 req->private_bio->bi_status = BLK_STS_IOERR;
1138 bio_endio(req->private_bio);
1142 drbd_submit_req_private_bio(struct drbd_request *req)
1144 struct drbd_device *device = req->device;
1145 struct bio *bio = req->private_bio;
1148 if (bio_op(bio) != REQ_OP_READ)
1149 type = DRBD_FAULT_DT_WR;
1150 else if (bio->bi_opf & REQ_RAHEAD)
1151 type = DRBD_FAULT_DT_RA;
1153 type = DRBD_FAULT_DT_RD;
1155 bio_set_dev(bio, device->ldev->backing_bdev);
1157 /* State may have changed since we grabbed our reference on the
1158 * ->ldev member. Double check, and short-circuit to endio.
1159 * In case the last activity log transaction failed to get on
1160 * stable storage, and this is a WRITE, we may not even submit
1162 if (get_ldev(device)) {
1163 if (drbd_insert_fault(device, type))
1165 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1166 drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1167 ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1168 else if (bio_op(bio) == REQ_OP_DISCARD)
1169 drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1171 submit_bio_noacct(bio);
1177 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1179 spin_lock_irq(&device->resource->req_lock);
1180 list_add_tail(&req->tl_requests, &device->submit.writes);
1181 list_add_tail(&req->req_pending_master_completion,
1182 &device->pending_master_completion[1 /* WRITE */]);
1183 spin_unlock_irq(&device->resource->req_lock);
1184 queue_work(device->submit.wq, &device->submit.worker);
1185 /* do_submit() may sleep internally on al_wait, too */
1186 wake_up(&device->al_wait);
1189 /* returns the new drbd_request pointer, if the caller is expected to
1190 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1191 * request on the submitter thread.
1192 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1194 static struct drbd_request *
1195 drbd_request_prepare(struct drbd_device *device, struct bio *bio)
1197 const int rw = bio_data_dir(bio);
1198 struct drbd_request *req;
1200 /* allocate outside of all locks; */
1201 req = drbd_req_new(device, bio);
1204 /* only pass the error to the upper layers.
1205 * if user cannot handle io errors, that's not our business. */
1206 drbd_err(device, "could not kmalloc() req\n");
1207 bio->bi_status = BLK_STS_RESOURCE;
1209 return ERR_PTR(-ENOMEM);
1212 /* Update disk stats */
1213 req->start_jif = bio_start_io_acct(req->master_bio);
1215 if (!get_ldev(device)) {
1216 bio_put(req->private_bio);
1217 req->private_bio = NULL;
1220 /* process discards always from our submitter thread */
1221 if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1222 bio_op(bio) == REQ_OP_DISCARD)
1223 goto queue_for_submitter_thread;
1225 if (rw == WRITE && req->private_bio && req->i.size
1226 && !test_bit(AL_SUSPENDED, &device->flags)) {
1227 if (!drbd_al_begin_io_fastpath(device, &req->i))
1228 goto queue_for_submitter_thread;
1229 req->rq_state |= RQ_IN_ACT_LOG;
1230 req->in_actlog_jif = jiffies;
1234 queue_for_submitter_thread:
1235 atomic_inc(&device->ap_actlog_cnt);
1236 drbd_queue_write(device, req);
1240 /* Require at least one path to current data.
1241 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1242 * We would not allow to read what was written,
1243 * we would not have bumped the data generation uuids,
1244 * we would cause data divergence for all the wrong reasons.
1246 * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1247 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1248 * and queues for retry later.
1250 static bool may_do_writes(struct drbd_device *device)
1252 const union drbd_dev_state s = device->state;
1253 return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1256 struct drbd_plug_cb {
1257 struct blk_plug_cb cb;
1258 struct drbd_request *most_recent_req;
1259 /* do we need more? */
1262 static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1264 struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1265 struct drbd_resource *resource = plug->cb.data;
1266 struct drbd_request *req = plug->most_recent_req;
1272 spin_lock_irq(&resource->req_lock);
1273 /* In case the sender did not process it yet, raise the flag to
1274 * have it followed with P_UNPLUG_REMOTE just after. */
1275 req->rq_state |= RQ_UNPLUG;
1276 /* but also queue a generic unplug */
1277 drbd_queue_unplug(req->device);
1278 kref_put(&req->kref, drbd_req_destroy);
1279 spin_unlock_irq(&resource->req_lock);
1282 static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1284 /* A lot of text to say
1285 * return (struct drbd_plug_cb*)blk_check_plugged(); */
1286 struct drbd_plug_cb *plug;
1287 struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1290 plug = container_of(cb, struct drbd_plug_cb, cb);
1296 static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1298 struct drbd_request *tmp = plug->most_recent_req;
1299 /* Will be sent to some peer.
1300 * Remember to tag it with UNPLUG_REMOTE on unplug */
1301 kref_get(&req->kref);
1302 plug->most_recent_req = req;
1304 kref_put(&tmp->kref, drbd_req_destroy);
1307 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1309 struct drbd_resource *resource = device->resource;
1310 const int rw = bio_data_dir(req->master_bio);
1311 struct bio_and_error m = { NULL, };
1312 bool no_remote = false;
1313 bool submit_private_bio = false;
1315 spin_lock_irq(&resource->req_lock);
1317 /* This may temporarily give up the req_lock,
1318 * but will re-aquire it before it returns here.
1319 * Needs to be before the check on drbd_suspended() */
1320 complete_conflicting_writes(req);
1321 /* no more giving up req_lock from now on! */
1323 /* check for congestion, and potentially stop sending
1324 * full data updates, but start sending "dirty bits" only. */
1325 maybe_pull_ahead(device);
1329 if (drbd_suspended(device)) {
1330 /* push back and retry: */
1331 req->rq_state |= RQ_POSTPONED;
1332 if (req->private_bio) {
1333 bio_put(req->private_bio);
1334 req->private_bio = NULL;
1340 /* We fail READ early, if we can not serve it.
1341 * We must do this before req is registered on any lists.
1342 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1344 if (!do_remote_read(req) && !req->private_bio)
1348 /* which transfer log epoch does this belong to? */
1349 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1351 /* no point in adding empty flushes to the transfer log,
1352 * they are mapped to drbd barriers already. */
1353 if (likely(req->i.size!=0)) {
1355 first_peer_device(device)->connection->current_tle_writes++;
1357 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1361 if (req->private_bio && !may_do_writes(device)) {
1362 bio_put(req->private_bio);
1363 req->private_bio = NULL;
1367 if (!drbd_process_write_request(req))
1370 /* We either have a private_bio, or we can read from remote.
1371 * Otherwise we had done the goto nodata above. */
1372 if (req->private_bio == NULL) {
1373 _req_mod(req, TO_BE_SENT);
1374 _req_mod(req, QUEUE_FOR_NET_READ);
1379 if (no_remote == false) {
1380 struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1382 drbd_update_plug(plug, req);
1385 /* If it took the fast path in drbd_request_prepare, add it here.
1386 * The slow path has added it already. */
1387 if (list_empty(&req->req_pending_master_completion))
1388 list_add_tail(&req->req_pending_master_completion,
1389 &device->pending_master_completion[rw == WRITE]);
1390 if (req->private_bio) {
1391 /* needs to be marked within the same spinlock */
1392 req->pre_submit_jif = jiffies;
1393 list_add_tail(&req->req_pending_local,
1394 &device->pending_completion[rw == WRITE]);
1395 _req_mod(req, TO_BE_SUBMITTED);
1396 /* but we need to give up the spinlock to submit */
1397 submit_private_bio = true;
1398 } else if (no_remote) {
1400 if (__ratelimit(&drbd_ratelimit_state))
1401 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1402 (unsigned long long)req->i.sector, req->i.size >> 9);
1403 /* A write may have been queued for send_oos, however.
1404 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1408 drbd_req_put_completion_ref(req, &m, 1);
1409 spin_unlock_irq(&resource->req_lock);
1411 /* Even though above is a kref_put(), this is safe.
1412 * As long as we still need to submit our private bio,
1413 * we hold a completion ref, and the request cannot disappear.
1414 * If however this request did not even have a private bio to submit
1415 * (e.g. remote read), req may already be invalid now.
1416 * That's why we cannot check on req->private_bio. */
1417 if (submit_private_bio)
1418 drbd_submit_req_private_bio(req);
1420 complete_master_bio(device, &m);
1423 void __drbd_make_request(struct drbd_device *device, struct bio *bio)
1425 struct drbd_request *req = drbd_request_prepare(device, bio);
1426 if (IS_ERR_OR_NULL(req))
1428 drbd_send_and_submit(device, req);
1431 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1433 struct blk_plug plug;
1434 struct drbd_request *req, *tmp;
1436 blk_start_plug(&plug);
1437 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1438 const int rw = bio_data_dir(req->master_bio);
1440 if (rw == WRITE /* rw != WRITE should not even end up here! */
1441 && req->private_bio && req->i.size
1442 && !test_bit(AL_SUSPENDED, &device->flags)) {
1443 if (!drbd_al_begin_io_fastpath(device, &req->i))
1446 req->rq_state |= RQ_IN_ACT_LOG;
1447 req->in_actlog_jif = jiffies;
1448 atomic_dec(&device->ap_actlog_cnt);
1451 list_del_init(&req->tl_requests);
1452 drbd_send_and_submit(device, req);
1454 blk_finish_plug(&plug);
1457 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1458 struct list_head *incoming,
1459 struct list_head *pending,
1460 struct list_head *later)
1462 struct drbd_request *req;
1466 spin_lock_irq(&device->al_lock);
1467 while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1468 err = drbd_al_begin_io_nonblock(device, &req->i);
1469 if (err == -ENOBUFS)
1474 list_move_tail(&req->tl_requests, later);
1476 list_move_tail(&req->tl_requests, pending);
1478 spin_unlock_irq(&device->al_lock);
1480 wake_up(&device->al_wait);
1481 return !list_empty(pending);
1484 static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1486 struct blk_plug plug;
1487 struct drbd_request *req;
1489 blk_start_plug(&plug);
1490 while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1491 req->rq_state |= RQ_IN_ACT_LOG;
1492 req->in_actlog_jif = jiffies;
1493 atomic_dec(&device->ap_actlog_cnt);
1494 list_del_init(&req->tl_requests);
1495 drbd_send_and_submit(device, req);
1497 blk_finish_plug(&plug);
1500 void do_submit(struct work_struct *ws)
1502 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1503 LIST_HEAD(incoming); /* from drbd_make_request() */
1504 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */
1505 LIST_HEAD(busy); /* blocked by resync requests */
1507 /* grab new incoming requests */
1508 spin_lock_irq(&device->resource->req_lock);
1509 list_splice_tail_init(&device->submit.writes, &incoming);
1510 spin_unlock_irq(&device->resource->req_lock);
1515 /* move used-to-be-busy back to front of incoming */
1516 list_splice_init(&busy, &incoming);
1517 submit_fast_path(device, &incoming);
1518 if (list_empty(&incoming))
1522 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1524 list_splice_init(&busy, &incoming);
1525 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1526 if (!list_empty(&pending))
1531 /* If all currently "hot" activity log extents are kept busy by
1532 * incoming requests, we still must not totally starve new
1533 * requests to "cold" extents.
1534 * Something left on &incoming means there had not been
1535 * enough update slots available, and the activity log
1536 * has been marked as "starving".
1538 * Try again now, without looking for new requests,
1539 * effectively blocking all new requests until we made
1540 * at least _some_ progress with what we currently have.
1542 if (!list_empty(&incoming))
1545 /* Nothing moved to pending, but nothing left
1546 * on incoming: all moved to busy!
1547 * Grab new and iterate. */
1548 spin_lock_irq(&device->resource->req_lock);
1549 list_splice_tail_init(&device->submit.writes, &incoming);
1550 spin_unlock_irq(&device->resource->req_lock);
1552 finish_wait(&device->al_wait, &wait);
1554 /* If the transaction was full, before all incoming requests
1555 * had been processed, skip ahead to commit, and iterate
1556 * without splicing in more incoming requests from upper layers.
1558 * Else, if all incoming have been processed,
1559 * they have become either "pending" (to be submitted after
1560 * next transaction commit) or "busy" (blocked by resync).
1562 * Maybe more was queued, while we prepared the transaction?
1563 * Try to stuff those into this transaction as well.
1564 * Be strictly non-blocking here,
1565 * we already have something to commit.
1567 * Commit if we don't make any more progres.
1570 while (list_empty(&incoming)) {
1571 LIST_HEAD(more_pending);
1572 LIST_HEAD(more_incoming);
1575 /* It is ok to look outside the lock,
1576 * it's only an optimization anyways */
1577 if (list_empty(&device->submit.writes))
1580 spin_lock_irq(&device->resource->req_lock);
1581 list_splice_tail_init(&device->submit.writes, &more_incoming);
1582 spin_unlock_irq(&device->resource->req_lock);
1584 if (list_empty(&more_incoming))
1587 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1589 list_splice_tail_init(&more_pending, &pending);
1590 list_splice_tail_init(&more_incoming, &incoming);
1595 drbd_al_begin_io_commit(device);
1596 send_and_submit_pending(device, &pending);
1600 blk_qc_t drbd_submit_bio(struct bio *bio)
1602 struct drbd_device *device = bio->bi_bdev->bd_disk->private_data;
1604 blk_queue_split(&bio);
1607 * what we "blindly" assume:
1609 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1612 __drbd_make_request(device, bio);
1613 return BLK_QC_T_NONE;
1616 static bool net_timeout_reached(struct drbd_request *net_req,
1617 struct drbd_connection *connection,
1618 unsigned long now, unsigned long ent,
1619 unsigned int ko_count, unsigned int timeout)
1621 struct drbd_device *device = net_req->device;
1623 if (!time_after(now, net_req->pre_send_jif + ent))
1626 if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1629 if (net_req->rq_state & RQ_NET_PENDING) {
1630 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1631 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1635 /* We received an ACK already (or are using protocol A),
1636 * but are waiting for the epoch closing barrier ack.
1637 * Check if we sent the barrier already. We should not blame the peer
1638 * for being unresponsive, if we did not even ask it yet. */
1639 if (net_req->epoch == connection->send.current_epoch_nr) {
1641 "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1642 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1646 /* Worst case: we may have been blocked for whatever reason, then
1647 * suddenly are able to send a lot of requests (and epoch separating
1648 * barriers) in quick succession.
1649 * The timestamp of the net_req may be much too old and not correspond
1650 * to the sending time of the relevant unack'ed barrier packet, so
1651 * would trigger a spurious timeout. The latest barrier packet may
1652 * have a too recent timestamp to trigger the timeout, potentially miss
1653 * a timeout. Right now we don't have a place to conveniently store
1655 * But in this particular situation, the application requests are still
1656 * completed to upper layers, DRBD should still "feel" responsive.
1657 * No need yet to kill this connection, it may still recover.
1658 * If not, eventually we will have queued enough into the network for
1659 * us to block. From that point of view, the timestamp of the last sent
1660 * barrier packet is relevant enough.
1662 if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1663 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",
1664 connection->send.last_sent_barrier_jif, now,
1665 jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1671 /* A request is considered timed out, if
1672 * - we have some effective timeout from the configuration,
1673 * with some state restrictions applied,
1674 * - the oldest request is waiting for a response from the network
1675 * resp. the local disk,
1676 * - the oldest request is in fact older than the effective timeout,
1677 * - the connection was established (resp. disk was attached)
1678 * for longer than the timeout already.
1679 * Note that for 32bit jiffies and very stable connections/disks,
1680 * we may have a wrap around, which is catched by
1681 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1683 * Side effect: once per 32bit wrap-around interval, which means every
1684 * ~198 days with 250 HZ, we have a window where the timeout would need
1685 * to expire twice (worst case) to become effective. Good enough.
1688 void request_timer_fn(struct timer_list *t)
1690 struct drbd_device *device = from_timer(device, t, request_timer);
1691 struct drbd_connection *connection = first_peer_device(device)->connection;
1692 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1693 struct net_conf *nc;
1694 unsigned long oldest_submit_jif;
1695 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1697 unsigned int ko_count = 0, timeout = 0;
1700 nc = rcu_dereference(connection->net_conf);
1701 if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1702 ko_count = nc->ko_count;
1703 timeout = nc->timeout;
1706 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1707 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1713 ent = timeout * HZ/10 * ko_count;
1714 et = min_not_zero(dt, ent);
1717 return; /* Recurring timer stopped */
1722 spin_lock_irq(&device->resource->req_lock);
1723 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1724 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1726 /* maybe the oldest request waiting for the peer is in fact still
1727 * blocking in tcp sendmsg. That's ok, though, that's handled via the
1728 * socket send timeout, requesting a ping, and bumping ko-count in
1729 * we_should_drop_the_connection().
1732 /* check the oldest request we did successfully sent,
1733 * but which is still waiting for an ACK. */
1734 req_peer = connection->req_ack_pending;
1736 /* if we don't have such request (e.g. protocoll A)
1737 * check the oldest requests which is still waiting on its epoch
1738 * closing barrier ack. */
1740 req_peer = connection->req_not_net_done;
1742 /* evaluate the oldest peer request only in one timer! */
1743 if (req_peer && req_peer->device != device)
1746 /* do we have something to evaluate? */
1747 if (req_peer == NULL && req_write == NULL && req_read == NULL)
1751 (req_write && req_read)
1752 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1753 ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1754 : req_write ? req_write->pre_submit_jif
1755 : req_read ? req_read->pre_submit_jif : now;
1757 if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1758 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1760 if (dt && oldest_submit_jif != now &&
1761 time_after(now, oldest_submit_jif + dt) &&
1762 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1763 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1764 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1767 /* Reschedule timer for the nearest not already expired timeout.
1768 * Fallback to now + min(effective network timeout, disk timeout). */
1769 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1770 ? req_peer->pre_send_jif + ent : now + et;
1771 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1772 ? oldest_submit_jif + dt : now + et;
1773 nt = time_before(ent, dt) ? ent : dt;
1775 spin_unlock_irq(&device->resource->req_lock);
1776 mod_timer(&device->request_timer, nt);