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 drbd_req_make_private_bio(req, bio_src);
34 req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
35 | (bio_op(bio_src) == REQ_OP_WRITE_SAME ? RQ_WSAME : 0)
36 | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0)
37 | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
39 req->master_bio = bio_src;
42 drbd_clear_interval(&req->i);
43 req->i.sector = bio_src->bi_iter.bi_sector;
44 req->i.size = bio_src->bi_iter.bi_size;
46 req->i.waiting = false;
48 INIT_LIST_HEAD(&req->tl_requests);
49 INIT_LIST_HEAD(&req->w.list);
50 INIT_LIST_HEAD(&req->req_pending_master_completion);
51 INIT_LIST_HEAD(&req->req_pending_local);
53 /* one reference to be put by __drbd_make_request */
54 atomic_set(&req->completion_ref, 1);
55 /* one kref as long as completion_ref > 0 */
56 kref_init(&req->kref);
60 static void drbd_remove_request_interval(struct rb_root *root,
61 struct drbd_request *req)
63 struct drbd_device *device = req->device;
64 struct drbd_interval *i = &req->i;
66 drbd_remove_interval(root, i);
68 /* Wake up any processes waiting for this request to complete. */
70 wake_up(&device->misc_wait);
73 void drbd_req_destroy(struct kref *kref)
75 struct drbd_request *req = container_of(kref, struct drbd_request, kref);
76 struct drbd_device *device = req->device;
77 const unsigned s = req->rq_state;
79 if ((req->master_bio && !(s & RQ_POSTPONED)) ||
80 atomic_read(&req->completion_ref) ||
81 (s & RQ_LOCAL_PENDING) ||
82 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
83 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
84 s, atomic_read(&req->completion_ref));
88 /* If called from mod_rq_state (expected normal case) or
89 * drbd_send_and_submit (the less likely normal path), this holds the
90 * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
91 * though it may be still empty (never added to the transfer log).
93 * If called from do_retry(), we do NOT hold the req_lock, but we are
94 * still allowed to unconditionally list_del(&req->tl_requests),
95 * because it will be on a local on-stack list only. */
96 list_del_init(&req->tl_requests);
98 /* finally remove the request from the conflict detection
99 * respective block_id verification interval tree. */
100 if (!drbd_interval_empty(&req->i)) {
101 struct rb_root *root;
104 root = &device->write_requests;
106 root = &device->read_requests;
107 drbd_remove_request_interval(root, req);
108 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
109 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
110 s, (unsigned long long)req->i.sector, req->i.size);
112 /* if it was a write, we may have to set the corresponding
113 * bit(s) out-of-sync first. If it had a local part, we need to
114 * release the reference to the activity log. */
116 /* Set out-of-sync unless both OK flags are set
117 * (local only or remote failed).
118 * Other places where we set out-of-sync:
119 * READ with local io-error */
121 /* There is a special case:
122 * we may notice late that IO was suspended,
123 * and postpone, or schedule for retry, a write,
124 * before it even was submitted or sent.
125 * In that case we do not want to touch the bitmap at all.
127 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
128 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
129 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
131 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
132 drbd_set_in_sync(device, req->i.sector, req->i.size);
135 /* one might be tempted to move the drbd_al_complete_io
136 * to the local io completion callback drbd_request_endio.
137 * but, if this was a mirror write, we may only
138 * drbd_al_complete_io after this is RQ_NET_DONE,
139 * otherwise the extent could be dropped from the al
140 * before it has actually been written on the peer.
141 * if we crash before our peer knows about the request,
142 * but after the extent has been dropped from the al,
143 * we would forget to resync the corresponding extent.
145 if (s & RQ_IN_ACT_LOG) {
146 if (get_ldev_if_state(device, D_FAILED)) {
147 drbd_al_complete_io(device, &req->i);
149 } else if (__ratelimit(&drbd_ratelimit_state)) {
150 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
151 "but my Disk seems to have failed :(\n",
152 (unsigned long long) req->i.sector, req->i.size);
157 mempool_free(req, &drbd_request_mempool);
160 static void wake_all_senders(struct drbd_connection *connection)
162 wake_up(&connection->sender_work.q_wait);
165 /* must hold resource->req_lock */
166 void start_new_tl_epoch(struct drbd_connection *connection)
168 /* no point closing an epoch, if it is empty, anyways. */
169 if (connection->current_tle_writes == 0)
172 connection->current_tle_writes = 0;
173 atomic_inc(&connection->current_tle_nr);
174 wake_all_senders(connection);
177 void complete_master_bio(struct drbd_device *device,
178 struct bio_and_error *m)
180 if (unlikely(m->error))
181 m->bio->bi_status = errno_to_blk_status(m->error);
187 /* Helper for __req_mod().
188 * Set m->bio to the master bio, if it is fit to be completed,
189 * or leave it alone (it is initialized to NULL in __req_mod),
190 * if it has already been completed, or cannot be completed yet.
191 * If m->bio is set, the error status to be returned is placed in m->error.
194 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
196 const unsigned s = req->rq_state;
197 struct drbd_device *device = req->device;
200 /* we must not complete the master bio, while it is
201 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
202 * not yet acknowledged by the peer
203 * not yet completed by the local io subsystem
204 * these flags may get cleared in any order by
207 * the bio_endio completion callbacks.
209 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
210 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
211 (s & RQ_COMPLETION_SUSP)) {
212 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
216 if (!req->master_bio) {
217 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
222 * figure out whether to report success or failure.
224 * report success when at least one of the operations succeeded.
225 * or, to put the other way,
226 * only report failure, when both operations failed.
228 * what to do about the failures is handled elsewhere.
229 * what we need to do here is just: complete the master_bio.
231 * local completion error, if any, has been stored as ERR_PTR
232 * in private_bio within drbd_request_endio.
234 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
235 error = PTR_ERR(req->private_bio);
237 /* Before we can signal completion to the upper layers,
238 * we may need to close the current transfer log epoch.
239 * We are within the request lock, so we can simply compare
240 * the request epoch number with the current transfer log
241 * epoch number. If they match, increase the current_tle_nr,
242 * and reset the transfer log epoch write_cnt.
244 if (op_is_write(bio_op(req->master_bio)) &&
245 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
246 start_new_tl_epoch(first_peer_device(device)->connection);
248 /* Update disk stats */
249 bio_end_io_acct(req->master_bio, req->start_jif);
252 * have it be pushed back to the retry work queue,
253 * so it will re-enter __drbd_make_request(),
254 * and be re-assigned to a suitable local or remote path,
255 * or failed if we do not have access to good data anymore.
257 * Unless it was failed early by __drbd_make_request(),
258 * because no path was available, in which case
259 * it was not even added to the transfer_log.
261 * read-ahead may fail, and will not be retried.
263 * WRITE should have used all available paths already.
266 bio_op(req->master_bio) == REQ_OP_READ &&
267 !(req->master_bio->bi_opf & REQ_RAHEAD) &&
268 !list_empty(&req->tl_requests))
269 req->rq_state |= RQ_POSTPONED;
271 if (!(req->rq_state & RQ_POSTPONED)) {
272 m->error = ok ? 0 : (error ?: -EIO);
273 m->bio = req->master_bio;
274 req->master_bio = NULL;
275 /* We leave it in the tree, to be able to verify later
276 * write-acks in protocol != C during resync.
277 * But we mark it as "complete", so it won't be counted as
278 * conflict in a multi-primary setup. */
279 req->i.completed = true;
283 wake_up(&device->misc_wait);
285 /* Either we are about to complete to upper layers,
286 * or we will restart this request.
287 * In either case, the request object will be destroyed soon,
288 * so better remove it from all lists. */
289 list_del_init(&req->req_pending_master_completion);
292 /* still holds resource->req_lock */
293 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
295 struct drbd_device *device = req->device;
296 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
301 if (!atomic_sub_and_test(put, &req->completion_ref))
304 drbd_req_complete(req, m);
306 /* local completion may still come in later,
307 * we need to keep the req object around. */
308 if (req->rq_state & RQ_LOCAL_ABORTED)
311 if (req->rq_state & RQ_POSTPONED) {
312 /* don't destroy the req object just yet,
313 * but queue it for retry */
314 drbd_restart_request(req);
318 kref_put(&req->kref, drbd_req_destroy);
321 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
323 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
326 if (connection->req_next == NULL)
327 connection->req_next = req;
330 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
332 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
335 if (connection->req_next != req)
337 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
338 const unsigned s = req->rq_state;
339 if (s & RQ_NET_QUEUED)
342 if (&req->tl_requests == &connection->transfer_log)
344 connection->req_next = req;
347 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
349 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
352 if (connection->req_ack_pending == NULL)
353 connection->req_ack_pending = req;
356 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
358 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
361 if (connection->req_ack_pending != req)
363 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
364 const unsigned s = req->rq_state;
365 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
368 if (&req->tl_requests == &connection->transfer_log)
370 connection->req_ack_pending = req;
373 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
375 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
378 if (connection->req_not_net_done == NULL)
379 connection->req_not_net_done = req;
382 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
384 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
387 if (connection->req_not_net_done != req)
389 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
390 const unsigned s = req->rq_state;
391 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
394 if (&req->tl_requests == &connection->transfer_log)
396 connection->req_not_net_done = req;
399 /* I'd like this to be the only place that manipulates
400 * req->completion_ref and req->kref. */
401 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
404 struct drbd_device *device = req->device;
405 struct drbd_peer_device *peer_device = first_peer_device(device);
406 unsigned s = req->rq_state;
409 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
410 set |= RQ_COMPLETION_SUSP;
414 req->rq_state &= ~clear;
415 req->rq_state |= set;
418 if (req->rq_state == s)
421 /* intent: get references */
423 kref_get(&req->kref);
425 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
426 atomic_inc(&req->completion_ref);
428 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
429 inc_ap_pending(device);
430 atomic_inc(&req->completion_ref);
433 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
434 atomic_inc(&req->completion_ref);
435 set_if_null_req_next(peer_device, req);
438 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
439 kref_get(&req->kref); /* wait for the DONE */
441 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
442 /* potentially already completed in the ack_receiver thread */
443 if (!(s & RQ_NET_DONE)) {
444 atomic_add(req->i.size >> 9, &device->ap_in_flight);
445 set_if_null_req_not_net_done(peer_device, req);
447 if (req->rq_state & RQ_NET_PENDING)
448 set_if_null_req_ack_pending(peer_device, req);
451 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
452 atomic_inc(&req->completion_ref);
454 /* progress: put references */
456 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
459 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
460 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
464 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
465 if (req->rq_state & RQ_LOCAL_ABORTED)
466 kref_put(&req->kref, drbd_req_destroy);
469 list_del_init(&req->req_pending_local);
472 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
473 dec_ap_pending(device);
475 req->acked_jif = jiffies;
476 advance_conn_req_ack_pending(peer_device, req);
479 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
481 advance_conn_req_next(peer_device, req);
484 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
486 atomic_sub(req->i.size >> 9, &device->ap_in_flight);
487 if (s & RQ_EXP_BARR_ACK)
488 kref_put(&req->kref, drbd_req_destroy);
489 req->net_done_jif = jiffies;
491 /* in ahead/behind mode, or just in case,
492 * before we finally destroy this request,
493 * the caching pointers must not reference it anymore */
494 advance_conn_req_next(peer_device, req);
495 advance_conn_req_ack_pending(peer_device, req);
496 advance_conn_req_not_net_done(peer_device, req);
499 /* potentially complete and destroy */
501 /* If we made progress, retry conflicting peer requests, if any. */
503 wake_up(&device->misc_wait);
505 drbd_req_put_completion_ref(req, m, c_put);
506 kref_put(&req->kref, drbd_req_destroy);
509 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
511 char b[BDEVNAME_SIZE];
513 if (!__ratelimit(&drbd_ratelimit_state))
516 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
517 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
518 (unsigned long long)req->i.sector,
520 bdevname(device->ldev->backing_bdev, b));
523 /* Helper for HANDED_OVER_TO_NETWORK.
524 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
525 * Is it also still "PENDING"?
526 * --> If so, clear PENDING and set NET_OK below.
527 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
528 * (and we must not set RQ_NET_OK) */
529 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
531 return (req->rq_state &
532 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
533 == (RQ_WRITE|RQ_NET_PENDING);
536 /* obviously this could be coded as many single functions
537 * instead of one huge switch,
538 * or by putting the code directly in the respective locations
539 * (as it has been before).
541 * but having it this way
542 * enforces that it is all in this one place, where it is easier to audit,
543 * it makes it obvious that whatever "event" "happens" to a request should
544 * happen "atomically" within the req_lock,
545 * and it enforces that we have to think in a very structured manner
546 * about the "events" that may happen to a request during its life time ...
548 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
549 struct bio_and_error *m)
551 struct drbd_device *const device = req->device;
552 struct drbd_peer_device *const peer_device = first_peer_device(device);
553 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
562 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
565 /* does not happen...
566 * initialization done in drbd_req_new
571 case TO_BE_SENT: /* via network */
572 /* reached via __drbd_make_request
573 * and from w_read_retry_remote */
574 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
576 nc = rcu_dereference(connection->net_conf);
577 p = nc->wire_protocol;
580 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
581 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
582 mod_rq_state(req, m, 0, RQ_NET_PENDING);
585 case TO_BE_SUBMITTED: /* locally */
586 /* reached via __drbd_make_request */
587 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
588 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
592 if (req->rq_state & RQ_WRITE)
593 device->writ_cnt += req->i.size >> 9;
595 device->read_cnt += req->i.size >> 9;
597 mod_rq_state(req, m, RQ_LOCAL_PENDING,
598 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
602 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
605 case WRITE_COMPLETED_WITH_ERROR:
606 drbd_report_io_error(device, req);
607 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
608 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
611 case READ_COMPLETED_WITH_ERROR:
612 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
613 drbd_report_io_error(device, req);
614 __drbd_chk_io_error(device, DRBD_READ_ERROR);
616 case READ_AHEAD_COMPLETED_WITH_ERROR:
617 /* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
618 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
621 case DISCARD_COMPLETED_NOTSUPP:
622 case DISCARD_COMPLETED_WITH_ERROR:
623 /* I'd rather not detach from local disk just because it
624 * failed a REQ_OP_DISCARD. */
625 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
628 case QUEUE_FOR_NET_READ:
631 * or target area marked as invalid,
632 * or just got an io-error. */
633 /* from __drbd_make_request
634 * or from bio_endio during read io-error recovery */
636 /* So we can verify the handle in the answer packet.
637 * Corresponding drbd_remove_request_interval is in
638 * drbd_req_complete() */
639 D_ASSERT(device, drbd_interval_empty(&req->i));
640 drbd_insert_interval(&device->read_requests, &req->i);
642 set_bit(UNPLUG_REMOTE, &device->flags);
644 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
645 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
646 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
647 req->w.cb = w_send_read_req;
648 drbd_queue_work(&connection->sender_work,
652 case QUEUE_FOR_NET_WRITE:
653 /* assert something? */
654 /* from __drbd_make_request only */
656 /* Corresponding drbd_remove_request_interval is in
657 * drbd_req_complete() */
658 D_ASSERT(device, drbd_interval_empty(&req->i));
659 drbd_insert_interval(&device->write_requests, &req->i);
662 * In case the req ended up on the transfer log before being
663 * queued on the worker, it could lead to this request being
664 * missed during cleanup after connection loss.
665 * So we have to do both operations here,
666 * within the same lock that protects the transfer log.
668 * _req_add_to_epoch(req); this has to be after the
669 * _maybe_start_new_epoch(req); which happened in
670 * __drbd_make_request, because we now may set the bit
671 * again ourselves to close the current epoch.
673 * Add req to the (now) current epoch (barrier). */
675 /* otherwise we may lose an unplug, which may cause some remote
676 * io-scheduler timeout to expire, increasing maximum latency,
677 * hurting performance. */
678 set_bit(UNPLUG_REMOTE, &device->flags);
680 /* queue work item to send data */
681 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
682 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
683 req->w.cb = w_send_dblock;
684 drbd_queue_work(&connection->sender_work,
687 /* close the epoch, in case it outgrew the limit */
689 nc = rcu_dereference(connection->net_conf);
690 p = nc->max_epoch_size;
692 if (connection->current_tle_writes >= p)
693 start_new_tl_epoch(connection);
697 case QUEUE_FOR_SEND_OOS:
698 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
699 req->w.cb = w_send_out_of_sync;
700 drbd_queue_work(&connection->sender_work,
704 case READ_RETRY_REMOTE_CANCELED:
707 /* real cleanup will be done from tl_clear. just update flags
708 * so it is no longer marked as on the worker queue */
709 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
712 case HANDED_OVER_TO_NETWORK:
713 /* assert something? */
714 if (is_pending_write_protocol_A(req))
715 /* this is what is dangerous about protocol A:
716 * pretend it was successfully written on the peer. */
717 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
718 RQ_NET_SENT|RQ_NET_OK);
720 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
721 /* It is still not yet RQ_NET_DONE until the
722 * corresponding epoch barrier got acked as well,
723 * so we know what to dirty on connection loss. */
726 case OOS_HANDED_TO_NETWORK:
727 /* Was not set PENDING, no longer QUEUED, so is now DONE
728 * as far as this connection is concerned. */
729 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
732 case CONNECTION_LOST_WHILE_PENDING:
733 /* transfer log cleanup after connection loss */
735 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
739 case CONFLICT_RESOLVED:
740 /* for superseded conflicting writes of multiple primaries,
741 * there is no need to keep anything in the tl, potential
742 * node crashes are covered by the activity log.
744 * If this request had been marked as RQ_POSTPONED before,
745 * it will actually not be completed, but "restarted",
746 * resubmitted from the retry worker context. */
747 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
748 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
749 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
752 case WRITE_ACKED_BY_PEER_AND_SIS:
753 req->rq_state |= RQ_NET_SIS;
754 case WRITE_ACKED_BY_PEER:
755 /* Normal operation protocol C: successfully written on peer.
756 * During resync, even in protocol != C,
757 * we requested an explicit write ack anyways.
758 * Which means we cannot even assert anything here.
759 * Nothing more to do here.
760 * We want to keep the tl in place for all protocols, to cater
761 * for volatile write-back caches on lower level devices. */
763 case RECV_ACKED_BY_PEER:
764 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
765 /* protocol B; pretends to be successfully written on peer.
766 * see also notes above in HANDED_OVER_TO_NETWORK about
769 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
773 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
774 /* If this node has already detected the write conflict, the
775 * worker will be waiting on misc_wait. Wake it up once this
776 * request has completed locally.
778 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
779 req->rq_state |= RQ_POSTPONED;
781 wake_up(&device->misc_wait);
782 /* Do not clear RQ_NET_PENDING. This request will make further
783 * progress via restart_conflicting_writes() or
784 * fail_postponed_requests(). Hopefully. */
788 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
791 case FAIL_FROZEN_DISK_IO:
792 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
794 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
797 case RESTART_FROZEN_DISK_IO:
798 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
802 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
806 if (bio_data_dir(req->master_bio) == WRITE)
809 get_ldev(device); /* always succeeds in this call path */
810 req->w.cb = w_restart_disk_io;
811 drbd_queue_work(&connection->sender_work,
816 /* Simply complete (local only) READs. */
817 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
818 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
822 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
823 before the connection loss (B&C only); only P_BARRIER_ACK
824 (or the local completion?) was missing when we suspended.
825 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
826 During connection handshake, we ensure that the peer was not rebooted. */
827 if (!(req->rq_state & RQ_NET_OK)) {
828 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
829 * in that case we must not set RQ_NET_PENDING. */
831 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
833 /* w.cb expected to be w_send_dblock, or w_send_read_req */
834 drbd_queue_work(&connection->sender_work,
836 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
837 } /* else: FIXME can this happen? */
840 fallthrough; /* to BARRIER_ACKED */
843 /* barrier ack for READ requests does not make sense */
844 if (!(req->rq_state & RQ_WRITE))
847 if (req->rq_state & RQ_NET_PENDING) {
848 /* barrier came in before all requests were acked.
849 * this is bad, because if the connection is lost now,
850 * we won't be able to clean them up... */
851 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
853 /* Allowed to complete requests, even while suspended.
854 * As this is called for all requests within a matching epoch,
855 * we need to filter, and only set RQ_NET_DONE for those that
856 * have actually been on the wire. */
857 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
858 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
862 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
863 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
866 case QUEUE_AS_DRBD_BARRIER:
867 start_new_tl_epoch(connection);
868 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
875 /* we may do a local read if:
876 * - we are consistent (of course),
877 * - or we are generally inconsistent,
878 * BUT we are still/already IN SYNC for this area.
879 * since size may be bigger than BM_BLOCK_SIZE,
880 * we may need to check several bits.
882 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
884 unsigned long sbnr, ebnr;
885 sector_t esector, nr_sectors;
887 if (device->state.disk == D_UP_TO_DATE)
889 if (device->state.disk != D_INCONSISTENT)
891 esector = sector + (size >> 9) - 1;
892 nr_sectors = get_capacity(device->vdisk);
893 D_ASSERT(device, sector < nr_sectors);
894 D_ASSERT(device, esector < nr_sectors);
896 sbnr = BM_SECT_TO_BIT(sector);
897 ebnr = BM_SECT_TO_BIT(esector);
899 return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
902 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
903 enum drbd_read_balancing rbm)
905 struct backing_dev_info *bdi;
909 case RB_CONGESTED_REMOTE:
910 bdi = device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
911 return bdi_read_congested(bdi);
912 case RB_LEAST_PENDING:
913 return atomic_read(&device->local_cnt) >
914 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
915 case RB_32K_STRIPING: /* stripe_shift = 15 */
916 case RB_64K_STRIPING:
917 case RB_128K_STRIPING:
918 case RB_256K_STRIPING:
919 case RB_512K_STRIPING:
920 case RB_1M_STRIPING: /* stripe_shift = 20 */
921 stripe_shift = (rbm - RB_32K_STRIPING + 15);
922 return (sector >> (stripe_shift - 9)) & 1;
924 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
925 case RB_PREFER_REMOTE:
927 case RB_PREFER_LOCAL:
934 * complete_conflicting_writes - wait for any conflicting write requests
936 * The write_requests tree contains all active write requests which we
937 * currently know about. Wait for any requests to complete which conflict with
940 * Only way out: remove the conflicting intervals from the tree.
942 static void complete_conflicting_writes(struct drbd_request *req)
945 struct drbd_device *device = req->device;
946 struct drbd_interval *i;
947 sector_t sector = req->i.sector;
948 int size = req->i.size;
951 drbd_for_each_overlap(i, &device->write_requests, sector, size) {
952 /* Ignore, if already completed to upper layers. */
955 /* Handle the first found overlap. After the schedule
956 * we have to restart the tree walk. */
962 /* Indicate to wake up device->misc_wait on progress. */
963 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
965 spin_unlock_irq(&device->resource->req_lock);
967 spin_lock_irq(&device->resource->req_lock);
969 finish_wait(&device->misc_wait, &wait);
972 /* called within req_lock */
973 static void maybe_pull_ahead(struct drbd_device *device)
975 struct drbd_connection *connection = first_peer_device(device)->connection;
977 bool congested = false;
978 enum drbd_on_congestion on_congestion;
981 nc = rcu_dereference(connection->net_conf);
982 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
984 if (on_congestion == OC_BLOCK ||
985 connection->agreed_pro_version < 96)
988 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
989 return; /* nothing to do ... */
991 /* If I don't even have good local storage, we can not reasonably try
992 * to pull ahead of the peer. We also need the local reference to make
993 * sure device->act_log is there.
995 if (!get_ldev_if_state(device, D_UP_TO_DATE))
999 atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1000 drbd_info(device, "Congestion-fill threshold reached\n");
1004 if (device->act_log->used >= nc->cong_extents) {
1005 drbd_info(device, "Congestion-extents threshold reached\n");
1010 /* start a new epoch for non-mirrored writes */
1011 start_new_tl_epoch(first_peer_device(device)->connection);
1013 if (on_congestion == OC_PULL_AHEAD)
1014 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1015 else /*nc->on_congestion == OC_DISCONNECT */
1016 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1021 /* If this returns false, and req->private_bio is still set,
1022 * this should be submitted locally.
1024 * If it returns false, but req->private_bio is not set,
1025 * we do not have access to good data :(
1027 * Otherwise, this destroys req->private_bio, if any,
1030 static bool do_remote_read(struct drbd_request *req)
1032 struct drbd_device *device = req->device;
1033 enum drbd_read_balancing rbm;
1035 if (req->private_bio) {
1036 if (!drbd_may_do_local_read(device,
1037 req->i.sector, req->i.size)) {
1038 bio_put(req->private_bio);
1039 req->private_bio = NULL;
1044 if (device->state.pdsk != D_UP_TO_DATE)
1047 if (req->private_bio == NULL)
1050 /* TODO: improve read balancing decisions, take into account drbd
1051 * protocol, pending requests etc. */
1054 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1057 if (rbm == RB_PREFER_LOCAL && req->private_bio)
1058 return false; /* submit locally */
1060 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1061 if (req->private_bio) {
1062 bio_put(req->private_bio);
1063 req->private_bio = NULL;
1072 bool drbd_should_do_remote(union drbd_dev_state s)
1074 return s.pdsk == D_UP_TO_DATE ||
1075 (s.pdsk >= D_INCONSISTENT &&
1076 s.conn >= C_WF_BITMAP_T &&
1078 /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1079 That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1083 static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1085 return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1086 /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1087 since we enter state C_AHEAD only if proto >= 96 */
1090 /* returns number of connections (== 1, for drbd 8.4)
1091 * expected to actually write this data,
1092 * which does NOT include those that we are L_AHEAD for. */
1093 static int drbd_process_write_request(struct drbd_request *req)
1095 struct drbd_device *device = req->device;
1096 int remote, send_oos;
1098 remote = drbd_should_do_remote(device->state);
1099 send_oos = drbd_should_send_out_of_sync(device->state);
1101 /* Need to replicate writes. Unless it is an empty flush,
1102 * which is better mapped to a DRBD P_BARRIER packet,
1103 * also for drbd wire protocol compatibility reasons.
1104 * If this was a flush, just start a new epoch.
1105 * Unless the current epoch was empty anyways, or we are not currently
1106 * replicating, in which case there is no point. */
1107 if (unlikely(req->i.size == 0)) {
1108 /* The only size==0 bios we expect are empty flushes. */
1109 D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1111 _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1115 if (!remote && !send_oos)
1118 D_ASSERT(device, !(remote && send_oos));
1121 _req_mod(req, TO_BE_SENT);
1122 _req_mod(req, QUEUE_FOR_NET_WRITE);
1123 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1124 _req_mod(req, QUEUE_FOR_SEND_OOS);
1129 static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1131 int err = drbd_issue_discard_or_zero_out(req->device,
1132 req->i.sector, req->i.size >> 9, flags);
1134 req->private_bio->bi_status = BLK_STS_IOERR;
1135 bio_endio(req->private_bio);
1139 drbd_submit_req_private_bio(struct drbd_request *req)
1141 struct drbd_device *device = req->device;
1142 struct bio *bio = req->private_bio;
1145 if (bio_op(bio) != REQ_OP_READ)
1146 type = DRBD_FAULT_DT_WR;
1147 else if (bio->bi_opf & REQ_RAHEAD)
1148 type = DRBD_FAULT_DT_RA;
1150 type = DRBD_FAULT_DT_RD;
1152 bio_set_dev(bio, device->ldev->backing_bdev);
1154 /* State may have changed since we grabbed our reference on the
1155 * ->ldev member. Double check, and short-circuit to endio.
1156 * In case the last activity log transaction failed to get on
1157 * stable storage, and this is a WRITE, we may not even submit
1159 if (get_ldev(device)) {
1160 if (drbd_insert_fault(device, type))
1162 else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1163 drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1164 ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1165 else if (bio_op(bio) == REQ_OP_DISCARD)
1166 drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1168 submit_bio_noacct(bio);
1174 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1176 spin_lock_irq(&device->resource->req_lock);
1177 list_add_tail(&req->tl_requests, &device->submit.writes);
1178 list_add_tail(&req->req_pending_master_completion,
1179 &device->pending_master_completion[1 /* WRITE */]);
1180 spin_unlock_irq(&device->resource->req_lock);
1181 queue_work(device->submit.wq, &device->submit.worker);
1182 /* do_submit() may sleep internally on al_wait, too */
1183 wake_up(&device->al_wait);
1186 /* returns the new drbd_request pointer, if the caller is expected to
1187 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1188 * request on the submitter thread.
1189 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1191 static struct drbd_request *
1192 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1194 const int rw = bio_data_dir(bio);
1195 struct drbd_request *req;
1197 /* allocate outside of all locks; */
1198 req = drbd_req_new(device, bio);
1201 /* only pass the error to the upper layers.
1202 * if user cannot handle io errors, that's not our business. */
1203 drbd_err(device, "could not kmalloc() req\n");
1204 bio->bi_status = BLK_STS_RESOURCE;
1206 return ERR_PTR(-ENOMEM);
1209 /* Update disk stats */
1210 req->start_jif = bio_start_io_acct(req->master_bio);
1212 if (!get_ldev(device)) {
1213 bio_put(req->private_bio);
1214 req->private_bio = NULL;
1217 /* process discards always from our submitter thread */
1218 if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1219 bio_op(bio) == REQ_OP_DISCARD)
1220 goto queue_for_submitter_thread;
1222 if (rw == WRITE && req->private_bio && req->i.size
1223 && !test_bit(AL_SUSPENDED, &device->flags)) {
1224 if (!drbd_al_begin_io_fastpath(device, &req->i))
1225 goto queue_for_submitter_thread;
1226 req->rq_state |= RQ_IN_ACT_LOG;
1227 req->in_actlog_jif = jiffies;
1231 queue_for_submitter_thread:
1232 atomic_inc(&device->ap_actlog_cnt);
1233 drbd_queue_write(device, req);
1237 /* Require at least one path to current data.
1238 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1239 * We would not allow to read what was written,
1240 * we would not have bumped the data generation uuids,
1241 * we would cause data divergence for all the wrong reasons.
1243 * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1244 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1245 * and queues for retry later.
1247 static bool may_do_writes(struct drbd_device *device)
1249 const union drbd_dev_state s = device->state;
1250 return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1253 struct drbd_plug_cb {
1254 struct blk_plug_cb cb;
1255 struct drbd_request *most_recent_req;
1256 /* do we need more? */
1259 static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1261 struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1262 struct drbd_resource *resource = plug->cb.data;
1263 struct drbd_request *req = plug->most_recent_req;
1269 spin_lock_irq(&resource->req_lock);
1270 /* In case the sender did not process it yet, raise the flag to
1271 * have it followed with P_UNPLUG_REMOTE just after. */
1272 req->rq_state |= RQ_UNPLUG;
1273 /* but also queue a generic unplug */
1274 drbd_queue_unplug(req->device);
1275 kref_put(&req->kref, drbd_req_destroy);
1276 spin_unlock_irq(&resource->req_lock);
1279 static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1281 /* A lot of text to say
1282 * return (struct drbd_plug_cb*)blk_check_plugged(); */
1283 struct drbd_plug_cb *plug;
1284 struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1287 plug = container_of(cb, struct drbd_plug_cb, cb);
1293 static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1295 struct drbd_request *tmp = plug->most_recent_req;
1296 /* Will be sent to some peer.
1297 * Remember to tag it with UNPLUG_REMOTE on unplug */
1298 kref_get(&req->kref);
1299 plug->most_recent_req = req;
1301 kref_put(&tmp->kref, drbd_req_destroy);
1304 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1306 struct drbd_resource *resource = device->resource;
1307 const int rw = bio_data_dir(req->master_bio);
1308 struct bio_and_error m = { NULL, };
1309 bool no_remote = false;
1310 bool submit_private_bio = false;
1312 spin_lock_irq(&resource->req_lock);
1314 /* This may temporarily give up the req_lock,
1315 * but will re-aquire it before it returns here.
1316 * Needs to be before the check on drbd_suspended() */
1317 complete_conflicting_writes(req);
1318 /* no more giving up req_lock from now on! */
1320 /* check for congestion, and potentially stop sending
1321 * full data updates, but start sending "dirty bits" only. */
1322 maybe_pull_ahead(device);
1326 if (drbd_suspended(device)) {
1327 /* push back and retry: */
1328 req->rq_state |= RQ_POSTPONED;
1329 if (req->private_bio) {
1330 bio_put(req->private_bio);
1331 req->private_bio = NULL;
1337 /* We fail READ early, if we can not serve it.
1338 * We must do this before req is registered on any lists.
1339 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1341 if (!do_remote_read(req) && !req->private_bio)
1345 /* which transfer log epoch does this belong to? */
1346 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1348 /* no point in adding empty flushes to the transfer log,
1349 * they are mapped to drbd barriers already. */
1350 if (likely(req->i.size!=0)) {
1352 first_peer_device(device)->connection->current_tle_writes++;
1354 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1358 if (req->private_bio && !may_do_writes(device)) {
1359 bio_put(req->private_bio);
1360 req->private_bio = NULL;
1364 if (!drbd_process_write_request(req))
1367 /* We either have a private_bio, or we can read from remote.
1368 * Otherwise we had done the goto nodata above. */
1369 if (req->private_bio == NULL) {
1370 _req_mod(req, TO_BE_SENT);
1371 _req_mod(req, QUEUE_FOR_NET_READ);
1376 if (no_remote == false) {
1377 struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1379 drbd_update_plug(plug, req);
1382 /* If it took the fast path in drbd_request_prepare, add it here.
1383 * The slow path has added it already. */
1384 if (list_empty(&req->req_pending_master_completion))
1385 list_add_tail(&req->req_pending_master_completion,
1386 &device->pending_master_completion[rw == WRITE]);
1387 if (req->private_bio) {
1388 /* needs to be marked within the same spinlock */
1389 req->pre_submit_jif = jiffies;
1390 list_add_tail(&req->req_pending_local,
1391 &device->pending_completion[rw == WRITE]);
1392 _req_mod(req, TO_BE_SUBMITTED);
1393 /* but we need to give up the spinlock to submit */
1394 submit_private_bio = true;
1395 } else if (no_remote) {
1397 if (__ratelimit(&drbd_ratelimit_state))
1398 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1399 (unsigned long long)req->i.sector, req->i.size >> 9);
1400 /* A write may have been queued for send_oos, however.
1401 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1405 drbd_req_put_completion_ref(req, &m, 1);
1406 spin_unlock_irq(&resource->req_lock);
1408 /* Even though above is a kref_put(), this is safe.
1409 * As long as we still need to submit our private bio,
1410 * we hold a completion ref, and the request cannot disappear.
1411 * If however this request did not even have a private bio to submit
1412 * (e.g. remote read), req may already be invalid now.
1413 * That's why we cannot check on req->private_bio. */
1414 if (submit_private_bio)
1415 drbd_submit_req_private_bio(req);
1417 complete_master_bio(device, &m);
1420 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1422 struct drbd_request *req = drbd_request_prepare(device, bio, start_jif);
1423 if (IS_ERR_OR_NULL(req))
1425 drbd_send_and_submit(device, req);
1428 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1430 struct blk_plug plug;
1431 struct drbd_request *req, *tmp;
1433 blk_start_plug(&plug);
1434 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1435 const int rw = bio_data_dir(req->master_bio);
1437 if (rw == WRITE /* rw != WRITE should not even end up here! */
1438 && req->private_bio && req->i.size
1439 && !test_bit(AL_SUSPENDED, &device->flags)) {
1440 if (!drbd_al_begin_io_fastpath(device, &req->i))
1443 req->rq_state |= RQ_IN_ACT_LOG;
1444 req->in_actlog_jif = jiffies;
1445 atomic_dec(&device->ap_actlog_cnt);
1448 list_del_init(&req->tl_requests);
1449 drbd_send_and_submit(device, req);
1451 blk_finish_plug(&plug);
1454 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1455 struct list_head *incoming,
1456 struct list_head *pending,
1457 struct list_head *later)
1459 struct drbd_request *req;
1463 spin_lock_irq(&device->al_lock);
1464 while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1465 err = drbd_al_begin_io_nonblock(device, &req->i);
1466 if (err == -ENOBUFS)
1471 list_move_tail(&req->tl_requests, later);
1473 list_move_tail(&req->tl_requests, pending);
1475 spin_unlock_irq(&device->al_lock);
1477 wake_up(&device->al_wait);
1478 return !list_empty(pending);
1481 static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1483 struct blk_plug plug;
1484 struct drbd_request *req;
1486 blk_start_plug(&plug);
1487 while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1488 req->rq_state |= RQ_IN_ACT_LOG;
1489 req->in_actlog_jif = jiffies;
1490 atomic_dec(&device->ap_actlog_cnt);
1491 list_del_init(&req->tl_requests);
1492 drbd_send_and_submit(device, req);
1494 blk_finish_plug(&plug);
1497 void do_submit(struct work_struct *ws)
1499 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1500 LIST_HEAD(incoming); /* from drbd_make_request() */
1501 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */
1502 LIST_HEAD(busy); /* blocked by resync requests */
1504 /* grab new incoming requests */
1505 spin_lock_irq(&device->resource->req_lock);
1506 list_splice_tail_init(&device->submit.writes, &incoming);
1507 spin_unlock_irq(&device->resource->req_lock);
1512 /* move used-to-be-busy back to front of incoming */
1513 list_splice_init(&busy, &incoming);
1514 submit_fast_path(device, &incoming);
1515 if (list_empty(&incoming))
1519 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1521 list_splice_init(&busy, &incoming);
1522 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1523 if (!list_empty(&pending))
1528 /* If all currently "hot" activity log extents are kept busy by
1529 * incoming requests, we still must not totally starve new
1530 * requests to "cold" extents.
1531 * Something left on &incoming means there had not been
1532 * enough update slots available, and the activity log
1533 * has been marked as "starving".
1535 * Try again now, without looking for new requests,
1536 * effectively blocking all new requests until we made
1537 * at least _some_ progress with what we currently have.
1539 if (!list_empty(&incoming))
1542 /* Nothing moved to pending, but nothing left
1543 * on incoming: all moved to busy!
1544 * Grab new and iterate. */
1545 spin_lock_irq(&device->resource->req_lock);
1546 list_splice_tail_init(&device->submit.writes, &incoming);
1547 spin_unlock_irq(&device->resource->req_lock);
1549 finish_wait(&device->al_wait, &wait);
1551 /* If the transaction was full, before all incoming requests
1552 * had been processed, skip ahead to commit, and iterate
1553 * without splicing in more incoming requests from upper layers.
1555 * Else, if all incoming have been processed,
1556 * they have become either "pending" (to be submitted after
1557 * next transaction commit) or "busy" (blocked by resync).
1559 * Maybe more was queued, while we prepared the transaction?
1560 * Try to stuff those into this transaction as well.
1561 * Be strictly non-blocking here,
1562 * we already have something to commit.
1564 * Commit if we don't make any more progres.
1567 while (list_empty(&incoming)) {
1568 LIST_HEAD(more_pending);
1569 LIST_HEAD(more_incoming);
1572 /* It is ok to look outside the lock,
1573 * it's only an optimization anyways */
1574 if (list_empty(&device->submit.writes))
1577 spin_lock_irq(&device->resource->req_lock);
1578 list_splice_tail_init(&device->submit.writes, &more_incoming);
1579 spin_unlock_irq(&device->resource->req_lock);
1581 if (list_empty(&more_incoming))
1584 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1586 list_splice_tail_init(&more_pending, &pending);
1587 list_splice_tail_init(&more_incoming, &incoming);
1592 drbd_al_begin_io_commit(device);
1593 send_and_submit_pending(device, &pending);
1597 blk_qc_t drbd_submit_bio(struct bio *bio)
1599 struct drbd_device *device = bio->bi_disk->private_data;
1600 unsigned long start_jif;
1602 blk_queue_split(&bio);
1604 start_jif = jiffies;
1607 * what we "blindly" assume:
1609 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1612 __drbd_make_request(device, bio, start_jif);
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);