4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 drbd is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
15 drbd is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with drbd; see the file COPYING. If not, write to
22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
26 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/drbd.h>
34 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size);
36 /* Update disk stats at start of I/O request */
37 static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
39 const int rw = bio_data_dir(bio);
41 cpu = part_stat_lock();
42 part_round_stats(cpu, &mdev->vdisk->part0);
43 part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
44 part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
45 (void) cpu; /* The macro invocations above want the cpu argument, I do not like
46 the compiler warning about cpu only assigned but never used... */
47 part_inc_in_flight(&mdev->vdisk->part0, rw);
51 /* Update disk stats when completing request upwards */
52 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
54 int rw = bio_data_dir(req->master_bio);
55 unsigned long duration = jiffies - req->start_time;
57 cpu = part_stat_lock();
58 part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
59 part_round_stats(cpu, &mdev->vdisk->part0);
60 part_dec_in_flight(&mdev->vdisk->part0, rw);
64 static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
67 struct drbd_request *req;
69 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
73 drbd_req_make_private_bio(req, bio_src);
74 req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
76 req->master_bio = bio_src;
79 drbd_clear_interval(&req->i);
80 req->i.sector = bio_src->bi_sector;
81 req->i.size = bio_src->bi_size;
83 req->i.waiting = false;
85 INIT_LIST_HEAD(&req->tl_requests);
86 INIT_LIST_HEAD(&req->w.list);
91 static void drbd_req_free(struct drbd_request *req)
93 mempool_free(req, drbd_request_mempool);
96 /* rw is bio_data_dir(), only READ or WRITE */
97 static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
99 const unsigned long s = req->rq_state;
101 /* remove it from the transfer log.
102 * well, only if it had been there in the first
103 * place... if it had not (local only or conflicting
104 * and never sent), it should still be "empty" as
105 * initialized in drbd_req_new(), so we can list_del() it
106 * here unconditionally */
107 list_del(&req->tl_requests);
109 /* if it was a write, we may have to set the corresponding
110 * bit(s) out-of-sync first. If it had a local part, we need to
111 * release the reference to the activity log. */
113 /* Set out-of-sync unless both OK flags are set
114 * (local only or remote failed).
115 * Other places where we set out-of-sync:
116 * READ with local io-error */
117 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
118 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
120 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
121 drbd_set_in_sync(mdev, req->i.sector, req->i.size);
123 /* one might be tempted to move the drbd_al_complete_io
124 * to the local io completion callback drbd_request_endio.
125 * but, if this was a mirror write, we may only
126 * drbd_al_complete_io after this is RQ_NET_DONE,
127 * otherwise the extent could be dropped from the al
128 * before it has actually been written on the peer.
129 * if we crash before our peer knows about the request,
130 * but after the extent has been dropped from the al,
131 * we would forget to resync the corresponding extent.
133 if (s & RQ_LOCAL_MASK) {
134 if (get_ldev_if_state(mdev, D_FAILED)) {
135 if (s & RQ_IN_ACT_LOG)
136 drbd_al_complete_io(mdev, &req->i);
138 } else if (__ratelimit(&drbd_ratelimit_state)) {
139 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
140 "but my Disk seems to have failed :(\n",
141 (unsigned long long) req->i.sector, req->i.size);
149 static void queue_barrier(struct drbd_conf *mdev)
151 struct drbd_tl_epoch *b;
152 struct drbd_tconn *tconn = mdev->tconn;
154 /* We are within the req_lock. Once we queued the barrier for sending,
155 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
156 * barrier/epoch object is added. This is the only place this bit is
157 * set. It indicates that the barrier for this epoch is already queued,
158 * and no new epoch has been created yet. */
159 if (test_bit(CREATE_BARRIER, &tconn->flags))
162 b = tconn->newest_tle;
163 b->w.cb = w_send_barrier;
165 /* inc_ap_pending done here, so we won't
166 * get imbalanced on connection loss.
167 * dec_ap_pending will be done in got_BarrierAck
168 * or (on connection loss) in tl_clear. */
169 inc_ap_pending(mdev);
170 drbd_queue_work(&tconn->data.work, &b->w);
171 set_bit(CREATE_BARRIER, &tconn->flags);
174 static void _about_to_complete_local_write(struct drbd_conf *mdev,
175 struct drbd_request *req)
177 const unsigned long s = req->rq_state;
179 /* Before we can signal completion to the upper layers,
180 * we may need to close the current epoch.
181 * We can skip this, if this request has not even been sent, because we
182 * did not have a fully established connection yet/anymore, during
183 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
185 if (mdev->state.conn >= C_CONNECTED &&
186 (s & RQ_NET_SENT) != 0 &&
187 req->epoch == mdev->tconn->newest_tle->br_number)
191 void complete_master_bio(struct drbd_conf *mdev,
192 struct bio_and_error *m)
194 bio_endio(m->bio, m->error);
199 static void drbd_remove_request_interval(struct rb_root *root,
200 struct drbd_request *req)
202 struct drbd_conf *mdev = req->w.mdev;
203 struct drbd_interval *i = &req->i;
205 drbd_remove_interval(root, i);
207 /* Wake up any processes waiting for this request to complete. */
209 wake_up(&mdev->misc_wait);
212 /* Helper for __req_mod().
213 * Set m->bio to the master bio, if it is fit to be completed,
214 * or leave it alone (it is initialized to NULL in __req_mod),
215 * if it has already been completed, or cannot be completed yet.
216 * If m->bio is set, the error status to be returned is placed in m->error.
218 void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
220 const unsigned long s = req->rq_state;
221 struct drbd_conf *mdev = req->w.mdev;
222 int rw = req->rq_state & RQ_WRITE ? WRITE : READ;
224 /* we must not complete the master bio, while it is
225 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
226 * not yet acknowledged by the peer
227 * not yet completed by the local io subsystem
228 * these flags may get cleared in any order by
231 * the bio_endio completion callbacks.
233 if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
235 if (req->i.waiting) {
236 /* Retry all conflicting peer requests. */
237 wake_up(&mdev->misc_wait);
239 if (s & RQ_NET_QUEUED)
241 if (s & RQ_NET_PENDING)
244 if (req->master_bio) {
245 /* this is DATA_RECEIVED (remote read)
246 * or protocol C P_WRITE_ACK
247 * or protocol B P_RECV_ACK
248 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
249 * or canceled or failed,
250 * or killed from the transfer log due to connection loss.
254 * figure out whether to report success or failure.
256 * report success when at least one of the operations succeeded.
257 * or, to put the other way,
258 * only report failure, when both operations failed.
260 * what to do about the failures is handled elsewhere.
261 * what we need to do here is just: complete the master_bio.
263 * local completion error, if any, has been stored as ERR_PTR
264 * in private_bio within drbd_request_endio.
266 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
267 int error = PTR_ERR(req->private_bio);
269 /* remove the request from the conflict detection
270 * respective block_id verification hash */
271 if (!drbd_interval_empty(&req->i)) {
272 struct rb_root *root;
275 root = &mdev->write_requests;
277 root = &mdev->read_requests;
278 drbd_remove_request_interval(root, req);
279 } else if (!(s & RQ_POSTPONED))
280 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
282 /* for writes we need to do some extra housekeeping */
284 _about_to_complete_local_write(mdev, req);
286 /* Update disk stats */
287 _drbd_end_io_acct(mdev, req);
289 if (!(s & RQ_POSTPONED)) {
290 m->error = ok ? 0 : (error ?: -EIO);
291 m->bio = req->master_bio;
293 req->master_bio = NULL;
296 if (s & RQ_LOCAL_PENDING)
299 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
300 /* this is disconnected (local only) operation,
301 * or protocol A, B, or C P_BARRIER_ACK,
302 * or killed from the transfer log due to connection loss. */
303 _req_is_done(mdev, req, rw);
305 /* else: network part and not DONE yet. that is
306 * protocol A, B, or C, barrier ack still pending... */
309 static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
311 struct drbd_conf *mdev = req->w.mdev;
313 if (!drbd_suspended(mdev))
314 _req_may_be_done(req, m);
317 /* obviously this could be coded as many single functions
318 * instead of one huge switch,
319 * or by putting the code directly in the respective locations
320 * (as it has been before).
322 * but having it this way
323 * enforces that it is all in this one place, where it is easier to audit,
324 * it makes it obvious that whatever "event" "happens" to a request should
325 * happen "atomically" within the req_lock,
326 * and it enforces that we have to think in a very structured manner
327 * about the "events" that may happen to a request during its life time ...
329 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
330 struct bio_and_error *m)
332 struct drbd_conf *mdev = req->w.mdev;
341 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
344 /* does not happen...
345 * initialization done in drbd_req_new
350 case TO_BE_SENT: /* via network */
351 /* reached via __drbd_make_request
352 * and from w_read_retry_remote */
353 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
354 req->rq_state |= RQ_NET_PENDING;
356 nc = rcu_dereference(mdev->tconn->net_conf);
357 p = nc->wire_protocol;
360 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
361 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
362 inc_ap_pending(mdev);
365 case TO_BE_SUBMITTED: /* locally */
366 /* reached via __drbd_make_request */
367 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
368 req->rq_state |= RQ_LOCAL_PENDING;
372 if (req->rq_state & RQ_WRITE)
373 mdev->writ_cnt += req->i.size >> 9;
375 mdev->read_cnt += req->i.size >> 9;
377 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
378 req->rq_state &= ~RQ_LOCAL_PENDING;
380 _req_may_be_done_not_susp(req, m);
385 req->rq_state |= RQ_LOCAL_ABORTED;
386 if (req->rq_state & RQ_WRITE)
387 _req_may_be_done_not_susp(req, m);
389 goto goto_queue_for_net_read;
392 case WRITE_COMPLETED_WITH_ERROR:
393 req->rq_state |= RQ_LOCAL_COMPLETED;
394 req->rq_state &= ~RQ_LOCAL_PENDING;
396 __drbd_chk_io_error(mdev, false);
397 _req_may_be_done_not_susp(req, m);
401 case READ_AHEAD_COMPLETED_WITH_ERROR:
402 /* it is legal to fail READA */
403 req->rq_state |= RQ_LOCAL_COMPLETED;
404 req->rq_state &= ~RQ_LOCAL_PENDING;
405 _req_may_be_done_not_susp(req, m);
409 case READ_COMPLETED_WITH_ERROR:
410 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
412 req->rq_state |= RQ_LOCAL_COMPLETED;
413 req->rq_state &= ~RQ_LOCAL_PENDING;
415 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
417 __drbd_chk_io_error(mdev, false);
420 goto_queue_for_net_read:
422 /* no point in retrying if there is no good remote data,
423 * or we have no connection. */
424 if (mdev->state.pdsk != D_UP_TO_DATE) {
425 _req_may_be_done_not_susp(req, m);
429 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
430 req->rq_state |= RQ_NET_PENDING;
431 inc_ap_pending(mdev);
432 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
434 case QUEUE_FOR_NET_READ:
435 /* READ or READA, and
437 * or target area marked as invalid,
438 * or just got an io-error. */
439 /* from __drbd_make_request
440 * or from bio_endio during read io-error recovery */
442 /* so we can verify the handle in the answer packet
443 * corresponding hlist_del is in _req_may_be_done() */
444 D_ASSERT(drbd_interval_empty(&req->i));
445 drbd_insert_interval(&mdev->read_requests, &req->i);
447 set_bit(UNPLUG_REMOTE, &mdev->flags);
449 D_ASSERT(req->rq_state & RQ_NET_PENDING);
450 req->rq_state |= RQ_NET_QUEUED;
451 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
452 ? w_read_retry_remote
454 drbd_queue_work(&mdev->tconn->data.work, &req->w);
457 case QUEUE_FOR_NET_WRITE:
458 /* assert something? */
459 /* from __drbd_make_request only */
461 /* corresponding hlist_del is in _req_may_be_done() */
462 D_ASSERT(drbd_interval_empty(&req->i));
463 drbd_insert_interval(&mdev->write_requests, &req->i);
466 * In case the req ended up on the transfer log before being
467 * queued on the worker, it could lead to this request being
468 * missed during cleanup after connection loss.
469 * So we have to do both operations here,
470 * within the same lock that protects the transfer log.
472 * _req_add_to_epoch(req); this has to be after the
473 * _maybe_start_new_epoch(req); which happened in
474 * __drbd_make_request, because we now may set the bit
475 * again ourselves to close the current epoch.
477 * Add req to the (now) current epoch (barrier). */
479 /* otherwise we may lose an unplug, which may cause some remote
480 * io-scheduler timeout to expire, increasing maximum latency,
481 * hurting performance. */
482 set_bit(UNPLUG_REMOTE, &mdev->flags);
484 /* see __drbd_make_request,
485 * just after it grabs the req_lock */
486 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->tconn->flags) == 0);
488 req->epoch = mdev->tconn->newest_tle->br_number;
490 /* increment size of current epoch */
491 mdev->tconn->newest_tle->n_writes++;
493 /* queue work item to send data */
494 D_ASSERT(req->rq_state & RQ_NET_PENDING);
495 req->rq_state |= RQ_NET_QUEUED;
496 req->w.cb = w_send_dblock;
497 drbd_queue_work(&mdev->tconn->data.work, &req->w);
499 /* close the epoch, in case it outgrew the limit */
501 nc = rcu_dereference(mdev->tconn->net_conf);
502 p = nc->max_epoch_size;
504 if (mdev->tconn->newest_tle->n_writes >= p)
509 case QUEUE_FOR_SEND_OOS:
510 req->rq_state |= RQ_NET_QUEUED;
511 req->w.cb = w_send_out_of_sync;
512 drbd_queue_work(&mdev->tconn->data.work, &req->w);
515 case OOS_HANDED_TO_NETWORK:
516 /* actually the same */
518 /* treat it the same */
520 /* real cleanup will be done from tl_clear. just update flags
521 * so it is no longer marked as on the worker queue */
522 req->rq_state &= ~RQ_NET_QUEUED;
523 /* if we did it right, tl_clear should be scheduled only after
524 * this, so this should not be necessary! */
525 _req_may_be_done_not_susp(req, m);
528 case HANDED_OVER_TO_NETWORK:
529 /* assert something? */
530 if (bio_data_dir(req->master_bio) == WRITE)
531 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
533 if (bio_data_dir(req->master_bio) == WRITE &&
534 !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
535 /* this is what is dangerous about protocol A:
536 * pretend it was successfully written on the peer. */
537 if (req->rq_state & RQ_NET_PENDING) {
538 dec_ap_pending(mdev);
539 req->rq_state &= ~RQ_NET_PENDING;
540 req->rq_state |= RQ_NET_OK;
541 } /* else: neg-ack was faster... */
542 /* it is still not yet RQ_NET_DONE until the
543 * corresponding epoch barrier got acked as well,
544 * so we know what to dirty on connection loss */
546 req->rq_state &= ~RQ_NET_QUEUED;
547 req->rq_state |= RQ_NET_SENT;
548 /* because _drbd_send_zc_bio could sleep, and may want to
549 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
550 * "COMPLETED_OK" events came in, once we return from
551 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
552 * whether it is done already, and end it. */
553 _req_may_be_done_not_susp(req, m);
556 case READ_RETRY_REMOTE_CANCELED:
557 req->rq_state &= ~RQ_NET_QUEUED;
558 /* fall through, in case we raced with drbd_disconnect */
559 case CONNECTION_LOST_WHILE_PENDING:
560 /* transfer log cleanup after connection loss */
561 /* assert something? */
562 if (req->rq_state & RQ_NET_PENDING)
563 dec_ap_pending(mdev);
565 p = !(req->rq_state & RQ_WRITE) && req->rq_state & RQ_NET_PENDING;
567 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
568 req->rq_state |= RQ_NET_DONE;
569 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
570 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
572 /* if it is still queued, we may not complete it here.
573 * it will be canceled soon. */
574 if (!(req->rq_state & RQ_NET_QUEUED)) {
576 goto goto_read_retry_local;
577 _req_may_be_done(req, m); /* Allowed while state.susp */
581 case WRITE_ACKED_BY_PEER_AND_SIS:
582 req->rq_state |= RQ_NET_SIS;
584 /* for discarded conflicting writes of multiple primaries,
585 * there is no need to keep anything in the tl, potential
586 * node crashes are covered by the activity log. */
587 req->rq_state |= RQ_NET_DONE;
589 case WRITE_ACKED_BY_PEER:
590 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
591 /* protocol C; successfully written on peer.
592 * Nothing to do here.
593 * We want to keep the tl in place for all protocols, to cater
594 * for volatile write-back caches on lower level devices.
596 * A barrier request is expected to have forced all prior
597 * requests onto stable storage, so completion of a barrier
598 * request could set NET_DONE right here, and not wait for the
599 * P_BARRIER_ACK, but that is an unnecessary optimization. */
602 /* this makes it effectively the same as for: */
603 case RECV_ACKED_BY_PEER:
604 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
605 /* protocol B; pretends to be successfully written on peer.
606 * see also notes above in HANDED_OVER_TO_NETWORK about
609 req->rq_state |= RQ_NET_OK;
610 D_ASSERT(req->rq_state & RQ_NET_PENDING);
611 dec_ap_pending(mdev);
612 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
613 req->rq_state &= ~RQ_NET_PENDING;
614 _req_may_be_done_not_susp(req, m);
618 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
619 /* If this node has already detected the write conflict, the
620 * worker will be waiting on misc_wait. Wake it up once this
621 * request has completed locally.
623 D_ASSERT(req->rq_state & RQ_NET_PENDING);
624 req->rq_state |= RQ_POSTPONED;
625 _req_may_be_done_not_susp(req, m);
629 /* assert something? */
630 if (req->rq_state & RQ_NET_PENDING) {
631 dec_ap_pending(mdev);
632 if (req->rq_state & RQ_WRITE)
633 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
635 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
637 req->rq_state |= RQ_NET_DONE;
639 if (!(req->rq_state & RQ_WRITE))
640 goto goto_read_retry_local;
642 _req_may_be_done_not_susp(req, m);
643 /* else: done by HANDED_OVER_TO_NETWORK */
646 goto_read_retry_local:
647 if (!drbd_may_do_local_read(mdev, req->i.sector, req->i.size)) {
648 _req_may_be_done_not_susp(req, m);
651 D_ASSERT(!(req->rq_state & RQ_LOCAL_PENDING));
652 req->rq_state |= RQ_LOCAL_PENDING;
655 req->w.cb = w_restart_disk_io;
656 drbd_queue_work(&mdev->tconn->data.work, &req->w);
659 case FAIL_FROZEN_DISK_IO:
660 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
663 _req_may_be_done(req, m); /* Allowed while state.susp */
666 case RESTART_FROZEN_DISK_IO:
667 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
670 req->rq_state &= ~RQ_LOCAL_COMPLETED;
673 if (bio_data_dir(req->master_bio) == WRITE)
677 req->w.cb = w_restart_disk_io;
678 drbd_queue_work(&mdev->tconn->data.work, &req->w);
682 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
683 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
684 Trowing them out of the TL here by pretending we got a BARRIER_ACK
685 We ensure that the peer was not rebooted */
686 if (!(req->rq_state & RQ_NET_OK)) {
688 drbd_queue_work(&mdev->tconn->data.work, &req->w);
689 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
693 /* else, fall through to BARRIER_ACKED */
696 if (!(req->rq_state & RQ_WRITE))
699 if (req->rq_state & RQ_NET_PENDING) {
700 /* barrier came in before all requests were acked.
701 * this is bad, because if the connection is lost now,
702 * we won't be able to clean them up... */
703 dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
704 list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests);
706 if ((req->rq_state & RQ_NET_MASK) != 0) {
707 req->rq_state |= RQ_NET_DONE;
708 if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK)))
709 atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
711 _req_may_be_done(req, m); /* Allowed while state.susp */
715 D_ASSERT(req->rq_state & RQ_NET_PENDING);
716 dec_ap_pending(mdev);
717 req->rq_state &= ~RQ_NET_PENDING;
718 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
719 _req_may_be_done_not_susp(req, m);
726 /* we may do a local read if:
727 * - we are consistent (of course),
728 * - or we are generally inconsistent,
729 * BUT we are still/already IN SYNC for this area.
730 * since size may be bigger than BM_BLOCK_SIZE,
731 * we may need to check several bits.
733 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
735 unsigned long sbnr, ebnr;
736 sector_t esector, nr_sectors;
738 if (mdev->state.disk == D_UP_TO_DATE)
740 if (mdev->state.disk != D_INCONSISTENT)
742 esector = sector + (size >> 9) - 1;
743 nr_sectors = drbd_get_capacity(mdev->this_bdev);
744 D_ASSERT(sector < nr_sectors);
745 D_ASSERT(esector < nr_sectors);
747 sbnr = BM_SECT_TO_BIT(sector);
748 ebnr = BM_SECT_TO_BIT(esector);
750 return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
753 static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector)
755 enum drbd_read_balancing rbm;
756 struct backing_dev_info *bdi;
759 if (mdev->state.pdsk < D_UP_TO_DATE)
763 rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
767 case RB_CONGESTED_REMOTE:
768 bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
769 return bdi_read_congested(bdi);
770 case RB_LEAST_PENDING:
771 return atomic_read(&mdev->local_cnt) >
772 atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
773 case RB_32K_STRIPING: /* stripe_shift = 15 */
774 case RB_64K_STRIPING:
775 case RB_128K_STRIPING:
776 case RB_256K_STRIPING:
777 case RB_512K_STRIPING:
778 case RB_1M_STRIPING: /* stripe_shift = 20 */
779 stripe_shift = (rbm - RB_32K_STRIPING + 15);
780 return (sector >> (stripe_shift - 9)) & 1;
782 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
783 case RB_PREFER_REMOTE:
785 case RB_PREFER_LOCAL:
792 * complete_conflicting_writes - wait for any conflicting write requests
794 * The write_requests tree contains all active write requests which we
795 * currently know about. Wait for any requests to complete which conflict with
798 static int complete_conflicting_writes(struct drbd_conf *mdev,
799 sector_t sector, int size)
802 struct drbd_interval *i;
805 i = drbd_find_overlap(&mdev->write_requests, sector, size);
808 err = drbd_wait_misc(mdev, i);
814 int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
816 const int rw = bio_rw(bio);
817 const int size = bio->bi_size;
818 const sector_t sector = bio->bi_sector;
819 struct drbd_tl_epoch *b = NULL;
820 struct drbd_request *req;
822 int local, remote, send_oos = 0;
825 union drbd_dev_state s;
827 /* allocate outside of all locks; */
828 req = drbd_req_new(mdev, bio);
831 /* only pass the error to the upper layers.
832 * if user cannot handle io errors, that's not our business. */
833 dev_err(DEV, "could not kmalloc() req\n");
834 bio_endio(bio, -ENOMEM);
837 req->start_time = start_time;
839 local = get_ldev(mdev);
841 bio_put(req->private_bio); /* or we get a bio leak */
842 req->private_bio = NULL;
849 if (!drbd_may_do_local_read(mdev, sector, size) ||
850 remote_due_to_read_balancing(mdev, sector)) {
851 /* we could kick the syncer to
852 * sync this extent asap, wait for
853 * it, then continue locally.
854 * Or just issue the request remotely.
857 bio_put(req->private_bio);
858 req->private_bio = NULL;
862 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
865 /* If we have a disk, but a READA request is mapped to remote,
866 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
867 * Just fail that READA request right here.
869 * THINK: maybe fail all READA when not local?
870 * or make this configurable...
871 * if network is slow, READA won't do any good.
873 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
875 goto fail_and_free_req;
878 /* For WRITES going to the local disk, grab a reference on the target
879 * extent. This waits for any resync activity in the corresponding
880 * resync extent to finish, and, if necessary, pulls in the target
881 * extent into the activity log, which involves further disk io because
882 * of transactional on-disk meta data updates. */
883 if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
884 req->rq_state |= RQ_IN_ACT_LOG;
885 drbd_al_begin_io(mdev, &req->i);
889 remote = remote && drbd_should_do_remote(s);
890 send_oos = rw == WRITE && drbd_should_send_out_of_sync(s);
891 D_ASSERT(!(remote && send_oos));
893 if (!(local || remote) && !drbd_suspended(mdev)) {
894 if (__ratelimit(&drbd_ratelimit_state))
895 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
897 goto fail_free_complete;
900 /* For WRITE request, we have to make sure that we have an
901 * unused_spare_tle, in case we need to start a new epoch.
902 * I try to be smart and avoid to pre-allocate always "just in case",
903 * but there is a race between testing the bit and pointer outside the
904 * spinlock, and grabbing the spinlock.
905 * if we lost that race, we retry. */
906 if (rw == WRITE && (remote || send_oos) &&
907 mdev->tconn->unused_spare_tle == NULL &&
908 test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
910 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
912 dev_err(DEV, "Failed to alloc barrier.\n");
914 goto fail_free_complete;
918 /* GOOD, everything prepared, grab the spin_lock */
919 spin_lock_irq(&mdev->tconn->req_lock);
922 err = complete_conflicting_writes(mdev, sector, size);
924 if (err != -ERESTARTSYS)
925 _conn_request_state(mdev->tconn,
928 spin_unlock_irq(&mdev->tconn->req_lock);
930 goto fail_free_complete;
934 if (drbd_suspended(mdev)) {
935 /* If we got suspended, use the retry mechanism in
936 drbd_make_request() to restart processing of this
937 bio. In the next call to drbd_make_request
938 we sleep in inc_ap_bio() */
940 spin_unlock_irq(&mdev->tconn->req_lock);
941 goto fail_free_complete;
944 if (remote || send_oos) {
945 remote = drbd_should_do_remote(mdev->state);
946 send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
947 D_ASSERT(!(remote && send_oos));
949 if (!(remote || send_oos))
950 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
951 if (!(local || remote)) {
952 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
953 spin_unlock_irq(&mdev->tconn->req_lock);
955 goto fail_free_complete;
959 if (b && mdev->tconn->unused_spare_tle == NULL) {
960 mdev->tconn->unused_spare_tle = b;
963 if (rw == WRITE && (remote || send_oos) &&
964 mdev->tconn->unused_spare_tle == NULL &&
965 test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
966 /* someone closed the current epoch
967 * while we were grabbing the spinlock */
968 spin_unlock_irq(&mdev->tconn->req_lock);
969 goto allocate_barrier;
973 /* Update disk stats */
974 _drbd_start_io_acct(mdev, req, bio);
976 /* _maybe_start_new_epoch(mdev);
977 * If we need to generate a write barrier packet, we have to add the
978 * new epoch (barrier) object, and queue the barrier packet for sending,
979 * and queue the req's data after it _within the same lock_, otherwise
980 * we have race conditions were the reorder domains could be mixed up.
982 * Even read requests may start a new epoch and queue the corresponding
983 * barrier packet. To get the write ordering right, we only have to
984 * make sure that, if this is a write request and it triggered a
985 * barrier packet, this request is queued within the same spinlock. */
986 if ((remote || send_oos) && mdev->tconn->unused_spare_tle &&
987 test_and_clear_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
988 _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle);
989 mdev->tconn->unused_spare_tle = NULL;
991 D_ASSERT(!(remote && rw == WRITE &&
992 test_bit(CREATE_BARRIER, &mdev->tconn->flags)));
996 * Actually, 'local' may be wrong here already, since we may have failed
997 * to write to the meta data, and may become wrong anytime because of
998 * local io-error for some other request, which would lead to us
999 * "detaching" the local disk.
1001 * 'remote' may become wrong any time because the network could fail.
1003 * This is a harmless race condition, though, since it is handled
1004 * correctly at the appropriate places; so it just defers the failure
1005 * of the respective operation.
1008 /* mark them early for readability.
1009 * this just sets some state flags. */
1011 _req_mod(req, TO_BE_SENT);
1013 _req_mod(req, TO_BE_SUBMITTED);
1015 list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests);
1017 /* NOTE remote first: to get the concurrent write detection right,
1018 * we must register the request before start of local IO. */
1020 /* either WRITE and C_CONNECTED,
1021 * or READ, and no local disk,
1022 * or READ, but not in sync.
1024 _req_mod(req, (rw == WRITE)
1025 ? QUEUE_FOR_NET_WRITE
1026 : QUEUE_FOR_NET_READ);
1028 if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
1029 _req_mod(req, QUEUE_FOR_SEND_OOS);
1032 nc = rcu_dereference(mdev->tconn->net_conf);
1034 nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) {
1037 if (nc->cong_fill &&
1038 atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
1039 dev_info(DEV, "Congestion-fill threshold reached\n");
1043 if (mdev->act_log->used >= nc->cong_extents) {
1044 dev_info(DEV, "Congestion-extents threshold reached\n");
1049 queue_barrier(mdev); /* last barrier, after mirrored writes */
1051 if (nc->on_congestion == OC_PULL_AHEAD)
1052 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
1053 else /*nc->on_congestion == OC_DISCONNECT */
1054 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
1059 spin_unlock_irq(&mdev->tconn->req_lock);
1060 kfree(b); /* if someone else has beaten us to it... */
1063 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
1065 /* State may have changed since we grabbed our reference on the
1066 * mdev->ldev member. Double check, and short-circuit to endio.
1067 * In case the last activity log transaction failed to get on
1068 * stable storage, and this is a WRITE, we may not even submit
1070 if (get_ldev(mdev)) {
1071 if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
1072 : rw == READ ? DRBD_FAULT_DT_RD
1073 : DRBD_FAULT_DT_RA))
1074 bio_endio(req->private_bio, -EIO);
1076 generic_make_request(req->private_bio);
1079 bio_endio(req->private_bio, -EIO);
1085 if (req->rq_state & RQ_IN_ACT_LOG)
1086 drbd_al_complete_io(mdev, &req->i);
1089 bio_put(req->private_bio);
1090 req->private_bio = NULL;
1094 bio_endio(bio, err);
1103 int drbd_make_request(struct request_queue *q, struct bio *bio)
1105 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1106 unsigned long start_time;
1108 start_time = jiffies;
1111 * what we "blindly" assume:
1113 D_ASSERT(bio->bi_size > 0);
1114 D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1118 } while (__drbd_make_request(mdev, bio, start_time));
1123 /* This is called by bio_add_page().
1125 * q->max_hw_sectors and other global limits are already enforced there.
1127 * We need to call down to our lower level device,
1128 * in case it has special restrictions.
1130 * We also may need to enforce configured max-bio-bvecs limits.
1132 * As long as the BIO is empty we have to allow at least one bvec,
1133 * regardless of size and offset, so no need to ask lower levels.
1135 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1137 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1138 unsigned int bio_size = bvm->bi_size;
1139 int limit = DRBD_MAX_BIO_SIZE;
1142 if (bio_size && get_ldev(mdev)) {
1143 struct request_queue * const b =
1144 mdev->ldev->backing_bdev->bd_disk->queue;
1145 if (b->merge_bvec_fn) {
1146 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1147 limit = min(limit, backing_limit);
1154 void request_timer_fn(unsigned long data)
1156 struct drbd_conf *mdev = (struct drbd_conf *) data;
1157 struct drbd_tconn *tconn = mdev->tconn;
1158 struct drbd_request *req; /* oldest request */
1159 struct list_head *le;
1160 struct net_conf *nc;
1161 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1164 nc = rcu_dereference(tconn->net_conf);
1165 ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0;
1167 if (get_ldev(mdev)) {
1168 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1173 et = min_not_zero(dt, ent);
1175 if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED))
1176 return; /* Recurring timer stopped */
1178 spin_lock_irq(&tconn->req_lock);
1179 le = &tconn->oldest_tle->requests;
1180 if (list_empty(le)) {
1181 spin_unlock_irq(&tconn->req_lock);
1182 mod_timer(&mdev->request_timer, jiffies + et);
1187 req = list_entry(le, struct drbd_request, tl_requests);
1188 if (ent && req->rq_state & RQ_NET_PENDING) {
1189 if (time_is_before_eq_jiffies(req->start_time + ent)) {
1190 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1191 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1194 if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev) {
1195 if (time_is_before_eq_jiffies(req->start_time + dt)) {
1196 dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1197 __drbd_chk_io_error(mdev, 1);
1200 nt = (time_is_before_eq_jiffies(req->start_time + et) ? jiffies : req->start_time) + et;
1201 spin_unlock_irq(&tconn->req_lock);
1202 mod_timer(&mdev->request_timer, nt);