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_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
43 part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
44 (void) cpu; /* The macro invocations above want the cpu argument, I do not like
45 the compiler warning about cpu only assigned but never used... */
46 part_inc_in_flight(&mdev->vdisk->part0, rw);
50 /* Update disk stats when completing request upwards */
51 static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
53 int rw = bio_data_dir(req->master_bio);
54 unsigned long duration = jiffies - req->start_time;
56 cpu = part_stat_lock();
57 part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
58 part_round_stats(cpu, &mdev->vdisk->part0);
59 part_dec_in_flight(&mdev->vdisk->part0, rw);
63 static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
66 struct drbd_request *req;
68 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
72 drbd_req_make_private_bio(req, bio_src);
73 req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
75 req->master_bio = bio_src;
78 drbd_clear_interval(&req->i);
79 req->i.sector = bio_src->bi_sector;
80 req->i.size = bio_src->bi_size;
82 req->i.waiting = false;
84 INIT_LIST_HEAD(&req->tl_requests);
85 INIT_LIST_HEAD(&req->w.list);
90 static void drbd_req_free(struct drbd_request *req)
92 mempool_free(req, drbd_request_mempool);
95 /* rw is bio_data_dir(), only READ or WRITE */
96 static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
98 const unsigned long s = req->rq_state;
100 /* remove it from the transfer log.
101 * well, only if it had been there in the first
102 * place... if it had not (local only or conflicting
103 * and never sent), it should still be "empty" as
104 * initialized in drbd_req_new(), so we can list_del() it
105 * here unconditionally */
106 list_del(&req->tl_requests);
108 /* if it was a write, we may have to set the corresponding
109 * bit(s) out-of-sync first. If it had a local part, we need to
110 * release the reference to the activity log. */
112 /* Set out-of-sync unless both OK flags are set
113 * (local only or remote failed).
114 * Other places where we set out-of-sync:
115 * READ with local io-error */
116 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
117 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
119 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
120 drbd_set_in_sync(mdev, req->i.sector, req->i.size);
122 /* one might be tempted to move the drbd_al_complete_io
123 * to the local io completion callback drbd_request_endio.
124 * but, if this was a mirror write, we may only
125 * drbd_al_complete_io after this is RQ_NET_DONE,
126 * otherwise the extent could be dropped from the al
127 * before it has actually been written on the peer.
128 * if we crash before our peer knows about the request,
129 * but after the extent has been dropped from the al,
130 * we would forget to resync the corresponding extent.
132 if (s & RQ_LOCAL_MASK) {
133 if (get_ldev_if_state(mdev, D_FAILED)) {
134 if (s & RQ_IN_ACT_LOG)
135 drbd_al_complete_io(mdev, &req->i);
137 } else if (__ratelimit(&drbd_ratelimit_state)) {
138 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
139 "but my Disk seems to have failed :(\n",
140 (unsigned long long) req->i.sector, req->i.size);
148 static void queue_barrier(struct drbd_conf *mdev)
150 struct drbd_tl_epoch *b;
151 struct drbd_tconn *tconn = mdev->tconn;
153 /* We are within the req_lock. Once we queued the barrier for sending,
154 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
155 * barrier/epoch object is added. This is the only place this bit is
156 * set. It indicates that the barrier for this epoch is already queued,
157 * and no new epoch has been created yet. */
158 if (test_bit(CREATE_BARRIER, &tconn->flags))
161 b = tconn->newest_tle;
162 b->w.cb = w_send_barrier;
164 /* inc_ap_pending done here, so we won't
165 * get imbalanced on connection loss.
166 * dec_ap_pending will be done in got_BarrierAck
167 * or (on connection loss) in tl_clear. */
168 inc_ap_pending(mdev);
169 drbd_queue_work(&tconn->data.work, &b->w);
170 set_bit(CREATE_BARRIER, &tconn->flags);
173 static void _about_to_complete_local_write(struct drbd_conf *mdev,
174 struct drbd_request *req)
176 const unsigned long s = req->rq_state;
178 /* Before we can signal completion to the upper layers,
179 * we may need to close the current epoch.
180 * We can skip this, if this request has not even been sent, because we
181 * did not have a fully established connection yet/anymore, during
182 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
184 if (mdev->state.conn >= C_CONNECTED &&
185 (s & RQ_NET_SENT) != 0 &&
186 req->epoch == mdev->tconn->newest_tle->br_number)
190 void complete_master_bio(struct drbd_conf *mdev,
191 struct bio_and_error *m)
193 bio_endio(m->bio, m->error);
198 static void drbd_remove_request_interval(struct rb_root *root,
199 struct drbd_request *req)
201 struct drbd_conf *mdev = req->w.mdev;
202 struct drbd_interval *i = &req->i;
204 drbd_remove_interval(root, i);
206 /* Wake up any processes waiting for this request to complete. */
208 wake_up(&mdev->misc_wait);
211 /* Helper for __req_mod().
212 * Set m->bio to the master bio, if it is fit to be completed,
213 * or leave it alone (it is initialized to NULL in __req_mod),
214 * if it has already been completed, or cannot be completed yet.
215 * If m->bio is set, the error status to be returned is placed in m->error.
217 void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
219 const unsigned long s = req->rq_state;
220 struct drbd_conf *mdev = req->w.mdev;
221 int rw = req->rq_state & RQ_WRITE ? WRITE : READ;
223 /* we must not complete the master bio, while it is
224 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
225 * not yet acknowledged by the peer
226 * not yet completed by the local io subsystem
227 * these flags may get cleared in any order by
230 * the bio_endio completion callbacks.
232 if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
234 if (req->i.waiting) {
235 /* Retry all conflicting peer requests. */
236 wake_up(&mdev->misc_wait);
238 if (s & RQ_NET_QUEUED)
240 if (s & RQ_NET_PENDING)
243 if (req->master_bio) {
244 /* this is DATA_RECEIVED (remote read)
245 * or protocol C P_WRITE_ACK
246 * or protocol B P_RECV_ACK
247 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
248 * or canceled or failed,
249 * or killed from the transfer log due to connection loss.
253 * figure out whether to report success or failure.
255 * report success when at least one of the operations succeeded.
256 * or, to put the other way,
257 * only report failure, when both operations failed.
259 * what to do about the failures is handled elsewhere.
260 * what we need to do here is just: complete the master_bio.
262 * local completion error, if any, has been stored as ERR_PTR
263 * in private_bio within drbd_request_endio.
265 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
266 int error = PTR_ERR(req->private_bio);
268 /* remove the request from the conflict detection
269 * respective block_id verification hash */
270 if (!drbd_interval_empty(&req->i)) {
271 struct rb_root *root;
274 root = &mdev->write_requests;
276 root = &mdev->read_requests;
277 drbd_remove_request_interval(root, req);
278 } else if (!(s & RQ_POSTPONED))
279 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
281 /* for writes we need to do some extra housekeeping */
283 _about_to_complete_local_write(mdev, req);
285 /* Update disk stats */
286 _drbd_end_io_acct(mdev, req);
288 if (!(s & RQ_POSTPONED)) {
289 m->error = ok ? 0 : (error ?: -EIO);
290 m->bio = req->master_bio;
292 req->master_bio = NULL;
295 if (s & RQ_LOCAL_PENDING)
298 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
299 /* this is disconnected (local only) operation,
300 * or protocol A, B, or C P_BARRIER_ACK,
301 * or killed from the transfer log due to connection loss. */
302 _req_is_done(mdev, req, rw);
304 /* else: network part and not DONE yet. that is
305 * protocol A, B, or C, barrier ack still pending... */
308 static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
310 struct drbd_conf *mdev = req->w.mdev;
312 if (!drbd_suspended(mdev))
313 _req_may_be_done(req, m);
316 /* obviously this could be coded as many single functions
317 * instead of one huge switch,
318 * or by putting the code directly in the respective locations
319 * (as it has been before).
321 * but having it this way
322 * enforces that it is all in this one place, where it is easier to audit,
323 * it makes it obvious that whatever "event" "happens" to a request should
324 * happen "atomically" within the req_lock,
325 * and it enforces that we have to think in a very structured manner
326 * about the "events" that may happen to a request during its life time ...
328 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
329 struct bio_and_error *m)
331 struct drbd_conf *mdev = req->w.mdev;
340 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
343 /* does not happen...
344 * initialization done in drbd_req_new
349 case TO_BE_SENT: /* via network */
350 /* reached via __drbd_make_request
351 * and from w_read_retry_remote */
352 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
353 req->rq_state |= RQ_NET_PENDING;
355 nc = rcu_dereference(mdev->tconn->net_conf);
356 p = nc->wire_protocol;
359 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
360 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
361 inc_ap_pending(mdev);
364 case TO_BE_SUBMITTED: /* locally */
365 /* reached via __drbd_make_request */
366 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
367 req->rq_state |= RQ_LOCAL_PENDING;
371 if (req->rq_state & RQ_WRITE)
372 mdev->writ_cnt += req->i.size >> 9;
374 mdev->read_cnt += req->i.size >> 9;
376 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
377 req->rq_state &= ~RQ_LOCAL_PENDING;
379 _req_may_be_done_not_susp(req, m);
384 req->rq_state |= RQ_LOCAL_ABORTED;
385 if (req->rq_state & RQ_WRITE)
386 _req_may_be_done_not_susp(req, m);
388 goto goto_queue_for_net_read;
391 case WRITE_COMPLETED_WITH_ERROR:
392 req->rq_state |= RQ_LOCAL_COMPLETED;
393 req->rq_state &= ~RQ_LOCAL_PENDING;
395 __drbd_chk_io_error(mdev, false);
396 _req_may_be_done_not_susp(req, m);
400 case READ_AHEAD_COMPLETED_WITH_ERROR:
401 /* it is legal to fail READA */
402 req->rq_state |= RQ_LOCAL_COMPLETED;
403 req->rq_state &= ~RQ_LOCAL_PENDING;
404 _req_may_be_done_not_susp(req, m);
408 case READ_COMPLETED_WITH_ERROR:
409 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
411 req->rq_state |= RQ_LOCAL_COMPLETED;
412 req->rq_state &= ~RQ_LOCAL_PENDING;
414 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
416 __drbd_chk_io_error(mdev, false);
419 goto_queue_for_net_read:
421 /* no point in retrying if there is no good remote data,
422 * or we have no connection. */
423 if (mdev->state.pdsk != D_UP_TO_DATE) {
424 _req_may_be_done_not_susp(req, m);
428 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
429 req->rq_state |= RQ_NET_PENDING;
430 inc_ap_pending(mdev);
431 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
433 case QUEUE_FOR_NET_READ:
434 /* READ or READA, and
436 * or target area marked as invalid,
437 * or just got an io-error. */
438 /* from __drbd_make_request
439 * or from bio_endio during read io-error recovery */
441 /* so we can verify the handle in the answer packet
442 * corresponding hlist_del is in _req_may_be_done() */
443 D_ASSERT(drbd_interval_empty(&req->i));
444 drbd_insert_interval(&mdev->read_requests, &req->i);
446 set_bit(UNPLUG_REMOTE, &mdev->flags);
448 D_ASSERT(req->rq_state & RQ_NET_PENDING);
449 req->rq_state |= RQ_NET_QUEUED;
450 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
451 ? w_read_retry_remote
453 drbd_queue_work(&mdev->tconn->data.work, &req->w);
456 case QUEUE_FOR_NET_WRITE:
457 /* assert something? */
458 /* from __drbd_make_request only */
460 /* corresponding hlist_del is in _req_may_be_done() */
461 D_ASSERT(drbd_interval_empty(&req->i));
462 drbd_insert_interval(&mdev->write_requests, &req->i);
465 * In case the req ended up on the transfer log before being
466 * queued on the worker, it could lead to this request being
467 * missed during cleanup after connection loss.
468 * So we have to do both operations here,
469 * within the same lock that protects the transfer log.
471 * _req_add_to_epoch(req); this has to be after the
472 * _maybe_start_new_epoch(req); which happened in
473 * __drbd_make_request, because we now may set the bit
474 * again ourselves to close the current epoch.
476 * Add req to the (now) current epoch (barrier). */
478 /* otherwise we may lose an unplug, which may cause some remote
479 * io-scheduler timeout to expire, increasing maximum latency,
480 * hurting performance. */
481 set_bit(UNPLUG_REMOTE, &mdev->flags);
483 /* see __drbd_make_request,
484 * just after it grabs the req_lock */
485 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->tconn->flags) == 0);
487 req->epoch = mdev->tconn->newest_tle->br_number;
489 /* increment size of current epoch */
490 mdev->tconn->newest_tle->n_writes++;
492 /* queue work item to send data */
493 D_ASSERT(req->rq_state & RQ_NET_PENDING);
494 req->rq_state |= RQ_NET_QUEUED;
495 req->w.cb = w_send_dblock;
496 drbd_queue_work(&mdev->tconn->data.work, &req->w);
498 /* close the epoch, in case it outgrew the limit */
500 nc = rcu_dereference(mdev->tconn->net_conf);
501 p = nc->max_epoch_size;
503 if (mdev->tconn->newest_tle->n_writes >= p)
508 case QUEUE_FOR_SEND_OOS:
509 req->rq_state |= RQ_NET_QUEUED;
510 req->w.cb = w_send_out_of_sync;
511 drbd_queue_work(&mdev->tconn->data.work, &req->w);
514 case OOS_HANDED_TO_NETWORK:
515 /* actually the same */
517 /* treat it the same */
519 /* real cleanup will be done from tl_clear. just update flags
520 * so it is no longer marked as on the worker queue */
521 req->rq_state &= ~RQ_NET_QUEUED;
522 /* if we did it right, tl_clear should be scheduled only after
523 * this, so this should not be necessary! */
524 _req_may_be_done_not_susp(req, m);
527 case HANDED_OVER_TO_NETWORK:
528 /* assert something? */
529 if (bio_data_dir(req->master_bio) == WRITE)
530 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
532 if (bio_data_dir(req->master_bio) == WRITE &&
533 !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
534 /* this is what is dangerous about protocol A:
535 * pretend it was successfully written on the peer. */
536 if (req->rq_state & RQ_NET_PENDING) {
537 dec_ap_pending(mdev);
538 req->rq_state &= ~RQ_NET_PENDING;
539 req->rq_state |= RQ_NET_OK;
540 } /* else: neg-ack was faster... */
541 /* it is still not yet RQ_NET_DONE until the
542 * corresponding epoch barrier got acked as well,
543 * so we know what to dirty on connection loss */
545 req->rq_state &= ~RQ_NET_QUEUED;
546 req->rq_state |= RQ_NET_SENT;
547 /* because _drbd_send_zc_bio could sleep, and may want to
548 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
549 * "COMPLETED_OK" events came in, once we return from
550 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
551 * whether it is done already, and end it. */
552 _req_may_be_done_not_susp(req, m);
555 case READ_RETRY_REMOTE_CANCELED:
556 req->rq_state &= ~RQ_NET_QUEUED;
557 /* fall through, in case we raced with drbd_disconnect */
558 case CONNECTION_LOST_WHILE_PENDING:
559 /* transfer log cleanup after connection loss */
560 /* assert something? */
561 if (req->rq_state & RQ_NET_PENDING)
562 dec_ap_pending(mdev);
564 p = !(req->rq_state & RQ_WRITE) && req->rq_state & RQ_NET_PENDING;
566 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
567 req->rq_state |= RQ_NET_DONE;
568 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
569 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
571 /* if it is still queued, we may not complete it here.
572 * it will be canceled soon. */
573 if (!(req->rq_state & RQ_NET_QUEUED)) {
575 goto goto_read_retry_local;
576 _req_may_be_done(req, m); /* Allowed while state.susp */
580 case WRITE_ACKED_BY_PEER_AND_SIS:
581 req->rq_state |= RQ_NET_SIS;
583 /* for discarded conflicting writes of multiple primaries,
584 * there is no need to keep anything in the tl, potential
585 * node crashes are covered by the activity log. */
586 req->rq_state |= RQ_NET_DONE;
588 case WRITE_ACKED_BY_PEER:
589 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
590 /* protocol C; successfully written on peer.
591 * Nothing to do here.
592 * We want to keep the tl in place for all protocols, to cater
593 * for volatile write-back caches on lower level devices.
595 * A barrier request is expected to have forced all prior
596 * requests onto stable storage, so completion of a barrier
597 * request could set NET_DONE right here, and not wait for the
598 * P_BARRIER_ACK, but that is an unnecessary optimization. */
601 /* this makes it effectively the same as for: */
602 case RECV_ACKED_BY_PEER:
603 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
604 /* protocol B; pretends to be successfully written on peer.
605 * see also notes above in HANDED_OVER_TO_NETWORK about
608 req->rq_state |= RQ_NET_OK;
609 D_ASSERT(req->rq_state & RQ_NET_PENDING);
610 dec_ap_pending(mdev);
611 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
612 req->rq_state &= ~RQ_NET_PENDING;
613 _req_may_be_done_not_susp(req, m);
617 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
618 /* If this node has already detected the write conflict, the
619 * worker will be waiting on misc_wait. Wake it up once this
620 * request has completed locally.
622 D_ASSERT(req->rq_state & RQ_NET_PENDING);
623 req->rq_state |= RQ_POSTPONED;
624 _req_may_be_done_not_susp(req, m);
628 /* assert something? */
629 if (req->rq_state & RQ_NET_PENDING) {
630 dec_ap_pending(mdev);
631 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
633 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
635 req->rq_state |= RQ_NET_DONE;
637 if (!(req->rq_state & RQ_WRITE))
638 goto goto_read_retry_local;
640 _req_may_be_done_not_susp(req, m);
641 /* else: done by HANDED_OVER_TO_NETWORK */
644 goto_read_retry_local:
645 if (!drbd_may_do_local_read(mdev, req->i.sector, req->i.size)) {
646 _req_may_be_done_not_susp(req, m);
649 D_ASSERT(!(req->rq_state & RQ_LOCAL_PENDING));
650 req->rq_state |= RQ_LOCAL_PENDING;
653 req->w.cb = w_restart_disk_io;
654 drbd_queue_work(&mdev->tconn->data.work, &req->w);
657 case FAIL_FROZEN_DISK_IO:
658 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
661 _req_may_be_done(req, m); /* Allowed while state.susp */
664 case RESTART_FROZEN_DISK_IO:
665 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
668 req->rq_state &= ~RQ_LOCAL_COMPLETED;
671 if (bio_data_dir(req->master_bio) == WRITE)
675 req->w.cb = w_restart_disk_io;
676 drbd_queue_work(&mdev->tconn->data.work, &req->w);
680 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
681 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
682 Trowing them out of the TL here by pretending we got a BARRIER_ACK
683 We ensure that the peer was not rebooted */
684 if (!(req->rq_state & RQ_NET_OK)) {
686 drbd_queue_work(&mdev->tconn->data.work, &req->w);
687 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
691 /* else, fall through to BARRIER_ACKED */
694 if (!(req->rq_state & RQ_WRITE))
697 if (req->rq_state & RQ_NET_PENDING) {
698 /* barrier came in before all requests were acked.
699 * this is bad, because if the connection is lost now,
700 * we won't be able to clean them up... */
701 dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
702 list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests);
704 if ((req->rq_state & RQ_NET_MASK) != 0) {
705 req->rq_state |= RQ_NET_DONE;
706 if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK)))
707 atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
709 _req_may_be_done(req, m); /* Allowed while state.susp */
713 D_ASSERT(req->rq_state & RQ_NET_PENDING);
714 dec_ap_pending(mdev);
715 req->rq_state &= ~RQ_NET_PENDING;
716 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
717 _req_may_be_done_not_susp(req, m);
724 /* we may do a local read if:
725 * - we are consistent (of course),
726 * - or we are generally inconsistent,
727 * BUT we are still/already IN SYNC for this area.
728 * since size may be bigger than BM_BLOCK_SIZE,
729 * we may need to check several bits.
731 static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
733 unsigned long sbnr, ebnr;
734 sector_t esector, nr_sectors;
736 if (mdev->state.disk == D_UP_TO_DATE)
738 if (mdev->state.disk != D_INCONSISTENT)
740 esector = sector + (size >> 9) - 1;
741 nr_sectors = drbd_get_capacity(mdev->this_bdev);
742 D_ASSERT(sector < nr_sectors);
743 D_ASSERT(esector < nr_sectors);
745 sbnr = BM_SECT_TO_BIT(sector);
746 ebnr = BM_SECT_TO_BIT(esector);
748 return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
751 static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector)
753 enum drbd_read_balancing rbm;
754 struct backing_dev_info *bdi;
757 if (mdev->state.pdsk < D_UP_TO_DATE)
761 rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
765 case RB_CONGESTED_REMOTE:
766 bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
767 return bdi_read_congested(bdi);
768 case RB_LEAST_PENDING:
769 return atomic_read(&mdev->local_cnt) >
770 atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt);
771 case RB_32K_STRIPING: /* stripe_shift = 15 */
772 case RB_64K_STRIPING:
773 case RB_128K_STRIPING:
774 case RB_256K_STRIPING:
775 case RB_512K_STRIPING:
776 case RB_1M_STRIPING: /* stripe_shift = 20 */
777 stripe_shift = (rbm - RB_32K_STRIPING + 15);
778 return (sector >> (stripe_shift - 9)) & 1;
780 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
781 case RB_PREFER_REMOTE:
783 case RB_PREFER_LOCAL:
790 * complete_conflicting_writes - wait for any conflicting write requests
792 * The write_requests tree contains all active write requests which we
793 * currently know about. Wait for any requests to complete which conflict with
796 static int complete_conflicting_writes(struct drbd_conf *mdev,
797 sector_t sector, int size)
800 struct drbd_interval *i;
803 i = drbd_find_overlap(&mdev->write_requests, sector, size);
806 err = drbd_wait_misc(mdev, i);
812 int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
814 const int rw = bio_rw(bio);
815 const int size = bio->bi_size;
816 const sector_t sector = bio->bi_sector;
817 struct drbd_tl_epoch *b = NULL;
818 struct drbd_request *req;
820 int local, remote, send_oos = 0;
824 /* allocate outside of all locks; */
825 req = drbd_req_new(mdev, bio);
828 /* only pass the error to the upper layers.
829 * if user cannot handle io errors, that's not our business. */
830 dev_err(DEV, "could not kmalloc() req\n");
831 bio_endio(bio, -ENOMEM);
834 req->start_time = start_time;
836 local = get_ldev(mdev);
838 bio_put(req->private_bio); /* or we get a bio leak */
839 req->private_bio = NULL;
846 if (!drbd_may_do_local_read(mdev, sector, size) ||
847 remote_due_to_read_balancing(mdev, sector)) {
848 /* we could kick the syncer to
849 * sync this extent asap, wait for
850 * it, then continue locally.
851 * Or just issue the request remotely.
854 bio_put(req->private_bio);
855 req->private_bio = NULL;
859 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
862 /* If we have a disk, but a READA request is mapped to remote,
863 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
864 * Just fail that READA request right here.
866 * THINK: maybe fail all READA when not local?
867 * or make this configurable...
868 * if network is slow, READA won't do any good.
870 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
872 goto fail_and_free_req;
875 /* For WRITES going to the local disk, grab a reference on the target
876 * extent. This waits for any resync activity in the corresponding
877 * resync extent to finish, and, if necessary, pulls in the target
878 * extent into the activity log, which involves further disk io because
879 * of transactional on-disk meta data updates. */
880 if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
881 req->rq_state |= RQ_IN_ACT_LOG;
882 drbd_al_begin_io(mdev, &req->i);
885 remote = remote && drbd_should_do_remote(mdev->state);
886 send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
887 D_ASSERT(!(remote && send_oos));
889 if (!(local || remote) && !drbd_suspended(mdev)) {
890 if (__ratelimit(&drbd_ratelimit_state))
891 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
893 goto fail_free_complete;
896 /* For WRITE request, we have to make sure that we have an
897 * unused_spare_tle, in case we need to start a new epoch.
898 * I try to be smart and avoid to pre-allocate always "just in case",
899 * but there is a race between testing the bit and pointer outside the
900 * spinlock, and grabbing the spinlock.
901 * if we lost that race, we retry. */
902 if (rw == WRITE && (remote || send_oos) &&
903 mdev->tconn->unused_spare_tle == NULL &&
904 test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
906 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
908 dev_err(DEV, "Failed to alloc barrier.\n");
910 goto fail_free_complete;
914 /* GOOD, everything prepared, grab the spin_lock */
915 spin_lock_irq(&mdev->tconn->req_lock);
918 err = complete_conflicting_writes(mdev, sector, size);
920 if (err != -ERESTARTSYS)
921 _conn_request_state(mdev->tconn,
924 spin_unlock_irq(&mdev->tconn->req_lock);
926 goto fail_free_complete;
930 if (drbd_suspended(mdev)) {
931 /* If we got suspended, use the retry mechanism of
932 generic_make_request() to restart processing of this
933 bio. In the next call to drbd_make_request
934 we sleep in inc_ap_bio() */
936 spin_unlock_irq(&mdev->tconn->req_lock);
937 goto fail_free_complete;
940 if (remote || send_oos) {
941 remote = drbd_should_do_remote(mdev->state);
942 send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
943 D_ASSERT(!(remote && send_oos));
945 if (!(remote || send_oos))
946 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
947 if (!(local || remote)) {
948 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
949 spin_unlock_irq(&mdev->tconn->req_lock);
951 goto fail_free_complete;
955 if (b && mdev->tconn->unused_spare_tle == NULL) {
956 mdev->tconn->unused_spare_tle = b;
959 if (rw == WRITE && (remote || send_oos) &&
960 mdev->tconn->unused_spare_tle == NULL &&
961 test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
962 /* someone closed the current epoch
963 * while we were grabbing the spinlock */
964 spin_unlock_irq(&mdev->tconn->req_lock);
965 goto allocate_barrier;
969 /* Update disk stats */
970 _drbd_start_io_acct(mdev, req, bio);
972 /* _maybe_start_new_epoch(mdev);
973 * If we need to generate a write barrier packet, we have to add the
974 * new epoch (barrier) object, and queue the barrier packet for sending,
975 * and queue the req's data after it _within the same lock_, otherwise
976 * we have race conditions were the reorder domains could be mixed up.
978 * Even read requests may start a new epoch and queue the corresponding
979 * barrier packet. To get the write ordering right, we only have to
980 * make sure that, if this is a write request and it triggered a
981 * barrier packet, this request is queued within the same spinlock. */
982 if ((remote || send_oos) && mdev->tconn->unused_spare_tle &&
983 test_and_clear_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
984 _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle);
985 mdev->tconn->unused_spare_tle = NULL;
987 D_ASSERT(!(remote && rw == WRITE &&
988 test_bit(CREATE_BARRIER, &mdev->tconn->flags)));
992 * Actually, 'local' may be wrong here already, since we may have failed
993 * to write to the meta data, and may become wrong anytime because of
994 * local io-error for some other request, which would lead to us
995 * "detaching" the local disk.
997 * 'remote' may become wrong any time because the network could fail.
999 * This is a harmless race condition, though, since it is handled
1000 * correctly at the appropriate places; so it just defers the failure
1001 * of the respective operation.
1004 /* mark them early for readability.
1005 * this just sets some state flags. */
1007 _req_mod(req, TO_BE_SENT);
1009 _req_mod(req, TO_BE_SUBMITTED);
1011 list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests);
1013 /* NOTE remote first: to get the concurrent write detection right,
1014 * we must register the request before start of local IO. */
1016 /* either WRITE and C_CONNECTED,
1017 * or READ, and no local disk,
1018 * or READ, but not in sync.
1020 _req_mod(req, (rw == WRITE)
1021 ? QUEUE_FOR_NET_WRITE
1022 : QUEUE_FOR_NET_READ);
1024 if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
1025 _req_mod(req, QUEUE_FOR_SEND_OOS);
1028 nc = rcu_dereference(mdev->tconn->net_conf);
1030 nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) {
1033 if (nc->cong_fill &&
1034 atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
1035 dev_info(DEV, "Congestion-fill threshold reached\n");
1039 if (mdev->act_log->used >= nc->cong_extents) {
1040 dev_info(DEV, "Congestion-extents threshold reached\n");
1045 queue_barrier(mdev); /* last barrier, after mirrored writes */
1047 if (nc->on_congestion == OC_PULL_AHEAD)
1048 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
1049 else /*nc->on_congestion == OC_DISCONNECT */
1050 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
1055 spin_unlock_irq(&mdev->tconn->req_lock);
1056 kfree(b); /* if someone else has beaten us to it... */
1059 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
1061 /* State may have changed since we grabbed our reference on the
1062 * mdev->ldev member. Double check, and short-circuit to endio.
1063 * In case the last activity log transaction failed to get on
1064 * stable storage, and this is a WRITE, we may not even submit
1066 if (get_ldev(mdev)) {
1067 if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
1068 : rw == READ ? DRBD_FAULT_DT_RD
1069 : DRBD_FAULT_DT_RA))
1070 bio_endio(req->private_bio, -EIO);
1072 generic_make_request(req->private_bio);
1075 bio_endio(req->private_bio, -EIO);
1081 if (req->rq_state & RQ_IN_ACT_LOG)
1082 drbd_al_complete_io(mdev, &req->i);
1085 bio_put(req->private_bio);
1086 req->private_bio = NULL;
1090 bio_endio(bio, err);
1099 int drbd_make_request(struct request_queue *q, struct bio *bio)
1101 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1102 unsigned long start_time;
1104 start_time = jiffies;
1107 * what we "blindly" assume:
1109 D_ASSERT(bio->bi_size > 0);
1110 D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
1113 return __drbd_make_request(mdev, bio, start_time);
1116 /* This is called by bio_add_page().
1118 * q->max_hw_sectors and other global limits are already enforced there.
1120 * We need to call down to our lower level device,
1121 * in case it has special restrictions.
1123 * We also may need to enforce configured max-bio-bvecs limits.
1125 * As long as the BIO is empty we have to allow at least one bvec,
1126 * regardless of size and offset, so no need to ask lower levels.
1128 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1130 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1131 unsigned int bio_size = bvm->bi_size;
1132 int limit = DRBD_MAX_BIO_SIZE;
1135 if (bio_size && get_ldev(mdev)) {
1136 struct request_queue * const b =
1137 mdev->ldev->backing_bdev->bd_disk->queue;
1138 if (b->merge_bvec_fn) {
1139 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1140 limit = min(limit, backing_limit);
1147 void request_timer_fn(unsigned long data)
1149 struct drbd_conf *mdev = (struct drbd_conf *) data;
1150 struct drbd_tconn *tconn = mdev->tconn;
1151 struct drbd_request *req; /* oldest request */
1152 struct list_head *le;
1153 struct net_conf *nc;
1154 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1157 nc = rcu_dereference(tconn->net_conf);
1158 ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0;
1160 if (get_ldev(mdev)) {
1161 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1166 et = min_not_zero(dt, ent);
1168 if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED))
1169 return; /* Recurring timer stopped */
1171 spin_lock_irq(&tconn->req_lock);
1172 le = &tconn->oldest_tle->requests;
1173 if (list_empty(le)) {
1174 spin_unlock_irq(&tconn->req_lock);
1175 mod_timer(&mdev->request_timer, jiffies + et);
1180 req = list_entry(le, struct drbd_request, tl_requests);
1181 if (ent && req->rq_state & RQ_NET_PENDING) {
1182 if (time_is_before_eq_jiffies(req->start_time + ent)) {
1183 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1184 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1187 if (dt && req->rq_state & RQ_LOCAL_PENDING) {
1188 if (time_is_before_eq_jiffies(req->start_time + dt)) {
1189 dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1190 __drbd_chk_io_error(mdev, 1);
1193 nt = (time_is_before_eq_jiffies(req->start_time + et) ? jiffies : req->start_time) + et;
1194 spin_unlock_irq(&tconn->req_lock);
1195 mod_timer(&mdev->request_timer, nt);