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 /* Update disk stats at start of I/O request */
35 static void _drbd_start_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
, struct bio
*bio
)
37 const int rw
= bio_data_dir(bio
);
39 cpu
= part_stat_lock();
40 part_round_stats(cpu
, &mdev
->vdisk
->part0
);
41 part_stat_inc(cpu
, &mdev
->vdisk
->part0
, ios
[rw
]);
42 part_stat_add(cpu
, &mdev
->vdisk
->part0
, sectors
[rw
], bio_sectors(bio
));
43 part_inc_in_flight(&mdev
->vdisk
->part0
, rw
);
47 /* Update disk stats when completing request upwards */
48 static void _drbd_end_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
)
50 int rw
= bio_data_dir(req
->master_bio
);
51 unsigned long duration
= jiffies
- req
->start_time
;
53 cpu
= part_stat_lock();
54 part_stat_add(cpu
, &mdev
->vdisk
->part0
, ticks
[rw
], duration
);
55 part_round_stats(cpu
, &mdev
->vdisk
->part0
);
56 part_dec_in_flight(&mdev
->vdisk
->part0
, rw
);
60 static void _req_is_done(struct drbd_conf
*mdev
, struct drbd_request
*req
, const int rw
)
62 const unsigned long s
= req
->rq_state
;
64 /* remove it from the transfer log.
65 * well, only if it had been there in the first
66 * place... if it had not (local only or conflicting
67 * and never sent), it should still be "empty" as
68 * initialized in drbd_req_new(), so we can list_del() it
69 * here unconditionally */
70 list_del(&req
->tl_requests
);
72 /* if it was a write, we may have to set the corresponding
73 * bit(s) out-of-sync first. If it had a local part, we need to
74 * release the reference to the activity log. */
76 /* Set out-of-sync unless both OK flags are set
77 * (local only or remote failed).
78 * Other places where we set out-of-sync:
79 * READ with local io-error */
80 if (!(s
& RQ_NET_OK
) || !(s
& RQ_LOCAL_OK
))
81 drbd_set_out_of_sync(mdev
, req
->sector
, req
->size
);
83 if ((s
& RQ_NET_OK
) && (s
& RQ_LOCAL_OK
) && (s
& RQ_NET_SIS
))
84 drbd_set_in_sync(mdev
, req
->sector
, req
->size
);
86 /* one might be tempted to move the drbd_al_complete_io
87 * to the local io completion callback drbd_endio_pri.
88 * but, if this was a mirror write, we may only
89 * drbd_al_complete_io after this is RQ_NET_DONE,
90 * otherwise the extent could be dropped from the al
91 * before it has actually been written on the peer.
92 * if we crash before our peer knows about the request,
93 * but after the extent has been dropped from the al,
94 * we would forget to resync the corresponding extent.
96 if (s
& RQ_LOCAL_MASK
) {
97 if (get_ldev_if_state(mdev
, D_FAILED
)) {
98 if (s
& RQ_IN_ACT_LOG
)
99 drbd_al_complete_io(mdev
, req
->sector
);
101 } else if (__ratelimit(&drbd_ratelimit_state
)) {
102 dev_warn(DEV
, "Should have called drbd_al_complete_io(, %llu), "
103 "but my Disk seems to have failed :(\n",
104 (unsigned long long) req
->sector
);
112 static void queue_barrier(struct drbd_conf
*mdev
)
114 struct drbd_tl_epoch
*b
;
116 /* We are within the req_lock. Once we queued the barrier for sending,
117 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
118 * barrier/epoch object is added. This is the only place this bit is
119 * set. It indicates that the barrier for this epoch is already queued,
120 * and no new epoch has been created yet. */
121 if (test_bit(CREATE_BARRIER
, &mdev
->flags
))
124 b
= mdev
->newest_tle
;
125 b
->w
.cb
= w_send_barrier
;
126 /* inc_ap_pending done here, so we won't
127 * get imbalanced on connection loss.
128 * dec_ap_pending will be done in got_BarrierAck
129 * or (on connection loss) in tl_clear. */
130 inc_ap_pending(mdev
);
131 drbd_queue_work(&mdev
->data
.work
, &b
->w
);
132 set_bit(CREATE_BARRIER
, &mdev
->flags
);
135 static void _about_to_complete_local_write(struct drbd_conf
*mdev
,
136 struct drbd_request
*req
)
138 const unsigned long s
= req
->rq_state
;
139 struct drbd_request
*i
;
140 struct drbd_epoch_entry
*e
;
141 struct hlist_node
*n
;
142 struct hlist_head
*slot
;
144 /* Before we can signal completion to the upper layers,
145 * we may need to close the current epoch.
146 * We can skip this, if this request has not even been sent, because we
147 * did not have a fully established connection yet/anymore, during
148 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
150 if (mdev
->state
.conn
>= C_CONNECTED
&&
151 (s
& RQ_NET_SENT
) != 0 &&
152 req
->epoch
== mdev
->newest_tle
->br_number
)
155 /* we need to do the conflict detection stuff,
156 * if we have the ee_hash (two_primaries) and
157 * this has been on the network */
158 if ((s
& RQ_NET_DONE
) && mdev
->ee_hash
!= NULL
) {
159 const sector_t sector
= req
->sector
;
160 const int size
= req
->size
;
163 * there must be no conflicting requests, since
164 * they must have been failed on the spot */
165 #define OVERLAPS overlaps(sector, size, i->sector, i->size)
166 slot
= tl_hash_slot(mdev
, sector
);
167 hlist_for_each_entry(i
, n
, slot
, collision
) {
169 dev_alert(DEV
, "LOGIC BUG: completed: %p %llus +%u; "
170 "other: %p %llus +%u\n",
171 req
, (unsigned long long)sector
, size
,
172 i
, (unsigned long long)i
->sector
, i
->size
);
176 /* maybe "wake" those conflicting epoch entries
177 * that wait for this request to finish.
179 * currently, there can be only _one_ such ee
180 * (well, or some more, which would be pending
181 * P_DISCARD_ACK not yet sent by the asender...),
182 * since we block the receiver thread upon the
183 * first conflict detection, which will wait on
184 * misc_wait. maybe we want to assert that?
186 * anyways, if we found one,
187 * we just have to do a wake_up. */
189 #define OVERLAPS overlaps(sector, size, e->sector, e->size)
190 slot
= ee_hash_slot(mdev
, req
->sector
);
191 hlist_for_each_entry(e
, n
, slot
, collision
) {
193 wake_up(&mdev
->misc_wait
);
201 void complete_master_bio(struct drbd_conf
*mdev
,
202 struct bio_and_error
*m
)
204 bio_endio(m
->bio
, m
->error
);
208 /* Helper for __req_mod().
209 * Set m->bio to the master bio, if it is fit to be completed,
210 * or leave it alone (it is initialized to NULL in __req_mod),
211 * if it has already been completed, or cannot be completed yet.
212 * If m->bio is set, the error status to be returned is placed in m->error.
214 void _req_may_be_done(struct drbd_request
*req
, struct bio_and_error
*m
)
216 const unsigned long s
= req
->rq_state
;
217 struct drbd_conf
*mdev
= req
->mdev
;
218 int rw
= req
->rq_state
& RQ_WRITE
? WRITE
: READ
;
220 /* we must not complete the master bio, while it is
221 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
222 * not yet acknowledged by the peer
223 * not yet completed by the local io subsystem
224 * these flags may get cleared in any order by
227 * the bio_endio completion callbacks.
229 if (s
& RQ_NET_QUEUED
)
231 if (s
& RQ_NET_PENDING
)
233 if (s
& RQ_LOCAL_PENDING
&& !(s
& RQ_LOCAL_ABORTED
))
236 if (req
->master_bio
) {
237 /* this is data_received (remote read)
238 * or protocol C P_WRITE_ACK
239 * or protocol B P_RECV_ACK
240 * or protocol A "handed_over_to_network" (SendAck)
241 * or canceled or failed,
242 * or killed from the transfer log due to connection loss.
246 * figure out whether to report success or failure.
248 * report success when at least one of the operations succeeded.
249 * or, to put the other way,
250 * only report failure, when both operations failed.
252 * what to do about the failures is handled elsewhere.
253 * what we need to do here is just: complete the master_bio.
255 * local completion error, if any, has been stored as ERR_PTR
256 * in private_bio within drbd_endio_pri.
258 int ok
= (s
& RQ_LOCAL_OK
) || (s
& RQ_NET_OK
);
259 int error
= PTR_ERR(req
->private_bio
);
261 /* remove the request from the conflict detection
262 * respective block_id verification hash */
263 if (!hlist_unhashed(&req
->collision
))
264 hlist_del(&req
->collision
);
266 D_ASSERT((s
& (RQ_NET_MASK
& ~RQ_NET_DONE
)) == 0);
268 /* for writes we need to do some extra housekeeping */
270 _about_to_complete_local_write(mdev
, req
);
272 /* Update disk stats */
273 _drbd_end_io_acct(mdev
, req
);
275 m
->error
= ok
? 0 : (error
?: -EIO
);
276 m
->bio
= req
->master_bio
;
277 req
->master_bio
= NULL
;
280 if (s
& RQ_LOCAL_PENDING
)
283 if ((s
& RQ_NET_MASK
) == 0 || (s
& RQ_NET_DONE
)) {
284 /* this is disconnected (local only) operation,
285 * or protocol C P_WRITE_ACK,
286 * or protocol A or B P_BARRIER_ACK,
287 * or killed from the transfer log due to connection loss. */
288 _req_is_done(mdev
, req
, rw
);
290 /* else: network part and not DONE yet. that is
291 * protocol A or B, barrier ack still pending... */
294 static void _req_may_be_done_not_susp(struct drbd_request
*req
, struct bio_and_error
*m
)
296 struct drbd_conf
*mdev
= req
->mdev
;
298 if (!is_susp(mdev
->state
))
299 _req_may_be_done(req
, m
);
303 * checks whether there was an overlapping request
304 * or ee already registered.
306 * if so, return 1, in which case this request is completed on the spot,
307 * without ever being submitted or send.
309 * return 0 if it is ok to submit this request.
312 * paranoia: assume something above us is broken, and issues different write
313 * requests for the same block simultaneously...
315 * To ensure these won't be reordered differently on both nodes, resulting in
316 * diverging data sets, we discard the later one(s). Not that this is supposed
317 * to happen, but this is the rationale why we also have to check for
318 * conflicting requests with local origin, and why we have to do so regardless
319 * of whether we allowed multiple primaries.
321 * BTW, in case we only have one primary, the ee_hash is empty anyways, and the
322 * second hlist_for_each_entry becomes a noop. This is even simpler than to
323 * grab a reference on the net_conf, and check for the two_primaries flag...
325 static int _req_conflicts(struct drbd_request
*req
)
327 struct drbd_conf
*mdev
= req
->mdev
;
328 const sector_t sector
= req
->sector
;
329 const int size
= req
->size
;
330 struct drbd_request
*i
;
331 struct drbd_epoch_entry
*e
;
332 struct hlist_node
*n
;
333 struct hlist_head
*slot
;
335 D_ASSERT(hlist_unhashed(&req
->collision
));
337 if (!get_net_conf(mdev
))
341 ERR_IF (mdev
->tl_hash_s
== 0)
342 goto out_no_conflict
;
343 BUG_ON(mdev
->tl_hash
== NULL
);
345 #define OVERLAPS overlaps(i->sector, i->size, sector, size)
346 slot
= tl_hash_slot(mdev
, sector
);
347 hlist_for_each_entry(i
, n
, slot
, collision
) {
349 dev_alert(DEV
, "%s[%u] Concurrent local write detected! "
350 "[DISCARD L] new: %llus +%u; "
351 "pending: %llus +%u\n",
352 current
->comm
, current
->pid
,
353 (unsigned long long)sector
, size
,
354 (unsigned long long)i
->sector
, i
->size
);
359 if (mdev
->ee_hash_s
) {
360 /* now, check for overlapping requests with remote origin */
361 BUG_ON(mdev
->ee_hash
== NULL
);
363 #define OVERLAPS overlaps(e->sector, e->size, sector, size)
364 slot
= ee_hash_slot(mdev
, sector
);
365 hlist_for_each_entry(e
, n
, slot
, collision
) {
367 dev_alert(DEV
, "%s[%u] Concurrent remote write detected!"
368 " [DISCARD L] new: %llus +%u; "
369 "pending: %llus +%u\n",
370 current
->comm
, current
->pid
,
371 (unsigned long long)sector
, size
,
372 (unsigned long long)e
->sector
, e
->size
);
380 /* this is like it should be, and what we expected.
381 * our users do behave after all... */
390 /* obviously this could be coded as many single functions
391 * instead of one huge switch,
392 * or by putting the code directly in the respective locations
393 * (as it has been before).
395 * but having it this way
396 * enforces that it is all in this one place, where it is easier to audit,
397 * it makes it obvious that whatever "event" "happens" to a request should
398 * happen "atomically" within the req_lock,
399 * and it enforces that we have to think in a very structured manner
400 * about the "events" that may happen to a request during its life time ...
402 int __req_mod(struct drbd_request
*req
, enum drbd_req_event what
,
403 struct bio_and_error
*m
)
405 struct drbd_conf
*mdev
= req
->mdev
;
411 dev_err(DEV
, "LOGIC BUG in %s:%u\n", __FILE__
, __LINE__
);
414 /* does not happen...
415 * initialization done in drbd_req_new
420 case to_be_send
: /* via network */
421 /* reached via drbd_make_request_common
422 * and from w_read_retry_remote */
423 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
424 req
->rq_state
|= RQ_NET_PENDING
;
425 inc_ap_pending(mdev
);
428 case to_be_submitted
: /* locally */
429 /* reached via drbd_make_request_common */
430 D_ASSERT(!(req
->rq_state
& RQ_LOCAL_MASK
));
431 req
->rq_state
|= RQ_LOCAL_PENDING
;
435 if (req
->rq_state
& RQ_WRITE
)
436 mdev
->writ_cnt
+= req
->size
>>9;
438 mdev
->read_cnt
+= req
->size
>>9;
440 req
->rq_state
|= (RQ_LOCAL_COMPLETED
|RQ_LOCAL_OK
);
441 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
443 _req_may_be_done_not_susp(req
, m
);
447 req
->rq_state
|= RQ_LOCAL_ABORTED
;
448 if (req
->rq_state
& RQ_WRITE
)
449 _req_may_be_done_not_susp(req
, m
);
451 goto goto_queue_for_net_read
;
454 case write_completed_with_error
:
455 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
456 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
458 __drbd_chk_io_error(mdev
, DRBD_IO_ERROR
);
459 _req_may_be_done_not_susp(req
, m
);
462 case read_ahead_completed_with_error
:
463 /* it is legal to fail READA */
464 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
465 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
466 _req_may_be_done_not_susp(req
, m
);
469 case read_completed_with_error
:
470 drbd_set_out_of_sync(mdev
, req
->sector
, req
->size
);
472 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
473 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
475 if (req
->rq_state
& RQ_LOCAL_ABORTED
) {
476 _req_may_be_done(req
, m
);
480 __drbd_chk_io_error(mdev
, DRBD_IO_ERROR
);
482 goto_queue_for_net_read
:
484 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
486 /* no point in retrying if there is no good remote data,
487 * or we have no connection. */
488 if (mdev
->state
.pdsk
!= D_UP_TO_DATE
) {
489 _req_may_be_done_not_susp(req
, m
);
493 /* _req_mod(req,to_be_send); oops, recursion... */
494 req
->rq_state
|= RQ_NET_PENDING
;
495 inc_ap_pending(mdev
);
496 /* fall through: _req_mod(req,queue_for_net_read); */
498 case queue_for_net_read
:
499 /* READ or READA, and
501 * or target area marked as invalid,
502 * or just got an io-error. */
503 /* from drbd_make_request_common
504 * or from bio_endio during read io-error recovery */
506 /* so we can verify the handle in the answer packet
507 * corresponding hlist_del is in _req_may_be_done() */
508 hlist_add_head(&req
->collision
, ar_hash_slot(mdev
, req
->sector
));
510 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
512 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
513 req
->rq_state
|= RQ_NET_QUEUED
;
514 req
->w
.cb
= (req
->rq_state
& RQ_LOCAL_MASK
)
515 ? w_read_retry_remote
517 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
520 case queue_for_net_write
:
521 /* assert something? */
522 /* from drbd_make_request_common only */
524 hlist_add_head(&req
->collision
, tl_hash_slot(mdev
, req
->sector
));
525 /* corresponding hlist_del is in _req_may_be_done() */
528 * In case the req ended up on the transfer log before being
529 * queued on the worker, it could lead to this request being
530 * missed during cleanup after connection loss.
531 * So we have to do both operations here,
532 * within the same lock that protects the transfer log.
534 * _req_add_to_epoch(req); this has to be after the
535 * _maybe_start_new_epoch(req); which happened in
536 * drbd_make_request_common, because we now may set the bit
537 * again ourselves to close the current epoch.
539 * Add req to the (now) current epoch (barrier). */
541 /* otherwise we may lose an unplug, which may cause some remote
542 * io-scheduler timeout to expire, increasing maximum latency,
543 * hurting performance. */
544 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
546 /* see drbd_make_request_common,
547 * just after it grabs the req_lock */
548 D_ASSERT(test_bit(CREATE_BARRIER
, &mdev
->flags
) == 0);
550 req
->epoch
= mdev
->newest_tle
->br_number
;
552 /* increment size of current epoch */
553 mdev
->newest_tle
->n_writes
++;
555 /* queue work item to send data */
556 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
557 req
->rq_state
|= RQ_NET_QUEUED
;
558 req
->w
.cb
= w_send_dblock
;
559 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
561 /* close the epoch, in case it outgrew the limit */
562 if (mdev
->newest_tle
->n_writes
>= mdev
->net_conf
->max_epoch_size
)
567 case queue_for_send_oos
:
568 req
->rq_state
|= RQ_NET_QUEUED
;
569 req
->w
.cb
= w_send_oos
;
570 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
573 case read_retry_remote_canceled
:
576 /* real cleanup will be done from tl_clear. just update flags
577 * so it is no longer marked as on the worker queue */
578 req
->rq_state
&= ~RQ_NET_QUEUED
;
579 /* if we did it right, tl_clear should be scheduled only after
580 * this, so this should not be necessary! */
581 _req_may_be_done_not_susp(req
, m
);
584 case handed_over_to_network
:
585 /* assert something? */
586 if (bio_data_dir(req
->master_bio
) == WRITE
)
587 atomic_add(req
->size
>>9, &mdev
->ap_in_flight
);
589 if (bio_data_dir(req
->master_bio
) == WRITE
&&
590 mdev
->net_conf
->wire_protocol
== DRBD_PROT_A
) {
591 /* this is what is dangerous about protocol A:
592 * pretend it was successfully written on the peer. */
593 if (req
->rq_state
& RQ_NET_PENDING
) {
594 dec_ap_pending(mdev
);
595 req
->rq_state
&= ~RQ_NET_PENDING
;
596 req
->rq_state
|= RQ_NET_OK
;
597 } /* else: neg-ack was faster... */
598 /* it is still not yet RQ_NET_DONE until the
599 * corresponding epoch barrier got acked as well,
600 * so we know what to dirty on connection loss */
602 req
->rq_state
&= ~RQ_NET_QUEUED
;
603 req
->rq_state
|= RQ_NET_SENT
;
604 _req_may_be_done_not_susp(req
, m
);
607 case oos_handed_to_network
:
608 /* Was not set PENDING, no longer QUEUED, so is now DONE
609 * as far as this connection is concerned. */
610 req
->rq_state
&= ~RQ_NET_QUEUED
;
611 req
->rq_state
|= RQ_NET_DONE
;
612 _req_may_be_done_not_susp(req
, m
);
615 case connection_lost_while_pending
:
616 /* transfer log cleanup after connection loss */
617 /* assert something? */
618 if (req
->rq_state
& RQ_NET_PENDING
)
619 dec_ap_pending(mdev
);
620 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
621 req
->rq_state
|= RQ_NET_DONE
;
622 if (req
->rq_state
& RQ_NET_SENT
&& req
->rq_state
& RQ_WRITE
)
623 atomic_sub(req
->size
>>9, &mdev
->ap_in_flight
);
625 /* if it is still queued, we may not complete it here.
626 * it will be canceled soon. */
627 if (!(req
->rq_state
& RQ_NET_QUEUED
))
628 _req_may_be_done(req
, m
); /* Allowed while state.susp */
631 case conflict_discarded_by_peer
:
632 /* for discarded conflicting writes of multiple primaries,
633 * there is no need to keep anything in the tl, potential
634 * node crashes are covered by the activity log. */
635 if (what
== conflict_discarded_by_peer
)
636 dev_alert(DEV
, "Got DiscardAck packet %llus +%u!"
637 " DRBD is not a random data generator!\n",
638 (unsigned long long)req
->sector
, req
->size
);
639 req
->rq_state
|= RQ_NET_DONE
;
641 case write_acked_by_peer_and_sis
:
642 case write_acked_by_peer
:
643 if (what
== write_acked_by_peer_and_sis
)
644 req
->rq_state
|= RQ_NET_SIS
;
645 /* protocol C; successfully written on peer.
646 * Nothing more to do here.
647 * We want to keep the tl in place for all protocols, to cater
648 * for volatile write-back caches on lower level devices. */
650 case recv_acked_by_peer
:
651 /* protocol B; pretends to be successfully written on peer.
652 * see also notes above in handed_over_to_network about
654 req
->rq_state
|= RQ_NET_OK
;
655 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
656 dec_ap_pending(mdev
);
657 atomic_sub(req
->size
>>9, &mdev
->ap_in_flight
);
658 req
->rq_state
&= ~RQ_NET_PENDING
;
659 _req_may_be_done_not_susp(req
, m
);
663 /* assert something? */
664 if (req
->rq_state
& RQ_NET_PENDING
) {
665 dec_ap_pending(mdev
);
666 atomic_sub(req
->size
>>9, &mdev
->ap_in_flight
);
668 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
670 req
->rq_state
|= RQ_NET_DONE
;
671 _req_may_be_done_not_susp(req
, m
);
672 /* else: done by handed_over_to_network */
675 case fail_frozen_disk_io
:
676 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
679 _req_may_be_done(req
, m
); /* Allowed while state.susp */
682 case restart_frozen_disk_io
:
683 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
686 req
->rq_state
&= ~RQ_LOCAL_COMPLETED
;
689 if (bio_data_dir(req
->master_bio
) == WRITE
)
693 req
->w
.cb
= w_restart_disk_io
;
694 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
698 /* Simply complete (local only) READs. */
699 if (!(req
->rq_state
& RQ_WRITE
) && !req
->w
.cb
) {
700 _req_may_be_done(req
, m
);
704 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
705 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
706 Trowing them out of the TL here by pretending we got a BARRIER_ACK
707 We ensure that the peer was not rebooted */
708 if (!(req
->rq_state
& RQ_NET_OK
)) {
710 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
711 rv
= req
->rq_state
& RQ_WRITE
? MR_WRITE
: MR_READ
;
715 /* else, fall through to barrier_acked */
718 if (!(req
->rq_state
& RQ_WRITE
))
721 if (req
->rq_state
& RQ_NET_PENDING
) {
722 /* barrier came in before all requests have been acked.
723 * this is bad, because if the connection is lost now,
724 * we won't be able to clean them up... */
725 dev_err(DEV
, "FIXME (barrier_acked but pending)\n");
726 list_move(&req
->tl_requests
, &mdev
->out_of_sequence_requests
);
728 if ((req
->rq_state
& RQ_NET_MASK
) != 0) {
729 req
->rq_state
|= RQ_NET_DONE
;
730 if (mdev
->net_conf
->wire_protocol
== DRBD_PROT_A
)
731 atomic_sub(req
->size
>>9, &mdev
->ap_in_flight
);
733 _req_may_be_done(req
, m
); /* Allowed while state.susp */
737 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
738 dec_ap_pending(mdev
);
739 req
->rq_state
&= ~RQ_NET_PENDING
;
740 req
->rq_state
|= (RQ_NET_OK
|RQ_NET_DONE
);
741 _req_may_be_done_not_susp(req
, m
);
748 /* we may do a local read if:
749 * - we are consistent (of course),
750 * - or we are generally inconsistent,
751 * BUT we are still/already IN SYNC for this area.
752 * since size may be bigger than BM_BLOCK_SIZE,
753 * we may need to check several bits.
755 static int drbd_may_do_local_read(struct drbd_conf
*mdev
, sector_t sector
, int size
)
757 unsigned long sbnr
, ebnr
;
758 sector_t esector
, nr_sectors
;
760 if (mdev
->state
.disk
== D_UP_TO_DATE
)
762 if (mdev
->state
.disk
>= D_OUTDATED
)
764 if (mdev
->state
.disk
< D_INCONSISTENT
)
766 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
767 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
768 esector
= sector
+ (size
>> 9) - 1;
770 D_ASSERT(sector
< nr_sectors
);
771 D_ASSERT(esector
< nr_sectors
);
773 sbnr
= BM_SECT_TO_BIT(sector
);
774 ebnr
= BM_SECT_TO_BIT(esector
);
776 return 0 == drbd_bm_count_bits(mdev
, sbnr
, ebnr
);
779 static void maybe_pull_ahead(struct drbd_conf
*mdev
)
783 /* If I don't even have good local storage, we can not reasonably try
784 * to pull ahead of the peer. We also need the local reference to make
785 * sure mdev->act_log is there.
786 * Note: caller has to make sure that net_conf is there.
788 if (!get_ldev_if_state(mdev
, D_UP_TO_DATE
))
791 if (mdev
->net_conf
->cong_fill
&&
792 atomic_read(&mdev
->ap_in_flight
) >= mdev
->net_conf
->cong_fill
) {
793 dev_info(DEV
, "Congestion-fill threshold reached\n");
797 if (mdev
->act_log
->used
>= mdev
->net_conf
->cong_extents
) {
798 dev_info(DEV
, "Congestion-extents threshold reached\n");
803 queue_barrier(mdev
); /* last barrier, after mirrored writes */
805 if (mdev
->net_conf
->on_congestion
== OC_PULL_AHEAD
)
806 _drbd_set_state(_NS(mdev
, conn
, C_AHEAD
), 0, NULL
);
807 else /*mdev->net_conf->on_congestion == OC_DISCONNECT */
808 _drbd_set_state(_NS(mdev
, conn
, C_DISCONNECTING
), 0, NULL
);
813 static int drbd_make_request_common(struct drbd_conf
*mdev
, struct bio
*bio
, unsigned long start_time
)
815 const int rw
= bio_rw(bio
);
816 const int size
= bio
->bi_size
;
817 const sector_t sector
= bio
->bi_sector
;
818 struct drbd_tl_epoch
*b
= NULL
;
819 struct drbd_request
*req
;
820 int local
, remote
, send_oos
= 0;
825 /* allocate outside of all locks; */
826 req
= drbd_req_new(mdev
, bio
);
829 /* only pass the error to the upper layers.
830 * if user cannot handle io errors, that's not our business. */
831 dev_err(DEV
, "could not kmalloc() req\n");
832 bio_endio(bio
, -ENOMEM
);
835 req
->start_time
= start_time
;
837 local
= get_ldev(mdev
);
839 bio_put(req
->private_bio
); /* or we get a bio leak */
840 req
->private_bio
= NULL
;
843 /* Need to replicate writes. Unless it is an empty flush,
844 * which is better mapped to a DRBD P_BARRIER packet,
845 * also for drbd wire protocol compatibility reasons. */
846 if (unlikely(size
== 0)) {
847 /* The only size==0 bios we expect are empty flushes. */
848 D_ASSERT(bio
->bi_rw
& REQ_FLUSH
);
855 if (!drbd_may_do_local_read(mdev
, sector
, size
)) {
856 /* we could kick the syncer to
857 * sync this extent asap, wait for
858 * it, then continue locally.
859 * Or just issue the request remotely.
862 bio_put(req
->private_bio
);
863 req
->private_bio
= NULL
;
867 remote
= !local
&& mdev
->state
.pdsk
>= D_UP_TO_DATE
;
870 /* If we have a disk, but a READA request is mapped to remote,
871 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
872 * Just fail that READA request right here.
874 * THINK: maybe fail all READA when not local?
875 * or make this configurable...
876 * if network is slow, READA won't do any good.
878 if (rw
== READA
&& mdev
->state
.disk
>= D_INCONSISTENT
&& !local
) {
880 goto fail_and_free_req
;
883 /* For WRITES going to the local disk, grab a reference on the target
884 * extent. This waits for any resync activity in the corresponding
885 * resync extent to finish, and, if necessary, pulls in the target
886 * extent into the activity log, which involves further disk io because
887 * of transactional on-disk meta data updates.
888 * Empty flushes don't need to go into the activity log, they can only
889 * flush data for pending writes which are already in there. */
890 if (rw
== WRITE
&& local
&& size
891 && !test_bit(AL_SUSPENDED
, &mdev
->flags
)) {
892 req
->rq_state
|= RQ_IN_ACT_LOG
;
893 drbd_al_begin_io(mdev
, sector
);
897 remote
= remote
&& drbd_should_do_remote(s
);
898 send_oos
= rw
== WRITE
&& drbd_should_send_oos(s
);
899 D_ASSERT(!(remote
&& send_oos
));
901 if (!(local
|| remote
) && !is_susp(mdev
->state
)) {
902 if (__ratelimit(&drbd_ratelimit_state
))
903 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
904 goto fail_free_complete
;
907 /* For WRITE request, we have to make sure that we have an
908 * unused_spare_tle, in case we need to start a new epoch.
909 * I try to be smart and avoid to pre-allocate always "just in case",
910 * but there is a race between testing the bit and pointer outside the
911 * spinlock, and grabbing the spinlock.
912 * if we lost that race, we retry. */
913 if (rw
== WRITE
&& (remote
|| send_oos
) &&
914 mdev
->unused_spare_tle
== NULL
&&
915 test_bit(CREATE_BARRIER
, &mdev
->flags
)) {
917 b
= kmalloc(sizeof(struct drbd_tl_epoch
), GFP_NOIO
);
919 dev_err(DEV
, "Failed to alloc barrier.\n");
921 goto fail_free_complete
;
925 /* GOOD, everything prepared, grab the spin_lock */
926 spin_lock_irq(&mdev
->req_lock
);
928 if (is_susp(mdev
->state
)) {
929 /* If we got suspended, use the retry mechanism of
930 drbd_make_request() to restart processing of this
931 bio. In the next call to drbd_make_request
932 we sleep in inc_ap_bio() */
934 spin_unlock_irq(&mdev
->req_lock
);
935 goto fail_free_complete
;
938 if (remote
|| send_oos
) {
939 remote
= drbd_should_do_remote(mdev
->state
);
940 send_oos
= rw
== WRITE
&& drbd_should_send_oos(mdev
->state
);
941 D_ASSERT(!(remote
&& send_oos
));
943 if (!(remote
|| send_oos
))
944 dev_warn(DEV
, "lost connection while grabbing the req_lock!\n");
945 if (!(local
|| remote
)) {
946 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
947 spin_unlock_irq(&mdev
->req_lock
);
948 goto fail_free_complete
;
952 if (b
&& mdev
->unused_spare_tle
== NULL
) {
953 mdev
->unused_spare_tle
= b
;
956 if (rw
== WRITE
&& (remote
|| send_oos
) &&
957 mdev
->unused_spare_tle
== NULL
&&
958 test_bit(CREATE_BARRIER
, &mdev
->flags
)) {
959 /* someone closed the current epoch
960 * while we were grabbing the spinlock */
961 spin_unlock_irq(&mdev
->req_lock
);
962 goto allocate_barrier
;
966 /* Update disk stats */
967 _drbd_start_io_acct(mdev
, req
, bio
);
969 /* _maybe_start_new_epoch(mdev);
970 * If we need to generate a write barrier packet, we have to add the
971 * new epoch (barrier) object, and queue the barrier packet for sending,
972 * and queue the req's data after it _within the same lock_, otherwise
973 * we have race conditions were the reorder domains could be mixed up.
975 * Even read requests may start a new epoch and queue the corresponding
976 * barrier packet. To get the write ordering right, we only have to
977 * make sure that, if this is a write request and it triggered a
978 * barrier packet, this request is queued within the same spinlock. */
979 if ((remote
|| send_oos
) && mdev
->unused_spare_tle
&&
980 test_and_clear_bit(CREATE_BARRIER
, &mdev
->flags
)) {
981 _tl_add_barrier(mdev
, mdev
->unused_spare_tle
);
982 mdev
->unused_spare_tle
= NULL
;
984 D_ASSERT(!(remote
&& rw
== WRITE
&&
985 test_bit(CREATE_BARRIER
, &mdev
->flags
)));
989 * Actually, 'local' may be wrong here already, since we may have failed
990 * to write to the meta data, and may become wrong anytime because of
991 * local io-error for some other request, which would lead to us
992 * "detaching" the local disk.
994 * 'remote' may become wrong any time because the network could fail.
996 * This is a harmless race condition, though, since it is handled
997 * correctly at the appropriate places; so it just defers the failure
998 * of the respective operation.
1001 /* mark them early for readability.
1002 * this just sets some state flags. */
1004 _req_mod(req
, to_be_send
);
1006 _req_mod(req
, to_be_submitted
);
1008 /* check this request on the collision detection hash tables.
1009 * if we have a conflict, just complete it here.
1010 * THINK do we want to check reads, too? (I don't think so...) */
1011 if (rw
== WRITE
&& _req_conflicts(req
))
1012 goto fail_conflicting
;
1014 /* no point in adding empty flushes to the transfer log,
1015 * they are mapped to drbd barriers already. */
1016 if (likely(size
!=0))
1017 list_add_tail(&req
->tl_requests
, &mdev
->newest_tle
->requests
);
1019 /* NOTE remote first: to get the concurrent write detection right,
1020 * we must register the request before start of local IO. */
1022 /* either WRITE and C_CONNECTED,
1023 * or READ, and no local disk,
1024 * or READ, but not in sync.
1026 _req_mod(req
, (rw
== WRITE
)
1027 ? queue_for_net_write
1028 : queue_for_net_read
);
1030 if (send_oos
&& drbd_set_out_of_sync(mdev
, sector
, size
))
1031 _req_mod(req
, queue_for_send_oos
);
1034 mdev
->net_conf
->on_congestion
!= OC_BLOCK
&& mdev
->agreed_pro_version
>= 96)
1035 maybe_pull_ahead(mdev
);
1037 /* If this was a flush, queue a drbd barrier/start a new epoch.
1038 * Unless the current epoch was empty anyways, or we are not currently
1039 * replicating, in which case there is no point. */
1040 if (unlikely(bio
->bi_rw
& REQ_FLUSH
)
1041 && mdev
->newest_tle
->n_writes
1042 && drbd_should_do_remote(mdev
->state
))
1043 queue_barrier(mdev
);
1045 spin_unlock_irq(&mdev
->req_lock
);
1046 kfree(b
); /* if someone else has beaten us to it... */
1049 req
->private_bio
->bi_bdev
= mdev
->ldev
->backing_bdev
;
1051 /* State may have changed since we grabbed our reference on the
1052 * mdev->ldev member. Double check, and short-circuit to endio.
1053 * In case the last activity log transaction failed to get on
1054 * stable storage, and this is a WRITE, we may not even submit
1056 if (get_ldev(mdev
)) {
1057 if (drbd_insert_fault(mdev
, rw
== WRITE
? DRBD_FAULT_DT_WR
1058 : rw
== READ
? DRBD_FAULT_DT_RD
1059 : DRBD_FAULT_DT_RA
))
1060 bio_endio(req
->private_bio
, -EIO
);
1062 generic_make_request(req
->private_bio
);
1065 bio_endio(req
->private_bio
, -EIO
);
1071 /* this is a conflicting request.
1072 * even though it may have been only _partially_
1073 * overlapping with one of the currently pending requests,
1074 * without even submitting or sending it, we will
1075 * pretend that it was successfully served right now.
1077 _drbd_end_io_acct(mdev
, req
);
1078 spin_unlock_irq(&mdev
->req_lock
);
1080 dec_ap_pending(mdev
);
1081 /* THINK: do we want to fail it (-EIO), or pretend success?
1082 * this pretends success. */
1086 if (req
->rq_state
& RQ_IN_ACT_LOG
)
1087 drbd_al_complete_io(mdev
, sector
);
1090 bio_put(req
->private_bio
);
1091 req
->private_bio
= NULL
;
1095 bio_endio(bio
, err
);
1104 /* helper function for drbd_make_request
1105 * if we can determine just by the mdev (state) that this request will fail,
1107 * otherwise return 0
1109 static int drbd_fail_request_early(struct drbd_conf
*mdev
, int is_write
)
1111 if (mdev
->state
.role
!= R_PRIMARY
&&
1112 (!allow_oos
|| is_write
)) {
1113 if (__ratelimit(&drbd_ratelimit_state
)) {
1114 dev_err(DEV
, "Process %s[%u] tried to %s; "
1115 "since we are not in Primary state, "
1116 "we cannot allow this\n",
1117 current
->comm
, current
->pid
,
1118 is_write
? "WRITE" : "READ");
1126 void drbd_make_request(struct request_queue
*q
, struct bio
*bio
)
1128 unsigned int s_enr
, e_enr
;
1129 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1130 unsigned long start_time
;
1132 if (drbd_fail_request_early(mdev
, bio_data_dir(bio
) & WRITE
)) {
1133 bio_endio(bio
, -EPERM
);
1137 start_time
= jiffies
;
1140 * what we "blindly" assume:
1142 D_ASSERT((bio
->bi_size
& 0x1ff) == 0);
1144 /* to make some things easier, force alignment of requests within the
1145 * granularity of our hash tables */
1146 s_enr
= bio
->bi_sector
>> HT_SHIFT
;
1147 e_enr
= bio
->bi_size
? (bio
->bi_sector
+(bio
->bi_size
>>9)-1) >> HT_SHIFT
: s_enr
;
1149 if (likely(s_enr
== e_enr
)) {
1151 inc_ap_bio(mdev
, 1);
1152 } while (drbd_make_request_common(mdev
, bio
, start_time
));
1156 /* can this bio be split generically?
1157 * Maybe add our own split-arbitrary-bios function. */
1158 if (bio
->bi_vcnt
!= 1 || bio
->bi_idx
!= 0 || bio
->bi_size
> DRBD_MAX_BIO_SIZE
) {
1159 /* rather error out here than BUG in bio_split */
1160 dev_err(DEV
, "bio would need to, but cannot, be split: "
1161 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1162 bio
->bi_vcnt
, bio
->bi_idx
, bio
->bi_size
,
1163 (unsigned long long)bio
->bi_sector
);
1164 bio_endio(bio
, -EINVAL
);
1166 /* This bio crosses some boundary, so we have to split it. */
1167 struct bio_pair
*bp
;
1168 /* works for the "do not cross hash slot boundaries" case
1169 * e.g. sector 262269, size 4096
1170 * s_enr = 262269 >> 6 = 4097
1171 * e_enr = (262269+8-1) >> 6 = 4098
1173 * sps = 64, mask = 63
1174 * first_sectors = 64 - (262269 & 63) = 3
1176 const sector_t sect
= bio
->bi_sector
;
1177 const int sps
= 1 << HT_SHIFT
; /* sectors per slot */
1178 const int mask
= sps
- 1;
1179 const sector_t first_sectors
= sps
- (sect
& mask
);
1180 bp
= bio_split(bio
, first_sectors
);
1182 /* we need to get a "reference count" (ap_bio_cnt)
1183 * to avoid races with the disconnect/reconnect/suspend code.
1184 * In case we need to split the bio here, we need to get three references
1185 * atomically, otherwise we might deadlock when trying to submit the
1187 inc_ap_bio(mdev
, 3);
1189 D_ASSERT(e_enr
== s_enr
+ 1);
1191 while (drbd_make_request_common(mdev
, &bp
->bio1
, start_time
))
1192 inc_ap_bio(mdev
, 1);
1194 while (drbd_make_request_common(mdev
, &bp
->bio2
, start_time
))
1195 inc_ap_bio(mdev
, 1);
1199 bio_pair_release(bp
);
1203 /* This is called by bio_add_page(). With this function we reduce
1204 * the number of BIOs that span over multiple DRBD_MAX_BIO_SIZEs
1205 * units (was AL_EXTENTs).
1207 * we do the calculation within the lower 32bit of the byte offsets,
1208 * since we don't care for actual offset, but only check whether it
1209 * would cross "activity log extent" boundaries.
1211 * As long as the BIO is empty we have to allow at least one bvec,
1212 * regardless of size and offset. so the resulting bio may still
1213 * cross extent boundaries. those are dealt with (bio_split) in
1214 * drbd_make_request.
1216 int drbd_merge_bvec(struct request_queue
*q
, struct bvec_merge_data
*bvm
, struct bio_vec
*bvec
)
1218 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1219 unsigned int bio_offset
=
1220 (unsigned int)bvm
->bi_sector
<< 9; /* 32 bit */
1221 unsigned int bio_size
= bvm
->bi_size
;
1222 int limit
, backing_limit
;
1224 limit
= DRBD_MAX_BIO_SIZE
1225 - ((bio_offset
& (DRBD_MAX_BIO_SIZE
-1)) + bio_size
);
1228 if (bio_size
== 0) {
1229 if (limit
<= bvec
->bv_len
)
1230 limit
= bvec
->bv_len
;
1231 } else if (limit
&& get_ldev(mdev
)) {
1232 struct request_queue
* const b
=
1233 mdev
->ldev
->backing_bdev
->bd_disk
->queue
;
1234 if (b
->merge_bvec_fn
) {
1235 backing_limit
= b
->merge_bvec_fn(b
, bvm
, bvec
);
1236 limit
= min(limit
, backing_limit
);
1243 void request_timer_fn(unsigned long data
)
1245 struct drbd_conf
*mdev
= (struct drbd_conf
*) data
;
1246 struct drbd_request
*req
; /* oldest request */
1247 struct list_head
*le
;
1248 unsigned long ent
= 0, dt
= 0, et
, nt
; /* effective timeout = ko_count * timeout */
1251 if (get_net_conf(mdev
)) {
1252 if (mdev
->state
.conn
>= C_WF_REPORT_PARAMS
)
1253 ent
= mdev
->net_conf
->timeout
*HZ
/10
1254 * mdev
->net_conf
->ko_count
;
1257 if (get_ldev(mdev
)) { /* implicit state.disk >= D_INCONSISTENT */
1258 dt
= mdev
->ldev
->dc
.disk_timeout
* HZ
/ 10;
1261 et
= min_not_zero(dt
, ent
);
1264 return; /* Recurring timer stopped */
1268 spin_lock_irq(&mdev
->req_lock
);
1269 le
= &mdev
->oldest_tle
->requests
;
1270 if (list_empty(le
)) {
1271 spin_unlock_irq(&mdev
->req_lock
);
1272 mod_timer(&mdev
->request_timer
, now
+ et
);
1277 req
= list_entry(le
, struct drbd_request
, tl_requests
);
1279 /* The request is considered timed out, if
1280 * - we have some effective timeout from the configuration,
1281 * with above state restrictions applied,
1282 * - the oldest request is waiting for a response from the network
1283 * resp. the local disk,
1284 * - the oldest request is in fact older than the effective timeout,
1285 * - the connection was established (resp. disk was attached)
1286 * for longer than the timeout already.
1287 * Note that for 32bit jiffies and very stable connections/disks,
1288 * we may have a wrap around, which is catched by
1289 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1291 * Side effect: once per 32bit wrap-around interval, which means every
1292 * ~198 days with 250 HZ, we have a window where the timeout would need
1293 * to expire twice (worst case) to become effective. Good enough.
1295 if (ent
&& req
->rq_state
& RQ_NET_PENDING
&&
1296 time_after(now
, req
->start_time
+ ent
) &&
1297 !time_in_range(now
, mdev
->last_reconnect_jif
, mdev
->last_reconnect_jif
+ ent
)) {
1298 dev_warn(DEV
, "Remote failed to finish a request within ko-count * timeout\n");
1299 _drbd_set_state(_NS(mdev
, conn
, C_TIMEOUT
), CS_VERBOSE
| CS_HARD
, NULL
);
1301 if (dt
&& req
->rq_state
& RQ_LOCAL_PENDING
&&
1302 time_after(now
, req
->start_time
+ dt
) &&
1303 !time_in_range(now
, mdev
->last_reattach_jif
, mdev
->last_reattach_jif
+ dt
)) {
1304 dev_warn(DEV
, "Local backing device failed to meet the disk-timeout\n");
1305 __drbd_chk_io_error(mdev
, DRBD_FORCE_DETACH
);
1307 nt
= (time_after(now
, req
->start_time
+ et
) ? now
: req
->start_time
) + et
;
1308 spin_unlock_irq(&mdev
->req_lock
);
1309 mod_timer(&mdev
->request_timer
, nt
);