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_stat_inc(cpu
, &mdev
->vdisk
->part0
, ios
[rw
]);
41 part_stat_add(cpu
, &mdev
->vdisk
->part0
, sectors
[rw
], bio_sectors(bio
));
42 part_inc_in_flight(&mdev
->vdisk
->part0
, rw
);
46 /* Update disk stats when completing request upwards */
47 static void _drbd_end_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
)
49 int rw
= bio_data_dir(req
->master_bio
);
50 unsigned long duration
= jiffies
- req
->start_time
;
52 cpu
= part_stat_lock();
53 part_stat_add(cpu
, &mdev
->vdisk
->part0
, ticks
[rw
], duration
);
54 part_round_stats(cpu
, &mdev
->vdisk
->part0
);
55 part_dec_in_flight(&mdev
->vdisk
->part0
, rw
);
59 static struct drbd_request
*drbd_req_new(struct drbd_conf
*mdev
,
62 struct drbd_request
*req
;
64 req
= mempool_alloc(drbd_request_mempool
, GFP_NOIO
);
68 drbd_req_make_private_bio(req
, bio_src
);
69 req
->rq_state
= bio_data_dir(bio_src
) == WRITE
? RQ_WRITE
: 0;
71 req
->master_bio
= bio_src
;
74 drbd_clear_interval(&req
->i
);
75 req
->i
.sector
= bio_src
->bi_sector
;
76 req
->i
.size
= bio_src
->bi_size
;
78 req
->i
.waiting
= false;
80 INIT_LIST_HEAD(&req
->tl_requests
);
81 INIT_LIST_HEAD(&req
->w
.list
);
86 static void drbd_req_free(struct drbd_request
*req
)
88 mempool_free(req
, drbd_request_mempool
);
91 /* rw is bio_data_dir(), only READ or WRITE */
92 static void _req_is_done(struct drbd_conf
*mdev
, struct drbd_request
*req
, const int rw
)
94 const unsigned long s
= req
->rq_state
;
96 /* remove it from the transfer log.
97 * well, only if it had been there in the first
98 * place... if it had not (local only or conflicting
99 * and never sent), it should still be "empty" as
100 * initialized in drbd_req_new(), so we can list_del() it
101 * here unconditionally */
102 list_del(&req
->tl_requests
);
104 /* if it was a write, we may have to set the corresponding
105 * bit(s) out-of-sync first. If it had a local part, we need to
106 * release the reference to the activity log. */
108 /* Set out-of-sync unless both OK flags are set
109 * (local only or remote failed).
110 * Other places where we set out-of-sync:
111 * READ with local io-error */
112 if (!(s
& RQ_NET_OK
) || !(s
& RQ_LOCAL_OK
))
113 drbd_set_out_of_sync(mdev
, req
->i
.sector
, req
->i
.size
);
115 if ((s
& RQ_NET_OK
) && (s
& RQ_LOCAL_OK
) && (s
& RQ_NET_SIS
))
116 drbd_set_in_sync(mdev
, req
->i
.sector
, req
->i
.size
);
118 /* one might be tempted to move the drbd_al_complete_io
119 * to the local io completion callback drbd_request_endio.
120 * but, if this was a mirror write, we may only
121 * drbd_al_complete_io after this is RQ_NET_DONE,
122 * otherwise the extent could be dropped from the al
123 * before it has actually been written on the peer.
124 * if we crash before our peer knows about the request,
125 * but after the extent has been dropped from the al,
126 * we would forget to resync the corresponding extent.
128 if (s
& RQ_LOCAL_MASK
) {
129 if (get_ldev_if_state(mdev
, D_FAILED
)) {
130 if (s
& RQ_IN_ACT_LOG
)
131 drbd_al_complete_io(mdev
, &req
->i
);
133 } else if (__ratelimit(&drbd_ratelimit_state
)) {
134 dev_warn(DEV
, "Should have called drbd_al_complete_io(, %llu, %u), "
135 "but my Disk seems to have failed :(\n",
136 (unsigned long long) req
->i
.sector
, req
->i
.size
);
144 static void queue_barrier(struct drbd_conf
*mdev
)
146 struct drbd_tl_epoch
*b
;
148 /* We are within the req_lock. Once we queued the barrier for sending,
149 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
150 * barrier/epoch object is added. This is the only place this bit is
151 * set. It indicates that the barrier for this epoch is already queued,
152 * and no new epoch has been created yet. */
153 if (test_bit(CREATE_BARRIER
, &mdev
->flags
))
156 b
= mdev
->tconn
->newest_tle
;
157 b
->w
.cb
= w_send_barrier
;
159 /* inc_ap_pending done here, so we won't
160 * get imbalanced on connection loss.
161 * dec_ap_pending will be done in got_BarrierAck
162 * or (on connection loss) in tl_clear. */
163 inc_ap_pending(mdev
);
164 drbd_queue_work(&mdev
->tconn
->data
.work
, &b
->w
);
165 set_bit(CREATE_BARRIER
, &mdev
->flags
);
168 static void _about_to_complete_local_write(struct drbd_conf
*mdev
,
169 struct drbd_request
*req
)
171 const unsigned long s
= req
->rq_state
;
173 /* Before we can signal completion to the upper layers,
174 * we may need to close the current epoch.
175 * We can skip this, if this request has not even been sent, because we
176 * did not have a fully established connection yet/anymore, during
177 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
179 if (mdev
->state
.conn
>= C_CONNECTED
&&
180 (s
& RQ_NET_SENT
) != 0 &&
181 req
->epoch
== mdev
->tconn
->newest_tle
->br_number
)
185 void complete_master_bio(struct drbd_conf
*mdev
,
186 struct bio_and_error
*m
)
188 bio_endio(m
->bio
, m
->error
);
193 static void drbd_remove_request_interval(struct rb_root
*root
,
194 struct drbd_request
*req
)
196 struct drbd_conf
*mdev
= req
->w
.mdev
;
197 struct drbd_interval
*i
= &req
->i
;
199 drbd_remove_interval(root
, i
);
201 /* Wake up any processes waiting for this request to complete. */
203 wake_up(&mdev
->misc_wait
);
206 /* Helper for __req_mod().
207 * Set m->bio to the master bio, if it is fit to be completed,
208 * or leave it alone (it is initialized to NULL in __req_mod),
209 * if it has already been completed, or cannot be completed yet.
210 * If m->bio is set, the error status to be returned is placed in m->error.
212 void _req_may_be_done(struct drbd_request
*req
, struct bio_and_error
*m
)
214 const unsigned long s
= req
->rq_state
;
215 struct drbd_conf
*mdev
= req
->w
.mdev
;
216 int rw
= req
->rq_state
& RQ_WRITE
? WRITE
: READ
;
218 /* we must not complete the master bio, while it is
219 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
220 * not yet acknowledged by the peer
221 * not yet completed by the local io subsystem
222 * these flags may get cleared in any order by
225 * the bio_endio completion callbacks.
227 if (s
& RQ_LOCAL_PENDING
&& !(s
& RQ_LOCAL_ABORTED
))
229 if (req
->i
.waiting
) {
230 /* Retry all conflicting peer requests. */
231 wake_up(&mdev
->misc_wait
);
233 if (s
& RQ_NET_QUEUED
)
235 if (s
& RQ_NET_PENDING
)
238 if (req
->master_bio
) {
239 /* this is DATA_RECEIVED (remote read)
240 * or protocol C P_WRITE_ACK
241 * or protocol B P_RECV_ACK
242 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
243 * or canceled or failed,
244 * or killed from the transfer log due to connection loss.
248 * figure out whether to report success or failure.
250 * report success when at least one of the operations succeeded.
251 * or, to put the other way,
252 * only report failure, when both operations failed.
254 * what to do about the failures is handled elsewhere.
255 * what we need to do here is just: complete the master_bio.
257 * local completion error, if any, has been stored as ERR_PTR
258 * in private_bio within drbd_request_endio.
260 int ok
= (s
& RQ_LOCAL_OK
) || (s
& RQ_NET_OK
);
261 int error
= PTR_ERR(req
->private_bio
);
263 /* remove the request from the conflict detection
264 * respective block_id verification hash */
265 if (!drbd_interval_empty(&req
->i
)) {
266 struct rb_root
*root
;
269 root
= &mdev
->write_requests
;
271 root
= &mdev
->read_requests
;
272 drbd_remove_request_interval(root
, req
);
273 } else if (!(s
& RQ_POSTPONED
))
274 D_ASSERT((s
& (RQ_NET_MASK
& ~RQ_NET_DONE
)) == 0);
276 /* for writes we need to do some extra housekeeping */
278 _about_to_complete_local_write(mdev
, req
);
280 /* Update disk stats */
281 _drbd_end_io_acct(mdev
, req
);
283 if (!(s
& RQ_POSTPONED
)) {
284 m
->error
= ok
? 0 : (error
?: -EIO
);
285 m
->bio
= req
->master_bio
;
287 req
->master_bio
= NULL
;
290 if (s
& RQ_LOCAL_PENDING
)
293 if ((s
& RQ_NET_MASK
) == 0 || (s
& RQ_NET_DONE
)) {
294 /* this is disconnected (local only) operation,
295 * or protocol A, B, or C P_BARRIER_ACK,
296 * or killed from the transfer log due to connection loss. */
297 _req_is_done(mdev
, req
, rw
);
299 /* else: network part and not DONE yet. that is
300 * protocol A, B, or C, barrier ack still pending... */
303 static void _req_may_be_done_not_susp(struct drbd_request
*req
, struct bio_and_error
*m
)
305 struct drbd_conf
*mdev
= req
->w
.mdev
;
307 if (!drbd_suspended(mdev
))
308 _req_may_be_done(req
, m
);
311 /* obviously this could be coded as many single functions
312 * instead of one huge switch,
313 * or by putting the code directly in the respective locations
314 * (as it has been before).
316 * but having it this way
317 * enforces that it is all in this one place, where it is easier to audit,
318 * it makes it obvious that whatever "event" "happens" to a request should
319 * happen "atomically" within the req_lock,
320 * and it enforces that we have to think in a very structured manner
321 * about the "events" that may happen to a request during its life time ...
323 int __req_mod(struct drbd_request
*req
, enum drbd_req_event what
,
324 struct bio_and_error
*m
)
326 struct drbd_conf
*mdev
= req
->w
.mdev
;
335 dev_err(DEV
, "LOGIC BUG in %s:%u\n", __FILE__
, __LINE__
);
338 /* does not happen...
339 * initialization done in drbd_req_new
344 case TO_BE_SENT
: /* via network */
345 /* reached via __drbd_make_request
346 * and from w_read_retry_remote */
347 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
348 req
->rq_state
|= RQ_NET_PENDING
;
350 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
351 p
= nc
->wire_protocol
;
354 p
== DRBD_PROT_C
? RQ_EXP_WRITE_ACK
:
355 p
== DRBD_PROT_B
? RQ_EXP_RECEIVE_ACK
: 0;
356 inc_ap_pending(mdev
);
359 case TO_BE_SUBMITTED
: /* locally */
360 /* reached via __drbd_make_request */
361 D_ASSERT(!(req
->rq_state
& RQ_LOCAL_MASK
));
362 req
->rq_state
|= RQ_LOCAL_PENDING
;
366 if (req
->rq_state
& RQ_WRITE
)
367 mdev
->writ_cnt
+= req
->i
.size
>> 9;
369 mdev
->read_cnt
+= req
->i
.size
>> 9;
371 req
->rq_state
|= (RQ_LOCAL_COMPLETED
|RQ_LOCAL_OK
);
372 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
374 _req_may_be_done_not_susp(req
, m
);
379 req
->rq_state
|= RQ_LOCAL_ABORTED
;
380 if (req
->rq_state
& RQ_WRITE
)
381 _req_may_be_done_not_susp(req
, m
);
383 goto goto_queue_for_net_read
;
386 case WRITE_COMPLETED_WITH_ERROR
:
387 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
388 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
390 __drbd_chk_io_error(mdev
, false);
391 _req_may_be_done_not_susp(req
, m
);
395 case READ_AHEAD_COMPLETED_WITH_ERROR
:
396 /* it is legal to fail READA */
397 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
398 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
399 _req_may_be_done_not_susp(req
, m
);
403 case READ_COMPLETED_WITH_ERROR
:
404 drbd_set_out_of_sync(mdev
, req
->i
.sector
, req
->i
.size
);
406 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
407 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
409 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
411 __drbd_chk_io_error(mdev
, false);
414 goto_queue_for_net_read
:
416 /* no point in retrying if there is no good remote data,
417 * or we have no connection. */
418 if (mdev
->state
.pdsk
!= D_UP_TO_DATE
) {
419 _req_may_be_done_not_susp(req
, m
);
423 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
424 req
->rq_state
|= RQ_NET_PENDING
;
425 inc_ap_pending(mdev
);
426 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
428 case QUEUE_FOR_NET_READ
:
429 /* READ or READA, and
431 * or target area marked as invalid,
432 * or just got an io-error. */
433 /* from __drbd_make_request
434 * or from bio_endio during read io-error recovery */
436 /* so we can verify the handle in the answer packet
437 * corresponding hlist_del is in _req_may_be_done() */
438 D_ASSERT(drbd_interval_empty(&req
->i
));
439 drbd_insert_interval(&mdev
->read_requests
, &req
->i
);
441 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
443 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
444 req
->rq_state
|= RQ_NET_QUEUED
;
445 req
->w
.cb
= (req
->rq_state
& RQ_LOCAL_MASK
)
446 ? w_read_retry_remote
448 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
451 case QUEUE_FOR_NET_WRITE
:
452 /* assert something? */
453 /* from __drbd_make_request only */
455 /* corresponding hlist_del is in _req_may_be_done() */
456 D_ASSERT(drbd_interval_empty(&req
->i
));
457 drbd_insert_interval(&mdev
->write_requests
, &req
->i
);
460 * In case the req ended up on the transfer log before being
461 * queued on the worker, it could lead to this request being
462 * missed during cleanup after connection loss.
463 * So we have to do both operations here,
464 * within the same lock that protects the transfer log.
466 * _req_add_to_epoch(req); this has to be after the
467 * _maybe_start_new_epoch(req); which happened in
468 * __drbd_make_request, because we now may set the bit
469 * again ourselves to close the current epoch.
471 * Add req to the (now) current epoch (barrier). */
473 /* otherwise we may lose an unplug, which may cause some remote
474 * io-scheduler timeout to expire, increasing maximum latency,
475 * hurting performance. */
476 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
478 /* see __drbd_make_request,
479 * just after it grabs the req_lock */
480 D_ASSERT(test_bit(CREATE_BARRIER
, &mdev
->flags
) == 0);
482 req
->epoch
= mdev
->tconn
->newest_tle
->br_number
;
484 /* increment size of current epoch */
485 mdev
->tconn
->newest_tle
->n_writes
++;
487 /* queue work item to send data */
488 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
489 req
->rq_state
|= RQ_NET_QUEUED
;
490 req
->w
.cb
= w_send_dblock
;
491 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
493 /* close the epoch, in case it outgrew the limit */
495 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
496 p
= nc
->max_epoch_size
;
498 if (mdev
->tconn
->newest_tle
->n_writes
>= p
)
503 case QUEUE_FOR_SEND_OOS
:
504 req
->rq_state
|= RQ_NET_QUEUED
;
505 req
->w
.cb
= w_send_out_of_sync
;
506 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
509 case OOS_HANDED_TO_NETWORK
:
510 /* actually the same */
512 /* treat it the same */
514 /* real cleanup will be done from tl_clear. just update flags
515 * so it is no longer marked as on the worker queue */
516 req
->rq_state
&= ~RQ_NET_QUEUED
;
517 /* if we did it right, tl_clear should be scheduled only after
518 * this, so this should not be necessary! */
519 _req_may_be_done_not_susp(req
, m
);
522 case HANDED_OVER_TO_NETWORK
:
523 /* assert something? */
524 if (bio_data_dir(req
->master_bio
) == WRITE
)
525 atomic_add(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
527 if (bio_data_dir(req
->master_bio
) == WRITE
&&
528 !(req
->rq_state
& (RQ_EXP_RECEIVE_ACK
| RQ_EXP_WRITE_ACK
))) {
529 /* this is what is dangerous about protocol A:
530 * pretend it was successfully written on the peer. */
531 if (req
->rq_state
& RQ_NET_PENDING
) {
532 dec_ap_pending(mdev
);
533 req
->rq_state
&= ~RQ_NET_PENDING
;
534 req
->rq_state
|= RQ_NET_OK
;
535 } /* else: neg-ack was faster... */
536 /* it is still not yet RQ_NET_DONE until the
537 * corresponding epoch barrier got acked as well,
538 * so we know what to dirty on connection loss */
540 req
->rq_state
&= ~RQ_NET_QUEUED
;
541 req
->rq_state
|= RQ_NET_SENT
;
542 /* because _drbd_send_zc_bio could sleep, and may want to
543 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
544 * "COMPLETED_OK" events came in, once we return from
545 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
546 * whether it is done already, and end it. */
547 _req_may_be_done_not_susp(req
, m
);
550 case READ_RETRY_REMOTE_CANCELED
:
551 req
->rq_state
&= ~RQ_NET_QUEUED
;
552 /* fall through, in case we raced with drbd_disconnect */
553 case CONNECTION_LOST_WHILE_PENDING
:
554 /* transfer log cleanup after connection loss */
555 /* assert something? */
556 if (req
->rq_state
& RQ_NET_PENDING
)
557 dec_ap_pending(mdev
);
558 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
559 req
->rq_state
|= RQ_NET_DONE
;
560 if (req
->rq_state
& RQ_NET_SENT
&& req
->rq_state
& RQ_WRITE
)
561 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
563 /* if it is still queued, we may not complete it here.
564 * it will be canceled soon. */
565 if (!(req
->rq_state
& RQ_NET_QUEUED
))
566 _req_may_be_done(req
, m
); /* Allowed while state.susp */
569 case WRITE_ACKED_BY_PEER_AND_SIS
:
570 req
->rq_state
|= RQ_NET_SIS
;
572 /* for discarded conflicting writes of multiple primaries,
573 * there is no need to keep anything in the tl, potential
574 * node crashes are covered by the activity log. */
575 req
->rq_state
|= RQ_NET_DONE
;
577 case WRITE_ACKED_BY_PEER
:
578 D_ASSERT(req
->rq_state
& RQ_EXP_WRITE_ACK
);
579 /* protocol C; successfully written on peer.
580 * Nothing to do here.
581 * We want to keep the tl in place for all protocols, to cater
582 * for volatile write-back caches on lower level devices.
584 * A barrier request is expected to have forced all prior
585 * requests onto stable storage, so completion of a barrier
586 * request could set NET_DONE right here, and not wait for the
587 * P_BARRIER_ACK, but that is an unnecessary optimization. */
590 /* this makes it effectively the same as for: */
591 case RECV_ACKED_BY_PEER
:
592 D_ASSERT(req
->rq_state
& RQ_EXP_RECEIVE_ACK
);
593 /* protocol B; pretends to be successfully written on peer.
594 * see also notes above in HANDED_OVER_TO_NETWORK about
597 req
->rq_state
|= RQ_NET_OK
;
598 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
599 dec_ap_pending(mdev
);
600 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
601 req
->rq_state
&= ~RQ_NET_PENDING
;
602 _req_may_be_done_not_susp(req
, m
);
606 D_ASSERT(req
->rq_state
& RQ_EXP_WRITE_ACK
);
607 /* If this node has already detected the write conflict, the
608 * worker will be waiting on misc_wait. Wake it up once this
609 * request has completed locally.
611 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
612 req
->rq_state
|= RQ_POSTPONED
;
613 _req_may_be_done_not_susp(req
, m
);
617 /* assert something? */
618 if (req
->rq_state
& RQ_NET_PENDING
) {
619 dec_ap_pending(mdev
);
620 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
622 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
624 req
->rq_state
|= RQ_NET_DONE
;
625 _req_may_be_done_not_susp(req
, m
);
626 /* else: done by HANDED_OVER_TO_NETWORK */
629 case FAIL_FROZEN_DISK_IO
:
630 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
633 _req_may_be_done(req
, m
); /* Allowed while state.susp */
636 case RESTART_FROZEN_DISK_IO
:
637 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
640 req
->rq_state
&= ~RQ_LOCAL_COMPLETED
;
643 if (bio_data_dir(req
->master_bio
) == WRITE
)
647 req
->w
.cb
= w_restart_disk_io
;
648 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
652 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
653 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
654 Trowing them out of the TL here by pretending we got a BARRIER_ACK
655 We ensure that the peer was not rebooted */
656 if (!(req
->rq_state
& RQ_NET_OK
)) {
658 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
659 rv
= req
->rq_state
& RQ_WRITE
? MR_WRITE
: MR_READ
;
663 /* else, fall through to BARRIER_ACKED */
666 if (!(req
->rq_state
& RQ_WRITE
))
669 if (req
->rq_state
& RQ_NET_PENDING
) {
670 /* barrier came in before all requests were acked.
671 * this is bad, because if the connection is lost now,
672 * we won't be able to clean them up... */
673 dev_err(DEV
, "FIXME (BARRIER_ACKED but pending)\n");
674 list_move(&req
->tl_requests
, &mdev
->tconn
->out_of_sequence_requests
);
676 if ((req
->rq_state
& RQ_NET_MASK
) != 0) {
677 req
->rq_state
|= RQ_NET_DONE
;
678 if (!(req
->rq_state
& (RQ_EXP_RECEIVE_ACK
| RQ_EXP_WRITE_ACK
)))
679 atomic_sub(req
->i
.size
>>9, &mdev
->ap_in_flight
);
681 _req_may_be_done(req
, m
); /* Allowed while state.susp */
685 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
686 dec_ap_pending(mdev
);
687 req
->rq_state
&= ~RQ_NET_PENDING
;
688 req
->rq_state
|= (RQ_NET_OK
|RQ_NET_DONE
);
689 _req_may_be_done_not_susp(req
, m
);
696 /* we may do a local read if:
697 * - we are consistent (of course),
698 * - or we are generally inconsistent,
699 * BUT we are still/already IN SYNC for this area.
700 * since size may be bigger than BM_BLOCK_SIZE,
701 * we may need to check several bits.
703 static bool drbd_may_do_local_read(struct drbd_conf
*mdev
, sector_t sector
, int size
)
705 unsigned long sbnr
, ebnr
;
706 sector_t esector
, nr_sectors
;
708 if (mdev
->state
.disk
== D_UP_TO_DATE
)
710 if (mdev
->state
.disk
!= D_INCONSISTENT
)
712 esector
= sector
+ (size
>> 9) - 1;
713 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
714 D_ASSERT(sector
< nr_sectors
);
715 D_ASSERT(esector
< nr_sectors
);
717 sbnr
= BM_SECT_TO_BIT(sector
);
718 ebnr
= BM_SECT_TO_BIT(esector
);
720 return drbd_bm_count_bits(mdev
, sbnr
, ebnr
) == 0;
724 * complete_conflicting_writes - wait for any conflicting write requests
726 * The write_requests tree contains all active write requests which we
727 * currently know about. Wait for any requests to complete which conflict with
730 static int complete_conflicting_writes(struct drbd_conf
*mdev
,
731 sector_t sector
, int size
)
734 struct drbd_interval
*i
;
737 i
= drbd_find_overlap(&mdev
->write_requests
, sector
, size
);
740 err
= drbd_wait_misc(mdev
, i
);
746 int __drbd_make_request(struct drbd_conf
*mdev
, struct bio
*bio
, unsigned long start_time
)
748 const int rw
= bio_rw(bio
);
749 const int size
= bio
->bi_size
;
750 const sector_t sector
= bio
->bi_sector
;
751 struct drbd_tl_epoch
*b
= NULL
;
752 struct drbd_request
*req
;
754 int local
, remote
, send_oos
= 0;
758 /* allocate outside of all locks; */
759 req
= drbd_req_new(mdev
, bio
);
762 /* only pass the error to the upper layers.
763 * if user cannot handle io errors, that's not our business. */
764 dev_err(DEV
, "could not kmalloc() req\n");
765 bio_endio(bio
, -ENOMEM
);
768 req
->start_time
= start_time
;
770 local
= get_ldev(mdev
);
772 bio_put(req
->private_bio
); /* or we get a bio leak */
773 req
->private_bio
= NULL
;
780 if (!drbd_may_do_local_read(mdev
, sector
, size
)) {
781 /* we could kick the syncer to
782 * sync this extent asap, wait for
783 * it, then continue locally.
784 * Or just issue the request remotely.
787 bio_put(req
->private_bio
);
788 req
->private_bio
= NULL
;
792 remote
= !local
&& mdev
->state
.pdsk
>= D_UP_TO_DATE
;
795 /* If we have a disk, but a READA request is mapped to remote,
796 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
797 * Just fail that READA request right here.
799 * THINK: maybe fail all READA when not local?
800 * or make this configurable...
801 * if network is slow, READA won't do any good.
803 if (rw
== READA
&& mdev
->state
.disk
>= D_INCONSISTENT
&& !local
) {
805 goto fail_and_free_req
;
808 /* For WRITES going to the local disk, grab a reference on the target
809 * extent. This waits for any resync activity in the corresponding
810 * resync extent to finish, and, if necessary, pulls in the target
811 * extent into the activity log, which involves further disk io because
812 * of transactional on-disk meta data updates. */
813 if (rw
== WRITE
&& local
&& !test_bit(AL_SUSPENDED
, &mdev
->flags
)) {
814 req
->rq_state
|= RQ_IN_ACT_LOG
;
815 drbd_al_begin_io(mdev
, &req
->i
);
818 remote
= remote
&& drbd_should_do_remote(mdev
->state
);
819 send_oos
= rw
== WRITE
&& drbd_should_send_out_of_sync(mdev
->state
);
820 D_ASSERT(!(remote
&& send_oos
));
822 if (!(local
|| remote
) && !drbd_suspended(mdev
)) {
823 if (__ratelimit(&drbd_ratelimit_state
))
824 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
826 goto fail_free_complete
;
829 /* For WRITE request, we have to make sure that we have an
830 * unused_spare_tle, in case we need to start a new epoch.
831 * I try to be smart and avoid to pre-allocate always "just in case",
832 * but there is a race between testing the bit and pointer outside the
833 * spinlock, and grabbing the spinlock.
834 * if we lost that race, we retry. */
835 if (rw
== WRITE
&& (remote
|| send_oos
) &&
836 mdev
->tconn
->unused_spare_tle
== NULL
&&
837 test_bit(CREATE_BARRIER
, &mdev
->flags
)) {
839 b
= kmalloc(sizeof(struct drbd_tl_epoch
), GFP_NOIO
);
841 dev_err(DEV
, "Failed to alloc barrier.\n");
843 goto fail_free_complete
;
847 /* GOOD, everything prepared, grab the spin_lock */
848 spin_lock_irq(&mdev
->tconn
->req_lock
);
851 err
= complete_conflicting_writes(mdev
, sector
, size
);
853 if (err
!= -ERESTARTSYS
)
854 _conn_request_state(mdev
->tconn
,
857 spin_unlock_irq(&mdev
->tconn
->req_lock
);
859 goto fail_free_complete
;
863 if (drbd_suspended(mdev
)) {
864 /* If we got suspended, use the retry mechanism of
865 generic_make_request() to restart processing of this
866 bio. In the next call to drbd_make_request
867 we sleep in inc_ap_bio() */
869 spin_unlock_irq(&mdev
->tconn
->req_lock
);
870 goto fail_free_complete
;
873 if (remote
|| send_oos
) {
874 remote
= drbd_should_do_remote(mdev
->state
);
875 send_oos
= rw
== WRITE
&& drbd_should_send_out_of_sync(mdev
->state
);
876 D_ASSERT(!(remote
&& send_oos
));
878 if (!(remote
|| send_oos
))
879 dev_warn(DEV
, "lost connection while grabbing the req_lock!\n");
880 if (!(local
|| remote
)) {
881 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
882 spin_unlock_irq(&mdev
->tconn
->req_lock
);
884 goto fail_free_complete
;
888 if (b
&& mdev
->tconn
->unused_spare_tle
== NULL
) {
889 mdev
->tconn
->unused_spare_tle
= b
;
892 if (rw
== WRITE
&& (remote
|| send_oos
) &&
893 mdev
->tconn
->unused_spare_tle
== NULL
&&
894 test_bit(CREATE_BARRIER
, &mdev
->flags
)) {
895 /* someone closed the current epoch
896 * while we were grabbing the spinlock */
897 spin_unlock_irq(&mdev
->tconn
->req_lock
);
898 goto allocate_barrier
;
902 /* Update disk stats */
903 _drbd_start_io_acct(mdev
, req
, bio
);
905 /* _maybe_start_new_epoch(mdev);
906 * If we need to generate a write barrier packet, we have to add the
907 * new epoch (barrier) object, and queue the barrier packet for sending,
908 * and queue the req's data after it _within the same lock_, otherwise
909 * we have race conditions were the reorder domains could be mixed up.
911 * Even read requests may start a new epoch and queue the corresponding
912 * barrier packet. To get the write ordering right, we only have to
913 * make sure that, if this is a write request and it triggered a
914 * barrier packet, this request is queued within the same spinlock. */
915 if ((remote
|| send_oos
) && mdev
->tconn
->unused_spare_tle
&&
916 test_and_clear_bit(CREATE_BARRIER
, &mdev
->flags
)) {
917 _tl_add_barrier(mdev
->tconn
, mdev
->tconn
->unused_spare_tle
);
918 mdev
->tconn
->unused_spare_tle
= NULL
;
920 D_ASSERT(!(remote
&& rw
== WRITE
&&
921 test_bit(CREATE_BARRIER
, &mdev
->flags
)));
925 * Actually, 'local' may be wrong here already, since we may have failed
926 * to write to the meta data, and may become wrong anytime because of
927 * local io-error for some other request, which would lead to us
928 * "detaching" the local disk.
930 * 'remote' may become wrong any time because the network could fail.
932 * This is a harmless race condition, though, since it is handled
933 * correctly at the appropriate places; so it just defers the failure
934 * of the respective operation.
937 /* mark them early for readability.
938 * this just sets some state flags. */
940 _req_mod(req
, TO_BE_SENT
);
942 _req_mod(req
, TO_BE_SUBMITTED
);
944 list_add_tail(&req
->tl_requests
, &mdev
->tconn
->newest_tle
->requests
);
946 /* NOTE remote first: to get the concurrent write detection right,
947 * we must register the request before start of local IO. */
949 /* either WRITE and C_CONNECTED,
950 * or READ, and no local disk,
951 * or READ, but not in sync.
953 _req_mod(req
, (rw
== WRITE
)
954 ? QUEUE_FOR_NET_WRITE
955 : QUEUE_FOR_NET_READ
);
957 if (send_oos
&& drbd_set_out_of_sync(mdev
, sector
, size
))
958 _req_mod(req
, QUEUE_FOR_SEND_OOS
);
961 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
963 nc
->on_congestion
!= OC_BLOCK
&& mdev
->tconn
->agreed_pro_version
>= 96) {
967 atomic_read(&mdev
->ap_in_flight
) >= nc
->cong_fill
) {
968 dev_info(DEV
, "Congestion-fill threshold reached\n");
972 if (mdev
->act_log
->used
>= nc
->cong_extents
) {
973 dev_info(DEV
, "Congestion-extents threshold reached\n");
978 queue_barrier(mdev
); /* last barrier, after mirrored writes */
980 if (nc
->on_congestion
== OC_PULL_AHEAD
)
981 _drbd_set_state(_NS(mdev
, conn
, C_AHEAD
), 0, NULL
);
982 else /*nc->on_congestion == OC_DISCONNECT */
983 _drbd_set_state(_NS(mdev
, conn
, C_DISCONNECTING
), 0, NULL
);
988 spin_unlock_irq(&mdev
->tconn
->req_lock
);
989 kfree(b
); /* if someone else has beaten us to it... */
992 req
->private_bio
->bi_bdev
= mdev
->ldev
->backing_bdev
;
994 /* State may have changed since we grabbed our reference on the
995 * mdev->ldev member. Double check, and short-circuit to endio.
996 * In case the last activity log transaction failed to get on
997 * stable storage, and this is a WRITE, we may not even submit
999 if (get_ldev(mdev
)) {
1000 if (drbd_insert_fault(mdev
, rw
== WRITE
? DRBD_FAULT_DT_WR
1001 : rw
== READ
? DRBD_FAULT_DT_RD
1002 : DRBD_FAULT_DT_RA
))
1003 bio_endio(req
->private_bio
, -EIO
);
1005 generic_make_request(req
->private_bio
);
1008 bio_endio(req
->private_bio
, -EIO
);
1014 if (req
->rq_state
& RQ_IN_ACT_LOG
)
1015 drbd_al_complete_io(mdev
, &req
->i
);
1018 bio_put(req
->private_bio
);
1019 req
->private_bio
= NULL
;
1023 bio_endio(bio
, err
);
1032 int drbd_make_request(struct request_queue
*q
, struct bio
*bio
)
1034 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1035 unsigned long start_time
;
1037 start_time
= jiffies
;
1040 * what we "blindly" assume:
1042 D_ASSERT(bio
->bi_size
> 0);
1043 D_ASSERT(IS_ALIGNED(bio
->bi_size
, 512));
1046 return __drbd_make_request(mdev
, bio
, start_time
);
1049 /* This is called by bio_add_page().
1051 * q->max_hw_sectors and other global limits are already enforced there.
1053 * We need to call down to our lower level device,
1054 * in case it has special restrictions.
1056 * We also may need to enforce configured max-bio-bvecs limits.
1058 * As long as the BIO is empty we have to allow at least one bvec,
1059 * regardless of size and offset, so no need to ask lower levels.
1061 int drbd_merge_bvec(struct request_queue
*q
, struct bvec_merge_data
*bvm
, struct bio_vec
*bvec
)
1063 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1064 unsigned int bio_size
= bvm
->bi_size
;
1065 int limit
= DRBD_MAX_BIO_SIZE
;
1068 if (bio_size
&& get_ldev(mdev
)) {
1069 struct request_queue
* const b
=
1070 mdev
->ldev
->backing_bdev
->bd_disk
->queue
;
1071 if (b
->merge_bvec_fn
) {
1072 backing_limit
= b
->merge_bvec_fn(b
, bvm
, bvec
);
1073 limit
= min(limit
, backing_limit
);
1080 void request_timer_fn(unsigned long data
)
1082 struct drbd_conf
*mdev
= (struct drbd_conf
*) data
;
1083 struct drbd_tconn
*tconn
= mdev
->tconn
;
1084 struct drbd_request
*req
; /* oldest request */
1085 struct list_head
*le
;
1086 struct net_conf
*nc
;
1087 unsigned long ent
= 0, dt
= 0, et
, nt
; /* effective timeout = ko_count * timeout */
1090 nc
= rcu_dereference(tconn
->net_conf
);
1091 ent
= nc
? nc
->timeout
* HZ
/10 * nc
->ko_count
: 0;
1093 if (get_ldev(mdev
)) {
1094 dt
= rcu_dereference(mdev
->ldev
->disk_conf
)->disk_timeout
* HZ
/ 10;
1099 et
= min_not_zero(dt
, ent
);
1101 if (!et
|| (mdev
->state
.conn
< C_WF_REPORT_PARAMS
&& mdev
->state
.disk
<= D_FAILED
))
1102 return; /* Recurring timer stopped */
1104 spin_lock_irq(&tconn
->req_lock
);
1105 le
= &tconn
->oldest_tle
->requests
;
1106 if (list_empty(le
)) {
1107 spin_unlock_irq(&tconn
->req_lock
);
1108 mod_timer(&mdev
->request_timer
, jiffies
+ et
);
1113 req
= list_entry(le
, struct drbd_request
, tl_requests
);
1114 if (ent
&& req
->rq_state
& RQ_NET_PENDING
) {
1115 if (time_is_before_eq_jiffies(req
->start_time
+ ent
)) {
1116 dev_warn(DEV
, "Remote failed to finish a request within ko-count * timeout\n");
1117 _drbd_set_state(_NS(mdev
, conn
, C_TIMEOUT
), CS_VERBOSE
| CS_HARD
, NULL
);
1120 if (dt
&& req
->rq_state
& RQ_LOCAL_PENDING
) {
1121 if (time_is_before_eq_jiffies(req
->start_time
+ dt
)) {
1122 dev_warn(DEV
, "Local backing device failed to meet the disk-timeout\n");
1123 __drbd_chk_io_error(mdev
, 1);
1126 nt
= (time_is_before_eq_jiffies(req
->start_time
+ et
) ? jiffies
: req
->start_time
) + et
;
1127 spin_unlock_irq(&tconn
->req_lock
);
1128 mod_timer(&mdev
->request_timer
, nt
);