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 (void) cpu
; /* The macro invocations above want the cpu argument, I do not like
43 the compiler warning about cpu only assigned but never used... */
44 part_inc_in_flight(&mdev
->vdisk
->part0
, rw
);
48 /* Update disk stats when completing request upwards */
49 static void _drbd_end_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
)
51 int rw
= bio_data_dir(req
->master_bio
);
52 unsigned long duration
= jiffies
- req
->start_time
;
54 cpu
= part_stat_lock();
55 part_stat_add(cpu
, &mdev
->vdisk
->part0
, ticks
[rw
], duration
);
56 part_round_stats(cpu
, &mdev
->vdisk
->part0
);
57 part_dec_in_flight(&mdev
->vdisk
->part0
, rw
);
61 static struct drbd_request
*drbd_req_new(struct drbd_conf
*mdev
,
64 struct drbd_request
*req
;
66 req
= mempool_alloc(drbd_request_mempool
, GFP_NOIO
);
70 drbd_req_make_private_bio(req
, bio_src
);
71 req
->rq_state
= bio_data_dir(bio_src
) == WRITE
? RQ_WRITE
: 0;
73 req
->master_bio
= bio_src
;
76 drbd_clear_interval(&req
->i
);
77 req
->i
.sector
= bio_src
->bi_sector
;
78 req
->i
.size
= bio_src
->bi_size
;
80 req
->i
.waiting
= false;
82 INIT_LIST_HEAD(&req
->tl_requests
);
83 INIT_LIST_HEAD(&req
->w
.list
);
88 static void drbd_req_free(struct drbd_request
*req
)
90 mempool_free(req
, drbd_request_mempool
);
93 /* rw is bio_data_dir(), only READ or WRITE */
94 static void _req_is_done(struct drbd_conf
*mdev
, struct drbd_request
*req
, const int rw
)
96 const unsigned long s
= req
->rq_state
;
98 /* remove it from the transfer log.
99 * well, only if it had been there in the first
100 * place... if it had not (local only or conflicting
101 * and never sent), it should still be "empty" as
102 * initialized in drbd_req_new(), so we can list_del() it
103 * here unconditionally */
104 list_del(&req
->tl_requests
);
106 /* if it was a write, we may have to set the corresponding
107 * bit(s) out-of-sync first. If it had a local part, we need to
108 * release the reference to the activity log. */
110 /* Set out-of-sync unless both OK flags are set
111 * (local only or remote failed).
112 * Other places where we set out-of-sync:
113 * READ with local io-error */
114 if (!(s
& RQ_NET_OK
) || !(s
& RQ_LOCAL_OK
))
115 drbd_set_out_of_sync(mdev
, req
->i
.sector
, req
->i
.size
);
117 if ((s
& RQ_NET_OK
) && (s
& RQ_LOCAL_OK
) && (s
& RQ_NET_SIS
))
118 drbd_set_in_sync(mdev
, req
->i
.sector
, req
->i
.size
);
120 /* one might be tempted to move the drbd_al_complete_io
121 * to the local io completion callback drbd_request_endio.
122 * but, if this was a mirror write, we may only
123 * drbd_al_complete_io after this is RQ_NET_DONE,
124 * otherwise the extent could be dropped from the al
125 * before it has actually been written on the peer.
126 * if we crash before our peer knows about the request,
127 * but after the extent has been dropped from the al,
128 * we would forget to resync the corresponding extent.
130 if (s
& RQ_LOCAL_MASK
) {
131 if (get_ldev_if_state(mdev
, D_FAILED
)) {
132 if (s
& RQ_IN_ACT_LOG
)
133 drbd_al_complete_io(mdev
, &req
->i
);
135 } else if (__ratelimit(&drbd_ratelimit_state
)) {
136 dev_warn(DEV
, "Should have called drbd_al_complete_io(, %llu, %u), "
137 "but my Disk seems to have failed :(\n",
138 (unsigned long long) req
->i
.sector
, req
->i
.size
);
146 static void queue_barrier(struct drbd_conf
*mdev
)
148 struct drbd_tl_epoch
*b
;
149 struct drbd_tconn
*tconn
= mdev
->tconn
;
151 /* We are within the req_lock. Once we queued the barrier for sending,
152 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
153 * barrier/epoch object is added. This is the only place this bit is
154 * set. It indicates that the barrier for this epoch is already queued,
155 * and no new epoch has been created yet. */
156 if (test_bit(CREATE_BARRIER
, &tconn
->flags
))
159 b
= tconn
->newest_tle
;
160 b
->w
.cb
= w_send_barrier
;
162 /* inc_ap_pending done here, so we won't
163 * get imbalanced on connection loss.
164 * dec_ap_pending will be done in got_BarrierAck
165 * or (on connection loss) in tl_clear. */
166 inc_ap_pending(mdev
);
167 drbd_queue_work(&tconn
->data
.work
, &b
->w
);
168 set_bit(CREATE_BARRIER
, &tconn
->flags
);
171 static void _about_to_complete_local_write(struct drbd_conf
*mdev
,
172 struct drbd_request
*req
)
174 const unsigned long s
= req
->rq_state
;
176 /* Before we can signal completion to the upper layers,
177 * we may need to close the current epoch.
178 * We can skip this, if this request has not even been sent, because we
179 * did not have a fully established connection yet/anymore, during
180 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
182 if (mdev
->state
.conn
>= C_CONNECTED
&&
183 (s
& RQ_NET_SENT
) != 0 &&
184 req
->epoch
== mdev
->tconn
->newest_tle
->br_number
)
188 void complete_master_bio(struct drbd_conf
*mdev
,
189 struct bio_and_error
*m
)
191 bio_endio(m
->bio
, m
->error
);
196 static void drbd_remove_request_interval(struct rb_root
*root
,
197 struct drbd_request
*req
)
199 struct drbd_conf
*mdev
= req
->w
.mdev
;
200 struct drbd_interval
*i
= &req
->i
;
202 drbd_remove_interval(root
, i
);
204 /* Wake up any processes waiting for this request to complete. */
206 wake_up(&mdev
->misc_wait
);
209 /* Helper for __req_mod().
210 * Set m->bio to the master bio, if it is fit to be completed,
211 * or leave it alone (it is initialized to NULL in __req_mod),
212 * if it has already been completed, or cannot be completed yet.
213 * If m->bio is set, the error status to be returned is placed in m->error.
215 void _req_may_be_done(struct drbd_request
*req
, struct bio_and_error
*m
)
217 const unsigned long s
= req
->rq_state
;
218 struct drbd_conf
*mdev
= req
->w
.mdev
;
219 int rw
= req
->rq_state
& RQ_WRITE
? WRITE
: READ
;
221 /* we must not complete the master bio, while it is
222 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
223 * not yet acknowledged by the peer
224 * not yet completed by the local io subsystem
225 * these flags may get cleared in any order by
228 * the bio_endio completion callbacks.
230 if (s
& RQ_LOCAL_PENDING
&& !(s
& RQ_LOCAL_ABORTED
))
232 if (req
->i
.waiting
) {
233 /* Retry all conflicting peer requests. */
234 wake_up(&mdev
->misc_wait
);
236 if (s
& RQ_NET_QUEUED
)
238 if (s
& RQ_NET_PENDING
)
241 if (req
->master_bio
) {
242 /* this is DATA_RECEIVED (remote read)
243 * or protocol C P_WRITE_ACK
244 * or protocol B P_RECV_ACK
245 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
246 * or canceled or failed,
247 * or killed from the transfer log due to connection loss.
251 * figure out whether to report success or failure.
253 * report success when at least one of the operations succeeded.
254 * or, to put the other way,
255 * only report failure, when both operations failed.
257 * what to do about the failures is handled elsewhere.
258 * what we need to do here is just: complete the master_bio.
260 * local completion error, if any, has been stored as ERR_PTR
261 * in private_bio within drbd_request_endio.
263 int ok
= (s
& RQ_LOCAL_OK
) || (s
& RQ_NET_OK
);
264 int error
= PTR_ERR(req
->private_bio
);
266 /* remove the request from the conflict detection
267 * respective block_id verification hash */
268 if (!drbd_interval_empty(&req
->i
)) {
269 struct rb_root
*root
;
272 root
= &mdev
->write_requests
;
274 root
= &mdev
->read_requests
;
275 drbd_remove_request_interval(root
, req
);
276 } else if (!(s
& RQ_POSTPONED
))
277 D_ASSERT((s
& (RQ_NET_MASK
& ~RQ_NET_DONE
)) == 0);
279 /* for writes we need to do some extra housekeeping */
281 _about_to_complete_local_write(mdev
, req
);
283 /* Update disk stats */
284 _drbd_end_io_acct(mdev
, req
);
286 if (!(s
& RQ_POSTPONED
)) {
287 m
->error
= ok
? 0 : (error
?: -EIO
);
288 m
->bio
= req
->master_bio
;
290 req
->master_bio
= NULL
;
293 if (s
& RQ_LOCAL_PENDING
)
296 if ((s
& RQ_NET_MASK
) == 0 || (s
& RQ_NET_DONE
)) {
297 /* this is disconnected (local only) operation,
298 * or protocol A, B, or C P_BARRIER_ACK,
299 * or killed from the transfer log due to connection loss. */
300 _req_is_done(mdev
, req
, rw
);
302 /* else: network part and not DONE yet. that is
303 * protocol A, B, or C, barrier ack still pending... */
306 static void _req_may_be_done_not_susp(struct drbd_request
*req
, struct bio_and_error
*m
)
308 struct drbd_conf
*mdev
= req
->w
.mdev
;
310 if (!drbd_suspended(mdev
))
311 _req_may_be_done(req
, m
);
314 /* obviously this could be coded as many single functions
315 * instead of one huge switch,
316 * or by putting the code directly in the respective locations
317 * (as it has been before).
319 * but having it this way
320 * enforces that it is all in this one place, where it is easier to audit,
321 * it makes it obvious that whatever "event" "happens" to a request should
322 * happen "atomically" within the req_lock,
323 * and it enforces that we have to think in a very structured manner
324 * about the "events" that may happen to a request during its life time ...
326 int __req_mod(struct drbd_request
*req
, enum drbd_req_event what
,
327 struct bio_and_error
*m
)
329 struct drbd_conf
*mdev
= req
->w
.mdev
;
338 dev_err(DEV
, "LOGIC BUG in %s:%u\n", __FILE__
, __LINE__
);
341 /* does not happen...
342 * initialization done in drbd_req_new
347 case TO_BE_SENT
: /* via network */
348 /* reached via __drbd_make_request
349 * and from w_read_retry_remote */
350 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
351 req
->rq_state
|= RQ_NET_PENDING
;
353 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
354 p
= nc
->wire_protocol
;
357 p
== DRBD_PROT_C
? RQ_EXP_WRITE_ACK
:
358 p
== DRBD_PROT_B
? RQ_EXP_RECEIVE_ACK
: 0;
359 inc_ap_pending(mdev
);
362 case TO_BE_SUBMITTED
: /* locally */
363 /* reached via __drbd_make_request */
364 D_ASSERT(!(req
->rq_state
& RQ_LOCAL_MASK
));
365 req
->rq_state
|= RQ_LOCAL_PENDING
;
369 if (req
->rq_state
& RQ_WRITE
)
370 mdev
->writ_cnt
+= req
->i
.size
>> 9;
372 mdev
->read_cnt
+= req
->i
.size
>> 9;
374 req
->rq_state
|= (RQ_LOCAL_COMPLETED
|RQ_LOCAL_OK
);
375 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
377 _req_may_be_done_not_susp(req
, m
);
382 req
->rq_state
|= RQ_LOCAL_ABORTED
;
383 if (req
->rq_state
& RQ_WRITE
)
384 _req_may_be_done_not_susp(req
, m
);
386 goto goto_queue_for_net_read
;
389 case WRITE_COMPLETED_WITH_ERROR
:
390 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
391 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
393 __drbd_chk_io_error(mdev
, false);
394 _req_may_be_done_not_susp(req
, m
);
398 case READ_AHEAD_COMPLETED_WITH_ERROR
:
399 /* it is legal to fail READA */
400 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
401 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
402 _req_may_be_done_not_susp(req
, m
);
406 case READ_COMPLETED_WITH_ERROR
:
407 drbd_set_out_of_sync(mdev
, req
->i
.sector
, req
->i
.size
);
409 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
410 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
412 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
414 __drbd_chk_io_error(mdev
, false);
417 goto_queue_for_net_read
:
419 /* no point in retrying if there is no good remote data,
420 * or we have no connection. */
421 if (mdev
->state
.pdsk
!= D_UP_TO_DATE
) {
422 _req_may_be_done_not_susp(req
, m
);
426 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
427 req
->rq_state
|= RQ_NET_PENDING
;
428 inc_ap_pending(mdev
);
429 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
431 case QUEUE_FOR_NET_READ
:
432 /* READ or READA, and
434 * or target area marked as invalid,
435 * or just got an io-error. */
436 /* from __drbd_make_request
437 * or from bio_endio during read io-error recovery */
439 /* so we can verify the handle in the answer packet
440 * corresponding hlist_del is in _req_may_be_done() */
441 D_ASSERT(drbd_interval_empty(&req
->i
));
442 drbd_insert_interval(&mdev
->read_requests
, &req
->i
);
444 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
446 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
447 req
->rq_state
|= RQ_NET_QUEUED
;
448 req
->w
.cb
= (req
->rq_state
& RQ_LOCAL_MASK
)
449 ? w_read_retry_remote
451 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
454 case QUEUE_FOR_NET_WRITE
:
455 /* assert something? */
456 /* from __drbd_make_request only */
458 /* corresponding hlist_del is in _req_may_be_done() */
459 D_ASSERT(drbd_interval_empty(&req
->i
));
460 drbd_insert_interval(&mdev
->write_requests
, &req
->i
);
463 * In case the req ended up on the transfer log before being
464 * queued on the worker, it could lead to this request being
465 * missed during cleanup after connection loss.
466 * So we have to do both operations here,
467 * within the same lock that protects the transfer log.
469 * _req_add_to_epoch(req); this has to be after the
470 * _maybe_start_new_epoch(req); which happened in
471 * __drbd_make_request, because we now may set the bit
472 * again ourselves to close the current epoch.
474 * Add req to the (now) current epoch (barrier). */
476 /* otherwise we may lose an unplug, which may cause some remote
477 * io-scheduler timeout to expire, increasing maximum latency,
478 * hurting performance. */
479 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
481 /* see __drbd_make_request,
482 * just after it grabs the req_lock */
483 D_ASSERT(test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
) == 0);
485 req
->epoch
= mdev
->tconn
->newest_tle
->br_number
;
487 /* increment size of current epoch */
488 mdev
->tconn
->newest_tle
->n_writes
++;
490 /* queue work item to send data */
491 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
492 req
->rq_state
|= RQ_NET_QUEUED
;
493 req
->w
.cb
= w_send_dblock
;
494 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
496 /* close the epoch, in case it outgrew the limit */
498 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
499 p
= nc
->max_epoch_size
;
501 if (mdev
->tconn
->newest_tle
->n_writes
>= p
)
506 case QUEUE_FOR_SEND_OOS
:
507 req
->rq_state
|= RQ_NET_QUEUED
;
508 req
->w
.cb
= w_send_out_of_sync
;
509 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
512 case OOS_HANDED_TO_NETWORK
:
513 /* actually the same */
515 /* treat it the same */
517 /* real cleanup will be done from tl_clear. just update flags
518 * so it is no longer marked as on the worker queue */
519 req
->rq_state
&= ~RQ_NET_QUEUED
;
520 /* if we did it right, tl_clear should be scheduled only after
521 * this, so this should not be necessary! */
522 _req_may_be_done_not_susp(req
, m
);
525 case HANDED_OVER_TO_NETWORK
:
526 /* assert something? */
527 if (bio_data_dir(req
->master_bio
) == WRITE
)
528 atomic_add(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
530 if (bio_data_dir(req
->master_bio
) == WRITE
&&
531 !(req
->rq_state
& (RQ_EXP_RECEIVE_ACK
| RQ_EXP_WRITE_ACK
))) {
532 /* this is what is dangerous about protocol A:
533 * pretend it was successfully written on the peer. */
534 if (req
->rq_state
& RQ_NET_PENDING
) {
535 dec_ap_pending(mdev
);
536 req
->rq_state
&= ~RQ_NET_PENDING
;
537 req
->rq_state
|= RQ_NET_OK
;
538 } /* else: neg-ack was faster... */
539 /* it is still not yet RQ_NET_DONE until the
540 * corresponding epoch barrier got acked as well,
541 * so we know what to dirty on connection loss */
543 req
->rq_state
&= ~RQ_NET_QUEUED
;
544 req
->rq_state
|= RQ_NET_SENT
;
545 /* because _drbd_send_zc_bio could sleep, and may want to
546 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
547 * "COMPLETED_OK" events came in, once we return from
548 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
549 * whether it is done already, and end it. */
550 _req_may_be_done_not_susp(req
, m
);
553 case READ_RETRY_REMOTE_CANCELED
:
554 req
->rq_state
&= ~RQ_NET_QUEUED
;
555 /* fall through, in case we raced with drbd_disconnect */
556 case CONNECTION_LOST_WHILE_PENDING
:
557 /* transfer log cleanup after connection loss */
558 /* assert something? */
559 if (req
->rq_state
& RQ_NET_PENDING
)
560 dec_ap_pending(mdev
);
561 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
562 req
->rq_state
|= RQ_NET_DONE
;
563 if (req
->rq_state
& RQ_NET_SENT
&& req
->rq_state
& RQ_WRITE
)
564 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
566 /* if it is still queued, we may not complete it here.
567 * it will be canceled soon. */
568 if (!(req
->rq_state
& RQ_NET_QUEUED
))
569 _req_may_be_done(req
, m
); /* Allowed while state.susp */
572 case WRITE_ACKED_BY_PEER_AND_SIS
:
573 req
->rq_state
|= RQ_NET_SIS
;
575 /* for discarded conflicting writes of multiple primaries,
576 * there is no need to keep anything in the tl, potential
577 * node crashes are covered by the activity log. */
578 req
->rq_state
|= RQ_NET_DONE
;
580 case WRITE_ACKED_BY_PEER
:
581 D_ASSERT(req
->rq_state
& RQ_EXP_WRITE_ACK
);
582 /* protocol C; successfully written on peer.
583 * Nothing to do here.
584 * We want to keep the tl in place for all protocols, to cater
585 * for volatile write-back caches on lower level devices.
587 * A barrier request is expected to have forced all prior
588 * requests onto stable storage, so completion of a barrier
589 * request could set NET_DONE right here, and not wait for the
590 * P_BARRIER_ACK, but that is an unnecessary optimization. */
593 /* this makes it effectively the same as for: */
594 case RECV_ACKED_BY_PEER
:
595 D_ASSERT(req
->rq_state
& RQ_EXP_RECEIVE_ACK
);
596 /* protocol B; pretends to be successfully written on peer.
597 * see also notes above in HANDED_OVER_TO_NETWORK about
600 req
->rq_state
|= RQ_NET_OK
;
601 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
602 dec_ap_pending(mdev
);
603 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
604 req
->rq_state
&= ~RQ_NET_PENDING
;
605 _req_may_be_done_not_susp(req
, m
);
609 D_ASSERT(req
->rq_state
& RQ_EXP_WRITE_ACK
);
610 /* If this node has already detected the write conflict, the
611 * worker will be waiting on misc_wait. Wake it up once this
612 * request has completed locally.
614 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
615 req
->rq_state
|= RQ_POSTPONED
;
616 _req_may_be_done_not_susp(req
, m
);
620 /* assert something? */
621 if (req
->rq_state
& RQ_NET_PENDING
) {
622 dec_ap_pending(mdev
);
623 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
625 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
627 req
->rq_state
|= RQ_NET_DONE
;
628 _req_may_be_done_not_susp(req
, m
);
629 /* else: done by HANDED_OVER_TO_NETWORK */
632 case FAIL_FROZEN_DISK_IO
:
633 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
636 _req_may_be_done(req
, m
); /* Allowed while state.susp */
639 case RESTART_FROZEN_DISK_IO
:
640 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
643 req
->rq_state
&= ~RQ_LOCAL_COMPLETED
;
646 if (bio_data_dir(req
->master_bio
) == WRITE
)
650 req
->w
.cb
= w_restart_disk_io
;
651 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
655 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
656 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
657 Trowing them out of the TL here by pretending we got a BARRIER_ACK
658 We ensure that the peer was not rebooted */
659 if (!(req
->rq_state
& RQ_NET_OK
)) {
661 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
662 rv
= req
->rq_state
& RQ_WRITE
? MR_WRITE
: MR_READ
;
666 /* else, fall through to BARRIER_ACKED */
669 if (!(req
->rq_state
& RQ_WRITE
))
672 if (req
->rq_state
& RQ_NET_PENDING
) {
673 /* barrier came in before all requests were acked.
674 * this is bad, because if the connection is lost now,
675 * we won't be able to clean them up... */
676 dev_err(DEV
, "FIXME (BARRIER_ACKED but pending)\n");
677 list_move(&req
->tl_requests
, &mdev
->tconn
->out_of_sequence_requests
);
679 if ((req
->rq_state
& RQ_NET_MASK
) != 0) {
680 req
->rq_state
|= RQ_NET_DONE
;
681 if (!(req
->rq_state
& (RQ_EXP_RECEIVE_ACK
| RQ_EXP_WRITE_ACK
)))
682 atomic_sub(req
->i
.size
>>9, &mdev
->ap_in_flight
);
684 _req_may_be_done(req
, m
); /* Allowed while state.susp */
688 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
689 dec_ap_pending(mdev
);
690 req
->rq_state
&= ~RQ_NET_PENDING
;
691 req
->rq_state
|= (RQ_NET_OK
|RQ_NET_DONE
);
692 _req_may_be_done_not_susp(req
, m
);
699 /* we may do a local read if:
700 * - we are consistent (of course),
701 * - or we are generally inconsistent,
702 * BUT we are still/already IN SYNC for this area.
703 * since size may be bigger than BM_BLOCK_SIZE,
704 * we may need to check several bits.
706 static bool drbd_may_do_local_read(struct drbd_conf
*mdev
, sector_t sector
, int size
)
708 unsigned long sbnr
, ebnr
;
709 sector_t esector
, nr_sectors
;
711 if (mdev
->state
.disk
== D_UP_TO_DATE
)
713 if (mdev
->state
.disk
!= D_INCONSISTENT
)
715 esector
= sector
+ (size
>> 9) - 1;
716 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
717 D_ASSERT(sector
< nr_sectors
);
718 D_ASSERT(esector
< nr_sectors
);
720 sbnr
= BM_SECT_TO_BIT(sector
);
721 ebnr
= BM_SECT_TO_BIT(esector
);
723 return drbd_bm_count_bits(mdev
, sbnr
, ebnr
) == 0;
727 * complete_conflicting_writes - wait for any conflicting write requests
729 * The write_requests tree contains all active write requests which we
730 * currently know about. Wait for any requests to complete which conflict with
733 static int complete_conflicting_writes(struct drbd_conf
*mdev
,
734 sector_t sector
, int size
)
737 struct drbd_interval
*i
;
740 i
= drbd_find_overlap(&mdev
->write_requests
, sector
, size
);
743 err
= drbd_wait_misc(mdev
, i
);
749 int __drbd_make_request(struct drbd_conf
*mdev
, struct bio
*bio
, unsigned long start_time
)
751 const int rw
= bio_rw(bio
);
752 const int size
= bio
->bi_size
;
753 const sector_t sector
= bio
->bi_sector
;
754 struct drbd_tl_epoch
*b
= NULL
;
755 struct drbd_request
*req
;
757 int local
, remote
, send_oos
= 0;
761 /* allocate outside of all locks; */
762 req
= drbd_req_new(mdev
, bio
);
765 /* only pass the error to the upper layers.
766 * if user cannot handle io errors, that's not our business. */
767 dev_err(DEV
, "could not kmalloc() req\n");
768 bio_endio(bio
, -ENOMEM
);
771 req
->start_time
= start_time
;
773 local
= get_ldev(mdev
);
775 bio_put(req
->private_bio
); /* or we get a bio leak */
776 req
->private_bio
= NULL
;
783 if (!drbd_may_do_local_read(mdev
, sector
, size
)) {
784 /* we could kick the syncer to
785 * sync this extent asap, wait for
786 * it, then continue locally.
787 * Or just issue the request remotely.
790 bio_put(req
->private_bio
);
791 req
->private_bio
= NULL
;
795 remote
= !local
&& mdev
->state
.pdsk
>= D_UP_TO_DATE
;
798 /* If we have a disk, but a READA request is mapped to remote,
799 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
800 * Just fail that READA request right here.
802 * THINK: maybe fail all READA when not local?
803 * or make this configurable...
804 * if network is slow, READA won't do any good.
806 if (rw
== READA
&& mdev
->state
.disk
>= D_INCONSISTENT
&& !local
) {
808 goto fail_and_free_req
;
811 /* For WRITES going to the local disk, grab a reference on the target
812 * extent. This waits for any resync activity in the corresponding
813 * resync extent to finish, and, if necessary, pulls in the target
814 * extent into the activity log, which involves further disk io because
815 * of transactional on-disk meta data updates. */
816 if (rw
== WRITE
&& local
&& !test_bit(AL_SUSPENDED
, &mdev
->flags
)) {
817 req
->rq_state
|= RQ_IN_ACT_LOG
;
818 drbd_al_begin_io(mdev
, &req
->i
);
821 remote
= remote
&& drbd_should_do_remote(mdev
->state
);
822 send_oos
= rw
== WRITE
&& drbd_should_send_out_of_sync(mdev
->state
);
823 D_ASSERT(!(remote
&& send_oos
));
825 if (!(local
|| remote
) && !drbd_suspended(mdev
)) {
826 if (__ratelimit(&drbd_ratelimit_state
))
827 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
829 goto fail_free_complete
;
832 /* For WRITE request, we have to make sure that we have an
833 * unused_spare_tle, in case we need to start a new epoch.
834 * I try to be smart and avoid to pre-allocate always "just in case",
835 * but there is a race between testing the bit and pointer outside the
836 * spinlock, and grabbing the spinlock.
837 * if we lost that race, we retry. */
838 if (rw
== WRITE
&& (remote
|| send_oos
) &&
839 mdev
->tconn
->unused_spare_tle
== NULL
&&
840 test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)) {
842 b
= kmalloc(sizeof(struct drbd_tl_epoch
), GFP_NOIO
);
844 dev_err(DEV
, "Failed to alloc barrier.\n");
846 goto fail_free_complete
;
850 /* GOOD, everything prepared, grab the spin_lock */
851 spin_lock_irq(&mdev
->tconn
->req_lock
);
854 err
= complete_conflicting_writes(mdev
, sector
, size
);
856 if (err
!= -ERESTARTSYS
)
857 _conn_request_state(mdev
->tconn
,
860 spin_unlock_irq(&mdev
->tconn
->req_lock
);
862 goto fail_free_complete
;
866 if (drbd_suspended(mdev
)) {
867 /* If we got suspended, use the retry mechanism of
868 generic_make_request() to restart processing of this
869 bio. In the next call to drbd_make_request
870 we sleep in inc_ap_bio() */
872 spin_unlock_irq(&mdev
->tconn
->req_lock
);
873 goto fail_free_complete
;
876 if (remote
|| send_oos
) {
877 remote
= drbd_should_do_remote(mdev
->state
);
878 send_oos
= rw
== WRITE
&& drbd_should_send_out_of_sync(mdev
->state
);
879 D_ASSERT(!(remote
&& send_oos
));
881 if (!(remote
|| send_oos
))
882 dev_warn(DEV
, "lost connection while grabbing the req_lock!\n");
883 if (!(local
|| remote
)) {
884 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
885 spin_unlock_irq(&mdev
->tconn
->req_lock
);
887 goto fail_free_complete
;
891 if (b
&& mdev
->tconn
->unused_spare_tle
== NULL
) {
892 mdev
->tconn
->unused_spare_tle
= b
;
895 if (rw
== WRITE
&& (remote
|| send_oos
) &&
896 mdev
->tconn
->unused_spare_tle
== NULL
&&
897 test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)) {
898 /* someone closed the current epoch
899 * while we were grabbing the spinlock */
900 spin_unlock_irq(&mdev
->tconn
->req_lock
);
901 goto allocate_barrier
;
905 /* Update disk stats */
906 _drbd_start_io_acct(mdev
, req
, bio
);
908 /* _maybe_start_new_epoch(mdev);
909 * If we need to generate a write barrier packet, we have to add the
910 * new epoch (barrier) object, and queue the barrier packet for sending,
911 * and queue the req's data after it _within the same lock_, otherwise
912 * we have race conditions were the reorder domains could be mixed up.
914 * Even read requests may start a new epoch and queue the corresponding
915 * barrier packet. To get the write ordering right, we only have to
916 * make sure that, if this is a write request and it triggered a
917 * barrier packet, this request is queued within the same spinlock. */
918 if ((remote
|| send_oos
) && mdev
->tconn
->unused_spare_tle
&&
919 test_and_clear_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)) {
920 _tl_add_barrier(mdev
->tconn
, mdev
->tconn
->unused_spare_tle
);
921 mdev
->tconn
->unused_spare_tle
= NULL
;
923 D_ASSERT(!(remote
&& rw
== WRITE
&&
924 test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)));
928 * Actually, 'local' may be wrong here already, since we may have failed
929 * to write to the meta data, and may become wrong anytime because of
930 * local io-error for some other request, which would lead to us
931 * "detaching" the local disk.
933 * 'remote' may become wrong any time because the network could fail.
935 * This is a harmless race condition, though, since it is handled
936 * correctly at the appropriate places; so it just defers the failure
937 * of the respective operation.
940 /* mark them early for readability.
941 * this just sets some state flags. */
943 _req_mod(req
, TO_BE_SENT
);
945 _req_mod(req
, TO_BE_SUBMITTED
);
947 list_add_tail(&req
->tl_requests
, &mdev
->tconn
->newest_tle
->requests
);
949 /* NOTE remote first: to get the concurrent write detection right,
950 * we must register the request before start of local IO. */
952 /* either WRITE and C_CONNECTED,
953 * or READ, and no local disk,
954 * or READ, but not in sync.
956 _req_mod(req
, (rw
== WRITE
)
957 ? QUEUE_FOR_NET_WRITE
958 : QUEUE_FOR_NET_READ
);
960 if (send_oos
&& drbd_set_out_of_sync(mdev
, sector
, size
))
961 _req_mod(req
, QUEUE_FOR_SEND_OOS
);
964 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
966 nc
->on_congestion
!= OC_BLOCK
&& mdev
->tconn
->agreed_pro_version
>= 96) {
970 atomic_read(&mdev
->ap_in_flight
) >= nc
->cong_fill
) {
971 dev_info(DEV
, "Congestion-fill threshold reached\n");
975 if (mdev
->act_log
->used
>= nc
->cong_extents
) {
976 dev_info(DEV
, "Congestion-extents threshold reached\n");
981 queue_barrier(mdev
); /* last barrier, after mirrored writes */
983 if (nc
->on_congestion
== OC_PULL_AHEAD
)
984 _drbd_set_state(_NS(mdev
, conn
, C_AHEAD
), 0, NULL
);
985 else /*nc->on_congestion == OC_DISCONNECT */
986 _drbd_set_state(_NS(mdev
, conn
, C_DISCONNECTING
), 0, NULL
);
991 spin_unlock_irq(&mdev
->tconn
->req_lock
);
992 kfree(b
); /* if someone else has beaten us to it... */
995 req
->private_bio
->bi_bdev
= mdev
->ldev
->backing_bdev
;
997 /* State may have changed since we grabbed our reference on the
998 * mdev->ldev member. Double check, and short-circuit to endio.
999 * In case the last activity log transaction failed to get on
1000 * stable storage, and this is a WRITE, we may not even submit
1002 if (get_ldev(mdev
)) {
1003 if (drbd_insert_fault(mdev
, rw
== WRITE
? DRBD_FAULT_DT_WR
1004 : rw
== READ
? DRBD_FAULT_DT_RD
1005 : DRBD_FAULT_DT_RA
))
1006 bio_endio(req
->private_bio
, -EIO
);
1008 generic_make_request(req
->private_bio
);
1011 bio_endio(req
->private_bio
, -EIO
);
1017 if (req
->rq_state
& RQ_IN_ACT_LOG
)
1018 drbd_al_complete_io(mdev
, &req
->i
);
1021 bio_put(req
->private_bio
);
1022 req
->private_bio
= NULL
;
1026 bio_endio(bio
, err
);
1035 int drbd_make_request(struct request_queue
*q
, struct bio
*bio
)
1037 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1038 unsigned long start_time
;
1040 start_time
= jiffies
;
1043 * what we "blindly" assume:
1045 D_ASSERT(bio
->bi_size
> 0);
1046 D_ASSERT(IS_ALIGNED(bio
->bi_size
, 512));
1049 return __drbd_make_request(mdev
, bio
, start_time
);
1052 /* This is called by bio_add_page().
1054 * q->max_hw_sectors and other global limits are already enforced there.
1056 * We need to call down to our lower level device,
1057 * in case it has special restrictions.
1059 * We also may need to enforce configured max-bio-bvecs limits.
1061 * As long as the BIO is empty we have to allow at least one bvec,
1062 * regardless of size and offset, so no need to ask lower levels.
1064 int drbd_merge_bvec(struct request_queue
*q
, struct bvec_merge_data
*bvm
, struct bio_vec
*bvec
)
1066 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1067 unsigned int bio_size
= bvm
->bi_size
;
1068 int limit
= DRBD_MAX_BIO_SIZE
;
1071 if (bio_size
&& get_ldev(mdev
)) {
1072 struct request_queue
* const b
=
1073 mdev
->ldev
->backing_bdev
->bd_disk
->queue
;
1074 if (b
->merge_bvec_fn
) {
1075 backing_limit
= b
->merge_bvec_fn(b
, bvm
, bvec
);
1076 limit
= min(limit
, backing_limit
);
1083 void request_timer_fn(unsigned long data
)
1085 struct drbd_conf
*mdev
= (struct drbd_conf
*) data
;
1086 struct drbd_tconn
*tconn
= mdev
->tconn
;
1087 struct drbd_request
*req
; /* oldest request */
1088 struct list_head
*le
;
1089 struct net_conf
*nc
;
1090 unsigned long ent
= 0, dt
= 0, et
, nt
; /* effective timeout = ko_count * timeout */
1093 nc
= rcu_dereference(tconn
->net_conf
);
1094 ent
= nc
? nc
->timeout
* HZ
/10 * nc
->ko_count
: 0;
1096 if (get_ldev(mdev
)) {
1097 dt
= rcu_dereference(mdev
->ldev
->disk_conf
)->disk_timeout
* HZ
/ 10;
1102 et
= min_not_zero(dt
, ent
);
1104 if (!et
|| (mdev
->state
.conn
< C_WF_REPORT_PARAMS
&& mdev
->state
.disk
<= D_FAILED
))
1105 return; /* Recurring timer stopped */
1107 spin_lock_irq(&tconn
->req_lock
);
1108 le
= &tconn
->oldest_tle
->requests
;
1109 if (list_empty(le
)) {
1110 spin_unlock_irq(&tconn
->req_lock
);
1111 mod_timer(&mdev
->request_timer
, jiffies
+ et
);
1116 req
= list_entry(le
, struct drbd_request
, tl_requests
);
1117 if (ent
&& req
->rq_state
& RQ_NET_PENDING
) {
1118 if (time_is_before_eq_jiffies(req
->start_time
+ ent
)) {
1119 dev_warn(DEV
, "Remote failed to finish a request within ko-count * timeout\n");
1120 _drbd_set_state(_NS(mdev
, conn
, C_TIMEOUT
), CS_VERBOSE
| CS_HARD
, NULL
);
1123 if (dt
&& req
->rq_state
& RQ_LOCAL_PENDING
) {
1124 if (time_is_before_eq_jiffies(req
->start_time
+ dt
)) {
1125 dev_warn(DEV
, "Local backing device failed to meet the disk-timeout\n");
1126 __drbd_chk_io_error(mdev
, 1);
1129 nt
= (time_is_before_eq_jiffies(req
->start_time
+ et
) ? jiffies
: req
->start_time
) + et
;
1130 spin_unlock_irq(&tconn
->req_lock
);
1131 mod_timer(&mdev
->request_timer
, nt
);