4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 drbd is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
15 drbd is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with drbd; see the file COPYING. If not, write to
22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
26 #include <linux/module.h>
28 #include <linux/slab.h>
29 #include <linux/drbd.h>
34 static bool drbd_may_do_local_read(struct drbd_conf
*mdev
, sector_t sector
, int size
);
36 /* Update disk stats at start of I/O request */
37 static void _drbd_start_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
, struct bio
*bio
)
39 const int rw
= bio_data_dir(bio
);
41 cpu
= part_stat_lock();
42 part_round_stats(cpu
, &mdev
->vdisk
->part0
);
43 part_stat_inc(cpu
, &mdev
->vdisk
->part0
, ios
[rw
]);
44 part_stat_add(cpu
, &mdev
->vdisk
->part0
, sectors
[rw
], bio_sectors(bio
));
45 (void) cpu
; /* The macro invocations above want the cpu argument, I do not like
46 the compiler warning about cpu only assigned but never used... */
47 part_inc_in_flight(&mdev
->vdisk
->part0
, rw
);
51 /* Update disk stats when completing request upwards */
52 static void _drbd_end_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
)
54 int rw
= bio_data_dir(req
->master_bio
);
55 unsigned long duration
= jiffies
- req
->start_time
;
57 cpu
= part_stat_lock();
58 part_stat_add(cpu
, &mdev
->vdisk
->part0
, ticks
[rw
], duration
);
59 part_round_stats(cpu
, &mdev
->vdisk
->part0
);
60 part_dec_in_flight(&mdev
->vdisk
->part0
, rw
);
64 static struct drbd_request
*drbd_req_new(struct drbd_conf
*mdev
,
67 struct drbd_request
*req
;
69 req
= mempool_alloc(drbd_request_mempool
, GFP_NOIO
);
73 drbd_req_make_private_bio(req
, bio_src
);
74 req
->rq_state
= bio_data_dir(bio_src
) == WRITE
? RQ_WRITE
: 0;
76 req
->master_bio
= bio_src
;
79 drbd_clear_interval(&req
->i
);
80 req
->i
.sector
= bio_src
->bi_sector
;
81 req
->i
.size
= bio_src
->bi_size
;
83 req
->i
.waiting
= false;
85 INIT_LIST_HEAD(&req
->tl_requests
);
86 INIT_LIST_HEAD(&req
->w
.list
);
91 static void drbd_req_free(struct drbd_request
*req
)
93 mempool_free(req
, drbd_request_mempool
);
96 /* rw is bio_data_dir(), only READ or WRITE */
97 static void _req_is_done(struct drbd_conf
*mdev
, struct drbd_request
*req
, const int rw
)
99 const unsigned long s
= req
->rq_state
;
101 /* remove it from the transfer log.
102 * well, only if it had been there in the first
103 * place... if it had not (local only or conflicting
104 * and never sent), it should still be "empty" as
105 * initialized in drbd_req_new(), so we can list_del() it
106 * here unconditionally */
107 list_del_init(&req
->tl_requests
);
109 /* if it was a write, we may have to set the corresponding
110 * bit(s) out-of-sync first. If it had a local part, we need to
111 * release the reference to the activity log. */
113 /* Set out-of-sync unless both OK flags are set
114 * (local only or remote failed).
115 * Other places where we set out-of-sync:
116 * READ with local io-error */
117 if (!(s
& RQ_NET_OK
) || !(s
& RQ_LOCAL_OK
))
118 drbd_set_out_of_sync(mdev
, req
->i
.sector
, req
->i
.size
);
120 if ((s
& RQ_NET_OK
) && (s
& RQ_LOCAL_OK
) && (s
& RQ_NET_SIS
))
121 drbd_set_in_sync(mdev
, req
->i
.sector
, req
->i
.size
);
123 /* one might be tempted to move the drbd_al_complete_io
124 * to the local io completion callback drbd_request_endio.
125 * but, if this was a mirror write, we may only
126 * drbd_al_complete_io after this is RQ_NET_DONE,
127 * otherwise the extent could be dropped from the al
128 * before it has actually been written on the peer.
129 * if we crash before our peer knows about the request,
130 * but after the extent has been dropped from the al,
131 * we would forget to resync the corresponding extent.
133 if (s
& RQ_LOCAL_MASK
) {
134 if (get_ldev_if_state(mdev
, D_FAILED
)) {
135 if (s
& RQ_IN_ACT_LOG
)
136 drbd_al_complete_io(mdev
, &req
->i
);
138 } else if (__ratelimit(&drbd_ratelimit_state
)) {
139 dev_warn(DEV
, "Should have called drbd_al_complete_io(, %llu, %u), "
140 "but my Disk seems to have failed :(\n",
141 (unsigned long long) req
->i
.sector
, req
->i
.size
);
146 if (s
& RQ_POSTPONED
)
147 drbd_restart_write(req
);
152 static void queue_barrier(struct drbd_conf
*mdev
)
154 struct drbd_tl_epoch
*b
;
155 struct drbd_tconn
*tconn
= mdev
->tconn
;
157 /* We are within the req_lock. Once we queued the barrier for sending,
158 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
159 * barrier/epoch object is added. This is the only place this bit is
160 * set. It indicates that the barrier for this epoch is already queued,
161 * and no new epoch has been created yet. */
162 if (test_bit(CREATE_BARRIER
, &tconn
->flags
))
165 b
= tconn
->newest_tle
;
166 b
->w
.cb
= w_send_barrier
;
168 /* inc_ap_pending done here, so we won't
169 * get imbalanced on connection loss.
170 * dec_ap_pending will be done in got_BarrierAck
171 * or (on connection loss) in tl_clear. */
172 inc_ap_pending(mdev
);
173 drbd_queue_work(&tconn
->data
.work
, &b
->w
);
174 set_bit(CREATE_BARRIER
, &tconn
->flags
);
177 static void _about_to_complete_local_write(struct drbd_conf
*mdev
,
178 struct drbd_request
*req
)
180 const unsigned long s
= req
->rq_state
;
182 /* Before we can signal completion to the upper layers,
183 * we may need to close the current epoch.
184 * We can skip this, if this request has not even been sent, because we
185 * did not have a fully established connection yet/anymore, during
186 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
188 if (mdev
->state
.conn
>= C_CONNECTED
&&
189 (s
& RQ_NET_SENT
) != 0 &&
190 req
->epoch
== mdev
->tconn
->newest_tle
->br_number
)
194 void complete_master_bio(struct drbd_conf
*mdev
,
195 struct bio_and_error
*m
)
197 bio_endio(m
->bio
, m
->error
);
202 static void drbd_remove_request_interval(struct rb_root
*root
,
203 struct drbd_request
*req
)
205 struct drbd_conf
*mdev
= req
->w
.mdev
;
206 struct drbd_interval
*i
= &req
->i
;
208 drbd_remove_interval(root
, i
);
210 /* Wake up any processes waiting for this request to complete. */
212 wake_up(&mdev
->misc_wait
);
215 /* Helper for __req_mod().
216 * Set m->bio to the master bio, if it is fit to be completed,
217 * or leave it alone (it is initialized to NULL in __req_mod),
218 * if it has already been completed, or cannot be completed yet.
219 * If m->bio is set, the error status to be returned is placed in m->error.
221 void _req_may_be_done(struct drbd_request
*req
, struct bio_and_error
*m
)
223 const unsigned long s
= req
->rq_state
;
224 struct drbd_conf
*mdev
= req
->w
.mdev
;
225 int rw
= req
->rq_state
& RQ_WRITE
? WRITE
: READ
;
227 /* we must not complete the master bio, while it is
228 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
229 * not yet acknowledged by the peer
230 * not yet completed by the local io subsystem
231 * these flags may get cleared in any order by
234 * the bio_endio completion callbacks.
236 if (s
& RQ_LOCAL_PENDING
&& !(s
& RQ_LOCAL_ABORTED
))
238 if (req
->i
.waiting
) {
239 /* Retry all conflicting peer requests. */
240 wake_up(&mdev
->misc_wait
);
242 if (s
& RQ_NET_QUEUED
)
244 if (s
& RQ_NET_PENDING
)
247 if (req
->master_bio
) {
248 /* this is DATA_RECEIVED (remote read)
249 * or protocol C P_WRITE_ACK
250 * or protocol B P_RECV_ACK
251 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
252 * or canceled or failed,
253 * or killed from the transfer log due to connection loss.
257 * figure out whether to report success or failure.
259 * report success when at least one of the operations succeeded.
260 * or, to put the other way,
261 * only report failure, when both operations failed.
263 * what to do about the failures is handled elsewhere.
264 * what we need to do here is just: complete the master_bio.
266 * local completion error, if any, has been stored as ERR_PTR
267 * in private_bio within drbd_request_endio.
269 int ok
= (s
& RQ_LOCAL_OK
) || (s
& RQ_NET_OK
);
270 int error
= PTR_ERR(req
->private_bio
);
272 /* remove the request from the conflict detection
273 * respective block_id verification hash */
274 if (!drbd_interval_empty(&req
->i
)) {
275 struct rb_root
*root
;
278 root
= &mdev
->write_requests
;
280 root
= &mdev
->read_requests
;
281 drbd_remove_request_interval(root
, req
);
282 } else if (!(s
& RQ_POSTPONED
))
283 D_ASSERT((s
& (RQ_NET_MASK
& ~RQ_NET_DONE
)) == 0);
285 /* for writes we need to do some extra housekeeping */
287 _about_to_complete_local_write(mdev
, req
);
289 /* Update disk stats */
290 _drbd_end_io_acct(mdev
, req
);
292 if (!(s
& RQ_POSTPONED
)) {
293 m
->error
= ok
? 0 : (error
?: -EIO
);
294 m
->bio
= req
->master_bio
;
295 req
->master_bio
= NULL
;
297 /* Assert that this will be _req_is_done()
298 * with this very invokation. */
300 * what about (RQ_LOCAL_PENDING | RQ_LOCAL_ABORTED)?
302 D_ASSERT(!(s
& RQ_LOCAL_PENDING
));
303 D_ASSERT(s
& RQ_NET_DONE
);
307 if (s
& RQ_LOCAL_PENDING
)
310 if ((s
& RQ_NET_MASK
) == 0 || (s
& RQ_NET_DONE
)) {
311 /* this is disconnected (local only) operation,
312 * or protocol A, B, or C P_BARRIER_ACK,
313 * or killed from the transfer log due to connection loss. */
314 _req_is_done(mdev
, req
, rw
);
316 /* else: network part and not DONE yet. that is
317 * protocol A, B, or C, barrier ack still pending... */
320 static void _req_may_be_done_not_susp(struct drbd_request
*req
, struct bio_and_error
*m
)
322 struct drbd_conf
*mdev
= req
->w
.mdev
;
324 if (!drbd_suspended(mdev
))
325 _req_may_be_done(req
, m
);
328 /* obviously this could be coded as many single functions
329 * instead of one huge switch,
330 * or by putting the code directly in the respective locations
331 * (as it has been before).
333 * but having it this way
334 * enforces that it is all in this one place, where it is easier to audit,
335 * it makes it obvious that whatever "event" "happens" to a request should
336 * happen "atomically" within the req_lock,
337 * and it enforces that we have to think in a very structured manner
338 * about the "events" that may happen to a request during its life time ...
340 int __req_mod(struct drbd_request
*req
, enum drbd_req_event what
,
341 struct bio_and_error
*m
)
343 struct drbd_conf
*mdev
= req
->w
.mdev
;
352 dev_err(DEV
, "LOGIC BUG in %s:%u\n", __FILE__
, __LINE__
);
355 /* does not happen...
356 * initialization done in drbd_req_new
361 case TO_BE_SENT
: /* via network */
362 /* reached via __drbd_make_request
363 * and from w_read_retry_remote */
364 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
365 req
->rq_state
|= RQ_NET_PENDING
;
367 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
368 p
= nc
->wire_protocol
;
371 p
== DRBD_PROT_C
? RQ_EXP_WRITE_ACK
:
372 p
== DRBD_PROT_B
? RQ_EXP_RECEIVE_ACK
: 0;
373 inc_ap_pending(mdev
);
376 case TO_BE_SUBMITTED
: /* locally */
377 /* reached via __drbd_make_request */
378 D_ASSERT(!(req
->rq_state
& RQ_LOCAL_MASK
));
379 req
->rq_state
|= RQ_LOCAL_PENDING
;
383 if (req
->rq_state
& RQ_WRITE
)
384 mdev
->writ_cnt
+= req
->i
.size
>> 9;
386 mdev
->read_cnt
+= req
->i
.size
>> 9;
388 req
->rq_state
|= (RQ_LOCAL_COMPLETED
|RQ_LOCAL_OK
);
389 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
391 _req_may_be_done_not_susp(req
, m
);
396 req
->rq_state
|= RQ_LOCAL_ABORTED
;
397 if (req
->rq_state
& RQ_WRITE
)
398 _req_may_be_done_not_susp(req
, m
);
400 goto goto_queue_for_net_read
;
403 case WRITE_COMPLETED_WITH_ERROR
:
404 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
405 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
407 __drbd_chk_io_error(mdev
, false);
408 _req_may_be_done_not_susp(req
, m
);
412 case READ_AHEAD_COMPLETED_WITH_ERROR
:
413 /* it is legal to fail READA */
414 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
415 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
416 _req_may_be_done_not_susp(req
, m
);
420 case READ_COMPLETED_WITH_ERROR
:
421 drbd_set_out_of_sync(mdev
, req
->i
.sector
, req
->i
.size
);
423 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
424 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
426 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
428 __drbd_chk_io_error(mdev
, false);
431 goto_queue_for_net_read
:
433 /* no point in retrying if there is no good remote data,
434 * or we have no connection. */
435 if (mdev
->state
.pdsk
!= D_UP_TO_DATE
) {
436 _req_may_be_done_not_susp(req
, m
);
440 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
441 req
->rq_state
|= RQ_NET_PENDING
;
442 inc_ap_pending(mdev
);
443 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
445 case QUEUE_FOR_NET_READ
:
446 /* READ or READA, and
448 * or target area marked as invalid,
449 * or just got an io-error. */
450 /* from __drbd_make_request
451 * or from bio_endio during read io-error recovery */
453 /* so we can verify the handle in the answer packet
454 * corresponding hlist_del is in _req_may_be_done() */
455 D_ASSERT(drbd_interval_empty(&req
->i
));
456 drbd_insert_interval(&mdev
->read_requests
, &req
->i
);
458 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
460 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
461 req
->rq_state
|= RQ_NET_QUEUED
;
462 req
->w
.cb
= (req
->rq_state
& RQ_LOCAL_MASK
)
463 ? w_read_retry_remote
465 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
468 case QUEUE_FOR_NET_WRITE
:
469 /* assert something? */
470 /* from __drbd_make_request only */
472 /* corresponding hlist_del is in _req_may_be_done() */
473 D_ASSERT(drbd_interval_empty(&req
->i
));
474 drbd_insert_interval(&mdev
->write_requests
, &req
->i
);
477 * In case the req ended up on the transfer log before being
478 * queued on the worker, it could lead to this request being
479 * missed during cleanup after connection loss.
480 * So we have to do both operations here,
481 * within the same lock that protects the transfer log.
483 * _req_add_to_epoch(req); this has to be after the
484 * _maybe_start_new_epoch(req); which happened in
485 * __drbd_make_request, because we now may set the bit
486 * again ourselves to close the current epoch.
488 * Add req to the (now) current epoch (barrier). */
490 /* otherwise we may lose an unplug, which may cause some remote
491 * io-scheduler timeout to expire, increasing maximum latency,
492 * hurting performance. */
493 set_bit(UNPLUG_REMOTE
, &mdev
->flags
);
495 /* see __drbd_make_request,
496 * just after it grabs the req_lock */
497 D_ASSERT(test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
) == 0);
499 req
->epoch
= mdev
->tconn
->newest_tle
->br_number
;
501 /* increment size of current epoch */
502 mdev
->tconn
->newest_tle
->n_writes
++;
504 /* queue work item to send data */
505 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
506 req
->rq_state
|= RQ_NET_QUEUED
;
507 req
->w
.cb
= w_send_dblock
;
508 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
510 /* close the epoch, in case it outgrew the limit */
512 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
513 p
= nc
->max_epoch_size
;
515 if (mdev
->tconn
->newest_tle
->n_writes
>= p
)
520 case QUEUE_FOR_SEND_OOS
:
521 req
->rq_state
|= RQ_NET_QUEUED
;
522 req
->w
.cb
= w_send_out_of_sync
;
523 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
526 case OOS_HANDED_TO_NETWORK
:
527 /* actually the same */
529 /* treat it the same */
531 /* real cleanup will be done from tl_clear. just update flags
532 * so it is no longer marked as on the worker queue */
533 req
->rq_state
&= ~RQ_NET_QUEUED
;
534 /* if we did it right, tl_clear should be scheduled only after
535 * this, so this should not be necessary! */
536 _req_may_be_done_not_susp(req
, m
);
539 case HANDED_OVER_TO_NETWORK
:
540 /* assert something? */
541 if (bio_data_dir(req
->master_bio
) == WRITE
)
542 atomic_add(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
544 if (bio_data_dir(req
->master_bio
) == WRITE
&&
545 !(req
->rq_state
& (RQ_EXP_RECEIVE_ACK
| RQ_EXP_WRITE_ACK
))) {
546 /* this is what is dangerous about protocol A:
547 * pretend it was successfully written on the peer. */
548 if (req
->rq_state
& RQ_NET_PENDING
) {
549 dec_ap_pending(mdev
);
550 req
->rq_state
&= ~RQ_NET_PENDING
;
551 req
->rq_state
|= RQ_NET_OK
;
552 } /* else: neg-ack was faster... */
553 /* it is still not yet RQ_NET_DONE until the
554 * corresponding epoch barrier got acked as well,
555 * so we know what to dirty on connection loss */
557 req
->rq_state
&= ~RQ_NET_QUEUED
;
558 req
->rq_state
|= RQ_NET_SENT
;
559 /* because _drbd_send_zc_bio could sleep, and may want to
560 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
561 * "COMPLETED_OK" events came in, once we return from
562 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
563 * whether it is done already, and end it. */
564 _req_may_be_done_not_susp(req
, m
);
567 case READ_RETRY_REMOTE_CANCELED
:
568 req
->rq_state
&= ~RQ_NET_QUEUED
;
569 /* fall through, in case we raced with drbd_disconnect */
570 case CONNECTION_LOST_WHILE_PENDING
:
571 /* transfer log cleanup after connection loss */
572 /* assert something? */
573 if (req
->rq_state
& RQ_NET_PENDING
)
574 dec_ap_pending(mdev
);
576 p
= !(req
->rq_state
& RQ_WRITE
) && req
->rq_state
& RQ_NET_PENDING
;
578 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
579 req
->rq_state
|= RQ_NET_DONE
;
580 if (req
->rq_state
& RQ_NET_SENT
&& req
->rq_state
& RQ_WRITE
)
581 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
583 /* if it is still queued, we may not complete it here.
584 * it will be canceled soon. */
585 if (!(req
->rq_state
& RQ_NET_QUEUED
)) {
587 goto goto_read_retry_local
;
588 _req_may_be_done(req
, m
); /* Allowed while state.susp */
592 case WRITE_ACKED_BY_PEER_AND_SIS
:
593 req
->rq_state
|= RQ_NET_SIS
;
595 /* for discarded conflicting writes of multiple primaries,
596 * there is no need to keep anything in the tl, potential
597 * node crashes are covered by the activity log. */
598 req
->rq_state
|= RQ_NET_DONE
;
600 case WRITE_ACKED_BY_PEER
:
601 D_ASSERT(req
->rq_state
& RQ_EXP_WRITE_ACK
);
602 /* protocol C; successfully written on peer.
603 * Nothing to do here.
604 * We want to keep the tl in place for all protocols, to cater
605 * for volatile write-back caches on lower level devices.
607 * A barrier request is expected to have forced all prior
608 * requests onto stable storage, so completion of a barrier
609 * request could set NET_DONE right here, and not wait for the
610 * P_BARRIER_ACK, but that is an unnecessary optimization. */
613 /* this makes it effectively the same as for: */
614 case RECV_ACKED_BY_PEER
:
615 D_ASSERT(req
->rq_state
& RQ_EXP_RECEIVE_ACK
);
616 /* protocol B; pretends to be successfully written on peer.
617 * see also notes above in HANDED_OVER_TO_NETWORK about
620 req
->rq_state
|= RQ_NET_OK
;
621 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
622 dec_ap_pending(mdev
);
623 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
624 req
->rq_state
&= ~RQ_NET_PENDING
;
625 _req_may_be_done_not_susp(req
, m
);
629 D_ASSERT(req
->rq_state
& RQ_EXP_WRITE_ACK
);
630 /* If this node has already detected the write conflict, the
631 * worker will be waiting on misc_wait. Wake it up once this
632 * request has completed locally.
634 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
635 req
->rq_state
|= RQ_POSTPONED
;
636 _req_may_be_done_not_susp(req
, m
);
640 /* assert something? */
641 if (req
->rq_state
& RQ_NET_PENDING
) {
642 dec_ap_pending(mdev
);
643 if (req
->rq_state
& RQ_WRITE
)
644 atomic_sub(req
->i
.size
>> 9, &mdev
->ap_in_flight
);
646 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
648 req
->rq_state
|= RQ_NET_DONE
;
650 if (!(req
->rq_state
& RQ_WRITE
))
651 goto goto_read_retry_local
;
653 _req_may_be_done_not_susp(req
, m
);
654 /* else: done by HANDED_OVER_TO_NETWORK */
657 goto_read_retry_local
:
658 if (!drbd_may_do_local_read(mdev
, req
->i
.sector
, req
->i
.size
)) {
659 _req_may_be_done_not_susp(req
, m
);
662 D_ASSERT(!(req
->rq_state
& RQ_LOCAL_PENDING
));
663 req
->rq_state
|= RQ_LOCAL_PENDING
;
666 req
->w
.cb
= w_restart_disk_io
;
667 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
670 case FAIL_FROZEN_DISK_IO
:
671 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
674 _req_may_be_done(req
, m
); /* Allowed while state.susp */
677 case RESTART_FROZEN_DISK_IO
:
678 if (!(req
->rq_state
& RQ_LOCAL_COMPLETED
))
681 req
->rq_state
&= ~RQ_LOCAL_COMPLETED
;
684 if (bio_data_dir(req
->master_bio
) == WRITE
)
688 req
->w
.cb
= w_restart_disk_io
;
689 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
693 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
694 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
695 Trowing them out of the TL here by pretending we got a BARRIER_ACK
696 We ensure that the peer was not rebooted */
697 if (!(req
->rq_state
& RQ_NET_OK
)) {
699 drbd_queue_work(&mdev
->tconn
->data
.work
, &req
->w
);
700 rv
= req
->rq_state
& RQ_WRITE
? MR_WRITE
: MR_READ
;
704 /* else, fall through to BARRIER_ACKED */
707 if (!(req
->rq_state
& RQ_WRITE
))
710 if (req
->rq_state
& RQ_NET_PENDING
) {
711 /* barrier came in before all requests were acked.
712 * this is bad, because if the connection is lost now,
713 * we won't be able to clean them up... */
714 dev_err(DEV
, "FIXME (BARRIER_ACKED but pending)\n");
715 list_move(&req
->tl_requests
, &mdev
->tconn
->out_of_sequence_requests
);
717 if ((req
->rq_state
& RQ_NET_MASK
) != 0) {
718 req
->rq_state
|= RQ_NET_DONE
;
719 if (!(req
->rq_state
& (RQ_EXP_RECEIVE_ACK
| RQ_EXP_WRITE_ACK
)))
720 atomic_sub(req
->i
.size
>>9, &mdev
->ap_in_flight
);
722 _req_may_be_done(req
, m
); /* Allowed while state.susp */
726 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
727 dec_ap_pending(mdev
);
728 req
->rq_state
&= ~RQ_NET_PENDING
;
729 req
->rq_state
|= (RQ_NET_OK
|RQ_NET_DONE
);
730 _req_may_be_done_not_susp(req
, m
);
737 /* we may do a local read if:
738 * - we are consistent (of course),
739 * - or we are generally inconsistent,
740 * BUT we are still/already IN SYNC for this area.
741 * since size may be bigger than BM_BLOCK_SIZE,
742 * we may need to check several bits.
744 static bool drbd_may_do_local_read(struct drbd_conf
*mdev
, sector_t sector
, int size
)
746 unsigned long sbnr
, ebnr
;
747 sector_t esector
, nr_sectors
;
749 if (mdev
->state
.disk
== D_UP_TO_DATE
)
751 if (mdev
->state
.disk
!= D_INCONSISTENT
)
753 esector
= sector
+ (size
>> 9) - 1;
754 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
755 D_ASSERT(sector
< nr_sectors
);
756 D_ASSERT(esector
< nr_sectors
);
758 sbnr
= BM_SECT_TO_BIT(sector
);
759 ebnr
= BM_SECT_TO_BIT(esector
);
761 return drbd_bm_count_bits(mdev
, sbnr
, ebnr
) == 0;
764 static bool remote_due_to_read_balancing(struct drbd_conf
*mdev
, sector_t sector
)
766 enum drbd_read_balancing rbm
;
767 struct backing_dev_info
*bdi
;
770 if (mdev
->state
.pdsk
< D_UP_TO_DATE
)
774 rbm
= rcu_dereference(mdev
->ldev
->disk_conf
)->read_balancing
;
778 case RB_CONGESTED_REMOTE
:
779 bdi
= &mdev
->ldev
->backing_bdev
->bd_disk
->queue
->backing_dev_info
;
780 return bdi_read_congested(bdi
);
781 case RB_LEAST_PENDING
:
782 return atomic_read(&mdev
->local_cnt
) >
783 atomic_read(&mdev
->ap_pending_cnt
) + atomic_read(&mdev
->rs_pending_cnt
);
784 case RB_32K_STRIPING
: /* stripe_shift = 15 */
785 case RB_64K_STRIPING
:
786 case RB_128K_STRIPING
:
787 case RB_256K_STRIPING
:
788 case RB_512K_STRIPING
:
789 case RB_1M_STRIPING
: /* stripe_shift = 20 */
790 stripe_shift
= (rbm
- RB_32K_STRIPING
+ 15);
791 return (sector
>> (stripe_shift
- 9)) & 1;
793 return test_and_change_bit(READ_BALANCE_RR
, &mdev
->flags
);
794 case RB_PREFER_REMOTE
:
796 case RB_PREFER_LOCAL
:
803 * complete_conflicting_writes - wait for any conflicting write requests
805 * The write_requests tree contains all active write requests which we
806 * currently know about. Wait for any requests to complete which conflict with
809 static int complete_conflicting_writes(struct drbd_conf
*mdev
,
810 sector_t sector
, int size
)
813 struct drbd_interval
*i
;
816 i
= drbd_find_overlap(&mdev
->write_requests
, sector
, size
);
819 err
= drbd_wait_misc(mdev
, i
);
825 int __drbd_make_request(struct drbd_conf
*mdev
, struct bio
*bio
, unsigned long start_time
)
827 const int rw
= bio_rw(bio
);
828 const int size
= bio
->bi_size
;
829 const sector_t sector
= bio
->bi_sector
;
830 struct drbd_tl_epoch
*b
= NULL
;
831 struct drbd_request
*req
;
833 int local
, remote
, send_oos
= 0;
836 union drbd_dev_state s
;
838 /* allocate outside of all locks; */
839 req
= drbd_req_new(mdev
, bio
);
842 /* only pass the error to the upper layers.
843 * if user cannot handle io errors, that's not our business. */
844 dev_err(DEV
, "could not kmalloc() req\n");
845 bio_endio(bio
, -ENOMEM
);
848 req
->start_time
= start_time
;
850 local
= get_ldev(mdev
);
852 bio_put(req
->private_bio
); /* or we get a bio leak */
853 req
->private_bio
= NULL
;
860 if (!drbd_may_do_local_read(mdev
, sector
, size
) ||
861 remote_due_to_read_balancing(mdev
, sector
)) {
862 /* we could kick the syncer to
863 * sync this extent asap, wait for
864 * it, then continue locally.
865 * Or just issue the request remotely.
868 bio_put(req
->private_bio
);
869 req
->private_bio
= NULL
;
873 remote
= !local
&& mdev
->state
.pdsk
>= D_UP_TO_DATE
;
876 /* If we have a disk, but a READA request is mapped to remote,
877 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
878 * Just fail that READA request right here.
880 * THINK: maybe fail all READA when not local?
881 * or make this configurable...
882 * if network is slow, READA won't do any good.
884 if (rw
== READA
&& mdev
->state
.disk
>= D_INCONSISTENT
&& !local
) {
886 goto fail_and_free_req
;
889 /* For WRITES going to the local disk, grab a reference on the target
890 * extent. This waits for any resync activity in the corresponding
891 * resync extent to finish, and, if necessary, pulls in the target
892 * extent into the activity log, which involves further disk io because
893 * of transactional on-disk meta data updates. */
894 if (rw
== WRITE
&& local
&& !test_bit(AL_SUSPENDED
, &mdev
->flags
)) {
895 req
->rq_state
|= RQ_IN_ACT_LOG
;
896 drbd_al_begin_io(mdev
, &req
->i
);
900 remote
= remote
&& drbd_should_do_remote(s
);
901 send_oos
= rw
== WRITE
&& drbd_should_send_out_of_sync(s
);
902 D_ASSERT(!(remote
&& send_oos
));
904 if (!(local
|| remote
) && !drbd_suspended(mdev
)) {
905 if (__ratelimit(&drbd_ratelimit_state
))
906 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
908 goto fail_free_complete
;
911 /* For WRITE request, we have to make sure that we have an
912 * unused_spare_tle, in case we need to start a new epoch.
913 * I try to be smart and avoid to pre-allocate always "just in case",
914 * but there is a race between testing the bit and pointer outside the
915 * spinlock, and grabbing the spinlock.
916 * if we lost that race, we retry. */
917 if (rw
== WRITE
&& (remote
|| send_oos
) &&
918 mdev
->tconn
->unused_spare_tle
== NULL
&&
919 test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)) {
921 b
= kmalloc(sizeof(struct drbd_tl_epoch
), GFP_NOIO
);
923 dev_err(DEV
, "Failed to alloc barrier.\n");
925 goto fail_free_complete
;
929 /* GOOD, everything prepared, grab the spin_lock */
930 spin_lock_irq(&mdev
->tconn
->req_lock
);
933 err
= complete_conflicting_writes(mdev
, sector
, size
);
935 if (err
!= -ERESTARTSYS
)
936 _conn_request_state(mdev
->tconn
,
939 spin_unlock_irq(&mdev
->tconn
->req_lock
);
941 goto fail_free_complete
;
945 if (drbd_suspended(mdev
)) {
946 /* If we got suspended, use the retry mechanism in
947 drbd_make_request() to restart processing of this
948 bio. In the next call to drbd_make_request
949 we sleep in inc_ap_bio() */
951 spin_unlock_irq(&mdev
->tconn
->req_lock
);
952 goto fail_free_complete
;
955 if (remote
|| send_oos
) {
956 remote
= drbd_should_do_remote(mdev
->state
);
957 send_oos
= rw
== WRITE
&& drbd_should_send_out_of_sync(mdev
->state
);
958 D_ASSERT(!(remote
&& send_oos
));
960 if (!(remote
|| send_oos
))
961 dev_warn(DEV
, "lost connection while grabbing the req_lock!\n");
962 if (!(local
|| remote
)) {
963 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
964 spin_unlock_irq(&mdev
->tconn
->req_lock
);
966 goto fail_free_complete
;
970 if (b
&& mdev
->tconn
->unused_spare_tle
== NULL
) {
971 mdev
->tconn
->unused_spare_tle
= b
;
974 if (rw
== WRITE
&& (remote
|| send_oos
) &&
975 mdev
->tconn
->unused_spare_tle
== NULL
&&
976 test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)) {
977 /* someone closed the current epoch
978 * while we were grabbing the spinlock */
979 spin_unlock_irq(&mdev
->tconn
->req_lock
);
980 goto allocate_barrier
;
984 /* Update disk stats */
985 _drbd_start_io_acct(mdev
, req
, bio
);
987 /* _maybe_start_new_epoch(mdev);
988 * If we need to generate a write barrier packet, we have to add the
989 * new epoch (barrier) object, and queue the barrier packet for sending,
990 * and queue the req's data after it _within the same lock_, otherwise
991 * we have race conditions were the reorder domains could be mixed up.
993 * Even read requests may start a new epoch and queue the corresponding
994 * barrier packet. To get the write ordering right, we only have to
995 * make sure that, if this is a write request and it triggered a
996 * barrier packet, this request is queued within the same spinlock. */
997 if ((remote
|| send_oos
) && mdev
->tconn
->unused_spare_tle
&&
998 test_and_clear_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)) {
999 _tl_add_barrier(mdev
->tconn
, mdev
->tconn
->unused_spare_tle
);
1000 mdev
->tconn
->unused_spare_tle
= NULL
;
1002 D_ASSERT(!(remote
&& rw
== WRITE
&&
1003 test_bit(CREATE_BARRIER
, &mdev
->tconn
->flags
)));
1007 * Actually, 'local' may be wrong here already, since we may have failed
1008 * to write to the meta data, and may become wrong anytime because of
1009 * local io-error for some other request, which would lead to us
1010 * "detaching" the local disk.
1012 * 'remote' may become wrong any time because the network could fail.
1014 * This is a harmless race condition, though, since it is handled
1015 * correctly at the appropriate places; so it just defers the failure
1016 * of the respective operation.
1019 /* mark them early for readability.
1020 * this just sets some state flags. */
1022 _req_mod(req
, TO_BE_SENT
);
1024 _req_mod(req
, TO_BE_SUBMITTED
);
1026 list_add_tail(&req
->tl_requests
, &mdev
->tconn
->newest_tle
->requests
);
1028 /* NOTE remote first: to get the concurrent write detection right,
1029 * we must register the request before start of local IO. */
1031 /* either WRITE and C_CONNECTED,
1032 * or READ, and no local disk,
1033 * or READ, but not in sync.
1035 _req_mod(req
, (rw
== WRITE
)
1036 ? QUEUE_FOR_NET_WRITE
1037 : QUEUE_FOR_NET_READ
);
1039 if (send_oos
&& drbd_set_out_of_sync(mdev
, sector
, size
))
1040 _req_mod(req
, QUEUE_FOR_SEND_OOS
);
1043 nc
= rcu_dereference(mdev
->tconn
->net_conf
);
1045 nc
->on_congestion
!= OC_BLOCK
&& mdev
->tconn
->agreed_pro_version
>= 96) {
1048 if (nc
->cong_fill
&&
1049 atomic_read(&mdev
->ap_in_flight
) >= nc
->cong_fill
) {
1050 dev_info(DEV
, "Congestion-fill threshold reached\n");
1054 if (mdev
->act_log
->used
>= nc
->cong_extents
) {
1055 dev_info(DEV
, "Congestion-extents threshold reached\n");
1060 queue_barrier(mdev
); /* last barrier, after mirrored writes */
1062 if (nc
->on_congestion
== OC_PULL_AHEAD
)
1063 _drbd_set_state(_NS(mdev
, conn
, C_AHEAD
), 0, NULL
);
1064 else /*nc->on_congestion == OC_DISCONNECT */
1065 _drbd_set_state(_NS(mdev
, conn
, C_DISCONNECTING
), 0, NULL
);
1070 spin_unlock_irq(&mdev
->tconn
->req_lock
);
1071 kfree(b
); /* if someone else has beaten us to it... */
1074 req
->private_bio
->bi_bdev
= mdev
->ldev
->backing_bdev
;
1076 /* State may have changed since we grabbed our reference on the
1077 * mdev->ldev member. Double check, and short-circuit to endio.
1078 * In case the last activity log transaction failed to get on
1079 * stable storage, and this is a WRITE, we may not even submit
1081 if (get_ldev(mdev
)) {
1082 if (drbd_insert_fault(mdev
, rw
== WRITE
? DRBD_FAULT_DT_WR
1083 : rw
== READ
? DRBD_FAULT_DT_RD
1084 : DRBD_FAULT_DT_RA
))
1085 bio_endio(req
->private_bio
, -EIO
);
1087 generic_make_request(req
->private_bio
);
1090 bio_endio(req
->private_bio
, -EIO
);
1096 if (req
->rq_state
& RQ_IN_ACT_LOG
)
1097 drbd_al_complete_io(mdev
, &req
->i
);
1100 bio_put(req
->private_bio
);
1101 req
->private_bio
= NULL
;
1105 bio_endio(bio
, err
);
1114 int drbd_make_request(struct request_queue
*q
, struct bio
*bio
)
1116 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1117 unsigned long start_time
;
1119 start_time
= jiffies
;
1122 * what we "blindly" assume:
1124 D_ASSERT(bio
->bi_size
> 0);
1125 D_ASSERT(IS_ALIGNED(bio
->bi_size
, 512));
1129 } while (__drbd_make_request(mdev
, bio
, start_time
));
1134 /* This is called by bio_add_page().
1136 * q->max_hw_sectors and other global limits are already enforced there.
1138 * We need to call down to our lower level device,
1139 * in case it has special restrictions.
1141 * We also may need to enforce configured max-bio-bvecs limits.
1143 * As long as the BIO is empty we have to allow at least one bvec,
1144 * regardless of size and offset, so no need to ask lower levels.
1146 int drbd_merge_bvec(struct request_queue
*q
, struct bvec_merge_data
*bvm
, struct bio_vec
*bvec
)
1148 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1149 unsigned int bio_size
= bvm
->bi_size
;
1150 int limit
= DRBD_MAX_BIO_SIZE
;
1153 if (bio_size
&& get_ldev(mdev
)) {
1154 struct request_queue
* const b
=
1155 mdev
->ldev
->backing_bdev
->bd_disk
->queue
;
1156 if (b
->merge_bvec_fn
) {
1157 backing_limit
= b
->merge_bvec_fn(b
, bvm
, bvec
);
1158 limit
= min(limit
, backing_limit
);
1165 void request_timer_fn(unsigned long data
)
1167 struct drbd_conf
*mdev
= (struct drbd_conf
*) data
;
1168 struct drbd_tconn
*tconn
= mdev
->tconn
;
1169 struct drbd_request
*req
; /* oldest request */
1170 struct list_head
*le
;
1171 struct net_conf
*nc
;
1172 unsigned long ent
= 0, dt
= 0, et
, nt
; /* effective timeout = ko_count * timeout */
1175 nc
= rcu_dereference(tconn
->net_conf
);
1176 ent
= nc
? nc
->timeout
* HZ
/10 * nc
->ko_count
: 0;
1178 if (get_ldev(mdev
)) {
1179 dt
= rcu_dereference(mdev
->ldev
->disk_conf
)->disk_timeout
* HZ
/ 10;
1184 et
= min_not_zero(dt
, ent
);
1186 if (!et
|| (mdev
->state
.conn
< C_WF_REPORT_PARAMS
&& mdev
->state
.disk
<= D_FAILED
))
1187 return; /* Recurring timer stopped */
1189 spin_lock_irq(&tconn
->req_lock
);
1190 le
= &tconn
->oldest_tle
->requests
;
1191 if (list_empty(le
)) {
1192 spin_unlock_irq(&tconn
->req_lock
);
1193 mod_timer(&mdev
->request_timer
, jiffies
+ et
);
1198 req
= list_entry(le
, struct drbd_request
, tl_requests
);
1199 if (ent
&& req
->rq_state
& RQ_NET_PENDING
) {
1200 if (time_is_before_eq_jiffies(req
->start_time
+ ent
)) {
1201 dev_warn(DEV
, "Remote failed to finish a request within ko-count * timeout\n");
1202 _drbd_set_state(_NS(mdev
, conn
, C_TIMEOUT
), CS_VERBOSE
| CS_HARD
, NULL
);
1205 if (dt
&& req
->rq_state
& RQ_LOCAL_PENDING
&& req
->w
.mdev
== mdev
) {
1206 if (time_is_before_eq_jiffies(req
->start_time
+ dt
)) {
1207 dev_warn(DEV
, "Local backing device failed to meet the disk-timeout\n");
1208 __drbd_chk_io_error(mdev
, 1);
1211 nt
= (time_is_before_eq_jiffies(req
->start_time
+ et
) ? jiffies
: req
->start_time
) + et
;
1212 spin_unlock_irq(&tconn
->req_lock
);
1213 mod_timer(&mdev
->request_timer
, nt
);