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/autoconf.h>
27 #include <linux/module.h>
29 #include <linux/slab.h>
30 #include <linux/drbd.h>
32 #include "drbd_tracing.h"
36 /* Update disk stats at start of I/O request */
37 static void _drbd_start_io_acct(struct drbd_conf
*mdev
, struct drbd_request
*req
, struct bio
*bio
)
39 const int rw
= bio_data_dir(bio
);
41 cpu
= part_stat_lock();
42 part_stat_inc(cpu
, &mdev
->vdisk
->part0
, ios
[rw
]);
43 part_stat_add(cpu
, &mdev
->vdisk
->part0
, sectors
[rw
], bio_sectors(bio
));
45 mdev
->vdisk
->part0
.in_flight
[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
);
58 mdev
->vdisk
->part0
.in_flight
[rw
]--;
61 static void _req_is_done(struct drbd_conf
*mdev
, struct drbd_request
*req
, const int rw
)
63 const unsigned long s
= req
->rq_state
;
64 /* if it was a write, we may have to set the corresponding
65 * bit(s) out-of-sync first. If it had a local part, we need to
66 * release the reference to the activity log. */
68 /* remove it from the transfer log.
69 * well, only if it had been there in the first
70 * place... if it had not (local only or conflicting
71 * and never sent), it should still be "empty" as
72 * initialized in drbd_req_new(), so we can list_del() it
73 * here unconditionally */
74 list_del(&req
->tl_requests
);
75 /* Set out-of-sync unless both OK flags are set
76 * (local only or remote failed).
77 * Other places where we set out-of-sync:
78 * READ with local io-error */
79 if (!(s
& RQ_NET_OK
) || !(s
& RQ_LOCAL_OK
))
80 drbd_set_out_of_sync(mdev
, req
->sector
, req
->size
);
82 if ((s
& RQ_NET_OK
) && (s
& RQ_LOCAL_OK
) && (s
& RQ_NET_SIS
))
83 drbd_set_in_sync(mdev
, req
->sector
, req
->size
);
85 /* one might be tempted to move the drbd_al_complete_io
86 * to the local io completion callback drbd_endio_pri.
87 * but, if this was a mirror write, we may only
88 * drbd_al_complete_io after this is RQ_NET_DONE,
89 * otherwise the extent could be dropped from the al
90 * before it has actually been written on the peer.
91 * if we crash before our peer knows about the request,
92 * but after the extent has been dropped from the al,
93 * we would forget to resync the corresponding extent.
95 if (s
& RQ_LOCAL_MASK
) {
96 if (get_ldev_if_state(mdev
, D_FAILED
)) {
97 drbd_al_complete_io(mdev
, req
->sector
);
99 } else if (__ratelimit(&drbd_ratelimit_state
)) {
100 dev_warn(DEV
, "Should have called drbd_al_complete_io(, %llu), "
101 "but my Disk seems to have failed :(\n",
102 (unsigned long long) req
->sector
);
107 /* if it was a local io error, we want to notify our
108 * peer about that, and see if we need to
109 * detach the disk and stuff.
110 * to avoid allocating some special work
111 * struct, reuse the request. */
114 * why do we do this not when we detect the error,
115 * but delay it until it is "done", i.e. possibly
116 * until the next barrier ack? */
119 ((s
& RQ_LOCAL_MASK
) && !(s
& RQ_LOCAL_OK
))) {
120 if (!(req
->w
.list
.next
== LIST_POISON1
||
121 list_empty(&req
->w
.list
))) {
122 /* DEBUG ASSERT only; if this triggers, we
123 * probably corrupt the worker list here */
124 dev_err(DEV
, "req->w.list.next = %p\n", req
->w
.list
.next
);
125 dev_err(DEV
, "req->w.list.prev = %p\n", req
->w
.list
.prev
);
127 req
->w
.cb
= w_io_error
;
128 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
129 /* drbd_req_free() is done in w_io_error */
135 static void queue_barrier(struct drbd_conf
*mdev
)
137 struct drbd_tl_epoch
*b
;
139 /* We are within the req_lock. Once we queued the barrier for sending,
140 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
141 * barrier/epoch object is added. This is the only place this bit is
142 * set. It indicates that the barrier for this epoch is already queued,
143 * and no new epoch has been created yet. */
144 if (test_bit(CREATE_BARRIER
, &mdev
->flags
))
147 b
= mdev
->newest_tle
;
148 b
->w
.cb
= w_send_barrier
;
149 /* inc_ap_pending done here, so we won't
150 * get imbalanced on connection loss.
151 * dec_ap_pending will be done in got_BarrierAck
152 * or (on connection loss) in tl_clear. */
153 inc_ap_pending(mdev
);
154 drbd_queue_work(&mdev
->data
.work
, &b
->w
);
155 set_bit(CREATE_BARRIER
, &mdev
->flags
);
158 static void _about_to_complete_local_write(struct drbd_conf
*mdev
,
159 struct drbd_request
*req
)
161 const unsigned long s
= req
->rq_state
;
162 struct drbd_request
*i
;
163 struct drbd_epoch_entry
*e
;
164 struct hlist_node
*n
;
165 struct hlist_head
*slot
;
167 /* before we can signal completion to the upper layers,
168 * we may need to close the current epoch */
169 if (mdev
->state
.conn
>= C_CONNECTED
&&
170 req
->epoch
== mdev
->newest_tle
->br_number
)
173 /* we need to do the conflict detection stuff,
174 * if we have the ee_hash (two_primaries) and
175 * this has been on the network */
176 if ((s
& RQ_NET_DONE
) && mdev
->ee_hash
!= NULL
) {
177 const sector_t sector
= req
->sector
;
178 const int size
= req
->size
;
181 * there must be no conflicting requests, since
182 * they must have been failed on the spot */
183 #define OVERLAPS overlaps(sector, size, i->sector, i->size)
184 slot
= tl_hash_slot(mdev
, sector
);
185 hlist_for_each_entry(i
, n
, slot
, colision
) {
187 dev_alert(DEV
, "LOGIC BUG: completed: %p %llus +%u; "
188 "other: %p %llus +%u\n",
189 req
, (unsigned long long)sector
, size
,
190 i
, (unsigned long long)i
->sector
, i
->size
);
194 /* maybe "wake" those conflicting epoch entries
195 * that wait for this request to finish.
197 * currently, there can be only _one_ such ee
198 * (well, or some more, which would be pending
199 * P_DISCARD_ACK not yet sent by the asender...),
200 * since we block the receiver thread upon the
201 * first conflict detection, which will wait on
202 * misc_wait. maybe we want to assert that?
204 * anyways, if we found one,
205 * we just have to do a wake_up. */
207 #define OVERLAPS overlaps(sector, size, e->sector, e->size)
208 slot
= ee_hash_slot(mdev
, req
->sector
);
209 hlist_for_each_entry(e
, n
, slot
, colision
) {
211 wake_up(&mdev
->misc_wait
);
219 void complete_master_bio(struct drbd_conf
*mdev
,
220 struct bio_and_error
*m
)
222 trace_drbd_bio(mdev
, "Rq", m
->bio
, 1, NULL
);
223 bio_endio(m
->bio
, m
->error
);
227 /* Helper for __req_mod().
228 * Set m->bio to the master bio, if it is fit to be completed,
229 * or leave it alone (it is initialized to NULL in __req_mod),
230 * if it has already been completed, or cannot be completed yet.
231 * If m->bio is set, the error status to be returned is placed in m->error.
233 void _req_may_be_done(struct drbd_request
*req
, struct bio_and_error
*m
)
235 const unsigned long s
= req
->rq_state
;
236 struct drbd_conf
*mdev
= req
->mdev
;
237 /* only WRITES may end up here without a master bio (on barrier ack) */
238 int rw
= req
->master_bio
? bio_data_dir(req
->master_bio
) : WRITE
;
240 trace_drbd_req(req
, nothing
, "_req_may_be_done");
242 /* we must not complete the master bio, while it is
243 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
244 * not yet acknowledged by the peer
245 * not yet completed by the local io subsystem
246 * these flags may get cleared in any order by
249 * the bio_endio completion callbacks.
251 if (s
& RQ_NET_QUEUED
)
253 if (s
& RQ_NET_PENDING
)
255 if (s
& RQ_LOCAL_PENDING
)
258 if (req
->master_bio
) {
259 /* this is data_received (remote read)
260 * or protocol C P_WRITE_ACK
261 * or protocol B P_RECV_ACK
262 * or protocol A "handed_over_to_network" (SendAck)
263 * or canceled or failed,
264 * or killed from the transfer log due to connection loss.
268 * figure out whether to report success or failure.
270 * report success when at least one of the operations succeeded.
271 * or, to put the other way,
272 * only report failure, when both operations failed.
274 * what to do about the failures is handled elsewhere.
275 * what we need to do here is just: complete the master_bio.
277 * local completion error, if any, has been stored as ERR_PTR
278 * in private_bio within drbd_endio_pri.
280 int ok
= (s
& RQ_LOCAL_OK
) || (s
& RQ_NET_OK
);
281 int error
= PTR_ERR(req
->private_bio
);
283 /* remove the request from the conflict detection
284 * respective block_id verification hash */
285 if (!hlist_unhashed(&req
->colision
))
286 hlist_del(&req
->colision
);
288 D_ASSERT((s
& RQ_NET_MASK
) == 0);
290 /* for writes we need to do some extra housekeeping */
292 _about_to_complete_local_write(mdev
, req
);
294 /* Update disk stats */
295 _drbd_end_io_acct(mdev
, req
);
297 m
->error
= ok
? 0 : (error
?: -EIO
);
298 m
->bio
= req
->master_bio
;
299 req
->master_bio
= NULL
;
302 if ((s
& RQ_NET_MASK
) == 0 || (s
& RQ_NET_DONE
)) {
303 /* this is disconnected (local only) operation,
304 * or protocol C P_WRITE_ACK,
305 * or protocol A or B P_BARRIER_ACK,
306 * or killed from the transfer log due to connection loss. */
307 _req_is_done(mdev
, req
, rw
);
309 /* else: network part and not DONE yet. that is
310 * protocol A or B, barrier ack still pending... */
314 * checks whether there was an overlapping request
315 * or ee already registered.
317 * if so, return 1, in which case this request is completed on the spot,
318 * without ever being submitted or send.
320 * return 0 if it is ok to submit this request.
323 * paranoia: assume something above us is broken, and issues different write
324 * requests for the same block simultaneously...
326 * To ensure these won't be reordered differently on both nodes, resulting in
327 * diverging data sets, we discard the later one(s). Not that this is supposed
328 * to happen, but this is the rationale why we also have to check for
329 * conflicting requests with local origin, and why we have to do so regardless
330 * of whether we allowed multiple primaries.
332 * BTW, in case we only have one primary, the ee_hash is empty anyways, and the
333 * second hlist_for_each_entry becomes a noop. This is even simpler than to
334 * grab a reference on the net_conf, and check for the two_primaries flag...
336 static int _req_conflicts(struct drbd_request
*req
)
338 struct drbd_conf
*mdev
= req
->mdev
;
339 const sector_t sector
= req
->sector
;
340 const int size
= req
->size
;
341 struct drbd_request
*i
;
342 struct drbd_epoch_entry
*e
;
343 struct hlist_node
*n
;
344 struct hlist_head
*slot
;
346 D_ASSERT(hlist_unhashed(&req
->colision
));
348 if (!get_net_conf(mdev
))
352 ERR_IF (mdev
->tl_hash_s
== 0)
353 goto out_no_conflict
;
354 BUG_ON(mdev
->tl_hash
== NULL
);
356 #define OVERLAPS overlaps(i->sector, i->size, sector, size)
357 slot
= tl_hash_slot(mdev
, sector
);
358 hlist_for_each_entry(i
, n
, slot
, colision
) {
360 dev_alert(DEV
, "%s[%u] Concurrent local write detected! "
361 "[DISCARD L] new: %llus +%u; "
362 "pending: %llus +%u\n",
363 current
->comm
, current
->pid
,
364 (unsigned long long)sector
, size
,
365 (unsigned long long)i
->sector
, i
->size
);
370 if (mdev
->ee_hash_s
) {
371 /* now, check for overlapping requests with remote origin */
372 BUG_ON(mdev
->ee_hash
== NULL
);
374 #define OVERLAPS overlaps(e->sector, e->size, sector, size)
375 slot
= ee_hash_slot(mdev
, sector
);
376 hlist_for_each_entry(e
, n
, slot
, colision
) {
378 dev_alert(DEV
, "%s[%u] Concurrent remote write detected!"
379 " [DISCARD L] new: %llus +%u; "
380 "pending: %llus +%u\n",
381 current
->comm
, current
->pid
,
382 (unsigned long long)sector
, size
,
383 (unsigned long long)e
->sector
, e
->size
);
391 /* this is like it should be, and what we expected.
392 * our users do behave after all... */
401 /* obviously this could be coded as many single functions
402 * instead of one huge switch,
403 * or by putting the code directly in the respective locations
404 * (as it has been before).
406 * but having it this way
407 * enforces that it is all in this one place, where it is easier to audit,
408 * it makes it obvious that whatever "event" "happens" to a request should
409 * happen "atomically" within the req_lock,
410 * and it enforces that we have to think in a very structured manner
411 * about the "events" that may happen to a request during its life time ...
413 void __req_mod(struct drbd_request
*req
, enum drbd_req_event what
,
414 struct bio_and_error
*m
)
416 struct drbd_conf
*mdev
= req
->mdev
;
419 trace_drbd_req(req
, what
, NULL
);
423 dev_err(DEV
, "LOGIC BUG in %s:%u\n", __FILE__
, __LINE__
);
426 /* does not happen...
427 * initialization done in drbd_req_new
432 case to_be_send
: /* via network */
433 /* reached via drbd_make_request_common
434 * and from w_read_retry_remote */
435 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
436 req
->rq_state
|= RQ_NET_PENDING
;
437 inc_ap_pending(mdev
);
440 case to_be_submitted
: /* locally */
441 /* reached via drbd_make_request_common */
442 D_ASSERT(!(req
->rq_state
& RQ_LOCAL_MASK
));
443 req
->rq_state
|= RQ_LOCAL_PENDING
;
447 if (bio_data_dir(req
->master_bio
) == WRITE
)
448 mdev
->writ_cnt
+= req
->size
>>9;
450 mdev
->read_cnt
+= req
->size
>>9;
452 req
->rq_state
|= (RQ_LOCAL_COMPLETED
|RQ_LOCAL_OK
);
453 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
455 _req_may_be_done(req
, m
);
459 case write_completed_with_error
:
460 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
461 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
463 dev_alert(DEV
, "Local WRITE failed sec=%llus size=%u\n",
464 (unsigned long long)req
->sector
, req
->size
);
465 /* and now: check how to handle local io error. */
466 __drbd_chk_io_error(mdev
, FALSE
);
467 _req_may_be_done(req
, m
);
471 case read_ahead_completed_with_error
:
472 /* it is legal to fail READA */
473 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
474 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
475 _req_may_be_done(req
, m
);
479 case read_completed_with_error
:
480 drbd_set_out_of_sync(mdev
, req
->sector
, req
->size
);
482 req
->rq_state
|= RQ_LOCAL_COMPLETED
;
483 req
->rq_state
&= ~RQ_LOCAL_PENDING
;
485 dev_alert(DEV
, "Local READ failed sec=%llus size=%u\n",
486 (unsigned long long)req
->sector
, req
->size
);
487 /* _req_mod(req,to_be_send); oops, recursion... */
488 D_ASSERT(!(req
->rq_state
& RQ_NET_MASK
));
489 req
->rq_state
|= RQ_NET_PENDING
;
490 inc_ap_pending(mdev
);
492 __drbd_chk_io_error(mdev
, FALSE
);
494 /* NOTE: if we have no connection,
495 * or know the peer has no good data either,
496 * then we don't actually need to "queue_for_net_read",
497 * but we do so anyways, since the drbd_io_error()
498 * and the potential state change to "Diskless"
499 * needs to be done from process context */
501 /* fall through: _req_mod(req,queue_for_net_read); */
503 case queue_for_net_read
:
504 /* READ or READA, and
506 * or target area marked as invalid,
507 * or just got an io-error. */
508 /* from drbd_make_request_common
509 * or from bio_endio during read io-error recovery */
511 /* so we can verify the handle in the answer packet
512 * corresponding hlist_del is in _req_may_be_done() */
513 hlist_add_head(&req
->colision
, ar_hash_slot(mdev
, req
->sector
));
515 set_bit(UNPLUG_REMOTE
, &mdev
->flags
); /* why? */
517 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
518 req
->rq_state
|= RQ_NET_QUEUED
;
519 req
->w
.cb
= (req
->rq_state
& RQ_LOCAL_MASK
)
520 ? w_read_retry_remote
522 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
525 case queue_for_net_write
:
526 /* assert something? */
527 /* from drbd_make_request_common only */
529 hlist_add_head(&req
->colision
, tl_hash_slot(mdev
, req
->sector
));
530 /* corresponding hlist_del is in _req_may_be_done() */
533 * In case the req ended up on the transfer log before being
534 * queued on the worker, it could lead to this request being
535 * missed during cleanup after connection loss.
536 * So we have to do both operations here,
537 * within the same lock that protects the transfer log.
539 * _req_add_to_epoch(req); this has to be after the
540 * _maybe_start_new_epoch(req); which happened in
541 * drbd_make_request_common, because we now may set the bit
542 * again ourselves to close the current epoch.
544 * Add req to the (now) current epoch (barrier). */
546 /* see drbd_make_request_common,
547 * just after it grabs the req_lock */
548 D_ASSERT(test_bit(CREATE_BARRIER
, &mdev
->flags
) == 0);
550 req
->epoch
= mdev
->newest_tle
->br_number
;
551 list_add_tail(&req
->tl_requests
,
552 &mdev
->newest_tle
->requests
);
554 /* increment size of current epoch */
555 mdev
->newest_tle
->n_req
++;
557 /* queue work item to send data */
558 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
559 req
->rq_state
|= RQ_NET_QUEUED
;
560 req
->w
.cb
= w_send_dblock
;
561 drbd_queue_work(&mdev
->data
.work
, &req
->w
);
563 /* close the epoch, in case it outgrew the limit */
564 if (mdev
->newest_tle
->n_req
>= mdev
->net_conf
->max_epoch_size
)
570 /* treat it the same */
572 /* real cleanup will be done from tl_clear. just update flags
573 * so it is no longer marked as on the worker queue */
574 req
->rq_state
&= ~RQ_NET_QUEUED
;
575 /* if we did it right, tl_clear should be scheduled only after
576 * this, so this should not be necessary! */
577 _req_may_be_done(req
, m
);
580 case handed_over_to_network
:
581 /* assert something? */
582 if (bio_data_dir(req
->master_bio
) == WRITE
&&
583 mdev
->net_conf
->wire_protocol
== DRBD_PROT_A
) {
584 /* this is what is dangerous about protocol A:
585 * pretend it was successfully written on the peer. */
586 if (req
->rq_state
& RQ_NET_PENDING
) {
587 dec_ap_pending(mdev
);
588 req
->rq_state
&= ~RQ_NET_PENDING
;
589 req
->rq_state
|= RQ_NET_OK
;
590 } /* else: neg-ack was faster... */
591 /* it is still not yet RQ_NET_DONE until the
592 * corresponding epoch barrier got acked as well,
593 * so we know what to dirty on connection loss */
595 req
->rq_state
&= ~RQ_NET_QUEUED
;
596 req
->rq_state
|= RQ_NET_SENT
;
597 /* because _drbd_send_zc_bio could sleep, and may want to
598 * dereference the bio even after the "write_acked_by_peer" and
599 * "completed_ok" events came in, once we return from
600 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
601 * whether it is done already, and end it. */
602 _req_may_be_done(req
, m
);
605 case connection_lost_while_pending
:
606 /* transfer log cleanup after connection loss */
607 /* assert something? */
608 if (req
->rq_state
& RQ_NET_PENDING
)
609 dec_ap_pending(mdev
);
610 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
611 req
->rq_state
|= RQ_NET_DONE
;
612 /* if it is still queued, we may not complete it here.
613 * it will be canceled soon. */
614 if (!(req
->rq_state
& RQ_NET_QUEUED
))
615 _req_may_be_done(req
, m
);
618 case write_acked_by_peer_and_sis
:
619 req
->rq_state
|= RQ_NET_SIS
;
620 case conflict_discarded_by_peer
:
621 /* for discarded conflicting writes of multiple primaries,
622 * there is no need to keep anything in the tl, potential
623 * node crashes are covered by the activity log. */
624 if (what
== conflict_discarded_by_peer
)
625 dev_alert(DEV
, "Got DiscardAck packet %llus +%u!"
626 " DRBD is not a random data generator!\n",
627 (unsigned long long)req
->sector
, req
->size
);
628 req
->rq_state
|= RQ_NET_DONE
;
630 case write_acked_by_peer
:
631 /* protocol C; successfully written on peer.
632 * Nothing to do here.
633 * We want to keep the tl in place for all protocols, to cater
634 * for volatile write-back caches on lower level devices.
636 * A barrier request is expected to have forced all prior
637 * requests onto stable storage, so completion of a barrier
638 * request could set NET_DONE right here, and not wait for the
639 * P_BARRIER_ACK, but that is an unnecessary optimization. */
641 /* this makes it effectively the same as for: */
642 case recv_acked_by_peer
:
643 /* protocol B; pretends to be successfully written on peer.
644 * see also notes above in handed_over_to_network about
646 req
->rq_state
|= RQ_NET_OK
;
647 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
648 dec_ap_pending(mdev
);
649 req
->rq_state
&= ~RQ_NET_PENDING
;
650 _req_may_be_done(req
, m
);
654 /* assert something? */
655 if (req
->rq_state
& RQ_NET_PENDING
)
656 dec_ap_pending(mdev
);
657 req
->rq_state
&= ~(RQ_NET_OK
|RQ_NET_PENDING
);
659 req
->rq_state
|= RQ_NET_DONE
;
660 _req_may_be_done(req
, m
);
661 /* else: done by handed_over_to_network */
665 if (req
->rq_state
& RQ_NET_PENDING
) {
666 /* barrier came in before all requests have been acked.
667 * this is bad, because if the connection is lost now,
668 * we won't be able to clean them up... */
669 dev_err(DEV
, "FIXME (barrier_acked but pending)\n");
670 trace_drbd_req(req
, nothing
, "FIXME (barrier_acked but pending)");
671 list_move(&req
->tl_requests
, &mdev
->out_of_sequence_requests
);
673 D_ASSERT(req
->rq_state
& RQ_NET_SENT
);
674 req
->rq_state
|= RQ_NET_DONE
;
675 _req_may_be_done(req
, m
);
679 D_ASSERT(req
->rq_state
& RQ_NET_PENDING
);
680 dec_ap_pending(mdev
);
681 req
->rq_state
&= ~RQ_NET_PENDING
;
682 req
->rq_state
|= (RQ_NET_OK
|RQ_NET_DONE
);
683 _req_may_be_done(req
, m
);
688 /* we may do a local read if:
689 * - we are consistent (of course),
690 * - or we are generally inconsistent,
691 * BUT we are still/already IN SYNC for this area.
692 * since size may be bigger than BM_BLOCK_SIZE,
693 * we may need to check several bits.
695 static int drbd_may_do_local_read(struct drbd_conf
*mdev
, sector_t sector
, int size
)
697 unsigned long sbnr
, ebnr
;
698 sector_t esector
, nr_sectors
;
700 if (mdev
->state
.disk
== D_UP_TO_DATE
)
702 if (mdev
->state
.disk
>= D_OUTDATED
)
704 if (mdev
->state
.disk
< D_INCONSISTENT
)
706 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
707 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
708 esector
= sector
+ (size
>> 9) - 1;
710 D_ASSERT(sector
< nr_sectors
);
711 D_ASSERT(esector
< nr_sectors
);
713 sbnr
= BM_SECT_TO_BIT(sector
);
714 ebnr
= BM_SECT_TO_BIT(esector
);
716 return 0 == drbd_bm_count_bits(mdev
, sbnr
, ebnr
);
719 static int drbd_make_request_common(struct drbd_conf
*mdev
, struct bio
*bio
)
721 const int rw
= bio_rw(bio
);
722 const int size
= bio
->bi_size
;
723 const sector_t sector
= bio
->bi_sector
;
724 struct drbd_tl_epoch
*b
= NULL
;
725 struct drbd_request
*req
;
729 /* allocate outside of all locks; */
730 req
= drbd_req_new(mdev
, bio
);
733 /* only pass the error to the upper layers.
734 * if user cannot handle io errors, that's not our business. */
735 dev_err(DEV
, "could not kmalloc() req\n");
736 bio_endio(bio
, -ENOMEM
);
740 trace_drbd_bio(mdev
, "Rq", bio
, 0, req
);
742 local
= get_ldev(mdev
);
744 bio_put(req
->private_bio
); /* or we get a bio leak */
745 req
->private_bio
= NULL
;
752 if (!drbd_may_do_local_read(mdev
, sector
, size
)) {
753 /* we could kick the syncer to
754 * sync this extent asap, wait for
755 * it, then continue locally.
756 * Or just issue the request remotely.
759 bio_put(req
->private_bio
);
760 req
->private_bio
= NULL
;
764 remote
= !local
&& mdev
->state
.pdsk
>= D_UP_TO_DATE
;
767 /* If we have a disk, but a READA request is mapped to remote,
768 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
769 * Just fail that READA request right here.
771 * THINK: maybe fail all READA when not local?
772 * or make this configurable...
773 * if network is slow, READA won't do any good.
775 if (rw
== READA
&& mdev
->state
.disk
>= D_INCONSISTENT
&& !local
) {
777 goto fail_and_free_req
;
780 /* For WRITES going to the local disk, grab a reference on the target
781 * extent. This waits for any resync activity in the corresponding
782 * resync extent to finish, and, if necessary, pulls in the target
783 * extent into the activity log, which involves further disk io because
784 * of transactional on-disk meta data updates. */
785 if (rw
== WRITE
&& local
)
786 drbd_al_begin_io(mdev
, sector
);
788 remote
= remote
&& (mdev
->state
.pdsk
== D_UP_TO_DATE
||
789 (mdev
->state
.pdsk
== D_INCONSISTENT
&&
790 mdev
->state
.conn
>= C_CONNECTED
));
792 if (!(local
|| remote
)) {
793 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
794 goto fail_free_complete
;
797 /* For WRITE request, we have to make sure that we have an
798 * unused_spare_tle, in case we need to start a new epoch.
799 * I try to be smart and avoid to pre-allocate always "just in case",
800 * but there is a race between testing the bit and pointer outside the
801 * spinlock, and grabbing the spinlock.
802 * if we lost that race, we retry. */
803 if (rw
== WRITE
&& remote
&&
804 mdev
->unused_spare_tle
== NULL
&&
805 test_bit(CREATE_BARRIER
, &mdev
->flags
)) {
807 b
= kmalloc(sizeof(struct drbd_tl_epoch
), GFP_NOIO
);
809 dev_err(DEV
, "Failed to alloc barrier.\n");
811 goto fail_free_complete
;
815 /* GOOD, everything prepared, grab the spin_lock */
816 spin_lock_irq(&mdev
->req_lock
);
819 remote
= (mdev
->state
.pdsk
== D_UP_TO_DATE
||
820 (mdev
->state
.pdsk
== D_INCONSISTENT
&&
821 mdev
->state
.conn
>= C_CONNECTED
));
823 dev_warn(DEV
, "lost connection while grabbing the req_lock!\n");
824 if (!(local
|| remote
)) {
825 dev_err(DEV
, "IO ERROR: neither local nor remote disk\n");
826 spin_unlock_irq(&mdev
->req_lock
);
827 goto fail_free_complete
;
831 if (b
&& mdev
->unused_spare_tle
== NULL
) {
832 mdev
->unused_spare_tle
= b
;
835 if (rw
== WRITE
&& remote
&&
836 mdev
->unused_spare_tle
== NULL
&&
837 test_bit(CREATE_BARRIER
, &mdev
->flags
)) {
838 /* someone closed the current epoch
839 * while we were grabbing the spinlock */
840 spin_unlock_irq(&mdev
->req_lock
);
841 goto allocate_barrier
;
845 /* Update disk stats */
846 _drbd_start_io_acct(mdev
, req
, bio
);
848 /* _maybe_start_new_epoch(mdev);
849 * If we need to generate a write barrier packet, we have to add the
850 * new epoch (barrier) object, and queue the barrier packet for sending,
851 * and queue the req's data after it _within the same lock_, otherwise
852 * we have race conditions were the reorder domains could be mixed up.
854 * Even read requests may start a new epoch and queue the corresponding
855 * barrier packet. To get the write ordering right, we only have to
856 * make sure that, if this is a write request and it triggered a
857 * barrier packet, this request is queued within the same spinlock. */
858 if (remote
&& mdev
->unused_spare_tle
&&
859 test_and_clear_bit(CREATE_BARRIER
, &mdev
->flags
)) {
860 _tl_add_barrier(mdev
, mdev
->unused_spare_tle
);
861 mdev
->unused_spare_tle
= NULL
;
863 D_ASSERT(!(remote
&& rw
== WRITE
&&
864 test_bit(CREATE_BARRIER
, &mdev
->flags
)));
868 * Actually, 'local' may be wrong here already, since we may have failed
869 * to write to the meta data, and may become wrong anytime because of
870 * local io-error for some other request, which would lead to us
871 * "detaching" the local disk.
873 * 'remote' may become wrong any time because the network could fail.
875 * This is a harmless race condition, though, since it is handled
876 * correctly at the appropriate places; so it just defers the failure
877 * of the respective operation.
880 /* mark them early for readability.
881 * this just sets some state flags. */
883 _req_mod(req
, to_be_send
);
885 _req_mod(req
, to_be_submitted
);
887 /* check this request on the collision detection hash tables.
888 * if we have a conflict, just complete it here.
889 * THINK do we want to check reads, too? (I don't think so...) */
890 if (rw
== WRITE
&& _req_conflicts(req
)) {
891 /* this is a conflicting request.
892 * even though it may have been only _partially_
893 * overlapping with one of the currently pending requests,
894 * without even submitting or sending it, we will
895 * pretend that it was successfully served right now.
898 bio_put(req
->private_bio
);
899 req
->private_bio
= NULL
;
900 drbd_al_complete_io(mdev
, req
->sector
);
905 dec_ap_pending(mdev
);
906 _drbd_end_io_acct(mdev
, req
);
907 /* THINK: do we want to fail it (-EIO), or pretend success? */
908 bio_endio(req
->master_bio
, 0);
909 req
->master_bio
= NULL
;
915 /* NOTE remote first: to get the concurrent write detection right,
916 * we must register the request before start of local IO. */
918 /* either WRITE and C_CONNECTED,
919 * or READ, and no local disk,
920 * or READ, but not in sync.
922 _req_mod(req
, (rw
== WRITE
)
923 ? queue_for_net_write
924 : queue_for_net_read
);
926 spin_unlock_irq(&mdev
->req_lock
);
927 kfree(b
); /* if someone else has beaten us to it... */
930 req
->private_bio
->bi_bdev
= mdev
->ldev
->backing_bdev
;
932 trace_drbd_bio(mdev
, "Pri", req
->private_bio
, 0, NULL
);
934 if (FAULT_ACTIVE(mdev
, rw
== WRITE
? DRBD_FAULT_DT_WR
935 : rw
== READ
? DRBD_FAULT_DT_RD
937 bio_endio(req
->private_bio
, -EIO
);
939 generic_make_request(req
->private_bio
);
942 /* we need to plug ALWAYS since we possibly need to kick lo_dev.
943 * we plug after submit, so we won't miss an unplug event */
944 drbd_plug_device(mdev
);
949 if (rw
== WRITE
&& local
)
950 drbd_al_complete_io(mdev
, sector
);
953 bio_put(req
->private_bio
);
954 req
->private_bio
= NULL
;
965 /* helper function for drbd_make_request
966 * if we can determine just by the mdev (state) that this request will fail,
970 static int drbd_fail_request_early(struct drbd_conf
*mdev
, int is_write
)
973 if (mdev
->state
.conn
== C_DISCONNECTING
&&
974 mdev
->state
.disk
== D_DISKLESS
)
977 if (mdev
->state
.role
!= R_PRIMARY
&&
978 (!allow_oos
|| is_write
)) {
979 if (__ratelimit(&drbd_ratelimit_state
)) {
980 dev_err(DEV
, "Process %s[%u] tried to %s; "
981 "since we are not in Primary state, "
982 "we cannot allow this\n",
983 current
->comm
, current
->pid
,
984 is_write
? "WRITE" : "READ");
990 * Paranoia: we might have been primary, but sync target, or
991 * even diskless, then lost the connection.
992 * This should have been handled (panic? suspend?) somewhere
993 * else. But maybe it was not, so check again here.
994 * Caution: as long as we do not have a read/write lock on mdev,
995 * to serialize state changes, this is racy, since we may lose
996 * the connection *after* we test for the cstate.
998 if (mdev
->state
.disk
< D_UP_TO_DATE
&& mdev
->state
.pdsk
< D_UP_TO_DATE
) {
999 if (__ratelimit(&drbd_ratelimit_state
))
1000 dev_err(DEV
, "Sorry, I have no access to good data anymore.\n");
1007 int drbd_make_request_26(struct request_queue
*q
, struct bio
*bio
)
1009 unsigned int s_enr
, e_enr
;
1010 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1012 if (drbd_fail_request_early(mdev
, bio_data_dir(bio
) & WRITE
)) {
1013 bio_endio(bio
, -EPERM
);
1017 /* Reject barrier requests if we know the underlying device does
1019 * XXX: Need to get this info from peer as well some how so we
1020 * XXX: reject if EITHER side/data/metadata area does not support them.
1022 * because of those XXX, this is not yet enabled,
1023 * i.e. in drbd_init_set_defaults we set the NO_BARRIER_SUPP bit.
1025 if (unlikely(bio_rw_flagged(bio
, BIO_RW_BARRIER
) && test_bit(NO_BARRIER_SUPP
, &mdev
->flags
))) {
1026 /* dev_warn(DEV, "Rejecting barrier request as underlying device does not support\n"); */
1027 bio_endio(bio
, -EOPNOTSUPP
);
1032 * what we "blindly" assume:
1034 D_ASSERT(bio
->bi_size
> 0);
1035 D_ASSERT((bio
->bi_size
& 0x1ff) == 0);
1036 D_ASSERT(bio
->bi_idx
== 0);
1038 /* to make some things easier, force alignment of requests within the
1039 * granularity of our hash tables */
1040 s_enr
= bio
->bi_sector
>> HT_SHIFT
;
1041 e_enr
= (bio
->bi_sector
+(bio
->bi_size
>>9)-1) >> HT_SHIFT
;
1043 if (likely(s_enr
== e_enr
)) {
1044 inc_ap_bio(mdev
, 1);
1045 return drbd_make_request_common(mdev
, bio
);
1048 /* can this bio be split generically?
1049 * Maybe add our own split-arbitrary-bios function. */
1050 if (bio
->bi_vcnt
!= 1 || bio
->bi_idx
!= 0 || bio
->bi_size
> DRBD_MAX_SEGMENT_SIZE
) {
1051 /* rather error out here than BUG in bio_split */
1052 dev_err(DEV
, "bio would need to, but cannot, be split: "
1053 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1054 bio
->bi_vcnt
, bio
->bi_idx
, bio
->bi_size
,
1055 (unsigned long long)bio
->bi_sector
);
1056 bio_endio(bio
, -EINVAL
);
1058 /* This bio crosses some boundary, so we have to split it. */
1059 struct bio_pair
*bp
;
1060 /* works for the "do not cross hash slot boundaries" case
1061 * e.g. sector 262269, size 4096
1062 * s_enr = 262269 >> 6 = 4097
1063 * e_enr = (262269+8-1) >> 6 = 4098
1065 * sps = 64, mask = 63
1066 * first_sectors = 64 - (262269 & 63) = 3
1068 const sector_t sect
= bio
->bi_sector
;
1069 const int sps
= 1 << HT_SHIFT
; /* sectors per slot */
1070 const int mask
= sps
- 1;
1071 const sector_t first_sectors
= sps
- (sect
& mask
);
1073 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,28)
1078 /* we need to get a "reference count" (ap_bio_cnt)
1079 * to avoid races with the disconnect/reconnect/suspend code.
1080 * In case we need to split the bio here, we need to get two references
1081 * atomically, otherwise we might deadlock when trying to submit the
1083 inc_ap_bio(mdev
, 2);
1085 D_ASSERT(e_enr
== s_enr
+ 1);
1087 drbd_make_request_common(mdev
, &bp
->bio1
);
1088 drbd_make_request_common(mdev
, &bp
->bio2
);
1089 bio_pair_release(bp
);
1094 /* This is called by bio_add_page(). With this function we reduce
1095 * the number of BIOs that span over multiple DRBD_MAX_SEGMENT_SIZEs
1096 * units (was AL_EXTENTs).
1098 * we do the calculation within the lower 32bit of the byte offsets,
1099 * since we don't care for actual offset, but only check whether it
1100 * would cross "activity log extent" boundaries.
1102 * As long as the BIO is empty we have to allow at least one bvec,
1103 * regardless of size and offset. so the resulting bio may still
1104 * cross extent boundaries. those are dealt with (bio_split) in
1105 * drbd_make_request_26.
1107 int drbd_merge_bvec(struct request_queue
*q
, struct bvec_merge_data
*bvm
, struct bio_vec
*bvec
)
1109 struct drbd_conf
*mdev
= (struct drbd_conf
*) q
->queuedata
;
1110 unsigned int bio_offset
=
1111 (unsigned int)bvm
->bi_sector
<< 9; /* 32 bit */
1112 unsigned int bio_size
= bvm
->bi_size
;
1113 int limit
, backing_limit
;
1115 limit
= DRBD_MAX_SEGMENT_SIZE
1116 - ((bio_offset
& (DRBD_MAX_SEGMENT_SIZE
-1)) + bio_size
);
1119 if (bio_size
== 0) {
1120 if (limit
<= bvec
->bv_len
)
1121 limit
= bvec
->bv_len
;
1122 } else if (limit
&& get_ldev(mdev
)) {
1123 struct request_queue
* const b
=
1124 mdev
->ldev
->backing_bdev
->bd_disk
->queue
;
1125 if (b
->merge_bvec_fn
&& mdev
->ldev
->dc
.use_bmbv
) {
1126 backing_limit
= b
->merge_bvec_fn(b
, bvm
, bvec
);
1127 limit
= min(limit
, backing_limit
);