]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - drivers/block/drbd/drbd_req.c
drbd: fix potential deadlock during "restart" of conflicting writes
[mirror_ubuntu-bionic-kernel.git] / drivers / block / drbd / drbd_req.c
CommitLineData
b411b363
PR
1/*
2 drbd_req.c
3
4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
5
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>.
9
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)
13 any later version.
14
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.
19
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.
23
24 */
25
b411b363
PR
26#include <linux/module.h>
27
28#include <linux/slab.h>
29#include <linux/drbd.h>
30#include "drbd_int.h"
b411b363
PR
31#include "drbd_req.h"
32
33
57bcb6cf
PR
34static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size);
35
b411b363
PR
36/* Update disk stats at start of I/O request */
37static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
38{
39 const int rw = bio_data_dir(bio);
40 int cpu;
41 cpu = part_stat_lock();
72585d24 42 part_round_stats(cpu, &mdev->vdisk->part0);
b411b363
PR
43 part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
44 part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
376694a0
PR
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... */
753c8913 47 part_inc_in_flight(&mdev->vdisk->part0, rw);
b411b363 48 part_stat_unlock();
b411b363
PR
49}
50
51/* Update disk stats when completing request upwards */
52static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
53{
54 int rw = bio_data_dir(req->master_bio);
55 unsigned long duration = jiffies - req->start_time;
56 int cpu;
57 cpu = part_stat_lock();
58 part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
59 part_round_stats(cpu, &mdev->vdisk->part0);
753c8913 60 part_dec_in_flight(&mdev->vdisk->part0, rw);
b411b363 61 part_stat_unlock();
b411b363
PR
62}
63
9e204cdd
AG
64static struct drbd_request *drbd_req_new(struct drbd_conf *mdev,
65 struct bio *bio_src)
66{
67 struct drbd_request *req;
68
69 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
70 if (!req)
71 return NULL;
72
73 drbd_req_make_private_bio(req, bio_src);
74 req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
a21e9298 75 req->w.mdev = mdev;
9e204cdd
AG
76 req->master_bio = bio_src;
77 req->epoch = 0;
53840641 78
9e204cdd
AG
79 drbd_clear_interval(&req->i);
80 req->i.sector = bio_src->bi_sector;
81 req->i.size = bio_src->bi_size;
5e472264 82 req->i.local = true;
53840641
AG
83 req->i.waiting = false;
84
9e204cdd
AG
85 INIT_LIST_HEAD(&req->tl_requests);
86 INIT_LIST_HEAD(&req->w.list);
87
88 return req;
89}
90
91static void drbd_req_free(struct drbd_request *req)
92{
93 mempool_free(req, drbd_request_mempool);
94}
95
96/* rw is bio_data_dir(), only READ or WRITE */
b411b363
PR
97static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
98{
99 const unsigned long s = req->rq_state;
288f422e
PR
100
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 */
2312f0b3 107 list_del_init(&req->tl_requests);
288f422e 108
b411b363
PR
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. */
112 if (rw == WRITE) {
b411b363
PR
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))
ace652ac 118 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
b411b363
PR
119
120 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
ace652ac 121 drbd_set_in_sync(mdev, req->i.sector, req->i.size);
b411b363
PR
122
123 /* one might be tempted to move the drbd_al_complete_io
fcefa62e 124 * to the local io completion callback drbd_request_endio.
b411b363
PR
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.
132 */
133 if (s & RQ_LOCAL_MASK) {
134 if (get_ldev_if_state(mdev, D_FAILED)) {
0778286a 135 if (s & RQ_IN_ACT_LOG)
181286ad 136 drbd_al_complete_io(mdev, &req->i);
b411b363
PR
137 put_ldev(mdev);
138 } else if (__ratelimit(&drbd_ratelimit_state)) {
181286ad
LE
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);
b411b363
PR
142 }
143 }
144 }
145
2312f0b3
LE
146 if (s & RQ_POSTPONED)
147 drbd_restart_write(req);
148 else
149 drbd_req_free(req);
b411b363
PR
150}
151
152static void queue_barrier(struct drbd_conf *mdev)
153{
154 struct drbd_tl_epoch *b;
6936fcb4 155 struct drbd_tconn *tconn = mdev->tconn;
b411b363
PR
156
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. */
6936fcb4 162 if (test_bit(CREATE_BARRIER, &tconn->flags))
b411b363
PR
163 return;
164
6936fcb4 165 b = tconn->newest_tle;
b411b363 166 b->w.cb = w_send_barrier;
a21e9298 167 b->w.mdev = mdev;
b411b363
PR
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);
6936fcb4
PR
173 drbd_queue_work(&tconn->data.work, &b->w);
174 set_bit(CREATE_BARRIER, &tconn->flags);
b411b363
PR
175}
176
177static void _about_to_complete_local_write(struct drbd_conf *mdev,
178 struct drbd_request *req)
179{
180 const unsigned long s = req->rq_state;
b411b363 181
8a3c1044
LE
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.
187 */
188 if (mdev->state.conn >= C_CONNECTED &&
189 (s & RQ_NET_SENT) != 0 &&
87eeee41 190 req->epoch == mdev->tconn->newest_tle->br_number)
b411b363 191 queue_barrier(mdev);
b411b363
PR
192}
193
194void complete_master_bio(struct drbd_conf *mdev,
195 struct bio_and_error *m)
196{
b411b363
PR
197 bio_endio(m->bio, m->error);
198 dec_ap_bio(mdev);
199}
200
53840641
AG
201
202static void drbd_remove_request_interval(struct rb_root *root,
203 struct drbd_request *req)
204{
a21e9298 205 struct drbd_conf *mdev = req->w.mdev;
53840641
AG
206 struct drbd_interval *i = &req->i;
207
208 drbd_remove_interval(root, i);
209
210 /* Wake up any processes waiting for this request to complete. */
211 if (i->waiting)
212 wake_up(&mdev->misc_wait);
213}
214
b411b363
PR
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.
220 */
221void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
222{
223 const unsigned long s = req->rq_state;
a21e9298 224 struct drbd_conf *mdev = req->w.mdev;
cdfda633 225 int rw = req->rq_state & RQ_WRITE ? WRITE : READ;
b411b363 226
b411b363
PR
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
232 * the worker,
233 * the receiver,
234 * the bio_endio completion callbacks.
235 */
cdfda633 236 if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED))
7be8da07
AG
237 return;
238 if (req->i.waiting) {
239 /* Retry all conflicting peer requests. */
240 wake_up(&mdev->misc_wait);
241 }
b411b363
PR
242 if (s & RQ_NET_QUEUED)
243 return;
244 if (s & RQ_NET_PENDING)
245 return;
b411b363
PR
246
247 if (req->master_bio) {
8554df1c 248 /* this is DATA_RECEIVED (remote read)
b411b363
PR
249 * or protocol C P_WRITE_ACK
250 * or protocol B P_RECV_ACK
8554df1c 251 * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck)
b411b363
PR
252 * or canceled or failed,
253 * or killed from the transfer log due to connection loss.
254 */
255
256 /*
257 * figure out whether to report success or failure.
258 *
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.
262 *
263 * what to do about the failures is handled elsewhere.
264 * what we need to do here is just: complete the master_bio.
265 *
266 * local completion error, if any, has been stored as ERR_PTR
fcefa62e 267 * in private_bio within drbd_request_endio.
b411b363
PR
268 */
269 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
270 int error = PTR_ERR(req->private_bio);
271
272 /* remove the request from the conflict detection
273 * respective block_id verification hash */
dac1389c
AG
274 if (!drbd_interval_empty(&req->i)) {
275 struct rb_root *root;
276
dac1389c
AG
277 if (rw == WRITE)
278 root = &mdev->write_requests;
279 else
280 root = &mdev->read_requests;
53840641 281 drbd_remove_request_interval(root, req);
7be8da07 282 } else if (!(s & RQ_POSTPONED))
8825f7c3 283 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
b411b363
PR
284
285 /* for writes we need to do some extra housekeeping */
286 if (rw == WRITE)
287 _about_to_complete_local_write(mdev, req);
288
289 /* Update disk stats */
290 _drbd_end_io_acct(mdev, req);
291
7be8da07
AG
292 if (!(s & RQ_POSTPONED)) {
293 m->error = ok ? 0 : (error ?: -EIO);
294 m->bio = req->master_bio;
2312f0b3
LE
295 req->master_bio = NULL;
296 } else {
297 /* Assert that this will be _req_is_done()
298 * with this very invokation. */
299 /* FIXME:
300 * what about (RQ_LOCAL_PENDING | RQ_LOCAL_ABORTED)?
301 */
302 D_ASSERT(!(s & RQ_LOCAL_PENDING));
303 D_ASSERT(s & RQ_NET_DONE);
7be8da07 304 }
b411b363
PR
305 }
306
cdfda633
PR
307 if (s & RQ_LOCAL_PENDING)
308 return;
309
b411b363
PR
310 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
311 /* this is disconnected (local only) operation,
a209b4ae 312 * or protocol A, B, or C P_BARRIER_ACK,
b411b363
PR
313 * or killed from the transfer log due to connection loss. */
314 _req_is_done(mdev, req, rw);
315 }
316 /* else: network part and not DONE yet. that is
a209b4ae 317 * protocol A, B, or C, barrier ack still pending... */
b411b363
PR
318}
319
cfa03415
PR
320static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
321{
a21e9298 322 struct drbd_conf *mdev = req->w.mdev;
cfa03415 323
2aebfabb 324 if (!drbd_suspended(mdev))
cfa03415
PR
325 _req_may_be_done(req, m);
326}
327
b411b363
PR
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).
332 *
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 ...
339 */
2a80699f 340int __req_mod(struct drbd_request *req, enum drbd_req_event what,
b411b363
PR
341 struct bio_and_error *m)
342{
a21e9298 343 struct drbd_conf *mdev = req->w.mdev;
44ed167d 344 struct net_conf *nc;
303d1448 345 int p, rv = 0;
7be8da07
AG
346
347 if (m)
348 m->bio = NULL;
b411b363 349
b411b363
PR
350 switch (what) {
351 default:
352 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
353 break;
354
355 /* does not happen...
356 * initialization done in drbd_req_new
8554df1c 357 case CREATED:
b411b363
PR
358 break;
359 */
360
8554df1c 361 case TO_BE_SENT: /* via network */
7be8da07 362 /* reached via __drbd_make_request
b411b363
PR
363 * and from w_read_retry_remote */
364 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
365 req->rq_state |= RQ_NET_PENDING;
44ed167d
PR
366 rcu_read_lock();
367 nc = rcu_dereference(mdev->tconn->net_conf);
368 p = nc->wire_protocol;
369 rcu_read_unlock();
303d1448
PR
370 req->rq_state |=
371 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
372 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
b411b363
PR
373 inc_ap_pending(mdev);
374 break;
375
8554df1c 376 case TO_BE_SUBMITTED: /* locally */
7be8da07 377 /* reached via __drbd_make_request */
b411b363
PR
378 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
379 req->rq_state |= RQ_LOCAL_PENDING;
380 break;
381
8554df1c 382 case COMPLETED_OK:
cdfda633 383 if (req->rq_state & RQ_WRITE)
ace652ac 384 mdev->writ_cnt += req->i.size >> 9;
b411b363 385 else
ace652ac 386 mdev->read_cnt += req->i.size >> 9;
b411b363
PR
387
388 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
389 req->rq_state &= ~RQ_LOCAL_PENDING;
390
cfa03415 391 _req_may_be_done_not_susp(req, m);
b411b363
PR
392 put_ldev(mdev);
393 break;
394
cdfda633
PR
395 case ABORT_DISK_IO:
396 req->rq_state |= RQ_LOCAL_ABORTED;
397 if (req->rq_state & RQ_WRITE)
398 _req_may_be_done_not_susp(req, m);
399 else
400 goto goto_queue_for_net_read;
401 break;
402
8554df1c 403 case WRITE_COMPLETED_WITH_ERROR:
b411b363
PR
404 req->rq_state |= RQ_LOCAL_COMPLETED;
405 req->rq_state &= ~RQ_LOCAL_PENDING;
406
81e84650 407 __drbd_chk_io_error(mdev, false);
cfa03415 408 _req_may_be_done_not_susp(req, m);
b411b363
PR
409 put_ldev(mdev);
410 break;
411
8554df1c 412 case READ_AHEAD_COMPLETED_WITH_ERROR:
b411b363
PR
413 /* it is legal to fail READA */
414 req->rq_state |= RQ_LOCAL_COMPLETED;
415 req->rq_state &= ~RQ_LOCAL_PENDING;
cfa03415 416 _req_may_be_done_not_susp(req, m);
b411b363
PR
417 put_ldev(mdev);
418 break;
419
8554df1c 420 case READ_COMPLETED_WITH_ERROR:
ace652ac 421 drbd_set_out_of_sync(mdev, req->i.sector, req->i.size);
b411b363
PR
422
423 req->rq_state |= RQ_LOCAL_COMPLETED;
424 req->rq_state &= ~RQ_LOCAL_PENDING;
425
b411b363 426 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
b411b363 427
81e84650 428 __drbd_chk_io_error(mdev, false);
b411b363 429 put_ldev(mdev);
b411b363 430
cdfda633
PR
431 goto_queue_for_net_read:
432
d255e5ff
LE
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) {
cfa03415 436 _req_may_be_done_not_susp(req, m);
d255e5ff
LE
437 break;
438 }
439
8554df1c 440 /* _req_mod(req,TO_BE_SENT); oops, recursion... */
d255e5ff
LE
441 req->rq_state |= RQ_NET_PENDING;
442 inc_ap_pending(mdev);
8554df1c 443 /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */
b411b363 444
8554df1c 445 case QUEUE_FOR_NET_READ:
b411b363
PR
446 /* READ or READA, and
447 * no local disk,
448 * or target area marked as invalid,
449 * or just got an io-error. */
7be8da07 450 /* from __drbd_make_request
b411b363
PR
451 * or from bio_endio during read io-error recovery */
452
453 /* so we can verify the handle in the answer packet
454 * corresponding hlist_del is in _req_may_be_done() */
97ddb687 455 D_ASSERT(drbd_interval_empty(&req->i));
dac1389c 456 drbd_insert_interval(&mdev->read_requests, &req->i);
b411b363 457
83c38830 458 set_bit(UNPLUG_REMOTE, &mdev->flags);
b411b363
PR
459
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
464 : w_send_read_req;
e42325a5 465 drbd_queue_work(&mdev->tconn->data.work, &req->w);
b411b363
PR
466 break;
467
8554df1c 468 case QUEUE_FOR_NET_WRITE:
b411b363 469 /* assert something? */
7be8da07 470 /* from __drbd_make_request only */
b411b363 471
b411b363 472 /* corresponding hlist_del is in _req_may_be_done() */
97ddb687 473 D_ASSERT(drbd_interval_empty(&req->i));
de696716 474 drbd_insert_interval(&mdev->write_requests, &req->i);
b411b363
PR
475
476 /* NOTE
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.
482 *
483 * _req_add_to_epoch(req); this has to be after the
484 * _maybe_start_new_epoch(req); which happened in
7be8da07 485 * __drbd_make_request, because we now may set the bit
b411b363
PR
486 * again ourselves to close the current epoch.
487 *
488 * Add req to the (now) current epoch (barrier). */
489
83c38830
LE
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);
494
7be8da07 495 /* see __drbd_make_request,
b411b363 496 * just after it grabs the req_lock */
6936fcb4 497 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->tconn->flags) == 0);
b411b363 498
87eeee41 499 req->epoch = mdev->tconn->newest_tle->br_number;
b411b363
PR
500
501 /* increment size of current epoch */
87eeee41 502 mdev->tconn->newest_tle->n_writes++;
b411b363
PR
503
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;
e42325a5 508 drbd_queue_work(&mdev->tconn->data.work, &req->w);
b411b363
PR
509
510 /* close the epoch, in case it outgrew the limit */
44ed167d
PR
511 rcu_read_lock();
512 nc = rcu_dereference(mdev->tconn->net_conf);
513 p = nc->max_epoch_size;
514 rcu_read_unlock();
515 if (mdev->tconn->newest_tle->n_writes >= p)
b411b363
PR
516 queue_barrier(mdev);
517
518 break;
519
8554df1c 520 case QUEUE_FOR_SEND_OOS:
73a01a18 521 req->rq_state |= RQ_NET_QUEUED;
8f7bed77 522 req->w.cb = w_send_out_of_sync;
e42325a5 523 drbd_queue_work(&mdev->tconn->data.work, &req->w);
73a01a18
PR
524 break;
525
8554df1c 526 case OOS_HANDED_TO_NETWORK:
73a01a18 527 /* actually the same */
8554df1c 528 case SEND_CANCELED:
b411b363 529 /* treat it the same */
8554df1c 530 case SEND_FAILED:
b411b363
PR
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! */
cfa03415 536 _req_may_be_done_not_susp(req, m);
b411b363
PR
537 break;
538
8554df1c 539 case HANDED_OVER_TO_NETWORK:
b411b363 540 /* assert something? */
759fbdfb 541 if (bio_data_dir(req->master_bio) == WRITE)
ace652ac 542 atomic_add(req->i.size >> 9, &mdev->ap_in_flight);
759fbdfb 543
b411b363 544 if (bio_data_dir(req->master_bio) == WRITE &&
303d1448 545 !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
b411b363
PR
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 */
556 }
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
8554df1c
AG
560 * dereference the bio even after the "WRITE_ACKED_BY_PEER" and
561 * "COMPLETED_OK" events came in, once we return from
b411b363
PR
562 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
563 * whether it is done already, and end it. */
cfa03415 564 _req_may_be_done_not_susp(req, m);
b411b363
PR
565 break;
566
8554df1c 567 case READ_RETRY_REMOTE_CANCELED:
d255e5ff
LE
568 req->rq_state &= ~RQ_NET_QUEUED;
569 /* fall through, in case we raced with drbd_disconnect */
8554df1c 570 case CONNECTION_LOST_WHILE_PENDING:
b411b363
PR
571 /* transfer log cleanup after connection loss */
572 /* assert something? */
573 if (req->rq_state & RQ_NET_PENDING)
574 dec_ap_pending(mdev);
57bcb6cf
PR
575
576 p = !(req->rq_state & RQ_WRITE) && req->rq_state & RQ_NET_PENDING;
577
b411b363
PR
578 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
579 req->rq_state |= RQ_NET_DONE;
759fbdfb 580 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
ace652ac 581 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
759fbdfb 582
b411b363
PR
583 /* if it is still queued, we may not complete it here.
584 * it will be canceled soon. */
57bcb6cf
PR
585 if (!(req->rq_state & RQ_NET_QUEUED)) {
586 if (p)
587 goto goto_read_retry_local;
cfa03415 588 _req_may_be_done(req, m); /* Allowed while state.susp */
57bcb6cf 589 }
b411b363
PR
590 break;
591
8554df1c 592 case WRITE_ACKED_BY_PEER_AND_SIS:
b411b363 593 req->rq_state |= RQ_NET_SIS;
7be8da07 594 case DISCARD_WRITE:
b411b363
PR
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. */
b411b363
PR
598 req->rq_state |= RQ_NET_DONE;
599 /* fall through */
8554df1c 600 case WRITE_ACKED_BY_PEER:
303d1448 601 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
b411b363
PR
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.
606 *
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. */
611
303d1448 612 goto ack_common;
b411b363 613 /* this makes it effectively the same as for: */
8554df1c 614 case RECV_ACKED_BY_PEER:
303d1448 615 D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
b411b363 616 /* protocol B; pretends to be successfully written on peer.
8554df1c 617 * see also notes above in HANDED_OVER_TO_NETWORK about
b411b363 618 * protocol != C */
303d1448 619 ack_common:
b411b363
PR
620 req->rq_state |= RQ_NET_OK;
621 D_ASSERT(req->rq_state & RQ_NET_PENDING);
622 dec_ap_pending(mdev);
ace652ac 623 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
b411b363 624 req->rq_state &= ~RQ_NET_PENDING;
cfa03415 625 _req_may_be_done_not_susp(req, m);
b411b363
PR
626 break;
627
7be8da07 628 case POSTPONE_WRITE:
303d1448
PR
629 D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
630 /* If this node has already detected the write conflict, the
7be8da07
AG
631 * worker will be waiting on misc_wait. Wake it up once this
632 * request has completed locally.
633 */
634 D_ASSERT(req->rq_state & RQ_NET_PENDING);
635 req->rq_state |= RQ_POSTPONED;
636 _req_may_be_done_not_susp(req, m);
637 break;
638
8554df1c 639 case NEG_ACKED:
b411b363 640 /* assert something? */
759fbdfb 641 if (req->rq_state & RQ_NET_PENDING) {
b411b363 642 dec_ap_pending(mdev);
e8cdc343
PR
643 if (req->rq_state & RQ_WRITE)
644 atomic_sub(req->i.size >> 9, &mdev->ap_in_flight);
759fbdfb 645 }
b411b363
PR
646 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
647
648 req->rq_state |= RQ_NET_DONE;
380207d0 649
57bcb6cf 650 if (!(req->rq_state & RQ_WRITE))
380207d0
PR
651 goto goto_read_retry_local;
652
cfa03415 653 _req_may_be_done_not_susp(req, m);
8554df1c 654 /* else: done by HANDED_OVER_TO_NETWORK */
b411b363
PR
655 break;
656
380207d0 657 goto_read_retry_local:
57bcb6cf
PR
658 if (!drbd_may_do_local_read(mdev, req->i.sector, req->i.size)) {
659 _req_may_be_done_not_susp(req, m);
660 break;
661 }
662 D_ASSERT(!(req->rq_state & RQ_LOCAL_PENDING));
380207d0 663 req->rq_state |= RQ_LOCAL_PENDING;
57bcb6cf
PR
664
665 get_ldev(mdev);
666 req->w.cb = w_restart_disk_io;
667 drbd_queue_work(&mdev->tconn->data.work, &req->w);
380207d0
PR
668 break;
669
8554df1c 670 case FAIL_FROZEN_DISK_IO:
265be2d0
PR
671 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
672 break;
673
cfa03415 674 _req_may_be_done(req, m); /* Allowed while state.susp */
265be2d0
PR
675 break;
676
8554df1c 677 case RESTART_FROZEN_DISK_IO:
265be2d0
PR
678 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
679 break;
680
681 req->rq_state &= ~RQ_LOCAL_COMPLETED;
682
683 rv = MR_READ;
684 if (bio_data_dir(req->master_bio) == WRITE)
685 rv = MR_WRITE;
686
687 get_ldev(mdev);
688 req->w.cb = w_restart_disk_io;
e42325a5 689 drbd_queue_work(&mdev->tconn->data.work, &req->w);
265be2d0
PR
690 break;
691
8554df1c 692 case RESEND:
11b58e73 693 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
47ff2d0a 694 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
11b58e73 695 Trowing them out of the TL here by pretending we got a BARRIER_ACK
481c6f50 696 We ensure that the peer was not rebooted */
11b58e73
PR
697 if (!(req->rq_state & RQ_NET_OK)) {
698 if (req->w.cb) {
e42325a5 699 drbd_queue_work(&mdev->tconn->data.work, &req->w);
11b58e73
PR
700 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
701 }
702 break;
703 }
8554df1c 704 /* else, fall through to BARRIER_ACKED */
11b58e73 705
8554df1c 706 case BARRIER_ACKED:
288f422e
PR
707 if (!(req->rq_state & RQ_WRITE))
708 break;
709
b411b363 710 if (req->rq_state & RQ_NET_PENDING) {
a209b4ae 711 /* barrier came in before all requests were acked.
b411b363
PR
712 * this is bad, because if the connection is lost now,
713 * we won't be able to clean them up... */
8554df1c 714 dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
87eeee41 715 list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests);
b411b363 716 }
e636db5b
LE
717 if ((req->rq_state & RQ_NET_MASK) != 0) {
718 req->rq_state |= RQ_NET_DONE;
303d1448 719 if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK)))
89e58e75 720 atomic_sub(req->i.size>>9, &mdev->ap_in_flight);
e636db5b 721 }
cfa03415 722 _req_may_be_done(req, m); /* Allowed while state.susp */
b411b363
PR
723 break;
724
8554df1c 725 case DATA_RECEIVED:
b411b363
PR
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);
cfa03415 730 _req_may_be_done_not_susp(req, m);
b411b363
PR
731 break;
732 };
2a80699f
PR
733
734 return rv;
b411b363
PR
735}
736
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.
743 */
0da34df0 744static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
b411b363
PR
745{
746 unsigned long sbnr, ebnr;
747 sector_t esector, nr_sectors;
748
749 if (mdev->state.disk == D_UP_TO_DATE)
0da34df0 750 return true;
8c387def 751 if (mdev->state.disk != D_INCONSISTENT)
0da34df0 752 return false;
b411b363 753 esector = sector + (size >> 9) - 1;
8ca9844f 754 nr_sectors = drbd_get_capacity(mdev->this_bdev);
b411b363
PR
755 D_ASSERT(sector < nr_sectors);
756 D_ASSERT(esector < nr_sectors);
757
758 sbnr = BM_SECT_TO_BIT(sector);
759 ebnr = BM_SECT_TO_BIT(esector);
760
0da34df0 761 return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0;
b411b363
PR
762}
763
d60de03a 764static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector)
380207d0
PR
765{
766 enum drbd_read_balancing rbm;
767 struct backing_dev_info *bdi;
d60de03a 768 int stripe_shift;
380207d0
PR
769
770 if (mdev->state.pdsk < D_UP_TO_DATE)
771 return false;
772
773 rcu_read_lock();
774 rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing;
775 rcu_read_unlock();
776
777 switch (rbm) {
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);
d60de03a
PR
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;
380207d0
PR
792 case RB_ROUND_ROBIN:
793 return test_and_change_bit(READ_BALANCE_RR, &mdev->flags);
794 case RB_PREFER_REMOTE:
795 return true;
796 case RB_PREFER_LOCAL:
797 default:
798 return false;
799 }
800}
801
6024fece
AG
802/*
803 * complete_conflicting_writes - wait for any conflicting write requests
804 *
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
807 * the new one.
808 */
809static int complete_conflicting_writes(struct drbd_conf *mdev,
810 sector_t sector, int size)
811{
812 for(;;) {
6024fece 813 struct drbd_interval *i;
7be8da07 814 int err;
6024fece
AG
815
816 i = drbd_find_overlap(&mdev->write_requests, sector, size);
817 if (!i)
818 return 0;
7be8da07
AG
819 err = drbd_wait_misc(mdev, i);
820 if (err)
821 return err;
6024fece
AG
822 }
823}
824
7be8da07 825int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
b411b363
PR
826{
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;
44ed167d 832 struct net_conf *nc;
73a01a18 833 int local, remote, send_oos = 0;
6024fece 834 int err;
9a25a04c 835 int ret = 0;
81f44862 836 union drbd_dev_state s;
b411b363
PR
837
838 /* allocate outside of all locks; */
839 req = drbd_req_new(mdev, bio);
840 if (!req) {
841 dec_ap_bio(mdev);
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);
846 return 0;
847 }
aeda1cd6 848 req->start_time = start_time;
b411b363 849
b411b363
PR
850 local = get_ldev(mdev);
851 if (!local) {
852 bio_put(req->private_bio); /* or we get a bio leak */
853 req->private_bio = NULL;
854 }
855 if (rw == WRITE) {
856 remote = 1;
857 } else {
858 /* READ || READA */
859 if (local) {
57bcb6cf
PR
860 if (!drbd_may_do_local_read(mdev, sector, size) ||
861 remote_due_to_read_balancing(mdev, sector)) {
b411b363
PR
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.
866 */
867 local = 0;
868 bio_put(req->private_bio);
869 req->private_bio = NULL;
870 put_ldev(mdev);
871 }
872 }
873 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
874 }
875
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.
879 *
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.
883 */
884 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
885 err = -EWOULDBLOCK;
886 goto fail_and_free_req;
887 }
888
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. */
0778286a
PR
894 if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
895 req->rq_state |= RQ_IN_ACT_LOG;
181286ad 896 drbd_al_begin_io(mdev, &req->i);
0778286a 897 }
b411b363 898
81f44862
LE
899 s = mdev->state;
900 remote = remote && drbd_should_do_remote(s);
901 send_oos = rw == WRITE && drbd_should_send_out_of_sync(s);
3719094e 902 D_ASSERT(!(remote && send_oos));
b411b363 903
2aebfabb 904 if (!(local || remote) && !drbd_suspended(mdev)) {
fb2c7a10
LE
905 if (__ratelimit(&drbd_ratelimit_state))
906 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
6024fece 907 err = -EIO;
b411b363
PR
908 goto fail_free_complete;
909 }
910
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. */
73a01a18 917 if (rw == WRITE && (remote || send_oos) &&
87eeee41 918 mdev->tconn->unused_spare_tle == NULL &&
6936fcb4 919 test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
b411b363
PR
920allocate_barrier:
921 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
922 if (!b) {
923 dev_err(DEV, "Failed to alloc barrier.\n");
924 err = -ENOMEM;
925 goto fail_free_complete;
926 }
927 }
928
929 /* GOOD, everything prepared, grab the spin_lock */
87eeee41 930 spin_lock_irq(&mdev->tconn->req_lock);
b411b363 931
6024fece
AG
932 if (rw == WRITE) {
933 err = complete_conflicting_writes(mdev, sector, size);
934 if (err) {
7be8da07
AG
935 if (err != -ERESTARTSYS)
936 _conn_request_state(mdev->tconn,
937 NS(conn, C_TIMEOUT),
938 CS_HARD);
6024fece 939 spin_unlock_irq(&mdev->tconn->req_lock);
7be8da07 940 err = -EIO;
6024fece
AG
941 goto fail_free_complete;
942 }
943 }
944
2aebfabb 945 if (drbd_suspended(mdev)) {
69b6a3b1
PR
946 /* If we got suspended, use the retry mechanism in
947 drbd_make_request() to restart processing of this
2f58dcfc 948 bio. In the next call to drbd_make_request
9a25a04c
PR
949 we sleep in inc_ap_bio() */
950 ret = 1;
87eeee41 951 spin_unlock_irq(&mdev->tconn->req_lock);
9a25a04c
PR
952 goto fail_free_complete;
953 }
954
73a01a18 955 if (remote || send_oos) {
6a35c45f 956 remote = drbd_should_do_remote(mdev->state);
8f7bed77 957 send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state);
3719094e 958 D_ASSERT(!(remote && send_oos));
73a01a18
PR
959
960 if (!(remote || send_oos))
b411b363
PR
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");
87eeee41 964 spin_unlock_irq(&mdev->tconn->req_lock);
6024fece 965 err = -EIO;
b411b363
PR
966 goto fail_free_complete;
967 }
968 }
969
87eeee41
PR
970 if (b && mdev->tconn->unused_spare_tle == NULL) {
971 mdev->tconn->unused_spare_tle = b;
b411b363
PR
972 b = NULL;
973 }
73a01a18 974 if (rw == WRITE && (remote || send_oos) &&
87eeee41 975 mdev->tconn->unused_spare_tle == NULL &&
6936fcb4 976 test_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
b411b363
PR
977 /* someone closed the current epoch
978 * while we were grabbing the spinlock */
87eeee41 979 spin_unlock_irq(&mdev->tconn->req_lock);
b411b363
PR
980 goto allocate_barrier;
981 }
982
983
984 /* Update disk stats */
985 _drbd_start_io_acct(mdev, req, bio);
986
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.
992 *
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. */
87eeee41 997 if ((remote || send_oos) && mdev->tconn->unused_spare_tle &&
6936fcb4 998 test_and_clear_bit(CREATE_BARRIER, &mdev->tconn->flags)) {
2f5cdd0b 999 _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle);
87eeee41 1000 mdev->tconn->unused_spare_tle = NULL;
b411b363
PR
1001 } else {
1002 D_ASSERT(!(remote && rw == WRITE &&
6936fcb4 1003 test_bit(CREATE_BARRIER, &mdev->tconn->flags)));
b411b363
PR
1004 }
1005
1006 /* NOTE
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.
1011 *
1012 * 'remote' may become wrong any time because the network could fail.
1013 *
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.
1017 */
1018
1019 /* mark them early for readability.
1020 * this just sets some state flags. */
1021 if (remote)
8554df1c 1022 _req_mod(req, TO_BE_SENT);
b411b363 1023 if (local)
8554df1c 1024 _req_mod(req, TO_BE_SUBMITTED);
b411b363 1025
87eeee41 1026 list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests);
288f422e 1027
b411b363
PR
1028 /* NOTE remote first: to get the concurrent write detection right,
1029 * we must register the request before start of local IO. */
1030 if (remote) {
1031 /* either WRITE and C_CONNECTED,
1032 * or READ, and no local disk,
1033 * or READ, but not in sync.
1034 */
1035 _req_mod(req, (rw == WRITE)
8554df1c
AG
1036 ? QUEUE_FOR_NET_WRITE
1037 : QUEUE_FOR_NET_READ);
b411b363 1038 }
73a01a18 1039 if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
8554df1c 1040 _req_mod(req, QUEUE_FOR_SEND_OOS);
67531718 1041
44ed167d
PR
1042 rcu_read_lock();
1043 nc = rcu_dereference(mdev->tconn->net_conf);
73a01a18 1044 if (remote &&
44ed167d 1045 nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) {
67531718
PR
1046 int congested = 0;
1047
44ed167d
PR
1048 if (nc->cong_fill &&
1049 atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) {
67531718
PR
1050 dev_info(DEV, "Congestion-fill threshold reached\n");
1051 congested = 1;
1052 }
1053
44ed167d 1054 if (mdev->act_log->used >= nc->cong_extents) {
67531718
PR
1055 dev_info(DEV, "Congestion-extents threshold reached\n");
1056 congested = 1;
1057 }
1058
71c78cfb 1059 if (congested) {
039312b6 1060 queue_barrier(mdev); /* last barrier, after mirrored writes */
73a01a18 1061
44ed167d 1062 if (nc->on_congestion == OC_PULL_AHEAD)
67531718 1063 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
44ed167d 1064 else /*nc->on_congestion == OC_DISCONNECT */
67531718
PR
1065 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
1066 }
1067 }
44ed167d 1068 rcu_read_unlock();
67531718 1069
87eeee41 1070 spin_unlock_irq(&mdev->tconn->req_lock);
b411b363
PR
1071 kfree(b); /* if someone else has beaten us to it... */
1072
1073 if (local) {
1074 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
1075
6719fb03
LE
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
1080 * this bio. */
1081 if (get_ldev(mdev)) {
0cf9d27e
AG
1082 if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
1083 : rw == READ ? DRBD_FAULT_DT_RD
1084 : DRBD_FAULT_DT_RA))
6719fb03
LE
1085 bio_endio(req->private_bio, -EIO);
1086 else
1087 generic_make_request(req->private_bio);
1088 put_ldev(mdev);
1089 } else
b411b363 1090 bio_endio(req->private_bio, -EIO);
b411b363
PR
1091 }
1092
b411b363
PR
1093 return 0;
1094
1095fail_free_complete:
76727f68 1096 if (req->rq_state & RQ_IN_ACT_LOG)
181286ad 1097 drbd_al_complete_io(mdev, &req->i);
b411b363 1098fail_and_free_req:
57bcb6cf 1099 if (local) {
b411b363
PR
1100 bio_put(req->private_bio);
1101 req->private_bio = NULL;
1102 put_ldev(mdev);
1103 }
9a25a04c
PR
1104 if (!ret)
1105 bio_endio(bio, err);
1106
b411b363
PR
1107 drbd_req_free(req);
1108 dec_ap_bio(mdev);
1109 kfree(b);
1110
9a25a04c 1111 return ret;
b411b363
PR
1112}
1113
2f58dcfc 1114int drbd_make_request(struct request_queue *q, struct bio *bio)
b411b363 1115{
b411b363 1116 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
aeda1cd6 1117 unsigned long start_time;
b411b363 1118
aeda1cd6
PR
1119 start_time = jiffies;
1120
b411b363
PR
1121 /*
1122 * what we "blindly" assume:
1123 */
1124 D_ASSERT(bio->bi_size > 0);
c670a398 1125 D_ASSERT(IS_ALIGNED(bio->bi_size, 512));
b411b363 1126
69b6a3b1
PR
1127 do {
1128 inc_ap_bio(mdev);
1129 } while (__drbd_make_request(mdev, bio, start_time));
1130
1131 return 0;
b411b363
PR
1132}
1133
23361cf3
LE
1134/* This is called by bio_add_page().
1135 *
1136 * q->max_hw_sectors and other global limits are already enforced there.
b411b363 1137 *
23361cf3
LE
1138 * We need to call down to our lower level device,
1139 * in case it has special restrictions.
1140 *
1141 * We also may need to enforce configured max-bio-bvecs limits.
b411b363
PR
1142 *
1143 * As long as the BIO is empty we have to allow at least one bvec,
23361cf3 1144 * regardless of size and offset, so no need to ask lower levels.
b411b363
PR
1145 */
1146int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1147{
1148 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
b411b363 1149 unsigned int bio_size = bvm->bi_size;
23361cf3
LE
1150 int limit = DRBD_MAX_BIO_SIZE;
1151 int backing_limit;
1152
1153 if (bio_size && get_ldev(mdev)) {
b411b363
PR
1154 struct request_queue * const b =
1155 mdev->ldev->backing_bdev->bd_disk->queue;
a1c88d0d 1156 if (b->merge_bvec_fn) {
b411b363
PR
1157 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1158 limit = min(limit, backing_limit);
1159 }
1160 put_ldev(mdev);
1161 }
1162 return limit;
1163}
7fde2be9
PR
1164
1165void request_timer_fn(unsigned long data)
1166{
1167 struct drbd_conf *mdev = (struct drbd_conf *) data;
8b924f1d 1168 struct drbd_tconn *tconn = mdev->tconn;
7fde2be9
PR
1169 struct drbd_request *req; /* oldest request */
1170 struct list_head *le;
44ed167d 1171 struct net_conf *nc;
3b03ad59 1172 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
44ed167d
PR
1173
1174 rcu_read_lock();
1175 nc = rcu_dereference(tconn->net_conf);
cdfda633
PR
1176 ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0;
1177
1178 if (get_ldev(mdev)) {
1179 dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
1180 put_ldev(mdev);
1181 }
44ed167d 1182 rcu_read_unlock();
7fde2be9 1183
cdfda633
PR
1184 et = min_not_zero(dt, ent);
1185
1186 if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED))
7fde2be9
PR
1187 return; /* Recurring timer stopped */
1188
8b924f1d
PR
1189 spin_lock_irq(&tconn->req_lock);
1190 le = &tconn->oldest_tle->requests;
7fde2be9 1191 if (list_empty(le)) {
8b924f1d 1192 spin_unlock_irq(&tconn->req_lock);
7fde2be9
PR
1193 mod_timer(&mdev->request_timer, jiffies + et);
1194 return;
1195 }
1196
1197 le = le->prev;
1198 req = list_entry(le, struct drbd_request, tl_requests);
cdfda633
PR
1199 if (ent && req->rq_state & RQ_NET_PENDING) {
1200 if (time_is_before_eq_jiffies(req->start_time + ent)) {
7fde2be9 1201 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
cdfda633
PR
1202 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1203 }
1204 }
38a05c16 1205 if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev) {
cdfda633
PR
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);
7fde2be9 1209 }
7fde2be9 1210 }
3b03ad59 1211 nt = (time_is_before_eq_jiffies(req->start_time + et) ? jiffies : req->start_time) + et;
8b924f1d 1212 spin_unlock_irq(&tconn->req_lock);
3b03ad59 1213 mod_timer(&mdev->request_timer, nt);
7fde2be9 1214}