blob: 5c0c8be1bb0ae144e4d82ef9d3da257ca06af99d [file] [log] [blame]
Philipp Reisnerb411b362009-09-25 16:07:19 -07001/*
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
Philipp Reisnerb411b362009-09-25 16:07:19 -070026#include <linux/module.h>
27
28#include <linux/slab.h>
29#include <linux/drbd.h>
30#include "drbd_int.h"
Philipp Reisnerb411b362009-09-25 16:07:19 -070031#include "drbd_req.h"
32
33
34/* Update disk stats at start of I/O request */
35static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio)
36{
37 const int rw = bio_data_dir(bio);
38 int cpu;
39 cpu = part_stat_lock();
40 part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]);
41 part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio));
Philipp Reisner753c8912009-11-18 15:52:51 +010042 part_inc_in_flight(&mdev->vdisk->part0, rw);
Philipp Reisnerb411b362009-09-25 16:07:19 -070043 part_stat_unlock();
Philipp Reisnerb411b362009-09-25 16:07:19 -070044}
45
46/* Update disk stats when completing request upwards */
47static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req)
48{
49 int rw = bio_data_dir(req->master_bio);
50 unsigned long duration = jiffies - req->start_time;
51 int cpu;
52 cpu = part_stat_lock();
53 part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration);
54 part_round_stats(cpu, &mdev->vdisk->part0);
Philipp Reisner753c8912009-11-18 15:52:51 +010055 part_dec_in_flight(&mdev->vdisk->part0, rw);
Philipp Reisnerb411b362009-09-25 16:07:19 -070056 part_stat_unlock();
Philipp Reisnerb411b362009-09-25 16:07:19 -070057}
58
59static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw)
60{
61 const unsigned long s = req->rq_state;
Philipp Reisner288f4222010-05-27 15:07:43 +020062
63 /* remove it from the transfer log.
64 * well, only if it had been there in the first
65 * place... if it had not (local only or conflicting
66 * and never sent), it should still be "empty" as
67 * initialized in drbd_req_new(), so we can list_del() it
68 * here unconditionally */
69 list_del(&req->tl_requests);
70
Philipp Reisnerb411b362009-09-25 16:07:19 -070071 /* if it was a write, we may have to set the corresponding
72 * bit(s) out-of-sync first. If it had a local part, we need to
73 * release the reference to the activity log. */
74 if (rw == WRITE) {
Philipp Reisnerb411b362009-09-25 16:07:19 -070075 /* 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);
81
82 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
83 drbd_set_in_sync(mdev, req->sector, req->size);
84
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.
94 */
95 if (s & RQ_LOCAL_MASK) {
96 if (get_ldev_if_state(mdev, D_FAILED)) {
Philipp Reisner07782862010-08-31 12:00:50 +020097 if (s & RQ_IN_ACT_LOG)
98 drbd_al_complete_io(mdev, req->sector);
Philipp Reisnerb411b362009-09-25 16:07:19 -070099 put_ldev(mdev);
100 } else if (__ratelimit(&drbd_ratelimit_state)) {
101 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu), "
102 "but my Disk seems to have failed :(\n",
103 (unsigned long long) req->sector);
104 }
105 }
106 }
107
Philipp Reisner32fa7e92010-05-26 17:13:18 +0200108 drbd_req_free(req);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700109}
110
111static void queue_barrier(struct drbd_conf *mdev)
112{
113 struct drbd_tl_epoch *b;
114
115 /* We are within the req_lock. Once we queued the barrier for sending,
116 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
117 * barrier/epoch object is added. This is the only place this bit is
118 * set. It indicates that the barrier for this epoch is already queued,
119 * and no new epoch has been created yet. */
120 if (test_bit(CREATE_BARRIER, &mdev->flags))
121 return;
122
123 b = mdev->newest_tle;
124 b->w.cb = w_send_barrier;
125 /* inc_ap_pending done here, so we won't
126 * get imbalanced on connection loss.
127 * dec_ap_pending will be done in got_BarrierAck
128 * or (on connection loss) in tl_clear. */
129 inc_ap_pending(mdev);
130 drbd_queue_work(&mdev->data.work, &b->w);
131 set_bit(CREATE_BARRIER, &mdev->flags);
132}
133
134static void _about_to_complete_local_write(struct drbd_conf *mdev,
135 struct drbd_request *req)
136{
137 const unsigned long s = req->rq_state;
138 struct drbd_request *i;
139 struct drbd_epoch_entry *e;
140 struct hlist_node *n;
141 struct hlist_head *slot;
142
Lars Ellenberg8a3c1042010-12-05 14:11:14 +0100143 /* Before we can signal completion to the upper layers,
144 * we may need to close the current epoch.
145 * We can skip this, if this request has not even been sent, because we
146 * did not have a fully established connection yet/anymore, during
147 * bitmap exchange, or while we are C_AHEAD due to congestion policy.
148 */
149 if (mdev->state.conn >= C_CONNECTED &&
150 (s & RQ_NET_SENT) != 0 &&
Philipp Reisnerb411b362009-09-25 16:07:19 -0700151 req->epoch == mdev->newest_tle->br_number)
152 queue_barrier(mdev);
153
154 /* we need to do the conflict detection stuff,
155 * if we have the ee_hash (two_primaries) and
156 * this has been on the network */
157 if ((s & RQ_NET_DONE) && mdev->ee_hash != NULL) {
158 const sector_t sector = req->sector;
159 const int size = req->size;
160
161 /* ASSERT:
162 * there must be no conflicting requests, since
163 * they must have been failed on the spot */
164#define OVERLAPS overlaps(sector, size, i->sector, i->size)
165 slot = tl_hash_slot(mdev, sector);
166 hlist_for_each_entry(i, n, slot, colision) {
167 if (OVERLAPS) {
168 dev_alert(DEV, "LOGIC BUG: completed: %p %llus +%u; "
169 "other: %p %llus +%u\n",
170 req, (unsigned long long)sector, size,
171 i, (unsigned long long)i->sector, i->size);
172 }
173 }
174
175 /* maybe "wake" those conflicting epoch entries
176 * that wait for this request to finish.
177 *
178 * currently, there can be only _one_ such ee
179 * (well, or some more, which would be pending
180 * P_DISCARD_ACK not yet sent by the asender...),
181 * since we block the receiver thread upon the
182 * first conflict detection, which will wait on
183 * misc_wait. maybe we want to assert that?
184 *
185 * anyways, if we found one,
186 * we just have to do a wake_up. */
187#undef OVERLAPS
188#define OVERLAPS overlaps(sector, size, e->sector, e->size)
189 slot = ee_hash_slot(mdev, req->sector);
190 hlist_for_each_entry(e, n, slot, colision) {
191 if (OVERLAPS) {
192 wake_up(&mdev->misc_wait);
193 break;
194 }
195 }
196 }
197#undef OVERLAPS
198}
199
200void complete_master_bio(struct drbd_conf *mdev,
201 struct bio_and_error *m)
202{
Philipp Reisnerb411b362009-09-25 16:07:19 -0700203 bio_endio(m->bio, m->error);
204 dec_ap_bio(mdev);
205}
206
207/* Helper for __req_mod().
208 * Set m->bio to the master bio, if it is fit to be completed,
209 * or leave it alone (it is initialized to NULL in __req_mod),
210 * if it has already been completed, or cannot be completed yet.
211 * If m->bio is set, the error status to be returned is placed in m->error.
212 */
213void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
214{
215 const unsigned long s = req->rq_state;
216 struct drbd_conf *mdev = req->mdev;
217 /* only WRITES may end up here without a master bio (on barrier ack) */
218 int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE;
219
Philipp Reisnerb411b362009-09-25 16:07:19 -0700220 /* we must not complete the master bio, while it is
221 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
222 * not yet acknowledged by the peer
223 * not yet completed by the local io subsystem
224 * these flags may get cleared in any order by
225 * the worker,
226 * the receiver,
227 * the bio_endio completion callbacks.
228 */
229 if (s & RQ_NET_QUEUED)
230 return;
231 if (s & RQ_NET_PENDING)
232 return;
233 if (s & RQ_LOCAL_PENDING)
234 return;
235
236 if (req->master_bio) {
237 /* this is data_received (remote read)
238 * or protocol C P_WRITE_ACK
239 * or protocol B P_RECV_ACK
240 * or protocol A "handed_over_to_network" (SendAck)
241 * or canceled or failed,
242 * or killed from the transfer log due to connection loss.
243 */
244
245 /*
246 * figure out whether to report success or failure.
247 *
248 * report success when at least one of the operations succeeded.
249 * or, to put the other way,
250 * only report failure, when both operations failed.
251 *
252 * what to do about the failures is handled elsewhere.
253 * what we need to do here is just: complete the master_bio.
254 *
255 * local completion error, if any, has been stored as ERR_PTR
256 * in private_bio within drbd_endio_pri.
257 */
258 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
259 int error = PTR_ERR(req->private_bio);
260
261 /* remove the request from the conflict detection
262 * respective block_id verification hash */
263 if (!hlist_unhashed(&req->colision))
264 hlist_del(&req->colision);
265 else
Philipp Reisner8825f7c2010-10-21 17:21:19 +0200266 D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700267
268 /* for writes we need to do some extra housekeeping */
269 if (rw == WRITE)
270 _about_to_complete_local_write(mdev, req);
271
272 /* Update disk stats */
273 _drbd_end_io_acct(mdev, req);
274
275 m->error = ok ? 0 : (error ?: -EIO);
276 m->bio = req->master_bio;
277 req->master_bio = NULL;
278 }
279
280 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
281 /* this is disconnected (local only) operation,
282 * or protocol C P_WRITE_ACK,
283 * or protocol A or B P_BARRIER_ACK,
284 * or killed from the transfer log due to connection loss. */
285 _req_is_done(mdev, req, rw);
286 }
287 /* else: network part and not DONE yet. that is
288 * protocol A or B, barrier ack still pending... */
289}
290
Philipp Reisnercfa03412010-06-23 17:18:51 +0200291static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m)
292{
293 struct drbd_conf *mdev = req->mdev;
294
Philipp Reisnerfb22c402010-09-08 23:20:21 +0200295 if (!is_susp(mdev->state))
Philipp Reisnercfa03412010-06-23 17:18:51 +0200296 _req_may_be_done(req, m);
297}
298
Philipp Reisnerb411b362009-09-25 16:07:19 -0700299/*
300 * checks whether there was an overlapping request
301 * or ee already registered.
302 *
303 * if so, return 1, in which case this request is completed on the spot,
304 * without ever being submitted or send.
305 *
306 * return 0 if it is ok to submit this request.
307 *
308 * NOTE:
309 * paranoia: assume something above us is broken, and issues different write
310 * requests for the same block simultaneously...
311 *
312 * To ensure these won't be reordered differently on both nodes, resulting in
313 * diverging data sets, we discard the later one(s). Not that this is supposed
314 * to happen, but this is the rationale why we also have to check for
315 * conflicting requests with local origin, and why we have to do so regardless
316 * of whether we allowed multiple primaries.
317 *
318 * BTW, in case we only have one primary, the ee_hash is empty anyways, and the
319 * second hlist_for_each_entry becomes a noop. This is even simpler than to
320 * grab a reference on the net_conf, and check for the two_primaries flag...
321 */
322static int _req_conflicts(struct drbd_request *req)
323{
324 struct drbd_conf *mdev = req->mdev;
325 const sector_t sector = req->sector;
326 const int size = req->size;
327 struct drbd_request *i;
328 struct drbd_epoch_entry *e;
329 struct hlist_node *n;
330 struct hlist_head *slot;
331
332 D_ASSERT(hlist_unhashed(&req->colision));
333
334 if (!get_net_conf(mdev))
335 return 0;
336
337 /* BUG_ON */
338 ERR_IF (mdev->tl_hash_s == 0)
339 goto out_no_conflict;
340 BUG_ON(mdev->tl_hash == NULL);
341
342#define OVERLAPS overlaps(i->sector, i->size, sector, size)
343 slot = tl_hash_slot(mdev, sector);
344 hlist_for_each_entry(i, n, slot, colision) {
345 if (OVERLAPS) {
346 dev_alert(DEV, "%s[%u] Concurrent local write detected! "
347 "[DISCARD L] new: %llus +%u; "
348 "pending: %llus +%u\n",
349 current->comm, current->pid,
350 (unsigned long long)sector, size,
351 (unsigned long long)i->sector, i->size);
352 goto out_conflict;
353 }
354 }
355
356 if (mdev->ee_hash_s) {
357 /* now, check for overlapping requests with remote origin */
358 BUG_ON(mdev->ee_hash == NULL);
359#undef OVERLAPS
360#define OVERLAPS overlaps(e->sector, e->size, sector, size)
361 slot = ee_hash_slot(mdev, sector);
362 hlist_for_each_entry(e, n, slot, colision) {
363 if (OVERLAPS) {
364 dev_alert(DEV, "%s[%u] Concurrent remote write detected!"
365 " [DISCARD L] new: %llus +%u; "
366 "pending: %llus +%u\n",
367 current->comm, current->pid,
368 (unsigned long long)sector, size,
369 (unsigned long long)e->sector, e->size);
370 goto out_conflict;
371 }
372 }
373 }
374#undef OVERLAPS
375
376out_no_conflict:
377 /* this is like it should be, and what we expected.
378 * our users do behave after all... */
379 put_net_conf(mdev);
380 return 0;
381
382out_conflict:
383 put_net_conf(mdev);
384 return 1;
385}
386
387/* obviously this could be coded as many single functions
388 * instead of one huge switch,
389 * or by putting the code directly in the respective locations
390 * (as it has been before).
391 *
392 * but having it this way
393 * enforces that it is all in this one place, where it is easier to audit,
394 * it makes it obvious that whatever "event" "happens" to a request should
395 * happen "atomically" within the req_lock,
396 * and it enforces that we have to think in a very structured manner
397 * about the "events" that may happen to a request during its life time ...
398 */
Philipp Reisner2a806992010-06-09 14:07:43 +0200399int __req_mod(struct drbd_request *req, enum drbd_req_event what,
Philipp Reisnerb411b362009-09-25 16:07:19 -0700400 struct bio_and_error *m)
401{
402 struct drbd_conf *mdev = req->mdev;
Philipp Reisner2a806992010-06-09 14:07:43 +0200403 int rv = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700404 m->bio = NULL;
405
Philipp Reisnerb411b362009-09-25 16:07:19 -0700406 switch (what) {
407 default:
408 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
409 break;
410
411 /* does not happen...
412 * initialization done in drbd_req_new
413 case created:
414 break;
415 */
416
417 case to_be_send: /* via network */
418 /* reached via drbd_make_request_common
419 * and from w_read_retry_remote */
420 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
421 req->rq_state |= RQ_NET_PENDING;
422 inc_ap_pending(mdev);
423 break;
424
425 case to_be_submitted: /* locally */
426 /* reached via drbd_make_request_common */
427 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
428 req->rq_state |= RQ_LOCAL_PENDING;
429 break;
430
431 case completed_ok:
432 if (bio_data_dir(req->master_bio) == WRITE)
433 mdev->writ_cnt += req->size>>9;
434 else
435 mdev->read_cnt += req->size>>9;
436
437 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
438 req->rq_state &= ~RQ_LOCAL_PENDING;
439
Philipp Reisnercfa03412010-06-23 17:18:51 +0200440 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700441 put_ldev(mdev);
442 break;
443
444 case write_completed_with_error:
445 req->rq_state |= RQ_LOCAL_COMPLETED;
446 req->rq_state &= ~RQ_LOCAL_PENDING;
447
Andreas Gruenbacher81e84652010-12-09 15:03:57 +0100448 __drbd_chk_io_error(mdev, false);
Philipp Reisnercfa03412010-06-23 17:18:51 +0200449 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700450 put_ldev(mdev);
451 break;
452
453 case read_ahead_completed_with_error:
454 /* it is legal to fail READA */
455 req->rq_state |= RQ_LOCAL_COMPLETED;
456 req->rq_state &= ~RQ_LOCAL_PENDING;
Philipp Reisnercfa03412010-06-23 17:18:51 +0200457 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700458 put_ldev(mdev);
459 break;
460
461 case read_completed_with_error:
462 drbd_set_out_of_sync(mdev, req->sector, req->size);
463
464 req->rq_state |= RQ_LOCAL_COMPLETED;
465 req->rq_state &= ~RQ_LOCAL_PENDING;
466
Philipp Reisnerb411b362009-09-25 16:07:19 -0700467 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
Philipp Reisnerb411b362009-09-25 16:07:19 -0700468
Andreas Gruenbacher81e84652010-12-09 15:03:57 +0100469 __drbd_chk_io_error(mdev, false);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700470 put_ldev(mdev);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700471
Lars Ellenbergd255e5f2010-05-27 09:45:45 +0200472 /* no point in retrying if there is no good remote data,
473 * or we have no connection. */
474 if (mdev->state.pdsk != D_UP_TO_DATE) {
Philipp Reisnercfa03412010-06-23 17:18:51 +0200475 _req_may_be_done_not_susp(req, m);
Lars Ellenbergd255e5f2010-05-27 09:45:45 +0200476 break;
477 }
478
479 /* _req_mod(req,to_be_send); oops, recursion... */
480 req->rq_state |= RQ_NET_PENDING;
481 inc_ap_pending(mdev);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700482 /* fall through: _req_mod(req,queue_for_net_read); */
483
484 case queue_for_net_read:
485 /* READ or READA, and
486 * no local disk,
487 * or target area marked as invalid,
488 * or just got an io-error. */
489 /* from drbd_make_request_common
490 * or from bio_endio during read io-error recovery */
491
492 /* so we can verify the handle in the answer packet
493 * corresponding hlist_del is in _req_may_be_done() */
494 hlist_add_head(&req->colision, ar_hash_slot(mdev, req->sector));
495
Lars Ellenberg83c38832009-11-03 02:22:06 +0100496 set_bit(UNPLUG_REMOTE, &mdev->flags);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700497
498 D_ASSERT(req->rq_state & RQ_NET_PENDING);
499 req->rq_state |= RQ_NET_QUEUED;
500 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
501 ? w_read_retry_remote
502 : w_send_read_req;
503 drbd_queue_work(&mdev->data.work, &req->w);
504 break;
505
506 case queue_for_net_write:
507 /* assert something? */
508 /* from drbd_make_request_common only */
509
510 hlist_add_head(&req->colision, tl_hash_slot(mdev, req->sector));
511 /* corresponding hlist_del is in _req_may_be_done() */
512
513 /* NOTE
514 * In case the req ended up on the transfer log before being
515 * queued on the worker, it could lead to this request being
516 * missed during cleanup after connection loss.
517 * So we have to do both operations here,
518 * within the same lock that protects the transfer log.
519 *
520 * _req_add_to_epoch(req); this has to be after the
521 * _maybe_start_new_epoch(req); which happened in
522 * drbd_make_request_common, because we now may set the bit
523 * again ourselves to close the current epoch.
524 *
525 * Add req to the (now) current epoch (barrier). */
526
Lars Ellenberg83c38832009-11-03 02:22:06 +0100527 /* otherwise we may lose an unplug, which may cause some remote
528 * io-scheduler timeout to expire, increasing maximum latency,
529 * hurting performance. */
530 set_bit(UNPLUG_REMOTE, &mdev->flags);
531
Philipp Reisnerb411b362009-09-25 16:07:19 -0700532 /* see drbd_make_request_common,
533 * just after it grabs the req_lock */
534 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
535
536 req->epoch = mdev->newest_tle->br_number;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700537
538 /* increment size of current epoch */
Philipp Reisner7e602c02010-05-27 14:49:27 +0200539 mdev->newest_tle->n_writes++;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700540
541 /* queue work item to send data */
542 D_ASSERT(req->rq_state & RQ_NET_PENDING);
543 req->rq_state |= RQ_NET_QUEUED;
544 req->w.cb = w_send_dblock;
545 drbd_queue_work(&mdev->data.work, &req->w);
546
547 /* close the epoch, in case it outgrew the limit */
Philipp Reisner7e602c02010-05-27 14:49:27 +0200548 if (mdev->newest_tle->n_writes >= mdev->net_conf->max_epoch_size)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700549 queue_barrier(mdev);
550
551 break;
552
Philipp Reisner73a01a12010-10-27 14:33:00 +0200553 case queue_for_send_oos:
554 req->rq_state |= RQ_NET_QUEUED;
555 req->w.cb = w_send_oos;
556 drbd_queue_work(&mdev->data.work, &req->w);
557 break;
558
559 case oos_handed_to_network:
560 /* actually the same */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700561 case send_canceled:
562 /* treat it the same */
563 case send_failed:
564 /* real cleanup will be done from tl_clear. just update flags
565 * so it is no longer marked as on the worker queue */
566 req->rq_state &= ~RQ_NET_QUEUED;
567 /* if we did it right, tl_clear should be scheduled only after
568 * this, so this should not be necessary! */
Philipp Reisnercfa03412010-06-23 17:18:51 +0200569 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700570 break;
571
572 case handed_over_to_network:
573 /* assert something? */
Philipp Reisner759fbdf2010-10-26 16:02:27 +0200574 if (bio_data_dir(req->master_bio) == WRITE)
575 atomic_add(req->size>>9, &mdev->ap_in_flight);
576
Philipp Reisnerb411b362009-09-25 16:07:19 -0700577 if (bio_data_dir(req->master_bio) == WRITE &&
578 mdev->net_conf->wire_protocol == DRBD_PROT_A) {
579 /* this is what is dangerous about protocol A:
580 * pretend it was successfully written on the peer. */
581 if (req->rq_state & RQ_NET_PENDING) {
582 dec_ap_pending(mdev);
583 req->rq_state &= ~RQ_NET_PENDING;
584 req->rq_state |= RQ_NET_OK;
585 } /* else: neg-ack was faster... */
586 /* it is still not yet RQ_NET_DONE until the
587 * corresponding epoch barrier got acked as well,
588 * so we know what to dirty on connection loss */
589 }
590 req->rq_state &= ~RQ_NET_QUEUED;
591 req->rq_state |= RQ_NET_SENT;
592 /* because _drbd_send_zc_bio could sleep, and may want to
593 * dereference the bio even after the "write_acked_by_peer" and
594 * "completed_ok" events came in, once we return from
595 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
596 * whether it is done already, and end it. */
Philipp Reisnercfa03412010-06-23 17:18:51 +0200597 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700598 break;
599
Lars Ellenbergd255e5f2010-05-27 09:45:45 +0200600 case read_retry_remote_canceled:
601 req->rq_state &= ~RQ_NET_QUEUED;
602 /* fall through, in case we raced with drbd_disconnect */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700603 case connection_lost_while_pending:
604 /* transfer log cleanup after connection loss */
605 /* assert something? */
606 if (req->rq_state & RQ_NET_PENDING)
607 dec_ap_pending(mdev);
608 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
609 req->rq_state |= RQ_NET_DONE;
Philipp Reisner759fbdf2010-10-26 16:02:27 +0200610 if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE)
611 atomic_sub(req->size>>9, &mdev->ap_in_flight);
612
Philipp Reisnerb411b362009-09-25 16:07:19 -0700613 /* if it is still queued, we may not complete it here.
614 * it will be canceled soon. */
615 if (!(req->rq_state & RQ_NET_QUEUED))
Philipp Reisnercfa03412010-06-23 17:18:51 +0200616 _req_may_be_done(req, m); /* Allowed while state.susp */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700617 break;
618
619 case write_acked_by_peer_and_sis:
620 req->rq_state |= RQ_NET_SIS;
621 case conflict_discarded_by_peer:
622 /* for discarded conflicting writes of multiple primaries,
623 * there is no need to keep anything in the tl, potential
624 * node crashes are covered by the activity log. */
625 if (what == conflict_discarded_by_peer)
626 dev_alert(DEV, "Got DiscardAck packet %llus +%u!"
627 " DRBD is not a random data generator!\n",
628 (unsigned long long)req->sector, req->size);
629 req->rq_state |= RQ_NET_DONE;
630 /* fall through */
631 case write_acked_by_peer:
632 /* protocol C; successfully written on peer.
633 * Nothing to do here.
634 * We want to keep the tl in place for all protocols, to cater
635 * for volatile write-back caches on lower level devices.
636 *
637 * A barrier request is expected to have forced all prior
638 * requests onto stable storage, so completion of a barrier
639 * request could set NET_DONE right here, and not wait for the
640 * P_BARRIER_ACK, but that is an unnecessary optimization. */
641
642 /* this makes it effectively the same as for: */
643 case recv_acked_by_peer:
644 /* protocol B; pretends to be successfully written on peer.
645 * see also notes above in handed_over_to_network about
646 * protocol != C */
647 req->rq_state |= RQ_NET_OK;
648 D_ASSERT(req->rq_state & RQ_NET_PENDING);
649 dec_ap_pending(mdev);
Philipp Reisner759fbdf2010-10-26 16:02:27 +0200650 atomic_sub(req->size>>9, &mdev->ap_in_flight);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700651 req->rq_state &= ~RQ_NET_PENDING;
Philipp Reisnercfa03412010-06-23 17:18:51 +0200652 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700653 break;
654
655 case neg_acked:
656 /* assert something? */
Philipp Reisner759fbdf2010-10-26 16:02:27 +0200657 if (req->rq_state & RQ_NET_PENDING) {
Philipp Reisnerb411b362009-09-25 16:07:19 -0700658 dec_ap_pending(mdev);
Philipp Reisner759fbdf2010-10-26 16:02:27 +0200659 atomic_sub(req->size>>9, &mdev->ap_in_flight);
660 }
Philipp Reisnerb411b362009-09-25 16:07:19 -0700661 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
662
663 req->rq_state |= RQ_NET_DONE;
Philipp Reisnercfa03412010-06-23 17:18:51 +0200664 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700665 /* else: done by handed_over_to_network */
666 break;
667
Philipp Reisner265be2d2010-05-31 10:14:17 +0200668 case fail_frozen_disk_io:
669 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
670 break;
671
Philipp Reisnercfa03412010-06-23 17:18:51 +0200672 _req_may_be_done(req, m); /* Allowed while state.susp */
Philipp Reisner265be2d2010-05-31 10:14:17 +0200673 break;
674
675 case restart_frozen_disk_io:
676 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
677 break;
678
679 req->rq_state &= ~RQ_LOCAL_COMPLETED;
680
681 rv = MR_READ;
682 if (bio_data_dir(req->master_bio) == WRITE)
683 rv = MR_WRITE;
684
685 get_ldev(mdev);
686 req->w.cb = w_restart_disk_io;
687 drbd_queue_work(&mdev->data.work, &req->w);
688 break;
689
Philipp Reisner11b58e72010-05-12 17:08:26 +0200690 case resend:
691 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
Philipp Reisner47ff2d02010-06-18 13:56:57 +0200692 before the connection loss (B&C only); only P_BARRIER_ACK was missing.
Philipp Reisner11b58e72010-05-12 17:08:26 +0200693 Trowing them out of the TL here by pretending we got a BARRIER_ACK
Philipp Reisner481c6f52010-06-22 14:03:27 +0200694 We ensure that the peer was not rebooted */
Philipp Reisner11b58e72010-05-12 17:08:26 +0200695 if (!(req->rq_state & RQ_NET_OK)) {
696 if (req->w.cb) {
697 drbd_queue_work(&mdev->data.work, &req->w);
698 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
699 }
700 break;
701 }
702 /* else, fall through to barrier_acked */
703
Philipp Reisnerb411b362009-09-25 16:07:19 -0700704 case barrier_acked:
Philipp Reisner288f4222010-05-27 15:07:43 +0200705 if (!(req->rq_state & RQ_WRITE))
706 break;
707
Philipp Reisnerb411b362009-09-25 16:07:19 -0700708 if (req->rq_state & RQ_NET_PENDING) {
709 /* barrier came in before all requests have been acked.
710 * this is bad, because if the connection is lost now,
711 * we won't be able to clean them up... */
712 dev_err(DEV, "FIXME (barrier_acked but pending)\n");
Philipp Reisnerb411b362009-09-25 16:07:19 -0700713 list_move(&req->tl_requests, &mdev->out_of_sequence_requests);
714 }
Lars Ellenberge636db52011-01-21 17:10:37 +0100715 if ((req->rq_state & RQ_NET_MASK) != 0) {
716 req->rq_state |= RQ_NET_DONE;
717 if (mdev->net_conf->wire_protocol == DRBD_PROT_A)
718 atomic_sub(req->size>>9, &mdev->ap_in_flight);
719 }
Philipp Reisnercfa03412010-06-23 17:18:51 +0200720 _req_may_be_done(req, m); /* Allowed while state.susp */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700721 break;
722
723 case data_received:
724 D_ASSERT(req->rq_state & RQ_NET_PENDING);
725 dec_ap_pending(mdev);
726 req->rq_state &= ~RQ_NET_PENDING;
727 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
Philipp Reisnercfa03412010-06-23 17:18:51 +0200728 _req_may_be_done_not_susp(req, m);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700729 break;
730 };
Philipp Reisner2a806992010-06-09 14:07:43 +0200731
732 return rv;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700733}
734
735/* we may do a local read if:
736 * - we are consistent (of course),
737 * - or we are generally inconsistent,
738 * BUT we are still/already IN SYNC for this area.
739 * since size may be bigger than BM_BLOCK_SIZE,
740 * we may need to check several bits.
741 */
742static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
743{
744 unsigned long sbnr, ebnr;
745 sector_t esector, nr_sectors;
746
747 if (mdev->state.disk == D_UP_TO_DATE)
748 return 1;
749 if (mdev->state.disk >= D_OUTDATED)
750 return 0;
751 if (mdev->state.disk < D_INCONSISTENT)
752 return 0;
753 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
754 nr_sectors = drbd_get_capacity(mdev->this_bdev);
755 esector = sector + (size >> 9) - 1;
756
757 D_ASSERT(sector < nr_sectors);
758 D_ASSERT(esector < nr_sectors);
759
760 sbnr = BM_SECT_TO_BIT(sector);
761 ebnr = BM_SECT_TO_BIT(esector);
762
763 return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr);
764}
765
Philipp Reisneraeda1cd62010-11-09 17:45:06 +0100766static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time)
Philipp Reisnerb411b362009-09-25 16:07:19 -0700767{
768 const int rw = bio_rw(bio);
769 const int size = bio->bi_size;
770 const sector_t sector = bio->bi_sector;
771 struct drbd_tl_epoch *b = NULL;
772 struct drbd_request *req;
Philipp Reisner73a01a12010-10-27 14:33:00 +0200773 int local, remote, send_oos = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700774 int err = -EIO;
Philipp Reisner9a25a042010-05-10 16:42:23 +0200775 int ret = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700776
777 /* allocate outside of all locks; */
778 req = drbd_req_new(mdev, bio);
779 if (!req) {
780 dec_ap_bio(mdev);
781 /* only pass the error to the upper layers.
782 * if user cannot handle io errors, that's not our business. */
783 dev_err(DEV, "could not kmalloc() req\n");
784 bio_endio(bio, -ENOMEM);
785 return 0;
786 }
Philipp Reisneraeda1cd62010-11-09 17:45:06 +0100787 req->start_time = start_time;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700788
Philipp Reisnerb411b362009-09-25 16:07:19 -0700789 local = get_ldev(mdev);
790 if (!local) {
791 bio_put(req->private_bio); /* or we get a bio leak */
792 req->private_bio = NULL;
793 }
794 if (rw == WRITE) {
795 remote = 1;
796 } else {
797 /* READ || READA */
798 if (local) {
799 if (!drbd_may_do_local_read(mdev, sector, size)) {
800 /* we could kick the syncer to
801 * sync this extent asap, wait for
802 * it, then continue locally.
803 * Or just issue the request remotely.
804 */
805 local = 0;
806 bio_put(req->private_bio);
807 req->private_bio = NULL;
808 put_ldev(mdev);
809 }
810 }
811 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
812 }
813
814 /* If we have a disk, but a READA request is mapped to remote,
815 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
816 * Just fail that READA request right here.
817 *
818 * THINK: maybe fail all READA when not local?
819 * or make this configurable...
820 * if network is slow, READA won't do any good.
821 */
822 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
823 err = -EWOULDBLOCK;
824 goto fail_and_free_req;
825 }
826
827 /* For WRITES going to the local disk, grab a reference on the target
828 * extent. This waits for any resync activity in the corresponding
829 * resync extent to finish, and, if necessary, pulls in the target
830 * extent into the activity log, which involves further disk io because
831 * of transactional on-disk meta data updates. */
Philipp Reisner07782862010-08-31 12:00:50 +0200832 if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) {
833 req->rq_state |= RQ_IN_ACT_LOG;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700834 drbd_al_begin_io(mdev, sector);
Philipp Reisner07782862010-08-31 12:00:50 +0200835 }
Philipp Reisnerb411b362009-09-25 16:07:19 -0700836
Philipp Reisner6a35c452011-01-17 20:27:30 +0100837 remote = remote && drbd_should_do_remote(mdev->state);
838 send_oos = rw == WRITE && drbd_should_send_oos(mdev->state);
Philipp Reisner37190942010-11-10 12:08:37 +0100839 D_ASSERT(!(remote && send_oos));
Philipp Reisnerb411b362009-09-25 16:07:19 -0700840
Philipp Reisnerfb22c402010-09-08 23:20:21 +0200841 if (!(local || remote) && !is_susp(mdev->state)) {
Lars Ellenbergfb2c7a12010-10-19 12:08:13 +0200842 if (__ratelimit(&drbd_ratelimit_state))
843 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
Philipp Reisnerb411b362009-09-25 16:07:19 -0700844 goto fail_free_complete;
845 }
846
847 /* For WRITE request, we have to make sure that we have an
848 * unused_spare_tle, in case we need to start a new epoch.
849 * I try to be smart and avoid to pre-allocate always "just in case",
850 * but there is a race between testing the bit and pointer outside the
851 * spinlock, and grabbing the spinlock.
852 * if we lost that race, we retry. */
Philipp Reisner73a01a12010-10-27 14:33:00 +0200853 if (rw == WRITE && (remote || send_oos) &&
Philipp Reisnerb411b362009-09-25 16:07:19 -0700854 mdev->unused_spare_tle == NULL &&
855 test_bit(CREATE_BARRIER, &mdev->flags)) {
856allocate_barrier:
857 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
858 if (!b) {
859 dev_err(DEV, "Failed to alloc barrier.\n");
860 err = -ENOMEM;
861 goto fail_free_complete;
862 }
863 }
864
865 /* GOOD, everything prepared, grab the spin_lock */
866 spin_lock_irq(&mdev->req_lock);
867
Philipp Reisnerfb22c402010-09-08 23:20:21 +0200868 if (is_susp(mdev->state)) {
Philipp Reisner9a25a042010-05-10 16:42:23 +0200869 /* If we got suspended, use the retry mechanism of
870 generic_make_request() to restart processing of this
Andreas Gruenbacher2f58dcf2010-12-13 17:48:19 +0100871 bio. In the next call to drbd_make_request
Philipp Reisner9a25a042010-05-10 16:42:23 +0200872 we sleep in inc_ap_bio() */
873 ret = 1;
874 spin_unlock_irq(&mdev->req_lock);
875 goto fail_free_complete;
876 }
877
Philipp Reisner73a01a12010-10-27 14:33:00 +0200878 if (remote || send_oos) {
Philipp Reisner6a35c452011-01-17 20:27:30 +0100879 remote = drbd_should_do_remote(mdev->state);
880 send_oos = rw == WRITE && drbd_should_send_oos(mdev->state);
Philipp Reisner37190942010-11-10 12:08:37 +0100881 D_ASSERT(!(remote && send_oos));
Philipp Reisner73a01a12010-10-27 14:33:00 +0200882
883 if (!(remote || send_oos))
Philipp Reisnerb411b362009-09-25 16:07:19 -0700884 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
885 if (!(local || remote)) {
886 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
887 spin_unlock_irq(&mdev->req_lock);
888 goto fail_free_complete;
889 }
890 }
891
892 if (b && mdev->unused_spare_tle == NULL) {
893 mdev->unused_spare_tle = b;
894 b = NULL;
895 }
Philipp Reisner73a01a12010-10-27 14:33:00 +0200896 if (rw == WRITE && (remote || send_oos) &&
Philipp Reisnerb411b362009-09-25 16:07:19 -0700897 mdev->unused_spare_tle == NULL &&
898 test_bit(CREATE_BARRIER, &mdev->flags)) {
899 /* someone closed the current epoch
900 * while we were grabbing the spinlock */
901 spin_unlock_irq(&mdev->req_lock);
902 goto allocate_barrier;
903 }
904
905
906 /* Update disk stats */
907 _drbd_start_io_acct(mdev, req, bio);
908
909 /* _maybe_start_new_epoch(mdev);
910 * If we need to generate a write barrier packet, we have to add the
911 * new epoch (barrier) object, and queue the barrier packet for sending,
912 * and queue the req's data after it _within the same lock_, otherwise
913 * we have race conditions were the reorder domains could be mixed up.
914 *
915 * Even read requests may start a new epoch and queue the corresponding
916 * barrier packet. To get the write ordering right, we only have to
917 * make sure that, if this is a write request and it triggered a
918 * barrier packet, this request is queued within the same spinlock. */
Philipp Reisner73a01a12010-10-27 14:33:00 +0200919 if ((remote || send_oos) && mdev->unused_spare_tle &&
Philipp Reisnerb411b362009-09-25 16:07:19 -0700920 test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
921 _tl_add_barrier(mdev, mdev->unused_spare_tle);
922 mdev->unused_spare_tle = NULL;
923 } else {
924 D_ASSERT(!(remote && rw == WRITE &&
925 test_bit(CREATE_BARRIER, &mdev->flags)));
926 }
927
928 /* NOTE
929 * Actually, 'local' may be wrong here already, since we may have failed
930 * to write to the meta data, and may become wrong anytime because of
931 * local io-error for some other request, which would lead to us
932 * "detaching" the local disk.
933 *
934 * 'remote' may become wrong any time because the network could fail.
935 *
936 * This is a harmless race condition, though, since it is handled
937 * correctly at the appropriate places; so it just defers the failure
938 * of the respective operation.
939 */
940
941 /* mark them early for readability.
942 * this just sets some state flags. */
943 if (remote)
944 _req_mod(req, to_be_send);
945 if (local)
946 _req_mod(req, to_be_submitted);
947
948 /* check this request on the collision detection hash tables.
949 * if we have a conflict, just complete it here.
950 * THINK do we want to check reads, too? (I don't think so...) */
Lars Ellenbergd28fd092010-07-09 23:28:10 +0200951 if (rw == WRITE && _req_conflicts(req))
952 goto fail_conflicting;
Philipp Reisner288f4222010-05-27 15:07:43 +0200953
954 list_add_tail(&req->tl_requests, &mdev->newest_tle->requests);
955
Philipp Reisnerb411b362009-09-25 16:07:19 -0700956 /* NOTE remote first: to get the concurrent write detection right,
957 * we must register the request before start of local IO. */
958 if (remote) {
959 /* either WRITE and C_CONNECTED,
960 * or READ, and no local disk,
961 * or READ, but not in sync.
962 */
963 _req_mod(req, (rw == WRITE)
964 ? queue_for_net_write
965 : queue_for_net_read);
966 }
Philipp Reisner73a01a12010-10-27 14:33:00 +0200967 if (send_oos && drbd_set_out_of_sync(mdev, sector, size))
968 _req_mod(req, queue_for_send_oos);
Philipp Reisner67531712010-10-27 12:21:30 +0200969
Philipp Reisner73a01a12010-10-27 14:33:00 +0200970 if (remote &&
971 mdev->net_conf->on_congestion != OC_BLOCK && mdev->agreed_pro_version >= 96) {
Philipp Reisner67531712010-10-27 12:21:30 +0200972 int congested = 0;
973
974 if (mdev->net_conf->cong_fill &&
975 atomic_read(&mdev->ap_in_flight) >= mdev->net_conf->cong_fill) {
976 dev_info(DEV, "Congestion-fill threshold reached\n");
977 congested = 1;
978 }
979
980 if (mdev->act_log->used >= mdev->net_conf->cong_extents) {
981 dev_info(DEV, "Congestion-extents threshold reached\n");
982 congested = 1;
983 }
984
Philipp Reisner71c78cf2011-01-14 19:20:34 +0100985 if (congested) {
Philipp Reisner039312b2011-01-21 14:13:22 +0100986 queue_barrier(mdev); /* last barrier, after mirrored writes */
Philipp Reisner73a01a12010-10-27 14:33:00 +0200987
Philipp Reisner67531712010-10-27 12:21:30 +0200988 if (mdev->net_conf->on_congestion == OC_PULL_AHEAD)
989 _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL);
990 else /*mdev->net_conf->on_congestion == OC_DISCONNECT */
991 _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL);
992 }
993 }
994
Philipp Reisnerb411b362009-09-25 16:07:19 -0700995 spin_unlock_irq(&mdev->req_lock);
996 kfree(b); /* if someone else has beaten us to it... */
997
998 if (local) {
999 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
1000
Lars Ellenberg6719fb02010-10-18 23:04:07 +02001001 /* State may have changed since we grabbed our reference on the
1002 * mdev->ldev member. Double check, and short-circuit to endio.
1003 * In case the last activity log transaction failed to get on
1004 * stable storage, and this is a WRITE, we may not even submit
1005 * this bio. */
1006 if (get_ldev(mdev)) {
Andreas Gruenbacher0cf9d272010-12-07 10:43:29 +01001007 if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
1008 : rw == READ ? DRBD_FAULT_DT_RD
1009 : DRBD_FAULT_DT_RA))
Lars Ellenberg6719fb02010-10-18 23:04:07 +02001010 bio_endio(req->private_bio, -EIO);
1011 else
1012 generic_make_request(req->private_bio);
1013 put_ldev(mdev);
1014 } else
Philipp Reisnerb411b362009-09-25 16:07:19 -07001015 bio_endio(req->private_bio, -EIO);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001016 }
1017
Philipp Reisnerb411b362009-09-25 16:07:19 -07001018 return 0;
1019
Lars Ellenbergd28fd092010-07-09 23:28:10 +02001020fail_conflicting:
1021 /* this is a conflicting request.
1022 * even though it may have been only _partially_
1023 * overlapping with one of the currently pending requests,
1024 * without even submitting or sending it, we will
1025 * pretend that it was successfully served right now.
1026 */
1027 _drbd_end_io_acct(mdev, req);
1028 spin_unlock_irq(&mdev->req_lock);
1029 if (remote)
1030 dec_ap_pending(mdev);
1031 /* THINK: do we want to fail it (-EIO), or pretend success?
1032 * this pretends success. */
1033 err = 0;
1034
Philipp Reisnerb411b362009-09-25 16:07:19 -07001035fail_free_complete:
1036 if (rw == WRITE && local)
1037 drbd_al_complete_io(mdev, sector);
1038fail_and_free_req:
1039 if (local) {
1040 bio_put(req->private_bio);
1041 req->private_bio = NULL;
1042 put_ldev(mdev);
1043 }
Philipp Reisner9a25a042010-05-10 16:42:23 +02001044 if (!ret)
1045 bio_endio(bio, err);
1046
Philipp Reisnerb411b362009-09-25 16:07:19 -07001047 drbd_req_free(req);
1048 dec_ap_bio(mdev);
1049 kfree(b);
1050
Philipp Reisner9a25a042010-05-10 16:42:23 +02001051 return ret;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001052}
1053
1054/* helper function for drbd_make_request
1055 * if we can determine just by the mdev (state) that this request will fail,
1056 * return 1
1057 * otherwise return 0
1058 */
1059static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write)
1060{
Philipp Reisnerb411b362009-09-25 16:07:19 -07001061 if (mdev->state.role != R_PRIMARY &&
1062 (!allow_oos || is_write)) {
1063 if (__ratelimit(&drbd_ratelimit_state)) {
1064 dev_err(DEV, "Process %s[%u] tried to %s; "
1065 "since we are not in Primary state, "
1066 "we cannot allow this\n",
1067 current->comm, current->pid,
1068 is_write ? "WRITE" : "READ");
1069 }
1070 return 1;
1071 }
1072
Philipp Reisnerb411b362009-09-25 16:07:19 -07001073 return 0;
1074}
1075
Andreas Gruenbacher2f58dcf2010-12-13 17:48:19 +01001076int drbd_make_request(struct request_queue *q, struct bio *bio)
Philipp Reisnerb411b362009-09-25 16:07:19 -07001077{
1078 unsigned int s_enr, e_enr;
1079 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
Philipp Reisneraeda1cd62010-11-09 17:45:06 +01001080 unsigned long start_time;
Philipp Reisnerb411b362009-09-25 16:07:19 -07001081
1082 if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) {
1083 bio_endio(bio, -EPERM);
1084 return 0;
1085 }
1086
Philipp Reisneraeda1cd62010-11-09 17:45:06 +01001087 start_time = jiffies;
1088
Philipp Reisnerb411b362009-09-25 16:07:19 -07001089 /*
1090 * what we "blindly" assume:
1091 */
1092 D_ASSERT(bio->bi_size > 0);
1093 D_ASSERT((bio->bi_size & 0x1ff) == 0);
1094 D_ASSERT(bio->bi_idx == 0);
1095
1096 /* to make some things easier, force alignment of requests within the
1097 * granularity of our hash tables */
1098 s_enr = bio->bi_sector >> HT_SHIFT;
1099 e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT;
1100
1101 if (likely(s_enr == e_enr)) {
1102 inc_ap_bio(mdev, 1);
Philipp Reisneraeda1cd62010-11-09 17:45:06 +01001103 return drbd_make_request_common(mdev, bio, start_time);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001104 }
1105
1106 /* can this bio be split generically?
1107 * Maybe add our own split-arbitrary-bios function. */
Lars Ellenberg1816a2b2010-11-11 15:19:07 +01001108 if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_BIO_SIZE) {
Philipp Reisnerb411b362009-09-25 16:07:19 -07001109 /* rather error out here than BUG in bio_split */
1110 dev_err(DEV, "bio would need to, but cannot, be split: "
1111 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1112 bio->bi_vcnt, bio->bi_idx, bio->bi_size,
1113 (unsigned long long)bio->bi_sector);
1114 bio_endio(bio, -EINVAL);
1115 } else {
1116 /* This bio crosses some boundary, so we have to split it. */
1117 struct bio_pair *bp;
1118 /* works for the "do not cross hash slot boundaries" case
1119 * e.g. sector 262269, size 4096
1120 * s_enr = 262269 >> 6 = 4097
1121 * e_enr = (262269+8-1) >> 6 = 4098
1122 * HT_SHIFT = 6
1123 * sps = 64, mask = 63
1124 * first_sectors = 64 - (262269 & 63) = 3
1125 */
1126 const sector_t sect = bio->bi_sector;
1127 const int sps = 1 << HT_SHIFT; /* sectors per slot */
1128 const int mask = sps - 1;
1129 const sector_t first_sectors = sps - (sect & mask);
Or Gerlitz03567812011-01-13 10:43:40 +01001130 bp = bio_split(bio, first_sectors);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001131
1132 /* we need to get a "reference count" (ap_bio_cnt)
1133 * to avoid races with the disconnect/reconnect/suspend code.
Philipp Reisner9a25a042010-05-10 16:42:23 +02001134 * In case we need to split the bio here, we need to get three references
Philipp Reisnerb411b362009-09-25 16:07:19 -07001135 * atomically, otherwise we might deadlock when trying to submit the
1136 * second one! */
Philipp Reisner9a25a042010-05-10 16:42:23 +02001137 inc_ap_bio(mdev, 3);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001138
1139 D_ASSERT(e_enr == s_enr + 1);
1140
Philipp Reisneraeda1cd62010-11-09 17:45:06 +01001141 while (drbd_make_request_common(mdev, &bp->bio1, start_time))
Philipp Reisner9a25a042010-05-10 16:42:23 +02001142 inc_ap_bio(mdev, 1);
1143
Philipp Reisneraeda1cd62010-11-09 17:45:06 +01001144 while (drbd_make_request_common(mdev, &bp->bio2, start_time))
Philipp Reisner9a25a042010-05-10 16:42:23 +02001145 inc_ap_bio(mdev, 1);
1146
1147 dec_ap_bio(mdev);
1148
Philipp Reisnerb411b362009-09-25 16:07:19 -07001149 bio_pair_release(bp);
1150 }
1151 return 0;
1152}
1153
1154/* This is called by bio_add_page(). With this function we reduce
Lars Ellenberg1816a2b2010-11-11 15:19:07 +01001155 * the number of BIOs that span over multiple DRBD_MAX_BIO_SIZEs
Philipp Reisnerb411b362009-09-25 16:07:19 -07001156 * units (was AL_EXTENTs).
1157 *
1158 * we do the calculation within the lower 32bit of the byte offsets,
1159 * since we don't care for actual offset, but only check whether it
1160 * would cross "activity log extent" boundaries.
1161 *
1162 * As long as the BIO is empty we have to allow at least one bvec,
1163 * regardless of size and offset. so the resulting bio may still
1164 * cross extent boundaries. those are dealt with (bio_split) in
Andreas Gruenbacher2f58dcf2010-12-13 17:48:19 +01001165 * drbd_make_request.
Philipp Reisnerb411b362009-09-25 16:07:19 -07001166 */
1167int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1168{
1169 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1170 unsigned int bio_offset =
1171 (unsigned int)bvm->bi_sector << 9; /* 32 bit */
1172 unsigned int bio_size = bvm->bi_size;
1173 int limit, backing_limit;
1174
Lars Ellenberg1816a2b2010-11-11 15:19:07 +01001175 limit = DRBD_MAX_BIO_SIZE
1176 - ((bio_offset & (DRBD_MAX_BIO_SIZE-1)) + bio_size);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001177 if (limit < 0)
1178 limit = 0;
1179 if (bio_size == 0) {
1180 if (limit <= bvec->bv_len)
1181 limit = bvec->bv_len;
1182 } else if (limit && get_ldev(mdev)) {
1183 struct request_queue * const b =
1184 mdev->ldev->backing_bdev->bd_disk->queue;
Lars Ellenberga1c88d02010-05-14 19:16:41 +02001185 if (b->merge_bvec_fn) {
Philipp Reisnerb411b362009-09-25 16:07:19 -07001186 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1187 limit = min(limit, backing_limit);
1188 }
1189 put_ldev(mdev);
1190 }
1191 return limit;
1192}
Philipp Reisner7fde2be2011-03-01 11:08:28 +01001193
1194void request_timer_fn(unsigned long data)
1195{
1196 struct drbd_conf *mdev = (struct drbd_conf *) data;
1197 struct drbd_request *req; /* oldest request */
1198 struct list_head *le;
1199 unsigned long et = 0; /* effective timeout = ko_count * timeout */
1200
1201 if (get_net_conf(mdev)) {
1202 et = mdev->net_conf->timeout*HZ/10 * mdev->net_conf->ko_count;
1203 put_net_conf(mdev);
1204 }
1205 if (!et || mdev->state.conn < C_WF_REPORT_PARAMS)
1206 return; /* Recurring timer stopped */
1207
1208 spin_lock_irq(&mdev->req_lock);
1209 le = &mdev->oldest_tle->requests;
1210 if (list_empty(le)) {
1211 spin_unlock_irq(&mdev->req_lock);
1212 mod_timer(&mdev->request_timer, jiffies + et);
1213 return;
1214 }
1215
1216 le = le->prev;
1217 req = list_entry(le, struct drbd_request, tl_requests);
1218 if (time_is_before_eq_jiffies(req->start_time + et)) {
1219 if (req->rq_state & RQ_NET_PENDING) {
1220 dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1221 _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE, NULL);
1222 } else {
1223 dev_warn(DEV, "Local backing block device frozen?\n");
1224 mod_timer(&mdev->request_timer, jiffies + et);
1225 }
1226 } else {
1227 mod_timer(&mdev->request_timer, req->start_time + et);
1228 }
1229
1230 spin_unlock_irq(&mdev->req_lock);
1231}