blob: f761d98a4e90320998bd54099ea8c4caa56154ac [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;
62 /* if it was a write, we may have to set the corresponding
63 * bit(s) out-of-sync first. If it had a local part, we need to
64 * release the reference to the activity log. */
65 if (rw == WRITE) {
66 /* remove it from the transfer log.
67 * well, only if it had been there in the first
68 * place... if it had not (local only or conflicting
69 * and never sent), it should still be "empty" as
70 * initialized in drbd_req_new(), so we can list_del() it
71 * here unconditionally */
72 list_del(&req->tl_requests);
73 /* Set out-of-sync unless both OK flags are set
74 * (local only or remote failed).
75 * Other places where we set out-of-sync:
76 * READ with local io-error */
77 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
78 drbd_set_out_of_sync(mdev, req->sector, req->size);
79
80 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
81 drbd_set_in_sync(mdev, req->sector, req->size);
82
83 /* one might be tempted to move the drbd_al_complete_io
84 * to the local io completion callback drbd_endio_pri.
85 * but, if this was a mirror write, we may only
86 * drbd_al_complete_io after this is RQ_NET_DONE,
87 * otherwise the extent could be dropped from the al
88 * before it has actually been written on the peer.
89 * if we crash before our peer knows about the request,
90 * but after the extent has been dropped from the al,
91 * we would forget to resync the corresponding extent.
92 */
93 if (s & RQ_LOCAL_MASK) {
94 if (get_ldev_if_state(mdev, D_FAILED)) {
95 drbd_al_complete_io(mdev, req->sector);
96 put_ldev(mdev);
97 } else if (__ratelimit(&drbd_ratelimit_state)) {
98 dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu), "
99 "but my Disk seems to have failed :(\n",
100 (unsigned long long) req->sector);
101 }
102 }
103 }
104
Philipp Reisner32fa7e92010-05-26 17:13:18 +0200105 drbd_req_free(req);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700106}
107
108static void queue_barrier(struct drbd_conf *mdev)
109{
110 struct drbd_tl_epoch *b;
111
112 /* We are within the req_lock. Once we queued the barrier for sending,
113 * we set the CREATE_BARRIER bit. It is cleared as soon as a new
114 * barrier/epoch object is added. This is the only place this bit is
115 * set. It indicates that the barrier for this epoch is already queued,
116 * and no new epoch has been created yet. */
117 if (test_bit(CREATE_BARRIER, &mdev->flags))
118 return;
119
120 b = mdev->newest_tle;
121 b->w.cb = w_send_barrier;
122 /* inc_ap_pending done here, so we won't
123 * get imbalanced on connection loss.
124 * dec_ap_pending will be done in got_BarrierAck
125 * or (on connection loss) in tl_clear. */
126 inc_ap_pending(mdev);
127 drbd_queue_work(&mdev->data.work, &b->w);
128 set_bit(CREATE_BARRIER, &mdev->flags);
129}
130
131static void _about_to_complete_local_write(struct drbd_conf *mdev,
132 struct drbd_request *req)
133{
134 const unsigned long s = req->rq_state;
135 struct drbd_request *i;
136 struct drbd_epoch_entry *e;
137 struct hlist_node *n;
138 struct hlist_head *slot;
139
140 /* before we can signal completion to the upper layers,
141 * we may need to close the current epoch */
142 if (mdev->state.conn >= C_CONNECTED &&
143 req->epoch == mdev->newest_tle->br_number)
144 queue_barrier(mdev);
145
146 /* we need to do the conflict detection stuff,
147 * if we have the ee_hash (two_primaries) and
148 * this has been on the network */
149 if ((s & RQ_NET_DONE) && mdev->ee_hash != NULL) {
150 const sector_t sector = req->sector;
151 const int size = req->size;
152
153 /* ASSERT:
154 * there must be no conflicting requests, since
155 * they must have been failed on the spot */
156#define OVERLAPS overlaps(sector, size, i->sector, i->size)
157 slot = tl_hash_slot(mdev, sector);
158 hlist_for_each_entry(i, n, slot, colision) {
159 if (OVERLAPS) {
160 dev_alert(DEV, "LOGIC BUG: completed: %p %llus +%u; "
161 "other: %p %llus +%u\n",
162 req, (unsigned long long)sector, size,
163 i, (unsigned long long)i->sector, i->size);
164 }
165 }
166
167 /* maybe "wake" those conflicting epoch entries
168 * that wait for this request to finish.
169 *
170 * currently, there can be only _one_ such ee
171 * (well, or some more, which would be pending
172 * P_DISCARD_ACK not yet sent by the asender...),
173 * since we block the receiver thread upon the
174 * first conflict detection, which will wait on
175 * misc_wait. maybe we want to assert that?
176 *
177 * anyways, if we found one,
178 * we just have to do a wake_up. */
179#undef OVERLAPS
180#define OVERLAPS overlaps(sector, size, e->sector, e->size)
181 slot = ee_hash_slot(mdev, req->sector);
182 hlist_for_each_entry(e, n, slot, colision) {
183 if (OVERLAPS) {
184 wake_up(&mdev->misc_wait);
185 break;
186 }
187 }
188 }
189#undef OVERLAPS
190}
191
192void complete_master_bio(struct drbd_conf *mdev,
193 struct bio_and_error *m)
194{
Philipp Reisnerb411b362009-09-25 16:07:19 -0700195 bio_endio(m->bio, m->error);
196 dec_ap_bio(mdev);
197}
198
199/* Helper for __req_mod().
200 * Set m->bio to the master bio, if it is fit to be completed,
201 * or leave it alone (it is initialized to NULL in __req_mod),
202 * if it has already been completed, or cannot be completed yet.
203 * If m->bio is set, the error status to be returned is placed in m->error.
204 */
205void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
206{
207 const unsigned long s = req->rq_state;
208 struct drbd_conf *mdev = req->mdev;
209 /* only WRITES may end up here without a master bio (on barrier ack) */
210 int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE;
211
Philipp Reisnerb411b362009-09-25 16:07:19 -0700212 /* we must not complete the master bio, while it is
213 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
214 * not yet acknowledged by the peer
215 * not yet completed by the local io subsystem
216 * these flags may get cleared in any order by
217 * the worker,
218 * the receiver,
219 * the bio_endio completion callbacks.
220 */
221 if (s & RQ_NET_QUEUED)
222 return;
223 if (s & RQ_NET_PENDING)
224 return;
225 if (s & RQ_LOCAL_PENDING)
226 return;
227
228 if (req->master_bio) {
229 /* this is data_received (remote read)
230 * or protocol C P_WRITE_ACK
231 * or protocol B P_RECV_ACK
232 * or protocol A "handed_over_to_network" (SendAck)
233 * or canceled or failed,
234 * or killed from the transfer log due to connection loss.
235 */
236
237 /*
238 * figure out whether to report success or failure.
239 *
240 * report success when at least one of the operations succeeded.
241 * or, to put the other way,
242 * only report failure, when both operations failed.
243 *
244 * what to do about the failures is handled elsewhere.
245 * what we need to do here is just: complete the master_bio.
246 *
247 * local completion error, if any, has been stored as ERR_PTR
248 * in private_bio within drbd_endio_pri.
249 */
250 int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
251 int error = PTR_ERR(req->private_bio);
252
253 /* remove the request from the conflict detection
254 * respective block_id verification hash */
255 if (!hlist_unhashed(&req->colision))
256 hlist_del(&req->colision);
257 else
258 D_ASSERT((s & RQ_NET_MASK) == 0);
259
260 /* for writes we need to do some extra housekeeping */
261 if (rw == WRITE)
262 _about_to_complete_local_write(mdev, req);
263
264 /* Update disk stats */
265 _drbd_end_io_acct(mdev, req);
266
267 m->error = ok ? 0 : (error ?: -EIO);
268 m->bio = req->master_bio;
269 req->master_bio = NULL;
270 }
271
272 if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) {
273 /* this is disconnected (local only) operation,
274 * or protocol C P_WRITE_ACK,
275 * or protocol A or B P_BARRIER_ACK,
276 * or killed from the transfer log due to connection loss. */
277 _req_is_done(mdev, req, rw);
278 }
279 /* else: network part and not DONE yet. that is
280 * protocol A or B, barrier ack still pending... */
281}
282
283/*
284 * checks whether there was an overlapping request
285 * or ee already registered.
286 *
287 * if so, return 1, in which case this request is completed on the spot,
288 * without ever being submitted or send.
289 *
290 * return 0 if it is ok to submit this request.
291 *
292 * NOTE:
293 * paranoia: assume something above us is broken, and issues different write
294 * requests for the same block simultaneously...
295 *
296 * To ensure these won't be reordered differently on both nodes, resulting in
297 * diverging data sets, we discard the later one(s). Not that this is supposed
298 * to happen, but this is the rationale why we also have to check for
299 * conflicting requests with local origin, and why we have to do so regardless
300 * of whether we allowed multiple primaries.
301 *
302 * BTW, in case we only have one primary, the ee_hash is empty anyways, and the
303 * second hlist_for_each_entry becomes a noop. This is even simpler than to
304 * grab a reference on the net_conf, and check for the two_primaries flag...
305 */
306static int _req_conflicts(struct drbd_request *req)
307{
308 struct drbd_conf *mdev = req->mdev;
309 const sector_t sector = req->sector;
310 const int size = req->size;
311 struct drbd_request *i;
312 struct drbd_epoch_entry *e;
313 struct hlist_node *n;
314 struct hlist_head *slot;
315
316 D_ASSERT(hlist_unhashed(&req->colision));
317
318 if (!get_net_conf(mdev))
319 return 0;
320
321 /* BUG_ON */
322 ERR_IF (mdev->tl_hash_s == 0)
323 goto out_no_conflict;
324 BUG_ON(mdev->tl_hash == NULL);
325
326#define OVERLAPS overlaps(i->sector, i->size, sector, size)
327 slot = tl_hash_slot(mdev, sector);
328 hlist_for_each_entry(i, n, slot, colision) {
329 if (OVERLAPS) {
330 dev_alert(DEV, "%s[%u] Concurrent local write detected! "
331 "[DISCARD L] new: %llus +%u; "
332 "pending: %llus +%u\n",
333 current->comm, current->pid,
334 (unsigned long long)sector, size,
335 (unsigned long long)i->sector, i->size);
336 goto out_conflict;
337 }
338 }
339
340 if (mdev->ee_hash_s) {
341 /* now, check for overlapping requests with remote origin */
342 BUG_ON(mdev->ee_hash == NULL);
343#undef OVERLAPS
344#define OVERLAPS overlaps(e->sector, e->size, sector, size)
345 slot = ee_hash_slot(mdev, sector);
346 hlist_for_each_entry(e, n, slot, colision) {
347 if (OVERLAPS) {
348 dev_alert(DEV, "%s[%u] Concurrent remote write detected!"
349 " [DISCARD L] new: %llus +%u; "
350 "pending: %llus +%u\n",
351 current->comm, current->pid,
352 (unsigned long long)sector, size,
353 (unsigned long long)e->sector, e->size);
354 goto out_conflict;
355 }
356 }
357 }
358#undef OVERLAPS
359
360out_no_conflict:
361 /* this is like it should be, and what we expected.
362 * our users do behave after all... */
363 put_net_conf(mdev);
364 return 0;
365
366out_conflict:
367 put_net_conf(mdev);
368 return 1;
369}
370
371/* obviously this could be coded as many single functions
372 * instead of one huge switch,
373 * or by putting the code directly in the respective locations
374 * (as it has been before).
375 *
376 * but having it this way
377 * enforces that it is all in this one place, where it is easier to audit,
378 * it makes it obvious that whatever "event" "happens" to a request should
379 * happen "atomically" within the req_lock,
380 * and it enforces that we have to think in a very structured manner
381 * about the "events" that may happen to a request during its life time ...
382 */
383void __req_mod(struct drbd_request *req, enum drbd_req_event what,
384 struct bio_and_error *m)
385{
386 struct drbd_conf *mdev = req->mdev;
387 m->bio = NULL;
388
Philipp Reisnerb411b362009-09-25 16:07:19 -0700389 switch (what) {
390 default:
391 dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
392 break;
393
394 /* does not happen...
395 * initialization done in drbd_req_new
396 case created:
397 break;
398 */
399
400 case to_be_send: /* via network */
401 /* reached via drbd_make_request_common
402 * and from w_read_retry_remote */
403 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
404 req->rq_state |= RQ_NET_PENDING;
405 inc_ap_pending(mdev);
406 break;
407
408 case to_be_submitted: /* locally */
409 /* reached via drbd_make_request_common */
410 D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
411 req->rq_state |= RQ_LOCAL_PENDING;
412 break;
413
414 case completed_ok:
415 if (bio_data_dir(req->master_bio) == WRITE)
416 mdev->writ_cnt += req->size>>9;
417 else
418 mdev->read_cnt += req->size>>9;
419
420 req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
421 req->rq_state &= ~RQ_LOCAL_PENDING;
422
423 _req_may_be_done(req, m);
424 put_ldev(mdev);
425 break;
426
427 case write_completed_with_error:
428 req->rq_state |= RQ_LOCAL_COMPLETED;
429 req->rq_state &= ~RQ_LOCAL_PENDING;
430
Philipp Reisnerb411b362009-09-25 16:07:19 -0700431 __drbd_chk_io_error(mdev, FALSE);
432 _req_may_be_done(req, m);
433 put_ldev(mdev);
434 break;
435
436 case read_ahead_completed_with_error:
437 /* it is legal to fail READA */
438 req->rq_state |= RQ_LOCAL_COMPLETED;
439 req->rq_state &= ~RQ_LOCAL_PENDING;
440 _req_may_be_done(req, m);
441 put_ldev(mdev);
442 break;
443
444 case read_completed_with_error:
445 drbd_set_out_of_sync(mdev, req->sector, req->size);
446
447 req->rq_state |= RQ_LOCAL_COMPLETED;
448 req->rq_state &= ~RQ_LOCAL_PENDING;
449
Philipp Reisnerb411b362009-09-25 16:07:19 -0700450 D_ASSERT(!(req->rq_state & RQ_NET_MASK));
Philipp Reisnerb411b362009-09-25 16:07:19 -0700451
452 __drbd_chk_io_error(mdev, FALSE);
453 put_ldev(mdev);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700454
Lars Ellenbergd255e5f2010-05-27 09:45:45 +0200455 /* no point in retrying if there is no good remote data,
456 * or we have no connection. */
457 if (mdev->state.pdsk != D_UP_TO_DATE) {
458 _req_may_be_done(req, m);
459 break;
460 }
461
462 /* _req_mod(req,to_be_send); oops, recursion... */
463 req->rq_state |= RQ_NET_PENDING;
464 inc_ap_pending(mdev);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700465 /* fall through: _req_mod(req,queue_for_net_read); */
466
467 case queue_for_net_read:
468 /* READ or READA, and
469 * no local disk,
470 * or target area marked as invalid,
471 * or just got an io-error. */
472 /* from drbd_make_request_common
473 * or from bio_endio during read io-error recovery */
474
475 /* so we can verify the handle in the answer packet
476 * corresponding hlist_del is in _req_may_be_done() */
477 hlist_add_head(&req->colision, ar_hash_slot(mdev, req->sector));
478
Lars Ellenberg83c38832009-11-03 02:22:06 +0100479 set_bit(UNPLUG_REMOTE, &mdev->flags);
Philipp Reisnerb411b362009-09-25 16:07:19 -0700480
481 D_ASSERT(req->rq_state & RQ_NET_PENDING);
482 req->rq_state |= RQ_NET_QUEUED;
483 req->w.cb = (req->rq_state & RQ_LOCAL_MASK)
484 ? w_read_retry_remote
485 : w_send_read_req;
486 drbd_queue_work(&mdev->data.work, &req->w);
487 break;
488
489 case queue_for_net_write:
490 /* assert something? */
491 /* from drbd_make_request_common only */
492
493 hlist_add_head(&req->colision, tl_hash_slot(mdev, req->sector));
494 /* corresponding hlist_del is in _req_may_be_done() */
495
496 /* NOTE
497 * In case the req ended up on the transfer log before being
498 * queued on the worker, it could lead to this request being
499 * missed during cleanup after connection loss.
500 * So we have to do both operations here,
501 * within the same lock that protects the transfer log.
502 *
503 * _req_add_to_epoch(req); this has to be after the
504 * _maybe_start_new_epoch(req); which happened in
505 * drbd_make_request_common, because we now may set the bit
506 * again ourselves to close the current epoch.
507 *
508 * Add req to the (now) current epoch (barrier). */
509
Lars Ellenberg83c38832009-11-03 02:22:06 +0100510 /* otherwise we may lose an unplug, which may cause some remote
511 * io-scheduler timeout to expire, increasing maximum latency,
512 * hurting performance. */
513 set_bit(UNPLUG_REMOTE, &mdev->flags);
514
Philipp Reisnerb411b362009-09-25 16:07:19 -0700515 /* see drbd_make_request_common,
516 * just after it grabs the req_lock */
517 D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0);
518
519 req->epoch = mdev->newest_tle->br_number;
520 list_add_tail(&req->tl_requests,
521 &mdev->newest_tle->requests);
522
523 /* increment size of current epoch */
524 mdev->newest_tle->n_req++;
525
526 /* queue work item to send data */
527 D_ASSERT(req->rq_state & RQ_NET_PENDING);
528 req->rq_state |= RQ_NET_QUEUED;
529 req->w.cb = w_send_dblock;
530 drbd_queue_work(&mdev->data.work, &req->w);
531
532 /* close the epoch, in case it outgrew the limit */
533 if (mdev->newest_tle->n_req >= mdev->net_conf->max_epoch_size)
534 queue_barrier(mdev);
535
536 break;
537
538 case send_canceled:
539 /* treat it the same */
540 case send_failed:
541 /* real cleanup will be done from tl_clear. just update flags
542 * so it is no longer marked as on the worker queue */
543 req->rq_state &= ~RQ_NET_QUEUED;
544 /* if we did it right, tl_clear should be scheduled only after
545 * this, so this should not be necessary! */
546 _req_may_be_done(req, m);
547 break;
548
549 case handed_over_to_network:
550 /* assert something? */
551 if (bio_data_dir(req->master_bio) == WRITE &&
552 mdev->net_conf->wire_protocol == DRBD_PROT_A) {
553 /* this is what is dangerous about protocol A:
554 * pretend it was successfully written on the peer. */
555 if (req->rq_state & RQ_NET_PENDING) {
556 dec_ap_pending(mdev);
557 req->rq_state &= ~RQ_NET_PENDING;
558 req->rq_state |= RQ_NET_OK;
559 } /* else: neg-ack was faster... */
560 /* it is still not yet RQ_NET_DONE until the
561 * corresponding epoch barrier got acked as well,
562 * so we know what to dirty on connection loss */
563 }
564 req->rq_state &= ~RQ_NET_QUEUED;
565 req->rq_state |= RQ_NET_SENT;
566 /* because _drbd_send_zc_bio could sleep, and may want to
567 * dereference the bio even after the "write_acked_by_peer" and
568 * "completed_ok" events came in, once we return from
569 * _drbd_send_zc_bio (drbd_send_dblock), we have to check
570 * whether it is done already, and end it. */
571 _req_may_be_done(req, m);
572 break;
573
Lars Ellenbergd255e5f2010-05-27 09:45:45 +0200574 case read_retry_remote_canceled:
575 req->rq_state &= ~RQ_NET_QUEUED;
576 /* fall through, in case we raced with drbd_disconnect */
Philipp Reisnerb411b362009-09-25 16:07:19 -0700577 case connection_lost_while_pending:
578 /* transfer log cleanup after connection loss */
579 /* assert something? */
580 if (req->rq_state & RQ_NET_PENDING)
581 dec_ap_pending(mdev);
582 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
583 req->rq_state |= RQ_NET_DONE;
584 /* if it is still queued, we may not complete it here.
585 * it will be canceled soon. */
586 if (!(req->rq_state & RQ_NET_QUEUED))
587 _req_may_be_done(req, m);
588 break;
589
590 case write_acked_by_peer_and_sis:
591 req->rq_state |= RQ_NET_SIS;
592 case conflict_discarded_by_peer:
593 /* for discarded conflicting writes of multiple primaries,
594 * there is no need to keep anything in the tl, potential
595 * node crashes are covered by the activity log. */
596 if (what == conflict_discarded_by_peer)
597 dev_alert(DEV, "Got DiscardAck packet %llus +%u!"
598 " DRBD is not a random data generator!\n",
599 (unsigned long long)req->sector, req->size);
600 req->rq_state |= RQ_NET_DONE;
601 /* fall through */
602 case write_acked_by_peer:
603 /* protocol C; successfully written on peer.
604 * Nothing to do here.
605 * We want to keep the tl in place for all protocols, to cater
606 * for volatile write-back caches on lower level devices.
607 *
608 * A barrier request is expected to have forced all prior
609 * requests onto stable storage, so completion of a barrier
610 * request could set NET_DONE right here, and not wait for the
611 * P_BARRIER_ACK, but that is an unnecessary optimization. */
612
613 /* this makes it effectively the same as for: */
614 case recv_acked_by_peer:
615 /* protocol B; pretends to be successfully written on peer.
616 * see also notes above in handed_over_to_network about
617 * protocol != C */
618 req->rq_state |= RQ_NET_OK;
619 D_ASSERT(req->rq_state & RQ_NET_PENDING);
620 dec_ap_pending(mdev);
621 req->rq_state &= ~RQ_NET_PENDING;
622 _req_may_be_done(req, m);
623 break;
624
625 case neg_acked:
626 /* assert something? */
627 if (req->rq_state & RQ_NET_PENDING)
628 dec_ap_pending(mdev);
629 req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING);
630
631 req->rq_state |= RQ_NET_DONE;
632 _req_may_be_done(req, m);
633 /* else: done by handed_over_to_network */
634 break;
635
636 case barrier_acked:
637 if (req->rq_state & RQ_NET_PENDING) {
638 /* barrier came in before all requests have been acked.
639 * this is bad, because if the connection is lost now,
640 * we won't be able to clean them up... */
641 dev_err(DEV, "FIXME (barrier_acked but pending)\n");
Philipp Reisnerb411b362009-09-25 16:07:19 -0700642 list_move(&req->tl_requests, &mdev->out_of_sequence_requests);
643 }
644 D_ASSERT(req->rq_state & RQ_NET_SENT);
645 req->rq_state |= RQ_NET_DONE;
646 _req_may_be_done(req, m);
647 break;
648
649 case data_received:
650 D_ASSERT(req->rq_state & RQ_NET_PENDING);
651 dec_ap_pending(mdev);
652 req->rq_state &= ~RQ_NET_PENDING;
653 req->rq_state |= (RQ_NET_OK|RQ_NET_DONE);
654 _req_may_be_done(req, m);
655 break;
656 };
657}
658
659/* we may do a local read if:
660 * - we are consistent (of course),
661 * - or we are generally inconsistent,
662 * BUT we are still/already IN SYNC for this area.
663 * since size may be bigger than BM_BLOCK_SIZE,
664 * we may need to check several bits.
665 */
666static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size)
667{
668 unsigned long sbnr, ebnr;
669 sector_t esector, nr_sectors;
670
671 if (mdev->state.disk == D_UP_TO_DATE)
672 return 1;
673 if (mdev->state.disk >= D_OUTDATED)
674 return 0;
675 if (mdev->state.disk < D_INCONSISTENT)
676 return 0;
677 /* state.disk == D_INCONSISTENT We will have a look at the BitMap */
678 nr_sectors = drbd_get_capacity(mdev->this_bdev);
679 esector = sector + (size >> 9) - 1;
680
681 D_ASSERT(sector < nr_sectors);
682 D_ASSERT(esector < nr_sectors);
683
684 sbnr = BM_SECT_TO_BIT(sector);
685 ebnr = BM_SECT_TO_BIT(esector);
686
687 return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr);
688}
689
690static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio)
691{
692 const int rw = bio_rw(bio);
693 const int size = bio->bi_size;
694 const sector_t sector = bio->bi_sector;
695 struct drbd_tl_epoch *b = NULL;
696 struct drbd_request *req;
697 int local, remote;
698 int err = -EIO;
Philipp Reisner9a25a042010-05-10 16:42:23 +0200699 int ret = 0;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700700
701 /* allocate outside of all locks; */
702 req = drbd_req_new(mdev, bio);
703 if (!req) {
704 dec_ap_bio(mdev);
705 /* only pass the error to the upper layers.
706 * if user cannot handle io errors, that's not our business. */
707 dev_err(DEV, "could not kmalloc() req\n");
708 bio_endio(bio, -ENOMEM);
709 return 0;
710 }
711
Philipp Reisnerb411b362009-09-25 16:07:19 -0700712 local = get_ldev(mdev);
713 if (!local) {
714 bio_put(req->private_bio); /* or we get a bio leak */
715 req->private_bio = NULL;
716 }
717 if (rw == WRITE) {
718 remote = 1;
719 } else {
720 /* READ || READA */
721 if (local) {
722 if (!drbd_may_do_local_read(mdev, sector, size)) {
723 /* we could kick the syncer to
724 * sync this extent asap, wait for
725 * it, then continue locally.
726 * Or just issue the request remotely.
727 */
728 local = 0;
729 bio_put(req->private_bio);
730 req->private_bio = NULL;
731 put_ldev(mdev);
732 }
733 }
734 remote = !local && mdev->state.pdsk >= D_UP_TO_DATE;
735 }
736
737 /* If we have a disk, but a READA request is mapped to remote,
738 * we are R_PRIMARY, D_INCONSISTENT, SyncTarget.
739 * Just fail that READA request right here.
740 *
741 * THINK: maybe fail all READA when not local?
742 * or make this configurable...
743 * if network is slow, READA won't do any good.
744 */
745 if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) {
746 err = -EWOULDBLOCK;
747 goto fail_and_free_req;
748 }
749
750 /* For WRITES going to the local disk, grab a reference on the target
751 * extent. This waits for any resync activity in the corresponding
752 * resync extent to finish, and, if necessary, pulls in the target
753 * extent into the activity log, which involves further disk io because
754 * of transactional on-disk meta data updates. */
755 if (rw == WRITE && local)
756 drbd_al_begin_io(mdev, sector);
757
758 remote = remote && (mdev->state.pdsk == D_UP_TO_DATE ||
759 (mdev->state.pdsk == D_INCONSISTENT &&
760 mdev->state.conn >= C_CONNECTED));
761
Philipp Reisner9a25a042010-05-10 16:42:23 +0200762 if (!(local || remote) && !mdev->state.susp) {
Philipp Reisnerb411b362009-09-25 16:07:19 -0700763 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
764 goto fail_free_complete;
765 }
766
767 /* For WRITE request, we have to make sure that we have an
768 * unused_spare_tle, in case we need to start a new epoch.
769 * I try to be smart and avoid to pre-allocate always "just in case",
770 * but there is a race between testing the bit and pointer outside the
771 * spinlock, and grabbing the spinlock.
772 * if we lost that race, we retry. */
773 if (rw == WRITE && remote &&
774 mdev->unused_spare_tle == NULL &&
775 test_bit(CREATE_BARRIER, &mdev->flags)) {
776allocate_barrier:
777 b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO);
778 if (!b) {
779 dev_err(DEV, "Failed to alloc barrier.\n");
780 err = -ENOMEM;
781 goto fail_free_complete;
782 }
783 }
784
785 /* GOOD, everything prepared, grab the spin_lock */
786 spin_lock_irq(&mdev->req_lock);
787
Philipp Reisner9a25a042010-05-10 16:42:23 +0200788 if (mdev->state.susp) {
789 /* If we got suspended, use the retry mechanism of
790 generic_make_request() to restart processing of this
791 bio. In the next call to drbd_make_request_26
792 we sleep in inc_ap_bio() */
793 ret = 1;
794 spin_unlock_irq(&mdev->req_lock);
795 goto fail_free_complete;
796 }
797
Philipp Reisnerb411b362009-09-25 16:07:19 -0700798 if (remote) {
799 remote = (mdev->state.pdsk == D_UP_TO_DATE ||
800 (mdev->state.pdsk == D_INCONSISTENT &&
801 mdev->state.conn >= C_CONNECTED));
802 if (!remote)
803 dev_warn(DEV, "lost connection while grabbing the req_lock!\n");
804 if (!(local || remote)) {
805 dev_err(DEV, "IO ERROR: neither local nor remote disk\n");
806 spin_unlock_irq(&mdev->req_lock);
807 goto fail_free_complete;
808 }
809 }
810
811 if (b && mdev->unused_spare_tle == NULL) {
812 mdev->unused_spare_tle = b;
813 b = NULL;
814 }
815 if (rw == WRITE && remote &&
816 mdev->unused_spare_tle == NULL &&
817 test_bit(CREATE_BARRIER, &mdev->flags)) {
818 /* someone closed the current epoch
819 * while we were grabbing the spinlock */
820 spin_unlock_irq(&mdev->req_lock);
821 goto allocate_barrier;
822 }
823
824
825 /* Update disk stats */
826 _drbd_start_io_acct(mdev, req, bio);
827
828 /* _maybe_start_new_epoch(mdev);
829 * If we need to generate a write barrier packet, we have to add the
830 * new epoch (barrier) object, and queue the barrier packet for sending,
831 * and queue the req's data after it _within the same lock_, otherwise
832 * we have race conditions were the reorder domains could be mixed up.
833 *
834 * Even read requests may start a new epoch and queue the corresponding
835 * barrier packet. To get the write ordering right, we only have to
836 * make sure that, if this is a write request and it triggered a
837 * barrier packet, this request is queued within the same spinlock. */
838 if (remote && mdev->unused_spare_tle &&
839 test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) {
840 _tl_add_barrier(mdev, mdev->unused_spare_tle);
841 mdev->unused_spare_tle = NULL;
842 } else {
843 D_ASSERT(!(remote && rw == WRITE &&
844 test_bit(CREATE_BARRIER, &mdev->flags)));
845 }
846
847 /* NOTE
848 * Actually, 'local' may be wrong here already, since we may have failed
849 * to write to the meta data, and may become wrong anytime because of
850 * local io-error for some other request, which would lead to us
851 * "detaching" the local disk.
852 *
853 * 'remote' may become wrong any time because the network could fail.
854 *
855 * This is a harmless race condition, though, since it is handled
856 * correctly at the appropriate places; so it just defers the failure
857 * of the respective operation.
858 */
859
860 /* mark them early for readability.
861 * this just sets some state flags. */
862 if (remote)
863 _req_mod(req, to_be_send);
864 if (local)
865 _req_mod(req, to_be_submitted);
866
867 /* check this request on the collision detection hash tables.
868 * if we have a conflict, just complete it here.
869 * THINK do we want to check reads, too? (I don't think so...) */
870 if (rw == WRITE && _req_conflicts(req)) {
871 /* this is a conflicting request.
872 * even though it may have been only _partially_
873 * overlapping with one of the currently pending requests,
874 * without even submitting or sending it, we will
875 * pretend that it was successfully served right now.
876 */
877 if (local) {
878 bio_put(req->private_bio);
879 req->private_bio = NULL;
880 drbd_al_complete_io(mdev, req->sector);
881 put_ldev(mdev);
882 local = 0;
883 }
884 if (remote)
885 dec_ap_pending(mdev);
886 _drbd_end_io_acct(mdev, req);
887 /* THINK: do we want to fail it (-EIO), or pretend success? */
888 bio_endio(req->master_bio, 0);
889 req->master_bio = NULL;
890 dec_ap_bio(mdev);
891 drbd_req_free(req);
892 remote = 0;
893 }
894
895 /* NOTE remote first: to get the concurrent write detection right,
896 * we must register the request before start of local IO. */
897 if (remote) {
898 /* either WRITE and C_CONNECTED,
899 * or READ, and no local disk,
900 * or READ, but not in sync.
901 */
902 _req_mod(req, (rw == WRITE)
903 ? queue_for_net_write
904 : queue_for_net_read);
905 }
906 spin_unlock_irq(&mdev->req_lock);
907 kfree(b); /* if someone else has beaten us to it... */
908
909 if (local) {
910 req->private_bio->bi_bdev = mdev->ldev->backing_bdev;
911
Philipp Reisnerb411b362009-09-25 16:07:19 -0700912 if (FAULT_ACTIVE(mdev, rw == WRITE ? DRBD_FAULT_DT_WR
913 : rw == READ ? DRBD_FAULT_DT_RD
914 : DRBD_FAULT_DT_RA))
915 bio_endio(req->private_bio, -EIO);
916 else
917 generic_make_request(req->private_bio);
918 }
919
920 /* we need to plug ALWAYS since we possibly need to kick lo_dev.
921 * we plug after submit, so we won't miss an unplug event */
922 drbd_plug_device(mdev);
923
924 return 0;
925
926fail_free_complete:
927 if (rw == WRITE && local)
928 drbd_al_complete_io(mdev, sector);
929fail_and_free_req:
930 if (local) {
931 bio_put(req->private_bio);
932 req->private_bio = NULL;
933 put_ldev(mdev);
934 }
Philipp Reisner9a25a042010-05-10 16:42:23 +0200935 if (!ret)
936 bio_endio(bio, err);
937
Philipp Reisnerb411b362009-09-25 16:07:19 -0700938 drbd_req_free(req);
939 dec_ap_bio(mdev);
940 kfree(b);
941
Philipp Reisner9a25a042010-05-10 16:42:23 +0200942 return ret;
Philipp Reisnerb411b362009-09-25 16:07:19 -0700943}
944
945/* helper function for drbd_make_request
946 * if we can determine just by the mdev (state) that this request will fail,
947 * return 1
948 * otherwise return 0
949 */
950static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write)
951{
Philipp Reisnerb411b362009-09-25 16:07:19 -0700952 if (mdev->state.role != R_PRIMARY &&
953 (!allow_oos || is_write)) {
954 if (__ratelimit(&drbd_ratelimit_state)) {
955 dev_err(DEV, "Process %s[%u] tried to %s; "
956 "since we are not in Primary state, "
957 "we cannot allow this\n",
958 current->comm, current->pid,
959 is_write ? "WRITE" : "READ");
960 }
961 return 1;
962 }
963
964 /*
965 * Paranoia: we might have been primary, but sync target, or
966 * even diskless, then lost the connection.
967 * This should have been handled (panic? suspend?) somewhere
968 * else. But maybe it was not, so check again here.
969 * Caution: as long as we do not have a read/write lock on mdev,
970 * to serialize state changes, this is racy, since we may lose
971 * the connection *after* we test for the cstate.
972 */
973 if (mdev->state.disk < D_UP_TO_DATE && mdev->state.pdsk < D_UP_TO_DATE) {
974 if (__ratelimit(&drbd_ratelimit_state))
975 dev_err(DEV, "Sorry, I have no access to good data anymore.\n");
976 return 1;
977 }
978
979 return 0;
980}
981
982int drbd_make_request_26(struct request_queue *q, struct bio *bio)
983{
984 unsigned int s_enr, e_enr;
985 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
986
987 if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) {
988 bio_endio(bio, -EPERM);
989 return 0;
990 }
991
992 /* Reject barrier requests if we know the underlying device does
993 * not support them.
994 * XXX: Need to get this info from peer as well some how so we
995 * XXX: reject if EITHER side/data/metadata area does not support them.
996 *
997 * because of those XXX, this is not yet enabled,
998 * i.e. in drbd_init_set_defaults we set the NO_BARRIER_SUPP bit.
999 */
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001000 if (unlikely(bio->bi_rw & REQ_HARDBARRIER) && test_bit(NO_BARRIER_SUPP, &mdev->flags)) {
Philipp Reisnerb411b362009-09-25 16:07:19 -07001001 /* dev_warn(DEV, "Rejecting barrier request as underlying device does not support\n"); */
1002 bio_endio(bio, -EOPNOTSUPP);
1003 return 0;
1004 }
1005
1006 /*
1007 * what we "blindly" assume:
1008 */
1009 D_ASSERT(bio->bi_size > 0);
1010 D_ASSERT((bio->bi_size & 0x1ff) == 0);
1011 D_ASSERT(bio->bi_idx == 0);
1012
1013 /* to make some things easier, force alignment of requests within the
1014 * granularity of our hash tables */
1015 s_enr = bio->bi_sector >> HT_SHIFT;
1016 e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT;
1017
1018 if (likely(s_enr == e_enr)) {
1019 inc_ap_bio(mdev, 1);
1020 return drbd_make_request_common(mdev, bio);
1021 }
1022
1023 /* can this bio be split generically?
1024 * Maybe add our own split-arbitrary-bios function. */
1025 if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_SEGMENT_SIZE) {
1026 /* rather error out here than BUG in bio_split */
1027 dev_err(DEV, "bio would need to, but cannot, be split: "
1028 "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n",
1029 bio->bi_vcnt, bio->bi_idx, bio->bi_size,
1030 (unsigned long long)bio->bi_sector);
1031 bio_endio(bio, -EINVAL);
1032 } else {
1033 /* This bio crosses some boundary, so we have to split it. */
1034 struct bio_pair *bp;
1035 /* works for the "do not cross hash slot boundaries" case
1036 * e.g. sector 262269, size 4096
1037 * s_enr = 262269 >> 6 = 4097
1038 * e_enr = (262269+8-1) >> 6 = 4098
1039 * HT_SHIFT = 6
1040 * sps = 64, mask = 63
1041 * first_sectors = 64 - (262269 & 63) = 3
1042 */
1043 const sector_t sect = bio->bi_sector;
1044 const int sps = 1 << HT_SHIFT; /* sectors per slot */
1045 const int mask = sps - 1;
1046 const sector_t first_sectors = sps - (sect & mask);
1047 bp = bio_split(bio,
1048#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,28)
1049 bio_split_pool,
1050#endif
1051 first_sectors);
1052
1053 /* we need to get a "reference count" (ap_bio_cnt)
1054 * to avoid races with the disconnect/reconnect/suspend code.
Philipp Reisner9a25a042010-05-10 16:42:23 +02001055 * In case we need to split the bio here, we need to get three references
Philipp Reisnerb411b362009-09-25 16:07:19 -07001056 * atomically, otherwise we might deadlock when trying to submit the
1057 * second one! */
Philipp Reisner9a25a042010-05-10 16:42:23 +02001058 inc_ap_bio(mdev, 3);
Philipp Reisnerb411b362009-09-25 16:07:19 -07001059
1060 D_ASSERT(e_enr == s_enr + 1);
1061
Philipp Reisner9a25a042010-05-10 16:42:23 +02001062 while (drbd_make_request_common(mdev, &bp->bio1))
1063 inc_ap_bio(mdev, 1);
1064
1065 while (drbd_make_request_common(mdev, &bp->bio2))
1066 inc_ap_bio(mdev, 1);
1067
1068 dec_ap_bio(mdev);
1069
Philipp Reisnerb411b362009-09-25 16:07:19 -07001070 bio_pair_release(bp);
1071 }
1072 return 0;
1073}
1074
1075/* This is called by bio_add_page(). With this function we reduce
1076 * the number of BIOs that span over multiple DRBD_MAX_SEGMENT_SIZEs
1077 * units (was AL_EXTENTs).
1078 *
1079 * we do the calculation within the lower 32bit of the byte offsets,
1080 * since we don't care for actual offset, but only check whether it
1081 * would cross "activity log extent" boundaries.
1082 *
1083 * As long as the BIO is empty we have to allow at least one bvec,
1084 * regardless of size and offset. so the resulting bio may still
1085 * cross extent boundaries. those are dealt with (bio_split) in
1086 * drbd_make_request_26.
1087 */
1088int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1089{
1090 struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata;
1091 unsigned int bio_offset =
1092 (unsigned int)bvm->bi_sector << 9; /* 32 bit */
1093 unsigned int bio_size = bvm->bi_size;
1094 int limit, backing_limit;
1095
1096 limit = DRBD_MAX_SEGMENT_SIZE
1097 - ((bio_offset & (DRBD_MAX_SEGMENT_SIZE-1)) + bio_size);
1098 if (limit < 0)
1099 limit = 0;
1100 if (bio_size == 0) {
1101 if (limit <= bvec->bv_len)
1102 limit = bvec->bv_len;
1103 } else if (limit && get_ldev(mdev)) {
1104 struct request_queue * const b =
1105 mdev->ldev->backing_bdev->bd_disk->queue;
Lars Ellenberga1c88d02010-05-14 19:16:41 +02001106 if (b->merge_bvec_fn) {
Philipp Reisnerb411b362009-09-25 16:07:19 -07001107 backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1108 limit = min(limit, backing_limit);
1109 }
1110 put_ldev(mdev);
1111 }
1112 return limit;
1113}