]>
Commit | Line | Data |
---|---|---|
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 | ||
34 | /* Update disk stats at start of I/O request */ | |
35 | static 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)); | |
753c8913 | 42 | part_inc_in_flight(&mdev->vdisk->part0, rw); |
b411b363 | 43 | part_stat_unlock(); |
b411b363 PR |
44 | } |
45 | ||
46 | /* Update disk stats when completing request upwards */ | |
47 | static 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); | |
753c8913 | 55 | part_dec_in_flight(&mdev->vdisk->part0, rw); |
b411b363 | 56 | part_stat_unlock(); |
b411b363 PR |
57 | } |
58 | ||
59 | static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw) | |
60 | { | |
61 | const unsigned long s = req->rq_state; | |
288f422e PR |
62 | |
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 | ||
b411b363 PR |
71 | /* 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) { | |
b411b363 PR |
75 | /* Set out-of-sync unless both OK flags are set |
76 | * (local only or remote failed). | |
77 | * Other places where we set out-of-sync: | |
78 | * READ with local io-error */ | |
79 | if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) | |
80 | drbd_set_out_of_sync(mdev, req->sector, req->size); | |
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)) { | |
97 | drbd_al_complete_io(mdev, req->sector); | |
98 | put_ldev(mdev); | |
99 | } else if (__ratelimit(&drbd_ratelimit_state)) { | |
100 | dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu), " | |
101 | "but my Disk seems to have failed :(\n", | |
102 | (unsigned long long) req->sector); | |
103 | } | |
104 | } | |
105 | } | |
106 | ||
32fa7e91 | 107 | drbd_req_free(req); |
b411b363 PR |
108 | } |
109 | ||
110 | static void queue_barrier(struct drbd_conf *mdev) | |
111 | { | |
112 | struct drbd_tl_epoch *b; | |
113 | ||
114 | /* We are within the req_lock. Once we queued the barrier for sending, | |
115 | * we set the CREATE_BARRIER bit. It is cleared as soon as a new | |
116 | * barrier/epoch object is added. This is the only place this bit is | |
117 | * set. It indicates that the barrier for this epoch is already queued, | |
118 | * and no new epoch has been created yet. */ | |
119 | if (test_bit(CREATE_BARRIER, &mdev->flags)) | |
120 | return; | |
121 | ||
122 | b = mdev->newest_tle; | |
123 | b->w.cb = w_send_barrier; | |
124 | /* inc_ap_pending done here, so we won't | |
125 | * get imbalanced on connection loss. | |
126 | * dec_ap_pending will be done in got_BarrierAck | |
127 | * or (on connection loss) in tl_clear. */ | |
128 | inc_ap_pending(mdev); | |
129 | drbd_queue_work(&mdev->data.work, &b->w); | |
130 | set_bit(CREATE_BARRIER, &mdev->flags); | |
131 | } | |
132 | ||
133 | static void _about_to_complete_local_write(struct drbd_conf *mdev, | |
134 | struct drbd_request *req) | |
135 | { | |
136 | const unsigned long s = req->rq_state; | |
137 | struct drbd_request *i; | |
138 | struct drbd_epoch_entry *e; | |
139 | struct hlist_node *n; | |
140 | struct hlist_head *slot; | |
141 | ||
142 | /* before we can signal completion to the upper layers, | |
143 | * we may need to close the current epoch */ | |
144 | if (mdev->state.conn >= C_CONNECTED && | |
145 | req->epoch == mdev->newest_tle->br_number) | |
146 | queue_barrier(mdev); | |
147 | ||
148 | /* we need to do the conflict detection stuff, | |
149 | * if we have the ee_hash (two_primaries) and | |
150 | * this has been on the network */ | |
151 | if ((s & RQ_NET_DONE) && mdev->ee_hash != NULL) { | |
152 | const sector_t sector = req->sector; | |
153 | const int size = req->size; | |
154 | ||
155 | /* ASSERT: | |
156 | * there must be no conflicting requests, since | |
157 | * they must have been failed on the spot */ | |
158 | #define OVERLAPS overlaps(sector, size, i->sector, i->size) | |
159 | slot = tl_hash_slot(mdev, sector); | |
160 | hlist_for_each_entry(i, n, slot, colision) { | |
161 | if (OVERLAPS) { | |
162 | dev_alert(DEV, "LOGIC BUG: completed: %p %llus +%u; " | |
163 | "other: %p %llus +%u\n", | |
164 | req, (unsigned long long)sector, size, | |
165 | i, (unsigned long long)i->sector, i->size); | |
166 | } | |
167 | } | |
168 | ||
169 | /* maybe "wake" those conflicting epoch entries | |
170 | * that wait for this request to finish. | |
171 | * | |
172 | * currently, there can be only _one_ such ee | |
173 | * (well, or some more, which would be pending | |
174 | * P_DISCARD_ACK not yet sent by the asender...), | |
175 | * since we block the receiver thread upon the | |
176 | * first conflict detection, which will wait on | |
177 | * misc_wait. maybe we want to assert that? | |
178 | * | |
179 | * anyways, if we found one, | |
180 | * we just have to do a wake_up. */ | |
181 | #undef OVERLAPS | |
182 | #define OVERLAPS overlaps(sector, size, e->sector, e->size) | |
183 | slot = ee_hash_slot(mdev, req->sector); | |
184 | hlist_for_each_entry(e, n, slot, colision) { | |
185 | if (OVERLAPS) { | |
186 | wake_up(&mdev->misc_wait); | |
187 | break; | |
188 | } | |
189 | } | |
190 | } | |
191 | #undef OVERLAPS | |
192 | } | |
193 | ||
194 | void 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 | ||
201 | /* Helper for __req_mod(). | |
202 | * Set m->bio to the master bio, if it is fit to be completed, | |
203 | * or leave it alone (it is initialized to NULL in __req_mod), | |
204 | * if it has already been completed, or cannot be completed yet. | |
205 | * If m->bio is set, the error status to be returned is placed in m->error. | |
206 | */ | |
207 | void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m) | |
208 | { | |
209 | const unsigned long s = req->rq_state; | |
210 | struct drbd_conf *mdev = req->mdev; | |
211 | /* only WRITES may end up here without a master bio (on barrier ack) */ | |
212 | int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE; | |
213 | ||
b411b363 PR |
214 | /* we must not complete the master bio, while it is |
215 | * still being processed by _drbd_send_zc_bio (drbd_send_dblock) | |
216 | * not yet acknowledged by the peer | |
217 | * not yet completed by the local io subsystem | |
218 | * these flags may get cleared in any order by | |
219 | * the worker, | |
220 | * the receiver, | |
221 | * the bio_endio completion callbacks. | |
222 | */ | |
223 | if (s & RQ_NET_QUEUED) | |
224 | return; | |
225 | if (s & RQ_NET_PENDING) | |
226 | return; | |
227 | if (s & RQ_LOCAL_PENDING) | |
228 | return; | |
265be2d0 PR |
229 | if (mdev->state.susp) |
230 | return; | |
b411b363 PR |
231 | |
232 | if (req->master_bio) { | |
233 | /* this is data_received (remote read) | |
234 | * or protocol C P_WRITE_ACK | |
235 | * or protocol B P_RECV_ACK | |
236 | * or protocol A "handed_over_to_network" (SendAck) | |
237 | * or canceled or failed, | |
238 | * or killed from the transfer log due to connection loss. | |
239 | */ | |
240 | ||
241 | /* | |
242 | * figure out whether to report success or failure. | |
243 | * | |
244 | * report success when at least one of the operations succeeded. | |
245 | * or, to put the other way, | |
246 | * only report failure, when both operations failed. | |
247 | * | |
248 | * what to do about the failures is handled elsewhere. | |
249 | * what we need to do here is just: complete the master_bio. | |
250 | * | |
251 | * local completion error, if any, has been stored as ERR_PTR | |
252 | * in private_bio within drbd_endio_pri. | |
253 | */ | |
254 | int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); | |
255 | int error = PTR_ERR(req->private_bio); | |
256 | ||
257 | /* remove the request from the conflict detection | |
258 | * respective block_id verification hash */ | |
259 | if (!hlist_unhashed(&req->colision)) | |
260 | hlist_del(&req->colision); | |
261 | else | |
262 | D_ASSERT((s & RQ_NET_MASK) == 0); | |
263 | ||
264 | /* for writes we need to do some extra housekeeping */ | |
265 | if (rw == WRITE) | |
266 | _about_to_complete_local_write(mdev, req); | |
267 | ||
268 | /* Update disk stats */ | |
269 | _drbd_end_io_acct(mdev, req); | |
270 | ||
271 | m->error = ok ? 0 : (error ?: -EIO); | |
272 | m->bio = req->master_bio; | |
273 | req->master_bio = NULL; | |
274 | } | |
275 | ||
276 | if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) { | |
277 | /* this is disconnected (local only) operation, | |
278 | * or protocol C P_WRITE_ACK, | |
279 | * or protocol A or B P_BARRIER_ACK, | |
280 | * or killed from the transfer log due to connection loss. */ | |
281 | _req_is_done(mdev, req, rw); | |
282 | } | |
283 | /* else: network part and not DONE yet. that is | |
284 | * protocol A or B, barrier ack still pending... */ | |
285 | } | |
286 | ||
287 | /* | |
288 | * checks whether there was an overlapping request | |
289 | * or ee already registered. | |
290 | * | |
291 | * if so, return 1, in which case this request is completed on the spot, | |
292 | * without ever being submitted or send. | |
293 | * | |
294 | * return 0 if it is ok to submit this request. | |
295 | * | |
296 | * NOTE: | |
297 | * paranoia: assume something above us is broken, and issues different write | |
298 | * requests for the same block simultaneously... | |
299 | * | |
300 | * To ensure these won't be reordered differently on both nodes, resulting in | |
301 | * diverging data sets, we discard the later one(s). Not that this is supposed | |
302 | * to happen, but this is the rationale why we also have to check for | |
303 | * conflicting requests with local origin, and why we have to do so regardless | |
304 | * of whether we allowed multiple primaries. | |
305 | * | |
306 | * BTW, in case we only have one primary, the ee_hash is empty anyways, and the | |
307 | * second hlist_for_each_entry becomes a noop. This is even simpler than to | |
308 | * grab a reference on the net_conf, and check for the two_primaries flag... | |
309 | */ | |
310 | static int _req_conflicts(struct drbd_request *req) | |
311 | { | |
312 | struct drbd_conf *mdev = req->mdev; | |
313 | const sector_t sector = req->sector; | |
314 | const int size = req->size; | |
315 | struct drbd_request *i; | |
316 | struct drbd_epoch_entry *e; | |
317 | struct hlist_node *n; | |
318 | struct hlist_head *slot; | |
319 | ||
320 | D_ASSERT(hlist_unhashed(&req->colision)); | |
321 | ||
322 | if (!get_net_conf(mdev)) | |
323 | return 0; | |
324 | ||
325 | /* BUG_ON */ | |
326 | ERR_IF (mdev->tl_hash_s == 0) | |
327 | goto out_no_conflict; | |
328 | BUG_ON(mdev->tl_hash == NULL); | |
329 | ||
330 | #define OVERLAPS overlaps(i->sector, i->size, sector, size) | |
331 | slot = tl_hash_slot(mdev, sector); | |
332 | hlist_for_each_entry(i, n, slot, colision) { | |
333 | if (OVERLAPS) { | |
334 | dev_alert(DEV, "%s[%u] Concurrent local write detected! " | |
335 | "[DISCARD L] new: %llus +%u; " | |
336 | "pending: %llus +%u\n", | |
337 | current->comm, current->pid, | |
338 | (unsigned long long)sector, size, | |
339 | (unsigned long long)i->sector, i->size); | |
340 | goto out_conflict; | |
341 | } | |
342 | } | |
343 | ||
344 | if (mdev->ee_hash_s) { | |
345 | /* now, check for overlapping requests with remote origin */ | |
346 | BUG_ON(mdev->ee_hash == NULL); | |
347 | #undef OVERLAPS | |
348 | #define OVERLAPS overlaps(e->sector, e->size, sector, size) | |
349 | slot = ee_hash_slot(mdev, sector); | |
350 | hlist_for_each_entry(e, n, slot, colision) { | |
351 | if (OVERLAPS) { | |
352 | dev_alert(DEV, "%s[%u] Concurrent remote write detected!" | |
353 | " [DISCARD L] new: %llus +%u; " | |
354 | "pending: %llus +%u\n", | |
355 | current->comm, current->pid, | |
356 | (unsigned long long)sector, size, | |
357 | (unsigned long long)e->sector, e->size); | |
358 | goto out_conflict; | |
359 | } | |
360 | } | |
361 | } | |
362 | #undef OVERLAPS | |
363 | ||
364 | out_no_conflict: | |
365 | /* this is like it should be, and what we expected. | |
366 | * our users do behave after all... */ | |
367 | put_net_conf(mdev); | |
368 | return 0; | |
369 | ||
370 | out_conflict: | |
371 | put_net_conf(mdev); | |
372 | return 1; | |
373 | } | |
374 | ||
375 | /* obviously this could be coded as many single functions | |
376 | * instead of one huge switch, | |
377 | * or by putting the code directly in the respective locations | |
378 | * (as it has been before). | |
379 | * | |
380 | * but having it this way | |
381 | * enforces that it is all in this one place, where it is easier to audit, | |
382 | * it makes it obvious that whatever "event" "happens" to a request should | |
383 | * happen "atomically" within the req_lock, | |
384 | * and it enforces that we have to think in a very structured manner | |
385 | * about the "events" that may happen to a request during its life time ... | |
386 | */ | |
2a80699f | 387 | int __req_mod(struct drbd_request *req, enum drbd_req_event what, |
b411b363 PR |
388 | struct bio_and_error *m) |
389 | { | |
390 | struct drbd_conf *mdev = req->mdev; | |
2a80699f | 391 | int rv = 0; |
b411b363 PR |
392 | m->bio = NULL; |
393 | ||
b411b363 PR |
394 | switch (what) { |
395 | default: | |
396 | dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); | |
397 | break; | |
398 | ||
399 | /* does not happen... | |
400 | * initialization done in drbd_req_new | |
401 | case created: | |
402 | break; | |
403 | */ | |
404 | ||
405 | case to_be_send: /* via network */ | |
406 | /* reached via drbd_make_request_common | |
407 | * and from w_read_retry_remote */ | |
408 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); | |
409 | req->rq_state |= RQ_NET_PENDING; | |
410 | inc_ap_pending(mdev); | |
411 | break; | |
412 | ||
413 | case to_be_submitted: /* locally */ | |
414 | /* reached via drbd_make_request_common */ | |
415 | D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK)); | |
416 | req->rq_state |= RQ_LOCAL_PENDING; | |
417 | break; | |
418 | ||
419 | case completed_ok: | |
420 | if (bio_data_dir(req->master_bio) == WRITE) | |
421 | mdev->writ_cnt += req->size>>9; | |
422 | else | |
423 | mdev->read_cnt += req->size>>9; | |
424 | ||
425 | req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); | |
426 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
427 | ||
428 | _req_may_be_done(req, m); | |
429 | put_ldev(mdev); | |
430 | break; | |
431 | ||
432 | case write_completed_with_error: | |
433 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
434 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
435 | ||
b411b363 PR |
436 | __drbd_chk_io_error(mdev, FALSE); |
437 | _req_may_be_done(req, m); | |
438 | put_ldev(mdev); | |
439 | break; | |
440 | ||
441 | case read_ahead_completed_with_error: | |
442 | /* it is legal to fail READA */ | |
443 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
444 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
445 | _req_may_be_done(req, m); | |
446 | put_ldev(mdev); | |
447 | break; | |
448 | ||
449 | case read_completed_with_error: | |
450 | drbd_set_out_of_sync(mdev, req->sector, req->size); | |
451 | ||
452 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
453 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
454 | ||
b411b363 | 455 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); |
b411b363 PR |
456 | |
457 | __drbd_chk_io_error(mdev, FALSE); | |
458 | put_ldev(mdev); | |
b411b363 | 459 | |
d255e5ff LE |
460 | /* no point in retrying if there is no good remote data, |
461 | * or we have no connection. */ | |
462 | if (mdev->state.pdsk != D_UP_TO_DATE) { | |
463 | _req_may_be_done(req, m); | |
464 | break; | |
465 | } | |
466 | ||
467 | /* _req_mod(req,to_be_send); oops, recursion... */ | |
468 | req->rq_state |= RQ_NET_PENDING; | |
469 | inc_ap_pending(mdev); | |
b411b363 PR |
470 | /* fall through: _req_mod(req,queue_for_net_read); */ |
471 | ||
472 | case queue_for_net_read: | |
473 | /* READ or READA, and | |
474 | * no local disk, | |
475 | * or target area marked as invalid, | |
476 | * or just got an io-error. */ | |
477 | /* from drbd_make_request_common | |
478 | * or from bio_endio during read io-error recovery */ | |
479 | ||
480 | /* so we can verify the handle in the answer packet | |
481 | * corresponding hlist_del is in _req_may_be_done() */ | |
482 | hlist_add_head(&req->colision, ar_hash_slot(mdev, req->sector)); | |
483 | ||
83c38830 | 484 | set_bit(UNPLUG_REMOTE, &mdev->flags); |
b411b363 PR |
485 | |
486 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
487 | req->rq_state |= RQ_NET_QUEUED; | |
488 | req->w.cb = (req->rq_state & RQ_LOCAL_MASK) | |
489 | ? w_read_retry_remote | |
490 | : w_send_read_req; | |
491 | drbd_queue_work(&mdev->data.work, &req->w); | |
492 | break; | |
493 | ||
494 | case queue_for_net_write: | |
495 | /* assert something? */ | |
496 | /* from drbd_make_request_common only */ | |
497 | ||
498 | hlist_add_head(&req->colision, tl_hash_slot(mdev, req->sector)); | |
499 | /* corresponding hlist_del is in _req_may_be_done() */ | |
500 | ||
501 | /* NOTE | |
502 | * In case the req ended up on the transfer log before being | |
503 | * queued on the worker, it could lead to this request being | |
504 | * missed during cleanup after connection loss. | |
505 | * So we have to do both operations here, | |
506 | * within the same lock that protects the transfer log. | |
507 | * | |
508 | * _req_add_to_epoch(req); this has to be after the | |
509 | * _maybe_start_new_epoch(req); which happened in | |
510 | * drbd_make_request_common, because we now may set the bit | |
511 | * again ourselves to close the current epoch. | |
512 | * | |
513 | * Add req to the (now) current epoch (barrier). */ | |
514 | ||
83c38830 LE |
515 | /* otherwise we may lose an unplug, which may cause some remote |
516 | * io-scheduler timeout to expire, increasing maximum latency, | |
517 | * hurting performance. */ | |
518 | set_bit(UNPLUG_REMOTE, &mdev->flags); | |
519 | ||
b411b363 PR |
520 | /* see drbd_make_request_common, |
521 | * just after it grabs the req_lock */ | |
522 | D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0); | |
523 | ||
524 | req->epoch = mdev->newest_tle->br_number; | |
b411b363 PR |
525 | |
526 | /* increment size of current epoch */ | |
7e602c0a | 527 | mdev->newest_tle->n_writes++; |
b411b363 PR |
528 | |
529 | /* queue work item to send data */ | |
530 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
531 | req->rq_state |= RQ_NET_QUEUED; | |
532 | req->w.cb = w_send_dblock; | |
533 | drbd_queue_work(&mdev->data.work, &req->w); | |
534 | ||
535 | /* close the epoch, in case it outgrew the limit */ | |
7e602c0a | 536 | if (mdev->newest_tle->n_writes >= mdev->net_conf->max_epoch_size) |
b411b363 PR |
537 | queue_barrier(mdev); |
538 | ||
539 | break; | |
540 | ||
541 | case send_canceled: | |
542 | /* treat it the same */ | |
543 | case send_failed: | |
544 | /* real cleanup will be done from tl_clear. just update flags | |
545 | * so it is no longer marked as on the worker queue */ | |
546 | req->rq_state &= ~RQ_NET_QUEUED; | |
547 | /* if we did it right, tl_clear should be scheduled only after | |
548 | * this, so this should not be necessary! */ | |
549 | _req_may_be_done(req, m); | |
550 | break; | |
551 | ||
552 | case handed_over_to_network: | |
553 | /* assert something? */ | |
554 | if (bio_data_dir(req->master_bio) == WRITE && | |
555 | mdev->net_conf->wire_protocol == DRBD_PROT_A) { | |
556 | /* this is what is dangerous about protocol A: | |
557 | * pretend it was successfully written on the peer. */ | |
558 | if (req->rq_state & RQ_NET_PENDING) { | |
559 | dec_ap_pending(mdev); | |
560 | req->rq_state &= ~RQ_NET_PENDING; | |
561 | req->rq_state |= RQ_NET_OK; | |
562 | } /* else: neg-ack was faster... */ | |
563 | /* it is still not yet RQ_NET_DONE until the | |
564 | * corresponding epoch barrier got acked as well, | |
565 | * so we know what to dirty on connection loss */ | |
566 | } | |
567 | req->rq_state &= ~RQ_NET_QUEUED; | |
568 | req->rq_state |= RQ_NET_SENT; | |
569 | /* because _drbd_send_zc_bio could sleep, and may want to | |
570 | * dereference the bio even after the "write_acked_by_peer" and | |
571 | * "completed_ok" events came in, once we return from | |
572 | * _drbd_send_zc_bio (drbd_send_dblock), we have to check | |
573 | * whether it is done already, and end it. */ | |
574 | _req_may_be_done(req, m); | |
575 | break; | |
576 | ||
d255e5ff LE |
577 | case read_retry_remote_canceled: |
578 | req->rq_state &= ~RQ_NET_QUEUED; | |
579 | /* fall through, in case we raced with drbd_disconnect */ | |
b411b363 PR |
580 | case connection_lost_while_pending: |
581 | /* transfer log cleanup after connection loss */ | |
582 | /* assert something? */ | |
583 | if (req->rq_state & RQ_NET_PENDING) | |
584 | dec_ap_pending(mdev); | |
585 | req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING); | |
586 | req->rq_state |= RQ_NET_DONE; | |
587 | /* if it is still queued, we may not complete it here. | |
588 | * it will be canceled soon. */ | |
589 | if (!(req->rq_state & RQ_NET_QUEUED)) | |
590 | _req_may_be_done(req, m); | |
591 | break; | |
592 | ||
593 | case write_acked_by_peer_and_sis: | |
594 | req->rq_state |= RQ_NET_SIS; | |
595 | case conflict_discarded_by_peer: | |
596 | /* for discarded conflicting writes of multiple primaries, | |
597 | * there is no need to keep anything in the tl, potential | |
598 | * node crashes are covered by the activity log. */ | |
599 | if (what == conflict_discarded_by_peer) | |
600 | dev_alert(DEV, "Got DiscardAck packet %llus +%u!" | |
601 | " DRBD is not a random data generator!\n", | |
602 | (unsigned long long)req->sector, req->size); | |
603 | req->rq_state |= RQ_NET_DONE; | |
604 | /* fall through */ | |
605 | case write_acked_by_peer: | |
606 | /* protocol C; successfully written on peer. | |
607 | * Nothing to do here. | |
608 | * We want to keep the tl in place for all protocols, to cater | |
609 | * for volatile write-back caches on lower level devices. | |
610 | * | |
611 | * A barrier request is expected to have forced all prior | |
612 | * requests onto stable storage, so completion of a barrier | |
613 | * request could set NET_DONE right here, and not wait for the | |
614 | * P_BARRIER_ACK, but that is an unnecessary optimization. */ | |
615 | ||
616 | /* this makes it effectively the same as for: */ | |
617 | case recv_acked_by_peer: | |
618 | /* protocol B; pretends to be successfully written on peer. | |
619 | * see also notes above in handed_over_to_network about | |
620 | * protocol != C */ | |
621 | req->rq_state |= RQ_NET_OK; | |
622 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
623 | dec_ap_pending(mdev); | |
624 | req->rq_state &= ~RQ_NET_PENDING; | |
625 | _req_may_be_done(req, m); | |
626 | break; | |
627 | ||
628 | case neg_acked: | |
629 | /* assert something? */ | |
630 | if (req->rq_state & RQ_NET_PENDING) | |
631 | dec_ap_pending(mdev); | |
632 | req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING); | |
633 | ||
634 | req->rq_state |= RQ_NET_DONE; | |
635 | _req_may_be_done(req, m); | |
636 | /* else: done by handed_over_to_network */ | |
637 | break; | |
638 | ||
265be2d0 PR |
639 | case fail_frozen_disk_io: |
640 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) | |
641 | break; | |
642 | ||
643 | _req_may_be_done(req, m); | |
644 | break; | |
645 | ||
646 | case restart_frozen_disk_io: | |
647 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) | |
648 | break; | |
649 | ||
650 | req->rq_state &= ~RQ_LOCAL_COMPLETED; | |
651 | ||
652 | rv = MR_READ; | |
653 | if (bio_data_dir(req->master_bio) == WRITE) | |
654 | rv = MR_WRITE; | |
655 | ||
656 | get_ldev(mdev); | |
657 | req->w.cb = w_restart_disk_io; | |
658 | drbd_queue_work(&mdev->data.work, &req->w); | |
659 | break; | |
660 | ||
11b58e73 PR |
661 | case resend: |
662 | /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK | |
47ff2d0a | 663 | before the connection loss (B&C only); only P_BARRIER_ACK was missing. |
11b58e73 | 664 | Trowing them out of the TL here by pretending we got a BARRIER_ACK |
481c6f50 | 665 | We ensure that the peer was not rebooted */ |
11b58e73 PR |
666 | if (!(req->rq_state & RQ_NET_OK)) { |
667 | if (req->w.cb) { | |
668 | drbd_queue_work(&mdev->data.work, &req->w); | |
669 | rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; | |
670 | } | |
671 | break; | |
672 | } | |
673 | /* else, fall through to barrier_acked */ | |
674 | ||
b411b363 | 675 | case barrier_acked: |
288f422e PR |
676 | if (!(req->rq_state & RQ_WRITE)) |
677 | break; | |
678 | ||
b411b363 PR |
679 | if (req->rq_state & RQ_NET_PENDING) { |
680 | /* barrier came in before all requests have been acked. | |
681 | * this is bad, because if the connection is lost now, | |
682 | * we won't be able to clean them up... */ | |
683 | dev_err(DEV, "FIXME (barrier_acked but pending)\n"); | |
b411b363 PR |
684 | list_move(&req->tl_requests, &mdev->out_of_sequence_requests); |
685 | } | |
686 | D_ASSERT(req->rq_state & RQ_NET_SENT); | |
687 | req->rq_state |= RQ_NET_DONE; | |
688 | _req_may_be_done(req, m); | |
689 | break; | |
690 | ||
691 | case data_received: | |
692 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
693 | dec_ap_pending(mdev); | |
694 | req->rq_state &= ~RQ_NET_PENDING; | |
695 | req->rq_state |= (RQ_NET_OK|RQ_NET_DONE); | |
696 | _req_may_be_done(req, m); | |
697 | break; | |
698 | }; | |
2a80699f PR |
699 | |
700 | return rv; | |
b411b363 PR |
701 | } |
702 | ||
703 | /* we may do a local read if: | |
704 | * - we are consistent (of course), | |
705 | * - or we are generally inconsistent, | |
706 | * BUT we are still/already IN SYNC for this area. | |
707 | * since size may be bigger than BM_BLOCK_SIZE, | |
708 | * we may need to check several bits. | |
709 | */ | |
710 | static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size) | |
711 | { | |
712 | unsigned long sbnr, ebnr; | |
713 | sector_t esector, nr_sectors; | |
714 | ||
715 | if (mdev->state.disk == D_UP_TO_DATE) | |
716 | return 1; | |
717 | if (mdev->state.disk >= D_OUTDATED) | |
718 | return 0; | |
719 | if (mdev->state.disk < D_INCONSISTENT) | |
720 | return 0; | |
721 | /* state.disk == D_INCONSISTENT We will have a look at the BitMap */ | |
722 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
723 | esector = sector + (size >> 9) - 1; | |
724 | ||
725 | D_ASSERT(sector < nr_sectors); | |
726 | D_ASSERT(esector < nr_sectors); | |
727 | ||
728 | sbnr = BM_SECT_TO_BIT(sector); | |
729 | ebnr = BM_SECT_TO_BIT(esector); | |
730 | ||
731 | return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr); | |
732 | } | |
733 | ||
734 | static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio) | |
735 | { | |
736 | const int rw = bio_rw(bio); | |
737 | const int size = bio->bi_size; | |
738 | const sector_t sector = bio->bi_sector; | |
739 | struct drbd_tl_epoch *b = NULL; | |
740 | struct drbd_request *req; | |
741 | int local, remote; | |
742 | int err = -EIO; | |
9a25a04c | 743 | int ret = 0; |
b411b363 PR |
744 | |
745 | /* allocate outside of all locks; */ | |
746 | req = drbd_req_new(mdev, bio); | |
747 | if (!req) { | |
748 | dec_ap_bio(mdev); | |
749 | /* only pass the error to the upper layers. | |
750 | * if user cannot handle io errors, that's not our business. */ | |
751 | dev_err(DEV, "could not kmalloc() req\n"); | |
752 | bio_endio(bio, -ENOMEM); | |
753 | return 0; | |
754 | } | |
755 | ||
b411b363 PR |
756 | local = get_ldev(mdev); |
757 | if (!local) { | |
758 | bio_put(req->private_bio); /* or we get a bio leak */ | |
759 | req->private_bio = NULL; | |
760 | } | |
761 | if (rw == WRITE) { | |
762 | remote = 1; | |
763 | } else { | |
764 | /* READ || READA */ | |
765 | if (local) { | |
766 | if (!drbd_may_do_local_read(mdev, sector, size)) { | |
767 | /* we could kick the syncer to | |
768 | * sync this extent asap, wait for | |
769 | * it, then continue locally. | |
770 | * Or just issue the request remotely. | |
771 | */ | |
772 | local = 0; | |
773 | bio_put(req->private_bio); | |
774 | req->private_bio = NULL; | |
775 | put_ldev(mdev); | |
776 | } | |
777 | } | |
778 | remote = !local && mdev->state.pdsk >= D_UP_TO_DATE; | |
779 | } | |
780 | ||
781 | /* If we have a disk, but a READA request is mapped to remote, | |
782 | * we are R_PRIMARY, D_INCONSISTENT, SyncTarget. | |
783 | * Just fail that READA request right here. | |
784 | * | |
785 | * THINK: maybe fail all READA when not local? | |
786 | * or make this configurable... | |
787 | * if network is slow, READA won't do any good. | |
788 | */ | |
789 | if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) { | |
790 | err = -EWOULDBLOCK; | |
791 | goto fail_and_free_req; | |
792 | } | |
793 | ||
794 | /* For WRITES going to the local disk, grab a reference on the target | |
795 | * extent. This waits for any resync activity in the corresponding | |
796 | * resync extent to finish, and, if necessary, pulls in the target | |
797 | * extent into the activity log, which involves further disk io because | |
798 | * of transactional on-disk meta data updates. */ | |
799 | if (rw == WRITE && local) | |
800 | drbd_al_begin_io(mdev, sector); | |
801 | ||
802 | remote = remote && (mdev->state.pdsk == D_UP_TO_DATE || | |
803 | (mdev->state.pdsk == D_INCONSISTENT && | |
804 | mdev->state.conn >= C_CONNECTED)); | |
805 | ||
9a25a04c | 806 | if (!(local || remote) && !mdev->state.susp) { |
b411b363 PR |
807 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); |
808 | goto fail_free_complete; | |
809 | } | |
810 | ||
811 | /* For WRITE request, we have to make sure that we have an | |
812 | * unused_spare_tle, in case we need to start a new epoch. | |
813 | * I try to be smart and avoid to pre-allocate always "just in case", | |
814 | * but there is a race between testing the bit and pointer outside the | |
815 | * spinlock, and grabbing the spinlock. | |
816 | * if we lost that race, we retry. */ | |
817 | if (rw == WRITE && remote && | |
818 | mdev->unused_spare_tle == NULL && | |
819 | test_bit(CREATE_BARRIER, &mdev->flags)) { | |
820 | allocate_barrier: | |
821 | b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO); | |
822 | if (!b) { | |
823 | dev_err(DEV, "Failed to alloc barrier.\n"); | |
824 | err = -ENOMEM; | |
825 | goto fail_free_complete; | |
826 | } | |
827 | } | |
828 | ||
829 | /* GOOD, everything prepared, grab the spin_lock */ | |
830 | spin_lock_irq(&mdev->req_lock); | |
831 | ||
9a25a04c PR |
832 | if (mdev->state.susp) { |
833 | /* If we got suspended, use the retry mechanism of | |
834 | generic_make_request() to restart processing of this | |
835 | bio. In the next call to drbd_make_request_26 | |
836 | we sleep in inc_ap_bio() */ | |
837 | ret = 1; | |
838 | spin_unlock_irq(&mdev->req_lock); | |
839 | goto fail_free_complete; | |
840 | } | |
841 | ||
b411b363 PR |
842 | if (remote) { |
843 | remote = (mdev->state.pdsk == D_UP_TO_DATE || | |
844 | (mdev->state.pdsk == D_INCONSISTENT && | |
845 | mdev->state.conn >= C_CONNECTED)); | |
846 | if (!remote) | |
847 | dev_warn(DEV, "lost connection while grabbing the req_lock!\n"); | |
848 | if (!(local || remote)) { | |
849 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); | |
850 | spin_unlock_irq(&mdev->req_lock); | |
851 | goto fail_free_complete; | |
852 | } | |
853 | } | |
854 | ||
855 | if (b && mdev->unused_spare_tle == NULL) { | |
856 | mdev->unused_spare_tle = b; | |
857 | b = NULL; | |
858 | } | |
859 | if (rw == WRITE && remote && | |
860 | mdev->unused_spare_tle == NULL && | |
861 | test_bit(CREATE_BARRIER, &mdev->flags)) { | |
862 | /* someone closed the current epoch | |
863 | * while we were grabbing the spinlock */ | |
864 | spin_unlock_irq(&mdev->req_lock); | |
865 | goto allocate_barrier; | |
866 | } | |
867 | ||
868 | ||
869 | /* Update disk stats */ | |
870 | _drbd_start_io_acct(mdev, req, bio); | |
871 | ||
872 | /* _maybe_start_new_epoch(mdev); | |
873 | * If we need to generate a write barrier packet, we have to add the | |
874 | * new epoch (barrier) object, and queue the barrier packet for sending, | |
875 | * and queue the req's data after it _within the same lock_, otherwise | |
876 | * we have race conditions were the reorder domains could be mixed up. | |
877 | * | |
878 | * Even read requests may start a new epoch and queue the corresponding | |
879 | * barrier packet. To get the write ordering right, we only have to | |
880 | * make sure that, if this is a write request and it triggered a | |
881 | * barrier packet, this request is queued within the same spinlock. */ | |
882 | if (remote && mdev->unused_spare_tle && | |
883 | test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) { | |
884 | _tl_add_barrier(mdev, mdev->unused_spare_tle); | |
885 | mdev->unused_spare_tle = NULL; | |
886 | } else { | |
887 | D_ASSERT(!(remote && rw == WRITE && | |
888 | test_bit(CREATE_BARRIER, &mdev->flags))); | |
889 | } | |
890 | ||
891 | /* NOTE | |
892 | * Actually, 'local' may be wrong here already, since we may have failed | |
893 | * to write to the meta data, and may become wrong anytime because of | |
894 | * local io-error for some other request, which would lead to us | |
895 | * "detaching" the local disk. | |
896 | * | |
897 | * 'remote' may become wrong any time because the network could fail. | |
898 | * | |
899 | * This is a harmless race condition, though, since it is handled | |
900 | * correctly at the appropriate places; so it just defers the failure | |
901 | * of the respective operation. | |
902 | */ | |
903 | ||
904 | /* mark them early for readability. | |
905 | * this just sets some state flags. */ | |
906 | if (remote) | |
907 | _req_mod(req, to_be_send); | |
908 | if (local) | |
909 | _req_mod(req, to_be_submitted); | |
910 | ||
911 | /* check this request on the collision detection hash tables. | |
912 | * if we have a conflict, just complete it here. | |
913 | * THINK do we want to check reads, too? (I don't think so...) */ | |
914 | if (rw == WRITE && _req_conflicts(req)) { | |
915 | /* this is a conflicting request. | |
916 | * even though it may have been only _partially_ | |
917 | * overlapping with one of the currently pending requests, | |
918 | * without even submitting or sending it, we will | |
919 | * pretend that it was successfully served right now. | |
920 | */ | |
921 | if (local) { | |
922 | bio_put(req->private_bio); | |
923 | req->private_bio = NULL; | |
924 | drbd_al_complete_io(mdev, req->sector); | |
925 | put_ldev(mdev); | |
926 | local = 0; | |
927 | } | |
928 | if (remote) | |
929 | dec_ap_pending(mdev); | |
930 | _drbd_end_io_acct(mdev, req); | |
931 | /* THINK: do we want to fail it (-EIO), or pretend success? */ | |
932 | bio_endio(req->master_bio, 0); | |
933 | req->master_bio = NULL; | |
934 | dec_ap_bio(mdev); | |
935 | drbd_req_free(req); | |
936 | remote = 0; | |
937 | } | |
938 | ||
288f422e PR |
939 | |
940 | list_add_tail(&req->tl_requests, &mdev->newest_tle->requests); | |
941 | ||
b411b363 PR |
942 | /* NOTE remote first: to get the concurrent write detection right, |
943 | * we must register the request before start of local IO. */ | |
944 | if (remote) { | |
945 | /* either WRITE and C_CONNECTED, | |
946 | * or READ, and no local disk, | |
947 | * or READ, but not in sync. | |
948 | */ | |
949 | _req_mod(req, (rw == WRITE) | |
950 | ? queue_for_net_write | |
951 | : queue_for_net_read); | |
952 | } | |
953 | spin_unlock_irq(&mdev->req_lock); | |
954 | kfree(b); /* if someone else has beaten us to it... */ | |
955 | ||
956 | if (local) { | |
957 | req->private_bio->bi_bdev = mdev->ldev->backing_bdev; | |
958 | ||
b411b363 PR |
959 | if (FAULT_ACTIVE(mdev, rw == WRITE ? DRBD_FAULT_DT_WR |
960 | : rw == READ ? DRBD_FAULT_DT_RD | |
961 | : DRBD_FAULT_DT_RA)) | |
962 | bio_endio(req->private_bio, -EIO); | |
963 | else | |
964 | generic_make_request(req->private_bio); | |
965 | } | |
966 | ||
967 | /* we need to plug ALWAYS since we possibly need to kick lo_dev. | |
968 | * we plug after submit, so we won't miss an unplug event */ | |
969 | drbd_plug_device(mdev); | |
970 | ||
971 | return 0; | |
972 | ||
973 | fail_free_complete: | |
974 | if (rw == WRITE && local) | |
975 | drbd_al_complete_io(mdev, sector); | |
976 | fail_and_free_req: | |
977 | if (local) { | |
978 | bio_put(req->private_bio); | |
979 | req->private_bio = NULL; | |
980 | put_ldev(mdev); | |
981 | } | |
9a25a04c PR |
982 | if (!ret) |
983 | bio_endio(bio, err); | |
984 | ||
b411b363 PR |
985 | drbd_req_free(req); |
986 | dec_ap_bio(mdev); | |
987 | kfree(b); | |
988 | ||
9a25a04c | 989 | return ret; |
b411b363 PR |
990 | } |
991 | ||
992 | /* helper function for drbd_make_request | |
993 | * if we can determine just by the mdev (state) that this request will fail, | |
994 | * return 1 | |
995 | * otherwise return 0 | |
996 | */ | |
997 | static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write) | |
998 | { | |
b411b363 PR |
999 | if (mdev->state.role != R_PRIMARY && |
1000 | (!allow_oos || is_write)) { | |
1001 | if (__ratelimit(&drbd_ratelimit_state)) { | |
1002 | dev_err(DEV, "Process %s[%u] tried to %s; " | |
1003 | "since we are not in Primary state, " | |
1004 | "we cannot allow this\n", | |
1005 | current->comm, current->pid, | |
1006 | is_write ? "WRITE" : "READ"); | |
1007 | } | |
1008 | return 1; | |
1009 | } | |
1010 | ||
b411b363 PR |
1011 | return 0; |
1012 | } | |
1013 | ||
1014 | int drbd_make_request_26(struct request_queue *q, struct bio *bio) | |
1015 | { | |
1016 | unsigned int s_enr, e_enr; | |
1017 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; | |
1018 | ||
1019 | if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) { | |
1020 | bio_endio(bio, -EPERM); | |
1021 | return 0; | |
1022 | } | |
1023 | ||
1024 | /* Reject barrier requests if we know the underlying device does | |
1025 | * not support them. | |
1026 | * XXX: Need to get this info from peer as well some how so we | |
1027 | * XXX: reject if EITHER side/data/metadata area does not support them. | |
1028 | * | |
1029 | * because of those XXX, this is not yet enabled, | |
1030 | * i.e. in drbd_init_set_defaults we set the NO_BARRIER_SUPP bit. | |
1031 | */ | |
7b6d91da | 1032 | if (unlikely(bio->bi_rw & REQ_HARDBARRIER) && test_bit(NO_BARRIER_SUPP, &mdev->flags)) { |
b411b363 PR |
1033 | /* dev_warn(DEV, "Rejecting barrier request as underlying device does not support\n"); */ |
1034 | bio_endio(bio, -EOPNOTSUPP); | |
1035 | return 0; | |
1036 | } | |
1037 | ||
1038 | /* | |
1039 | * what we "blindly" assume: | |
1040 | */ | |
1041 | D_ASSERT(bio->bi_size > 0); | |
1042 | D_ASSERT((bio->bi_size & 0x1ff) == 0); | |
1043 | D_ASSERT(bio->bi_idx == 0); | |
1044 | ||
1045 | /* to make some things easier, force alignment of requests within the | |
1046 | * granularity of our hash tables */ | |
1047 | s_enr = bio->bi_sector >> HT_SHIFT; | |
1048 | e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT; | |
1049 | ||
1050 | if (likely(s_enr == e_enr)) { | |
1051 | inc_ap_bio(mdev, 1); | |
1052 | return drbd_make_request_common(mdev, bio); | |
1053 | } | |
1054 | ||
1055 | /* can this bio be split generically? | |
1056 | * Maybe add our own split-arbitrary-bios function. */ | |
1057 | if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_SEGMENT_SIZE) { | |
1058 | /* rather error out here than BUG in bio_split */ | |
1059 | dev_err(DEV, "bio would need to, but cannot, be split: " | |
1060 | "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n", | |
1061 | bio->bi_vcnt, bio->bi_idx, bio->bi_size, | |
1062 | (unsigned long long)bio->bi_sector); | |
1063 | bio_endio(bio, -EINVAL); | |
1064 | } else { | |
1065 | /* This bio crosses some boundary, so we have to split it. */ | |
1066 | struct bio_pair *bp; | |
1067 | /* works for the "do not cross hash slot boundaries" case | |
1068 | * e.g. sector 262269, size 4096 | |
1069 | * s_enr = 262269 >> 6 = 4097 | |
1070 | * e_enr = (262269+8-1) >> 6 = 4098 | |
1071 | * HT_SHIFT = 6 | |
1072 | * sps = 64, mask = 63 | |
1073 | * first_sectors = 64 - (262269 & 63) = 3 | |
1074 | */ | |
1075 | const sector_t sect = bio->bi_sector; | |
1076 | const int sps = 1 << HT_SHIFT; /* sectors per slot */ | |
1077 | const int mask = sps - 1; | |
1078 | const sector_t first_sectors = sps - (sect & mask); | |
1079 | bp = bio_split(bio, | |
1080 | #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,28) | |
1081 | bio_split_pool, | |
1082 | #endif | |
1083 | first_sectors); | |
1084 | ||
1085 | /* we need to get a "reference count" (ap_bio_cnt) | |
1086 | * to avoid races with the disconnect/reconnect/suspend code. | |
9a25a04c | 1087 | * In case we need to split the bio here, we need to get three references |
b411b363 PR |
1088 | * atomically, otherwise we might deadlock when trying to submit the |
1089 | * second one! */ | |
9a25a04c | 1090 | inc_ap_bio(mdev, 3); |
b411b363 PR |
1091 | |
1092 | D_ASSERT(e_enr == s_enr + 1); | |
1093 | ||
9a25a04c PR |
1094 | while (drbd_make_request_common(mdev, &bp->bio1)) |
1095 | inc_ap_bio(mdev, 1); | |
1096 | ||
1097 | while (drbd_make_request_common(mdev, &bp->bio2)) | |
1098 | inc_ap_bio(mdev, 1); | |
1099 | ||
1100 | dec_ap_bio(mdev); | |
1101 | ||
b411b363 PR |
1102 | bio_pair_release(bp); |
1103 | } | |
1104 | return 0; | |
1105 | } | |
1106 | ||
1107 | /* This is called by bio_add_page(). With this function we reduce | |
1108 | * the number of BIOs that span over multiple DRBD_MAX_SEGMENT_SIZEs | |
1109 | * units (was AL_EXTENTs). | |
1110 | * | |
1111 | * we do the calculation within the lower 32bit of the byte offsets, | |
1112 | * since we don't care for actual offset, but only check whether it | |
1113 | * would cross "activity log extent" boundaries. | |
1114 | * | |
1115 | * As long as the BIO is empty we have to allow at least one bvec, | |
1116 | * regardless of size and offset. so the resulting bio may still | |
1117 | * cross extent boundaries. those are dealt with (bio_split) in | |
1118 | * drbd_make_request_26. | |
1119 | */ | |
1120 | int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec) | |
1121 | { | |
1122 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; | |
1123 | unsigned int bio_offset = | |
1124 | (unsigned int)bvm->bi_sector << 9; /* 32 bit */ | |
1125 | unsigned int bio_size = bvm->bi_size; | |
1126 | int limit, backing_limit; | |
1127 | ||
1128 | limit = DRBD_MAX_SEGMENT_SIZE | |
1129 | - ((bio_offset & (DRBD_MAX_SEGMENT_SIZE-1)) + bio_size); | |
1130 | if (limit < 0) | |
1131 | limit = 0; | |
1132 | if (bio_size == 0) { | |
1133 | if (limit <= bvec->bv_len) | |
1134 | limit = bvec->bv_len; | |
1135 | } else if (limit && get_ldev(mdev)) { | |
1136 | struct request_queue * const b = | |
1137 | mdev->ldev->backing_bdev->bd_disk->queue; | |
a1c88d0d | 1138 | if (b->merge_bvec_fn) { |
b411b363 PR |
1139 | backing_limit = b->merge_bvec_fn(b, bvm, bvec); |
1140 | limit = min(limit, backing_limit); | |
1141 | } | |
1142 | put_ldev(mdev); | |
1143 | } | |
1144 | return limit; | |
1145 | } |