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 | ||
9e204cdd AG |
59 | static struct drbd_request *drbd_req_new(struct drbd_conf *mdev, |
60 | struct bio *bio_src) | |
61 | { | |
62 | struct drbd_request *req; | |
63 | ||
64 | req = mempool_alloc(drbd_request_mempool, GFP_NOIO); | |
65 | if (!req) | |
66 | return NULL; | |
67 | ||
68 | drbd_req_make_private_bio(req, bio_src); | |
69 | req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0; | |
70 | req->mdev = mdev; | |
71 | req->master_bio = bio_src; | |
72 | req->epoch = 0; | |
73 | drbd_clear_interval(&req->i); | |
74 | req->i.sector = bio_src->bi_sector; | |
75 | req->i.size = bio_src->bi_size; | |
76 | INIT_LIST_HEAD(&req->tl_requests); | |
77 | INIT_LIST_HEAD(&req->w.list); | |
78 | ||
79 | return req; | |
80 | } | |
81 | ||
82 | static void drbd_req_free(struct drbd_request *req) | |
83 | { | |
84 | mempool_free(req, drbd_request_mempool); | |
85 | } | |
86 | ||
87 | /* rw is bio_data_dir(), only READ or WRITE */ | |
b411b363 PR |
88 | static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw) |
89 | { | |
90 | const unsigned long s = req->rq_state; | |
288f422e PR |
91 | |
92 | /* remove it from the transfer log. | |
93 | * well, only if it had been there in the first | |
94 | * place... if it had not (local only or conflicting | |
95 | * and never sent), it should still be "empty" as | |
96 | * initialized in drbd_req_new(), so we can list_del() it | |
97 | * here unconditionally */ | |
98 | list_del(&req->tl_requests); | |
99 | ||
b411b363 PR |
100 | /* if it was a write, we may have to set the corresponding |
101 | * bit(s) out-of-sync first. If it had a local part, we need to | |
102 | * release the reference to the activity log. */ | |
103 | if (rw == WRITE) { | |
b411b363 PR |
104 | /* Set out-of-sync unless both OK flags are set |
105 | * (local only or remote failed). | |
106 | * Other places where we set out-of-sync: | |
107 | * READ with local io-error */ | |
108 | if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) | |
ace652ac | 109 | drbd_set_out_of_sync(mdev, req->i.sector, req->i.size); |
b411b363 PR |
110 | |
111 | if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) | |
ace652ac | 112 | drbd_set_in_sync(mdev, req->i.sector, req->i.size); |
b411b363 PR |
113 | |
114 | /* one might be tempted to move the drbd_al_complete_io | |
115 | * to the local io completion callback drbd_endio_pri. | |
116 | * but, if this was a mirror write, we may only | |
117 | * drbd_al_complete_io after this is RQ_NET_DONE, | |
118 | * otherwise the extent could be dropped from the al | |
119 | * before it has actually been written on the peer. | |
120 | * if we crash before our peer knows about the request, | |
121 | * but after the extent has been dropped from the al, | |
122 | * we would forget to resync the corresponding extent. | |
123 | */ | |
124 | if (s & RQ_LOCAL_MASK) { | |
125 | if (get_ldev_if_state(mdev, D_FAILED)) { | |
0778286a | 126 | if (s & RQ_IN_ACT_LOG) |
ace652ac | 127 | drbd_al_complete_io(mdev, req->i.sector); |
b411b363 PR |
128 | put_ldev(mdev); |
129 | } else if (__ratelimit(&drbd_ratelimit_state)) { | |
130 | dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu), " | |
131 | "but my Disk seems to have failed :(\n", | |
ace652ac | 132 | (unsigned long long) req->i.sector); |
b411b363 PR |
133 | } |
134 | } | |
135 | } | |
136 | ||
32fa7e91 | 137 | drbd_req_free(req); |
b411b363 PR |
138 | } |
139 | ||
140 | static void queue_barrier(struct drbd_conf *mdev) | |
141 | { | |
142 | struct drbd_tl_epoch *b; | |
143 | ||
144 | /* We are within the req_lock. Once we queued the barrier for sending, | |
145 | * we set the CREATE_BARRIER bit. It is cleared as soon as a new | |
146 | * barrier/epoch object is added. This is the only place this bit is | |
147 | * set. It indicates that the barrier for this epoch is already queued, | |
148 | * and no new epoch has been created yet. */ | |
149 | if (test_bit(CREATE_BARRIER, &mdev->flags)) | |
150 | return; | |
151 | ||
87eeee41 | 152 | b = mdev->tconn->newest_tle; |
b411b363 PR |
153 | b->w.cb = w_send_barrier; |
154 | /* inc_ap_pending done here, so we won't | |
155 | * get imbalanced on connection loss. | |
156 | * dec_ap_pending will be done in got_BarrierAck | |
157 | * or (on connection loss) in tl_clear. */ | |
158 | inc_ap_pending(mdev); | |
e42325a5 | 159 | drbd_queue_work(&mdev->tconn->data.work, &b->w); |
b411b363 PR |
160 | set_bit(CREATE_BARRIER, &mdev->flags); |
161 | } | |
162 | ||
163 | static void _about_to_complete_local_write(struct drbd_conf *mdev, | |
164 | struct drbd_request *req) | |
165 | { | |
166 | const unsigned long s = req->rq_state; | |
b411b363 | 167 | |
8a3c1044 LE |
168 | /* Before we can signal completion to the upper layers, |
169 | * we may need to close the current epoch. | |
170 | * We can skip this, if this request has not even been sent, because we | |
171 | * did not have a fully established connection yet/anymore, during | |
172 | * bitmap exchange, or while we are C_AHEAD due to congestion policy. | |
173 | */ | |
174 | if (mdev->state.conn >= C_CONNECTED && | |
175 | (s & RQ_NET_SENT) != 0 && | |
87eeee41 | 176 | req->epoch == mdev->tconn->newest_tle->br_number) |
b411b363 PR |
177 | queue_barrier(mdev); |
178 | ||
a500c2ef AG |
179 | /* Wake up any processes waiting for this request to complete. */ |
180 | if ((s & RQ_NET_DONE) && (s & RQ_COLLISION)) | |
181 | wake_up(&mdev->misc_wait); | |
b411b363 PR |
182 | } |
183 | ||
184 | void complete_master_bio(struct drbd_conf *mdev, | |
185 | struct bio_and_error *m) | |
186 | { | |
b411b363 PR |
187 | bio_endio(m->bio, m->error); |
188 | dec_ap_bio(mdev); | |
189 | } | |
190 | ||
191 | /* Helper for __req_mod(). | |
192 | * Set m->bio to the master bio, if it is fit to be completed, | |
193 | * or leave it alone (it is initialized to NULL in __req_mod), | |
194 | * if it has already been completed, or cannot be completed yet. | |
195 | * If m->bio is set, the error status to be returned is placed in m->error. | |
196 | */ | |
197 | void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m) | |
198 | { | |
199 | const unsigned long s = req->rq_state; | |
200 | struct drbd_conf *mdev = req->mdev; | |
201 | /* only WRITES may end up here without a master bio (on barrier ack) */ | |
202 | int rw = req->master_bio ? bio_data_dir(req->master_bio) : WRITE; | |
203 | ||
b411b363 PR |
204 | /* we must not complete the master bio, while it is |
205 | * still being processed by _drbd_send_zc_bio (drbd_send_dblock) | |
206 | * not yet acknowledged by the peer | |
207 | * not yet completed by the local io subsystem | |
208 | * these flags may get cleared in any order by | |
209 | * the worker, | |
210 | * the receiver, | |
211 | * the bio_endio completion callbacks. | |
212 | */ | |
213 | if (s & RQ_NET_QUEUED) | |
214 | return; | |
215 | if (s & RQ_NET_PENDING) | |
216 | return; | |
217 | if (s & RQ_LOCAL_PENDING) | |
218 | return; | |
219 | ||
220 | if (req->master_bio) { | |
8554df1c | 221 | /* this is DATA_RECEIVED (remote read) |
b411b363 PR |
222 | * or protocol C P_WRITE_ACK |
223 | * or protocol B P_RECV_ACK | |
8554df1c | 224 | * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck) |
b411b363 PR |
225 | * or canceled or failed, |
226 | * or killed from the transfer log due to connection loss. | |
227 | */ | |
228 | ||
229 | /* | |
230 | * figure out whether to report success or failure. | |
231 | * | |
232 | * report success when at least one of the operations succeeded. | |
233 | * or, to put the other way, | |
234 | * only report failure, when both operations failed. | |
235 | * | |
236 | * what to do about the failures is handled elsewhere. | |
237 | * what we need to do here is just: complete the master_bio. | |
238 | * | |
239 | * local completion error, if any, has been stored as ERR_PTR | |
240 | * in private_bio within drbd_endio_pri. | |
241 | */ | |
242 | int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); | |
243 | int error = PTR_ERR(req->private_bio); | |
244 | ||
245 | /* remove the request from the conflict detection | |
246 | * respective block_id verification hash */ | |
dac1389c AG |
247 | if (!drbd_interval_empty(&req->i)) { |
248 | struct rb_root *root; | |
249 | ||
dac1389c AG |
250 | if (rw == WRITE) |
251 | root = &mdev->write_requests; | |
252 | else | |
253 | root = &mdev->read_requests; | |
254 | drbd_remove_interval(root, &req->i); | |
de696716 | 255 | } else |
8825f7c3 | 256 | D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0); |
b411b363 PR |
257 | |
258 | /* for writes we need to do some extra housekeeping */ | |
259 | if (rw == WRITE) | |
260 | _about_to_complete_local_write(mdev, req); | |
261 | ||
262 | /* Update disk stats */ | |
263 | _drbd_end_io_acct(mdev, req); | |
264 | ||
265 | m->error = ok ? 0 : (error ?: -EIO); | |
266 | m->bio = req->master_bio; | |
267 | req->master_bio = NULL; | |
268 | } | |
269 | ||
270 | if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) { | |
271 | /* this is disconnected (local only) operation, | |
272 | * or protocol C P_WRITE_ACK, | |
273 | * or protocol A or B P_BARRIER_ACK, | |
274 | * or killed from the transfer log due to connection loss. */ | |
275 | _req_is_done(mdev, req, rw); | |
276 | } | |
277 | /* else: network part and not DONE yet. that is | |
278 | * protocol A or B, barrier ack still pending... */ | |
279 | } | |
280 | ||
cfa03415 PR |
281 | static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m) |
282 | { | |
283 | struct drbd_conf *mdev = req->mdev; | |
284 | ||
fb22c402 | 285 | if (!is_susp(mdev->state)) |
cfa03415 PR |
286 | _req_may_be_done(req, m); |
287 | } | |
288 | ||
b411b363 PR |
289 | /* |
290 | * checks whether there was an overlapping request | |
291 | * or ee already registered. | |
292 | * | |
293 | * if so, return 1, in which case this request is completed on the spot, | |
294 | * without ever being submitted or send. | |
295 | * | |
296 | * return 0 if it is ok to submit this request. | |
297 | * | |
298 | * NOTE: | |
299 | * paranoia: assume something above us is broken, and issues different write | |
300 | * requests for the same block simultaneously... | |
301 | * | |
302 | * To ensure these won't be reordered differently on both nodes, resulting in | |
303 | * diverging data sets, we discard the later one(s). Not that this is supposed | |
304 | * to happen, but this is the rationale why we also have to check for | |
305 | * conflicting requests with local origin, and why we have to do so regardless | |
306 | * of whether we allowed multiple primaries. | |
307 | * | |
bb3bfe96 | 308 | * In case we only have one primary, the epoch_entries tree is empty. |
b411b363 PR |
309 | */ |
310 | static int _req_conflicts(struct drbd_request *req) | |
311 | { | |
312 | struct drbd_conf *mdev = req->mdev; | |
ace652ac AG |
313 | const sector_t sector = req->i.sector; |
314 | const int size = req->i.size; | |
de696716 | 315 | struct drbd_interval *i; |
b411b363 | 316 | |
dac1389c | 317 | D_ASSERT(drbd_interval_empty(&req->i)); |
b411b363 | 318 | |
b2fb6dbe | 319 | if (!get_net_conf(mdev->tconn)) |
b411b363 PR |
320 | return 0; |
321 | ||
de696716 AG |
322 | i = drbd_find_overlap(&mdev->write_requests, sector, size); |
323 | if (i) { | |
324 | struct drbd_request *req2 = | |
325 | container_of(i, struct drbd_request, i); | |
326 | ||
327 | dev_alert(DEV, "%s[%u] Concurrent local write detected! " | |
328 | "[DISCARD L] new: %llus +%u; " | |
329 | "pending: %llus +%u\n", | |
330 | current->comm, current->pid, | |
331 | (unsigned long long)sector, size, | |
332 | (unsigned long long)req2->i.sector, req2->i.size); | |
333 | goto out_conflict; | |
b411b363 PR |
334 | } |
335 | ||
bb3bfe96 AG |
336 | if (!RB_EMPTY_ROOT(&mdev->epoch_entries)) { |
337 | /* check for overlapping requests with remote origin */ | |
8b946255 AG |
338 | i = drbd_find_overlap(&mdev->epoch_entries, sector, size); |
339 | if (i) { | |
340 | struct drbd_epoch_entry *e = | |
341 | container_of(i, struct drbd_epoch_entry, i); | |
342 | ||
343 | dev_alert(DEV, "%s[%u] Concurrent remote write detected!" | |
344 | " [DISCARD L] new: %llus +%u; " | |
345 | "pending: %llus +%u\n", | |
346 | current->comm, current->pid, | |
347 | (unsigned long long)sector, size, | |
348 | (unsigned long long)e->i.sector, e->i.size); | |
349 | goto out_conflict; | |
b411b363 PR |
350 | } |
351 | } | |
b411b363 | 352 | |
b411b363 PR |
353 | /* this is like it should be, and what we expected. |
354 | * our users do behave after all... */ | |
b2fb6dbe | 355 | put_net_conf(mdev->tconn); |
b411b363 PR |
356 | return 0; |
357 | ||
358 | out_conflict: | |
b2fb6dbe | 359 | put_net_conf(mdev->tconn); |
b411b363 PR |
360 | return 1; |
361 | } | |
362 | ||
363 | /* obviously this could be coded as many single functions | |
364 | * instead of one huge switch, | |
365 | * or by putting the code directly in the respective locations | |
366 | * (as it has been before). | |
367 | * | |
368 | * but having it this way | |
369 | * enforces that it is all in this one place, where it is easier to audit, | |
370 | * it makes it obvious that whatever "event" "happens" to a request should | |
371 | * happen "atomically" within the req_lock, | |
372 | * and it enforces that we have to think in a very structured manner | |
373 | * about the "events" that may happen to a request during its life time ... | |
374 | */ | |
2a80699f | 375 | int __req_mod(struct drbd_request *req, enum drbd_req_event what, |
b411b363 PR |
376 | struct bio_and_error *m) |
377 | { | |
378 | struct drbd_conf *mdev = req->mdev; | |
2a80699f | 379 | int rv = 0; |
b411b363 PR |
380 | m->bio = NULL; |
381 | ||
b411b363 PR |
382 | switch (what) { |
383 | default: | |
384 | dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); | |
385 | break; | |
386 | ||
387 | /* does not happen... | |
388 | * initialization done in drbd_req_new | |
8554df1c | 389 | case CREATED: |
b411b363 PR |
390 | break; |
391 | */ | |
392 | ||
8554df1c | 393 | case TO_BE_SENT: /* via network */ |
b411b363 PR |
394 | /* reached via drbd_make_request_common |
395 | * and from w_read_retry_remote */ | |
396 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); | |
397 | req->rq_state |= RQ_NET_PENDING; | |
398 | inc_ap_pending(mdev); | |
399 | break; | |
400 | ||
8554df1c | 401 | case TO_BE_SUBMITTED: /* locally */ |
b411b363 PR |
402 | /* reached via drbd_make_request_common */ |
403 | D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK)); | |
404 | req->rq_state |= RQ_LOCAL_PENDING; | |
405 | break; | |
406 | ||
8554df1c | 407 | case COMPLETED_OK: |
b411b363 | 408 | if (bio_data_dir(req->master_bio) == WRITE) |
ace652ac | 409 | mdev->writ_cnt += req->i.size >> 9; |
b411b363 | 410 | else |
ace652ac | 411 | mdev->read_cnt += req->i.size >> 9; |
b411b363 PR |
412 | |
413 | req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); | |
414 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
415 | ||
cfa03415 | 416 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
417 | put_ldev(mdev); |
418 | break; | |
419 | ||
8554df1c | 420 | case WRITE_COMPLETED_WITH_ERROR: |
b411b363 PR |
421 | req->rq_state |= RQ_LOCAL_COMPLETED; |
422 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
423 | ||
81e84650 | 424 | __drbd_chk_io_error(mdev, false); |
cfa03415 | 425 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
426 | put_ldev(mdev); |
427 | break; | |
428 | ||
8554df1c | 429 | case READ_AHEAD_COMPLETED_WITH_ERROR: |
b411b363 PR |
430 | /* it is legal to fail READA */ |
431 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
432 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
cfa03415 | 433 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
434 | put_ldev(mdev); |
435 | break; | |
436 | ||
8554df1c | 437 | case READ_COMPLETED_WITH_ERROR: |
ace652ac | 438 | drbd_set_out_of_sync(mdev, req->i.sector, req->i.size); |
b411b363 PR |
439 | |
440 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
441 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
442 | ||
b411b363 | 443 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); |
b411b363 | 444 | |
81e84650 | 445 | __drbd_chk_io_error(mdev, false); |
b411b363 | 446 | put_ldev(mdev); |
b411b363 | 447 | |
d255e5ff LE |
448 | /* no point in retrying if there is no good remote data, |
449 | * or we have no connection. */ | |
450 | if (mdev->state.pdsk != D_UP_TO_DATE) { | |
cfa03415 | 451 | _req_may_be_done_not_susp(req, m); |
d255e5ff LE |
452 | break; |
453 | } | |
454 | ||
8554df1c | 455 | /* _req_mod(req,TO_BE_SENT); oops, recursion... */ |
d255e5ff LE |
456 | req->rq_state |= RQ_NET_PENDING; |
457 | inc_ap_pending(mdev); | |
8554df1c | 458 | /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */ |
b411b363 | 459 | |
8554df1c | 460 | case QUEUE_FOR_NET_READ: |
b411b363 PR |
461 | /* READ or READA, and |
462 | * no local disk, | |
463 | * or target area marked as invalid, | |
464 | * or just got an io-error. */ | |
465 | /* from drbd_make_request_common | |
466 | * or from bio_endio during read io-error recovery */ | |
467 | ||
468 | /* so we can verify the handle in the answer packet | |
469 | * corresponding hlist_del is in _req_may_be_done() */ | |
dac1389c | 470 | drbd_insert_interval(&mdev->read_requests, &req->i); |
b411b363 | 471 | |
83c38830 | 472 | set_bit(UNPLUG_REMOTE, &mdev->flags); |
b411b363 PR |
473 | |
474 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
475 | req->rq_state |= RQ_NET_QUEUED; | |
476 | req->w.cb = (req->rq_state & RQ_LOCAL_MASK) | |
477 | ? w_read_retry_remote | |
478 | : w_send_read_req; | |
e42325a5 | 479 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
b411b363 PR |
480 | break; |
481 | ||
8554df1c | 482 | case QUEUE_FOR_NET_WRITE: |
b411b363 PR |
483 | /* assert something? */ |
484 | /* from drbd_make_request_common only */ | |
485 | ||
b411b363 | 486 | /* corresponding hlist_del is in _req_may_be_done() */ |
de696716 | 487 | drbd_insert_interval(&mdev->write_requests, &req->i); |
b411b363 PR |
488 | |
489 | /* NOTE | |
490 | * In case the req ended up on the transfer log before being | |
491 | * queued on the worker, it could lead to this request being | |
492 | * missed during cleanup after connection loss. | |
493 | * So we have to do both operations here, | |
494 | * within the same lock that protects the transfer log. | |
495 | * | |
496 | * _req_add_to_epoch(req); this has to be after the | |
497 | * _maybe_start_new_epoch(req); which happened in | |
498 | * drbd_make_request_common, because we now may set the bit | |
499 | * again ourselves to close the current epoch. | |
500 | * | |
501 | * Add req to the (now) current epoch (barrier). */ | |
502 | ||
83c38830 LE |
503 | /* otherwise we may lose an unplug, which may cause some remote |
504 | * io-scheduler timeout to expire, increasing maximum latency, | |
505 | * hurting performance. */ | |
506 | set_bit(UNPLUG_REMOTE, &mdev->flags); | |
507 | ||
b411b363 PR |
508 | /* see drbd_make_request_common, |
509 | * just after it grabs the req_lock */ | |
510 | D_ASSERT(test_bit(CREATE_BARRIER, &mdev->flags) == 0); | |
511 | ||
87eeee41 | 512 | req->epoch = mdev->tconn->newest_tle->br_number; |
b411b363 PR |
513 | |
514 | /* increment size of current epoch */ | |
87eeee41 | 515 | mdev->tconn->newest_tle->n_writes++; |
b411b363 PR |
516 | |
517 | /* queue work item to send data */ | |
518 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
519 | req->rq_state |= RQ_NET_QUEUED; | |
520 | req->w.cb = w_send_dblock; | |
e42325a5 | 521 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
b411b363 PR |
522 | |
523 | /* close the epoch, in case it outgrew the limit */ | |
87eeee41 | 524 | if (mdev->tconn->newest_tle->n_writes >= mdev->tconn->net_conf->max_epoch_size) |
b411b363 PR |
525 | queue_barrier(mdev); |
526 | ||
527 | break; | |
528 | ||
8554df1c | 529 | case QUEUE_FOR_SEND_OOS: |
73a01a18 PR |
530 | req->rq_state |= RQ_NET_QUEUED; |
531 | req->w.cb = w_send_oos; | |
e42325a5 | 532 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
73a01a18 PR |
533 | break; |
534 | ||
8554df1c | 535 | case OOS_HANDED_TO_NETWORK: |
73a01a18 | 536 | /* actually the same */ |
8554df1c | 537 | case SEND_CANCELED: |
b411b363 | 538 | /* treat it the same */ |
8554df1c | 539 | case SEND_FAILED: |
b411b363 PR |
540 | /* real cleanup will be done from tl_clear. just update flags |
541 | * so it is no longer marked as on the worker queue */ | |
542 | req->rq_state &= ~RQ_NET_QUEUED; | |
543 | /* if we did it right, tl_clear should be scheduled only after | |
544 | * this, so this should not be necessary! */ | |
cfa03415 | 545 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
546 | break; |
547 | ||
8554df1c | 548 | case HANDED_OVER_TO_NETWORK: |
b411b363 | 549 | /* assert something? */ |
759fbdfb | 550 | if (bio_data_dir(req->master_bio) == WRITE) |
ace652ac | 551 | atomic_add(req->i.size >> 9, &mdev->ap_in_flight); |
759fbdfb | 552 | |
b411b363 | 553 | if (bio_data_dir(req->master_bio) == WRITE && |
89e58e75 | 554 | mdev->tconn->net_conf->wire_protocol == DRBD_PROT_A) { |
b411b363 PR |
555 | /* this is what is dangerous about protocol A: |
556 | * pretend it was successfully written on the peer. */ | |
557 | if (req->rq_state & RQ_NET_PENDING) { | |
558 | dec_ap_pending(mdev); | |
559 | req->rq_state &= ~RQ_NET_PENDING; | |
560 | req->rq_state |= RQ_NET_OK; | |
561 | } /* else: neg-ack was faster... */ | |
562 | /* it is still not yet RQ_NET_DONE until the | |
563 | * corresponding epoch barrier got acked as well, | |
564 | * so we know what to dirty on connection loss */ | |
565 | } | |
566 | req->rq_state &= ~RQ_NET_QUEUED; | |
567 | req->rq_state |= RQ_NET_SENT; | |
568 | /* because _drbd_send_zc_bio could sleep, and may want to | |
8554df1c AG |
569 | * dereference the bio even after the "WRITE_ACKED_BY_PEER" and |
570 | * "COMPLETED_OK" events came in, once we return from | |
b411b363 PR |
571 | * _drbd_send_zc_bio (drbd_send_dblock), we have to check |
572 | * whether it is done already, and end it. */ | |
cfa03415 | 573 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
574 | break; |
575 | ||
8554df1c | 576 | case READ_RETRY_REMOTE_CANCELED: |
d255e5ff LE |
577 | req->rq_state &= ~RQ_NET_QUEUED; |
578 | /* fall through, in case we raced with drbd_disconnect */ | |
8554df1c | 579 | case CONNECTION_LOST_WHILE_PENDING: |
b411b363 PR |
580 | /* transfer log cleanup after connection loss */ |
581 | /* assert something? */ | |
582 | if (req->rq_state & RQ_NET_PENDING) | |
583 | dec_ap_pending(mdev); | |
584 | req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING); | |
585 | req->rq_state |= RQ_NET_DONE; | |
759fbdfb | 586 | if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE) |
ace652ac | 587 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); |
759fbdfb | 588 | |
b411b363 PR |
589 | /* if it is still queued, we may not complete it here. |
590 | * it will be canceled soon. */ | |
591 | if (!(req->rq_state & RQ_NET_QUEUED)) | |
cfa03415 | 592 | _req_may_be_done(req, m); /* Allowed while state.susp */ |
b411b363 PR |
593 | break; |
594 | ||
8554df1c | 595 | case WRITE_ACKED_BY_PEER_AND_SIS: |
b411b363 | 596 | req->rq_state |= RQ_NET_SIS; |
8554df1c | 597 | case CONFLICT_DISCARDED_BY_PEER: |
b411b363 PR |
598 | /* for discarded conflicting writes of multiple primaries, |
599 | * there is no need to keep anything in the tl, potential | |
600 | * node crashes are covered by the activity log. */ | |
8554df1c | 601 | if (what == CONFLICT_DISCARDED_BY_PEER) |
b411b363 PR |
602 | dev_alert(DEV, "Got DiscardAck packet %llus +%u!" |
603 | " DRBD is not a random data generator!\n", | |
ace652ac | 604 | (unsigned long long)req->i.sector, req->i.size); |
b411b363 PR |
605 | req->rq_state |= RQ_NET_DONE; |
606 | /* fall through */ | |
8554df1c | 607 | case WRITE_ACKED_BY_PEER: |
b411b363 PR |
608 | /* protocol C; successfully written on peer. |
609 | * Nothing to do here. | |
610 | * We want to keep the tl in place for all protocols, to cater | |
611 | * for volatile write-back caches on lower level devices. | |
612 | * | |
613 | * A barrier request is expected to have forced all prior | |
614 | * requests onto stable storage, so completion of a barrier | |
615 | * request could set NET_DONE right here, and not wait for the | |
616 | * P_BARRIER_ACK, but that is an unnecessary optimization. */ | |
617 | ||
618 | /* this makes it effectively the same as for: */ | |
8554df1c | 619 | case RECV_ACKED_BY_PEER: |
b411b363 | 620 | /* protocol B; pretends to be successfully written on peer. |
8554df1c | 621 | * see also notes above in HANDED_OVER_TO_NETWORK about |
b411b363 PR |
622 | * protocol != C */ |
623 | req->rq_state |= RQ_NET_OK; | |
624 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
625 | dec_ap_pending(mdev); | |
ace652ac | 626 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); |
b411b363 | 627 | req->rq_state &= ~RQ_NET_PENDING; |
cfa03415 | 628 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
629 | break; |
630 | ||
8554df1c | 631 | case NEG_ACKED: |
b411b363 | 632 | /* assert something? */ |
759fbdfb | 633 | if (req->rq_state & RQ_NET_PENDING) { |
b411b363 | 634 | dec_ap_pending(mdev); |
ace652ac | 635 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); |
759fbdfb | 636 | } |
b411b363 PR |
637 | req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING); |
638 | ||
639 | req->rq_state |= RQ_NET_DONE; | |
cfa03415 | 640 | _req_may_be_done_not_susp(req, m); |
8554df1c | 641 | /* else: done by HANDED_OVER_TO_NETWORK */ |
b411b363 PR |
642 | break; |
643 | ||
8554df1c | 644 | case FAIL_FROZEN_DISK_IO: |
265be2d0 PR |
645 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
646 | break; | |
647 | ||
cfa03415 | 648 | _req_may_be_done(req, m); /* Allowed while state.susp */ |
265be2d0 PR |
649 | break; |
650 | ||
8554df1c | 651 | case RESTART_FROZEN_DISK_IO: |
265be2d0 PR |
652 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
653 | break; | |
654 | ||
655 | req->rq_state &= ~RQ_LOCAL_COMPLETED; | |
656 | ||
657 | rv = MR_READ; | |
658 | if (bio_data_dir(req->master_bio) == WRITE) | |
659 | rv = MR_WRITE; | |
660 | ||
661 | get_ldev(mdev); | |
662 | req->w.cb = w_restart_disk_io; | |
e42325a5 | 663 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
265be2d0 PR |
664 | break; |
665 | ||
8554df1c | 666 | case RESEND: |
11b58e73 | 667 | /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK |
47ff2d0a | 668 | before the connection loss (B&C only); only P_BARRIER_ACK was missing. |
11b58e73 | 669 | Trowing them out of the TL here by pretending we got a BARRIER_ACK |
481c6f50 | 670 | We ensure that the peer was not rebooted */ |
11b58e73 PR |
671 | if (!(req->rq_state & RQ_NET_OK)) { |
672 | if (req->w.cb) { | |
e42325a5 | 673 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
11b58e73 PR |
674 | rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; |
675 | } | |
676 | break; | |
677 | } | |
8554df1c | 678 | /* else, fall through to BARRIER_ACKED */ |
11b58e73 | 679 | |
8554df1c | 680 | case BARRIER_ACKED: |
288f422e PR |
681 | if (!(req->rq_state & RQ_WRITE)) |
682 | break; | |
683 | ||
b411b363 PR |
684 | if (req->rq_state & RQ_NET_PENDING) { |
685 | /* barrier came in before all requests have been acked. | |
686 | * this is bad, because if the connection is lost now, | |
687 | * we won't be able to clean them up... */ | |
8554df1c | 688 | dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n"); |
87eeee41 | 689 | list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests); |
b411b363 | 690 | } |
e636db5b LE |
691 | if ((req->rq_state & RQ_NET_MASK) != 0) { |
692 | req->rq_state |= RQ_NET_DONE; | |
89e58e75 PR |
693 | if (mdev->tconn->net_conf->wire_protocol == DRBD_PROT_A) |
694 | atomic_sub(req->i.size>>9, &mdev->ap_in_flight); | |
e636db5b | 695 | } |
cfa03415 | 696 | _req_may_be_done(req, m); /* Allowed while state.susp */ |
b411b363 PR |
697 | break; |
698 | ||
8554df1c | 699 | case DATA_RECEIVED: |
b411b363 PR |
700 | D_ASSERT(req->rq_state & RQ_NET_PENDING); |
701 | dec_ap_pending(mdev); | |
702 | req->rq_state &= ~RQ_NET_PENDING; | |
703 | req->rq_state |= (RQ_NET_OK|RQ_NET_DONE); | |
cfa03415 | 704 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
705 | break; |
706 | }; | |
2a80699f PR |
707 | |
708 | return rv; | |
b411b363 PR |
709 | } |
710 | ||
711 | /* we may do a local read if: | |
712 | * - we are consistent (of course), | |
713 | * - or we are generally inconsistent, | |
714 | * BUT we are still/already IN SYNC for this area. | |
715 | * since size may be bigger than BM_BLOCK_SIZE, | |
716 | * we may need to check several bits. | |
717 | */ | |
718 | static int drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size) | |
719 | { | |
720 | unsigned long sbnr, ebnr; | |
721 | sector_t esector, nr_sectors; | |
722 | ||
723 | if (mdev->state.disk == D_UP_TO_DATE) | |
724 | return 1; | |
725 | if (mdev->state.disk >= D_OUTDATED) | |
726 | return 0; | |
727 | if (mdev->state.disk < D_INCONSISTENT) | |
728 | return 0; | |
729 | /* state.disk == D_INCONSISTENT We will have a look at the BitMap */ | |
730 | nr_sectors = drbd_get_capacity(mdev->this_bdev); | |
731 | esector = sector + (size >> 9) - 1; | |
732 | ||
733 | D_ASSERT(sector < nr_sectors); | |
734 | D_ASSERT(esector < nr_sectors); | |
735 | ||
736 | sbnr = BM_SECT_TO_BIT(sector); | |
737 | ebnr = BM_SECT_TO_BIT(esector); | |
738 | ||
739 | return 0 == drbd_bm_count_bits(mdev, sbnr, ebnr); | |
740 | } | |
741 | ||
aeda1cd6 | 742 | static int drbd_make_request_common(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time) |
b411b363 PR |
743 | { |
744 | const int rw = bio_rw(bio); | |
745 | const int size = bio->bi_size; | |
746 | const sector_t sector = bio->bi_sector; | |
747 | struct drbd_tl_epoch *b = NULL; | |
748 | struct drbd_request *req; | |
73a01a18 | 749 | int local, remote, send_oos = 0; |
b411b363 | 750 | int err = -EIO; |
9a25a04c | 751 | int ret = 0; |
b411b363 PR |
752 | |
753 | /* allocate outside of all locks; */ | |
754 | req = drbd_req_new(mdev, bio); | |
755 | if (!req) { | |
756 | dec_ap_bio(mdev); | |
757 | /* only pass the error to the upper layers. | |
758 | * if user cannot handle io errors, that's not our business. */ | |
759 | dev_err(DEV, "could not kmalloc() req\n"); | |
760 | bio_endio(bio, -ENOMEM); | |
761 | return 0; | |
762 | } | |
aeda1cd6 | 763 | req->start_time = start_time; |
b411b363 | 764 | |
b411b363 PR |
765 | local = get_ldev(mdev); |
766 | if (!local) { | |
767 | bio_put(req->private_bio); /* or we get a bio leak */ | |
768 | req->private_bio = NULL; | |
769 | } | |
770 | if (rw == WRITE) { | |
771 | remote = 1; | |
772 | } else { | |
773 | /* READ || READA */ | |
774 | if (local) { | |
775 | if (!drbd_may_do_local_read(mdev, sector, size)) { | |
776 | /* we could kick the syncer to | |
777 | * sync this extent asap, wait for | |
778 | * it, then continue locally. | |
779 | * Or just issue the request remotely. | |
780 | */ | |
781 | local = 0; | |
782 | bio_put(req->private_bio); | |
783 | req->private_bio = NULL; | |
784 | put_ldev(mdev); | |
785 | } | |
786 | } | |
787 | remote = !local && mdev->state.pdsk >= D_UP_TO_DATE; | |
788 | } | |
789 | ||
790 | /* If we have a disk, but a READA request is mapped to remote, | |
791 | * we are R_PRIMARY, D_INCONSISTENT, SyncTarget. | |
792 | * Just fail that READA request right here. | |
793 | * | |
794 | * THINK: maybe fail all READA when not local? | |
795 | * or make this configurable... | |
796 | * if network is slow, READA won't do any good. | |
797 | */ | |
798 | if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) { | |
799 | err = -EWOULDBLOCK; | |
800 | goto fail_and_free_req; | |
801 | } | |
802 | ||
803 | /* For WRITES going to the local disk, grab a reference on the target | |
804 | * extent. This waits for any resync activity in the corresponding | |
805 | * resync extent to finish, and, if necessary, pulls in the target | |
806 | * extent into the activity log, which involves further disk io because | |
807 | * of transactional on-disk meta data updates. */ | |
0778286a PR |
808 | if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) { |
809 | req->rq_state |= RQ_IN_ACT_LOG; | |
b411b363 | 810 | drbd_al_begin_io(mdev, sector); |
0778286a | 811 | } |
b411b363 | 812 | |
6a35c45f PR |
813 | remote = remote && drbd_should_do_remote(mdev->state); |
814 | send_oos = rw == WRITE && drbd_should_send_oos(mdev->state); | |
3719094e | 815 | D_ASSERT(!(remote && send_oos)); |
b411b363 | 816 | |
fb22c402 | 817 | if (!(local || remote) && !is_susp(mdev->state)) { |
fb2c7a10 LE |
818 | if (__ratelimit(&drbd_ratelimit_state)) |
819 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); | |
b411b363 PR |
820 | goto fail_free_complete; |
821 | } | |
822 | ||
823 | /* For WRITE request, we have to make sure that we have an | |
824 | * unused_spare_tle, in case we need to start a new epoch. | |
825 | * I try to be smart and avoid to pre-allocate always "just in case", | |
826 | * but there is a race between testing the bit and pointer outside the | |
827 | * spinlock, and grabbing the spinlock. | |
828 | * if we lost that race, we retry. */ | |
73a01a18 | 829 | if (rw == WRITE && (remote || send_oos) && |
87eeee41 | 830 | mdev->tconn->unused_spare_tle == NULL && |
b411b363 PR |
831 | test_bit(CREATE_BARRIER, &mdev->flags)) { |
832 | allocate_barrier: | |
833 | b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO); | |
834 | if (!b) { | |
835 | dev_err(DEV, "Failed to alloc barrier.\n"); | |
836 | err = -ENOMEM; | |
837 | goto fail_free_complete; | |
838 | } | |
839 | } | |
840 | ||
841 | /* GOOD, everything prepared, grab the spin_lock */ | |
87eeee41 | 842 | spin_lock_irq(&mdev->tconn->req_lock); |
b411b363 | 843 | |
fb22c402 | 844 | if (is_susp(mdev->state)) { |
9a25a04c PR |
845 | /* If we got suspended, use the retry mechanism of |
846 | generic_make_request() to restart processing of this | |
2f58dcfc | 847 | bio. In the next call to drbd_make_request |
9a25a04c PR |
848 | we sleep in inc_ap_bio() */ |
849 | ret = 1; | |
87eeee41 | 850 | spin_unlock_irq(&mdev->tconn->req_lock); |
9a25a04c PR |
851 | goto fail_free_complete; |
852 | } | |
853 | ||
73a01a18 | 854 | if (remote || send_oos) { |
6a35c45f PR |
855 | remote = drbd_should_do_remote(mdev->state); |
856 | send_oos = rw == WRITE && drbd_should_send_oos(mdev->state); | |
3719094e | 857 | D_ASSERT(!(remote && send_oos)); |
73a01a18 PR |
858 | |
859 | if (!(remote || send_oos)) | |
b411b363 PR |
860 | dev_warn(DEV, "lost connection while grabbing the req_lock!\n"); |
861 | if (!(local || remote)) { | |
862 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); | |
87eeee41 | 863 | spin_unlock_irq(&mdev->tconn->req_lock); |
b411b363 PR |
864 | goto fail_free_complete; |
865 | } | |
866 | } | |
867 | ||
87eeee41 PR |
868 | if (b && mdev->tconn->unused_spare_tle == NULL) { |
869 | mdev->tconn->unused_spare_tle = b; | |
b411b363 PR |
870 | b = NULL; |
871 | } | |
73a01a18 | 872 | if (rw == WRITE && (remote || send_oos) && |
87eeee41 | 873 | mdev->tconn->unused_spare_tle == NULL && |
b411b363 PR |
874 | test_bit(CREATE_BARRIER, &mdev->flags)) { |
875 | /* someone closed the current epoch | |
876 | * while we were grabbing the spinlock */ | |
87eeee41 | 877 | spin_unlock_irq(&mdev->tconn->req_lock); |
b411b363 PR |
878 | goto allocate_barrier; |
879 | } | |
880 | ||
881 | ||
882 | /* Update disk stats */ | |
883 | _drbd_start_io_acct(mdev, req, bio); | |
884 | ||
885 | /* _maybe_start_new_epoch(mdev); | |
886 | * If we need to generate a write barrier packet, we have to add the | |
887 | * new epoch (barrier) object, and queue the barrier packet for sending, | |
888 | * and queue the req's data after it _within the same lock_, otherwise | |
889 | * we have race conditions were the reorder domains could be mixed up. | |
890 | * | |
891 | * Even read requests may start a new epoch and queue the corresponding | |
892 | * barrier packet. To get the write ordering right, we only have to | |
893 | * make sure that, if this is a write request and it triggered a | |
894 | * barrier packet, this request is queued within the same spinlock. */ | |
87eeee41 | 895 | if ((remote || send_oos) && mdev->tconn->unused_spare_tle && |
b411b363 | 896 | test_and_clear_bit(CREATE_BARRIER, &mdev->flags)) { |
87eeee41 PR |
897 | _tl_add_barrier(mdev, mdev->tconn->unused_spare_tle); |
898 | mdev->tconn->unused_spare_tle = NULL; | |
b411b363 PR |
899 | } else { |
900 | D_ASSERT(!(remote && rw == WRITE && | |
901 | test_bit(CREATE_BARRIER, &mdev->flags))); | |
902 | } | |
903 | ||
904 | /* NOTE | |
905 | * Actually, 'local' may be wrong here already, since we may have failed | |
906 | * to write to the meta data, and may become wrong anytime because of | |
907 | * local io-error for some other request, which would lead to us | |
908 | * "detaching" the local disk. | |
909 | * | |
910 | * 'remote' may become wrong any time because the network could fail. | |
911 | * | |
912 | * This is a harmless race condition, though, since it is handled | |
913 | * correctly at the appropriate places; so it just defers the failure | |
914 | * of the respective operation. | |
915 | */ | |
916 | ||
917 | /* mark them early for readability. | |
918 | * this just sets some state flags. */ | |
919 | if (remote) | |
8554df1c | 920 | _req_mod(req, TO_BE_SENT); |
b411b363 | 921 | if (local) |
8554df1c | 922 | _req_mod(req, TO_BE_SUBMITTED); |
b411b363 PR |
923 | |
924 | /* check this request on the collision detection hash tables. | |
925 | * if we have a conflict, just complete it here. | |
926 | * THINK do we want to check reads, too? (I don't think so...) */ | |
d28fd092 LE |
927 | if (rw == WRITE && _req_conflicts(req)) |
928 | goto fail_conflicting; | |
288f422e | 929 | |
87eeee41 | 930 | list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests); |
288f422e | 931 | |
b411b363 PR |
932 | /* NOTE remote first: to get the concurrent write detection right, |
933 | * we must register the request before start of local IO. */ | |
934 | if (remote) { | |
935 | /* either WRITE and C_CONNECTED, | |
936 | * or READ, and no local disk, | |
937 | * or READ, but not in sync. | |
938 | */ | |
939 | _req_mod(req, (rw == WRITE) | |
8554df1c AG |
940 | ? QUEUE_FOR_NET_WRITE |
941 | : QUEUE_FOR_NET_READ); | |
b411b363 | 942 | } |
73a01a18 | 943 | if (send_oos && drbd_set_out_of_sync(mdev, sector, size)) |
8554df1c | 944 | _req_mod(req, QUEUE_FOR_SEND_OOS); |
67531718 | 945 | |
73a01a18 | 946 | if (remote && |
31890f4a | 947 | mdev->tconn->net_conf->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) { |
67531718 PR |
948 | int congested = 0; |
949 | ||
89e58e75 PR |
950 | if (mdev->tconn->net_conf->cong_fill && |
951 | atomic_read(&mdev->ap_in_flight) >= mdev->tconn->net_conf->cong_fill) { | |
67531718 PR |
952 | dev_info(DEV, "Congestion-fill threshold reached\n"); |
953 | congested = 1; | |
954 | } | |
955 | ||
89e58e75 | 956 | if (mdev->act_log->used >= mdev->tconn->net_conf->cong_extents) { |
67531718 PR |
957 | dev_info(DEV, "Congestion-extents threshold reached\n"); |
958 | congested = 1; | |
959 | } | |
960 | ||
71c78cfb | 961 | if (congested) { |
039312b6 | 962 | queue_barrier(mdev); /* last barrier, after mirrored writes */ |
73a01a18 | 963 | |
89e58e75 | 964 | if (mdev->tconn->net_conf->on_congestion == OC_PULL_AHEAD) |
67531718 | 965 | _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL); |
89e58e75 | 966 | else /*mdev->tconn->net_conf->on_congestion == OC_DISCONNECT */ |
67531718 PR |
967 | _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL); |
968 | } | |
969 | } | |
970 | ||
87eeee41 | 971 | spin_unlock_irq(&mdev->tconn->req_lock); |
b411b363 PR |
972 | kfree(b); /* if someone else has beaten us to it... */ |
973 | ||
974 | if (local) { | |
975 | req->private_bio->bi_bdev = mdev->ldev->backing_bdev; | |
976 | ||
6719fb03 LE |
977 | /* State may have changed since we grabbed our reference on the |
978 | * mdev->ldev member. Double check, and short-circuit to endio. | |
979 | * In case the last activity log transaction failed to get on | |
980 | * stable storage, and this is a WRITE, we may not even submit | |
981 | * this bio. */ | |
982 | if (get_ldev(mdev)) { | |
0cf9d27e AG |
983 | if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR |
984 | : rw == READ ? DRBD_FAULT_DT_RD | |
985 | : DRBD_FAULT_DT_RA)) | |
6719fb03 LE |
986 | bio_endio(req->private_bio, -EIO); |
987 | else | |
988 | generic_make_request(req->private_bio); | |
989 | put_ldev(mdev); | |
990 | } else | |
b411b363 | 991 | bio_endio(req->private_bio, -EIO); |
b411b363 PR |
992 | } |
993 | ||
b411b363 PR |
994 | return 0; |
995 | ||
d28fd092 LE |
996 | fail_conflicting: |
997 | /* this is a conflicting request. | |
998 | * even though it may have been only _partially_ | |
999 | * overlapping with one of the currently pending requests, | |
1000 | * without even submitting or sending it, we will | |
1001 | * pretend that it was successfully served right now. | |
1002 | */ | |
1003 | _drbd_end_io_acct(mdev, req); | |
87eeee41 | 1004 | spin_unlock_irq(&mdev->tconn->req_lock); |
d28fd092 LE |
1005 | if (remote) |
1006 | dec_ap_pending(mdev); | |
1007 | /* THINK: do we want to fail it (-EIO), or pretend success? | |
1008 | * this pretends success. */ | |
1009 | err = 0; | |
1010 | ||
b411b363 | 1011 | fail_free_complete: |
76727f68 | 1012 | if (req->rq_state & RQ_IN_ACT_LOG) |
b411b363 PR |
1013 | drbd_al_complete_io(mdev, sector); |
1014 | fail_and_free_req: | |
1015 | if (local) { | |
1016 | bio_put(req->private_bio); | |
1017 | req->private_bio = NULL; | |
1018 | put_ldev(mdev); | |
1019 | } | |
9a25a04c PR |
1020 | if (!ret) |
1021 | bio_endio(bio, err); | |
1022 | ||
b411b363 PR |
1023 | drbd_req_free(req); |
1024 | dec_ap_bio(mdev); | |
1025 | kfree(b); | |
1026 | ||
9a25a04c | 1027 | return ret; |
b411b363 PR |
1028 | } |
1029 | ||
1030 | /* helper function for drbd_make_request | |
1031 | * if we can determine just by the mdev (state) that this request will fail, | |
1032 | * return 1 | |
1033 | * otherwise return 0 | |
1034 | */ | |
1035 | static int drbd_fail_request_early(struct drbd_conf *mdev, int is_write) | |
1036 | { | |
b411b363 PR |
1037 | if (mdev->state.role != R_PRIMARY && |
1038 | (!allow_oos || is_write)) { | |
1039 | if (__ratelimit(&drbd_ratelimit_state)) { | |
1040 | dev_err(DEV, "Process %s[%u] tried to %s; " | |
1041 | "since we are not in Primary state, " | |
1042 | "we cannot allow this\n", | |
1043 | current->comm, current->pid, | |
1044 | is_write ? "WRITE" : "READ"); | |
1045 | } | |
1046 | return 1; | |
1047 | } | |
1048 | ||
b411b363 PR |
1049 | return 0; |
1050 | } | |
1051 | ||
2f58dcfc | 1052 | int drbd_make_request(struct request_queue *q, struct bio *bio) |
b411b363 PR |
1053 | { |
1054 | unsigned int s_enr, e_enr; | |
1055 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; | |
aeda1cd6 | 1056 | unsigned long start_time; |
b411b363 PR |
1057 | |
1058 | if (drbd_fail_request_early(mdev, bio_data_dir(bio) & WRITE)) { | |
1059 | bio_endio(bio, -EPERM); | |
1060 | return 0; | |
1061 | } | |
1062 | ||
aeda1cd6 PR |
1063 | start_time = jiffies; |
1064 | ||
b411b363 PR |
1065 | /* |
1066 | * what we "blindly" assume: | |
1067 | */ | |
1068 | D_ASSERT(bio->bi_size > 0); | |
1069 | D_ASSERT((bio->bi_size & 0x1ff) == 0); | |
1070 | D_ASSERT(bio->bi_idx == 0); | |
1071 | ||
1072 | /* to make some things easier, force alignment of requests within the | |
1073 | * granularity of our hash tables */ | |
1074 | s_enr = bio->bi_sector >> HT_SHIFT; | |
1075 | e_enr = (bio->bi_sector+(bio->bi_size>>9)-1) >> HT_SHIFT; | |
1076 | ||
1077 | if (likely(s_enr == e_enr)) { | |
1078 | inc_ap_bio(mdev, 1); | |
aeda1cd6 | 1079 | return drbd_make_request_common(mdev, bio, start_time); |
b411b363 PR |
1080 | } |
1081 | ||
1082 | /* can this bio be split generically? | |
1083 | * Maybe add our own split-arbitrary-bios function. */ | |
1816a2b4 | 1084 | if (bio->bi_vcnt != 1 || bio->bi_idx != 0 || bio->bi_size > DRBD_MAX_BIO_SIZE) { |
b411b363 PR |
1085 | /* rather error out here than BUG in bio_split */ |
1086 | dev_err(DEV, "bio would need to, but cannot, be split: " | |
1087 | "(vcnt=%u,idx=%u,size=%u,sector=%llu)\n", | |
1088 | bio->bi_vcnt, bio->bi_idx, bio->bi_size, | |
1089 | (unsigned long long)bio->bi_sector); | |
1090 | bio_endio(bio, -EINVAL); | |
1091 | } else { | |
1092 | /* This bio crosses some boundary, so we have to split it. */ | |
1093 | struct bio_pair *bp; | |
1094 | /* works for the "do not cross hash slot boundaries" case | |
1095 | * e.g. sector 262269, size 4096 | |
1096 | * s_enr = 262269 >> 6 = 4097 | |
1097 | * e_enr = (262269+8-1) >> 6 = 4098 | |
1098 | * HT_SHIFT = 6 | |
1099 | * sps = 64, mask = 63 | |
1100 | * first_sectors = 64 - (262269 & 63) = 3 | |
1101 | */ | |
1102 | const sector_t sect = bio->bi_sector; | |
1103 | const int sps = 1 << HT_SHIFT; /* sectors per slot */ | |
1104 | const int mask = sps - 1; | |
1105 | const sector_t first_sectors = sps - (sect & mask); | |
03567812 | 1106 | bp = bio_split(bio, first_sectors); |
b411b363 PR |
1107 | |
1108 | /* we need to get a "reference count" (ap_bio_cnt) | |
1109 | * to avoid races with the disconnect/reconnect/suspend code. | |
9a25a04c | 1110 | * In case we need to split the bio here, we need to get three references |
b411b363 PR |
1111 | * atomically, otherwise we might deadlock when trying to submit the |
1112 | * second one! */ | |
9a25a04c | 1113 | inc_ap_bio(mdev, 3); |
b411b363 PR |
1114 | |
1115 | D_ASSERT(e_enr == s_enr + 1); | |
1116 | ||
aeda1cd6 | 1117 | while (drbd_make_request_common(mdev, &bp->bio1, start_time)) |
9a25a04c PR |
1118 | inc_ap_bio(mdev, 1); |
1119 | ||
aeda1cd6 | 1120 | while (drbd_make_request_common(mdev, &bp->bio2, start_time)) |
9a25a04c PR |
1121 | inc_ap_bio(mdev, 1); |
1122 | ||
1123 | dec_ap_bio(mdev); | |
1124 | ||
b411b363 PR |
1125 | bio_pair_release(bp); |
1126 | } | |
1127 | return 0; | |
1128 | } | |
1129 | ||
1130 | /* This is called by bio_add_page(). With this function we reduce | |
1816a2b4 | 1131 | * the number of BIOs that span over multiple DRBD_MAX_BIO_SIZEs |
b411b363 PR |
1132 | * units (was AL_EXTENTs). |
1133 | * | |
1134 | * we do the calculation within the lower 32bit of the byte offsets, | |
1135 | * since we don't care for actual offset, but only check whether it | |
1136 | * would cross "activity log extent" boundaries. | |
1137 | * | |
1138 | * As long as the BIO is empty we have to allow at least one bvec, | |
1139 | * regardless of size and offset. so the resulting bio may still | |
1140 | * cross extent boundaries. those are dealt with (bio_split) in | |
2f58dcfc | 1141 | * drbd_make_request. |
b411b363 PR |
1142 | */ |
1143 | int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec) | |
1144 | { | |
1145 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; | |
1146 | unsigned int bio_offset = | |
1147 | (unsigned int)bvm->bi_sector << 9; /* 32 bit */ | |
1148 | unsigned int bio_size = bvm->bi_size; | |
1149 | int limit, backing_limit; | |
1150 | ||
1816a2b4 LE |
1151 | limit = DRBD_MAX_BIO_SIZE |
1152 | - ((bio_offset & (DRBD_MAX_BIO_SIZE-1)) + bio_size); | |
b411b363 PR |
1153 | if (limit < 0) |
1154 | limit = 0; | |
1155 | if (bio_size == 0) { | |
1156 | if (limit <= bvec->bv_len) | |
1157 | limit = bvec->bv_len; | |
1158 | } else if (limit && get_ldev(mdev)) { | |
1159 | struct request_queue * const b = | |
1160 | mdev->ldev->backing_bdev->bd_disk->queue; | |
a1c88d0d | 1161 | if (b->merge_bvec_fn) { |
b411b363 PR |
1162 | backing_limit = b->merge_bvec_fn(b, bvm, bvec); |
1163 | limit = min(limit, backing_limit); | |
1164 | } | |
1165 | put_ldev(mdev); | |
1166 | } | |
1167 | return limit; | |
1168 | } | |
7fde2be9 PR |
1169 | |
1170 | void request_timer_fn(unsigned long data) | |
1171 | { | |
1172 | struct drbd_conf *mdev = (struct drbd_conf *) data; | |
1173 | struct drbd_request *req; /* oldest request */ | |
1174 | struct list_head *le; | |
1175 | unsigned long et = 0; /* effective timeout = ko_count * timeout */ | |
1176 | ||
b2fb6dbe | 1177 | if (get_net_conf(mdev->tconn)) { |
89e58e75 | 1178 | et = mdev->tconn->net_conf->timeout*HZ/10 * mdev->tconn->net_conf->ko_count; |
b2fb6dbe | 1179 | put_net_conf(mdev->tconn); |
7fde2be9 PR |
1180 | } |
1181 | if (!et || mdev->state.conn < C_WF_REPORT_PARAMS) | |
1182 | return; /* Recurring timer stopped */ | |
1183 | ||
87eeee41 PR |
1184 | spin_lock_irq(&mdev->tconn->req_lock); |
1185 | le = &mdev->tconn->oldest_tle->requests; | |
7fde2be9 | 1186 | if (list_empty(le)) { |
87eeee41 | 1187 | spin_unlock_irq(&mdev->tconn->req_lock); |
7fde2be9 PR |
1188 | mod_timer(&mdev->request_timer, jiffies + et); |
1189 | return; | |
1190 | } | |
1191 | ||
1192 | le = le->prev; | |
1193 | req = list_entry(le, struct drbd_request, tl_requests); | |
1194 | if (time_is_before_eq_jiffies(req->start_time + et)) { | |
1195 | if (req->rq_state & RQ_NET_PENDING) { | |
1196 | dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n"); | |
1197 | _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE, NULL); | |
1198 | } else { | |
1199 | dev_warn(DEV, "Local backing block device frozen?\n"); | |
1200 | mod_timer(&mdev->request_timer, jiffies + et); | |
1201 | } | |
1202 | } else { | |
1203 | mod_timer(&mdev->request_timer, req->start_time + et); | |
1204 | } | |
1205 | ||
87eeee41 | 1206 | spin_unlock_irq(&mdev->tconn->req_lock); |
7fde2be9 | 1207 | } |