]>
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)); | |
376694a0 PR |
42 | (void) cpu; /* The macro invocations above want the cpu argument, I do not like |
43 | the compiler warning about cpu only assigned but never used... */ | |
753c8913 | 44 | part_inc_in_flight(&mdev->vdisk->part0, rw); |
b411b363 | 45 | part_stat_unlock(); |
b411b363 PR |
46 | } |
47 | ||
48 | /* Update disk stats when completing request upwards */ | |
49 | static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req) | |
50 | { | |
51 | int rw = bio_data_dir(req->master_bio); | |
52 | unsigned long duration = jiffies - req->start_time; | |
53 | int cpu; | |
54 | cpu = part_stat_lock(); | |
55 | part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration); | |
56 | part_round_stats(cpu, &mdev->vdisk->part0); | |
753c8913 | 57 | part_dec_in_flight(&mdev->vdisk->part0, rw); |
b411b363 | 58 | part_stat_unlock(); |
b411b363 PR |
59 | } |
60 | ||
9e204cdd AG |
61 | static struct drbd_request *drbd_req_new(struct drbd_conf *mdev, |
62 | struct bio *bio_src) | |
63 | { | |
64 | struct drbd_request *req; | |
65 | ||
66 | req = mempool_alloc(drbd_request_mempool, GFP_NOIO); | |
67 | if (!req) | |
68 | return NULL; | |
69 | ||
70 | drbd_req_make_private_bio(req, bio_src); | |
71 | req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0; | |
a21e9298 | 72 | req->w.mdev = mdev; |
9e204cdd AG |
73 | req->master_bio = bio_src; |
74 | req->epoch = 0; | |
53840641 | 75 | |
9e204cdd AG |
76 | drbd_clear_interval(&req->i); |
77 | req->i.sector = bio_src->bi_sector; | |
78 | req->i.size = bio_src->bi_size; | |
5e472264 | 79 | req->i.local = true; |
53840641 AG |
80 | req->i.waiting = false; |
81 | ||
9e204cdd AG |
82 | INIT_LIST_HEAD(&req->tl_requests); |
83 | INIT_LIST_HEAD(&req->w.list); | |
84 | ||
85 | return req; | |
86 | } | |
87 | ||
88 | static void drbd_req_free(struct drbd_request *req) | |
89 | { | |
90 | mempool_free(req, drbd_request_mempool); | |
91 | } | |
92 | ||
93 | /* rw is bio_data_dir(), only READ or WRITE */ | |
b411b363 PR |
94 | static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const int rw) |
95 | { | |
96 | const unsigned long s = req->rq_state; | |
288f422e PR |
97 | |
98 | /* remove it from the transfer log. | |
99 | * well, only if it had been there in the first | |
100 | * place... if it had not (local only or conflicting | |
101 | * and never sent), it should still be "empty" as | |
102 | * initialized in drbd_req_new(), so we can list_del() it | |
103 | * here unconditionally */ | |
104 | list_del(&req->tl_requests); | |
105 | ||
b411b363 PR |
106 | /* if it was a write, we may have to set the corresponding |
107 | * bit(s) out-of-sync first. If it had a local part, we need to | |
108 | * release the reference to the activity log. */ | |
109 | if (rw == WRITE) { | |
b411b363 PR |
110 | /* Set out-of-sync unless both OK flags are set |
111 | * (local only or remote failed). | |
112 | * Other places where we set out-of-sync: | |
113 | * READ with local io-error */ | |
114 | if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) | |
ace652ac | 115 | drbd_set_out_of_sync(mdev, req->i.sector, req->i.size); |
b411b363 PR |
116 | |
117 | if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) | |
ace652ac | 118 | drbd_set_in_sync(mdev, req->i.sector, req->i.size); |
b411b363 PR |
119 | |
120 | /* one might be tempted to move the drbd_al_complete_io | |
fcefa62e | 121 | * to the local io completion callback drbd_request_endio. |
b411b363 PR |
122 | * but, if this was a mirror write, we may only |
123 | * drbd_al_complete_io after this is RQ_NET_DONE, | |
124 | * otherwise the extent could be dropped from the al | |
125 | * before it has actually been written on the peer. | |
126 | * if we crash before our peer knows about the request, | |
127 | * but after the extent has been dropped from the al, | |
128 | * we would forget to resync the corresponding extent. | |
129 | */ | |
130 | if (s & RQ_LOCAL_MASK) { | |
131 | if (get_ldev_if_state(mdev, D_FAILED)) { | |
0778286a | 132 | if (s & RQ_IN_ACT_LOG) |
181286ad | 133 | drbd_al_complete_io(mdev, &req->i); |
b411b363 PR |
134 | put_ldev(mdev); |
135 | } else if (__ratelimit(&drbd_ratelimit_state)) { | |
181286ad LE |
136 | dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), " |
137 | "but my Disk seems to have failed :(\n", | |
138 | (unsigned long long) req->i.sector, req->i.size); | |
b411b363 PR |
139 | } |
140 | } | |
141 | } | |
142 | ||
32fa7e91 | 143 | drbd_req_free(req); |
b411b363 PR |
144 | } |
145 | ||
146 | static void queue_barrier(struct drbd_conf *mdev) | |
147 | { | |
148 | struct drbd_tl_epoch *b; | |
6936fcb4 | 149 | struct drbd_tconn *tconn = mdev->tconn; |
b411b363 PR |
150 | |
151 | /* We are within the req_lock. Once we queued the barrier for sending, | |
152 | * we set the CREATE_BARRIER bit. It is cleared as soon as a new | |
153 | * barrier/epoch object is added. This is the only place this bit is | |
154 | * set. It indicates that the barrier for this epoch is already queued, | |
155 | * and no new epoch has been created yet. */ | |
6936fcb4 | 156 | if (test_bit(CREATE_BARRIER, &tconn->flags)) |
b411b363 PR |
157 | return; |
158 | ||
6936fcb4 | 159 | b = tconn->newest_tle; |
b411b363 | 160 | b->w.cb = w_send_barrier; |
a21e9298 | 161 | b->w.mdev = mdev; |
b411b363 PR |
162 | /* inc_ap_pending done here, so we won't |
163 | * get imbalanced on connection loss. | |
164 | * dec_ap_pending will be done in got_BarrierAck | |
165 | * or (on connection loss) in tl_clear. */ | |
166 | inc_ap_pending(mdev); | |
6936fcb4 PR |
167 | drbd_queue_work(&tconn->data.work, &b->w); |
168 | set_bit(CREATE_BARRIER, &tconn->flags); | |
b411b363 PR |
169 | } |
170 | ||
171 | static void _about_to_complete_local_write(struct drbd_conf *mdev, | |
172 | struct drbd_request *req) | |
173 | { | |
174 | const unsigned long s = req->rq_state; | |
b411b363 | 175 | |
8a3c1044 LE |
176 | /* Before we can signal completion to the upper layers, |
177 | * we may need to close the current epoch. | |
178 | * We can skip this, if this request has not even been sent, because we | |
179 | * did not have a fully established connection yet/anymore, during | |
180 | * bitmap exchange, or while we are C_AHEAD due to congestion policy. | |
181 | */ | |
182 | if (mdev->state.conn >= C_CONNECTED && | |
183 | (s & RQ_NET_SENT) != 0 && | |
87eeee41 | 184 | req->epoch == mdev->tconn->newest_tle->br_number) |
b411b363 | 185 | queue_barrier(mdev); |
b411b363 PR |
186 | } |
187 | ||
188 | void complete_master_bio(struct drbd_conf *mdev, | |
189 | struct bio_and_error *m) | |
190 | { | |
b411b363 PR |
191 | bio_endio(m->bio, m->error); |
192 | dec_ap_bio(mdev); | |
193 | } | |
194 | ||
53840641 AG |
195 | |
196 | static void drbd_remove_request_interval(struct rb_root *root, | |
197 | struct drbd_request *req) | |
198 | { | |
a21e9298 | 199 | struct drbd_conf *mdev = req->w.mdev; |
53840641 AG |
200 | struct drbd_interval *i = &req->i; |
201 | ||
202 | drbd_remove_interval(root, i); | |
203 | ||
204 | /* Wake up any processes waiting for this request to complete. */ | |
205 | if (i->waiting) | |
206 | wake_up(&mdev->misc_wait); | |
207 | } | |
208 | ||
b411b363 PR |
209 | /* Helper for __req_mod(). |
210 | * Set m->bio to the master bio, if it is fit to be completed, | |
211 | * or leave it alone (it is initialized to NULL in __req_mod), | |
212 | * if it has already been completed, or cannot be completed yet. | |
213 | * If m->bio is set, the error status to be returned is placed in m->error. | |
214 | */ | |
215 | void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m) | |
216 | { | |
217 | const unsigned long s = req->rq_state; | |
a21e9298 | 218 | struct drbd_conf *mdev = req->w.mdev; |
cdfda633 | 219 | int rw = req->rq_state & RQ_WRITE ? WRITE : READ; |
b411b363 | 220 | |
b411b363 PR |
221 | /* we must not complete the master bio, while it is |
222 | * still being processed by _drbd_send_zc_bio (drbd_send_dblock) | |
223 | * not yet acknowledged by the peer | |
224 | * not yet completed by the local io subsystem | |
225 | * these flags may get cleared in any order by | |
226 | * the worker, | |
227 | * the receiver, | |
228 | * the bio_endio completion callbacks. | |
229 | */ | |
cdfda633 | 230 | if (s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) |
7be8da07 AG |
231 | return; |
232 | if (req->i.waiting) { | |
233 | /* Retry all conflicting peer requests. */ | |
234 | wake_up(&mdev->misc_wait); | |
235 | } | |
b411b363 PR |
236 | if (s & RQ_NET_QUEUED) |
237 | return; | |
238 | if (s & RQ_NET_PENDING) | |
239 | return; | |
b411b363 PR |
240 | |
241 | if (req->master_bio) { | |
8554df1c | 242 | /* this is DATA_RECEIVED (remote read) |
b411b363 PR |
243 | * or protocol C P_WRITE_ACK |
244 | * or protocol B P_RECV_ACK | |
8554df1c | 245 | * or protocol A "HANDED_OVER_TO_NETWORK" (SendAck) |
b411b363 PR |
246 | * or canceled or failed, |
247 | * or killed from the transfer log due to connection loss. | |
248 | */ | |
249 | ||
250 | /* | |
251 | * figure out whether to report success or failure. | |
252 | * | |
253 | * report success when at least one of the operations succeeded. | |
254 | * or, to put the other way, | |
255 | * only report failure, when both operations failed. | |
256 | * | |
257 | * what to do about the failures is handled elsewhere. | |
258 | * what we need to do here is just: complete the master_bio. | |
259 | * | |
260 | * local completion error, if any, has been stored as ERR_PTR | |
fcefa62e | 261 | * in private_bio within drbd_request_endio. |
b411b363 PR |
262 | */ |
263 | int ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); | |
264 | int error = PTR_ERR(req->private_bio); | |
265 | ||
266 | /* remove the request from the conflict detection | |
267 | * respective block_id verification hash */ | |
dac1389c AG |
268 | if (!drbd_interval_empty(&req->i)) { |
269 | struct rb_root *root; | |
270 | ||
dac1389c AG |
271 | if (rw == WRITE) |
272 | root = &mdev->write_requests; | |
273 | else | |
274 | root = &mdev->read_requests; | |
53840641 | 275 | drbd_remove_request_interval(root, req); |
7be8da07 | 276 | } else if (!(s & RQ_POSTPONED)) |
8825f7c3 | 277 | D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0); |
b411b363 PR |
278 | |
279 | /* for writes we need to do some extra housekeeping */ | |
280 | if (rw == WRITE) | |
281 | _about_to_complete_local_write(mdev, req); | |
282 | ||
283 | /* Update disk stats */ | |
284 | _drbd_end_io_acct(mdev, req); | |
285 | ||
7be8da07 AG |
286 | if (!(s & RQ_POSTPONED)) { |
287 | m->error = ok ? 0 : (error ?: -EIO); | |
288 | m->bio = req->master_bio; | |
289 | } | |
b411b363 PR |
290 | req->master_bio = NULL; |
291 | } | |
292 | ||
cdfda633 PR |
293 | if (s & RQ_LOCAL_PENDING) |
294 | return; | |
295 | ||
b411b363 PR |
296 | if ((s & RQ_NET_MASK) == 0 || (s & RQ_NET_DONE)) { |
297 | /* this is disconnected (local only) operation, | |
a209b4ae | 298 | * or protocol A, B, or C P_BARRIER_ACK, |
b411b363 PR |
299 | * or killed from the transfer log due to connection loss. */ |
300 | _req_is_done(mdev, req, rw); | |
301 | } | |
302 | /* else: network part and not DONE yet. that is | |
a209b4ae | 303 | * protocol A, B, or C, barrier ack still pending... */ |
b411b363 PR |
304 | } |
305 | ||
cfa03415 PR |
306 | static void _req_may_be_done_not_susp(struct drbd_request *req, struct bio_and_error *m) |
307 | { | |
a21e9298 | 308 | struct drbd_conf *mdev = req->w.mdev; |
cfa03415 | 309 | |
2aebfabb | 310 | if (!drbd_suspended(mdev)) |
cfa03415 PR |
311 | _req_may_be_done(req, m); |
312 | } | |
313 | ||
b411b363 PR |
314 | /* obviously this could be coded as many single functions |
315 | * instead of one huge switch, | |
316 | * or by putting the code directly in the respective locations | |
317 | * (as it has been before). | |
318 | * | |
319 | * but having it this way | |
320 | * enforces that it is all in this one place, where it is easier to audit, | |
321 | * it makes it obvious that whatever "event" "happens" to a request should | |
322 | * happen "atomically" within the req_lock, | |
323 | * and it enforces that we have to think in a very structured manner | |
324 | * about the "events" that may happen to a request during its life time ... | |
325 | */ | |
2a80699f | 326 | int __req_mod(struct drbd_request *req, enum drbd_req_event what, |
b411b363 PR |
327 | struct bio_and_error *m) |
328 | { | |
a21e9298 | 329 | struct drbd_conf *mdev = req->w.mdev; |
44ed167d | 330 | struct net_conf *nc; |
303d1448 | 331 | int p, rv = 0; |
7be8da07 AG |
332 | |
333 | if (m) | |
334 | m->bio = NULL; | |
b411b363 | 335 | |
b411b363 PR |
336 | switch (what) { |
337 | default: | |
338 | dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); | |
339 | break; | |
340 | ||
341 | /* does not happen... | |
342 | * initialization done in drbd_req_new | |
8554df1c | 343 | case CREATED: |
b411b363 PR |
344 | break; |
345 | */ | |
346 | ||
8554df1c | 347 | case TO_BE_SENT: /* via network */ |
7be8da07 | 348 | /* reached via __drbd_make_request |
b411b363 PR |
349 | * and from w_read_retry_remote */ |
350 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); | |
351 | req->rq_state |= RQ_NET_PENDING; | |
44ed167d PR |
352 | rcu_read_lock(); |
353 | nc = rcu_dereference(mdev->tconn->net_conf); | |
354 | p = nc->wire_protocol; | |
355 | rcu_read_unlock(); | |
303d1448 PR |
356 | req->rq_state |= |
357 | p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK : | |
358 | p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0; | |
b411b363 PR |
359 | inc_ap_pending(mdev); |
360 | break; | |
361 | ||
8554df1c | 362 | case TO_BE_SUBMITTED: /* locally */ |
7be8da07 | 363 | /* reached via __drbd_make_request */ |
b411b363 PR |
364 | D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK)); |
365 | req->rq_state |= RQ_LOCAL_PENDING; | |
366 | break; | |
367 | ||
8554df1c | 368 | case COMPLETED_OK: |
cdfda633 | 369 | if (req->rq_state & RQ_WRITE) |
ace652ac | 370 | mdev->writ_cnt += req->i.size >> 9; |
b411b363 | 371 | else |
ace652ac | 372 | mdev->read_cnt += req->i.size >> 9; |
b411b363 PR |
373 | |
374 | req->rq_state |= (RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); | |
375 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
376 | ||
cfa03415 | 377 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
378 | put_ldev(mdev); |
379 | break; | |
380 | ||
cdfda633 PR |
381 | case ABORT_DISK_IO: |
382 | req->rq_state |= RQ_LOCAL_ABORTED; | |
383 | if (req->rq_state & RQ_WRITE) | |
384 | _req_may_be_done_not_susp(req, m); | |
385 | else | |
386 | goto goto_queue_for_net_read; | |
387 | break; | |
388 | ||
8554df1c | 389 | case WRITE_COMPLETED_WITH_ERROR: |
b411b363 PR |
390 | req->rq_state |= RQ_LOCAL_COMPLETED; |
391 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
392 | ||
81e84650 | 393 | __drbd_chk_io_error(mdev, false); |
cfa03415 | 394 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
395 | put_ldev(mdev); |
396 | break; | |
397 | ||
8554df1c | 398 | case READ_AHEAD_COMPLETED_WITH_ERROR: |
b411b363 PR |
399 | /* it is legal to fail READA */ |
400 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
401 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
cfa03415 | 402 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
403 | put_ldev(mdev); |
404 | break; | |
405 | ||
8554df1c | 406 | case READ_COMPLETED_WITH_ERROR: |
ace652ac | 407 | drbd_set_out_of_sync(mdev, req->i.sector, req->i.size); |
b411b363 PR |
408 | |
409 | req->rq_state |= RQ_LOCAL_COMPLETED; | |
410 | req->rq_state &= ~RQ_LOCAL_PENDING; | |
411 | ||
b411b363 | 412 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); |
b411b363 | 413 | |
81e84650 | 414 | __drbd_chk_io_error(mdev, false); |
b411b363 | 415 | put_ldev(mdev); |
b411b363 | 416 | |
cdfda633 PR |
417 | goto_queue_for_net_read: |
418 | ||
d255e5ff LE |
419 | /* no point in retrying if there is no good remote data, |
420 | * or we have no connection. */ | |
421 | if (mdev->state.pdsk != D_UP_TO_DATE) { | |
cfa03415 | 422 | _req_may_be_done_not_susp(req, m); |
d255e5ff LE |
423 | break; |
424 | } | |
425 | ||
8554df1c | 426 | /* _req_mod(req,TO_BE_SENT); oops, recursion... */ |
d255e5ff LE |
427 | req->rq_state |= RQ_NET_PENDING; |
428 | inc_ap_pending(mdev); | |
8554df1c | 429 | /* fall through: _req_mod(req,QUEUE_FOR_NET_READ); */ |
b411b363 | 430 | |
8554df1c | 431 | case QUEUE_FOR_NET_READ: |
b411b363 PR |
432 | /* READ or READA, and |
433 | * no local disk, | |
434 | * or target area marked as invalid, | |
435 | * or just got an io-error. */ | |
7be8da07 | 436 | /* from __drbd_make_request |
b411b363 PR |
437 | * or from bio_endio during read io-error recovery */ |
438 | ||
439 | /* so we can verify the handle in the answer packet | |
440 | * corresponding hlist_del is in _req_may_be_done() */ | |
97ddb687 | 441 | D_ASSERT(drbd_interval_empty(&req->i)); |
dac1389c | 442 | drbd_insert_interval(&mdev->read_requests, &req->i); |
b411b363 | 443 | |
83c38830 | 444 | set_bit(UNPLUG_REMOTE, &mdev->flags); |
b411b363 PR |
445 | |
446 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
447 | req->rq_state |= RQ_NET_QUEUED; | |
448 | req->w.cb = (req->rq_state & RQ_LOCAL_MASK) | |
449 | ? w_read_retry_remote | |
450 | : w_send_read_req; | |
e42325a5 | 451 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
b411b363 PR |
452 | break; |
453 | ||
8554df1c | 454 | case QUEUE_FOR_NET_WRITE: |
b411b363 | 455 | /* assert something? */ |
7be8da07 | 456 | /* from __drbd_make_request only */ |
b411b363 | 457 | |
b411b363 | 458 | /* corresponding hlist_del is in _req_may_be_done() */ |
97ddb687 | 459 | D_ASSERT(drbd_interval_empty(&req->i)); |
de696716 | 460 | drbd_insert_interval(&mdev->write_requests, &req->i); |
b411b363 PR |
461 | |
462 | /* NOTE | |
463 | * In case the req ended up on the transfer log before being | |
464 | * queued on the worker, it could lead to this request being | |
465 | * missed during cleanup after connection loss. | |
466 | * So we have to do both operations here, | |
467 | * within the same lock that protects the transfer log. | |
468 | * | |
469 | * _req_add_to_epoch(req); this has to be after the | |
470 | * _maybe_start_new_epoch(req); which happened in | |
7be8da07 | 471 | * __drbd_make_request, because we now may set the bit |
b411b363 PR |
472 | * again ourselves to close the current epoch. |
473 | * | |
474 | * Add req to the (now) current epoch (barrier). */ | |
475 | ||
83c38830 LE |
476 | /* otherwise we may lose an unplug, which may cause some remote |
477 | * io-scheduler timeout to expire, increasing maximum latency, | |
478 | * hurting performance. */ | |
479 | set_bit(UNPLUG_REMOTE, &mdev->flags); | |
480 | ||
7be8da07 | 481 | /* see __drbd_make_request, |
b411b363 | 482 | * just after it grabs the req_lock */ |
6936fcb4 | 483 | D_ASSERT(test_bit(CREATE_BARRIER, &mdev->tconn->flags) == 0); |
b411b363 | 484 | |
87eeee41 | 485 | req->epoch = mdev->tconn->newest_tle->br_number; |
b411b363 PR |
486 | |
487 | /* increment size of current epoch */ | |
87eeee41 | 488 | mdev->tconn->newest_tle->n_writes++; |
b411b363 PR |
489 | |
490 | /* queue work item to send data */ | |
491 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
492 | req->rq_state |= RQ_NET_QUEUED; | |
493 | req->w.cb = w_send_dblock; | |
e42325a5 | 494 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
b411b363 PR |
495 | |
496 | /* close the epoch, in case it outgrew the limit */ | |
44ed167d PR |
497 | rcu_read_lock(); |
498 | nc = rcu_dereference(mdev->tconn->net_conf); | |
499 | p = nc->max_epoch_size; | |
500 | rcu_read_unlock(); | |
501 | if (mdev->tconn->newest_tle->n_writes >= p) | |
b411b363 PR |
502 | queue_barrier(mdev); |
503 | ||
504 | break; | |
505 | ||
8554df1c | 506 | case QUEUE_FOR_SEND_OOS: |
73a01a18 | 507 | req->rq_state |= RQ_NET_QUEUED; |
8f7bed77 | 508 | req->w.cb = w_send_out_of_sync; |
e42325a5 | 509 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
73a01a18 PR |
510 | break; |
511 | ||
8554df1c | 512 | case OOS_HANDED_TO_NETWORK: |
73a01a18 | 513 | /* actually the same */ |
8554df1c | 514 | case SEND_CANCELED: |
b411b363 | 515 | /* treat it the same */ |
8554df1c | 516 | case SEND_FAILED: |
b411b363 PR |
517 | /* real cleanup will be done from tl_clear. just update flags |
518 | * so it is no longer marked as on the worker queue */ | |
519 | req->rq_state &= ~RQ_NET_QUEUED; | |
520 | /* if we did it right, tl_clear should be scheduled only after | |
521 | * this, so this should not be necessary! */ | |
cfa03415 | 522 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
523 | break; |
524 | ||
8554df1c | 525 | case HANDED_OVER_TO_NETWORK: |
b411b363 | 526 | /* assert something? */ |
759fbdfb | 527 | if (bio_data_dir(req->master_bio) == WRITE) |
ace652ac | 528 | atomic_add(req->i.size >> 9, &mdev->ap_in_flight); |
759fbdfb | 529 | |
b411b363 | 530 | if (bio_data_dir(req->master_bio) == WRITE && |
303d1448 | 531 | !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) { |
b411b363 PR |
532 | /* this is what is dangerous about protocol A: |
533 | * pretend it was successfully written on the peer. */ | |
534 | if (req->rq_state & RQ_NET_PENDING) { | |
535 | dec_ap_pending(mdev); | |
536 | req->rq_state &= ~RQ_NET_PENDING; | |
537 | req->rq_state |= RQ_NET_OK; | |
538 | } /* else: neg-ack was faster... */ | |
539 | /* it is still not yet RQ_NET_DONE until the | |
540 | * corresponding epoch barrier got acked as well, | |
541 | * so we know what to dirty on connection loss */ | |
542 | } | |
543 | req->rq_state &= ~RQ_NET_QUEUED; | |
544 | req->rq_state |= RQ_NET_SENT; | |
545 | /* because _drbd_send_zc_bio could sleep, and may want to | |
8554df1c AG |
546 | * dereference the bio even after the "WRITE_ACKED_BY_PEER" and |
547 | * "COMPLETED_OK" events came in, once we return from | |
b411b363 PR |
548 | * _drbd_send_zc_bio (drbd_send_dblock), we have to check |
549 | * whether it is done already, and end it. */ | |
cfa03415 | 550 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
551 | break; |
552 | ||
8554df1c | 553 | case READ_RETRY_REMOTE_CANCELED: |
d255e5ff LE |
554 | req->rq_state &= ~RQ_NET_QUEUED; |
555 | /* fall through, in case we raced with drbd_disconnect */ | |
8554df1c | 556 | case CONNECTION_LOST_WHILE_PENDING: |
b411b363 PR |
557 | /* transfer log cleanup after connection loss */ |
558 | /* assert something? */ | |
559 | if (req->rq_state & RQ_NET_PENDING) | |
560 | dec_ap_pending(mdev); | |
561 | req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING); | |
562 | req->rq_state |= RQ_NET_DONE; | |
759fbdfb | 563 | if (req->rq_state & RQ_NET_SENT && req->rq_state & RQ_WRITE) |
ace652ac | 564 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); |
759fbdfb | 565 | |
380207d0 PR |
566 | if (!(req->rq_state & RQ_WRITE) && |
567 | mdev->state.disk == D_UP_TO_DATE && | |
568 | !IS_ERR_OR_NULL(req->private_bio)) | |
569 | goto goto_read_retry_local; | |
570 | ||
b411b363 PR |
571 | /* if it is still queued, we may not complete it here. |
572 | * it will be canceled soon. */ | |
573 | if (!(req->rq_state & RQ_NET_QUEUED)) | |
cfa03415 | 574 | _req_may_be_done(req, m); /* Allowed while state.susp */ |
b411b363 PR |
575 | break; |
576 | ||
8554df1c | 577 | case WRITE_ACKED_BY_PEER_AND_SIS: |
b411b363 | 578 | req->rq_state |= RQ_NET_SIS; |
7be8da07 | 579 | case DISCARD_WRITE: |
b411b363 PR |
580 | /* for discarded conflicting writes of multiple primaries, |
581 | * there is no need to keep anything in the tl, potential | |
582 | * node crashes are covered by the activity log. */ | |
b411b363 PR |
583 | req->rq_state |= RQ_NET_DONE; |
584 | /* fall through */ | |
8554df1c | 585 | case WRITE_ACKED_BY_PEER: |
303d1448 | 586 | D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK); |
b411b363 PR |
587 | /* protocol C; successfully written on peer. |
588 | * Nothing to do here. | |
589 | * We want to keep the tl in place for all protocols, to cater | |
590 | * for volatile write-back caches on lower level devices. | |
591 | * | |
592 | * A barrier request is expected to have forced all prior | |
593 | * requests onto stable storage, so completion of a barrier | |
594 | * request could set NET_DONE right here, and not wait for the | |
595 | * P_BARRIER_ACK, but that is an unnecessary optimization. */ | |
596 | ||
303d1448 | 597 | goto ack_common; |
b411b363 | 598 | /* this makes it effectively the same as for: */ |
8554df1c | 599 | case RECV_ACKED_BY_PEER: |
303d1448 | 600 | D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK); |
b411b363 | 601 | /* protocol B; pretends to be successfully written on peer. |
8554df1c | 602 | * see also notes above in HANDED_OVER_TO_NETWORK about |
b411b363 | 603 | * protocol != C */ |
303d1448 | 604 | ack_common: |
b411b363 PR |
605 | req->rq_state |= RQ_NET_OK; |
606 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
607 | dec_ap_pending(mdev); | |
ace652ac | 608 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); |
b411b363 | 609 | req->rq_state &= ~RQ_NET_PENDING; |
cfa03415 | 610 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
611 | break; |
612 | ||
7be8da07 | 613 | case POSTPONE_WRITE: |
303d1448 PR |
614 | D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK); |
615 | /* If this node has already detected the write conflict, the | |
7be8da07 AG |
616 | * worker will be waiting on misc_wait. Wake it up once this |
617 | * request has completed locally. | |
618 | */ | |
619 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
620 | req->rq_state |= RQ_POSTPONED; | |
621 | _req_may_be_done_not_susp(req, m); | |
622 | break; | |
623 | ||
8554df1c | 624 | case NEG_ACKED: |
b411b363 | 625 | /* assert something? */ |
759fbdfb | 626 | if (req->rq_state & RQ_NET_PENDING) { |
b411b363 | 627 | dec_ap_pending(mdev); |
ace652ac | 628 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); |
759fbdfb | 629 | } |
b411b363 PR |
630 | req->rq_state &= ~(RQ_NET_OK|RQ_NET_PENDING); |
631 | ||
632 | req->rq_state |= RQ_NET_DONE; | |
380207d0 PR |
633 | |
634 | if (!(req->rq_state & RQ_WRITE) && | |
635 | mdev->state.disk == D_UP_TO_DATE && | |
636 | !IS_ERR_OR_NULL(req->private_bio)) | |
637 | goto goto_read_retry_local; | |
638 | ||
cfa03415 | 639 | _req_may_be_done_not_susp(req, m); |
8554df1c | 640 | /* else: done by HANDED_OVER_TO_NETWORK */ |
b411b363 PR |
641 | break; |
642 | ||
380207d0 PR |
643 | goto_read_retry_local: |
644 | req->rq_state |= RQ_LOCAL_PENDING; | |
645 | req->private_bio->bi_bdev = mdev->ldev->backing_bdev; | |
646 | generic_make_request(req->private_bio); | |
647 | break; | |
648 | ||
8554df1c | 649 | case FAIL_FROZEN_DISK_IO: |
265be2d0 PR |
650 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
651 | break; | |
652 | ||
cfa03415 | 653 | _req_may_be_done(req, m); /* Allowed while state.susp */ |
265be2d0 PR |
654 | break; |
655 | ||
8554df1c | 656 | case RESTART_FROZEN_DISK_IO: |
265be2d0 PR |
657 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
658 | break; | |
659 | ||
660 | req->rq_state &= ~RQ_LOCAL_COMPLETED; | |
661 | ||
662 | rv = MR_READ; | |
663 | if (bio_data_dir(req->master_bio) == WRITE) | |
664 | rv = MR_WRITE; | |
665 | ||
666 | get_ldev(mdev); | |
667 | req->w.cb = w_restart_disk_io; | |
e42325a5 | 668 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
265be2d0 PR |
669 | break; |
670 | ||
8554df1c | 671 | case RESEND: |
11b58e73 | 672 | /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK |
47ff2d0a | 673 | before the connection loss (B&C only); only P_BARRIER_ACK was missing. |
11b58e73 | 674 | Trowing them out of the TL here by pretending we got a BARRIER_ACK |
481c6f50 | 675 | We ensure that the peer was not rebooted */ |
11b58e73 PR |
676 | if (!(req->rq_state & RQ_NET_OK)) { |
677 | if (req->w.cb) { | |
e42325a5 | 678 | drbd_queue_work(&mdev->tconn->data.work, &req->w); |
11b58e73 PR |
679 | rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; |
680 | } | |
681 | break; | |
682 | } | |
8554df1c | 683 | /* else, fall through to BARRIER_ACKED */ |
11b58e73 | 684 | |
8554df1c | 685 | case BARRIER_ACKED: |
288f422e PR |
686 | if (!(req->rq_state & RQ_WRITE)) |
687 | break; | |
688 | ||
b411b363 | 689 | if (req->rq_state & RQ_NET_PENDING) { |
a209b4ae | 690 | /* barrier came in before all requests were acked. |
b411b363 PR |
691 | * this is bad, because if the connection is lost now, |
692 | * we won't be able to clean them up... */ | |
8554df1c | 693 | dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n"); |
87eeee41 | 694 | list_move(&req->tl_requests, &mdev->tconn->out_of_sequence_requests); |
b411b363 | 695 | } |
e636db5b LE |
696 | if ((req->rq_state & RQ_NET_MASK) != 0) { |
697 | req->rq_state |= RQ_NET_DONE; | |
303d1448 | 698 | if (!(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) |
89e58e75 | 699 | atomic_sub(req->i.size>>9, &mdev->ap_in_flight); |
e636db5b | 700 | } |
cfa03415 | 701 | _req_may_be_done(req, m); /* Allowed while state.susp */ |
b411b363 PR |
702 | break; |
703 | ||
8554df1c | 704 | case DATA_RECEIVED: |
b411b363 PR |
705 | D_ASSERT(req->rq_state & RQ_NET_PENDING); |
706 | dec_ap_pending(mdev); | |
707 | req->rq_state &= ~RQ_NET_PENDING; | |
708 | req->rq_state |= (RQ_NET_OK|RQ_NET_DONE); | |
380207d0 PR |
709 | if (!IS_ERR_OR_NULL(req->private_bio)) { |
710 | bio_put(req->private_bio); | |
711 | req->private_bio = NULL; | |
712 | put_ldev(mdev); | |
713 | } | |
cfa03415 | 714 | _req_may_be_done_not_susp(req, m); |
b411b363 PR |
715 | break; |
716 | }; | |
2a80699f PR |
717 | |
718 | return rv; | |
b411b363 PR |
719 | } |
720 | ||
721 | /* we may do a local read if: | |
722 | * - we are consistent (of course), | |
723 | * - or we are generally inconsistent, | |
724 | * BUT we are still/already IN SYNC for this area. | |
725 | * since size may be bigger than BM_BLOCK_SIZE, | |
726 | * we may need to check several bits. | |
727 | */ | |
0da34df0 | 728 | static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size) |
b411b363 PR |
729 | { |
730 | unsigned long sbnr, ebnr; | |
731 | sector_t esector, nr_sectors; | |
732 | ||
733 | if (mdev->state.disk == D_UP_TO_DATE) | |
0da34df0 | 734 | return true; |
8c387def | 735 | if (mdev->state.disk != D_INCONSISTENT) |
0da34df0 | 736 | return false; |
b411b363 | 737 | esector = sector + (size >> 9) - 1; |
8ca9844f | 738 | nr_sectors = drbd_get_capacity(mdev->this_bdev); |
b411b363 PR |
739 | D_ASSERT(sector < nr_sectors); |
740 | D_ASSERT(esector < nr_sectors); | |
741 | ||
742 | sbnr = BM_SECT_TO_BIT(sector); | |
743 | ebnr = BM_SECT_TO_BIT(esector); | |
744 | ||
0da34df0 | 745 | return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0; |
b411b363 PR |
746 | } |
747 | ||
380207d0 PR |
748 | static bool remote_due_to_read_balancing(struct drbd_conf *mdev) |
749 | { | |
750 | enum drbd_read_balancing rbm; | |
751 | struct backing_dev_info *bdi; | |
752 | ||
753 | if (mdev->state.pdsk < D_UP_TO_DATE) | |
754 | return false; | |
755 | ||
756 | rcu_read_lock(); | |
757 | rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing; | |
758 | rcu_read_unlock(); | |
759 | ||
760 | switch (rbm) { | |
761 | case RB_CONGESTED_REMOTE: | |
762 | bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info; | |
763 | return bdi_read_congested(bdi); | |
764 | case RB_LEAST_PENDING: | |
765 | return atomic_read(&mdev->local_cnt) > | |
766 | atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt); | |
767 | case RB_ROUND_ROBIN: | |
768 | return test_and_change_bit(READ_BALANCE_RR, &mdev->flags); | |
769 | case RB_PREFER_REMOTE: | |
770 | return true; | |
771 | case RB_PREFER_LOCAL: | |
772 | default: | |
773 | return false; | |
774 | } | |
775 | } | |
776 | ||
6024fece AG |
777 | /* |
778 | * complete_conflicting_writes - wait for any conflicting write requests | |
779 | * | |
780 | * The write_requests tree contains all active write requests which we | |
781 | * currently know about. Wait for any requests to complete which conflict with | |
782 | * the new one. | |
783 | */ | |
784 | static int complete_conflicting_writes(struct drbd_conf *mdev, | |
785 | sector_t sector, int size) | |
786 | { | |
787 | for(;;) { | |
6024fece | 788 | struct drbd_interval *i; |
7be8da07 | 789 | int err; |
6024fece AG |
790 | |
791 | i = drbd_find_overlap(&mdev->write_requests, sector, size); | |
792 | if (!i) | |
793 | return 0; | |
7be8da07 AG |
794 | err = drbd_wait_misc(mdev, i); |
795 | if (err) | |
796 | return err; | |
6024fece AG |
797 | } |
798 | } | |
799 | ||
7be8da07 | 800 | int __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time) |
b411b363 PR |
801 | { |
802 | const int rw = bio_rw(bio); | |
803 | const int size = bio->bi_size; | |
804 | const sector_t sector = bio->bi_sector; | |
805 | struct drbd_tl_epoch *b = NULL; | |
806 | struct drbd_request *req; | |
44ed167d | 807 | struct net_conf *nc; |
73a01a18 | 808 | int local, remote, send_oos = 0; |
6024fece | 809 | int err; |
9a25a04c | 810 | int ret = 0; |
b411b363 PR |
811 | |
812 | /* allocate outside of all locks; */ | |
813 | req = drbd_req_new(mdev, bio); | |
814 | if (!req) { | |
815 | dec_ap_bio(mdev); | |
816 | /* only pass the error to the upper layers. | |
817 | * if user cannot handle io errors, that's not our business. */ | |
818 | dev_err(DEV, "could not kmalloc() req\n"); | |
819 | bio_endio(bio, -ENOMEM); | |
820 | return 0; | |
821 | } | |
aeda1cd6 | 822 | req->start_time = start_time; |
b411b363 | 823 | |
b411b363 PR |
824 | local = get_ldev(mdev); |
825 | if (!local) { | |
826 | bio_put(req->private_bio); /* or we get a bio leak */ | |
827 | req->private_bio = NULL; | |
828 | } | |
829 | if (rw == WRITE) { | |
830 | remote = 1; | |
831 | } else { | |
832 | /* READ || READA */ | |
833 | if (local) { | |
834 | if (!drbd_may_do_local_read(mdev, sector, size)) { | |
835 | /* we could kick the syncer to | |
836 | * sync this extent asap, wait for | |
837 | * it, then continue locally. | |
838 | * Or just issue the request remotely. | |
839 | */ | |
840 | local = 0; | |
841 | bio_put(req->private_bio); | |
842 | req->private_bio = NULL; | |
843 | put_ldev(mdev); | |
380207d0 PR |
844 | } else if (remote_due_to_read_balancing(mdev)) { |
845 | /* Keep the private bio in case we need it | |
846 | for a local retry */ | |
847 | local = 0; | |
b411b363 PR |
848 | } |
849 | } | |
850 | remote = !local && mdev->state.pdsk >= D_UP_TO_DATE; | |
851 | } | |
852 | ||
853 | /* If we have a disk, but a READA request is mapped to remote, | |
854 | * we are R_PRIMARY, D_INCONSISTENT, SyncTarget. | |
855 | * Just fail that READA request right here. | |
856 | * | |
857 | * THINK: maybe fail all READA when not local? | |
858 | * or make this configurable... | |
859 | * if network is slow, READA won't do any good. | |
860 | */ | |
861 | if (rw == READA && mdev->state.disk >= D_INCONSISTENT && !local) { | |
862 | err = -EWOULDBLOCK; | |
863 | goto fail_and_free_req; | |
864 | } | |
865 | ||
866 | /* For WRITES going to the local disk, grab a reference on the target | |
867 | * extent. This waits for any resync activity in the corresponding | |
868 | * resync extent to finish, and, if necessary, pulls in the target | |
869 | * extent into the activity log, which involves further disk io because | |
870 | * of transactional on-disk meta data updates. */ | |
0778286a PR |
871 | if (rw == WRITE && local && !test_bit(AL_SUSPENDED, &mdev->flags)) { |
872 | req->rq_state |= RQ_IN_ACT_LOG; | |
181286ad | 873 | drbd_al_begin_io(mdev, &req->i); |
0778286a | 874 | } |
b411b363 | 875 | |
6a35c45f | 876 | remote = remote && drbd_should_do_remote(mdev->state); |
8f7bed77 | 877 | send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state); |
3719094e | 878 | D_ASSERT(!(remote && send_oos)); |
b411b363 | 879 | |
2aebfabb | 880 | if (!(local || remote) && !drbd_suspended(mdev)) { |
fb2c7a10 LE |
881 | if (__ratelimit(&drbd_ratelimit_state)) |
882 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); | |
6024fece | 883 | err = -EIO; |
b411b363 PR |
884 | goto fail_free_complete; |
885 | } | |
886 | ||
887 | /* For WRITE request, we have to make sure that we have an | |
888 | * unused_spare_tle, in case we need to start a new epoch. | |
889 | * I try to be smart and avoid to pre-allocate always "just in case", | |
890 | * but there is a race between testing the bit and pointer outside the | |
891 | * spinlock, and grabbing the spinlock. | |
892 | * if we lost that race, we retry. */ | |
73a01a18 | 893 | if (rw == WRITE && (remote || send_oos) && |
87eeee41 | 894 | mdev->tconn->unused_spare_tle == NULL && |
6936fcb4 | 895 | test_bit(CREATE_BARRIER, &mdev->tconn->flags)) { |
b411b363 PR |
896 | allocate_barrier: |
897 | b = kmalloc(sizeof(struct drbd_tl_epoch), GFP_NOIO); | |
898 | if (!b) { | |
899 | dev_err(DEV, "Failed to alloc barrier.\n"); | |
900 | err = -ENOMEM; | |
901 | goto fail_free_complete; | |
902 | } | |
903 | } | |
904 | ||
905 | /* GOOD, everything prepared, grab the spin_lock */ | |
87eeee41 | 906 | spin_lock_irq(&mdev->tconn->req_lock); |
b411b363 | 907 | |
6024fece AG |
908 | if (rw == WRITE) { |
909 | err = complete_conflicting_writes(mdev, sector, size); | |
910 | if (err) { | |
7be8da07 AG |
911 | if (err != -ERESTARTSYS) |
912 | _conn_request_state(mdev->tconn, | |
913 | NS(conn, C_TIMEOUT), | |
914 | CS_HARD); | |
6024fece | 915 | spin_unlock_irq(&mdev->tconn->req_lock); |
7be8da07 | 916 | err = -EIO; |
6024fece AG |
917 | goto fail_free_complete; |
918 | } | |
919 | } | |
920 | ||
2aebfabb | 921 | if (drbd_suspended(mdev)) { |
9a25a04c PR |
922 | /* If we got suspended, use the retry mechanism of |
923 | generic_make_request() to restart processing of this | |
2f58dcfc | 924 | bio. In the next call to drbd_make_request |
9a25a04c PR |
925 | we sleep in inc_ap_bio() */ |
926 | ret = 1; | |
87eeee41 | 927 | spin_unlock_irq(&mdev->tconn->req_lock); |
9a25a04c PR |
928 | goto fail_free_complete; |
929 | } | |
930 | ||
73a01a18 | 931 | if (remote || send_oos) { |
6a35c45f | 932 | remote = drbd_should_do_remote(mdev->state); |
8f7bed77 | 933 | send_oos = rw == WRITE && drbd_should_send_out_of_sync(mdev->state); |
3719094e | 934 | D_ASSERT(!(remote && send_oos)); |
73a01a18 PR |
935 | |
936 | if (!(remote || send_oos)) | |
b411b363 PR |
937 | dev_warn(DEV, "lost connection while grabbing the req_lock!\n"); |
938 | if (!(local || remote)) { | |
939 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); | |
87eeee41 | 940 | spin_unlock_irq(&mdev->tconn->req_lock); |
6024fece | 941 | err = -EIO; |
b411b363 PR |
942 | goto fail_free_complete; |
943 | } | |
944 | } | |
945 | ||
87eeee41 PR |
946 | if (b && mdev->tconn->unused_spare_tle == NULL) { |
947 | mdev->tconn->unused_spare_tle = b; | |
b411b363 PR |
948 | b = NULL; |
949 | } | |
73a01a18 | 950 | if (rw == WRITE && (remote || send_oos) && |
87eeee41 | 951 | mdev->tconn->unused_spare_tle == NULL && |
6936fcb4 | 952 | test_bit(CREATE_BARRIER, &mdev->tconn->flags)) { |
b411b363 PR |
953 | /* someone closed the current epoch |
954 | * while we were grabbing the spinlock */ | |
87eeee41 | 955 | spin_unlock_irq(&mdev->tconn->req_lock); |
b411b363 PR |
956 | goto allocate_barrier; |
957 | } | |
958 | ||
959 | ||
960 | /* Update disk stats */ | |
961 | _drbd_start_io_acct(mdev, req, bio); | |
962 | ||
963 | /* _maybe_start_new_epoch(mdev); | |
964 | * If we need to generate a write barrier packet, we have to add the | |
965 | * new epoch (barrier) object, and queue the barrier packet for sending, | |
966 | * and queue the req's data after it _within the same lock_, otherwise | |
967 | * we have race conditions were the reorder domains could be mixed up. | |
968 | * | |
969 | * Even read requests may start a new epoch and queue the corresponding | |
970 | * barrier packet. To get the write ordering right, we only have to | |
971 | * make sure that, if this is a write request and it triggered a | |
972 | * barrier packet, this request is queued within the same spinlock. */ | |
87eeee41 | 973 | if ((remote || send_oos) && mdev->tconn->unused_spare_tle && |
6936fcb4 | 974 | test_and_clear_bit(CREATE_BARRIER, &mdev->tconn->flags)) { |
2f5cdd0b | 975 | _tl_add_barrier(mdev->tconn, mdev->tconn->unused_spare_tle); |
87eeee41 | 976 | mdev->tconn->unused_spare_tle = NULL; |
b411b363 PR |
977 | } else { |
978 | D_ASSERT(!(remote && rw == WRITE && | |
6936fcb4 | 979 | test_bit(CREATE_BARRIER, &mdev->tconn->flags))); |
b411b363 PR |
980 | } |
981 | ||
982 | /* NOTE | |
983 | * Actually, 'local' may be wrong here already, since we may have failed | |
984 | * to write to the meta data, and may become wrong anytime because of | |
985 | * local io-error for some other request, which would lead to us | |
986 | * "detaching" the local disk. | |
987 | * | |
988 | * 'remote' may become wrong any time because the network could fail. | |
989 | * | |
990 | * This is a harmless race condition, though, since it is handled | |
991 | * correctly at the appropriate places; so it just defers the failure | |
992 | * of the respective operation. | |
993 | */ | |
994 | ||
995 | /* mark them early for readability. | |
996 | * this just sets some state flags. */ | |
997 | if (remote) | |
8554df1c | 998 | _req_mod(req, TO_BE_SENT); |
b411b363 | 999 | if (local) |
8554df1c | 1000 | _req_mod(req, TO_BE_SUBMITTED); |
b411b363 | 1001 | |
87eeee41 | 1002 | list_add_tail(&req->tl_requests, &mdev->tconn->newest_tle->requests); |
288f422e | 1003 | |
b411b363 PR |
1004 | /* NOTE remote first: to get the concurrent write detection right, |
1005 | * we must register the request before start of local IO. */ | |
1006 | if (remote) { | |
1007 | /* either WRITE and C_CONNECTED, | |
1008 | * or READ, and no local disk, | |
1009 | * or READ, but not in sync. | |
1010 | */ | |
1011 | _req_mod(req, (rw == WRITE) | |
8554df1c AG |
1012 | ? QUEUE_FOR_NET_WRITE |
1013 | : QUEUE_FOR_NET_READ); | |
b411b363 | 1014 | } |
73a01a18 | 1015 | if (send_oos && drbd_set_out_of_sync(mdev, sector, size)) |
8554df1c | 1016 | _req_mod(req, QUEUE_FOR_SEND_OOS); |
67531718 | 1017 | |
44ed167d PR |
1018 | rcu_read_lock(); |
1019 | nc = rcu_dereference(mdev->tconn->net_conf); | |
73a01a18 | 1020 | if (remote && |
44ed167d | 1021 | nc->on_congestion != OC_BLOCK && mdev->tconn->agreed_pro_version >= 96) { |
67531718 PR |
1022 | int congested = 0; |
1023 | ||
44ed167d PR |
1024 | if (nc->cong_fill && |
1025 | atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) { | |
67531718 PR |
1026 | dev_info(DEV, "Congestion-fill threshold reached\n"); |
1027 | congested = 1; | |
1028 | } | |
1029 | ||
44ed167d | 1030 | if (mdev->act_log->used >= nc->cong_extents) { |
67531718 PR |
1031 | dev_info(DEV, "Congestion-extents threshold reached\n"); |
1032 | congested = 1; | |
1033 | } | |
1034 | ||
71c78cfb | 1035 | if (congested) { |
039312b6 | 1036 | queue_barrier(mdev); /* last barrier, after mirrored writes */ |
73a01a18 | 1037 | |
44ed167d | 1038 | if (nc->on_congestion == OC_PULL_AHEAD) |
67531718 | 1039 | _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL); |
44ed167d | 1040 | else /*nc->on_congestion == OC_DISCONNECT */ |
67531718 PR |
1041 | _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL); |
1042 | } | |
1043 | } | |
44ed167d | 1044 | rcu_read_unlock(); |
67531718 | 1045 | |
87eeee41 | 1046 | spin_unlock_irq(&mdev->tconn->req_lock); |
b411b363 PR |
1047 | kfree(b); /* if someone else has beaten us to it... */ |
1048 | ||
1049 | if (local) { | |
1050 | req->private_bio->bi_bdev = mdev->ldev->backing_bdev; | |
1051 | ||
6719fb03 LE |
1052 | /* State may have changed since we grabbed our reference on the |
1053 | * mdev->ldev member. Double check, and short-circuit to endio. | |
1054 | * In case the last activity log transaction failed to get on | |
1055 | * stable storage, and this is a WRITE, we may not even submit | |
1056 | * this bio. */ | |
1057 | if (get_ldev(mdev)) { | |
0cf9d27e AG |
1058 | if (drbd_insert_fault(mdev, rw == WRITE ? DRBD_FAULT_DT_WR |
1059 | : rw == READ ? DRBD_FAULT_DT_RD | |
1060 | : DRBD_FAULT_DT_RA)) | |
6719fb03 LE |
1061 | bio_endio(req->private_bio, -EIO); |
1062 | else | |
1063 | generic_make_request(req->private_bio); | |
1064 | put_ldev(mdev); | |
1065 | } else | |
b411b363 | 1066 | bio_endio(req->private_bio, -EIO); |
b411b363 PR |
1067 | } |
1068 | ||
b411b363 PR |
1069 | return 0; |
1070 | ||
1071 | fail_free_complete: | |
76727f68 | 1072 | if (req->rq_state & RQ_IN_ACT_LOG) |
181286ad | 1073 | drbd_al_complete_io(mdev, &req->i); |
b411b363 | 1074 | fail_and_free_req: |
380207d0 | 1075 | if (!IS_ERR_OR_NULL(req->private_bio)) { |
b411b363 PR |
1076 | bio_put(req->private_bio); |
1077 | req->private_bio = NULL; | |
1078 | put_ldev(mdev); | |
1079 | } | |
9a25a04c PR |
1080 | if (!ret) |
1081 | bio_endio(bio, err); | |
1082 | ||
b411b363 PR |
1083 | drbd_req_free(req); |
1084 | dec_ap_bio(mdev); | |
1085 | kfree(b); | |
1086 | ||
9a25a04c | 1087 | return ret; |
b411b363 PR |
1088 | } |
1089 | ||
2f58dcfc | 1090 | int drbd_make_request(struct request_queue *q, struct bio *bio) |
b411b363 | 1091 | { |
b411b363 | 1092 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; |
aeda1cd6 | 1093 | unsigned long start_time; |
b411b363 | 1094 | |
aeda1cd6 PR |
1095 | start_time = jiffies; |
1096 | ||
b411b363 PR |
1097 | /* |
1098 | * what we "blindly" assume: | |
1099 | */ | |
1100 | D_ASSERT(bio->bi_size > 0); | |
c670a398 | 1101 | D_ASSERT(IS_ALIGNED(bio->bi_size, 512)); |
b411b363 | 1102 | |
23361cf3 LE |
1103 | inc_ap_bio(mdev); |
1104 | return __drbd_make_request(mdev, bio, start_time); | |
b411b363 PR |
1105 | } |
1106 | ||
23361cf3 LE |
1107 | /* This is called by bio_add_page(). |
1108 | * | |
1109 | * q->max_hw_sectors and other global limits are already enforced there. | |
b411b363 | 1110 | * |
23361cf3 LE |
1111 | * We need to call down to our lower level device, |
1112 | * in case it has special restrictions. | |
1113 | * | |
1114 | * We also may need to enforce configured max-bio-bvecs limits. | |
b411b363 PR |
1115 | * |
1116 | * As long as the BIO is empty we have to allow at least one bvec, | |
23361cf3 | 1117 | * regardless of size and offset, so no need to ask lower levels. |
b411b363 PR |
1118 | */ |
1119 | int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec) | |
1120 | { | |
1121 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; | |
b411b363 | 1122 | unsigned int bio_size = bvm->bi_size; |
23361cf3 LE |
1123 | int limit = DRBD_MAX_BIO_SIZE; |
1124 | int backing_limit; | |
1125 | ||
1126 | if (bio_size && get_ldev(mdev)) { | |
b411b363 PR |
1127 | struct request_queue * const b = |
1128 | mdev->ldev->backing_bdev->bd_disk->queue; | |
a1c88d0d | 1129 | if (b->merge_bvec_fn) { |
b411b363 PR |
1130 | backing_limit = b->merge_bvec_fn(b, bvm, bvec); |
1131 | limit = min(limit, backing_limit); | |
1132 | } | |
1133 | put_ldev(mdev); | |
1134 | } | |
1135 | return limit; | |
1136 | } | |
7fde2be9 PR |
1137 | |
1138 | void request_timer_fn(unsigned long data) | |
1139 | { | |
1140 | struct drbd_conf *mdev = (struct drbd_conf *) data; | |
8b924f1d | 1141 | struct drbd_tconn *tconn = mdev->tconn; |
7fde2be9 PR |
1142 | struct drbd_request *req; /* oldest request */ |
1143 | struct list_head *le; | |
44ed167d | 1144 | struct net_conf *nc; |
3b03ad59 | 1145 | unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */ |
44ed167d PR |
1146 | |
1147 | rcu_read_lock(); | |
1148 | nc = rcu_dereference(tconn->net_conf); | |
cdfda633 PR |
1149 | ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0; |
1150 | ||
1151 | if (get_ldev(mdev)) { | |
1152 | dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10; | |
1153 | put_ldev(mdev); | |
1154 | } | |
44ed167d | 1155 | rcu_read_unlock(); |
7fde2be9 | 1156 | |
cdfda633 PR |
1157 | et = min_not_zero(dt, ent); |
1158 | ||
1159 | if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED)) | |
7fde2be9 PR |
1160 | return; /* Recurring timer stopped */ |
1161 | ||
8b924f1d PR |
1162 | spin_lock_irq(&tconn->req_lock); |
1163 | le = &tconn->oldest_tle->requests; | |
7fde2be9 | 1164 | if (list_empty(le)) { |
8b924f1d | 1165 | spin_unlock_irq(&tconn->req_lock); |
7fde2be9 PR |
1166 | mod_timer(&mdev->request_timer, jiffies + et); |
1167 | return; | |
1168 | } | |
1169 | ||
1170 | le = le->prev; | |
1171 | req = list_entry(le, struct drbd_request, tl_requests); | |
cdfda633 PR |
1172 | if (ent && req->rq_state & RQ_NET_PENDING) { |
1173 | if (time_is_before_eq_jiffies(req->start_time + ent)) { | |
7fde2be9 | 1174 | dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n"); |
cdfda633 PR |
1175 | _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL); |
1176 | } | |
1177 | } | |
1178 | if (dt && req->rq_state & RQ_LOCAL_PENDING) { | |
1179 | if (time_is_before_eq_jiffies(req->start_time + dt)) { | |
1180 | dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n"); | |
1181 | __drbd_chk_io_error(mdev, 1); | |
7fde2be9 | 1182 | } |
7fde2be9 | 1183 | } |
3b03ad59 | 1184 | nt = (time_is_before_eq_jiffies(req->start_time + et) ? jiffies : req->start_time) + et; |
8b924f1d | 1185 | spin_unlock_irq(&tconn->req_lock); |
3b03ad59 | 1186 | mod_timer(&mdev->request_timer, nt); |
7fde2be9 | 1187 | } |