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