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 | ||
57bcb6cf PR |
34 | static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size); |
35 | ||
b411b363 PR |
36 | /* Update disk stats at start of I/O request */ |
37 | static void _drbd_start_io_acct(struct drbd_conf *mdev, struct drbd_request *req, struct bio *bio) | |
38 | { | |
39 | const int rw = bio_data_dir(bio); | |
40 | int cpu; | |
41 | cpu = part_stat_lock(); | |
72585d24 | 42 | part_round_stats(cpu, &mdev->vdisk->part0); |
b411b363 PR |
43 | part_stat_inc(cpu, &mdev->vdisk->part0, ios[rw]); |
44 | part_stat_add(cpu, &mdev->vdisk->part0, sectors[rw], bio_sectors(bio)); | |
376694a0 PR |
45 | (void) cpu; /* The macro invocations above want the cpu argument, I do not like |
46 | the compiler warning about cpu only assigned but never used... */ | |
753c8913 | 47 | part_inc_in_flight(&mdev->vdisk->part0, rw); |
b411b363 | 48 | part_stat_unlock(); |
b411b363 PR |
49 | } |
50 | ||
51 | /* Update disk stats when completing request upwards */ | |
52 | static void _drbd_end_io_acct(struct drbd_conf *mdev, struct drbd_request *req) | |
53 | { | |
54 | int rw = bio_data_dir(req->master_bio); | |
55 | unsigned long duration = jiffies - req->start_time; | |
56 | int cpu; | |
57 | cpu = part_stat_lock(); | |
58 | part_stat_add(cpu, &mdev->vdisk->part0, ticks[rw], duration); | |
59 | part_round_stats(cpu, &mdev->vdisk->part0); | |
753c8913 | 60 | part_dec_in_flight(&mdev->vdisk->part0, rw); |
b411b363 | 61 | part_stat_unlock(); |
b411b363 PR |
62 | } |
63 | ||
9e204cdd AG |
64 | static struct drbd_request *drbd_req_new(struct drbd_conf *mdev, |
65 | struct bio *bio_src) | |
66 | { | |
67 | struct drbd_request *req; | |
68 | ||
69 | req = mempool_alloc(drbd_request_mempool, GFP_NOIO); | |
70 | if (!req) | |
71 | return NULL; | |
72 | ||
73 | drbd_req_make_private_bio(req, bio_src); | |
74 | req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0; | |
a21e9298 | 75 | req->w.mdev = mdev; |
9e204cdd AG |
76 | req->master_bio = bio_src; |
77 | req->epoch = 0; | |
53840641 | 78 | |
9e204cdd AG |
79 | drbd_clear_interval(&req->i); |
80 | req->i.sector = bio_src->bi_sector; | |
81 | req->i.size = bio_src->bi_size; | |
5e472264 | 82 | req->i.local = true; |
53840641 AG |
83 | req->i.waiting = false; |
84 | ||
9e204cdd AG |
85 | INIT_LIST_HEAD(&req->tl_requests); |
86 | INIT_LIST_HEAD(&req->w.list); | |
87 | ||
a0d856df | 88 | /* one reference to be put by __drbd_make_request */ |
b406777e | 89 | atomic_set(&req->completion_ref, 1); |
a0d856df | 90 | /* one kref as long as completion_ref > 0 */ |
b406777e | 91 | kref_init(&req->kref); |
9e204cdd AG |
92 | return req; |
93 | } | |
94 | ||
9a278a79 | 95 | void drbd_req_destroy(struct kref *kref) |
b411b363 | 96 | { |
b406777e LE |
97 | struct drbd_request *req = container_of(kref, struct drbd_request, kref); |
98 | struct drbd_conf *mdev = req->w.mdev; | |
a0d856df LE |
99 | const unsigned s = req->rq_state; |
100 | ||
101 | if ((req->master_bio && !(s & RQ_POSTPONED)) || | |
102 | atomic_read(&req->completion_ref) || | |
103 | (s & RQ_LOCAL_PENDING) || | |
104 | ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) { | |
105 | dev_err(DEV, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n", | |
106 | s, atomic_read(&req->completion_ref)); | |
107 | return; | |
108 | } | |
288f422e PR |
109 | |
110 | /* remove it from the transfer log. | |
111 | * well, only if it had been there in the first | |
112 | * place... if it had not (local only or conflicting | |
113 | * and never sent), it should still be "empty" as | |
114 | * initialized in drbd_req_new(), so we can list_del() it | |
115 | * here unconditionally */ | |
2312f0b3 | 116 | list_del_init(&req->tl_requests); |
288f422e | 117 | |
b411b363 PR |
118 | /* if it was a write, we may have to set the corresponding |
119 | * bit(s) out-of-sync first. If it had a local part, we need to | |
120 | * release the reference to the activity log. */ | |
b406777e | 121 | if (s & RQ_WRITE) { |
b411b363 PR |
122 | /* Set out-of-sync unless both OK flags are set |
123 | * (local only or remote failed). | |
124 | * Other places where we set out-of-sync: | |
125 | * READ with local io-error */ | |
d7644018 PR |
126 | if (!(s & RQ_POSTPONED)) { |
127 | if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK)) | |
128 | drbd_set_out_of_sync(mdev, req->i.sector, req->i.size); | |
b411b363 | 129 | |
d7644018 PR |
130 | if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS)) |
131 | drbd_set_in_sync(mdev, req->i.sector, req->i.size); | |
132 | } | |
b411b363 PR |
133 | |
134 | /* one might be tempted to move the drbd_al_complete_io | |
fcefa62e | 135 | * to the local io completion callback drbd_request_endio. |
b411b363 PR |
136 | * but, if this was a mirror write, we may only |
137 | * drbd_al_complete_io after this is RQ_NET_DONE, | |
138 | * otherwise the extent could be dropped from the al | |
139 | * before it has actually been written on the peer. | |
140 | * if we crash before our peer knows about the request, | |
141 | * but after the extent has been dropped from the al, | |
142 | * we would forget to resync the corresponding extent. | |
143 | */ | |
76590cd1 | 144 | if (s & RQ_IN_ACT_LOG) { |
b411b363 | 145 | if (get_ldev_if_state(mdev, D_FAILED)) { |
76590cd1 | 146 | drbd_al_complete_io(mdev, &req->i); |
b411b363 PR |
147 | put_ldev(mdev); |
148 | } else if (__ratelimit(&drbd_ratelimit_state)) { | |
181286ad LE |
149 | dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), " |
150 | "but my Disk seems to have failed :(\n", | |
151 | (unsigned long long) req->i.sector, req->i.size); | |
b411b363 PR |
152 | } |
153 | } | |
154 | } | |
155 | ||
9a278a79 | 156 | mempool_free(req, drbd_request_mempool); |
b411b363 PR |
157 | } |
158 | ||
b6dd1a89 LE |
159 | static void wake_all_senders(struct drbd_tconn *tconn) { |
160 | wake_up(&tconn->sender_work.q_wait); | |
b411b363 PR |
161 | } |
162 | ||
b6dd1a89 LE |
163 | /* must hold resource->req_lock */ |
164 | static void start_new_tl_epoch(struct drbd_tconn *tconn) | |
b411b363 | 165 | { |
99b4d8fe LE |
166 | /* no point closing an epoch, if it is empty, anyways. */ |
167 | if (tconn->current_tle_writes == 0) | |
168 | return; | |
169 | ||
b6dd1a89 LE |
170 | tconn->current_tle_writes = 0; |
171 | atomic_inc(&tconn->current_tle_nr); | |
172 | wake_all_senders(tconn); | |
b411b363 PR |
173 | } |
174 | ||
175 | void complete_master_bio(struct drbd_conf *mdev, | |
176 | struct bio_and_error *m) | |
177 | { | |
b411b363 PR |
178 | bio_endio(m->bio, m->error); |
179 | dec_ap_bio(mdev); | |
180 | } | |
181 | ||
53840641 AG |
182 | |
183 | static void drbd_remove_request_interval(struct rb_root *root, | |
184 | struct drbd_request *req) | |
185 | { | |
a21e9298 | 186 | struct drbd_conf *mdev = req->w.mdev; |
53840641 AG |
187 | struct drbd_interval *i = &req->i; |
188 | ||
189 | drbd_remove_interval(root, i); | |
190 | ||
191 | /* Wake up any processes waiting for this request to complete. */ | |
192 | if (i->waiting) | |
193 | wake_up(&mdev->misc_wait); | |
194 | } | |
195 | ||
b411b363 PR |
196 | /* Helper for __req_mod(). |
197 | * Set m->bio to the master bio, if it is fit to be completed, | |
198 | * or leave it alone (it is initialized to NULL in __req_mod), | |
199 | * if it has already been completed, or cannot be completed yet. | |
200 | * If m->bio is set, the error status to be returned is placed in m->error. | |
201 | */ | |
6870ca6d | 202 | static |
a0d856df | 203 | void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m) |
b411b363 | 204 | { |
a0d856df | 205 | const unsigned s = req->rq_state; |
a21e9298 | 206 | struct drbd_conf *mdev = req->w.mdev; |
a0d856df LE |
207 | int rw; |
208 | int error, ok; | |
b411b363 | 209 | |
b411b363 PR |
210 | /* we must not complete the master bio, while it is |
211 | * still being processed by _drbd_send_zc_bio (drbd_send_dblock) | |
212 | * not yet acknowledged by the peer | |
213 | * not yet completed by the local io subsystem | |
214 | * these flags may get cleared in any order by | |
215 | * the worker, | |
216 | * the receiver, | |
217 | * the bio_endio completion callbacks. | |
218 | */ | |
a0d856df LE |
219 | if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) || |
220 | (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) || | |
221 | (s & RQ_COMPLETION_SUSP)) { | |
222 | dev_err(DEV, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s); | |
b411b363 | 223 | return; |
a0d856df LE |
224 | } |
225 | ||
226 | if (!req->master_bio) { | |
227 | dev_err(DEV, "drbd_req_complete: Logic BUG, master_bio == NULL!\n"); | |
b411b363 | 228 | return; |
a0d856df | 229 | } |
b411b363 | 230 | |
a0d856df | 231 | rw = bio_rw(req->master_bio); |
b406777e | 232 | |
a0d856df LE |
233 | /* |
234 | * figure out whether to report success or failure. | |
235 | * | |
236 | * report success when at least one of the operations succeeded. | |
237 | * or, to put the other way, | |
238 | * only report failure, when both operations failed. | |
239 | * | |
240 | * what to do about the failures is handled elsewhere. | |
241 | * what we need to do here is just: complete the master_bio. | |
242 | * | |
243 | * local completion error, if any, has been stored as ERR_PTR | |
244 | * in private_bio within drbd_request_endio. | |
245 | */ | |
246 | ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK); | |
247 | error = PTR_ERR(req->private_bio); | |
4439c400 | 248 | |
a0d856df LE |
249 | /* remove the request from the conflict detection |
250 | * respective block_id verification hash */ | |
251 | if (!drbd_interval_empty(&req->i)) { | |
252 | struct rb_root *root; | |
b411b363 | 253 | |
a0d856df LE |
254 | if (rw == WRITE) |
255 | root = &mdev->write_requests; | |
256 | else | |
257 | root = &mdev->read_requests; | |
258 | drbd_remove_request_interval(root, req); | |
259 | } else if (!(s & RQ_POSTPONED)) | |
260 | D_ASSERT((s & (RQ_NET_MASK & ~RQ_NET_DONE)) == 0); | |
261 | ||
262 | /* Before we can signal completion to the upper layers, | |
263 | * we may need to close the current transfer log epoch. | |
264 | * We are within the request lock, so we can simply compare | |
265 | * the request epoch number with the current transfer log | |
266 | * epoch number. If they match, increase the current_tle_nr, | |
267 | * and reset the transfer log epoch write_cnt. | |
268 | */ | |
269 | if (rw == WRITE && | |
270 | req->epoch == atomic_read(&mdev->tconn->current_tle_nr)) | |
271 | start_new_tl_epoch(mdev->tconn); | |
b411b363 | 272 | |
a0d856df LE |
273 | /* Update disk stats */ |
274 | _drbd_end_io_acct(mdev, req); | |
b411b363 | 275 | |
a0d856df LE |
276 | /* If READ failed, |
277 | * have it be pushed back to the retry work queue, | |
278 | * so it will re-enter __drbd_make_request(), | |
279 | * and be re-assigned to a suitable local or remote path, | |
280 | * or failed if we do not have access to good data anymore. | |
281 | * | |
282 | * Unless it was failed early by __drbd_make_request(), | |
283 | * because no path was available, in which case | |
284 | * it was not even added to the transfer_log. | |
285 | * | |
286 | * READA may fail, and will not be retried. | |
287 | * | |
288 | * WRITE should have used all available paths already. | |
289 | */ | |
290 | if (!ok && rw == READ && !list_empty(&req->tl_requests)) | |
291 | req->rq_state |= RQ_POSTPONED; | |
292 | ||
293 | if (!(req->rq_state & RQ_POSTPONED)) { | |
294 | m->error = ok ? 0 : (error ?: -EIO); | |
295 | m->bio = req->master_bio; | |
296 | req->master_bio = NULL; | |
b411b363 | 297 | } |
b411b363 PR |
298 | } |
299 | ||
a0d856df | 300 | static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put) |
cfa03415 | 301 | { |
a21e9298 | 302 | struct drbd_conf *mdev = req->w.mdev; |
a0d856df LE |
303 | D_ASSERT(m || (req->rq_state & RQ_POSTPONED)); |
304 | ||
305 | if (!atomic_sub_and_test(put, &req->completion_ref)) | |
306 | return 0; | |
307 | ||
a0d856df | 308 | drbd_req_complete(req, m); |
9a278a79 LE |
309 | |
310 | if (req->rq_state & RQ_POSTPONED) { | |
311 | /* don't destroy the req object just yet, | |
312 | * but queue it for retry */ | |
313 | drbd_restart_request(req); | |
314 | return 0; | |
315 | } | |
316 | ||
a0d856df LE |
317 | return 1; |
318 | } | |
319 | ||
320 | /* I'd like this to be the only place that manipulates | |
321 | * req->completion_ref and req->kref. */ | |
322 | static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m, | |
323 | int clear, int set) | |
324 | { | |
325 | struct drbd_conf *mdev = req->w.mdev; | |
326 | unsigned s = req->rq_state; | |
327 | int c_put = 0; | |
328 | int k_put = 0; | |
329 | ||
5af2e8ce PR |
330 | if (drbd_suspended(mdev) && !((s | clear) & RQ_COMPLETION_SUSP)) |
331 | set |= RQ_COMPLETION_SUSP; | |
332 | ||
a0d856df LE |
333 | /* apply */ |
334 | ||
335 | req->rq_state &= ~clear; | |
336 | req->rq_state |= set; | |
337 | ||
338 | /* no change? */ | |
339 | if (req->rq_state == s) | |
340 | return; | |
341 | ||
342 | /* intent: get references */ | |
343 | ||
344 | if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING)) | |
345 | atomic_inc(&req->completion_ref); | |
346 | ||
347 | if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) { | |
348 | inc_ap_pending(mdev); | |
349 | atomic_inc(&req->completion_ref); | |
350 | } | |
351 | ||
352 | if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) | |
353 | atomic_inc(&req->completion_ref); | |
354 | ||
355 | if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK)) | |
356 | kref_get(&req->kref); /* wait for the DONE */ | |
357 | ||
358 | if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) | |
359 | atomic_add(req->i.size >> 9, &mdev->ap_in_flight); | |
360 | ||
5af2e8ce PR |
361 | if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP)) |
362 | atomic_inc(&req->completion_ref); | |
363 | ||
a0d856df LE |
364 | /* progress: put references */ |
365 | ||
366 | if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP)) | |
367 | ++c_put; | |
368 | ||
369 | if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) { | |
370 | D_ASSERT(req->rq_state & RQ_LOCAL_PENDING); | |
371 | /* local completion may still come in later, | |
372 | * we need to keep the req object around. */ | |
373 | kref_get(&req->kref); | |
374 | ++c_put; | |
375 | } | |
376 | ||
377 | if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) { | |
378 | if (req->rq_state & RQ_LOCAL_ABORTED) | |
379 | ++k_put; | |
380 | else | |
381 | ++c_put; | |
382 | } | |
383 | ||
384 | if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) { | |
385 | dec_ap_pending(mdev); | |
386 | ++c_put; | |
387 | } | |
388 | ||
389 | if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) | |
390 | ++c_put; | |
391 | ||
392 | if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) { | |
393 | if (req->rq_state & RQ_NET_SENT) | |
394 | atomic_sub(req->i.size >> 9, &mdev->ap_in_flight); | |
395 | ++k_put; | |
396 | } | |
397 | ||
398 | /* potentially complete and destroy */ | |
399 | ||
400 | if (k_put || c_put) { | |
401 | /* Completion does it's own kref_put. If we are going to | |
402 | * kref_sub below, we need req to be still around then. */ | |
403 | int at_least = k_put + !!c_put; | |
404 | int refcount = atomic_read(&req->kref.refcount); | |
405 | if (refcount < at_least) | |
406 | dev_err(DEV, | |
407 | "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n", | |
408 | s, req->rq_state, refcount, at_least); | |
409 | } | |
410 | ||
411 | /* If we made progress, retry conflicting peer requests, if any. */ | |
412 | if (req->i.waiting) | |
413 | wake_up(&mdev->misc_wait); | |
414 | ||
415 | if (c_put) | |
416 | k_put += drbd_req_put_completion_ref(req, m, c_put); | |
417 | if (k_put) | |
418 | kref_sub(&req->kref, k_put, drbd_req_destroy); | |
cfa03415 PR |
419 | } |
420 | ||
b411b363 PR |
421 | /* obviously this could be coded as many single functions |
422 | * instead of one huge switch, | |
423 | * or by putting the code directly in the respective locations | |
424 | * (as it has been before). | |
425 | * | |
426 | * but having it this way | |
427 | * enforces that it is all in this one place, where it is easier to audit, | |
428 | * it makes it obvious that whatever "event" "happens" to a request should | |
429 | * happen "atomically" within the req_lock, | |
430 | * and it enforces that we have to think in a very structured manner | |
431 | * about the "events" that may happen to a request during its life time ... | |
432 | */ | |
2a80699f | 433 | int __req_mod(struct drbd_request *req, enum drbd_req_event what, |
b411b363 PR |
434 | struct bio_and_error *m) |
435 | { | |
a21e9298 | 436 | struct drbd_conf *mdev = req->w.mdev; |
44ed167d | 437 | struct net_conf *nc; |
303d1448 | 438 | int p, rv = 0; |
7be8da07 AG |
439 | |
440 | if (m) | |
441 | m->bio = NULL; | |
b411b363 | 442 | |
b411b363 PR |
443 | switch (what) { |
444 | default: | |
445 | dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__); | |
446 | break; | |
447 | ||
448 | /* does not happen... | |
449 | * initialization done in drbd_req_new | |
8554df1c | 450 | case CREATED: |
b411b363 PR |
451 | break; |
452 | */ | |
453 | ||
8554df1c | 454 | case TO_BE_SENT: /* via network */ |
7be8da07 | 455 | /* reached via __drbd_make_request |
b411b363 PR |
456 | * and from w_read_retry_remote */ |
457 | D_ASSERT(!(req->rq_state & RQ_NET_MASK)); | |
44ed167d PR |
458 | rcu_read_lock(); |
459 | nc = rcu_dereference(mdev->tconn->net_conf); | |
460 | p = nc->wire_protocol; | |
461 | rcu_read_unlock(); | |
303d1448 PR |
462 | req->rq_state |= |
463 | p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK : | |
464 | p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0; | |
a0d856df | 465 | mod_rq_state(req, m, 0, RQ_NET_PENDING); |
b411b363 PR |
466 | break; |
467 | ||
8554df1c | 468 | case TO_BE_SUBMITTED: /* locally */ |
7be8da07 | 469 | /* reached via __drbd_make_request */ |
b411b363 | 470 | D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK)); |
a0d856df | 471 | mod_rq_state(req, m, 0, RQ_LOCAL_PENDING); |
b411b363 PR |
472 | break; |
473 | ||
8554df1c | 474 | case COMPLETED_OK: |
cdfda633 | 475 | if (req->rq_state & RQ_WRITE) |
ace652ac | 476 | mdev->writ_cnt += req->i.size >> 9; |
b411b363 | 477 | else |
ace652ac | 478 | mdev->read_cnt += req->i.size >> 9; |
b411b363 | 479 | |
a0d856df LE |
480 | mod_rq_state(req, m, RQ_LOCAL_PENDING, |
481 | RQ_LOCAL_COMPLETED|RQ_LOCAL_OK); | |
b411b363 PR |
482 | break; |
483 | ||
cdfda633 | 484 | case ABORT_DISK_IO: |
a0d856df | 485 | mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED); |
b411b363 PR |
486 | break; |
487 | ||
8554df1c | 488 | case READ_COMPLETED_WITH_ERROR: |
ace652ac | 489 | drbd_set_out_of_sync(mdev, req->i.sector, req->i.size); |
a0d856df LE |
490 | /* fall through. */ |
491 | case WRITE_COMPLETED_WITH_ERROR: | |
0c849666 | 492 | __drbd_chk_io_error(mdev, DRBD_IO_ERROR); |
a0d856df LE |
493 | /* fall through. */ |
494 | case READ_AHEAD_COMPLETED_WITH_ERROR: | |
495 | /* it is legal to fail READA, no __drbd_chk_io_error in that case. */ | |
496 | mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED); | |
4439c400 | 497 | break; |
b411b363 | 498 | |
8554df1c | 499 | case QUEUE_FOR_NET_READ: |
b411b363 PR |
500 | /* READ or READA, and |
501 | * no local disk, | |
502 | * or target area marked as invalid, | |
503 | * or just got an io-error. */ | |
7be8da07 | 504 | /* from __drbd_make_request |
b411b363 PR |
505 | * or from bio_endio during read io-error recovery */ |
506 | ||
6870ca6d LE |
507 | /* So we can verify the handle in the answer packet. |
508 | * Corresponding drbd_remove_request_interval is in | |
a0d856df | 509 | * drbd_req_complete() */ |
97ddb687 | 510 | D_ASSERT(drbd_interval_empty(&req->i)); |
dac1389c | 511 | drbd_insert_interval(&mdev->read_requests, &req->i); |
b411b363 | 512 | |
83c38830 | 513 | set_bit(UNPLUG_REMOTE, &mdev->flags); |
b411b363 PR |
514 | |
515 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
4439c400 | 516 | D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0); |
a0d856df | 517 | mod_rq_state(req, m, 0, RQ_NET_QUEUED); |
4439c400 | 518 | req->w.cb = w_send_read_req; |
d5b27b01 | 519 | drbd_queue_work(&mdev->tconn->sender_work, &req->w); |
b411b363 PR |
520 | break; |
521 | ||
8554df1c | 522 | case QUEUE_FOR_NET_WRITE: |
b411b363 | 523 | /* assert something? */ |
7be8da07 | 524 | /* from __drbd_make_request only */ |
b411b363 | 525 | |
6870ca6d | 526 | /* Corresponding drbd_remove_request_interval is in |
a0d856df | 527 | * drbd_req_complete() */ |
97ddb687 | 528 | D_ASSERT(drbd_interval_empty(&req->i)); |
de696716 | 529 | drbd_insert_interval(&mdev->write_requests, &req->i); |
b411b363 PR |
530 | |
531 | /* NOTE | |
532 | * In case the req ended up on the transfer log before being | |
533 | * queued on the worker, it could lead to this request being | |
534 | * missed during cleanup after connection loss. | |
535 | * So we have to do both operations here, | |
536 | * within the same lock that protects the transfer log. | |
537 | * | |
538 | * _req_add_to_epoch(req); this has to be after the | |
539 | * _maybe_start_new_epoch(req); which happened in | |
7be8da07 | 540 | * __drbd_make_request, because we now may set the bit |
b411b363 PR |
541 | * again ourselves to close the current epoch. |
542 | * | |
543 | * Add req to the (now) current epoch (barrier). */ | |
544 | ||
83c38830 LE |
545 | /* otherwise we may lose an unplug, which may cause some remote |
546 | * io-scheduler timeout to expire, increasing maximum latency, | |
547 | * hurting performance. */ | |
548 | set_bit(UNPLUG_REMOTE, &mdev->flags); | |
549 | ||
b411b363 PR |
550 | /* queue work item to send data */ |
551 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
a0d856df | 552 | mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK); |
b411b363 | 553 | req->w.cb = w_send_dblock; |
d5b27b01 | 554 | drbd_queue_work(&mdev->tconn->sender_work, &req->w); |
b411b363 PR |
555 | |
556 | /* close the epoch, in case it outgrew the limit */ | |
44ed167d PR |
557 | rcu_read_lock(); |
558 | nc = rcu_dereference(mdev->tconn->net_conf); | |
559 | p = nc->max_epoch_size; | |
560 | rcu_read_unlock(); | |
b6dd1a89 LE |
561 | if (mdev->tconn->current_tle_writes >= p) |
562 | start_new_tl_epoch(mdev->tconn); | |
b411b363 PR |
563 | |
564 | break; | |
565 | ||
8554df1c | 566 | case QUEUE_FOR_SEND_OOS: |
a0d856df | 567 | mod_rq_state(req, m, 0, RQ_NET_QUEUED); |
8f7bed77 | 568 | req->w.cb = w_send_out_of_sync; |
d5b27b01 | 569 | drbd_queue_work(&mdev->tconn->sender_work, &req->w); |
73a01a18 PR |
570 | break; |
571 | ||
ea9d6729 | 572 | case READ_RETRY_REMOTE_CANCELED: |
8554df1c | 573 | case SEND_CANCELED: |
8554df1c | 574 | case SEND_FAILED: |
b411b363 PR |
575 | /* real cleanup will be done from tl_clear. just update flags |
576 | * so it is no longer marked as on the worker queue */ | |
a0d856df | 577 | mod_rq_state(req, m, RQ_NET_QUEUED, 0); |
b411b363 PR |
578 | break; |
579 | ||
8554df1c | 580 | case HANDED_OVER_TO_NETWORK: |
b411b363 PR |
581 | /* assert something? */ |
582 | if (bio_data_dir(req->master_bio) == WRITE && | |
303d1448 | 583 | !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) { |
b411b363 PR |
584 | /* this is what is dangerous about protocol A: |
585 | * pretend it was successfully written on the peer. */ | |
a0d856df LE |
586 | if (req->rq_state & RQ_NET_PENDING) |
587 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); | |
588 | /* else: neg-ack was faster... */ | |
b411b363 PR |
589 | /* it is still not yet RQ_NET_DONE until the |
590 | * corresponding epoch barrier got acked as well, | |
591 | * so we know what to dirty on connection loss */ | |
592 | } | |
a0d856df | 593 | mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT); |
27a434fe LE |
594 | break; |
595 | ||
596 | case OOS_HANDED_TO_NETWORK: | |
597 | /* Was not set PENDING, no longer QUEUED, so is now DONE | |
598 | * as far as this connection is concerned. */ | |
a0d856df | 599 | mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE); |
b411b363 PR |
600 | break; |
601 | ||
8554df1c | 602 | case CONNECTION_LOST_WHILE_PENDING: |
b411b363 | 603 | /* transfer log cleanup after connection loss */ |
a0d856df LE |
604 | mod_rq_state(req, m, |
605 | RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP, | |
606 | RQ_NET_DONE); | |
b411b363 PR |
607 | break; |
608 | ||
d4dabbe2 LE |
609 | case CONFLICT_RESOLVED: |
610 | /* for superseded conflicting writes of multiple primaries, | |
b411b363 | 611 | * there is no need to keep anything in the tl, potential |
934722a2 LE |
612 | * node crashes are covered by the activity log. |
613 | * | |
614 | * If this request had been marked as RQ_POSTPONED before, | |
d4dabbe2 | 615 | * it will actually not be completed, but "restarted", |
934722a2 LE |
616 | * resubmitted from the retry worker context. */ |
617 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
618 | D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK); | |
619 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK); | |
620 | break; | |
621 | ||
0afd569a | 622 | case WRITE_ACKED_BY_PEER_AND_SIS: |
934722a2 | 623 | req->rq_state |= RQ_NET_SIS; |
8554df1c | 624 | case WRITE_ACKED_BY_PEER: |
303d1448 | 625 | D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK); |
b411b363 | 626 | /* protocol C; successfully written on peer. |
0afd569a | 627 | * Nothing more to do here. |
b411b363 | 628 | * We want to keep the tl in place for all protocols, to cater |
0afd569a | 629 | * for volatile write-back caches on lower level devices. */ |
b411b363 | 630 | |
303d1448 | 631 | goto ack_common; |
8554df1c | 632 | case RECV_ACKED_BY_PEER: |
303d1448 | 633 | D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK); |
b411b363 | 634 | /* protocol B; pretends to be successfully written on peer. |
8554df1c | 635 | * see also notes above in HANDED_OVER_TO_NETWORK about |
b411b363 | 636 | * protocol != C */ |
303d1448 | 637 | ack_common: |
b411b363 | 638 | D_ASSERT(req->rq_state & RQ_NET_PENDING); |
a0d856df | 639 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK); |
b411b363 PR |
640 | break; |
641 | ||
7be8da07 | 642 | case POSTPONE_WRITE: |
303d1448 PR |
643 | D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK); |
644 | /* If this node has already detected the write conflict, the | |
7be8da07 AG |
645 | * worker will be waiting on misc_wait. Wake it up once this |
646 | * request has completed locally. | |
647 | */ | |
648 | D_ASSERT(req->rq_state & RQ_NET_PENDING); | |
649 | req->rq_state |= RQ_POSTPONED; | |
a0d856df LE |
650 | if (req->i.waiting) |
651 | wake_up(&mdev->misc_wait); | |
652 | /* Do not clear RQ_NET_PENDING. This request will make further | |
653 | * progress via restart_conflicting_writes() or | |
654 | * fail_postponed_requests(). Hopefully. */ | |
7be8da07 AG |
655 | break; |
656 | ||
8554df1c | 657 | case NEG_ACKED: |
46e21bba | 658 | mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0); |
b411b363 PR |
659 | break; |
660 | ||
8554df1c | 661 | case FAIL_FROZEN_DISK_IO: |
265be2d0 PR |
662 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
663 | break; | |
a0d856df | 664 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); |
265be2d0 PR |
665 | break; |
666 | ||
8554df1c | 667 | case RESTART_FROZEN_DISK_IO: |
265be2d0 PR |
668 | if (!(req->rq_state & RQ_LOCAL_COMPLETED)) |
669 | break; | |
670 | ||
a0d856df LE |
671 | mod_rq_state(req, m, |
672 | RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED, | |
673 | RQ_LOCAL_PENDING); | |
265be2d0 PR |
674 | |
675 | rv = MR_READ; | |
676 | if (bio_data_dir(req->master_bio) == WRITE) | |
677 | rv = MR_WRITE; | |
678 | ||
a0d856df | 679 | get_ldev(mdev); /* always succeeds in this call path */ |
265be2d0 | 680 | req->w.cb = w_restart_disk_io; |
d5b27b01 | 681 | drbd_queue_work(&mdev->tconn->sender_work, &req->w); |
265be2d0 PR |
682 | break; |
683 | ||
8554df1c | 684 | case RESEND: |
8a0bab2a PR |
685 | /* Simply complete (local only) READs. */ |
686 | if (!(req->rq_state & RQ_WRITE) && !req->w.cb) { | |
687 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0); | |
688 | break; | |
689 | } | |
690 | ||
11b58e73 | 691 | /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK |
a0d856df LE |
692 | before the connection loss (B&C only); only P_BARRIER_ACK |
693 | (or the local completion?) was missing when we suspended. | |
6870ca6d LE |
694 | Throwing them out of the TL here by pretending we got a BARRIER_ACK. |
695 | During connection handshake, we ensure that the peer was not rebooted. */ | |
11b58e73 | 696 | if (!(req->rq_state & RQ_NET_OK)) { |
a0d856df LE |
697 | /* FIXME could this possibly be a req->w.cb == w_send_out_of_sync? |
698 | * in that case we must not set RQ_NET_PENDING. */ | |
699 | ||
700 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING); | |
11b58e73 | 701 | if (req->w.cb) { |
d5b27b01 | 702 | drbd_queue_work(&mdev->tconn->sender_work, &req->w); |
11b58e73 | 703 | rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ; |
a0d856df | 704 | } /* else: FIXME can this happen? */ |
11b58e73 PR |
705 | break; |
706 | } | |
8554df1c | 707 | /* else, fall through to BARRIER_ACKED */ |
11b58e73 | 708 | |
8554df1c | 709 | case BARRIER_ACKED: |
a0d856df | 710 | /* barrier ack for READ requests does not make sense */ |
288f422e PR |
711 | if (!(req->rq_state & RQ_WRITE)) |
712 | break; | |
713 | ||
b411b363 | 714 | if (req->rq_state & RQ_NET_PENDING) { |
a209b4ae | 715 | /* barrier came in before all requests were acked. |
b411b363 PR |
716 | * this is bad, because if the connection is lost now, |
717 | * we won't be able to clean them up... */ | |
8554df1c | 718 | dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n"); |
b411b363 | 719 | } |
a0d856df LE |
720 | /* Allowed to complete requests, even while suspended. |
721 | * As this is called for all requests within a matching epoch, | |
722 | * we need to filter, and only set RQ_NET_DONE for those that | |
723 | * have actually been on the wire. */ | |
724 | mod_rq_state(req, m, RQ_COMPLETION_SUSP, | |
725 | (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0); | |
b411b363 PR |
726 | break; |
727 | ||
8554df1c | 728 | case DATA_RECEIVED: |
b411b363 | 729 | D_ASSERT(req->rq_state & RQ_NET_PENDING); |
a0d856df | 730 | mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE); |
b411b363 PR |
731 | break; |
732 | }; | |
2a80699f PR |
733 | |
734 | return rv; | |
b411b363 PR |
735 | } |
736 | ||
737 | /* we may do a local read if: | |
738 | * - we are consistent (of course), | |
739 | * - or we are generally inconsistent, | |
740 | * BUT we are still/already IN SYNC for this area. | |
741 | * since size may be bigger than BM_BLOCK_SIZE, | |
742 | * we may need to check several bits. | |
743 | */ | |
0da34df0 | 744 | static bool drbd_may_do_local_read(struct drbd_conf *mdev, sector_t sector, int size) |
b411b363 PR |
745 | { |
746 | unsigned long sbnr, ebnr; | |
747 | sector_t esector, nr_sectors; | |
748 | ||
749 | if (mdev->state.disk == D_UP_TO_DATE) | |
0da34df0 | 750 | return true; |
8c387def | 751 | if (mdev->state.disk != D_INCONSISTENT) |
0da34df0 | 752 | return false; |
b411b363 | 753 | esector = sector + (size >> 9) - 1; |
8ca9844f | 754 | nr_sectors = drbd_get_capacity(mdev->this_bdev); |
b411b363 PR |
755 | D_ASSERT(sector < nr_sectors); |
756 | D_ASSERT(esector < nr_sectors); | |
757 | ||
758 | sbnr = BM_SECT_TO_BIT(sector); | |
759 | ebnr = BM_SECT_TO_BIT(esector); | |
760 | ||
0da34df0 | 761 | return drbd_bm_count_bits(mdev, sbnr, ebnr) == 0; |
b411b363 PR |
762 | } |
763 | ||
5da9c836 LE |
764 | static bool remote_due_to_read_balancing(struct drbd_conf *mdev, sector_t sector, |
765 | enum drbd_read_balancing rbm) | |
380207d0 | 766 | { |
380207d0 | 767 | struct backing_dev_info *bdi; |
d60de03a | 768 | int stripe_shift; |
380207d0 | 769 | |
380207d0 PR |
770 | switch (rbm) { |
771 | case RB_CONGESTED_REMOTE: | |
772 | bdi = &mdev->ldev->backing_bdev->bd_disk->queue->backing_dev_info; | |
773 | return bdi_read_congested(bdi); | |
774 | case RB_LEAST_PENDING: | |
775 | return atomic_read(&mdev->local_cnt) > | |
776 | atomic_read(&mdev->ap_pending_cnt) + atomic_read(&mdev->rs_pending_cnt); | |
d60de03a PR |
777 | case RB_32K_STRIPING: /* stripe_shift = 15 */ |
778 | case RB_64K_STRIPING: | |
779 | case RB_128K_STRIPING: | |
780 | case RB_256K_STRIPING: | |
781 | case RB_512K_STRIPING: | |
782 | case RB_1M_STRIPING: /* stripe_shift = 20 */ | |
783 | stripe_shift = (rbm - RB_32K_STRIPING + 15); | |
784 | return (sector >> (stripe_shift - 9)) & 1; | |
380207d0 PR |
785 | case RB_ROUND_ROBIN: |
786 | return test_and_change_bit(READ_BALANCE_RR, &mdev->flags); | |
787 | case RB_PREFER_REMOTE: | |
788 | return true; | |
789 | case RB_PREFER_LOCAL: | |
790 | default: | |
791 | return false; | |
792 | } | |
793 | } | |
794 | ||
6024fece AG |
795 | /* |
796 | * complete_conflicting_writes - wait for any conflicting write requests | |
797 | * | |
798 | * The write_requests tree contains all active write requests which we | |
799 | * currently know about. Wait for any requests to complete which conflict with | |
800 | * the new one. | |
648e46b5 LE |
801 | * |
802 | * Only way out: remove the conflicting intervals from the tree. | |
6024fece | 803 | */ |
648e46b5 | 804 | static void complete_conflicting_writes(struct drbd_request *req) |
6024fece | 805 | { |
648e46b5 LE |
806 | DEFINE_WAIT(wait); |
807 | struct drbd_conf *mdev = req->w.mdev; | |
808 | struct drbd_interval *i; | |
809 | sector_t sector = req->i.sector; | |
810 | int size = req->i.size; | |
811 | ||
812 | i = drbd_find_overlap(&mdev->write_requests, sector, size); | |
813 | if (!i) | |
814 | return; | |
6024fece | 815 | |
648e46b5 LE |
816 | for (;;) { |
817 | prepare_to_wait(&mdev->misc_wait, &wait, TASK_UNINTERRUPTIBLE); | |
6024fece AG |
818 | i = drbd_find_overlap(&mdev->write_requests, sector, size); |
819 | if (!i) | |
648e46b5 LE |
820 | break; |
821 | /* Indicate to wake up device->misc_wait on progress. */ | |
822 | i->waiting = true; | |
823 | spin_unlock_irq(&mdev->tconn->req_lock); | |
824 | schedule(); | |
825 | spin_lock_irq(&mdev->tconn->req_lock); | |
6024fece | 826 | } |
648e46b5 | 827 | finish_wait(&mdev->misc_wait, &wait); |
6024fece AG |
828 | } |
829 | ||
5da9c836 | 830 | /* called within req_lock and rcu_read_lock() */ |
3b9ef85e | 831 | static void maybe_pull_ahead(struct drbd_conf *mdev) |
5da9c836 LE |
832 | { |
833 | struct drbd_tconn *tconn = mdev->tconn; | |
834 | struct net_conf *nc; | |
835 | bool congested = false; | |
836 | enum drbd_on_congestion on_congestion; | |
837 | ||
838 | nc = rcu_dereference(tconn->net_conf); | |
839 | on_congestion = nc ? nc->on_congestion : OC_BLOCK; | |
840 | if (on_congestion == OC_BLOCK || | |
841 | tconn->agreed_pro_version < 96) | |
3b9ef85e LE |
842 | return; |
843 | ||
844 | /* If I don't even have good local storage, we can not reasonably try | |
845 | * to pull ahead of the peer. We also need the local reference to make | |
846 | * sure mdev->act_log is there. | |
847 | */ | |
848 | if (!get_ldev_if_state(mdev, D_UP_TO_DATE)) | |
849 | return; | |
5da9c836 LE |
850 | |
851 | if (nc->cong_fill && | |
852 | atomic_read(&mdev->ap_in_flight) >= nc->cong_fill) { | |
853 | dev_info(DEV, "Congestion-fill threshold reached\n"); | |
854 | congested = true; | |
855 | } | |
856 | ||
857 | if (mdev->act_log->used >= nc->cong_extents) { | |
858 | dev_info(DEV, "Congestion-extents threshold reached\n"); | |
859 | congested = true; | |
860 | } | |
861 | ||
862 | if (congested) { | |
99b4d8fe LE |
863 | /* start a new epoch for non-mirrored writes */ |
864 | start_new_tl_epoch(mdev->tconn); | |
5da9c836 LE |
865 | |
866 | if (on_congestion == OC_PULL_AHEAD) | |
867 | _drbd_set_state(_NS(mdev, conn, C_AHEAD), 0, NULL); | |
868 | else /*nc->on_congestion == OC_DISCONNECT */ | |
869 | _drbd_set_state(_NS(mdev, conn, C_DISCONNECTING), 0, NULL); | |
870 | } | |
3b9ef85e | 871 | put_ldev(mdev); |
5da9c836 LE |
872 | } |
873 | ||
874 | /* If this returns false, and req->private_bio is still set, | |
875 | * this should be submitted locally. | |
876 | * | |
877 | * If it returns false, but req->private_bio is not set, | |
878 | * we do not have access to good data :( | |
879 | * | |
880 | * Otherwise, this destroys req->private_bio, if any, | |
881 | * and returns true. | |
882 | */ | |
883 | static bool do_remote_read(struct drbd_request *req) | |
884 | { | |
885 | struct drbd_conf *mdev = req->w.mdev; | |
886 | enum drbd_read_balancing rbm; | |
887 | ||
888 | if (req->private_bio) { | |
889 | if (!drbd_may_do_local_read(mdev, | |
890 | req->i.sector, req->i.size)) { | |
891 | bio_put(req->private_bio); | |
892 | req->private_bio = NULL; | |
893 | put_ldev(mdev); | |
894 | } | |
895 | } | |
896 | ||
897 | if (mdev->state.pdsk != D_UP_TO_DATE) | |
898 | return false; | |
899 | ||
a0d856df LE |
900 | if (req->private_bio == NULL) |
901 | return true; | |
902 | ||
5da9c836 LE |
903 | /* TODO: improve read balancing decisions, take into account drbd |
904 | * protocol, pending requests etc. */ | |
905 | ||
906 | rcu_read_lock(); | |
907 | rbm = rcu_dereference(mdev->ldev->disk_conf)->read_balancing; | |
908 | rcu_read_unlock(); | |
909 | ||
910 | if (rbm == RB_PREFER_LOCAL && req->private_bio) | |
911 | return false; /* submit locally */ | |
912 | ||
5da9c836 LE |
913 | if (remote_due_to_read_balancing(mdev, req->i.sector, rbm)) { |
914 | if (req->private_bio) { | |
915 | bio_put(req->private_bio); | |
916 | req->private_bio = NULL; | |
917 | put_ldev(mdev); | |
918 | } | |
919 | return true; | |
920 | } | |
921 | ||
922 | return false; | |
923 | } | |
924 | ||
925 | /* returns number of connections (== 1, for drbd 8.4) | |
926 | * expected to actually write this data, | |
927 | * which does NOT include those that we are L_AHEAD for. */ | |
928 | static int drbd_process_write_request(struct drbd_request *req) | |
929 | { | |
930 | struct drbd_conf *mdev = req->w.mdev; | |
931 | int remote, send_oos; | |
932 | ||
933 | rcu_read_lock(); | |
934 | remote = drbd_should_do_remote(mdev->state); | |
935 | if (remote) { | |
3b9ef85e | 936 | maybe_pull_ahead(mdev); |
5da9c836 LE |
937 | remote = drbd_should_do_remote(mdev->state); |
938 | } | |
939 | send_oos = drbd_should_send_out_of_sync(mdev->state); | |
940 | rcu_read_unlock(); | |
941 | ||
519b6d3e LE |
942 | /* Need to replicate writes. Unless it is an empty flush, |
943 | * which is better mapped to a DRBD P_BARRIER packet, | |
944 | * also for drbd wire protocol compatibility reasons. | |
945 | * If this was a flush, just start a new epoch. | |
946 | * Unless the current epoch was empty anyways, or we are not currently | |
947 | * replicating, in which case there is no point. */ | |
948 | if (unlikely(req->i.size == 0)) { | |
949 | /* The only size==0 bios we expect are empty flushes. */ | |
950 | D_ASSERT(req->master_bio->bi_rw & REQ_FLUSH); | |
99b4d8fe | 951 | if (remote) |
519b6d3e LE |
952 | start_new_tl_epoch(mdev->tconn); |
953 | return 0; | |
954 | } | |
955 | ||
5da9c836 LE |
956 | if (!remote && !send_oos) |
957 | return 0; | |
958 | ||
959 | D_ASSERT(!(remote && send_oos)); | |
960 | ||
961 | if (remote) { | |
962 | _req_mod(req, TO_BE_SENT); | |
963 | _req_mod(req, QUEUE_FOR_NET_WRITE); | |
964 | } else if (drbd_set_out_of_sync(mdev, req->i.sector, req->i.size)) | |
965 | _req_mod(req, QUEUE_FOR_SEND_OOS); | |
966 | ||
967 | return remote; | |
968 | } | |
969 | ||
970 | static void | |
971 | drbd_submit_req_private_bio(struct drbd_request *req) | |
972 | { | |
973 | struct drbd_conf *mdev = req->w.mdev; | |
974 | struct bio *bio = req->private_bio; | |
975 | const int rw = bio_rw(bio); | |
976 | ||
977 | bio->bi_bdev = mdev->ldev->backing_bdev; | |
978 | ||
979 | /* State may have changed since we grabbed our reference on the | |
980 | * ->ldev member. Double check, and short-circuit to endio. | |
981 | * In case the last activity log transaction failed to get on | |
982 | * stable storage, and this is a WRITE, we may not even submit | |
983 | * this bio. */ | |
984 | if (get_ldev(mdev)) { | |
985 | if (drbd_insert_fault(mdev, | |
986 | rw == WRITE ? DRBD_FAULT_DT_WR | |
987 | : rw == READ ? DRBD_FAULT_DT_RD | |
988 | : DRBD_FAULT_DT_RA)) | |
989 | bio_endio(bio, -EIO); | |
990 | else | |
991 | generic_make_request(bio); | |
992 | put_ldev(mdev); | |
993 | } else | |
994 | bio_endio(bio, -EIO); | |
995 | } | |
996 | ||
5df69ece | 997 | void __drbd_make_request(struct drbd_conf *mdev, struct bio *bio, unsigned long start_time) |
b411b363 PR |
998 | { |
999 | const int rw = bio_rw(bio); | |
5da9c836 | 1000 | struct bio_and_error m = { NULL, }; |
b411b363 | 1001 | struct drbd_request *req; |
5da9c836 | 1002 | bool no_remote = false; |
b411b363 PR |
1003 | |
1004 | /* allocate outside of all locks; */ | |
1005 | req = drbd_req_new(mdev, bio); | |
1006 | if (!req) { | |
1007 | dec_ap_bio(mdev); | |
1008 | /* only pass the error to the upper layers. | |
1009 | * if user cannot handle io errors, that's not our business. */ | |
1010 | dev_err(DEV, "could not kmalloc() req\n"); | |
1011 | bio_endio(bio, -ENOMEM); | |
5df69ece | 1012 | return; |
b411b363 | 1013 | } |
aeda1cd6 | 1014 | req->start_time = start_time; |
b411b363 | 1015 | |
5da9c836 LE |
1016 | if (!get_ldev(mdev)) { |
1017 | bio_put(req->private_bio); | |
b411b363 PR |
1018 | req->private_bio = NULL; |
1019 | } | |
b411b363 PR |
1020 | |
1021 | /* For WRITES going to the local disk, grab a reference on the target | |
1022 | * extent. This waits for any resync activity in the corresponding | |
1023 | * resync extent to finish, and, if necessary, pulls in the target | |
1024 | * extent into the activity log, which involves further disk io because | |
519b6d3e LE |
1025 | * of transactional on-disk meta data updates. |
1026 | * Empty flushes don't need to go into the activity log, they can only | |
1027 | * flush data for pending writes which are already in there. */ | |
1028 | if (rw == WRITE && req->private_bio && req->i.size | |
5da9c836 | 1029 | && !test_bit(AL_SUSPENDED, &mdev->flags)) { |
0778286a | 1030 | req->rq_state |= RQ_IN_ACT_LOG; |
181286ad | 1031 | drbd_al_begin_io(mdev, &req->i); |
0778286a | 1032 | } |
b411b363 | 1033 | |
87eeee41 | 1034 | spin_lock_irq(&mdev->tconn->req_lock); |
6024fece | 1035 | if (rw == WRITE) { |
648e46b5 LE |
1036 | /* This may temporarily give up the req_lock, |
1037 | * but will re-aquire it before it returns here. | |
1038 | * Needs to be before the check on drbd_suspended() */ | |
1039 | complete_conflicting_writes(req); | |
6024fece AG |
1040 | } |
1041 | ||
5da9c836 | 1042 | /* no more giving up req_lock from now on! */ |
73a01a18 | 1043 | |
5da9c836 LE |
1044 | if (drbd_suspended(mdev)) { |
1045 | /* push back and retry: */ | |
1046 | req->rq_state |= RQ_POSTPONED; | |
1047 | if (req->private_bio) { | |
1048 | bio_put(req->private_bio); | |
1049 | req->private_bio = NULL; | |
d7644018 | 1050 | put_ldev(mdev); |
b411b363 | 1051 | } |
5da9c836 | 1052 | goto out; |
b411b363 PR |
1053 | } |
1054 | ||
b411b363 PR |
1055 | /* Update disk stats */ |
1056 | _drbd_start_io_acct(mdev, req, bio); | |
1057 | ||
5da9c836 LE |
1058 | /* We fail READ/READA early, if we can not serve it. |
1059 | * We must do this before req is registered on any lists. | |
a0d856df | 1060 | * Otherwise, drbd_req_complete() will queue failed READ for retry. */ |
5da9c836 LE |
1061 | if (rw != WRITE) { |
1062 | if (!do_remote_read(req) && !req->private_bio) | |
1063 | goto nodata; | |
1064 | } | |
b411b363 | 1065 | |
b6dd1a89 LE |
1066 | /* which transfer log epoch does this belong to? */ |
1067 | req->epoch = atomic_read(&mdev->tconn->current_tle_nr); | |
b6dd1a89 | 1068 | |
519b6d3e LE |
1069 | /* no point in adding empty flushes to the transfer log, |
1070 | * they are mapped to drbd barriers already. */ | |
99b4d8fe LE |
1071 | if (likely(req->i.size!=0)) { |
1072 | if (rw == WRITE) | |
1073 | mdev->tconn->current_tle_writes++; | |
1074 | ||
519b6d3e | 1075 | list_add_tail(&req->tl_requests, &mdev->tconn->transfer_log); |
99b4d8fe | 1076 | } |
288f422e | 1077 | |
5da9c836 LE |
1078 | if (rw == WRITE) { |
1079 | if (!drbd_process_write_request(req)) | |
1080 | no_remote = true; | |
1081 | } else { | |
1082 | /* We either have a private_bio, or we can read from remote. | |
1083 | * Otherwise we had done the goto nodata above. */ | |
1084 | if (req->private_bio == NULL) { | |
1085 | _req_mod(req, TO_BE_SENT); | |
1086 | _req_mod(req, QUEUE_FOR_NET_READ); | |
1087 | } else | |
1088 | no_remote = true; | |
b411b363 | 1089 | } |
67531718 | 1090 | |
5da9c836 LE |
1091 | if (req->private_bio) { |
1092 | /* needs to be marked within the same spinlock */ | |
1093 | _req_mod(req, TO_BE_SUBMITTED); | |
1094 | /* but we need to give up the spinlock to submit */ | |
1095 | spin_unlock_irq(&mdev->tconn->req_lock); | |
1096 | drbd_submit_req_private_bio(req); | |
a0d856df | 1097 | spin_lock_irq(&mdev->tconn->req_lock); |
5da9c836 LE |
1098 | } else if (no_remote) { |
1099 | nodata: | |
1100 | if (__ratelimit(&drbd_ratelimit_state)) | |
1101 | dev_err(DEV, "IO ERROR: neither local nor remote disk\n"); | |
1102 | /* A write may have been queued for send_oos, however. | |
a0d856df | 1103 | * So we can not simply free it, we must go through drbd_req_put_completion_ref() */ |
67531718 PR |
1104 | } |
1105 | ||
5da9c836 | 1106 | out: |
a0d856df LE |
1107 | if (drbd_req_put_completion_ref(req, &m, 1)) |
1108 | kref_put(&req->kref, drbd_req_destroy); | |
87eeee41 | 1109 | spin_unlock_irq(&mdev->tconn->req_lock); |
b411b363 | 1110 | |
5da9c836 LE |
1111 | if (m.bio) |
1112 | complete_master_bio(mdev, &m); | |
5df69ece | 1113 | return; |
b411b363 PR |
1114 | } |
1115 | ||
2f58dcfc | 1116 | int drbd_make_request(struct request_queue *q, struct bio *bio) |
b411b363 | 1117 | { |
b411b363 | 1118 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; |
aeda1cd6 | 1119 | unsigned long start_time; |
b411b363 | 1120 | |
aeda1cd6 PR |
1121 | start_time = jiffies; |
1122 | ||
b411b363 PR |
1123 | /* |
1124 | * what we "blindly" assume: | |
1125 | */ | |
c670a398 | 1126 | D_ASSERT(IS_ALIGNED(bio->bi_size, 512)); |
b411b363 | 1127 | |
5df69ece LE |
1128 | inc_ap_bio(mdev); |
1129 | __drbd_make_request(mdev, bio, start_time); | |
69b6a3b1 PR |
1130 | |
1131 | return 0; | |
b411b363 PR |
1132 | } |
1133 | ||
23361cf3 LE |
1134 | /* This is called by bio_add_page(). |
1135 | * | |
1136 | * q->max_hw_sectors and other global limits are already enforced there. | |
b411b363 | 1137 | * |
23361cf3 LE |
1138 | * We need to call down to our lower level device, |
1139 | * in case it has special restrictions. | |
1140 | * | |
1141 | * We also may need to enforce configured max-bio-bvecs limits. | |
b411b363 PR |
1142 | * |
1143 | * As long as the BIO is empty we have to allow at least one bvec, | |
23361cf3 | 1144 | * regardless of size and offset, so no need to ask lower levels. |
b411b363 PR |
1145 | */ |
1146 | int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec) | |
1147 | { | |
1148 | struct drbd_conf *mdev = (struct drbd_conf *) q->queuedata; | |
b411b363 | 1149 | unsigned int bio_size = bvm->bi_size; |
23361cf3 LE |
1150 | int limit = DRBD_MAX_BIO_SIZE; |
1151 | int backing_limit; | |
1152 | ||
1153 | if (bio_size && get_ldev(mdev)) { | |
b411b363 PR |
1154 | struct request_queue * const b = |
1155 | mdev->ldev->backing_bdev->bd_disk->queue; | |
a1c88d0d | 1156 | if (b->merge_bvec_fn) { |
b411b363 PR |
1157 | backing_limit = b->merge_bvec_fn(b, bvm, bvec); |
1158 | limit = min(limit, backing_limit); | |
1159 | } | |
1160 | put_ldev(mdev); | |
1161 | } | |
1162 | return limit; | |
1163 | } | |
7fde2be9 | 1164 | |
b6dd1a89 LE |
1165 | struct drbd_request *find_oldest_request(struct drbd_tconn *tconn) |
1166 | { | |
1167 | /* Walk the transfer log, | |
1168 | * and find the oldest not yet completed request */ | |
1169 | struct drbd_request *r; | |
1170 | list_for_each_entry(r, &tconn->transfer_log, tl_requests) { | |
b406777e | 1171 | if (atomic_read(&r->completion_ref)) |
b6dd1a89 LE |
1172 | return r; |
1173 | } | |
1174 | return NULL; | |
1175 | } | |
1176 | ||
7fde2be9 PR |
1177 | void request_timer_fn(unsigned long data) |
1178 | { | |
1179 | struct drbd_conf *mdev = (struct drbd_conf *) data; | |
8b924f1d | 1180 | struct drbd_tconn *tconn = mdev->tconn; |
7fde2be9 | 1181 | struct drbd_request *req; /* oldest request */ |
44ed167d | 1182 | struct net_conf *nc; |
3b03ad59 | 1183 | unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */ |
07be15b1 | 1184 | unsigned long now; |
44ed167d PR |
1185 | |
1186 | rcu_read_lock(); | |
1187 | nc = rcu_dereference(tconn->net_conf); | |
07be15b1 LE |
1188 | if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS) |
1189 | ent = nc->timeout * HZ/10 * nc->ko_count; | |
cdfda633 | 1190 | |
07be15b1 | 1191 | if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */ |
cdfda633 PR |
1192 | dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10; |
1193 | put_ldev(mdev); | |
1194 | } | |
44ed167d | 1195 | rcu_read_unlock(); |
7fde2be9 | 1196 | |
cdfda633 PR |
1197 | et = min_not_zero(dt, ent); |
1198 | ||
07be15b1 | 1199 | if (!et) |
7fde2be9 PR |
1200 | return; /* Recurring timer stopped */ |
1201 | ||
07be15b1 LE |
1202 | now = jiffies; |
1203 | ||
8b924f1d | 1204 | spin_lock_irq(&tconn->req_lock); |
b6dd1a89 LE |
1205 | req = find_oldest_request(tconn); |
1206 | if (!req) { | |
8b924f1d | 1207 | spin_unlock_irq(&tconn->req_lock); |
07be15b1 | 1208 | mod_timer(&mdev->request_timer, now + et); |
7fde2be9 PR |
1209 | return; |
1210 | } | |
1211 | ||
07be15b1 LE |
1212 | /* The request is considered timed out, if |
1213 | * - we have some effective timeout from the configuration, | |
1214 | * with above state restrictions applied, | |
1215 | * - the oldest request is waiting for a response from the network | |
1216 | * resp. the local disk, | |
1217 | * - the oldest request is in fact older than the effective timeout, | |
1218 | * - the connection was established (resp. disk was attached) | |
1219 | * for longer than the timeout already. | |
1220 | * Note that for 32bit jiffies and very stable connections/disks, | |
1221 | * we may have a wrap around, which is catched by | |
1222 | * !time_in_range(now, last_..._jif, last_..._jif + timeout). | |
1223 | * | |
1224 | * Side effect: once per 32bit wrap-around interval, which means every | |
1225 | * ~198 days with 250 HZ, we have a window where the timeout would need | |
1226 | * to expire twice (worst case) to become effective. Good enough. | |
1227 | */ | |
1228 | if (ent && req->rq_state & RQ_NET_PENDING && | |
1229 | time_after(now, req->start_time + ent) && | |
1230 | !time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) { | |
1231 | dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n"); | |
1232 | _drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL); | |
cdfda633 | 1233 | } |
07be15b1 LE |
1234 | if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev && |
1235 | time_after(now, req->start_time + dt) && | |
1236 | !time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) { | |
1237 | dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n"); | |
0c849666 | 1238 | __drbd_chk_io_error(mdev, DRBD_FORCE_DETACH); |
7fde2be9 | 1239 | } |
07be15b1 | 1240 | nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et; |
8b924f1d | 1241 | spin_unlock_irq(&tconn->req_lock); |
3b03ad59 | 1242 | mod_timer(&mdev->request_timer, nt); |
7fde2be9 | 1243 | } |