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
26 #include <linux/module.h>
27
28 #include <linux/slab.h>
29 #include <linux/drbd.h>
30 #include "drbd_int.h"
31 #include "drbd_req.h"
32
33
34 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
35
36 /* Update disk stats at start of I/O request */
_drbd_start_io_acct(struct drbd_device * device,struct drbd_request * req)37 static void _drbd_start_io_acct(struct drbd_device *device, struct drbd_request *req)
38 {
39 generic_start_io_acct(bio_data_dir(req->master_bio), req->i.size >> 9,
40 &device->vdisk->part0);
41 }
42
43 /* Update disk stats when completing request upwards */
_drbd_end_io_acct(struct drbd_device * device,struct drbd_request * req)44 static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
45 {
46 generic_end_io_acct(bio_data_dir(req->master_bio),
47 &device->vdisk->part0, req->start_jif);
48 }
49
drbd_req_new(struct drbd_device * device,struct bio * bio_src)50 static struct drbd_request *drbd_req_new(struct drbd_device *device,
51 struct bio *bio_src)
52 {
53 struct drbd_request *req;
54
55 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
56 if (!req)
57 return NULL;
58 memset(req, 0, sizeof(*req));
59
60 drbd_req_make_private_bio(req, bio_src);
61 req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
62 req->device = device;
63 req->master_bio = bio_src;
64 req->epoch = 0;
65
66 drbd_clear_interval(&req->i);
67 req->i.sector = bio_src->bi_iter.bi_sector;
68 req->i.size = bio_src->bi_iter.bi_size;
69 req->i.local = true;
70 req->i.waiting = false;
71
72 INIT_LIST_HEAD(&req->tl_requests);
73 INIT_LIST_HEAD(&req->w.list);
74 INIT_LIST_HEAD(&req->req_pending_master_completion);
75 INIT_LIST_HEAD(&req->req_pending_local);
76
77 /* one reference to be put by __drbd_make_request */
78 atomic_set(&req->completion_ref, 1);
79 /* one kref as long as completion_ref > 0 */
80 kref_init(&req->kref);
81 return req;
82 }
83
drbd_remove_request_interval(struct rb_root * root,struct drbd_request * req)84 static void drbd_remove_request_interval(struct rb_root *root,
85 struct drbd_request *req)
86 {
87 struct drbd_device *device = req->device;
88 struct drbd_interval *i = &req->i;
89
90 drbd_remove_interval(root, i);
91
92 /* Wake up any processes waiting for this request to complete. */
93 if (i->waiting)
94 wake_up(&device->misc_wait);
95 }
96
drbd_req_destroy(struct kref * kref)97 void drbd_req_destroy(struct kref *kref)
98 {
99 struct drbd_request *req = container_of(kref, struct drbd_request, kref);
100 struct drbd_device *device = req->device;
101 const unsigned s = req->rq_state;
102
103 if ((req->master_bio && !(s & RQ_POSTPONED)) ||
104 atomic_read(&req->completion_ref) ||
105 (s & RQ_LOCAL_PENDING) ||
106 ((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
107 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
108 s, atomic_read(&req->completion_ref));
109 return;
110 }
111
112 /* If called from mod_rq_state (expected normal case) or
113 * drbd_send_and_submit (the less likely normal path), this holds the
114 * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
115 * though it may be still empty (never added to the transfer log).
116 *
117 * If called from do_retry(), we do NOT hold the req_lock, but we are
118 * still allowed to unconditionally list_del(&req->tl_requests),
119 * because it will be on a local on-stack list only. */
120 list_del_init(&req->tl_requests);
121
122 /* finally remove the request from the conflict detection
123 * respective block_id verification interval tree. */
124 if (!drbd_interval_empty(&req->i)) {
125 struct rb_root *root;
126
127 if (s & RQ_WRITE)
128 root = &device->write_requests;
129 else
130 root = &device->read_requests;
131 drbd_remove_request_interval(root, req);
132 } else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
133 drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
134 s, (unsigned long long)req->i.sector, req->i.size);
135
136 /* if it was a write, we may have to set the corresponding
137 * bit(s) out-of-sync first. If it had a local part, we need to
138 * release the reference to the activity log. */
139 if (s & RQ_WRITE) {
140 /* Set out-of-sync unless both OK flags are set
141 * (local only or remote failed).
142 * Other places where we set out-of-sync:
143 * READ with local io-error */
144
145 /* There is a special case:
146 * we may notice late that IO was suspended,
147 * and postpone, or schedule for retry, a write,
148 * before it even was submitted or sent.
149 * In that case we do not want to touch the bitmap at all.
150 */
151 if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
152 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
153 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
154
155 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
156 drbd_set_in_sync(device, req->i.sector, req->i.size);
157 }
158
159 /* one might be tempted to move the drbd_al_complete_io
160 * to the local io completion callback drbd_request_endio.
161 * but, if this was a mirror write, we may only
162 * drbd_al_complete_io after this is RQ_NET_DONE,
163 * otherwise the extent could be dropped from the al
164 * before it has actually been written on the peer.
165 * if we crash before our peer knows about the request,
166 * but after the extent has been dropped from the al,
167 * we would forget to resync the corresponding extent.
168 */
169 if (s & RQ_IN_ACT_LOG) {
170 if (get_ldev_if_state(device, D_FAILED)) {
171 drbd_al_complete_io(device, &req->i);
172 put_ldev(device);
173 } else if (__ratelimit(&drbd_ratelimit_state)) {
174 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
175 "but my Disk seems to have failed :(\n",
176 (unsigned long long) req->i.sector, req->i.size);
177 }
178 }
179 }
180
181 mempool_free(req, drbd_request_mempool);
182 }
183
wake_all_senders(struct drbd_connection * connection)184 static void wake_all_senders(struct drbd_connection *connection)
185 {
186 wake_up(&connection->sender_work.q_wait);
187 }
188
189 /* must hold resource->req_lock */
start_new_tl_epoch(struct drbd_connection * connection)190 void start_new_tl_epoch(struct drbd_connection *connection)
191 {
192 /* no point closing an epoch, if it is empty, anyways. */
193 if (connection->current_tle_writes == 0)
194 return;
195
196 connection->current_tle_writes = 0;
197 atomic_inc(&connection->current_tle_nr);
198 wake_all_senders(connection);
199 }
200
complete_master_bio(struct drbd_device * device,struct bio_and_error * m)201 void complete_master_bio(struct drbd_device *device,
202 struct bio_and_error *m)
203 {
204 m->bio->bi_error = m->error;
205 bio_endio(m->bio);
206 dec_ap_bio(device);
207 }
208
209
210 /* Helper for __req_mod().
211 * Set m->bio to the master bio, if it is fit to be completed,
212 * or leave it alone (it is initialized to NULL in __req_mod),
213 * if it has already been completed, or cannot be completed yet.
214 * If m->bio is set, the error status to be returned is placed in m->error.
215 */
216 static
drbd_req_complete(struct drbd_request * req,struct bio_and_error * m)217 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
218 {
219 const unsigned s = req->rq_state;
220 struct drbd_device *device = req->device;
221 int rw;
222 int error, ok;
223
224 /* we must not complete the master bio, while it is
225 * still being processed by _drbd_send_zc_bio (drbd_send_dblock)
226 * not yet acknowledged by the peer
227 * not yet completed by the local io subsystem
228 * these flags may get cleared in any order by
229 * the worker,
230 * the receiver,
231 * the bio_endio completion callbacks.
232 */
233 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
234 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
235 (s & RQ_COMPLETION_SUSP)) {
236 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
237 return;
238 }
239
240 if (!req->master_bio) {
241 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
242 return;
243 }
244
245 rw = bio_rw(req->master_bio);
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
258 * in private_bio within drbd_request_endio.
259 */
260 ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
261 error = PTR_ERR(req->private_bio);
262
263 /* Before we can signal completion to the upper layers,
264 * we may need to close the current transfer log epoch.
265 * We are within the request lock, so we can simply compare
266 * the request epoch number with the current transfer log
267 * epoch number. If they match, increase the current_tle_nr,
268 * and reset the transfer log epoch write_cnt.
269 */
270 if (rw == WRITE &&
271 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
272 start_new_tl_epoch(first_peer_device(device)->connection);
273
274 /* Update disk stats */
275 _drbd_end_io_acct(device, req);
276
277 /* If READ failed,
278 * have it be pushed back to the retry work queue,
279 * so it will re-enter __drbd_make_request(),
280 * and be re-assigned to a suitable local or remote path,
281 * or failed if we do not have access to good data anymore.
282 *
283 * Unless it was failed early by __drbd_make_request(),
284 * because no path was available, in which case
285 * it was not even added to the transfer_log.
286 *
287 * READA may fail, and will not be retried.
288 *
289 * WRITE should have used all available paths already.
290 */
291 if (!ok && rw == READ && !list_empty(&req->tl_requests))
292 req->rq_state |= RQ_POSTPONED;
293
294 if (!(req->rq_state & RQ_POSTPONED)) {
295 m->error = ok ? 0 : (error ?: -EIO);
296 m->bio = req->master_bio;
297 req->master_bio = NULL;
298 /* We leave it in the tree, to be able to verify later
299 * write-acks in protocol != C during resync.
300 * But we mark it as "complete", so it won't be counted as
301 * conflict in a multi-primary setup. */
302 req->i.completed = true;
303 }
304
305 if (req->i.waiting)
306 wake_up(&device->misc_wait);
307
308 /* Either we are about to complete to upper layers,
309 * or we will restart this request.
310 * In either case, the request object will be destroyed soon,
311 * so better remove it from all lists. */
312 list_del_init(&req->req_pending_master_completion);
313 }
314
315 /* still holds resource->req_lock */
drbd_req_put_completion_ref(struct drbd_request * req,struct bio_and_error * m,int put)316 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
317 {
318 struct drbd_device *device = req->device;
319 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
320
321 if (!atomic_sub_and_test(put, &req->completion_ref))
322 return 0;
323
324 drbd_req_complete(req, m);
325
326 if (req->rq_state & RQ_POSTPONED) {
327 /* don't destroy the req object just yet,
328 * but queue it for retry */
329 drbd_restart_request(req);
330 return 0;
331 }
332
333 return 1;
334 }
335
set_if_null_req_next(struct drbd_peer_device * peer_device,struct drbd_request * req)336 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
337 {
338 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
339 if (!connection)
340 return;
341 if (connection->req_next == NULL)
342 connection->req_next = req;
343 }
344
advance_conn_req_next(struct drbd_peer_device * peer_device,struct drbd_request * req)345 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
346 {
347 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
348 if (!connection)
349 return;
350 if (connection->req_next != req)
351 return;
352 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
353 const unsigned s = req->rq_state;
354 if (s & RQ_NET_QUEUED)
355 break;
356 }
357 if (&req->tl_requests == &connection->transfer_log)
358 req = NULL;
359 connection->req_next = req;
360 }
361
set_if_null_req_ack_pending(struct drbd_peer_device * peer_device,struct drbd_request * req)362 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
363 {
364 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
365 if (!connection)
366 return;
367 if (connection->req_ack_pending == NULL)
368 connection->req_ack_pending = req;
369 }
370
advance_conn_req_ack_pending(struct drbd_peer_device * peer_device,struct drbd_request * req)371 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
372 {
373 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
374 if (!connection)
375 return;
376 if (connection->req_ack_pending != req)
377 return;
378 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
379 const unsigned s = req->rq_state;
380 if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
381 break;
382 }
383 if (&req->tl_requests == &connection->transfer_log)
384 req = NULL;
385 connection->req_ack_pending = req;
386 }
387
set_if_null_req_not_net_done(struct drbd_peer_device * peer_device,struct drbd_request * req)388 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
389 {
390 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
391 if (!connection)
392 return;
393 if (connection->req_not_net_done == NULL)
394 connection->req_not_net_done = req;
395 }
396
advance_conn_req_not_net_done(struct drbd_peer_device * peer_device,struct drbd_request * req)397 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
398 {
399 struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
400 if (!connection)
401 return;
402 if (connection->req_not_net_done != req)
403 return;
404 list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
405 const unsigned s = req->rq_state;
406 if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
407 break;
408 }
409 if (&req->tl_requests == &connection->transfer_log)
410 req = NULL;
411 connection->req_not_net_done = req;
412 }
413
414 /* I'd like this to be the only place that manipulates
415 * req->completion_ref and req->kref. */
mod_rq_state(struct drbd_request * req,struct bio_and_error * m,int clear,int set)416 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
417 int clear, int set)
418 {
419 struct drbd_device *device = req->device;
420 struct drbd_peer_device *peer_device = first_peer_device(device);
421 unsigned s = req->rq_state;
422 int c_put = 0;
423 int k_put = 0;
424
425 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
426 set |= RQ_COMPLETION_SUSP;
427
428 /* apply */
429
430 req->rq_state &= ~clear;
431 req->rq_state |= set;
432
433 /* no change? */
434 if (req->rq_state == s)
435 return;
436
437 /* intent: get references */
438
439 if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
440 atomic_inc(&req->completion_ref);
441
442 if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
443 inc_ap_pending(device);
444 atomic_inc(&req->completion_ref);
445 }
446
447 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
448 atomic_inc(&req->completion_ref);
449 set_if_null_req_next(peer_device, req);
450 }
451
452 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
453 kref_get(&req->kref); /* wait for the DONE */
454
455 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
456 /* potentially already completed in the asender thread */
457 if (!(s & RQ_NET_DONE)) {
458 atomic_add(req->i.size >> 9, &device->ap_in_flight);
459 set_if_null_req_not_net_done(peer_device, req);
460 }
461 if (s & RQ_NET_PENDING)
462 set_if_null_req_ack_pending(peer_device, req);
463 }
464
465 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
466 atomic_inc(&req->completion_ref);
467
468 /* progress: put references */
469
470 if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
471 ++c_put;
472
473 if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
474 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
475 /* local completion may still come in later,
476 * we need to keep the req object around. */
477 kref_get(&req->kref);
478 ++c_put;
479 }
480
481 if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
482 if (req->rq_state & RQ_LOCAL_ABORTED)
483 ++k_put;
484 else
485 ++c_put;
486 list_del_init(&req->req_pending_local);
487 }
488
489 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
490 dec_ap_pending(device);
491 ++c_put;
492 req->acked_jif = jiffies;
493 advance_conn_req_ack_pending(peer_device, req);
494 }
495
496 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
497 ++c_put;
498 advance_conn_req_next(peer_device, req);
499 }
500
501 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
502 if (s & RQ_NET_SENT)
503 atomic_sub(req->i.size >> 9, &device->ap_in_flight);
504 if (s & RQ_EXP_BARR_ACK)
505 ++k_put;
506 req->net_done_jif = jiffies;
507
508 /* in ahead/behind mode, or just in case,
509 * before we finally destroy this request,
510 * the caching pointers must not reference it anymore */
511 advance_conn_req_next(peer_device, req);
512 advance_conn_req_ack_pending(peer_device, req);
513 advance_conn_req_not_net_done(peer_device, req);
514 }
515
516 /* potentially complete and destroy */
517
518 if (k_put || c_put) {
519 /* Completion does it's own kref_put. If we are going to
520 * kref_sub below, we need req to be still around then. */
521 int at_least = k_put + !!c_put;
522 int refcount = atomic_read(&req->kref.refcount);
523 if (refcount < at_least)
524 drbd_err(device,
525 "mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
526 s, req->rq_state, refcount, at_least);
527 }
528
529 /* If we made progress, retry conflicting peer requests, if any. */
530 if (req->i.waiting)
531 wake_up(&device->misc_wait);
532
533 if (c_put)
534 k_put += drbd_req_put_completion_ref(req, m, c_put);
535 if (k_put)
536 kref_sub(&req->kref, k_put, drbd_req_destroy);
537 }
538
drbd_report_io_error(struct drbd_device * device,struct drbd_request * req)539 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
540 {
541 char b[BDEVNAME_SIZE];
542
543 if (!__ratelimit(&drbd_ratelimit_state))
544 return;
545
546 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
547 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
548 (unsigned long long)req->i.sector,
549 req->i.size >> 9,
550 bdevname(device->ldev->backing_bdev, b));
551 }
552
553 /* Helper for HANDED_OVER_TO_NETWORK.
554 * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
555 * Is it also still "PENDING"?
556 * --> If so, clear PENDING and set NET_OK below.
557 * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
558 * (and we must not set RQ_NET_OK) */
is_pending_write_protocol_A(struct drbd_request * req)559 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
560 {
561 return (req->rq_state &
562 (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
563 == (RQ_WRITE|RQ_NET_PENDING);
564 }
565
566 /* obviously this could be coded as many single functions
567 * instead of one huge switch,
568 * or by putting the code directly in the respective locations
569 * (as it has been before).
570 *
571 * but having it this way
572 * enforces that it is all in this one place, where it is easier to audit,
573 * it makes it obvious that whatever "event" "happens" to a request should
574 * happen "atomically" within the req_lock,
575 * and it enforces that we have to think in a very structured manner
576 * about the "events" that may happen to a request during its life time ...
577 */
__req_mod(struct drbd_request * req,enum drbd_req_event what,struct bio_and_error * m)578 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
579 struct bio_and_error *m)
580 {
581 struct drbd_device *const device = req->device;
582 struct drbd_peer_device *const peer_device = first_peer_device(device);
583 struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
584 struct net_conf *nc;
585 int p, rv = 0;
586
587 if (m)
588 m->bio = NULL;
589
590 switch (what) {
591 default:
592 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
593 break;
594
595 /* does not happen...
596 * initialization done in drbd_req_new
597 case CREATED:
598 break;
599 */
600
601 case TO_BE_SENT: /* via network */
602 /* reached via __drbd_make_request
603 * and from w_read_retry_remote */
604 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
605 rcu_read_lock();
606 nc = rcu_dereference(connection->net_conf);
607 p = nc->wire_protocol;
608 rcu_read_unlock();
609 req->rq_state |=
610 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
611 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
612 mod_rq_state(req, m, 0, RQ_NET_PENDING);
613 break;
614
615 case TO_BE_SUBMITTED: /* locally */
616 /* reached via __drbd_make_request */
617 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
618 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
619 break;
620
621 case COMPLETED_OK:
622 if (req->rq_state & RQ_WRITE)
623 device->writ_cnt += req->i.size >> 9;
624 else
625 device->read_cnt += req->i.size >> 9;
626
627 mod_rq_state(req, m, RQ_LOCAL_PENDING,
628 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
629 break;
630
631 case ABORT_DISK_IO:
632 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
633 break;
634
635 case WRITE_COMPLETED_WITH_ERROR:
636 drbd_report_io_error(device, req);
637 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
638 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
639 break;
640
641 case READ_COMPLETED_WITH_ERROR:
642 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
643 drbd_report_io_error(device, req);
644 __drbd_chk_io_error(device, DRBD_READ_ERROR);
645 /* fall through. */
646 case READ_AHEAD_COMPLETED_WITH_ERROR:
647 /* it is legal to fail READA, no __drbd_chk_io_error in that case. */
648 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
649 break;
650
651 case DISCARD_COMPLETED_NOTSUPP:
652 case DISCARD_COMPLETED_WITH_ERROR:
653 /* I'd rather not detach from local disk just because it
654 * failed a REQ_DISCARD. */
655 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
656 break;
657
658 case QUEUE_FOR_NET_READ:
659 /* READ or READA, and
660 * no local disk,
661 * or target area marked as invalid,
662 * or just got an io-error. */
663 /* from __drbd_make_request
664 * or from bio_endio during read io-error recovery */
665
666 /* So we can verify the handle in the answer packet.
667 * Corresponding drbd_remove_request_interval is in
668 * drbd_req_complete() */
669 D_ASSERT(device, drbd_interval_empty(&req->i));
670 drbd_insert_interval(&device->read_requests, &req->i);
671
672 set_bit(UNPLUG_REMOTE, &device->flags);
673
674 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
675 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
676 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
677 req->w.cb = w_send_read_req;
678 drbd_queue_work(&connection->sender_work,
679 &req->w);
680 break;
681
682 case QUEUE_FOR_NET_WRITE:
683 /* assert something? */
684 /* from __drbd_make_request only */
685
686 /* Corresponding drbd_remove_request_interval is in
687 * drbd_req_complete() */
688 D_ASSERT(device, drbd_interval_empty(&req->i));
689 drbd_insert_interval(&device->write_requests, &req->i);
690
691 /* NOTE
692 * In case the req ended up on the transfer log before being
693 * queued on the worker, it could lead to this request being
694 * missed during cleanup after connection loss.
695 * So we have to do both operations here,
696 * within the same lock that protects the transfer log.
697 *
698 * _req_add_to_epoch(req); this has to be after the
699 * _maybe_start_new_epoch(req); which happened in
700 * __drbd_make_request, because we now may set the bit
701 * again ourselves to close the current epoch.
702 *
703 * Add req to the (now) current epoch (barrier). */
704
705 /* otherwise we may lose an unplug, which may cause some remote
706 * io-scheduler timeout to expire, increasing maximum latency,
707 * hurting performance. */
708 set_bit(UNPLUG_REMOTE, &device->flags);
709
710 /* queue work item to send data */
711 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
712 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
713 req->w.cb = w_send_dblock;
714 drbd_queue_work(&connection->sender_work,
715 &req->w);
716
717 /* close the epoch, in case it outgrew the limit */
718 rcu_read_lock();
719 nc = rcu_dereference(connection->net_conf);
720 p = nc->max_epoch_size;
721 rcu_read_unlock();
722 if (connection->current_tle_writes >= p)
723 start_new_tl_epoch(connection);
724
725 break;
726
727 case QUEUE_FOR_SEND_OOS:
728 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
729 req->w.cb = w_send_out_of_sync;
730 drbd_queue_work(&connection->sender_work,
731 &req->w);
732 break;
733
734 case READ_RETRY_REMOTE_CANCELED:
735 case SEND_CANCELED:
736 case SEND_FAILED:
737 /* real cleanup will be done from tl_clear. just update flags
738 * so it is no longer marked as on the worker queue */
739 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
740 break;
741
742 case HANDED_OVER_TO_NETWORK:
743 /* assert something? */
744 if (is_pending_write_protocol_A(req))
745 /* this is what is dangerous about protocol A:
746 * pretend it was successfully written on the peer. */
747 mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
748 RQ_NET_SENT|RQ_NET_OK);
749 else
750 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
751 /* It is still not yet RQ_NET_DONE until the
752 * corresponding epoch barrier got acked as well,
753 * so we know what to dirty on connection loss. */
754 break;
755
756 case OOS_HANDED_TO_NETWORK:
757 /* Was not set PENDING, no longer QUEUED, so is now DONE
758 * as far as this connection is concerned. */
759 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
760 break;
761
762 case CONNECTION_LOST_WHILE_PENDING:
763 /* transfer log cleanup after connection loss */
764 mod_rq_state(req, m,
765 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
766 RQ_NET_DONE);
767 break;
768
769 case CONFLICT_RESOLVED:
770 /* for superseded conflicting writes of multiple primaries,
771 * there is no need to keep anything in the tl, potential
772 * node crashes are covered by the activity log.
773 *
774 * If this request had been marked as RQ_POSTPONED before,
775 * it will actually not be completed, but "restarted",
776 * resubmitted from the retry worker context. */
777 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
778 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
779 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
780 break;
781
782 case WRITE_ACKED_BY_PEER_AND_SIS:
783 req->rq_state |= RQ_NET_SIS;
784 case WRITE_ACKED_BY_PEER:
785 /* Normal operation protocol C: successfully written on peer.
786 * During resync, even in protocol != C,
787 * we requested an explicit write ack anyways.
788 * Which means we cannot even assert anything here.
789 * Nothing more to do here.
790 * We want to keep the tl in place for all protocols, to cater
791 * for volatile write-back caches on lower level devices. */
792 goto ack_common;
793 case RECV_ACKED_BY_PEER:
794 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
795 /* protocol B; pretends to be successfully written on peer.
796 * see also notes above in HANDED_OVER_TO_NETWORK about
797 * protocol != C */
798 ack_common:
799 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
800 break;
801
802 case POSTPONE_WRITE:
803 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
804 /* If this node has already detected the write conflict, the
805 * worker will be waiting on misc_wait. Wake it up once this
806 * request has completed locally.
807 */
808 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
809 req->rq_state |= RQ_POSTPONED;
810 if (req->i.waiting)
811 wake_up(&device->misc_wait);
812 /* Do not clear RQ_NET_PENDING. This request will make further
813 * progress via restart_conflicting_writes() or
814 * fail_postponed_requests(). Hopefully. */
815 break;
816
817 case NEG_ACKED:
818 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
819 break;
820
821 case FAIL_FROZEN_DISK_IO:
822 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
823 break;
824 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
825 break;
826
827 case RESTART_FROZEN_DISK_IO:
828 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
829 break;
830
831 mod_rq_state(req, m,
832 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
833 RQ_LOCAL_PENDING);
834
835 rv = MR_READ;
836 if (bio_data_dir(req->master_bio) == WRITE)
837 rv = MR_WRITE;
838
839 get_ldev(device); /* always succeeds in this call path */
840 req->w.cb = w_restart_disk_io;
841 drbd_queue_work(&connection->sender_work,
842 &req->w);
843 break;
844
845 case RESEND:
846 /* Simply complete (local only) READs. */
847 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
848 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
849 break;
850 }
851
852 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
853 before the connection loss (B&C only); only P_BARRIER_ACK
854 (or the local completion?) was missing when we suspended.
855 Throwing them out of the TL here by pretending we got a BARRIER_ACK.
856 During connection handshake, we ensure that the peer was not rebooted. */
857 if (!(req->rq_state & RQ_NET_OK)) {
858 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
859 * in that case we must not set RQ_NET_PENDING. */
860
861 mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
862 if (req->w.cb) {
863 /* w.cb expected to be w_send_dblock, or w_send_read_req */
864 drbd_queue_work(&connection->sender_work,
865 &req->w);
866 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
867 } /* else: FIXME can this happen? */
868 break;
869 }
870 /* else, fall through to BARRIER_ACKED */
871
872 case BARRIER_ACKED:
873 /* barrier ack for READ requests does not make sense */
874 if (!(req->rq_state & RQ_WRITE))
875 break;
876
877 if (req->rq_state & RQ_NET_PENDING) {
878 /* barrier came in before all requests were acked.
879 * this is bad, because if the connection is lost now,
880 * we won't be able to clean them up... */
881 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
882 }
883 /* Allowed to complete requests, even while suspended.
884 * As this is called for all requests within a matching epoch,
885 * we need to filter, and only set RQ_NET_DONE for those that
886 * have actually been on the wire. */
887 mod_rq_state(req, m, RQ_COMPLETION_SUSP,
888 (req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
889 break;
890
891 case DATA_RECEIVED:
892 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
893 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
894 break;
895
896 case QUEUE_AS_DRBD_BARRIER:
897 start_new_tl_epoch(connection);
898 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
899 break;
900 };
901
902 return rv;
903 }
904
905 /* we may do a local read if:
906 * - we are consistent (of course),
907 * - or we are generally inconsistent,
908 * BUT we are still/already IN SYNC for this area.
909 * since size may be bigger than BM_BLOCK_SIZE,
910 * we may need to check several bits.
911 */
drbd_may_do_local_read(struct drbd_device * device,sector_t sector,int size)912 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
913 {
914 unsigned long sbnr, ebnr;
915 sector_t esector, nr_sectors;
916
917 if (device->state.disk == D_UP_TO_DATE)
918 return true;
919 if (device->state.disk != D_INCONSISTENT)
920 return false;
921 esector = sector + (size >> 9) - 1;
922 nr_sectors = drbd_get_capacity(device->this_bdev);
923 D_ASSERT(device, sector < nr_sectors);
924 D_ASSERT(device, esector < nr_sectors);
925
926 sbnr = BM_SECT_TO_BIT(sector);
927 ebnr = BM_SECT_TO_BIT(esector);
928
929 return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
930 }
931
remote_due_to_read_balancing(struct drbd_device * device,sector_t sector,enum drbd_read_balancing rbm)932 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
933 enum drbd_read_balancing rbm)
934 {
935 struct backing_dev_info *bdi;
936 int stripe_shift;
937
938 switch (rbm) {
939 case RB_CONGESTED_REMOTE:
940 bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
941 return bdi_read_congested(bdi);
942 case RB_LEAST_PENDING:
943 return atomic_read(&device->local_cnt) >
944 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
945 case RB_32K_STRIPING: /* stripe_shift = 15 */
946 case RB_64K_STRIPING:
947 case RB_128K_STRIPING:
948 case RB_256K_STRIPING:
949 case RB_512K_STRIPING:
950 case RB_1M_STRIPING: /* stripe_shift = 20 */
951 stripe_shift = (rbm - RB_32K_STRIPING + 15);
952 return (sector >> (stripe_shift - 9)) & 1;
953 case RB_ROUND_ROBIN:
954 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
955 case RB_PREFER_REMOTE:
956 return true;
957 case RB_PREFER_LOCAL:
958 default:
959 return false;
960 }
961 }
962
963 /*
964 * complete_conflicting_writes - wait for any conflicting write requests
965 *
966 * The write_requests tree contains all active write requests which we
967 * currently know about. Wait for any requests to complete which conflict with
968 * the new one.
969 *
970 * Only way out: remove the conflicting intervals from the tree.
971 */
complete_conflicting_writes(struct drbd_request * req)972 static void complete_conflicting_writes(struct drbd_request *req)
973 {
974 DEFINE_WAIT(wait);
975 struct drbd_device *device = req->device;
976 struct drbd_interval *i;
977 sector_t sector = req->i.sector;
978 int size = req->i.size;
979
980 i = drbd_find_overlap(&device->write_requests, sector, size);
981 if (!i)
982 return;
983
984 for (;;) {
985 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
986 i = drbd_find_overlap(&device->write_requests, sector, size);
987 if (!i)
988 break;
989 /* Indicate to wake up device->misc_wait on progress. */
990 i->waiting = true;
991 spin_unlock_irq(&device->resource->req_lock);
992 schedule();
993 spin_lock_irq(&device->resource->req_lock);
994 }
995 finish_wait(&device->misc_wait, &wait);
996 }
997
998 /* called within req_lock and rcu_read_lock() */
maybe_pull_ahead(struct drbd_device * device)999 static void maybe_pull_ahead(struct drbd_device *device)
1000 {
1001 struct drbd_connection *connection = first_peer_device(device)->connection;
1002 struct net_conf *nc;
1003 bool congested = false;
1004 enum drbd_on_congestion on_congestion;
1005
1006 rcu_read_lock();
1007 nc = rcu_dereference(connection->net_conf);
1008 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
1009 rcu_read_unlock();
1010 if (on_congestion == OC_BLOCK ||
1011 connection->agreed_pro_version < 96)
1012 return;
1013
1014 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1015 return; /* nothing to do ... */
1016
1017 /* If I don't even have good local storage, we can not reasonably try
1018 * to pull ahead of the peer. We also need the local reference to make
1019 * sure device->act_log is there.
1020 */
1021 if (!get_ldev_if_state(device, D_UP_TO_DATE))
1022 return;
1023
1024 if (nc->cong_fill &&
1025 atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1026 drbd_info(device, "Congestion-fill threshold reached\n");
1027 congested = true;
1028 }
1029
1030 if (device->act_log->used >= nc->cong_extents) {
1031 drbd_info(device, "Congestion-extents threshold reached\n");
1032 congested = true;
1033 }
1034
1035 if (congested) {
1036 /* start a new epoch for non-mirrored writes */
1037 start_new_tl_epoch(first_peer_device(device)->connection);
1038
1039 if (on_congestion == OC_PULL_AHEAD)
1040 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1041 else /*nc->on_congestion == OC_DISCONNECT */
1042 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1043 }
1044 put_ldev(device);
1045 }
1046
1047 /* If this returns false, and req->private_bio is still set,
1048 * this should be submitted locally.
1049 *
1050 * If it returns false, but req->private_bio is not set,
1051 * we do not have access to good data :(
1052 *
1053 * Otherwise, this destroys req->private_bio, if any,
1054 * and returns true.
1055 */
do_remote_read(struct drbd_request * req)1056 static bool do_remote_read(struct drbd_request *req)
1057 {
1058 struct drbd_device *device = req->device;
1059 enum drbd_read_balancing rbm;
1060
1061 if (req->private_bio) {
1062 if (!drbd_may_do_local_read(device,
1063 req->i.sector, req->i.size)) {
1064 bio_put(req->private_bio);
1065 req->private_bio = NULL;
1066 put_ldev(device);
1067 }
1068 }
1069
1070 if (device->state.pdsk != D_UP_TO_DATE)
1071 return false;
1072
1073 if (req->private_bio == NULL)
1074 return true;
1075
1076 /* TODO: improve read balancing decisions, take into account drbd
1077 * protocol, pending requests etc. */
1078
1079 rcu_read_lock();
1080 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1081 rcu_read_unlock();
1082
1083 if (rbm == RB_PREFER_LOCAL && req->private_bio)
1084 return false; /* submit locally */
1085
1086 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1087 if (req->private_bio) {
1088 bio_put(req->private_bio);
1089 req->private_bio = NULL;
1090 put_ldev(device);
1091 }
1092 return true;
1093 }
1094
1095 return false;
1096 }
1097
1098 /* returns number of connections (== 1, for drbd 8.4)
1099 * expected to actually write this data,
1100 * which does NOT include those that we are L_AHEAD for. */
drbd_process_write_request(struct drbd_request * req)1101 static int drbd_process_write_request(struct drbd_request *req)
1102 {
1103 struct drbd_device *device = req->device;
1104 int remote, send_oos;
1105
1106 remote = drbd_should_do_remote(device->state);
1107 send_oos = drbd_should_send_out_of_sync(device->state);
1108
1109 /* Need to replicate writes. Unless it is an empty flush,
1110 * which is better mapped to a DRBD P_BARRIER packet,
1111 * also for drbd wire protocol compatibility reasons.
1112 * If this was a flush, just start a new epoch.
1113 * Unless the current epoch was empty anyways, or we are not currently
1114 * replicating, in which case there is no point. */
1115 if (unlikely(req->i.size == 0)) {
1116 /* The only size==0 bios we expect are empty flushes. */
1117 D_ASSERT(device, req->master_bio->bi_rw & REQ_FLUSH);
1118 if (remote)
1119 _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1120 return remote;
1121 }
1122
1123 if (!remote && !send_oos)
1124 return 0;
1125
1126 D_ASSERT(device, !(remote && send_oos));
1127
1128 if (remote) {
1129 _req_mod(req, TO_BE_SENT);
1130 _req_mod(req, QUEUE_FOR_NET_WRITE);
1131 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1132 _req_mod(req, QUEUE_FOR_SEND_OOS);
1133
1134 return remote;
1135 }
1136
1137 static void
drbd_submit_req_private_bio(struct drbd_request * req)1138 drbd_submit_req_private_bio(struct drbd_request *req)
1139 {
1140 struct drbd_device *device = req->device;
1141 struct bio *bio = req->private_bio;
1142 const int rw = bio_rw(bio);
1143
1144 bio->bi_bdev = device->ldev->backing_bdev;
1145
1146 /* State may have changed since we grabbed our reference on the
1147 * ->ldev member. Double check, and short-circuit to endio.
1148 * In case the last activity log transaction failed to get on
1149 * stable storage, and this is a WRITE, we may not even submit
1150 * this bio. */
1151 if (get_ldev(device)) {
1152 req->pre_submit_jif = jiffies;
1153 if (drbd_insert_fault(device,
1154 rw == WRITE ? DRBD_FAULT_DT_WR
1155 : rw == READ ? DRBD_FAULT_DT_RD
1156 : DRBD_FAULT_DT_RA))
1157 bio_io_error(bio);
1158 else
1159 generic_make_request(bio);
1160 put_ldev(device);
1161 } else
1162 bio_io_error(bio);
1163 }
1164
drbd_queue_write(struct drbd_device * device,struct drbd_request * req)1165 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1166 {
1167 spin_lock_irq(&device->resource->req_lock);
1168 list_add_tail(&req->tl_requests, &device->submit.writes);
1169 list_add_tail(&req->req_pending_master_completion,
1170 &device->pending_master_completion[1 /* WRITE */]);
1171 spin_unlock_irq(&device->resource->req_lock);
1172 queue_work(device->submit.wq, &device->submit.worker);
1173 /* do_submit() may sleep internally on al_wait, too */
1174 wake_up(&device->al_wait);
1175 }
1176
1177 /* returns the new drbd_request pointer, if the caller is expected to
1178 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1179 * request on the submitter thread.
1180 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1181 */
1182 static struct drbd_request *
drbd_request_prepare(struct drbd_device * device,struct bio * bio,unsigned long start_jif)1183 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1184 {
1185 const int rw = bio_data_dir(bio);
1186 struct drbd_request *req;
1187
1188 /* allocate outside of all locks; */
1189 req = drbd_req_new(device, bio);
1190 if (!req) {
1191 dec_ap_bio(device);
1192 /* only pass the error to the upper layers.
1193 * if user cannot handle io errors, that's not our business. */
1194 drbd_err(device, "could not kmalloc() req\n");
1195 bio->bi_error = -ENOMEM;
1196 bio_endio(bio);
1197 return ERR_PTR(-ENOMEM);
1198 }
1199 req->start_jif = start_jif;
1200
1201 if (!get_ldev(device)) {
1202 bio_put(req->private_bio);
1203 req->private_bio = NULL;
1204 }
1205
1206 /* Update disk stats */
1207 _drbd_start_io_acct(device, req);
1208
1209 if (rw == WRITE && req->private_bio && req->i.size
1210 && !test_bit(AL_SUSPENDED, &device->flags)) {
1211 if (!drbd_al_begin_io_fastpath(device, &req->i)) {
1212 atomic_inc(&device->ap_actlog_cnt);
1213 drbd_queue_write(device, req);
1214 return NULL;
1215 }
1216 req->rq_state |= RQ_IN_ACT_LOG;
1217 req->in_actlog_jif = jiffies;
1218 }
1219
1220 return req;
1221 }
1222
drbd_send_and_submit(struct drbd_device * device,struct drbd_request * req)1223 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1224 {
1225 struct drbd_resource *resource = device->resource;
1226 const int rw = bio_rw(req->master_bio);
1227 struct bio_and_error m = { NULL, };
1228 bool no_remote = false;
1229 bool submit_private_bio = false;
1230
1231 spin_lock_irq(&resource->req_lock);
1232 if (rw == WRITE) {
1233 /* This may temporarily give up the req_lock,
1234 * but will re-aquire it before it returns here.
1235 * Needs to be before the check on drbd_suspended() */
1236 complete_conflicting_writes(req);
1237 /* no more giving up req_lock from now on! */
1238
1239 /* check for congestion, and potentially stop sending
1240 * full data updates, but start sending "dirty bits" only. */
1241 maybe_pull_ahead(device);
1242 }
1243
1244
1245 if (drbd_suspended(device)) {
1246 /* push back and retry: */
1247 req->rq_state |= RQ_POSTPONED;
1248 if (req->private_bio) {
1249 bio_put(req->private_bio);
1250 req->private_bio = NULL;
1251 put_ldev(device);
1252 }
1253 goto out;
1254 }
1255
1256 /* We fail READ/READA early, if we can not serve it.
1257 * We must do this before req is registered on any lists.
1258 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1259 if (rw != WRITE) {
1260 if (!do_remote_read(req) && !req->private_bio)
1261 goto nodata;
1262 }
1263
1264 /* which transfer log epoch does this belong to? */
1265 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1266
1267 /* no point in adding empty flushes to the transfer log,
1268 * they are mapped to drbd barriers already. */
1269 if (likely(req->i.size!=0)) {
1270 if (rw == WRITE)
1271 first_peer_device(device)->connection->current_tle_writes++;
1272
1273 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1274 }
1275
1276 if (rw == WRITE) {
1277 if (!drbd_process_write_request(req))
1278 no_remote = true;
1279 } else {
1280 /* We either have a private_bio, or we can read from remote.
1281 * Otherwise we had done the goto nodata above. */
1282 if (req->private_bio == NULL) {
1283 _req_mod(req, TO_BE_SENT);
1284 _req_mod(req, QUEUE_FOR_NET_READ);
1285 } else
1286 no_remote = true;
1287 }
1288
1289 /* If it took the fast path in drbd_request_prepare, add it here.
1290 * The slow path has added it already. */
1291 if (list_empty(&req->req_pending_master_completion))
1292 list_add_tail(&req->req_pending_master_completion,
1293 &device->pending_master_completion[rw == WRITE]);
1294 if (req->private_bio) {
1295 /* needs to be marked within the same spinlock */
1296 list_add_tail(&req->req_pending_local,
1297 &device->pending_completion[rw == WRITE]);
1298 _req_mod(req, TO_BE_SUBMITTED);
1299 /* but we need to give up the spinlock to submit */
1300 submit_private_bio = true;
1301 } else if (no_remote) {
1302 nodata:
1303 if (__ratelimit(&drbd_ratelimit_state))
1304 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1305 (unsigned long long)req->i.sector, req->i.size >> 9);
1306 /* A write may have been queued for send_oos, however.
1307 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1308 }
1309
1310 out:
1311 if (drbd_req_put_completion_ref(req, &m, 1))
1312 kref_put(&req->kref, drbd_req_destroy);
1313 spin_unlock_irq(&resource->req_lock);
1314
1315 /* Even though above is a kref_put(), this is safe.
1316 * As long as we still need to submit our private bio,
1317 * we hold a completion ref, and the request cannot disappear.
1318 * If however this request did not even have a private bio to submit
1319 * (e.g. remote read), req may already be invalid now.
1320 * That's why we cannot check on req->private_bio. */
1321 if (submit_private_bio)
1322 drbd_submit_req_private_bio(req);
1323 if (m.bio)
1324 complete_master_bio(device, &m);
1325 }
1326
__drbd_make_request(struct drbd_device * device,struct bio * bio,unsigned long start_jif)1327 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1328 {
1329 struct drbd_request *req = drbd_request_prepare(device, bio, start_jif);
1330 if (IS_ERR_OR_NULL(req))
1331 return;
1332 drbd_send_and_submit(device, req);
1333 }
1334
submit_fast_path(struct drbd_device * device,struct list_head * incoming)1335 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1336 {
1337 struct drbd_request *req, *tmp;
1338 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1339 const int rw = bio_data_dir(req->master_bio);
1340
1341 if (rw == WRITE /* rw != WRITE should not even end up here! */
1342 && req->private_bio && req->i.size
1343 && !test_bit(AL_SUSPENDED, &device->flags)) {
1344 if (!drbd_al_begin_io_fastpath(device, &req->i))
1345 continue;
1346
1347 req->rq_state |= RQ_IN_ACT_LOG;
1348 req->in_actlog_jif = jiffies;
1349 atomic_dec(&device->ap_actlog_cnt);
1350 }
1351
1352 list_del_init(&req->tl_requests);
1353 drbd_send_and_submit(device, req);
1354 }
1355 }
1356
prepare_al_transaction_nonblock(struct drbd_device * device,struct list_head * incoming,struct list_head * pending,struct list_head * later)1357 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1358 struct list_head *incoming,
1359 struct list_head *pending,
1360 struct list_head *later)
1361 {
1362 struct drbd_request *req, *tmp;
1363 int wake = 0;
1364 int err;
1365
1366 spin_lock_irq(&device->al_lock);
1367 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1368 err = drbd_al_begin_io_nonblock(device, &req->i);
1369 if (err == -ENOBUFS)
1370 break;
1371 if (err == -EBUSY)
1372 wake = 1;
1373 if (err)
1374 list_move_tail(&req->tl_requests, later);
1375 else
1376 list_move_tail(&req->tl_requests, pending);
1377 }
1378 spin_unlock_irq(&device->al_lock);
1379 if (wake)
1380 wake_up(&device->al_wait);
1381 return !list_empty(pending);
1382 }
1383
send_and_submit_pending(struct drbd_device * device,struct list_head * pending)1384 void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1385 {
1386 struct drbd_request *req, *tmp;
1387
1388 list_for_each_entry_safe(req, tmp, pending, tl_requests) {
1389 req->rq_state |= RQ_IN_ACT_LOG;
1390 req->in_actlog_jif = jiffies;
1391 atomic_dec(&device->ap_actlog_cnt);
1392 list_del_init(&req->tl_requests);
1393 drbd_send_and_submit(device, req);
1394 }
1395 }
1396
do_submit(struct work_struct * ws)1397 void do_submit(struct work_struct *ws)
1398 {
1399 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1400 LIST_HEAD(incoming); /* from drbd_make_request() */
1401 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */
1402 LIST_HEAD(busy); /* blocked by resync requests */
1403
1404 /* grab new incoming requests */
1405 spin_lock_irq(&device->resource->req_lock);
1406 list_splice_tail_init(&device->submit.writes, &incoming);
1407 spin_unlock_irq(&device->resource->req_lock);
1408
1409 for (;;) {
1410 DEFINE_WAIT(wait);
1411
1412 /* move used-to-be-busy back to front of incoming */
1413 list_splice_init(&busy, &incoming);
1414 submit_fast_path(device, &incoming);
1415 if (list_empty(&incoming))
1416 break;
1417
1418 for (;;) {
1419 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1420
1421 list_splice_init(&busy, &incoming);
1422 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1423 if (!list_empty(&pending))
1424 break;
1425
1426 schedule();
1427
1428 /* If all currently "hot" activity log extents are kept busy by
1429 * incoming requests, we still must not totally starve new
1430 * requests to "cold" extents.
1431 * Something left on &incoming means there had not been
1432 * enough update slots available, and the activity log
1433 * has been marked as "starving".
1434 *
1435 * Try again now, without looking for new requests,
1436 * effectively blocking all new requests until we made
1437 * at least _some_ progress with what we currently have.
1438 */
1439 if (!list_empty(&incoming))
1440 continue;
1441
1442 /* Nothing moved to pending, but nothing left
1443 * on incoming: all moved to busy!
1444 * Grab new and iterate. */
1445 spin_lock_irq(&device->resource->req_lock);
1446 list_splice_tail_init(&device->submit.writes, &incoming);
1447 spin_unlock_irq(&device->resource->req_lock);
1448 }
1449 finish_wait(&device->al_wait, &wait);
1450
1451 /* If the transaction was full, before all incoming requests
1452 * had been processed, skip ahead to commit, and iterate
1453 * without splicing in more incoming requests from upper layers.
1454 *
1455 * Else, if all incoming have been processed,
1456 * they have become either "pending" (to be submitted after
1457 * next transaction commit) or "busy" (blocked by resync).
1458 *
1459 * Maybe more was queued, while we prepared the transaction?
1460 * Try to stuff those into this transaction as well.
1461 * Be strictly non-blocking here,
1462 * we already have something to commit.
1463 *
1464 * Commit if we don't make any more progres.
1465 */
1466
1467 while (list_empty(&incoming)) {
1468 LIST_HEAD(more_pending);
1469 LIST_HEAD(more_incoming);
1470 bool made_progress;
1471
1472 /* It is ok to look outside the lock,
1473 * it's only an optimization anyways */
1474 if (list_empty(&device->submit.writes))
1475 break;
1476
1477 spin_lock_irq(&device->resource->req_lock);
1478 list_splice_tail_init(&device->submit.writes, &more_incoming);
1479 spin_unlock_irq(&device->resource->req_lock);
1480
1481 if (list_empty(&more_incoming))
1482 break;
1483
1484 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1485
1486 list_splice_tail_init(&more_pending, &pending);
1487 list_splice_tail_init(&more_incoming, &incoming);
1488 if (!made_progress)
1489 break;
1490 }
1491
1492 drbd_al_begin_io_commit(device);
1493 send_and_submit_pending(device, &pending);
1494 }
1495 }
1496
drbd_make_request(struct request_queue * q,struct bio * bio)1497 blk_qc_t drbd_make_request(struct request_queue *q, struct bio *bio)
1498 {
1499 struct drbd_device *device = (struct drbd_device *) q->queuedata;
1500 unsigned long start_jif;
1501
1502 blk_queue_split(q, &bio, q->bio_split);
1503
1504 start_jif = jiffies;
1505
1506 /*
1507 * what we "blindly" assume:
1508 */
1509 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1510
1511 inc_ap_bio(device);
1512 __drbd_make_request(device, bio, start_jif);
1513 return BLK_QC_T_NONE;
1514 }
1515
request_timer_fn(unsigned long data)1516 void request_timer_fn(unsigned long data)
1517 {
1518 struct drbd_device *device = (struct drbd_device *) data;
1519 struct drbd_connection *connection = first_peer_device(device)->connection;
1520 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1521 struct net_conf *nc;
1522 unsigned long oldest_submit_jif;
1523 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1524 unsigned long now;
1525
1526 rcu_read_lock();
1527 nc = rcu_dereference(connection->net_conf);
1528 if (nc && device->state.conn >= C_WF_REPORT_PARAMS)
1529 ent = nc->timeout * HZ/10 * nc->ko_count;
1530
1531 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1532 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1533 put_ldev(device);
1534 }
1535 rcu_read_unlock();
1536
1537 et = min_not_zero(dt, ent);
1538
1539 if (!et)
1540 return; /* Recurring timer stopped */
1541
1542 now = jiffies;
1543 nt = now + et;
1544
1545 spin_lock_irq(&device->resource->req_lock);
1546 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1547 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1548 req_peer = connection->req_not_net_done;
1549 /* maybe the oldest request waiting for the peer is in fact still
1550 * blocking in tcp sendmsg */
1551 if (!req_peer && connection->req_next && connection->req_next->pre_send_jif)
1552 req_peer = connection->req_next;
1553
1554 /* evaluate the oldest peer request only in one timer! */
1555 if (req_peer && req_peer->device != device)
1556 req_peer = NULL;
1557
1558 /* do we have something to evaluate? */
1559 if (req_peer == NULL && req_write == NULL && req_read == NULL)
1560 goto out;
1561
1562 oldest_submit_jif =
1563 (req_write && req_read)
1564 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1565 ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1566 : req_write ? req_write->pre_submit_jif
1567 : req_read ? req_read->pre_submit_jif : now;
1568
1569 /* The request is considered timed out, if
1570 * - we have some effective timeout from the configuration,
1571 * with above state restrictions applied,
1572 * - the oldest request is waiting for a response from the network
1573 * resp. the local disk,
1574 * - the oldest request is in fact older than the effective timeout,
1575 * - the connection was established (resp. disk was attached)
1576 * for longer than the timeout already.
1577 * Note that for 32bit jiffies and very stable connections/disks,
1578 * we may have a wrap around, which is catched by
1579 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1580 *
1581 * Side effect: once per 32bit wrap-around interval, which means every
1582 * ~198 days with 250 HZ, we have a window where the timeout would need
1583 * to expire twice (worst case) to become effective. Good enough.
1584 */
1585 if (ent && req_peer &&
1586 time_after(now, req_peer->pre_send_jif + ent) &&
1587 !time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) {
1588 drbd_warn(device, "Remote failed to finish a request within ko-count * timeout\n");
1589 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1590 }
1591 if (dt && oldest_submit_jif != now &&
1592 time_after(now, oldest_submit_jif + dt) &&
1593 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1594 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1595 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1596 }
1597
1598 /* Reschedule timer for the nearest not already expired timeout.
1599 * Fallback to now + min(effective network timeout, disk timeout). */
1600 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1601 ? req_peer->pre_send_jif + ent : now + et;
1602 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1603 ? oldest_submit_jif + dt : now + et;
1604 nt = time_before(ent, dt) ? ent : dt;
1605 out:
1606 spin_unlock_irq(&device->resource->req_lock);
1607 mod_timer(&device->request_timer, nt);
1608 }
1609