]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - drivers/block/drbd/drbd_req.c
drbd: report sizes if rejecting too small peer disk
[mirror_ubuntu-bionic-kernel.git] / drivers / block / drbd / drbd_req.c
CommitLineData
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
b30ab791 34static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
57bcb6cf 35
b411b363 36/* Update disk stats at start of I/O request */
b30ab791 37static void _drbd_start_io_acct(struct drbd_device *device, struct drbd_request *req)
b411b363 38{
24480854
GZ
39 generic_start_io_acct(bio_data_dir(req->master_bio), req->i.size >> 9,
40 &device->vdisk->part0);
b411b363
PR
41}
42
43/* Update disk stats when completing request upwards */
b30ab791 44static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
b411b363 45{
24480854
GZ
46 generic_end_io_acct(bio_data_dir(req->master_bio),
47 &device->vdisk->part0, req->start_jif);
b411b363
PR
48}
49
b30ab791 50static struct drbd_request *drbd_req_new(struct drbd_device *device,
9e204cdd
AG
51 struct bio *bio_src)
52{
53 struct drbd_request *req;
54
23fe8f8b 55 req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
9e204cdd
AG
56 if (!req)
57 return NULL;
23fe8f8b 58 memset(req, 0, sizeof(*req));
9e204cdd
AG
59
60 drbd_req_make_private_bio(req, bio_src);
61 req->rq_state = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
84b8c06b 62 req->device = device;
9e204cdd
AG
63 req->master_bio = bio_src;
64 req->epoch = 0;
53840641 65
9e204cdd 66 drbd_clear_interval(&req->i);
4f024f37
KO
67 req->i.sector = bio_src->bi_iter.bi_sector;
68 req->i.size = bio_src->bi_iter.bi_size;
5e472264 69 req->i.local = true;
53840641
AG
70 req->i.waiting = false;
71
9e204cdd
AG
72 INIT_LIST_HEAD(&req->tl_requests);
73 INIT_LIST_HEAD(&req->w.list);
844a6ae7
LE
74 INIT_LIST_HEAD(&req->req_pending_master_completion);
75 INIT_LIST_HEAD(&req->req_pending_local);
9e204cdd 76
a0d856df 77 /* one reference to be put by __drbd_make_request */
b406777e 78 atomic_set(&req->completion_ref, 1);
a0d856df 79 /* one kref as long as completion_ref > 0 */
b406777e 80 kref_init(&req->kref);
9e204cdd
AG
81 return req;
82}
83
08d0dabf
LE
84static 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
9a278a79 97void drbd_req_destroy(struct kref *kref)
b411b363 98{
b406777e 99 struct drbd_request *req = container_of(kref, struct drbd_request, kref);
84b8c06b 100 struct drbd_device *device = req->device;
a0d856df
LE
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))) {
d0180171 107 drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
a0d856df
LE
108 s, atomic_read(&req->completion_ref));
109 return;
110 }
288f422e 111
844a6ae7
LE
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. */
2312f0b3 120 list_del_init(&req->tl_requests);
288f422e 121
08d0dabf
LE
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
b411b363
PR
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. */
b406777e 139 if (s & RQ_WRITE) {
b411b363
PR
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 */
b411b363 144
70f17b6b
LE
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) {
d7644018 152 if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
b30ab791 153 drbd_set_out_of_sync(device, req->i.sector, req->i.size);
b411b363 154
d7644018 155 if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
b30ab791 156 drbd_set_in_sync(device, req->i.sector, req->i.size);
d7644018 157 }
b411b363
PR
158
159 /* one might be tempted to move the drbd_al_complete_io
fcefa62e 160 * to the local io completion callback drbd_request_endio.
b411b363
PR
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 */
76590cd1 169 if (s & RQ_IN_ACT_LOG) {
b30ab791
AG
170 if (get_ldev_if_state(device, D_FAILED)) {
171 drbd_al_complete_io(device, &req->i);
172 put_ldev(device);
b411b363 173 } else if (__ratelimit(&drbd_ratelimit_state)) {
d0180171 174 drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
181286ad
LE
175 "but my Disk seems to have failed :(\n",
176 (unsigned long long) req->i.sector, req->i.size);
b411b363
PR
177 }
178 }
179 }
180
9a278a79 181 mempool_free(req, drbd_request_mempool);
b411b363
PR
182}
183
bde89a9e
AG
184static void wake_all_senders(struct drbd_connection *connection)
185{
186 wake_up(&connection->sender_work.q_wait);
b411b363
PR
187}
188
b6dd1a89 189/* must hold resource->req_lock */
bde89a9e 190void start_new_tl_epoch(struct drbd_connection *connection)
b411b363 191{
99b4d8fe 192 /* no point closing an epoch, if it is empty, anyways. */
bde89a9e 193 if (connection->current_tle_writes == 0)
99b4d8fe 194 return;
b411b363 195
bde89a9e
AG
196 connection->current_tle_writes = 0;
197 atomic_inc(&connection->current_tle_nr);
198 wake_all_senders(connection);
b411b363
PR
199}
200
b30ab791 201void complete_master_bio(struct drbd_device *device,
b411b363
PR
202 struct bio_and_error *m)
203{
4246a0b6
CH
204 m->bio->bi_error = m->error;
205 bio_endio(m->bio);
b30ab791 206 dec_ap_bio(device);
b411b363
PR
207}
208
53840641 209
b411b363
PR
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 */
6870ca6d 216static
a0d856df 217void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
b411b363 218{
a0d856df 219 const unsigned s = req->rq_state;
84b8c06b 220 struct drbd_device *device = req->device;
a0d856df
LE
221 int rw;
222 int error, ok;
b411b363 223
b411b363
PR
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 */
a0d856df
LE
233 if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
234 (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
235 (s & RQ_COMPLETION_SUSP)) {
d0180171 236 drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
b411b363 237 return;
a0d856df
LE
238 }
239
240 if (!req->master_bio) {
d0180171 241 drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
b411b363 242 return;
a0d856df 243 }
b411b363 244
a0d856df 245 rw = bio_rw(req->master_bio);
b411b363 246
a0d856df
LE
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);
b411b363 262
a0d856df
LE
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 &&
a6b32bc3
AG
271 req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
272 start_new_tl_epoch(first_peer_device(device)->connection);
b411b363 273
a0d856df 274 /* Update disk stats */
b30ab791 275 _drbd_end_io_acct(device, req);
b411b363 276
a0d856df
LE
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;
b411b363 293
a0d856df 294 if (!(req->rq_state & RQ_POSTPONED)) {
b411b363
PR
295 m->error = ok ? 0 : (error ?: -EIO);
296 m->bio = req->master_bio;
297 req->master_bio = NULL;
08d0dabf
LE
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;
b411b363 303 }
08d0dabf
LE
304
305 if (req->i.waiting)
306 wake_up(&device->misc_wait);
844a6ae7
LE
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);
b411b363 313}
b411b363 314
844a6ae7 315/* still holds resource->req_lock */
a0d856df 316static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
cfa03415 317{
84b8c06b 318 struct drbd_device *device = req->device;
0b0ba1ef 319 D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
a0d856df
LE
320
321 if (!atomic_sub_and_test(put, &req->completion_ref))
322 return 0;
2b4dd36f 323
a0d856df 324 drbd_req_complete(req, m);
9a278a79
LE
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;
b411b363 331 }
9a278a79 332
a0d856df 333 return 1;
b411b363
PR
334}
335
7753a4c1
LE
336static 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
345static 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
362static 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
371static 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
388static 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
397static 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
a0d856df
LE
414/* I'd like this to be the only place that manipulates
415 * req->completion_ref and req->kref. */
416static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
417 int clear, int set)
cfa03415 418{
84b8c06b 419 struct drbd_device *device = req->device;
7753a4c1 420 struct drbd_peer_device *peer_device = first_peer_device(device);
a0d856df
LE
421 unsigned s = req->rq_state;
422 int c_put = 0;
423 int k_put = 0;
cfa03415 424
b30ab791 425 if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
5af2e8ce 426 set |= RQ_COMPLETION_SUSP;
cfa03415 427
a0d856df 428 /* apply */
b411b363 429
a0d856df
LE
430 req->rq_state &= ~clear;
431 req->rq_state |= set;
b411b363 432
a0d856df
LE
433 /* no change? */
434 if (req->rq_state == s)
435 return;
b411b363 436
a0d856df
LE
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)) {
b30ab791 443 inc_ap_pending(device);
a0d856df 444 atomic_inc(&req->completion_ref);
b411b363
PR
445 }
446
e5f891b2 447 if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
a0d856df 448 atomic_inc(&req->completion_ref);
7753a4c1 449 set_if_null_req_next(peer_device, req);
e5f891b2 450 }
a0d856df
LE
451
452 if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
453 kref_get(&req->kref); /* wait for the DONE */
454
e5f891b2 455 if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
668700b4 456 /* potentially already completed in the ack_receiver thread */
7753a4c1 457 if (!(s & RQ_NET_DONE)) {
e5f891b2 458 atomic_add(req->i.size >> 9, &device->ap_in_flight);
7753a4c1
LE
459 set_if_null_req_not_net_done(peer_device, req);
460 }
f85d9f2d 461 if (req->rq_state & RQ_NET_PENDING)
7753a4c1 462 set_if_null_req_ack_pending(peer_device, req);
e5f891b2 463 }
a0d856df 464
5af2e8ce
PR
465 if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
466 atomic_inc(&req->completion_ref);
467
a0d856df
LE
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)) {
0b0ba1ef 474 D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
a0d856df
LE
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;
b411b363 479 }
b411b363 480
a0d856df
LE
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;
844a6ae7 486 list_del_init(&req->req_pending_local);
a0d856df 487 }
b411b363 488
a0d856df 489 if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
b30ab791 490 dec_ap_pending(device);
a0d856df 491 ++c_put;
e5f891b2 492 req->acked_jif = jiffies;
7753a4c1 493 advance_conn_req_ack_pending(peer_device, req);
a0d856df
LE
494 }
495
7753a4c1 496 if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
a0d856df 497 ++c_put;
7753a4c1
LE
498 advance_conn_req_next(peer_device, req);
499 }
a0d856df 500
e5f891b2
LE
501 if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
502 if (s & RQ_NET_SENT)
b30ab791 503 atomic_sub(req->i.size >> 9, &device->ap_in_flight);
e5f891b2
LE
504 if (s & RQ_EXP_BARR_ACK)
505 ++k_put;
506 req->net_done_jif = jiffies;
7753a4c1
LE
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);
a0d856df
LE
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)
d0180171 524 drbd_err(device,
a0d856df
LE
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)
b30ab791 531 wake_up(&device->misc_wait);
a0d856df
LE
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);
b411b363
PR
537}
538
b30ab791 539static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
ccae7868
LE
540{
541 char b[BDEVNAME_SIZE];
542
42839f65 543 if (!__ratelimit(&drbd_ratelimit_state))
ccae7868
LE
544 return;
545
d0180171 546 drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
ccae7868 547 (req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
42839f65
LE
548 (unsigned long long)req->i.sector,
549 req->i.size >> 9,
b30ab791 550 bdevname(device->ldev->backing_bdev, b));
ccae7868
LE
551}
552
e5f891b2
LE
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) */
559static 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
b411b363
PR
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 */
2a80699f 578int __req_mod(struct drbd_request *req, enum drbd_req_event what,
b411b363
PR
579 struct bio_and_error *m)
580{
44a4d551
LE
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;
44ed167d 584 struct net_conf *nc;
303d1448 585 int p, rv = 0;
7be8da07
AG
586
587 if (m)
588 m->bio = NULL;
b411b363 589
b411b363
PR
590 switch (what) {
591 default:
d0180171 592 drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
b411b363
PR
593 break;
594
595 /* does not happen...
596 * initialization done in drbd_req_new
8554df1c 597 case CREATED:
b411b363
PR
598 break;
599 */
600
8554df1c 601 case TO_BE_SENT: /* via network */
7be8da07 602 /* reached via __drbd_make_request
b411b363 603 * and from w_read_retry_remote */
0b0ba1ef 604 D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
44ed167d 605 rcu_read_lock();
44a4d551 606 nc = rcu_dereference(connection->net_conf);
44ed167d
PR
607 p = nc->wire_protocol;
608 rcu_read_unlock();
303d1448
PR
609 req->rq_state |=
610 p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
611 p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
a0d856df 612 mod_rq_state(req, m, 0, RQ_NET_PENDING);
b411b363
PR
613 break;
614
8554df1c 615 case TO_BE_SUBMITTED: /* locally */
7be8da07 616 /* reached via __drbd_make_request */
0b0ba1ef 617 D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
a0d856df 618 mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
b411b363
PR
619 break;
620
8554df1c 621 case COMPLETED_OK:
2b4dd36f 622 if (req->rq_state & RQ_WRITE)
b30ab791 623 device->writ_cnt += req->i.size >> 9;
b411b363 624 else
b30ab791 625 device->read_cnt += req->i.size >> 9;
b411b363 626
a0d856df
LE
627 mod_rq_state(req, m, RQ_LOCAL_PENDING,
628 RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
b411b363
PR
629 break;
630
cdfda633 631 case ABORT_DISK_IO:
a0d856df 632 mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
2b4dd36f
PR
633 break;
634
edc9f5eb 635 case WRITE_COMPLETED_WITH_ERROR:
b30ab791
AG
636 drbd_report_io_error(device, req);
637 __drbd_chk_io_error(device, DRBD_WRITE_ERROR);
edc9f5eb 638 mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
b411b363
PR
639 break;
640
8554df1c 641 case READ_COMPLETED_WITH_ERROR:
b30ab791
AG
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);
a0d856df
LE
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);
2f632aeb
LE
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);
4439c400 656 break;
b411b363 657
8554df1c 658 case QUEUE_FOR_NET_READ:
b411b363
PR
659 /* READ or READA, and
660 * no local disk,
661 * or target area marked as invalid,
662 * or just got an io-error. */
7be8da07 663 /* from __drbd_make_request
b411b363
PR
664 * or from bio_endio during read io-error recovery */
665
6870ca6d
LE
666 /* So we can verify the handle in the answer packet.
667 * Corresponding drbd_remove_request_interval is in
a0d856df 668 * drbd_req_complete() */
0b0ba1ef 669 D_ASSERT(device, drbd_interval_empty(&req->i));
b30ab791 670 drbd_insert_interval(&device->read_requests, &req->i);
b411b363 671
b30ab791 672 set_bit(UNPLUG_REMOTE, &device->flags);
b411b363 673
0b0ba1ef
AG
674 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
675 D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
a0d856df 676 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
4439c400 677 req->w.cb = w_send_read_req;
44a4d551 678 drbd_queue_work(&connection->sender_work,
84b8c06b 679 &req->w);
b411b363
PR
680 break;
681
8554df1c 682 case QUEUE_FOR_NET_WRITE:
b411b363 683 /* assert something? */
7be8da07 684 /* from __drbd_make_request only */
b411b363 685
6870ca6d 686 /* Corresponding drbd_remove_request_interval is in
a0d856df 687 * drbd_req_complete() */
0b0ba1ef 688 D_ASSERT(device, drbd_interval_empty(&req->i));
b30ab791 689 drbd_insert_interval(&device->write_requests, &req->i);
b411b363
PR
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
7be8da07 700 * __drbd_make_request, because we now may set the bit
b411b363
PR
701 * again ourselves to close the current epoch.
702 *
703 * Add req to the (now) current epoch (barrier). */
704
83c38830
LE
705 /* otherwise we may lose an unplug, which may cause some remote
706 * io-scheduler timeout to expire, increasing maximum latency,
707 * hurting performance. */
b30ab791 708 set_bit(UNPLUG_REMOTE, &device->flags);
b411b363
PR
709
710 /* queue work item to send data */
0b0ba1ef 711 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
a0d856df 712 mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
b411b363 713 req->w.cb = w_send_dblock;
44a4d551 714 drbd_queue_work(&connection->sender_work,
84b8c06b 715 &req->w);
b411b363
PR
716
717 /* close the epoch, in case it outgrew the limit */
44ed167d 718 rcu_read_lock();
44a4d551 719 nc = rcu_dereference(connection->net_conf);
44ed167d
PR
720 p = nc->max_epoch_size;
721 rcu_read_unlock();
44a4d551
LE
722 if (connection->current_tle_writes >= p)
723 start_new_tl_epoch(connection);
b411b363
PR
724
725 break;
726
8554df1c 727 case QUEUE_FOR_SEND_OOS:
a0d856df 728 mod_rq_state(req, m, 0, RQ_NET_QUEUED);
8f7bed77 729 req->w.cb = w_send_out_of_sync;
44a4d551 730 drbd_queue_work(&connection->sender_work,
84b8c06b 731 &req->w);
73a01a18
PR
732 break;
733
ea9d6729 734 case READ_RETRY_REMOTE_CANCELED:
8554df1c 735 case SEND_CANCELED:
8554df1c 736 case SEND_FAILED:
b411b363
PR
737 /* real cleanup will be done from tl_clear. just update flags
738 * so it is no longer marked as on the worker queue */
a0d856df 739 mod_rq_state(req, m, RQ_NET_QUEUED, 0);
b411b363
PR
740 break;
741
8554df1c 742 case HANDED_OVER_TO_NETWORK:
b411b363 743 /* assert something? */
e5f891b2 744 if (is_pending_write_protocol_A(req))
b411b363
PR
745 /* this is what is dangerous about protocol A:
746 * pretend it was successfully written on the peer. */
e5f891b2
LE
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. */
6d49e101
LE
754 break;
755
27a434fe 756 case OOS_HANDED_TO_NETWORK:
6d49e101
LE
757 /* Was not set PENDING, no longer QUEUED, so is now DONE
758 * as far as this connection is concerned. */
a0d856df 759 mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
b411b363
PR
760 break;
761
8554df1c 762 case CONNECTION_LOST_WHILE_PENDING:
b411b363 763 /* transfer log cleanup after connection loss */
a0d856df
LE
764 mod_rq_state(req, m,
765 RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
766 RQ_NET_DONE);
b411b363
PR
767 break;
768
d4dabbe2
LE
769 case CONFLICT_RESOLVED:
770 /* for superseded conflicting writes of multiple primaries,
b411b363 771 * there is no need to keep anything in the tl, potential
934722a2
LE
772 * node crashes are covered by the activity log.
773 *
774 * If this request had been marked as RQ_POSTPONED before,
d4dabbe2 775 * it will actually not be completed, but "restarted",
934722a2 776 * resubmitted from the retry worker context. */
0b0ba1ef
AG
777 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
778 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
934722a2
LE
779 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
780 break;
781
0afd569a 782 case WRITE_ACKED_BY_PEER_AND_SIS:
934722a2 783 req->rq_state |= RQ_NET_SIS;
8554df1c 784 case WRITE_ACKED_BY_PEER:
08d0dabf
LE
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.
d64957c9 789 * Nothing more to do here.
b411b363 790 * We want to keep the tl in place for all protocols, to cater
d64957c9 791 * for volatile write-back caches on lower level devices. */
303d1448 792 goto ack_common;
8554df1c 793 case RECV_ACKED_BY_PEER:
0b0ba1ef 794 D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
b411b363 795 /* protocol B; pretends to be successfully written on peer.
8554df1c 796 * see also notes above in HANDED_OVER_TO_NETWORK about
b411b363 797 * protocol != C */
303d1448 798 ack_common:
a0d856df 799 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
b411b363
PR
800 break;
801
7be8da07 802 case POSTPONE_WRITE:
0b0ba1ef 803 D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
303d1448 804 /* If this node has already detected the write conflict, the
7be8da07
AG
805 * worker will be waiting on misc_wait. Wake it up once this
806 * request has completed locally.
807 */
0b0ba1ef 808 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
7be8da07 809 req->rq_state |= RQ_POSTPONED;
a0d856df 810 if (req->i.waiting)
b30ab791 811 wake_up(&device->misc_wait);
a0d856df
LE
812 /* Do not clear RQ_NET_PENDING. This request will make further
813 * progress via restart_conflicting_writes() or
814 * fail_postponed_requests(). Hopefully. */
7be8da07 815 break;
b411b363 816
8554df1c 817 case NEG_ACKED:
46e21bba 818 mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
b411b363
PR
819 break;
820
8554df1c 821 case FAIL_FROZEN_DISK_IO:
265be2d0
PR
822 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
823 break;
a0d856df 824 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
265be2d0
PR
825 break;
826
8554df1c 827 case RESTART_FROZEN_DISK_IO:
265be2d0
PR
828 if (!(req->rq_state & RQ_LOCAL_COMPLETED))
829 break;
830
a0d856df
LE
831 mod_rq_state(req, m,
832 RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
833 RQ_LOCAL_PENDING);
265be2d0
PR
834
835 rv = MR_READ;
836 if (bio_data_dir(req->master_bio) == WRITE)
837 rv = MR_WRITE;
838
b30ab791 839 get_ldev(device); /* always succeeds in this call path */
265be2d0 840 req->w.cb = w_restart_disk_io;
44a4d551 841 drbd_queue_work(&connection->sender_work,
84b8c06b 842 &req->w);
265be2d0
PR
843 break;
844
8554df1c 845 case RESEND:
509fc019
PR
846 /* Simply complete (local only) READs. */
847 if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
8a0bab2a 848 mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
509fc019
PR
849 break;
850 }
851
11b58e73 852 /* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
a0d856df
LE
853 before the connection loss (B&C only); only P_BARRIER_ACK
854 (or the local completion?) was missing when we suspended.
6870ca6d
LE
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. */
11b58e73 857 if (!(req->rq_state & RQ_NET_OK)) {
84b8c06b 858 /* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
a0d856df
LE
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);
11b58e73 862 if (req->w.cb) {
44a4d551
LE
863 /* w.cb expected to be w_send_dblock, or w_send_read_req */
864 drbd_queue_work(&connection->sender_work,
84b8c06b 865 &req->w);
11b58e73 866 rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
a0d856df 867 } /* else: FIXME can this happen? */
11b58e73
PR
868 break;
869 }
8554df1c 870 /* else, fall through to BARRIER_ACKED */
11b58e73 871
8554df1c 872 case BARRIER_ACKED:
a0d856df 873 /* barrier ack for READ requests does not make sense */
288f422e
PR
874 if (!(req->rq_state & RQ_WRITE))
875 break;
876
b411b363 877 if (req->rq_state & RQ_NET_PENDING) {
a209b4ae 878 /* barrier came in before all requests were acked.
b411b363
PR
879 * this is bad, because if the connection is lost now,
880 * we won't be able to clean them up... */
d0180171 881 drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
b411b363 882 }
a0d856df
LE
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);
b411b363
PR
889 break;
890
8554df1c 891 case DATA_RECEIVED:
0b0ba1ef 892 D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
a0d856df 893 mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
b411b363 894 break;
7074e4a7
LE
895
896 case QUEUE_AS_DRBD_BARRIER:
44a4d551 897 start_new_tl_epoch(connection);
7074e4a7
LE
898 mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
899 break;
b411b363 900 };
2a80699f
PR
901
902 return rv;
b411b363
PR
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 */
b30ab791 912static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
b411b363
PR
913{
914 unsigned long sbnr, ebnr;
915 sector_t esector, nr_sectors;
916
b30ab791 917 if (device->state.disk == D_UP_TO_DATE)
0da34df0 918 return true;
b30ab791 919 if (device->state.disk != D_INCONSISTENT)
0da34df0 920 return false;
b411b363 921 esector = sector + (size >> 9) - 1;
b30ab791 922 nr_sectors = drbd_get_capacity(device->this_bdev);
0b0ba1ef
AG
923 D_ASSERT(device, sector < nr_sectors);
924 D_ASSERT(device, esector < nr_sectors);
b411b363
PR
925
926 sbnr = BM_SECT_TO_BIT(sector);
927 ebnr = BM_SECT_TO_BIT(esector);
928
b30ab791 929 return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
b411b363
PR
930}
931
b30ab791 932static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
5da9c836 933 enum drbd_read_balancing rbm)
380207d0 934{
380207d0 935 struct backing_dev_info *bdi;
d60de03a 936 int stripe_shift;
380207d0 937
380207d0
PR
938 switch (rbm) {
939 case RB_CONGESTED_REMOTE:
b30ab791 940 bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
380207d0
PR
941 return bdi_read_congested(bdi);
942 case RB_LEAST_PENDING:
b30ab791
AG
943 return atomic_read(&device->local_cnt) >
944 atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
d60de03a
PR
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;
380207d0 953 case RB_ROUND_ROBIN:
b30ab791 954 return test_and_change_bit(READ_BALANCE_RR, &device->flags);
380207d0
PR
955 case RB_PREFER_REMOTE:
956 return true;
957 case RB_PREFER_LOCAL:
958 default:
959 return false;
960 }
961}
962
6024fece
AG
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.
648e46b5
LE
969 *
970 * Only way out: remove the conflicting intervals from the tree.
6024fece 971 */
648e46b5 972static void complete_conflicting_writes(struct drbd_request *req)
6024fece 973{
648e46b5 974 DEFINE_WAIT(wait);
84b8c06b 975 struct drbd_device *device = req->device;
648e46b5
LE
976 struct drbd_interval *i;
977 sector_t sector = req->i.sector;
978 int size = req->i.size;
979
648e46b5 980 for (;;) {
1b228c98
LE
981 drbd_for_each_overlap(i, &device->write_requests, sector, size) {
982 /* Ignore, if already completed to upper layers. */
983 if (i->completed)
984 continue;
985 /* Handle the first found overlap. After the schedule
986 * we have to restart the tree walk. */
648e46b5 987 break;
1b228c98
LE
988 }
989 if (!i) /* if any */
990 break;
991
648e46b5 992 /* Indicate to wake up device->misc_wait on progress. */
1b228c98 993 prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
648e46b5 994 i->waiting = true;
0500813f 995 spin_unlock_irq(&device->resource->req_lock);
648e46b5 996 schedule();
0500813f 997 spin_lock_irq(&device->resource->req_lock);
6024fece 998 }
b30ab791 999 finish_wait(&device->misc_wait, &wait);
b411b363
PR
1000}
1001
5da9c836 1002/* called within req_lock and rcu_read_lock() */
b30ab791 1003static void maybe_pull_ahead(struct drbd_device *device)
0d5934e3 1004{
a6b32bc3 1005 struct drbd_connection *connection = first_peer_device(device)->connection;
5da9c836
LE
1006 struct net_conf *nc;
1007 bool congested = false;
1008 enum drbd_on_congestion on_congestion;
1009
607f25e5 1010 rcu_read_lock();
bde89a9e 1011 nc = rcu_dereference(connection->net_conf);
5da9c836 1012 on_congestion = nc ? nc->on_congestion : OC_BLOCK;
607f25e5 1013 rcu_read_unlock();
5da9c836 1014 if (on_congestion == OC_BLOCK ||
bde89a9e 1015 connection->agreed_pro_version < 96)
3b9ef85e 1016 return;
0d5934e3 1017
0c066bc3
LE
1018 if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1019 return; /* nothing to do ... */
1020
0d5934e3
LE
1021 /* If I don't even have good local storage, we can not reasonably try
1022 * to pull ahead of the peer. We also need the local reference to make
b30ab791 1023 * sure device->act_log is there.
0d5934e3 1024 */
b30ab791 1025 if (!get_ldev_if_state(device, D_UP_TO_DATE))
0d5934e3
LE
1026 return;
1027
5da9c836 1028 if (nc->cong_fill &&
b30ab791 1029 atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
d0180171 1030 drbd_info(device, "Congestion-fill threshold reached\n");
5da9c836 1031 congested = true;
0d5934e3
LE
1032 }
1033
b30ab791 1034 if (device->act_log->used >= nc->cong_extents) {
d0180171 1035 drbd_info(device, "Congestion-extents threshold reached\n");
5da9c836 1036 congested = true;
0d5934e3
LE
1037 }
1038
1039 if (congested) {
99b4d8fe 1040 /* start a new epoch for non-mirrored writes */
a6b32bc3 1041 start_new_tl_epoch(first_peer_device(device)->connection);
0d5934e3 1042
5da9c836 1043 if (on_congestion == OC_PULL_AHEAD)
b30ab791 1044 _drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
5da9c836 1045 else /*nc->on_congestion == OC_DISCONNECT */
b30ab791 1046 _drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
0d5934e3 1047 }
b30ab791 1048 put_ldev(device);
0d5934e3
LE
1049}
1050
5da9c836
LE
1051/* If this returns false, and req->private_bio is still set,
1052 * this should be submitted locally.
1053 *
1054 * If it returns false, but req->private_bio is not set,
1055 * we do not have access to good data :(
1056 *
1057 * Otherwise, this destroys req->private_bio, if any,
1058 * and returns true.
1059 */
1060static bool do_remote_read(struct drbd_request *req)
1061{
84b8c06b 1062 struct drbd_device *device = req->device;
5da9c836
LE
1063 enum drbd_read_balancing rbm;
1064
1065 if (req->private_bio) {
b30ab791 1066 if (!drbd_may_do_local_read(device,
5da9c836
LE
1067 req->i.sector, req->i.size)) {
1068 bio_put(req->private_bio);
1069 req->private_bio = NULL;
b30ab791 1070 put_ldev(device);
5da9c836
LE
1071 }
1072 }
1073
b30ab791 1074 if (device->state.pdsk != D_UP_TO_DATE)
5da9c836
LE
1075 return false;
1076
a0d856df
LE
1077 if (req->private_bio == NULL)
1078 return true;
1079
5da9c836
LE
1080 /* TODO: improve read balancing decisions, take into account drbd
1081 * protocol, pending requests etc. */
1082
1083 rcu_read_lock();
b30ab791 1084 rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
5da9c836
LE
1085 rcu_read_unlock();
1086
1087 if (rbm == RB_PREFER_LOCAL && req->private_bio)
1088 return false; /* submit locally */
1089
b30ab791 1090 if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
5da9c836
LE
1091 if (req->private_bio) {
1092 bio_put(req->private_bio);
1093 req->private_bio = NULL;
b30ab791 1094 put_ldev(device);
5da9c836
LE
1095 }
1096 return true;
1097 }
1098
1099 return false;
1100}
1101
2e9ffde6
AG
1102bool drbd_should_do_remote(union drbd_dev_state s)
1103{
1104 return s.pdsk == D_UP_TO_DATE ||
1105 (s.pdsk >= D_INCONSISTENT &&
1106 s.conn >= C_WF_BITMAP_T &&
1107 s.conn < C_AHEAD);
1108 /* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1109 That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1110 states. */
1111}
1112
1113static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1114{
1115 return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1116 /* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1117 since we enter state C_AHEAD only if proto >= 96 */
1118}
1119
5da9c836
LE
1120/* returns number of connections (== 1, for drbd 8.4)
1121 * expected to actually write this data,
1122 * which does NOT include those that we are L_AHEAD for. */
1123static int drbd_process_write_request(struct drbd_request *req)
1124{
84b8c06b 1125 struct drbd_device *device = req->device;
5da9c836
LE
1126 int remote, send_oos;
1127
b30ab791
AG
1128 remote = drbd_should_do_remote(device->state);
1129 send_oos = drbd_should_send_out_of_sync(device->state);
5da9c836 1130
519b6d3e
LE
1131 /* Need to replicate writes. Unless it is an empty flush,
1132 * which is better mapped to a DRBD P_BARRIER packet,
1133 * also for drbd wire protocol compatibility reasons.
1134 * If this was a flush, just start a new epoch.
1135 * Unless the current epoch was empty anyways, or we are not currently
1136 * replicating, in which case there is no point. */
1137 if (unlikely(req->i.size == 0)) {
1138 /* The only size==0 bios we expect are empty flushes. */
28a8f0d3 1139 D_ASSERT(device, req->master_bio->bi_rw & REQ_PREFLUSH);
99b4d8fe 1140 if (remote)
7074e4a7
LE
1141 _req_mod(req, QUEUE_AS_DRBD_BARRIER);
1142 return remote;
519b6d3e
LE
1143 }
1144
5da9c836
LE
1145 if (!remote && !send_oos)
1146 return 0;
1147
0b0ba1ef 1148 D_ASSERT(device, !(remote && send_oos));
5da9c836
LE
1149
1150 if (remote) {
1151 _req_mod(req, TO_BE_SENT);
1152 _req_mod(req, QUEUE_FOR_NET_WRITE);
b30ab791 1153 } else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
5da9c836
LE
1154 _req_mod(req, QUEUE_FOR_SEND_OOS);
1155
1156 return remote;
1157}
1158
7435e901
LE
1159static void drbd_process_discard_req(struct drbd_request *req)
1160{
1161 int err = drbd_issue_discard_or_zero_out(req->device,
1162 req->i.sector, req->i.size >> 9, true);
1163
1164 if (err)
1165 req->private_bio->bi_error = -EIO;
1166 bio_endio(req->private_bio);
1167}
1168
5da9c836
LE
1169static void
1170drbd_submit_req_private_bio(struct drbd_request *req)
1171{
84b8c06b 1172 struct drbd_device *device = req->device;
5da9c836
LE
1173 struct bio *bio = req->private_bio;
1174 const int rw = bio_rw(bio);
1175
b30ab791 1176 bio->bi_bdev = device->ldev->backing_bdev;
5da9c836
LE
1177
1178 /* State may have changed since we grabbed our reference on the
1179 * ->ldev member. Double check, and short-circuit to endio.
1180 * In case the last activity log transaction failed to get on
1181 * stable storage, and this is a WRITE, we may not even submit
1182 * this bio. */
b30ab791
AG
1183 if (get_ldev(device)) {
1184 if (drbd_insert_fault(device,
5da9c836
LE
1185 rw == WRITE ? DRBD_FAULT_DT_WR
1186 : rw == READ ? DRBD_FAULT_DT_RD
1187 : DRBD_FAULT_DT_RA))
4246a0b6 1188 bio_io_error(bio);
7435e901
LE
1189 else if (bio_op(bio) == REQ_OP_DISCARD)
1190 drbd_process_discard_req(req);
5da9c836
LE
1191 else
1192 generic_make_request(bio);
b30ab791 1193 put_ldev(device);
5da9c836 1194 } else
4246a0b6 1195 bio_io_error(bio);
5da9c836
LE
1196}
1197
b30ab791 1198static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
779b3fe4 1199{
844a6ae7 1200 spin_lock_irq(&device->resource->req_lock);
b30ab791 1201 list_add_tail(&req->tl_requests, &device->submit.writes);
844a6ae7
LE
1202 list_add_tail(&req->req_pending_master_completion,
1203 &device->pending_master_completion[1 /* WRITE */]);
1204 spin_unlock_irq(&device->resource->req_lock);
b30ab791 1205 queue_work(device->submit.wq, &device->submit.worker);
f5b90b6b
LE
1206 /* do_submit() may sleep internally on al_wait, too */
1207 wake_up(&device->al_wait);
779b3fe4
LE
1208}
1209
6d9febe2
LE
1210/* returns the new drbd_request pointer, if the caller is expected to
1211 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1212 * request on the submitter thread.
1213 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1214 */
01cd2636 1215static struct drbd_request *
e5f891b2 1216drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
b411b363 1217{
6d9febe2 1218 const int rw = bio_data_dir(bio);
b411b363 1219 struct drbd_request *req;
b411b363
PR
1220
1221 /* allocate outside of all locks; */
b30ab791 1222 req = drbd_req_new(device, bio);
b411b363 1223 if (!req) {
b30ab791 1224 dec_ap_bio(device);
b411b363
PR
1225 /* only pass the error to the upper layers.
1226 * if user cannot handle io errors, that's not our business. */
d0180171 1227 drbd_err(device, "could not kmalloc() req\n");
4246a0b6
CH
1228 bio->bi_error = -ENOMEM;
1229 bio_endio(bio);
6d9febe2 1230 return ERR_PTR(-ENOMEM);
b411b363 1231 }
e5f891b2 1232 req->start_jif = start_jif;
b411b363 1233
b30ab791 1234 if (!get_ldev(device)) {
5da9c836 1235 bio_put(req->private_bio);
b411b363
PR
1236 req->private_bio = NULL;
1237 }
b411b363 1238
7e8c288f 1239 /* Update disk stats */
b30ab791 1240 _drbd_start_io_acct(device, req);
7e8c288f 1241
7435e901
LE
1242 /* process discards always from our submitter thread */
1243 if (bio_op(bio) & REQ_OP_DISCARD)
1244 goto queue_for_submitter_thread;
1245
519b6d3e 1246 if (rw == WRITE && req->private_bio && req->i.size
b30ab791 1247 && !test_bit(AL_SUSPENDED, &device->flags)) {
7435e901
LE
1248 if (!drbd_al_begin_io_fastpath(device, &req->i))
1249 goto queue_for_submitter_thread;
0778286a 1250 req->rq_state |= RQ_IN_ACT_LOG;
e5f891b2 1251 req->in_actlog_jif = jiffies;
0778286a 1252 }
6d9febe2 1253 return req;
7435e901
LE
1254
1255 queue_for_submitter_thread:
1256 atomic_inc(&device->ap_actlog_cnt);
1257 drbd_queue_write(device, req);
1258 return NULL;
6d9febe2
LE
1259}
1260
0ead5cca
LE
1261/* Require at least one path to current data.
1262 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1263 * We would not allow to read what was written,
1264 * we would not have bumped the data generation uuids,
1265 * we would cause data divergence for all the wrong reasons.
1266 *
1267 * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1268 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1269 * and queues for retry later.
1270 */
1271static bool may_do_writes(struct drbd_device *device)
1272{
1273 const union drbd_dev_state s = device->state;
1274 return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1275}
1276
b30ab791 1277static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
6d9febe2 1278{
35b5ed5b 1279 struct drbd_resource *resource = device->resource;
6d9febe2
LE
1280 const int rw = bio_rw(req->master_bio);
1281 struct bio_and_error m = { NULL, };
1282 bool no_remote = false;
35b5ed5b 1283 bool submit_private_bio = false;
6d9febe2 1284
35b5ed5b 1285 spin_lock_irq(&resource->req_lock);
6024fece 1286 if (rw == WRITE) {
648e46b5
LE
1287 /* This may temporarily give up the req_lock,
1288 * but will re-aquire it before it returns here.
1289 * Needs to be before the check on drbd_suspended() */
1290 complete_conflicting_writes(req);
607f25e5
LE
1291 /* no more giving up req_lock from now on! */
1292
1293 /* check for congestion, and potentially stop sending
1294 * full data updates, but start sending "dirty bits" only. */
b30ab791 1295 maybe_pull_ahead(device);
b411b363
PR
1296 }
1297
9a25a04c 1298
b30ab791 1299 if (drbd_suspended(device)) {
5da9c836
LE
1300 /* push back and retry: */
1301 req->rq_state |= RQ_POSTPONED;
1302 if (req->private_bio) {
1303 bio_put(req->private_bio);
1304 req->private_bio = NULL;
b30ab791 1305 put_ldev(device);
b411b363 1306 }
5da9c836 1307 goto out;
b411b363
PR
1308 }
1309
5da9c836
LE
1310 /* We fail READ/READA early, if we can not serve it.
1311 * We must do this before req is registered on any lists.
a0d856df 1312 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
5da9c836
LE
1313 if (rw != WRITE) {
1314 if (!do_remote_read(req) && !req->private_bio)
1315 goto nodata;
b411b363
PR
1316 }
1317
b6dd1a89 1318 /* which transfer log epoch does this belong to? */
a6b32bc3 1319 req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
288f422e 1320
227f052f
LE
1321 /* no point in adding empty flushes to the transfer log,
1322 * they are mapped to drbd barriers already. */
99b4d8fe
LE
1323 if (likely(req->i.size!=0)) {
1324 if (rw == WRITE)
a6b32bc3 1325 first_peer_device(device)->connection->current_tle_writes++;
288f422e 1326
a6b32bc3 1327 list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
b411b363 1328 }
67531718 1329
5da9c836 1330 if (rw == WRITE) {
0ead5cca
LE
1331 if (req->private_bio && !may_do_writes(device)) {
1332 bio_put(req->private_bio);
1333 req->private_bio = NULL;
1334 put_ldev(device);
1335 goto nodata;
1336 }
5da9c836
LE
1337 if (!drbd_process_write_request(req))
1338 no_remote = true;
1339 } else {
1340 /* We either have a private_bio, or we can read from remote.
1341 * Otherwise we had done the goto nodata above. */
1342 if (req->private_bio == NULL) {
1343 _req_mod(req, TO_BE_SENT);
1344 _req_mod(req, QUEUE_FOR_NET_READ);
6719fb03 1345 } else
5da9c836 1346 no_remote = true;
b411b363
PR
1347 }
1348
844a6ae7
LE
1349 /* If it took the fast path in drbd_request_prepare, add it here.
1350 * The slow path has added it already. */
1351 if (list_empty(&req->req_pending_master_completion))
1352 list_add_tail(&req->req_pending_master_completion,
1353 &device->pending_master_completion[rw == WRITE]);
5da9c836
LE
1354 if (req->private_bio) {
1355 /* needs to be marked within the same spinlock */
05cbbb39 1356 req->pre_submit_jif = jiffies;
844a6ae7
LE
1357 list_add_tail(&req->req_pending_local,
1358 &device->pending_completion[rw == WRITE]);
5da9c836
LE
1359 _req_mod(req, TO_BE_SUBMITTED);
1360 /* but we need to give up the spinlock to submit */
35b5ed5b 1361 submit_private_bio = true;
5da9c836
LE
1362 } else if (no_remote) {
1363nodata:
1364 if (__ratelimit(&drbd_ratelimit_state))
d0180171 1365 drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
42839f65 1366 (unsigned long long)req->i.sector, req->i.size >> 9);
5da9c836 1367 /* A write may have been queued for send_oos, however.
a0d856df 1368 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
b411b363 1369 }
b411b363 1370
5da9c836 1371out:
a0d856df
LE
1372 if (drbd_req_put_completion_ref(req, &m, 1))
1373 kref_put(&req->kref, drbd_req_destroy);
35b5ed5b
LE
1374 spin_unlock_irq(&resource->req_lock);
1375
1376 /* Even though above is a kref_put(), this is safe.
1377 * As long as we still need to submit our private bio,
1378 * we hold a completion ref, and the request cannot disappear.
1379 * If however this request did not even have a private bio to submit
1380 * (e.g. remote read), req may already be invalid now.
1381 * That's why we cannot check on req->private_bio. */
1382 if (submit_private_bio)
1383 drbd_submit_req_private_bio(req);
5da9c836 1384 if (m.bio)
b30ab791 1385 complete_master_bio(device, &m);
6d9febe2
LE
1386}
1387
e5f891b2 1388void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
6d9febe2 1389{
e5f891b2 1390 struct drbd_request *req = drbd_request_prepare(device, bio, start_jif);
6d9febe2
LE
1391 if (IS_ERR_OR_NULL(req))
1392 return;
b30ab791 1393 drbd_send_and_submit(device, req);
b411b363
PR
1394}
1395
b30ab791 1396static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
113fef9e 1397{
08a1ddab
LE
1398 struct drbd_request *req, *tmp;
1399 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1400 const int rw = bio_data_dir(req->master_bio);
113fef9e 1401
08a1ddab
LE
1402 if (rw == WRITE /* rw != WRITE should not even end up here! */
1403 && req->private_bio && req->i.size
b30ab791
AG
1404 && !test_bit(AL_SUSPENDED, &device->flags)) {
1405 if (!drbd_al_begin_io_fastpath(device, &req->i))
08a1ddab
LE
1406 continue;
1407
1408 req->rq_state |= RQ_IN_ACT_LOG;
e5f891b2 1409 req->in_actlog_jif = jiffies;
ad3fee79 1410 atomic_dec(&device->ap_actlog_cnt);
08a1ddab
LE
1411 }
1412
1413 list_del_init(&req->tl_requests);
b30ab791 1414 drbd_send_and_submit(device, req);
113fef9e 1415 }
113fef9e
LE
1416}
1417
b30ab791 1418static bool prepare_al_transaction_nonblock(struct drbd_device *device,
08a1ddab 1419 struct list_head *incoming,
f5b90b6b
LE
1420 struct list_head *pending,
1421 struct list_head *later)
08a1ddab
LE
1422{
1423 struct drbd_request *req, *tmp;
1424 int wake = 0;
1425 int err;
1426
b30ab791 1427 spin_lock_irq(&device->al_lock);
08a1ddab 1428 list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
b30ab791 1429 err = drbd_al_begin_io_nonblock(device, &req->i);
f5b90b6b
LE
1430 if (err == -ENOBUFS)
1431 break;
08a1ddab
LE
1432 if (err == -EBUSY)
1433 wake = 1;
1434 if (err)
f5b90b6b
LE
1435 list_move_tail(&req->tl_requests, later);
1436 else
1437 list_move_tail(&req->tl_requests, pending);
08a1ddab 1438 }
b30ab791 1439 spin_unlock_irq(&device->al_lock);
08a1ddab 1440 if (wake)
b30ab791 1441 wake_up(&device->al_wait);
08a1ddab
LE
1442 return !list_empty(pending);
1443}
113fef9e 1444
f5b90b6b
LE
1445void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1446{
1447 struct drbd_request *req, *tmp;
1448
1449 list_for_each_entry_safe(req, tmp, pending, tl_requests) {
1450 req->rq_state |= RQ_IN_ACT_LOG;
1451 req->in_actlog_jif = jiffies;
1452 atomic_dec(&device->ap_actlog_cnt);
1453 list_del_init(&req->tl_requests);
1454 drbd_send_and_submit(device, req);
1455 }
1456}
1457
113fef9e
LE
1458void do_submit(struct work_struct *ws)
1459{
b30ab791 1460 struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
f5b90b6b
LE
1461 LIST_HEAD(incoming); /* from drbd_make_request() */
1462 LIST_HEAD(pending); /* to be submitted after next AL-transaction commit */
1463 LIST_HEAD(busy); /* blocked by resync requests */
1464
1465 /* grab new incoming requests */
1466 spin_lock_irq(&device->resource->req_lock);
1467 list_splice_tail_init(&device->submit.writes, &incoming);
1468 spin_unlock_irq(&device->resource->req_lock);
113fef9e 1469
08a1ddab 1470 for (;;) {
f5b90b6b 1471 DEFINE_WAIT(wait);
113fef9e 1472
f5b90b6b
LE
1473 /* move used-to-be-busy back to front of incoming */
1474 list_splice_init(&busy, &incoming);
b30ab791 1475 submit_fast_path(device, &incoming);
08a1ddab
LE
1476 if (list_empty(&incoming))
1477 break;
1478
45ad07b3 1479 for (;;) {
f5b90b6b
LE
1480 prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1481
1482 list_splice_init(&busy, &incoming);
1483 prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1484 if (!list_empty(&pending))
1485 break;
1486
1487 schedule();
1488
1489 /* If all currently "hot" activity log extents are kept busy by
1490 * incoming requests, we still must not totally starve new
1491 * requests to "cold" extents.
1492 * Something left on &incoming means there had not been
1493 * enough update slots available, and the activity log
1494 * has been marked as "starving".
1495 *
1496 * Try again now, without looking for new requests,
1497 * effectively blocking all new requests until we made
1498 * at least _some_ progress with what we currently have.
1499 */
1500 if (!list_empty(&incoming))
1501 continue;
1502
1503 /* Nothing moved to pending, but nothing left
1504 * on incoming: all moved to busy!
1505 * Grab new and iterate. */
1506 spin_lock_irq(&device->resource->req_lock);
1507 list_splice_tail_init(&device->submit.writes, &incoming);
1508 spin_unlock_irq(&device->resource->req_lock);
1509 }
1510 finish_wait(&device->al_wait, &wait);
1511
1512 /* If the transaction was full, before all incoming requests
1513 * had been processed, skip ahead to commit, and iterate
1514 * without splicing in more incoming requests from upper layers.
1515 *
1516 * Else, if all incoming have been processed,
1517 * they have become either "pending" (to be submitted after
1518 * next transaction commit) or "busy" (blocked by resync).
1519 *
1520 * Maybe more was queued, while we prepared the transaction?
1521 * Try to stuff those into this transaction as well.
1522 * Be strictly non-blocking here,
1523 * we already have something to commit.
1524 *
1525 * Commit if we don't make any more progres.
1526 */
1527
1528 while (list_empty(&incoming)) {
45ad07b3
LE
1529 LIST_HEAD(more_pending);
1530 LIST_HEAD(more_incoming);
1531 bool made_progress;
1532
1533 /* It is ok to look outside the lock,
1534 * it's only an optimization anyways */
b30ab791 1535 if (list_empty(&device->submit.writes))
45ad07b3
LE
1536 break;
1537
844a6ae7 1538 spin_lock_irq(&device->resource->req_lock);
b30ab791 1539 list_splice_tail_init(&device->submit.writes, &more_incoming);
844a6ae7 1540 spin_unlock_irq(&device->resource->req_lock);
45ad07b3
LE
1541
1542 if (list_empty(&more_incoming))
1543 break;
1544
f5b90b6b 1545 made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
45ad07b3
LE
1546
1547 list_splice_tail_init(&more_pending, &pending);
1548 list_splice_tail_init(&more_incoming, &incoming);
45ad07b3
LE
1549 if (!made_progress)
1550 break;
1551 }
08a1ddab 1552
f5b90b6b
LE
1553 drbd_al_begin_io_commit(device);
1554 send_and_submit_pending(device, &pending);
113fef9e
LE
1555 }
1556}
1557
dece1635 1558blk_qc_t drbd_make_request(struct request_queue *q, struct bio *bio)
b411b363 1559{
b30ab791 1560 struct drbd_device *device = (struct drbd_device *) q->queuedata;
e5f891b2 1561 unsigned long start_jif;
b411b363 1562
54efd50b
KO
1563 blk_queue_split(q, &bio, q->bio_split);
1564
e5f891b2 1565 start_jif = jiffies;
aeda1cd6 1566
b411b363
PR
1567 /*
1568 * what we "blindly" assume:
1569 */
0b0ba1ef 1570 D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
b411b363 1571
b30ab791 1572 inc_ap_bio(device);
e5f891b2 1573 __drbd_make_request(device, bio, start_jif);
dece1635 1574 return BLK_QC_T_NONE;
b411b363
PR
1575}
1576
84d34f2f
LE
1577static bool net_timeout_reached(struct drbd_request *net_req,
1578 struct drbd_connection *connection,
1579 unsigned long now, unsigned long ent,
1580 unsigned int ko_count, unsigned int timeout)
1581{
1582 struct drbd_device *device = net_req->device;
1583
1584 if (!time_after(now, net_req->pre_send_jif + ent))
1585 return false;
1586
1587 if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1588 return false;
1589
1590 if (net_req->rq_state & RQ_NET_PENDING) {
1591 drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1592 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1593 return true;
1594 }
1595
1596 /* We received an ACK already (or are using protocol A),
1597 * but are waiting for the epoch closing barrier ack.
1598 * Check if we sent the barrier already. We should not blame the peer
1599 * for being unresponsive, if we did not even ask it yet. */
1600 if (net_req->epoch == connection->send.current_epoch_nr) {
1601 drbd_warn(device,
1602 "We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1603 jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1604 return false;
1605 }
1606
1607 /* Worst case: we may have been blocked for whatever reason, then
1608 * suddenly are able to send a lot of requests (and epoch separating
1609 * barriers) in quick succession.
1610 * The timestamp of the net_req may be much too old and not correspond
1611 * to the sending time of the relevant unack'ed barrier packet, so
1612 * would trigger a spurious timeout. The latest barrier packet may
1613 * have a too recent timestamp to trigger the timeout, potentially miss
1614 * a timeout. Right now we don't have a place to conveniently store
1615 * these timestamps.
1616 * But in this particular situation, the application requests are still
1617 * completed to upper layers, DRBD should still "feel" responsive.
1618 * No need yet to kill this connection, it may still recover.
1619 * If not, eventually we will have queued enough into the network for
1620 * us to block. From that point of view, the timestamp of the last sent
1621 * barrier packet is relevant enough.
1622 */
1623 if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1624 drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1625 connection->send.last_sent_barrier_jif, now,
1626 jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1627 return true;
1628 }
1629 return false;
1630}
1631
1632/* A request is considered timed out, if
1633 * - we have some effective timeout from the configuration,
1634 * with some state restrictions applied,
1635 * - the oldest request is waiting for a response from the network
1636 * resp. the local disk,
1637 * - the oldest request is in fact older than the effective timeout,
1638 * - the connection was established (resp. disk was attached)
1639 * for longer than the timeout already.
1640 * Note that for 32bit jiffies and very stable connections/disks,
1641 * we may have a wrap around, which is catched by
1642 * !time_in_range(now, last_..._jif, last_..._jif + timeout).
1643 *
1644 * Side effect: once per 32bit wrap-around interval, which means every
1645 * ~198 days with 250 HZ, we have a window where the timeout would need
1646 * to expire twice (worst case) to become effective. Good enough.
1647 */
1648
7fde2be9
PR
1649void request_timer_fn(unsigned long data)
1650{
b30ab791 1651 struct drbd_device *device = (struct drbd_device *) data;
a6b32bc3 1652 struct drbd_connection *connection = first_peer_device(device)->connection;
7753a4c1 1653 struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
44ed167d 1654 struct net_conf *nc;
7753a4c1 1655 unsigned long oldest_submit_jif;
dfa8bedb 1656 unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
ba280c09 1657 unsigned long now;
84d34f2f 1658 unsigned int ko_count = 0, timeout = 0;
7fde2be9 1659
44ed167d 1660 rcu_read_lock();
bde89a9e 1661 nc = rcu_dereference(connection->net_conf);
84d34f2f
LE
1662 if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1663 ko_count = nc->ko_count;
1664 timeout = nc->timeout;
1665 }
cdfda633 1666
b30ab791
AG
1667 if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1668 dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1669 put_ldev(device);
dfa8bedb 1670 }
44ed167d 1671 rcu_read_unlock();
7fde2be9 1672
84d34f2f
LE
1673
1674 ent = timeout * HZ/10 * ko_count;
dfa8bedb
PR
1675 et = min_not_zero(dt, ent);
1676
ba280c09 1677 if (!et)
7fde2be9
PR
1678 return; /* Recurring timer stopped */
1679
ba280c09 1680 now = jiffies;
7753a4c1 1681 nt = now + et;
ba280c09 1682
0500813f 1683 spin_lock_irq(&device->resource->req_lock);
7753a4c1
LE
1684 req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1685 req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
84d34f2f 1686
7753a4c1 1687 /* maybe the oldest request waiting for the peer is in fact still
84d34f2f
LE
1688 * blocking in tcp sendmsg. That's ok, though, that's handled via the
1689 * socket send timeout, requesting a ping, and bumping ko-count in
1690 * we_should_drop_the_connection().
1691 */
1692
1693 /* check the oldest request we did successfully sent,
1694 * but which is still waiting for an ACK. */
1695 req_peer = connection->req_ack_pending;
1696
1697 /* if we don't have such request (e.g. protocoll A)
1698 * check the oldest requests which is still waiting on its epoch
1699 * closing barrier ack. */
1700 if (!req_peer)
1701 req_peer = connection->req_not_net_done;
7753a4c1
LE
1702
1703 /* evaluate the oldest peer request only in one timer! */
1704 if (req_peer && req_peer->device != device)
1705 req_peer = NULL;
1706
1707 /* do we have something to evaluate? */
1708 if (req_peer == NULL && req_write == NULL && req_read == NULL)
1709 goto out;
1710
1711 oldest_submit_jif =
1712 (req_write && req_read)
1713 ? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1714 ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1715 : req_write ? req_write->pre_submit_jif
1716 : req_read ? req_read->pre_submit_jif : now;
7fde2be9 1717
84d34f2f 1718 if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
9581f97a 1719 _conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
84d34f2f 1720
7753a4c1
LE
1721 if (dt && oldest_submit_jif != now &&
1722 time_after(now, oldest_submit_jif + dt) &&
b30ab791 1723 !time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
d0180171 1724 drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
b30ab791 1725 __drbd_chk_io_error(device, DRBD_FORCE_DETACH);
dfa8bedb 1726 }
08535466
LE
1727
1728 /* Reschedule timer for the nearest not already expired timeout.
1729 * Fallback to now + min(effective network timeout, disk timeout). */
7753a4c1
LE
1730 ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1731 ? req_peer->pre_send_jif + ent : now + et;
1732 dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1733 ? oldest_submit_jif + dt : now + et;
08535466 1734 nt = time_before(ent, dt) ? ent : dt;
7753a4c1 1735out:
8d4ba3f0 1736 spin_unlock_irq(&device->resource->req_lock);
b30ab791 1737 mod_timer(&device->request_timer, nt);
7fde2be9 1738}