xref: /linux/drivers/block/drbd/drbd_req.c (revision 00a6d7b6762c27d441e9ac8faff36384bc0fc180)
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 */
37 static void _drbd_start_io_acct(struct drbd_device *device, struct drbd_request *req)
38 {
39 	const int rw = bio_data_dir(req->master_bio);
40 	int cpu;
41 	cpu = part_stat_lock();
42 	part_round_stats(cpu, &device->vdisk->part0);
43 	part_stat_inc(cpu, &device->vdisk->part0, ios[rw]);
44 	part_stat_add(cpu, &device->vdisk->part0, sectors[rw], req->i.size >> 9);
45 	(void) cpu; /* The macro invocations above want the cpu argument, I do not like
46 		       the compiler warning about cpu only assigned but never used... */
47 	part_inc_in_flight(&device->vdisk->part0, rw);
48 	part_stat_unlock();
49 }
50 
51 /* Update disk stats when completing request upwards */
52 static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
53 {
54 	int rw = bio_data_dir(req->master_bio);
55 	unsigned long duration = jiffies - req->start_time;
56 	int cpu;
57 	cpu = part_stat_lock();
58 	part_stat_add(cpu, &device->vdisk->part0, ticks[rw], duration);
59 	part_round_stats(cpu, &device->vdisk->part0);
60 	part_dec_in_flight(&device->vdisk->part0, rw);
61 	part_stat_unlock();
62 }
63 
64 static struct drbd_request *drbd_req_new(struct drbd_device *device,
65 					       struct bio *bio_src)
66 {
67 	struct drbd_request *req;
68 
69 	req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
70 	if (!req)
71 		return NULL;
72 
73 	drbd_req_make_private_bio(req, bio_src);
74 	req->rq_state    = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
75 	req->device   = device;
76 	req->master_bio  = bio_src;
77 	req->epoch       = 0;
78 
79 	drbd_clear_interval(&req->i);
80 	req->i.sector     = bio_src->bi_iter.bi_sector;
81 	req->i.size      = bio_src->bi_iter.bi_size;
82 	req->i.local = true;
83 	req->i.waiting = false;
84 
85 	INIT_LIST_HEAD(&req->tl_requests);
86 	INIT_LIST_HEAD(&req->w.list);
87 
88 	/* one reference to be put by __drbd_make_request */
89 	atomic_set(&req->completion_ref, 1);
90 	/* one kref as long as completion_ref > 0 */
91 	kref_init(&req->kref);
92 	return req;
93 }
94 
95 void drbd_req_destroy(struct kref *kref)
96 {
97 	struct drbd_request *req = container_of(kref, struct drbd_request, kref);
98 	struct drbd_device *device = req->device;
99 	const unsigned s = req->rq_state;
100 
101 	if ((req->master_bio && !(s & RQ_POSTPONED)) ||
102 		atomic_read(&req->completion_ref) ||
103 		(s & RQ_LOCAL_PENDING) ||
104 		((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
105 		drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
106 				s, atomic_read(&req->completion_ref));
107 		return;
108 	}
109 
110 	/* remove it from the transfer log.
111 	 * well, only if it had been there in the first
112 	 * place... if it had not (local only or conflicting
113 	 * and never sent), it should still be "empty" as
114 	 * initialized in drbd_req_new(), so we can list_del() it
115 	 * here unconditionally */
116 	list_del_init(&req->tl_requests);
117 
118 	/* if it was a write, we may have to set the corresponding
119 	 * bit(s) out-of-sync first. If it had a local part, we need to
120 	 * release the reference to the activity log. */
121 	if (s & RQ_WRITE) {
122 		/* Set out-of-sync unless both OK flags are set
123 		 * (local only or remote failed).
124 		 * Other places where we set out-of-sync:
125 		 * READ with local io-error */
126 
127 		/* There is a special case:
128 		 * we may notice late that IO was suspended,
129 		 * and postpone, or schedule for retry, a write,
130 		 * before it even was submitted or sent.
131 		 * In that case we do not want to touch the bitmap at all.
132 		 */
133 		if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
134 			if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
135 				drbd_set_out_of_sync(device, req->i.sector, req->i.size);
136 
137 			if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
138 				drbd_set_in_sync(device, req->i.sector, req->i.size);
139 		}
140 
141 		/* one might be tempted to move the drbd_al_complete_io
142 		 * to the local io completion callback drbd_request_endio.
143 		 * but, if this was a mirror write, we may only
144 		 * drbd_al_complete_io after this is RQ_NET_DONE,
145 		 * otherwise the extent could be dropped from the al
146 		 * before it has actually been written on the peer.
147 		 * if we crash before our peer knows about the request,
148 		 * but after the extent has been dropped from the al,
149 		 * we would forget to resync the corresponding extent.
150 		 */
151 		if (s & RQ_IN_ACT_LOG) {
152 			if (get_ldev_if_state(device, D_FAILED)) {
153 				drbd_al_complete_io(device, &req->i);
154 				put_ldev(device);
155 			} else if (__ratelimit(&drbd_ratelimit_state)) {
156 				drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
157 					 "but my Disk seems to have failed :(\n",
158 					 (unsigned long long) req->i.sector, req->i.size);
159 			}
160 		}
161 	}
162 
163 	mempool_free(req, drbd_request_mempool);
164 }
165 
166 static void wake_all_senders(struct drbd_connection *connection)
167 {
168 	wake_up(&connection->sender_work.q_wait);
169 }
170 
171 /* must hold resource->req_lock */
172 void start_new_tl_epoch(struct drbd_connection *connection)
173 {
174 	/* no point closing an epoch, if it is empty, anyways. */
175 	if (connection->current_tle_writes == 0)
176 		return;
177 
178 	connection->current_tle_writes = 0;
179 	atomic_inc(&connection->current_tle_nr);
180 	wake_all_senders(connection);
181 }
182 
183 void complete_master_bio(struct drbd_device *device,
184 		struct bio_and_error *m)
185 {
186 	bio_endio(m->bio, m->error);
187 	dec_ap_bio(device);
188 }
189 
190 
191 static void drbd_remove_request_interval(struct rb_root *root,
192 					 struct drbd_request *req)
193 {
194 	struct drbd_device *device = req->device;
195 	struct drbd_interval *i = &req->i;
196 
197 	drbd_remove_interval(root, i);
198 
199 	/* Wake up any processes waiting for this request to complete.  */
200 	if (i->waiting)
201 		wake_up(&device->misc_wait);
202 }
203 
204 /* Helper for __req_mod().
205  * Set m->bio to the master bio, if it is fit to be completed,
206  * or leave it alone (it is initialized to NULL in __req_mod),
207  * if it has already been completed, or cannot be completed yet.
208  * If m->bio is set, the error status to be returned is placed in m->error.
209  */
210 static
211 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
212 {
213 	const unsigned s = req->rq_state;
214 	struct drbd_device *device = req->device;
215 	int rw;
216 	int error, ok;
217 
218 	/* we must not complete the master bio, while it is
219 	 *	still being processed by _drbd_send_zc_bio (drbd_send_dblock)
220 	 *	not yet acknowledged by the peer
221 	 *	not yet completed by the local io subsystem
222 	 * these flags may get cleared in any order by
223 	 *	the worker,
224 	 *	the receiver,
225 	 *	the bio_endio completion callbacks.
226 	 */
227 	if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
228 	    (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
229 	    (s & RQ_COMPLETION_SUSP)) {
230 		drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
231 		return;
232 	}
233 
234 	if (!req->master_bio) {
235 		drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
236 		return;
237 	}
238 
239 	rw = bio_rw(req->master_bio);
240 
241 	/*
242 	 * figure out whether to report success or failure.
243 	 *
244 	 * report success when at least one of the operations succeeded.
245 	 * or, to put the other way,
246 	 * only report failure, when both operations failed.
247 	 *
248 	 * what to do about the failures is handled elsewhere.
249 	 * what we need to do here is just: complete the master_bio.
250 	 *
251 	 * local completion error, if any, has been stored as ERR_PTR
252 	 * in private_bio within drbd_request_endio.
253 	 */
254 	ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
255 	error = PTR_ERR(req->private_bio);
256 
257 	/* remove the request from the conflict detection
258 	 * respective block_id verification hash */
259 	if (!drbd_interval_empty(&req->i)) {
260 		struct rb_root *root;
261 
262 		if (rw == WRITE)
263 			root = &device->write_requests;
264 		else
265 			root = &device->read_requests;
266 		drbd_remove_request_interval(root, req);
267 	}
268 
269 	/* Before we can signal completion to the upper layers,
270 	 * we may need to close the current transfer log epoch.
271 	 * We are within the request lock, so we can simply compare
272 	 * the request epoch number with the current transfer log
273 	 * epoch number.  If they match, increase the current_tle_nr,
274 	 * and reset the transfer log epoch write_cnt.
275 	 */
276 	if (rw == WRITE &&
277 	    req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
278 		start_new_tl_epoch(first_peer_device(device)->connection);
279 
280 	/* Update disk stats */
281 	_drbd_end_io_acct(device, req);
282 
283 	/* If READ failed,
284 	 * have it be pushed back to the retry work queue,
285 	 * so it will re-enter __drbd_make_request(),
286 	 * and be re-assigned to a suitable local or remote path,
287 	 * or failed if we do not have access to good data anymore.
288 	 *
289 	 * Unless it was failed early by __drbd_make_request(),
290 	 * because no path was available, in which case
291 	 * it was not even added to the transfer_log.
292 	 *
293 	 * READA may fail, and will not be retried.
294 	 *
295 	 * WRITE should have used all available paths already.
296 	 */
297 	if (!ok && rw == READ && !list_empty(&req->tl_requests))
298 		req->rq_state |= RQ_POSTPONED;
299 
300 	if (!(req->rq_state & RQ_POSTPONED)) {
301 		m->error = ok ? 0 : (error ?: -EIO);
302 		m->bio = req->master_bio;
303 		req->master_bio = NULL;
304 	}
305 }
306 
307 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
308 {
309 	struct drbd_device *device = req->device;
310 	D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
311 
312 	if (!atomic_sub_and_test(put, &req->completion_ref))
313 		return 0;
314 
315 	drbd_req_complete(req, m);
316 
317 	if (req->rq_state & RQ_POSTPONED) {
318 		/* don't destroy the req object just yet,
319 		 * but queue it for retry */
320 		drbd_restart_request(req);
321 		return 0;
322 	}
323 
324 	return 1;
325 }
326 
327 /* I'd like this to be the only place that manipulates
328  * req->completion_ref and req->kref. */
329 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
330 		int clear, int set)
331 {
332 	struct drbd_device *device = req->device;
333 	unsigned s = req->rq_state;
334 	int c_put = 0;
335 	int k_put = 0;
336 
337 	if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
338 		set |= RQ_COMPLETION_SUSP;
339 
340 	/* apply */
341 
342 	req->rq_state &= ~clear;
343 	req->rq_state |= set;
344 
345 	/* no change? */
346 	if (req->rq_state == s)
347 		return;
348 
349 	/* intent: get references */
350 
351 	if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
352 		atomic_inc(&req->completion_ref);
353 
354 	if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
355 		inc_ap_pending(device);
356 		atomic_inc(&req->completion_ref);
357 	}
358 
359 	if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
360 		atomic_inc(&req->completion_ref);
361 
362 	if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
363 		kref_get(&req->kref); /* wait for the DONE */
364 
365 	if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
366 		atomic_add(req->i.size >> 9, &device->ap_in_flight);
367 
368 	if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
369 		atomic_inc(&req->completion_ref);
370 
371 	/* progress: put references */
372 
373 	if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
374 		++c_put;
375 
376 	if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
377 		D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
378 		/* local completion may still come in later,
379 		 * we need to keep the req object around. */
380 		kref_get(&req->kref);
381 		++c_put;
382 	}
383 
384 	if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
385 		if (req->rq_state & RQ_LOCAL_ABORTED)
386 			++k_put;
387 		else
388 			++c_put;
389 	}
390 
391 	if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
392 		dec_ap_pending(device);
393 		++c_put;
394 	}
395 
396 	if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
397 		++c_put;
398 
399 	if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
400 		if (req->rq_state & RQ_NET_SENT)
401 			atomic_sub(req->i.size >> 9, &device->ap_in_flight);
402 		++k_put;
403 	}
404 
405 	/* potentially complete and destroy */
406 
407 	if (k_put || c_put) {
408 		/* Completion does it's own kref_put.  If we are going to
409 		 * kref_sub below, we need req to be still around then. */
410 		int at_least = k_put + !!c_put;
411 		int refcount = atomic_read(&req->kref.refcount);
412 		if (refcount < at_least)
413 			drbd_err(device,
414 				"mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
415 				s, req->rq_state, refcount, at_least);
416 	}
417 
418 	/* If we made progress, retry conflicting peer requests, if any. */
419 	if (req->i.waiting)
420 		wake_up(&device->misc_wait);
421 
422 	if (c_put)
423 		k_put += drbd_req_put_completion_ref(req, m, c_put);
424 	if (k_put)
425 		kref_sub(&req->kref, k_put, drbd_req_destroy);
426 }
427 
428 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
429 {
430         char b[BDEVNAME_SIZE];
431 
432 	if (!__ratelimit(&drbd_ratelimit_state))
433 		return;
434 
435 	drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
436 			(req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
437 			(unsigned long long)req->i.sector,
438 			req->i.size >> 9,
439 			bdevname(device->ldev->backing_bdev, b));
440 }
441 
442 /* obviously this could be coded as many single functions
443  * instead of one huge switch,
444  * or by putting the code directly in the respective locations
445  * (as it has been before).
446  *
447  * but having it this way
448  *  enforces that it is all in this one place, where it is easier to audit,
449  *  it makes it obvious that whatever "event" "happens" to a request should
450  *  happen "atomically" within the req_lock,
451  *  and it enforces that we have to think in a very structured manner
452  *  about the "events" that may happen to a request during its life time ...
453  */
454 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
455 		struct bio_and_error *m)
456 {
457 	struct drbd_device *device = req->device;
458 	struct net_conf *nc;
459 	int p, rv = 0;
460 
461 	if (m)
462 		m->bio = NULL;
463 
464 	switch (what) {
465 	default:
466 		drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
467 		break;
468 
469 	/* does not happen...
470 	 * initialization done in drbd_req_new
471 	case CREATED:
472 		break;
473 		*/
474 
475 	case TO_BE_SENT: /* via network */
476 		/* reached via __drbd_make_request
477 		 * and from w_read_retry_remote */
478 		D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
479 		rcu_read_lock();
480 		nc = rcu_dereference(first_peer_device(device)->connection->net_conf);
481 		p = nc->wire_protocol;
482 		rcu_read_unlock();
483 		req->rq_state |=
484 			p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
485 			p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
486 		mod_rq_state(req, m, 0, RQ_NET_PENDING);
487 		break;
488 
489 	case TO_BE_SUBMITTED: /* locally */
490 		/* reached via __drbd_make_request */
491 		D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
492 		mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
493 		break;
494 
495 	case COMPLETED_OK:
496 		if (req->rq_state & RQ_WRITE)
497 			device->writ_cnt += req->i.size >> 9;
498 		else
499 			device->read_cnt += req->i.size >> 9;
500 
501 		mod_rq_state(req, m, RQ_LOCAL_PENDING,
502 				RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
503 		break;
504 
505 	case ABORT_DISK_IO:
506 		mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
507 		break;
508 
509 	case WRITE_COMPLETED_WITH_ERROR:
510 		drbd_report_io_error(device, req);
511 		__drbd_chk_io_error(device, DRBD_WRITE_ERROR);
512 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
513 		break;
514 
515 	case READ_COMPLETED_WITH_ERROR:
516 		drbd_set_out_of_sync(device, req->i.sector, req->i.size);
517 		drbd_report_io_error(device, req);
518 		__drbd_chk_io_error(device, DRBD_READ_ERROR);
519 		/* fall through. */
520 	case READ_AHEAD_COMPLETED_WITH_ERROR:
521 		/* it is legal to fail READA, no __drbd_chk_io_error in that case. */
522 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
523 		break;
524 
525 	case QUEUE_FOR_NET_READ:
526 		/* READ or READA, and
527 		 * no local disk,
528 		 * or target area marked as invalid,
529 		 * or just got an io-error. */
530 		/* from __drbd_make_request
531 		 * or from bio_endio during read io-error recovery */
532 
533 		/* So we can verify the handle in the answer packet.
534 		 * Corresponding drbd_remove_request_interval is in
535 		 * drbd_req_complete() */
536 		D_ASSERT(device, drbd_interval_empty(&req->i));
537 		drbd_insert_interval(&device->read_requests, &req->i);
538 
539 		set_bit(UNPLUG_REMOTE, &device->flags);
540 
541 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
542 		D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
543 		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
544 		req->w.cb = w_send_read_req;
545 		drbd_queue_work(&first_peer_device(device)->connection->sender_work,
546 				&req->w);
547 		break;
548 
549 	case QUEUE_FOR_NET_WRITE:
550 		/* assert something? */
551 		/* from __drbd_make_request only */
552 
553 		/* Corresponding drbd_remove_request_interval is in
554 		 * drbd_req_complete() */
555 		D_ASSERT(device, drbd_interval_empty(&req->i));
556 		drbd_insert_interval(&device->write_requests, &req->i);
557 
558 		/* NOTE
559 		 * In case the req ended up on the transfer log before being
560 		 * queued on the worker, it could lead to this request being
561 		 * missed during cleanup after connection loss.
562 		 * So we have to do both operations here,
563 		 * within the same lock that protects the transfer log.
564 		 *
565 		 * _req_add_to_epoch(req); this has to be after the
566 		 * _maybe_start_new_epoch(req); which happened in
567 		 * __drbd_make_request, because we now may set the bit
568 		 * again ourselves to close the current epoch.
569 		 *
570 		 * Add req to the (now) current epoch (barrier). */
571 
572 		/* otherwise we may lose an unplug, which may cause some remote
573 		 * io-scheduler timeout to expire, increasing maximum latency,
574 		 * hurting performance. */
575 		set_bit(UNPLUG_REMOTE, &device->flags);
576 
577 		/* queue work item to send data */
578 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
579 		mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
580 		req->w.cb =  w_send_dblock;
581 		drbd_queue_work(&first_peer_device(device)->connection->sender_work,
582 				&req->w);
583 
584 		/* close the epoch, in case it outgrew the limit */
585 		rcu_read_lock();
586 		nc = rcu_dereference(first_peer_device(device)->connection->net_conf);
587 		p = nc->max_epoch_size;
588 		rcu_read_unlock();
589 		if (first_peer_device(device)->connection->current_tle_writes >= p)
590 			start_new_tl_epoch(first_peer_device(device)->connection);
591 
592 		break;
593 
594 	case QUEUE_FOR_SEND_OOS:
595 		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
596 		req->w.cb =  w_send_out_of_sync;
597 		drbd_queue_work(&first_peer_device(device)->connection->sender_work,
598 				&req->w);
599 		break;
600 
601 	case READ_RETRY_REMOTE_CANCELED:
602 	case SEND_CANCELED:
603 	case SEND_FAILED:
604 		/* real cleanup will be done from tl_clear.  just update flags
605 		 * so it is no longer marked as on the worker queue */
606 		mod_rq_state(req, m, RQ_NET_QUEUED, 0);
607 		break;
608 
609 	case HANDED_OVER_TO_NETWORK:
610 		/* assert something? */
611 		if (bio_data_dir(req->master_bio) == WRITE &&
612 		    !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
613 			/* this is what is dangerous about protocol A:
614 			 * pretend it was successfully written on the peer. */
615 			if (req->rq_state & RQ_NET_PENDING)
616 				mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
617 			/* else: neg-ack was faster... */
618 			/* it is still not yet RQ_NET_DONE until the
619 			 * corresponding epoch barrier got acked as well,
620 			 * so we know what to dirty on connection loss */
621 		}
622 		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
623 		break;
624 
625 	case OOS_HANDED_TO_NETWORK:
626 		/* Was not set PENDING, no longer QUEUED, so is now DONE
627 		 * as far as this connection is concerned. */
628 		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
629 		break;
630 
631 	case CONNECTION_LOST_WHILE_PENDING:
632 		/* transfer log cleanup after connection loss */
633 		mod_rq_state(req, m,
634 				RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
635 				RQ_NET_DONE);
636 		break;
637 
638 	case CONFLICT_RESOLVED:
639 		/* for superseded conflicting writes of multiple primaries,
640 		 * there is no need to keep anything in the tl, potential
641 		 * node crashes are covered by the activity log.
642 		 *
643 		 * If this request had been marked as RQ_POSTPONED before,
644 		 * it will actually not be completed, but "restarted",
645 		 * resubmitted from the retry worker context. */
646 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
647 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
648 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
649 		break;
650 
651 	case WRITE_ACKED_BY_PEER_AND_SIS:
652 		req->rq_state |= RQ_NET_SIS;
653 	case WRITE_ACKED_BY_PEER:
654 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
655 		/* protocol C; successfully written on peer.
656 		 * Nothing more to do here.
657 		 * We want to keep the tl in place for all protocols, to cater
658 		 * for volatile write-back caches on lower level devices. */
659 
660 		goto ack_common;
661 	case RECV_ACKED_BY_PEER:
662 		D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
663 		/* protocol B; pretends to be successfully written on peer.
664 		 * see also notes above in HANDED_OVER_TO_NETWORK about
665 		 * protocol != C */
666 	ack_common:
667 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
668 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
669 		break;
670 
671 	case POSTPONE_WRITE:
672 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
673 		/* If this node has already detected the write conflict, the
674 		 * worker will be waiting on misc_wait.  Wake it up once this
675 		 * request has completed locally.
676 		 */
677 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
678 		req->rq_state |= RQ_POSTPONED;
679 		if (req->i.waiting)
680 			wake_up(&device->misc_wait);
681 		/* Do not clear RQ_NET_PENDING. This request will make further
682 		 * progress via restart_conflicting_writes() or
683 		 * fail_postponed_requests(). Hopefully. */
684 		break;
685 
686 	case NEG_ACKED:
687 		mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
688 		break;
689 
690 	case FAIL_FROZEN_DISK_IO:
691 		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
692 			break;
693 		mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
694 		break;
695 
696 	case RESTART_FROZEN_DISK_IO:
697 		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
698 			break;
699 
700 		mod_rq_state(req, m,
701 				RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
702 				RQ_LOCAL_PENDING);
703 
704 		rv = MR_READ;
705 		if (bio_data_dir(req->master_bio) == WRITE)
706 			rv = MR_WRITE;
707 
708 		get_ldev(device); /* always succeeds in this call path */
709 		req->w.cb = w_restart_disk_io;
710 		drbd_queue_work(&first_peer_device(device)->connection->sender_work,
711 				&req->w);
712 		break;
713 
714 	case RESEND:
715 		/* Simply complete (local only) READs. */
716 		if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
717 			mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
718 			break;
719 		}
720 
721 		/* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
722 		   before the connection loss (B&C only); only P_BARRIER_ACK
723 		   (or the local completion?) was missing when we suspended.
724 		   Throwing them out of the TL here by pretending we got a BARRIER_ACK.
725 		   During connection handshake, we ensure that the peer was not rebooted. */
726 		if (!(req->rq_state & RQ_NET_OK)) {
727 			/* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
728 			 * in that case we must not set RQ_NET_PENDING. */
729 
730 			mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
731 			if (req->w.cb) {
732 				drbd_queue_work(&first_peer_device(device)->connection->sender_work,
733 						&req->w);
734 				rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
735 			} /* else: FIXME can this happen? */
736 			break;
737 		}
738 		/* else, fall through to BARRIER_ACKED */
739 
740 	case BARRIER_ACKED:
741 		/* barrier ack for READ requests does not make sense */
742 		if (!(req->rq_state & RQ_WRITE))
743 			break;
744 
745 		if (req->rq_state & RQ_NET_PENDING) {
746 			/* barrier came in before all requests were acked.
747 			 * this is bad, because if the connection is lost now,
748 			 * we won't be able to clean them up... */
749 			drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
750 		}
751 		/* Allowed to complete requests, even while suspended.
752 		 * As this is called for all requests within a matching epoch,
753 		 * we need to filter, and only set RQ_NET_DONE for those that
754 		 * have actually been on the wire. */
755 		mod_rq_state(req, m, RQ_COMPLETION_SUSP,
756 				(req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
757 		break;
758 
759 	case DATA_RECEIVED:
760 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
761 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
762 		break;
763 
764 	case QUEUE_AS_DRBD_BARRIER:
765 		start_new_tl_epoch(first_peer_device(device)->connection);
766 		mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
767 		break;
768 	};
769 
770 	return rv;
771 }
772 
773 /* we may do a local read if:
774  * - we are consistent (of course),
775  * - or we are generally inconsistent,
776  *   BUT we are still/already IN SYNC for this area.
777  *   since size may be bigger than BM_BLOCK_SIZE,
778  *   we may need to check several bits.
779  */
780 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
781 {
782 	unsigned long sbnr, ebnr;
783 	sector_t esector, nr_sectors;
784 
785 	if (device->state.disk == D_UP_TO_DATE)
786 		return true;
787 	if (device->state.disk != D_INCONSISTENT)
788 		return false;
789 	esector = sector + (size >> 9) - 1;
790 	nr_sectors = drbd_get_capacity(device->this_bdev);
791 	D_ASSERT(device, sector  < nr_sectors);
792 	D_ASSERT(device, esector < nr_sectors);
793 
794 	sbnr = BM_SECT_TO_BIT(sector);
795 	ebnr = BM_SECT_TO_BIT(esector);
796 
797 	return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
798 }
799 
800 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
801 		enum drbd_read_balancing rbm)
802 {
803 	struct backing_dev_info *bdi;
804 	int stripe_shift;
805 
806 	switch (rbm) {
807 	case RB_CONGESTED_REMOTE:
808 		bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
809 		return bdi_read_congested(bdi);
810 	case RB_LEAST_PENDING:
811 		return atomic_read(&device->local_cnt) >
812 			atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
813 	case RB_32K_STRIPING:  /* stripe_shift = 15 */
814 	case RB_64K_STRIPING:
815 	case RB_128K_STRIPING:
816 	case RB_256K_STRIPING:
817 	case RB_512K_STRIPING:
818 	case RB_1M_STRIPING:   /* stripe_shift = 20 */
819 		stripe_shift = (rbm - RB_32K_STRIPING + 15);
820 		return (sector >> (stripe_shift - 9)) & 1;
821 	case RB_ROUND_ROBIN:
822 		return test_and_change_bit(READ_BALANCE_RR, &device->flags);
823 	case RB_PREFER_REMOTE:
824 		return true;
825 	case RB_PREFER_LOCAL:
826 	default:
827 		return false;
828 	}
829 }
830 
831 /*
832  * complete_conflicting_writes  -  wait for any conflicting write requests
833  *
834  * The write_requests tree contains all active write requests which we
835  * currently know about.  Wait for any requests to complete which conflict with
836  * the new one.
837  *
838  * Only way out: remove the conflicting intervals from the tree.
839  */
840 static void complete_conflicting_writes(struct drbd_request *req)
841 {
842 	DEFINE_WAIT(wait);
843 	struct drbd_device *device = req->device;
844 	struct drbd_interval *i;
845 	sector_t sector = req->i.sector;
846 	int size = req->i.size;
847 
848 	i = drbd_find_overlap(&device->write_requests, sector, size);
849 	if (!i)
850 		return;
851 
852 	for (;;) {
853 		prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
854 		i = drbd_find_overlap(&device->write_requests, sector, size);
855 		if (!i)
856 			break;
857 		/* Indicate to wake up device->misc_wait on progress.  */
858 		i->waiting = true;
859 		spin_unlock_irq(&device->resource->req_lock);
860 		schedule();
861 		spin_lock_irq(&device->resource->req_lock);
862 	}
863 	finish_wait(&device->misc_wait, &wait);
864 }
865 
866 /* called within req_lock and rcu_read_lock() */
867 static void maybe_pull_ahead(struct drbd_device *device)
868 {
869 	struct drbd_connection *connection = first_peer_device(device)->connection;
870 	struct net_conf *nc;
871 	bool congested = false;
872 	enum drbd_on_congestion on_congestion;
873 
874 	rcu_read_lock();
875 	nc = rcu_dereference(connection->net_conf);
876 	on_congestion = nc ? nc->on_congestion : OC_BLOCK;
877 	rcu_read_unlock();
878 	if (on_congestion == OC_BLOCK ||
879 	    connection->agreed_pro_version < 96)
880 		return;
881 
882 	/* If I don't even have good local storage, we can not reasonably try
883 	 * to pull ahead of the peer. We also need the local reference to make
884 	 * sure device->act_log is there.
885 	 */
886 	if (!get_ldev_if_state(device, D_UP_TO_DATE))
887 		return;
888 
889 	if (nc->cong_fill &&
890 	    atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
891 		drbd_info(device, "Congestion-fill threshold reached\n");
892 		congested = true;
893 	}
894 
895 	if (device->act_log->used >= nc->cong_extents) {
896 		drbd_info(device, "Congestion-extents threshold reached\n");
897 		congested = true;
898 	}
899 
900 	if (congested) {
901 		/* start a new epoch for non-mirrored writes */
902 		start_new_tl_epoch(first_peer_device(device)->connection);
903 
904 		if (on_congestion == OC_PULL_AHEAD)
905 			_drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
906 		else  /*nc->on_congestion == OC_DISCONNECT */
907 			_drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
908 	}
909 	put_ldev(device);
910 }
911 
912 /* If this returns false, and req->private_bio is still set,
913  * this should be submitted locally.
914  *
915  * If it returns false, but req->private_bio is not set,
916  * we do not have access to good data :(
917  *
918  * Otherwise, this destroys req->private_bio, if any,
919  * and returns true.
920  */
921 static bool do_remote_read(struct drbd_request *req)
922 {
923 	struct drbd_device *device = req->device;
924 	enum drbd_read_balancing rbm;
925 
926 	if (req->private_bio) {
927 		if (!drbd_may_do_local_read(device,
928 					req->i.sector, req->i.size)) {
929 			bio_put(req->private_bio);
930 			req->private_bio = NULL;
931 			put_ldev(device);
932 		}
933 	}
934 
935 	if (device->state.pdsk != D_UP_TO_DATE)
936 		return false;
937 
938 	if (req->private_bio == NULL)
939 		return true;
940 
941 	/* TODO: improve read balancing decisions, take into account drbd
942 	 * protocol, pending requests etc. */
943 
944 	rcu_read_lock();
945 	rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
946 	rcu_read_unlock();
947 
948 	if (rbm == RB_PREFER_LOCAL && req->private_bio)
949 		return false; /* submit locally */
950 
951 	if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
952 		if (req->private_bio) {
953 			bio_put(req->private_bio);
954 			req->private_bio = NULL;
955 			put_ldev(device);
956 		}
957 		return true;
958 	}
959 
960 	return false;
961 }
962 
963 /* returns number of connections (== 1, for drbd 8.4)
964  * expected to actually write this data,
965  * which does NOT include those that we are L_AHEAD for. */
966 static int drbd_process_write_request(struct drbd_request *req)
967 {
968 	struct drbd_device *device = req->device;
969 	int remote, send_oos;
970 
971 	remote = drbd_should_do_remote(device->state);
972 	send_oos = drbd_should_send_out_of_sync(device->state);
973 
974 	/* Need to replicate writes.  Unless it is an empty flush,
975 	 * which is better mapped to a DRBD P_BARRIER packet,
976 	 * also for drbd wire protocol compatibility reasons.
977 	 * If this was a flush, just start a new epoch.
978 	 * Unless the current epoch was empty anyways, or we are not currently
979 	 * replicating, in which case there is no point. */
980 	if (unlikely(req->i.size == 0)) {
981 		/* The only size==0 bios we expect are empty flushes. */
982 		D_ASSERT(device, req->master_bio->bi_rw & REQ_FLUSH);
983 		if (remote)
984 			_req_mod(req, QUEUE_AS_DRBD_BARRIER);
985 		return remote;
986 	}
987 
988 	if (!remote && !send_oos)
989 		return 0;
990 
991 	D_ASSERT(device, !(remote && send_oos));
992 
993 	if (remote) {
994 		_req_mod(req, TO_BE_SENT);
995 		_req_mod(req, QUEUE_FOR_NET_WRITE);
996 	} else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
997 		_req_mod(req, QUEUE_FOR_SEND_OOS);
998 
999 	return remote;
1000 }
1001 
1002 static void
1003 drbd_submit_req_private_bio(struct drbd_request *req)
1004 {
1005 	struct drbd_device *device = req->device;
1006 	struct bio *bio = req->private_bio;
1007 	const int rw = bio_rw(bio);
1008 
1009 	bio->bi_bdev = device->ldev->backing_bdev;
1010 
1011 	/* State may have changed since we grabbed our reference on the
1012 	 * ->ldev member. Double check, and short-circuit to endio.
1013 	 * In case the last activity log transaction failed to get on
1014 	 * stable storage, and this is a WRITE, we may not even submit
1015 	 * this bio. */
1016 	if (get_ldev(device)) {
1017 		if (drbd_insert_fault(device,
1018 				      rw == WRITE ? DRBD_FAULT_DT_WR
1019 				    : rw == READ  ? DRBD_FAULT_DT_RD
1020 				    :               DRBD_FAULT_DT_RA))
1021 			bio_endio(bio, -EIO);
1022 		else
1023 			generic_make_request(bio);
1024 		put_ldev(device);
1025 	} else
1026 		bio_endio(bio, -EIO);
1027 }
1028 
1029 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1030 {
1031 	spin_lock(&device->submit.lock);
1032 	list_add_tail(&req->tl_requests, &device->submit.writes);
1033 	spin_unlock(&device->submit.lock);
1034 	queue_work(device->submit.wq, &device->submit.worker);
1035 }
1036 
1037 /* returns the new drbd_request pointer, if the caller is expected to
1038  * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1039  * request on the submitter thread.
1040  * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1041  */
1042 static struct drbd_request *
1043 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_time)
1044 {
1045 	const int rw = bio_data_dir(bio);
1046 	struct drbd_request *req;
1047 
1048 	/* allocate outside of all locks; */
1049 	req = drbd_req_new(device, bio);
1050 	if (!req) {
1051 		dec_ap_bio(device);
1052 		/* only pass the error to the upper layers.
1053 		 * if user cannot handle io errors, that's not our business. */
1054 		drbd_err(device, "could not kmalloc() req\n");
1055 		bio_endio(bio, -ENOMEM);
1056 		return ERR_PTR(-ENOMEM);
1057 	}
1058 	req->start_time = start_time;
1059 
1060 	if (!get_ldev(device)) {
1061 		bio_put(req->private_bio);
1062 		req->private_bio = NULL;
1063 	}
1064 
1065 	/* Update disk stats */
1066 	_drbd_start_io_acct(device, req);
1067 
1068 	if (rw == WRITE && req->private_bio && req->i.size
1069 	&& !test_bit(AL_SUSPENDED, &device->flags)) {
1070 		if (!drbd_al_begin_io_fastpath(device, &req->i)) {
1071 			drbd_queue_write(device, req);
1072 			return NULL;
1073 		}
1074 		req->rq_state |= RQ_IN_ACT_LOG;
1075 	}
1076 
1077 	return req;
1078 }
1079 
1080 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1081 {
1082 	const int rw = bio_rw(req->master_bio);
1083 	struct bio_and_error m = { NULL, };
1084 	bool no_remote = false;
1085 
1086 	spin_lock_irq(&device->resource->req_lock);
1087 	if (rw == WRITE) {
1088 		/* This may temporarily give up the req_lock,
1089 		 * but will re-aquire it before it returns here.
1090 		 * Needs to be before the check on drbd_suspended() */
1091 		complete_conflicting_writes(req);
1092 		/* no more giving up req_lock from now on! */
1093 
1094 		/* check for congestion, and potentially stop sending
1095 		 * full data updates, but start sending "dirty bits" only. */
1096 		maybe_pull_ahead(device);
1097 	}
1098 
1099 
1100 	if (drbd_suspended(device)) {
1101 		/* push back and retry: */
1102 		req->rq_state |= RQ_POSTPONED;
1103 		if (req->private_bio) {
1104 			bio_put(req->private_bio);
1105 			req->private_bio = NULL;
1106 			put_ldev(device);
1107 		}
1108 		goto out;
1109 	}
1110 
1111 	/* We fail READ/READA early, if we can not serve it.
1112 	 * We must do this before req is registered on any lists.
1113 	 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1114 	if (rw != WRITE) {
1115 		if (!do_remote_read(req) && !req->private_bio)
1116 			goto nodata;
1117 	}
1118 
1119 	/* which transfer log epoch does this belong to? */
1120 	req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1121 
1122 	/* no point in adding empty flushes to the transfer log,
1123 	 * they are mapped to drbd barriers already. */
1124 	if (likely(req->i.size!=0)) {
1125 		if (rw == WRITE)
1126 			first_peer_device(device)->connection->current_tle_writes++;
1127 
1128 		list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1129 	}
1130 
1131 	if (rw == WRITE) {
1132 		if (!drbd_process_write_request(req))
1133 			no_remote = true;
1134 	} else {
1135 		/* We either have a private_bio, or we can read from remote.
1136 		 * Otherwise we had done the goto nodata above. */
1137 		if (req->private_bio == NULL) {
1138 			_req_mod(req, TO_BE_SENT);
1139 			_req_mod(req, QUEUE_FOR_NET_READ);
1140 		} else
1141 			no_remote = true;
1142 	}
1143 
1144 	if (req->private_bio) {
1145 		/* needs to be marked within the same spinlock */
1146 		_req_mod(req, TO_BE_SUBMITTED);
1147 		/* but we need to give up the spinlock to submit */
1148 		spin_unlock_irq(&device->resource->req_lock);
1149 		drbd_submit_req_private_bio(req);
1150 		spin_lock_irq(&device->resource->req_lock);
1151 	} else if (no_remote) {
1152 nodata:
1153 		if (__ratelimit(&drbd_ratelimit_state))
1154 			drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1155 					(unsigned long long)req->i.sector, req->i.size >> 9);
1156 		/* A write may have been queued for send_oos, however.
1157 		 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1158 	}
1159 
1160 out:
1161 	if (drbd_req_put_completion_ref(req, &m, 1))
1162 		kref_put(&req->kref, drbd_req_destroy);
1163 	spin_unlock_irq(&device->resource->req_lock);
1164 
1165 	if (m.bio)
1166 		complete_master_bio(device, &m);
1167 }
1168 
1169 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_time)
1170 {
1171 	struct drbd_request *req = drbd_request_prepare(device, bio, start_time);
1172 	if (IS_ERR_OR_NULL(req))
1173 		return;
1174 	drbd_send_and_submit(device, req);
1175 }
1176 
1177 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1178 {
1179 	struct drbd_request *req, *tmp;
1180 	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1181 		const int rw = bio_data_dir(req->master_bio);
1182 
1183 		if (rw == WRITE /* rw != WRITE should not even end up here! */
1184 		&& req->private_bio && req->i.size
1185 		&& !test_bit(AL_SUSPENDED, &device->flags)) {
1186 			if (!drbd_al_begin_io_fastpath(device, &req->i))
1187 				continue;
1188 
1189 			req->rq_state |= RQ_IN_ACT_LOG;
1190 		}
1191 
1192 		list_del_init(&req->tl_requests);
1193 		drbd_send_and_submit(device, req);
1194 	}
1195 }
1196 
1197 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1198 					    struct list_head *incoming,
1199 					    struct list_head *pending)
1200 {
1201 	struct drbd_request *req, *tmp;
1202 	int wake = 0;
1203 	int err;
1204 
1205 	spin_lock_irq(&device->al_lock);
1206 	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1207 		err = drbd_al_begin_io_nonblock(device, &req->i);
1208 		if (err == -EBUSY)
1209 			wake = 1;
1210 		if (err)
1211 			continue;
1212 		req->rq_state |= RQ_IN_ACT_LOG;
1213 		list_move_tail(&req->tl_requests, pending);
1214 	}
1215 	spin_unlock_irq(&device->al_lock);
1216 	if (wake)
1217 		wake_up(&device->al_wait);
1218 
1219 	return !list_empty(pending);
1220 }
1221 
1222 void do_submit(struct work_struct *ws)
1223 {
1224 	struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1225 	LIST_HEAD(incoming);
1226 	LIST_HEAD(pending);
1227 	struct drbd_request *req, *tmp;
1228 
1229 	for (;;) {
1230 		spin_lock(&device->submit.lock);
1231 		list_splice_tail_init(&device->submit.writes, &incoming);
1232 		spin_unlock(&device->submit.lock);
1233 
1234 		submit_fast_path(device, &incoming);
1235 		if (list_empty(&incoming))
1236 			break;
1237 
1238 		wait_event(device->al_wait, prepare_al_transaction_nonblock(device, &incoming, &pending));
1239 		/* Maybe more was queued, while we prepared the transaction?
1240 		 * Try to stuff them into this transaction as well.
1241 		 * Be strictly non-blocking here, no wait_event, we already
1242 		 * have something to commit.
1243 		 * Stop if we don't make any more progres.
1244 		 */
1245 		for (;;) {
1246 			LIST_HEAD(more_pending);
1247 			LIST_HEAD(more_incoming);
1248 			bool made_progress;
1249 
1250 			/* It is ok to look outside the lock,
1251 			 * it's only an optimization anyways */
1252 			if (list_empty(&device->submit.writes))
1253 				break;
1254 
1255 			spin_lock(&device->submit.lock);
1256 			list_splice_tail_init(&device->submit.writes, &more_incoming);
1257 			spin_unlock(&device->submit.lock);
1258 
1259 			if (list_empty(&more_incoming))
1260 				break;
1261 
1262 			made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending);
1263 
1264 			list_splice_tail_init(&more_pending, &pending);
1265 			list_splice_tail_init(&more_incoming, &incoming);
1266 
1267 			if (!made_progress)
1268 				break;
1269 		}
1270 		drbd_al_begin_io_commit(device, false);
1271 
1272 		list_for_each_entry_safe(req, tmp, &pending, tl_requests) {
1273 			list_del_init(&req->tl_requests);
1274 			drbd_send_and_submit(device, req);
1275 		}
1276 	}
1277 }
1278 
1279 void drbd_make_request(struct request_queue *q, struct bio *bio)
1280 {
1281 	struct drbd_device *device = (struct drbd_device *) q->queuedata;
1282 	unsigned long start_time;
1283 
1284 	start_time = jiffies;
1285 
1286 	/*
1287 	 * what we "blindly" assume:
1288 	 */
1289 	D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1290 
1291 	inc_ap_bio(device);
1292 	__drbd_make_request(device, bio, start_time);
1293 }
1294 
1295 /* This is called by bio_add_page().
1296  *
1297  * q->max_hw_sectors and other global limits are already enforced there.
1298  *
1299  * We need to call down to our lower level device,
1300  * in case it has special restrictions.
1301  *
1302  * We also may need to enforce configured max-bio-bvecs limits.
1303  *
1304  * As long as the BIO is empty we have to allow at least one bvec,
1305  * regardless of size and offset, so no need to ask lower levels.
1306  */
1307 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1308 {
1309 	struct drbd_device *device = (struct drbd_device *) q->queuedata;
1310 	unsigned int bio_size = bvm->bi_size;
1311 	int limit = DRBD_MAX_BIO_SIZE;
1312 	int backing_limit;
1313 
1314 	if (bio_size && get_ldev(device)) {
1315 		unsigned int max_hw_sectors = queue_max_hw_sectors(q);
1316 		struct request_queue * const b =
1317 			device->ldev->backing_bdev->bd_disk->queue;
1318 		if (b->merge_bvec_fn) {
1319 			backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1320 			limit = min(limit, backing_limit);
1321 		}
1322 		put_ldev(device);
1323 		if ((limit >> 9) > max_hw_sectors)
1324 			limit = max_hw_sectors << 9;
1325 	}
1326 	return limit;
1327 }
1328 
1329 static struct drbd_request *find_oldest_request(struct drbd_connection *connection)
1330 {
1331 	/* Walk the transfer log,
1332 	 * and find the oldest not yet completed request */
1333 	struct drbd_request *r;
1334 	list_for_each_entry(r, &connection->transfer_log, tl_requests) {
1335 		if (atomic_read(&r->completion_ref))
1336 			return r;
1337 	}
1338 	return NULL;
1339 }
1340 
1341 void request_timer_fn(unsigned long data)
1342 {
1343 	struct drbd_device *device = (struct drbd_device *) data;
1344 	struct drbd_connection *connection = first_peer_device(device)->connection;
1345 	struct drbd_request *req; /* oldest request */
1346 	struct net_conf *nc;
1347 	unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1348 	unsigned long now;
1349 
1350 	rcu_read_lock();
1351 	nc = rcu_dereference(connection->net_conf);
1352 	if (nc && device->state.conn >= C_WF_REPORT_PARAMS)
1353 		ent = nc->timeout * HZ/10 * nc->ko_count;
1354 
1355 	if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1356 		dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1357 		put_ldev(device);
1358 	}
1359 	rcu_read_unlock();
1360 
1361 	et = min_not_zero(dt, ent);
1362 
1363 	if (!et)
1364 		return; /* Recurring timer stopped */
1365 
1366 	now = jiffies;
1367 
1368 	spin_lock_irq(&device->resource->req_lock);
1369 	req = find_oldest_request(connection);
1370 	if (!req) {
1371 		spin_unlock_irq(&device->resource->req_lock);
1372 		mod_timer(&device->request_timer, now + et);
1373 		return;
1374 	}
1375 
1376 	/* The request is considered timed out, if
1377 	 * - we have some effective timeout from the configuration,
1378 	 *   with above state restrictions applied,
1379 	 * - the oldest request is waiting for a response from the network
1380 	 *   resp. the local disk,
1381 	 * - the oldest request is in fact older than the effective timeout,
1382 	 * - the connection was established (resp. disk was attached)
1383 	 *   for longer than the timeout already.
1384 	 * Note that for 32bit jiffies and very stable connections/disks,
1385 	 * we may have a wrap around, which is catched by
1386 	 *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1387 	 *
1388 	 * Side effect: once per 32bit wrap-around interval, which means every
1389 	 * ~198 days with 250 HZ, we have a window where the timeout would need
1390 	 * to expire twice (worst case) to become effective. Good enough.
1391 	 */
1392 	if (ent && req->rq_state & RQ_NET_PENDING &&
1393 		 time_after(now, req->start_time + ent) &&
1394 		!time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) {
1395 		drbd_warn(device, "Remote failed to finish a request within ko-count * timeout\n");
1396 		_drbd_set_state(_NS(device, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1397 	}
1398 	if (dt && req->rq_state & RQ_LOCAL_PENDING && req->device == device &&
1399 		 time_after(now, req->start_time + dt) &&
1400 		!time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1401 		drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1402 		__drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1403 	}
1404 	nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
1405 	spin_unlock_irq(&connection->resource->req_lock);
1406 	mod_timer(&device->request_timer, nt);
1407 }
1408