xref: /openbmc/linux/drivers/block/drbd/drbd_req.c (revision 52fb57e7)
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 	generic_start_io_acct(bio_data_dir(req->master_bio), req->i.size >> 9,
40 			      &device->vdisk->part0);
41 }
42 
43 /* Update disk stats when completing request upwards */
44 static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
45 {
46 	generic_end_io_acct(bio_data_dir(req->master_bio),
47 			    &device->vdisk->part0, req->start_jif);
48 }
49 
50 static struct drbd_request *drbd_req_new(struct drbd_device *device,
51 					       struct bio *bio_src)
52 {
53 	struct drbd_request *req;
54 
55 	req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
56 	if (!req)
57 		return NULL;
58 	memset(req, 0, sizeof(*req));
59 
60 	drbd_req_make_private_bio(req, bio_src);
61 	req->rq_state    = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
62 	req->device   = device;
63 	req->master_bio  = bio_src;
64 	req->epoch       = 0;
65 
66 	drbd_clear_interval(&req->i);
67 	req->i.sector     = bio_src->bi_iter.bi_sector;
68 	req->i.size      = bio_src->bi_iter.bi_size;
69 	req->i.local = true;
70 	req->i.waiting = false;
71 
72 	INIT_LIST_HEAD(&req->tl_requests);
73 	INIT_LIST_HEAD(&req->w.list);
74 	INIT_LIST_HEAD(&req->req_pending_master_completion);
75 	INIT_LIST_HEAD(&req->req_pending_local);
76 
77 	/* one reference to be put by __drbd_make_request */
78 	atomic_set(&req->completion_ref, 1);
79 	/* one kref as long as completion_ref > 0 */
80 	kref_init(&req->kref);
81 	return req;
82 }
83 
84 static void drbd_remove_request_interval(struct rb_root *root,
85 					 struct drbd_request *req)
86 {
87 	struct drbd_device *device = req->device;
88 	struct drbd_interval *i = &req->i;
89 
90 	drbd_remove_interval(root, i);
91 
92 	/* Wake up any processes waiting for this request to complete.  */
93 	if (i->waiting)
94 		wake_up(&device->misc_wait);
95 }
96 
97 void drbd_req_destroy(struct kref *kref)
98 {
99 	struct drbd_request *req = container_of(kref, struct drbd_request, kref);
100 	struct drbd_device *device = req->device;
101 	const unsigned s = req->rq_state;
102 
103 	if ((req->master_bio && !(s & RQ_POSTPONED)) ||
104 		atomic_read(&req->completion_ref) ||
105 		(s & RQ_LOCAL_PENDING) ||
106 		((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
107 		drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
108 				s, atomic_read(&req->completion_ref));
109 		return;
110 	}
111 
112 	/* If called from mod_rq_state (expected normal case) or
113 	 * drbd_send_and_submit (the less likely normal path), this holds the
114 	 * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
115 	 * though it may be still empty (never added to the transfer log).
116 	 *
117 	 * If called from do_retry(), we do NOT hold the req_lock, but we are
118 	 * still allowed to unconditionally list_del(&req->tl_requests),
119 	 * because it will be on a local on-stack list only. */
120 	list_del_init(&req->tl_requests);
121 
122 	/* finally remove the request from the conflict detection
123 	 * respective block_id verification interval tree. */
124 	if (!drbd_interval_empty(&req->i)) {
125 		struct rb_root *root;
126 
127 		if (s & RQ_WRITE)
128 			root = &device->write_requests;
129 		else
130 			root = &device->read_requests;
131 		drbd_remove_request_interval(root, req);
132 	} else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
133 		drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
134 			s, (unsigned long long)req->i.sector, req->i.size);
135 
136 	/* if it was a write, we may have to set the corresponding
137 	 * bit(s) out-of-sync first. If it had a local part, we need to
138 	 * release the reference to the activity log. */
139 	if (s & RQ_WRITE) {
140 		/* Set out-of-sync unless both OK flags are set
141 		 * (local only or remote failed).
142 		 * Other places where we set out-of-sync:
143 		 * READ with local io-error */
144 
145 		/* There is a special case:
146 		 * we may notice late that IO was suspended,
147 		 * and postpone, or schedule for retry, a write,
148 		 * before it even was submitted or sent.
149 		 * In that case we do not want to touch the bitmap at all.
150 		 */
151 		if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
152 			if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
153 				drbd_set_out_of_sync(device, req->i.sector, req->i.size);
154 
155 			if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
156 				drbd_set_in_sync(device, req->i.sector, req->i.size);
157 		}
158 
159 		/* one might be tempted to move the drbd_al_complete_io
160 		 * to the local io completion callback drbd_request_endio.
161 		 * but, if this was a mirror write, we may only
162 		 * drbd_al_complete_io after this is RQ_NET_DONE,
163 		 * otherwise the extent could be dropped from the al
164 		 * before it has actually been written on the peer.
165 		 * if we crash before our peer knows about the request,
166 		 * but after the extent has been dropped from the al,
167 		 * we would forget to resync the corresponding extent.
168 		 */
169 		if (s & RQ_IN_ACT_LOG) {
170 			if (get_ldev_if_state(device, D_FAILED)) {
171 				drbd_al_complete_io(device, &req->i);
172 				put_ldev(device);
173 			} else if (__ratelimit(&drbd_ratelimit_state)) {
174 				drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
175 					 "but my Disk seems to have failed :(\n",
176 					 (unsigned long long) req->i.sector, req->i.size);
177 			}
178 		}
179 	}
180 
181 	mempool_free(req, drbd_request_mempool);
182 }
183 
184 static void wake_all_senders(struct drbd_connection *connection)
185 {
186 	wake_up(&connection->sender_work.q_wait);
187 }
188 
189 /* must hold resource->req_lock */
190 void start_new_tl_epoch(struct drbd_connection *connection)
191 {
192 	/* no point closing an epoch, if it is empty, anyways. */
193 	if (connection->current_tle_writes == 0)
194 		return;
195 
196 	connection->current_tle_writes = 0;
197 	atomic_inc(&connection->current_tle_nr);
198 	wake_all_senders(connection);
199 }
200 
201 void complete_master_bio(struct drbd_device *device,
202 		struct bio_and_error *m)
203 {
204 	bio_endio(m->bio, m->error);
205 	dec_ap_bio(device);
206 }
207 
208 
209 /* Helper for __req_mod().
210  * Set m->bio to the master bio, if it is fit to be completed,
211  * or leave it alone (it is initialized to NULL in __req_mod),
212  * if it has already been completed, or cannot be completed yet.
213  * If m->bio is set, the error status to be returned is placed in m->error.
214  */
215 static
216 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
217 {
218 	const unsigned s = req->rq_state;
219 	struct drbd_device *device = req->device;
220 	int rw;
221 	int error, ok;
222 
223 	/* we must not complete the master bio, while it is
224 	 *	still being processed by _drbd_send_zc_bio (drbd_send_dblock)
225 	 *	not yet acknowledged by the peer
226 	 *	not yet completed by the local io subsystem
227 	 * these flags may get cleared in any order by
228 	 *	the worker,
229 	 *	the receiver,
230 	 *	the bio_endio completion callbacks.
231 	 */
232 	if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
233 	    (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
234 	    (s & RQ_COMPLETION_SUSP)) {
235 		drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
236 		return;
237 	}
238 
239 	if (!req->master_bio) {
240 		drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
241 		return;
242 	}
243 
244 	rw = bio_rw(req->master_bio);
245 
246 	/*
247 	 * figure out whether to report success or failure.
248 	 *
249 	 * report success when at least one of the operations succeeded.
250 	 * or, to put the other way,
251 	 * only report failure, when both operations failed.
252 	 *
253 	 * what to do about the failures is handled elsewhere.
254 	 * what we need to do here is just: complete the master_bio.
255 	 *
256 	 * local completion error, if any, has been stored as ERR_PTR
257 	 * in private_bio within drbd_request_endio.
258 	 */
259 	ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
260 	error = PTR_ERR(req->private_bio);
261 
262 	/* Before we can signal completion to the upper layers,
263 	 * we may need to close the current transfer log epoch.
264 	 * We are within the request lock, so we can simply compare
265 	 * the request epoch number with the current transfer log
266 	 * epoch number.  If they match, increase the current_tle_nr,
267 	 * and reset the transfer log epoch write_cnt.
268 	 */
269 	if (rw == WRITE &&
270 	    req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
271 		start_new_tl_epoch(first_peer_device(device)->connection);
272 
273 	/* Update disk stats */
274 	_drbd_end_io_acct(device, req);
275 
276 	/* If READ failed,
277 	 * have it be pushed back to the retry work queue,
278 	 * so it will re-enter __drbd_make_request(),
279 	 * and be re-assigned to a suitable local or remote path,
280 	 * or failed if we do not have access to good data anymore.
281 	 *
282 	 * Unless it was failed early by __drbd_make_request(),
283 	 * because no path was available, in which case
284 	 * it was not even added to the transfer_log.
285 	 *
286 	 * READA may fail, and will not be retried.
287 	 *
288 	 * WRITE should have used all available paths already.
289 	 */
290 	if (!ok && rw == READ && !list_empty(&req->tl_requests))
291 		req->rq_state |= RQ_POSTPONED;
292 
293 	if (!(req->rq_state & RQ_POSTPONED)) {
294 		m->error = ok ? 0 : (error ?: -EIO);
295 		m->bio = req->master_bio;
296 		req->master_bio = NULL;
297 		/* We leave it in the tree, to be able to verify later
298 		 * write-acks in protocol != C during resync.
299 		 * But we mark it as "complete", so it won't be counted as
300 		 * conflict in a multi-primary setup. */
301 		req->i.completed = true;
302 	}
303 
304 	if (req->i.waiting)
305 		wake_up(&device->misc_wait);
306 
307 	/* Either we are about to complete to upper layers,
308 	 * or we will restart this request.
309 	 * In either case, the request object will be destroyed soon,
310 	 * so better remove it from all lists. */
311 	list_del_init(&req->req_pending_master_completion);
312 }
313 
314 /* still holds resource->req_lock */
315 static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
316 {
317 	struct drbd_device *device = req->device;
318 	D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
319 
320 	if (!atomic_sub_and_test(put, &req->completion_ref))
321 		return 0;
322 
323 	drbd_req_complete(req, m);
324 
325 	if (req->rq_state & RQ_POSTPONED) {
326 		/* don't destroy the req object just yet,
327 		 * but queue it for retry */
328 		drbd_restart_request(req);
329 		return 0;
330 	}
331 
332 	return 1;
333 }
334 
335 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
336 {
337 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
338 	if (!connection)
339 		return;
340 	if (connection->req_next == NULL)
341 		connection->req_next = req;
342 }
343 
344 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
345 {
346 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
347 	if (!connection)
348 		return;
349 	if (connection->req_next != req)
350 		return;
351 	list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
352 		const unsigned s = req->rq_state;
353 		if (s & RQ_NET_QUEUED)
354 			break;
355 	}
356 	if (&req->tl_requests == &connection->transfer_log)
357 		req = NULL;
358 	connection->req_next = req;
359 }
360 
361 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
362 {
363 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
364 	if (!connection)
365 		return;
366 	if (connection->req_ack_pending == NULL)
367 		connection->req_ack_pending = req;
368 }
369 
370 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
371 {
372 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
373 	if (!connection)
374 		return;
375 	if (connection->req_ack_pending != req)
376 		return;
377 	list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
378 		const unsigned s = req->rq_state;
379 		if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING))
380 			break;
381 	}
382 	if (&req->tl_requests == &connection->transfer_log)
383 		req = NULL;
384 	connection->req_ack_pending = req;
385 }
386 
387 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
388 {
389 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
390 	if (!connection)
391 		return;
392 	if (connection->req_not_net_done == NULL)
393 		connection->req_not_net_done = req;
394 }
395 
396 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
397 {
398 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
399 	if (!connection)
400 		return;
401 	if (connection->req_not_net_done != req)
402 		return;
403 	list_for_each_entry_continue(req, &connection->transfer_log, tl_requests) {
404 		const unsigned s = req->rq_state;
405 		if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE))
406 			break;
407 	}
408 	if (&req->tl_requests == &connection->transfer_log)
409 		req = NULL;
410 	connection->req_not_net_done = req;
411 }
412 
413 /* I'd like this to be the only place that manipulates
414  * req->completion_ref and req->kref. */
415 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
416 		int clear, int set)
417 {
418 	struct drbd_device *device = req->device;
419 	struct drbd_peer_device *peer_device = first_peer_device(device);
420 	unsigned s = req->rq_state;
421 	int c_put = 0;
422 	int k_put = 0;
423 
424 	if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
425 		set |= RQ_COMPLETION_SUSP;
426 
427 	/* apply */
428 
429 	req->rq_state &= ~clear;
430 	req->rq_state |= set;
431 
432 	/* no change? */
433 	if (req->rq_state == s)
434 		return;
435 
436 	/* intent: get references */
437 
438 	if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
439 		atomic_inc(&req->completion_ref);
440 
441 	if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
442 		inc_ap_pending(device);
443 		atomic_inc(&req->completion_ref);
444 	}
445 
446 	if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
447 		atomic_inc(&req->completion_ref);
448 		set_if_null_req_next(peer_device, req);
449 	}
450 
451 	if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
452 		kref_get(&req->kref); /* wait for the DONE */
453 
454 	if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
455 		/* potentially already completed in the asender thread */
456 		if (!(s & RQ_NET_DONE)) {
457 			atomic_add(req->i.size >> 9, &device->ap_in_flight);
458 			set_if_null_req_not_net_done(peer_device, req);
459 		}
460 		if (s & RQ_NET_PENDING)
461 			set_if_null_req_ack_pending(peer_device, req);
462 	}
463 
464 	if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
465 		atomic_inc(&req->completion_ref);
466 
467 	/* progress: put references */
468 
469 	if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
470 		++c_put;
471 
472 	if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
473 		D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
474 		/* local completion may still come in later,
475 		 * we need to keep the req object around. */
476 		kref_get(&req->kref);
477 		++c_put;
478 	}
479 
480 	if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
481 		if (req->rq_state & RQ_LOCAL_ABORTED)
482 			++k_put;
483 		else
484 			++c_put;
485 		list_del_init(&req->req_pending_local);
486 	}
487 
488 	if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
489 		dec_ap_pending(device);
490 		++c_put;
491 		req->acked_jif = jiffies;
492 		advance_conn_req_ack_pending(peer_device, req);
493 	}
494 
495 	if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
496 		++c_put;
497 		advance_conn_req_next(peer_device, req);
498 	}
499 
500 	if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
501 		if (s & RQ_NET_SENT)
502 			atomic_sub(req->i.size >> 9, &device->ap_in_flight);
503 		if (s & RQ_EXP_BARR_ACK)
504 			++k_put;
505 		req->net_done_jif = jiffies;
506 
507 		/* in ahead/behind mode, or just in case,
508 		 * before we finally destroy this request,
509 		 * the caching pointers must not reference it anymore */
510 		advance_conn_req_next(peer_device, req);
511 		advance_conn_req_ack_pending(peer_device, req);
512 		advance_conn_req_not_net_done(peer_device, req);
513 	}
514 
515 	/* potentially complete and destroy */
516 
517 	if (k_put || c_put) {
518 		/* Completion does it's own kref_put.  If we are going to
519 		 * kref_sub below, we need req to be still around then. */
520 		int at_least = k_put + !!c_put;
521 		int refcount = atomic_read(&req->kref.refcount);
522 		if (refcount < at_least)
523 			drbd_err(device,
524 				"mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
525 				s, req->rq_state, refcount, at_least);
526 	}
527 
528 	/* If we made progress, retry conflicting peer requests, if any. */
529 	if (req->i.waiting)
530 		wake_up(&device->misc_wait);
531 
532 	if (c_put)
533 		k_put += drbd_req_put_completion_ref(req, m, c_put);
534 	if (k_put)
535 		kref_sub(&req->kref, k_put, drbd_req_destroy);
536 }
537 
538 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
539 {
540         char b[BDEVNAME_SIZE];
541 
542 	if (!__ratelimit(&drbd_ratelimit_state))
543 		return;
544 
545 	drbd_warn(device, "local %s IO error sector %llu+%u on %s\n",
546 			(req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
547 			(unsigned long long)req->i.sector,
548 			req->i.size >> 9,
549 			bdevname(device->ldev->backing_bdev, b));
550 }
551 
552 /* Helper for HANDED_OVER_TO_NETWORK.
553  * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
554  * Is it also still "PENDING"?
555  * --> If so, clear PENDING and set NET_OK below.
556  * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
557  * (and we must not set RQ_NET_OK) */
558 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
559 {
560 	return (req->rq_state &
561 		   (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
562 		== (RQ_WRITE|RQ_NET_PENDING);
563 }
564 
565 /* obviously this could be coded as many single functions
566  * instead of one huge switch,
567  * or by putting the code directly in the respective locations
568  * (as it has been before).
569  *
570  * but having it this way
571  *  enforces that it is all in this one place, where it is easier to audit,
572  *  it makes it obvious that whatever "event" "happens" to a request should
573  *  happen "atomically" within the req_lock,
574  *  and it enforces that we have to think in a very structured manner
575  *  about the "events" that may happen to a request during its life time ...
576  */
577 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
578 		struct bio_and_error *m)
579 {
580 	struct drbd_device *const device = req->device;
581 	struct drbd_peer_device *const peer_device = first_peer_device(device);
582 	struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
583 	struct net_conf *nc;
584 	int p, rv = 0;
585 
586 	if (m)
587 		m->bio = NULL;
588 
589 	switch (what) {
590 	default:
591 		drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
592 		break;
593 
594 	/* does not happen...
595 	 * initialization done in drbd_req_new
596 	case CREATED:
597 		break;
598 		*/
599 
600 	case TO_BE_SENT: /* via network */
601 		/* reached via __drbd_make_request
602 		 * and from w_read_retry_remote */
603 		D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
604 		rcu_read_lock();
605 		nc = rcu_dereference(connection->net_conf);
606 		p = nc->wire_protocol;
607 		rcu_read_unlock();
608 		req->rq_state |=
609 			p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
610 			p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
611 		mod_rq_state(req, m, 0, RQ_NET_PENDING);
612 		break;
613 
614 	case TO_BE_SUBMITTED: /* locally */
615 		/* reached via __drbd_make_request */
616 		D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
617 		mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
618 		break;
619 
620 	case COMPLETED_OK:
621 		if (req->rq_state & RQ_WRITE)
622 			device->writ_cnt += req->i.size >> 9;
623 		else
624 			device->read_cnt += req->i.size >> 9;
625 
626 		mod_rq_state(req, m, RQ_LOCAL_PENDING,
627 				RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
628 		break;
629 
630 	case ABORT_DISK_IO:
631 		mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
632 		break;
633 
634 	case WRITE_COMPLETED_WITH_ERROR:
635 		drbd_report_io_error(device, req);
636 		__drbd_chk_io_error(device, DRBD_WRITE_ERROR);
637 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
638 		break;
639 
640 	case READ_COMPLETED_WITH_ERROR:
641 		drbd_set_out_of_sync(device, req->i.sector, req->i.size);
642 		drbd_report_io_error(device, req);
643 		__drbd_chk_io_error(device, DRBD_READ_ERROR);
644 		/* fall through. */
645 	case READ_AHEAD_COMPLETED_WITH_ERROR:
646 		/* it is legal to fail READA, no __drbd_chk_io_error in that case. */
647 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
648 		break;
649 
650 	case DISCARD_COMPLETED_NOTSUPP:
651 	case DISCARD_COMPLETED_WITH_ERROR:
652 		/* I'd rather not detach from local disk just because it
653 		 * failed a REQ_DISCARD. */
654 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
655 		break;
656 
657 	case QUEUE_FOR_NET_READ:
658 		/* READ or READA, and
659 		 * no local disk,
660 		 * or target area marked as invalid,
661 		 * or just got an io-error. */
662 		/* from __drbd_make_request
663 		 * or from bio_endio during read io-error recovery */
664 
665 		/* So we can verify the handle in the answer packet.
666 		 * Corresponding drbd_remove_request_interval is in
667 		 * drbd_req_complete() */
668 		D_ASSERT(device, drbd_interval_empty(&req->i));
669 		drbd_insert_interval(&device->read_requests, &req->i);
670 
671 		set_bit(UNPLUG_REMOTE, &device->flags);
672 
673 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
674 		D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
675 		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
676 		req->w.cb = w_send_read_req;
677 		drbd_queue_work(&connection->sender_work,
678 				&req->w);
679 		break;
680 
681 	case QUEUE_FOR_NET_WRITE:
682 		/* assert something? */
683 		/* from __drbd_make_request only */
684 
685 		/* Corresponding drbd_remove_request_interval is in
686 		 * drbd_req_complete() */
687 		D_ASSERT(device, drbd_interval_empty(&req->i));
688 		drbd_insert_interval(&device->write_requests, &req->i);
689 
690 		/* NOTE
691 		 * In case the req ended up on the transfer log before being
692 		 * queued on the worker, it could lead to this request being
693 		 * missed during cleanup after connection loss.
694 		 * So we have to do both operations here,
695 		 * within the same lock that protects the transfer log.
696 		 *
697 		 * _req_add_to_epoch(req); this has to be after the
698 		 * _maybe_start_new_epoch(req); which happened in
699 		 * __drbd_make_request, because we now may set the bit
700 		 * again ourselves to close the current epoch.
701 		 *
702 		 * Add req to the (now) current epoch (barrier). */
703 
704 		/* otherwise we may lose an unplug, which may cause some remote
705 		 * io-scheduler timeout to expire, increasing maximum latency,
706 		 * hurting performance. */
707 		set_bit(UNPLUG_REMOTE, &device->flags);
708 
709 		/* queue work item to send data */
710 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
711 		mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
712 		req->w.cb =  w_send_dblock;
713 		drbd_queue_work(&connection->sender_work,
714 				&req->w);
715 
716 		/* close the epoch, in case it outgrew the limit */
717 		rcu_read_lock();
718 		nc = rcu_dereference(connection->net_conf);
719 		p = nc->max_epoch_size;
720 		rcu_read_unlock();
721 		if (connection->current_tle_writes >= p)
722 			start_new_tl_epoch(connection);
723 
724 		break;
725 
726 	case QUEUE_FOR_SEND_OOS:
727 		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
728 		req->w.cb =  w_send_out_of_sync;
729 		drbd_queue_work(&connection->sender_work,
730 				&req->w);
731 		break;
732 
733 	case READ_RETRY_REMOTE_CANCELED:
734 	case SEND_CANCELED:
735 	case SEND_FAILED:
736 		/* real cleanup will be done from tl_clear.  just update flags
737 		 * so it is no longer marked as on the worker queue */
738 		mod_rq_state(req, m, RQ_NET_QUEUED, 0);
739 		break;
740 
741 	case HANDED_OVER_TO_NETWORK:
742 		/* assert something? */
743 		if (is_pending_write_protocol_A(req))
744 			/* this is what is dangerous about protocol A:
745 			 * pretend it was successfully written on the peer. */
746 			mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
747 						RQ_NET_SENT|RQ_NET_OK);
748 		else
749 			mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
750 		/* It is still not yet RQ_NET_DONE until the
751 		 * corresponding epoch barrier got acked as well,
752 		 * so we know what to dirty on connection loss. */
753 		break;
754 
755 	case OOS_HANDED_TO_NETWORK:
756 		/* Was not set PENDING, no longer QUEUED, so is now DONE
757 		 * as far as this connection is concerned. */
758 		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
759 		break;
760 
761 	case CONNECTION_LOST_WHILE_PENDING:
762 		/* transfer log cleanup after connection loss */
763 		mod_rq_state(req, m,
764 				RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
765 				RQ_NET_DONE);
766 		break;
767 
768 	case CONFLICT_RESOLVED:
769 		/* for superseded conflicting writes of multiple primaries,
770 		 * there is no need to keep anything in the tl, potential
771 		 * node crashes are covered by the activity log.
772 		 *
773 		 * If this request had been marked as RQ_POSTPONED before,
774 		 * it will actually not be completed, but "restarted",
775 		 * resubmitted from the retry worker context. */
776 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
777 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
778 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
779 		break;
780 
781 	case WRITE_ACKED_BY_PEER_AND_SIS:
782 		req->rq_state |= RQ_NET_SIS;
783 	case WRITE_ACKED_BY_PEER:
784 		/* Normal operation protocol C: successfully written on peer.
785 		 * During resync, even in protocol != C,
786 		 * we requested an explicit write ack anyways.
787 		 * Which means we cannot even assert anything here.
788 		 * Nothing more to do here.
789 		 * We want to keep the tl in place for all protocols, to cater
790 		 * for volatile write-back caches on lower level devices. */
791 		goto ack_common;
792 	case RECV_ACKED_BY_PEER:
793 		D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
794 		/* protocol B; pretends to be successfully written on peer.
795 		 * see also notes above in HANDED_OVER_TO_NETWORK about
796 		 * protocol != C */
797 	ack_common:
798 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
799 		break;
800 
801 	case POSTPONE_WRITE:
802 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
803 		/* If this node has already detected the write conflict, the
804 		 * worker will be waiting on misc_wait.  Wake it up once this
805 		 * request has completed locally.
806 		 */
807 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
808 		req->rq_state |= RQ_POSTPONED;
809 		if (req->i.waiting)
810 			wake_up(&device->misc_wait);
811 		/* Do not clear RQ_NET_PENDING. This request will make further
812 		 * progress via restart_conflicting_writes() or
813 		 * fail_postponed_requests(). Hopefully. */
814 		break;
815 
816 	case NEG_ACKED:
817 		mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
818 		break;
819 
820 	case FAIL_FROZEN_DISK_IO:
821 		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
822 			break;
823 		mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
824 		break;
825 
826 	case RESTART_FROZEN_DISK_IO:
827 		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
828 			break;
829 
830 		mod_rq_state(req, m,
831 				RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
832 				RQ_LOCAL_PENDING);
833 
834 		rv = MR_READ;
835 		if (bio_data_dir(req->master_bio) == WRITE)
836 			rv = MR_WRITE;
837 
838 		get_ldev(device); /* always succeeds in this call path */
839 		req->w.cb = w_restart_disk_io;
840 		drbd_queue_work(&connection->sender_work,
841 				&req->w);
842 		break;
843 
844 	case RESEND:
845 		/* Simply complete (local only) READs. */
846 		if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
847 			mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
848 			break;
849 		}
850 
851 		/* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
852 		   before the connection loss (B&C only); only P_BARRIER_ACK
853 		   (or the local completion?) was missing when we suspended.
854 		   Throwing them out of the TL here by pretending we got a BARRIER_ACK.
855 		   During connection handshake, we ensure that the peer was not rebooted. */
856 		if (!(req->rq_state & RQ_NET_OK)) {
857 			/* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
858 			 * in that case we must not set RQ_NET_PENDING. */
859 
860 			mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
861 			if (req->w.cb) {
862 				/* w.cb expected to be w_send_dblock, or w_send_read_req */
863 				drbd_queue_work(&connection->sender_work,
864 						&req->w);
865 				rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
866 			} /* else: FIXME can this happen? */
867 			break;
868 		}
869 		/* else, fall through to BARRIER_ACKED */
870 
871 	case BARRIER_ACKED:
872 		/* barrier ack for READ requests does not make sense */
873 		if (!(req->rq_state & RQ_WRITE))
874 			break;
875 
876 		if (req->rq_state & RQ_NET_PENDING) {
877 			/* barrier came in before all requests were acked.
878 			 * this is bad, because if the connection is lost now,
879 			 * we won't be able to clean them up... */
880 			drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
881 		}
882 		/* Allowed to complete requests, even while suspended.
883 		 * As this is called for all requests within a matching epoch,
884 		 * we need to filter, and only set RQ_NET_DONE for those that
885 		 * have actually been on the wire. */
886 		mod_rq_state(req, m, RQ_COMPLETION_SUSP,
887 				(req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
888 		break;
889 
890 	case DATA_RECEIVED:
891 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
892 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
893 		break;
894 
895 	case QUEUE_AS_DRBD_BARRIER:
896 		start_new_tl_epoch(connection);
897 		mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
898 		break;
899 	};
900 
901 	return rv;
902 }
903 
904 /* we may do a local read if:
905  * - we are consistent (of course),
906  * - or we are generally inconsistent,
907  *   BUT we are still/already IN SYNC for this area.
908  *   since size may be bigger than BM_BLOCK_SIZE,
909  *   we may need to check several bits.
910  */
911 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
912 {
913 	unsigned long sbnr, ebnr;
914 	sector_t esector, nr_sectors;
915 
916 	if (device->state.disk == D_UP_TO_DATE)
917 		return true;
918 	if (device->state.disk != D_INCONSISTENT)
919 		return false;
920 	esector = sector + (size >> 9) - 1;
921 	nr_sectors = drbd_get_capacity(device->this_bdev);
922 	D_ASSERT(device, sector  < nr_sectors);
923 	D_ASSERT(device, esector < nr_sectors);
924 
925 	sbnr = BM_SECT_TO_BIT(sector);
926 	ebnr = BM_SECT_TO_BIT(esector);
927 
928 	return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
929 }
930 
931 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
932 		enum drbd_read_balancing rbm)
933 {
934 	struct backing_dev_info *bdi;
935 	int stripe_shift;
936 
937 	switch (rbm) {
938 	case RB_CONGESTED_REMOTE:
939 		bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
940 		return bdi_read_congested(bdi);
941 	case RB_LEAST_PENDING:
942 		return atomic_read(&device->local_cnt) >
943 			atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
944 	case RB_32K_STRIPING:  /* stripe_shift = 15 */
945 	case RB_64K_STRIPING:
946 	case RB_128K_STRIPING:
947 	case RB_256K_STRIPING:
948 	case RB_512K_STRIPING:
949 	case RB_1M_STRIPING:   /* stripe_shift = 20 */
950 		stripe_shift = (rbm - RB_32K_STRIPING + 15);
951 		return (sector >> (stripe_shift - 9)) & 1;
952 	case RB_ROUND_ROBIN:
953 		return test_and_change_bit(READ_BALANCE_RR, &device->flags);
954 	case RB_PREFER_REMOTE:
955 		return true;
956 	case RB_PREFER_LOCAL:
957 	default:
958 		return false;
959 	}
960 }
961 
962 /*
963  * complete_conflicting_writes  -  wait for any conflicting write requests
964  *
965  * The write_requests tree contains all active write requests which we
966  * currently know about.  Wait for any requests to complete which conflict with
967  * the new one.
968  *
969  * Only way out: remove the conflicting intervals from the tree.
970  */
971 static void complete_conflicting_writes(struct drbd_request *req)
972 {
973 	DEFINE_WAIT(wait);
974 	struct drbd_device *device = req->device;
975 	struct drbd_interval *i;
976 	sector_t sector = req->i.sector;
977 	int size = req->i.size;
978 
979 	i = drbd_find_overlap(&device->write_requests, sector, size);
980 	if (!i)
981 		return;
982 
983 	for (;;) {
984 		prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
985 		i = drbd_find_overlap(&device->write_requests, sector, size);
986 		if (!i)
987 			break;
988 		/* Indicate to wake up device->misc_wait on progress.  */
989 		i->waiting = true;
990 		spin_unlock_irq(&device->resource->req_lock);
991 		schedule();
992 		spin_lock_irq(&device->resource->req_lock);
993 	}
994 	finish_wait(&device->misc_wait, &wait);
995 }
996 
997 /* called within req_lock and rcu_read_lock() */
998 static void maybe_pull_ahead(struct drbd_device *device)
999 {
1000 	struct drbd_connection *connection = first_peer_device(device)->connection;
1001 	struct net_conf *nc;
1002 	bool congested = false;
1003 	enum drbd_on_congestion on_congestion;
1004 
1005 	rcu_read_lock();
1006 	nc = rcu_dereference(connection->net_conf);
1007 	on_congestion = nc ? nc->on_congestion : OC_BLOCK;
1008 	rcu_read_unlock();
1009 	if (on_congestion == OC_BLOCK ||
1010 	    connection->agreed_pro_version < 96)
1011 		return;
1012 
1013 	if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1014 		return; /* nothing to do ... */
1015 
1016 	/* If I don't even have good local storage, we can not reasonably try
1017 	 * to pull ahead of the peer. We also need the local reference to make
1018 	 * sure device->act_log is there.
1019 	 */
1020 	if (!get_ldev_if_state(device, D_UP_TO_DATE))
1021 		return;
1022 
1023 	if (nc->cong_fill &&
1024 	    atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1025 		drbd_info(device, "Congestion-fill threshold reached\n");
1026 		congested = true;
1027 	}
1028 
1029 	if (device->act_log->used >= nc->cong_extents) {
1030 		drbd_info(device, "Congestion-extents threshold reached\n");
1031 		congested = true;
1032 	}
1033 
1034 	if (congested) {
1035 		/* start a new epoch for non-mirrored writes */
1036 		start_new_tl_epoch(first_peer_device(device)->connection);
1037 
1038 		if (on_congestion == OC_PULL_AHEAD)
1039 			_drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1040 		else  /*nc->on_congestion == OC_DISCONNECT */
1041 			_drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1042 	}
1043 	put_ldev(device);
1044 }
1045 
1046 /* If this returns false, and req->private_bio is still set,
1047  * this should be submitted locally.
1048  *
1049  * If it returns false, but req->private_bio is not set,
1050  * we do not have access to good data :(
1051  *
1052  * Otherwise, this destroys req->private_bio, if any,
1053  * and returns true.
1054  */
1055 static bool do_remote_read(struct drbd_request *req)
1056 {
1057 	struct drbd_device *device = req->device;
1058 	enum drbd_read_balancing rbm;
1059 
1060 	if (req->private_bio) {
1061 		if (!drbd_may_do_local_read(device,
1062 					req->i.sector, req->i.size)) {
1063 			bio_put(req->private_bio);
1064 			req->private_bio = NULL;
1065 			put_ldev(device);
1066 		}
1067 	}
1068 
1069 	if (device->state.pdsk != D_UP_TO_DATE)
1070 		return false;
1071 
1072 	if (req->private_bio == NULL)
1073 		return true;
1074 
1075 	/* TODO: improve read balancing decisions, take into account drbd
1076 	 * protocol, pending requests etc. */
1077 
1078 	rcu_read_lock();
1079 	rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1080 	rcu_read_unlock();
1081 
1082 	if (rbm == RB_PREFER_LOCAL && req->private_bio)
1083 		return false; /* submit locally */
1084 
1085 	if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1086 		if (req->private_bio) {
1087 			bio_put(req->private_bio);
1088 			req->private_bio = NULL;
1089 			put_ldev(device);
1090 		}
1091 		return true;
1092 	}
1093 
1094 	return false;
1095 }
1096 
1097 /* returns number of connections (== 1, for drbd 8.4)
1098  * expected to actually write this data,
1099  * which does NOT include those that we are L_AHEAD for. */
1100 static int drbd_process_write_request(struct drbd_request *req)
1101 {
1102 	struct drbd_device *device = req->device;
1103 	int remote, send_oos;
1104 
1105 	remote = drbd_should_do_remote(device->state);
1106 	send_oos = drbd_should_send_out_of_sync(device->state);
1107 
1108 	/* Need to replicate writes.  Unless it is an empty flush,
1109 	 * which is better mapped to a DRBD P_BARRIER packet,
1110 	 * also for drbd wire protocol compatibility reasons.
1111 	 * If this was a flush, just start a new epoch.
1112 	 * Unless the current epoch was empty anyways, or we are not currently
1113 	 * replicating, in which case there is no point. */
1114 	if (unlikely(req->i.size == 0)) {
1115 		/* The only size==0 bios we expect are empty flushes. */
1116 		D_ASSERT(device, req->master_bio->bi_rw & REQ_FLUSH);
1117 		if (remote)
1118 			_req_mod(req, QUEUE_AS_DRBD_BARRIER);
1119 		return remote;
1120 	}
1121 
1122 	if (!remote && !send_oos)
1123 		return 0;
1124 
1125 	D_ASSERT(device, !(remote && send_oos));
1126 
1127 	if (remote) {
1128 		_req_mod(req, TO_BE_SENT);
1129 		_req_mod(req, QUEUE_FOR_NET_WRITE);
1130 	} else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
1131 		_req_mod(req, QUEUE_FOR_SEND_OOS);
1132 
1133 	return remote;
1134 }
1135 
1136 static void
1137 drbd_submit_req_private_bio(struct drbd_request *req)
1138 {
1139 	struct drbd_device *device = req->device;
1140 	struct bio *bio = req->private_bio;
1141 	const int rw = bio_rw(bio);
1142 
1143 	bio->bi_bdev = device->ldev->backing_bdev;
1144 
1145 	/* State may have changed since we grabbed our reference on the
1146 	 * ->ldev member. Double check, and short-circuit to endio.
1147 	 * In case the last activity log transaction failed to get on
1148 	 * stable storage, and this is a WRITE, we may not even submit
1149 	 * this bio. */
1150 	if (get_ldev(device)) {
1151 		req->pre_submit_jif = jiffies;
1152 		if (drbd_insert_fault(device,
1153 				      rw == WRITE ? DRBD_FAULT_DT_WR
1154 				    : rw == READ  ? DRBD_FAULT_DT_RD
1155 				    :               DRBD_FAULT_DT_RA))
1156 			bio_endio(bio, -EIO);
1157 		else
1158 			generic_make_request(bio);
1159 		put_ldev(device);
1160 	} else
1161 		bio_endio(bio, -EIO);
1162 }
1163 
1164 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1165 {
1166 	spin_lock_irq(&device->resource->req_lock);
1167 	list_add_tail(&req->tl_requests, &device->submit.writes);
1168 	list_add_tail(&req->req_pending_master_completion,
1169 			&device->pending_master_completion[1 /* WRITE */]);
1170 	spin_unlock_irq(&device->resource->req_lock);
1171 	queue_work(device->submit.wq, &device->submit.worker);
1172 	/* do_submit() may sleep internally on al_wait, too */
1173 	wake_up(&device->al_wait);
1174 }
1175 
1176 /* returns the new drbd_request pointer, if the caller is expected to
1177  * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1178  * request on the submitter thread.
1179  * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1180  */
1181 static struct drbd_request *
1182 drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1183 {
1184 	const int rw = bio_data_dir(bio);
1185 	struct drbd_request *req;
1186 
1187 	/* allocate outside of all locks; */
1188 	req = drbd_req_new(device, bio);
1189 	if (!req) {
1190 		dec_ap_bio(device);
1191 		/* only pass the error to the upper layers.
1192 		 * if user cannot handle io errors, that's not our business. */
1193 		drbd_err(device, "could not kmalloc() req\n");
1194 		bio_endio(bio, -ENOMEM);
1195 		return ERR_PTR(-ENOMEM);
1196 	}
1197 	req->start_jif = start_jif;
1198 
1199 	if (!get_ldev(device)) {
1200 		bio_put(req->private_bio);
1201 		req->private_bio = NULL;
1202 	}
1203 
1204 	/* Update disk stats */
1205 	_drbd_start_io_acct(device, req);
1206 
1207 	if (rw == WRITE && req->private_bio && req->i.size
1208 	&& !test_bit(AL_SUSPENDED, &device->flags)) {
1209 		if (!drbd_al_begin_io_fastpath(device, &req->i)) {
1210 			atomic_inc(&device->ap_actlog_cnt);
1211 			drbd_queue_write(device, req);
1212 			return NULL;
1213 		}
1214 		req->rq_state |= RQ_IN_ACT_LOG;
1215 		req->in_actlog_jif = jiffies;
1216 	}
1217 
1218 	return req;
1219 }
1220 
1221 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1222 {
1223 	struct drbd_resource *resource = device->resource;
1224 	const int rw = bio_rw(req->master_bio);
1225 	struct bio_and_error m = { NULL, };
1226 	bool no_remote = false;
1227 	bool submit_private_bio = false;
1228 
1229 	spin_lock_irq(&resource->req_lock);
1230 	if (rw == WRITE) {
1231 		/* This may temporarily give up the req_lock,
1232 		 * but will re-aquire it before it returns here.
1233 		 * Needs to be before the check on drbd_suspended() */
1234 		complete_conflicting_writes(req);
1235 		/* no more giving up req_lock from now on! */
1236 
1237 		/* check for congestion, and potentially stop sending
1238 		 * full data updates, but start sending "dirty bits" only. */
1239 		maybe_pull_ahead(device);
1240 	}
1241 
1242 
1243 	if (drbd_suspended(device)) {
1244 		/* push back and retry: */
1245 		req->rq_state |= RQ_POSTPONED;
1246 		if (req->private_bio) {
1247 			bio_put(req->private_bio);
1248 			req->private_bio = NULL;
1249 			put_ldev(device);
1250 		}
1251 		goto out;
1252 	}
1253 
1254 	/* We fail READ/READA early, if we can not serve it.
1255 	 * We must do this before req is registered on any lists.
1256 	 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1257 	if (rw != WRITE) {
1258 		if (!do_remote_read(req) && !req->private_bio)
1259 			goto nodata;
1260 	}
1261 
1262 	/* which transfer log epoch does this belong to? */
1263 	req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1264 
1265 	/* no point in adding empty flushes to the transfer log,
1266 	 * they are mapped to drbd barriers already. */
1267 	if (likely(req->i.size!=0)) {
1268 		if (rw == WRITE)
1269 			first_peer_device(device)->connection->current_tle_writes++;
1270 
1271 		list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1272 	}
1273 
1274 	if (rw == WRITE) {
1275 		if (!drbd_process_write_request(req))
1276 			no_remote = true;
1277 	} else {
1278 		/* We either have a private_bio, or we can read from remote.
1279 		 * Otherwise we had done the goto nodata above. */
1280 		if (req->private_bio == NULL) {
1281 			_req_mod(req, TO_BE_SENT);
1282 			_req_mod(req, QUEUE_FOR_NET_READ);
1283 		} else
1284 			no_remote = true;
1285 	}
1286 
1287 	/* If it took the fast path in drbd_request_prepare, add it here.
1288 	 * The slow path has added it already. */
1289 	if (list_empty(&req->req_pending_master_completion))
1290 		list_add_tail(&req->req_pending_master_completion,
1291 			&device->pending_master_completion[rw == WRITE]);
1292 	if (req->private_bio) {
1293 		/* needs to be marked within the same spinlock */
1294 		list_add_tail(&req->req_pending_local,
1295 			&device->pending_completion[rw == WRITE]);
1296 		_req_mod(req, TO_BE_SUBMITTED);
1297 		/* but we need to give up the spinlock to submit */
1298 		submit_private_bio = true;
1299 	} else if (no_remote) {
1300 nodata:
1301 		if (__ratelimit(&drbd_ratelimit_state))
1302 			drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1303 					(unsigned long long)req->i.sector, req->i.size >> 9);
1304 		/* A write may have been queued for send_oos, however.
1305 		 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1306 	}
1307 
1308 out:
1309 	if (drbd_req_put_completion_ref(req, &m, 1))
1310 		kref_put(&req->kref, drbd_req_destroy);
1311 	spin_unlock_irq(&resource->req_lock);
1312 
1313 	/* Even though above is a kref_put(), this is safe.
1314 	 * As long as we still need to submit our private bio,
1315 	 * we hold a completion ref, and the request cannot disappear.
1316 	 * If however this request did not even have a private bio to submit
1317 	 * (e.g. remote read), req may already be invalid now.
1318 	 * That's why we cannot check on req->private_bio. */
1319 	if (submit_private_bio)
1320 		drbd_submit_req_private_bio(req);
1321 	if (m.bio)
1322 		complete_master_bio(device, &m);
1323 }
1324 
1325 void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_jif)
1326 {
1327 	struct drbd_request *req = drbd_request_prepare(device, bio, start_jif);
1328 	if (IS_ERR_OR_NULL(req))
1329 		return;
1330 	drbd_send_and_submit(device, req);
1331 }
1332 
1333 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1334 {
1335 	struct drbd_request *req, *tmp;
1336 	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1337 		const int rw = bio_data_dir(req->master_bio);
1338 
1339 		if (rw == WRITE /* rw != WRITE should not even end up here! */
1340 		&& req->private_bio && req->i.size
1341 		&& !test_bit(AL_SUSPENDED, &device->flags)) {
1342 			if (!drbd_al_begin_io_fastpath(device, &req->i))
1343 				continue;
1344 
1345 			req->rq_state |= RQ_IN_ACT_LOG;
1346 			req->in_actlog_jif = jiffies;
1347 			atomic_dec(&device->ap_actlog_cnt);
1348 		}
1349 
1350 		list_del_init(&req->tl_requests);
1351 		drbd_send_and_submit(device, req);
1352 	}
1353 }
1354 
1355 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1356 					    struct list_head *incoming,
1357 					    struct list_head *pending,
1358 					    struct list_head *later)
1359 {
1360 	struct drbd_request *req, *tmp;
1361 	int wake = 0;
1362 	int err;
1363 
1364 	spin_lock_irq(&device->al_lock);
1365 	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1366 		err = drbd_al_begin_io_nonblock(device, &req->i);
1367 		if (err == -ENOBUFS)
1368 			break;
1369 		if (err == -EBUSY)
1370 			wake = 1;
1371 		if (err)
1372 			list_move_tail(&req->tl_requests, later);
1373 		else
1374 			list_move_tail(&req->tl_requests, pending);
1375 	}
1376 	spin_unlock_irq(&device->al_lock);
1377 	if (wake)
1378 		wake_up(&device->al_wait);
1379 	return !list_empty(pending);
1380 }
1381 
1382 void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1383 {
1384 	struct drbd_request *req, *tmp;
1385 
1386 	list_for_each_entry_safe(req, tmp, pending, tl_requests) {
1387 		req->rq_state |= RQ_IN_ACT_LOG;
1388 		req->in_actlog_jif = jiffies;
1389 		atomic_dec(&device->ap_actlog_cnt);
1390 		list_del_init(&req->tl_requests);
1391 		drbd_send_and_submit(device, req);
1392 	}
1393 }
1394 
1395 void do_submit(struct work_struct *ws)
1396 {
1397 	struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1398 	LIST_HEAD(incoming);	/* from drbd_make_request() */
1399 	LIST_HEAD(pending);	/* to be submitted after next AL-transaction commit */
1400 	LIST_HEAD(busy);	/* blocked by resync requests */
1401 
1402 	/* grab new incoming requests */
1403 	spin_lock_irq(&device->resource->req_lock);
1404 	list_splice_tail_init(&device->submit.writes, &incoming);
1405 	spin_unlock_irq(&device->resource->req_lock);
1406 
1407 	for (;;) {
1408 		DEFINE_WAIT(wait);
1409 
1410 		/* move used-to-be-busy back to front of incoming */
1411 		list_splice_init(&busy, &incoming);
1412 		submit_fast_path(device, &incoming);
1413 		if (list_empty(&incoming))
1414 			break;
1415 
1416 		for (;;) {
1417 			prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1418 
1419 			list_splice_init(&busy, &incoming);
1420 			prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1421 			if (!list_empty(&pending))
1422 				break;
1423 
1424 			schedule();
1425 
1426 			/* If all currently "hot" activity log extents are kept busy by
1427 			 * incoming requests, we still must not totally starve new
1428 			 * requests to "cold" extents.
1429 			 * Something left on &incoming means there had not been
1430 			 * enough update slots available, and the activity log
1431 			 * has been marked as "starving".
1432 			 *
1433 			 * Try again now, without looking for new requests,
1434 			 * effectively blocking all new requests until we made
1435 			 * at least _some_ progress with what we currently have.
1436 			 */
1437 			if (!list_empty(&incoming))
1438 				continue;
1439 
1440 			/* Nothing moved to pending, but nothing left
1441 			 * on incoming: all moved to busy!
1442 			 * Grab new and iterate. */
1443 			spin_lock_irq(&device->resource->req_lock);
1444 			list_splice_tail_init(&device->submit.writes, &incoming);
1445 			spin_unlock_irq(&device->resource->req_lock);
1446 		}
1447 		finish_wait(&device->al_wait, &wait);
1448 
1449 		/* If the transaction was full, before all incoming requests
1450 		 * had been processed, skip ahead to commit, and iterate
1451 		 * without splicing in more incoming requests from upper layers.
1452 		 *
1453 		 * Else, if all incoming have been processed,
1454 		 * they have become either "pending" (to be submitted after
1455 		 * next transaction commit) or "busy" (blocked by resync).
1456 		 *
1457 		 * Maybe more was queued, while we prepared the transaction?
1458 		 * Try to stuff those into this transaction as well.
1459 		 * Be strictly non-blocking here,
1460 		 * we already have something to commit.
1461 		 *
1462 		 * Commit if we don't make any more progres.
1463 		 */
1464 
1465 		while (list_empty(&incoming)) {
1466 			LIST_HEAD(more_pending);
1467 			LIST_HEAD(more_incoming);
1468 			bool made_progress;
1469 
1470 			/* It is ok to look outside the lock,
1471 			 * it's only an optimization anyways */
1472 			if (list_empty(&device->submit.writes))
1473 				break;
1474 
1475 			spin_lock_irq(&device->resource->req_lock);
1476 			list_splice_tail_init(&device->submit.writes, &more_incoming);
1477 			spin_unlock_irq(&device->resource->req_lock);
1478 
1479 			if (list_empty(&more_incoming))
1480 				break;
1481 
1482 			made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1483 
1484 			list_splice_tail_init(&more_pending, &pending);
1485 			list_splice_tail_init(&more_incoming, &incoming);
1486 			if (!made_progress)
1487 				break;
1488 		}
1489 
1490 		drbd_al_begin_io_commit(device);
1491 		send_and_submit_pending(device, &pending);
1492 	}
1493 }
1494 
1495 void drbd_make_request(struct request_queue *q, struct bio *bio)
1496 {
1497 	struct drbd_device *device = (struct drbd_device *) q->queuedata;
1498 	unsigned long start_jif;
1499 
1500 	start_jif = jiffies;
1501 
1502 	/*
1503 	 * what we "blindly" assume:
1504 	 */
1505 	D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1506 
1507 	inc_ap_bio(device);
1508 	__drbd_make_request(device, bio, start_jif);
1509 }
1510 
1511 /* This is called by bio_add_page().
1512  *
1513  * q->max_hw_sectors and other global limits are already enforced there.
1514  *
1515  * We need to call down to our lower level device,
1516  * in case it has special restrictions.
1517  *
1518  * We also may need to enforce configured max-bio-bvecs limits.
1519  *
1520  * As long as the BIO is empty we have to allow at least one bvec,
1521  * regardless of size and offset, so no need to ask lower levels.
1522  */
1523 int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
1524 {
1525 	struct drbd_device *device = (struct drbd_device *) q->queuedata;
1526 	unsigned int bio_size = bvm->bi_size;
1527 	int limit = DRBD_MAX_BIO_SIZE;
1528 	int backing_limit;
1529 
1530 	if (bio_size && get_ldev(device)) {
1531 		unsigned int max_hw_sectors = queue_max_hw_sectors(q);
1532 		struct request_queue * const b =
1533 			device->ldev->backing_bdev->bd_disk->queue;
1534 		if (b->merge_bvec_fn) {
1535 			bvm->bi_bdev = device->ldev->backing_bdev;
1536 			backing_limit = b->merge_bvec_fn(b, bvm, bvec);
1537 			limit = min(limit, backing_limit);
1538 		}
1539 		put_ldev(device);
1540 		if ((limit >> 9) > max_hw_sectors)
1541 			limit = max_hw_sectors << 9;
1542 	}
1543 	return limit;
1544 }
1545 
1546 void request_timer_fn(unsigned long data)
1547 {
1548 	struct drbd_device *device = (struct drbd_device *) data;
1549 	struct drbd_connection *connection = first_peer_device(device)->connection;
1550 	struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1551 	struct net_conf *nc;
1552 	unsigned long oldest_submit_jif;
1553 	unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1554 	unsigned long now;
1555 
1556 	rcu_read_lock();
1557 	nc = rcu_dereference(connection->net_conf);
1558 	if (nc && device->state.conn >= C_WF_REPORT_PARAMS)
1559 		ent = nc->timeout * HZ/10 * nc->ko_count;
1560 
1561 	if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1562 		dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1563 		put_ldev(device);
1564 	}
1565 	rcu_read_unlock();
1566 
1567 	et = min_not_zero(dt, ent);
1568 
1569 	if (!et)
1570 		return; /* Recurring timer stopped */
1571 
1572 	now = jiffies;
1573 	nt = now + et;
1574 
1575 	spin_lock_irq(&device->resource->req_lock);
1576 	req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1577 	req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1578 	req_peer = connection->req_not_net_done;
1579 	/* maybe the oldest request waiting for the peer is in fact still
1580 	 * blocking in tcp sendmsg */
1581 	if (!req_peer && connection->req_next && connection->req_next->pre_send_jif)
1582 		req_peer = connection->req_next;
1583 
1584 	/* evaluate the oldest peer request only in one timer! */
1585 	if (req_peer && req_peer->device != device)
1586 		req_peer = NULL;
1587 
1588 	/* do we have something to evaluate? */
1589 	if (req_peer == NULL && req_write == NULL && req_read == NULL)
1590 		goto out;
1591 
1592 	oldest_submit_jif =
1593 		(req_write && req_read)
1594 		? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1595 		  ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1596 		: req_write ? req_write->pre_submit_jif
1597 		: req_read ? req_read->pre_submit_jif : now;
1598 
1599 	/* The request is considered timed out, if
1600 	 * - we have some effective timeout from the configuration,
1601 	 *   with above state restrictions applied,
1602 	 * - the oldest request is waiting for a response from the network
1603 	 *   resp. the local disk,
1604 	 * - the oldest request is in fact older than the effective timeout,
1605 	 * - the connection was established (resp. disk was attached)
1606 	 *   for longer than the timeout already.
1607 	 * Note that for 32bit jiffies and very stable connections/disks,
1608 	 * we may have a wrap around, which is catched by
1609 	 *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1610 	 *
1611 	 * Side effect: once per 32bit wrap-around interval, which means every
1612 	 * ~198 days with 250 HZ, we have a window where the timeout would need
1613 	 * to expire twice (worst case) to become effective. Good enough.
1614 	 */
1615 	if (ent && req_peer &&
1616 		 time_after(now, req_peer->pre_send_jif + ent) &&
1617 		!time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) {
1618 		drbd_warn(device, "Remote failed to finish a request within ko-count * timeout\n");
1619 		_conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1620 	}
1621 	if (dt && oldest_submit_jif != now &&
1622 		 time_after(now, oldest_submit_jif + dt) &&
1623 		!time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1624 		drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1625 		__drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1626 	}
1627 
1628 	/* Reschedule timer for the nearest not already expired timeout.
1629 	 * Fallback to now + min(effective network timeout, disk timeout). */
1630 	ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1631 		? req_peer->pre_send_jif + ent : now + et;
1632 	dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1633 		? oldest_submit_jif + dt : now + et;
1634 	nt = time_before(ent, dt) ? ent : dt;
1635 out:
1636 	spin_unlock_irq(&device->resource->req_lock);
1637 	mod_timer(&device->request_timer, nt);
1638 }
1639