1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3    drbd_req.c
4 
5    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 
7    Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
8    Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
9    Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 
11 
12  */
13 
14 #include <linux/module.h>
15 
16 #include <linux/slab.h>
17 #include <linux/drbd.h>
18 #include "drbd_int.h"
19 #include "drbd_req.h"
20 
21 
22 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
23 
drbd_req_new(struct drbd_device * device,struct bio * bio_src)24 static struct drbd_request *drbd_req_new(struct drbd_device *device, struct bio *bio_src)
25 {
26 	struct drbd_request *req;
27 
28 	req = mempool_alloc(&drbd_request_mempool, GFP_NOIO);
29 	if (!req)
30 		return NULL;
31 	memset(req, 0, sizeof(*req));
32 
33 	req->rq_state = (bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0)
34 		      | (bio_op(bio_src) == REQ_OP_WRITE_ZEROES ? RQ_ZEROES : 0)
35 		      | (bio_op(bio_src) == REQ_OP_DISCARD ? RQ_UNMAP : 0);
36 	req->device = device;
37 	req->master_bio = bio_src;
38 	req->epoch = 0;
39 
40 	drbd_clear_interval(&req->i);
41 	req->i.sector     = bio_src->bi_iter.bi_sector;
42 	req->i.size      = bio_src->bi_iter.bi_size;
43 	req->i.local = true;
44 	req->i.waiting = false;
45 
46 	INIT_LIST_HEAD(&req->tl_requests);
47 	INIT_LIST_HEAD(&req->w.list);
48 	INIT_LIST_HEAD(&req->req_pending_master_completion);
49 	INIT_LIST_HEAD(&req->req_pending_local);
50 
51 	/* one reference to be put by __drbd_make_request */
52 	atomic_set(&req->completion_ref, 1);
53 	/* one kref as long as completion_ref > 0 */
54 	kref_init(&req->kref);
55 	return req;
56 }
57 
drbd_remove_request_interval(struct rb_root * root,struct drbd_request * req)58 static void drbd_remove_request_interval(struct rb_root *root,
59 					 struct drbd_request *req)
60 {
61 	struct drbd_device *device = req->device;
62 	struct drbd_interval *i = &req->i;
63 
64 	drbd_remove_interval(root, i);
65 
66 	/* Wake up any processes waiting for this request to complete.  */
67 	if (i->waiting)
68 		wake_up(&device->misc_wait);
69 }
70 
drbd_req_destroy(struct kref * kref)71 void drbd_req_destroy(struct kref *kref)
72 {
73 	struct drbd_request *req = container_of(kref, struct drbd_request, kref);
74 	struct drbd_device *device = req->device;
75 	const unsigned s = req->rq_state;
76 
77 	if ((req->master_bio && !(s & RQ_POSTPONED)) ||
78 		atomic_read(&req->completion_ref) ||
79 		(s & RQ_LOCAL_PENDING) ||
80 		((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
81 		drbd_err(device, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
82 				s, atomic_read(&req->completion_ref));
83 		return;
84 	}
85 
86 	/* If called from mod_rq_state (expected normal case) or
87 	 * drbd_send_and_submit (the less likely normal path), this holds the
88 	 * req_lock, and req->tl_requests will typicaly be on ->transfer_log,
89 	 * though it may be still empty (never added to the transfer log).
90 	 *
91 	 * If called from do_retry(), we do NOT hold the req_lock, but we are
92 	 * still allowed to unconditionally list_del(&req->tl_requests),
93 	 * because it will be on a local on-stack list only. */
94 	list_del_init(&req->tl_requests);
95 
96 	/* finally remove the request from the conflict detection
97 	 * respective block_id verification interval tree. */
98 	if (!drbd_interval_empty(&req->i)) {
99 		struct rb_root *root;
100 
101 		if (s & RQ_WRITE)
102 			root = &device->write_requests;
103 		else
104 			root = &device->read_requests;
105 		drbd_remove_request_interval(root, req);
106 	} else if (s & (RQ_NET_MASK & ~RQ_NET_DONE) && req->i.size != 0)
107 		drbd_err(device, "drbd_req_destroy: Logic BUG: interval empty, but: rq_state=0x%x, sect=%llu, size=%u\n",
108 			s, (unsigned long long)req->i.sector, req->i.size);
109 
110 	/* if it was a write, we may have to set the corresponding
111 	 * bit(s) out-of-sync first. If it had a local part, we need to
112 	 * release the reference to the activity log. */
113 	if (s & RQ_WRITE) {
114 		/* Set out-of-sync unless both OK flags are set
115 		 * (local only or remote failed).
116 		 * Other places where we set out-of-sync:
117 		 * READ with local io-error */
118 
119 		/* There is a special case:
120 		 * we may notice late that IO was suspended,
121 		 * and postpone, or schedule for retry, a write,
122 		 * before it even was submitted or sent.
123 		 * In that case we do not want to touch the bitmap at all.
124 		 */
125 		struct drbd_peer_device *peer_device = first_peer_device(device);
126 		if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
127 			if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
128 				drbd_set_out_of_sync(peer_device, req->i.sector, req->i.size);
129 
130 			if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
131 				drbd_set_in_sync(peer_device, req->i.sector, req->i.size);
132 		}
133 
134 		/* one might be tempted to move the drbd_al_complete_io
135 		 * to the local io completion callback drbd_request_endio.
136 		 * but, if this was a mirror write, we may only
137 		 * drbd_al_complete_io after this is RQ_NET_DONE,
138 		 * otherwise the extent could be dropped from the al
139 		 * before it has actually been written on the peer.
140 		 * if we crash before our peer knows about the request,
141 		 * but after the extent has been dropped from the al,
142 		 * we would forget to resync the corresponding extent.
143 		 */
144 		if (s & RQ_IN_ACT_LOG) {
145 			if (get_ldev_if_state(device, D_FAILED)) {
146 				drbd_al_complete_io(device, &req->i);
147 				put_ldev(device);
148 			} else if (drbd_ratelimit()) {
149 				drbd_warn(device, "Should have called drbd_al_complete_io(, %llu, %u), "
150 					 "but my Disk seems to have failed :(\n",
151 					 (unsigned long long) req->i.sector, req->i.size);
152 			}
153 		}
154 	}
155 
156 	mempool_free(req, &drbd_request_mempool);
157 }
158 
wake_all_senders(struct drbd_connection * connection)159 static void wake_all_senders(struct drbd_connection *connection)
160 {
161 	wake_up(&connection->sender_work.q_wait);
162 }
163 
164 /* must hold resource->req_lock */
start_new_tl_epoch(struct drbd_connection * connection)165 void start_new_tl_epoch(struct drbd_connection *connection)
166 {
167 	/* no point closing an epoch, if it is empty, anyways. */
168 	if (connection->current_tle_writes == 0)
169 		return;
170 
171 	connection->current_tle_writes = 0;
172 	atomic_inc(&connection->current_tle_nr);
173 	wake_all_senders(connection);
174 }
175 
complete_master_bio(struct drbd_device * device,struct bio_and_error * m)176 void complete_master_bio(struct drbd_device *device,
177 		struct bio_and_error *m)
178 {
179 	if (unlikely(m->error))
180 		m->bio->bi_status = errno_to_blk_status(m->error);
181 	bio_endio(m->bio);
182 	dec_ap_bio(device);
183 }
184 
185 
186 /* Helper for __req_mod().
187  * Set m->bio to the master bio, if it is fit to be completed,
188  * or leave it alone (it is initialized to NULL in __req_mod),
189  * if it has already been completed, or cannot be completed yet.
190  * If m->bio is set, the error status to be returned is placed in m->error.
191  */
192 static
drbd_req_complete(struct drbd_request * req,struct bio_and_error * m)193 void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
194 {
195 	const unsigned s = req->rq_state;
196 	struct drbd_device *device = req->device;
197 	int error, ok;
198 
199 	/* we must not complete the master bio, while it is
200 	 *	still being processed by _drbd_send_zc_bio (drbd_send_dblock)
201 	 *	not yet acknowledged by the peer
202 	 *	not yet completed by the local io subsystem
203 	 * these flags may get cleared in any order by
204 	 *	the worker,
205 	 *	the receiver,
206 	 *	the bio_endio completion callbacks.
207 	 */
208 	if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
209 	    (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
210 	    (s & RQ_COMPLETION_SUSP)) {
211 		drbd_err(device, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
212 		return;
213 	}
214 
215 	if (!req->master_bio) {
216 		drbd_err(device, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
217 		return;
218 	}
219 
220 	/*
221 	 * figure out whether to report success or failure.
222 	 *
223 	 * report success when at least one of the operations succeeded.
224 	 * or, to put the other way,
225 	 * only report failure, when both operations failed.
226 	 *
227 	 * what to do about the failures is handled elsewhere.
228 	 * what we need to do here is just: complete the master_bio.
229 	 *
230 	 * local completion error, if any, has been stored as ERR_PTR
231 	 * in private_bio within drbd_request_endio.
232 	 */
233 	ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
234 	error = PTR_ERR(req->private_bio);
235 
236 	/* Before we can signal completion to the upper layers,
237 	 * we may need to close the current transfer log epoch.
238 	 * We are within the request lock, so we can simply compare
239 	 * the request epoch number with the current transfer log
240 	 * epoch number.  If they match, increase the current_tle_nr,
241 	 * and reset the transfer log epoch write_cnt.
242 	 */
243 	if (op_is_write(bio_op(req->master_bio)) &&
244 	    req->epoch == atomic_read(&first_peer_device(device)->connection->current_tle_nr))
245 		start_new_tl_epoch(first_peer_device(device)->connection);
246 
247 	/* Update disk stats */
248 	bio_end_io_acct(req->master_bio, req->start_jif);
249 
250 	/* If READ failed,
251 	 * have it be pushed back to the retry work queue,
252 	 * so it will re-enter __drbd_make_request(),
253 	 * and be re-assigned to a suitable local or remote path,
254 	 * or failed if we do not have access to good data anymore.
255 	 *
256 	 * Unless it was failed early by __drbd_make_request(),
257 	 * because no path was available, in which case
258 	 * it was not even added to the transfer_log.
259 	 *
260 	 * read-ahead may fail, and will not be retried.
261 	 *
262 	 * WRITE should have used all available paths already.
263 	 */
264 	if (!ok &&
265 	    bio_op(req->master_bio) == REQ_OP_READ &&
266 	    !(req->master_bio->bi_opf & REQ_RAHEAD) &&
267 	    !list_empty(&req->tl_requests))
268 		req->rq_state |= RQ_POSTPONED;
269 
270 	if (!(req->rq_state & RQ_POSTPONED)) {
271 		m->error = ok ? 0 : (error ?: -EIO);
272 		m->bio = req->master_bio;
273 		req->master_bio = NULL;
274 		/* We leave it in the tree, to be able to verify later
275 		 * write-acks in protocol != C during resync.
276 		 * But we mark it as "complete", so it won't be counted as
277 		 * conflict in a multi-primary setup. */
278 		req->i.completed = true;
279 	}
280 
281 	if (req->i.waiting)
282 		wake_up(&device->misc_wait);
283 
284 	/* Either we are about to complete to upper layers,
285 	 * or we will restart this request.
286 	 * In either case, the request object will be destroyed soon,
287 	 * so better remove it from all lists. */
288 	list_del_init(&req->req_pending_master_completion);
289 }
290 
291 /* still holds resource->req_lock */
drbd_req_put_completion_ref(struct drbd_request * req,struct bio_and_error * m,int put)292 static void drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
293 {
294 	struct drbd_device *device = req->device;
295 	D_ASSERT(device, m || (req->rq_state & RQ_POSTPONED));
296 
297 	if (!put)
298 		return;
299 
300 	if (!atomic_sub_and_test(put, &req->completion_ref))
301 		return;
302 
303 	drbd_req_complete(req, m);
304 
305 	/* local completion may still come in later,
306 	 * we need to keep the req object around. */
307 	if (req->rq_state & RQ_LOCAL_ABORTED)
308 		return;
309 
310 	if (req->rq_state & RQ_POSTPONED) {
311 		/* don't destroy the req object just yet,
312 		 * but queue it for retry */
313 		drbd_restart_request(req);
314 		return;
315 	}
316 
317 	kref_put(&req->kref, drbd_req_destroy);
318 }
319 
set_if_null_req_next(struct drbd_peer_device * peer_device,struct drbd_request * req)320 static void set_if_null_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
321 {
322 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
323 	if (!connection)
324 		return;
325 	if (connection->req_next == NULL)
326 		connection->req_next = req;
327 }
328 
advance_conn_req_next(struct drbd_peer_device * peer_device,struct drbd_request * req)329 static void advance_conn_req_next(struct drbd_peer_device *peer_device, struct drbd_request *req)
330 {
331 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
332 	struct drbd_request *iter = req;
333 	if (!connection)
334 		return;
335 	if (connection->req_next != req)
336 		return;
337 
338 	req = NULL;
339 	list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) {
340 		const unsigned int s = iter->rq_state;
341 
342 		if (s & RQ_NET_QUEUED) {
343 			req = iter;
344 			break;
345 		}
346 	}
347 	connection->req_next = req;
348 }
349 
set_if_null_req_ack_pending(struct drbd_peer_device * peer_device,struct drbd_request * req)350 static void set_if_null_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
351 {
352 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
353 	if (!connection)
354 		return;
355 	if (connection->req_ack_pending == NULL)
356 		connection->req_ack_pending = req;
357 }
358 
advance_conn_req_ack_pending(struct drbd_peer_device * peer_device,struct drbd_request * req)359 static void advance_conn_req_ack_pending(struct drbd_peer_device *peer_device, struct drbd_request *req)
360 {
361 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
362 	struct drbd_request *iter = req;
363 	if (!connection)
364 		return;
365 	if (connection->req_ack_pending != req)
366 		return;
367 
368 	req = NULL;
369 	list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) {
370 		const unsigned int s = iter->rq_state;
371 
372 		if ((s & RQ_NET_SENT) && (s & RQ_NET_PENDING)) {
373 			req = iter;
374 			break;
375 		}
376 	}
377 	connection->req_ack_pending = req;
378 }
379 
set_if_null_req_not_net_done(struct drbd_peer_device * peer_device,struct drbd_request * req)380 static void set_if_null_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
381 {
382 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
383 	if (!connection)
384 		return;
385 	if (connection->req_not_net_done == NULL)
386 		connection->req_not_net_done = req;
387 }
388 
advance_conn_req_not_net_done(struct drbd_peer_device * peer_device,struct drbd_request * req)389 static void advance_conn_req_not_net_done(struct drbd_peer_device *peer_device, struct drbd_request *req)
390 {
391 	struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
392 	struct drbd_request *iter = req;
393 	if (!connection)
394 		return;
395 	if (connection->req_not_net_done != req)
396 		return;
397 
398 	req = NULL;
399 	list_for_each_entry_continue(iter, &connection->transfer_log, tl_requests) {
400 		const unsigned int s = iter->rq_state;
401 
402 		if ((s & RQ_NET_SENT) && !(s & RQ_NET_DONE)) {
403 			req = iter;
404 			break;
405 		}
406 	}
407 	connection->req_not_net_done = req;
408 }
409 
410 /* I'd like this to be the only place that manipulates
411  * req->completion_ref and req->kref. */
mod_rq_state(struct drbd_request * req,struct bio_and_error * m,int clear,int set)412 static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
413 		int clear, int set)
414 {
415 	struct drbd_device *device = req->device;
416 	struct drbd_peer_device *peer_device = first_peer_device(device);
417 	unsigned s = req->rq_state;
418 	int c_put = 0;
419 
420 	if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
421 		set |= RQ_COMPLETION_SUSP;
422 
423 	/* apply */
424 
425 	req->rq_state &= ~clear;
426 	req->rq_state |= set;
427 
428 	/* no change? */
429 	if (req->rq_state == s)
430 		return;
431 
432 	/* intent: get references */
433 
434 	kref_get(&req->kref);
435 
436 	if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
437 		atomic_inc(&req->completion_ref);
438 
439 	if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
440 		inc_ap_pending(device);
441 		atomic_inc(&req->completion_ref);
442 	}
443 
444 	if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED)) {
445 		atomic_inc(&req->completion_ref);
446 		set_if_null_req_next(peer_device, req);
447 	}
448 
449 	if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
450 		kref_get(&req->kref); /* wait for the DONE */
451 
452 	if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT)) {
453 		/* potentially already completed in the ack_receiver thread */
454 		if (!(s & RQ_NET_DONE)) {
455 			atomic_add(req->i.size >> 9, &device->ap_in_flight);
456 			set_if_null_req_not_net_done(peer_device, req);
457 		}
458 		if (req->rq_state & RQ_NET_PENDING)
459 			set_if_null_req_ack_pending(peer_device, req);
460 	}
461 
462 	if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
463 		atomic_inc(&req->completion_ref);
464 
465 	/* progress: put references */
466 
467 	if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
468 		++c_put;
469 
470 	if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
471 		D_ASSERT(device, req->rq_state & RQ_LOCAL_PENDING);
472 		++c_put;
473 	}
474 
475 	if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
476 		if (req->rq_state & RQ_LOCAL_ABORTED)
477 			kref_put(&req->kref, drbd_req_destroy);
478 		else
479 			++c_put;
480 		list_del_init(&req->req_pending_local);
481 	}
482 
483 	if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
484 		dec_ap_pending(device);
485 		++c_put;
486 		req->acked_jif = jiffies;
487 		advance_conn_req_ack_pending(peer_device, req);
488 	}
489 
490 	if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED)) {
491 		++c_put;
492 		advance_conn_req_next(peer_device, req);
493 	}
494 
495 	if (!(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
496 		if (s & RQ_NET_SENT)
497 			atomic_sub(req->i.size >> 9, &device->ap_in_flight);
498 		if (s & RQ_EXP_BARR_ACK)
499 			kref_put(&req->kref, drbd_req_destroy);
500 		req->net_done_jif = jiffies;
501 
502 		/* in ahead/behind mode, or just in case,
503 		 * before we finally destroy this request,
504 		 * the caching pointers must not reference it anymore */
505 		advance_conn_req_next(peer_device, req);
506 		advance_conn_req_ack_pending(peer_device, req);
507 		advance_conn_req_not_net_done(peer_device, req);
508 	}
509 
510 	/* potentially complete and destroy */
511 
512 	/* If we made progress, retry conflicting peer requests, if any. */
513 	if (req->i.waiting)
514 		wake_up(&device->misc_wait);
515 
516 	drbd_req_put_completion_ref(req, m, c_put);
517 	kref_put(&req->kref, drbd_req_destroy);
518 }
519 
drbd_report_io_error(struct drbd_device * device,struct drbd_request * req)520 static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
521 {
522 	if (!drbd_ratelimit())
523 		return;
524 
525 	drbd_warn(device, "local %s IO error sector %llu+%u on %pg\n",
526 			(req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
527 			(unsigned long long)req->i.sector,
528 			req->i.size >> 9,
529 			device->ldev->backing_bdev);
530 }
531 
532 /* Helper for HANDED_OVER_TO_NETWORK.
533  * Is this a protocol A write (neither WRITE_ACK nor RECEIVE_ACK expected)?
534  * Is it also still "PENDING"?
535  * --> If so, clear PENDING and set NET_OK below.
536  * If it is a protocol A write, but not RQ_PENDING anymore, neg-ack was faster
537  * (and we must not set RQ_NET_OK) */
is_pending_write_protocol_A(struct drbd_request * req)538 static inline bool is_pending_write_protocol_A(struct drbd_request *req)
539 {
540 	return (req->rq_state &
541 		   (RQ_WRITE|RQ_NET_PENDING|RQ_EXP_WRITE_ACK|RQ_EXP_RECEIVE_ACK))
542 		== (RQ_WRITE|RQ_NET_PENDING);
543 }
544 
545 /* obviously this could be coded as many single functions
546  * instead of one huge switch,
547  * or by putting the code directly in the respective locations
548  * (as it has been before).
549  *
550  * but having it this way
551  *  enforces that it is all in this one place, where it is easier to audit,
552  *  it makes it obvious that whatever "event" "happens" to a request should
553  *  happen "atomically" within the req_lock,
554  *  and it enforces that we have to think in a very structured manner
555  *  about the "events" that may happen to a request during its life time ...
556  *
557  *
558  * peer_device == NULL means local disk
559  */
__req_mod(struct drbd_request * req,enum drbd_req_event what,struct drbd_peer_device * peer_device,struct bio_and_error * m)560 int __req_mod(struct drbd_request *req, enum drbd_req_event what,
561 		struct drbd_peer_device *peer_device,
562 		struct bio_and_error *m)
563 {
564 	struct drbd_device *const device = req->device;
565 	struct drbd_connection *const connection = peer_device ? peer_device->connection : NULL;
566 	struct net_conf *nc;
567 	int p, rv = 0;
568 
569 	if (m)
570 		m->bio = NULL;
571 
572 	switch (what) {
573 	default:
574 		drbd_err(device, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
575 		break;
576 
577 	/* does not happen...
578 	 * initialization done in drbd_req_new
579 	case CREATED:
580 		break;
581 		*/
582 
583 	case TO_BE_SENT: /* via network */
584 		/* reached via __drbd_make_request
585 		 * and from w_read_retry_remote */
586 		D_ASSERT(device, !(req->rq_state & RQ_NET_MASK));
587 		rcu_read_lock();
588 		nc = rcu_dereference(connection->net_conf);
589 		p = nc->wire_protocol;
590 		rcu_read_unlock();
591 		req->rq_state |=
592 			p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
593 			p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
594 		mod_rq_state(req, m, 0, RQ_NET_PENDING);
595 		break;
596 
597 	case TO_BE_SUBMITTED: /* locally */
598 		/* reached via __drbd_make_request */
599 		D_ASSERT(device, !(req->rq_state & RQ_LOCAL_MASK));
600 		mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
601 		break;
602 
603 	case COMPLETED_OK:
604 		if (req->rq_state & RQ_WRITE)
605 			device->writ_cnt += req->i.size >> 9;
606 		else
607 			device->read_cnt += req->i.size >> 9;
608 
609 		mod_rq_state(req, m, RQ_LOCAL_PENDING,
610 				RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
611 		break;
612 
613 	case ABORT_DISK_IO:
614 		mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
615 		break;
616 
617 	case WRITE_COMPLETED_WITH_ERROR:
618 		drbd_report_io_error(device, req);
619 		__drbd_chk_io_error(device, DRBD_WRITE_ERROR);
620 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
621 		break;
622 
623 	case READ_COMPLETED_WITH_ERROR:
624 		drbd_set_out_of_sync(peer_device, req->i.sector, req->i.size);
625 		drbd_report_io_error(device, req);
626 		__drbd_chk_io_error(device, DRBD_READ_ERROR);
627 		fallthrough;
628 	case READ_AHEAD_COMPLETED_WITH_ERROR:
629 		/* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
630 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
631 		break;
632 
633 	case DISCARD_COMPLETED_NOTSUPP:
634 	case DISCARD_COMPLETED_WITH_ERROR:
635 		/* I'd rather not detach from local disk just because it
636 		 * failed a REQ_OP_DISCARD. */
637 		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
638 		break;
639 
640 	case QUEUE_FOR_NET_READ:
641 		/* READ, and
642 		 * no local disk,
643 		 * or target area marked as invalid,
644 		 * or just got an io-error. */
645 		/* from __drbd_make_request
646 		 * or from bio_endio during read io-error recovery */
647 
648 		/* So we can verify the handle in the answer packet.
649 		 * Corresponding drbd_remove_request_interval is in
650 		 * drbd_req_complete() */
651 		D_ASSERT(device, drbd_interval_empty(&req->i));
652 		drbd_insert_interval(&device->read_requests, &req->i);
653 
654 		set_bit(UNPLUG_REMOTE, &device->flags);
655 
656 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
657 		D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
658 		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
659 		req->w.cb = w_send_read_req;
660 		drbd_queue_work(&connection->sender_work,
661 				&req->w);
662 		break;
663 
664 	case QUEUE_FOR_NET_WRITE:
665 		/* assert something? */
666 		/* from __drbd_make_request only */
667 
668 		/* Corresponding drbd_remove_request_interval is in
669 		 * drbd_req_complete() */
670 		D_ASSERT(device, drbd_interval_empty(&req->i));
671 		drbd_insert_interval(&device->write_requests, &req->i);
672 
673 		/* NOTE
674 		 * In case the req ended up on the transfer log before being
675 		 * queued on the worker, it could lead to this request being
676 		 * missed during cleanup after connection loss.
677 		 * So we have to do both operations here,
678 		 * within the same lock that protects the transfer log.
679 		 *
680 		 * _req_add_to_epoch(req); this has to be after the
681 		 * _maybe_start_new_epoch(req); which happened in
682 		 * __drbd_make_request, because we now may set the bit
683 		 * again ourselves to close the current epoch.
684 		 *
685 		 * Add req to the (now) current epoch (barrier). */
686 
687 		/* otherwise we may lose an unplug, which may cause some remote
688 		 * io-scheduler timeout to expire, increasing maximum latency,
689 		 * hurting performance. */
690 		set_bit(UNPLUG_REMOTE, &device->flags);
691 
692 		/* queue work item to send data */
693 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
694 		mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
695 		req->w.cb =  w_send_dblock;
696 		drbd_queue_work(&connection->sender_work,
697 				&req->w);
698 
699 		/* close the epoch, in case it outgrew the limit */
700 		rcu_read_lock();
701 		nc = rcu_dereference(connection->net_conf);
702 		p = nc->max_epoch_size;
703 		rcu_read_unlock();
704 		if (connection->current_tle_writes >= p)
705 			start_new_tl_epoch(connection);
706 
707 		break;
708 
709 	case QUEUE_FOR_SEND_OOS:
710 		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
711 		req->w.cb =  w_send_out_of_sync;
712 		drbd_queue_work(&connection->sender_work,
713 				&req->w);
714 		break;
715 
716 	case READ_RETRY_REMOTE_CANCELED:
717 	case SEND_CANCELED:
718 	case SEND_FAILED:
719 		/* real cleanup will be done from tl_clear.  just update flags
720 		 * so it is no longer marked as on the worker queue */
721 		mod_rq_state(req, m, RQ_NET_QUEUED, 0);
722 		break;
723 
724 	case HANDED_OVER_TO_NETWORK:
725 		/* assert something? */
726 		if (is_pending_write_protocol_A(req))
727 			/* this is what is dangerous about protocol A:
728 			 * pretend it was successfully written on the peer. */
729 			mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
730 						RQ_NET_SENT|RQ_NET_OK);
731 		else
732 			mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
733 		/* It is still not yet RQ_NET_DONE until the
734 		 * corresponding epoch barrier got acked as well,
735 		 * so we know what to dirty on connection loss. */
736 		break;
737 
738 	case OOS_HANDED_TO_NETWORK:
739 		/* Was not set PENDING, no longer QUEUED, so is now DONE
740 		 * as far as this connection is concerned. */
741 		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
742 		break;
743 
744 	case CONNECTION_LOST_WHILE_PENDING:
745 		/* transfer log cleanup after connection loss */
746 		mod_rq_state(req, m,
747 				RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
748 				RQ_NET_DONE);
749 		break;
750 
751 	case CONFLICT_RESOLVED:
752 		/* for superseded conflicting writes of multiple primaries,
753 		 * there is no need to keep anything in the tl, potential
754 		 * node crashes are covered by the activity log.
755 		 *
756 		 * If this request had been marked as RQ_POSTPONED before,
757 		 * it will actually not be completed, but "restarted",
758 		 * resubmitted from the retry worker context. */
759 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
760 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
761 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
762 		break;
763 
764 	case WRITE_ACKED_BY_PEER_AND_SIS:
765 		req->rq_state |= RQ_NET_SIS;
766 		fallthrough;
767 	case WRITE_ACKED_BY_PEER:
768 		/* Normal operation protocol C: successfully written on peer.
769 		 * During resync, even in protocol != C,
770 		 * we requested an explicit write ack anyways.
771 		 * Which means we cannot even assert anything here.
772 		 * Nothing more to do here.
773 		 * We want to keep the tl in place for all protocols, to cater
774 		 * for volatile write-back caches on lower level devices. */
775 		goto ack_common;
776 	case RECV_ACKED_BY_PEER:
777 		D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
778 		/* protocol B; pretends to be successfully written on peer.
779 		 * see also notes above in HANDED_OVER_TO_NETWORK about
780 		 * protocol != C */
781 	ack_common:
782 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
783 		break;
784 
785 	case POSTPONE_WRITE:
786 		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
787 		/* If this node has already detected the write conflict, the
788 		 * worker will be waiting on misc_wait.  Wake it up once this
789 		 * request has completed locally.
790 		 */
791 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
792 		req->rq_state |= RQ_POSTPONED;
793 		if (req->i.waiting)
794 			wake_up(&device->misc_wait);
795 		/* Do not clear RQ_NET_PENDING. This request will make further
796 		 * progress via restart_conflicting_writes() or
797 		 * fail_postponed_requests(). Hopefully. */
798 		break;
799 
800 	case NEG_ACKED:
801 		mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
802 		break;
803 
804 	case FAIL_FROZEN_DISK_IO:
805 		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
806 			break;
807 		mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
808 		break;
809 
810 	case RESTART_FROZEN_DISK_IO:
811 		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
812 			break;
813 
814 		mod_rq_state(req, m,
815 				RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
816 				RQ_LOCAL_PENDING);
817 
818 		rv = MR_READ;
819 		if (bio_data_dir(req->master_bio) == WRITE)
820 			rv = MR_WRITE;
821 
822 		get_ldev(device); /* always succeeds in this call path */
823 		req->w.cb = w_restart_disk_io;
824 		drbd_queue_work(&connection->sender_work,
825 				&req->w);
826 		break;
827 
828 	case RESEND:
829 		/* Simply complete (local only) READs. */
830 		if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
831 			mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
832 			break;
833 		}
834 
835 		/* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
836 		   before the connection loss (B&C only); only P_BARRIER_ACK
837 		   (or the local completion?) was missing when we suspended.
838 		   Throwing them out of the TL here by pretending we got a BARRIER_ACK.
839 		   During connection handshake, we ensure that the peer was not rebooted. */
840 		if (!(req->rq_state & RQ_NET_OK)) {
841 			/* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
842 			 * in that case we must not set RQ_NET_PENDING. */
843 
844 			mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
845 			if (req->w.cb) {
846 				/* w.cb expected to be w_send_dblock, or w_send_read_req */
847 				drbd_queue_work(&connection->sender_work,
848 						&req->w);
849 				rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
850 			} /* else: FIXME can this happen? */
851 			break;
852 		}
853 		fallthrough;	/* to BARRIER_ACKED */
854 
855 	case BARRIER_ACKED:
856 		/* barrier ack for READ requests does not make sense */
857 		if (!(req->rq_state & RQ_WRITE))
858 			break;
859 
860 		if (req->rq_state & RQ_NET_PENDING) {
861 			/* barrier came in before all requests were acked.
862 			 * this is bad, because if the connection is lost now,
863 			 * we won't be able to clean them up... */
864 			drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
865 		}
866 		/* Allowed to complete requests, even while suspended.
867 		 * As this is called for all requests within a matching epoch,
868 		 * we need to filter, and only set RQ_NET_DONE for those that
869 		 * have actually been on the wire. */
870 		mod_rq_state(req, m, RQ_COMPLETION_SUSP,
871 				(req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
872 		break;
873 
874 	case DATA_RECEIVED:
875 		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
876 		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
877 		break;
878 
879 	case QUEUE_AS_DRBD_BARRIER:
880 		start_new_tl_epoch(connection);
881 		mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
882 		break;
883 	}
884 
885 	return rv;
886 }
887 
888 /* we may do a local read if:
889  * - we are consistent (of course),
890  * - or we are generally inconsistent,
891  *   BUT we are still/already IN SYNC for this area.
892  *   since size may be bigger than BM_BLOCK_SIZE,
893  *   we may need to check several bits.
894  */
drbd_may_do_local_read(struct drbd_device * device,sector_t sector,int size)895 static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
896 {
897 	unsigned long sbnr, ebnr;
898 	sector_t esector, nr_sectors;
899 
900 	if (device->state.disk == D_UP_TO_DATE)
901 		return true;
902 	if (device->state.disk != D_INCONSISTENT)
903 		return false;
904 	esector = sector + (size >> 9) - 1;
905 	nr_sectors = get_capacity(device->vdisk);
906 	D_ASSERT(device, sector  < nr_sectors);
907 	D_ASSERT(device, esector < nr_sectors);
908 
909 	sbnr = BM_SECT_TO_BIT(sector);
910 	ebnr = BM_SECT_TO_BIT(esector);
911 
912 	return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
913 }
914 
remote_due_to_read_balancing(struct drbd_device * device,sector_t sector,enum drbd_read_balancing rbm)915 static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
916 		enum drbd_read_balancing rbm)
917 {
918 	int stripe_shift;
919 
920 	switch (rbm) {
921 	case RB_CONGESTED_REMOTE:
922 		return false;
923 	case RB_LEAST_PENDING:
924 		return atomic_read(&device->local_cnt) >
925 			atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
926 	case RB_32K_STRIPING:  /* stripe_shift = 15 */
927 	case RB_64K_STRIPING:
928 	case RB_128K_STRIPING:
929 	case RB_256K_STRIPING:
930 	case RB_512K_STRIPING:
931 	case RB_1M_STRIPING:   /* stripe_shift = 20 */
932 		stripe_shift = (rbm - RB_32K_STRIPING + 15);
933 		return (sector >> (stripe_shift - 9)) & 1;
934 	case RB_ROUND_ROBIN:
935 		return test_and_change_bit(READ_BALANCE_RR, &device->flags);
936 	case RB_PREFER_REMOTE:
937 		return true;
938 	case RB_PREFER_LOCAL:
939 	default:
940 		return false;
941 	}
942 }
943 
944 /*
945  * complete_conflicting_writes  -  wait for any conflicting write requests
946  *
947  * The write_requests tree contains all active write requests which we
948  * currently know about.  Wait for any requests to complete which conflict with
949  * the new one.
950  *
951  * Only way out: remove the conflicting intervals from the tree.
952  */
complete_conflicting_writes(struct drbd_request * req)953 static void complete_conflicting_writes(struct drbd_request *req)
954 {
955 	DEFINE_WAIT(wait);
956 	struct drbd_device *device = req->device;
957 	struct drbd_interval *i;
958 	sector_t sector = req->i.sector;
959 	int size = req->i.size;
960 
961 	for (;;) {
962 		drbd_for_each_overlap(i, &device->write_requests, sector, size) {
963 			/* Ignore, if already completed to upper layers. */
964 			if (i->completed)
965 				continue;
966 			/* Handle the first found overlap.  After the schedule
967 			 * we have to restart the tree walk. */
968 			break;
969 		}
970 		if (!i)	/* if any */
971 			break;
972 
973 		/* Indicate to wake up device->misc_wait on progress.  */
974 		prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
975 		i->waiting = true;
976 		spin_unlock_irq(&device->resource->req_lock);
977 		schedule();
978 		spin_lock_irq(&device->resource->req_lock);
979 	}
980 	finish_wait(&device->misc_wait, &wait);
981 }
982 
983 /* called within req_lock */
maybe_pull_ahead(struct drbd_device * device)984 static void maybe_pull_ahead(struct drbd_device *device)
985 {
986 	struct drbd_connection *connection = first_peer_device(device)->connection;
987 	struct net_conf *nc;
988 	bool congested = false;
989 	enum drbd_on_congestion on_congestion;
990 
991 	rcu_read_lock();
992 	nc = rcu_dereference(connection->net_conf);
993 	on_congestion = nc ? nc->on_congestion : OC_BLOCK;
994 	rcu_read_unlock();
995 	if (on_congestion == OC_BLOCK ||
996 	    connection->agreed_pro_version < 96)
997 		return;
998 
999 	if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1000 		return; /* nothing to do ... */
1001 
1002 	/* If I don't even have good local storage, we can not reasonably try
1003 	 * to pull ahead of the peer. We also need the local reference to make
1004 	 * sure device->act_log is there.
1005 	 */
1006 	if (!get_ldev_if_state(device, D_UP_TO_DATE))
1007 		return;
1008 
1009 	if (nc->cong_fill &&
1010 	    atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1011 		drbd_info(device, "Congestion-fill threshold reached\n");
1012 		congested = true;
1013 	}
1014 
1015 	if (device->act_log->used >= nc->cong_extents) {
1016 		drbd_info(device, "Congestion-extents threshold reached\n");
1017 		congested = true;
1018 	}
1019 
1020 	if (congested) {
1021 		/* start a new epoch for non-mirrored writes */
1022 		start_new_tl_epoch(first_peer_device(device)->connection);
1023 
1024 		if (on_congestion == OC_PULL_AHEAD)
1025 			_drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1026 		else  /*nc->on_congestion == OC_DISCONNECT */
1027 			_drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1028 	}
1029 	put_ldev(device);
1030 }
1031 
1032 /* If this returns false, and req->private_bio is still set,
1033  * this should be submitted locally.
1034  *
1035  * If it returns false, but req->private_bio is not set,
1036  * we do not have access to good data :(
1037  *
1038  * Otherwise, this destroys req->private_bio, if any,
1039  * and returns true.
1040  */
do_remote_read(struct drbd_request * req)1041 static bool do_remote_read(struct drbd_request *req)
1042 {
1043 	struct drbd_device *device = req->device;
1044 	enum drbd_read_balancing rbm;
1045 
1046 	if (req->private_bio) {
1047 		if (!drbd_may_do_local_read(device,
1048 					req->i.sector, req->i.size)) {
1049 			bio_put(req->private_bio);
1050 			req->private_bio = NULL;
1051 			put_ldev(device);
1052 		}
1053 	}
1054 
1055 	if (device->state.pdsk != D_UP_TO_DATE)
1056 		return false;
1057 
1058 	if (req->private_bio == NULL)
1059 		return true;
1060 
1061 	/* TODO: improve read balancing decisions, take into account drbd
1062 	 * protocol, pending requests etc. */
1063 
1064 	rcu_read_lock();
1065 	rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1066 	rcu_read_unlock();
1067 
1068 	if (rbm == RB_PREFER_LOCAL && req->private_bio)
1069 		return false; /* submit locally */
1070 
1071 	if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1072 		if (req->private_bio) {
1073 			bio_put(req->private_bio);
1074 			req->private_bio = NULL;
1075 			put_ldev(device);
1076 		}
1077 		return true;
1078 	}
1079 
1080 	return false;
1081 }
1082 
drbd_should_do_remote(union drbd_dev_state s)1083 bool drbd_should_do_remote(union drbd_dev_state s)
1084 {
1085 	return s.pdsk == D_UP_TO_DATE ||
1086 		(s.pdsk >= D_INCONSISTENT &&
1087 		 s.conn >= C_WF_BITMAP_T &&
1088 		 s.conn < C_AHEAD);
1089 	/* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1090 	   That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1091 	   states. */
1092 }
1093 
drbd_should_send_out_of_sync(union drbd_dev_state s)1094 static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1095 {
1096 	return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1097 	/* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1098 	   since we enter state C_AHEAD only if proto >= 96 */
1099 }
1100 
1101 /* returns number of connections (== 1, for drbd 8.4)
1102  * expected to actually write this data,
1103  * which does NOT include those that we are L_AHEAD for. */
drbd_process_write_request(struct drbd_request * req)1104 static int drbd_process_write_request(struct drbd_request *req)
1105 {
1106 	struct drbd_device *device = req->device;
1107 	struct drbd_peer_device *peer_device = first_peer_device(device);
1108 	int remote, send_oos;
1109 
1110 	remote = drbd_should_do_remote(device->state);
1111 	send_oos = drbd_should_send_out_of_sync(device->state);
1112 
1113 	/* Need to replicate writes.  Unless it is an empty flush,
1114 	 * which is better mapped to a DRBD P_BARRIER packet,
1115 	 * also for drbd wire protocol compatibility reasons.
1116 	 * If this was a flush, just start a new epoch.
1117 	 * Unless the current epoch was empty anyways, or we are not currently
1118 	 * replicating, in which case there is no point. */
1119 	if (unlikely(req->i.size == 0)) {
1120 		/* The only size==0 bios we expect are empty flushes. */
1121 		D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1122 		if (remote)
1123 			_req_mod(req, QUEUE_AS_DRBD_BARRIER, peer_device);
1124 		return remote;
1125 	}
1126 
1127 	if (!remote && !send_oos)
1128 		return 0;
1129 
1130 	D_ASSERT(device, !(remote && send_oos));
1131 
1132 	if (remote) {
1133 		_req_mod(req, TO_BE_SENT, peer_device);
1134 		_req_mod(req, QUEUE_FOR_NET_WRITE, peer_device);
1135 	} else if (drbd_set_out_of_sync(peer_device, req->i.sector, req->i.size))
1136 		_req_mod(req, QUEUE_FOR_SEND_OOS, peer_device);
1137 
1138 	return remote;
1139 }
1140 
drbd_process_discard_or_zeroes_req(struct drbd_request * req,int flags)1141 static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1142 {
1143 	int err = drbd_issue_discard_or_zero_out(req->device,
1144 				req->i.sector, req->i.size >> 9, flags);
1145 	if (err)
1146 		req->private_bio->bi_status = BLK_STS_IOERR;
1147 	bio_endio(req->private_bio);
1148 }
1149 
1150 static void
drbd_submit_req_private_bio(struct drbd_request * req)1151 drbd_submit_req_private_bio(struct drbd_request *req)
1152 {
1153 	struct drbd_device *device = req->device;
1154 	struct bio *bio = req->private_bio;
1155 	unsigned int type;
1156 
1157 	if (bio_op(bio) != REQ_OP_READ)
1158 		type = DRBD_FAULT_DT_WR;
1159 	else if (bio->bi_opf & REQ_RAHEAD)
1160 		type = DRBD_FAULT_DT_RA;
1161 	else
1162 		type = DRBD_FAULT_DT_RD;
1163 
1164 	/* State may have changed since we grabbed our reference on the
1165 	 * ->ldev member. Double check, and short-circuit to endio.
1166 	 * In case the last activity log transaction failed to get on
1167 	 * stable storage, and this is a WRITE, we may not even submit
1168 	 * this bio. */
1169 	if (get_ldev(device)) {
1170 		if (drbd_insert_fault(device, type))
1171 			bio_io_error(bio);
1172 		else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1173 			drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1174 			    ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1175 		else if (bio_op(bio) == REQ_OP_DISCARD)
1176 			drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1177 		else
1178 			submit_bio_noacct(bio);
1179 		put_ldev(device);
1180 	} else
1181 		bio_io_error(bio);
1182 }
1183 
drbd_queue_write(struct drbd_device * device,struct drbd_request * req)1184 static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1185 {
1186 	spin_lock_irq(&device->resource->req_lock);
1187 	list_add_tail(&req->tl_requests, &device->submit.writes);
1188 	list_add_tail(&req->req_pending_master_completion,
1189 			&device->pending_master_completion[1 /* WRITE */]);
1190 	spin_unlock_irq(&device->resource->req_lock);
1191 	queue_work(device->submit.wq, &device->submit.worker);
1192 	/* do_submit() may sleep internally on al_wait, too */
1193 	wake_up(&device->al_wait);
1194 }
1195 
1196 /* returns the new drbd_request pointer, if the caller is expected to
1197  * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1198  * request on the submitter thread.
1199  * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1200  */
1201 static struct drbd_request *
drbd_request_prepare(struct drbd_device * device,struct bio * bio)1202 drbd_request_prepare(struct drbd_device *device, struct bio *bio)
1203 {
1204 	const int rw = bio_data_dir(bio);
1205 	struct drbd_request *req;
1206 
1207 	/* allocate outside of all locks; */
1208 	req = drbd_req_new(device, bio);
1209 	if (!req) {
1210 		dec_ap_bio(device);
1211 		/* only pass the error to the upper layers.
1212 		 * if user cannot handle io errors, that's not our business. */
1213 		drbd_err(device, "could not kmalloc() req\n");
1214 		bio->bi_status = BLK_STS_RESOURCE;
1215 		bio_endio(bio);
1216 		return ERR_PTR(-ENOMEM);
1217 	}
1218 
1219 	/* Update disk stats */
1220 	req->start_jif = bio_start_io_acct(req->master_bio);
1221 
1222 	if (get_ldev(device)) {
1223 		req->private_bio = bio_alloc_clone(device->ldev->backing_bdev,
1224 						   bio, GFP_NOIO,
1225 						   &drbd_io_bio_set);
1226 		req->private_bio->bi_private = req;
1227 		req->private_bio->bi_end_io = drbd_request_endio;
1228 	}
1229 
1230 	/* process discards always from our submitter thread */
1231 	if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1232 	    bio_op(bio) == REQ_OP_DISCARD)
1233 		goto queue_for_submitter_thread;
1234 
1235 	if (rw == WRITE && req->private_bio && req->i.size
1236 	&& !test_bit(AL_SUSPENDED, &device->flags)) {
1237 		if (!drbd_al_begin_io_fastpath(device, &req->i))
1238 			goto queue_for_submitter_thread;
1239 		req->rq_state |= RQ_IN_ACT_LOG;
1240 		req->in_actlog_jif = jiffies;
1241 	}
1242 	return req;
1243 
1244  queue_for_submitter_thread:
1245 	atomic_inc(&device->ap_actlog_cnt);
1246 	drbd_queue_write(device, req);
1247 	return NULL;
1248 }
1249 
1250 /* Require at least one path to current data.
1251  * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1252  * We would not allow to read what was written,
1253  * we would not have bumped the data generation uuids,
1254  * we would cause data divergence for all the wrong reasons.
1255  *
1256  * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1257  * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1258  * and queues for retry later.
1259  */
may_do_writes(struct drbd_device * device)1260 static bool may_do_writes(struct drbd_device *device)
1261 {
1262 	const union drbd_dev_state s = device->state;
1263 	return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1264 }
1265 
1266 struct drbd_plug_cb {
1267 	struct blk_plug_cb cb;
1268 	struct drbd_request *most_recent_req;
1269 	/* do we need more? */
1270 };
1271 
drbd_unplug(struct blk_plug_cb * cb,bool from_schedule)1272 static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1273 {
1274 	struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1275 	struct drbd_resource *resource = plug->cb.data;
1276 	struct drbd_request *req = plug->most_recent_req;
1277 
1278 	kfree(cb);
1279 	if (!req)
1280 		return;
1281 
1282 	spin_lock_irq(&resource->req_lock);
1283 	/* In case the sender did not process it yet, raise the flag to
1284 	 * have it followed with P_UNPLUG_REMOTE just after. */
1285 	req->rq_state |= RQ_UNPLUG;
1286 	/* but also queue a generic unplug */
1287 	drbd_queue_unplug(req->device);
1288 	kref_put(&req->kref, drbd_req_destroy);
1289 	spin_unlock_irq(&resource->req_lock);
1290 }
1291 
drbd_check_plugged(struct drbd_resource * resource)1292 static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1293 {
1294 	/* A lot of text to say
1295 	 * return (struct drbd_plug_cb*)blk_check_plugged(); */
1296 	struct drbd_plug_cb *plug;
1297 	struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1298 
1299 	if (cb)
1300 		plug = container_of(cb, struct drbd_plug_cb, cb);
1301 	else
1302 		plug = NULL;
1303 	return plug;
1304 }
1305 
drbd_update_plug(struct drbd_plug_cb * plug,struct drbd_request * req)1306 static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1307 {
1308 	struct drbd_request *tmp = plug->most_recent_req;
1309 	/* Will be sent to some peer.
1310 	 * Remember to tag it with UNPLUG_REMOTE on unplug */
1311 	kref_get(&req->kref);
1312 	plug->most_recent_req = req;
1313 	if (tmp)
1314 		kref_put(&tmp->kref, drbd_req_destroy);
1315 }
1316 
drbd_send_and_submit(struct drbd_device * device,struct drbd_request * req)1317 static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1318 {
1319 	struct drbd_resource *resource = device->resource;
1320 	struct drbd_peer_device *peer_device = first_peer_device(device);
1321 	const int rw = bio_data_dir(req->master_bio);
1322 	struct bio_and_error m = { NULL, };
1323 	bool no_remote = false;
1324 	bool submit_private_bio = false;
1325 
1326 	spin_lock_irq(&resource->req_lock);
1327 	if (rw == WRITE) {
1328 		/* This may temporarily give up the req_lock,
1329 		 * but will re-aquire it before it returns here.
1330 		 * Needs to be before the check on drbd_suspended() */
1331 		complete_conflicting_writes(req);
1332 		/* no more giving up req_lock from now on! */
1333 
1334 		/* check for congestion, and potentially stop sending
1335 		 * full data updates, but start sending "dirty bits" only. */
1336 		maybe_pull_ahead(device);
1337 	}
1338 
1339 
1340 	if (drbd_suspended(device)) {
1341 		/* push back and retry: */
1342 		req->rq_state |= RQ_POSTPONED;
1343 		if (req->private_bio) {
1344 			bio_put(req->private_bio);
1345 			req->private_bio = NULL;
1346 			put_ldev(device);
1347 		}
1348 		goto out;
1349 	}
1350 
1351 	/* We fail READ early, if we can not serve it.
1352 	 * We must do this before req is registered on any lists.
1353 	 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1354 	if (rw != WRITE) {
1355 		if (!do_remote_read(req) && !req->private_bio)
1356 			goto nodata;
1357 	}
1358 
1359 	/* which transfer log epoch does this belong to? */
1360 	req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1361 
1362 	/* no point in adding empty flushes to the transfer log,
1363 	 * they are mapped to drbd barriers already. */
1364 	if (likely(req->i.size!=0)) {
1365 		if (rw == WRITE)
1366 			first_peer_device(device)->connection->current_tle_writes++;
1367 
1368 		list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1369 	}
1370 
1371 	if (rw == WRITE) {
1372 		if (req->private_bio && !may_do_writes(device)) {
1373 			bio_put(req->private_bio);
1374 			req->private_bio = NULL;
1375 			put_ldev(device);
1376 			goto nodata;
1377 		}
1378 		if (!drbd_process_write_request(req))
1379 			no_remote = true;
1380 	} else {
1381 		/* We either have a private_bio, or we can read from remote.
1382 		 * Otherwise we had done the goto nodata above. */
1383 		if (req->private_bio == NULL) {
1384 			_req_mod(req, TO_BE_SENT, peer_device);
1385 			_req_mod(req, QUEUE_FOR_NET_READ, peer_device);
1386 		} else
1387 			no_remote = true;
1388 	}
1389 
1390 	if (no_remote == false) {
1391 		struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1392 		if (plug)
1393 			drbd_update_plug(plug, req);
1394 	}
1395 
1396 	/* If it took the fast path in drbd_request_prepare, add it here.
1397 	 * The slow path has added it already. */
1398 	if (list_empty(&req->req_pending_master_completion))
1399 		list_add_tail(&req->req_pending_master_completion,
1400 			&device->pending_master_completion[rw == WRITE]);
1401 	if (req->private_bio) {
1402 		/* needs to be marked within the same spinlock */
1403 		req->pre_submit_jif = jiffies;
1404 		list_add_tail(&req->req_pending_local,
1405 			&device->pending_completion[rw == WRITE]);
1406 		_req_mod(req, TO_BE_SUBMITTED, NULL);
1407 		/* but we need to give up the spinlock to submit */
1408 		submit_private_bio = true;
1409 	} else if (no_remote) {
1410 nodata:
1411 		if (drbd_ratelimit())
1412 			drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1413 					(unsigned long long)req->i.sector, req->i.size >> 9);
1414 		/* A write may have been queued for send_oos, however.
1415 		 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1416 	}
1417 
1418 out:
1419 	drbd_req_put_completion_ref(req, &m, 1);
1420 	spin_unlock_irq(&resource->req_lock);
1421 
1422 	/* Even though above is a kref_put(), this is safe.
1423 	 * As long as we still need to submit our private bio,
1424 	 * we hold a completion ref, and the request cannot disappear.
1425 	 * If however this request did not even have a private bio to submit
1426 	 * (e.g. remote read), req may already be invalid now.
1427 	 * That's why we cannot check on req->private_bio. */
1428 	if (submit_private_bio)
1429 		drbd_submit_req_private_bio(req);
1430 	if (m.bio)
1431 		complete_master_bio(device, &m);
1432 }
1433 
__drbd_make_request(struct drbd_device * device,struct bio * bio)1434 void __drbd_make_request(struct drbd_device *device, struct bio *bio)
1435 {
1436 	struct drbd_request *req = drbd_request_prepare(device, bio);
1437 	if (IS_ERR_OR_NULL(req))
1438 		return;
1439 	drbd_send_and_submit(device, req);
1440 }
1441 
submit_fast_path(struct drbd_device * device,struct list_head * incoming)1442 static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1443 {
1444 	struct blk_plug plug;
1445 	struct drbd_request *req, *tmp;
1446 
1447 	blk_start_plug(&plug);
1448 	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1449 		const int rw = bio_data_dir(req->master_bio);
1450 
1451 		if (rw == WRITE /* rw != WRITE should not even end up here! */
1452 		&& req->private_bio && req->i.size
1453 		&& !test_bit(AL_SUSPENDED, &device->flags)) {
1454 			if (!drbd_al_begin_io_fastpath(device, &req->i))
1455 				continue;
1456 
1457 			req->rq_state |= RQ_IN_ACT_LOG;
1458 			req->in_actlog_jif = jiffies;
1459 			atomic_dec(&device->ap_actlog_cnt);
1460 		}
1461 
1462 		list_del_init(&req->tl_requests);
1463 		drbd_send_and_submit(device, req);
1464 	}
1465 	blk_finish_plug(&plug);
1466 }
1467 
prepare_al_transaction_nonblock(struct drbd_device * device,struct list_head * incoming,struct list_head * pending,struct list_head * later)1468 static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1469 					    struct list_head *incoming,
1470 					    struct list_head *pending,
1471 					    struct list_head *later)
1472 {
1473 	struct drbd_request *req;
1474 	int wake = 0;
1475 	int err;
1476 
1477 	spin_lock_irq(&device->al_lock);
1478 	while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1479 		err = drbd_al_begin_io_nonblock(device, &req->i);
1480 		if (err == -ENOBUFS)
1481 			break;
1482 		if (err == -EBUSY)
1483 			wake = 1;
1484 		if (err)
1485 			list_move_tail(&req->tl_requests, later);
1486 		else
1487 			list_move_tail(&req->tl_requests, pending);
1488 	}
1489 	spin_unlock_irq(&device->al_lock);
1490 	if (wake)
1491 		wake_up(&device->al_wait);
1492 	return !list_empty(pending);
1493 }
1494 
send_and_submit_pending(struct drbd_device * device,struct list_head * pending)1495 static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1496 {
1497 	struct blk_plug plug;
1498 	struct drbd_request *req;
1499 
1500 	blk_start_plug(&plug);
1501 	while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1502 		req->rq_state |= RQ_IN_ACT_LOG;
1503 		req->in_actlog_jif = jiffies;
1504 		atomic_dec(&device->ap_actlog_cnt);
1505 		list_del_init(&req->tl_requests);
1506 		drbd_send_and_submit(device, req);
1507 	}
1508 	blk_finish_plug(&plug);
1509 }
1510 
do_submit(struct work_struct * ws)1511 void do_submit(struct work_struct *ws)
1512 {
1513 	struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1514 	LIST_HEAD(incoming);	/* from drbd_make_request() */
1515 	LIST_HEAD(pending);	/* to be submitted after next AL-transaction commit */
1516 	LIST_HEAD(busy);	/* blocked by resync requests */
1517 
1518 	/* grab new incoming requests */
1519 	spin_lock_irq(&device->resource->req_lock);
1520 	list_splice_tail_init(&device->submit.writes, &incoming);
1521 	spin_unlock_irq(&device->resource->req_lock);
1522 
1523 	for (;;) {
1524 		DEFINE_WAIT(wait);
1525 
1526 		/* move used-to-be-busy back to front of incoming */
1527 		list_splice_init(&busy, &incoming);
1528 		submit_fast_path(device, &incoming);
1529 		if (list_empty(&incoming))
1530 			break;
1531 
1532 		for (;;) {
1533 			prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1534 
1535 			list_splice_init(&busy, &incoming);
1536 			prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1537 			if (!list_empty(&pending))
1538 				break;
1539 
1540 			schedule();
1541 
1542 			/* If all currently "hot" activity log extents are kept busy by
1543 			 * incoming requests, we still must not totally starve new
1544 			 * requests to "cold" extents.
1545 			 * Something left on &incoming means there had not been
1546 			 * enough update slots available, and the activity log
1547 			 * has been marked as "starving".
1548 			 *
1549 			 * Try again now, without looking for new requests,
1550 			 * effectively blocking all new requests until we made
1551 			 * at least _some_ progress with what we currently have.
1552 			 */
1553 			if (!list_empty(&incoming))
1554 				continue;
1555 
1556 			/* Nothing moved to pending, but nothing left
1557 			 * on incoming: all moved to busy!
1558 			 * Grab new and iterate. */
1559 			spin_lock_irq(&device->resource->req_lock);
1560 			list_splice_tail_init(&device->submit.writes, &incoming);
1561 			spin_unlock_irq(&device->resource->req_lock);
1562 		}
1563 		finish_wait(&device->al_wait, &wait);
1564 
1565 		/* If the transaction was full, before all incoming requests
1566 		 * had been processed, skip ahead to commit, and iterate
1567 		 * without splicing in more incoming requests from upper layers.
1568 		 *
1569 		 * Else, if all incoming have been processed,
1570 		 * they have become either "pending" (to be submitted after
1571 		 * next transaction commit) or "busy" (blocked by resync).
1572 		 *
1573 		 * Maybe more was queued, while we prepared the transaction?
1574 		 * Try to stuff those into this transaction as well.
1575 		 * Be strictly non-blocking here,
1576 		 * we already have something to commit.
1577 		 *
1578 		 * Commit if we don't make any more progres.
1579 		 */
1580 
1581 		while (list_empty(&incoming)) {
1582 			LIST_HEAD(more_pending);
1583 			LIST_HEAD(more_incoming);
1584 			bool made_progress;
1585 
1586 			/* It is ok to look outside the lock,
1587 			 * it's only an optimization anyways */
1588 			if (list_empty(&device->submit.writes))
1589 				break;
1590 
1591 			spin_lock_irq(&device->resource->req_lock);
1592 			list_splice_tail_init(&device->submit.writes, &more_incoming);
1593 			spin_unlock_irq(&device->resource->req_lock);
1594 
1595 			if (list_empty(&more_incoming))
1596 				break;
1597 
1598 			made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1599 
1600 			list_splice_tail_init(&more_pending, &pending);
1601 			list_splice_tail_init(&more_incoming, &incoming);
1602 			if (!made_progress)
1603 				break;
1604 		}
1605 
1606 		drbd_al_begin_io_commit(device);
1607 		send_and_submit_pending(device, &pending);
1608 	}
1609 }
1610 
drbd_submit_bio(struct bio * bio)1611 void drbd_submit_bio(struct bio *bio)
1612 {
1613 	struct drbd_device *device = bio->bi_bdev->bd_disk->private_data;
1614 
1615 	bio = bio_split_to_limits(bio);
1616 	if (!bio)
1617 		return;
1618 
1619 	/*
1620 	 * what we "blindly" assume:
1621 	 */
1622 	D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1623 
1624 	inc_ap_bio(device);
1625 	__drbd_make_request(device, bio);
1626 }
1627 
net_timeout_reached(struct drbd_request * net_req,struct drbd_connection * connection,unsigned long now,unsigned long ent,unsigned int ko_count,unsigned int timeout)1628 static bool net_timeout_reached(struct drbd_request *net_req,
1629 		struct drbd_connection *connection,
1630 		unsigned long now, unsigned long ent,
1631 		unsigned int ko_count, unsigned int timeout)
1632 {
1633 	struct drbd_device *device = net_req->device;
1634 
1635 	if (!time_after(now, net_req->pre_send_jif + ent))
1636 		return false;
1637 
1638 	if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1639 		return false;
1640 
1641 	if (net_req->rq_state & RQ_NET_PENDING) {
1642 		drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1643 			jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1644 		return true;
1645 	}
1646 
1647 	/* We received an ACK already (or are using protocol A),
1648 	 * but are waiting for the epoch closing barrier ack.
1649 	 * Check if we sent the barrier already.  We should not blame the peer
1650 	 * for being unresponsive, if we did not even ask it yet. */
1651 	if (net_req->epoch == connection->send.current_epoch_nr) {
1652 		drbd_warn(device,
1653 			"We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1654 			jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1655 		return false;
1656 	}
1657 
1658 	/* Worst case: we may have been blocked for whatever reason, then
1659 	 * suddenly are able to send a lot of requests (and epoch separating
1660 	 * barriers) in quick succession.
1661 	 * The timestamp of the net_req may be much too old and not correspond
1662 	 * to the sending time of the relevant unack'ed barrier packet, so
1663 	 * would trigger a spurious timeout.  The latest barrier packet may
1664 	 * have a too recent timestamp to trigger the timeout, potentially miss
1665 	 * a timeout.  Right now we don't have a place to conveniently store
1666 	 * these timestamps.
1667 	 * But in this particular situation, the application requests are still
1668 	 * completed to upper layers, DRBD should still "feel" responsive.
1669 	 * No need yet to kill this connection, it may still recover.
1670 	 * If not, eventually we will have queued enough into the network for
1671 	 * us to block. From that point of view, the timestamp of the last sent
1672 	 * barrier packet is relevant enough.
1673 	 */
1674 	if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1675 		drbd_warn(device, "Remote failed to answer a P_BARRIER (sent at %lu jif; now=%lu jif) within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1676 			connection->send.last_sent_barrier_jif, now,
1677 			jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1678 		return true;
1679 	}
1680 	return false;
1681 }
1682 
1683 /* A request is considered timed out, if
1684  * - we have some effective timeout from the configuration,
1685  *   with some state restrictions applied,
1686  * - the oldest request is waiting for a response from the network
1687  *   resp. the local disk,
1688  * - the oldest request is in fact older than the effective timeout,
1689  * - the connection was established (resp. disk was attached)
1690  *   for longer than the timeout already.
1691  * Note that for 32bit jiffies and very stable connections/disks,
1692  * we may have a wrap around, which is catched by
1693  *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1694  *
1695  * Side effect: once per 32bit wrap-around interval, which means every
1696  * ~198 days with 250 HZ, we have a window where the timeout would need
1697  * to expire twice (worst case) to become effective. Good enough.
1698  */
1699 
request_timer_fn(struct timer_list * t)1700 void request_timer_fn(struct timer_list *t)
1701 {
1702 	struct drbd_device *device = from_timer(device, t, request_timer);
1703 	struct drbd_connection *connection = first_peer_device(device)->connection;
1704 	struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1705 	struct net_conf *nc;
1706 	unsigned long oldest_submit_jif;
1707 	unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1708 	unsigned long now;
1709 	unsigned int ko_count = 0, timeout = 0;
1710 
1711 	rcu_read_lock();
1712 	nc = rcu_dereference(connection->net_conf);
1713 	if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1714 		ko_count = nc->ko_count;
1715 		timeout = nc->timeout;
1716 	}
1717 
1718 	if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1719 		dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1720 		put_ldev(device);
1721 	}
1722 	rcu_read_unlock();
1723 
1724 
1725 	ent = timeout * HZ/10 * ko_count;
1726 	et = min_not_zero(dt, ent);
1727 
1728 	if (!et)
1729 		return; /* Recurring timer stopped */
1730 
1731 	now = jiffies;
1732 	nt = now + et;
1733 
1734 	spin_lock_irq(&device->resource->req_lock);
1735 	req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1736 	req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1737 
1738 	/* maybe the oldest request waiting for the peer is in fact still
1739 	 * blocking in tcp sendmsg.  That's ok, though, that's handled via the
1740 	 * socket send timeout, requesting a ping, and bumping ko-count in
1741 	 * we_should_drop_the_connection().
1742 	 */
1743 
1744 	/* check the oldest request we did successfully sent,
1745 	 * but which is still waiting for an ACK. */
1746 	req_peer = connection->req_ack_pending;
1747 
1748 	/* if we don't have such request (e.g. protocoll A)
1749 	 * check the oldest requests which is still waiting on its epoch
1750 	 * closing barrier ack. */
1751 	if (!req_peer)
1752 		req_peer = connection->req_not_net_done;
1753 
1754 	/* evaluate the oldest peer request only in one timer! */
1755 	if (req_peer && req_peer->device != device)
1756 		req_peer = NULL;
1757 
1758 	/* do we have something to evaluate? */
1759 	if (req_peer == NULL && req_write == NULL && req_read == NULL)
1760 		goto out;
1761 
1762 	oldest_submit_jif =
1763 		(req_write && req_read)
1764 		? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1765 		  ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1766 		: req_write ? req_write->pre_submit_jif
1767 		: req_read ? req_read->pre_submit_jif : now;
1768 
1769 	if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1770 		_conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1771 
1772 	if (dt && oldest_submit_jif != now &&
1773 		 time_after(now, oldest_submit_jif + dt) &&
1774 		!time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1775 		drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1776 		__drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1777 	}
1778 
1779 	/* Reschedule timer for the nearest not already expired timeout.
1780 	 * Fallback to now + min(effective network timeout, disk timeout). */
1781 	ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1782 		? req_peer->pre_send_jif + ent : now + et;
1783 	dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1784 		? oldest_submit_jif + dt : now + et;
1785 	nt = time_before(ent, dt) ? ent : dt;
1786 out:
1787 	spin_unlock_irq(&device->resource->req_lock);
1788 	mod_timer(&device->request_timer, nt);
1789 }
1790