drbd_req.c 42.1 KB
Newer Older
P
Philipp Reisner 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
/*
   drbd_req.c

   This file is part of DRBD by Philipp Reisner and Lars Ellenberg.

   Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
   Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
   Copyright (C) 2002-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.

   drbd is free software; you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
   the Free Software Foundation; either version 2, or (at your option)
   any later version.

   drbd is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with drbd; see the file COPYING.  If not, write to
   the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.

 */

#include <linux/module.h>

#include <linux/slab.h>
#include <linux/drbd.h>
#include "drbd_int.h"
#include "drbd_req.h"


34
static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
35

P
Philipp Reisner 已提交
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)
P
Philipp Reisner 已提交
38
{
39
	const int rw = bio_data_dir(req->master_bio);
P
Philipp Reisner 已提交
40 41
	int cpu;
	cpu = part_stat_lock();
42 43 44
	part_round_stats(cpu, &device->vdisk->part0);
	part_stat_inc(cpu, &device->vdisk->part0, ios[rw]);
	part_stat_add(cpu, &device->vdisk->part0, sectors[rw], req->i.size >> 9);
P
Philipp Reisner 已提交
45 46
	(void) cpu; /* The macro invocations above want the cpu argument, I do not like
		       the compiler warning about cpu only assigned but never used... */
47
	part_inc_in_flight(&device->vdisk->part0, rw);
P
Philipp Reisner 已提交
48 49 50 51
	part_stat_unlock();
}

/* Update disk stats when completing request upwards */
52
static void _drbd_end_io_acct(struct drbd_device *device, struct drbd_request *req)
P
Philipp Reisner 已提交
53 54 55 56 57
{
	int rw = bio_data_dir(req->master_bio);
	unsigned long duration = jiffies - req->start_time;
	int cpu;
	cpu = part_stat_lock();
58 59 60
	part_stat_add(cpu, &device->vdisk->part0, ticks[rw], duration);
	part_round_stats(cpu, &device->vdisk->part0);
	part_dec_in_flight(&device->vdisk->part0, rw);
P
Philipp Reisner 已提交
61 62 63
	part_stat_unlock();
}

64
static struct drbd_request *drbd_req_new(struct drbd_device *device,
65 66 67 68 69 70 71 72 73 74
					       struct bio *bio_src)
{
	struct drbd_request *req;

	req = mempool_alloc(drbd_request_mempool, GFP_NOIO);
	if (!req)
		return NULL;

	drbd_req_make_private_bio(req, bio_src);
	req->rq_state    = bio_data_dir(bio_src) == WRITE ? RQ_WRITE : 0;
75
	req->w.device      = device;
76 77
	req->master_bio  = bio_src;
	req->epoch       = 0;
78

79
	drbd_clear_interval(&req->i);
80 81
	req->i.sector     = bio_src->bi_iter.bi_sector;
	req->i.size      = bio_src->bi_iter.bi_size;
82
	req->i.local = true;
83 84
	req->i.waiting = false;

85 86 87
	INIT_LIST_HEAD(&req->tl_requests);
	INIT_LIST_HEAD(&req->w.list);

88
	/* one reference to be put by __drbd_make_request */
89
	atomic_set(&req->completion_ref, 1);
90
	/* one kref as long as completion_ref > 0 */
91
	kref_init(&req->kref);
92 93 94
	return req;
}

95
void drbd_req_destroy(struct kref *kref)
P
Philipp Reisner 已提交
96
{
97
	struct drbd_request *req = container_of(kref, struct drbd_request, kref);
98
	struct drbd_device *device = req->w.device;
99 100 101 102 103 104 105 106 107 108
	const unsigned s = req->rq_state;

	if ((req->master_bio && !(s & RQ_POSTPONED)) ||
		atomic_read(&req->completion_ref) ||
		(s & RQ_LOCAL_PENDING) ||
		((s & RQ_NET_MASK) && !(s & RQ_NET_DONE))) {
		dev_err(DEV, "drbd_req_destroy: Logic BUG rq_state = 0x%x, completion_ref = %d\n",
				s, atomic_read(&req->completion_ref));
		return;
	}
109 110 111 112 113 114 115

	/* remove it from the transfer log.
	 * well, only if it had been there in the first
	 * place... if it had not (local only or conflicting
	 * and never sent), it should still be "empty" as
	 * initialized in drbd_req_new(), so we can list_del() it
	 * here unconditionally */
116
	list_del_init(&req->tl_requests);
117

P
Philipp Reisner 已提交
118 119 120
	/* if it was a write, we may have to set the corresponding
	 * bit(s) out-of-sync first. If it had a local part, we need to
	 * release the reference to the activity log. */
121
	if (s & RQ_WRITE) {
P
Philipp Reisner 已提交
122 123 124 125 126
		/* Set out-of-sync unless both OK flags are set
		 * (local only or remote failed).
		 * Other places where we set out-of-sync:
		 * READ with local io-error */

127 128 129 130 131 132 133
		/* There is a special case:
		 * we may notice late that IO was suspended,
		 * and postpone, or schedule for retry, a write,
		 * before it even was submitted or sent.
		 * In that case we do not want to touch the bitmap at all.
		 */
		if ((s & (RQ_POSTPONED|RQ_LOCAL_MASK|RQ_NET_MASK)) != RQ_POSTPONED) {
P
Philipp Reisner 已提交
134
			if (!(s & RQ_NET_OK) || !(s & RQ_LOCAL_OK))
135
				drbd_set_out_of_sync(device, req->i.sector, req->i.size);
P
Philipp Reisner 已提交
136

P
Philipp Reisner 已提交
137
			if ((s & RQ_NET_OK) && (s & RQ_LOCAL_OK) && (s & RQ_NET_SIS))
138
				drbd_set_in_sync(device, req->i.sector, req->i.size);
P
Philipp Reisner 已提交
139
		}
P
Philipp Reisner 已提交
140 141

		/* one might be tempted to move the drbd_al_complete_io
142
		 * to the local io completion callback drbd_request_endio.
P
Philipp Reisner 已提交
143 144 145 146 147 148 149 150
		 * but, if this was a mirror write, we may only
		 * drbd_al_complete_io after this is RQ_NET_DONE,
		 * otherwise the extent could be dropped from the al
		 * before it has actually been written on the peer.
		 * if we crash before our peer knows about the request,
		 * but after the extent has been dropped from the al,
		 * we would forget to resync the corresponding extent.
		 */
P
Philipp Reisner 已提交
151
		if (s & RQ_IN_ACT_LOG) {
152 153 154
			if (get_ldev_if_state(device, D_FAILED)) {
				drbd_al_complete_io(device, &req->i);
				put_ldev(device);
P
Philipp Reisner 已提交
155
			} else if (__ratelimit(&drbd_ratelimit_state)) {
156 157 158
				dev_warn(DEV, "Should have called drbd_al_complete_io(, %llu, %u), "
					 "but my Disk seems to have failed :(\n",
					 (unsigned long long) req->i.sector, req->i.size);
P
Philipp Reisner 已提交
159 160 161 162
			}
		}
	}

163
	mempool_free(req, drbd_request_mempool);
P
Philipp Reisner 已提交
164 165
}

166 167 168
static void wake_all_senders(struct drbd_connection *connection)
{
	wake_up(&connection->sender_work.q_wait);
P
Philipp Reisner 已提交
169 170
}

171
/* must hold resource->req_lock */
172
void start_new_tl_epoch(struct drbd_connection *connection)
P
Philipp Reisner 已提交
173
{
174
	/* no point closing an epoch, if it is empty, anyways. */
175
	if (connection->current_tle_writes == 0)
176
		return;
P
Philipp Reisner 已提交
177

178 179 180
	connection->current_tle_writes = 0;
	atomic_inc(&connection->current_tle_nr);
	wake_all_senders(connection);
P
Philipp Reisner 已提交
181 182
}

183
void complete_master_bio(struct drbd_device *device,
P
Philipp Reisner 已提交
184 185 186
		struct bio_and_error *m)
{
	bio_endio(m->bio, m->error);
187
	dec_ap_bio(device);
P
Philipp Reisner 已提交
188 189
}

190 191 192 193

static void drbd_remove_request_interval(struct rb_root *root,
					 struct drbd_request *req)
{
194
	struct drbd_device *device = req->w.device;
195 196 197 198 199 200
	struct drbd_interval *i = &req->i;

	drbd_remove_interval(root, i);

	/* Wake up any processes waiting for this request to complete.  */
	if (i->waiting)
201
		wake_up(&device->misc_wait);
202 203
}

P
Philipp Reisner 已提交
204 205 206 207 208 209
/* Helper for __req_mod().
 * Set m->bio to the master bio, if it is fit to be completed,
 * or leave it alone (it is initialized to NULL in __req_mod),
 * if it has already been completed, or cannot be completed yet.
 * If m->bio is set, the error status to be returned is placed in m->error.
 */
210
static
211
void drbd_req_complete(struct drbd_request *req, struct bio_and_error *m)
P
Philipp Reisner 已提交
212
{
213
	const unsigned s = req->rq_state;
214
	struct drbd_device *device = req->w.device;
215 216
	int rw;
	int error, ok;
P
Philipp Reisner 已提交
217 218 219 220 221 222 223 224 225 226

	/* we must not complete the master bio, while it is
	 *	still being processed by _drbd_send_zc_bio (drbd_send_dblock)
	 *	not yet acknowledged by the peer
	 *	not yet completed by the local io subsystem
	 * these flags may get cleared in any order by
	 *	the worker,
	 *	the receiver,
	 *	the bio_endio completion callbacks.
	 */
227 228 229 230
	if ((s & RQ_LOCAL_PENDING && !(s & RQ_LOCAL_ABORTED)) ||
	    (s & RQ_NET_QUEUED) || (s & RQ_NET_PENDING) ||
	    (s & RQ_COMPLETION_SUSP)) {
		dev_err(DEV, "drbd_req_complete: Logic BUG rq_state = 0x%x\n", s);
P
Philipp Reisner 已提交
231
		return;
232 233 234 235
	}

	if (!req->master_bio) {
		dev_err(DEV, "drbd_req_complete: Logic BUG, master_bio == NULL!\n");
P
Philipp Reisner 已提交
236
		return;
237
	}
P
Philipp Reisner 已提交
238

239
	rw = bio_rw(req->master_bio);
P
Philipp Reisner 已提交
240

241 242 243 244 245 246 247 248 249 250 251 252 253 254 255
	/*
	 * figure out whether to report success or failure.
	 *
	 * report success when at least one of the operations succeeded.
	 * or, to put the other way,
	 * only report failure, when both operations failed.
	 *
	 * what to do about the failures is handled elsewhere.
	 * what we need to do here is just: complete the master_bio.
	 *
	 * local completion error, if any, has been stored as ERR_PTR
	 * in private_bio within drbd_request_endio.
	 */
	ok = (s & RQ_LOCAL_OK) || (s & RQ_NET_OK);
	error = PTR_ERR(req->private_bio);
P
Philipp Reisner 已提交
256

257 258 259 260
	/* remove the request from the conflict detection
	 * respective block_id verification hash */
	if (!drbd_interval_empty(&req->i)) {
		struct rb_root *root;
P
Philipp Reisner 已提交
261 262

		if (rw == WRITE)
263
			root = &device->write_requests;
264
		else
265
			root = &device->read_requests;
266
		drbd_remove_request_interval(root, req);
267
	}
P
Philipp Reisner 已提交
268

269 270 271 272 273 274 275 276
	/* Before we can signal completion to the upper layers,
	 * we may need to close the current transfer log epoch.
	 * We are within the request lock, so we can simply compare
	 * the request epoch number with the current transfer log
	 * epoch number.  If they match, increase the current_tle_nr,
	 * and reset the transfer log epoch write_cnt.
	 */
	if (rw == WRITE &&
277 278
	    req->epoch == atomic_read(&device->connection->current_tle_nr))
		start_new_tl_epoch(device->connection);
P
Philipp Reisner 已提交
279

280
	/* Update disk stats */
281
	_drbd_end_io_acct(device, req);
P
Philipp Reisner 已提交
282

283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
	/* If READ failed,
	 * have it be pushed back to the retry work queue,
	 * so it will re-enter __drbd_make_request(),
	 * and be re-assigned to a suitable local or remote path,
	 * or failed if we do not have access to good data anymore.
	 *
	 * Unless it was failed early by __drbd_make_request(),
	 * because no path was available, in which case
	 * it was not even added to the transfer_log.
	 *
	 * READA may fail, and will not be retried.
	 *
	 * WRITE should have used all available paths already.
	 */
	if (!ok && rw == READ && !list_empty(&req->tl_requests))
		req->rq_state |= RQ_POSTPONED;
P
Philipp Reisner 已提交
299

300
	if (!(req->rq_state & RQ_POSTPONED)) {
P
Philipp Reisner 已提交
301 302 303 304 305 306
		m->error = ok ? 0 : (error ?: -EIO);
		m->bio = req->master_bio;
		req->master_bio = NULL;
	}
}

307
static int drbd_req_put_completion_ref(struct drbd_request *req, struct bio_and_error *m, int put)
308
{
309
	struct drbd_device *device = req->w.device;
310 311 312 313
	D_ASSERT(m || (req->rq_state & RQ_POSTPONED));

	if (!atomic_sub_and_test(put, &req->completion_ref))
		return 0;
314

315
	drbd_req_complete(req, m);
316 317 318 319 320 321

	if (req->rq_state & RQ_POSTPONED) {
		/* don't destroy the req object just yet,
		 * but queue it for retry */
		drbd_restart_request(req);
		return 0;
P
Philipp Reisner 已提交
322
	}
323

324
	return 1;
P
Philipp Reisner 已提交
325 326
}

327 328 329 330
/* I'd like this to be the only place that manipulates
 * req->completion_ref and req->kref. */
static void mod_rq_state(struct drbd_request *req, struct bio_and_error *m,
		int clear, int set)
331
{
332
	struct drbd_device *device = req->w.device;
333 334 335
	unsigned s = req->rq_state;
	int c_put = 0;
	int k_put = 0;
336

337
	if (drbd_suspended(device) && !((s | clear) & RQ_COMPLETION_SUSP))
338
		set |= RQ_COMPLETION_SUSP;
339

340
	/* apply */
P
Philipp Reisner 已提交
341

342 343
	req->rq_state &= ~clear;
	req->rq_state |= set;
P
Philipp Reisner 已提交
344

345 346 347
	/* no change? */
	if (req->rq_state == s)
		return;
P
Philipp Reisner 已提交
348

349 350 351 352 353 354
	/* intent: get references */

	if (!(s & RQ_LOCAL_PENDING) && (set & RQ_LOCAL_PENDING))
		atomic_inc(&req->completion_ref);

	if (!(s & RQ_NET_PENDING) && (set & RQ_NET_PENDING)) {
355
		inc_ap_pending(device);
356
		atomic_inc(&req->completion_ref);
P
Philipp Reisner 已提交
357 358
	}

359 360 361 362 363 364 365
	if (!(s & RQ_NET_QUEUED) && (set & RQ_NET_QUEUED))
		atomic_inc(&req->completion_ref);

	if (!(s & RQ_EXP_BARR_ACK) && (set & RQ_EXP_BARR_ACK))
		kref_get(&req->kref); /* wait for the DONE */

	if (!(s & RQ_NET_SENT) && (set & RQ_NET_SENT))
366
		atomic_add(req->i.size >> 9, &device->ap_in_flight);
367

368 369 370
	if (!(s & RQ_COMPLETION_SUSP) && (set & RQ_COMPLETION_SUSP))
		atomic_inc(&req->completion_ref);

371 372 373 374 375 376 377 378 379 380 381
	/* progress: put references */

	if ((s & RQ_COMPLETION_SUSP) && (clear & RQ_COMPLETION_SUSP))
		++c_put;

	if (!(s & RQ_LOCAL_ABORTED) && (set & RQ_LOCAL_ABORTED)) {
		D_ASSERT(req->rq_state & RQ_LOCAL_PENDING);
		/* local completion may still come in later,
		 * we need to keep the req object around. */
		kref_get(&req->kref);
		++c_put;
P
Philipp Reisner 已提交
382 383
	}

384 385 386 387 388 389
	if ((s & RQ_LOCAL_PENDING) && (clear & RQ_LOCAL_PENDING)) {
		if (req->rq_state & RQ_LOCAL_ABORTED)
			++k_put;
		else
			++c_put;
	}
P
Philipp Reisner 已提交
390

391
	if ((s & RQ_NET_PENDING) && (clear & RQ_NET_PENDING)) {
392
		dec_ap_pending(device);
393 394 395 396 397 398 399 400
		++c_put;
	}

	if ((s & RQ_NET_QUEUED) && (clear & RQ_NET_QUEUED))
		++c_put;

	if ((s & RQ_EXP_BARR_ACK) && !(s & RQ_NET_DONE) && (set & RQ_NET_DONE)) {
		if (req->rq_state & RQ_NET_SENT)
401
			atomic_sub(req->i.size >> 9, &device->ap_in_flight);
402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
		++k_put;
	}

	/* potentially complete and destroy */

	if (k_put || c_put) {
		/* Completion does it's own kref_put.  If we are going to
		 * kref_sub below, we need req to be still around then. */
		int at_least = k_put + !!c_put;
		int refcount = atomic_read(&req->kref.refcount);
		if (refcount < at_least)
			dev_err(DEV,
				"mod_rq_state: Logic BUG: %x -> %x: refcount = %d, should be >= %d\n",
				s, req->rq_state, refcount, at_least);
	}

	/* If we made progress, retry conflicting peer requests, if any. */
	if (req->i.waiting)
420
		wake_up(&device->misc_wait);
421 422 423 424 425

	if (c_put)
		k_put += drbd_req_put_completion_ref(req, m, c_put);
	if (k_put)
		kref_sub(&req->kref, k_put, drbd_req_destroy);
P
Philipp Reisner 已提交
426 427
}

428
static void drbd_report_io_error(struct drbd_device *device, struct drbd_request *req)
429 430 431
{
        char b[BDEVNAME_SIZE];

432
	if (!__ratelimit(&drbd_ratelimit_state))
433 434 435 436
		return;

	dev_warn(DEV, "local %s IO error sector %llu+%u on %s\n",
			(req->rq_state & RQ_WRITE) ? "WRITE" : "READ",
437 438
			(unsigned long long)req->i.sector,
			req->i.size >> 9,
439
			bdevname(device->ldev->backing_bdev, b));
440 441
}

P
Philipp Reisner 已提交
442 443 444 445 446 447 448 449 450 451 452 453
/* obviously this could be coded as many single functions
 * instead of one huge switch,
 * or by putting the code directly in the respective locations
 * (as it has been before).
 *
 * but having it this way
 *  enforces that it is all in this one place, where it is easier to audit,
 *  it makes it obvious that whatever "event" "happens" to a request should
 *  happen "atomically" within the req_lock,
 *  and it enforces that we have to think in a very structured manner
 *  about the "events" that may happen to a request during its life time ...
 */
454
int __req_mod(struct drbd_request *req, enum drbd_req_event what,
P
Philipp Reisner 已提交
455 456
		struct bio_and_error *m)
{
457
	struct drbd_device *device = req->w.device;
458
	struct net_conf *nc;
459
	int p, rv = 0;
460 461 462

	if (m)
		m->bio = NULL;
P
Philipp Reisner 已提交
463 464 465 466 467 468 469 470

	switch (what) {
	default:
		dev_err(DEV, "LOGIC BUG in %s:%u\n", __FILE__ , __LINE__);
		break;

	/* does not happen...
	 * initialization done in drbd_req_new
471
	case CREATED:
P
Philipp Reisner 已提交
472 473 474
		break;
		*/

475
	case TO_BE_SENT: /* via network */
476
		/* reached via __drbd_make_request
P
Philipp Reisner 已提交
477 478
		 * and from w_read_retry_remote */
		D_ASSERT(!(req->rq_state & RQ_NET_MASK));
479
		rcu_read_lock();
480
		nc = rcu_dereference(device->connection->net_conf);
481 482
		p = nc->wire_protocol;
		rcu_read_unlock();
483 484 485
		req->rq_state |=
			p == DRBD_PROT_C ? RQ_EXP_WRITE_ACK :
			p == DRBD_PROT_B ? RQ_EXP_RECEIVE_ACK : 0;
486
		mod_rq_state(req, m, 0, RQ_NET_PENDING);
P
Philipp Reisner 已提交
487 488
		break;

489
	case TO_BE_SUBMITTED: /* locally */
490
		/* reached via __drbd_make_request */
P
Philipp Reisner 已提交
491
		D_ASSERT(!(req->rq_state & RQ_LOCAL_MASK));
492
		mod_rq_state(req, m, 0, RQ_LOCAL_PENDING);
P
Philipp Reisner 已提交
493 494
		break;

495
	case COMPLETED_OK:
496
		if (req->rq_state & RQ_WRITE)
497
			device->writ_cnt += req->i.size >> 9;
P
Philipp Reisner 已提交
498
		else
499
			device->read_cnt += req->i.size >> 9;
P
Philipp Reisner 已提交
500

501 502
		mod_rq_state(req, m, RQ_LOCAL_PENDING,
				RQ_LOCAL_COMPLETED|RQ_LOCAL_OK);
P
Philipp Reisner 已提交
503 504
		break;

505
	case ABORT_DISK_IO:
506
		mod_rq_state(req, m, 0, RQ_LOCAL_ABORTED);
507 508
		break;

509
	case WRITE_COMPLETED_WITH_ERROR:
510 511
		drbd_report_io_error(device, req);
		__drbd_chk_io_error(device, DRBD_WRITE_ERROR);
512
		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
P
Philipp Reisner 已提交
513 514
		break;

515
	case READ_COMPLETED_WITH_ERROR:
516 517 518
		drbd_set_out_of_sync(device, req->i.sector, req->i.size);
		drbd_report_io_error(device, req);
		__drbd_chk_io_error(device, DRBD_READ_ERROR);
519 520 521 522
		/* fall through. */
	case READ_AHEAD_COMPLETED_WITH_ERROR:
		/* it is legal to fail READA, no __drbd_chk_io_error in that case. */
		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
523
		break;
P
Philipp Reisner 已提交
524

525
	case QUEUE_FOR_NET_READ:
P
Philipp Reisner 已提交
526 527 528 529
		/* READ or READA, and
		 * no local disk,
		 * or target area marked as invalid,
		 * or just got an io-error. */
530
		/* from __drbd_make_request
P
Philipp Reisner 已提交
531 532
		 * or from bio_endio during read io-error recovery */

533 534
		/* So we can verify the handle in the answer packet.
		 * Corresponding drbd_remove_request_interval is in
535
		 * drbd_req_complete() */
536
		D_ASSERT(drbd_interval_empty(&req->i));
537
		drbd_insert_interval(&device->read_requests, &req->i);
P
Philipp Reisner 已提交
538

539
		set_bit(UNPLUG_REMOTE, &device->flags);
P
Philipp Reisner 已提交
540 541

		D_ASSERT(req->rq_state & RQ_NET_PENDING);
542
		D_ASSERT((req->rq_state & RQ_LOCAL_MASK) == 0);
543
		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
544
		req->w.cb = w_send_read_req;
545
		drbd_queue_work(&device->connection->sender_work, &req->w);
P
Philipp Reisner 已提交
546 547
		break;

548
	case QUEUE_FOR_NET_WRITE:
P
Philipp Reisner 已提交
549
		/* assert something? */
550
		/* from __drbd_make_request only */
P
Philipp Reisner 已提交
551

552
		/* Corresponding drbd_remove_request_interval is in
553
		 * drbd_req_complete() */
554
		D_ASSERT(drbd_interval_empty(&req->i));
555
		drbd_insert_interval(&device->write_requests, &req->i);
P
Philipp Reisner 已提交
556 557 558 559 560 561 562 563 564 565

		/* NOTE
		 * In case the req ended up on the transfer log before being
		 * queued on the worker, it could lead to this request being
		 * missed during cleanup after connection loss.
		 * So we have to do both operations here,
		 * within the same lock that protects the transfer log.
		 *
		 * _req_add_to_epoch(req); this has to be after the
		 * _maybe_start_new_epoch(req); which happened in
566
		 * __drbd_make_request, because we now may set the bit
P
Philipp Reisner 已提交
567 568 569 570
		 * again ourselves to close the current epoch.
		 *
		 * Add req to the (now) current epoch (barrier). */

571 572 573
		/* otherwise we may lose an unplug, which may cause some remote
		 * io-scheduler timeout to expire, increasing maximum latency,
		 * hurting performance. */
574
		set_bit(UNPLUG_REMOTE, &device->flags);
P
Philipp Reisner 已提交
575 576 577

		/* queue work item to send data */
		D_ASSERT(req->rq_state & RQ_NET_PENDING);
578
		mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
P
Philipp Reisner 已提交
579
		req->w.cb =  w_send_dblock;
580
		drbd_queue_work(&device->connection->sender_work, &req->w);
P
Philipp Reisner 已提交
581 582

		/* close the epoch, in case it outgrew the limit */
583
		rcu_read_lock();
584
		nc = rcu_dereference(device->connection->net_conf);
585 586
		p = nc->max_epoch_size;
		rcu_read_unlock();
587 588
		if (device->connection->current_tle_writes >= p)
			start_new_tl_epoch(device->connection);
P
Philipp Reisner 已提交
589 590 591

		break;

592
	case QUEUE_FOR_SEND_OOS:
593
		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
594
		req->w.cb =  w_send_out_of_sync;
595
		drbd_queue_work(&device->connection->sender_work, &req->w);
596 597
		break;

598
	case READ_RETRY_REMOTE_CANCELED:
599 600
	case SEND_CANCELED:
	case SEND_FAILED:
P
Philipp Reisner 已提交
601 602
		/* real cleanup will be done from tl_clear.  just update flags
		 * so it is no longer marked as on the worker queue */
603
		mod_rq_state(req, m, RQ_NET_QUEUED, 0);
P
Philipp Reisner 已提交
604 605
		break;

606
	case HANDED_OVER_TO_NETWORK:
P
Philipp Reisner 已提交
607 608
		/* assert something? */
		if (bio_data_dir(req->master_bio) == WRITE &&
609
		    !(req->rq_state & (RQ_EXP_RECEIVE_ACK | RQ_EXP_WRITE_ACK))) {
P
Philipp Reisner 已提交
610 611
			/* this is what is dangerous about protocol A:
			 * pretend it was successfully written on the peer. */
612 613 614
			if (req->rq_state & RQ_NET_PENDING)
				mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
			/* else: neg-ack was faster... */
P
Philipp Reisner 已提交
615 616 617 618
			/* it is still not yet RQ_NET_DONE until the
			 * corresponding epoch barrier got acked as well,
			 * so we know what to dirty on connection loss */
		}
619
		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
620 621
		break;

622
	case OOS_HANDED_TO_NETWORK:
623 624
		/* Was not set PENDING, no longer QUEUED, so is now DONE
		 * as far as this connection is concerned. */
625
		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
P
Philipp Reisner 已提交
626 627
		break;

628
	case CONNECTION_LOST_WHILE_PENDING:
P
Philipp Reisner 已提交
629
		/* transfer log cleanup after connection loss */
630 631 632
		mod_rq_state(req, m,
				RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
				RQ_NET_DONE);
P
Philipp Reisner 已提交
633 634
		break;

635 636
	case CONFLICT_RESOLVED:
		/* for superseded conflicting writes of multiple primaries,
P
Philipp Reisner 已提交
637
		 * there is no need to keep anything in the tl, potential
638 639 640
		 * node crashes are covered by the activity log.
		 *
		 * If this request had been marked as RQ_POSTPONED before,
641
		 * it will actually not be completed, but "restarted",
642 643 644 645 646 647
		 * resubmitted from the retry worker context. */
		D_ASSERT(req->rq_state & RQ_NET_PENDING);
		D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
		break;

648
	case WRITE_ACKED_BY_PEER_AND_SIS:
649
		req->rq_state |= RQ_NET_SIS;
650
	case WRITE_ACKED_BY_PEER:
651
		D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
P
Philipp Reisner 已提交
652
		/* protocol C; successfully written on peer.
653
		 * Nothing more to do here.
P
Philipp Reisner 已提交
654
		 * We want to keep the tl in place for all protocols, to cater
655
		 * for volatile write-back caches on lower level devices. */
P
Philipp Reisner 已提交
656

657
		goto ack_common;
658
	case RECV_ACKED_BY_PEER:
659
		D_ASSERT(req->rq_state & RQ_EXP_RECEIVE_ACK);
P
Philipp Reisner 已提交
660
		/* protocol B; pretends to be successfully written on peer.
661
		 * see also notes above in HANDED_OVER_TO_NETWORK about
P
Philipp Reisner 已提交
662
		 * protocol != C */
663
	ack_common:
P
Philipp Reisner 已提交
664
		D_ASSERT(req->rq_state & RQ_NET_PENDING);
665
		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
P
Philipp Reisner 已提交
666 667
		break;

668
	case POSTPONE_WRITE:
669 670
		D_ASSERT(req->rq_state & RQ_EXP_WRITE_ACK);
		/* If this node has already detected the write conflict, the
671 672 673 674 675
		 * worker will be waiting on misc_wait.  Wake it up once this
		 * request has completed locally.
		 */
		D_ASSERT(req->rq_state & RQ_NET_PENDING);
		req->rq_state |= RQ_POSTPONED;
676
		if (req->i.waiting)
677
			wake_up(&device->misc_wait);
678 679 680
		/* Do not clear RQ_NET_PENDING. This request will make further
		 * progress via restart_conflicting_writes() or
		 * fail_postponed_requests(). Hopefully. */
681
		break;
P
Philipp Reisner 已提交
682

683
	case NEG_ACKED:
684
		mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
P
Philipp Reisner 已提交
685 686
		break;

687
	case FAIL_FROZEN_DISK_IO:
688 689
		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
			break;
690
		mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
691 692
		break;

693
	case RESTART_FROZEN_DISK_IO:
694 695 696
		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
			break;

697 698 699
		mod_rq_state(req, m,
				RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
				RQ_LOCAL_PENDING);
700 701 702 703 704

		rv = MR_READ;
		if (bio_data_dir(req->master_bio) == WRITE)
			rv = MR_WRITE;

705
		get_ldev(device); /* always succeeds in this call path */
706
		req->w.cb = w_restart_disk_io;
707
		drbd_queue_work(&device->connection->sender_work, &req->w);
708 709
		break;

710
	case RESEND:
711 712
		/* Simply complete (local only) READs. */
		if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
713
			mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
714 715 716
			break;
		}

717
		/* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
718 719
		   before the connection loss (B&C only); only P_BARRIER_ACK
		   (or the local completion?) was missing when we suspended.
720 721
		   Throwing them out of the TL here by pretending we got a BARRIER_ACK.
		   During connection handshake, we ensure that the peer was not rebooted. */
722
		if (!(req->rq_state & RQ_NET_OK)) {
723 724 725 726
			/* FIXME could this possibly be a req->w.cb == w_send_out_of_sync?
			 * in that case we must not set RQ_NET_PENDING. */

			mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
727
			if (req->w.cb) {
728
				drbd_queue_work(&device->connection->sender_work, &req->w);
729
				rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
730
			} /* else: FIXME can this happen? */
731 732
			break;
		}
733
		/* else, fall through to BARRIER_ACKED */
734

735
	case BARRIER_ACKED:
736
		/* barrier ack for READ requests does not make sense */
737 738 739
		if (!(req->rq_state & RQ_WRITE))
			break;

P
Philipp Reisner 已提交
740
		if (req->rq_state & RQ_NET_PENDING) {
741
			/* barrier came in before all requests were acked.
P
Philipp Reisner 已提交
742 743
			 * this is bad, because if the connection is lost now,
			 * we won't be able to clean them up... */
744
			dev_err(DEV, "FIXME (BARRIER_ACKED but pending)\n");
P
Philipp Reisner 已提交
745
		}
746 747 748 749 750 751
		/* Allowed to complete requests, even while suspended.
		 * As this is called for all requests within a matching epoch,
		 * we need to filter, and only set RQ_NET_DONE for those that
		 * have actually been on the wire. */
		mod_rq_state(req, m, RQ_COMPLETION_SUSP,
				(req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
P
Philipp Reisner 已提交
752 753
		break;

754
	case DATA_RECEIVED:
P
Philipp Reisner 已提交
755
		D_ASSERT(req->rq_state & RQ_NET_PENDING);
756
		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
P
Philipp Reisner 已提交
757
		break;
758 759

	case QUEUE_AS_DRBD_BARRIER:
760
		start_new_tl_epoch(device->connection);
761 762
		mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
		break;
P
Philipp Reisner 已提交
763
	};
764 765

	return rv;
P
Philipp Reisner 已提交
766 767 768 769 770 771 772 773 774
}

/* we may do a local read if:
 * - we are consistent (of course),
 * - or we are generally inconsistent,
 *   BUT we are still/already IN SYNC for this area.
 *   since size may be bigger than BM_BLOCK_SIZE,
 *   we may need to check several bits.
 */
775
static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
P
Philipp Reisner 已提交
776 777 778 779
{
	unsigned long sbnr, ebnr;
	sector_t esector, nr_sectors;

780
	if (device->state.disk == D_UP_TO_DATE)
781
		return true;
782
	if (device->state.disk != D_INCONSISTENT)
783
		return false;
P
Philipp Reisner 已提交
784
	esector = sector + (size >> 9) - 1;
785
	nr_sectors = drbd_get_capacity(device->this_bdev);
P
Philipp Reisner 已提交
786 787 788 789 790 791
	D_ASSERT(sector  < nr_sectors);
	D_ASSERT(esector < nr_sectors);

	sbnr = BM_SECT_TO_BIT(sector);
	ebnr = BM_SECT_TO_BIT(esector);

792
	return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
P
Philipp Reisner 已提交
793 794
}

795
static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
796
		enum drbd_read_balancing rbm)
797 798
{
	struct backing_dev_info *bdi;
799
	int stripe_shift;
800 801 802

	switch (rbm) {
	case RB_CONGESTED_REMOTE:
803
		bdi = &device->ldev->backing_bdev->bd_disk->queue->backing_dev_info;
804 805
		return bdi_read_congested(bdi);
	case RB_LEAST_PENDING:
806 807
		return atomic_read(&device->local_cnt) >
			atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
808 809 810 811 812 813 814 815
	case RB_32K_STRIPING:  /* stripe_shift = 15 */
	case RB_64K_STRIPING:
	case RB_128K_STRIPING:
	case RB_256K_STRIPING:
	case RB_512K_STRIPING:
	case RB_1M_STRIPING:   /* stripe_shift = 20 */
		stripe_shift = (rbm - RB_32K_STRIPING + 15);
		return (sector >> (stripe_shift - 9)) & 1;
816
	case RB_ROUND_ROBIN:
817
		return test_and_change_bit(READ_BALANCE_RR, &device->flags);
818 819 820 821 822 823 824 825
	case RB_PREFER_REMOTE:
		return true;
	case RB_PREFER_LOCAL:
	default:
		return false;
	}
}

826 827 828 829 830 831
/*
 * complete_conflicting_writes  -  wait for any conflicting write requests
 *
 * The write_requests tree contains all active write requests which we
 * currently know about.  Wait for any requests to complete which conflict with
 * the new one.
832 833
 *
 * Only way out: remove the conflicting intervals from the tree.
834
 */
835
static void complete_conflicting_writes(struct drbd_request *req)
836
{
837
	DEFINE_WAIT(wait);
838
	struct drbd_device *device = req->w.device;
839 840 841 842
	struct drbd_interval *i;
	sector_t sector = req->i.sector;
	int size = req->i.size;

843
	i = drbd_find_overlap(&device->write_requests, sector, size);
844 845
	if (!i)
		return;
846

847
	for (;;) {
848 849
		prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
		i = drbd_find_overlap(&device->write_requests, sector, size);
850
		if (!i)
851 852 853
			break;
		/* Indicate to wake up device->misc_wait on progress.  */
		i->waiting = true;
854
		spin_unlock_irq(&device->connection->req_lock);
855
		schedule();
856
		spin_lock_irq(&device->connection->req_lock);
857
	}
858
	finish_wait(&device->misc_wait, &wait);
P
Philipp Reisner 已提交
859 860
}

861
/* called within req_lock and rcu_read_lock() */
862
static void maybe_pull_ahead(struct drbd_device *device)
863
{
864
	struct drbd_connection *connection = device->connection;
865 866 867 868
	struct net_conf *nc;
	bool congested = false;
	enum drbd_on_congestion on_congestion;

869
	rcu_read_lock();
870
	nc = rcu_dereference(connection->net_conf);
871
	on_congestion = nc ? nc->on_congestion : OC_BLOCK;
872
	rcu_read_unlock();
873
	if (on_congestion == OC_BLOCK ||
874
	    connection->agreed_pro_version < 96)
875
		return;
876 877 878

	/* If I don't even have good local storage, we can not reasonably try
	 * to pull ahead of the peer. We also need the local reference to make
879
	 * sure device->act_log is there.
880
	 */
881
	if (!get_ldev_if_state(device, D_UP_TO_DATE))
882 883
		return;

884
	if (nc->cong_fill &&
885
	    atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
886
		dev_info(DEV, "Congestion-fill threshold reached\n");
887
		congested = true;
888 889
	}

890
	if (device->act_log->used >= nc->cong_extents) {
891
		dev_info(DEV, "Congestion-extents threshold reached\n");
892
		congested = true;
893 894 895
	}

	if (congested) {
896
		/* start a new epoch for non-mirrored writes */
897
		start_new_tl_epoch(device->connection);
898

899
		if (on_congestion == OC_PULL_AHEAD)
900
			_drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
901
		else  /*nc->on_congestion == OC_DISCONNECT */
902
			_drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
903
	}
904
	put_ldev(device);
905 906
}

907 908 909 910 911 912 913 914 915 916 917
/* If this returns false, and req->private_bio is still set,
 * this should be submitted locally.
 *
 * If it returns false, but req->private_bio is not set,
 * we do not have access to good data :(
 *
 * Otherwise, this destroys req->private_bio, if any,
 * and returns true.
 */
static bool do_remote_read(struct drbd_request *req)
{
918
	struct drbd_device *device = req->w.device;
919 920 921
	enum drbd_read_balancing rbm;

	if (req->private_bio) {
922
		if (!drbd_may_do_local_read(device,
923 924 925
					req->i.sector, req->i.size)) {
			bio_put(req->private_bio);
			req->private_bio = NULL;
926
			put_ldev(device);
927 928 929
		}
	}

930
	if (device->state.pdsk != D_UP_TO_DATE)
931 932
		return false;

933 934 935
	if (req->private_bio == NULL)
		return true;

936 937 938 939
	/* TODO: improve read balancing decisions, take into account drbd
	 * protocol, pending requests etc. */

	rcu_read_lock();
940
	rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
941 942 943 944 945
	rcu_read_unlock();

	if (rbm == RB_PREFER_LOCAL && req->private_bio)
		return false; /* submit locally */

946
	if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
947 948 949
		if (req->private_bio) {
			bio_put(req->private_bio);
			req->private_bio = NULL;
950
			put_ldev(device);
951 952 953 954 955 956 957 958 959 960 961 962
		}
		return true;
	}

	return false;
}

/* returns number of connections (== 1, for drbd 8.4)
 * expected to actually write this data,
 * which does NOT include those that we are L_AHEAD for. */
static int drbd_process_write_request(struct drbd_request *req)
{
963
	struct drbd_device *device = req->w.device;
964 965
	int remote, send_oos;

966 967
	remote = drbd_should_do_remote(device->state);
	send_oos = drbd_should_send_out_of_sync(device->state);
968

969 970 971 972 973 974 975 976 977
	/* Need to replicate writes.  Unless it is an empty flush,
	 * which is better mapped to a DRBD P_BARRIER packet,
	 * also for drbd wire protocol compatibility reasons.
	 * If this was a flush, just start a new epoch.
	 * Unless the current epoch was empty anyways, or we are not currently
	 * replicating, in which case there is no point. */
	if (unlikely(req->i.size == 0)) {
		/* The only size==0 bios we expect are empty flushes. */
		D_ASSERT(req->master_bio->bi_rw & REQ_FLUSH);
978
		if (remote)
979 980
			_req_mod(req, QUEUE_AS_DRBD_BARRIER);
		return remote;
981 982
	}

983 984 985 986 987 988 989 990
	if (!remote && !send_oos)
		return 0;

	D_ASSERT(!(remote && send_oos));

	if (remote) {
		_req_mod(req, TO_BE_SENT);
		_req_mod(req, QUEUE_FOR_NET_WRITE);
991
	} else if (drbd_set_out_of_sync(device, req->i.sector, req->i.size))
992 993 994 995 996 997 998 999
		_req_mod(req, QUEUE_FOR_SEND_OOS);

	return remote;
}

static void
drbd_submit_req_private_bio(struct drbd_request *req)
{
1000
	struct drbd_device *device = req->w.device;
1001 1002 1003
	struct bio *bio = req->private_bio;
	const int rw = bio_rw(bio);

1004
	bio->bi_bdev = device->ldev->backing_bdev;
1005 1006 1007 1008 1009 1010

	/* State may have changed since we grabbed our reference on the
	 * ->ldev member. Double check, and short-circuit to endio.
	 * In case the last activity log transaction failed to get on
	 * stable storage, and this is a WRITE, we may not even submit
	 * this bio. */
1011 1012
	if (get_ldev(device)) {
		if (drbd_insert_fault(device,
1013 1014 1015 1016 1017 1018
				      rw == WRITE ? DRBD_FAULT_DT_WR
				    : rw == READ  ? DRBD_FAULT_DT_RD
				    :               DRBD_FAULT_DT_RA))
			bio_endio(bio, -EIO);
		else
			generic_make_request(bio);
1019
		put_ldev(device);
1020 1021 1022 1023
	} else
		bio_endio(bio, -EIO);
}

1024
static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1025
{
1026 1027 1028 1029
	spin_lock(&device->submit.lock);
	list_add_tail(&req->tl_requests, &device->submit.writes);
	spin_unlock(&device->submit.lock);
	queue_work(device->submit.wq, &device->submit.worker);
1030 1031
}

1032 1033 1034 1035 1036
/* returns the new drbd_request pointer, if the caller is expected to
 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
 * request on the submitter thread.
 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
 */
1037
static struct drbd_request *
1038
drbd_request_prepare(struct drbd_device *device, struct bio *bio, unsigned long start_time)
P
Philipp Reisner 已提交
1039
{
1040
	const int rw = bio_data_dir(bio);
P
Philipp Reisner 已提交
1041 1042 1043
	struct drbd_request *req;

	/* allocate outside of all locks; */
1044
	req = drbd_req_new(device, bio);
P
Philipp Reisner 已提交
1045
	if (!req) {
1046
		dec_ap_bio(device);
P
Philipp Reisner 已提交
1047 1048 1049 1050
		/* only pass the error to the upper layers.
		 * if user cannot handle io errors, that's not our business. */
		dev_err(DEV, "could not kmalloc() req\n");
		bio_endio(bio, -ENOMEM);
1051
		return ERR_PTR(-ENOMEM);
P
Philipp Reisner 已提交
1052
	}
1053
	req->start_time = start_time;
P
Philipp Reisner 已提交
1054

1055
	if (!get_ldev(device)) {
1056
		bio_put(req->private_bio);
P
Philipp Reisner 已提交
1057 1058 1059
		req->private_bio = NULL;
	}

1060
	/* Update disk stats */
1061
	_drbd_start_io_acct(device, req);
1062

1063
	if (rw == WRITE && req->private_bio && req->i.size
1064 1065 1066
	&& !test_bit(AL_SUSPENDED, &device->flags)) {
		if (!drbd_al_begin_io_fastpath(device, &req->i)) {
			drbd_queue_write(device, req);
1067 1068
			return NULL;
		}
1069 1070
		req->rq_state |= RQ_IN_ACT_LOG;
	}
P
Philipp Reisner 已提交
1071

1072 1073 1074
	return req;
}

1075
static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1076 1077 1078 1079 1080
{
	const int rw = bio_rw(req->master_bio);
	struct bio_and_error m = { NULL, };
	bool no_remote = false;

1081
	spin_lock_irq(&device->connection->req_lock);
1082
	if (rw == WRITE) {
1083 1084 1085 1086
		/* This may temporarily give up the req_lock,
		 * but will re-aquire it before it returns here.
		 * Needs to be before the check on drbd_suspended() */
		complete_conflicting_writes(req);
1087 1088 1089 1090
		/* no more giving up req_lock from now on! */

		/* check for congestion, and potentially stop sending
		 * full data updates, but start sending "dirty bits" only. */
1091
		maybe_pull_ahead(device);
P
Philipp Reisner 已提交
1092 1093
	}

1094

1095
	if (drbd_suspended(device)) {
1096 1097 1098 1099 1100
		/* push back and retry: */
		req->rq_state |= RQ_POSTPONED;
		if (req->private_bio) {
			bio_put(req->private_bio);
			req->private_bio = NULL;
1101
			put_ldev(device);
P
Philipp Reisner 已提交
1102
		}
1103
		goto out;
P
Philipp Reisner 已提交
1104 1105
	}

1106 1107
	/* We fail READ/READA early, if we can not serve it.
	 * We must do this before req is registered on any lists.
1108
	 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1109 1110 1111
	if (rw != WRITE) {
		if (!do_remote_read(req) && !req->private_bio)
			goto nodata;
P
Philipp Reisner 已提交
1112 1113
	}

1114
	/* which transfer log epoch does this belong to? */
1115
	req->epoch = atomic_read(&device->connection->current_tle_nr);
1116

1117 1118
	/* no point in adding empty flushes to the transfer log,
	 * they are mapped to drbd barriers already. */
1119 1120
	if (likely(req->i.size!=0)) {
		if (rw == WRITE)
1121
			device->connection->current_tle_writes++;
1122

1123
		list_add_tail(&req->tl_requests, &device->connection->transfer_log);
P
Philipp Reisner 已提交
1124
	}
1125

1126 1127 1128 1129 1130 1131 1132 1133 1134
	if (rw == WRITE) {
		if (!drbd_process_write_request(req))
			no_remote = true;
	} else {
		/* We either have a private_bio, or we can read from remote.
		 * Otherwise we had done the goto nodata above. */
		if (req->private_bio == NULL) {
			_req_mod(req, TO_BE_SENT);
			_req_mod(req, QUEUE_FOR_NET_READ);
1135
		} else
1136
			no_remote = true;
P
Philipp Reisner 已提交
1137 1138
	}

1139 1140 1141 1142
	if (req->private_bio) {
		/* needs to be marked within the same spinlock */
		_req_mod(req, TO_BE_SUBMITTED);
		/* but we need to give up the spinlock to submit */
1143
		spin_unlock_irq(&device->connection->req_lock);
1144
		drbd_submit_req_private_bio(req);
1145
		spin_lock_irq(&device->connection->req_lock);
1146 1147 1148
	} else if (no_remote) {
nodata:
		if (__ratelimit(&drbd_ratelimit_state))
1149 1150
			dev_err(DEV, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
					(unsigned long long)req->i.sector, req->i.size >> 9);
1151
		/* A write may have been queued for send_oos, however.
1152
		 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
P
Philipp Reisner 已提交
1153 1154
	}

1155
out:
1156 1157
	if (drbd_req_put_completion_ref(req, &m, 1))
		kref_put(&req->kref, drbd_req_destroy);
1158
	spin_unlock_irq(&device->connection->req_lock);
P
Philipp Reisner 已提交
1159

1160
	if (m.bio)
1161
		complete_master_bio(device, &m);
1162 1163
}

1164
void __drbd_make_request(struct drbd_device *device, struct bio *bio, unsigned long start_time)
1165
{
1166
	struct drbd_request *req = drbd_request_prepare(device, bio, start_time);
1167 1168
	if (IS_ERR_OR_NULL(req))
		return;
1169
	drbd_send_and_submit(device, req);
P
Philipp Reisner 已提交
1170 1171
}

1172
static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1173
{
1174 1175 1176
	struct drbd_request *req, *tmp;
	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
		const int rw = bio_data_dir(req->master_bio);
1177

1178 1179
		if (rw == WRITE /* rw != WRITE should not even end up here! */
		&& req->private_bio && req->i.size
1180 1181
		&& !test_bit(AL_SUSPENDED, &device->flags)) {
			if (!drbd_al_begin_io_fastpath(device, &req->i))
1182 1183 1184 1185 1186 1187
				continue;

			req->rq_state |= RQ_IN_ACT_LOG;
		}

		list_del_init(&req->tl_requests);
1188
		drbd_send_and_submit(device, req);
1189 1190 1191
	}
}

1192
static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1193 1194 1195 1196 1197 1198 1199
					    struct list_head *incoming,
					    struct list_head *pending)
{
	struct drbd_request *req, *tmp;
	int wake = 0;
	int err;

1200
	spin_lock_irq(&device->al_lock);
1201
	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1202
		err = drbd_al_begin_io_nonblock(device, &req->i);
1203 1204 1205 1206 1207 1208 1209
		if (err == -EBUSY)
			wake = 1;
		if (err)
			continue;
		req->rq_state |= RQ_IN_ACT_LOG;
		list_move_tail(&req->tl_requests, pending);
	}
1210
	spin_unlock_irq(&device->al_lock);
1211
	if (wake)
1212
		wake_up(&device->al_wait);
1213 1214 1215

	return !list_empty(pending);
}
1216 1217 1218

void do_submit(struct work_struct *ws)
{
1219
	struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1220 1221
	LIST_HEAD(incoming);
	LIST_HEAD(pending);
1222 1223
	struct drbd_request *req, *tmp;

1224
	for (;;) {
1225 1226 1227
		spin_lock(&device->submit.lock);
		list_splice_tail_init(&device->submit.writes, &incoming);
		spin_unlock(&device->submit.lock);
1228

1229
		submit_fast_path(device, &incoming);
1230 1231 1232
		if (list_empty(&incoming))
			break;

1233
		wait_event(device->al_wait, prepare_al_transaction_nonblock(device, &incoming, &pending));
1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246
		/* Maybe more was queued, while we prepared the transaction?
		 * Try to stuff them into this transaction as well.
		 * Be strictly non-blocking here, no wait_event, we already
		 * have something to commit.
		 * Stop if we don't make any more progres.
		 */
		for (;;) {
			LIST_HEAD(more_pending);
			LIST_HEAD(more_incoming);
			bool made_progress;

			/* It is ok to look outside the lock,
			 * it's only an optimization anyways */
1247
			if (list_empty(&device->submit.writes))
1248 1249
				break;

1250 1251 1252
			spin_lock(&device->submit.lock);
			list_splice_tail_init(&device->submit.writes, &more_incoming);
			spin_unlock(&device->submit.lock);
1253 1254 1255 1256

			if (list_empty(&more_incoming))
				break;

1257
			made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending);
1258 1259 1260 1261 1262 1263 1264

			list_splice_tail_init(&more_pending, &pending);
			list_splice_tail_init(&more_incoming, &incoming);

			if (!made_progress)
				break;
		}
1265
		drbd_al_begin_io_commit(device, false);
1266 1267 1268

		list_for_each_entry_safe(req, tmp, &pending, tl_requests) {
			list_del_init(&req->tl_requests);
1269
			drbd_send_and_submit(device, req);
1270
		}
1271 1272 1273
	}
}

1274
void drbd_make_request(struct request_queue *q, struct bio *bio)
P
Philipp Reisner 已提交
1275
{
1276
	struct drbd_device *device = (struct drbd_device *) q->queuedata;
1277
	unsigned long start_time;
P
Philipp Reisner 已提交
1278

1279 1280
	start_time = jiffies;

P
Philipp Reisner 已提交
1281 1282 1283
	/*
	 * what we "blindly" assume:
	 */
1284
	D_ASSERT(IS_ALIGNED(bio->bi_iter.bi_size, 512));
P
Philipp Reisner 已提交
1285

1286 1287
	inc_ap_bio(device);
	__drbd_make_request(device, bio, start_time);
P
Philipp Reisner 已提交
1288 1289
}

1290 1291 1292
/* This is called by bio_add_page().
 *
 * q->max_hw_sectors and other global limits are already enforced there.
P
Philipp Reisner 已提交
1293
 *
1294 1295 1296 1297
 * We need to call down to our lower level device,
 * in case it has special restrictions.
 *
 * We also may need to enforce configured max-bio-bvecs limits.
P
Philipp Reisner 已提交
1298 1299
 *
 * As long as the BIO is empty we have to allow at least one bvec,
1300
 * regardless of size and offset, so no need to ask lower levels.
P
Philipp Reisner 已提交
1301 1302 1303
 */
int drbd_merge_bvec(struct request_queue *q, struct bvec_merge_data *bvm, struct bio_vec *bvec)
{
1304
	struct drbd_device *device = (struct drbd_device *) q->queuedata;
P
Philipp Reisner 已提交
1305
	unsigned int bio_size = bvm->bi_size;
1306 1307 1308
	int limit = DRBD_MAX_BIO_SIZE;
	int backing_limit;

1309
	if (bio_size && get_ldev(device)) {
1310
		unsigned int max_hw_sectors = queue_max_hw_sectors(q);
P
Philipp Reisner 已提交
1311
		struct request_queue * const b =
1312
			device->ldev->backing_bdev->bd_disk->queue;
1313
		if (b->merge_bvec_fn) {
P
Philipp Reisner 已提交
1314 1315 1316
			backing_limit = b->merge_bvec_fn(b, bvm, bvec);
			limit = min(limit, backing_limit);
		}
1317
		put_ldev(device);
1318 1319
		if ((limit >> 9) > max_hw_sectors)
			limit = max_hw_sectors << 9;
P
Philipp Reisner 已提交
1320 1321 1322
	}
	return limit;
}
1323

1324
static struct drbd_request *find_oldest_request(struct drbd_connection *connection)
1325 1326 1327 1328
{
	/* Walk the transfer log,
	 * and find the oldest not yet completed request */
	struct drbd_request *r;
1329
	list_for_each_entry(r, &connection->transfer_log, tl_requests) {
1330
		if (atomic_read(&r->completion_ref))
1331 1332 1333 1334 1335
			return r;
	}
	return NULL;
}

1336 1337
void request_timer_fn(unsigned long data)
{
1338
	struct drbd_device *device = (struct drbd_device *) data;
1339
	struct drbd_connection *connection = device->connection;
1340
	struct drbd_request *req; /* oldest request */
1341
	struct net_conf *nc;
1342
	unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1343
	unsigned long now;
1344

1345
	rcu_read_lock();
1346
	nc = rcu_dereference(connection->net_conf);
1347
	if (nc && device->state.conn >= C_WF_REPORT_PARAMS)
1348
		ent = nc->timeout * HZ/10 * nc->ko_count;
1349

1350 1351 1352
	if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
		dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
		put_ldev(device);
1353
	}
1354
	rcu_read_unlock();
1355

1356 1357
	et = min_not_zero(dt, ent);

1358
	if (!et)
1359 1360
		return; /* Recurring timer stopped */

1361 1362
	now = jiffies;

1363 1364
	spin_lock_irq(&connection->req_lock);
	req = find_oldest_request(connection);
1365
	if (!req) {
1366
		spin_unlock_irq(&connection->req_lock);
1367
		mod_timer(&device->request_timer, now + et);
1368 1369 1370
		return;
	}

1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388
	/* The request is considered timed out, if
	 * - we have some effective timeout from the configuration,
	 *   with above state restrictions applied,
	 * - the oldest request is waiting for a response from the network
	 *   resp. the local disk,
	 * - the oldest request is in fact older than the effective timeout,
	 * - the connection was established (resp. disk was attached)
	 *   for longer than the timeout already.
	 * Note that for 32bit jiffies and very stable connections/disks,
	 * we may have a wrap around, which is catched by
	 *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
	 *
	 * Side effect: once per 32bit wrap-around interval, which means every
	 * ~198 days with 250 HZ, we have a window where the timeout would need
	 * to expire twice (worst case) to become effective. Good enough.
	 */
	if (ent && req->rq_state & RQ_NET_PENDING &&
		 time_after(now, req->start_time + ent) &&
1389
		!time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent)) {
1390
		dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
1391
		_drbd_set_state(_NS(device, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
1392
	}
1393
	if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.device == device &&
1394
		 time_after(now, req->start_time + dt) &&
1395
		!time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1396
		dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
1397
		__drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1398
	}
1399
	nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
1400
	spin_unlock_irq(&connection->req_lock);
1401
	mod_timer(&device->request_timer, nt);
1402
}