drbd_req.c

   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
  22static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size);
  23
  24static 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
  58static 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
  71void 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
 159static 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 */
 165void 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
 176void 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 */
 192static
 193void 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 */
 292static 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
 320static 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
 329static 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
 350static 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
 359static 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
 380static 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
 389static 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. */
 412static 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
 520static 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) */
 538static 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 */
 560int __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(first_peer_device(device),
 625				req->i.sector, req->i.size);
 626		drbd_report_io_error(device, req);
 627		__drbd_chk_io_error(device, DRBD_READ_ERROR);
 628		fallthrough;
 629	case READ_AHEAD_COMPLETED_WITH_ERROR:
 630		/* it is legal to fail read-ahead, no __drbd_chk_io_error in that case. */
 631		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 632		break;
 633
 634	case DISCARD_COMPLETED_NOTSUPP:
 635	case DISCARD_COMPLETED_WITH_ERROR:
 636		/* I'd rather not detach from local disk just because it
 637		 * failed a REQ_OP_DISCARD. */
 638		mod_rq_state(req, m, RQ_LOCAL_PENDING, RQ_LOCAL_COMPLETED);
 639		break;
 640
 641	case QUEUE_FOR_NET_READ:
 642		/* READ, and
 643		 * no local disk,
 644		 * or target area marked as invalid,
 645		 * or just got an io-error. */
 646		/* from __drbd_make_request
 647		 * or from bio_endio during read io-error recovery */
 648
 649		/* So we can verify the handle in the answer packet.
 650		 * Corresponding drbd_remove_request_interval is in
 651		 * drbd_req_complete() */
 652		D_ASSERT(device, drbd_interval_empty(&req->i));
 653		drbd_insert_interval(&device->read_requests, &req->i);
 654
 655		set_bit(UNPLUG_REMOTE, &device->flags);
 656
 657		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 658		D_ASSERT(device, (req->rq_state & RQ_LOCAL_MASK) == 0);
 659		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 660		req->w.cb = w_send_read_req;
 661		drbd_queue_work(&connection->sender_work,
 662				&req->w);
 663		break;
 664
 665	case QUEUE_FOR_NET_WRITE:
 666		/* assert something? */
 667		/* from __drbd_make_request only */
 668
 669		/* Corresponding drbd_remove_request_interval is in
 670		 * drbd_req_complete() */
 671		D_ASSERT(device, drbd_interval_empty(&req->i));
 672		drbd_insert_interval(&device->write_requests, &req->i);
 673
 674		/* NOTE
 675		 * In case the req ended up on the transfer log before being
 676		 * queued on the worker, it could lead to this request being
 677		 * missed during cleanup after connection loss.
 678		 * So we have to do both operations here,
 679		 * within the same lock that protects the transfer log.
 680		 *
 681		 * _req_add_to_epoch(req); this has to be after the
 682		 * _maybe_start_new_epoch(req); which happened in
 683		 * __drbd_make_request, because we now may set the bit
 684		 * again ourselves to close the current epoch.
 685		 *
 686		 * Add req to the (now) current epoch (barrier). */
 687
 688		/* otherwise we may lose an unplug, which may cause some remote
 689		 * io-scheduler timeout to expire, increasing maximum latency,
 690		 * hurting performance. */
 691		set_bit(UNPLUG_REMOTE, &device->flags);
 692
 693		/* queue work item to send data */
 694		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 695		mod_rq_state(req, m, 0, RQ_NET_QUEUED|RQ_EXP_BARR_ACK);
 696		req->w.cb =  w_send_dblock;
 697		drbd_queue_work(&connection->sender_work,
 698				&req->w);
 699
 700		/* close the epoch, in case it outgrew the limit */
 701		rcu_read_lock();
 702		nc = rcu_dereference(connection->net_conf);
 703		p = nc->max_epoch_size;
 704		rcu_read_unlock();
 705		if (connection->current_tle_writes >= p)
 706			start_new_tl_epoch(connection);
 707
 708		break;
 709
 710	case QUEUE_FOR_SEND_OOS:
 711		mod_rq_state(req, m, 0, RQ_NET_QUEUED);
 712		req->w.cb =  w_send_out_of_sync;
 713		drbd_queue_work(&connection->sender_work,
 714				&req->w);
 715		break;
 716
 717	case READ_RETRY_REMOTE_CANCELED:
 718	case SEND_CANCELED:
 719	case SEND_FAILED:
 720		/* real cleanup will be done from tl_clear.  just update flags
 721		 * so it is no longer marked as on the worker queue */
 722		mod_rq_state(req, m, RQ_NET_QUEUED, 0);
 723		break;
 724
 725	case HANDED_OVER_TO_NETWORK:
 726		/* assert something? */
 727		if (is_pending_write_protocol_A(req))
 728			/* this is what is dangerous about protocol A:
 729			 * pretend it was successfully written on the peer. */
 730			mod_rq_state(req, m, RQ_NET_QUEUED|RQ_NET_PENDING,
 731						RQ_NET_SENT|RQ_NET_OK);
 732		else
 733			mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_SENT);
 734		/* It is still not yet RQ_NET_DONE until the
 735		 * corresponding epoch barrier got acked as well,
 736		 * so we know what to dirty on connection loss. */
 737		break;
 738
 739	case OOS_HANDED_TO_NETWORK:
 740		/* Was not set PENDING, no longer QUEUED, so is now DONE
 741		 * as far as this connection is concerned. */
 742		mod_rq_state(req, m, RQ_NET_QUEUED, RQ_NET_DONE);
 743		break;
 744
 745	case CONNECTION_LOST_WHILE_PENDING:
 746		/* transfer log cleanup after connection loss */
 747		mod_rq_state(req, m,
 748				RQ_NET_OK|RQ_NET_PENDING|RQ_COMPLETION_SUSP,
 749				RQ_NET_DONE);
 750		break;
 751
 752	case CONFLICT_RESOLVED:
 753		/* for superseded conflicting writes of multiple primaries,
 754		 * there is no need to keep anything in the tl, potential
 755		 * node crashes are covered by the activity log.
 756		 *
 757		 * If this request had been marked as RQ_POSTPONED before,
 758		 * it will actually not be completed, but "restarted",
 759		 * resubmitted from the retry worker context. */
 760		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 761		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 762		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_DONE|RQ_NET_OK);
 763		break;
 764
 765	case WRITE_ACKED_BY_PEER_AND_SIS:
 766		req->rq_state |= RQ_NET_SIS;
 767		fallthrough;
 768	case WRITE_ACKED_BY_PEER:
 769		/* Normal operation protocol C: successfully written on peer.
 770		 * During resync, even in protocol != C,
 771		 * we requested an explicit write ack anyways.
 772		 * Which means we cannot even assert anything here.
 773		 * Nothing more to do here.
 774		 * We want to keep the tl in place for all protocols, to cater
 775		 * for volatile write-back caches on lower level devices. */
 776		goto ack_common;
 777	case RECV_ACKED_BY_PEER:
 778		D_ASSERT(device, req->rq_state & RQ_EXP_RECEIVE_ACK);
 779		/* protocol B; pretends to be successfully written on peer.
 780		 * see also notes above in HANDED_OVER_TO_NETWORK about
 781		 * protocol != C */
 782	ack_common:
 783		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK);
 784		break;
 785
 786	case POSTPONE_WRITE:
 787		D_ASSERT(device, req->rq_state & RQ_EXP_WRITE_ACK);
 788		/* If this node has already detected the write conflict, the
 789		 * worker will be waiting on misc_wait.  Wake it up once this
 790		 * request has completed locally.
 791		 */
 792		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 793		req->rq_state |= RQ_POSTPONED;
 794		if (req->i.waiting)
 795			wake_up(&device->misc_wait);
 796		/* Do not clear RQ_NET_PENDING. This request will make further
 797		 * progress via restart_conflicting_writes() or
 798		 * fail_postponed_requests(). Hopefully. */
 799		break;
 800
 801	case NEG_ACKED:
 802		mod_rq_state(req, m, RQ_NET_OK|RQ_NET_PENDING, 0);
 803		break;
 804
 805	case FAIL_FROZEN_DISK_IO:
 806		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 807			break;
 808		mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 809		break;
 810
 811	case RESTART_FROZEN_DISK_IO:
 812		if (!(req->rq_state & RQ_LOCAL_COMPLETED))
 813			break;
 814
 815		mod_rq_state(req, m,
 816				RQ_COMPLETION_SUSP|RQ_LOCAL_COMPLETED,
 817				RQ_LOCAL_PENDING);
 818
 819		rv = MR_READ;
 820		if (bio_data_dir(req->master_bio) == WRITE)
 821			rv = MR_WRITE;
 822
 823		get_ldev(device); /* always succeeds in this call path */
 824		req->w.cb = w_restart_disk_io;
 825		drbd_queue_work(&connection->sender_work,
 826				&req->w);
 827		break;
 828
 829	case RESEND:
 830		/* Simply complete (local only) READs. */
 831		if (!(req->rq_state & RQ_WRITE) && !req->w.cb) {
 832			mod_rq_state(req, m, RQ_COMPLETION_SUSP, 0);
 833			break;
 834		}
 835
 836		/* If RQ_NET_OK is already set, we got a P_WRITE_ACK or P_RECV_ACK
 837		   before the connection loss (B&C only); only P_BARRIER_ACK
 838		   (or the local completion?) was missing when we suspended.
 839		   Throwing them out of the TL here by pretending we got a BARRIER_ACK.
 840		   During connection handshake, we ensure that the peer was not rebooted. */
 841		if (!(req->rq_state & RQ_NET_OK)) {
 842			/* FIXME could this possibly be a req->dw.cb == w_send_out_of_sync?
 843			 * in that case we must not set RQ_NET_PENDING. */
 844
 845			mod_rq_state(req, m, RQ_COMPLETION_SUSP, RQ_NET_QUEUED|RQ_NET_PENDING);
 846			if (req->w.cb) {
 847				/* w.cb expected to be w_send_dblock, or w_send_read_req */
 848				drbd_queue_work(&connection->sender_work,
 849						&req->w);
 850				rv = req->rq_state & RQ_WRITE ? MR_WRITE : MR_READ;
 851			} /* else: FIXME can this happen? */
 852			break;
 853		}
 854		fallthrough;	/* to BARRIER_ACKED */
 855
 856	case BARRIER_ACKED:
 857		/* barrier ack for READ requests does not make sense */
 858		if (!(req->rq_state & RQ_WRITE))
 859			break;
 860
 861		if (req->rq_state & RQ_NET_PENDING) {
 862			/* barrier came in before all requests were acked.
 863			 * this is bad, because if the connection is lost now,
 864			 * we won't be able to clean them up... */
 865			drbd_err(device, "FIXME (BARRIER_ACKED but pending)\n");
 866		}
 867		/* Allowed to complete requests, even while suspended.
 868		 * As this is called for all requests within a matching epoch,
 869		 * we need to filter, and only set RQ_NET_DONE for those that
 870		 * have actually been on the wire. */
 871		mod_rq_state(req, m, RQ_COMPLETION_SUSP,
 872				(req->rq_state & RQ_NET_MASK) ? RQ_NET_DONE : 0);
 873		break;
 874
 875	case DATA_RECEIVED:
 876		D_ASSERT(device, req->rq_state & RQ_NET_PENDING);
 877		mod_rq_state(req, m, RQ_NET_PENDING, RQ_NET_OK|RQ_NET_DONE);
 878		break;
 879
 880	case QUEUE_AS_DRBD_BARRIER:
 881		start_new_tl_epoch(connection);
 882		mod_rq_state(req, m, 0, RQ_NET_OK|RQ_NET_DONE);
 883		break;
 884	}
 885
 886	return rv;
 887}
 888
 889/* we may do a local read if:
 890 * - we are consistent (of course),
 891 * - or we are generally inconsistent,
 892 *   BUT we are still/already IN SYNC for this area.
 893 *   since size may be bigger than BM_BLOCK_SIZE,
 894 *   we may need to check several bits.
 895 */
 896static bool drbd_may_do_local_read(struct drbd_device *device, sector_t sector, int size)
 897{
 898	unsigned long sbnr, ebnr;
 899	sector_t esector, nr_sectors;
 900
 901	if (device->state.disk == D_UP_TO_DATE)
 902		return true;
 903	if (device->state.disk != D_INCONSISTENT)
 904		return false;
 905	esector = sector + (size >> 9) - 1;
 906	nr_sectors = get_capacity(device->vdisk);
 907	D_ASSERT(device, sector  < nr_sectors);
 908	D_ASSERT(device, esector < nr_sectors);
 909
 910	sbnr = BM_SECT_TO_BIT(sector);
 911	ebnr = BM_SECT_TO_BIT(esector);
 912
 913	return drbd_bm_count_bits(device, sbnr, ebnr) == 0;
 914}
 915
 916static bool remote_due_to_read_balancing(struct drbd_device *device, sector_t sector,
 917		enum drbd_read_balancing rbm)
 918{
 919	int stripe_shift;
 920
 921	switch (rbm) {
 922	case RB_CONGESTED_REMOTE:
 923		return false;
 924	case RB_LEAST_PENDING:
 925		return atomic_read(&device->local_cnt) >
 926			atomic_read(&device->ap_pending_cnt) + atomic_read(&device->rs_pending_cnt);
 927	case RB_32K_STRIPING:  /* stripe_shift = 15 */
 928	case RB_64K_STRIPING:
 929	case RB_128K_STRIPING:
 930	case RB_256K_STRIPING:
 931	case RB_512K_STRIPING:
 932	case RB_1M_STRIPING:   /* stripe_shift = 20 */
 933		stripe_shift = (rbm - RB_32K_STRIPING + 15);
 934		return (sector >> (stripe_shift - 9)) & 1;
 935	case RB_ROUND_ROBIN:
 936		return test_and_change_bit(READ_BALANCE_RR, &device->flags);
 937	case RB_PREFER_REMOTE:
 938		return true;
 939	case RB_PREFER_LOCAL:
 940	default:
 941		return false;
 942	}
 943}
 944
 945/*
 946 * complete_conflicting_writes  -  wait for any conflicting write requests
 947 *
 948 * The write_requests tree contains all active write requests which we
 949 * currently know about.  Wait for any requests to complete which conflict with
 950 * the new one.
 951 *
 952 * Only way out: remove the conflicting intervals from the tree.
 953 */
 954static void complete_conflicting_writes(struct drbd_request *req)
 955{
 956	DEFINE_WAIT(wait);
 957	struct drbd_device *device = req->device;
 958	struct drbd_interval *i;
 959	sector_t sector = req->i.sector;
 960	int size = req->i.size;
 961
 962	for (;;) {
 963		drbd_for_each_overlap(i, &device->write_requests, sector, size) {
 964			/* Ignore, if already completed to upper layers. */
 965			if (i->completed)
 966				continue;
 967			/* Handle the first found overlap.  After the schedule
 968			 * we have to restart the tree walk. */
 969			break;
 970		}
 971		if (!i)	/* if any */
 972			break;
 973
 974		/* Indicate to wake up device->misc_wait on progress.  */
 975		prepare_to_wait(&device->misc_wait, &wait, TASK_UNINTERRUPTIBLE);
 976		i->waiting = true;
 977		spin_unlock_irq(&device->resource->req_lock);
 978		schedule();
 979		spin_lock_irq(&device->resource->req_lock);
 980	}
 981	finish_wait(&device->misc_wait, &wait);
 982}
 983
 984/* called within req_lock */
 985static void maybe_pull_ahead(struct drbd_device *device)
 986{
 987	struct drbd_connection *connection = first_peer_device(device)->connection;
 988	struct net_conf *nc;
 989	bool congested = false;
 990	enum drbd_on_congestion on_congestion;
 991
 992	rcu_read_lock();
 993	nc = rcu_dereference(connection->net_conf);
 994	on_congestion = nc ? nc->on_congestion : OC_BLOCK;
 995	rcu_read_unlock();
 996	if (on_congestion == OC_BLOCK ||
 997	    connection->agreed_pro_version < 96)
 998		return;
 999
1000	if (on_congestion == OC_PULL_AHEAD && device->state.conn == C_AHEAD)
1001		return; /* nothing to do ... */
1002
1003	/* If I don't even have good local storage, we can not reasonably try
1004	 * to pull ahead of the peer. We also need the local reference to make
1005	 * sure device->act_log is there.
1006	 */
1007	if (!get_ldev_if_state(device, D_UP_TO_DATE))
1008		return;
1009
1010	if (nc->cong_fill &&
1011	    atomic_read(&device->ap_in_flight) >= nc->cong_fill) {
1012		drbd_info(device, "Congestion-fill threshold reached\n");
1013		congested = true;
1014	}
1015
1016	if (device->act_log->used >= nc->cong_extents) {
1017		drbd_info(device, "Congestion-extents threshold reached\n");
1018		congested = true;
1019	}
1020
1021	if (congested) {
1022		/* start a new epoch for non-mirrored writes */
1023		start_new_tl_epoch(first_peer_device(device)->connection);
1024
1025		if (on_congestion == OC_PULL_AHEAD)
1026			_drbd_set_state(_NS(device, conn, C_AHEAD), 0, NULL);
1027		else  /*nc->on_congestion == OC_DISCONNECT */
1028			_drbd_set_state(_NS(device, conn, C_DISCONNECTING), 0, NULL);
1029	}
1030	put_ldev(device);
1031}
1032
1033/* If this returns false, and req->private_bio is still set,
1034 * this should be submitted locally.
1035 *
1036 * If it returns false, but req->private_bio is not set,
1037 * we do not have access to good data :(
1038 *
1039 * Otherwise, this destroys req->private_bio, if any,
1040 * and returns true.
1041 */
1042static bool do_remote_read(struct drbd_request *req)
1043{
1044	struct drbd_device *device = req->device;
1045	enum drbd_read_balancing rbm;
1046
1047	if (req->private_bio) {
1048		if (!drbd_may_do_local_read(device,
1049					req->i.sector, req->i.size)) {
1050			bio_put(req->private_bio);
1051			req->private_bio = NULL;
1052			put_ldev(device);
1053		}
1054	}
1055
1056	if (device->state.pdsk != D_UP_TO_DATE)
1057		return false;
1058
1059	if (req->private_bio == NULL)
1060		return true;
1061
1062	/* TODO: improve read balancing decisions, take into account drbd
1063	 * protocol, pending requests etc. */
1064
1065	rcu_read_lock();
1066	rbm = rcu_dereference(device->ldev->disk_conf)->read_balancing;
1067	rcu_read_unlock();
1068
1069	if (rbm == RB_PREFER_LOCAL && req->private_bio)
1070		return false; /* submit locally */
1071
1072	if (remote_due_to_read_balancing(device, req->i.sector, rbm)) {
1073		if (req->private_bio) {
1074			bio_put(req->private_bio);
1075			req->private_bio = NULL;
1076			put_ldev(device);
1077		}
1078		return true;
1079	}
1080
1081	return false;
1082}
1083
1084bool drbd_should_do_remote(union drbd_dev_state s)
1085{
1086	return s.pdsk == D_UP_TO_DATE ||
1087		(s.pdsk >= D_INCONSISTENT &&
1088		 s.conn >= C_WF_BITMAP_T &&
1089		 s.conn < C_AHEAD);
1090	/* Before proto 96 that was >= CONNECTED instead of >= C_WF_BITMAP_T.
1091	   That is equivalent since before 96 IO was frozen in the C_WF_BITMAP*
1092	   states. */
1093}
1094
1095static bool drbd_should_send_out_of_sync(union drbd_dev_state s)
1096{
1097	return s.conn == C_AHEAD || s.conn == C_WF_BITMAP_S;
1098	/* pdsk = D_INCONSISTENT as a consequence. Protocol 96 check not necessary
1099	   since we enter state C_AHEAD only if proto >= 96 */
1100}
1101
1102/* returns number of connections (== 1, for drbd 8.4)
1103 * expected to actually write this data,
1104 * which does NOT include those that we are L_AHEAD for. */
1105static int drbd_process_write_request(struct drbd_request *req)
1106{
1107	struct drbd_device *device = req->device;
1108	struct drbd_peer_device *peer_device = first_peer_device(device);
1109	int remote, send_oos;
1110
1111	remote = drbd_should_do_remote(device->state);
1112	send_oos = drbd_should_send_out_of_sync(device->state);
1113
1114	/* Need to replicate writes.  Unless it is an empty flush,
1115	 * which is better mapped to a DRBD P_BARRIER packet,
1116	 * also for drbd wire protocol compatibility reasons.
1117	 * If this was a flush, just start a new epoch.
1118	 * Unless the current epoch was empty anyways, or we are not currently
1119	 * replicating, in which case there is no point. */
1120	if (unlikely(req->i.size == 0)) {
1121		/* The only size==0 bios we expect are empty flushes. */
1122		D_ASSERT(device, req->master_bio->bi_opf & REQ_PREFLUSH);
1123		if (remote)
1124			_req_mod(req, QUEUE_AS_DRBD_BARRIER, peer_device);
1125		return remote;
1126	}
1127
1128	if (!remote && !send_oos)
1129		return 0;
1130
1131	D_ASSERT(device, !(remote && send_oos));
1132
1133	if (remote) {
1134		_req_mod(req, TO_BE_SENT, peer_device);
1135		_req_mod(req, QUEUE_FOR_NET_WRITE, peer_device);
1136	} else if (drbd_set_out_of_sync(peer_device, req->i.sector, req->i.size))
1137		_req_mod(req, QUEUE_FOR_SEND_OOS, peer_device);
1138
1139	return remote;
1140}
1141
1142static void drbd_process_discard_or_zeroes_req(struct drbd_request *req, int flags)
1143{
1144	int err = drbd_issue_discard_or_zero_out(req->device,
1145				req->i.sector, req->i.size >> 9, flags);
1146	if (err)
1147		req->private_bio->bi_status = BLK_STS_IOERR;
1148	bio_endio(req->private_bio);
1149}
1150
1151static void
1152drbd_submit_req_private_bio(struct drbd_request *req)
1153{
1154	struct drbd_device *device = req->device;
1155	struct bio *bio = req->private_bio;
1156	unsigned int type;
1157
1158	if (bio_op(bio) != REQ_OP_READ)
1159		type = DRBD_FAULT_DT_WR;
1160	else if (bio->bi_opf & REQ_RAHEAD)
1161		type = DRBD_FAULT_DT_RA;
1162	else
1163		type = DRBD_FAULT_DT_RD;
1164
1165	/* State may have changed since we grabbed our reference on the
1166	 * ->ldev member. Double check, and short-circuit to endio.
1167	 * In case the last activity log transaction failed to get on
1168	 * stable storage, and this is a WRITE, we may not even submit
1169	 * this bio. */
1170	if (get_ldev(device)) {
1171		if (drbd_insert_fault(device, type))
1172			bio_io_error(bio);
1173		else if (bio_op(bio) == REQ_OP_WRITE_ZEROES)
1174			drbd_process_discard_or_zeroes_req(req, EE_ZEROOUT |
1175			    ((bio->bi_opf & REQ_NOUNMAP) ? 0 : EE_TRIM));
1176		else if (bio_op(bio) == REQ_OP_DISCARD)
1177			drbd_process_discard_or_zeroes_req(req, EE_TRIM);
1178		else
1179			submit_bio_noacct(bio);
1180		put_ldev(device);
1181	} else
1182		bio_io_error(bio);
1183}
1184
1185static void drbd_queue_write(struct drbd_device *device, struct drbd_request *req)
1186{
1187	spin_lock_irq(&device->resource->req_lock);
1188	list_add_tail(&req->tl_requests, &device->submit.writes);
1189	list_add_tail(&req->req_pending_master_completion,
1190			&device->pending_master_completion[1 /* WRITE */]);
1191	spin_unlock_irq(&device->resource->req_lock);
1192	queue_work(device->submit.wq, &device->submit.worker);
1193	/* do_submit() may sleep internally on al_wait, too */
1194	wake_up(&device->al_wait);
1195}
1196
1197/* returns the new drbd_request pointer, if the caller is expected to
1198 * drbd_send_and_submit() it (to save latency), or NULL if we queued the
1199 * request on the submitter thread.
1200 * Returns ERR_PTR(-ENOMEM) if we cannot allocate a drbd_request.
1201 */
1202static struct drbd_request *
1203drbd_request_prepare(struct drbd_device *device, struct bio *bio)
1204{
1205	const int rw = bio_data_dir(bio);
1206	struct drbd_request *req;
1207
1208	/* allocate outside of all locks; */
1209	req = drbd_req_new(device, bio);
1210	if (!req) {
1211		dec_ap_bio(device);
1212		/* only pass the error to the upper layers.
1213		 * if user cannot handle io errors, that's not our business. */
1214		drbd_err(device, "could not kmalloc() req\n");
1215		bio->bi_status = BLK_STS_RESOURCE;
1216		bio_endio(bio);
1217		return ERR_PTR(-ENOMEM);
1218	}
1219
1220	/* Update disk stats */
1221	req->start_jif = bio_start_io_acct(req->master_bio);
1222
1223	if (get_ldev(device)) {
1224		req->private_bio = bio_alloc_clone(device->ldev->backing_bdev,
1225						   bio, GFP_NOIO,
1226						   &drbd_io_bio_set);
1227		req->private_bio->bi_private = req;
1228		req->private_bio->bi_end_io = drbd_request_endio;
1229	}
1230
1231	/* process discards always from our submitter thread */
1232	if (bio_op(bio) == REQ_OP_WRITE_ZEROES ||
1233	    bio_op(bio) == REQ_OP_DISCARD)
1234		goto queue_for_submitter_thread;
1235
1236	if (rw == WRITE && req->private_bio && req->i.size
1237	&& !test_bit(AL_SUSPENDED, &device->flags)) {
1238		if (!drbd_al_begin_io_fastpath(device, &req->i))
1239			goto queue_for_submitter_thread;
1240		req->rq_state |= RQ_IN_ACT_LOG;
1241		req->in_actlog_jif = jiffies;
1242	}
1243	return req;
1244
1245 queue_for_submitter_thread:
1246	atomic_inc(&device->ap_actlog_cnt);
1247	drbd_queue_write(device, req);
1248	return NULL;
1249}
1250
1251/* Require at least one path to current data.
1252 * We don't want to allow writes on C_STANDALONE D_INCONSISTENT:
1253 * We would not allow to read what was written,
1254 * we would not have bumped the data generation uuids,
1255 * we would cause data divergence for all the wrong reasons.
1256 *
1257 * If we don't see at least one D_UP_TO_DATE, we will fail this request,
1258 * which either returns EIO, or, if OND_SUSPEND_IO is set, suspends IO,
1259 * and queues for retry later.
1260 */
1261static bool may_do_writes(struct drbd_device *device)
1262{
1263	const union drbd_dev_state s = device->state;
1264	return s.disk == D_UP_TO_DATE || s.pdsk == D_UP_TO_DATE;
1265}
1266
1267struct drbd_plug_cb {
1268	struct blk_plug_cb cb;
1269	struct drbd_request *most_recent_req;
1270	/* do we need more? */
1271};
1272
1273static void drbd_unplug(struct blk_plug_cb *cb, bool from_schedule)
1274{
1275	struct drbd_plug_cb *plug = container_of(cb, struct drbd_plug_cb, cb);
1276	struct drbd_resource *resource = plug->cb.data;
1277	struct drbd_request *req = plug->most_recent_req;
1278
1279	kfree(cb);
1280	if (!req)
1281		return;
1282
1283	spin_lock_irq(&resource->req_lock);
1284	/* In case the sender did not process it yet, raise the flag to
1285	 * have it followed with P_UNPLUG_REMOTE just after. */
1286	req->rq_state |= RQ_UNPLUG;
1287	/* but also queue a generic unplug */
1288	drbd_queue_unplug(req->device);
1289	kref_put(&req->kref, drbd_req_destroy);
1290	spin_unlock_irq(&resource->req_lock);
1291}
1292
1293static struct drbd_plug_cb* drbd_check_plugged(struct drbd_resource *resource)
1294{
1295	/* A lot of text to say
1296	 * return (struct drbd_plug_cb*)blk_check_plugged(); */
1297	struct drbd_plug_cb *plug;
1298	struct blk_plug_cb *cb = blk_check_plugged(drbd_unplug, resource, sizeof(*plug));
1299
1300	if (cb)
1301		plug = container_of(cb, struct drbd_plug_cb, cb);
1302	else
1303		plug = NULL;
1304	return plug;
1305}
1306
1307static void drbd_update_plug(struct drbd_plug_cb *plug, struct drbd_request *req)
1308{
1309	struct drbd_request *tmp = plug->most_recent_req;
1310	/* Will be sent to some peer.
1311	 * Remember to tag it with UNPLUG_REMOTE on unplug */
1312	kref_get(&req->kref);
1313	plug->most_recent_req = req;
1314	if (tmp)
1315		kref_put(&tmp->kref, drbd_req_destroy);
1316}
1317
1318static void drbd_send_and_submit(struct drbd_device *device, struct drbd_request *req)
1319{
1320	struct drbd_resource *resource = device->resource;
1321	struct drbd_peer_device *peer_device = first_peer_device(device);
1322	const int rw = bio_data_dir(req->master_bio);
1323	struct bio_and_error m = { NULL, };
1324	bool no_remote = false;
1325	bool submit_private_bio = false;
1326
1327	spin_lock_irq(&resource->req_lock);
1328	if (rw == WRITE) {
1329		/* This may temporarily give up the req_lock,
1330		 * but will re-aquire it before it returns here.
1331		 * Needs to be before the check on drbd_suspended() */
1332		complete_conflicting_writes(req);
1333		/* no more giving up req_lock from now on! */
1334
1335		/* check for congestion, and potentially stop sending
1336		 * full data updates, but start sending "dirty bits" only. */
1337		maybe_pull_ahead(device);
1338	}
1339
1340
1341	if (drbd_suspended(device)) {
1342		/* push back and retry: */
1343		req->rq_state |= RQ_POSTPONED;
1344		if (req->private_bio) {
1345			bio_put(req->private_bio);
1346			req->private_bio = NULL;
1347			put_ldev(device);
1348		}
1349		goto out;
1350	}
1351
1352	/* We fail READ early, if we can not serve it.
1353	 * We must do this before req is registered on any lists.
1354	 * Otherwise, drbd_req_complete() will queue failed READ for retry. */
1355	if (rw != WRITE) {
1356		if (!do_remote_read(req) && !req->private_bio)
1357			goto nodata;
1358	}
1359
1360	/* which transfer log epoch does this belong to? */
1361	req->epoch = atomic_read(&first_peer_device(device)->connection->current_tle_nr);
1362
1363	/* no point in adding empty flushes to the transfer log,
1364	 * they are mapped to drbd barriers already. */
1365	if (likely(req->i.size!=0)) {
1366		if (rw == WRITE)
1367			first_peer_device(device)->connection->current_tle_writes++;
1368
1369		list_add_tail(&req->tl_requests, &first_peer_device(device)->connection->transfer_log);
1370	}
1371
1372	if (rw == WRITE) {
1373		if (req->private_bio && !may_do_writes(device)) {
1374			bio_put(req->private_bio);
1375			req->private_bio = NULL;
1376			put_ldev(device);
1377			goto nodata;
1378		}
1379		if (!drbd_process_write_request(req))
1380			no_remote = true;
1381	} else {
1382		/* We either have a private_bio, or we can read from remote.
1383		 * Otherwise we had done the goto nodata above. */
1384		if (req->private_bio == NULL) {
1385			_req_mod(req, TO_BE_SENT, peer_device);
1386			_req_mod(req, QUEUE_FOR_NET_READ, peer_device);
1387		} else
1388			no_remote = true;
1389	}
1390
1391	if (no_remote == false) {
1392		struct drbd_plug_cb *plug = drbd_check_plugged(resource);
1393		if (plug)
1394			drbd_update_plug(plug, req);
1395	}
1396
1397	/* If it took the fast path in drbd_request_prepare, add it here.
1398	 * The slow path has added it already. */
1399	if (list_empty(&req->req_pending_master_completion))
1400		list_add_tail(&req->req_pending_master_completion,
1401			&device->pending_master_completion[rw == WRITE]);
1402	if (req->private_bio) {
1403		/* needs to be marked within the same spinlock */
1404		req->pre_submit_jif = jiffies;
1405		list_add_tail(&req->req_pending_local,
1406			&device->pending_completion[rw == WRITE]);
1407		_req_mod(req, TO_BE_SUBMITTED, NULL);
1408		/* but we need to give up the spinlock to submit */
1409		submit_private_bio = true;
1410	} else if (no_remote) {
1411nodata:
1412		if (drbd_ratelimit())
1413			drbd_err(device, "IO ERROR: neither local nor remote data, sector %llu+%u\n",
1414					(unsigned long long)req->i.sector, req->i.size >> 9);
1415		/* A write may have been queued for send_oos, however.
1416		 * So we can not simply free it, we must go through drbd_req_put_completion_ref() */
1417	}
1418
1419out:
1420	drbd_req_put_completion_ref(req, &m, 1);
1421	spin_unlock_irq(&resource->req_lock);
1422
1423	/* Even though above is a kref_put(), this is safe.
1424	 * As long as we still need to submit our private bio,
1425	 * we hold a completion ref, and the request cannot disappear.
1426	 * If however this request did not even have a private bio to submit
1427	 * (e.g. remote read), req may already be invalid now.
1428	 * That's why we cannot check on req->private_bio. */
1429	if (submit_private_bio)
1430		drbd_submit_req_private_bio(req);
1431	if (m.bio)
1432		complete_master_bio(device, &m);
1433}
1434
1435void __drbd_make_request(struct drbd_device *device, struct bio *bio)
1436{
1437	struct drbd_request *req = drbd_request_prepare(device, bio);
1438	if (IS_ERR_OR_NULL(req))
1439		return;
1440	drbd_send_and_submit(device, req);
1441}
1442
1443static void submit_fast_path(struct drbd_device *device, struct list_head *incoming)
1444{
1445	struct blk_plug plug;
1446	struct drbd_request *req, *tmp;
1447
1448	blk_start_plug(&plug);
1449	list_for_each_entry_safe(req, tmp, incoming, tl_requests) {
1450		const int rw = bio_data_dir(req->master_bio);
1451
1452		if (rw == WRITE /* rw != WRITE should not even end up here! */
1453		&& req->private_bio && req->i.size
1454		&& !test_bit(AL_SUSPENDED, &device->flags)) {
1455			if (!drbd_al_begin_io_fastpath(device, &req->i))
1456				continue;
1457
1458			req->rq_state |= RQ_IN_ACT_LOG;
1459			req->in_actlog_jif = jiffies;
1460			atomic_dec(&device->ap_actlog_cnt);
1461		}
1462
1463		list_del_init(&req->tl_requests);
1464		drbd_send_and_submit(device, req);
1465	}
1466	blk_finish_plug(&plug);
1467}
1468
1469static bool prepare_al_transaction_nonblock(struct drbd_device *device,
1470					    struct list_head *incoming,
1471					    struct list_head *pending,
1472					    struct list_head *later)
1473{
1474	struct drbd_request *req;
1475	int wake = 0;
1476	int err;
1477
1478	spin_lock_irq(&device->al_lock);
1479	while ((req = list_first_entry_or_null(incoming, struct drbd_request, tl_requests))) {
1480		err = drbd_al_begin_io_nonblock(device, &req->i);
1481		if (err == -ENOBUFS)
1482			break;
1483		if (err == -EBUSY)
1484			wake = 1;
1485		if (err)
1486			list_move_tail(&req->tl_requests, later);
1487		else
1488			list_move_tail(&req->tl_requests, pending);
1489	}
1490	spin_unlock_irq(&device->al_lock);
1491	if (wake)
1492		wake_up(&device->al_wait);
1493	return !list_empty(pending);
1494}
1495
1496static void send_and_submit_pending(struct drbd_device *device, struct list_head *pending)
1497{
1498	struct blk_plug plug;
1499	struct drbd_request *req;
1500
1501	blk_start_plug(&plug);
1502	while ((req = list_first_entry_or_null(pending, struct drbd_request, tl_requests))) {
1503		req->rq_state |= RQ_IN_ACT_LOG;
1504		req->in_actlog_jif = jiffies;
1505		atomic_dec(&device->ap_actlog_cnt);
1506		list_del_init(&req->tl_requests);
1507		drbd_send_and_submit(device, req);
1508	}
1509	blk_finish_plug(&plug);
1510}
1511
1512void do_submit(struct work_struct *ws)
1513{
1514	struct drbd_device *device = container_of(ws, struct drbd_device, submit.worker);
1515	LIST_HEAD(incoming);	/* from drbd_make_request() */
1516	LIST_HEAD(pending);	/* to be submitted after next AL-transaction commit */
1517	LIST_HEAD(busy);	/* blocked by resync requests */
1518
1519	/* grab new incoming requests */
1520	spin_lock_irq(&device->resource->req_lock);
1521	list_splice_tail_init(&device->submit.writes, &incoming);
1522	spin_unlock_irq(&device->resource->req_lock);
1523
1524	for (;;) {
1525		DEFINE_WAIT(wait);
1526
1527		/* move used-to-be-busy back to front of incoming */
1528		list_splice_init(&busy, &incoming);
1529		submit_fast_path(device, &incoming);
1530		if (list_empty(&incoming))
1531			break;
1532
1533		for (;;) {
1534			prepare_to_wait(&device->al_wait, &wait, TASK_UNINTERRUPTIBLE);
1535
1536			list_splice_init(&busy, &incoming);
1537			prepare_al_transaction_nonblock(device, &incoming, &pending, &busy);
1538			if (!list_empty(&pending))
1539				break;
1540
1541			schedule();
1542
1543			/* If all currently "hot" activity log extents are kept busy by
1544			 * incoming requests, we still must not totally starve new
1545			 * requests to "cold" extents.
1546			 * Something left on &incoming means there had not been
1547			 * enough update slots available, and the activity log
1548			 * has been marked as "starving".
1549			 *
1550			 * Try again now, without looking for new requests,
1551			 * effectively blocking all new requests until we made
1552			 * at least _some_ progress with what we currently have.
1553			 */
1554			if (!list_empty(&incoming))
1555				continue;
1556
1557			/* Nothing moved to pending, but nothing left
1558			 * on incoming: all moved to busy!
1559			 * Grab new and iterate. */
1560			spin_lock_irq(&device->resource->req_lock);
1561			list_splice_tail_init(&device->submit.writes, &incoming);
1562			spin_unlock_irq(&device->resource->req_lock);
1563		}
1564		finish_wait(&device->al_wait, &wait);
1565
1566		/* If the transaction was full, before all incoming requests
1567		 * had been processed, skip ahead to commit, and iterate
1568		 * without splicing in more incoming requests from upper layers.
1569		 *
1570		 * Else, if all incoming have been processed,
1571		 * they have become either "pending" (to be submitted after
1572		 * next transaction commit) or "busy" (blocked by resync).
1573		 *
1574		 * Maybe more was queued, while we prepared the transaction?
1575		 * Try to stuff those into this transaction as well.
1576		 * Be strictly non-blocking here,
1577		 * we already have something to commit.
1578		 *
1579		 * Commit if we don't make any more progres.
1580		 */
1581
1582		while (list_empty(&incoming)) {
1583			LIST_HEAD(more_pending);
1584			LIST_HEAD(more_incoming);
1585			bool made_progress;
1586
1587			/* It is ok to look outside the lock,
1588			 * it's only an optimization anyways */
1589			if (list_empty(&device->submit.writes))
1590				break;
1591
1592			spin_lock_irq(&device->resource->req_lock);
1593			list_splice_tail_init(&device->submit.writes, &more_incoming);
1594			spin_unlock_irq(&device->resource->req_lock);
1595
1596			if (list_empty(&more_incoming))
1597				break;
1598
1599			made_progress = prepare_al_transaction_nonblock(device, &more_incoming, &more_pending, &busy);
1600
1601			list_splice_tail_init(&more_pending, &pending);
1602			list_splice_tail_init(&more_incoming, &incoming);
1603			if (!made_progress)
1604				break;
1605		}
1606
1607		drbd_al_begin_io_commit(device);
1608		send_and_submit_pending(device, &pending);
1609	}
1610}
1611
1612void drbd_submit_bio(struct bio *bio)
1613{
1614	struct drbd_device *device = bio->bi_bdev->bd_disk->private_data;
1615
1616	bio = bio_split_to_limits(bio);
1617	if (!bio)
1618		return;
1619
1620	/*
1621	 * what we "blindly" assume:
1622	 */
1623	D_ASSERT(device, IS_ALIGNED(bio->bi_iter.bi_size, 512));
1624
1625	inc_ap_bio(device);
1626	__drbd_make_request(device, bio);
1627}
1628
1629static bool net_timeout_reached(struct drbd_request *net_req,
1630		struct drbd_connection *connection,
1631		unsigned long now, unsigned long ent,
1632		unsigned int ko_count, unsigned int timeout)
1633{
1634	struct drbd_device *device = net_req->device;
1635
1636	if (!time_after(now, net_req->pre_send_jif + ent))
1637		return false;
1638
1639	if (time_in_range(now, connection->last_reconnect_jif, connection->last_reconnect_jif + ent))
1640		return false;
1641
1642	if (net_req->rq_state & RQ_NET_PENDING) {
1643		drbd_warn(device, "Remote failed to finish a request within %ums > ko-count (%u) * timeout (%u * 0.1s)\n",
1644			jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1645		return true;
1646	}
1647
1648	/* We received an ACK already (or are using protocol A),
1649	 * but are waiting for the epoch closing barrier ack.
1650	 * Check if we sent the barrier already.  We should not blame the peer
1651	 * for being unresponsive, if we did not even ask it yet. */
1652	if (net_req->epoch == connection->send.current_epoch_nr) {
1653		drbd_warn(device,
1654			"We did not send a P_BARRIER for %ums > ko-count (%u) * timeout (%u * 0.1s); drbd kernel thread blocked?\n",
1655			jiffies_to_msecs(now - net_req->pre_send_jif), ko_count, timeout);
1656		return false;
1657	}
1658
1659	/* Worst case: we may have been blocked for whatever reason, then
1660	 * suddenly are able to send a lot of requests (and epoch separating
1661	 * barriers) in quick succession.
1662	 * The timestamp of the net_req may be much too old and not correspond
1663	 * to the sending time of the relevant unack'ed barrier packet, so
1664	 * would trigger a spurious timeout.  The latest barrier packet may
1665	 * have a too recent timestamp to trigger the timeout, potentially miss
1666	 * a timeout.  Right now we don't have a place to conveniently store
1667	 * these timestamps.
1668	 * But in this particular situation, the application requests are still
1669	 * completed to upper layers, DRBD should still "feel" responsive.
1670	 * No need yet to kill this connection, it may still recover.
1671	 * If not, eventually we will have queued enough into the network for
1672	 * us to block. From that point of view, the timestamp of the last sent
1673	 * barrier packet is relevant enough.
1674	 */
1675	if (time_after(now, connection->send.last_sent_barrier_jif + ent)) {
1676		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",
1677			connection->send.last_sent_barrier_jif, now,
1678			jiffies_to_msecs(now - connection->send.last_sent_barrier_jif), ko_count, timeout);
1679		return true;
1680	}
1681	return false;
1682}
1683
1684/* A request is considered timed out, if
1685 * - we have some effective timeout from the configuration,
1686 *   with some state restrictions applied,
1687 * - the oldest request is waiting for a response from the network
1688 *   resp. the local disk,
1689 * - the oldest request is in fact older than the effective timeout,
1690 * - the connection was established (resp. disk was attached)
1691 *   for longer than the timeout already.
1692 * Note that for 32bit jiffies and very stable connections/disks,
1693 * we may have a wrap around, which is catched by
1694 *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
1695 *
1696 * Side effect: once per 32bit wrap-around interval, which means every
1697 * ~198 days with 250 HZ, we have a window where the timeout would need
1698 * to expire twice (worst case) to become effective. Good enough.
1699 */
1700
1701void request_timer_fn(struct timer_list *t)
1702{
1703	struct drbd_device *device = timer_container_of(device, t,
1704							request_timer);
1705	struct drbd_connection *connection = first_peer_device(device)->connection;
1706	struct drbd_request *req_read, *req_write, *req_peer; /* oldest request */
1707	struct net_conf *nc;
1708	unsigned long oldest_submit_jif;
1709	unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
1710	unsigned long now;
1711	unsigned int ko_count = 0, timeout = 0;
1712
1713	rcu_read_lock();
1714	nc = rcu_dereference(connection->net_conf);
1715	if (nc && device->state.conn >= C_WF_REPORT_PARAMS) {
1716		ko_count = nc->ko_count;
1717		timeout = nc->timeout;
1718	}
1719
1720	if (get_ldev(device)) { /* implicit state.disk >= D_INCONSISTENT */
1721		dt = rcu_dereference(device->ldev->disk_conf)->disk_timeout * HZ / 10;
1722		put_ldev(device);
1723	}
1724	rcu_read_unlock();
1725
1726
1727	ent = timeout * HZ/10 * ko_count;
1728	et = min_not_zero(dt, ent);
1729
1730	if (!et)
1731		return; /* Recurring timer stopped */
1732
1733	now = jiffies;
1734	nt = now + et;
1735
1736	spin_lock_irq(&device->resource->req_lock);
1737	req_read = list_first_entry_or_null(&device->pending_completion[0], struct drbd_request, req_pending_local);
1738	req_write = list_first_entry_or_null(&device->pending_completion[1], struct drbd_request, req_pending_local);
1739
1740	/* maybe the oldest request waiting for the peer is in fact still
1741	 * blocking in tcp sendmsg.  That's ok, though, that's handled via the
1742	 * socket send timeout, requesting a ping, and bumping ko-count in
1743	 * we_should_drop_the_connection().
1744	 */
1745
1746	/* check the oldest request we did successfully sent,
1747	 * but which is still waiting for an ACK. */
1748	req_peer = connection->req_ack_pending;
1749
1750	/* if we don't have such request (e.g. protocoll A)
1751	 * check the oldest requests which is still waiting on its epoch
1752	 * closing barrier ack. */
1753	if (!req_peer)
1754		req_peer = connection->req_not_net_done;
1755
1756	/* evaluate the oldest peer request only in one timer! */
1757	if (req_peer && req_peer->device != device)
1758		req_peer = NULL;
1759
1760	/* do we have something to evaluate? */
1761	if (req_peer == NULL && req_write == NULL && req_read == NULL)
1762		goto out;
1763
1764	oldest_submit_jif =
1765		(req_write && req_read)
1766		? ( time_before(req_write->pre_submit_jif, req_read->pre_submit_jif)
1767		  ? req_write->pre_submit_jif : req_read->pre_submit_jif )
1768		: req_write ? req_write->pre_submit_jif
1769		: req_read ? req_read->pre_submit_jif : now;
1770
1771	if (ent && req_peer && net_timeout_reached(req_peer, connection, now, ent, ko_count, timeout))
1772		_conn_request_state(connection, NS(conn, C_TIMEOUT), CS_VERBOSE | CS_HARD);
1773
1774	if (dt && oldest_submit_jif != now &&
1775		 time_after(now, oldest_submit_jif + dt) &&
1776		!time_in_range(now, device->last_reattach_jif, device->last_reattach_jif + dt)) {
1777		drbd_warn(device, "Local backing device failed to meet the disk-timeout\n");
1778		__drbd_chk_io_error(device, DRBD_FORCE_DETACH);
1779	}
1780
1781	/* Reschedule timer for the nearest not already expired timeout.
1782	 * Fallback to now + min(effective network timeout, disk timeout). */
1783	ent = (ent && req_peer && time_before(now, req_peer->pre_send_jif + ent))
1784		? req_peer->pre_send_jif + ent : now + et;
1785	dt = (dt && oldest_submit_jif != now && time_before(now, oldest_submit_jif + dt))
1786		? oldest_submit_jif + dt : now + et;
1787	nt = time_before(ent, dt) ? ent : dt;
1788out:
1789	spin_unlock_irq(&device->resource->req_lock);
1790	mod_timer(&device->request_timer, nt);
1791}