xfs_log.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_mount.h"
  13#include "xfs_errortag.h"
  14#include "xfs_error.h"
  15#include "xfs_trans.h"
  16#include "xfs_trans_priv.h"
  17#include "xfs_log.h"
  18#include "xfs_log_priv.h"
  19#include "xfs_trace.h"
  20#include "xfs_sysfs.h"
  21#include "xfs_sb.h"
  22#include "xfs_health.h"
  23#include "xfs_zone_alloc.h"
  24
  25struct kmem_cache	*xfs_log_ticket_cache;
  26
  27/* Local miscellaneous function prototypes */
  28STATIC struct xlog *
  29xlog_alloc_log(
  30	struct xfs_mount	*mp,
  31	struct xfs_buftarg	*log_target,
  32	xfs_daddr_t		blk_offset,
  33	int			num_bblks);
  34STATIC void
  35xlog_dealloc_log(
  36	struct xlog		*log);
  37
  38/* local state machine functions */
  39STATIC void xlog_state_done_syncing(
  40	struct xlog_in_core	*iclog);
  41STATIC void xlog_state_do_callback(
  42	struct xlog		*log);
  43STATIC int
  44xlog_state_get_iclog_space(
  45	struct xlog		*log,
  46	int			len,
  47	struct xlog_in_core	**iclog,
  48	struct xlog_ticket	*ticket,
  49	int			*logoffsetp);
  50STATIC void
  51xlog_sync(
  52	struct xlog		*log,
  53	struct xlog_in_core	*iclog,
  54	struct xlog_ticket	*ticket);
  55#if defined(DEBUG)
  56STATIC void
  57xlog_verify_iclog(
  58	struct xlog		*log,
  59	struct xlog_in_core	*iclog,
  60	int			count);
  61STATIC void
  62xlog_verify_tail_lsn(
  63	struct xlog		*log,
  64	struct xlog_in_core	*iclog);
  65#else
  66#define xlog_verify_iclog(a,b,c)
  67#define xlog_verify_tail_lsn(a,b)
  68#endif
  69
  70STATIC int
  71xlog_iclogs_empty(
  72	struct xlog		*log);
  73
  74static int
  75xfs_log_cover(struct xfs_mount *);
  76
  77/*
  78 * We need to make sure the buffer pointer returned is naturally aligned for the
  79 * biggest basic data type we put into it. We have already accounted for this
  80 * padding when sizing the buffer.
  81 *
  82 * However, this padding does not get written into the log, and hence we have to
  83 * track the space used by the log vectors separately to prevent log space hangs
  84 * due to inaccurate accounting (i.e. a leak) of the used log space through the
  85 * CIL context ticket.
  86 *
  87 * We also add space for the xlog_op_header that describes this region in the
  88 * log. This prepends the data region we return to the caller to copy their data
  89 * into, so do all the static initialisation of the ophdr now. Because the ophdr
  90 * is not 8 byte aligned, we have to be careful to ensure that we align the
  91 * start of the buffer such that the region we return to the call is 8 byte
  92 * aligned and packed against the tail of the ophdr.
  93 */
  94void *
  95xlog_prepare_iovec(
  96	struct xfs_log_vec	*lv,
  97	struct xfs_log_iovec	**vecp,
  98	uint			type)
  99{
 100	struct xfs_log_iovec	*vec = *vecp;
 101	struct xlog_op_header	*oph;
 102	uint32_t		len;
 103	void			*buf;
 104
 105	if (vec) {
 106		ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
 107		vec++;
 108	} else {
 109		vec = &lv->lv_iovecp[0];
 110	}
 111
 112	len = lv->lv_buf_used + sizeof(struct xlog_op_header);
 113	if (!IS_ALIGNED(len, sizeof(uint64_t))) {
 114		lv->lv_buf_used = round_up(len, sizeof(uint64_t)) -
 115					sizeof(struct xlog_op_header);
 116	}
 117
 118	vec->i_type = type;
 119	vec->i_addr = lv->lv_buf + lv->lv_buf_used;
 120
 121	oph = vec->i_addr;
 122	oph->oh_clientid = XFS_TRANSACTION;
 123	oph->oh_res2 = 0;
 124	oph->oh_flags = 0;
 125
 126	buf = vec->i_addr + sizeof(struct xlog_op_header);
 127	ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
 128
 129	*vecp = vec;
 130	return buf;
 131}
 132
 133static inline void
 134xlog_grant_sub_space(
 135	struct xlog_grant_head	*head,
 136	int64_t			bytes)
 137{
 138	atomic64_sub(bytes, &head->grant);
 139}
 140
 141static inline void
 142xlog_grant_add_space(
 143	struct xlog_grant_head	*head,
 144	int64_t			bytes)
 145{
 146	atomic64_add(bytes, &head->grant);
 147}
 148
 149static void
 150xlog_grant_head_init(
 151	struct xlog_grant_head	*head)
 152{
 153	atomic64_set(&head->grant, 0);
 154	INIT_LIST_HEAD(&head->waiters);
 155	spin_lock_init(&head->lock);
 156}
 157
 158void
 159xlog_grant_return_space(
 160	struct xlog	*log,
 161	xfs_lsn_t	old_head,
 162	xfs_lsn_t	new_head)
 163{
 164	int64_t		diff = xlog_lsn_sub(log, new_head, old_head);
 165
 166	xlog_grant_sub_space(&log->l_reserve_head, diff);
 167	xlog_grant_sub_space(&log->l_write_head, diff);
 168}
 169
 170/*
 171 * Return the space in the log between the tail and the head.  In the case where
 172 * we have overrun available reservation space, return 0. The memory barrier
 173 * pairs with the smp_wmb() in xlog_cil_ail_insert() to ensure that grant head
 174 * vs tail space updates are seen in the correct order and hence avoid
 175 * transients as space is transferred from the grant heads to the AIL on commit
 176 * completion.
 177 */
 178static uint64_t
 179xlog_grant_space_left(
 180	struct xlog		*log,
 181	struct xlog_grant_head	*head)
 182{
 183	int64_t			free_bytes;
 184
 185	smp_rmb();	/* paired with smp_wmb in xlog_cil_ail_insert() */
 186	free_bytes = log->l_logsize - READ_ONCE(log->l_tail_space) -
 187			atomic64_read(&head->grant);
 188	if (free_bytes > 0)
 189		return free_bytes;
 190	return 0;
 191}
 192
 193STATIC void
 194xlog_grant_head_wake_all(
 195	struct xlog_grant_head	*head)
 196{
 197	struct xlog_ticket	*tic;
 198
 199	spin_lock(&head->lock);
 200	list_for_each_entry(tic, &head->waiters, t_queue)
 201		wake_up_process(tic->t_task);
 202	spin_unlock(&head->lock);
 203}
 204
 205static inline int
 206xlog_ticket_reservation(
 207	struct xlog		*log,
 208	struct xlog_grant_head	*head,
 209	struct xlog_ticket	*tic)
 210{
 211	if (head == &log->l_write_head) {
 212		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 213		return tic->t_unit_res;
 214	}
 215
 216	if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 217		return tic->t_unit_res * tic->t_cnt;
 218
 219	return tic->t_unit_res;
 220}
 221
 222STATIC bool
 223xlog_grant_head_wake(
 224	struct xlog		*log,
 225	struct xlog_grant_head	*head,
 226	int			*free_bytes)
 227{
 228	struct xlog_ticket	*tic;
 229	int			need_bytes;
 230
 231	list_for_each_entry(tic, &head->waiters, t_queue) {
 232		need_bytes = xlog_ticket_reservation(log, head, tic);
 233		if (*free_bytes < need_bytes)
 234			return false;
 235
 236		*free_bytes -= need_bytes;
 237		trace_xfs_log_grant_wake_up(log, tic);
 238		wake_up_process(tic->t_task);
 239	}
 240
 241	return true;
 242}
 243
 244STATIC int
 245xlog_grant_head_wait(
 246	struct xlog		*log,
 247	struct xlog_grant_head	*head,
 248	struct xlog_ticket	*tic,
 249	int			need_bytes) __releases(&head->lock)
 250					    __acquires(&head->lock)
 251{
 252	list_add_tail(&tic->t_queue, &head->waiters);
 253
 254	do {
 255		if (xlog_is_shutdown(log))
 256			goto shutdown;
 257
 258		__set_current_state(TASK_UNINTERRUPTIBLE);
 259		spin_unlock(&head->lock);
 260
 261		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 262
 263		/* Push on the AIL to free up all the log space. */
 264		xfs_ail_push_all(log->l_ailp);
 265
 266		trace_xfs_log_grant_sleep(log, tic);
 267		schedule();
 268		trace_xfs_log_grant_wake(log, tic);
 269
 270		spin_lock(&head->lock);
 271		if (xlog_is_shutdown(log))
 272			goto shutdown;
 273	} while (xlog_grant_space_left(log, head) < need_bytes);
 274
 275	list_del_init(&tic->t_queue);
 276	return 0;
 277shutdown:
 278	list_del_init(&tic->t_queue);
 279	return -EIO;
 280}
 281
 282/*
 283 * Atomically get the log space required for a log ticket.
 284 *
 285 * Once a ticket gets put onto head->waiters, it will only return after the
 286 * needed reservation is satisfied.
 287 *
 288 * This function is structured so that it has a lock free fast path. This is
 289 * necessary because every new transaction reservation will come through this
 290 * path. Hence any lock will be globally hot if we take it unconditionally on
 291 * every pass.
 292 *
 293 * As tickets are only ever moved on and off head->waiters under head->lock, we
 294 * only need to take that lock if we are going to add the ticket to the queue
 295 * and sleep. We can avoid taking the lock if the ticket was never added to
 296 * head->waiters because the t_queue list head will be empty and we hold the
 297 * only reference to it so it can safely be checked unlocked.
 298 */
 299STATIC int
 300xlog_grant_head_check(
 301	struct xlog		*log,
 302	struct xlog_grant_head	*head,
 303	struct xlog_ticket	*tic,
 304	int			*need_bytes)
 305{
 306	int			free_bytes;
 307	int			error = 0;
 308
 309	ASSERT(!xlog_in_recovery(log));
 310
 311	/*
 312	 * If there are other waiters on the queue then give them a chance at
 313	 * logspace before us.  Wake up the first waiters, if we do not wake
 314	 * up all the waiters then go to sleep waiting for more free space,
 315	 * otherwise try to get some space for this transaction.
 316	 */
 317	*need_bytes = xlog_ticket_reservation(log, head, tic);
 318	free_bytes = xlog_grant_space_left(log, head);
 319	if (!list_empty_careful(&head->waiters)) {
 320		spin_lock(&head->lock);
 321		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 322		    free_bytes < *need_bytes) {
 323			error = xlog_grant_head_wait(log, head, tic,
 324						     *need_bytes);
 325		}
 326		spin_unlock(&head->lock);
 327	} else if (free_bytes < *need_bytes) {
 328		spin_lock(&head->lock);
 329		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 330		spin_unlock(&head->lock);
 331	}
 332
 333	return error;
 334}
 335
 336bool
 337xfs_log_writable(
 338	struct xfs_mount	*mp)
 339{
 340	/*
 341	 * Do not write to the log on norecovery mounts, if the data or log
 342	 * devices are read-only, or if the filesystem is shutdown. Read-only
 343	 * mounts allow internal writes for log recovery and unmount purposes,
 344	 * so don't restrict that case.
 345	 */
 346	if (xfs_has_norecovery(mp))
 347		return false;
 348	if (xfs_readonly_buftarg(mp->m_ddev_targp))
 349		return false;
 350	if (xfs_readonly_buftarg(mp->m_log->l_targ))
 351		return false;
 352	if (xlog_is_shutdown(mp->m_log))
 353		return false;
 354	return true;
 355}
 356
 357/*
 358 * Replenish the byte reservation required by moving the grant write head.
 359 */
 360int
 361xfs_log_regrant(
 362	struct xfs_mount	*mp,
 363	struct xlog_ticket	*tic)
 364{
 365	struct xlog		*log = mp->m_log;
 366	int			need_bytes;
 367	int			error = 0;
 368
 369	if (xlog_is_shutdown(log))
 370		return -EIO;
 371
 372	XFS_STATS_INC(mp, xs_try_logspace);
 373
 374	/*
 375	 * This is a new transaction on the ticket, so we need to change the
 376	 * transaction ID so that the next transaction has a different TID in
 377	 * the log. Just add one to the existing tid so that we can see chains
 378	 * of rolling transactions in the log easily.
 379	 */
 380	tic->t_tid++;
 381	tic->t_curr_res = tic->t_unit_res;
 382	if (tic->t_cnt > 0)
 383		return 0;
 384
 385	trace_xfs_log_regrant(log, tic);
 386
 387	error = xlog_grant_head_check(log, &log->l_write_head, tic,
 388				      &need_bytes);
 389	if (error)
 390		goto out_error;
 391
 392	xlog_grant_add_space(&log->l_write_head, need_bytes);
 393	trace_xfs_log_regrant_exit(log, tic);
 394	return 0;
 395
 396out_error:
 397	/*
 398	 * If we are failing, make sure the ticket doesn't have any current
 399	 * reservations.  We don't want to add this back when the ticket/
 400	 * transaction gets cancelled.
 401	 */
 402	tic->t_curr_res = 0;
 403	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
 404	return error;
 405}
 406
 407/*
 408 * Reserve log space and return a ticket corresponding to the reservation.
 409 *
 410 * Each reservation is going to reserve extra space for a log record header.
 411 * When writes happen to the on-disk log, we don't subtract the length of the
 412 * log record header from any reservation.  By wasting space in each
 413 * reservation, we prevent over allocation problems.
 414 */
 415int
 416xfs_log_reserve(
 417	struct xfs_mount	*mp,
 418	int			unit_bytes,
 419	int			cnt,
 420	struct xlog_ticket	**ticp,
 421	bool			permanent)
 422{
 423	struct xlog		*log = mp->m_log;
 424	struct xlog_ticket	*tic;
 425	int			need_bytes;
 426	int			error = 0;
 427
 428	if (xlog_is_shutdown(log))
 429		return -EIO;
 430
 431	XFS_STATS_INC(mp, xs_try_logspace);
 432
 433	ASSERT(*ticp == NULL);
 434	tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
 435	*ticp = tic;
 436	trace_xfs_log_reserve(log, tic);
 437	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 438				      &need_bytes);
 439	if (error)
 440		goto out_error;
 441
 442	xlog_grant_add_space(&log->l_reserve_head, need_bytes);
 443	xlog_grant_add_space(&log->l_write_head, need_bytes);
 444	trace_xfs_log_reserve_exit(log, tic);
 445	return 0;
 446
 447out_error:
 448	/*
 449	 * If we are failing, make sure the ticket doesn't have any current
 450	 * reservations.  We don't want to add this back when the ticket/
 451	 * transaction gets cancelled.
 452	 */
 453	tic->t_curr_res = 0;
 454	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
 455	return error;
 456}
 457
 458/*
 459 * Run all the pending iclog callbacks and wake log force waiters and iclog
 460 * space waiters so they can process the newly set shutdown state. We really
 461 * don't care what order we process callbacks here because the log is shut down
 462 * and so state cannot change on disk anymore. However, we cannot wake waiters
 463 * until the callbacks have been processed because we may be in unmount and
 464 * we must ensure that all AIL operations the callbacks perform have completed
 465 * before we tear down the AIL.
 466 *
 467 * We avoid processing actively referenced iclogs so that we don't run callbacks
 468 * while the iclog owner might still be preparing the iclog for IO submssion.
 469 * These will be caught by xlog_state_iclog_release() and call this function
 470 * again to process any callbacks that may have been added to that iclog.
 471 */
 472static void
 473xlog_state_shutdown_callbacks(
 474	struct xlog		*log)
 475{
 476	struct xlog_in_core	*iclog;
 477	LIST_HEAD(cb_list);
 478
 479	iclog = log->l_iclog;
 480	do {
 481		if (atomic_read(&iclog->ic_refcnt)) {
 482			/* Reference holder will re-run iclog callbacks. */
 483			continue;
 484		}
 485		list_splice_init(&iclog->ic_callbacks, &cb_list);
 486		spin_unlock(&log->l_icloglock);
 487
 488		xlog_cil_process_committed(&cb_list);
 489
 490		spin_lock(&log->l_icloglock);
 491		wake_up_all(&iclog->ic_write_wait);
 492		wake_up_all(&iclog->ic_force_wait);
 493	} while ((iclog = iclog->ic_next) != log->l_iclog);
 494
 495	wake_up_all(&log->l_flush_wait);
 496}
 497
 498/*
 499 * Flush iclog to disk if this is the last reference to the given iclog and the
 500 * it is in the WANT_SYNC state.
 501 *
 502 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
 503 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
 504 * written to stable storage, and implies that a commit record is contained
 505 * within the iclog. We need to ensure that the log tail does not move beyond
 506 * the tail that the first commit record in the iclog ordered against, otherwise
 507 * correct recovery of that checkpoint becomes dependent on future operations
 508 * performed on this iclog.
 509 *
 510 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
 511 * current tail into iclog. Once the iclog tail is set, future operations must
 512 * not modify it, otherwise they potentially violate ordering constraints for
 513 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
 514 * the iclog will get zeroed on activation of the iclog after sync, so we
 515 * always capture the tail lsn on the iclog on the first NEED_FUA release
 516 * regardless of the number of active reference counts on this iclog.
 517 */
 518int
 519xlog_state_release_iclog(
 520	struct xlog		*log,
 521	struct xlog_in_core	*iclog,
 522	struct xlog_ticket	*ticket)
 523{
 524	bool			last_ref;
 525
 526	lockdep_assert_held(&log->l_icloglock);
 527
 528	trace_xlog_iclog_release(iclog, _RET_IP_);
 529	/*
 530	 * Grabbing the current log tail needs to be atomic w.r.t. the writing
 531	 * of the tail LSN into the iclog so we guarantee that the log tail does
 532	 * not move between the first time we know that the iclog needs to be
 533	 * made stable and when we eventually submit it.
 534	 */
 535	if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
 536	     (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
 537	    !iclog->ic_header->h_tail_lsn) {
 538		iclog->ic_header->h_tail_lsn =
 539				cpu_to_be64(atomic64_read(&log->l_tail_lsn));
 540	}
 541
 542	last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
 543
 544	if (xlog_is_shutdown(log)) {
 545		/*
 546		 * If there are no more references to this iclog, process the
 547		 * pending iclog callbacks that were waiting on the release of
 548		 * this iclog.
 549		 */
 550		if (last_ref)
 551			xlog_state_shutdown_callbacks(log);
 552		return -EIO;
 553	}
 554
 555	if (!last_ref)
 556		return 0;
 557
 558	if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
 559		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 560		return 0;
 561	}
 562
 563	iclog->ic_state = XLOG_STATE_SYNCING;
 564	xlog_verify_tail_lsn(log, iclog);
 565	trace_xlog_iclog_syncing(iclog, _RET_IP_);
 566
 567	spin_unlock(&log->l_icloglock);
 568	xlog_sync(log, iclog, ticket);
 569	spin_lock(&log->l_icloglock);
 570	return 0;
 571}
 572
 573/*
 574 * Mount a log filesystem
 575 *
 576 * mp		- ubiquitous xfs mount point structure
 577 * log_target	- buftarg of on-disk log device
 578 * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
 579 * num_bblocks	- Number of BBSIZE blocks in on-disk log
 580 *
 581 * Return error or zero.
 582 */
 583int
 584xfs_log_mount(
 585	xfs_mount_t		*mp,
 586	struct xfs_buftarg	*log_target,
 587	xfs_daddr_t		blk_offset,
 588	int			num_bblks)
 589{
 590	struct xlog		*log;
 591	int			error = 0;
 592	int			min_logfsbs;
 593
 594	if (!xfs_has_norecovery(mp)) {
 595		xfs_notice(mp, "Mounting V%d Filesystem %pU",
 596			   XFS_SB_VERSION_NUM(&mp->m_sb),
 597			   &mp->m_sb.sb_uuid);
 598	} else {
 599		xfs_notice(mp,
 600"Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
 601			   XFS_SB_VERSION_NUM(&mp->m_sb),
 602			   &mp->m_sb.sb_uuid);
 603		ASSERT(xfs_is_readonly(mp));
 604	}
 605
 606	log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 607	if (IS_ERR(log)) {
 608		error = PTR_ERR(log);
 609		goto out;
 610	}
 611	mp->m_log = log;
 612
 613	/*
 614	 * Now that we have set up the log and it's internal geometry
 615	 * parameters, we can validate the given log space and drop a critical
 616	 * message via syslog if the log size is too small. A log that is too
 617	 * small can lead to unexpected situations in transaction log space
 618	 * reservation stage. The superblock verifier has already validated all
 619	 * the other log geometry constraints, so we don't have to check those
 620	 * here.
 621	 *
 622	 * Note: For v4 filesystems, we can't just reject the mount if the
 623	 * validation fails.  This would mean that people would have to
 624	 * downgrade their kernel just to remedy the situation as there is no
 625	 * way to grow the log (short of black magic surgery with xfs_db).
 626	 *
 627	 * We can, however, reject mounts for V5 format filesystems, as the
 628	 * mkfs binary being used to make the filesystem should never create a
 629	 * filesystem with a log that is too small.
 630	 */
 631	min_logfsbs = xfs_log_calc_minimum_size(mp);
 632	if (mp->m_sb.sb_logblocks < min_logfsbs) {
 633		xfs_warn(mp,
 634		"Log size %d blocks too small, minimum size is %d blocks",
 635			 mp->m_sb.sb_logblocks, min_logfsbs);
 636
 637		/*
 638		 * Log check errors are always fatal on v5; or whenever bad
 639		 * metadata leads to a crash.
 640		 */
 641		if (xfs_has_crc(mp)) {
 642			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 643			ASSERT(0);
 644			error = -EINVAL;
 645			goto out_free_log;
 646		}
 647		xfs_crit(mp, "Log size out of supported range.");
 648		xfs_crit(mp,
 649"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 650	}
 651
 652	/*
 653	 * Initialize the AIL now we have a log.
 654	 */
 655	error = xfs_trans_ail_init(mp);
 656	if (error) {
 657		xfs_warn(mp, "AIL initialisation failed: error %d", error);
 658		goto out_free_log;
 659	}
 660	log->l_ailp = mp->m_ail;
 661
 662	/*
 663	 * skip log recovery on a norecovery mount.  pretend it all
 664	 * just worked.
 665	 */
 666	if (!xfs_has_norecovery(mp)) {
 667		error = xlog_recover(log);
 668		if (error) {
 669			xfs_warn(mp, "log mount/recovery failed: error %d",
 670				error);
 671			xlog_recover_cancel(log);
 672			goto out_destroy_ail;
 673		}
 674	}
 675
 676	error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 677			       "log");
 678	if (error)
 679		goto out_destroy_ail;
 680
 681	/* Normal transactions can now occur */
 682	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
 683
 684	/*
 685	 * Now the log has been fully initialised and we know were our
 686	 * space grant counters are, we can initialise the permanent ticket
 687	 * needed for delayed logging to work.
 688	 */
 689	xlog_cil_init_post_recovery(log);
 690
 691	return 0;
 692
 693out_destroy_ail:
 694	xfs_trans_ail_destroy(mp);
 695out_free_log:
 696	xlog_dealloc_log(log);
 697out:
 698	return error;
 699}
 700
 701/*
 702 * Finish the recovery of the file system.  This is separate from the
 703 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 704 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 705 * here.
 706 *
 707 * If we finish recovery successfully, start the background log work. If we are
 708 * not doing recovery, then we have a RO filesystem and we don't need to start
 709 * it.
 710 */
 711int
 712xfs_log_mount_finish(
 713	struct xfs_mount	*mp)
 714{
 715	struct xlog		*log = mp->m_log;
 716	int			error = 0;
 717
 718	if (xfs_has_norecovery(mp)) {
 719		ASSERT(xfs_is_readonly(mp));
 720		return 0;
 721	}
 722
 723	/*
 724	 * During the second phase of log recovery, we need iget and
 725	 * iput to behave like they do for an active filesystem.
 726	 * xfs_fs_drop_inode needs to be able to prevent the deletion
 727	 * of inodes before we're done replaying log items on those
 728	 * inodes.  Turn it off immediately after recovery finishes
 729	 * so that we don't leak the quota inodes if subsequent mount
 730	 * activities fail.
 731	 *
 732	 * We let all inodes involved in redo item processing end up on
 733	 * the LRU instead of being evicted immediately so that if we do
 734	 * something to an unlinked inode, the irele won't cause
 735	 * premature truncation and freeing of the inode, which results
 736	 * in log recovery failure.  We have to evict the unreferenced
 737	 * lru inodes after clearing SB_ACTIVE because we don't
 738	 * otherwise clean up the lru if there's a subsequent failure in
 739	 * xfs_mountfs, which leads to us leaking the inodes if nothing
 740	 * else (e.g. quotacheck) references the inodes before the
 741	 * mount failure occurs.
 742	 */
 743	mp->m_super->s_flags |= SB_ACTIVE;
 744	xfs_log_work_queue(mp);
 745	if (xlog_recovery_needed(log))
 746		error = xlog_recover_finish(log);
 747	mp->m_super->s_flags &= ~SB_ACTIVE;
 748	evict_inodes(mp->m_super);
 749
 750	/*
 751	 * Drain the buffer LRU after log recovery. This is required for v4
 752	 * filesystems to avoid leaving around buffers with NULL verifier ops,
 753	 * but we do it unconditionally to make sure we're always in a clean
 754	 * cache state after mount.
 755	 *
 756	 * Don't push in the error case because the AIL may have pending intents
 757	 * that aren't removed until recovery is cancelled.
 758	 */
 759	if (xlog_recovery_needed(log)) {
 760		if (!error) {
 761			xfs_log_force(mp, XFS_LOG_SYNC);
 762			xfs_ail_push_all_sync(mp->m_ail);
 763		}
 764		xfs_notice(mp, "Ending recovery (logdev: %s)",
 765				mp->m_logname ? mp->m_logname : "internal");
 766	} else {
 767		xfs_info(mp, "Ending clean mount");
 768	}
 769	xfs_buftarg_drain(mp->m_ddev_targp);
 770
 771	clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
 772
 773	/* Make sure the log is dead if we're returning failure. */
 774	ASSERT(!error || xlog_is_shutdown(log));
 775
 776	return error;
 777}
 778
 779/*
 780 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 781 * the log.
 782 */
 783void
 784xfs_log_mount_cancel(
 785	struct xfs_mount	*mp)
 786{
 787	xlog_recover_cancel(mp->m_log);
 788	xfs_log_unmount(mp);
 789}
 790
 791/*
 792 * Flush out the iclog to disk ensuring that device caches are flushed and
 793 * the iclog hits stable storage before any completion waiters are woken.
 794 */
 795static inline int
 796xlog_force_iclog(
 797	struct xlog_in_core	*iclog)
 798{
 799	atomic_inc(&iclog->ic_refcnt);
 800	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
 801	if (iclog->ic_state == XLOG_STATE_ACTIVE)
 802		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
 803	return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
 804}
 805
 806/*
 807 * Cycle all the iclogbuf locks to make sure all log IO completion
 808 * is done before we tear down these buffers.
 809 */
 810static void
 811xlog_wait_iclog_completion(struct xlog *log)
 812{
 813	int		i;
 814	struct xlog_in_core	*iclog = log->l_iclog;
 815
 816	for (i = 0; i < log->l_iclog_bufs; i++) {
 817		down(&iclog->ic_sema);
 818		up(&iclog->ic_sema);
 819		iclog = iclog->ic_next;
 820	}
 821}
 822
 823/*
 824 * Wait for the iclog and all prior iclogs to be written disk as required by the
 825 * log force state machine. Waiting on ic_force_wait ensures iclog completions
 826 * have been ordered and callbacks run before we are woken here, hence
 827 * guaranteeing that all the iclogs up to this one are on stable storage.
 828 */
 829int
 830xlog_wait_on_iclog(
 831	struct xlog_in_core	*iclog)
 832		__releases(iclog->ic_log->l_icloglock)
 833{
 834	struct xlog		*log = iclog->ic_log;
 835
 836	trace_xlog_iclog_wait_on(iclog, _RET_IP_);
 837	if (!xlog_is_shutdown(log) &&
 838	    iclog->ic_state != XLOG_STATE_ACTIVE &&
 839	    iclog->ic_state != XLOG_STATE_DIRTY) {
 840		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
 841		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
 842	} else {
 843		spin_unlock(&log->l_icloglock);
 844	}
 845
 846	if (xlog_is_shutdown(log))
 847		return -EIO;
 848	return 0;
 849}
 850
 851/*
 852 * Write out an unmount record using the ticket provided. We have to account for
 853 * the data space used in the unmount ticket as this write is not done from a
 854 * transaction context that has already done the accounting for us.
 855 */
 856static int
 857xlog_write_unmount_record(
 858	struct xlog		*log,
 859	struct xlog_ticket	*ticket)
 860{
 861	struct  {
 862		struct xlog_op_header ophdr;
 863		struct xfs_unmount_log_format ulf;
 864	} unmount_rec = {
 865		.ophdr = {
 866			.oh_clientid = XFS_LOG,
 867			.oh_tid = cpu_to_be32(ticket->t_tid),
 868			.oh_flags = XLOG_UNMOUNT_TRANS,
 869		},
 870		.ulf = {
 871			.magic = XLOG_UNMOUNT_TYPE,
 872		},
 873	};
 874	struct xfs_log_iovec reg = {
 875		.i_addr = &unmount_rec,
 876		.i_len = sizeof(unmount_rec),
 877		.i_type = XLOG_REG_TYPE_UNMOUNT,
 878	};
 879	struct xfs_log_vec vec = {
 880		.lv_niovecs = 1,
 881		.lv_iovecp = &reg,
 882	};
 883	LIST_HEAD(lv_chain);
 884	list_add(&vec.lv_list, &lv_chain);
 885
 886	BUILD_BUG_ON((sizeof(struct xlog_op_header) +
 887		      sizeof(struct xfs_unmount_log_format)) !=
 888							sizeof(unmount_rec));
 889
 890	/* account for space used by record data */
 891	ticket->t_curr_res -= sizeof(unmount_rec);
 892
 893	return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
 894}
 895
 896/*
 897 * Mark the filesystem clean by writing an unmount record to the head of the
 898 * log.
 899 */
 900static void
 901xlog_unmount_write(
 902	struct xlog		*log)
 903{
 904	struct xfs_mount	*mp = log->l_mp;
 905	struct xlog_in_core	*iclog;
 906	struct xlog_ticket	*tic = NULL;
 907	int			error;
 908
 909	error = xfs_log_reserve(mp, 600, 1, &tic, 0);
 910	if (error)
 911		goto out_err;
 912
 913	error = xlog_write_unmount_record(log, tic);
 914	/*
 915	 * At this point, we're umounting anyway, so there's no point in
 916	 * transitioning log state to shutdown. Just continue...
 917	 */
 918out_err:
 919	if (error)
 920		xfs_alert(mp, "%s: unmount record failed", __func__);
 921
 922	spin_lock(&log->l_icloglock);
 923	iclog = log->l_iclog;
 924	error = xlog_force_iclog(iclog);
 925	xlog_wait_on_iclog(iclog);
 926
 927	if (tic) {
 928		trace_xfs_log_umount_write(log, tic);
 929		xfs_log_ticket_ungrant(log, tic);
 930	}
 931}
 932
 933static void
 934xfs_log_unmount_verify_iclog(
 935	struct xlog		*log)
 936{
 937	struct xlog_in_core	*iclog = log->l_iclog;
 938
 939	do {
 940		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 941		ASSERT(iclog->ic_offset == 0);
 942	} while ((iclog = iclog->ic_next) != log->l_iclog);
 943}
 944
 945/*
 946 * Unmount record used to have a string "Unmount filesystem--" in the
 947 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 948 * We just write the magic number now since that particular field isn't
 949 * currently architecture converted and "Unmount" is a bit foo.
 950 * As far as I know, there weren't any dependencies on the old behaviour.
 951 */
 952static void
 953xfs_log_unmount_write(
 954	struct xfs_mount	*mp)
 955{
 956	struct xlog		*log = mp->m_log;
 957
 958	if (!xfs_log_writable(mp))
 959		return;
 960
 961	xfs_log_force(mp, XFS_LOG_SYNC);
 962
 963	if (xlog_is_shutdown(log))
 964		return;
 965
 966	/*
 967	 * If we think the summary counters are bad, avoid writing the unmount
 968	 * record to force log recovery at next mount, after which the summary
 969	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
 970	 * more details.
 971	 */
 972	if (xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS) ||
 973	    XFS_TEST_ERROR(mp, XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
 974		xfs_alert(mp, "%s: will fix summary counters at next mount",
 975				__func__);
 976		return;
 977	}
 978
 979	xfs_log_unmount_verify_iclog(log);
 980	xlog_unmount_write(log);
 981}
 982
 983/*
 984 * Empty the log for unmount/freeze.
 985 *
 986 * To do this, we first need to shut down the background log work so it is not
 987 * trying to cover the log as we clean up. We then need to unpin all objects in
 988 * the log so we can then flush them out. Once they have completed their IO and
 989 * run the callbacks removing themselves from the AIL, we can cover the log.
 990 */
 991int
 992xfs_log_quiesce(
 993	struct xfs_mount	*mp)
 994{
 995	/*
 996	 * Clear log incompat features since we're quiescing the log.  Report
 997	 * failures, though it's not fatal to have a higher log feature
 998	 * protection level than the log contents actually require.
 999	 */
1000	if (xfs_clear_incompat_log_features(mp)) {
1001		int error;
1002
1003		error = xfs_sync_sb(mp, false);
1004		if (error)
1005			xfs_warn(mp,
1006	"Failed to clear log incompat features on quiesce");
1007	}
1008
1009	cancel_delayed_work_sync(&mp->m_log->l_work);
1010	xfs_log_force(mp, XFS_LOG_SYNC);
1011
1012	/*
1013	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1014	 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1015	 * xfs_buf_iowait() cannot be used because it was pushed with the
1016	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1017	 * the IO to complete.
1018	 */
1019	xfs_ail_push_all_sync(mp->m_ail);
1020	xfs_buftarg_wait(mp->m_ddev_targp);
1021	xfs_buf_lock(mp->m_sb_bp);
1022	xfs_buf_unlock(mp->m_sb_bp);
1023
1024	return xfs_log_cover(mp);
1025}
1026
1027void
1028xfs_log_clean(
1029	struct xfs_mount	*mp)
1030{
1031	xfs_log_quiesce(mp);
1032	xfs_log_unmount_write(mp);
1033}
1034
1035/*
1036 * Shut down and release the AIL and Log.
1037 *
1038 * During unmount, we need to ensure we flush all the dirty metadata objects
1039 * from the AIL so that the log is empty before we write the unmount record to
1040 * the log. Once this is done, we can tear down the AIL and the log.
1041 */
1042void
1043xfs_log_unmount(
1044	struct xfs_mount	*mp)
1045{
1046	xfs_log_clean(mp);
1047
1048	/*
1049	 * If shutdown has come from iclog IO context, the log
1050	 * cleaning will have been skipped and so we need to wait
1051	 * for the iclog to complete shutdown processing before we
1052	 * tear anything down.
1053	 */
1054	xlog_wait_iclog_completion(mp->m_log);
1055
1056	xfs_buftarg_drain(mp->m_ddev_targp);
1057
1058	xfs_trans_ail_destroy(mp);
1059
1060	xfs_sysfs_del(&mp->m_log->l_kobj);
1061
1062	xlog_dealloc_log(mp->m_log);
1063}
1064
1065void
1066xfs_log_item_init(
1067	struct xfs_mount	*mp,
1068	struct xfs_log_item	*item,
1069	int			type,
1070	const struct xfs_item_ops *ops)
1071{
1072	item->li_log = mp->m_log;
1073	item->li_ailp = mp->m_ail;
1074	item->li_type = type;
1075	item->li_ops = ops;
1076	item->li_lv = NULL;
1077
1078	INIT_LIST_HEAD(&item->li_ail);
1079	INIT_LIST_HEAD(&item->li_cil);
1080	INIT_LIST_HEAD(&item->li_bio_list);
1081	INIT_LIST_HEAD(&item->li_trans);
1082}
1083
1084/*
1085 * Wake up processes waiting for log space after we have moved the log tail.
1086 */
1087void
1088xfs_log_space_wake(
1089	struct xfs_mount	*mp)
1090{
1091	struct xlog		*log = mp->m_log;
1092	int			free_bytes;
1093
1094	if (xlog_is_shutdown(log))
1095		return;
1096
1097	if (!list_empty_careful(&log->l_write_head.waiters)) {
1098		ASSERT(!xlog_in_recovery(log));
1099
1100		spin_lock(&log->l_write_head.lock);
1101		free_bytes = xlog_grant_space_left(log, &log->l_write_head);
1102		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1103		spin_unlock(&log->l_write_head.lock);
1104	}
1105
1106	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1107		ASSERT(!xlog_in_recovery(log));
1108
1109		spin_lock(&log->l_reserve_head.lock);
1110		free_bytes = xlog_grant_space_left(log, &log->l_reserve_head);
1111		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1112		spin_unlock(&log->l_reserve_head.lock);
1113	}
1114}
1115
1116/*
1117 * Determine if we have a transaction that has gone to disk that needs to be
1118 * covered. To begin the transition to the idle state firstly the log needs to
1119 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1120 * we start attempting to cover the log.
1121 *
1122 * Only if we are then in a state where covering is needed, the caller is
1123 * informed that dummy transactions are required to move the log into the idle
1124 * state.
1125 *
1126 * If there are any items in the AIl or CIL, then we do not want to attempt to
1127 * cover the log as we may be in a situation where there isn't log space
1128 * available to run a dummy transaction and this can lead to deadlocks when the
1129 * tail of the log is pinned by an item that is modified in the CIL.  Hence
1130 * there's no point in running a dummy transaction at this point because we
1131 * can't start trying to idle the log until both the CIL and AIL are empty.
1132 */
1133static bool
1134xfs_log_need_covered(
1135	struct xfs_mount	*mp)
1136{
1137	struct xlog		*log = mp->m_log;
1138	bool			needed = false;
1139
1140	if (!xlog_cil_empty(log))
1141		return false;
1142
1143	spin_lock(&log->l_icloglock);
1144	switch (log->l_covered_state) {
1145	case XLOG_STATE_COVER_DONE:
1146	case XLOG_STATE_COVER_DONE2:
1147	case XLOG_STATE_COVER_IDLE:
1148		break;
1149	case XLOG_STATE_COVER_NEED:
1150	case XLOG_STATE_COVER_NEED2:
1151		if (xfs_ail_min_lsn(log->l_ailp))
1152			break;
1153		if (!xlog_iclogs_empty(log))
1154			break;
1155
1156		needed = true;
1157		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1158			log->l_covered_state = XLOG_STATE_COVER_DONE;
1159		else
1160			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1161		break;
1162	default:
1163		needed = true;
1164		break;
1165	}
1166	spin_unlock(&log->l_icloglock);
1167	return needed;
1168}
1169
1170/*
1171 * Explicitly cover the log. This is similar to background log covering but
1172 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1173 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1174 * must all be empty.
1175 */
1176static int
1177xfs_log_cover(
1178	struct xfs_mount	*mp)
1179{
1180	int			error = 0;
1181	bool			need_covered;
1182
1183	if (!xlog_is_shutdown(mp->m_log)) {
1184		ASSERT(xlog_cil_empty(mp->m_log));
1185		ASSERT(xlog_iclogs_empty(mp->m_log));
1186		ASSERT(!xfs_ail_min_lsn(mp->m_log->l_ailp));
1187	}
1188
1189	if (!xfs_log_writable(mp))
1190		return 0;
1191
1192	/*
1193	 * xfs_log_need_covered() is not idempotent because it progresses the
1194	 * state machine if the log requires covering. Therefore, we must call
1195	 * this function once and use the result until we've issued an sb sync.
1196	 * Do so first to make that abundantly clear.
1197	 *
1198	 * Fall into the covering sequence if the log needs covering or the
1199	 * mount has lazy superblock accounting to sync to disk. The sb sync
1200	 * used for covering accumulates the in-core counters, so covering
1201	 * handles this for us.
1202	 */
1203	need_covered = xfs_log_need_covered(mp);
1204	if (!need_covered && !xfs_has_lazysbcount(mp))
1205		return 0;
1206
1207	/*
1208	 * To cover the log, commit the superblock twice (at most) in
1209	 * independent checkpoints. The first serves as a reference for the
1210	 * tail pointer. The sync transaction and AIL push empties the AIL and
1211	 * updates the in-core tail to the LSN of the first checkpoint. The
1212	 * second commit updates the on-disk tail with the in-core LSN,
1213	 * covering the log. Push the AIL one more time to leave it empty, as
1214	 * we found it.
1215	 */
1216	do {
1217		error = xfs_sync_sb(mp, true);
1218		if (error)
1219			break;
1220		xfs_ail_push_all_sync(mp->m_ail);
1221	} while (xfs_log_need_covered(mp));
1222
1223	return error;
1224}
1225
1226static void
1227xlog_ioend_work(
1228	struct work_struct	*work)
1229{
1230	struct xlog_in_core     *iclog =
1231		container_of(work, struct xlog_in_core, ic_end_io_work);
1232	struct xlog		*log = iclog->ic_log;
1233	int			error;
1234
1235	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1236#ifdef DEBUG
1237	/* treat writes with injected CRC errors as failed */
1238	if (iclog->ic_fail_crc)
1239		error = -EIO;
1240#endif
1241
1242	/*
1243	 * Race to shutdown the filesystem if we see an error.
1244	 */
1245	if (error || XFS_TEST_ERROR(log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1246		xfs_alert(log->l_mp, "log I/O error %d", error);
1247		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1248	}
1249
1250	xlog_state_done_syncing(iclog);
1251	bio_uninit(&iclog->ic_bio);
1252
1253	/*
1254	 * Drop the lock to signal that we are done. Nothing references the
1255	 * iclog after this, so an unmount waiting on this lock can now tear it
1256	 * down safely. As such, it is unsafe to reference the iclog after the
1257	 * unlock as we could race with it being freed.
1258	 */
1259	up(&iclog->ic_sema);
1260}
1261
1262/*
1263 * Return size of each in-core log record buffer.
1264 *
1265 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1266 *
1267 * If the filesystem blocksize is too large, we may need to choose a
1268 * larger size since the directory code currently logs entire blocks.
1269 */
1270STATIC void
1271xlog_get_iclog_buffer_size(
1272	struct xfs_mount	*mp,
1273	struct xlog		*log)
1274{
1275	if (mp->m_logbufs <= 0)
1276		mp->m_logbufs = XLOG_MAX_ICLOGS;
1277	if (mp->m_logbsize <= 0)
1278		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1279
1280	log->l_iclog_bufs = mp->m_logbufs;
1281	log->l_iclog_size = mp->m_logbsize;
1282
1283	/*
1284	 * Combined size of the log record headers.  The first 32k cycles
1285	 * are stored directly in the xlog_rec_header, the rest in the
1286	 * variable number of xlog_rec_ext_headers at its end.
1287	 */
1288	log->l_iclog_hsize = struct_size(log->l_iclog->ic_header, h_ext,
1289		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE) - 1);
1290}
1291
1292void
1293xfs_log_work_queue(
1294	struct xfs_mount        *mp)
1295{
1296	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1297				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1298}
1299
1300/*
1301 * Clear the log incompat flags if we have the opportunity.
1302 *
1303 * This only happens if we're about to log the second dummy transaction as part
1304 * of covering the log.
1305 */
1306static inline void
1307xlog_clear_incompat(
1308	struct xlog		*log)
1309{
1310	struct xfs_mount	*mp = log->l_mp;
1311
1312	if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1313				XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1314		return;
1315
1316	if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1317		return;
1318
1319	xfs_clear_incompat_log_features(mp);
1320}
1321
1322/*
1323 * Every sync period we need to unpin all items in the AIL and push them to
1324 * disk. If there is nothing dirty, then we might need to cover the log to
1325 * indicate that the filesystem is idle.
1326 */
1327static void
1328xfs_log_worker(
1329	struct work_struct	*work)
1330{
1331	struct xlog		*log = container_of(to_delayed_work(work),
1332						struct xlog, l_work);
1333	struct xfs_mount	*mp = log->l_mp;
1334
1335	/* dgc: errors ignored - not fatal and nowhere to report them */
1336	if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1337		/*
1338		 * Dump a transaction into the log that contains no real change.
1339		 * This is needed to stamp the current tail LSN into the log
1340		 * during the covering operation.
1341		 *
1342		 * We cannot use an inode here for this - that will push dirty
1343		 * state back up into the VFS and then periodic inode flushing
1344		 * will prevent log covering from making progress. Hence we
1345		 * synchronously log the superblock instead to ensure the
1346		 * superblock is immediately unpinned and can be written back.
1347		 */
1348		xlog_clear_incompat(log);
1349		xfs_sync_sb(mp, true);
1350	} else
1351		xfs_log_force(mp, 0);
1352
1353	/* start pushing all the metadata that is currently dirty */
1354	xfs_ail_push_all(mp->m_ail);
1355
1356	/* queue us up again */
1357	xfs_log_work_queue(mp);
1358}
1359
1360/*
1361 * This routine initializes some of the log structure for a given mount point.
1362 * Its primary purpose is to fill in enough, so recovery can occur.  However,
1363 * some other stuff may be filled in too.
1364 */
1365STATIC struct xlog *
1366xlog_alloc_log(
1367	struct xfs_mount	*mp,
1368	struct xfs_buftarg	*log_target,
1369	xfs_daddr_t		blk_offset,
1370	int			num_bblks)
1371{
1372	struct xlog		*log;
1373	struct xlog_in_core	**iclogp;
1374	struct xlog_in_core	*iclog, *prev_iclog = NULL;
1375	int			i;
1376	int			error = -ENOMEM;
1377	uint			log2_size = 0;
1378
1379	log = kzalloc(sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1380	if (!log) {
1381		xfs_warn(mp, "Log allocation failed: No memory!");
1382		goto out;
1383	}
1384
1385	log->l_mp	   = mp;
1386	log->l_targ	   = log_target;
1387	log->l_logsize     = BBTOB(num_bblks);
1388	log->l_logBBstart  = blk_offset;
1389	log->l_logBBsize   = num_bblks;
1390	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1391	set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1392	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1393	INIT_LIST_HEAD(&log->r_dfops);
1394
1395	log->l_prev_block  = -1;
1396	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1397	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1398	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1399
1400	if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1401		log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1402	else if (mp->m_sb.sb_logsectsize > 0)
1403		log->l_iclog_roundoff = mp->m_sb.sb_logsectsize;
1404	else
1405		log->l_iclog_roundoff = BBSIZE;
1406
1407	xlog_grant_head_init(&log->l_reserve_head);
1408	xlog_grant_head_init(&log->l_write_head);
1409
1410	error = -EFSCORRUPTED;
1411	if (xfs_has_sector(mp)) {
1412	        log2_size = mp->m_sb.sb_logsectlog;
1413		if (log2_size < BBSHIFT) {
1414			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1415				log2_size, BBSHIFT);
1416			goto out_free_log;
1417		}
1418
1419	        log2_size -= BBSHIFT;
1420		if (log2_size > mp->m_sectbb_log) {
1421			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1422				log2_size, mp->m_sectbb_log);
1423			goto out_free_log;
1424		}
1425
1426		/* for larger sector sizes, must have v2 or external log */
1427		if (log2_size && log->l_logBBstart > 0 &&
1428			    !xfs_has_logv2(mp)) {
1429			xfs_warn(mp,
1430		"log sector size (0x%x) invalid for configuration.",
1431				log2_size);
1432			goto out_free_log;
1433		}
1434	}
1435	log->l_sectBBsize = 1 << log2_size;
1436
1437	xlog_get_iclog_buffer_size(mp, log);
1438
1439	spin_lock_init(&log->l_icloglock);
1440	init_waitqueue_head(&log->l_flush_wait);
1441
1442	iclogp = &log->l_iclog;
1443	ASSERT(log->l_iclog_size >= 4096);
1444	for (i = 0; i < log->l_iclog_bufs; i++) {
1445		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1446				sizeof(struct bio_vec);
1447
1448		iclog = kzalloc(sizeof(*iclog) + bvec_size,
1449				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1450		if (!iclog)
1451			goto out_free_iclog;
1452
1453		*iclogp = iclog;
1454		iclog->ic_prev = prev_iclog;
1455		prev_iclog = iclog;
1456
1457		iclog->ic_header = kvzalloc(log->l_iclog_size,
1458				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1459		if (!iclog->ic_header)
1460			goto out_free_iclog;
1461		iclog->ic_header->h_magicno =
1462			cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1463		iclog->ic_header->h_version = cpu_to_be32(
1464			xfs_has_logv2(log->l_mp) ? 2 : 1);
1465		iclog->ic_header->h_size = cpu_to_be32(log->l_iclog_size);
1466		iclog->ic_header->h_fmt = cpu_to_be32(XLOG_FMT);
1467		memcpy(&iclog->ic_header->h_fs_uuid, &mp->m_sb.sb_uuid,
1468			sizeof(iclog->ic_header->h_fs_uuid));
1469
1470		iclog->ic_datap = (void *)iclog->ic_header + log->l_iclog_hsize;
1471		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1472		iclog->ic_state = XLOG_STATE_ACTIVE;
1473		iclog->ic_log = log;
1474		atomic_set(&iclog->ic_refcnt, 0);
1475		INIT_LIST_HEAD(&iclog->ic_callbacks);
1476
1477		init_waitqueue_head(&iclog->ic_force_wait);
1478		init_waitqueue_head(&iclog->ic_write_wait);
1479		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1480		sema_init(&iclog->ic_sema, 1);
1481
1482		iclogp = &iclog->ic_next;
1483	}
1484	*iclogp = log->l_iclog;			/* complete ring */
1485	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1486
1487	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1488			XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_PERCPU),
1489			0, mp->m_super->s_id);
1490	if (!log->l_ioend_workqueue)
1491		goto out_free_iclog;
1492
1493	error = xlog_cil_init(log);
1494	if (error)
1495		goto out_destroy_workqueue;
1496	return log;
1497
1498out_destroy_workqueue:
1499	destroy_workqueue(log->l_ioend_workqueue);
1500out_free_iclog:
1501	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1502		prev_iclog = iclog->ic_next;
1503		kvfree(iclog->ic_header);
1504		kfree(iclog);
1505		if (prev_iclog == log->l_iclog)
1506			break;
1507	}
1508out_free_log:
1509	kfree(log);
1510out:
1511	return ERR_PTR(error);
1512}	/* xlog_alloc_log */
1513
1514/*
1515 * Stamp cycle number in every block
1516 */
1517STATIC void
1518xlog_pack_data(
1519	struct xlog		*log,
1520	struct xlog_in_core	*iclog,
1521	int			roundoff)
1522{
1523	struct xlog_rec_header	*rhead = iclog->ic_header;
1524	__be32			cycle_lsn = CYCLE_LSN_DISK(rhead->h_lsn);
1525	char			*dp = iclog->ic_datap;
1526	int			i;
1527
1528	for (i = 0; i < BTOBB(iclog->ic_offset + roundoff); i++) {
1529		*xlog_cycle_data(rhead, i) = *(__be32 *)dp;
1530		*(__be32 *)dp = cycle_lsn;
1531		dp += BBSIZE;
1532	}
1533
1534	for (i = 0; i < (log->l_iclog_hsize >> BBSHIFT) - 1; i++)
1535		rhead->h_ext[i].xh_cycle = cycle_lsn;
1536}
1537
1538/*
1539 * Calculate the checksum for a log buffer.
1540 *
1541 * This is a little more complicated than it should be because the various
1542 * headers and the actual data are non-contiguous.
1543 */
1544__le32
1545xlog_cksum(
1546	struct xlog		*log,
1547	struct xlog_rec_header	*rhead,
1548	char			*dp,
1549	unsigned int		hdrsize,
1550	unsigned int		size)
1551{
1552	uint32_t		crc;
1553
1554	/* first generate the crc for the record header ... */
1555	crc = xfs_start_cksum_update((char *)rhead, hdrsize,
1556			      offsetof(struct xlog_rec_header, h_crc));
1557
1558	/* ... then for additional cycle data for v2 logs ... */
1559	if (xfs_has_logv2(log->l_mp)) {
1560		int		xheads, i;
1561
1562		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE) - 1;
1563		for (i = 0; i < xheads; i++)
1564			crc = crc32c(crc, &rhead->h_ext[i], XLOG_REC_EXT_SIZE);
1565	}
1566
1567	/* ... and finally for the payload */
1568	crc = crc32c(crc, dp, size);
1569
1570	return xfs_end_cksum(crc);
1571}
1572
1573static void
1574xlog_bio_end_io(
1575	struct bio		*bio)
1576{
1577	struct xlog_in_core	*iclog = bio->bi_private;
1578
1579	queue_work(iclog->ic_log->l_ioend_workqueue,
1580		   &iclog->ic_end_io_work);
1581}
1582
1583STATIC void
1584xlog_write_iclog(
1585	struct xlog		*log,
1586	struct xlog_in_core	*iclog,
1587	uint64_t		bno,
1588	unsigned int		count)
1589{
1590	ASSERT(bno < log->l_logBBsize);
1591	trace_xlog_iclog_write(iclog, _RET_IP_);
1592
1593	/*
1594	 * We lock the iclogbufs here so that we can serialise against I/O
1595	 * completion during unmount.  We might be processing a shutdown
1596	 * triggered during unmount, and that can occur asynchronously to the
1597	 * unmount thread, and hence we need to ensure that completes before
1598	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1599	 * across the log IO to archieve that.
1600	 */
1601	down(&iclog->ic_sema);
1602	if (xlog_is_shutdown(log)) {
1603		/*
1604		 * It would seem logical to return EIO here, but we rely on
1605		 * the log state machine to propagate I/O errors instead of
1606		 * doing it here.  We kick of the state machine and unlock
1607		 * the buffer manually, the code needs to be kept in sync
1608		 * with the I/O completion path.
1609		 */
1610		goto sync;
1611	}
1612
1613	/*
1614	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1615	 * IOs coming immediately after this one. This prevents the block layer
1616	 * writeback throttle from throttling log writes behind background
1617	 * metadata writeback and causing priority inversions.
1618	 */
1619	bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1620		 howmany(count, PAGE_SIZE),
1621		 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1622	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1623	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1624	iclog->ic_bio.bi_private = iclog;
1625
1626	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1627		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1628		/*
1629		 * For external log devices, we also need to flush the data
1630		 * device cache first to ensure all metadata writeback covered
1631		 * by the LSN in this iclog is on stable storage. This is slow,
1632		 * but it *must* complete before we issue the external log IO.
1633		 *
1634		 * If the flush fails, we cannot conclude that past metadata
1635		 * writeback from the log succeeded.  Repeating the flush is
1636		 * not possible, hence we must shut down with log IO error to
1637		 * avoid shutdown re-entering this path and erroring out again.
1638		 */
1639		if (log->l_targ != log->l_mp->m_ddev_targp &&
1640		    blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev))
1641			goto shutdown;
1642	}
1643	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1644		iclog->ic_bio.bi_opf |= REQ_FUA;
1645
1646	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1647
1648	if (is_vmalloc_addr(iclog->ic_header)) {
1649		if (!bio_add_vmalloc(&iclog->ic_bio, iclog->ic_header, count))
1650			goto shutdown;
1651	} else {
1652		bio_add_virt_nofail(&iclog->ic_bio, iclog->ic_header, count);
1653	}
1654
1655	/*
1656	 * If this log buffer would straddle the end of the log we will have
1657	 * to split it up into two bios, so that we can continue at the start.
1658	 */
1659	if (bno + BTOBB(count) > log->l_logBBsize) {
1660		struct bio *split;
1661
1662		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1663				  GFP_NOIO, &fs_bio_set);
1664		bio_chain(split, &iclog->ic_bio);
1665		submit_bio(split);
1666
1667		/* restart at logical offset zero for the remainder */
1668		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1669	}
1670
1671	submit_bio(&iclog->ic_bio);
1672	return;
1673shutdown:
1674	xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1675sync:
1676	xlog_state_done_syncing(iclog);
1677	up(&iclog->ic_sema);
1678}
1679
1680/*
1681 * We need to bump cycle number for the part of the iclog that is
1682 * written to the start of the log. Watch out for the header magic
1683 * number case, though.
1684 */
1685static void
1686xlog_split_iclog(
1687	struct xlog		*log,
1688	void			*data,
1689	uint64_t		bno,
1690	unsigned int		count)
1691{
1692	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1693	unsigned int		i;
1694
1695	for (i = split_offset; i < count; i += BBSIZE) {
1696		uint32_t cycle = get_unaligned_be32(data + i);
1697
1698		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1699			cycle++;
1700		put_unaligned_be32(cycle, data + i);
1701	}
1702}
1703
1704static int
1705xlog_calc_iclog_size(
1706	struct xlog		*log,
1707	struct xlog_in_core	*iclog,
1708	uint32_t		*roundoff)
1709{
1710	uint32_t		count_init, count;
1711
1712	/* Add for LR header */
1713	count_init = log->l_iclog_hsize + iclog->ic_offset;
1714	count = roundup(count_init, log->l_iclog_roundoff);
1715
1716	*roundoff = count - count_init;
1717
1718	ASSERT(count >= count_init);
1719	ASSERT(*roundoff < log->l_iclog_roundoff);
1720	return count;
1721}
1722
1723/*
1724 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1725 * fashion.  Previously, we should have moved the current iclog
1726 * ptr in the log to point to the next available iclog.  This allows further
1727 * write to continue while this code syncs out an iclog ready to go.
1728 * Before an in-core log can be written out, the data section must be scanned
1729 * to save away the 1st word of each BBSIZE block into the header.  We replace
1730 * it with the current cycle count.  Each BBSIZE block is tagged with the
1731 * cycle count because there in an implicit assumption that drives will
1732 * guarantee that entire 512 byte blocks get written at once.  In other words,
1733 * we can't have part of a 512 byte block written and part not written.  By
1734 * tagging each block, we will know which blocks are valid when recovering
1735 * after an unclean shutdown.
1736 *
1737 * This routine is single threaded on the iclog.  No other thread can be in
1738 * this routine with the same iclog.  Changing contents of iclog can there-
1739 * fore be done without grabbing the state machine lock.  Updating the global
1740 * log will require grabbing the lock though.
1741 *
1742 * The entire log manager uses a logical block numbering scheme.  Only
1743 * xlog_write_iclog knows about the fact that the log may not start with
1744 * block zero on a given device.
1745 */
1746STATIC void
1747xlog_sync(
1748	struct xlog		*log,
1749	struct xlog_in_core	*iclog,
1750	struct xlog_ticket	*ticket)
1751{
1752	unsigned int		count;		/* byte count of bwrite */
1753	unsigned int		roundoff;       /* roundoff to BB or stripe */
1754	uint64_t		bno;
1755	unsigned int		size;
1756
1757	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1758	trace_xlog_iclog_sync(iclog, _RET_IP_);
1759
1760	count = xlog_calc_iclog_size(log, iclog, &roundoff);
1761
1762	/*
1763	 * If we have a ticket, account for the roundoff via the ticket
1764	 * reservation to avoid touching the hot grant heads needlessly.
1765	 * Otherwise, we have to move grant heads directly.
1766	 */
1767	if (ticket) {
1768		ticket->t_curr_res -= roundoff;
1769	} else {
1770		xlog_grant_add_space(&log->l_reserve_head, roundoff);
1771		xlog_grant_add_space(&log->l_write_head, roundoff);
1772	}
1773
1774	/* put cycle number in every block */
1775	xlog_pack_data(log, iclog, roundoff);
1776
1777	/* real byte length */
1778	size = iclog->ic_offset;
1779	if (xfs_has_logv2(log->l_mp))
1780		size += roundoff;
1781	iclog->ic_header->h_len = cpu_to_be32(size);
1782
1783	XFS_STATS_INC(log->l_mp, xs_log_writes);
1784	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1785
1786	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header->h_lsn));
1787
1788	/* Do we need to split this write into 2 parts? */
1789	if (bno + BTOBB(count) > log->l_logBBsize)
1790		xlog_split_iclog(log, iclog->ic_header, bno, count);
1791
1792	/* calculcate the checksum */
1793	iclog->ic_header->h_crc = xlog_cksum(log, iclog->ic_header,
1794			iclog->ic_datap, XLOG_REC_SIZE, size);
1795	/*
1796	 * Intentionally corrupt the log record CRC based on the error injection
1797	 * frequency, if defined. This facilitates testing log recovery in the
1798	 * event of torn writes. Hence, set the IOABORT state to abort the log
1799	 * write on I/O completion and shutdown the fs. The subsequent mount
1800	 * detects the bad CRC and attempts to recover.
1801	 */
1802#ifdef DEBUG
1803	if (XFS_TEST_ERROR(log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1804		iclog->ic_header->h_crc &= cpu_to_le32(0xAAAAAAAA);
1805		iclog->ic_fail_crc = true;
1806		xfs_warn(log->l_mp,
1807	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1808			 be64_to_cpu(iclog->ic_header->h_lsn));
1809	}
1810#endif
1811	xlog_verify_iclog(log, iclog, count);
1812	xlog_write_iclog(log, iclog, bno, count);
1813}
1814
1815/*
1816 * Deallocate a log structure
1817 */
1818STATIC void
1819xlog_dealloc_log(
1820	struct xlog		*log)
1821{
1822	struct xlog_in_core	*iclog, *next_iclog;
1823	int			i;
1824
1825	/*
1826	 * Destroy the CIL after waiting for iclog IO completion because an
1827	 * iclog EIO error will try to shut down the log, which accesses the
1828	 * CIL to wake up the waiters.
1829	 */
1830	xlog_cil_destroy(log);
1831
1832	iclog = log->l_iclog;
1833	for (i = 0; i < log->l_iclog_bufs; i++) {
1834		next_iclog = iclog->ic_next;
1835		kvfree(iclog->ic_header);
1836		kfree(iclog);
1837		iclog = next_iclog;
1838	}
1839
1840	log->l_mp->m_log = NULL;
1841	destroy_workqueue(log->l_ioend_workqueue);
1842	kfree(log);
1843}
1844
1845/*
1846 * Update counters atomically now that memcpy is done.
1847 */
1848static inline void
1849xlog_state_finish_copy(
1850	struct xlog		*log,
1851	struct xlog_in_core	*iclog,
1852	int			record_cnt,
1853	int			copy_bytes)
1854{
1855	lockdep_assert_held(&log->l_icloglock);
1856
1857	be32_add_cpu(&iclog->ic_header->h_num_logops, record_cnt);
1858	iclog->ic_offset += copy_bytes;
1859}
1860
1861/*
1862 * print out info relating to regions written which consume
1863 * the reservation
1864 */
1865void
1866xlog_print_tic_res(
1867	struct xfs_mount	*mp,
1868	struct xlog_ticket	*ticket)
1869{
1870	xfs_warn(mp, "ticket reservation summary:");
1871	xfs_warn(mp, "  unit res    = %d bytes", ticket->t_unit_res);
1872	xfs_warn(mp, "  current res = %d bytes", ticket->t_curr_res);
1873	xfs_warn(mp, "  original count  = %d", ticket->t_ocnt);
1874	xfs_warn(mp, "  remaining count = %d", ticket->t_cnt);
1875}
1876
1877/*
1878 * Print a summary of the transaction.
1879 */
1880void
1881xlog_print_trans(
1882	struct xfs_trans	*tp)
1883{
1884	struct xfs_mount	*mp = tp->t_mountp;
1885	struct xfs_log_item	*lip;
1886
1887	/* dump core transaction and ticket info */
1888	xfs_warn(mp, "transaction summary:");
1889	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
1890	xfs_warn(mp, "  log count = %d", tp->t_log_count);
1891	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
1892
1893	xlog_print_tic_res(mp, tp->t_ticket);
1894
1895	/* dump each log item */
1896	list_for_each_entry(lip, &tp->t_items, li_trans) {
1897		struct xfs_log_vec	*lv = lip->li_lv;
1898		struct xfs_log_iovec	*vec;
1899		int			i;
1900
1901		xfs_warn(mp, "log item: ");
1902		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
1903		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
1904		if (!lv)
1905			continue;
1906		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
1907		xfs_warn(mp, "  alloc_size = %d", lv->lv_alloc_size);
1908		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
1909		xfs_warn(mp, "  buf used= %d", lv->lv_buf_used);
1910
1911		/* dump each iovec for the log item */
1912		vec = lv->lv_iovecp;
1913		for (i = 0; i < lv->lv_niovecs; i++) {
1914			int dumplen = min(vec->i_len, 32);
1915
1916			xfs_warn(mp, "  iovec[%d]", i);
1917			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
1918			xfs_warn(mp, "    len	= %d", vec->i_len);
1919			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
1920			xfs_hex_dump(vec->i_addr, dumplen);
1921
1922			vec++;
1923		}
1924	}
1925}
1926
1927static inline void
1928xlog_write_iovec(
1929	struct xlog_in_core	*iclog,
1930	uint32_t		*log_offset,
1931	void			*data,
1932	uint32_t		write_len,
1933	int			*bytes_left,
1934	uint32_t		*record_cnt,
1935	uint32_t		*data_cnt)
1936{
1937	ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
1938	ASSERT(*log_offset % sizeof(int32_t) == 0);
1939	ASSERT(write_len % sizeof(int32_t) == 0);
1940
1941	memcpy(iclog->ic_datap + *log_offset, data, write_len);
1942	*log_offset += write_len;
1943	*bytes_left -= write_len;
1944	(*record_cnt)++;
1945	*data_cnt += write_len;
1946}
1947
1948/*
1949 * Write log vectors into a single iclog which is guaranteed by the caller
1950 * to have enough space to write the entire log vector into.
1951 */
1952static void
1953xlog_write_full(
1954	struct xfs_log_vec	*lv,
1955	struct xlog_ticket	*ticket,
1956	struct xlog_in_core	*iclog,
1957	uint32_t		*log_offset,
1958	uint32_t		*len,
1959	uint32_t		*record_cnt,
1960	uint32_t		*data_cnt)
1961{
1962	int			index;
1963
1964	ASSERT(*log_offset + *len <= iclog->ic_size ||
1965		iclog->ic_state == XLOG_STATE_WANT_SYNC);
1966
1967	/*
1968	 * Ordered log vectors have no regions to write so this
1969	 * loop will naturally skip them.
1970	 */
1971	for (index = 0; index < lv->lv_niovecs; index++) {
1972		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
1973		struct xlog_op_header	*ophdr = reg->i_addr;
1974
1975		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1976		xlog_write_iovec(iclog, log_offset, reg->i_addr,
1977				reg->i_len, len, record_cnt, data_cnt);
1978	}
1979}
1980
1981static int
1982xlog_write_get_more_iclog_space(
1983	struct xlog_ticket	*ticket,
1984	struct xlog_in_core	**iclogp,
1985	uint32_t		*log_offset,
1986	uint32_t		len,
1987	uint32_t		*record_cnt,
1988	uint32_t		*data_cnt)
1989{
1990	struct xlog_in_core	*iclog = *iclogp;
1991	struct xlog		*log = iclog->ic_log;
1992	int			error;
1993
1994	spin_lock(&log->l_icloglock);
1995	ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
1996	xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1997	error = xlog_state_release_iclog(log, iclog, ticket);
1998	spin_unlock(&log->l_icloglock);
1999	if (error)
2000		return error;
2001
2002	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2003					log_offset);
2004	if (error)
2005		return error;
2006	*record_cnt = 0;
2007	*data_cnt = 0;
2008	*iclogp = iclog;
2009	return 0;
2010}
2011
2012/*
2013 * Write log vectors into a single iclog which is smaller than the current chain
2014 * length. We write until we cannot fit a full record into the remaining space
2015 * and then stop. We return the log vector that is to be written that cannot
2016 * wholly fit in the iclog.
2017 */
2018static int
2019xlog_write_partial(
2020	struct xfs_log_vec	*lv,
2021	struct xlog_ticket	*ticket,
2022	struct xlog_in_core	**iclogp,
2023	uint32_t		*log_offset,
2024	uint32_t		*len,
2025	uint32_t		*record_cnt,
2026	uint32_t		*data_cnt)
2027{
2028	struct xlog_in_core	*iclog = *iclogp;
2029	struct xlog_op_header	*ophdr;
2030	int			index = 0;
2031	uint32_t		rlen;
2032	int			error;
2033
2034	/* walk the logvec, copying until we run out of space in the iclog */
2035	for (index = 0; index < lv->lv_niovecs; index++) {
2036		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2037		uint32_t		reg_offset = 0;
2038
2039		/*
2040		 * The first region of a continuation must have a non-zero
2041		 * length otherwise log recovery will just skip over it and
2042		 * start recovering from the next opheader it finds. Because we
2043		 * mark the next opheader as a continuation, recovery will then
2044		 * incorrectly add the continuation to the previous region and
2045		 * that breaks stuff.
2046		 *
2047		 * Hence if there isn't space for region data after the
2048		 * opheader, then we need to start afresh with a new iclog.
2049		 */
2050		if (iclog->ic_size - *log_offset <=
2051					sizeof(struct xlog_op_header)) {
2052			error = xlog_write_get_more_iclog_space(ticket,
2053					&iclog, log_offset, *len, record_cnt,
2054					data_cnt);
2055			if (error)
2056				return error;
2057		}
2058
2059		ophdr = reg->i_addr;
2060		rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2061
2062		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2063		ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2064		if (rlen != reg->i_len)
2065			ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2066
2067		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2068				rlen, len, record_cnt, data_cnt);
2069
2070		/* If we wrote the whole region, move to the next. */
2071		if (rlen == reg->i_len)
2072			continue;
2073
2074		/*
2075		 * We now have a partially written iovec, but it can span
2076		 * multiple iclogs so we loop here. First we release the iclog
2077		 * we currently have, then we get a new iclog and add a new
2078		 * opheader. Then we continue copying from where we were until
2079		 * we either complete the iovec or fill the iclog. If we
2080		 * complete the iovec, then we increment the index and go right
2081		 * back to the top of the outer loop. if we fill the iclog, we
2082		 * run the inner loop again.
2083		 *
2084		 * This is complicated by the tail of a region using all the
2085		 * space in an iclog and hence requiring us to release the iclog
2086		 * and get a new one before returning to the outer loop. We must
2087		 * always guarantee that we exit this inner loop with at least
2088		 * space for log transaction opheaders left in the current
2089		 * iclog, hence we cannot just terminate the loop at the end
2090		 * of the of the continuation. So we loop while there is no
2091		 * space left in the current iclog, and check for the end of the
2092		 * continuation after getting a new iclog.
2093		 */
2094		do {
2095			/*
2096			 * Ensure we include the continuation opheader in the
2097			 * space we need in the new iclog by adding that size
2098			 * to the length we require. This continuation opheader
2099			 * needs to be accounted to the ticket as the space it
2100			 * consumes hasn't been accounted to the lv we are
2101			 * writing.
2102			 */
2103			error = xlog_write_get_more_iclog_space(ticket,
2104					&iclog, log_offset,
2105					*len + sizeof(struct xlog_op_header),
2106					record_cnt, data_cnt);
2107			if (error)
2108				return error;
2109
2110			ophdr = iclog->ic_datap + *log_offset;
2111			ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2112			ophdr->oh_clientid = XFS_TRANSACTION;
2113			ophdr->oh_res2 = 0;
2114			ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2115
2116			ticket->t_curr_res -= sizeof(struct xlog_op_header);
2117			*log_offset += sizeof(struct xlog_op_header);
2118			*data_cnt += sizeof(struct xlog_op_header);
2119
2120			/*
2121			 * If rlen fits in the iclog, then end the region
2122			 * continuation. Otherwise we're going around again.
2123			 */
2124			reg_offset += rlen;
2125			rlen = reg->i_len - reg_offset;
2126			if (rlen <= iclog->ic_size - *log_offset)
2127				ophdr->oh_flags |= XLOG_END_TRANS;
2128			else
2129				ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2130
2131			rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2132			ophdr->oh_len = cpu_to_be32(rlen);
2133
2134			xlog_write_iovec(iclog, log_offset,
2135					reg->i_addr + reg_offset,
2136					rlen, len, record_cnt, data_cnt);
2137
2138		} while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2139	}
2140
2141	/*
2142	 * No more iovecs remain in this logvec so return the next log vec to
2143	 * the caller so it can go back to fast path copying.
2144	 */
2145	*iclogp = iclog;
2146	return 0;
2147}
2148
2149/*
2150 * Write some region out to in-core log
2151 *
2152 * This will be called when writing externally provided regions or when
2153 * writing out a commit record for a given transaction.
2154 *
2155 * General algorithm:
2156 *	1. Find total length of this write.  This may include adding to the
2157 *		lengths passed in.
2158 *	2. Check whether we violate the tickets reservation.
2159 *	3. While writing to this iclog
2160 *	    A. Reserve as much space in this iclog as can get
2161 *	    B. If this is first write, save away start lsn
2162 *	    C. While writing this region:
2163 *		1. If first write of transaction, write start record
2164 *		2. Write log operation header (header per region)
2165 *		3. Find out if we can fit entire region into this iclog
2166 *		4. Potentially, verify destination memcpy ptr
2167 *		5. Memcpy (partial) region
2168 *		6. If partial copy, release iclog; otherwise, continue
2169 *			copying more regions into current iclog
2170 *	4. Mark want sync bit (in simulation mode)
2171 *	5. Release iclog for potential flush to on-disk log.
2172 *
2173 * ERRORS:
2174 * 1.	Panic if reservation is overrun.  This should never happen since
2175 *	reservation amounts are generated internal to the filesystem.
2176 * NOTES:
2177 * 1. Tickets are single threaded data structures.
2178 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2179 *	syncing routine.  When a single log_write region needs to span
2180 *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2181 *	on all log operation writes which don't contain the end of the
2182 *	region.  The XLOG_END_TRANS bit is used for the in-core log
2183 *	operation which contains the end of the continued log_write region.
2184 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2185 *	we don't really know exactly how much space will be used.  As a result,
2186 *	we don't update ic_offset until the end when we know exactly how many
2187 *	bytes have been written out.
2188 */
2189int
2190xlog_write(
2191	struct xlog		*log,
2192	struct xfs_cil_ctx	*ctx,
2193	struct list_head	*lv_chain,
2194	struct xlog_ticket	*ticket,
2195	uint32_t		len)
2196
2197{
2198	struct xlog_in_core	*iclog = NULL;
2199	struct xfs_log_vec	*lv;
2200	uint32_t		record_cnt = 0;
2201	uint32_t		data_cnt = 0;
2202	int			error = 0;
2203	int			log_offset;
2204
2205	if (ticket->t_curr_res < 0) {
2206		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2207		     "ctx ticket reservation ran out. Need to up reservation");
2208		xlog_print_tic_res(log->l_mp, ticket);
2209		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2210	}
2211
2212	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2213					   &log_offset);
2214	if (error)
2215		return error;
2216
2217	ASSERT(log_offset <= iclog->ic_size - 1);
2218
2219	/*
2220	 * If we have a context pointer, pass it the first iclog we are
2221	 * writing to so it can record state needed for iclog write
2222	 * ordering.
2223	 */
2224	if (ctx)
2225		xlog_cil_set_ctx_write_state(ctx, iclog);
2226
2227	list_for_each_entry(lv, lv_chain, lv_list) {
2228		/*
2229		 * If the entire log vec does not fit in the iclog, punt it to
2230		 * the partial copy loop which can handle this case.
2231		 */
2232		if (lv->lv_niovecs &&
2233		    lv->lv_bytes > iclog->ic_size - log_offset) {
2234			error = xlog_write_partial(lv, ticket, &iclog,
2235					&log_offset, &len, &record_cnt,
2236					&data_cnt);
2237			if (error) {
2238				/*
2239				 * We have no iclog to release, so just return
2240				 * the error immediately.
2241				 */
2242				return error;
2243			}
2244		} else {
2245			xlog_write_full(lv, ticket, iclog, &log_offset,
2246					 &len, &record_cnt, &data_cnt);
2247		}
2248	}
2249	ASSERT(len == 0);
2250
2251	/*
2252	 * We've already been guaranteed that the last writes will fit inside
2253	 * the current iclog, and hence it will already have the space used by
2254	 * those writes accounted to it. Hence we do not need to update the
2255	 * iclog with the number of bytes written here.
2256	 */
2257	spin_lock(&log->l_icloglock);
2258	xlog_state_finish_copy(log, iclog, record_cnt, 0);
2259	error = xlog_state_release_iclog(log, iclog, ticket);
2260	spin_unlock(&log->l_icloglock);
2261
2262	return error;
2263}
2264
2265static void
2266xlog_state_activate_iclog(
2267	struct xlog_in_core	*iclog,
2268	int			*iclogs_changed)
2269{
2270	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2271	trace_xlog_iclog_activate(iclog, _RET_IP_);
2272
2273	/*
2274	 * If the number of ops in this iclog indicate it just contains the
2275	 * dummy transaction, we can change state into IDLE (the second time
2276	 * around). Otherwise we should change the state into NEED a dummy.
2277	 * We don't need to cover the dummy.
2278	 */
2279	if (*iclogs_changed == 0 &&
2280	    iclog->ic_header->h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2281		*iclogs_changed = 1;
2282	} else {
2283		/*
2284		 * We have two dirty iclogs so start over.  This could also be
2285		 * num of ops indicating this is not the dummy going out.
2286		 */
2287		*iclogs_changed = 2;
2288	}
2289
2290	iclog->ic_state	= XLOG_STATE_ACTIVE;
2291	iclog->ic_offset = 0;
2292	iclog->ic_header->h_num_logops = 0;
2293	memset(iclog->ic_header->h_cycle_data, 0,
2294		sizeof(iclog->ic_header->h_cycle_data));
2295	iclog->ic_header->h_lsn = 0;
2296	iclog->ic_header->h_tail_lsn = 0;
2297}
2298
2299/*
2300 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2301 * ACTIVE after iclog I/O has completed.
2302 */
2303static void
2304xlog_state_activate_iclogs(
2305	struct xlog		*log,
2306	int			*iclogs_changed)
2307{
2308	struct xlog_in_core	*iclog = log->l_iclog;
2309
2310	do {
2311		if (iclog->ic_state == XLOG_STATE_DIRTY)
2312			xlog_state_activate_iclog(iclog, iclogs_changed);
2313		/*
2314		 * The ordering of marking iclogs ACTIVE must be maintained, so
2315		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2316		 */
2317		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2318			break;
2319	} while ((iclog = iclog->ic_next) != log->l_iclog);
2320}
2321
2322static int
2323xlog_covered_state(
2324	int			prev_state,
2325	int			iclogs_changed)
2326{
2327	/*
2328	 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2329	 * wrote the first covering record (DONE). We go to IDLE if we just
2330	 * wrote the second covering record (DONE2) and remain in IDLE until a
2331	 * non-covering write occurs.
2332	 */
2333	switch (prev_state) {
2334	case XLOG_STATE_COVER_IDLE:
2335		if (iclogs_changed == 1)
2336			return XLOG_STATE_COVER_IDLE;
2337		fallthrough;
2338	case XLOG_STATE_COVER_NEED:
2339	case XLOG_STATE_COVER_NEED2:
2340		break;
2341	case XLOG_STATE_COVER_DONE:
2342		if (iclogs_changed == 1)
2343			return XLOG_STATE_COVER_NEED2;
2344		break;
2345	case XLOG_STATE_COVER_DONE2:
2346		if (iclogs_changed == 1)
2347			return XLOG_STATE_COVER_IDLE;
2348		break;
2349	default:
2350		ASSERT(0);
2351	}
2352
2353	return XLOG_STATE_COVER_NEED;
2354}
2355
2356STATIC void
2357xlog_state_clean_iclog(
2358	struct xlog		*log,
2359	struct xlog_in_core	*dirty_iclog)
2360{
2361	int			iclogs_changed = 0;
2362
2363	trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2364
2365	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2366
2367	xlog_state_activate_iclogs(log, &iclogs_changed);
2368	wake_up_all(&dirty_iclog->ic_force_wait);
2369
2370	if (iclogs_changed) {
2371		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2372				iclogs_changed);
2373	}
2374}
2375
2376STATIC xfs_lsn_t
2377xlog_get_lowest_lsn(
2378	struct xlog		*log)
2379{
2380	struct xlog_in_core	*iclog = log->l_iclog;
2381	xfs_lsn_t		lowest_lsn = 0, lsn;
2382
2383	do {
2384		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2385		    iclog->ic_state == XLOG_STATE_DIRTY)
2386			continue;
2387
2388		lsn = be64_to_cpu(iclog->ic_header->h_lsn);
2389		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2390			lowest_lsn = lsn;
2391	} while ((iclog = iclog->ic_next) != log->l_iclog);
2392
2393	return lowest_lsn;
2394}
2395
2396/*
2397 * Return true if we need to stop processing, false to continue to the next
2398 * iclog. The caller will need to run callbacks if the iclog is returned in the
2399 * XLOG_STATE_CALLBACK state.
2400 */
2401static bool
2402xlog_state_iodone_process_iclog(
2403	struct xlog		*log,
2404	struct xlog_in_core	*iclog)
2405{
2406	xfs_lsn_t		lowest_lsn;
2407	xfs_lsn_t		header_lsn;
2408
2409	switch (iclog->ic_state) {
2410	case XLOG_STATE_ACTIVE:
2411	case XLOG_STATE_DIRTY:
2412		/*
2413		 * Skip all iclogs in the ACTIVE & DIRTY states:
2414		 */
2415		return false;
2416	case XLOG_STATE_DONE_SYNC:
2417		/*
2418		 * Now that we have an iclog that is in the DONE_SYNC state, do
2419		 * one more check here to see if we have chased our tail around.
2420		 * If this is not the lowest lsn iclog, then we will leave it
2421		 * for another completion to process.
2422		 */
2423		header_lsn = be64_to_cpu(iclog->ic_header->h_lsn);
2424		lowest_lsn = xlog_get_lowest_lsn(log);
2425		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2426			return false;
2427		/*
2428		 * If there are no callbacks on this iclog, we can mark it clean
2429		 * immediately and return. Otherwise we need to run the
2430		 * callbacks.
2431		 */
2432		if (list_empty(&iclog->ic_callbacks)) {
2433			xlog_state_clean_iclog(log, iclog);
2434			return false;
2435		}
2436		trace_xlog_iclog_callback(iclog, _RET_IP_);
2437		iclog->ic_state = XLOG_STATE_CALLBACK;
2438		return false;
2439	default:
2440		/*
2441		 * Can only perform callbacks in order.  Since this iclog is not
2442		 * in the DONE_SYNC state, we skip the rest and just try to
2443		 * clean up.
2444		 */
2445		return true;
2446	}
2447}
2448
2449/*
2450 * Loop over all the iclogs, running attached callbacks on them. Return true if
2451 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2452 * to handle transient shutdown state here at all because
2453 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2454 * cleanup of the callbacks.
2455 */
2456static bool
2457xlog_state_do_iclog_callbacks(
2458	struct xlog		*log)
2459		__releases(&log->l_icloglock)
2460		__acquires(&log->l_icloglock)
2461{
2462	struct xlog_in_core	*first_iclog = log->l_iclog;
2463	struct xlog_in_core	*iclog = first_iclog;
2464	bool			ran_callback = false;
2465
2466	do {
2467		LIST_HEAD(cb_list);
2468
2469		if (xlog_state_iodone_process_iclog(log, iclog))
2470			break;
2471		if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2472			iclog = iclog->ic_next;
2473			continue;
2474		}
2475		list_splice_init(&iclog->ic_callbacks, &cb_list);
2476		spin_unlock(&log->l_icloglock);
2477
2478		trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2479		xlog_cil_process_committed(&cb_list);
2480		trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2481		ran_callback = true;
2482
2483		spin_lock(&log->l_icloglock);
2484		xlog_state_clean_iclog(log, iclog);
2485		iclog = iclog->ic_next;
2486	} while (iclog != first_iclog);
2487
2488	return ran_callback;
2489}
2490
2491
2492/*
2493 * Loop running iclog completion callbacks until there are no more iclogs in a
2494 * state that can run callbacks.
2495 */
2496STATIC void
2497xlog_state_do_callback(
2498	struct xlog		*log)
2499{
2500	int			flushcnt = 0;
2501	int			repeats = 0;
2502
2503	spin_lock(&log->l_icloglock);
2504	while (xlog_state_do_iclog_callbacks(log)) {
2505		if (xlog_is_shutdown(log))
2506			break;
2507
2508		if (++repeats > 5000) {
2509			flushcnt += repeats;
2510			repeats = 0;
2511			xfs_warn(log->l_mp,
2512				"%s: possible infinite loop (%d iterations)",
2513				__func__, flushcnt);
2514		}
2515	}
2516
2517	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2518		wake_up_all(&log->l_flush_wait);
2519
2520	spin_unlock(&log->l_icloglock);
2521}
2522
2523
2524/*
2525 * Finish transitioning this iclog to the dirty state.
2526 *
2527 * Callbacks could take time, so they are done outside the scope of the
2528 * global state machine log lock.
2529 */
2530STATIC void
2531xlog_state_done_syncing(
2532	struct xlog_in_core	*iclog)
2533{
2534	struct xlog		*log = iclog->ic_log;
2535
2536	spin_lock(&log->l_icloglock);
2537	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2538	trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2539
2540	/*
2541	 * If we got an error, either on the first buffer, or in the case of
2542	 * split log writes, on the second, we shut down the file system and
2543	 * no iclogs should ever be attempted to be written to disk again.
2544	 */
2545	if (!xlog_is_shutdown(log)) {
2546		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2547		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2548	}
2549
2550	/*
2551	 * Someone could be sleeping prior to writing out the next
2552	 * iclog buffer, we wake them all, one will get to do the
2553	 * I/O, the others get to wait for the result.
2554	 */
2555	wake_up_all(&iclog->ic_write_wait);
2556	spin_unlock(&log->l_icloglock);
2557	xlog_state_do_callback(log);
2558}
2559
2560/*
2561 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2562 * sleep.  We wait on the flush queue on the head iclog as that should be
2563 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2564 * we will wait here and all new writes will sleep until a sync completes.
2565 *
2566 * The in-core logs are used in a circular fashion. They are not used
2567 * out-of-order even when an iclog past the head is free.
2568 *
2569 * return:
2570 *	* log_offset where xlog_write() can start writing into the in-core
2571 *		log's data space.
2572 *	* in-core log pointer to which xlog_write() should write.
2573 *	* boolean indicating this is a continued write to an in-core log.
2574 *		If this is the last write, then the in-core log's offset field
2575 *		needs to be incremented, depending on the amount of data which
2576 *		is copied.
2577 */
2578STATIC int
2579xlog_state_get_iclog_space(
2580	struct xlog		*log,
2581	int			len,
2582	struct xlog_in_core	**iclogp,
2583	struct xlog_ticket	*ticket,
2584	int			*logoffsetp)
2585{
2586	int			log_offset;
2587	struct xlog_rec_header	*head;
2588	struct xlog_in_core	*iclog;
2589
2590restart:
2591	spin_lock(&log->l_icloglock);
2592	if (xlog_is_shutdown(log)) {
2593		spin_unlock(&log->l_icloglock);
2594		return -EIO;
2595	}
2596
2597	iclog = log->l_iclog;
2598	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2599		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2600
2601		/* Wait for log writes to have flushed */
2602		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2603		goto restart;
2604	}
2605
2606	head = iclog->ic_header;
2607
2608	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2609	log_offset = iclog->ic_offset;
2610
2611	trace_xlog_iclog_get_space(iclog, _RET_IP_);
2612
2613	/* On the 1st write to an iclog, figure out lsn.  This works
2614	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2615	 * committing to.  If the offset is set, that's how many blocks
2616	 * must be written.
2617	 */
2618	if (log_offset == 0) {
2619		ticket->t_curr_res -= log->l_iclog_hsize;
2620		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2621		head->h_lsn = cpu_to_be64(
2622			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2623		ASSERT(log->l_curr_block >= 0);
2624	}
2625
2626	/* If there is enough room to write everything, then do it.  Otherwise,
2627	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2628	 * bit is on, so this will get flushed out.  Don't update ic_offset
2629	 * until you know exactly how many bytes get copied.  Therefore, wait
2630	 * until later to update ic_offset.
2631	 *
2632	 * xlog_write() algorithm assumes that at least 2 xlog_op_header's
2633	 * can fit into remaining data section.
2634	 */
2635	if (iclog->ic_size - iclog->ic_offset <
2636	    2 * sizeof(struct xlog_op_header)) {
2637		int		error = 0;
2638
2639		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2640
2641		/*
2642		 * If we are the only one writing to this iclog, sync it to
2643		 * disk.  We need to do an atomic compare and decrement here to
2644		 * avoid racing with concurrent atomic_dec_and_lock() calls in
2645		 * xlog_state_release_iclog() when there is more than one
2646		 * reference to the iclog.
2647		 */
2648		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2649			error = xlog_state_release_iclog(log, iclog, ticket);
2650		spin_unlock(&log->l_icloglock);
2651		if (error)
2652			return error;
2653		goto restart;
2654	}
2655
2656	/* Do we have enough room to write the full amount in the remainder
2657	 * of this iclog?  Or must we continue a write on the next iclog and
2658	 * mark this iclog as completely taken?  In the case where we switch
2659	 * iclogs (to mark it taken), this particular iclog will release/sync
2660	 * to disk in xlog_write().
2661	 */
2662	if (len <= iclog->ic_size - iclog->ic_offset)
2663		iclog->ic_offset += len;
2664	else
2665		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2666	*iclogp = iclog;
2667
2668	ASSERT(iclog->ic_offset <= iclog->ic_size);
2669	spin_unlock(&log->l_icloglock);
2670
2671	*logoffsetp = log_offset;
2672	return 0;
2673}
2674
2675/*
2676 * The first cnt-1 times a ticket goes through here we don't need to move the
2677 * grant write head because the permanent reservation has reserved cnt times the
2678 * unit amount.  Release part of current permanent unit reservation and reset
2679 * current reservation to be one units worth.  Also move grant reservation head
2680 * forward.
2681 */
2682void
2683xfs_log_ticket_regrant(
2684	struct xlog		*log,
2685	struct xlog_ticket	*ticket)
2686{
2687	trace_xfs_log_ticket_regrant(log, ticket);
2688
2689	if (ticket->t_cnt > 0)
2690		ticket->t_cnt--;
2691
2692	xlog_grant_sub_space(&log->l_reserve_head, ticket->t_curr_res);
2693	xlog_grant_sub_space(&log->l_write_head, ticket->t_curr_res);
2694	ticket->t_curr_res = ticket->t_unit_res;
2695
2696	trace_xfs_log_ticket_regrant_sub(log, ticket);
2697
2698	/* just return if we still have some of the pre-reserved space */
2699	if (!ticket->t_cnt) {
2700		xlog_grant_add_space(&log->l_reserve_head, ticket->t_unit_res);
2701		trace_xfs_log_ticket_regrant_exit(log, ticket);
2702	}
2703
2704	xfs_log_ticket_put(ticket);
2705}
2706
2707/*
2708 * Give back the space left from a reservation.
2709 *
2710 * All the information we need to make a correct determination of space left
2711 * is present.  For non-permanent reservations, things are quite easy.  The
2712 * count should have been decremented to zero.  We only need to deal with the
2713 * space remaining in the current reservation part of the ticket.  If the
2714 * ticket contains a permanent reservation, there may be left over space which
2715 * needs to be released.  A count of N means that N-1 refills of the current
2716 * reservation can be done before we need to ask for more space.  The first
2717 * one goes to fill up the first current reservation.  Once we run out of
2718 * space, the count will stay at zero and the only space remaining will be
2719 * in the current reservation field.
2720 */
2721void
2722xfs_log_ticket_ungrant(
2723	struct xlog		*log,
2724	struct xlog_ticket	*ticket)
2725{
2726	int			bytes;
2727
2728	trace_xfs_log_ticket_ungrant(log, ticket);
2729
2730	if (ticket->t_cnt > 0)
2731		ticket->t_cnt--;
2732
2733	trace_xfs_log_ticket_ungrant_sub(log, ticket);
2734
2735	/*
2736	 * If this is a permanent reservation ticket, we may be able to free
2737	 * up more space based on the remaining count.
2738	 */
2739	bytes = ticket->t_curr_res;
2740	if (ticket->t_cnt > 0) {
2741		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2742		bytes += ticket->t_unit_res*ticket->t_cnt;
2743	}
2744
2745	xlog_grant_sub_space(&log->l_reserve_head, bytes);
2746	xlog_grant_sub_space(&log->l_write_head, bytes);
2747
2748	trace_xfs_log_ticket_ungrant_exit(log, ticket);
2749
2750	xfs_log_space_wake(log->l_mp);
2751	xfs_log_ticket_put(ticket);
2752}
2753
2754/*
2755 * This routine will mark the current iclog in the ring as WANT_SYNC and move
2756 * the current iclog pointer to the next iclog in the ring.
2757 */
2758void
2759xlog_state_switch_iclogs(
2760	struct xlog		*log,
2761	struct xlog_in_core	*iclog,
2762	int			eventual_size)
2763{
2764	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2765	assert_spin_locked(&log->l_icloglock);
2766	trace_xlog_iclog_switch(iclog, _RET_IP_);
2767
2768	if (!eventual_size)
2769		eventual_size = iclog->ic_offset;
2770	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2771	iclog->ic_header->h_prev_block = cpu_to_be32(log->l_prev_block);
2772	log->l_prev_block = log->l_curr_block;
2773	log->l_prev_cycle = log->l_curr_cycle;
2774
2775	/* roll log?: ic_offset changed later */
2776	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2777
2778	/* Round up to next log-sunit */
2779	if (log->l_iclog_roundoff > BBSIZE) {
2780		uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
2781		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2782	}
2783
2784	if (log->l_curr_block >= log->l_logBBsize) {
2785		/*
2786		 * Rewind the current block before the cycle is bumped to make
2787		 * sure that the combined LSN never transiently moves forward
2788		 * when the log wraps to the next cycle. This is to support the
2789		 * unlocked sample of these fields from xlog_valid_lsn(). Most
2790		 * other cases should acquire l_icloglock.
2791		 */
2792		log->l_curr_block -= log->l_logBBsize;
2793		ASSERT(log->l_curr_block >= 0);
2794		smp_wmb();
2795		log->l_curr_cycle++;
2796		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2797			log->l_curr_cycle++;
2798	}
2799	ASSERT(iclog == log->l_iclog);
2800	log->l_iclog = iclog->ic_next;
2801}
2802
2803/*
2804 * Force the iclog to disk and check if the iclog has been completed before
2805 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
2806 * pmem) or fast async storage because we drop the icloglock to issue the IO.
2807 * If completion has already occurred, tell the caller so that it can avoid an
2808 * unnecessary wait on the iclog.
2809 */
2810static int
2811xlog_force_and_check_iclog(
2812	struct xlog_in_core	*iclog,
2813	bool			*completed)
2814{
2815	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header->h_lsn);
2816	int			error;
2817
2818	*completed = false;
2819	error = xlog_force_iclog(iclog);
2820	if (error)
2821		return error;
2822
2823	/*
2824	 * If the iclog has already been completed and reused the header LSN
2825	 * will have been rewritten by completion
2826	 */
2827	if (be64_to_cpu(iclog->ic_header->h_lsn) != lsn)
2828		*completed = true;
2829	return 0;
2830}
2831
2832/*
2833 * Write out all data in the in-core log as of this exact moment in time.
2834 *
2835 * Data may be written to the in-core log during this call.  However,
2836 * we don't guarantee this data will be written out.  A change from past
2837 * implementation means this routine will *not* write out zero length LRs.
2838 *
2839 * Basically, we try and perform an intelligent scan of the in-core logs.
2840 * If we determine there is no flushable data, we just return.  There is no
2841 * flushable data if:
2842 *
2843 *	1. the current iclog is active and has no data; the previous iclog
2844 *		is in the active or dirty state.
2845 *	2. the current iclog is dirty, and the previous iclog is in the
2846 *		active or dirty state.
2847 *
2848 * We may sleep if:
2849 *
2850 *	1. the current iclog is not in the active nor dirty state.
2851 *	2. the current iclog dirty, and the previous iclog is not in the
2852 *		active nor dirty state.
2853 *	3. the current iclog is active, and there is another thread writing
2854 *		to this particular iclog.
2855 *	4. a) the current iclog is active and has no other writers
2856 *	   b) when we return from flushing out this iclog, it is still
2857 *		not in the active nor dirty state.
2858 */
2859int
2860xfs_log_force(
2861	struct xfs_mount	*mp,
2862	uint			flags)
2863{
2864	struct xlog		*log = mp->m_log;
2865	struct xlog_in_core	*iclog;
2866
2867	XFS_STATS_INC(mp, xs_log_force);
2868	trace_xfs_log_force(mp, 0, _RET_IP_);
2869
2870	xlog_cil_force(log);
2871
2872	spin_lock(&log->l_icloglock);
2873	if (xlog_is_shutdown(log))
2874		goto out_error;
2875
2876	iclog = log->l_iclog;
2877	trace_xlog_iclog_force(iclog, _RET_IP_);
2878
2879	if (iclog->ic_state == XLOG_STATE_DIRTY ||
2880	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
2881	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
2882		/*
2883		 * If the head is dirty or (active and empty), then we need to
2884		 * look at the previous iclog.
2885		 *
2886		 * If the previous iclog is active or dirty we are done.  There
2887		 * is nothing to sync out. Otherwise, we attach ourselves to the
2888		 * previous iclog and go to sleep.
2889		 */
2890		iclog = iclog->ic_prev;
2891	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2892		if (atomic_read(&iclog->ic_refcnt) == 0) {
2893			/* We have exclusive access to this iclog. */
2894			bool	completed;
2895
2896			if (xlog_force_and_check_iclog(iclog, &completed))
2897				goto out_error;
2898
2899			if (completed)
2900				goto out_unlock;
2901		} else {
2902			/*
2903			 * Someone else is still writing to this iclog, so we
2904			 * need to ensure that when they release the iclog it
2905			 * gets synced immediately as we may be waiting on it.
2906			 */
2907			xlog_state_switch_iclogs(log, iclog, 0);
2908		}
2909	}
2910
2911	/*
2912	 * The iclog we are about to wait on may contain the checkpoint pushed
2913	 * by the above xlog_cil_force() call, but it may not have been pushed
2914	 * to disk yet. Like the ACTIVE case above, we need to make sure caches
2915	 * are flushed when this iclog is written.
2916	 */
2917	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
2918		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
2919
2920	if (flags & XFS_LOG_SYNC)
2921		return xlog_wait_on_iclog(iclog);
2922out_unlock:
2923	spin_unlock(&log->l_icloglock);
2924	return 0;
2925out_error:
2926	spin_unlock(&log->l_icloglock);
2927	return -EIO;
2928}
2929
2930/*
2931 * Force the log to a specific LSN.
2932 *
2933 * If an iclog with that lsn can be found:
2934 *	If it is in the DIRTY state, just return.
2935 *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2936 *		state and go to sleep or return.
2937 *	If it is in any other state, go to sleep or return.
2938 *
2939 * Synchronous forces are implemented with a wait queue.  All callers trying
2940 * to force a given lsn to disk must wait on the queue attached to the
2941 * specific in-core log.  When given in-core log finally completes its write
2942 * to disk, that thread will wake up all threads waiting on the queue.
2943 */
2944static int
2945xlog_force_lsn(
2946	struct xlog		*log,
2947	xfs_lsn_t		lsn,
2948	uint			flags,
2949	int			*log_flushed,
2950	bool			already_slept)
2951{
2952	struct xlog_in_core	*iclog;
2953	bool			completed;
2954
2955	spin_lock(&log->l_icloglock);
2956	if (xlog_is_shutdown(log))
2957		goto out_error;
2958
2959	iclog = log->l_iclog;
2960	while (be64_to_cpu(iclog->ic_header->h_lsn) != lsn) {
2961		trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
2962		iclog = iclog->ic_next;
2963		if (iclog == log->l_iclog)
2964			goto out_unlock;
2965	}
2966
2967	switch (iclog->ic_state) {
2968	case XLOG_STATE_ACTIVE:
2969		/*
2970		 * We sleep here if we haven't already slept (e.g. this is the
2971		 * first time we've looked at the correct iclog buf) and the
2972		 * buffer before us is going to be sync'ed.  The reason for this
2973		 * is that if we are doing sync transactions here, by waiting
2974		 * for the previous I/O to complete, we can allow a few more
2975		 * transactions into this iclog before we close it down.
2976		 *
2977		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
2978		 * refcnt so we can release the log (which drops the ref count).
2979		 * The state switch keeps new transaction commits from using
2980		 * this buffer.  When the current commits finish writing into
2981		 * the buffer, the refcount will drop to zero and the buffer
2982		 * will go out then.
2983		 */
2984		if (!already_slept &&
2985		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
2986		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
2987			xlog_wait(&iclog->ic_prev->ic_write_wait,
2988					&log->l_icloglock);
2989			return -EAGAIN;
2990		}
2991		if (xlog_force_and_check_iclog(iclog, &completed))
2992			goto out_error;
2993		if (log_flushed)
2994			*log_flushed = 1;
2995		if (completed)
2996			goto out_unlock;
2997		break;
2998	case XLOG_STATE_WANT_SYNC:
2999		/*
3000		 * This iclog may contain the checkpoint pushed by the
3001		 * xlog_cil_force_seq() call, but there are other writers still
3002		 * accessing it so it hasn't been pushed to disk yet. Like the
3003		 * ACTIVE case above, we need to make sure caches are flushed
3004		 * when this iclog is written.
3005		 */
3006		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3007		break;
3008	default:
3009		/*
3010		 * The entire checkpoint was written by the CIL force and is on
3011		 * its way to disk already. It will be stable when it
3012		 * completes, so we don't need to manipulate caches here at all.
3013		 * We just need to wait for completion if necessary.
3014		 */
3015		break;
3016	}
3017
3018	if (flags & XFS_LOG_SYNC)
3019		return xlog_wait_on_iclog(iclog);
3020out_unlock:
3021	spin_unlock(&log->l_icloglock);
3022	return 0;
3023out_error:
3024	spin_unlock(&log->l_icloglock);
3025	return -EIO;
3026}
3027
3028/*
3029 * Force the log to a specific checkpoint sequence.
3030 *
3031 * First force the CIL so that all the required changes have been flushed to the
3032 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3033 * the iclog that needs to be flushed to stable storage. If the caller needs
3034 * a synchronous log force, we will wait on the iclog with the LSN returned by
3035 * xlog_cil_force_seq() to be completed.
3036 */
3037int
3038xfs_log_force_seq(
3039	struct xfs_mount	*mp,
3040	xfs_csn_t		seq,
3041	uint			flags,
3042	int			*log_flushed)
3043{
3044	struct xlog		*log = mp->m_log;
3045	xfs_lsn_t		lsn;
3046	int			ret;
3047	ASSERT(seq != 0);
3048
3049	XFS_STATS_INC(mp, xs_log_force);
3050	trace_xfs_log_force(mp, seq, _RET_IP_);
3051
3052	lsn = xlog_cil_force_seq(log, seq);
3053	if (lsn == NULLCOMMITLSN)
3054		return 0;
3055
3056	ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3057	if (ret == -EAGAIN) {
3058		XFS_STATS_INC(mp, xs_log_force_sleep);
3059		ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3060	}
3061	return ret;
3062}
3063
3064/*
3065 * Free a used ticket when its refcount falls to zero.
3066 */
3067void
3068xfs_log_ticket_put(
3069	struct xlog_ticket	*ticket)
3070{
3071	ASSERT(atomic_read(&ticket->t_ref) > 0);
3072	if (atomic_dec_and_test(&ticket->t_ref))
3073		kmem_cache_free(xfs_log_ticket_cache, ticket);
3074}
3075
3076struct xlog_ticket *
3077xfs_log_ticket_get(
3078	struct xlog_ticket	*ticket)
3079{
3080	ASSERT(atomic_read(&ticket->t_ref) > 0);
3081	atomic_inc(&ticket->t_ref);
3082	return ticket;
3083}
3084
3085/*
3086 * Figure out the total log space unit (in bytes) that would be
3087 * required for a log ticket.
3088 */
3089static int
3090xlog_calc_unit_res(
3091	struct xlog		*log,
3092	int			unit_bytes,
3093	int			*niclogs)
3094{
3095	int			iclog_space;
3096	uint			num_headers;
3097
3098	/*
3099	 * Permanent reservations have up to 'cnt'-1 active log operations
3100	 * in the log.  A unit in this case is the amount of space for one
3101	 * of these log operations.  Normal reservations have a cnt of 1
3102	 * and their unit amount is the total amount of space required.
3103	 *
3104	 * The following lines of code account for non-transaction data
3105	 * which occupy space in the on-disk log.
3106	 *
3107	 * Normal form of a transaction is:
3108	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3109	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3110	 *
3111	 * We need to account for all the leadup data and trailer data
3112	 * around the transaction data.
3113	 * And then we need to account for the worst case in terms of using
3114	 * more space.
3115	 * The worst case will happen if:
3116	 * - the placement of the transaction happens to be such that the
3117	 *   roundoff is at its maximum
3118	 * - the transaction data is synced before the commit record is synced
3119	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3120	 *   Therefore the commit record is in its own Log Record.
3121	 *   This can happen as the commit record is called with its
3122	 *   own region to xlog_write().
3123	 *   This then means that in the worst case, roundoff can happen for
3124	 *   the commit-rec as well.
3125	 *   The commit-rec is smaller than padding in this scenario and so it is
3126	 *   not added separately.
3127	 */
3128
3129	/* for trans header */
3130	unit_bytes += sizeof(struct xlog_op_header);
3131	unit_bytes += sizeof(struct xfs_trans_header);
3132
3133	/* for start-rec */
3134	unit_bytes += sizeof(struct xlog_op_header);
3135
3136	/*
3137	 * for LR headers - the space for data in an iclog is the size minus
3138	 * the space used for the headers. If we use the iclog size, then we
3139	 * undercalculate the number of headers required.
3140	 *
3141	 * Furthermore - the addition of op headers for split-recs might
3142	 * increase the space required enough to require more log and op
3143	 * headers, so take that into account too.
3144	 *
3145	 * IMPORTANT: This reservation makes the assumption that if this
3146	 * transaction is the first in an iclog and hence has the LR headers
3147	 * accounted to it, then the remaining space in the iclog is
3148	 * exclusively for this transaction.  i.e. if the transaction is larger
3149	 * than the iclog, it will be the only thing in that iclog.
3150	 * Fundamentally, this means we must pass the entire log vector to
3151	 * xlog_write to guarantee this.
3152	 */
3153	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3154	num_headers = howmany(unit_bytes, iclog_space);
3155
3156	/* for split-recs - ophdrs added when data split over LRs */
3157	unit_bytes += sizeof(struct xlog_op_header) * num_headers;
3158
3159	/* add extra header reservations if we overrun */
3160	while (!num_headers ||
3161	       howmany(unit_bytes, iclog_space) > num_headers) {
3162		unit_bytes += sizeof(struct xlog_op_header);
3163		num_headers++;
3164	}
3165	unit_bytes += log->l_iclog_hsize * num_headers;
3166
3167	/* for commit-rec LR header - note: padding will subsume the ophdr */
3168	unit_bytes += log->l_iclog_hsize;
3169
3170	/* roundoff padding for transaction data and one for commit record */
3171	unit_bytes += 2 * log->l_iclog_roundoff;
3172
3173	if (niclogs)
3174		*niclogs = num_headers;
3175	return unit_bytes;
3176}
3177
3178int
3179xfs_log_calc_unit_res(
3180	struct xfs_mount	*mp,
3181	int			unit_bytes)
3182{
3183	return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3184}
3185
3186/*
3187 * Allocate and initialise a new log ticket.
3188 */
3189struct xlog_ticket *
3190xlog_ticket_alloc(
3191	struct xlog		*log,
3192	int			unit_bytes,
3193	int			cnt,
3194	bool			permanent)
3195{
3196	struct xlog_ticket	*tic;
3197	int			unit_res;
3198
3199	tic = kmem_cache_zalloc(xfs_log_ticket_cache,
3200			GFP_KERNEL | __GFP_NOFAIL);
3201
3202	unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3203
3204	atomic_set(&tic->t_ref, 1);
3205	tic->t_task		= current;
3206	INIT_LIST_HEAD(&tic->t_queue);
3207	tic->t_unit_res		= unit_res;
3208	tic->t_curr_res		= unit_res;
3209	tic->t_cnt		= cnt;
3210	tic->t_ocnt		= cnt;
3211	tic->t_tid		= get_random_u32();
3212	if (permanent)
3213		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3214
3215	return tic;
3216}
3217
3218#if defined(DEBUG)
3219static void
3220xlog_verify_dump_tail(
3221	struct xlog		*log,
3222	struct xlog_in_core	*iclog)
3223{
3224	xfs_alert(log->l_mp,
3225"ran out of log space tail 0x%llx/0x%llx, head lsn 0x%llx, head 0x%x/0x%x, prev head 0x%x/0x%x",
3226			iclog ? be64_to_cpu(iclog->ic_header->h_tail_lsn) : -1,
3227			atomic64_read(&log->l_tail_lsn),
3228			log->l_ailp->ail_head_lsn,
3229			log->l_curr_cycle, log->l_curr_block,
3230			log->l_prev_cycle, log->l_prev_block);
3231	xfs_alert(log->l_mp,
3232"write grant 0x%llx, reserve grant 0x%llx, tail_space 0x%llx, size 0x%x, iclog flags 0x%x",
3233			atomic64_read(&log->l_write_head.grant),
3234			atomic64_read(&log->l_reserve_head.grant),
3235			log->l_tail_space, log->l_logsize,
3236			iclog ? iclog->ic_flags : -1);
3237}
3238
3239/* Check if the new iclog will fit in the log. */
3240STATIC void
3241xlog_verify_tail_lsn(
3242	struct xlog		*log,
3243	struct xlog_in_core	*iclog)
3244{
3245	xfs_lsn_t	tail_lsn = be64_to_cpu(iclog->ic_header->h_tail_lsn);
3246	int		blocks;
3247
3248	if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3249		blocks = log->l_logBBsize -
3250				(log->l_prev_block - BLOCK_LSN(tail_lsn));
3251		if (blocks < BTOBB(iclog->ic_offset) +
3252					BTOBB(log->l_iclog_hsize)) {
3253			xfs_emerg(log->l_mp,
3254					"%s: ran out of log space", __func__);
3255			xlog_verify_dump_tail(log, iclog);
3256		}
3257		return;
3258	}
3259
3260	if (CYCLE_LSN(tail_lsn) + 1 != log->l_prev_cycle) {
3261		xfs_emerg(log->l_mp, "%s: head has wrapped tail.", __func__);
3262		xlog_verify_dump_tail(log, iclog);
3263		return;
3264	}
3265	if (BLOCK_LSN(tail_lsn) == log->l_prev_block) {
3266		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3267		xlog_verify_dump_tail(log, iclog);
3268		return;
3269	}
3270
3271	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3272	if (blocks < BTOBB(iclog->ic_offset) + 1) {
3273		xfs_emerg(log->l_mp, "%s: ran out of iclog space", __func__);
3274		xlog_verify_dump_tail(log, iclog);
3275	}
3276}
3277
3278/*
3279 * Perform a number of checks on the iclog before writing to disk.
3280 *
3281 * 1. Make sure the iclogs are still circular
3282 * 2. Make sure we have a good magic number
3283 * 3. Make sure we don't have magic numbers in the data
3284 * 4. Check fields of each log operation header for:
3285 *	A. Valid client identifier
3286 *	B. tid ptr value falls in valid ptr space (user space code)
3287 *	C. Length in log record header is correct according to the
3288 *		individual operation headers within record.
3289 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3290 *	log, check the preceding blocks of the physical log to make sure all
3291 *	the cycle numbers agree with the current cycle number.
3292 */
3293STATIC void
3294xlog_verify_iclog(
3295	struct xlog		*log,
3296	struct xlog_in_core	*iclog,
3297	int			count)
3298{
3299	struct xlog_rec_header	*rhead = iclog->ic_header;
3300	struct xlog_in_core	*icptr;
3301	void			*base_ptr, *ptr;
3302	ptrdiff_t		field_offset;
3303	uint8_t			clientid;
3304	int			len, i, op_len;
3305	int			idx;
3306
3307	/* check validity of iclog pointers */
3308	spin_lock(&log->l_icloglock);
3309	icptr = log->l_iclog;
3310	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3311		ASSERT(icptr);
3312
3313	if (icptr != log->l_iclog)
3314		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3315	spin_unlock(&log->l_icloglock);
3316
3317	/* check log magic numbers */
3318	if (rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3319		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3320
3321	base_ptr = ptr = rhead;
3322	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3323		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3324			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3325				__func__);
3326	}
3327
3328	/* check fields */
3329	len = be32_to_cpu(rhead->h_num_logops);
3330	base_ptr = ptr = iclog->ic_datap;
3331	for (i = 0; i < len; i++) {
3332		struct xlog_op_header	*ophead = ptr;
3333		void			*p = &ophead->oh_clientid;
3334
3335		/* clientid is only 1 byte */
3336		field_offset = p - base_ptr;
3337		if (field_offset & 0x1ff) {
3338			clientid = ophead->oh_clientid;
3339		} else {
3340			idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3341			clientid = xlog_get_client_id(*xlog_cycle_data(rhead, idx));
3342		}
3343		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3344			xfs_warn(log->l_mp,
3345				"%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3346				__func__, i, clientid, ophead,
3347				(unsigned long)field_offset);
3348		}
3349
3350		/* check length */
3351		p = &ophead->oh_len;
3352		field_offset = p - base_ptr;
3353		if (field_offset & 0x1ff) {
3354			op_len = be32_to_cpu(ophead->oh_len);
3355		} else {
3356			idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3357			op_len = be32_to_cpu(*xlog_cycle_data(rhead, idx));
3358		}
3359		ptr += sizeof(struct xlog_op_header) + op_len;
3360	}
3361}
3362#endif
3363
3364/*
3365 * Perform a forced shutdown on the log.
3366 *
3367 * This can be called from low level log code to trigger a shutdown, or from the
3368 * high level mount shutdown code when the mount shuts down.
3369 *
3370 * Our main objectives here are to make sure that:
3371 *	a. if the shutdown was not due to a log IO error, flush the logs to
3372 *	   disk. Anything modified after this is ignored.
3373 *	b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3374 *	   parties to find out. Nothing new gets queued after this is done.
3375 *	c. Tasks sleeping on log reservations, pinned objects and
3376 *	   other resources get woken up.
3377 *	d. The mount is also marked as shut down so that log triggered shutdowns
3378 *	   still behave the same as if they called xfs_forced_shutdown().
3379 *
3380 * Return true if the shutdown cause was a log IO error and we actually shut the
3381 * log down.
3382 */
3383bool
3384xlog_force_shutdown(
3385	struct xlog	*log,
3386	uint32_t	shutdown_flags)
3387{
3388	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3389
3390	if (!log)
3391		return false;
3392
3393	/*
3394	 * Ensure that there is only ever one log shutdown being processed.
3395	 * If we allow the log force below on a second pass after shutting
3396	 * down the log, we risk deadlocking the CIL push as it may require
3397	 * locks on objects the current shutdown context holds (e.g. taking
3398	 * buffer locks to abort buffers on last unpin of buf log items).
3399	 */
3400	if (test_and_set_bit(XLOG_SHUTDOWN_STARTED, &log->l_opstate))
3401		return false;
3402
3403	/*
3404	 * Flush all the completed transactions to disk before marking the log
3405	 * being shut down. We need to do this first as shutting down the log
3406	 * before the force will prevent the log force from flushing the iclogs
3407	 * to disk.
3408	 *
3409	 * When we are in recovery, there are no transactions to flush, and
3410	 * we don't want to touch the log because we don't want to perturb the
3411	 * current head/tail for future recovery attempts. Hence we need to
3412	 * avoid a log force in this case.
3413	 *
3414	 * If we are shutting down due to a log IO error, then we must avoid
3415	 * trying to write the log as that may just result in more IO errors and
3416	 * an endless shutdown/force loop.
3417	 */
3418	if (!log_error && !xlog_in_recovery(log))
3419		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3420
3421	/*
3422	 * Atomically set the shutdown state. If the shutdown state is already
3423	 * set, there someone else is performing the shutdown and so we are done
3424	 * here. This should never happen because we should only ever get called
3425	 * once by the first shutdown caller.
3426	 *
3427	 * Much of the log state machine transitions assume that shutdown state
3428	 * cannot change once they hold the log->l_icloglock. Hence we need to
3429	 * hold that lock here, even though we use the atomic test_and_set_bit()
3430	 * operation to set the shutdown state.
3431	 */
3432	spin_lock(&log->l_icloglock);
3433	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3434		spin_unlock(&log->l_icloglock);
3435		ASSERT(0);
3436		return false;
3437	}
3438	spin_unlock(&log->l_icloglock);
3439
3440	/*
3441	 * If this log shutdown also sets the mount shutdown state, issue a
3442	 * shutdown warning message.
3443	 */
3444	if (!xfs_set_shutdown(log->l_mp)) {
3445		xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3446"Filesystem has been shut down due to log error (0x%x).",
3447				shutdown_flags);
3448		xfs_alert(log->l_mp,
3449"Please unmount the filesystem and rectify the problem(s).");
3450		if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3451			xfs_stack_trace();
3452	}
3453
3454	/*
3455	 * We don't want anybody waiting for log reservations after this. That
3456	 * means we have to wake up everybody queued up on reserveq as well as
3457	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3458	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3459	 * action is protected by the grant locks.
3460	 */
3461	xlog_grant_head_wake_all(&log->l_reserve_head);
3462	xlog_grant_head_wake_all(&log->l_write_head);
3463
3464	/*
3465	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3466	 * as if the log writes were completed. The abort handling in the log
3467	 * item committed callback functions will do this again under lock to
3468	 * avoid races.
3469	 */
3470	spin_lock(&log->l_cilp->xc_push_lock);
3471	wake_up_all(&log->l_cilp->xc_start_wait);
3472	wake_up_all(&log->l_cilp->xc_commit_wait);
3473	spin_unlock(&log->l_cilp->xc_push_lock);
3474
3475	spin_lock(&log->l_icloglock);
3476	xlog_state_shutdown_callbacks(log);
3477	spin_unlock(&log->l_icloglock);
3478
3479	wake_up_var(&log->l_opstate);
3480	if (IS_ENABLED(CONFIG_XFS_RT) && xfs_has_zoned(log->l_mp))
3481		xfs_zoned_wake_all(log->l_mp);
3482
3483	return log_error;
3484}
3485
3486STATIC int
3487xlog_iclogs_empty(
3488	struct xlog		*log)
3489{
3490	struct xlog_in_core	*iclog = log->l_iclog;
3491
3492	do {
3493		/* endianness does not matter here, zero is zero in
3494		 * any language.
3495		 */
3496		if (iclog->ic_header->h_num_logops)
3497			return 0;
3498		iclog = iclog->ic_next;
3499	} while (iclog != log->l_iclog);
3500
3501	return 1;
3502}
3503
3504/*
3505 * Verify that an LSN stamped into a piece of metadata is valid. This is
3506 * intended for use in read verifiers on v5 superblocks.
3507 */
3508bool
3509xfs_log_check_lsn(
3510	struct xfs_mount	*mp,
3511	xfs_lsn_t		lsn)
3512{
3513	struct xlog		*log = mp->m_log;
3514	bool			valid;
3515
3516	/*
3517	 * norecovery mode skips mount-time log processing and unconditionally
3518	 * resets the in-core LSN. We can't validate in this mode, but
3519	 * modifications are not allowed anyways so just return true.
3520	 */
3521	if (xfs_has_norecovery(mp))
3522		return true;
3523
3524	/*
3525	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3526	 * handled by recovery and thus safe to ignore here.
3527	 */
3528	if (lsn == NULLCOMMITLSN)
3529		return true;
3530
3531	valid = xlog_valid_lsn(mp->m_log, lsn);
3532
3533	/* warn the user about what's gone wrong before verifier failure */
3534	if (!valid) {
3535		spin_lock(&log->l_icloglock);
3536		xfs_warn(mp,
3537"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3538"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3539			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3540			 log->l_curr_cycle, log->l_curr_block);
3541		spin_unlock(&log->l_icloglock);
3542	}
3543
3544	return valid;
3545}