Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_shared.h"
  21#include "xfs_format.h"
  22#include "xfs_log_format.h"
  23#include "xfs_trans_resv.h"
  24#include "xfs_mount.h"
  25#include "xfs_errortag.h"
  26#include "xfs_error.h"
  27#include "xfs_trans.h"
  28#include "xfs_trans_priv.h"
  29#include "xfs_log.h"
  30#include "xfs_log_priv.h"
  31#include "xfs_log_recover.h"
  32#include "xfs_inode.h"
  33#include "xfs_trace.h"
  34#include "xfs_fsops.h"
  35#include "xfs_cksum.h"
  36#include "xfs_sysfs.h"
  37#include "xfs_sb.h"
 
  38
  39kmem_zone_t	*xfs_log_ticket_zone;
  40
  41/* Local miscellaneous function prototypes */
  42STATIC int
  43xlog_commit_record(
  44	struct xlog		*log,
  45	struct xlog_ticket	*ticket,
  46	struct xlog_in_core	**iclog,
  47	xfs_lsn_t		*commitlsnp);
  48
  49STATIC struct xlog *
  50xlog_alloc_log(
  51	struct xfs_mount	*mp,
  52	struct xfs_buftarg	*log_target,
  53	xfs_daddr_t		blk_offset,
  54	int			num_bblks);
  55STATIC int
  56xlog_space_left(
  57	struct xlog		*log,
  58	atomic64_t		*head);
  59STATIC int
  60xlog_sync(
  61	struct xlog		*log,
  62	struct xlog_in_core	*iclog);
  63STATIC void
  64xlog_dealloc_log(
  65	struct xlog		*log);
  66
  67/* local state machine functions */
  68STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  69STATIC void
  70xlog_state_do_callback(
  71	struct xlog		*log,
  72	int			aborted,
  73	struct xlog_in_core	*iclog);
 
 
  74STATIC int
  75xlog_state_get_iclog_space(
  76	struct xlog		*log,
  77	int			len,
  78	struct xlog_in_core	**iclog,
  79	struct xlog_ticket	*ticket,
  80	int			*continued_write,
  81	int			*logoffsetp);
  82STATIC int
  83xlog_state_release_iclog(
  84	struct xlog		*log,
  85	struct xlog_in_core	*iclog);
  86STATIC void
  87xlog_state_switch_iclogs(
  88	struct xlog		*log,
  89	struct xlog_in_core	*iclog,
  90	int			eventual_size);
  91STATIC void
  92xlog_state_want_sync(
  93	struct xlog		*log,
  94	struct xlog_in_core	*iclog);
  95
  96STATIC void
  97xlog_grant_push_ail(
  98	struct xlog		*log,
  99	int			need_bytes);
 100STATIC void
 101xlog_regrant_reserve_log_space(
 102	struct xlog		*log,
 103	struct xlog_ticket	*ticket);
 104STATIC void
 105xlog_ungrant_log_space(
 106	struct xlog		*log,
 107	struct xlog_ticket	*ticket);
 108
 109#if defined(DEBUG)
 110STATIC void
 111xlog_verify_dest_ptr(
 112	struct xlog		*log,
 113	void			*ptr);
 114STATIC void
 115xlog_verify_grant_tail(
 116	struct xlog *log);
 117STATIC void
 118xlog_verify_iclog(
 119	struct xlog		*log,
 120	struct xlog_in_core	*iclog,
 121	int			count,
 122	bool                    syncing);
 123STATIC void
 124xlog_verify_tail_lsn(
 125	struct xlog		*log,
 126	struct xlog_in_core	*iclog,
 127	xfs_lsn_t		tail_lsn);
 128#else
 129#define xlog_verify_dest_ptr(a,b)
 130#define xlog_verify_grant_tail(a)
 131#define xlog_verify_iclog(a,b,c,d)
 132#define xlog_verify_tail_lsn(a,b,c)
 133#endif
 134
 135STATIC int
 136xlog_iclogs_empty(
 137	struct xlog		*log);
 138
 139static void
 140xlog_grant_sub_space(
 141	struct xlog		*log,
 142	atomic64_t		*head,
 143	int			bytes)
 144{
 145	int64_t	head_val = atomic64_read(head);
 146	int64_t new, old;
 147
 148	do {
 149		int	cycle, space;
 150
 151		xlog_crack_grant_head_val(head_val, &cycle, &space);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 152
 153		space -= bytes;
 154		if (space < 0) {
 155			space += log->l_logsize;
 156			cycle--;
 157		}
 158
 159		old = head_val;
 160		new = xlog_assign_grant_head_val(cycle, space);
 161		head_val = atomic64_cmpxchg(head, old, new);
 162	} while (head_val != old);
 163}
 164
 165static void
 166xlog_grant_add_space(
 167	struct xlog		*log,
 168	atomic64_t		*head,
 169	int			bytes)
 170{
 171	int64_t	head_val = atomic64_read(head);
 172	int64_t new, old;
 173
 174	do {
 175		int		tmp;
 176		int		cycle, space;
 177
 178		xlog_crack_grant_head_val(head_val, &cycle, &space);
 
 
 179
 180		tmp = log->l_logsize - space;
 181		if (tmp > bytes)
 182			space += bytes;
 183		else {
 184			space = bytes - tmp;
 185			cycle++;
 186		}
 187
 188		old = head_val;
 189		new = xlog_assign_grant_head_val(cycle, space);
 190		head_val = atomic64_cmpxchg(head, old, new);
 191	} while (head_val != old);
 
 
 192}
 193
 194STATIC void
 195xlog_grant_head_init(
 196	struct xlog_grant_head	*head)
 197{
 198	xlog_assign_grant_head(&head->grant, 1, 0);
 199	INIT_LIST_HEAD(&head->waiters);
 200	spin_lock_init(&head->lock);
 201}
 202
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 203STATIC void
 204xlog_grant_head_wake_all(
 205	struct xlog_grant_head	*head)
 206{
 207	struct xlog_ticket	*tic;
 208
 209	spin_lock(&head->lock);
 210	list_for_each_entry(tic, &head->waiters, t_queue)
 211		wake_up_process(tic->t_task);
 212	spin_unlock(&head->lock);
 213}
 214
 215static inline int
 216xlog_ticket_reservation(
 217	struct xlog		*log,
 218	struct xlog_grant_head	*head,
 219	struct xlog_ticket	*tic)
 220{
 221	if (head == &log->l_write_head) {
 222		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 223		return tic->t_unit_res;
 224	} else {
 225		if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 226			return tic->t_unit_res * tic->t_cnt;
 227		else
 228			return tic->t_unit_res;
 229	}
 
 
 
 
 
 230}
 231
 232STATIC bool
 233xlog_grant_head_wake(
 234	struct xlog		*log,
 235	struct xlog_grant_head	*head,
 236	int			*free_bytes)
 237{
 238	struct xlog_ticket	*tic;
 239	int			need_bytes;
 240
 241	list_for_each_entry(tic, &head->waiters, t_queue) {
 242		need_bytes = xlog_ticket_reservation(log, head, tic);
 243		if (*free_bytes < need_bytes)
 244			return false;
 245
 246		*free_bytes -= need_bytes;
 247		trace_xfs_log_grant_wake_up(log, tic);
 248		wake_up_process(tic->t_task);
 249	}
 250
 251	return true;
 252}
 253
 254STATIC int
 255xlog_grant_head_wait(
 256	struct xlog		*log,
 257	struct xlog_grant_head	*head,
 258	struct xlog_ticket	*tic,
 259	int			need_bytes) __releases(&head->lock)
 260					    __acquires(&head->lock)
 261{
 262	list_add_tail(&tic->t_queue, &head->waiters);
 263
 264	do {
 265		if (XLOG_FORCED_SHUTDOWN(log))
 266			goto shutdown;
 267		xlog_grant_push_ail(log, need_bytes);
 268
 269		__set_current_state(TASK_UNINTERRUPTIBLE);
 270		spin_unlock(&head->lock);
 271
 272		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 273
 
 
 
 274		trace_xfs_log_grant_sleep(log, tic);
 275		schedule();
 276		trace_xfs_log_grant_wake(log, tic);
 277
 278		spin_lock(&head->lock);
 279		if (XLOG_FORCED_SHUTDOWN(log))
 280			goto shutdown;
 281	} while (xlog_space_left(log, &head->grant) < need_bytes);
 282
 283	list_del_init(&tic->t_queue);
 284	return 0;
 285shutdown:
 286	list_del_init(&tic->t_queue);
 287	return -EIO;
 288}
 289
 290/*
 291 * Atomically get the log space required for a log ticket.
 292 *
 293 * Once a ticket gets put onto head->waiters, it will only return after the
 294 * needed reservation is satisfied.
 295 *
 296 * This function is structured so that it has a lock free fast path. This is
 297 * necessary because every new transaction reservation will come through this
 298 * path. Hence any lock will be globally hot if we take it unconditionally on
 299 * every pass.
 300 *
 301 * As tickets are only ever moved on and off head->waiters under head->lock, we
 302 * only need to take that lock if we are going to add the ticket to the queue
 303 * and sleep. We can avoid taking the lock if the ticket was never added to
 304 * head->waiters because the t_queue list head will be empty and we hold the
 305 * only reference to it so it can safely be checked unlocked.
 306 */
 307STATIC int
 308xlog_grant_head_check(
 309	struct xlog		*log,
 310	struct xlog_grant_head	*head,
 311	struct xlog_ticket	*tic,
 312	int			*need_bytes)
 313{
 314	int			free_bytes;
 315	int			error = 0;
 316
 317	ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 318
 319	/*
 320	 * If there are other waiters on the queue then give them a chance at
 321	 * logspace before us.  Wake up the first waiters, if we do not wake
 322	 * up all the waiters then go to sleep waiting for more free space,
 323	 * otherwise try to get some space for this transaction.
 324	 */
 325	*need_bytes = xlog_ticket_reservation(log, head, tic);
 326	free_bytes = xlog_space_left(log, &head->grant);
 327	if (!list_empty_careful(&head->waiters)) {
 328		spin_lock(&head->lock);
 329		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 330		    free_bytes < *need_bytes) {
 331			error = xlog_grant_head_wait(log, head, tic,
 332						     *need_bytes);
 333		}
 334		spin_unlock(&head->lock);
 335	} else if (free_bytes < *need_bytes) {
 336		spin_lock(&head->lock);
 337		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 338		spin_unlock(&head->lock);
 339	}
 340
 341	return error;
 342}
 343
 344static void
 345xlog_tic_reset_res(xlog_ticket_t *tic)
 346{
 347	tic->t_res_num = 0;
 348	tic->t_res_arr_sum = 0;
 349	tic->t_res_num_ophdrs = 0;
 350}
 351
 352static void
 353xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 354{
 355	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 356		/* add to overflow and start again */
 357		tic->t_res_o_flow += tic->t_res_arr_sum;
 358		tic->t_res_num = 0;
 359		tic->t_res_arr_sum = 0;
 360	}
 361
 362	tic->t_res_arr[tic->t_res_num].r_len = len;
 363	tic->t_res_arr[tic->t_res_num].r_type = type;
 364	tic->t_res_arr_sum += len;
 365	tic->t_res_num++;
 
 
 
 
 366}
 367
 368/*
 369 * Replenish the byte reservation required by moving the grant write head.
 370 */
 371int
 372xfs_log_regrant(
 373	struct xfs_mount	*mp,
 374	struct xlog_ticket	*tic)
 375{
 376	struct xlog		*log = mp->m_log;
 377	int			need_bytes;
 378	int			error = 0;
 379
 380	if (XLOG_FORCED_SHUTDOWN(log))
 381		return -EIO;
 382
 383	XFS_STATS_INC(mp, xs_try_logspace);
 384
 385	/*
 386	 * This is a new transaction on the ticket, so we need to change the
 387	 * transaction ID so that the next transaction has a different TID in
 388	 * the log. Just add one to the existing tid so that we can see chains
 389	 * of rolling transactions in the log easily.
 390	 */
 391	tic->t_tid++;
 392
 393	xlog_grant_push_ail(log, tic->t_unit_res);
 394
 395	tic->t_curr_res = tic->t_unit_res;
 396	xlog_tic_reset_res(tic);
 397
 398	if (tic->t_cnt > 0)
 399		return 0;
 400
 401	trace_xfs_log_regrant(log, tic);
 402
 403	error = xlog_grant_head_check(log, &log->l_write_head, tic,
 404				      &need_bytes);
 405	if (error)
 406		goto out_error;
 407
 408	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 409	trace_xfs_log_regrant_exit(log, tic);
 410	xlog_verify_grant_tail(log);
 411	return 0;
 412
 413out_error:
 414	/*
 415	 * If we are failing, make sure the ticket doesn't have any current
 416	 * reservations.  We don't want to add this back when the ticket/
 417	 * transaction gets cancelled.
 418	 */
 419	tic->t_curr_res = 0;
 420	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
 421	return error;
 422}
 423
 424/*
 425 * Reserve log space and return a ticket corresponding the reservation.
 426 *
 427 * Each reservation is going to reserve extra space for a log record header.
 428 * When writes happen to the on-disk log, we don't subtract the length of the
 429 * log record header from any reservation.  By wasting space in each
 430 * reservation, we prevent over allocation problems.
 431 */
 432int
 433xfs_log_reserve(
 434	struct xfs_mount	*mp,
 435	int		 	unit_bytes,
 436	int		 	cnt,
 437	struct xlog_ticket	**ticp,
 438	uint8_t		 	client,
 439	bool			permanent)
 440{
 441	struct xlog		*log = mp->m_log;
 442	struct xlog_ticket	*tic;
 443	int			need_bytes;
 444	int			error = 0;
 445
 446	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 447
 448	if (XLOG_FORCED_SHUTDOWN(log))
 449		return -EIO;
 450
 451	XFS_STATS_INC(mp, xs_try_logspace);
 452
 453	ASSERT(*ticp == NULL);
 454	tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 455				KM_SLEEP | KM_MAYFAIL);
 456	if (!tic)
 457		return -ENOMEM;
 458
 459	*ticp = tic;
 460
 461	xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 462					    : tic->t_unit_res);
 463
 464	trace_xfs_log_reserve(log, tic);
 465
 466	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 467				      &need_bytes);
 468	if (error)
 469		goto out_error;
 470
 471	xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 472	xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 473	trace_xfs_log_reserve_exit(log, tic);
 474	xlog_verify_grant_tail(log);
 475	return 0;
 476
 477out_error:
 478	/*
 479	 * If we are failing, make sure the ticket doesn't have any current
 480	 * reservations.  We don't want to add this back when the ticket/
 481	 * transaction gets cancelled.
 482	 */
 483	tic->t_curr_res = 0;
 484	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
 485	return error;
 486}
 487
 488
 489/*
 490 * NOTES:
 491 *
 492 *	1. currblock field gets updated at startup and after in-core logs
 493 *		marked as with WANT_SYNC.
 494 */
 495
 496/*
 497 * This routine is called when a user of a log manager ticket is done with
 498 * the reservation.  If the ticket was ever used, then a commit record for
 499 * the associated transaction is written out as a log operation header with
 500 * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 501 * a given ticket.  If the ticket was one with a permanent reservation, then
 502 * a few operations are done differently.  Permanent reservation tickets by
 503 * default don't release the reservation.  They just commit the current
 504 * transaction with the belief that the reservation is still needed.  A flag
 505 * must be passed in before permanent reservations are actually released.
 506 * When these type of tickets are not released, they need to be set into
 507 * the inited state again.  By doing this, a start record will be written
 508 * out when the next write occurs.
 509 */
 510xfs_lsn_t
 511xfs_log_done(
 512	struct xfs_mount	*mp,
 513	struct xlog_ticket	*ticket,
 514	struct xlog_in_core	**iclog,
 515	bool			regrant)
 516{
 517	struct xlog		*log = mp->m_log;
 518	xfs_lsn_t		lsn = 0;
 519
 520	if (XLOG_FORCED_SHUTDOWN(log) ||
 521	    /*
 522	     * If nothing was ever written, don't write out commit record.
 523	     * If we get an error, just continue and give back the log ticket.
 524	     */
 525	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 526	     (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 527		lsn = (xfs_lsn_t) -1;
 528		regrant = false;
 529	}
 530
 531
 532	if (!regrant) {
 533		trace_xfs_log_done_nonperm(log, ticket);
 534
 535		/*
 536		 * Release ticket if not permanent reservation or a specific
 537		 * request has been made to release a permanent reservation.
 538		 */
 539		xlog_ungrant_log_space(log, ticket);
 540	} else {
 541		trace_xfs_log_done_perm(log, ticket);
 
 542
 543		xlog_regrant_reserve_log_space(log, ticket);
 544		/* If this ticket was a permanent reservation and we aren't
 545		 * trying to release it, reset the inited flags; so next time
 546		 * we write, a start record will be written out.
 547		 */
 548		ticket->t_flags |= XLOG_TIC_INITED;
 549	}
 550
 551	xfs_log_ticket_put(ticket);
 552	return lsn;
 
 
 
 
 553}
 554
 555/*
 556 * Attaches a new iclog I/O completion callback routine during
 557 * transaction commit.  If the log is in error state, a non-zero
 558 * return code is handed back and the caller is responsible for
 559 * executing the callback at an appropriate time.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 560 */
 561int
 562xfs_log_notify(
 
 563	struct xlog_in_core	*iclog,
 564	xfs_log_callback_t	*cb)
 565{
 566	int	abortflg;
 567
 568	spin_lock(&iclog->ic_callback_lock);
 569	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 570	if (!abortflg) {
 571		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 572			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 573		cb->cb_next = NULL;
 574		*(iclog->ic_callback_tail) = cb;
 575		iclog->ic_callback_tail = &(cb->cb_next);
 
 
 
 
 
 
 576	}
 577	spin_unlock(&iclog->ic_callback_lock);
 578	return abortflg;
 579}
 580
 581int
 582xfs_log_release_iclog(
 583	struct xfs_mount	*mp,
 584	struct xlog_in_core	*iclog)
 585{
 586	if (xlog_state_release_iclog(mp->m_log, iclog)) {
 587		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 
 
 
 588		return -EIO;
 589	}
 590
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 591	return 0;
 592}
 593
 594/*
 595 * Mount a log filesystem
 596 *
 597 * mp		- ubiquitous xfs mount point structure
 598 * log_target	- buftarg of on-disk log device
 599 * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
 600 * num_bblocks	- Number of BBSIZE blocks in on-disk log
 601 *
 602 * Return error or zero.
 603 */
 604int
 605xfs_log_mount(
 606	xfs_mount_t	*mp,
 607	xfs_buftarg_t	*log_target,
 608	xfs_daddr_t	blk_offset,
 609	int		num_bblks)
 610{
 611	bool		fatal = xfs_sb_version_hascrc(&mp->m_sb);
 612	int		error = 0;
 613	int		min_logfsbs;
 614
 615	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 616		xfs_notice(mp, "Mounting V%d Filesystem",
 617			   XFS_SB_VERSION_NUM(&mp->m_sb));
 
 618	} else {
 619		xfs_notice(mp,
 620"Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 621			   XFS_SB_VERSION_NUM(&mp->m_sb));
 622		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 
 623	}
 624
 625	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 626	if (IS_ERR(mp->m_log)) {
 627		error = PTR_ERR(mp->m_log);
 628		goto out;
 629	}
 
 630
 631	/*
 632	 * Validate the given log space and drop a critical message via syslog
 633	 * if the log size is too small that would lead to some unexpected
 634	 * situations in transaction log space reservation stage.
 
 
 
 
 635	 *
 636	 * Note: we can't just reject the mount if the validation fails.  This
 637	 * would mean that people would have to downgrade their kernel just to
 638	 * remedy the situation as there is no way to grow the log (short of
 639	 * black magic surgery with xfs_db).
 640	 *
 641	 * We can, however, reject mounts for CRC format filesystems, as the
 642	 * mkfs binary being used to make the filesystem should never create a
 643	 * filesystem with a log that is too small.
 644	 */
 645	min_logfsbs = xfs_log_calc_minimum_size(mp);
 646
 647	if (mp->m_sb.sb_logblocks < min_logfsbs) {
 648		xfs_warn(mp,
 649		"Log size %d blocks too small, minimum size is %d blocks",
 650			 mp->m_sb.sb_logblocks, min_logfsbs);
 651		error = -EINVAL;
 652	} else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 653		xfs_warn(mp,
 654		"Log size %d blocks too large, maximum size is %lld blocks",
 655			 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 656		error = -EINVAL;
 657	} else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 658		xfs_warn(mp,
 659		"log size %lld bytes too large, maximum size is %lld bytes",
 660			 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 661			 XFS_MAX_LOG_BYTES);
 662		error = -EINVAL;
 663	} else if (mp->m_sb.sb_logsunit > 1 &&
 664		   mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
 665		xfs_warn(mp,
 666		"log stripe unit %u bytes must be a multiple of block size",
 667			 mp->m_sb.sb_logsunit);
 668		error = -EINVAL;
 669		fatal = true;
 670	}
 671	if (error) {
 672		/*
 673		 * Log check errors are always fatal on v5; or whenever bad
 674		 * metadata leads to a crash.
 675		 */
 676		if (fatal) {
 677			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 678			ASSERT(0);
 
 679			goto out_free_log;
 680		}
 681		xfs_crit(mp, "Log size out of supported range.");
 682		xfs_crit(mp,
 683"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 684	}
 685
 686	/*
 687	 * Initialize the AIL now we have a log.
 688	 */
 689	error = xfs_trans_ail_init(mp);
 690	if (error) {
 691		xfs_warn(mp, "AIL initialisation failed: error %d", error);
 692		goto out_free_log;
 693	}
 694	mp->m_log->l_ailp = mp->m_ail;
 695
 696	/*
 697	 * skip log recovery on a norecovery mount.  pretend it all
 698	 * just worked.
 699	 */
 700	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 701		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 702
 703		if (readonly)
 704			mp->m_flags &= ~XFS_MOUNT_RDONLY;
 705
 706		error = xlog_recover(mp->m_log);
 707
 708		if (readonly)
 709			mp->m_flags |= XFS_MOUNT_RDONLY;
 710		if (error) {
 711			xfs_warn(mp, "log mount/recovery failed: error %d",
 712				error);
 713			xlog_recover_cancel(mp->m_log);
 714			goto out_destroy_ail;
 715		}
 716	}
 717
 718	error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 719			       "log");
 720	if (error)
 721		goto out_destroy_ail;
 722
 723	/* Normal transactions can now occur */
 724	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 725
 726	/*
 727	 * Now the log has been fully initialised and we know were our
 728	 * space grant counters are, we can initialise the permanent ticket
 729	 * needed for delayed logging to work.
 730	 */
 731	xlog_cil_init_post_recovery(mp->m_log);
 732
 733	return 0;
 734
 735out_destroy_ail:
 736	xfs_trans_ail_destroy(mp);
 737out_free_log:
 738	xlog_dealloc_log(mp->m_log);
 739out:
 740	return error;
 741}
 742
 743/*
 744 * Finish the recovery of the file system.  This is separate from the
 745 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 746 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 747 * here.
 748 *
 749 * If we finish recovery successfully, start the background log work. If we are
 750 * not doing recovery, then we have a RO filesystem and we don't need to start
 751 * it.
 752 */
 753int
 754xfs_log_mount_finish(
 755	struct xfs_mount	*mp)
 756{
 757	int	error = 0;
 758	bool	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 759	bool	recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
 760
 761	if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 762		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 763		return 0;
 764	} else if (readonly) {
 765		/* Allow unlinked processing to proceed */
 766		mp->m_flags &= ~XFS_MOUNT_RDONLY;
 767	}
 768
 769	/*
 770	 * During the second phase of log recovery, we need iget and
 771	 * iput to behave like they do for an active filesystem.
 772	 * xfs_fs_drop_inode needs to be able to prevent the deletion
 773	 * of inodes before we're done replaying log items on those
 774	 * inodes.  Turn it off immediately after recovery finishes
 775	 * so that we don't leak the quota inodes if subsequent mount
 776	 * activities fail.
 777	 *
 778	 * We let all inodes involved in redo item processing end up on
 779	 * the LRU instead of being evicted immediately so that if we do
 780	 * something to an unlinked inode, the irele won't cause
 781	 * premature truncation and freeing of the inode, which results
 782	 * in log recovery failure.  We have to evict the unreferenced
 783	 * lru inodes after clearing SB_ACTIVE because we don't
 784	 * otherwise clean up the lru if there's a subsequent failure in
 785	 * xfs_mountfs, which leads to us leaking the inodes if nothing
 786	 * else (e.g. quotacheck) references the inodes before the
 787	 * mount failure occurs.
 788	 */
 789	mp->m_super->s_flags |= SB_ACTIVE;
 790	error = xlog_recover_finish(mp->m_log);
 791	if (!error)
 792		xfs_log_work_queue(mp);
 793	mp->m_super->s_flags &= ~SB_ACTIVE;
 794	evict_inodes(mp->m_super);
 795
 796	/*
 797	 * Drain the buffer LRU after log recovery. This is required for v4
 798	 * filesystems to avoid leaving around buffers with NULL verifier ops,
 799	 * but we do it unconditionally to make sure we're always in a clean
 800	 * cache state after mount.
 801	 *
 802	 * Don't push in the error case because the AIL may have pending intents
 803	 * that aren't removed until recovery is cancelled.
 804	 */
 805	if (!error && recovered) {
 806		xfs_log_force(mp, XFS_LOG_SYNC);
 807		xfs_ail_push_all_sync(mp->m_ail);
 
 
 
 
 
 
 808	}
 809	xfs_wait_buftarg(mp->m_ddev_targp);
 
 
 810
 811	if (readonly)
 812		mp->m_flags |= XFS_MOUNT_RDONLY;
 813
 814	return error;
 815}
 816
 817/*
 818 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 819 * the log.
 820 */
 821int
 822xfs_log_mount_cancel(
 823	struct xfs_mount	*mp)
 824{
 825	int			error;
 826
 827	error = xlog_recover_cancel(mp->m_log);
 828	xfs_log_unmount(mp);
 
 829
 830	return error;
 
 
 
 
 
 
 
 
 
 
 
 
 831}
 832
 833/*
 834 * Final log writes as part of unmount.
 835 *
 836 * Mark the filesystem clean as unmount happens.  Note that during relocation
 837 * this routine needs to be executed as part of source-bag while the
 838 * deallocation must not be done until source-end.
 839 */
 
 
 
 
 
 
 
 
 
 
 
 
 840
 841/*
 842 * Unmount record used to have a string "Unmount filesystem--" in the
 843 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 844 * We just write the magic number now since that particular field isn't
 845 * currently architecture converted and "Unmount" is a bit foo.
 846 * As far as I know, there weren't any dependencies on the old behaviour.
 847 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 848
 
 
 
 
 
 849static int
 850xfs_log_unmount_write(xfs_mount_t *mp)
 
 
 851{
 852	struct xlog	 *log = mp->m_log;
 853	xlog_in_core_t	 *iclog;
 854#ifdef DEBUG
 855	xlog_in_core_t	 *first_iclog;
 856#endif
 857	xlog_ticket_t	*tic = NULL;
 858	xfs_lsn_t	 lsn;
 859	int		 error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 860
 
 861	/*
 862	 * Don't write out unmount record on norecovery mounts or ro devices.
 863	 * Or, if we are doing a forced umount (typically because of IO errors).
 864	 */
 865	if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
 866	    xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
 867		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 868		return 0;
 869	}
 870
 871	error = xfs_log_force(mp, XFS_LOG_SYNC);
 872	ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 873
 874#ifdef DEBUG
 875	first_iclog = iclog = log->l_iclog;
 876	do {
 877		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 878			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 879			ASSERT(iclog->ic_offset == 0);
 880		}
 881		iclog = iclog->ic_next;
 882	} while (iclog != first_iclog);
 883#endif
 884	if (! (XLOG_FORCED_SHUTDOWN(log))) {
 885		error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 886		if (!error) {
 887			/* the data section must be 32 bit size aligned */
 888			struct {
 889			    uint16_t magic;
 890			    uint16_t pad1;
 891			    uint32_t pad2; /* may as well make it 64 bits */
 892			} magic = {
 893				.magic = XLOG_UNMOUNT_TYPE,
 894			};
 895			struct xfs_log_iovec reg = {
 896				.i_addr = &magic,
 897				.i_len = sizeof(magic),
 898				.i_type = XLOG_REG_TYPE_UNMOUNT,
 899			};
 900			struct xfs_log_vec vec = {
 901				.lv_niovecs = 1,
 902				.lv_iovecp = &reg,
 903			};
 904
 905			/* remove inited flag, and account for space used */
 906			tic->t_flags = 0;
 907			tic->t_curr_res -= sizeof(magic);
 908			error = xlog_write(log, &vec, tic, &lsn,
 909					   NULL, XLOG_UNMOUNT_TRANS);
 910			/*
 911			 * At this point, we're umounting anyway,
 912			 * so there's no point in transitioning log state
 913			 * to IOERROR. Just continue...
 914			 */
 915		}
 916
 917		if (error)
 918			xfs_alert(mp, "%s: unmount record failed", __func__);
 
 
 
 919
 
 
 
 
 
 920
 921		spin_lock(&log->l_icloglock);
 922		iclog = log->l_iclog;
 923		atomic_inc(&iclog->ic_refcnt);
 924		xlog_state_want_sync(log, iclog);
 925		spin_unlock(&log->l_icloglock);
 926		error = xlog_state_release_iclog(log, iclog);
 927
 928		spin_lock(&log->l_icloglock);
 929		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 930		      iclog->ic_state == XLOG_STATE_DIRTY)) {
 931			if (!XLOG_FORCED_SHUTDOWN(log)) {
 932				xlog_wait(&iclog->ic_force_wait,
 933							&log->l_icloglock);
 934			} else {
 935				spin_unlock(&log->l_icloglock);
 936			}
 937		} else {
 938			spin_unlock(&log->l_icloglock);
 939		}
 940		if (tic) {
 941			trace_xfs_log_umount_write(log, tic);
 942			xlog_ungrant_log_space(log, tic);
 943			xfs_log_ticket_put(tic);
 944		}
 945	} else {
 946		/*
 947		 * We're already in forced_shutdown mode, couldn't
 948		 * even attempt to write out the unmount transaction.
 949		 *
 950		 * Go through the motions of sync'ing and releasing
 951		 * the iclog, even though no I/O will actually happen,
 952		 * we need to wait for other log I/Os that may already
 953		 * be in progress.  Do this as a separate section of
 954		 * code so we'll know if we ever get stuck here that
 955		 * we're in this odd situation of trying to unmount
 956		 * a file system that went into forced_shutdown as
 957		 * the result of an unmount..
 958		 */
 959		spin_lock(&log->l_icloglock);
 960		iclog = log->l_iclog;
 961		atomic_inc(&iclog->ic_refcnt);
 962
 963		xlog_state_want_sync(log, iclog);
 964		spin_unlock(&log->l_icloglock);
 965		error =  xlog_state_release_iclog(log, iclog);
 966
 967		spin_lock(&log->l_icloglock);
 968
 969		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 970			|| iclog->ic_state == XLOG_STATE_DIRTY
 971			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
 972
 973				xlog_wait(&iclog->ic_force_wait,
 974							&log->l_icloglock);
 975		} else {
 976			spin_unlock(&log->l_icloglock);
 977		}
 
 
 
 
 
 
 978	}
 979
 980	return error;
 981}	/* xfs_log_unmount_write */
 
 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 write the unmount
 990 * record.
 991 */
 992void
 993xfs_log_quiesce(
 994	struct xfs_mount	*mp)
 995{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 996	cancel_delayed_work_sync(&mp->m_log->l_work);
 997	xfs_log_force(mp, XFS_LOG_SYNC);
 998
 999	/*
1000	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1001	 * will push it, xfs_wait_buftarg() will not wait for it. Further,
1002	 * xfs_buf_iowait() cannot be used because it was pushed with the
1003	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1004	 * the IO to complete.
1005	 */
1006	xfs_ail_push_all_sync(mp->m_ail);
1007	xfs_wait_buftarg(mp->m_ddev_targp);
1008	xfs_buf_lock(mp->m_sb_bp);
1009	xfs_buf_unlock(mp->m_sb_bp);
1010
 
 
 
 
 
 
 
 
1011	xfs_log_unmount_write(mp);
1012}
1013
1014/*
1015 * Shut down and release the AIL and Log.
1016 *
1017 * During unmount, we need to ensure we flush all the dirty metadata objects
1018 * from the AIL so that the log is empty before we write the unmount record to
1019 * the log. Once this is done, we can tear down the AIL and the log.
1020 */
1021void
1022xfs_log_unmount(
1023	struct xfs_mount	*mp)
1024{
1025	xfs_log_quiesce(mp);
 
 
 
 
 
 
 
 
 
 
1026
1027	xfs_trans_ail_destroy(mp);
1028
1029	xfs_sysfs_del(&mp->m_log->l_kobj);
1030
1031	xlog_dealloc_log(mp->m_log);
1032}
1033
1034void
1035xfs_log_item_init(
1036	struct xfs_mount	*mp,
1037	struct xfs_log_item	*item,
1038	int			type,
1039	const struct xfs_item_ops *ops)
1040{
1041	item->li_mountp = mp;
1042	item->li_ailp = mp->m_ail;
1043	item->li_type = type;
1044	item->li_ops = ops;
1045	item->li_lv = NULL;
1046
1047	INIT_LIST_HEAD(&item->li_ail);
1048	INIT_LIST_HEAD(&item->li_cil);
1049	INIT_LIST_HEAD(&item->li_bio_list);
 
1050}
1051
1052/*
1053 * Wake up processes waiting for log space after we have moved the log tail.
1054 */
1055void
1056xfs_log_space_wake(
1057	struct xfs_mount	*mp)
1058{
1059	struct xlog		*log = mp->m_log;
1060	int			free_bytes;
1061
1062	if (XLOG_FORCED_SHUTDOWN(log))
1063		return;
1064
1065	if (!list_empty_careful(&log->l_write_head.waiters)) {
1066		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1067
1068		spin_lock(&log->l_write_head.lock);
1069		free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1070		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1071		spin_unlock(&log->l_write_head.lock);
1072	}
1073
1074	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1075		ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1076
1077		spin_lock(&log->l_reserve_head.lock);
1078		free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1079		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1080		spin_unlock(&log->l_reserve_head.lock);
1081	}
1082}
1083
1084/*
1085 * Determine if we have a transaction that has gone to disk that needs to be
1086 * covered. To begin the transition to the idle state firstly the log needs to
1087 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1088 * we start attempting to cover the log.
1089 *
1090 * Only if we are then in a state where covering is needed, the caller is
1091 * informed that dummy transactions are required to move the log into the idle
1092 * state.
1093 *
1094 * If there are any items in the AIl or CIL, then we do not want to attempt to
1095 * cover the log as we may be in a situation where there isn't log space
1096 * available to run a dummy transaction and this can lead to deadlocks when the
1097 * tail of the log is pinned by an item that is modified in the CIL.  Hence
1098 * there's no point in running a dummy transaction at this point because we
1099 * can't start trying to idle the log until both the CIL and AIL are empty.
1100 */
1101static int
1102xfs_log_need_covered(xfs_mount_t *mp)
 
1103{
1104	struct xlog	*log = mp->m_log;
1105	int		needed = 0;
1106
1107	if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1108		return 0;
1109
1110	if (!xlog_cil_empty(log))
1111		return 0;
1112
1113	spin_lock(&log->l_icloglock);
1114	switch (log->l_covered_state) {
1115	case XLOG_STATE_COVER_DONE:
1116	case XLOG_STATE_COVER_DONE2:
1117	case XLOG_STATE_COVER_IDLE:
1118		break;
1119	case XLOG_STATE_COVER_NEED:
1120	case XLOG_STATE_COVER_NEED2:
1121		if (xfs_ail_min_lsn(log->l_ailp))
1122			break;
1123		if (!xlog_iclogs_empty(log))
1124			break;
1125
1126		needed = 1;
1127		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1128			log->l_covered_state = XLOG_STATE_COVER_DONE;
1129		else
1130			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1131		break;
1132	default:
1133		needed = 1;
1134		break;
1135	}
1136	spin_unlock(&log->l_icloglock);
1137	return needed;
1138}
1139
1140/*
1141 * We may be holding the log iclog lock upon entering this routine.
 
 
 
1142 */
1143xfs_lsn_t
1144xlog_assign_tail_lsn_locked(
1145	struct xfs_mount	*mp)
1146{
1147	struct xlog		*log = mp->m_log;
1148	struct xfs_log_item	*lip;
1149	xfs_lsn_t		tail_lsn;
 
 
 
1150
1151	assert_spin_locked(&mp->m_ail->ail_lock);
 
1152
1153	/*
1154	 * To make sure we always have a valid LSN for the log tail we keep
1155	 * track of the last LSN which was committed in log->l_last_sync_lsn,
1156	 * and use that when the AIL was empty.
 
 
 
 
 
 
1157	 */
1158	lip = xfs_ail_min(mp->m_ail);
1159	if (lip)
1160		tail_lsn = lip->li_lsn;
1161	else
1162		tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1163	trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1164	atomic64_set(&log->l_tail_lsn, tail_lsn);
1165	return tail_lsn;
1166}
1167
1168xfs_lsn_t
1169xlog_assign_tail_lsn(
1170	struct xfs_mount	*mp)
1171{
1172	xfs_lsn_t		tail_lsn;
1173
1174	spin_lock(&mp->m_ail->ail_lock);
1175	tail_lsn = xlog_assign_tail_lsn_locked(mp);
1176	spin_unlock(&mp->m_ail->ail_lock);
1177
1178	return tail_lsn;
1179}
 
 
 
 
 
 
 
 
 
 
 
 
 
1180
1181/*
1182 * Return the space in the log between the tail and the head.  The head
1183 * is passed in the cycle/bytes formal parms.  In the special case where
1184 * the reserve head has wrapped passed the tail, this calculation is no
1185 * longer valid.  In this case, just return 0 which means there is no space
1186 * in the log.  This works for all places where this function is called
1187 * with the reserve head.  Of course, if the write head were to ever
1188 * wrap the tail, we should blow up.  Rather than catch this case here,
1189 * we depend on other ASSERTions in other parts of the code.   XXXmiken
1190 *
1191 * This code also handles the case where the reservation head is behind
1192 * the tail.  The details of this case are described below, but the end
1193 * result is that we return the size of the log as the amount of space left.
1194 */
1195STATIC int
1196xlog_space_left(
1197	struct xlog	*log,
1198	atomic64_t	*head)
1199{
1200	int		free_bytes;
1201	int		tail_bytes;
1202	int		tail_cycle;
1203	int		head_cycle;
1204	int		head_bytes;
1205
1206	xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1207	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1208	tail_bytes = BBTOB(tail_bytes);
1209	if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1210		free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1211	else if (tail_cycle + 1 < head_cycle)
1212		return 0;
1213	else if (tail_cycle < head_cycle) {
1214		ASSERT(tail_cycle == (head_cycle - 1));
1215		free_bytes = tail_bytes - head_bytes;
1216	} else {
1217		/*
1218		 * The reservation head is behind the tail.
1219		 * In this case we just want to return the size of the
1220		 * log as the amount of space left.
1221		 */
1222		xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1223		xfs_alert(log->l_mp,
1224			  "  tail_cycle = %d, tail_bytes = %d",
1225			  tail_cycle, tail_bytes);
1226		xfs_alert(log->l_mp,
1227			  "  GH   cycle = %d, GH   bytes = %d",
1228			  head_cycle, head_bytes);
1229		ASSERT(0);
1230		free_bytes = log->l_logsize;
1231	}
1232	return free_bytes;
1233}
1234
1235
1236/*
1237 * Log function which is called when an io completes.
1238 *
1239 * The log manager needs its own routine, in order to control what
1240 * happens with the buffer after the write completes.
1241 */
1242static void
1243xlog_iodone(xfs_buf_t *bp)
 
1244{
1245	struct xlog_in_core	*iclog = bp->b_log_item;
1246	struct xlog		*l = iclog->ic_log;
1247	int			aborted = 0;
 
 
 
 
 
 
 
 
1248
1249	/*
1250	 * Race to shutdown the filesystem if we see an error or the iclog is in
1251	 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1252	 * CRC errors into log recovery.
1253	 */
1254	if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR) ||
1255	    iclog->ic_state & XLOG_STATE_IOABORT) {
1256		if (iclog->ic_state & XLOG_STATE_IOABORT)
1257			iclog->ic_state &= ~XLOG_STATE_IOABORT;
1258
1259		xfs_buf_ioerror_alert(bp, __func__);
1260		xfs_buf_stale(bp);
1261		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1262		/*
1263		 * This flag will be propagated to the trans-committed
1264		 * callback routines to let them know that the log-commit
1265		 * didn't succeed.
1266		 */
1267		aborted = XFS_LI_ABORTED;
1268	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1269		aborted = XFS_LI_ABORTED;
1270	}
1271
1272	/* log I/O is always issued ASYNC */
1273	ASSERT(bp->b_flags & XBF_ASYNC);
1274	xlog_state_done_syncing(iclog, aborted);
1275
1276	/*
1277	 * drop the buffer lock now that we are done. Nothing references
1278	 * the buffer after this, so an unmount waiting on this lock can now
1279	 * tear it down safely. As such, it is unsafe to reference the buffer
1280	 * (bp) after the unlock as we could race with it being freed.
1281	 */
1282	xfs_buf_unlock(bp);
1283}
1284
1285/*
1286 * Return size of each in-core log record buffer.
1287 *
1288 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1289 *
1290 * If the filesystem blocksize is too large, we may need to choose a
1291 * larger size since the directory code currently logs entire blocks.
1292 */
1293
1294STATIC void
1295xlog_get_iclog_buffer_size(
1296	struct xfs_mount	*mp,
1297	struct xlog		*log)
1298{
1299	int size;
1300	int xhdrs;
1301
1302	if (mp->m_logbufs <= 0)
1303		log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1304	else
1305		log->l_iclog_bufs = mp->m_logbufs;
 
 
 
1306
1307	/*
1308	 * Buffer size passed in from mount system call.
1309	 */
1310	if (mp->m_logbsize > 0) {
1311		size = log->l_iclog_size = mp->m_logbsize;
1312		log->l_iclog_size_log = 0;
1313		while (size != 1) {
1314			log->l_iclog_size_log++;
1315			size >>= 1;
1316		}
1317
1318		if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1319			/* # headers = size / 32k
1320			 * one header holds cycles from 32k of data
1321			 */
1322
1323			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1324			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1325				xhdrs++;
1326			log->l_iclog_hsize = xhdrs << BBSHIFT;
1327			log->l_iclog_heads = xhdrs;
1328		} else {
1329			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1330			log->l_iclog_hsize = BBSIZE;
1331			log->l_iclog_heads = 1;
1332		}
1333		goto done;
1334	}
1335
1336	/* All machines use 32kB buffers by default. */
1337	log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1338	log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1339
1340	/* the default log size is 16k or 32k which is one header sector */
1341	log->l_iclog_hsize = BBSIZE;
1342	log->l_iclog_heads = 1;
1343
1344done:
1345	/* are we being asked to make the sizes selected above visible? */
1346	if (mp->m_logbufs == 0)
1347		mp->m_logbufs = log->l_iclog_bufs;
1348	if (mp->m_logbsize == 0)
1349		mp->m_logbsize = log->l_iclog_size;
1350}	/* xlog_get_iclog_buffer_size */
1351
1352
1353void
1354xfs_log_work_queue(
1355	struct xfs_mount        *mp)
1356{
1357	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1358				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1359}
1360
1361/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1362 * Every sync period we need to unpin all items in the AIL and push them to
1363 * disk. If there is nothing dirty, then we might need to cover the log to
1364 * indicate that the filesystem is idle.
1365 */
1366static void
1367xfs_log_worker(
1368	struct work_struct	*work)
1369{
1370	struct xlog		*log = container_of(to_delayed_work(work),
1371						struct xlog, l_work);
1372	struct xfs_mount	*mp = log->l_mp;
1373
1374	/* dgc: errors ignored - not fatal and nowhere to report them */
1375	if (xfs_log_need_covered(mp)) {
1376		/*
1377		 * Dump a transaction into the log that contains no real change.
1378		 * This is needed to stamp the current tail LSN into the log
1379		 * during the covering operation.
1380		 *
1381		 * We cannot use an inode here for this - that will push dirty
1382		 * state back up into the VFS and then periodic inode flushing
1383		 * will prevent log covering from making progress. Hence we
1384		 * synchronously log the superblock instead to ensure the
1385		 * superblock is immediately unpinned and can be written back.
1386		 */
 
1387		xfs_sync_sb(mp, true);
1388	} else
1389		xfs_log_force(mp, 0);
1390
1391	/* start pushing all the metadata that is currently dirty */
1392	xfs_ail_push_all(mp->m_ail);
1393
1394	/* queue us up again */
1395	xfs_log_work_queue(mp);
1396}
1397
1398/*
1399 * This routine initializes some of the log structure for a given mount point.
1400 * Its primary purpose is to fill in enough, so recovery can occur.  However,
1401 * some other stuff may be filled in too.
1402 */
1403STATIC struct xlog *
1404xlog_alloc_log(
1405	struct xfs_mount	*mp,
1406	struct xfs_buftarg	*log_target,
1407	xfs_daddr_t		blk_offset,
1408	int			num_bblks)
1409{
1410	struct xlog		*log;
1411	xlog_rec_header_t	*head;
1412	xlog_in_core_t		**iclogp;
1413	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1414	xfs_buf_t		*bp;
1415	int			i;
1416	int			error = -ENOMEM;
1417	uint			log2_size = 0;
1418
1419	log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1420	if (!log) {
1421		xfs_warn(mp, "Log allocation failed: No memory!");
1422		goto out;
1423	}
1424
1425	log->l_mp	   = mp;
1426	log->l_targ	   = log_target;
1427	log->l_logsize     = BBTOB(num_bblks);
1428	log->l_logBBstart  = blk_offset;
1429	log->l_logBBsize   = num_bblks;
1430	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1431	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1432	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
 
1433
1434	log->l_prev_block  = -1;
1435	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1436	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1437	xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1438	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1439
 
 
 
 
 
1440	xlog_grant_head_init(&log->l_reserve_head);
1441	xlog_grant_head_init(&log->l_write_head);
1442
1443	error = -EFSCORRUPTED;
1444	if (xfs_sb_version_hassector(&mp->m_sb)) {
1445	        log2_size = mp->m_sb.sb_logsectlog;
1446		if (log2_size < BBSHIFT) {
1447			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1448				log2_size, BBSHIFT);
1449			goto out_free_log;
1450		}
1451
1452	        log2_size -= BBSHIFT;
1453		if (log2_size > mp->m_sectbb_log) {
1454			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1455				log2_size, mp->m_sectbb_log);
1456			goto out_free_log;
1457		}
1458
1459		/* for larger sector sizes, must have v2 or external log */
1460		if (log2_size && log->l_logBBstart > 0 &&
1461			    !xfs_sb_version_haslogv2(&mp->m_sb)) {
1462			xfs_warn(mp,
1463		"log sector size (0x%x) invalid for configuration.",
1464				log2_size);
1465			goto out_free_log;
1466		}
1467	}
1468	log->l_sectBBsize = 1 << log2_size;
1469
1470	xlog_get_iclog_buffer_size(mp, log);
1471
1472	/*
1473	 * Use a NULL block for the extra log buffer used during splits so that
1474	 * it will trigger errors if we ever try to do IO on it without first
1475	 * having set it up properly.
1476	 */
1477	error = -ENOMEM;
1478	bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1479			   BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1480	if (!bp)
1481		goto out_free_log;
1482
1483	/*
1484	 * The iclogbuf buffer locks are held over IO but we are not going to do
1485	 * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1486	 * when appropriately.
1487	 */
1488	ASSERT(xfs_buf_islocked(bp));
1489	xfs_buf_unlock(bp);
1490
1491	/* use high priority wq for log I/O completion */
1492	bp->b_ioend_wq = mp->m_log_workqueue;
1493	bp->b_iodone = xlog_iodone;
1494	log->l_xbuf = bp;
1495
1496	spin_lock_init(&log->l_icloglock);
1497	init_waitqueue_head(&log->l_flush_wait);
1498
1499	iclogp = &log->l_iclog;
1500	/*
1501	 * The amount of memory to allocate for the iclog structure is
1502	 * rather funky due to the way the structure is defined.  It is
1503	 * done this way so that we can use different sizes for machines
1504	 * with different amounts of memory.  See the definition of
1505	 * xlog_in_core_t in xfs_log_priv.h for details.
1506	 */
1507	ASSERT(log->l_iclog_size >= 4096);
1508	for (i=0; i < log->l_iclog_bufs; i++) {
1509		*iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1510		if (!*iclogp)
 
 
 
 
1511			goto out_free_iclog;
1512
1513		iclog = *iclogp;
1514		iclog->ic_prev = prev_iclog;
1515		prev_iclog = iclog;
1516
1517		bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1518					  BTOBB(log->l_iclog_size),
1519					  XBF_NO_IOACCT);
1520		if (!bp)
1521			goto out_free_iclog;
1522
1523		ASSERT(xfs_buf_islocked(bp));
1524		xfs_buf_unlock(bp);
1525
1526		/* use high priority wq for log I/O completion */
1527		bp->b_ioend_wq = mp->m_log_workqueue;
1528		bp->b_iodone = xlog_iodone;
1529		iclog->ic_bp = bp;
1530		iclog->ic_data = bp->b_addr;
1531#ifdef DEBUG
1532		log->l_iclog_bak[i] = &iclog->ic_header;
1533#endif
1534		head = &iclog->ic_header;
1535		memset(head, 0, sizeof(xlog_rec_header_t));
1536		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1537		head->h_version = cpu_to_be32(
1538			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1539		head->h_size = cpu_to_be32(log->l_iclog_size);
1540		/* new fields */
1541		head->h_fmt = cpu_to_be32(XLOG_FMT);
1542		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1543
1544		iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1545		iclog->ic_state = XLOG_STATE_ACTIVE;
1546		iclog->ic_log = log;
1547		atomic_set(&iclog->ic_refcnt, 0);
1548		spin_lock_init(&iclog->ic_callback_lock);
1549		iclog->ic_callback_tail = &(iclog->ic_callback);
1550		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1551
1552		init_waitqueue_head(&iclog->ic_force_wait);
1553		init_waitqueue_head(&iclog->ic_write_wait);
 
 
1554
1555		iclogp = &iclog->ic_next;
1556	}
1557	*iclogp = log->l_iclog;			/* complete ring */
1558	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1559
 
 
 
 
 
 
 
1560	error = xlog_cil_init(log);
1561	if (error)
1562		goto out_free_iclog;
1563	return log;
1564
 
 
1565out_free_iclog:
1566	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1567		prev_iclog = iclog->ic_next;
1568		if (iclog->ic_bp)
1569			xfs_buf_free(iclog->ic_bp);
1570		kmem_free(iclog);
 
1571	}
1572	spinlock_destroy(&log->l_icloglock);
1573	xfs_buf_free(log->l_xbuf);
1574out_free_log:
1575	kmem_free(log);
1576out:
1577	return ERR_PTR(error);
1578}	/* xlog_alloc_log */
1579
1580
1581/*
1582 * Write out the commit record of a transaction associated with the given
1583 * ticket.  Return the lsn of the commit record.
1584 */
1585STATIC int
1586xlog_commit_record(
1587	struct xlog		*log,
1588	struct xlog_ticket	*ticket,
1589	struct xlog_in_core	**iclog,
1590	xfs_lsn_t		*commitlsnp)
1591{
1592	struct xfs_mount *mp = log->l_mp;
1593	int	error;
1594	struct xfs_log_iovec reg = {
1595		.i_addr = NULL,
1596		.i_len = 0,
1597		.i_type = XLOG_REG_TYPE_COMMIT,
1598	};
1599	struct xfs_log_vec vec = {
1600		.lv_niovecs = 1,
1601		.lv_iovecp = &reg,
1602	};
1603
1604	ASSERT_ALWAYS(iclog);
1605	error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1606					XLOG_COMMIT_TRANS);
1607	if (error)
1608		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1609	return error;
1610}
1611
1612/*
1613 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1614 * log space.  This code pushes on the lsn which would supposedly free up
1615 * the 25% which we want to leave free.  We may need to adopt a policy which
1616 * pushes on an lsn which is further along in the log once we reach the high
1617 * water mark.  In this manner, we would be creating a low water mark.
1618 */
1619STATIC void
1620xlog_grant_push_ail(
1621	struct xlog	*log,
1622	int		need_bytes)
1623{
1624	xfs_lsn_t	threshold_lsn = 0;
1625	xfs_lsn_t	last_sync_lsn;
1626	int		free_blocks;
1627	int		free_bytes;
1628	int		threshold_block;
1629	int		threshold_cycle;
1630	int		free_threshold;
1631
1632	ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1633
1634	free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1635	free_blocks = BTOBBT(free_bytes);
1636
1637	/*
1638	 * Set the threshold for the minimum number of free blocks in the
1639	 * log to the maximum of what the caller needs, one quarter of the
1640	 * log, and 256 blocks.
1641	 */
1642	free_threshold = BTOBB(need_bytes);
1643	free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1644	free_threshold = MAX(free_threshold, 256);
1645	if (free_blocks >= free_threshold)
1646		return;
1647
1648	xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1649						&threshold_block);
1650	threshold_block += free_threshold;
1651	if (threshold_block >= log->l_logBBsize) {
1652		threshold_block -= log->l_logBBsize;
1653		threshold_cycle += 1;
1654	}
1655	threshold_lsn = xlog_assign_lsn(threshold_cycle,
1656					threshold_block);
1657	/*
1658	 * Don't pass in an lsn greater than the lsn of the last
1659	 * log record known to be on disk. Use a snapshot of the last sync lsn
1660	 * so that it doesn't change between the compare and the set.
1661	 */
1662	last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1663	if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1664		threshold_lsn = last_sync_lsn;
1665
1666	/*
1667	 * Get the transaction layer to kick the dirty buffers out to
1668	 * disk asynchronously. No point in trying to do this if
1669	 * the filesystem is shutting down.
1670	 */
1671	if (!XLOG_FORCED_SHUTDOWN(log))
1672		xfs_ail_push(log->l_ailp, threshold_lsn);
1673}
1674
1675/*
1676 * Stamp cycle number in every block
1677 */
1678STATIC void
1679xlog_pack_data(
1680	struct xlog		*log,
1681	struct xlog_in_core	*iclog,
1682	int			roundoff)
1683{
1684	int			i, j, k;
1685	int			size = iclog->ic_offset + roundoff;
1686	__be32			cycle_lsn;
1687	char			*dp;
1688
1689	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1690
1691	dp = iclog->ic_datap;
1692	for (i = 0; i < BTOBB(size); i++) {
1693		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1694			break;
1695		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1696		*(__be32 *)dp = cycle_lsn;
1697		dp += BBSIZE;
1698	}
1699
1700	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1701		xlog_in_core_2_t *xhdr = iclog->ic_data;
1702
1703		for ( ; i < BTOBB(size); i++) {
1704			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1705			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1706			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1707			*(__be32 *)dp = cycle_lsn;
1708			dp += BBSIZE;
1709		}
1710
1711		for (i = 1; i < log->l_iclog_heads; i++)
1712			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1713	}
1714}
1715
1716/*
1717 * Calculate the checksum for a log buffer.
1718 *
1719 * This is a little more complicated than it should be because the various
1720 * headers and the actual data are non-contiguous.
1721 */
1722__le32
1723xlog_cksum(
1724	struct xlog		*log,
1725	struct xlog_rec_header	*rhead,
1726	char			*dp,
1727	int			size)
1728{
1729	uint32_t		crc;
1730
1731	/* first generate the crc for the record header ... */
1732	crc = xfs_start_cksum_update((char *)rhead,
1733			      sizeof(struct xlog_rec_header),
1734			      offsetof(struct xlog_rec_header, h_crc));
1735
1736	/* ... then for additional cycle data for v2 logs ... */
1737	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1738		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1739		int		i;
1740		int		xheads;
1741
1742		xheads = size / XLOG_HEADER_CYCLE_SIZE;
1743		if (size % XLOG_HEADER_CYCLE_SIZE)
1744			xheads++;
1745
1746		for (i = 1; i < xheads; i++) {
1747			crc = crc32c(crc, &xhdr[i].hic_xheader,
1748				     sizeof(struct xlog_rec_ext_header));
1749		}
1750	}
1751
1752	/* ... and finally for the payload */
1753	crc = crc32c(crc, dp, size);
1754
1755	return xfs_end_cksum(crc);
1756}
1757
1758/*
1759 * The bdstrat callback function for log bufs. This gives us a central
1760 * place to trap bufs in case we get hit by a log I/O error and need to
1761 * shutdown. Actually, in practice, even when we didn't get a log error,
1762 * we transition the iclogs to IOERROR state *after* flushing all existing
1763 * iclogs to disk. This is because we don't want anymore new transactions to be
1764 * started or completed afterwards.
1765 *
1766 * We lock the iclogbufs here so that we can serialise against IO completion
1767 * during unmount. We might be processing a shutdown triggered during unmount,
1768 * and that can occur asynchronously to the unmount thread, and hence we need to
1769 * ensure that completes before tearing down the iclogbufs. Hence we need to
1770 * hold the buffer lock across the log IO to acheive that.
1771 */
1772STATIC int
1773xlog_bdstrat(
1774	struct xfs_buf		*bp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1775{
1776	struct xlog_in_core	*iclog = bp->b_log_item;
 
1777
1778	xfs_buf_lock(bp);
1779	if (iclog->ic_state & XLOG_STATE_IOERROR) {
1780		xfs_buf_ioerror(bp, -EIO);
1781		xfs_buf_stale(bp);
1782		xfs_buf_ioend(bp);
 
 
 
 
 
1783		/*
1784		 * It would seem logical to return EIO here, but we rely on
1785		 * the log state machine to propagate I/O errors instead of
1786		 * doing it here. Similarly, IO completion will unlock the
1787		 * buffer, so we don't do it here.
 
1788		 */
1789		return 0;
1790	}
1791
1792	xfs_buf_submit(bp);
1793	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1794}
1795
1796/*
1797 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1798 * fashion.  Previously, we should have moved the current iclog
1799 * ptr in the log to point to the next available iclog.  This allows further
1800 * write to continue while this code syncs out an iclog ready to go.
1801 * Before an in-core log can be written out, the data section must be scanned
1802 * to save away the 1st word of each BBSIZE block into the header.  We replace
1803 * it with the current cycle count.  Each BBSIZE block is tagged with the
1804 * cycle count because there in an implicit assumption that drives will
1805 * guarantee that entire 512 byte blocks get written at once.  In other words,
1806 * we can't have part of a 512 byte block written and part not written.  By
1807 * tagging each block, we will know which blocks are valid when recovering
1808 * after an unclean shutdown.
1809 *
1810 * This routine is single threaded on the iclog.  No other thread can be in
1811 * this routine with the same iclog.  Changing contents of iclog can there-
1812 * fore be done without grabbing the state machine lock.  Updating the global
1813 * log will require grabbing the lock though.
1814 *
1815 * The entire log manager uses a logical block numbering scheme.  Only
1816 * log_sync (and then only bwrite()) know about the fact that the log may
1817 * not start with block zero on a given device.  The log block start offset
1818 * is added immediately before calling bwrite().
1819 */
1820
1821STATIC int
1822xlog_sync(
1823	struct xlog		*log,
1824	struct xlog_in_core	*iclog)
 
1825{
1826	xfs_buf_t	*bp;
1827	int		i;
1828	uint		count;		/* byte count of bwrite */
1829	uint		count_init;	/* initial count before roundup */
1830	int		roundoff;       /* roundoff to BB or stripe */
1831	int		split = 0;	/* split write into two regions */
1832	int		error;
1833	int		v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1834	int		size;
1835
1836	XFS_STATS_INC(log->l_mp, xs_log_writes);
1837	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
 
1838
1839	/* Add for LR header */
1840	count_init = log->l_iclog_hsize + iclog->ic_offset;
1841
1842	/* Round out the log write size */
1843	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1844		/* we have a v2 stripe unit to use */
1845		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
 
 
 
1846	} else {
1847		count = BBTOB(BTOBB(count_init));
 
1848	}
1849	roundoff = count - count_init;
1850	ASSERT(roundoff >= 0);
1851	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1852                roundoff < log->l_mp->m_sb.sb_logsunit)
1853		|| 
1854		(log->l_mp->m_sb.sb_logsunit <= 1 && 
1855		 roundoff < BBTOB(1)));
1856
1857	/* move grant heads by roundoff in sync */
1858	xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1859	xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1860
1861	/* put cycle number in every block */
1862	xlog_pack_data(log, iclog, roundoff); 
1863
1864	/* real byte length */
1865	size = iclog->ic_offset;
1866	if (v2)
1867		size += roundoff;
1868	iclog->ic_header.h_len = cpu_to_be32(size);
1869
1870	bp = iclog->ic_bp;
1871	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1872
1873	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1874
1875	/* Do we need to split this write into 2 parts? */
1876	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1877		char		*dptr;
1878
1879		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1880		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1881		iclog->ic_bwritecnt = 2;
1882
1883		/*
1884		 * Bump the cycle numbers at the start of each block in the
1885		 * part of the iclog that ends up in the buffer that gets
1886		 * written to the start of the log.
1887		 *
1888		 * Watch out for the header magic number case, though.
1889		 */
1890		dptr = (char *)&iclog->ic_header + count;
1891		for (i = 0; i < split; i += BBSIZE) {
1892			uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1893			if (++cycle == XLOG_HEADER_MAGIC_NUM)
1894				cycle++;
1895			*(__be32 *)dptr = cpu_to_be32(cycle);
1896
1897			dptr += BBSIZE;
1898		}
1899	} else {
1900		iclog->ic_bwritecnt = 1;
1901	}
1902
1903	/* calculcate the checksum */
1904	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1905					    iclog->ic_datap, size);
1906	/*
1907	 * Intentionally corrupt the log record CRC based on the error injection
1908	 * frequency, if defined. This facilitates testing log recovery in the
1909	 * event of torn writes. Hence, set the IOABORT state to abort the log
1910	 * write on I/O completion and shutdown the fs. The subsequent mount
1911	 * detects the bad CRC and attempts to recover.
1912	 */
 
1913	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1914		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1915		iclog->ic_state |= XLOG_STATE_IOABORT;
1916		xfs_warn(log->l_mp,
1917	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1918			 be64_to_cpu(iclog->ic_header.h_lsn));
1919	}
1920
1921	bp->b_io_length = BTOBB(count);
1922	bp->b_log_item = iclog;
1923	bp->b_flags &= ~XBF_FLUSH;
1924	bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1925
1926	/*
1927	 * Flush the data device before flushing the log to make sure all meta
1928	 * data written back from the AIL actually made it to disk before
1929	 * stamping the new log tail LSN into the log buffer.  For an external
1930	 * log we need to issue the flush explicitly, and unfortunately
1931	 * synchronously here; for an internal log we can simply use the block
1932	 * layer state machine for preflushes.
1933	 */
1934	if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1935		xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1936	else
1937		bp->b_flags |= XBF_FLUSH;
1938
1939	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1940	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1941
1942	xlog_verify_iclog(log, iclog, count, true);
1943
1944	/* account for log which doesn't start at block #0 */
1945	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1946
1947	/*
1948	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1949	 * is shutting down.
1950	 */
1951	error = xlog_bdstrat(bp);
1952	if (error) {
1953		xfs_buf_ioerror_alert(bp, "xlog_sync");
1954		return error;
1955	}
1956	if (split) {
1957		bp = iclog->ic_log->l_xbuf;
1958		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1959		xfs_buf_associate_memory(bp,
1960				(char *)&iclog->ic_header + count, split);
1961		bp->b_log_item = iclog;
1962		bp->b_flags &= ~XBF_FLUSH;
1963		bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1964
1965		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1966		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1967
1968		/* account for internal log which doesn't start at block #0 */
1969		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1970		error = xlog_bdstrat(bp);
1971		if (error) {
1972			xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1973			return error;
1974		}
1975	}
1976	return 0;
1977}	/* xlog_sync */
1978
1979/*
1980 * Deallocate a log structure
1981 */
1982STATIC void
1983xlog_dealloc_log(
1984	struct xlog	*log)
1985{
1986	xlog_in_core_t	*iclog, *next_iclog;
1987	int		i;
1988
1989	xlog_cil_destroy(log);
1990
1991	/*
1992	 * Cycle all the iclogbuf locks to make sure all log IO completion
1993	 * is done before we tear down these buffers.
 
1994	 */
1995	iclog = log->l_iclog;
1996	for (i = 0; i < log->l_iclog_bufs; i++) {
1997		xfs_buf_lock(iclog->ic_bp);
1998		xfs_buf_unlock(iclog->ic_bp);
1999		iclog = iclog->ic_next;
2000	}
2001
2002	/*
2003	 * Always need to ensure that the extra buffer does not point to memory
2004	 * owned by another log buffer before we free it. Also, cycle the lock
2005	 * first to ensure we've completed IO on it.
2006	 */
2007	xfs_buf_lock(log->l_xbuf);
2008	xfs_buf_unlock(log->l_xbuf);
2009	xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
2010	xfs_buf_free(log->l_xbuf);
2011
2012	iclog = log->l_iclog;
2013	for (i = 0; i < log->l_iclog_bufs; i++) {
2014		xfs_buf_free(iclog->ic_bp);
2015		next_iclog = iclog->ic_next;
2016		kmem_free(iclog);
 
2017		iclog = next_iclog;
2018	}
2019	spinlock_destroy(&log->l_icloglock);
2020
2021	log->l_mp->m_log = NULL;
2022	kmem_free(log);
2023}	/* xlog_dealloc_log */
 
2024
2025/*
2026 * Update counters atomically now that memcpy is done.
2027 */
2028/* ARGSUSED */
2029static inline void
2030xlog_state_finish_copy(
2031	struct xlog		*log,
2032	struct xlog_in_core	*iclog,
2033	int			record_cnt,
2034	int			copy_bytes)
2035{
2036	spin_lock(&log->l_icloglock);
2037
2038	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2039	iclog->ic_offset += copy_bytes;
2040
2041	spin_unlock(&log->l_icloglock);
2042}	/* xlog_state_finish_copy */
2043
2044
2045
2046
2047/*
2048 * print out info relating to regions written which consume
2049 * the reservation
2050 */
2051void
2052xlog_print_tic_res(
2053	struct xfs_mount	*mp,
2054	struct xlog_ticket	*ticket)
2055{
2056	uint i;
2057	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2058
2059	/* match with XLOG_REG_TYPE_* in xfs_log.h */
2060#define REG_TYPE_STR(type, str)	[XLOG_REG_TYPE_##type] = str
2061	static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2062	    REG_TYPE_STR(BFORMAT, "bformat"),
2063	    REG_TYPE_STR(BCHUNK, "bchunk"),
2064	    REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2065	    REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2066	    REG_TYPE_STR(IFORMAT, "iformat"),
2067	    REG_TYPE_STR(ICORE, "icore"),
2068	    REG_TYPE_STR(IEXT, "iext"),
2069	    REG_TYPE_STR(IBROOT, "ibroot"),
2070	    REG_TYPE_STR(ILOCAL, "ilocal"),
2071	    REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2072	    REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2073	    REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2074	    REG_TYPE_STR(QFORMAT, "qformat"),
2075	    REG_TYPE_STR(DQUOT, "dquot"),
2076	    REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2077	    REG_TYPE_STR(LRHEADER, "LR header"),
2078	    REG_TYPE_STR(UNMOUNT, "unmount"),
2079	    REG_TYPE_STR(COMMIT, "commit"),
2080	    REG_TYPE_STR(TRANSHDR, "trans header"),
2081	    REG_TYPE_STR(ICREATE, "inode create")
2082	};
2083#undef REG_TYPE_STR
2084
2085	xfs_warn(mp, "ticket reservation summary:");
2086	xfs_warn(mp, "  unit res    = %d bytes",
2087		 ticket->t_unit_res);
2088	xfs_warn(mp, "  current res = %d bytes",
2089		 ticket->t_curr_res);
2090	xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2091		 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2092	xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2093		 ticket->t_res_num_ophdrs, ophdr_spc);
2094	xfs_warn(mp, "  ophdr + reg = %u bytes",
2095		 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2096	xfs_warn(mp, "  num regions = %u",
2097		 ticket->t_res_num);
2098
2099	for (i = 0; i < ticket->t_res_num; i++) {
2100		uint r_type = ticket->t_res_arr[i].r_type;
2101		xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2102			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2103			    "bad-rtype" : res_type_str[r_type]),
2104			    ticket->t_res_arr[i].r_len);
2105	}
2106}
2107
2108/*
2109 * Print a summary of the transaction.
2110 */
2111void
2112xlog_print_trans(
2113	struct xfs_trans		*tp)
2114{
2115	struct xfs_mount		*mp = tp->t_mountp;
2116	struct xfs_log_item_desc	*lidp;
2117
2118	/* dump core transaction and ticket info */
2119	xfs_warn(mp, "transaction summary:");
2120	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2121	xfs_warn(mp, "  log count = %d", tp->t_log_count);
2122	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2123
2124	xlog_print_tic_res(mp, tp->t_ticket);
2125
2126	/* dump each log item */
2127	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
2128		struct xfs_log_item	*lip = lidp->lid_item;
2129		struct xfs_log_vec	*lv = lip->li_lv;
2130		struct xfs_log_iovec	*vec;
2131		int			i;
2132
2133		xfs_warn(mp, "log item: ");
2134		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
2135		xfs_warn(mp, "  flags	= 0x%x", lip->li_flags);
2136		if (!lv)
2137			continue;
2138		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
2139		xfs_warn(mp, "  size	= %d", lv->lv_size);
2140		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
2141		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
2142
2143		/* dump each iovec for the log item */
2144		vec = lv->lv_iovecp;
2145		for (i = 0; i < lv->lv_niovecs; i++) {
2146			int dumplen = min(vec->i_len, 32);
2147
2148			xfs_warn(mp, "  iovec[%d]", i);
2149			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
2150			xfs_warn(mp, "    len	= %d", vec->i_len);
2151			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2152			xfs_hex_dump(vec->i_addr, dumplen);
2153
2154			vec++;
2155		}
2156	}
2157}
2158
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2159/*
2160 * Calculate the potential space needed by the log vector.  Each region gets
2161 * its own xlog_op_header_t and may need to be double word aligned.
2162 */
2163static int
2164xlog_write_calc_vec_length(
 
2165	struct xlog_ticket	*ticket,
2166	struct xfs_log_vec	*log_vector)
 
 
 
 
2167{
2168	struct xfs_log_vec	*lv;
2169	int			headers = 0;
2170	int			len = 0;
2171	int			i;
2172
2173	/* acct for start rec of xact */
2174	if (ticket->t_flags & XLOG_TIC_INITED)
2175		headers++;
2176
2177	for (lv = log_vector; lv; lv = lv->lv_next) {
2178		/* we don't write ordered log vectors */
2179		if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2180			continue;
2181
2182		headers += lv->lv_niovecs;
 
2183
2184		for (i = 0; i < lv->lv_niovecs; i++) {
2185			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
 
 
 
 
 
2186
2187			len += vecp->i_len;
2188			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2189		}
2190	}
2191
2192	ticket->t_res_num_ophdrs += headers;
2193	len += headers * sizeof(struct xlog_op_header);
2194
2195	return len;
2196}
2197
2198/*
2199 * If first write for transaction, insert start record  We can't be trying to
2200 * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2201 */
2202static int
2203xlog_write_start_rec(
2204	struct xlog_op_header	*ophdr,
2205	struct xlog_ticket	*ticket)
2206{
2207	if (!(ticket->t_flags & XLOG_TIC_INITED))
2208		return 0;
2209
2210	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
2211	ophdr->oh_clientid = ticket->t_clientid;
2212	ophdr->oh_len = 0;
2213	ophdr->oh_flags = XLOG_START_TRANS;
2214	ophdr->oh_res2 = 0;
2215
2216	ticket->t_flags &= ~XLOG_TIC_INITED;
2217
2218	return sizeof(struct xlog_op_header);
2219}
2220
2221static xlog_op_header_t *
2222xlog_write_setup_ophdr(
2223	struct xlog		*log,
2224	struct xlog_op_header	*ophdr,
2225	struct xlog_ticket	*ticket,
2226	uint			flags)
 
 
 
 
2227{
2228	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2229	ophdr->oh_clientid = ticket->t_clientid;
2230	ophdr->oh_res2 = 0;
2231
2232	/* are we copying a commit or unmount record? */
2233	ophdr->oh_flags = flags;
2234
2235	/*
2236	 * We've seen logs corrupted with bad transaction client ids.  This
2237	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2238	 * and shut down the filesystem.
2239	 */
2240	switch (ophdr->oh_clientid)  {
2241	case XFS_TRANSACTION:
2242	case XFS_VOLUME:
2243	case XFS_LOG:
2244		break;
2245	default:
2246		xfs_warn(log->l_mp,
2247			"Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2248			ophdr->oh_clientid, ticket);
2249		return NULL;
2250	}
2251
2252	return ophdr;
 
 
 
 
 
 
 
2253}
2254
2255/*
2256 * Set up the parameters of the region copy into the log. This has
2257 * to handle region write split across multiple log buffers - this
2258 * state is kept external to this function so that this code can
2259 * be written in an obvious, self documenting manner.
2260 */
2261static int
2262xlog_write_setup_copy(
 
2263	struct xlog_ticket	*ticket,
2264	struct xlog_op_header	*ophdr,
2265	int			space_available,
2266	int			space_required,
2267	int			*copy_off,
2268	int			*copy_len,
2269	int			*last_was_partial_copy,
2270	int			*bytes_consumed)
2271{
2272	int			still_to_copy;
2273
2274	still_to_copy = space_required - *bytes_consumed;
2275	*copy_off = *bytes_consumed;
2276
2277	if (still_to_copy <= space_available) {
2278		/* write of region completes here */
2279		*copy_len = still_to_copy;
2280		ophdr->oh_len = cpu_to_be32(*copy_len);
2281		if (*last_was_partial_copy)
2282			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2283		*last_was_partial_copy = 0;
2284		*bytes_consumed = 0;
2285		return 0;
2286	}
2287
2288	/* partial write of region, needs extra log op header reservation */
2289	*copy_len = space_available;
2290	ophdr->oh_len = cpu_to_be32(*copy_len);
2291	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2292	if (*last_was_partial_copy)
2293		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2294	*bytes_consumed += *copy_len;
2295	(*last_was_partial_copy)++;
2296
2297	/* account for new log op header */
2298	ticket->t_curr_res -= sizeof(struct xlog_op_header);
2299	ticket->t_res_num_ophdrs++;
2300
2301	return sizeof(struct xlog_op_header);
2302}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2303
2304static int
2305xlog_write_copy_finish(
2306	struct xlog		*log,
2307	struct xlog_in_core	*iclog,
2308	uint			flags,
2309	int			*record_cnt,
2310	int			*data_cnt,
2311	int			*partial_copy,
2312	int			*partial_copy_len,
2313	int			log_offset,
2314	struct xlog_in_core	**commit_iclog)
2315{
2316	if (*partial_copy) {
2317		/*
2318		 * This iclog has already been marked WANT_SYNC by
2319		 * xlog_state_get_iclog_space.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2320		 */
2321		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2322		*record_cnt = 0;
2323		*data_cnt = 0;
2324		return xlog_state_release_iclog(log, iclog);
2325	}
2326
2327	*partial_copy = 0;
2328	*partial_copy_len = 0;
2329
2330	if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2331		/* no more space in this iclog - push it. */
2332		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2333		*record_cnt = 0;
2334		*data_cnt = 0;
 
2335
2336		spin_lock(&log->l_icloglock);
2337		xlog_state_want_sync(log, iclog);
2338		spin_unlock(&log->l_icloglock);
 
 
 
 
 
 
2339
2340		if (!commit_iclog)
2341			return xlog_state_release_iclog(log, iclog);
2342		ASSERT(flags & XLOG_COMMIT_TRANS);
2343		*commit_iclog = iclog;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2344	}
2345
 
 
 
 
 
2346	return 0;
2347}
2348
2349/*
2350 * Write some region out to in-core log
2351 *
2352 * This will be called when writing externally provided regions or when
2353 * writing out a commit record for a given transaction.
2354 *
2355 * General algorithm:
2356 *	1. Find total length of this write.  This may include adding to the
2357 *		lengths passed in.
2358 *	2. Check whether we violate the tickets reservation.
2359 *	3. While writing to this iclog
2360 *	    A. Reserve as much space in this iclog as can get
2361 *	    B. If this is first write, save away start lsn
2362 *	    C. While writing this region:
2363 *		1. If first write of transaction, write start record
2364 *		2. Write log operation header (header per region)
2365 *		3. Find out if we can fit entire region into this iclog
2366 *		4. Potentially, verify destination memcpy ptr
2367 *		5. Memcpy (partial) region
2368 *		6. If partial copy, release iclog; otherwise, continue
2369 *			copying more regions into current iclog
2370 *	4. Mark want sync bit (in simulation mode)
2371 *	5. Release iclog for potential flush to on-disk log.
2372 *
2373 * ERRORS:
2374 * 1.	Panic if reservation is overrun.  This should never happen since
2375 *	reservation amounts are generated internal to the filesystem.
2376 * NOTES:
2377 * 1. Tickets are single threaded data structures.
2378 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2379 *	syncing routine.  When a single log_write region needs to span
2380 *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2381 *	on all log operation writes which don't contain the end of the
2382 *	region.  The XLOG_END_TRANS bit is used for the in-core log
2383 *	operation which contains the end of the continued log_write region.
2384 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2385 *	we don't really know exactly how much space will be used.  As a result,
2386 *	we don't update ic_offset until the end when we know exactly how many
2387 *	bytes have been written out.
2388 */
2389int
2390xlog_write(
2391	struct xlog		*log,
2392	struct xfs_log_vec	*log_vector,
 
2393	struct xlog_ticket	*ticket,
2394	xfs_lsn_t		*start_lsn,
2395	struct xlog_in_core	**commit_iclog,
2396	uint			flags)
2397{
2398	struct xlog_in_core	*iclog = NULL;
2399	struct xfs_log_iovec	*vecp;
2400	struct xfs_log_vec	*lv;
2401	int			len;
2402	int			index;
2403	int			partial_copy = 0;
2404	int			partial_copy_len = 0;
2405	int			contwr = 0;
2406	int			record_cnt = 0;
2407	int			data_cnt = 0;
2408	int			error;
2409
2410	*start_lsn = 0;
2411
2412	len = xlog_write_calc_vec_length(ticket, log_vector);
2413
2414	/*
2415	 * Region headers and bytes are already accounted for.
2416	 * We only need to take into account start records and
2417	 * split regions in this function.
2418	 */
2419	if (ticket->t_flags & XLOG_TIC_INITED)
2420		ticket->t_curr_res -= sizeof(xlog_op_header_t);
2421
2422	/*
2423	 * Commit record headers need to be accounted for. These
2424	 * come in as separate writes so are easy to detect.
2425	 */
2426	if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2427		ticket->t_curr_res -= sizeof(xlog_op_header_t);
2428
2429	if (ticket->t_curr_res < 0) {
2430		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2431		     "ctx ticket reservation ran out. Need to up reservation");
2432		xlog_print_tic_res(log->l_mp, ticket);
2433		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2434	}
2435
2436	index = 0;
2437	lv = log_vector;
2438	vecp = lv->lv_iovecp;
2439	while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2440		void		*ptr;
2441		int		log_offset;
2442
2443		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2444						   &contwr, &log_offset);
2445		if (error)
2446			return error;
2447
2448		ASSERT(log_offset <= iclog->ic_size - 1);
2449		ptr = iclog->ic_datap + log_offset;
2450
2451		/* start_lsn is the first lsn written to. That's all we need. */
2452		if (!*start_lsn)
2453			*start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
 
 
 
 
2454
 
2455		/*
2456		 * This loop writes out as many regions as can fit in the amount
2457		 * of space which was allocated by xlog_state_get_iclog_space().
2458		 */
2459		while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2460			struct xfs_log_iovec	*reg;
2461			struct xlog_op_header	*ophdr;
2462			int			start_rec_copy;
2463			int			copy_len;
2464			int			copy_off;
2465			bool			ordered = false;
2466
2467			/* ordered log vectors have no regions to write */
2468			if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2469				ASSERT(lv->lv_niovecs == 0);
2470				ordered = true;
2471				goto next_lv;
2472			}
2473
2474			reg = &vecp[index];
2475			ASSERT(reg->i_len % sizeof(int32_t) == 0);
2476			ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2477
2478			start_rec_copy = xlog_write_start_rec(ptr, ticket);
2479			if (start_rec_copy) {
2480				record_cnt++;
2481				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2482						   start_rec_copy);
2483			}
2484
2485			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2486			if (!ophdr)
2487				return -EIO;
2488
2489			xlog_write_adv_cnt(&ptr, &len, &log_offset,
2490					   sizeof(struct xlog_op_header));
2491
2492			len += xlog_write_setup_copy(ticket, ophdr,
2493						     iclog->ic_size-log_offset,
2494						     reg->i_len,
2495						     &copy_off, &copy_len,
2496						     &partial_copy,
2497						     &partial_copy_len);
2498			xlog_verify_dest_ptr(log, ptr);
2499
2500			/*
2501			 * Copy region.
2502			 *
2503			 * Unmount records just log an opheader, so can have
2504			 * empty payloads with no data region to copy. Hence we
2505			 * only copy the payload if the vector says it has data
2506			 * to copy.
2507			 */
2508			ASSERT(copy_len >= 0);
2509			if (copy_len > 0) {
2510				memcpy(ptr, reg->i_addr + copy_off, copy_len);
2511				xlog_write_adv_cnt(&ptr, &len, &log_offset,
2512						   copy_len);
2513			}
2514			copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2515			record_cnt++;
2516			data_cnt += contwr ? copy_len : 0;
2517
2518			error = xlog_write_copy_finish(log, iclog, flags,
2519						       &record_cnt, &data_cnt,
2520						       &partial_copy,
2521						       &partial_copy_len,
2522						       log_offset,
2523						       commit_iclog);
2524			if (error)
2525				return error;
2526
2527			/*
2528			 * if we had a partial copy, we need to get more iclog
2529			 * space but we don't want to increment the region
2530			 * index because there is still more is this region to
2531			 * write.
2532			 *
2533			 * If we completed writing this region, and we flushed
2534			 * the iclog (indicated by resetting of the record
2535			 * count), then we also need to get more log space. If
2536			 * this was the last record, though, we are done and
2537			 * can just return.
2538			 */
2539			if (partial_copy)
2540				break;
2541
2542			if (++index == lv->lv_niovecs) {
2543next_lv:
2544				lv = lv->lv_next;
2545				index = 0;
2546				if (lv)
2547					vecp = lv->lv_iovecp;
2548			}
2549			if (record_cnt == 0 && !ordered) {
2550				if (!lv)
2551					return 0;
2552				break;
2553			}
 
 
 
2554		}
2555	}
2556
2557	ASSERT(len == 0);
2558
2559	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2560	if (!commit_iclog)
2561		return xlog_state_release_iclog(log, iclog);
 
 
 
 
 
 
 
2562
2563	ASSERT(flags & XLOG_COMMIT_TRANS);
2564	*commit_iclog = iclog;
2565	return 0;
2566}
2567
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2568
2569/*****************************************************************************
2570 *
2571 *		State Machine functions
2572 *
2573 *****************************************************************************
2574 */
 
 
2575
2576/* Clean iclogs starting from the head.  This ordering must be
2577 * maintained, so an iclog doesn't become ACTIVE beyond one that
2578 * is SYNCING.  This is also required to maintain the notion that we use
2579 * a ordered wait queue to hold off would be writers to the log when every
2580 * iclog is trying to sync to disk.
2581 *
2582 * State Change: DIRTY -> ACTIVE
2583 */
2584STATIC void
2585xlog_state_clean_log(
2586	struct xlog *log)
 
2587{
2588	xlog_in_core_t	*iclog;
2589	int changed = 0;
2590
2591	iclog = log->l_iclog;
2592	do {
2593		if (iclog->ic_state == XLOG_STATE_DIRTY) {
2594			iclog->ic_state	= XLOG_STATE_ACTIVE;
2595			iclog->ic_offset       = 0;
2596			ASSERT(iclog->ic_callback == NULL);
2597			/*
2598			 * If the number of ops in this iclog indicate it just
2599			 * contains the dummy transaction, we can
2600			 * change state into IDLE (the second time around).
2601			 * Otherwise we should change the state into
2602			 * NEED a dummy.
2603			 * We don't need to cover the dummy.
2604			 */
2605			if (!changed &&
2606			   (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2607			   		XLOG_COVER_OPS)) {
2608				changed = 1;
2609			} else {
2610				/*
2611				 * We have two dirty iclogs so start over
2612				 * This could also be num of ops indicates
2613				 * this is not the dummy going out.
2614				 */
2615				changed = 2;
2616			}
2617			iclog->ic_header.h_num_logops = 0;
2618			memset(iclog->ic_header.h_cycle_data, 0,
2619			      sizeof(iclog->ic_header.h_cycle_data));
2620			iclog->ic_header.h_lsn = 0;
2621		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2622			/* do nothing */;
2623		else
2624			break;	/* stop cleaning */
2625		iclog = iclog->ic_next;
2626	} while (iclog != log->l_iclog);
2627
2628	/* log is locked when we are called */
 
 
 
 
2629	/*
2630	 * Change state for the dummy log recording.
2631	 * We usually go to NEED. But we go to NEED2 if the changed indicates
2632	 * we are done writing the dummy record.
2633	 * If we are done with the second dummy recored (DONE2), then
2634	 * we go to IDLE.
2635	 */
2636	if (changed) {
2637		switch (log->l_covered_state) {
2638		case XLOG_STATE_COVER_IDLE:
2639		case XLOG_STATE_COVER_NEED:
2640		case XLOG_STATE_COVER_NEED2:
2641			log->l_covered_state = XLOG_STATE_COVER_NEED;
2642			break;
 
 
 
 
 
 
 
 
 
 
 
2643
2644		case XLOG_STATE_COVER_DONE:
2645			if (changed == 1)
2646				log->l_covered_state = XLOG_STATE_COVER_NEED2;
2647			else
2648				log->l_covered_state = XLOG_STATE_COVER_NEED;
2649			break;
2650
2651		case XLOG_STATE_COVER_DONE2:
2652			if (changed == 1)
2653				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2654			else
2655				log->l_covered_state = XLOG_STATE_COVER_NEED;
2656			break;
2657
2658		default:
2659			ASSERT(0);
2660		}
 
 
 
 
 
 
 
2661	}
2662}	/* xlog_state_clean_log */
2663
2664STATIC xfs_lsn_t
2665xlog_get_lowest_lsn(
2666	struct xlog	*log)
2667{
2668	xlog_in_core_t  *lsn_log;
2669	xfs_lsn_t	lowest_lsn, lsn;
2670
2671	lsn_log = log->l_iclog;
2672	lowest_lsn = 0;
2673	do {
2674	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2675		lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2676		if ((lsn && !lowest_lsn) ||
2677		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
 
 
2678			lowest_lsn = lsn;
2679		}
2680	    }
2681	    lsn_log = lsn_log->ic_next;
2682	} while (lsn_log != log->l_iclog);
2683	return lowest_lsn;
2684}
2685
2686
2687STATIC void
2688xlog_state_do_callback(
 
 
 
 
2689	struct xlog		*log,
2690	int			aborted,
2691	struct xlog_in_core	*ciclog)
2692{
2693	xlog_in_core_t	   *iclog;
2694	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2695						 * processed all iclogs once */
2696	xfs_log_callback_t *cb, *cb_next;
2697	int		   flushcnt = 0;
2698	xfs_lsn_t	   lowest_lsn;
2699	int		   ioerrors;	/* counter: iclogs with errors */
2700	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2701	int		   funcdidcallbacks; /* flag: function did callbacks */
2702	int		   repeats;	/* for issuing console warnings if
2703					 * looping too many times */
2704	int		   wake = 0;
2705
2706	spin_lock(&log->l_icloglock);
2707	first_iclog = iclog = log->l_iclog;
2708	ioerrors = 0;
2709	funcdidcallbacks = 0;
2710	repeats = 0;
2711
2712	do {
2713		/*
2714		 * Scan all iclogs starting with the one pointed to by the
2715		 * log.  Reset this starting point each time the log is
2716		 * unlocked (during callbacks).
2717		 *
2718		 * Keep looping through iclogs until one full pass is made
2719		 * without running any callbacks.
2720		 */
2721		first_iclog = log->l_iclog;
2722		iclog = log->l_iclog;
2723		loopdidcallbacks = 0;
2724		repeats++;
2725
2726		do {
2727
2728			/* skip all iclogs in the ACTIVE & DIRTY states */
2729			if (iclog->ic_state &
2730			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2731				iclog = iclog->ic_next;
2732				continue;
2733			}
2734
2735			/*
2736			 * Between marking a filesystem SHUTDOWN and stopping
2737			 * the log, we do flush all iclogs to disk (if there
2738			 * wasn't a log I/O error). So, we do want things to
2739			 * go smoothly in case of just a SHUTDOWN  w/o a
2740			 * LOG_IO_ERROR.
2741			 */
2742			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2743				/*
2744				 * Can only perform callbacks in order.  Since
2745				 * this iclog is not in the DONE_SYNC/
2746				 * DO_CALLBACK state, we skip the rest and
2747				 * just try to clean up.  If we set our iclog
2748				 * to DO_CALLBACK, we will not process it when
2749				 * we retry since a previous iclog is in the
2750				 * CALLBACK and the state cannot change since
2751				 * we are holding the l_icloglock.
2752				 */
2753				if (!(iclog->ic_state &
2754					(XLOG_STATE_DONE_SYNC |
2755						 XLOG_STATE_DO_CALLBACK))) {
2756					if (ciclog && (ciclog->ic_state ==
2757							XLOG_STATE_DONE_SYNC)) {
2758						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2759					}
2760					break;
2761				}
2762				/*
2763				 * We now have an iclog that is in either the
2764				 * DO_CALLBACK or DONE_SYNC states. The other
2765				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2766				 * caught by the above if and are going to
2767				 * clean (i.e. we aren't doing their callbacks)
2768				 * see the above if.
2769				 */
2770
2771				/*
2772				 * We will do one more check here to see if we
2773				 * have chased our tail around.
2774				 */
2775
2776				lowest_lsn = xlog_get_lowest_lsn(log);
2777				if (lowest_lsn &&
2778				    XFS_LSN_CMP(lowest_lsn,
2779						be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2780					iclog = iclog->ic_next;
2781					continue; /* Leave this iclog for
2782						   * another thread */
2783				}
2784
2785				iclog->ic_state = XLOG_STATE_CALLBACK;
2786
2787
2788				/*
2789				 * Completion of a iclog IO does not imply that
2790				 * a transaction has completed, as transactions
2791				 * can be large enough to span many iclogs. We
2792				 * cannot change the tail of the log half way
2793				 * through a transaction as this may be the only
2794				 * transaction in the log and moving th etail to
2795				 * point to the middle of it will prevent
2796				 * recovery from finding the start of the
2797				 * transaction. Hence we should only update the
2798				 * last_sync_lsn if this iclog contains
2799				 * transaction completion callbacks on it.
2800				 *
2801				 * We have to do this before we drop the
2802				 * icloglock to ensure we are the only one that
2803				 * can update it.
2804				 */
2805				ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2806					be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2807				if (iclog->ic_callback)
2808					atomic64_set(&log->l_last_sync_lsn,
2809						be64_to_cpu(iclog->ic_header.h_lsn));
2810
2811			} else
2812				ioerrors++;
2813
2814			spin_unlock(&log->l_icloglock);
 
 
 
 
 
 
 
2815
2816			/*
2817			 * Keep processing entries in the callback list until
2818			 * we come around and it is empty.  We need to
2819			 * atomically see that the list is empty and change the
2820			 * state to DIRTY so that we don't miss any more
2821			 * callbacks being added.
2822			 */
2823			spin_lock(&iclog->ic_callback_lock);
2824			cb = iclog->ic_callback;
2825			while (cb) {
2826				iclog->ic_callback_tail = &(iclog->ic_callback);
2827				iclog->ic_callback = NULL;
2828				spin_unlock(&iclog->ic_callback_lock);
2829
2830				/* perform callbacks in the order given */
2831				for (; cb; cb = cb_next) {
2832					cb_next = cb->cb_next;
2833					cb->cb_func(cb->cb_arg, aborted);
2834				}
2835				spin_lock(&iclog->ic_callback_lock);
2836				cb = iclog->ic_callback;
2837			}
2838
2839			loopdidcallbacks++;
2840			funcdidcallbacks++;
 
 
2841
2842			spin_lock(&log->l_icloglock);
2843			ASSERT(iclog->ic_callback == NULL);
2844			spin_unlock(&iclog->ic_callback_lock);
2845			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2846				iclog->ic_state = XLOG_STATE_DIRTY;
2847
2848			/*
2849			 * Transition from DIRTY to ACTIVE if applicable.
2850			 * NOP if STATE_IOERROR.
2851			 */
2852			xlog_state_clean_log(log);
2853
2854			/* wake up threads waiting in xfs_log_force() */
2855			wake_up_all(&iclog->ic_force_wait);
 
 
 
 
 
 
 
 
2856
2857			iclog = iclog->ic_next;
2858		} while (first_iclog != iclog);
 
 
2859
2860		if (repeats > 5000) {
2861			flushcnt += repeats;
2862			repeats = 0;
2863			xfs_warn(log->l_mp,
2864				"%s: possible infinite loop (%d iterations)",
2865				__func__, flushcnt);
2866		}
2867	} while (!ioerrors && loopdidcallbacks);
2868
2869#ifdef DEBUG
2870	/*
2871	 * Make one last gasp attempt to see if iclogs are being left in limbo.
2872	 * If the above loop finds an iclog earlier than the current iclog and
2873	 * in one of the syncing states, the current iclog is put into
2874	 * DO_CALLBACK and the callbacks are deferred to the completion of the
2875	 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2876	 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2877	 * states.
2878	 *
2879	 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2880	 * for ic_state == SYNCING.
2881	 */
2882	if (funcdidcallbacks) {
2883		first_iclog = iclog = log->l_iclog;
2884		do {
2885			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2886			/*
2887			 * Terminate the loop if iclogs are found in states
2888			 * which will cause other threads to clean up iclogs.
2889			 *
2890			 * SYNCING - i/o completion will go through logs
2891			 * DONE_SYNC - interrupt thread should be waiting for
2892			 *              l_icloglock
2893			 * IOERROR - give up hope all ye who enter here
2894			 */
2895			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2896			    iclog->ic_state & XLOG_STATE_SYNCING ||
2897			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2898			    iclog->ic_state == XLOG_STATE_IOERROR )
2899				break;
2900			iclog = iclog->ic_next;
2901		} while (first_iclog != iclog);
2902	}
2903#endif
2904
2905	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2906		wake = 1;
2907	spin_unlock(&log->l_icloglock);
2908
2909	if (wake)
2910		wake_up_all(&log->l_flush_wait);
 
 
2911}
2912
2913
2914/*
2915 * Finish transitioning this iclog to the dirty state.
2916 *
2917 * Make sure that we completely execute this routine only when this is
2918 * the last call to the iclog.  There is a good chance that iclog flushes,
2919 * when we reach the end of the physical log, get turned into 2 separate
2920 * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2921 * routine.  By using the reference count bwritecnt, we guarantee that only
2922 * the second completion goes through.
2923 *
2924 * Callbacks could take time, so they are done outside the scope of the
2925 * global state machine log lock.
2926 */
2927STATIC void
2928xlog_state_done_syncing(
2929	xlog_in_core_t	*iclog,
2930	int		aborted)
2931{
2932	struct xlog	   *log = iclog->ic_log;
2933
2934	spin_lock(&log->l_icloglock);
2935
2936	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2937	       iclog->ic_state == XLOG_STATE_IOERROR);
2938	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2939	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2940
2941
2942	/*
2943	 * If we got an error, either on the first buffer, or in the case of
2944	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2945	 * and none should ever be attempted to be written to disk
2946	 * again.
2947	 */
2948	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2949		if (--iclog->ic_bwritecnt == 1) {
2950			spin_unlock(&log->l_icloglock);
2951			return;
2952		}
2953		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2954	}
2955
2956	/*
2957	 * Someone could be sleeping prior to writing out the next
2958	 * iclog buffer, we wake them all, one will get to do the
2959	 * I/O, the others get to wait for the result.
2960	 */
2961	wake_up_all(&iclog->ic_write_wait);
2962	spin_unlock(&log->l_icloglock);
2963	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2964}	/* xlog_state_done_syncing */
2965
2966
2967/*
2968 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2969 * sleep.  We wait on the flush queue on the head iclog as that should be
2970 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2971 * we will wait here and all new writes will sleep until a sync completes.
2972 *
2973 * The in-core logs are used in a circular fashion. They are not used
2974 * out-of-order even when an iclog past the head is free.
2975 *
2976 * return:
2977 *	* log_offset where xlog_write() can start writing into the in-core
2978 *		log's data space.
2979 *	* in-core log pointer to which xlog_write() should write.
2980 *	* boolean indicating this is a continued write to an in-core log.
2981 *		If this is the last write, then the in-core log's offset field
2982 *		needs to be incremented, depending on the amount of data which
2983 *		is copied.
2984 */
2985STATIC int
2986xlog_state_get_iclog_space(
2987	struct xlog		*log,
2988	int			len,
2989	struct xlog_in_core	**iclogp,
2990	struct xlog_ticket	*ticket,
2991	int			*continued_write,
2992	int			*logoffsetp)
2993{
2994	int		  log_offset;
2995	xlog_rec_header_t *head;
2996	xlog_in_core_t	  *iclog;
2997	int		  error;
2998
2999restart:
3000	spin_lock(&log->l_icloglock);
3001	if (XLOG_FORCED_SHUTDOWN(log)) {
3002		spin_unlock(&log->l_icloglock);
3003		return -EIO;
3004	}
3005
3006	iclog = log->l_iclog;
3007	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3008		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3009
3010		/* Wait for log writes to have flushed */
3011		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3012		goto restart;
3013	}
3014
3015	head = &iclog->ic_header;
3016
3017	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
3018	log_offset = iclog->ic_offset;
3019
 
 
3020	/* On the 1st write to an iclog, figure out lsn.  This works
3021	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3022	 * committing to.  If the offset is set, that's how many blocks
3023	 * must be written.
3024	 */
3025	if (log_offset == 0) {
3026		ticket->t_curr_res -= log->l_iclog_hsize;
3027		xlog_tic_add_region(ticket,
3028				    log->l_iclog_hsize,
3029				    XLOG_REG_TYPE_LRHEADER);
3030		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3031		head->h_lsn = cpu_to_be64(
3032			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3033		ASSERT(log->l_curr_block >= 0);
3034	}
3035
3036	/* If there is enough room to write everything, then do it.  Otherwise,
3037	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3038	 * bit is on, so this will get flushed out.  Don't update ic_offset
3039	 * until you know exactly how many bytes get copied.  Therefore, wait
3040	 * until later to update ic_offset.
3041	 *
3042	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3043	 * can fit into remaining data section.
3044	 */
3045	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
 
 
3046		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3047
3048		/*
3049		 * If I'm the only one writing to this iclog, sync it to disk.
3050		 * We need to do an atomic compare and decrement here to avoid
3051		 * racing with concurrent atomic_dec_and_lock() calls in
3052		 * xlog_state_release_iclog() when there is more than one
3053		 * reference to the iclog.
3054		 */
3055		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3056			/* we are the only one */
3057			spin_unlock(&log->l_icloglock);
3058			error = xlog_state_release_iclog(log, iclog);
3059			if (error)
3060				return error;
3061		} else {
3062			spin_unlock(&log->l_icloglock);
3063		}
3064		goto restart;
3065	}
3066
3067	/* Do we have enough room to write the full amount in the remainder
3068	 * of this iclog?  Or must we continue a write on the next iclog and
3069	 * mark this iclog as completely taken?  In the case where we switch
3070	 * iclogs (to mark it taken), this particular iclog will release/sync
3071	 * to disk in xlog_write().
3072	 */
3073	if (len <= iclog->ic_size - iclog->ic_offset) {
3074		*continued_write = 0;
3075		iclog->ic_offset += len;
3076	} else {
3077		*continued_write = 1;
3078		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3079	}
3080	*iclogp = iclog;
3081
3082	ASSERT(iclog->ic_offset <= iclog->ic_size);
3083	spin_unlock(&log->l_icloglock);
3084
3085	*logoffsetp = log_offset;
3086	return 0;
3087}	/* xlog_state_get_iclog_space */
3088
3089/* The first cnt-1 times through here we don't need to
3090 * move the grant write head because the permanent
3091 * reservation has reserved cnt times the unit amount.
3092 * Release part of current permanent unit reservation and
3093 * reset current reservation to be one units worth.  Also
3094 * move grant reservation head forward.
3095 */
3096STATIC void
3097xlog_regrant_reserve_log_space(
3098	struct xlog		*log,
3099	struct xlog_ticket	*ticket)
3100{
3101	trace_xfs_log_regrant_reserve_enter(log, ticket);
3102
3103	if (ticket->t_cnt > 0)
3104		ticket->t_cnt--;
3105
3106	xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3107					ticket->t_curr_res);
3108	xlog_grant_sub_space(log, &log->l_write_head.grant,
3109					ticket->t_curr_res);
3110	ticket->t_curr_res = ticket->t_unit_res;
3111	xlog_tic_reset_res(ticket);
3112
3113	trace_xfs_log_regrant_reserve_sub(log, ticket);
3114
3115	/* just return if we still have some of the pre-reserved space */
3116	if (ticket->t_cnt > 0)
3117		return;
3118
3119	xlog_grant_add_space(log, &log->l_reserve_head.grant,
3120					ticket->t_unit_res);
3121
3122	trace_xfs_log_regrant_reserve_exit(log, ticket);
3123
3124	ticket->t_curr_res = ticket->t_unit_res;
3125	xlog_tic_reset_res(ticket);
3126}	/* xlog_regrant_reserve_log_space */
3127
 
 
3128
3129/*
3130 * Give back the space left from a reservation.
3131 *
3132 * All the information we need to make a correct determination of space left
3133 * is present.  For non-permanent reservations, things are quite easy.  The
3134 * count should have been decremented to zero.  We only need to deal with the
3135 * space remaining in the current reservation part of the ticket.  If the
3136 * ticket contains a permanent reservation, there may be left over space which
3137 * needs to be released.  A count of N means that N-1 refills of the current
3138 * reservation can be done before we need to ask for more space.  The first
3139 * one goes to fill up the first current reservation.  Once we run out of
3140 * space, the count will stay at zero and the only space remaining will be
3141 * in the current reservation field.
3142 */
3143STATIC void
3144xlog_ungrant_log_space(
3145	struct xlog		*log,
3146	struct xlog_ticket	*ticket)
3147{
3148	int	bytes;
 
 
3149
3150	if (ticket->t_cnt > 0)
3151		ticket->t_cnt--;
3152
3153	trace_xfs_log_ungrant_enter(log, ticket);
3154	trace_xfs_log_ungrant_sub(log, ticket);
3155
3156	/*
3157	 * If this is a permanent reservation ticket, we may be able to free
3158	 * up more space based on the remaining count.
3159	 */
3160	bytes = ticket->t_curr_res;
3161	if (ticket->t_cnt > 0) {
3162		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3163		bytes += ticket->t_unit_res*ticket->t_cnt;
3164	}
3165
3166	xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3167	xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3168
3169	trace_xfs_log_ungrant_exit(log, ticket);
3170
3171	xfs_log_space_wake(log->l_mp);
 
3172}
3173
3174/*
3175 * Flush iclog to disk if this is the last reference to the given iclog and
3176 * the WANT_SYNC bit is set.
3177 *
3178 * When this function is entered, the iclog is not necessarily in the
3179 * WANT_SYNC state.  It may be sitting around waiting to get filled.
3180 *
3181 *
3182 */
3183STATIC int
3184xlog_state_release_iclog(
3185	struct xlog		*log,
3186	struct xlog_in_core	*iclog)
3187{
3188	int		sync = 0;	/* do we sync? */
3189
3190	if (iclog->ic_state & XLOG_STATE_IOERROR)
3191		return -EIO;
3192
3193	ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3194	if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3195		return 0;
3196
3197	if (iclog->ic_state & XLOG_STATE_IOERROR) {
3198		spin_unlock(&log->l_icloglock);
3199		return -EIO;
3200	}
3201	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3202	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
3203
3204	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3205		/* update tail before writing to iclog */
3206		xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3207		sync++;
3208		iclog->ic_state = XLOG_STATE_SYNCING;
3209		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3210		xlog_verify_tail_lsn(log, iclog, tail_lsn);
3211		/* cycle incremented when incrementing curr_block */
3212	}
3213	spin_unlock(&log->l_icloglock);
3214
3215	/*
3216	 * We let the log lock go, so it's possible that we hit a log I/O
3217	 * error or some other SHUTDOWN condition that marks the iclog
3218	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3219	 * this iclog has consistent data, so we ignore IOERROR
3220	 * flags after this point.
3221	 */
3222	if (sync)
3223		return xlog_sync(log, iclog);
3224	return 0;
3225}	/* xlog_state_release_iclog */
3226
3227
3228/*
3229 * This routine will mark the current iclog in the ring as WANT_SYNC
3230 * and move the current iclog pointer to the next iclog in the ring.
3231 * When this routine is called from xlog_state_get_iclog_space(), the
3232 * exact size of the iclog has not yet been determined.  All we know is
3233 * that every data block.  We have run out of space in this log record.
3234 */
3235STATIC void
3236xlog_state_switch_iclogs(
3237	struct xlog		*log,
3238	struct xlog_in_core	*iclog,
3239	int			eventual_size)
3240{
3241	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 
 
 
3242	if (!eventual_size)
3243		eventual_size = iclog->ic_offset;
3244	iclog->ic_state = XLOG_STATE_WANT_SYNC;
3245	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3246	log->l_prev_block = log->l_curr_block;
3247	log->l_prev_cycle = log->l_curr_cycle;
3248
3249	/* roll log?: ic_offset changed later */
3250	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3251
3252	/* Round up to next log-sunit */
3253	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3254	    log->l_mp->m_sb.sb_logsunit > 1) {
3255		uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3256		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3257	}
3258
3259	if (log->l_curr_block >= log->l_logBBsize) {
3260		/*
3261		 * Rewind the current block before the cycle is bumped to make
3262		 * sure that the combined LSN never transiently moves forward
3263		 * when the log wraps to the next cycle. This is to support the
3264		 * unlocked sample of these fields from xlog_valid_lsn(). Most
3265		 * other cases should acquire l_icloglock.
3266		 */
3267		log->l_curr_block -= log->l_logBBsize;
3268		ASSERT(log->l_curr_block >= 0);
3269		smp_wmb();
3270		log->l_curr_cycle++;
3271		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3272			log->l_curr_cycle++;
3273	}
3274	ASSERT(iclog == log->l_iclog);
3275	log->l_iclog = iclog->ic_next;
3276}	/* xlog_state_switch_iclogs */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3277
3278/*
3279 * Write out all data in the in-core log as of this exact moment in time.
3280 *
3281 * Data may be written to the in-core log during this call.  However,
3282 * we don't guarantee this data will be written out.  A change from past
3283 * implementation means this routine will *not* write out zero length LRs.
3284 *
3285 * Basically, we try and perform an intelligent scan of the in-core logs.
3286 * If we determine there is no flushable data, we just return.  There is no
3287 * flushable data if:
3288 *
3289 *	1. the current iclog is active and has no data; the previous iclog
3290 *		is in the active or dirty state.
3291 *	2. the current iclog is drity, and the previous iclog is in the
3292 *		active or dirty state.
3293 *
3294 * We may sleep if:
3295 *
3296 *	1. the current iclog is not in the active nor dirty state.
3297 *	2. the current iclog dirty, and the previous iclog is not in the
3298 *		active nor dirty state.
3299 *	3. the current iclog is active, and there is another thread writing
3300 *		to this particular iclog.
3301 *	4. a) the current iclog is active and has no other writers
3302 *	   b) when we return from flushing out this iclog, it is still
3303 *		not in the active nor dirty state.
3304 */
3305int
3306xfs_log_force(
3307	struct xfs_mount	*mp,
3308	uint			flags)
3309{
3310	struct xlog		*log = mp->m_log;
3311	struct xlog_in_core	*iclog;
3312	xfs_lsn_t		lsn;
3313
3314	XFS_STATS_INC(mp, xs_log_force);
3315	trace_xfs_log_force(mp, 0, _RET_IP_);
3316
3317	xlog_cil_force(log);
3318
3319	spin_lock(&log->l_icloglock);
3320	iclog = log->l_iclog;
3321	if (iclog->ic_state & XLOG_STATE_IOERROR)
3322		goto out_error;
3323
 
 
 
3324	if (iclog->ic_state == XLOG_STATE_DIRTY ||
3325	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
3326	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3327		/*
3328		 * If the head is dirty or (active and empty), then we need to
3329		 * look at the previous iclog.
3330		 *
3331		 * If the previous iclog is active or dirty we are done.  There
3332		 * is nothing to sync out. Otherwise, we attach ourselves to the
3333		 * previous iclog and go to sleep.
3334		 */
3335		iclog = iclog->ic_prev;
3336		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3337		    iclog->ic_state == XLOG_STATE_DIRTY)
3338			goto out_unlock;
3339	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3340		if (atomic_read(&iclog->ic_refcnt) == 0) {
3341			/*
3342			 * We are the only one with access to this iclog.
3343			 *
3344			 * Flush it out now.  There should be a roundoff of zero
3345			 * to show that someone has already taken care of the
3346			 * roundoff from the previous sync.
3347			 */
3348			atomic_inc(&iclog->ic_refcnt);
3349			lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3350			xlog_state_switch_iclogs(log, iclog, 0);
3351			spin_unlock(&log->l_icloglock);
3352
3353			if (xlog_state_release_iclog(log, iclog))
3354				return -EIO;
3355
3356			spin_lock(&log->l_icloglock);
3357			if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3358			    iclog->ic_state == XLOG_STATE_DIRTY)
3359				goto out_unlock;
3360		} else {
3361			/*
3362			 * Someone else is writing to this iclog.
3363			 *
3364			 * Use its call to flush out the data.  However, the
3365			 * other thread may not force out this LR, so we mark
3366			 * it WANT_SYNC.
3367			 */
3368			xlog_state_switch_iclogs(log, iclog, 0);
3369		}
3370	} else {
3371		/*
3372		 * If the head iclog is not active nor dirty, we just attach
3373		 * ourselves to the head and go to sleep if necessary.
3374		 */
3375		;
3376	}
3377
3378	if (!(flags & XFS_LOG_SYNC))
3379		goto out_unlock;
3380
3381	if (iclog->ic_state & XLOG_STATE_IOERROR)
3382		goto out_error;
3383	XFS_STATS_INC(mp, xs_log_force_sleep);
3384	xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3385	if (iclog->ic_state & XLOG_STATE_IOERROR)
3386		return -EIO;
3387	return 0;
3388
 
 
3389out_unlock:
3390	spin_unlock(&log->l_icloglock);
3391	return 0;
3392out_error:
3393	spin_unlock(&log->l_icloglock);
3394	return -EIO;
3395}
3396
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3397static int
3398__xfs_log_force_lsn(
3399	struct xfs_mount	*mp,
3400	xfs_lsn_t		lsn,
3401	uint			flags,
3402	int			*log_flushed,
3403	bool			already_slept)
3404{
3405	struct xlog		*log = mp->m_log;
3406	struct xlog_in_core	*iclog;
 
3407
3408	spin_lock(&log->l_icloglock);
3409	iclog = log->l_iclog;
3410	if (iclog->ic_state & XLOG_STATE_IOERROR)
3411		goto out_error;
3412
 
3413	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
 
3414		iclog = iclog->ic_next;
3415		if (iclog == log->l_iclog)
3416			goto out_unlock;
3417	}
3418
3419	if (iclog->ic_state == XLOG_STATE_DIRTY)
3420		goto out_unlock;
3421
3422	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3423		/*
3424		 * We sleep here if we haven't already slept (e.g. this is the
3425		 * first time we've looked at the correct iclog buf) and the
3426		 * buffer before us is going to be sync'ed.  The reason for this
3427		 * is that if we are doing sync transactions here, by waiting
3428		 * for the previous I/O to complete, we can allow a few more
3429		 * transactions into this iclog before we close it down.
3430		 *
3431		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3432		 * refcnt so we can release the log (which drops the ref count).
3433		 * The state switch keeps new transaction commits from using
3434		 * this buffer.  When the current commits finish writing into
3435		 * the buffer, the refcount will drop to zero and the buffer
3436		 * will go out then.
3437		 */
3438		if (!already_slept &&
3439		    (iclog->ic_prev->ic_state &
3440		     (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3441			ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3442
3443			XFS_STATS_INC(mp, xs_log_force_sleep);
3444
3445			xlog_wait(&iclog->ic_prev->ic_write_wait,
3446					&log->l_icloglock);
3447			return -EAGAIN;
3448		}
3449		atomic_inc(&iclog->ic_refcnt);
3450		xlog_state_switch_iclogs(log, iclog, 0);
3451		spin_unlock(&log->l_icloglock);
3452		if (xlog_state_release_iclog(log, iclog))
3453			return -EIO;
3454		if (log_flushed)
3455			*log_flushed = 1;
3456		spin_lock(&log->l_icloglock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3457	}
3458
3459	if (!(flags & XFS_LOG_SYNC) ||
3460	    (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3461		goto out_unlock;
3462
3463	if (iclog->ic_state & XLOG_STATE_IOERROR)
3464		goto out_error;
3465
3466	XFS_STATS_INC(mp, xs_log_force_sleep);
3467	xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3468	if (iclog->ic_state & XLOG_STATE_IOERROR)
3469		return -EIO;
3470	return 0;
3471
3472out_unlock:
3473	spin_unlock(&log->l_icloglock);
3474	return 0;
3475out_error:
3476	spin_unlock(&log->l_icloglock);
3477	return -EIO;
3478}
3479
3480/*
3481 * Force the in-core log to disk for a specific LSN.
3482 *
3483 * Find in-core log with lsn.
3484 *	If it is in the DIRTY state, just return.
3485 *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3486 *		state and go to sleep or return.
3487 *	If it is in any other state, go to sleep or return.
3488 *
3489 * Synchronous forces are implemented with a wait queue.  All callers trying
3490 * to force a given lsn to disk must wait on the queue attached to the
3491 * specific in-core log.  When given in-core log finally completes its write
3492 * to disk, that thread will wake up all threads waiting on the queue.
 
3493 */
3494int
3495xfs_log_force_lsn(
3496	struct xfs_mount	*mp,
3497	xfs_lsn_t		lsn,
3498	uint			flags,
3499	int			*log_flushed)
3500{
 
 
3501	int			ret;
3502	ASSERT(lsn != 0);
3503
3504	XFS_STATS_INC(mp, xs_log_force);
3505	trace_xfs_log_force(mp, lsn, _RET_IP_);
3506
3507	lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3508	if (lsn == NULLCOMMITLSN)
3509		return 0;
3510
3511	ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3512	if (ret == -EAGAIN)
3513		ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3514	return ret;
3515}
3516
3517/*
3518 * Called when we want to mark the current iclog as being ready to sync to
3519 * disk.
3520 */
3521STATIC void
3522xlog_state_want_sync(
3523	struct xlog		*log,
3524	struct xlog_in_core	*iclog)
3525{
3526	assert_spin_locked(&log->l_icloglock);
3527
3528	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3529		xlog_state_switch_iclogs(log, iclog, 0);
3530	} else {
3531		ASSERT(iclog->ic_state &
3532			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3533	}
 
3534}
3535
3536
3537/*****************************************************************************
3538 *
3539 *		TICKET functions
3540 *
3541 *****************************************************************************
3542 */
3543
3544/*
3545 * Free a used ticket when its refcount falls to zero.
3546 */
3547void
3548xfs_log_ticket_put(
3549	xlog_ticket_t	*ticket)
3550{
3551	ASSERT(atomic_read(&ticket->t_ref) > 0);
3552	if (atomic_dec_and_test(&ticket->t_ref))
3553		kmem_zone_free(xfs_log_ticket_zone, ticket);
3554}
3555
3556xlog_ticket_t *
3557xfs_log_ticket_get(
3558	xlog_ticket_t	*ticket)
3559{
3560	ASSERT(atomic_read(&ticket->t_ref) > 0);
3561	atomic_inc(&ticket->t_ref);
3562	return ticket;
3563}
3564
3565/*
3566 * Figure out the total log space unit (in bytes) that would be
3567 * required for a log ticket.
3568 */
3569int
3570xfs_log_calc_unit_res(
3571	struct xfs_mount	*mp,
3572	int			unit_bytes)
 
3573{
3574	struct xlog		*log = mp->m_log;
3575	int			iclog_space;
3576	uint			num_headers;
3577
3578	/*
3579	 * Permanent reservations have up to 'cnt'-1 active log operations
3580	 * in the log.  A unit in this case is the amount of space for one
3581	 * of these log operations.  Normal reservations have a cnt of 1
3582	 * and their unit amount is the total amount of space required.
3583	 *
3584	 * The following lines of code account for non-transaction data
3585	 * which occupy space in the on-disk log.
3586	 *
3587	 * Normal form of a transaction is:
3588	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3589	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3590	 *
3591	 * We need to account for all the leadup data and trailer data
3592	 * around the transaction data.
3593	 * And then we need to account for the worst case in terms of using
3594	 * more space.
3595	 * The worst case will happen if:
3596	 * - the placement of the transaction happens to be such that the
3597	 *   roundoff is at its maximum
3598	 * - the transaction data is synced before the commit record is synced
3599	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3600	 *   Therefore the commit record is in its own Log Record.
3601	 *   This can happen as the commit record is called with its
3602	 *   own region to xlog_write().
3603	 *   This then means that in the worst case, roundoff can happen for
3604	 *   the commit-rec as well.
3605	 *   The commit-rec is smaller than padding in this scenario and so it is
3606	 *   not added separately.
3607	 */
3608
3609	/* for trans header */
3610	unit_bytes += sizeof(xlog_op_header_t);
3611	unit_bytes += sizeof(xfs_trans_header_t);
3612
3613	/* for start-rec */
3614	unit_bytes += sizeof(xlog_op_header_t);
3615
3616	/*
3617	 * for LR headers - the space for data in an iclog is the size minus
3618	 * the space used for the headers. If we use the iclog size, then we
3619	 * undercalculate the number of headers required.
3620	 *
3621	 * Furthermore - the addition of op headers for split-recs might
3622	 * increase the space required enough to require more log and op
3623	 * headers, so take that into account too.
3624	 *
3625	 * IMPORTANT: This reservation makes the assumption that if this
3626	 * transaction is the first in an iclog and hence has the LR headers
3627	 * accounted to it, then the remaining space in the iclog is
3628	 * exclusively for this transaction.  i.e. if the transaction is larger
3629	 * than the iclog, it will be the only thing in that iclog.
3630	 * Fundamentally, this means we must pass the entire log vector to
3631	 * xlog_write to guarantee this.
3632	 */
3633	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3634	num_headers = howmany(unit_bytes, iclog_space);
3635
3636	/* for split-recs - ophdrs added when data split over LRs */
3637	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3638
3639	/* add extra header reservations if we overrun */
3640	while (!num_headers ||
3641	       howmany(unit_bytes, iclog_space) > num_headers) {
3642		unit_bytes += sizeof(xlog_op_header_t);
3643		num_headers++;
3644	}
3645	unit_bytes += log->l_iclog_hsize * num_headers;
3646
3647	/* for commit-rec LR header - note: padding will subsume the ophdr */
3648	unit_bytes += log->l_iclog_hsize;
3649
3650	/* for roundoff padding for transaction data and one for commit record */
3651	if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3652		/* log su roundoff */
3653		unit_bytes += 2 * mp->m_sb.sb_logsunit;
3654	} else {
3655		/* BB roundoff */
3656		unit_bytes += 2 * BBSIZE;
3657        }
3658
 
 
3659	return unit_bytes;
3660}
3661
 
 
 
 
 
 
 
 
3662/*
3663 * Allocate and initialise a new log ticket.
3664 */
3665struct xlog_ticket *
3666xlog_ticket_alloc(
3667	struct xlog		*log,
3668	int			unit_bytes,
3669	int			cnt,
3670	char			client,
3671	bool			permanent,
3672	xfs_km_flags_t		alloc_flags)
3673{
3674	struct xlog_ticket	*tic;
3675	int			unit_res;
3676
3677	tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3678	if (!tic)
3679		return NULL;
3680
3681	unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3682
3683	atomic_set(&tic->t_ref, 1);
3684	tic->t_task		= current;
3685	INIT_LIST_HEAD(&tic->t_queue);
3686	tic->t_unit_res		= unit_res;
3687	tic->t_curr_res		= unit_res;
3688	tic->t_cnt		= cnt;
3689	tic->t_ocnt		= cnt;
3690	tic->t_tid		= prandom_u32();
3691	tic->t_clientid		= client;
3692	tic->t_flags		= XLOG_TIC_INITED;
3693	if (permanent)
3694		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3695
3696	xlog_tic_reset_res(tic);
3697
3698	return tic;
3699}
3700
3701
3702/******************************************************************************
3703 *
3704 *		Log debug routines
3705 *
3706 ******************************************************************************
3707 */
3708#if defined(DEBUG)
3709/*
3710 * Make sure that the destination ptr is within the valid data region of
3711 * one of the iclogs.  This uses backup pointers stored in a different
3712 * part of the log in case we trash the log structure.
3713 */
3714STATIC void
3715xlog_verify_dest_ptr(
3716	struct xlog	*log,
3717	void		*ptr)
3718{
3719	int i;
3720	int good_ptr = 0;
3721
3722	for (i = 0; i < log->l_iclog_bufs; i++) {
3723		if (ptr >= log->l_iclog_bak[i] &&
3724		    ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3725			good_ptr++;
3726	}
3727
3728	if (!good_ptr)
3729		xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3730}
3731
3732/*
3733 * Check to make sure the grant write head didn't just over lap the tail.  If
3734 * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3735 * the cycles differ by exactly one and check the byte count.
3736 *
3737 * This check is run unlocked, so can give false positives. Rather than assert
3738 * on failures, use a warn-once flag and a panic tag to allow the admin to
3739 * determine if they want to panic the machine when such an error occurs. For
3740 * debug kernels this will have the same effect as using an assert but, unlinke
3741 * an assert, it can be turned off at runtime.
3742 */
3743STATIC void
3744xlog_verify_grant_tail(
3745	struct xlog	*log)
3746{
3747	int		tail_cycle, tail_blocks;
3748	int		cycle, space;
3749
3750	xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3751	xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3752	if (tail_cycle != cycle) {
3753		if (cycle - 1 != tail_cycle &&
3754		    !(log->l_flags & XLOG_TAIL_WARN)) {
3755			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3756				"%s: cycle - 1 != tail_cycle", __func__);
3757			log->l_flags |= XLOG_TAIL_WARN;
3758		}
3759
3760		if (space > BBTOB(tail_blocks) &&
3761		    !(log->l_flags & XLOG_TAIL_WARN)) {
3762			xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3763				"%s: space > BBTOB(tail_blocks)", __func__);
3764			log->l_flags |= XLOG_TAIL_WARN;
3765		}
3766	}
3767}
3768
3769/* check if it will fit */
3770STATIC void
3771xlog_verify_tail_lsn(
3772	struct xlog		*log,
3773	struct xlog_in_core	*iclog,
3774	xfs_lsn_t		tail_lsn)
3775{
3776    int blocks;
 
3777
3778    if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3779	blocks =
3780	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3781	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3782		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3783    } else {
3784	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
 
 
 
 
3785
3786	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
 
 
 
 
 
3787		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
 
 
 
3788
3789	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3790	if (blocks < BTOBB(iclog->ic_offset) + 1)
3791		xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3792    }
3793}	/* xlog_verify_tail_lsn */
 
3794
3795/*
3796 * Perform a number of checks on the iclog before writing to disk.
3797 *
3798 * 1. Make sure the iclogs are still circular
3799 * 2. Make sure we have a good magic number
3800 * 3. Make sure we don't have magic numbers in the data
3801 * 4. Check fields of each log operation header for:
3802 *	A. Valid client identifier
3803 *	B. tid ptr value falls in valid ptr space (user space code)
3804 *	C. Length in log record header is correct according to the
3805 *		individual operation headers within record.
3806 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3807 *	log, check the preceding blocks of the physical log to make sure all
3808 *	the cycle numbers agree with the current cycle number.
3809 */
3810STATIC void
3811xlog_verify_iclog(
3812	struct xlog		*log,
3813	struct xlog_in_core	*iclog,
3814	int			count,
3815	bool                    syncing)
3816{
3817	xlog_op_header_t	*ophead;
3818	xlog_in_core_t		*icptr;
3819	xlog_in_core_2_t	*xhdr;
3820	void			*base_ptr, *ptr, *p;
3821	ptrdiff_t		field_offset;
3822	uint8_t			clientid;
3823	int			len, i, j, k, op_len;
3824	int			idx;
3825
3826	/* check validity of iclog pointers */
3827	spin_lock(&log->l_icloglock);
3828	icptr = log->l_iclog;
3829	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3830		ASSERT(icptr);
3831
3832	if (icptr != log->l_iclog)
3833		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3834	spin_unlock(&log->l_icloglock);
3835
3836	/* check log magic numbers */
3837	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3838		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3839
3840	base_ptr = ptr = &iclog->ic_header;
3841	p = &iclog->ic_header;
3842	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3843		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3844			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3845				__func__);
3846	}
3847
3848	/* check fields */
3849	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3850	base_ptr = ptr = iclog->ic_datap;
3851	ophead = ptr;
3852	xhdr = iclog->ic_data;
3853	for (i = 0; i < len; i++) {
3854		ophead = ptr;
3855
3856		/* clientid is only 1 byte */
3857		p = &ophead->oh_clientid;
3858		field_offset = p - base_ptr;
3859		if (!syncing || (field_offset & 0x1ff)) {
3860			clientid = ophead->oh_clientid;
3861		} else {
3862			idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3863			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3864				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3865				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3866				clientid = xlog_get_client_id(
3867					xhdr[j].hic_xheader.xh_cycle_data[k]);
3868			} else {
3869				clientid = xlog_get_client_id(
3870					iclog->ic_header.h_cycle_data[idx]);
3871			}
3872		}
3873		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3874			xfs_warn(log->l_mp,
3875				"%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3876				__func__, clientid, ophead,
3877				(unsigned long)field_offset);
 
3878
3879		/* check length */
3880		p = &ophead->oh_len;
3881		field_offset = p - base_ptr;
3882		if (!syncing || (field_offset & 0x1ff)) {
3883			op_len = be32_to_cpu(ophead->oh_len);
3884		} else {
3885			idx = BTOBBT((uintptr_t)&ophead->oh_len -
3886				    (uintptr_t)iclog->ic_datap);
3887			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3888				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3889				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3890				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3891			} else {
3892				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3893			}
3894		}
3895		ptr += sizeof(xlog_op_header_t) + op_len;
3896	}
3897}	/* xlog_verify_iclog */
3898#endif
3899
3900/*
3901 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3902 */
3903STATIC int
3904xlog_state_ioerror(
3905	struct xlog	*log)
3906{
3907	xlog_in_core_t	*iclog, *ic;
3908
3909	iclog = log->l_iclog;
3910	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3911		/*
3912		 * Mark all the incore logs IOERROR.
3913		 * From now on, no log flushes will result.
3914		 */
3915		ic = iclog;
3916		do {
3917			ic->ic_state = XLOG_STATE_IOERROR;
3918			ic = ic->ic_next;
3919		} while (ic != iclog);
3920		return 0;
3921	}
3922	/*
3923	 * Return non-zero, if state transition has already happened.
3924	 */
3925	return 1;
3926}
 
3927
3928/*
3929 * This is called from xfs_force_shutdown, when we're forcibly
3930 * shutting down the filesystem, typically because of an IO error.
 
 
 
3931 * Our main objectives here are to make sure that:
3932 *	a. if !logerror, flush the logs to disk. Anything modified
3933 *	   after this is ignored.
3934 *	b. the filesystem gets marked 'SHUTDOWN' for all interested
3935 *	   parties to find out, 'atomically'.
3936 *	c. those who're sleeping on log reservations, pinned objects and
3937 *	    other resources get woken up, and be told the bad news.
3938 *	d. nothing new gets queued up after (b) and (c) are done.
3939 *
3940 * Note: for the !logerror case we need to flush the regions held in memory out
3941 * to disk first. This needs to be done before the log is marked as shutdown,
3942 * otherwise the iclog writes will fail.
3943 */
3944int
3945xfs_log_force_umount(
3946	struct xfs_mount	*mp,
3947	int			logerror)
3948{
3949	struct xlog	*log;
3950	int		retval;
3951
3952	log = mp->m_log;
 
3953
3954	/*
3955	 * If this happens during log recovery, don't worry about
3956	 * locking; the log isn't open for business yet.
 
 
 
3957	 */
3958	if (!log ||
3959	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3960		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3961		if (mp->m_sb_bp)
3962			mp->m_sb_bp->b_flags |= XBF_DONE;
3963		return 0;
3964	}
3965
3966	/*
3967	 * Somebody could've already done the hard work for us.
3968	 * No need to get locks for this.
3969	 */
3970	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3971		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3972		return 1;
3973	}
3974
3975	/*
3976	 * Flush all the completed transactions to disk before marking the log
3977	 * being shut down. We need to do it in this order to ensure that
3978	 * completed operations are safely on disk before we shut down, and that
3979	 * we don't have to issue any buffer IO after the shutdown flags are set
3980	 * to guarantee this.
3981	 */
3982	if (!logerror)
3983		xfs_log_force(mp, XFS_LOG_SYNC);
 
 
 
 
 
 
 
 
3984
3985	/*
3986	 * mark the filesystem and the as in a shutdown state and wake
3987	 * everybody up to tell them the bad news.
 
 
 
 
 
 
 
3988	 */
3989	spin_lock(&log->l_icloglock);
3990	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3991	if (mp->m_sb_bp)
3992		mp->m_sb_bp->b_flags |= XBF_DONE;
 
 
 
3993
3994	/*
3995	 * Mark the log and the iclogs with IO error flags to prevent any
3996	 * further log IO from being issued or completed.
3997	 */
3998	log->l_flags |= XLOG_IO_ERROR;
3999	retval = xlog_state_ioerror(log);
4000	spin_unlock(&log->l_icloglock);
 
 
 
 
 
 
4001
4002	/*
4003	 * We don't want anybody waiting for log reservations after this. That
4004	 * means we have to wake up everybody queued up on reserveq as well as
4005	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
4006	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
4007	 * action is protected by the grant locks.
4008	 */
4009	xlog_grant_head_wake_all(&log->l_reserve_head);
4010	xlog_grant_head_wake_all(&log->l_write_head);
4011
4012	/*
4013	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4014	 * as if the log writes were completed. The abort handling in the log
4015	 * item committed callback functions will do this again under lock to
4016	 * avoid races.
4017	 */
 
 
4018	wake_up_all(&log->l_cilp->xc_commit_wait);
4019	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4020
4021#ifdef XFSERRORDEBUG
4022	{
4023		xlog_in_core_t	*iclog;
4024
4025		spin_lock(&log->l_icloglock);
4026		iclog = log->l_iclog;
4027		do {
4028			ASSERT(iclog->ic_callback == 0);
4029			iclog = iclog->ic_next;
4030		} while (iclog != log->l_iclog);
4031		spin_unlock(&log->l_icloglock);
4032	}
4033#endif
4034	/* return non-zero if log IOERROR transition had already happened */
4035	return retval;
4036}
4037
4038STATIC int
4039xlog_iclogs_empty(
4040	struct xlog	*log)
4041{
4042	xlog_in_core_t	*iclog;
4043
4044	iclog = log->l_iclog;
4045	do {
4046		/* endianness does not matter here, zero is zero in
4047		 * any language.
4048		 */
4049		if (iclog->ic_header.h_num_logops)
4050			return 0;
4051		iclog = iclog->ic_next;
4052	} while (iclog != log->l_iclog);
4053	return 1;
4054}
4055
4056/*
4057 * Verify that an LSN stamped into a piece of metadata is valid. This is
4058 * intended for use in read verifiers on v5 superblocks.
4059 */
4060bool
4061xfs_log_check_lsn(
4062	struct xfs_mount	*mp,
4063	xfs_lsn_t		lsn)
4064{
4065	struct xlog		*log = mp->m_log;
4066	bool			valid;
4067
4068	/*
4069	 * norecovery mode skips mount-time log processing and unconditionally
4070	 * resets the in-core LSN. We can't validate in this mode, but
4071	 * modifications are not allowed anyways so just return true.
4072	 */
4073	if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4074		return true;
4075
4076	/*
4077	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4078	 * handled by recovery and thus safe to ignore here.
4079	 */
4080	if (lsn == NULLCOMMITLSN)
4081		return true;
4082
4083	valid = xlog_valid_lsn(mp->m_log, lsn);
4084
4085	/* warn the user about what's gone wrong before verifier failure */
4086	if (!valid) {
4087		spin_lock(&log->l_icloglock);
4088		xfs_warn(mp,
4089"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4090"Please unmount and run xfs_repair (>= v4.3) to resolve.",
4091			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4092			 log->l_curr_cycle, log->l_curr_block);
4093		spin_unlock(&log->l_icloglock);
4094	}
4095
4096	return valid;
4097}
v6.13.7
   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
  24struct kmem_cache	*xfs_log_ticket_cache;
  25
  26/* Local miscellaneous function prototypes */
 
 
 
 
 
 
 
  27STATIC struct xlog *
  28xlog_alloc_log(
  29	struct xfs_mount	*mp,
  30	struct xfs_buftarg	*log_target,
  31	xfs_daddr_t		blk_offset,
  32	int			num_bblks);
 
 
 
 
 
 
 
 
  33STATIC void
  34xlog_dealloc_log(
  35	struct xlog		*log);
  36
  37/* local state machine functions */
  38STATIC void xlog_state_done_syncing(
 
 
 
 
  39	struct xlog_in_core	*iclog);
  40STATIC void xlog_state_do_callback(
  41	struct xlog		*log);
  42STATIC int
  43xlog_state_get_iclog_space(
  44	struct xlog		*log,
  45	int			len,
  46	struct xlog_in_core	**iclog,
  47	struct xlog_ticket	*ticket,
 
  48	int			*logoffsetp);
 
 
 
 
  49STATIC void
  50xlog_sync(
  51	struct xlog		*log,
  52	struct xlog_in_core	*iclog,
 
 
 
 
 
 
 
 
 
 
 
 
 
  53	struct xlog_ticket	*ticket);
 
 
 
 
 
  54#if defined(DEBUG)
  55STATIC void
 
 
 
 
 
 
 
  56xlog_verify_iclog(
  57	struct xlog		*log,
  58	struct xlog_in_core	*iclog,
  59	int			count);
 
  60STATIC void
  61xlog_verify_tail_lsn(
  62	struct xlog		*log,
  63	struct xlog_in_core	*iclog);
 
  64#else
  65#define xlog_verify_iclog(a,b,c)
  66#define xlog_verify_tail_lsn(a,b)
 
 
  67#endif
  68
  69STATIC int
  70xlog_iclogs_empty(
  71	struct xlog		*log);
  72
  73static int
  74xfs_log_cover(struct xfs_mount *);
 
 
 
 
 
 
 
 
 
  75
  76/*
  77 * We need to make sure the buffer pointer returned is naturally aligned for the
  78 * biggest basic data type we put into it. We have already accounted for this
  79 * padding when sizing the buffer.
  80 *
  81 * However, this padding does not get written into the log, and hence we have to
  82 * track the space used by the log vectors separately to prevent log space hangs
  83 * due to inaccurate accounting (i.e. a leak) of the used log space through the
  84 * CIL context ticket.
  85 *
  86 * We also add space for the xlog_op_header that describes this region in the
  87 * log. This prepends the data region we return to the caller to copy their data
  88 * into, so do all the static initialisation of the ophdr now. Because the ophdr
  89 * is not 8 byte aligned, we have to be careful to ensure that we align the
  90 * start of the buffer such that the region we return to the call is 8 byte
  91 * aligned and packed against the tail of the ophdr.
  92 */
  93void *
  94xlog_prepare_iovec(
  95	struct xfs_log_vec	*lv,
  96	struct xfs_log_iovec	**vecp,
  97	uint			type)
  98{
  99	struct xfs_log_iovec	*vec = *vecp;
 100	struct xlog_op_header	*oph;
 101	uint32_t		len;
 102	void			*buf;
 103
 104	if (vec) {
 105		ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
 106		vec++;
 107	} else {
 108		vec = &lv->lv_iovecp[0];
 109	}
 110
 111	len = lv->lv_buf_len + sizeof(struct xlog_op_header);
 112	if (!IS_ALIGNED(len, sizeof(uint64_t))) {
 113		lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
 114					sizeof(struct xlog_op_header);
 115	}
 116
 117	vec->i_type = type;
 118	vec->i_addr = lv->lv_buf + lv->lv_buf_len;
 
 
 
 119
 120	oph = vec->i_addr;
 121	oph->oh_clientid = XFS_TRANSACTION;
 122	oph->oh_res2 = 0;
 123	oph->oh_flags = 0;
 
 
 
 
 124
 125	buf = vec->i_addr + sizeof(struct xlog_op_header);
 126	ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
 
 127
 128	*vecp = vec;
 129	return buf;
 130}
 131
 132static inline void
 133xlog_grant_sub_space(
 134	struct xlog_grant_head	*head,
 135	int64_t			bytes)
 136{
 137	atomic64_sub(bytes, &head->grant);
 138}
 139
 140static inline void
 141xlog_grant_add_space(
 142	struct xlog_grant_head	*head,
 143	int64_t			bytes)
 144{
 145	atomic64_add(bytes, &head->grant);
 146}
 147
 148static void
 149xlog_grant_head_init(
 150	struct xlog_grant_head	*head)
 151{
 152	atomic64_set(&head->grant, 0);
 153	INIT_LIST_HEAD(&head->waiters);
 154	spin_lock_init(&head->lock);
 155}
 156
 157void
 158xlog_grant_return_space(
 159	struct xlog	*log,
 160	xfs_lsn_t	old_head,
 161	xfs_lsn_t	new_head)
 162{
 163	int64_t		diff = xlog_lsn_sub(log, new_head, old_head);
 164
 165	xlog_grant_sub_space(&log->l_reserve_head, diff);
 166	xlog_grant_sub_space(&log->l_write_head, diff);
 167}
 168
 169/*
 170 * Return the space in the log between the tail and the head.  In the case where
 171 * we have overrun available reservation space, return 0. The memory barrier
 172 * pairs with the smp_wmb() in xlog_cil_ail_insert() to ensure that grant head
 173 * vs tail space updates are seen in the correct order and hence avoid
 174 * transients as space is transferred from the grant heads to the AIL on commit
 175 * completion.
 176 */
 177static uint64_t
 178xlog_grant_space_left(
 179	struct xlog		*log,
 180	struct xlog_grant_head	*head)
 181{
 182	int64_t			free_bytes;
 183
 184	smp_rmb();	/* paired with smp_wmb in xlog_cil_ail_insert() */
 185	free_bytes = log->l_logsize - READ_ONCE(log->l_tail_space) -
 186			atomic64_read(&head->grant);
 187	if (free_bytes > 0)
 188		return free_bytes;
 189	return 0;
 190}
 191
 192STATIC void
 193xlog_grant_head_wake_all(
 194	struct xlog_grant_head	*head)
 195{
 196	struct xlog_ticket	*tic;
 197
 198	spin_lock(&head->lock);
 199	list_for_each_entry(tic, &head->waiters, t_queue)
 200		wake_up_process(tic->t_task);
 201	spin_unlock(&head->lock);
 202}
 203
 204static inline int
 205xlog_ticket_reservation(
 206	struct xlog		*log,
 207	struct xlog_grant_head	*head,
 208	struct xlog_ticket	*tic)
 209{
 210	if (head == &log->l_write_head) {
 211		ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 212		return tic->t_unit_res;
 
 
 
 
 
 213	}
 214
 215	if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 216		return tic->t_unit_res * tic->t_cnt;
 217
 218	return tic->t_unit_res;
 219}
 220
 221STATIC bool
 222xlog_grant_head_wake(
 223	struct xlog		*log,
 224	struct xlog_grant_head	*head,
 225	int			*free_bytes)
 226{
 227	struct xlog_ticket	*tic;
 228	int			need_bytes;
 229
 230	list_for_each_entry(tic, &head->waiters, t_queue) {
 231		need_bytes = xlog_ticket_reservation(log, head, tic);
 232		if (*free_bytes < need_bytes)
 233			return false;
 234
 235		*free_bytes -= need_bytes;
 236		trace_xfs_log_grant_wake_up(log, tic);
 237		wake_up_process(tic->t_task);
 238	}
 239
 240	return true;
 241}
 242
 243STATIC int
 244xlog_grant_head_wait(
 245	struct xlog		*log,
 246	struct xlog_grant_head	*head,
 247	struct xlog_ticket	*tic,
 248	int			need_bytes) __releases(&head->lock)
 249					    __acquires(&head->lock)
 250{
 251	list_add_tail(&tic->t_queue, &head->waiters);
 252
 253	do {
 254		if (xlog_is_shutdown(log))
 255			goto shutdown;
 
 256
 257		__set_current_state(TASK_UNINTERRUPTIBLE);
 258		spin_unlock(&head->lock);
 259
 260		XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 261
 262		/* Push on the AIL to free up all the log space. */
 263		xfs_ail_push_all(log->l_ailp);
 264
 265		trace_xfs_log_grant_sleep(log, tic);
 266		schedule();
 267		trace_xfs_log_grant_wake(log, tic);
 268
 269		spin_lock(&head->lock);
 270		if (xlog_is_shutdown(log))
 271			goto shutdown;
 272	} while (xlog_grant_space_left(log, head) < need_bytes);
 273
 274	list_del_init(&tic->t_queue);
 275	return 0;
 276shutdown:
 277	list_del_init(&tic->t_queue);
 278	return -EIO;
 279}
 280
 281/*
 282 * Atomically get the log space required for a log ticket.
 283 *
 284 * Once a ticket gets put onto head->waiters, it will only return after the
 285 * needed reservation is satisfied.
 286 *
 287 * This function is structured so that it has a lock free fast path. This is
 288 * necessary because every new transaction reservation will come through this
 289 * path. Hence any lock will be globally hot if we take it unconditionally on
 290 * every pass.
 291 *
 292 * As tickets are only ever moved on and off head->waiters under head->lock, we
 293 * only need to take that lock if we are going to add the ticket to the queue
 294 * and sleep. We can avoid taking the lock if the ticket was never added to
 295 * head->waiters because the t_queue list head will be empty and we hold the
 296 * only reference to it so it can safely be checked unlocked.
 297 */
 298STATIC int
 299xlog_grant_head_check(
 300	struct xlog		*log,
 301	struct xlog_grant_head	*head,
 302	struct xlog_ticket	*tic,
 303	int			*need_bytes)
 304{
 305	int			free_bytes;
 306	int			error = 0;
 307
 308	ASSERT(!xlog_in_recovery(log));
 309
 310	/*
 311	 * If there are other waiters on the queue then give them a chance at
 312	 * logspace before us.  Wake up the first waiters, if we do not wake
 313	 * up all the waiters then go to sleep waiting for more free space,
 314	 * otherwise try to get some space for this transaction.
 315	 */
 316	*need_bytes = xlog_ticket_reservation(log, head, tic);
 317	free_bytes = xlog_grant_space_left(log, head);
 318	if (!list_empty_careful(&head->waiters)) {
 319		spin_lock(&head->lock);
 320		if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 321		    free_bytes < *need_bytes) {
 322			error = xlog_grant_head_wait(log, head, tic,
 323						     *need_bytes);
 324		}
 325		spin_unlock(&head->lock);
 326	} else if (free_bytes < *need_bytes) {
 327		spin_lock(&head->lock);
 328		error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 329		spin_unlock(&head->lock);
 330	}
 331
 332	return error;
 333}
 334
 335bool
 336xfs_log_writable(
 337	struct xfs_mount	*mp)
 
 
 
 
 
 
 
 338{
 339	/*
 340	 * Do not write to the log on norecovery mounts, if the data or log
 341	 * devices are read-only, or if the filesystem is shutdown. Read-only
 342	 * mounts allow internal writes for log recovery and unmount purposes,
 343	 * so don't restrict that case.
 344	 */
 345	if (xfs_has_norecovery(mp))
 346		return false;
 347	if (xfs_readonly_buftarg(mp->m_ddev_targp))
 348		return false;
 349	if (xfs_readonly_buftarg(mp->m_log->l_targ))
 350		return false;
 351	if (xlog_is_shutdown(mp->m_log))
 352		return false;
 353	return true;
 354}
 355
 356/*
 357 * Replenish the byte reservation required by moving the grant write head.
 358 */
 359int
 360xfs_log_regrant(
 361	struct xfs_mount	*mp,
 362	struct xlog_ticket	*tic)
 363{
 364	struct xlog		*log = mp->m_log;
 365	int			need_bytes;
 366	int			error = 0;
 367
 368	if (xlog_is_shutdown(log))
 369		return -EIO;
 370
 371	XFS_STATS_INC(mp, xs_try_logspace);
 372
 373	/*
 374	 * This is a new transaction on the ticket, so we need to change the
 375	 * transaction ID so that the next transaction has a different TID in
 376	 * the log. Just add one to the existing tid so that we can see chains
 377	 * of rolling transactions in the log easily.
 378	 */
 379	tic->t_tid++;
 
 
 
 380	tic->t_curr_res = tic->t_unit_res;
 
 
 381	if (tic->t_cnt > 0)
 382		return 0;
 383
 384	trace_xfs_log_regrant(log, tic);
 385
 386	error = xlog_grant_head_check(log, &log->l_write_head, tic,
 387				      &need_bytes);
 388	if (error)
 389		goto out_error;
 390
 391	xlog_grant_add_space(&log->l_write_head, need_bytes);
 392	trace_xfs_log_regrant_exit(log, tic);
 
 393	return 0;
 394
 395out_error:
 396	/*
 397	 * If we are failing, make sure the ticket doesn't have any current
 398	 * reservations.  We don't want to add this back when the ticket/
 399	 * transaction gets cancelled.
 400	 */
 401	tic->t_curr_res = 0;
 402	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
 403	return error;
 404}
 405
 406/*
 407 * Reserve log space and return a ticket corresponding to the reservation.
 408 *
 409 * Each reservation is going to reserve extra space for a log record header.
 410 * When writes happen to the on-disk log, we don't subtract the length of the
 411 * log record header from any reservation.  By wasting space in each
 412 * reservation, we prevent over allocation problems.
 413 */
 414int
 415xfs_log_reserve(
 416	struct xfs_mount	*mp,
 417	int			unit_bytes,
 418	int			cnt,
 419	struct xlog_ticket	**ticp,
 
 420	bool			permanent)
 421{
 422	struct xlog		*log = mp->m_log;
 423	struct xlog_ticket	*tic;
 424	int			need_bytes;
 425	int			error = 0;
 426
 427	if (xlog_is_shutdown(log))
 
 
 428		return -EIO;
 429
 430	XFS_STATS_INC(mp, xs_try_logspace);
 431
 432	ASSERT(*ticp == NULL);
 433	tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
 
 
 
 
 434	*ticp = tic;
 
 
 
 
 435	trace_xfs_log_reserve(log, tic);
 
 436	error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 437				      &need_bytes);
 438	if (error)
 439		goto out_error;
 440
 441	xlog_grant_add_space(&log->l_reserve_head, need_bytes);
 442	xlog_grant_add_space(&log->l_write_head, need_bytes);
 443	trace_xfs_log_reserve_exit(log, tic);
 
 444	return 0;
 445
 446out_error:
 447	/*
 448	 * If we are failing, make sure the ticket doesn't have any current
 449	 * reservations.  We don't want to add this back when the ticket/
 450	 * transaction gets cancelled.
 451	 */
 452	tic->t_curr_res = 0;
 453	tic->t_cnt = 0;	/* ungrant will give back unit_res * t_cnt. */
 454	return error;
 455}
 456
 
 
 
 
 
 
 
 
 457/*
 458 * Run all the pending iclog callbacks and wake log force waiters and iclog
 459 * space waiters so they can process the newly set shutdown state. We really
 460 * don't care what order we process callbacks here because the log is shut down
 461 * and so state cannot change on disk anymore. However, we cannot wake waiters
 462 * until the callbacks have been processed because we may be in unmount and
 463 * we must ensure that all AIL operations the callbacks perform have completed
 464 * before we tear down the AIL.
 465 *
 466 * We avoid processing actively referenced iclogs so that we don't run callbacks
 467 * while the iclog owner might still be preparing the iclog for IO submssion.
 468 * These will be caught by xlog_state_iclog_release() and call this function
 469 * again to process any callbacks that may have been added to that iclog.
 470 */
 471static void
 472xlog_state_shutdown_callbacks(
 473	struct xlog		*log)
 
 
 
 474{
 475	struct xlog_in_core	*iclog;
 476	LIST_HEAD(cb_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 477
 478	iclog = log->l_iclog;
 479	do {
 480		if (atomic_read(&iclog->ic_refcnt)) {
 481			/* Reference holder will re-run iclog callbacks. */
 482			continue;
 483		}
 484		list_splice_init(&iclog->ic_callbacks, &cb_list);
 485		spin_unlock(&log->l_icloglock);
 486
 487		xlog_cil_process_committed(&cb_list);
 
 
 
 
 
 
 488
 489		spin_lock(&log->l_icloglock);
 490		wake_up_all(&iclog->ic_write_wait);
 491		wake_up_all(&iclog->ic_force_wait);
 492	} while ((iclog = iclog->ic_next) != log->l_iclog);
 493
 494	wake_up_all(&log->l_flush_wait);
 495}
 496
 497/*
 498 * Flush iclog to disk if this is the last reference to the given iclog and the
 499 * it is in the WANT_SYNC state.
 500 *
 501 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
 502 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
 503 * written to stable storage, and implies that a commit record is contained
 504 * within the iclog. We need to ensure that the log tail does not move beyond
 505 * the tail that the first commit record in the iclog ordered against, otherwise
 506 * correct recovery of that checkpoint becomes dependent on future operations
 507 * performed on this iclog.
 508 *
 509 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
 510 * current tail into iclog. Once the iclog tail is set, future operations must
 511 * not modify it, otherwise they potentially violate ordering constraints for
 512 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
 513 * the iclog will get zeroed on activation of the iclog after sync, so we
 514 * always capture the tail lsn on the iclog on the first NEED_FUA release
 515 * regardless of the number of active reference counts on this iclog.
 516 */
 517int
 518xlog_state_release_iclog(
 519	struct xlog		*log,
 520	struct xlog_in_core	*iclog,
 521	struct xlog_ticket	*ticket)
 522{
 523	bool			last_ref;
 524
 525	lockdep_assert_held(&log->l_icloglock);
 526
 527	trace_xlog_iclog_release(iclog, _RET_IP_);
 528	/*
 529	 * Grabbing the current log tail needs to be atomic w.r.t. the writing
 530	 * of the tail LSN into the iclog so we guarantee that the log tail does
 531	 * not move between the first time we know that the iclog needs to be
 532	 * made stable and when we eventually submit it.
 533	 */
 534	if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
 535	     (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
 536	    !iclog->ic_header.h_tail_lsn) {
 537		iclog->ic_header.h_tail_lsn =
 538				cpu_to_be64(atomic64_read(&log->l_tail_lsn));
 539	}
 
 
 
 540
 541	last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
 542
 543	if (xlog_is_shutdown(log)) {
 544		/*
 545		 * If there are no more references to this iclog, process the
 546		 * pending iclog callbacks that were waiting on the release of
 547		 * this iclog.
 548		 */
 549		if (last_ref)
 550			xlog_state_shutdown_callbacks(log);
 551		return -EIO;
 552	}
 553
 554	if (!last_ref)
 555		return 0;
 556
 557	if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
 558		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 559		return 0;
 560	}
 561
 562	iclog->ic_state = XLOG_STATE_SYNCING;
 563	xlog_verify_tail_lsn(log, iclog);
 564	trace_xlog_iclog_syncing(iclog, _RET_IP_);
 565
 566	spin_unlock(&log->l_icloglock);
 567	xlog_sync(log, iclog, ticket);
 568	spin_lock(&log->l_icloglock);
 569	return 0;
 570}
 571
 572/*
 573 * Mount a log filesystem
 574 *
 575 * mp		- ubiquitous xfs mount point structure
 576 * log_target	- buftarg of on-disk log device
 577 * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
 578 * num_bblocks	- Number of BBSIZE blocks in on-disk log
 579 *
 580 * Return error or zero.
 581 */
 582int
 583xfs_log_mount(
 584	xfs_mount_t		*mp,
 585	struct xfs_buftarg	*log_target,
 586	xfs_daddr_t		blk_offset,
 587	int			num_bblks)
 588{
 589	struct xlog		*log;
 590	int			error = 0;
 591	int			min_logfsbs;
 592
 593	if (!xfs_has_norecovery(mp)) {
 594		xfs_notice(mp, "Mounting V%d Filesystem %pU",
 595			   XFS_SB_VERSION_NUM(&mp->m_sb),
 596			   &mp->m_sb.sb_uuid);
 597	} else {
 598		xfs_notice(mp,
 599"Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
 600			   XFS_SB_VERSION_NUM(&mp->m_sb),
 601			   &mp->m_sb.sb_uuid);
 602		ASSERT(xfs_is_readonly(mp));
 603	}
 604
 605	log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 606	if (IS_ERR(log)) {
 607		error = PTR_ERR(log);
 608		goto out;
 609	}
 610	mp->m_log = log;
 611
 612	/*
 613	 * Now that we have set up the log and it's internal geometry
 614	 * parameters, we can validate the given log space and drop a critical
 615	 * message via syslog if the log size is too small. A log that is too
 616	 * small can lead to unexpected situations in transaction log space
 617	 * reservation stage. The superblock verifier has already validated all
 618	 * the other log geometry constraints, so we don't have to check those
 619	 * here.
 620	 *
 621	 * Note: For v4 filesystems, we can't just reject the mount if the
 622	 * validation fails.  This would mean that people would have to
 623	 * downgrade their kernel just to remedy the situation as there is no
 624	 * way to grow the log (short of black magic surgery with xfs_db).
 625	 *
 626	 * We can, however, reject mounts for V5 format filesystems, as the
 627	 * mkfs binary being used to make the filesystem should never create a
 628	 * filesystem with a log that is too small.
 629	 */
 630	min_logfsbs = xfs_log_calc_minimum_size(mp);
 
 631	if (mp->m_sb.sb_logblocks < min_logfsbs) {
 632		xfs_warn(mp,
 633		"Log size %d blocks too small, minimum size is %d blocks",
 634			 mp->m_sb.sb_logblocks, min_logfsbs);
 635
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 636		/*
 637		 * Log check errors are always fatal on v5; or whenever bad
 638		 * metadata leads to a crash.
 639		 */
 640		if (xfs_has_crc(mp)) {
 641			xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 642			ASSERT(0);
 643			error = -EINVAL;
 644			goto out_free_log;
 645		}
 646		xfs_crit(mp, "Log size out of supported range.");
 647		xfs_crit(mp,
 648"Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 649	}
 650
 651	/*
 652	 * Initialize the AIL now we have a log.
 653	 */
 654	error = xfs_trans_ail_init(mp);
 655	if (error) {
 656		xfs_warn(mp, "AIL initialisation failed: error %d", error);
 657		goto out_free_log;
 658	}
 659	log->l_ailp = mp->m_ail;
 660
 661	/*
 662	 * skip log recovery on a norecovery mount.  pretend it all
 663	 * just worked.
 664	 */
 665	if (!xfs_has_norecovery(mp)) {
 666		error = xlog_recover(log);
 
 
 
 
 
 
 
 
 667		if (error) {
 668			xfs_warn(mp, "log mount/recovery failed: error %d",
 669				error);
 670			xlog_recover_cancel(log);
 671			goto out_destroy_ail;
 672		}
 673	}
 674
 675	error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 676			       "log");
 677	if (error)
 678		goto out_destroy_ail;
 679
 680	/* Normal transactions can now occur */
 681	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
 682
 683	/*
 684	 * Now the log has been fully initialised and we know were our
 685	 * space grant counters are, we can initialise the permanent ticket
 686	 * needed for delayed logging to work.
 687	 */
 688	xlog_cil_init_post_recovery(log);
 689
 690	return 0;
 691
 692out_destroy_ail:
 693	xfs_trans_ail_destroy(mp);
 694out_free_log:
 695	xlog_dealloc_log(log);
 696out:
 697	return error;
 698}
 699
 700/*
 701 * Finish the recovery of the file system.  This is separate from the
 702 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 703 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 704 * here.
 705 *
 706 * If we finish recovery successfully, start the background log work. If we are
 707 * not doing recovery, then we have a RO filesystem and we don't need to start
 708 * it.
 709 */
 710int
 711xfs_log_mount_finish(
 712	struct xfs_mount	*mp)
 713{
 714	struct xlog		*log = mp->m_log;
 715	int			error = 0;
 
 716
 717	if (xfs_has_norecovery(mp)) {
 718		ASSERT(xfs_is_readonly(mp));
 719		return 0;
 
 
 
 720	}
 721
 722	/*
 723	 * During the second phase of log recovery, we need iget and
 724	 * iput to behave like they do for an active filesystem.
 725	 * xfs_fs_drop_inode needs to be able to prevent the deletion
 726	 * of inodes before we're done replaying log items on those
 727	 * inodes.  Turn it off immediately after recovery finishes
 728	 * so that we don't leak the quota inodes if subsequent mount
 729	 * activities fail.
 730	 *
 731	 * We let all inodes involved in redo item processing end up on
 732	 * the LRU instead of being evicted immediately so that if we do
 733	 * something to an unlinked inode, the irele won't cause
 734	 * premature truncation and freeing of the inode, which results
 735	 * in log recovery failure.  We have to evict the unreferenced
 736	 * lru inodes after clearing SB_ACTIVE because we don't
 737	 * otherwise clean up the lru if there's a subsequent failure in
 738	 * xfs_mountfs, which leads to us leaking the inodes if nothing
 739	 * else (e.g. quotacheck) references the inodes before the
 740	 * mount failure occurs.
 741	 */
 742	mp->m_super->s_flags |= SB_ACTIVE;
 743	xfs_log_work_queue(mp);
 744	if (xlog_recovery_needed(log))
 745		error = xlog_recover_finish(log);
 746	mp->m_super->s_flags &= ~SB_ACTIVE;
 747	evict_inodes(mp->m_super);
 748
 749	/*
 750	 * Drain the buffer LRU after log recovery. This is required for v4
 751	 * filesystems to avoid leaving around buffers with NULL verifier ops,
 752	 * but we do it unconditionally to make sure we're always in a clean
 753	 * cache state after mount.
 754	 *
 755	 * Don't push in the error case because the AIL may have pending intents
 756	 * that aren't removed until recovery is cancelled.
 757	 */
 758	if (xlog_recovery_needed(log)) {
 759		if (!error) {
 760			xfs_log_force(mp, XFS_LOG_SYNC);
 761			xfs_ail_push_all_sync(mp->m_ail);
 762		}
 763		xfs_notice(mp, "Ending recovery (logdev: %s)",
 764				mp->m_logname ? mp->m_logname : "internal");
 765	} else {
 766		xfs_info(mp, "Ending clean mount");
 767	}
 768	xfs_buftarg_drain(mp->m_ddev_targp);
 769
 770	clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
 771
 772	/* Make sure the log is dead if we're returning failure. */
 773	ASSERT(!error || xlog_is_shutdown(log));
 774
 775	return error;
 776}
 777
 778/*
 779 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 780 * the log.
 781 */
 782void
 783xfs_log_mount_cancel(
 784	struct xfs_mount	*mp)
 785{
 786	xlog_recover_cancel(mp->m_log);
 
 
 787	xfs_log_unmount(mp);
 788}
 789
 790/*
 791 * Flush out the iclog to disk ensuring that device caches are flushed and
 792 * the iclog hits stable storage before any completion waiters are woken.
 793 */
 794static inline int
 795xlog_force_iclog(
 796	struct xlog_in_core	*iclog)
 797{
 798	atomic_inc(&iclog->ic_refcnt);
 799	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
 800	if (iclog->ic_state == XLOG_STATE_ACTIVE)
 801		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
 802	return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
 803}
 804
 805/*
 806 * Cycle all the iclogbuf locks to make sure all log IO completion
 807 * is done before we tear down these buffers.
 
 
 
 808 */
 809static void
 810xlog_wait_iclog_completion(struct xlog *log)
 811{
 812	int		i;
 813	struct xlog_in_core	*iclog = log->l_iclog;
 814
 815	for (i = 0; i < log->l_iclog_bufs; i++) {
 816		down(&iclog->ic_sema);
 817		up(&iclog->ic_sema);
 818		iclog = iclog->ic_next;
 819	}
 820}
 821
 822/*
 823 * Wait for the iclog and all prior iclogs to be written disk as required by the
 824 * log force state machine. Waiting on ic_force_wait ensures iclog completions
 825 * have been ordered and callbacks run before we are woken here, hence
 826 * guaranteeing that all the iclogs up to this one are on stable storage.
 
 827 */
 828int
 829xlog_wait_on_iclog(
 830	struct xlog_in_core	*iclog)
 831		__releases(iclog->ic_log->l_icloglock)
 832{
 833	struct xlog		*log = iclog->ic_log;
 834
 835	trace_xlog_iclog_wait_on(iclog, _RET_IP_);
 836	if (!xlog_is_shutdown(log) &&
 837	    iclog->ic_state != XLOG_STATE_ACTIVE &&
 838	    iclog->ic_state != XLOG_STATE_DIRTY) {
 839		XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
 840		xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
 841	} else {
 842		spin_unlock(&log->l_icloglock);
 843	}
 844
 845	if (xlog_is_shutdown(log))
 846		return -EIO;
 847	return 0;
 848}
 849
 850/*
 851 * Write out an unmount record using the ticket provided. We have to account for
 852 * the data space used in the unmount ticket as this write is not done from a
 853 * transaction context that has already done the accounting for us.
 854 */
 855static int
 856xlog_write_unmount_record(
 857	struct xlog		*log,
 858	struct xlog_ticket	*ticket)
 859{
 860	struct  {
 861		struct xlog_op_header ophdr;
 862		struct xfs_unmount_log_format ulf;
 863	} unmount_rec = {
 864		.ophdr = {
 865			.oh_clientid = XFS_LOG,
 866			.oh_tid = cpu_to_be32(ticket->t_tid),
 867			.oh_flags = XLOG_UNMOUNT_TRANS,
 868		},
 869		.ulf = {
 870			.magic = XLOG_UNMOUNT_TYPE,
 871		},
 872	};
 873	struct xfs_log_iovec reg = {
 874		.i_addr = &unmount_rec,
 875		.i_len = sizeof(unmount_rec),
 876		.i_type = XLOG_REG_TYPE_UNMOUNT,
 877	};
 878	struct xfs_log_vec vec = {
 879		.lv_niovecs = 1,
 880		.lv_iovecp = &reg,
 881	};
 882	LIST_HEAD(lv_chain);
 883	list_add(&vec.lv_list, &lv_chain);
 884
 885	BUILD_BUG_ON((sizeof(struct xlog_op_header) +
 886		      sizeof(struct xfs_unmount_log_format)) !=
 887							sizeof(unmount_rec));
 888
 889	/* account for space used by record data */
 890	ticket->t_curr_res -= sizeof(unmount_rec);
 891
 892	return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
 893}
 894
 895/*
 896 * Mark the filesystem clean by writing an unmount record to the head of the
 897 * log.
 898 */
 899static void
 900xlog_unmount_write(
 901	struct xlog		*log)
 902{
 903	struct xfs_mount	*mp = log->l_mp;
 904	struct xlog_in_core	*iclog;
 905	struct xlog_ticket	*tic = NULL;
 906	int			error;
 907
 908	error = xfs_log_reserve(mp, 600, 1, &tic, 0);
 909	if (error)
 910		goto out_err;
 911
 912	error = xlog_write_unmount_record(log, tic);
 913	/*
 914	 * At this point, we're umounting anyway, so there's no point in
 915	 * transitioning log state to shutdown. Just continue...
 916	 */
 917out_err:
 918	if (error)
 919		xfs_alert(mp, "%s: unmount record failed", __func__);
 
 
 
 
 
 920
 921	spin_lock(&log->l_icloglock);
 922	iclog = log->l_iclog;
 923	error = xlog_force_iclog(iclog);
 924	xlog_wait_on_iclog(iclog);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 925
 926	if (tic) {
 927		trace_xfs_log_umount_write(log, tic);
 928		xfs_log_ticket_ungrant(log, tic);
 929	}
 930}
 931
 932static void
 933xfs_log_unmount_verify_iclog(
 934	struct xlog		*log)
 935{
 936	struct xlog_in_core	*iclog = log->l_iclog;
 937
 938	do {
 939		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 940		ASSERT(iclog->ic_offset == 0);
 941	} while ((iclog = iclog->ic_next) != log->l_iclog);
 942}
 
 943
 944/*
 945 * Unmount record used to have a string "Unmount filesystem--" in the
 946 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 947 * We just write the magic number now since that particular field isn't
 948 * currently architecture converted and "Unmount" is a bit foo.
 949 * As far as I know, there weren't any dependencies on the old behaviour.
 950 */
 951static void
 952xfs_log_unmount_write(
 953	struct xfs_mount	*mp)
 954{
 955	struct xlog		*log = mp->m_log;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 956
 957	if (!xfs_log_writable(mp))
 958		return;
 
 959
 960	xfs_log_force(mp, XFS_LOG_SYNC);
 961
 962	if (xlog_is_shutdown(log))
 963		return;
 
 964
 965	/*
 966	 * If we think the summary counters are bad, avoid writing the unmount
 967	 * record to force log recovery at next mount, after which the summary
 968	 * counters will be recalculated.  Refer to xlog_check_unmount_rec for
 969	 * more details.
 970	 */
 971	if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
 972			XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
 973		xfs_alert(mp, "%s: will fix summary counters at next mount",
 974				__func__);
 975		return;
 976	}
 977
 978	xfs_log_unmount_verify_iclog(log);
 979	xlog_unmount_write(log);
 980}
 981
 982/*
 983 * Empty the log for unmount/freeze.
 984 *
 985 * To do this, we first need to shut down the background log work so it is not
 986 * trying to cover the log as we clean up. We then need to unpin all objects in
 987 * the log so we can then flush them out. Once they have completed their IO and
 988 * run the callbacks removing themselves from the AIL, we can cover the log.
 
 989 */
 990int
 991xfs_log_quiesce(
 992	struct xfs_mount	*mp)
 993{
 994	/*
 995	 * Clear log incompat features since we're quiescing the log.  Report
 996	 * failures, though it's not fatal to have a higher log feature
 997	 * protection level than the log contents actually require.
 998	 */
 999	if (xfs_clear_incompat_log_features(mp)) {
1000		int error;
1001
1002		error = xfs_sync_sb(mp, false);
1003		if (error)
1004			xfs_warn(mp,
1005	"Failed to clear log incompat features on quiesce");
1006	}
1007
1008	cancel_delayed_work_sync(&mp->m_log->l_work);
1009	xfs_log_force(mp, XFS_LOG_SYNC);
1010
1011	/*
1012	 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1013	 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1014	 * xfs_buf_iowait() cannot be used because it was pushed with the
1015	 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1016	 * the IO to complete.
1017	 */
1018	xfs_ail_push_all_sync(mp->m_ail);
1019	xfs_buftarg_wait(mp->m_ddev_targp);
1020	xfs_buf_lock(mp->m_sb_bp);
1021	xfs_buf_unlock(mp->m_sb_bp);
1022
1023	return xfs_log_cover(mp);
1024}
1025
1026void
1027xfs_log_clean(
1028	struct xfs_mount	*mp)
1029{
1030	xfs_log_quiesce(mp);
1031	xfs_log_unmount_write(mp);
1032}
1033
1034/*
1035 * Shut down and release the AIL and Log.
1036 *
1037 * During unmount, we need to ensure we flush all the dirty metadata objects
1038 * from the AIL so that the log is empty before we write the unmount record to
1039 * the log. Once this is done, we can tear down the AIL and the log.
1040 */
1041void
1042xfs_log_unmount(
1043	struct xfs_mount	*mp)
1044{
1045	xfs_log_clean(mp);
1046
1047	/*
1048	 * If shutdown has come from iclog IO context, the log
1049	 * cleaning will have been skipped and so we need to wait
1050	 * for the iclog to complete shutdown processing before we
1051	 * tear anything down.
1052	 */
1053	xlog_wait_iclog_completion(mp->m_log);
1054
1055	xfs_buftarg_drain(mp->m_ddev_targp);
1056
1057	xfs_trans_ail_destroy(mp);
1058
1059	xfs_sysfs_del(&mp->m_log->l_kobj);
1060
1061	xlog_dealloc_log(mp->m_log);
1062}
1063
1064void
1065xfs_log_item_init(
1066	struct xfs_mount	*mp,
1067	struct xfs_log_item	*item,
1068	int			type,
1069	const struct xfs_item_ops *ops)
1070{
1071	item->li_log = mp->m_log;
1072	item->li_ailp = mp->m_ail;
1073	item->li_type = type;
1074	item->li_ops = ops;
1075	item->li_lv = NULL;
1076
1077	INIT_LIST_HEAD(&item->li_ail);
1078	INIT_LIST_HEAD(&item->li_cil);
1079	INIT_LIST_HEAD(&item->li_bio_list);
1080	INIT_LIST_HEAD(&item->li_trans);
1081}
1082
1083/*
1084 * Wake up processes waiting for log space after we have moved the log tail.
1085 */
1086void
1087xfs_log_space_wake(
1088	struct xfs_mount	*mp)
1089{
1090	struct xlog		*log = mp->m_log;
1091	int			free_bytes;
1092
1093	if (xlog_is_shutdown(log))
1094		return;
1095
1096	if (!list_empty_careful(&log->l_write_head.waiters)) {
1097		ASSERT(!xlog_in_recovery(log));
1098
1099		spin_lock(&log->l_write_head.lock);
1100		free_bytes = xlog_grant_space_left(log, &log->l_write_head);
1101		xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1102		spin_unlock(&log->l_write_head.lock);
1103	}
1104
1105	if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1106		ASSERT(!xlog_in_recovery(log));
1107
1108		spin_lock(&log->l_reserve_head.lock);
1109		free_bytes = xlog_grant_space_left(log, &log->l_reserve_head);
1110		xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1111		spin_unlock(&log->l_reserve_head.lock);
1112	}
1113}
1114
1115/*
1116 * Determine if we have a transaction that has gone to disk that needs to be
1117 * covered. To begin the transition to the idle state firstly the log needs to
1118 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1119 * we start attempting to cover the log.
1120 *
1121 * Only if we are then in a state where covering is needed, the caller is
1122 * informed that dummy transactions are required to move the log into the idle
1123 * state.
1124 *
1125 * If there are any items in the AIl or CIL, then we do not want to attempt to
1126 * cover the log as we may be in a situation where there isn't log space
1127 * available to run a dummy transaction and this can lead to deadlocks when the
1128 * tail of the log is pinned by an item that is modified in the CIL.  Hence
1129 * there's no point in running a dummy transaction at this point because we
1130 * can't start trying to idle the log until both the CIL and AIL are empty.
1131 */
1132static bool
1133xfs_log_need_covered(
1134	struct xfs_mount	*mp)
1135{
1136	struct xlog		*log = mp->m_log;
1137	bool			needed = false;
 
 
 
1138
1139	if (!xlog_cil_empty(log))
1140		return false;
1141
1142	spin_lock(&log->l_icloglock);
1143	switch (log->l_covered_state) {
1144	case XLOG_STATE_COVER_DONE:
1145	case XLOG_STATE_COVER_DONE2:
1146	case XLOG_STATE_COVER_IDLE:
1147		break;
1148	case XLOG_STATE_COVER_NEED:
1149	case XLOG_STATE_COVER_NEED2:
1150		if (xfs_ail_min_lsn(log->l_ailp))
1151			break;
1152		if (!xlog_iclogs_empty(log))
1153			break;
1154
1155		needed = true;
1156		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1157			log->l_covered_state = XLOG_STATE_COVER_DONE;
1158		else
1159			log->l_covered_state = XLOG_STATE_COVER_DONE2;
1160		break;
1161	default:
1162		needed = true;
1163		break;
1164	}
1165	spin_unlock(&log->l_icloglock);
1166	return needed;
1167}
1168
1169/*
1170 * Explicitly cover the log. This is similar to background log covering but
1171 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1172 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1173 * must all be empty.
1174 */
1175static int
1176xfs_log_cover(
1177	struct xfs_mount	*mp)
1178{
1179	int			error = 0;
1180	bool			need_covered;
1181
1182	ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1183	        !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1184		xlog_is_shutdown(mp->m_log));
1185
1186	if (!xfs_log_writable(mp))
1187		return 0;
1188
1189	/*
1190	 * xfs_log_need_covered() is not idempotent because it progresses the
1191	 * state machine if the log requires covering. Therefore, we must call
1192	 * this function once and use the result until we've issued an sb sync.
1193	 * Do so first to make that abundantly clear.
1194	 *
1195	 * Fall into the covering sequence if the log needs covering or the
1196	 * mount has lazy superblock accounting to sync to disk. The sb sync
1197	 * used for covering accumulates the in-core counters, so covering
1198	 * handles this for us.
1199	 */
1200	need_covered = xfs_log_need_covered(mp);
1201	if (!need_covered && !xfs_has_lazysbcount(mp))
1202		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1203
1204	/*
1205	 * To cover the log, commit the superblock twice (at most) in
1206	 * independent checkpoints. The first serves as a reference for the
1207	 * tail pointer. The sync transaction and AIL push empties the AIL and
1208	 * updates the in-core tail to the LSN of the first checkpoint. The
1209	 * second commit updates the on-disk tail with the in-core LSN,
1210	 * covering the log. Push the AIL one more time to leave it empty, as
1211	 * we found it.
1212	 */
1213	do {
1214		error = xfs_sync_sb(mp, true);
1215		if (error)
1216			break;
1217		xfs_ail_push_all_sync(mp->m_ail);
1218	} while (xfs_log_need_covered(mp));
1219
1220	return error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1221}
1222
 
 
 
 
 
 
 
1223static void
1224xlog_ioend_work(
1225	struct work_struct	*work)
1226{
1227	struct xlog_in_core     *iclog =
1228		container_of(work, struct xlog_in_core, ic_end_io_work);
1229	struct xlog		*log = iclog->ic_log;
1230	int			error;
1231
1232	error = blk_status_to_errno(iclog->ic_bio.bi_status);
1233#ifdef DEBUG
1234	/* treat writes with injected CRC errors as failed */
1235	if (iclog->ic_fail_crc)
1236		error = -EIO;
1237#endif
1238
1239	/*
1240	 * Race to shutdown the filesystem if we see an error.
 
 
1241	 */
1242	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1243		xfs_alert(log->l_mp, "log I/O error %d", error);
1244		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 
 
 
 
 
 
 
 
 
 
 
 
 
1245	}
1246
1247	xlog_state_done_syncing(iclog);
1248	bio_uninit(&iclog->ic_bio);
 
1249
1250	/*
1251	 * Drop the lock to signal that we are done. Nothing references the
1252	 * iclog after this, so an unmount waiting on this lock can now tear it
1253	 * down safely. As such, it is unsafe to reference the iclog after the
1254	 * unlock as we could race with it being freed.
1255	 */
1256	up(&iclog->ic_sema);
1257}
1258
1259/*
1260 * Return size of each in-core log record buffer.
1261 *
1262 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1263 *
1264 * If the filesystem blocksize is too large, we may need to choose a
1265 * larger size since the directory code currently logs entire blocks.
1266 */
 
1267STATIC void
1268xlog_get_iclog_buffer_size(
1269	struct xfs_mount	*mp,
1270	struct xlog		*log)
1271{
 
 
 
1272	if (mp->m_logbufs <= 0)
1273		mp->m_logbufs = XLOG_MAX_ICLOGS;
1274	if (mp->m_logbsize <= 0)
1275		mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1276
1277	log->l_iclog_bufs = mp->m_logbufs;
1278	log->l_iclog_size = mp->m_logbsize;
1279
1280	/*
1281	 * # headers = size / 32k - one header holds cycles from 32k of data.
1282	 */
1283	log->l_iclog_heads =
1284		DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1285	log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1286}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1287
1288void
1289xfs_log_work_queue(
1290	struct xfs_mount        *mp)
1291{
1292	queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1293				msecs_to_jiffies(xfs_syncd_centisecs * 10));
1294}
1295
1296/*
1297 * Clear the log incompat flags if we have the opportunity.
1298 *
1299 * This only happens if we're about to log the second dummy transaction as part
1300 * of covering the log.
1301 */
1302static inline void
1303xlog_clear_incompat(
1304	struct xlog		*log)
1305{
1306	struct xfs_mount	*mp = log->l_mp;
1307
1308	if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1309				XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1310		return;
1311
1312	if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1313		return;
1314
1315	xfs_clear_incompat_log_features(mp);
1316}
1317
1318/*
1319 * Every sync period we need to unpin all items in the AIL and push them to
1320 * disk. If there is nothing dirty, then we might need to cover the log to
1321 * indicate that the filesystem is idle.
1322 */
1323static void
1324xfs_log_worker(
1325	struct work_struct	*work)
1326{
1327	struct xlog		*log = container_of(to_delayed_work(work),
1328						struct xlog, l_work);
1329	struct xfs_mount	*mp = log->l_mp;
1330
1331	/* dgc: errors ignored - not fatal and nowhere to report them */
1332	if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1333		/*
1334		 * Dump a transaction into the log that contains no real change.
1335		 * This is needed to stamp the current tail LSN into the log
1336		 * during the covering operation.
1337		 *
1338		 * We cannot use an inode here for this - that will push dirty
1339		 * state back up into the VFS and then periodic inode flushing
1340		 * will prevent log covering from making progress. Hence we
1341		 * synchronously log the superblock instead to ensure the
1342		 * superblock is immediately unpinned and can be written back.
1343		 */
1344		xlog_clear_incompat(log);
1345		xfs_sync_sb(mp, true);
1346	} else
1347		xfs_log_force(mp, 0);
1348
1349	/* start pushing all the metadata that is currently dirty */
1350	xfs_ail_push_all(mp->m_ail);
1351
1352	/* queue us up again */
1353	xfs_log_work_queue(mp);
1354}
1355
1356/*
1357 * This routine initializes some of the log structure for a given mount point.
1358 * Its primary purpose is to fill in enough, so recovery can occur.  However,
1359 * some other stuff may be filled in too.
1360 */
1361STATIC struct xlog *
1362xlog_alloc_log(
1363	struct xfs_mount	*mp,
1364	struct xfs_buftarg	*log_target,
1365	xfs_daddr_t		blk_offset,
1366	int			num_bblks)
1367{
1368	struct xlog		*log;
1369	xlog_rec_header_t	*head;
1370	xlog_in_core_t		**iclogp;
1371	xlog_in_core_t		*iclog, *prev_iclog=NULL;
 
1372	int			i;
1373	int			error = -ENOMEM;
1374	uint			log2_size = 0;
1375
1376	log = kzalloc(sizeof(struct xlog), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1377	if (!log) {
1378		xfs_warn(mp, "Log allocation failed: No memory!");
1379		goto out;
1380	}
1381
1382	log->l_mp	   = mp;
1383	log->l_targ	   = log_target;
1384	log->l_logsize     = BBTOB(num_bblks);
1385	log->l_logBBstart  = blk_offset;
1386	log->l_logBBsize   = num_bblks;
1387	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1388	set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1389	INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1390	INIT_LIST_HEAD(&log->r_dfops);
1391
1392	log->l_prev_block  = -1;
1393	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1394	xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
 
1395	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1396
1397	if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1398		log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1399	else
1400		log->l_iclog_roundoff = BBSIZE;
1401
1402	xlog_grant_head_init(&log->l_reserve_head);
1403	xlog_grant_head_init(&log->l_write_head);
1404
1405	error = -EFSCORRUPTED;
1406	if (xfs_has_sector(mp)) {
1407	        log2_size = mp->m_sb.sb_logsectlog;
1408		if (log2_size < BBSHIFT) {
1409			xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1410				log2_size, BBSHIFT);
1411			goto out_free_log;
1412		}
1413
1414	        log2_size -= BBSHIFT;
1415		if (log2_size > mp->m_sectbb_log) {
1416			xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1417				log2_size, mp->m_sectbb_log);
1418			goto out_free_log;
1419		}
1420
1421		/* for larger sector sizes, must have v2 or external log */
1422		if (log2_size && log->l_logBBstart > 0 &&
1423			    !xfs_has_logv2(mp)) {
1424			xfs_warn(mp,
1425		"log sector size (0x%x) invalid for configuration.",
1426				log2_size);
1427			goto out_free_log;
1428		}
1429	}
1430	log->l_sectBBsize = 1 << log2_size;
1431
1432	xlog_get_iclog_buffer_size(mp, log);
1433
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1434	spin_lock_init(&log->l_icloglock);
1435	init_waitqueue_head(&log->l_flush_wait);
1436
1437	iclogp = &log->l_iclog;
1438	/*
1439	 * The amount of memory to allocate for the iclog structure is
1440	 * rather funky due to the way the structure is defined.  It is
1441	 * done this way so that we can use different sizes for machines
1442	 * with different amounts of memory.  See the definition of
1443	 * xlog_in_core_t in xfs_log_priv.h for details.
1444	 */
1445	ASSERT(log->l_iclog_size >= 4096);
1446	for (i = 0; i < log->l_iclog_bufs; i++) {
1447		size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1448				sizeof(struct bio_vec);
1449
1450		iclog = kzalloc(sizeof(*iclog) + bvec_size,
1451				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1452		if (!iclog)
1453			goto out_free_iclog;
1454
1455		*iclogp = iclog;
1456		iclog->ic_prev = prev_iclog;
1457		prev_iclog = iclog;
1458
1459		iclog->ic_data = kvzalloc(log->l_iclog_size,
1460				GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1461		if (!iclog->ic_data)
 
1462			goto out_free_iclog;
 
 
 
 
 
 
 
 
 
 
 
 
1463		head = &iclog->ic_header;
1464		memset(head, 0, sizeof(xlog_rec_header_t));
1465		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1466		head->h_version = cpu_to_be32(
1467			xfs_has_logv2(log->l_mp) ? 2 : 1);
1468		head->h_size = cpu_to_be32(log->l_iclog_size);
1469		/* new fields */
1470		head->h_fmt = cpu_to_be32(XLOG_FMT);
1471		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1472
1473		iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1474		iclog->ic_state = XLOG_STATE_ACTIVE;
1475		iclog->ic_log = log;
1476		atomic_set(&iclog->ic_refcnt, 0);
1477		INIT_LIST_HEAD(&iclog->ic_callbacks);
1478		iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
 
1479
1480		init_waitqueue_head(&iclog->ic_force_wait);
1481		init_waitqueue_head(&iclog->ic_write_wait);
1482		INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1483		sema_init(&iclog->ic_sema, 1);
1484
1485		iclogp = &iclog->ic_next;
1486	}
1487	*iclogp = log->l_iclog;			/* complete ring */
1488	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1489
1490	log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1491			XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1492				    WQ_HIGHPRI),
1493			0, mp->m_super->s_id);
1494	if (!log->l_ioend_workqueue)
1495		goto out_free_iclog;
1496
1497	error = xlog_cil_init(log);
1498	if (error)
1499		goto out_destroy_workqueue;
1500	return log;
1501
1502out_destroy_workqueue:
1503	destroy_workqueue(log->l_ioend_workqueue);
1504out_free_iclog:
1505	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1506		prev_iclog = iclog->ic_next;
1507		kvfree(iclog->ic_data);
1508		kfree(iclog);
1509		if (prev_iclog == log->l_iclog)
1510			break;
1511	}
 
 
1512out_free_log:
1513	kfree(log);
1514out:
1515	return ERR_PTR(error);
1516}	/* xlog_alloc_log */
1517
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1518/*
1519 * Stamp cycle number in every block
1520 */
1521STATIC void
1522xlog_pack_data(
1523	struct xlog		*log,
1524	struct xlog_in_core	*iclog,
1525	int			roundoff)
1526{
1527	int			i, j, k;
1528	int			size = iclog->ic_offset + roundoff;
1529	__be32			cycle_lsn;
1530	char			*dp;
1531
1532	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1533
1534	dp = iclog->ic_datap;
1535	for (i = 0; i < BTOBB(size); i++) {
1536		if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1537			break;
1538		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1539		*(__be32 *)dp = cycle_lsn;
1540		dp += BBSIZE;
1541	}
1542
1543	if (xfs_has_logv2(log->l_mp)) {
1544		xlog_in_core_2_t *xhdr = iclog->ic_data;
1545
1546		for ( ; i < BTOBB(size); i++) {
1547			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1548			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1549			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1550			*(__be32 *)dp = cycle_lsn;
1551			dp += BBSIZE;
1552		}
1553
1554		for (i = 1; i < log->l_iclog_heads; i++)
1555			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1556	}
1557}
1558
1559/*
1560 * Calculate the checksum for a log buffer.
1561 *
1562 * This is a little more complicated than it should be because the various
1563 * headers and the actual data are non-contiguous.
1564 */
1565__le32
1566xlog_cksum(
1567	struct xlog		*log,
1568	struct xlog_rec_header	*rhead,
1569	char			*dp,
1570	int			size)
1571{
1572	uint32_t		crc;
1573
1574	/* first generate the crc for the record header ... */
1575	crc = xfs_start_cksum_update((char *)rhead,
1576			      sizeof(struct xlog_rec_header),
1577			      offsetof(struct xlog_rec_header, h_crc));
1578
1579	/* ... then for additional cycle data for v2 logs ... */
1580	if (xfs_has_logv2(log->l_mp)) {
1581		union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1582		int		i;
1583		int		xheads;
1584
1585		xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
 
 
1586
1587		for (i = 1; i < xheads; i++) {
1588			crc = crc32c(crc, &xhdr[i].hic_xheader,
1589				     sizeof(struct xlog_rec_ext_header));
1590		}
1591	}
1592
1593	/* ... and finally for the payload */
1594	crc = crc32c(crc, dp, size);
1595
1596	return xfs_end_cksum(crc);
1597}
1598
1599static void
1600xlog_bio_end_io(
1601	struct bio		*bio)
1602{
1603	struct xlog_in_core	*iclog = bio->bi_private;
1604
1605	queue_work(iclog->ic_log->l_ioend_workqueue,
1606		   &iclog->ic_end_io_work);
1607}
1608
1609static int
1610xlog_map_iclog_data(
1611	struct bio		*bio,
1612	void			*data,
1613	size_t			count)
1614{
1615	do {
1616		struct page	*page = kmem_to_page(data);
1617		unsigned int	off = offset_in_page(data);
1618		size_t		len = min_t(size_t, count, PAGE_SIZE - off);
1619
1620		if (bio_add_page(bio, page, len, off) != len)
1621			return -EIO;
1622
1623		data += len;
1624		count -= len;
1625	} while (count);
1626
1627	return 0;
1628}
1629
1630STATIC void
1631xlog_write_iclog(
1632	struct xlog		*log,
1633	struct xlog_in_core	*iclog,
1634	uint64_t		bno,
1635	unsigned int		count)
1636{
1637	ASSERT(bno < log->l_logBBsize);
1638	trace_xlog_iclog_write(iclog, _RET_IP_);
1639
1640	/*
1641	 * We lock the iclogbufs here so that we can serialise against I/O
1642	 * completion during unmount.  We might be processing a shutdown
1643	 * triggered during unmount, and that can occur asynchronously to the
1644	 * unmount thread, and hence we need to ensure that completes before
1645	 * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1646	 * across the log IO to archieve that.
1647	 */
1648	down(&iclog->ic_sema);
1649	if (xlog_is_shutdown(log)) {
1650		/*
1651		 * It would seem logical to return EIO here, but we rely on
1652		 * the log state machine to propagate I/O errors instead of
1653		 * doing it here.  We kick of the state machine and unlock
1654		 * the buffer manually, the code needs to be kept in sync
1655		 * with the I/O completion path.
1656		 */
1657		goto sync;
1658	}
1659
1660	/*
1661	 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1662	 * IOs coming immediately after this one. This prevents the block layer
1663	 * writeback throttle from throttling log writes behind background
1664	 * metadata writeback and causing priority inversions.
1665	 */
1666	bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1667		 howmany(count, PAGE_SIZE),
1668		 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1669	iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1670	iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1671	iclog->ic_bio.bi_private = iclog;
1672
1673	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1674		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1675		/*
1676		 * For external log devices, we also need to flush the data
1677		 * device cache first to ensure all metadata writeback covered
1678		 * by the LSN in this iclog is on stable storage. This is slow,
1679		 * but it *must* complete before we issue the external log IO.
1680		 *
1681		 * If the flush fails, we cannot conclude that past metadata
1682		 * writeback from the log succeeded.  Repeating the flush is
1683		 * not possible, hence we must shut down with log IO error to
1684		 * avoid shutdown re-entering this path and erroring out again.
1685		 */
1686		if (log->l_targ != log->l_mp->m_ddev_targp &&
1687		    blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev))
1688			goto shutdown;
1689	}
1690	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1691		iclog->ic_bio.bi_opf |= REQ_FUA;
1692
1693	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1694
1695	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count))
1696		goto shutdown;
1697
1698	if (is_vmalloc_addr(iclog->ic_data))
1699		flush_kernel_vmap_range(iclog->ic_data, count);
1700
1701	/*
1702	 * If this log buffer would straddle the end of the log we will have
1703	 * to split it up into two bios, so that we can continue at the start.
1704	 */
1705	if (bno + BTOBB(count) > log->l_logBBsize) {
1706		struct bio *split;
1707
1708		split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1709				  GFP_NOIO, &fs_bio_set);
1710		bio_chain(split, &iclog->ic_bio);
1711		submit_bio(split);
1712
1713		/* restart at logical offset zero for the remainder */
1714		iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1715	}
1716
1717	submit_bio(&iclog->ic_bio);
1718	return;
1719shutdown:
1720	xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1721sync:
1722	xlog_state_done_syncing(iclog);
1723	up(&iclog->ic_sema);
1724}
1725
1726/*
1727 * We need to bump cycle number for the part of the iclog that is
1728 * written to the start of the log. Watch out for the header magic
1729 * number case, though.
1730 */
1731static void
1732xlog_split_iclog(
1733	struct xlog		*log,
1734	void			*data,
1735	uint64_t		bno,
1736	unsigned int		count)
1737{
1738	unsigned int		split_offset = BBTOB(log->l_logBBsize - bno);
1739	unsigned int		i;
1740
1741	for (i = split_offset; i < count; i += BBSIZE) {
1742		uint32_t cycle = get_unaligned_be32(data + i);
1743
1744		if (++cycle == XLOG_HEADER_MAGIC_NUM)
1745			cycle++;
1746		put_unaligned_be32(cycle, data + i);
1747	}
1748}
1749
1750static int
1751xlog_calc_iclog_size(
1752	struct xlog		*log,
1753	struct xlog_in_core	*iclog,
1754	uint32_t		*roundoff)
1755{
1756	uint32_t		count_init, count;
1757
1758	/* Add for LR header */
1759	count_init = log->l_iclog_hsize + iclog->ic_offset;
1760	count = roundup(count_init, log->l_iclog_roundoff);
1761
1762	*roundoff = count - count_init;
1763
1764	ASSERT(count >= count_init);
1765	ASSERT(*roundoff < log->l_iclog_roundoff);
1766	return count;
1767}
1768
1769/*
1770 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1771 * fashion.  Previously, we should have moved the current iclog
1772 * ptr in the log to point to the next available iclog.  This allows further
1773 * write to continue while this code syncs out an iclog ready to go.
1774 * Before an in-core log can be written out, the data section must be scanned
1775 * to save away the 1st word of each BBSIZE block into the header.  We replace
1776 * it with the current cycle count.  Each BBSIZE block is tagged with the
1777 * cycle count because there in an implicit assumption that drives will
1778 * guarantee that entire 512 byte blocks get written at once.  In other words,
1779 * we can't have part of a 512 byte block written and part not written.  By
1780 * tagging each block, we will know which blocks are valid when recovering
1781 * after an unclean shutdown.
1782 *
1783 * This routine is single threaded on the iclog.  No other thread can be in
1784 * this routine with the same iclog.  Changing contents of iclog can there-
1785 * fore be done without grabbing the state machine lock.  Updating the global
1786 * log will require grabbing the lock though.
1787 *
1788 * The entire log manager uses a logical block numbering scheme.  Only
1789 * xlog_write_iclog knows about the fact that the log may not start with
1790 * block zero on a given device.
 
1791 */
1792STATIC void
 
1793xlog_sync(
1794	struct xlog		*log,
1795	struct xlog_in_core	*iclog,
1796	struct xlog_ticket	*ticket)
1797{
1798	unsigned int		count;		/* byte count of bwrite */
1799	unsigned int		roundoff;       /* roundoff to BB or stripe */
1800	uint64_t		bno;
1801	unsigned int		size;
 
 
 
 
 
1802
 
1803	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1804	trace_xlog_iclog_sync(iclog, _RET_IP_);
1805
1806	count = xlog_calc_iclog_size(log, iclog, &roundoff);
 
1807
1808	/*
1809	 * If we have a ticket, account for the roundoff via the ticket
1810	 * reservation to avoid touching the hot grant heads needlessly.
1811	 * Otherwise, we have to move grant heads directly.
1812	 */
1813	if (ticket) {
1814		ticket->t_curr_res -= roundoff;
1815	} else {
1816		xlog_grant_add_space(&log->l_reserve_head, roundoff);
1817		xlog_grant_add_space(&log->l_write_head, roundoff);
1818	}
 
 
 
 
 
 
 
 
 
 
 
1819
1820	/* put cycle number in every block */
1821	xlog_pack_data(log, iclog, roundoff);
1822
1823	/* real byte length */
1824	size = iclog->ic_offset;
1825	if (xfs_has_logv2(log->l_mp))
1826		size += roundoff;
1827	iclog->ic_header.h_len = cpu_to_be32(size);
1828
1829	XFS_STATS_INC(log->l_mp, xs_log_writes);
 
 
1830	XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1831
1832	bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
 
 
 
 
 
 
1833
1834	/* Do we need to split this write into 2 parts? */
1835	if (bno + BTOBB(count) > log->l_logBBsize)
1836		xlog_split_iclog(log, &iclog->ic_header, bno, count);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1837
1838	/* calculcate the checksum */
1839	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1840					    iclog->ic_datap, size);
1841	/*
1842	 * Intentionally corrupt the log record CRC based on the error injection
1843	 * frequency, if defined. This facilitates testing log recovery in the
1844	 * event of torn writes. Hence, set the IOABORT state to abort the log
1845	 * write on I/O completion and shutdown the fs. The subsequent mount
1846	 * detects the bad CRC and attempts to recover.
1847	 */
1848#ifdef DEBUG
1849	if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1850		iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1851		iclog->ic_fail_crc = true;
1852		xfs_warn(log->l_mp,
1853	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1854			 be64_to_cpu(iclog->ic_header.h_lsn));
1855	}
1856#endif
1857	xlog_verify_iclog(log, iclog, count);
1858	xlog_write_iclog(log, iclog, bno, count);
1859}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1860
1861/*
1862 * Deallocate a log structure
1863 */
1864STATIC void
1865xlog_dealloc_log(
1866	struct xlog	*log)
1867{
1868	xlog_in_core_t	*iclog, *next_iclog;
1869	int		i;
1870
 
 
1871	/*
1872	 * Destroy the CIL after waiting for iclog IO completion because an
1873	 * iclog EIO error will try to shut down the log, which accesses the
1874	 * CIL to wake up the waiters.
1875	 */
1876	xlog_cil_destroy(log);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1877
1878	iclog = log->l_iclog;
1879	for (i = 0; i < log->l_iclog_bufs; i++) {
 
1880		next_iclog = iclog->ic_next;
1881		kvfree(iclog->ic_data);
1882		kfree(iclog);
1883		iclog = next_iclog;
1884	}
 
1885
1886	log->l_mp->m_log = NULL;
1887	destroy_workqueue(log->l_ioend_workqueue);
1888	kfree(log);
1889}
1890
1891/*
1892 * Update counters atomically now that memcpy is done.
1893 */
 
1894static inline void
1895xlog_state_finish_copy(
1896	struct xlog		*log,
1897	struct xlog_in_core	*iclog,
1898	int			record_cnt,
1899	int			copy_bytes)
1900{
1901	lockdep_assert_held(&log->l_icloglock);
1902
1903	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1904	iclog->ic_offset += copy_bytes;
1905}
 
 
 
 
 
1906
1907/*
1908 * print out info relating to regions written which consume
1909 * the reservation
1910 */
1911void
1912xlog_print_tic_res(
1913	struct xfs_mount	*mp,
1914	struct xlog_ticket	*ticket)
1915{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1916	xfs_warn(mp, "ticket reservation summary:");
1917	xfs_warn(mp, "  unit res    = %d bytes", ticket->t_unit_res);
1918	xfs_warn(mp, "  current res = %d bytes", ticket->t_curr_res);
1919	xfs_warn(mp, "  original count  = %d", ticket->t_ocnt);
1920	xfs_warn(mp, "  remaining count = %d", ticket->t_cnt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1921}
1922
1923/*
1924 * Print a summary of the transaction.
1925 */
1926void
1927xlog_print_trans(
1928	struct xfs_trans	*tp)
1929{
1930	struct xfs_mount	*mp = tp->t_mountp;
1931	struct xfs_log_item	*lip;
1932
1933	/* dump core transaction and ticket info */
1934	xfs_warn(mp, "transaction summary:");
1935	xfs_warn(mp, "  log res   = %d", tp->t_log_res);
1936	xfs_warn(mp, "  log count = %d", tp->t_log_count);
1937	xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
1938
1939	xlog_print_tic_res(mp, tp->t_ticket);
1940
1941	/* dump each log item */
1942	list_for_each_entry(lip, &tp->t_items, li_trans) {
 
1943		struct xfs_log_vec	*lv = lip->li_lv;
1944		struct xfs_log_iovec	*vec;
1945		int			i;
1946
1947		xfs_warn(mp, "log item: ");
1948		xfs_warn(mp, "  type	= 0x%x", lip->li_type);
1949		xfs_warn(mp, "  flags	= 0x%lx", lip->li_flags);
1950		if (!lv)
1951			continue;
1952		xfs_warn(mp, "  niovecs	= %d", lv->lv_niovecs);
1953		xfs_warn(mp, "  size	= %d", lv->lv_size);
1954		xfs_warn(mp, "  bytes	= %d", lv->lv_bytes);
1955		xfs_warn(mp, "  buf len	= %d", lv->lv_buf_len);
1956
1957		/* dump each iovec for the log item */
1958		vec = lv->lv_iovecp;
1959		for (i = 0; i < lv->lv_niovecs; i++) {
1960			int dumplen = min(vec->i_len, 32);
1961
1962			xfs_warn(mp, "  iovec[%d]", i);
1963			xfs_warn(mp, "    type	= 0x%x", vec->i_type);
1964			xfs_warn(mp, "    len	= %d", vec->i_len);
1965			xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
1966			xfs_hex_dump(vec->i_addr, dumplen);
1967
1968			vec++;
1969		}
1970	}
1971}
1972
1973static inline void
1974xlog_write_iovec(
1975	struct xlog_in_core	*iclog,
1976	uint32_t		*log_offset,
1977	void			*data,
1978	uint32_t		write_len,
1979	int			*bytes_left,
1980	uint32_t		*record_cnt,
1981	uint32_t		*data_cnt)
1982{
1983	ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
1984	ASSERT(*log_offset % sizeof(int32_t) == 0);
1985	ASSERT(write_len % sizeof(int32_t) == 0);
1986
1987	memcpy(iclog->ic_datap + *log_offset, data, write_len);
1988	*log_offset += write_len;
1989	*bytes_left -= write_len;
1990	(*record_cnt)++;
1991	*data_cnt += write_len;
1992}
1993
1994/*
1995 * Write log vectors into a single iclog which is guaranteed by the caller
1996 * to have enough space to write the entire log vector into.
1997 */
1998static void
1999xlog_write_full(
2000	struct xfs_log_vec	*lv,
2001	struct xlog_ticket	*ticket,
2002	struct xlog_in_core	*iclog,
2003	uint32_t		*log_offset,
2004	uint32_t		*len,
2005	uint32_t		*record_cnt,
2006	uint32_t		*data_cnt)
2007{
2008	int			index;
 
 
 
 
 
 
 
 
 
 
 
 
2009
2010	ASSERT(*log_offset + *len <= iclog->ic_size ||
2011		iclog->ic_state == XLOG_STATE_WANT_SYNC);
2012
2013	/*
2014	 * Ordered log vectors have no regions to write so this
2015	 * loop will naturally skip them.
2016	 */
2017	for (index = 0; index < lv->lv_niovecs; index++) {
2018		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2019		struct xlog_op_header	*ophdr = reg->i_addr;
2020
2021		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2022		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2023				reg->i_len, len, record_cnt, data_cnt);
2024	}
 
 
 
 
 
2025}
2026
 
 
 
 
2027static int
2028xlog_write_get_more_iclog_space(
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2029	struct xlog_ticket	*ticket,
2030	struct xlog_in_core	**iclogp,
2031	uint32_t		*log_offset,
2032	uint32_t		len,
2033	uint32_t		*record_cnt,
2034	uint32_t		*data_cnt)
2035{
2036	struct xlog_in_core	*iclog = *iclogp;
2037	struct xlog		*log = iclog->ic_log;
2038	int			error;
2039
2040	spin_lock(&log->l_icloglock);
2041	ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2042	xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2043	error = xlog_state_release_iclog(log, iclog, ticket);
2044	spin_unlock(&log->l_icloglock);
2045	if (error)
2046		return error;
 
 
 
 
 
 
 
 
 
 
 
 
2047
2048	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2049					log_offset);
2050	if (error)
2051		return error;
2052	*record_cnt = 0;
2053	*data_cnt = 0;
2054	*iclogp = iclog;
2055	return 0;
2056}
2057
2058/*
2059 * Write log vectors into a single iclog which is smaller than the current chain
2060 * length. We write until we cannot fit a full record into the remaining space
2061 * and then stop. We return the log vector that is to be written that cannot
2062 * wholly fit in the iclog.
2063 */
2064static int
2065xlog_write_partial(
2066	struct xfs_log_vec	*lv,
2067	struct xlog_ticket	*ticket,
2068	struct xlog_in_core	**iclogp,
2069	uint32_t		*log_offset,
2070	uint32_t		*len,
2071	uint32_t		*record_cnt,
2072	uint32_t		*data_cnt)
2073{
2074	struct xlog_in_core	*iclog = *iclogp;
2075	struct xlog_op_header	*ophdr;
2076	int			index = 0;
2077	uint32_t		rlen;
2078	int			error;
 
 
 
 
 
 
 
 
 
 
 
 
2079
2080	/* walk the logvec, copying until we run out of space in the iclog */
2081	for (index = 0; index < lv->lv_niovecs; index++) {
2082		struct xfs_log_iovec	*reg = &lv->lv_iovecp[index];
2083		uint32_t		reg_offset = 0;
 
 
 
 
 
 
 
 
2084
2085		/*
2086		 * The first region of a continuation must have a non-zero
2087		 * length otherwise log recovery will just skip over it and
2088		 * start recovering from the next opheader it finds. Because we
2089		 * mark the next opheader as a continuation, recovery will then
2090		 * incorrectly add the continuation to the previous region and
2091		 * that breaks stuff.
2092		 *
2093		 * Hence if there isn't space for region data after the
2094		 * opheader, then we need to start afresh with a new iclog.
2095		 */
2096		if (iclog->ic_size - *log_offset <=
2097					sizeof(struct xlog_op_header)) {
2098			error = xlog_write_get_more_iclog_space(ticket,
2099					&iclog, log_offset, *len, record_cnt,
2100					data_cnt);
2101			if (error)
2102				return error;
2103		}
2104
2105		ophdr = reg->i_addr;
2106		rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2107
2108		ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2109		ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2110		if (rlen != reg->i_len)
2111			ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2112
2113		xlog_write_iovec(iclog, log_offset, reg->i_addr,
2114				rlen, len, record_cnt, data_cnt);
2115
2116		/* If we wrote the whole region, move to the next. */
2117		if (rlen == reg->i_len)
2118			continue;
2119
 
 
 
 
 
 
 
 
 
 
 
 
 
2120		/*
2121		 * We now have a partially written iovec, but it can span
2122		 * multiple iclogs so we loop here. First we release the iclog
2123		 * we currently have, then we get a new iclog and add a new
2124		 * opheader. Then we continue copying from where we were until
2125		 * we either complete the iovec or fill the iclog. If we
2126		 * complete the iovec, then we increment the index and go right
2127		 * back to the top of the outer loop. if we fill the iclog, we
2128		 * run the inner loop again.
2129		 *
2130		 * This is complicated by the tail of a region using all the
2131		 * space in an iclog and hence requiring us to release the iclog
2132		 * and get a new one before returning to the outer loop. We must
2133		 * always guarantee that we exit this inner loop with at least
2134		 * space for log transaction opheaders left in the current
2135		 * iclog, hence we cannot just terminate the loop at the end
2136		 * of the of the continuation. So we loop while there is no
2137		 * space left in the current iclog, and check for the end of the
2138		 * continuation after getting a new iclog.
2139		 */
2140		do {
2141			/*
2142			 * Ensure we include the continuation opheader in the
2143			 * space we need in the new iclog by adding that size
2144			 * to the length we require. This continuation opheader
2145			 * needs to be accounted to the ticket as the space it
2146			 * consumes hasn't been accounted to the lv we are
2147			 * writing.
2148			 */
2149			error = xlog_write_get_more_iclog_space(ticket,
2150					&iclog, log_offset,
2151					*len + sizeof(struct xlog_op_header),
2152					record_cnt, data_cnt);
2153			if (error)
2154				return error;
2155
2156			ophdr = iclog->ic_datap + *log_offset;
2157			ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2158			ophdr->oh_clientid = XFS_TRANSACTION;
2159			ophdr->oh_res2 = 0;
2160			ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2161
2162			ticket->t_curr_res -= sizeof(struct xlog_op_header);
2163			*log_offset += sizeof(struct xlog_op_header);
2164			*data_cnt += sizeof(struct xlog_op_header);
2165
2166			/*
2167			 * If rlen fits in the iclog, then end the region
2168			 * continuation. Otherwise we're going around again.
2169			 */
2170			reg_offset += rlen;
2171			rlen = reg->i_len - reg_offset;
2172			if (rlen <= iclog->ic_size - *log_offset)
2173				ophdr->oh_flags |= XLOG_END_TRANS;
2174			else
2175				ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2176
2177			rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2178			ophdr->oh_len = cpu_to_be32(rlen);
2179
2180			xlog_write_iovec(iclog, log_offset,
2181					reg->i_addr + reg_offset,
2182					rlen, len, record_cnt, data_cnt);
2183
2184		} while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2185	}
2186
2187	/*
2188	 * No more iovecs remain in this logvec so return the next log vec to
2189	 * the caller so it can go back to fast path copying.
2190	 */
2191	*iclogp = iclog;
2192	return 0;
2193}
2194
2195/*
2196 * Write some region out to in-core log
2197 *
2198 * This will be called when writing externally provided regions or when
2199 * writing out a commit record for a given transaction.
2200 *
2201 * General algorithm:
2202 *	1. Find total length of this write.  This may include adding to the
2203 *		lengths passed in.
2204 *	2. Check whether we violate the tickets reservation.
2205 *	3. While writing to this iclog
2206 *	    A. Reserve as much space in this iclog as can get
2207 *	    B. If this is first write, save away start lsn
2208 *	    C. While writing this region:
2209 *		1. If first write of transaction, write start record
2210 *		2. Write log operation header (header per region)
2211 *		3. Find out if we can fit entire region into this iclog
2212 *		4. Potentially, verify destination memcpy ptr
2213 *		5. Memcpy (partial) region
2214 *		6. If partial copy, release iclog; otherwise, continue
2215 *			copying more regions into current iclog
2216 *	4. Mark want sync bit (in simulation mode)
2217 *	5. Release iclog for potential flush to on-disk log.
2218 *
2219 * ERRORS:
2220 * 1.	Panic if reservation is overrun.  This should never happen since
2221 *	reservation amounts are generated internal to the filesystem.
2222 * NOTES:
2223 * 1. Tickets are single threaded data structures.
2224 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2225 *	syncing routine.  When a single log_write region needs to span
2226 *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2227 *	on all log operation writes which don't contain the end of the
2228 *	region.  The XLOG_END_TRANS bit is used for the in-core log
2229 *	operation which contains the end of the continued log_write region.
2230 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2231 *	we don't really know exactly how much space will be used.  As a result,
2232 *	we don't update ic_offset until the end when we know exactly how many
2233 *	bytes have been written out.
2234 */
2235int
2236xlog_write(
2237	struct xlog		*log,
2238	struct xfs_cil_ctx	*ctx,
2239	struct list_head	*lv_chain,
2240	struct xlog_ticket	*ticket,
2241	uint32_t		len)
2242
 
2243{
2244	struct xlog_in_core	*iclog = NULL;
 
2245	struct xfs_log_vec	*lv;
2246	uint32_t		record_cnt = 0;
2247	uint32_t		data_cnt = 0;
2248	int			error = 0;
2249	int			log_offset;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2250
2251	if (ticket->t_curr_res < 0) {
2252		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2253		     "ctx ticket reservation ran out. Need to up reservation");
2254		xlog_print_tic_res(log->l_mp, ticket);
2255		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2256	}
2257
2258	error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2259					   &log_offset);
2260	if (error)
2261		return error;
 
 
 
 
 
 
 
2262
2263	ASSERT(log_offset <= iclog->ic_size - 1);
 
2264
2265	/*
2266	 * If we have a context pointer, pass it the first iclog we are
2267	 * writing to so it can record state needed for iclog write
2268	 * ordering.
2269	 */
2270	if (ctx)
2271		xlog_cil_set_ctx_write_state(ctx, iclog);
2272
2273	list_for_each_entry(lv, lv_chain, lv_list) {
2274		/*
2275		 * If the entire log vec does not fit in the iclog, punt it to
2276		 * the partial copy loop which can handle this case.
2277		 */
2278		if (lv->lv_niovecs &&
2279		    lv->lv_bytes > iclog->ic_size - log_offset) {
2280			error = xlog_write_partial(lv, ticket, &iclog,
2281					&log_offset, &len, &record_cnt,
2282					&data_cnt);
2283			if (error) {
2284				/*
2285				 * We have no iclog to release, so just return
2286				 * the error immediately.
2287				 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2288				return error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2289			}
2290		} else {
2291			xlog_write_full(lv, ticket, iclog, &log_offset,
2292					 &len, &record_cnt, &data_cnt);
2293		}
2294	}
 
2295	ASSERT(len == 0);
2296
2297	/*
2298	 * We've already been guaranteed that the last writes will fit inside
2299	 * the current iclog, and hence it will already have the space used by
2300	 * those writes accounted to it. Hence we do not need to update the
2301	 * iclog with the number of bytes written here.
2302	 */
2303	spin_lock(&log->l_icloglock);
2304	xlog_state_finish_copy(log, iclog, record_cnt, 0);
2305	error = xlog_state_release_iclog(log, iclog, ticket);
2306	spin_unlock(&log->l_icloglock);
2307
2308	return error;
 
 
2309}
2310
2311static void
2312xlog_state_activate_iclog(
2313	struct xlog_in_core	*iclog,
2314	int			*iclogs_changed)
2315{
2316	ASSERT(list_empty_careful(&iclog->ic_callbacks));
2317	trace_xlog_iclog_activate(iclog, _RET_IP_);
2318
2319	/*
2320	 * If the number of ops in this iclog indicate it just contains the
2321	 * dummy transaction, we can change state into IDLE (the second time
2322	 * around). Otherwise we should change the state into NEED a dummy.
2323	 * We don't need to cover the dummy.
2324	 */
2325	if (*iclogs_changed == 0 &&
2326	    iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2327		*iclogs_changed = 1;
2328	} else {
2329		/*
2330		 * We have two dirty iclogs so start over.  This could also be
2331		 * num of ops indicating this is not the dummy going out.
2332		 */
2333		*iclogs_changed = 2;
2334	}
2335
2336	iclog->ic_state	= XLOG_STATE_ACTIVE;
2337	iclog->ic_offset = 0;
2338	iclog->ic_header.h_num_logops = 0;
2339	memset(iclog->ic_header.h_cycle_data, 0,
2340		sizeof(iclog->ic_header.h_cycle_data));
2341	iclog->ic_header.h_lsn = 0;
2342	iclog->ic_header.h_tail_lsn = 0;
2343}
2344
2345/*
2346 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2347 * ACTIVE after iclog I/O has completed.
 
 
 
 
2348 */
2349static void
2350xlog_state_activate_iclogs(
2351	struct xlog		*log,
2352	int			*iclogs_changed)
2353{
2354	struct xlog_in_core	*iclog = log->l_iclog;
 
2355
 
2356	do {
2357		if (iclog->ic_state == XLOG_STATE_DIRTY)
2358			xlog_state_activate_iclog(iclog, iclogs_changed);
2359		/*
2360		 * The ordering of marking iclogs ACTIVE must be maintained, so
2361		 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2362		 */
2363		else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2364			break;
2365	} while ((iclog = iclog->ic_next) != log->l_iclog);
2366}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2367
2368static int
2369xlog_covered_state(
2370	int			prev_state,
2371	int			iclogs_changed)
2372{
2373	/*
2374	 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2375	 * wrote the first covering record (DONE). We go to IDLE if we just
2376	 * wrote the second covering record (DONE2) and remain in IDLE until a
2377	 * non-covering write occurs.
2378	 */
2379	switch (prev_state) {
2380	case XLOG_STATE_COVER_IDLE:
2381		if (iclogs_changed == 1)
2382			return XLOG_STATE_COVER_IDLE;
2383		fallthrough;
2384	case XLOG_STATE_COVER_NEED:
2385	case XLOG_STATE_COVER_NEED2:
2386		break;
2387	case XLOG_STATE_COVER_DONE:
2388		if (iclogs_changed == 1)
2389			return XLOG_STATE_COVER_NEED2;
2390		break;
2391	case XLOG_STATE_COVER_DONE2:
2392		if (iclogs_changed == 1)
2393			return XLOG_STATE_COVER_IDLE;
2394		break;
2395	default:
2396		ASSERT(0);
2397	}
2398
2399	return XLOG_STATE_COVER_NEED;
2400}
 
 
 
 
2401
2402STATIC void
2403xlog_state_clean_iclog(
2404	struct xlog		*log,
2405	struct xlog_in_core	*dirty_iclog)
2406{
2407	int			iclogs_changed = 0;
2408
2409	trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2410
2411	dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2412
2413	xlog_state_activate_iclogs(log, &iclogs_changed);
2414	wake_up_all(&dirty_iclog->ic_force_wait);
2415
2416	if (iclogs_changed) {
2417		log->l_covered_state = xlog_covered_state(log->l_covered_state,
2418				iclogs_changed);
2419	}
2420}
2421
2422STATIC xfs_lsn_t
2423xlog_get_lowest_lsn(
2424	struct xlog		*log)
2425{
2426	struct xlog_in_core	*iclog = log->l_iclog;
2427	xfs_lsn_t		lowest_lsn = 0, lsn;
2428
 
 
2429	do {
2430		if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2431		    iclog->ic_state == XLOG_STATE_DIRTY)
2432			continue;
2433
2434		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2435		if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2436			lowest_lsn = lsn;
2437	} while ((iclog = iclog->ic_next) != log->l_iclog);
2438
 
 
2439	return lowest_lsn;
2440}
2441
2442/*
2443 * Return true if we need to stop processing, false to continue to the next
2444 * iclog. The caller will need to run callbacks if the iclog is returned in the
2445 * XLOG_STATE_CALLBACK state.
2446 */
2447static bool
2448xlog_state_iodone_process_iclog(
2449	struct xlog		*log,
2450	struct xlog_in_core	*iclog)
 
2451{
2452	xfs_lsn_t		lowest_lsn;
2453	xfs_lsn_t		header_lsn;
 
 
 
 
 
 
 
 
 
 
2454
2455	switch (iclog->ic_state) {
2456	case XLOG_STATE_ACTIVE:
2457	case XLOG_STATE_DIRTY:
 
 
 
 
2458		/*
2459		 * Skip all iclogs in the ACTIVE & DIRTY states:
 
 
 
 
 
2460		 */
2461		return false;
2462	case XLOG_STATE_DONE_SYNC:
2463		/*
2464		 * Now that we have an iclog that is in the DONE_SYNC state, do
2465		 * one more check here to see if we have chased our tail around.
2466		 * If this is not the lowest lsn iclog, then we will leave it
2467		 * for another completion to process.
2468		 */
2469		header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2470		lowest_lsn = xlog_get_lowest_lsn(log);
2471		if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2472			return false;
2473		/*
2474		 * If there are no callbacks on this iclog, we can mark it clean
2475		 * immediately and return. Otherwise we need to run the
2476		 * callbacks.
2477		 */
2478		if (list_empty(&iclog->ic_callbacks)) {
2479			xlog_state_clean_iclog(log, iclog);
2480			return false;
2481		}
2482		trace_xlog_iclog_callback(iclog, _RET_IP_);
2483		iclog->ic_state = XLOG_STATE_CALLBACK;
2484		return false;
2485	default:
2486		/*
2487		 * Can only perform callbacks in order.  Since this iclog is not
2488		 * in the DONE_SYNC state, we skip the rest and just try to
2489		 * clean up.
2490		 */
2491		return true;
2492	}
2493}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2494
2495/*
2496 * Loop over all the iclogs, running attached callbacks on them. Return true if
2497 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2498 * to handle transient shutdown state here at all because
2499 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2500 * cleanup of the callbacks.
2501 */
2502static bool
2503xlog_state_do_iclog_callbacks(
2504	struct xlog		*log)
2505		__releases(&log->l_icloglock)
2506		__acquires(&log->l_icloglock)
2507{
2508	struct xlog_in_core	*first_iclog = log->l_iclog;
2509	struct xlog_in_core	*iclog = first_iclog;
2510	bool			ran_callback = false;
 
 
 
 
 
 
2511
2512	do {
2513		LIST_HEAD(cb_list);
2514
2515		if (xlog_state_iodone_process_iclog(log, iclog))
2516			break;
2517		if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2518			iclog = iclog->ic_next;
2519			continue;
2520		}
2521		list_splice_init(&iclog->ic_callbacks, &cb_list);
2522		spin_unlock(&log->l_icloglock);
2523
2524		trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2525		xlog_cil_process_committed(&cb_list);
2526		trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2527		ran_callback = true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2528
2529		spin_lock(&log->l_icloglock);
2530		xlog_state_clean_iclog(log, iclog);
2531		iclog = iclog->ic_next;
2532	} while (iclog != first_iclog);
2533
2534	return ran_callback;
2535}
 
 
 
2536
 
 
 
 
 
2537
2538/*
2539 * Loop running iclog completion callbacks until there are no more iclogs in a
2540 * state that can run callbacks.
2541 */
2542STATIC void
2543xlog_state_do_callback(
2544	struct xlog		*log)
2545{
2546	int			flushcnt = 0;
2547	int			repeats = 0;
2548
2549	spin_lock(&log->l_icloglock);
2550	while (xlog_state_do_iclog_callbacks(log)) {
2551		if (xlog_is_shutdown(log))
2552			break;
2553
2554		if (++repeats > 5000) {
2555			flushcnt += repeats;
2556			repeats = 0;
2557			xfs_warn(log->l_mp,
2558				"%s: possible infinite loop (%d iterations)",
2559				__func__, flushcnt);
2560		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2561	}
 
2562
2563	if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
 
 
 
 
2564		wake_up_all(&log->l_flush_wait);
2565
2566	spin_unlock(&log->l_icloglock);
2567}
2568
2569
2570/*
2571 * Finish transitioning this iclog to the dirty state.
2572 *
 
 
 
 
 
 
 
2573 * Callbacks could take time, so they are done outside the scope of the
2574 * global state machine log lock.
2575 */
2576STATIC void
2577xlog_state_done_syncing(
2578	struct xlog_in_core	*iclog)
 
2579{
2580	struct xlog		*log = iclog->ic_log;
2581
2582	spin_lock(&log->l_icloglock);
 
 
 
2583	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2584	trace_xlog_iclog_sync_done(iclog, _RET_IP_);
 
2585
2586	/*
2587	 * If we got an error, either on the first buffer, or in the case of
2588	 * split log writes, on the second, we shut down the file system and
2589	 * no iclogs should ever be attempted to be written to disk again.
2590	 */
2591	if (!xlog_is_shutdown(log)) {
2592		ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
 
 
 
 
2593		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2594	}
2595
2596	/*
2597	 * Someone could be sleeping prior to writing out the next
2598	 * iclog buffer, we wake them all, one will get to do the
2599	 * I/O, the others get to wait for the result.
2600	 */
2601	wake_up_all(&iclog->ic_write_wait);
2602	spin_unlock(&log->l_icloglock);
2603	xlog_state_do_callback(log);
2604}
 
2605
2606/*
2607 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2608 * sleep.  We wait on the flush queue on the head iclog as that should be
2609 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2610 * we will wait here and all new writes will sleep until a sync completes.
2611 *
2612 * The in-core logs are used in a circular fashion. They are not used
2613 * out-of-order even when an iclog past the head is free.
2614 *
2615 * return:
2616 *	* log_offset where xlog_write() can start writing into the in-core
2617 *		log's data space.
2618 *	* in-core log pointer to which xlog_write() should write.
2619 *	* boolean indicating this is a continued write to an in-core log.
2620 *		If this is the last write, then the in-core log's offset field
2621 *		needs to be incremented, depending on the amount of data which
2622 *		is copied.
2623 */
2624STATIC int
2625xlog_state_get_iclog_space(
2626	struct xlog		*log,
2627	int			len,
2628	struct xlog_in_core	**iclogp,
2629	struct xlog_ticket	*ticket,
 
2630	int			*logoffsetp)
2631{
2632	int		  log_offset;
2633	xlog_rec_header_t *head;
2634	xlog_in_core_t	  *iclog;
 
2635
2636restart:
2637	spin_lock(&log->l_icloglock);
2638	if (xlog_is_shutdown(log)) {
2639		spin_unlock(&log->l_icloglock);
2640		return -EIO;
2641	}
2642
2643	iclog = log->l_iclog;
2644	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2645		XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2646
2647		/* Wait for log writes to have flushed */
2648		xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2649		goto restart;
2650	}
2651
2652	head = &iclog->ic_header;
2653
2654	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2655	log_offset = iclog->ic_offset;
2656
2657	trace_xlog_iclog_get_space(iclog, _RET_IP_);
2658
2659	/* On the 1st write to an iclog, figure out lsn.  This works
2660	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2661	 * committing to.  If the offset is set, that's how many blocks
2662	 * must be written.
2663	 */
2664	if (log_offset == 0) {
2665		ticket->t_curr_res -= log->l_iclog_hsize;
 
 
 
2666		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2667		head->h_lsn = cpu_to_be64(
2668			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2669		ASSERT(log->l_curr_block >= 0);
2670	}
2671
2672	/* If there is enough room to write everything, then do it.  Otherwise,
2673	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2674	 * bit is on, so this will get flushed out.  Don't update ic_offset
2675	 * until you know exactly how many bytes get copied.  Therefore, wait
2676	 * until later to update ic_offset.
2677	 *
2678	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2679	 * can fit into remaining data section.
2680	 */
2681	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2682		int		error = 0;
2683
2684		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2685
2686		/*
2687		 * If we are the only one writing to this iclog, sync it to
2688		 * disk.  We need to do an atomic compare and decrement here to
2689		 * avoid racing with concurrent atomic_dec_and_lock() calls in
2690		 * xlog_state_release_iclog() when there is more than one
2691		 * reference to the iclog.
2692		 */
2693		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2694			error = xlog_state_release_iclog(log, iclog, ticket);
2695		spin_unlock(&log->l_icloglock);
2696		if (error)
2697			return error;
 
 
 
 
2698		goto restart;
2699	}
2700
2701	/* Do we have enough room to write the full amount in the remainder
2702	 * of this iclog?  Or must we continue a write on the next iclog and
2703	 * mark this iclog as completely taken?  In the case where we switch
2704	 * iclogs (to mark it taken), this particular iclog will release/sync
2705	 * to disk in xlog_write().
2706	 */
2707	if (len <= iclog->ic_size - iclog->ic_offset)
 
2708		iclog->ic_offset += len;
2709	else
 
2710		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
 
2711	*iclogp = iclog;
2712
2713	ASSERT(iclog->ic_offset <= iclog->ic_size);
2714	spin_unlock(&log->l_icloglock);
2715
2716	*logoffsetp = log_offset;
2717	return 0;
2718}
2719
2720/*
2721 * The first cnt-1 times a ticket goes through here we don't need to move the
2722 * grant write head because the permanent reservation has reserved cnt times the
2723 * unit amount.  Release part of current permanent unit reservation and reset
2724 * current reservation to be one units worth.  Also move grant reservation head
2725 * forward.
2726 */
2727void
2728xfs_log_ticket_regrant(
2729	struct xlog		*log,
2730	struct xlog_ticket	*ticket)
2731{
2732	trace_xfs_log_ticket_regrant(log, ticket);
2733
2734	if (ticket->t_cnt > 0)
2735		ticket->t_cnt--;
2736
2737	xlog_grant_sub_space(&log->l_reserve_head, ticket->t_curr_res);
2738	xlog_grant_sub_space(&log->l_write_head, ticket->t_curr_res);
 
 
2739	ticket->t_curr_res = ticket->t_unit_res;
 
2740
2741	trace_xfs_log_ticket_regrant_sub(log, ticket);
2742
2743	/* just return if we still have some of the pre-reserved space */
2744	if (!ticket->t_cnt) {
2745		xlog_grant_add_space(&log->l_reserve_head, ticket->t_unit_res);
2746		trace_xfs_log_ticket_regrant_exit(log, ticket);
 
 
2747
2748		ticket->t_curr_res = ticket->t_unit_res;
2749	}
 
 
 
2750
2751	xfs_log_ticket_put(ticket);
2752}
2753
2754/*
2755 * Give back the space left from a reservation.
2756 *
2757 * All the information we need to make a correct determination of space left
2758 * is present.  For non-permanent reservations, things are quite easy.  The
2759 * count should have been decremented to zero.  We only need to deal with the
2760 * space remaining in the current reservation part of the ticket.  If the
2761 * ticket contains a permanent reservation, there may be left over space which
2762 * needs to be released.  A count of N means that N-1 refills of the current
2763 * reservation can be done before we need to ask for more space.  The first
2764 * one goes to fill up the first current reservation.  Once we run out of
2765 * space, the count will stay at zero and the only space remaining will be
2766 * in the current reservation field.
2767 */
2768void
2769xfs_log_ticket_ungrant(
2770	struct xlog		*log,
2771	struct xlog_ticket	*ticket)
2772{
2773	int			bytes;
2774
2775	trace_xfs_log_ticket_ungrant(log, ticket);
2776
2777	if (ticket->t_cnt > 0)
2778		ticket->t_cnt--;
2779
2780	trace_xfs_log_ticket_ungrant_sub(log, ticket);
 
2781
2782	/*
2783	 * If this is a permanent reservation ticket, we may be able to free
2784	 * up more space based on the remaining count.
2785	 */
2786	bytes = ticket->t_curr_res;
2787	if (ticket->t_cnt > 0) {
2788		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2789		bytes += ticket->t_unit_res*ticket->t_cnt;
2790	}
2791
2792	xlog_grant_sub_space(&log->l_reserve_head, bytes);
2793	xlog_grant_sub_space(&log->l_write_head, bytes);
2794
2795	trace_xfs_log_ticket_ungrant_exit(log, ticket);
2796
2797	xfs_log_space_wake(log->l_mp);
2798	xfs_log_ticket_put(ticket);
2799}
2800
2801/*
2802 * This routine will mark the current iclog in the ring as WANT_SYNC and move
2803 * the current iclog pointer to the next iclog in the ring.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2804 */
2805void
2806xlog_state_switch_iclogs(
2807	struct xlog		*log,
2808	struct xlog_in_core	*iclog,
2809	int			eventual_size)
2810{
2811	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2812	assert_spin_locked(&log->l_icloglock);
2813	trace_xlog_iclog_switch(iclog, _RET_IP_);
2814
2815	if (!eventual_size)
2816		eventual_size = iclog->ic_offset;
2817	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2818	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2819	log->l_prev_block = log->l_curr_block;
2820	log->l_prev_cycle = log->l_curr_cycle;
2821
2822	/* roll log?: ic_offset changed later */
2823	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2824
2825	/* Round up to next log-sunit */
2826	if (log->l_iclog_roundoff > BBSIZE) {
2827		uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
 
2828		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2829	}
2830
2831	if (log->l_curr_block >= log->l_logBBsize) {
2832		/*
2833		 * Rewind the current block before the cycle is bumped to make
2834		 * sure that the combined LSN never transiently moves forward
2835		 * when the log wraps to the next cycle. This is to support the
2836		 * unlocked sample of these fields from xlog_valid_lsn(). Most
2837		 * other cases should acquire l_icloglock.
2838		 */
2839		log->l_curr_block -= log->l_logBBsize;
2840		ASSERT(log->l_curr_block >= 0);
2841		smp_wmb();
2842		log->l_curr_cycle++;
2843		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2844			log->l_curr_cycle++;
2845	}
2846	ASSERT(iclog == log->l_iclog);
2847	log->l_iclog = iclog->ic_next;
2848}
2849
2850/*
2851 * Force the iclog to disk and check if the iclog has been completed before
2852 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
2853 * pmem) or fast async storage because we drop the icloglock to issue the IO.
2854 * If completion has already occurred, tell the caller so that it can avoid an
2855 * unnecessary wait on the iclog.
2856 */
2857static int
2858xlog_force_and_check_iclog(
2859	struct xlog_in_core	*iclog,
2860	bool			*completed)
2861{
2862	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2863	int			error;
2864
2865	*completed = false;
2866	error = xlog_force_iclog(iclog);
2867	if (error)
2868		return error;
2869
2870	/*
2871	 * If the iclog has already been completed and reused the header LSN
2872	 * will have been rewritten by completion
2873	 */
2874	if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
2875		*completed = true;
2876	return 0;
2877}
2878
2879/*
2880 * Write out all data in the in-core log as of this exact moment in time.
2881 *
2882 * Data may be written to the in-core log during this call.  However,
2883 * we don't guarantee this data will be written out.  A change from past
2884 * implementation means this routine will *not* write out zero length LRs.
2885 *
2886 * Basically, we try and perform an intelligent scan of the in-core logs.
2887 * If we determine there is no flushable data, we just return.  There is no
2888 * flushable data if:
2889 *
2890 *	1. the current iclog is active and has no data; the previous iclog
2891 *		is in the active or dirty state.
2892 *	2. the current iclog is drity, and the previous iclog is in the
2893 *		active or dirty state.
2894 *
2895 * We may sleep if:
2896 *
2897 *	1. the current iclog is not in the active nor dirty state.
2898 *	2. the current iclog dirty, and the previous iclog is not in the
2899 *		active nor dirty state.
2900 *	3. the current iclog is active, and there is another thread writing
2901 *		to this particular iclog.
2902 *	4. a) the current iclog is active and has no other writers
2903 *	   b) when we return from flushing out this iclog, it is still
2904 *		not in the active nor dirty state.
2905 */
2906int
2907xfs_log_force(
2908	struct xfs_mount	*mp,
2909	uint			flags)
2910{
2911	struct xlog		*log = mp->m_log;
2912	struct xlog_in_core	*iclog;
 
2913
2914	XFS_STATS_INC(mp, xs_log_force);
2915	trace_xfs_log_force(mp, 0, _RET_IP_);
2916
2917	xlog_cil_force(log);
2918
2919	spin_lock(&log->l_icloglock);
2920	if (xlog_is_shutdown(log))
 
2921		goto out_error;
2922
2923	iclog = log->l_iclog;
2924	trace_xlog_iclog_force(iclog, _RET_IP_);
2925
2926	if (iclog->ic_state == XLOG_STATE_DIRTY ||
2927	    (iclog->ic_state == XLOG_STATE_ACTIVE &&
2928	     atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
2929		/*
2930		 * If the head is dirty or (active and empty), then we need to
2931		 * look at the previous iclog.
2932		 *
2933		 * If the previous iclog is active or dirty we are done.  There
2934		 * is nothing to sync out. Otherwise, we attach ourselves to the
2935		 * previous iclog and go to sleep.
2936		 */
2937		iclog = iclog->ic_prev;
 
 
 
2938	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2939		if (atomic_read(&iclog->ic_refcnt) == 0) {
2940			/* We have exclusive access to this iclog. */
2941			bool	completed;
 
 
 
 
 
 
 
 
 
2942
2943			if (xlog_force_and_check_iclog(iclog, &completed))
2944				goto out_error;
2945
2946			if (completed)
 
 
2947				goto out_unlock;
2948		} else {
2949			/*
2950			 * Someone else is still writing to this iclog, so we
2951			 * need to ensure that when they release the iclog it
2952			 * gets synced immediately as we may be waiting on it.
 
 
2953			 */
2954			xlog_state_switch_iclogs(log, iclog, 0);
2955		}
 
 
 
 
 
 
2956	}
2957
2958	/*
2959	 * The iclog we are about to wait on may contain the checkpoint pushed
2960	 * by the above xlog_cil_force() call, but it may not have been pushed
2961	 * to disk yet. Like the ACTIVE case above, we need to make sure caches
2962	 * are flushed when this iclog is written.
2963	 */
2964	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
2965		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
 
 
2966
2967	if (flags & XFS_LOG_SYNC)
2968		return xlog_wait_on_iclog(iclog);
2969out_unlock:
2970	spin_unlock(&log->l_icloglock);
2971	return 0;
2972out_error:
2973	spin_unlock(&log->l_icloglock);
2974	return -EIO;
2975}
2976
2977/*
2978 * Force the log to a specific LSN.
2979 *
2980 * If an iclog with that lsn can be found:
2981 *	If it is in the DIRTY state, just return.
2982 *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2983 *		state and go to sleep or return.
2984 *	If it is in any other state, go to sleep or return.
2985 *
2986 * Synchronous forces are implemented with a wait queue.  All callers trying
2987 * to force a given lsn to disk must wait on the queue attached to the
2988 * specific in-core log.  When given in-core log finally completes its write
2989 * to disk, that thread will wake up all threads waiting on the queue.
2990 */
2991static int
2992xlog_force_lsn(
2993	struct xlog		*log,
2994	xfs_lsn_t		lsn,
2995	uint			flags,
2996	int			*log_flushed,
2997	bool			already_slept)
2998{
 
2999	struct xlog_in_core	*iclog;
3000	bool			completed;
3001
3002	spin_lock(&log->l_icloglock);
3003	if (xlog_is_shutdown(log))
 
3004		goto out_error;
3005
3006	iclog = log->l_iclog;
3007	while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3008		trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3009		iclog = iclog->ic_next;
3010		if (iclog == log->l_iclog)
3011			goto out_unlock;
3012	}
3013
3014	switch (iclog->ic_state) {
3015	case XLOG_STATE_ACTIVE:
 
 
3016		/*
3017		 * We sleep here if we haven't already slept (e.g. this is the
3018		 * first time we've looked at the correct iclog buf) and the
3019		 * buffer before us is going to be sync'ed.  The reason for this
3020		 * is that if we are doing sync transactions here, by waiting
3021		 * for the previous I/O to complete, we can allow a few more
3022		 * transactions into this iclog before we close it down.
3023		 *
3024		 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3025		 * refcnt so we can release the log (which drops the ref count).
3026		 * The state switch keeps new transaction commits from using
3027		 * this buffer.  When the current commits finish writing into
3028		 * the buffer, the refcount will drop to zero and the buffer
3029		 * will go out then.
3030		 */
3031		if (!already_slept &&
3032		    (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3033		     iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
 
 
 
 
3034			xlog_wait(&iclog->ic_prev->ic_write_wait,
3035					&log->l_icloglock);
3036			return -EAGAIN;
3037		}
3038		if (xlog_force_and_check_iclog(iclog, &completed))
3039			goto out_error;
 
 
 
3040		if (log_flushed)
3041			*log_flushed = 1;
3042		if (completed)
3043			goto out_unlock;
3044		break;
3045	case XLOG_STATE_WANT_SYNC:
3046		/*
3047		 * This iclog may contain the checkpoint pushed by the
3048		 * xlog_cil_force_seq() call, but there are other writers still
3049		 * accessing it so it hasn't been pushed to disk yet. Like the
3050		 * ACTIVE case above, we need to make sure caches are flushed
3051		 * when this iclog is written.
3052		 */
3053		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3054		break;
3055	default:
3056		/*
3057		 * The entire checkpoint was written by the CIL force and is on
3058		 * its way to disk already. It will be stable when it
3059		 * completes, so we don't need to manipulate caches here at all.
3060		 * We just need to wait for completion if necessary.
3061		 */
3062		break;
3063	}
3064
3065	if (flags & XFS_LOG_SYNC)
3066		return xlog_wait_on_iclog(iclog);
 
 
 
 
 
 
 
 
 
 
 
3067out_unlock:
3068	spin_unlock(&log->l_icloglock);
3069	return 0;
3070out_error:
3071	spin_unlock(&log->l_icloglock);
3072	return -EIO;
3073}
3074
3075/*
3076 * Force the log to a specific checkpoint sequence.
 
 
 
 
 
 
3077 *
3078 * First force the CIL so that all the required changes have been flushed to the
3079 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3080 * the iclog that needs to be flushed to stable storage. If the caller needs
3081 * a synchronous log force, we will wait on the iclog with the LSN returned by
3082 * xlog_cil_force_seq() to be completed.
3083 */
3084int
3085xfs_log_force_seq(
3086	struct xfs_mount	*mp,
3087	xfs_csn_t		seq,
3088	uint			flags,
3089	int			*log_flushed)
3090{
3091	struct xlog		*log = mp->m_log;
3092	xfs_lsn_t		lsn;
3093	int			ret;
3094	ASSERT(seq != 0);
3095
3096	XFS_STATS_INC(mp, xs_log_force);
3097	trace_xfs_log_force(mp, seq, _RET_IP_);
3098
3099	lsn = xlog_cil_force_seq(log, seq);
3100	if (lsn == NULLCOMMITLSN)
3101		return 0;
3102
3103	ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3104	if (ret == -EAGAIN) {
3105		XFS_STATS_INC(mp, xs_log_force_sleep);
3106		ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3107	}
3108	return ret;
3109}
3110
 
 
 
 
 
 
 
 
3111/*
3112 * Free a used ticket when its refcount falls to zero.
3113 */
3114void
3115xfs_log_ticket_put(
3116	xlog_ticket_t	*ticket)
3117{
3118	ASSERT(atomic_read(&ticket->t_ref) > 0);
3119	if (atomic_dec_and_test(&ticket->t_ref))
3120		kmem_cache_free(xfs_log_ticket_cache, ticket);
3121}
3122
3123xlog_ticket_t *
3124xfs_log_ticket_get(
3125	xlog_ticket_t	*ticket)
3126{
3127	ASSERT(atomic_read(&ticket->t_ref) > 0);
3128	atomic_inc(&ticket->t_ref);
3129	return ticket;
3130}
3131
3132/*
3133 * Figure out the total log space unit (in bytes) that would be
3134 * required for a log ticket.
3135 */
3136static int
3137xlog_calc_unit_res(
3138	struct xlog		*log,
3139	int			unit_bytes,
3140	int			*niclogs)
3141{
 
3142	int			iclog_space;
3143	uint			num_headers;
3144
3145	/*
3146	 * Permanent reservations have up to 'cnt'-1 active log operations
3147	 * in the log.  A unit in this case is the amount of space for one
3148	 * of these log operations.  Normal reservations have a cnt of 1
3149	 * and their unit amount is the total amount of space required.
3150	 *
3151	 * The following lines of code account for non-transaction data
3152	 * which occupy space in the on-disk log.
3153	 *
3154	 * Normal form of a transaction is:
3155	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3156	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3157	 *
3158	 * We need to account for all the leadup data and trailer data
3159	 * around the transaction data.
3160	 * And then we need to account for the worst case in terms of using
3161	 * more space.
3162	 * The worst case will happen if:
3163	 * - the placement of the transaction happens to be such that the
3164	 *   roundoff is at its maximum
3165	 * - the transaction data is synced before the commit record is synced
3166	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3167	 *   Therefore the commit record is in its own Log Record.
3168	 *   This can happen as the commit record is called with its
3169	 *   own region to xlog_write().
3170	 *   This then means that in the worst case, roundoff can happen for
3171	 *   the commit-rec as well.
3172	 *   The commit-rec is smaller than padding in this scenario and so it is
3173	 *   not added separately.
3174	 */
3175
3176	/* for trans header */
3177	unit_bytes += sizeof(xlog_op_header_t);
3178	unit_bytes += sizeof(xfs_trans_header_t);
3179
3180	/* for start-rec */
3181	unit_bytes += sizeof(xlog_op_header_t);
3182
3183	/*
3184	 * for LR headers - the space for data in an iclog is the size minus
3185	 * the space used for the headers. If we use the iclog size, then we
3186	 * undercalculate the number of headers required.
3187	 *
3188	 * Furthermore - the addition of op headers for split-recs might
3189	 * increase the space required enough to require more log and op
3190	 * headers, so take that into account too.
3191	 *
3192	 * IMPORTANT: This reservation makes the assumption that if this
3193	 * transaction is the first in an iclog and hence has the LR headers
3194	 * accounted to it, then the remaining space in the iclog is
3195	 * exclusively for this transaction.  i.e. if the transaction is larger
3196	 * than the iclog, it will be the only thing in that iclog.
3197	 * Fundamentally, this means we must pass the entire log vector to
3198	 * xlog_write to guarantee this.
3199	 */
3200	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3201	num_headers = howmany(unit_bytes, iclog_space);
3202
3203	/* for split-recs - ophdrs added when data split over LRs */
3204	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3205
3206	/* add extra header reservations if we overrun */
3207	while (!num_headers ||
3208	       howmany(unit_bytes, iclog_space) > num_headers) {
3209		unit_bytes += sizeof(xlog_op_header_t);
3210		num_headers++;
3211	}
3212	unit_bytes += log->l_iclog_hsize * num_headers;
3213
3214	/* for commit-rec LR header - note: padding will subsume the ophdr */
3215	unit_bytes += log->l_iclog_hsize;
3216
3217	/* roundoff padding for transaction data and one for commit record */
3218	unit_bytes += 2 * log->l_iclog_roundoff;
 
 
 
 
 
 
3219
3220	if (niclogs)
3221		*niclogs = num_headers;
3222	return unit_bytes;
3223}
3224
3225int
3226xfs_log_calc_unit_res(
3227	struct xfs_mount	*mp,
3228	int			unit_bytes)
3229{
3230	return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3231}
3232
3233/*
3234 * Allocate and initialise a new log ticket.
3235 */
3236struct xlog_ticket *
3237xlog_ticket_alloc(
3238	struct xlog		*log,
3239	int			unit_bytes,
3240	int			cnt,
3241	bool			permanent)
 
 
3242{
3243	struct xlog_ticket	*tic;
3244	int			unit_res;
3245
3246	tic = kmem_cache_zalloc(xfs_log_ticket_cache,
3247			GFP_KERNEL | __GFP_NOFAIL);
 
3248
3249	unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3250
3251	atomic_set(&tic->t_ref, 1);
3252	tic->t_task		= current;
3253	INIT_LIST_HEAD(&tic->t_queue);
3254	tic->t_unit_res		= unit_res;
3255	tic->t_curr_res		= unit_res;
3256	tic->t_cnt		= cnt;
3257	tic->t_ocnt		= cnt;
3258	tic->t_tid		= get_random_u32();
 
 
3259	if (permanent)
3260		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3261
 
 
3262	return tic;
3263}
3264
 
 
 
 
 
 
 
3265#if defined(DEBUG)
3266static void
3267xlog_verify_dump_tail(
3268	struct xlog		*log,
3269	struct xlog_in_core	*iclog)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3270{
3271	xfs_alert(log->l_mp,
3272"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",
3273			iclog ? be64_to_cpu(iclog->ic_header.h_tail_lsn) : -1,
3274			atomic64_read(&log->l_tail_lsn),
3275			log->l_ailp->ail_head_lsn,
3276			log->l_curr_cycle, log->l_curr_block,
3277			log->l_prev_cycle, log->l_prev_block);
3278	xfs_alert(log->l_mp,
3279"write grant 0x%llx, reserve grant 0x%llx, tail_space 0x%llx, size 0x%x, iclog flags 0x%x",
3280			atomic64_read(&log->l_write_head.grant),
3281			atomic64_read(&log->l_reserve_head.grant),
3282			log->l_tail_space, log->l_logsize,
3283			iclog ? iclog->ic_flags : -1);
 
 
 
 
 
 
 
3284}
3285
3286/* Check if the new iclog will fit in the log. */
3287STATIC void
3288xlog_verify_tail_lsn(
3289	struct xlog		*log,
3290	struct xlog_in_core	*iclog)
 
3291{
3292	xfs_lsn_t	tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3293	int		blocks;
3294
3295	if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3296		blocks = log->l_logBBsize -
3297				(log->l_prev_block - BLOCK_LSN(tail_lsn));
3298		if (blocks < BTOBB(iclog->ic_offset) +
3299					BTOBB(log->l_iclog_hsize)) {
3300			xfs_emerg(log->l_mp,
3301					"%s: ran out of log space", __func__);
3302			xlog_verify_dump_tail(log, iclog);
3303		}
3304		return;
3305	}
3306
3307	if (CYCLE_LSN(tail_lsn) + 1 != log->l_prev_cycle) {
3308		xfs_emerg(log->l_mp, "%s: head has wrapped tail.", __func__);
3309		xlog_verify_dump_tail(log, iclog);
3310		return;
3311	}
3312	if (BLOCK_LSN(tail_lsn) == log->l_prev_block) {
3313		xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3314		xlog_verify_dump_tail(log, iclog);
3315		return;
3316	}
3317
3318	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3319	if (blocks < BTOBB(iclog->ic_offset) + 1) {
3320		xfs_emerg(log->l_mp, "%s: ran out of iclog space", __func__);
3321		xlog_verify_dump_tail(log, iclog);
3322	}
3323}
3324
3325/*
3326 * Perform a number of checks on the iclog before writing to disk.
3327 *
3328 * 1. Make sure the iclogs are still circular
3329 * 2. Make sure we have a good magic number
3330 * 3. Make sure we don't have magic numbers in the data
3331 * 4. Check fields of each log operation header for:
3332 *	A. Valid client identifier
3333 *	B. tid ptr value falls in valid ptr space (user space code)
3334 *	C. Length in log record header is correct according to the
3335 *		individual operation headers within record.
3336 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3337 *	log, check the preceding blocks of the physical log to make sure all
3338 *	the cycle numbers agree with the current cycle number.
3339 */
3340STATIC void
3341xlog_verify_iclog(
3342	struct xlog		*log,
3343	struct xlog_in_core	*iclog,
3344	int			count)
 
3345{
3346	xlog_op_header_t	*ophead;
3347	xlog_in_core_t		*icptr;
3348	xlog_in_core_2_t	*xhdr;
3349	void			*base_ptr, *ptr, *p;
3350	ptrdiff_t		field_offset;
3351	uint8_t			clientid;
3352	int			len, i, j, k, op_len;
3353	int			idx;
3354
3355	/* check validity of iclog pointers */
3356	spin_lock(&log->l_icloglock);
3357	icptr = log->l_iclog;
3358	for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3359		ASSERT(icptr);
3360
3361	if (icptr != log->l_iclog)
3362		xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3363	spin_unlock(&log->l_icloglock);
3364
3365	/* check log magic numbers */
3366	if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3367		xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3368
3369	base_ptr = ptr = &iclog->ic_header;
3370	p = &iclog->ic_header;
3371	for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3372		if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3373			xfs_emerg(log->l_mp, "%s: unexpected magic num",
3374				__func__);
3375	}
3376
3377	/* check fields */
3378	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3379	base_ptr = ptr = iclog->ic_datap;
3380	ophead = ptr;
3381	xhdr = iclog->ic_data;
3382	for (i = 0; i < len; i++) {
3383		ophead = ptr;
3384
3385		/* clientid is only 1 byte */
3386		p = &ophead->oh_clientid;
3387		field_offset = p - base_ptr;
3388		if (field_offset & 0x1ff) {
3389			clientid = ophead->oh_clientid;
3390		} else {
3391			idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3392			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3393				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3394				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3395				clientid = xlog_get_client_id(
3396					xhdr[j].hic_xheader.xh_cycle_data[k]);
3397			} else {
3398				clientid = xlog_get_client_id(
3399					iclog->ic_header.h_cycle_data[idx]);
3400			}
3401		}
3402		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3403			xfs_warn(log->l_mp,
3404				"%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3405				__func__, i, clientid, ophead,
3406				(unsigned long)field_offset);
3407		}
3408
3409		/* check length */
3410		p = &ophead->oh_len;
3411		field_offset = p - base_ptr;
3412		if (field_offset & 0x1ff) {
3413			op_len = be32_to_cpu(ophead->oh_len);
3414		} else {
3415			idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
 
3416			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3417				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3418				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3419				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3420			} else {
3421				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3422			}
3423		}
3424		ptr += sizeof(xlog_op_header_t) + op_len;
3425	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3426}
3427#endif
3428
3429/*
3430 * Perform a forced shutdown on the log.
3431 *
3432 * This can be called from low level log code to trigger a shutdown, or from the
3433 * high level mount shutdown code when the mount shuts down.
3434 *
3435 * Our main objectives here are to make sure that:
3436 *	a. if the shutdown was not due to a log IO error, flush the logs to
3437 *	   disk. Anything modified after this is ignored.
3438 *	b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3439 *	   parties to find out. Nothing new gets queued after this is done.
3440 *	c. Tasks sleeping on log reservations, pinned objects and
3441 *	   other resources get woken up.
3442 *	d. The mount is also marked as shut down so that log triggered shutdowns
3443 *	   still behave the same as if they called xfs_forced_shutdown().
3444 *
3445 * Return true if the shutdown cause was a log IO error and we actually shut the
3446 * log down.
3447 */
3448bool
3449xlog_force_shutdown(
3450	struct xlog	*log,
3451	uint32_t	shutdown_flags)
3452{
3453	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
 
3454
3455	if (!log)
3456		return false;
3457
3458	/*
3459	 * Ensure that there is only ever one log shutdown being processed.
3460	 * If we allow the log force below on a second pass after shutting
3461	 * down the log, we risk deadlocking the CIL push as it may require
3462	 * locks on objects the current shutdown context holds (e.g. taking
3463	 * buffer locks to abort buffers on last unpin of buf log items).
3464	 */
3465	if (test_and_set_bit(XLOG_SHUTDOWN_STARTED, &log->l_opstate))
3466		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3467
3468	/*
3469	 * Flush all the completed transactions to disk before marking the log
3470	 * being shut down. We need to do this first as shutting down the log
3471	 * before the force will prevent the log force from flushing the iclogs
3472	 * to disk.
3473	 *
3474	 * When we are in recovery, there are no transactions to flush, and
3475	 * we don't want to touch the log because we don't want to perturb the
3476	 * current head/tail for future recovery attempts. Hence we need to
3477	 * avoid a log force in this case.
3478	 *
3479	 * If we are shutting down due to a log IO error, then we must avoid
3480	 * trying to write the log as that may just result in more IO errors and
3481	 * an endless shutdown/force loop.
3482	 */
3483	if (!log_error && !xlog_in_recovery(log))
3484		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3485
3486	/*
3487	 * Atomically set the shutdown state. If the shutdown state is already
3488	 * set, there someone else is performing the shutdown and so we are done
3489	 * here. This should never happen because we should only ever get called
3490	 * once by the first shutdown caller.
3491	 *
3492	 * Much of the log state machine transitions assume that shutdown state
3493	 * cannot change once they hold the log->l_icloglock. Hence we need to
3494	 * hold that lock here, even though we use the atomic test_and_set_bit()
3495	 * operation to set the shutdown state.
3496	 */
3497	spin_lock(&log->l_icloglock);
3498	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3499		spin_unlock(&log->l_icloglock);
3500		ASSERT(0);
3501		return false;
3502	}
3503	spin_unlock(&log->l_icloglock);
3504
3505	/*
3506	 * If this log shutdown also sets the mount shutdown state, issue a
3507	 * shutdown warning message.
3508	 */
3509	if (!xfs_set_shutdown(log->l_mp)) {
3510		xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3511"Filesystem has been shut down due to log error (0x%x).",
3512				shutdown_flags);
3513		xfs_alert(log->l_mp,
3514"Please unmount the filesystem and rectify the problem(s).");
3515		if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3516			xfs_stack_trace();
3517	}
3518
3519	/*
3520	 * We don't want anybody waiting for log reservations after this. That
3521	 * means we have to wake up everybody queued up on reserveq as well as
3522	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3523	 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3524	 * action is protected by the grant locks.
3525	 */
3526	xlog_grant_head_wake_all(&log->l_reserve_head);
3527	xlog_grant_head_wake_all(&log->l_write_head);
3528
3529	/*
3530	 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3531	 * as if the log writes were completed. The abort handling in the log
3532	 * item committed callback functions will do this again under lock to
3533	 * avoid races.
3534	 */
3535	spin_lock(&log->l_cilp->xc_push_lock);
3536	wake_up_all(&log->l_cilp->xc_start_wait);
3537	wake_up_all(&log->l_cilp->xc_commit_wait);
3538	spin_unlock(&log->l_cilp->xc_push_lock);
3539
3540	spin_lock(&log->l_icloglock);
3541	xlog_state_shutdown_callbacks(log);
3542	spin_unlock(&log->l_icloglock);
3543
3544	wake_up_var(&log->l_opstate);
3545	return log_error;
 
 
 
 
 
 
 
 
 
3546}
3547
3548STATIC int
3549xlog_iclogs_empty(
3550	struct xlog	*log)
3551{
3552	xlog_in_core_t	*iclog;
3553
3554	iclog = log->l_iclog;
3555	do {
3556		/* endianness does not matter here, zero is zero in
3557		 * any language.
3558		 */
3559		if (iclog->ic_header.h_num_logops)
3560			return 0;
3561		iclog = iclog->ic_next;
3562	} while (iclog != log->l_iclog);
3563	return 1;
3564}
3565
3566/*
3567 * Verify that an LSN stamped into a piece of metadata is valid. This is
3568 * intended for use in read verifiers on v5 superblocks.
3569 */
3570bool
3571xfs_log_check_lsn(
3572	struct xfs_mount	*mp,
3573	xfs_lsn_t		lsn)
3574{
3575	struct xlog		*log = mp->m_log;
3576	bool			valid;
3577
3578	/*
3579	 * norecovery mode skips mount-time log processing and unconditionally
3580	 * resets the in-core LSN. We can't validate in this mode, but
3581	 * modifications are not allowed anyways so just return true.
3582	 */
3583	if (xfs_has_norecovery(mp))
3584		return true;
3585
3586	/*
3587	 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3588	 * handled by recovery and thus safe to ignore here.
3589	 */
3590	if (lsn == NULLCOMMITLSN)
3591		return true;
3592
3593	valid = xlog_valid_lsn(mp->m_log, lsn);
3594
3595	/* warn the user about what's gone wrong before verifier failure */
3596	if (!valid) {
3597		spin_lock(&log->l_icloglock);
3598		xfs_warn(mp,
3599"Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3600"Please unmount and run xfs_repair (>= v4.3) to resolve.",
3601			 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3602			 log->l_curr_cycle, log->l_curr_block);
3603		spin_unlock(&log->l_icloglock);
3604	}
3605
3606	return valid;
3607}