xfs_trans.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
   4 * Copyright (C) 2010 Red Hat, Inc.
   5 * All Rights Reserved.
   6 */
   7#include "xfs.h"
   8#include "xfs_fs.h"
   9#include "xfs_shared.h"
  10#include "xfs_format.h"
  11#include "xfs_log_format.h"
  12#include "xfs_trans_resv.h"
  13#include "xfs_mount.h"
  14#include "xfs_extent_busy.h"
  15#include "xfs_quota.h"
  16#include "xfs_trans.h"
  17#include "xfs_trans_priv.h"
  18#include "xfs_log.h"
  19#include "xfs_log_priv.h"
  20#include "xfs_trace.h"
  21#include "xfs_error.h"
  22#include "xfs_defer.h"
  23#include "xfs_inode.h"
  24#include "xfs_dquot_item.h"
  25#include "xfs_dquot.h"
  26#include "xfs_icache.h"
  27#include "xfs_rtbitmap.h"
  28#include "xfs_rtgroup.h"
  29#include "xfs_sb.h"
  30
  31struct kmem_cache	*xfs_trans_cache;
  32
  33#if defined(CONFIG_TRACEPOINTS)
  34static void
  35xfs_trans_trace_reservations(
  36	struct xfs_mount	*mp)
  37{
  38	struct xfs_trans_res	*res;
  39	struct xfs_trans_res	*end_res;
  40	int			i;
  41
  42	res = (struct xfs_trans_res *)M_RES(mp);
  43	end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);
  44	for (i = 0; res < end_res; i++, res++)
  45		trace_xfs_trans_resv_calc(mp, i, res);
  46}
  47#else
  48# define xfs_trans_trace_reservations(mp)
  49#endif
  50
  51/*
  52 * Initialize the precomputed transaction reservation values
  53 * in the mount structure.
  54 */
  55void
  56xfs_trans_init(
  57	struct xfs_mount	*mp)
  58{
  59	xfs_trans_resv_calc(mp, M_RES(mp));
  60	xfs_trans_trace_reservations(mp);
  61}
  62
  63/*
  64 * Free the transaction structure.  If there is more clean up
  65 * to do when the structure is freed, add it here.
  66 */
  67STATIC void
  68xfs_trans_free(
  69	struct xfs_trans	*tp)
  70{
  71	xfs_extent_busy_sort(&tp->t_busy);
  72	xfs_extent_busy_clear(&tp->t_busy, false);
  73
  74	trace_xfs_trans_free(tp, _RET_IP_);
  75	xfs_trans_clear_context(tp);
  76	if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
  77		sb_end_intwrite(tp->t_mountp->m_super);
  78	xfs_trans_free_dqinfo(tp);
  79	kmem_cache_free(xfs_trans_cache, tp);
  80}
  81
  82/*
  83 * This is called to create a new transaction which will share the
  84 * permanent log reservation of the given transaction.  The remaining
  85 * unused block and rt extent reservations are also inherited.  This
  86 * implies that the original transaction is no longer allowed to allocate
  87 * blocks.  Locks and log items, however, are no inherited.  They must
  88 * be added to the new transaction explicitly.
  89 */
  90STATIC struct xfs_trans *
  91xfs_trans_dup(
  92	struct xfs_trans	*tp)
  93{
  94	struct xfs_trans	*ntp;
  95
  96	trace_xfs_trans_dup(tp, _RET_IP_);
  97
  98	ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
  99
 100	/*
 101	 * Initialize the new transaction structure.
 102	 */
 103	ntp->t_mountp = tp->t_mountp;
 104	INIT_LIST_HEAD(&ntp->t_items);
 105	INIT_LIST_HEAD(&ntp->t_busy);
 106	INIT_LIST_HEAD(&ntp->t_dfops);
 107	ntp->t_highest_agno = NULLAGNUMBER;
 108
 109	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
 110	ASSERT(tp->t_ticket != NULL);
 111
 112	ntp->t_flags = XFS_TRANS_PERM_LOG_RES |
 113		       (tp->t_flags & XFS_TRANS_RESERVE) |
 114		       (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |
 115		       (tp->t_flags & XFS_TRANS_RES_FDBLKS);
 116	/* We gave our writer reference to the new transaction */
 117	tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;
 118	ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);
 119
 120	ASSERT(tp->t_blk_res >= tp->t_blk_res_used);
 121	ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;
 122	tp->t_blk_res = tp->t_blk_res_used;
 123
 124	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
 125	tp->t_rtx_res = tp->t_rtx_res_used;
 126
 127	xfs_trans_switch_context(tp, ntp);
 128
 129	/* move deferred ops over to the new tp */
 130	xfs_defer_move(ntp, tp);
 131
 132	xfs_trans_dup_dqinfo(tp, ntp);
 133	return ntp;
 134}
 135
 136/*
 137 * This is called to reserve free disk blocks and log space for the given
 138 * transaction before allocating any resources within the transaction.
 139 *
 140 * This will return ENOSPC if there are not enough blocks available.
 141 * It will sleep waiting for available log space.
 142 *
 143 * This does not do quota reservations. That typically is done by the caller
 144 * afterwards.
 145 */
 146static int
 147xfs_trans_reserve(
 148	struct xfs_trans	*tp,
 149	struct xfs_trans_res	*resp,
 150	uint			blocks,
 151	uint			rtextents)
 152{
 153	struct xfs_mount	*mp = tp->t_mountp;
 154	int			error = 0;
 155	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 156
 157	ASSERT(resp->tr_logres > 0);
 158
 159	/*
 160	 * Attempt to reserve the needed disk blocks by decrementing the number
 161	 * needed from the number available.  This will fail if the count would
 162	 * go below zero.
 163	 */
 164	if (blocks > 0) {
 165		error = xfs_dec_fdblocks(mp, blocks, rsvd);
 166		if (error != 0)
 167			return -ENOSPC;
 168		tp->t_blk_res += blocks;
 169	}
 170
 171	/*
 172	 * Reserve the log space needed for this transaction.
 173	 */
 174	if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES)
 175		tp->t_flags |= XFS_TRANS_PERM_LOG_RES;
 176	error = xfs_log_reserve(mp, resp->tr_logres, resp->tr_logcount,
 177			&tp->t_ticket, (tp->t_flags & XFS_TRANS_PERM_LOG_RES));
 178	if (error)
 179		goto undo_blocks;
 180
 181	tp->t_log_res = resp->tr_logres;
 182	tp->t_log_count = resp->tr_logcount;
 183
 184	/*
 185	 * Attempt to reserve the needed realtime extents by decrementing the
 186	 * number needed from the number available.  This will fail if the
 187	 * count would go below zero.
 188	 */
 189	if (rtextents > 0) {
 190		error = xfs_dec_frextents(mp, rtextents);
 191		if (error) {
 192			error = -ENOSPC;
 193			goto undo_log;
 194		}
 195		tp->t_rtx_res += rtextents;
 196	}
 197
 198	return 0;
 199
 200undo_log:
 201	xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
 202	tp->t_ticket = NULL;
 203	tp->t_log_res = 0;
 204	tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;
 205undo_blocks:
 206	if (blocks > 0) {
 207		xfs_add_fdblocks(mp, blocks);
 208		tp->t_blk_res = 0;
 209	}
 210	return error;
 211}
 212
 213static struct xfs_trans *
 214__xfs_trans_alloc(
 215	struct xfs_mount	*mp,
 216	uint			flags)
 217{
 218	struct xfs_trans	*tp;
 219
 220	ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) || xfs_has_lazysbcount(mp));
 221
 222	tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);
 223	if (!(flags & XFS_TRANS_NO_WRITECOUNT))
 224		sb_start_intwrite(mp->m_super);
 225	xfs_trans_set_context(tp);
 226	tp->t_flags = flags;
 227	tp->t_mountp = mp;
 228	INIT_LIST_HEAD(&tp->t_items);
 229	INIT_LIST_HEAD(&tp->t_busy);
 230	INIT_LIST_HEAD(&tp->t_dfops);
 231	tp->t_highest_agno = NULLAGNUMBER;
 232	return tp;
 233}
 234
 235int
 236xfs_trans_alloc(
 237	struct xfs_mount	*mp,
 238	struct xfs_trans_res	*resp,
 239	uint			blocks,
 240	uint			rtextents,
 241	uint			flags,
 242	struct xfs_trans	**tpp)
 243{
 244	struct xfs_trans	*tp;
 245	bool			want_retry = true;
 246	int			error;
 247
 248	ASSERT(resp->tr_logres > 0);
 249
 250	/*
 251	 * Allocate the handle before we do our freeze accounting and setting up
 252	 * GFP_NOFS allocation context so that we avoid lockdep false positives
 253	 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
 254	 */
 255retry:
 256	tp = __xfs_trans_alloc(mp, flags);
 257	WARN_ON(mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);
 258	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
 259	if (error == -ENOSPC && want_retry) {
 260		xfs_trans_cancel(tp);
 261
 262		/*
 263		 * We weren't able to reserve enough space for the transaction.
 264		 * Flush the other speculative space allocations to free space.
 265		 * Do not perform a synchronous scan because callers can hold
 266		 * other locks.
 267		 */
 268		error = xfs_blockgc_flush_all(mp);
 269		if (error)
 270			return error;
 271		want_retry = false;
 272		goto retry;
 273	}
 274	if (error) {
 275		xfs_trans_cancel(tp);
 276		return error;
 277	}
 278
 279	trace_xfs_trans_alloc(tp, _RET_IP_);
 280
 281	*tpp = tp;
 282	return 0;
 283}
 284
 285/*
 286 * Create an empty transaction with no reservation.  This is a defensive
 287 * mechanism for routines that query metadata without actually modifying them --
 288 * if the metadata being queried is somehow cross-linked (think a btree block
 289 * pointer that points higher in the tree), we risk deadlock.  However, blocks
 290 * grabbed as part of a transaction can be re-grabbed.  The verifiers will
 291 * notice the corrupt block and the operation will fail back to userspace
 292 * without deadlocking.
 293 *
 294 * Note the zero-length reservation; this transaction MUST be cancelled without
 295 * any dirty data.
 296 *
 297 * Callers should obtain freeze protection to avoid a conflict with fs freezing
 298 * where we can be grabbing buffers at the same time that freeze is trying to
 299 * drain the buffer LRU list.
 300 */
 301struct xfs_trans *
 302xfs_trans_alloc_empty(
 303	struct xfs_mount		*mp)
 304{
 305	return __xfs_trans_alloc(mp, XFS_TRANS_NO_WRITECOUNT);
 306}
 307
 308/*
 309 * Record the indicated change to the given field for application
 310 * to the file system's superblock when the transaction commits.
 311 * For now, just store the change in the transaction structure.
 312 *
 313 * Mark the transaction structure to indicate that the superblock
 314 * needs to be updated before committing.
 315 *
 316 * Because we may not be keeping track of allocated/free inodes and
 317 * used filesystem blocks in the superblock, we do not mark the
 318 * superblock dirty in this transaction if we modify these fields.
 319 * We still need to update the transaction deltas so that they get
 320 * applied to the incore superblock, but we don't want them to
 321 * cause the superblock to get locked and logged if these are the
 322 * only fields in the superblock that the transaction modifies.
 323 */
 324void
 325xfs_trans_mod_sb(
 326	xfs_trans_t	*tp,
 327	uint		field,
 328	int64_t		delta)
 329{
 330	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);
 331	xfs_mount_t	*mp = tp->t_mountp;
 332
 333	switch (field) {
 334	case XFS_TRANS_SB_ICOUNT:
 335		tp->t_icount_delta += delta;
 336		if (xfs_has_lazysbcount(mp))
 337			flags &= ~XFS_TRANS_SB_DIRTY;
 338		break;
 339	case XFS_TRANS_SB_IFREE:
 340		tp->t_ifree_delta += delta;
 341		if (xfs_has_lazysbcount(mp))
 342			flags &= ~XFS_TRANS_SB_DIRTY;
 343		break;
 344	case XFS_TRANS_SB_FDBLOCKS:
 345		/*
 346		 * Track the number of blocks allocated in the transaction.
 347		 * Make sure it does not exceed the number reserved. If so,
 348		 * shutdown as this can lead to accounting inconsistency.
 349		 */
 350		if (delta < 0) {
 351			tp->t_blk_res_used += (uint)-delta;
 352			if (tp->t_blk_res_used > tp->t_blk_res)
 353				xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
 354		} else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {
 355			int64_t	blkres_delta;
 356
 357			/*
 358			 * Return freed blocks directly to the reservation
 359			 * instead of the global pool, being careful not to
 360			 * overflow the trans counter. This is used to preserve
 361			 * reservation across chains of transaction rolls that
 362			 * repeatedly free and allocate blocks.
 363			 */
 364			blkres_delta = min_t(int64_t, delta,
 365					     UINT_MAX - tp->t_blk_res);
 366			tp->t_blk_res += blkres_delta;
 367			delta -= blkres_delta;
 368		}
 369		tp->t_fdblocks_delta += delta;
 370		if (xfs_has_lazysbcount(mp))
 371			flags &= ~XFS_TRANS_SB_DIRTY;
 372		break;
 373	case XFS_TRANS_SB_RES_FDBLOCKS:
 374		/*
 375		 * The allocation has already been applied to the
 376		 * in-core superblock's counter.  This should only
 377		 * be applied to the on-disk superblock.
 378		 */
 379		tp->t_res_fdblocks_delta += delta;
 380		if (xfs_has_lazysbcount(mp))
 381			flags &= ~XFS_TRANS_SB_DIRTY;
 382		break;
 383	case XFS_TRANS_SB_FREXTENTS:
 384		/*
 385		 * Track the number of blocks allocated in the
 386		 * transaction.  Make sure it does not exceed the
 387		 * number reserved.
 388		 */
 389		if (delta < 0) {
 390			tp->t_rtx_res_used += (uint)-delta;
 391			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);
 392		}
 393		tp->t_frextents_delta += delta;
 394		if (xfs_has_rtgroups(mp))
 395			flags &= ~XFS_TRANS_SB_DIRTY;
 396		break;
 397	case XFS_TRANS_SB_RES_FREXTENTS:
 398		/*
 399		 * The allocation has already been applied to the
 400		 * in-core superblock's counter.  This should only
 401		 * be applied to the on-disk superblock.
 402		 */
 403		ASSERT(delta < 0);
 404		tp->t_res_frextents_delta += delta;
 405		if (xfs_has_rtgroups(mp))
 406			flags &= ~XFS_TRANS_SB_DIRTY;
 407		break;
 408	case XFS_TRANS_SB_DBLOCKS:
 409		tp->t_dblocks_delta += delta;
 410		break;
 411	case XFS_TRANS_SB_AGCOUNT:
 412		ASSERT(delta > 0);
 413		tp->t_agcount_delta += delta;
 414		break;
 415	case XFS_TRANS_SB_IMAXPCT:
 416		tp->t_imaxpct_delta += delta;
 417		break;
 418	case XFS_TRANS_SB_REXTSIZE:
 419		tp->t_rextsize_delta += delta;
 420		break;
 421	case XFS_TRANS_SB_RBMBLOCKS:
 422		tp->t_rbmblocks_delta += delta;
 423		break;
 424	case XFS_TRANS_SB_RBLOCKS:
 425		tp->t_rblocks_delta += delta;
 426		break;
 427	case XFS_TRANS_SB_REXTENTS:
 428		tp->t_rextents_delta += delta;
 429		break;
 430	case XFS_TRANS_SB_REXTSLOG:
 431		tp->t_rextslog_delta += delta;
 432		break;
 433	case XFS_TRANS_SB_RGCOUNT:
 434		ASSERT(delta > 0);
 435		tp->t_rgcount_delta += delta;
 436		break;
 437	default:
 438		ASSERT(0);
 439		return;
 440	}
 441
 442	tp->t_flags |= flags;
 443}
 444
 445/*
 446 * xfs_trans_apply_sb_deltas() is called from the commit code
 447 * to bring the superblock buffer into the current transaction
 448 * and modify it as requested by earlier calls to xfs_trans_mod_sb().
 449 *
 450 * For now we just look at each field allowed to change and change
 451 * it if necessary.
 452 */
 453STATIC void
 454xfs_trans_apply_sb_deltas(
 455	struct xfs_trans	*tp)
 456{
 457	struct xfs_mount	*mp = tp->t_mountp;
 458	struct xfs_buf		*bp = xfs_trans_getsb(tp);
 459	struct xfs_dsb		*sbp = bp->b_addr;
 460	int			whole = 0;
 461
 462	/*
 463	 * Only update the superblock counters if we are logging them
 464	 */
 465	if (!xfs_has_lazysbcount(mp)) {
 466		if (tp->t_icount_delta)
 467			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);
 468		if (tp->t_ifree_delta)
 469			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);
 470		if (tp->t_fdblocks_delta)
 471			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);
 472		if (tp->t_res_fdblocks_delta)
 473			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);
 474	}
 475
 476	/*
 477	 * sb_frextents was added to the lazy sb counters when the rt groups
 478	 * feature was introduced.  This is possible because we know that all
 479	 * kernels supporting rtgroups will also recompute frextents from the
 480	 * realtime bitmap.
 481	 *
 482	 * For older file systems, updating frextents requires careful handling
 483	 * because we cannot rely on log recovery in older kernels to recompute
 484	 * the value from the rtbitmap.  This means that the ondisk frextents
 485	 * must be consistent with the rtbitmap.
 486	 *
 487	 * Therefore, log the frextents change to the ondisk superblock and
 488	 * update the incore superblock so that future calls to xfs_log_sb
 489	 * write the correct value ondisk.
 490	 */
 491	if ((tp->t_frextents_delta || tp->t_res_frextents_delta) &&
 492	    !xfs_has_rtgroups(mp)) {
 493		int64_t			rtxdelta;
 494
 495		rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta;
 496
 497		spin_lock(&mp->m_sb_lock);
 498		be64_add_cpu(&sbp->sb_frextents, rtxdelta);
 499		mp->m_sb.sb_frextents += rtxdelta;
 500		spin_unlock(&mp->m_sb_lock);
 501	}
 502
 503	if (tp->t_dblocks_delta) {
 504		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);
 505		mp->m_ddev_targp->bt_nr_sectors +=
 506			XFS_FSB_TO_BB(mp, tp->t_dblocks_delta);
 507		whole = 1;
 508	}
 509	if (tp->t_agcount_delta) {
 510		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);
 511		whole = 1;
 512	}
 513	if (tp->t_imaxpct_delta) {
 514		sbp->sb_imax_pct += tp->t_imaxpct_delta;
 515		whole = 1;
 516	}
 517	if (tp->t_rextsize_delta) {
 518		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);
 519
 520		/*
 521		 * Because the ondisk sb records rtgroup size in units of rt
 522		 * extents, any time we update the rt extent size we have to
 523		 * recompute the ondisk rtgroup block log.  The incore values
 524		 * will be recomputed in xfs_trans_unreserve_and_mod_sb.
 525		 */
 526		if (xfs_has_rtgroups(mp)) {
 527			sbp->sb_rgblklog = xfs_compute_rgblklog(
 528						be32_to_cpu(sbp->sb_rgextents),
 529						be32_to_cpu(sbp->sb_rextsize));
 530		}
 531		whole = 1;
 532	}
 533	if (tp->t_rbmblocks_delta) {
 534		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);
 535		whole = 1;
 536	}
 537	if (tp->t_rblocks_delta) {
 538		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);
 539		mp->m_rtdev_targp->bt_nr_sectors +=
 540			XFS_FSB_TO_BB(mp, tp->t_rblocks_delta);
 541		whole = 1;
 542	}
 543	if (tp->t_rextents_delta) {
 544		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);
 545		whole = 1;
 546	}
 547	if (tp->t_rextslog_delta) {
 548		sbp->sb_rextslog += tp->t_rextslog_delta;
 549		whole = 1;
 550	}
 551	if (tp->t_rgcount_delta) {
 552		be32_add_cpu(&sbp->sb_rgcount, tp->t_rgcount_delta);
 553		whole = 1;
 554	}
 555
 556	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);
 557	if (whole)
 558		/*
 559		 * Log the whole thing, the fields are noncontiguous.
 560		 */
 561		xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);
 562	else
 563		/*
 564		 * Since all the modifiable fields are contiguous, we
 565		 * can get away with this.
 566		 */
 567		xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),
 568				  offsetof(struct xfs_dsb, sb_frextents) +
 569				  sizeof(sbp->sb_frextents) - 1);
 570}
 571
 572/*
 573 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and
 574 * apply superblock counter changes to the in-core superblock.  The
 575 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT
 576 * applied to the in-core superblock.  The idea is that that has already been
 577 * done.
 578 *
 579 * If we are not logging superblock counters, then the inode allocated/free and
 580 * used block counts are not updated in the on disk superblock. In this case,
 581 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we
 582 * still need to update the incore superblock with the changes.
 583 *
 584 * Deltas for the inode count are +/-64, hence we use a large batch size of 128
 585 * so we don't need to take the counter lock on every update.
 586 */
 587#define XFS_ICOUNT_BATCH	128
 588
 589void
 590xfs_trans_unreserve_and_mod_sb(
 591	struct xfs_trans	*tp)
 592{
 593	struct xfs_mount	*mp = tp->t_mountp;
 594	int64_t			blkdelta = tp->t_blk_res;
 595	int64_t			rtxdelta = tp->t_rtx_res;
 596	int64_t			idelta = 0;
 597	int64_t			ifreedelta = 0;
 598
 599	/*
 600	 * Calculate the deltas.
 601	 *
 602	 * t_fdblocks_delta and t_frextents_delta can be positive or negative:
 603	 *
 604	 *  - positive values indicate blocks freed in the transaction.
 605	 *  - negative values indicate blocks allocated in the transaction
 606	 *
 607	 * Negative values can only happen if the transaction has a block
 608	 * reservation that covers the allocated block.  The end result is
 609	 * that the calculated delta values must always be positive and we
 610	 * can only put back previous allocated or reserved blocks here.
 611	 */
 612	ASSERT(tp->t_blk_res || tp->t_fdblocks_delta >= 0);
 613	if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
 614	        blkdelta += tp->t_fdblocks_delta;
 615		ASSERT(blkdelta >= 0);
 616	}
 617
 618	ASSERT(tp->t_rtx_res || tp->t_frextents_delta >= 0);
 619	if (xfs_has_rtgroups(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
 620		rtxdelta += tp->t_frextents_delta;
 621		ASSERT(rtxdelta >= 0);
 622	}
 623
 624	if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {
 625		idelta = tp->t_icount_delta;
 626		ifreedelta = tp->t_ifree_delta;
 627	}
 628
 629	/* apply the per-cpu counters */
 630	if (blkdelta)
 631		xfs_add_fdblocks(mp, blkdelta);
 632
 633	if (idelta)
 634		percpu_counter_add_batch(&mp->m_icount, idelta,
 635					 XFS_ICOUNT_BATCH);
 636
 637	if (ifreedelta)
 638		percpu_counter_add(&mp->m_ifree, ifreedelta);
 639
 640	if (rtxdelta)
 641		xfs_add_frextents(mp, rtxdelta);
 642
 643	if (!(tp->t_flags & XFS_TRANS_SB_DIRTY))
 644		return;
 645
 646	/* apply remaining deltas */
 647	spin_lock(&mp->m_sb_lock);
 648	mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;
 649	mp->m_sb.sb_icount += idelta;
 650	mp->m_sb.sb_ifree += ifreedelta;
 651	/*
 652	 * Do not touch sb_frextents here because it is handled in
 653	 * xfs_trans_apply_sb_deltas for file systems where it isn't a lazy
 654	 * counter anyway.
 655	 */
 656	mp->m_sb.sb_dblocks += tp->t_dblocks_delta;
 657	mp->m_sb.sb_agcount += tp->t_agcount_delta;
 658	mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;
 659	if (tp->t_rextsize_delta)
 660		xfs_mount_sb_set_rextsize(mp, &mp->m_sb,
 661				mp->m_sb.sb_rextsize + tp->t_rextsize_delta);
 662	mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;
 663	mp->m_sb.sb_rblocks += tp->t_rblocks_delta;
 664	mp->m_sb.sb_rextents += tp->t_rextents_delta;
 665	mp->m_sb.sb_rextslog += tp->t_rextslog_delta;
 666	mp->m_sb.sb_rgcount += tp->t_rgcount_delta;
 667	spin_unlock(&mp->m_sb_lock);
 668
 669	/*
 670	 * Debug checks outside of the spinlock so they don't lock up the
 671	 * machine if they fail.
 672	 */
 673	ASSERT(mp->m_sb.sb_imax_pct >= 0);
 674	ASSERT(mp->m_sb.sb_rextslog >= 0);
 675}
 676
 677/* Add the given log item to the transaction's list of log items. */
 678void
 679xfs_trans_add_item(
 680	struct xfs_trans	*tp,
 681	struct xfs_log_item	*lip)
 682{
 683	ASSERT(lip->li_log == tp->t_mountp->m_log);
 684	ASSERT(lip->li_ailp == tp->t_mountp->m_ail);
 685	ASSERT(list_empty(&lip->li_trans));
 686	ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));
 687
 688	list_add_tail(&lip->li_trans, &tp->t_items);
 689	trace_xfs_trans_add_item(tp, _RET_IP_);
 690}
 691
 692/*
 693 * Unlink the log item from the transaction. the log item is no longer
 694 * considered dirty in this transaction, as the linked transaction has
 695 * finished, either by abort or commit completion.
 696 */
 697void
 698xfs_trans_del_item(
 699	struct xfs_log_item	*lip)
 700{
 701	clear_bit(XFS_LI_DIRTY, &lip->li_flags);
 702	list_del_init(&lip->li_trans);
 703}
 704
 705/* Detach and unlock all of the items in a transaction */
 706static void
 707xfs_trans_free_items(
 708	struct xfs_trans	*tp,
 709	bool			abort)
 710{
 711	struct xfs_log_item	*lip, *next;
 712
 713	trace_xfs_trans_free_items(tp, _RET_IP_);
 714
 715	list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {
 716		xfs_trans_del_item(lip);
 717		if (abort) {
 718			trace_xfs_trans_free_abort(lip);
 719			set_bit(XFS_LI_ABORTED, &lip->li_flags);
 720		}
 721		if (lip->li_ops->iop_release)
 722			lip->li_ops->iop_release(lip);
 723	}
 724}
 725
 726/*
 727 * Sort transaction items prior to running precommit operations. This will
 728 * attempt to order the items such that they will always be locked in the same
 729 * order. Items that have no sort function are moved to the end of the list
 730 * and so are locked last.
 731 *
 732 * This may need refinement as different types of objects add sort functions.
 733 *
 734 * Function is more complex than it needs to be because we are comparing 64 bit
 735 * values and the function only returns 32 bit values.
 736 */
 737static int
 738xfs_trans_precommit_sort(
 739	void			*unused_arg,
 740	const struct list_head	*a,
 741	const struct list_head	*b)
 742{
 743	struct xfs_log_item	*lia = container_of(a,
 744					struct xfs_log_item, li_trans);
 745	struct xfs_log_item	*lib = container_of(b,
 746					struct xfs_log_item, li_trans);
 747	int64_t			diff;
 748
 749	/*
 750	 * If both items are non-sortable, leave them alone. If only one is
 751	 * sortable, move the non-sortable item towards the end of the list.
 752	 */
 753	if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort)
 754		return 0;
 755	if (!lia->li_ops->iop_sort)
 756		return 1;
 757	if (!lib->li_ops->iop_sort)
 758		return -1;
 759
 760	diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib);
 761	if (diff < 0)
 762		return -1;
 763	if (diff > 0)
 764		return 1;
 765	return 0;
 766}
 767
 768/*
 769 * Run transaction precommit functions.
 770 *
 771 * If there is an error in any of the callouts, then stop immediately and
 772 * trigger a shutdown to abort the transaction. There is no recovery possible
 773 * from errors at this point as the transaction is dirty....
 774 */
 775static int
 776xfs_trans_run_precommits(
 777	struct xfs_trans	*tp)
 778{
 779	struct xfs_mount	*mp = tp->t_mountp;
 780	struct xfs_log_item	*lip, *n;
 781	int			error = 0;
 782
 783	/*
 784	 * Sort the item list to avoid ABBA deadlocks with other transactions
 785	 * running precommit operations that lock multiple shared items such as
 786	 * inode cluster buffers.
 787	 */
 788	list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort);
 789
 790	/*
 791	 * Precommit operations can remove the log item from the transaction
 792	 * if the log item exists purely to delay modifications until they
 793	 * can be ordered against other operations. Hence we have to use
 794	 * list_for_each_entry_safe() here.
 795	 */
 796	list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) {
 797		if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))
 798			continue;
 799		if (lip->li_ops->iop_precommit) {
 800			error = lip->li_ops->iop_precommit(tp, lip);
 801			if (error)
 802				break;
 803		}
 804	}
 805	if (error)
 806		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
 807	return error;
 808}
 809
 810/*
 811 * Commit the given transaction to the log.
 812 *
 813 * XFS disk error handling mechanism is not based on a typical
 814 * transaction abort mechanism. Logically after the filesystem
 815 * gets marked 'SHUTDOWN', we can't let any new transactions
 816 * be durable - ie. committed to disk - because some metadata might
 817 * be inconsistent. In such cases, this returns an error, and the
 818 * caller may assume that all locked objects joined to the transaction
 819 * have already been unlocked as if the commit had succeeded.
 820 * Do not reference the transaction structure after this call.
 821 */
 822static int
 823__xfs_trans_commit(
 824	struct xfs_trans	*tp,
 825	bool			regrant)
 826{
 827	struct xfs_mount	*mp = tp->t_mountp;
 828	struct xlog		*log = mp->m_log;
 829	xfs_csn_t		commit_seq = 0;
 830	int			error = 0;
 831	int			sync = tp->t_flags & XFS_TRANS_SYNC;
 832
 833	trace_xfs_trans_commit(tp, _RET_IP_);
 834
 835	/*
 836	 * Commit per-transaction changes that are not already tracked through
 837	 * log items.  This can add dirty log items to the transaction.
 838	 */
 839	if (tp->t_flags & XFS_TRANS_SB_DIRTY)
 840		xfs_trans_apply_sb_deltas(tp);
 841	xfs_trans_apply_dquot_deltas(tp);
 842
 843	error = xfs_trans_run_precommits(tp);
 844	if (error)
 845		goto out_unreserve;
 846
 847	/*
 848	 * If there is nothing to be logged by the transaction,
 849	 * then unlock all of the items associated with the
 850	 * transaction and free the transaction structure.
 851	 * Also make sure to return any reserved blocks to
 852	 * the free pool.
 853	 */
 854	if (!(tp->t_flags & XFS_TRANS_DIRTY))
 855		goto out_unreserve;
 856
 857	/*
 858	 * We must check against log shutdown here because we cannot abort log
 859	 * items and leave them dirty, inconsistent and unpinned in memory while
 860	 * the log is active. This leaves them open to being written back to
 861	 * disk, and that will lead to on-disk corruption.
 862	 */
 863	if (xlog_is_shutdown(log)) {
 864		error = -EIO;
 865		goto out_unreserve;
 866	}
 867
 868	ASSERT(tp->t_ticket != NULL);
 869
 870	xlog_cil_commit(log, tp, &commit_seq, regrant);
 871
 872	xfs_trans_free(tp);
 873
 874	/*
 875	 * If the transaction needs to be synchronous, then force the
 876	 * log out now and wait for it.
 877	 */
 878	if (sync) {
 879		error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);
 880		XFS_STATS_INC(mp, xs_trans_sync);
 881	} else {
 882		XFS_STATS_INC(mp, xs_trans_async);
 883	}
 884
 885	return error;
 886
 887out_unreserve:
 888	xfs_trans_unreserve_and_mod_sb(tp);
 889
 890	/*
 891	 * It is indeed possible for the transaction to be not dirty but
 892	 * the dqinfo portion to be.  All that means is that we have some
 893	 * (non-persistent) quota reservations that need to be unreserved.
 894	 */
 895	xfs_trans_unreserve_and_mod_dquots(tp, true);
 896	if (tp->t_ticket) {
 897		if (regrant && !xlog_is_shutdown(log))
 898			xfs_log_ticket_regrant(log, tp->t_ticket);
 899		else
 900			xfs_log_ticket_ungrant(log, tp->t_ticket);
 901		tp->t_ticket = NULL;
 902	}
 903	xfs_trans_free_items(tp, !!error);
 904	xfs_trans_free(tp);
 905
 906	XFS_STATS_INC(mp, xs_trans_empty);
 907	return error;
 908}
 909
 910int
 911xfs_trans_commit(
 912	struct xfs_trans	*tp)
 913{
 914	/*
 915	 * Finish deferred items on final commit. Only permanent transactions
 916	 * should ever have deferred ops.
 917	 */
 918	WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&
 919		     !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));
 920	if (tp->t_flags & XFS_TRANS_PERM_LOG_RES) {
 921		int error = xfs_defer_finish_noroll(&tp);
 922		if (error) {
 923			xfs_trans_cancel(tp);
 924			return error;
 925		}
 926	}
 927
 928	return __xfs_trans_commit(tp, false);
 929}
 930
 931/*
 932 * Unlock all of the transaction's items and free the transaction.  If the
 933 * transaction is dirty, we must shut down the filesystem because there is no
 934 * way to restore them to their previous state.
 935 *
 936 * If the transaction has made a log reservation, make sure to release it as
 937 * well.
 938 *
 939 * This is a high level function (equivalent to xfs_trans_commit()) and so can
 940 * be called after the transaction has effectively been aborted due to the mount
 941 * being shut down. However, if the mount has not been shut down and the
 942 * transaction is dirty we will shut the mount down and, in doing so, that
 943 * guarantees that the log is shut down, too. Hence we don't need to be as
 944 * careful with shutdown state and dirty items here as we need to be in
 945 * xfs_trans_commit().
 946 */
 947void
 948xfs_trans_cancel(
 949	struct xfs_trans	*tp)
 950{
 951	struct xfs_mount	*mp = tp->t_mountp;
 952	struct xlog		*log = mp->m_log;
 953	bool			dirty = (tp->t_flags & XFS_TRANS_DIRTY);
 954
 955	trace_xfs_trans_cancel(tp, _RET_IP_);
 956
 957	/*
 958	 * It's never valid to cancel a transaction with deferred ops attached,
 959	 * because the transaction is effectively dirty.  Complain about this
 960	 * loudly before freeing the in-memory defer items and shutting down the
 961	 * filesystem.
 962	 */
 963	if (!list_empty(&tp->t_dfops)) {
 964		ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
 965		dirty = true;
 966		xfs_defer_cancel(tp);
 967	}
 968
 969	/*
 970	 * See if the caller is relying on us to shut down the filesystem. We
 971	 * only want an error report if there isn't already a shutdown in
 972	 * progress, so we only need to check against the mount shutdown state
 973	 * here.
 974	 */
 975	if (dirty && !xfs_is_shutdown(mp)) {
 976		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);
 977		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
 978	}
 979#ifdef DEBUG
 980	/* Log items need to be consistent until the log is shut down. */
 981	if (!dirty && !xlog_is_shutdown(log)) {
 982		struct xfs_log_item *lip;
 983
 984		list_for_each_entry(lip, &tp->t_items, li_trans)
 985			ASSERT(!xlog_item_is_intent_done(lip));
 986	}
 987#endif
 988	xfs_trans_unreserve_and_mod_sb(tp);
 989	xfs_trans_unreserve_and_mod_dquots(tp, false);
 990
 991	if (tp->t_ticket) {
 992		xfs_log_ticket_ungrant(log, tp->t_ticket);
 993		tp->t_ticket = NULL;
 994	}
 995
 996	xfs_trans_free_items(tp, dirty);
 997	xfs_trans_free(tp);
 998}
 999
1000/*
1001 * Roll from one trans in the sequence of PERMANENT transactions to the next:
1002 * permanent transactions are only flushed out when committed with
1003 * xfs_trans_commit(), but we still want as soon as possible to let chunks of it
1004 * go to the log.  So we commit the chunk we've been working on and get a new
1005 * transaction to continue.
1006 */
1007int
1008xfs_trans_roll(
1009	struct xfs_trans	**tpp)
1010{
1011	struct xfs_trans	*tp = *tpp;
1012	unsigned int		log_res = tp->t_log_res;
1013	unsigned int		log_count = tp->t_log_count;
1014	int			error;
1015
1016	trace_xfs_trans_roll(tp, _RET_IP_);
1017
1018	ASSERT(log_res > 0);
1019
1020	/*
1021	 * Copy the critical parameters from one trans to the next.
1022	 */
1023	*tpp = xfs_trans_dup(tp);
1024
1025	/*
1026	 * Commit the current transaction.
1027	 *
1028	 * If this commit failed, then it'd just unlock those items that are not
1029	 * marked ihold. That also means that a filesystem shutdown is in
1030	 * progress.  The caller takes the responsibility to cancel the
1031	 * duplicate transaction that gets returned.
1032	 */
1033	error = __xfs_trans_commit(tp, true);
1034	if (error)
1035		return error;
1036
1037	/*
1038	 * Reserve space in the log for the next transaction.
1039	 *
1040	 * This also pushes items in the AIL out to disk if they are taking up
1041	 * space at the tail of the log that we want to use.  This requires that
1042	 * either nothing be locked across this call, or that anything that is
1043	 * locked be logged in the prior and the next transactions.
1044	 */
1045	tp = *tpp;
1046	error = xfs_log_regrant(tp->t_mountp, tp->t_ticket);
1047	if (error)
1048		return error;
1049	tp->t_log_res = log_res;
1050	tp->t_log_count = log_count;
1051	return 0;
1052}
1053
1054/*
1055 * Allocate an transaction, lock and join the inode to it, and reserve quota.
1056 *
1057 * The caller must ensure that the on-disk dquots attached to this inode have
1058 * already been allocated and initialized.  The caller is responsible for
1059 * releasing ILOCK_EXCL if a new transaction is returned.
1060 */
1061int
1062xfs_trans_alloc_inode(
1063	struct xfs_inode	*ip,
1064	struct xfs_trans_res	*resv,
1065	unsigned int		dblocks,
1066	unsigned int		rblocks,
1067	bool			force,
1068	struct xfs_trans	**tpp)
1069{
1070	struct xfs_trans	*tp;
1071	struct xfs_mount	*mp = ip->i_mount;
1072	bool			retried = false;
1073	int			error;
1074
1075retry:
1076	error = xfs_trans_alloc(mp, resv, dblocks,
1077			xfs_extlen_to_rtxlen(mp, rblocks),
1078			force ? XFS_TRANS_RESERVE : 0, &tp);
1079	if (error)
1080		return error;
1081
1082	xfs_ilock(ip, XFS_ILOCK_EXCL);
1083	xfs_trans_ijoin(tp, ip, 0);
1084
1085	error = xfs_qm_dqattach_locked(ip, false);
1086	if (error) {
1087		/* Caller should have allocated the dquots! */
1088		ASSERT(error != -ENOENT);
1089		goto out_cancel;
1090	}
1091
1092	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);
1093	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1094		xfs_trans_cancel(tp);
1095		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1096		xfs_blockgc_free_quota(ip, 0);
1097		retried = true;
1098		goto retry;
1099	}
1100	if (error)
1101		goto out_cancel;
1102
1103	*tpp = tp;
1104	return 0;
1105
1106out_cancel:
1107	xfs_trans_cancel(tp);
1108	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1109	return error;
1110}
1111
1112/*
1113 * Try to reserve more blocks for a transaction.
1114 *
1115 * This is for callers that need to attach resources to a transaction, scan
1116 * those resources to determine the space reservation requirements, and then
1117 * modify the attached resources.  In other words, online repair.  This can
1118 * fail due to ENOSPC, so the caller must be able to cancel the transaction
1119 * without shutting down the fs.
1120 */
1121int
1122xfs_trans_reserve_more(
1123	struct xfs_trans	*tp,
1124	unsigned int		blocks,
1125	unsigned int		rtextents)
1126{
1127	bool			rsvd = tp->t_flags & XFS_TRANS_RESERVE;
1128
1129	if (blocks && xfs_dec_fdblocks(tp->t_mountp, blocks, rsvd))
1130		return -ENOSPC;
1131	if (rtextents && xfs_dec_frextents(tp->t_mountp, rtextents)) {
1132		if (blocks)
1133			xfs_add_fdblocks(tp->t_mountp, blocks);
1134		return -ENOSPC;
1135	}
1136	tp->t_blk_res += blocks;
1137	tp->t_rtx_res += rtextents;
1138	return 0;
1139}
1140
1141/*
1142 * Try to reserve more blocks and file quota for a transaction.  Same
1143 * conditions of usage as xfs_trans_reserve_more.
1144 */
1145int
1146xfs_trans_reserve_more_inode(
1147	struct xfs_trans	*tp,
1148	struct xfs_inode	*ip,
1149	unsigned int		dblocks,
1150	unsigned int		rblocks,
1151	bool			force_quota)
1152{
1153	struct xfs_mount	*mp = ip->i_mount;
1154	unsigned int		rtx = xfs_extlen_to_rtxlen(mp, rblocks);
1155	int			error;
1156
1157	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
1158
1159	error = xfs_trans_reserve_more(tp, dblocks, rtx);
1160	if (error)
1161		return error;
1162
1163	if (!XFS_IS_QUOTA_ON(mp) || xfs_is_quota_inode(&mp->m_sb, ip->i_ino))
1164		return 0;
1165
1166	if (tp->t_flags & XFS_TRANS_RESERVE)
1167		force_quota = true;
1168
1169	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks,
1170			force_quota);
1171	if (!error)
1172		return 0;
1173
1174	/* Quota failed, give back the new reservation. */
1175	xfs_add_fdblocks(mp, dblocks);
1176	tp->t_blk_res -= dblocks;
1177	xfs_add_frextents(mp, rtx);
1178	tp->t_rtx_res -= rtx;
1179	return error;
1180}
1181
1182/*
1183 * Allocate an transaction in preparation for inode creation by reserving quota
1184 * against the given dquots.  Callers are not required to hold any inode locks.
1185 */
1186int
1187xfs_trans_alloc_icreate(
1188	struct xfs_mount	*mp,
1189	struct xfs_trans_res	*resv,
1190	struct xfs_dquot	*udqp,
1191	struct xfs_dquot	*gdqp,
1192	struct xfs_dquot	*pdqp,
1193	unsigned int		dblocks,
1194	struct xfs_trans	**tpp)
1195{
1196	struct xfs_trans	*tp;
1197	bool			retried = false;
1198	int			error;
1199
1200retry:
1201	error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);
1202	if (error)
1203		return error;
1204
1205	error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);
1206	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1207		xfs_trans_cancel(tp);
1208		xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1209		retried = true;
1210		goto retry;
1211	}
1212	if (error) {
1213		xfs_trans_cancel(tp);
1214		return error;
1215	}
1216
1217	*tpp = tp;
1218	return 0;
1219}
1220
1221/*
1222 * Allocate an transaction, lock and join the inode to it, and reserve quota
1223 * in preparation for inode attribute changes that include uid, gid, or prid
1224 * changes.
1225 *
1226 * The caller must ensure that the on-disk dquots attached to this inode have
1227 * already been allocated and initialized.  The ILOCK will be dropped when the
1228 * transaction is committed or cancelled.
1229 */
1230int
1231xfs_trans_alloc_ichange(
1232	struct xfs_inode	*ip,
1233	struct xfs_dquot	*new_udqp,
1234	struct xfs_dquot	*new_gdqp,
1235	struct xfs_dquot	*new_pdqp,
1236	bool			force,
1237	struct xfs_trans	**tpp)
1238{
1239	struct xfs_trans	*tp;
1240	struct xfs_mount	*mp = ip->i_mount;
1241	struct xfs_dquot	*udqp;
1242	struct xfs_dquot	*gdqp;
1243	struct xfs_dquot	*pdqp;
1244	bool			retried = false;
1245	int			error;
1246
1247retry:
1248	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1249	if (error)
1250		return error;
1251
1252	xfs_ilock(ip, XFS_ILOCK_EXCL);
1253	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1254
1255	if (xfs_is_metadir_inode(ip))
1256		goto out;
1257
1258	error = xfs_qm_dqattach_locked(ip, false);
1259	if (error) {
1260		/* Caller should have allocated the dquots! */
1261		ASSERT(error != -ENOENT);
1262		goto out_cancel;
1263	}
1264
1265	/*
1266	 * For each quota type, skip quota reservations if the inode's dquots
1267	 * now match the ones that came from the caller, or the caller didn't
1268	 * pass one in.  The inode's dquots can change if we drop the ILOCK to
1269	 * perform a blockgc scan, so we must preserve the caller's arguments.
1270	 */
1271	udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;
1272	gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;
1273	pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;
1274	if (udqp || gdqp || pdqp) {
1275		xfs_filblks_t	dblocks, rblocks;
1276		unsigned int	qflags = XFS_QMOPT_RES_REGBLKS;
1277		bool		isrt = XFS_IS_REALTIME_INODE(ip);
1278
1279		if (force)
1280			qflags |= XFS_QMOPT_FORCE_RES;
1281
1282		if (isrt) {
1283			error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1284			if (error)
1285				goto out_cancel;
1286		}
1287
1288		xfs_inode_count_blocks(tp, ip, &dblocks, &rblocks);
1289
1290		if (isrt)
1291			rblocks += ip->i_delayed_blks;
1292		else
1293			dblocks += ip->i_delayed_blks;
1294
1295		/*
1296		 * Reserve enough quota to handle blocks on disk and reserved
1297		 * for a delayed allocation.  We'll actually transfer the
1298		 * delalloc reservation between dquots at chown time, even
1299		 * though that part is only semi-transactional.
1300		 */
1301		error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1302				pdqp, dblocks, 1, qflags);
1303		if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1304			xfs_trans_cancel(tp);
1305			xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1306			retried = true;
1307			goto retry;
1308		}
1309		if (error)
1310			goto out_cancel;
1311
1312		/* Do the same for realtime. */
1313		qflags = XFS_QMOPT_RES_RTBLKS | (qflags & XFS_QMOPT_FORCE_RES);
1314		error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,
1315				pdqp, rblocks, 0, qflags);
1316		if ((error == -EDQUOT || error == -ENOSPC) && !retried) {
1317			xfs_trans_cancel(tp);
1318			xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);
1319			retried = true;
1320			goto retry;
1321		}
1322		if (error)
1323			goto out_cancel;
1324	}
1325
1326out:
1327	*tpp = tp;
1328	return 0;
1329
1330out_cancel:
1331	xfs_trans_cancel(tp);
1332	return error;
1333}
1334
1335/*
1336 * Allocate an transaction, lock and join the directory and child inodes to it,
1337 * and reserve quota for a directory update.  If there isn't sufficient space,
1338 * @dblocks will be set to zero for a reservationless directory update and
1339 * @nospace_error will be set to a negative errno describing the space
1340 * constraint we hit.
1341 *
1342 * The caller must ensure that the on-disk dquots attached to this inode have
1343 * already been allocated and initialized.  The ILOCKs will be dropped when the
1344 * transaction is committed or cancelled.
1345 *
1346 * Caller is responsible for unlocking the inodes manually upon return
1347 */
1348int
1349xfs_trans_alloc_dir(
1350	struct xfs_inode	*dp,
1351	struct xfs_trans_res	*resv,
1352	struct xfs_inode	*ip,
1353	unsigned int		*dblocks,
1354	struct xfs_trans	**tpp,
1355	int			*nospace_error)
1356{
1357	struct xfs_trans	*tp;
1358	struct xfs_mount	*mp = ip->i_mount;
1359	unsigned int		resblks;
1360	bool			retried = false;
1361	int			error;
1362
1363retry:
1364	*nospace_error = 0;
1365	resblks = *dblocks;
1366	error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1367	if (error == -ENOSPC) {
1368		*nospace_error = error;
1369		resblks = 0;
1370		error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);
1371	}
1372	if (error)
1373		return error;
1374
1375	xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
1376
1377	xfs_trans_ijoin(tp, dp, 0);
1378	xfs_trans_ijoin(tp, ip, 0);
1379
1380	error = xfs_qm_dqattach_locked(dp, false);
1381	if (error) {
1382		/* Caller should have allocated the dquots! */
1383		ASSERT(error != -ENOENT);
1384		goto out_cancel;
1385	}
1386
1387	error = xfs_qm_dqattach_locked(ip, false);
1388	if (error) {
1389		/* Caller should have allocated the dquots! */
1390		ASSERT(error != -ENOENT);
1391		goto out_cancel;
1392	}
1393
1394	if (resblks == 0)
1395		goto done;
1396
1397	error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false);
1398	if (error == -EDQUOT || error == -ENOSPC) {
1399		if (!retried) {
1400			xfs_trans_cancel(tp);
1401			xfs_iunlock(dp, XFS_ILOCK_EXCL);
1402			if (dp != ip)
1403				xfs_iunlock(ip, XFS_ILOCK_EXCL);
1404			xfs_blockgc_free_quota(dp, 0);
1405			retried = true;
1406			goto retry;
1407		}
1408
1409		*nospace_error = error;
1410		resblks = 0;
1411		error = 0;
1412	}
1413	if (error)
1414		goto out_cancel;
1415
1416done:
1417	*tpp = tp;
1418	*dblocks = resblks;
1419	return 0;
1420
1421out_cancel:
1422	xfs_trans_cancel(tp);
1423	xfs_iunlock(dp, XFS_ILOCK_EXCL);
1424	if (dp != ip)
1425		xfs_iunlock(ip, XFS_ILOCK_EXCL);
1426	return error;
1427}