brintos

brintos / linux-shallow public Read only

0
0
Text · 37.3 KiB · bdf3704 Raw
1387 lines · c
1// SPDX-License-Identifier: GPL-2.02/*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 29struct kmem_cache	*xfs_trans_cache;30 31#if defined(CONFIG_TRACEPOINTS)32static void33xfs_trans_trace_reservations(34	struct xfs_mount	*mp)35{36	struct xfs_trans_res	*res;37	struct xfs_trans_res	*end_res;38	int			i;39 40	res = (struct xfs_trans_res *)M_RES(mp);41	end_res = (struct xfs_trans_res *)(M_RES(mp) + 1);42	for (i = 0; res < end_res; i++, res++)43		trace_xfs_trans_resv_calc(mp, i, res);44}45#else46# define xfs_trans_trace_reservations(mp)47#endif48 49/*50 * Initialize the precomputed transaction reservation values51 * in the mount structure.52 */53void54xfs_trans_init(55	struct xfs_mount	*mp)56{57	xfs_trans_resv_calc(mp, M_RES(mp));58	xfs_trans_trace_reservations(mp);59}60 61/*62 * Free the transaction structure.  If there is more clean up63 * to do when the structure is freed, add it here.64 */65STATIC void66xfs_trans_free(67	struct xfs_trans	*tp)68{69	xfs_extent_busy_sort(&tp->t_busy);70	xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);71 72	trace_xfs_trans_free(tp, _RET_IP_);73	xfs_trans_clear_context(tp);74	if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))75		sb_end_intwrite(tp->t_mountp->m_super);76	xfs_trans_free_dqinfo(tp);77	kmem_cache_free(xfs_trans_cache, tp);78}79 80/*81 * This is called to create a new transaction which will share the82 * permanent log reservation of the given transaction.  The remaining83 * unused block and rt extent reservations are also inherited.  This84 * implies that the original transaction is no longer allowed to allocate85 * blocks.  Locks and log items, however, are no inherited.  They must86 * be added to the new transaction explicitly.87 */88STATIC struct xfs_trans *89xfs_trans_dup(90	struct xfs_trans	*tp)91{92	struct xfs_trans	*ntp;93 94	trace_xfs_trans_dup(tp, _RET_IP_);95 96	ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);97 98	/*99	 * Initialize the new transaction structure.100	 */101	ntp->t_magic = XFS_TRANS_HEADER_MAGIC;102	ntp->t_mountp = tp->t_mountp;103	INIT_LIST_HEAD(&ntp->t_items);104	INIT_LIST_HEAD(&ntp->t_busy);105	INIT_LIST_HEAD(&ntp->t_dfops);106	ntp->t_highest_agno = NULLAGNUMBER;107 108	ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);109	ASSERT(tp->t_ticket != NULL);110 111	ntp->t_flags = XFS_TRANS_PERM_LOG_RES |112		       (tp->t_flags & XFS_TRANS_RESERVE) |113		       (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) |114		       (tp->t_flags & XFS_TRANS_RES_FDBLKS);115	/* We gave our writer reference to the new transaction */116	tp->t_flags |= XFS_TRANS_NO_WRITECOUNT;117	ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket);118 119	ASSERT(tp->t_blk_res >= tp->t_blk_res_used);120	ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used;121	tp->t_blk_res = tp->t_blk_res_used;122 123	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;124	tp->t_rtx_res = tp->t_rtx_res_used;125 126	xfs_trans_switch_context(tp, ntp);127 128	/* move deferred ops over to the new tp */129	xfs_defer_move(ntp, tp);130 131	xfs_trans_dup_dqinfo(tp, ntp);132	return ntp;133}134 135/*136 * This is called to reserve free disk blocks and log space for the137 * given transaction.  This must be done before allocating any resources138 * 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 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which143 * is used by long running transactions.  If any one of the reservations144 * fails then they will all be backed out.145 *146 * This does not do quota reservations. That typically is done by the147 * caller afterwards.148 */149static int150xfs_trans_reserve(151	struct xfs_trans	*tp,152	struct xfs_trans_res	*resp,153	uint			blocks,154	uint			rtextents)155{156	struct xfs_mount	*mp = tp->t_mountp;157	int			error = 0;158	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;159 160	/*161	 * Attempt to reserve the needed disk blocks by decrementing162	 * the number needed from the number available.  This will163	 * fail if the count would go below zero.164	 */165	if (blocks > 0) {166		error = xfs_dec_fdblocks(mp, blocks, rsvd);167		if (error != 0)168			return -ENOSPC;169		tp->t_blk_res += blocks;170	}171 172	/*173	 * Reserve the log space needed for this transaction.174	 */175	if (resp->tr_logres > 0) {176		bool	permanent = false;177 178		ASSERT(tp->t_log_res == 0 ||179		       tp->t_log_res == resp->tr_logres);180		ASSERT(tp->t_log_count == 0 ||181		       tp->t_log_count == resp->tr_logcount);182 183		if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) {184			tp->t_flags |= XFS_TRANS_PERM_LOG_RES;185			permanent = true;186		} else {187			ASSERT(tp->t_ticket == NULL);188			ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES));189		}190 191		if (tp->t_ticket != NULL) {192			ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES);193			error = xfs_log_regrant(mp, tp->t_ticket);194		} else {195			error = xfs_log_reserve(mp, resp->tr_logres,196						resp->tr_logcount,197						&tp->t_ticket, permanent);198		}199 200		if (error)201			goto undo_blocks;202 203		tp->t_log_res = resp->tr_logres;204		tp->t_log_count = resp->tr_logcount;205	}206 207	/*208	 * Attempt to reserve the needed realtime extents by decrementing209	 * the number needed from the number available.  This will210	 * fail if the count would go below zero.211	 */212	if (rtextents > 0) {213		error = xfs_dec_frextents(mp, rtextents);214		if (error) {215			error = -ENOSPC;216			goto undo_log;217		}218		tp->t_rtx_res += rtextents;219	}220 221	return 0;222 223	/*224	 * Error cases jump to one of these labels to undo any225	 * reservations which have already been performed.226	 */227undo_log:228	if (resp->tr_logres > 0) {229		xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);230		tp->t_ticket = NULL;231		tp->t_log_res = 0;232		tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES;233	}234 235undo_blocks:236	if (blocks > 0) {237		xfs_add_fdblocks(mp, blocks);238		tp->t_blk_res = 0;239	}240	return error;241}242 243int244xfs_trans_alloc(245	struct xfs_mount	*mp,246	struct xfs_trans_res	*resp,247	uint			blocks,248	uint			rtextents,249	uint			flags,250	struct xfs_trans	**tpp)251{252	struct xfs_trans	*tp;253	bool			want_retry = true;254	int			error;255 256	/*257	 * Allocate the handle before we do our freeze accounting and setting up258	 * GFP_NOFS allocation context so that we avoid lockdep false positives259	 * by doing GFP_KERNEL allocations inside sb_start_intwrite().260	 */261retry:262	tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL);263	if (!(flags & XFS_TRANS_NO_WRITECOUNT))264		sb_start_intwrite(mp->m_super);265	xfs_trans_set_context(tp);266 267	/*268	 * Zero-reservation ("empty") transactions can't modify anything, so269	 * they're allowed to run while we're frozen.270	 */271	WARN_ON(resp->tr_logres > 0 &&272		mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE);273	ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) ||274	       xfs_has_lazysbcount(mp));275 276	tp->t_magic = XFS_TRANS_HEADER_MAGIC;277	tp->t_flags = flags;278	tp->t_mountp = mp;279	INIT_LIST_HEAD(&tp->t_items);280	INIT_LIST_HEAD(&tp->t_busy);281	INIT_LIST_HEAD(&tp->t_dfops);282	tp->t_highest_agno = NULLAGNUMBER;283 284	error = xfs_trans_reserve(tp, resp, blocks, rtextents);285	if (error == -ENOSPC && want_retry) {286		xfs_trans_cancel(tp);287 288		/*289		 * We weren't able to reserve enough space for the transaction.290		 * Flush the other speculative space allocations to free space.291		 * Do not perform a synchronous scan because callers can hold292		 * other locks.293		 */294		error = xfs_blockgc_flush_all(mp);295		if (error)296			return error;297		want_retry = false;298		goto retry;299	}300	if (error) {301		xfs_trans_cancel(tp);302		return error;303	}304 305	trace_xfs_trans_alloc(tp, _RET_IP_);306 307	*tpp = tp;308	return 0;309}310 311/*312 * Create an empty transaction with no reservation.  This is a defensive313 * mechanism for routines that query metadata without actually modifying them --314 * if the metadata being queried is somehow cross-linked (think a btree block315 * pointer that points higher in the tree), we risk deadlock.  However, blocks316 * grabbed as part of a transaction can be re-grabbed.  The verifiers will317 * notice the corrupt block and the operation will fail back to userspace318 * without deadlocking.319 *320 * Note the zero-length reservation; this transaction MUST be cancelled without321 * any dirty data.322 *323 * Callers should obtain freeze protection to avoid a conflict with fs freezing324 * where we can be grabbing buffers at the same time that freeze is trying to325 * drain the buffer LRU list.326 */327int328xfs_trans_alloc_empty(329	struct xfs_mount		*mp,330	struct xfs_trans		**tpp)331{332	struct xfs_trans_res		resv = {0};333 334	return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp);335}336 337/*338 * Record the indicated change to the given field for application339 * to the file system's superblock when the transaction commits.340 * For now, just store the change in the transaction structure.341 *342 * Mark the transaction structure to indicate that the superblock343 * needs to be updated before committing.344 *345 * Because we may not be keeping track of allocated/free inodes and346 * used filesystem blocks in the superblock, we do not mark the347 * superblock dirty in this transaction if we modify these fields.348 * We still need to update the transaction deltas so that they get349 * applied to the incore superblock, but we don't want them to350 * cause the superblock to get locked and logged if these are the351 * only fields in the superblock that the transaction modifies.352 */353void354xfs_trans_mod_sb(355	xfs_trans_t	*tp,356	uint		field,357	int64_t		delta)358{359	uint32_t	flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY);360	xfs_mount_t	*mp = tp->t_mountp;361 362	switch (field) {363	case XFS_TRANS_SB_ICOUNT:364		tp->t_icount_delta += delta;365		if (xfs_has_lazysbcount(mp))366			flags &= ~XFS_TRANS_SB_DIRTY;367		break;368	case XFS_TRANS_SB_IFREE:369		tp->t_ifree_delta += delta;370		if (xfs_has_lazysbcount(mp))371			flags &= ~XFS_TRANS_SB_DIRTY;372		break;373	case XFS_TRANS_SB_FDBLOCKS:374		/*375		 * Track the number of blocks allocated in the transaction.376		 * Make sure it does not exceed the number reserved. If so,377		 * shutdown as this can lead to accounting inconsistency.378		 */379		if (delta < 0) {380			tp->t_blk_res_used += (uint)-delta;381			if (tp->t_blk_res_used > tp->t_blk_res)382				xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);383		} else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) {384			int64_t	blkres_delta;385 386			/*387			 * Return freed blocks directly to the reservation388			 * instead of the global pool, being careful not to389			 * overflow the trans counter. This is used to preserve390			 * reservation across chains of transaction rolls that391			 * repeatedly free and allocate blocks.392			 */393			blkres_delta = min_t(int64_t, delta,394					     UINT_MAX - tp->t_blk_res);395			tp->t_blk_res += blkres_delta;396			delta -= blkres_delta;397		}398		tp->t_fdblocks_delta += delta;399		if (xfs_has_lazysbcount(mp))400			flags &= ~XFS_TRANS_SB_DIRTY;401		break;402	case XFS_TRANS_SB_RES_FDBLOCKS:403		/*404		 * The allocation has already been applied to the405		 * in-core superblock's counter.  This should only406		 * be applied to the on-disk superblock.407		 */408		tp->t_res_fdblocks_delta += delta;409		if (xfs_has_lazysbcount(mp))410			flags &= ~XFS_TRANS_SB_DIRTY;411		break;412	case XFS_TRANS_SB_FREXTENTS:413		/*414		 * Track the number of blocks allocated in the415		 * transaction.  Make sure it does not exceed the416		 * number reserved.417		 */418		if (delta < 0) {419			tp->t_rtx_res_used += (uint)-delta;420			ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res);421		}422		tp->t_frextents_delta += delta;423		break;424	case XFS_TRANS_SB_RES_FREXTENTS:425		/*426		 * The allocation has already been applied to the427		 * in-core superblock's counter.  This should only428		 * be applied to the on-disk superblock.429		 */430		ASSERT(delta < 0);431		tp->t_res_frextents_delta += delta;432		break;433	case XFS_TRANS_SB_DBLOCKS:434		tp->t_dblocks_delta += delta;435		break;436	case XFS_TRANS_SB_AGCOUNT:437		ASSERT(delta > 0);438		tp->t_agcount_delta += delta;439		break;440	case XFS_TRANS_SB_IMAXPCT:441		tp->t_imaxpct_delta += delta;442		break;443	case XFS_TRANS_SB_REXTSIZE:444		tp->t_rextsize_delta += delta;445		break;446	case XFS_TRANS_SB_RBMBLOCKS:447		tp->t_rbmblocks_delta += delta;448		break;449	case XFS_TRANS_SB_RBLOCKS:450		tp->t_rblocks_delta += delta;451		break;452	case XFS_TRANS_SB_REXTENTS:453		tp->t_rextents_delta += delta;454		break;455	case XFS_TRANS_SB_REXTSLOG:456		tp->t_rextslog_delta += delta;457		break;458	default:459		ASSERT(0);460		return;461	}462 463	tp->t_flags |= flags;464}465 466/*467 * xfs_trans_apply_sb_deltas() is called from the commit code468 * to bring the superblock buffer into the current transaction469 * and modify it as requested by earlier calls to xfs_trans_mod_sb().470 *471 * For now we just look at each field allowed to change and change472 * it if necessary.473 */474STATIC void475xfs_trans_apply_sb_deltas(476	xfs_trans_t	*tp)477{478	struct xfs_dsb	*sbp;479	struct xfs_buf	*bp;480	int		whole = 0;481 482	bp = xfs_trans_getsb(tp);483	sbp = bp->b_addr;484 485	/*486	 * Only update the superblock counters if we are logging them487	 */488	if (!xfs_has_lazysbcount((tp->t_mountp))) {489		if (tp->t_icount_delta)490			be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta);491		if (tp->t_ifree_delta)492			be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta);493		if (tp->t_fdblocks_delta)494			be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta);495		if (tp->t_res_fdblocks_delta)496			be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta);497	}498 499	/*500	 * Updating frextents requires careful handling because it does not501	 * behave like the lazysb counters because we cannot rely on log502	 * recovery in older kenels to recompute the value from the rtbitmap.503	 * This means that the ondisk frextents must be consistent with the504	 * rtbitmap.505	 *506	 * Therefore, log the frextents change to the ondisk superblock and507	 * update the incore superblock so that future calls to xfs_log_sb508	 * write the correct value ondisk.509	 *510	 * Don't touch m_frextents because it includes incore reservations,511	 * and those are handled by the unreserve function.512	 */513	if (tp->t_frextents_delta || tp->t_res_frextents_delta) {514		struct xfs_mount	*mp = tp->t_mountp;515		int64_t			rtxdelta;516 517		rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta;518 519		spin_lock(&mp->m_sb_lock);520		be64_add_cpu(&sbp->sb_frextents, rtxdelta);521		mp->m_sb.sb_frextents += rtxdelta;522		spin_unlock(&mp->m_sb_lock);523	}524 525	if (tp->t_dblocks_delta) {526		be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta);527		whole = 1;528	}529	if (tp->t_agcount_delta) {530		be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta);531		whole = 1;532	}533	if (tp->t_imaxpct_delta) {534		sbp->sb_imax_pct += tp->t_imaxpct_delta;535		whole = 1;536	}537	if (tp->t_rextsize_delta) {538		be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta);539		whole = 1;540	}541	if (tp->t_rbmblocks_delta) {542		be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta);543		whole = 1;544	}545	if (tp->t_rblocks_delta) {546		be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta);547		whole = 1;548	}549	if (tp->t_rextents_delta) {550		be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta);551		whole = 1;552	}553	if (tp->t_rextslog_delta) {554		sbp->sb_rextslog += tp->t_rextslog_delta;555		whole = 1;556	}557 558	xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF);559	if (whole)560		/*561		 * Log the whole thing, the fields are noncontiguous.562		 */563		xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1);564	else565		/*566		 * Since all the modifiable fields are contiguous, we567		 * can get away with this.568		 */569		xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount),570				  offsetof(struct xfs_dsb, sb_frextents) +571				  sizeof(sbp->sb_frextents) - 1);572}573 574/*575 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and576 * apply superblock counter changes to the in-core superblock.  The577 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT578 * applied to the in-core superblock.  The idea is that that has already been579 * done.580 *581 * If we are not logging superblock counters, then the inode allocated/free and582 * used block counts are not updated in the on disk superblock. In this case,583 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we584 * still need to update the incore superblock with the changes.585 *586 * Deltas for the inode count are +/-64, hence we use a large batch size of 128587 * so we don't need to take the counter lock on every update.588 */589#define XFS_ICOUNT_BATCH	128590 591void592xfs_trans_unreserve_and_mod_sb(593	struct xfs_trans	*tp)594{595	struct xfs_mount	*mp = tp->t_mountp;596	int64_t			blkdelta = tp->t_blk_res;597	int64_t			rtxdelta = tp->t_rtx_res;598	int64_t			idelta = 0;599	int64_t			ifreedelta = 0;600 601	/*602	 * Calculate the deltas.603	 *604	 * t_fdblocks_delta and t_frextents_delta can be positive or negative:605	 *606	 *  - positive values indicate blocks freed in the transaction.607	 *  - negative values indicate blocks allocated in the transaction608	 *609	 * Negative values can only happen if the transaction has a block610	 * reservation that covers the allocated block.  The end result is611	 * that the calculated delta values must always be positive and we612	 * can only put back previous allocated or reserved blocks here.613	 */614	ASSERT(tp->t_blk_res || tp->t_fdblocks_delta >= 0);615	if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {616	        blkdelta += tp->t_fdblocks_delta;617		ASSERT(blkdelta >= 0);618	}619 620	ASSERT(tp->t_rtx_res || tp->t_frextents_delta >= 0);621	if (tp->t_flags & XFS_TRANS_SB_DIRTY) {622		rtxdelta += tp->t_frextents_delta;623		ASSERT(rtxdelta >= 0);624	}625 626	if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) {627		idelta = tp->t_icount_delta;628		ifreedelta = tp->t_ifree_delta;629	}630 631	/* apply the per-cpu counters */632	if (blkdelta)633		xfs_add_fdblocks(mp, blkdelta);634 635	if (idelta)636		percpu_counter_add_batch(&mp->m_icount, idelta,637					 XFS_ICOUNT_BATCH);638 639	if (ifreedelta)640		percpu_counter_add(&mp->m_ifree, ifreedelta);641 642	if (rtxdelta)643		xfs_add_frextents(mp, rtxdelta);644 645	if (!(tp->t_flags & XFS_TRANS_SB_DIRTY))646		return;647 648	/* apply remaining deltas */649	spin_lock(&mp->m_sb_lock);650	mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta;651	mp->m_sb.sb_icount += idelta;652	mp->m_sb.sb_ifree += ifreedelta;653	/*654	 * Do not touch sb_frextents here because we are dealing with incore655	 * reservation.  sb_frextents is not part of the lazy sb counters so it656	 * must be consistent with the ondisk rtbitmap and must never include657	 * incore reservations.658	 */659	mp->m_sb.sb_dblocks += tp->t_dblocks_delta;660	mp->m_sb.sb_agcount += tp->t_agcount_delta;661	mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta;662	mp->m_sb.sb_rextsize += tp->t_rextsize_delta;663	if (tp->t_rextsize_delta) {664		mp->m_rtxblklog = log2_if_power2(mp->m_sb.sb_rextsize);665		mp->m_rtxblkmask = mask64_if_power2(mp->m_sb.sb_rextsize);666	}667	mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta;668	mp->m_sb.sb_rblocks += tp->t_rblocks_delta;669	mp->m_sb.sb_rextents += tp->t_rextents_delta;670	mp->m_sb.sb_rextslog += tp->t_rextslog_delta;671	spin_unlock(&mp->m_sb_lock);672 673	/*674	 * Debug checks outside of the spinlock so they don't lock up the675	 * machine if they fail.676	 */677	ASSERT(mp->m_sb.sb_imax_pct >= 0);678	ASSERT(mp->m_sb.sb_rextslog >= 0);679}680 681/* Add the given log item to the transaction's list of log items. */682void683xfs_trans_add_item(684	struct xfs_trans	*tp,685	struct xfs_log_item	*lip)686{687	ASSERT(lip->li_log == tp->t_mountp->m_log);688	ASSERT(lip->li_ailp == tp->t_mountp->m_ail);689	ASSERT(list_empty(&lip->li_trans));690	ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags));691 692	list_add_tail(&lip->li_trans, &tp->t_items);693	trace_xfs_trans_add_item(tp, _RET_IP_);694}695 696/*697 * Unlink the log item from the transaction. the log item is no longer698 * considered dirty in this transaction, as the linked transaction has699 * finished, either by abort or commit completion.700 */701void702xfs_trans_del_item(703	struct xfs_log_item	*lip)704{705	clear_bit(XFS_LI_DIRTY, &lip->li_flags);706	list_del_init(&lip->li_trans);707}708 709/* Detach and unlock all of the items in a transaction */710static void711xfs_trans_free_items(712	struct xfs_trans	*tp,713	bool			abort)714{715	struct xfs_log_item	*lip, *next;716 717	trace_xfs_trans_free_items(tp, _RET_IP_);718 719	list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) {720		xfs_trans_del_item(lip);721		if (abort)722			set_bit(XFS_LI_ABORTED, &lip->li_flags);723		if (lip->li_ops->iop_release)724			lip->li_ops->iop_release(lip);725	}726}727 728/*729 * Sort transaction items prior to running precommit operations. This will730 * attempt to order the items such that they will always be locked in the same731 * order. Items that have no sort function are moved to the end of the list732 * and so are locked last.733 *734 * This may need refinement as different types of objects add sort functions.735 *736 * Function is more complex than it needs to be because we are comparing 64 bit737 * values and the function only returns 32 bit values.738 */739static int740xfs_trans_precommit_sort(741	void			*unused_arg,742	const struct list_head	*a,743	const struct list_head	*b)744{745	struct xfs_log_item	*lia = container_of(a,746					struct xfs_log_item, li_trans);747	struct xfs_log_item	*lib = container_of(b,748					struct xfs_log_item, li_trans);749	int64_t			diff;750 751	/*752	 * If both items are non-sortable, leave them alone. If only one is753	 * sortable, move the non-sortable item towards the end of the list.754	 */755	if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort)756		return 0;757	if (!lia->li_ops->iop_sort)758		return 1;759	if (!lib->li_ops->iop_sort)760		return -1;761 762	diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib);763	if (diff < 0)764		return -1;765	if (diff > 0)766		return 1;767	return 0;768}769 770/*771 * Run transaction precommit functions.772 *773 * If there is an error in any of the callouts, then stop immediately and774 * trigger a shutdown to abort the transaction. There is no recovery possible775 * from errors at this point as the transaction is dirty....776 */777static int778xfs_trans_run_precommits(779	struct xfs_trans	*tp)780{781	struct xfs_mount	*mp = tp->t_mountp;782	struct xfs_log_item	*lip, *n;783	int			error = 0;784 785	/*786	 * Sort the item list to avoid ABBA deadlocks with other transactions787	 * running precommit operations that lock multiple shared items such as788	 * inode cluster buffers.789	 */790	list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort);791 792	/*793	 * Precommit operations can remove the log item from the transaction794	 * if the log item exists purely to delay modifications until they795	 * can be ordered against other operations. Hence we have to use796	 * list_for_each_entry_safe() here.797	 */798	list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) {799		if (!test_bit(XFS_LI_DIRTY, &lip->li_flags))800			continue;801		if (lip->li_ops->iop_precommit) {802			error = lip->li_ops->iop_precommit(tp, lip);803			if (error)804				break;805		}806	}807	if (error)808		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);809	return error;810}811 812/*813 * Commit the given transaction to the log.814 *815 * XFS disk error handling mechanism is not based on a typical816 * transaction abort mechanism. Logically after the filesystem817 * gets marked 'SHUTDOWN', we can't let any new transactions818 * be durable - ie. committed to disk - because some metadata might819 * be inconsistent. In such cases, this returns an error, and the820 * caller may assume that all locked objects joined to the transaction821 * have already been unlocked as if the commit had succeeded.822 * Do not reference the transaction structure after this call.823 */824static int825__xfs_trans_commit(826	struct xfs_trans	*tp,827	bool			regrant)828{829	struct xfs_mount	*mp = tp->t_mountp;830	struct xlog		*log = mp->m_log;831	xfs_csn_t		commit_seq = 0;832	int			error = 0;833	int			sync = tp->t_flags & XFS_TRANS_SYNC;834 835	trace_xfs_trans_commit(tp, _RET_IP_);836 837	error = xfs_trans_run_precommits(tp);838	if (error) {839		if (tp->t_flags & XFS_TRANS_PERM_LOG_RES)840			xfs_defer_cancel(tp);841		goto out_unreserve;842	}843 844	/*845	 * Finish deferred items on final commit. Only permanent transactions846	 * should ever have deferred ops.847	 */848	WARN_ON_ONCE(!list_empty(&tp->t_dfops) &&849		     !(tp->t_flags & XFS_TRANS_PERM_LOG_RES));850	if (!regrant && (tp->t_flags & XFS_TRANS_PERM_LOG_RES)) {851		error = xfs_defer_finish_noroll(&tp);852		if (error)853			goto out_unreserve;854 855		/* Run precommits from final tx in defer chain. */856		error = xfs_trans_run_precommits(tp);857		if (error)858			goto out_unreserve;859	}860 861	/*862	 * If there is nothing to be logged by the transaction,863	 * then unlock all of the items associated with the864	 * transaction and free the transaction structure.865	 * Also make sure to return any reserved blocks to866	 * the free pool.867	 */868	if (!(tp->t_flags & XFS_TRANS_DIRTY))869		goto out_unreserve;870 871	/*872	 * We must check against log shutdown here because we cannot abort log873	 * items and leave them dirty, inconsistent and unpinned in memory while874	 * the log is active. This leaves them open to being written back to875	 * disk, and that will lead to on-disk corruption.876	 */877	if (xlog_is_shutdown(log)) {878		error = -EIO;879		goto out_unreserve;880	}881 882	ASSERT(tp->t_ticket != NULL);883 884	/*885	 * If we need to update the superblock, then do it now.886	 */887	if (tp->t_flags & XFS_TRANS_SB_DIRTY)888		xfs_trans_apply_sb_deltas(tp);889	xfs_trans_apply_dquot_deltas(tp);890 891	xlog_cil_commit(log, tp, &commit_seq, regrant);892 893	xfs_trans_free(tp);894 895	/*896	 * If the transaction needs to be synchronous, then force the897	 * log out now and wait for it.898	 */899	if (sync) {900		error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL);901		XFS_STATS_INC(mp, xs_trans_sync);902	} else {903		XFS_STATS_INC(mp, xs_trans_async);904	}905 906	return error;907 908out_unreserve:909	xfs_trans_unreserve_and_mod_sb(tp);910 911	/*912	 * It is indeed possible for the transaction to be not dirty but913	 * the dqinfo portion to be.  All that means is that we have some914	 * (non-persistent) quota reservations that need to be unreserved.915	 */916	xfs_trans_unreserve_and_mod_dquots(tp);917	if (tp->t_ticket) {918		if (regrant && !xlog_is_shutdown(log))919			xfs_log_ticket_regrant(log, tp->t_ticket);920		else921			xfs_log_ticket_ungrant(log, tp->t_ticket);922		tp->t_ticket = NULL;923	}924	xfs_trans_free_items(tp, !!error);925	xfs_trans_free(tp);926 927	XFS_STATS_INC(mp, xs_trans_empty);928	return error;929}930 931int932xfs_trans_commit(933	struct xfs_trans	*tp)934{935	return __xfs_trans_commit(tp, false);936}937 938/*939 * Unlock all of the transaction's items and free the transaction.  If the940 * transaction is dirty, we must shut down the filesystem because there is no941 * way to restore them to their previous state.942 *943 * If the transaction has made a log reservation, make sure to release it as944 * well.945 *946 * This is a high level function (equivalent to xfs_trans_commit()) and so can947 * be called after the transaction has effectively been aborted due to the mount948 * being shut down. However, if the mount has not been shut down and the949 * transaction is dirty we will shut the mount down and, in doing so, that950 * guarantees that the log is shut down, too. Hence we don't need to be as951 * careful with shutdown state and dirty items here as we need to be in952 * xfs_trans_commit().953 */954void955xfs_trans_cancel(956	struct xfs_trans	*tp)957{958	struct xfs_mount	*mp = tp->t_mountp;959	struct xlog		*log = mp->m_log;960	bool			dirty = (tp->t_flags & XFS_TRANS_DIRTY);961 962	trace_xfs_trans_cancel(tp, _RET_IP_);963 964	/*965	 * It's never valid to cancel a transaction with deferred ops attached,966	 * because the transaction is effectively dirty.  Complain about this967	 * loudly before freeing the in-memory defer items and shutting down the968	 * filesystem.969	 */970	if (!list_empty(&tp->t_dfops)) {971		ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);972		dirty = true;973		xfs_defer_cancel(tp);974	}975 976	/*977	 * See if the caller is relying on us to shut down the filesystem. We978	 * only want an error report if there isn't already a shutdown in979	 * progress, so we only need to check against the mount shutdown state980	 * here.981	 */982	if (dirty && !xfs_is_shutdown(mp)) {983		XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp);984		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);985	}986#ifdef DEBUG987	/* Log items need to be consistent until the log is shut down. */988	if (!dirty && !xlog_is_shutdown(log)) {989		struct xfs_log_item *lip;990 991		list_for_each_entry(lip, &tp->t_items, li_trans)992			ASSERT(!xlog_item_is_intent_done(lip));993	}994#endif995	xfs_trans_unreserve_and_mod_sb(tp);996	xfs_trans_unreserve_and_mod_dquots(tp);997 998	if (tp->t_ticket) {999		xfs_log_ticket_ungrant(log, tp->t_ticket);1000		tp->t_ticket = NULL;1001	}1002 1003	xfs_trans_free_items(tp, dirty);1004	xfs_trans_free(tp);1005}1006 1007/*1008 * Roll from one trans in the sequence of PERMANENT transactions to1009 * the next: permanent transactions are only flushed out when1010 * committed with xfs_trans_commit(), but we still want as soon1011 * as possible to let chunks of it go to the log. So we commit the1012 * chunk we've been working on and get a new transaction to continue.1013 */1014int1015xfs_trans_roll(1016	struct xfs_trans	**tpp)1017{1018	struct xfs_trans	*trans = *tpp;1019	struct xfs_trans_res	tres;1020	int			error;1021 1022	trace_xfs_trans_roll(trans, _RET_IP_);1023 1024	/*1025	 * Copy the critical parameters from one trans to the next.1026	 */1027	tres.tr_logres = trans->t_log_res;1028	tres.tr_logcount = trans->t_log_count;1029 1030	*tpp = xfs_trans_dup(trans);1031 1032	/*1033	 * Commit the current transaction.1034	 * If this commit failed, then it'd just unlock those items that1035	 * are not marked ihold. That also means that a filesystem shutdown1036	 * is in progress. The caller takes the responsibility to cancel1037	 * the duplicate transaction that gets returned.1038	 */1039	error = __xfs_trans_commit(trans, true);1040	if (error)1041		return error;1042 1043	/*1044	 * Reserve space in the log for the next transaction.1045	 * This also pushes items in the "AIL", the list of logged items,1046	 * out to disk if they are taking up space at the tail of the log1047	 * that we want to use.  This requires that either nothing be locked1048	 * across this call, or that anything that is locked be logged in1049	 * the prior and the next transactions.1050	 */1051	tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;1052	return xfs_trans_reserve(*tpp, &tres, 0, 0);1053}1054 1055/*1056 * Allocate an transaction, lock and join the inode to it, and reserve quota.1057 *1058 * The caller must ensure that the on-disk dquots attached to this inode have1059 * already been allocated and initialized.  The caller is responsible for1060 * releasing ILOCK_EXCL if a new transaction is returned.1061 */1062int1063xfs_trans_alloc_inode(1064	struct xfs_inode	*ip,1065	struct xfs_trans_res	*resv,1066	unsigned int		dblocks,1067	unsigned int		rblocks,1068	bool			force,1069	struct xfs_trans	**tpp)1070{1071	struct xfs_trans	*tp;1072	struct xfs_mount	*mp = ip->i_mount;1073	bool			retried = false;1074	int			error;1075 1076retry:1077	error = xfs_trans_alloc(mp, resv, dblocks,1078			xfs_extlen_to_rtxlen(mp, rblocks),1079			force ? XFS_TRANS_RESERVE : 0, &tp);1080	if (error)1081		return error;1082 1083	xfs_ilock(ip, XFS_ILOCK_EXCL);1084	xfs_trans_ijoin(tp, ip, 0);1085 1086	error = xfs_qm_dqattach_locked(ip, false);1087	if (error) {1088		/* Caller should have allocated the dquots! */1089		ASSERT(error != -ENOENT);1090		goto out_cancel;1091	}1092 1093	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force);1094	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {1095		xfs_trans_cancel(tp);1096		xfs_iunlock(ip, XFS_ILOCK_EXCL);1097		xfs_blockgc_free_quota(ip, 0);1098		retried = true;1099		goto retry;1100	}1101	if (error)1102		goto out_cancel;1103 1104	*tpp = tp;1105	return 0;1106 1107out_cancel:1108	xfs_trans_cancel(tp);1109	xfs_iunlock(ip, XFS_ILOCK_EXCL);1110	return error;1111}1112 1113/*1114 * Try to reserve more blocks for a transaction.1115 *1116 * This is for callers that need to attach resources to a transaction, scan1117 * those resources to determine the space reservation requirements, and then1118 * modify the attached resources.  In other words, online repair.  This can1119 * fail due to ENOSPC, so the caller must be able to cancel the transaction1120 * without shutting down the fs.1121 */1122int1123xfs_trans_reserve_more(1124	struct xfs_trans	*tp,1125	unsigned int		blocks,1126	unsigned int		rtextents)1127{1128	struct xfs_trans_res	resv = { };1129 1130	return xfs_trans_reserve(tp, &resv, blocks, rtextents);1131}1132 1133/*1134 * Try to reserve more blocks and file quota for a transaction.  Same1135 * conditions of usage as xfs_trans_reserve_more.1136 */1137int1138xfs_trans_reserve_more_inode(1139	struct xfs_trans	*tp,1140	struct xfs_inode	*ip,1141	unsigned int		dblocks,1142	unsigned int		rblocks,1143	bool			force_quota)1144{1145	struct xfs_trans_res	resv = { };1146	struct xfs_mount	*mp = ip->i_mount;1147	unsigned int		rtx = xfs_extlen_to_rtxlen(mp, rblocks);1148	int			error;1149 1150	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);1151 1152	error = xfs_trans_reserve(tp, &resv, dblocks, rtx);1153	if (error)1154		return error;1155 1156	if (!XFS_IS_QUOTA_ON(mp) || xfs_is_quota_inode(&mp->m_sb, ip->i_ino))1157		return 0;1158 1159	if (tp->t_flags & XFS_TRANS_RESERVE)1160		force_quota = true;1161 1162	error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks,1163			force_quota);1164	if (!error)1165		return 0;1166 1167	/* Quota failed, give back the new reservation. */1168	xfs_add_fdblocks(mp, dblocks);1169	tp->t_blk_res -= dblocks;1170	xfs_add_frextents(mp, rtx);1171	tp->t_rtx_res -= rtx;1172	return error;1173}1174 1175/*1176 * Allocate an transaction in preparation for inode creation by reserving quota1177 * against the given dquots.  Callers are not required to hold any inode locks.1178 */1179int1180xfs_trans_alloc_icreate(1181	struct xfs_mount	*mp,1182	struct xfs_trans_res	*resv,1183	struct xfs_dquot	*udqp,1184	struct xfs_dquot	*gdqp,1185	struct xfs_dquot	*pdqp,1186	unsigned int		dblocks,1187	struct xfs_trans	**tpp)1188{1189	struct xfs_trans	*tp;1190	bool			retried = false;1191	int			error;1192 1193retry:1194	error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp);1195	if (error)1196		return error;1197 1198	error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks);1199	if ((error == -EDQUOT || error == -ENOSPC) && !retried) {1200		xfs_trans_cancel(tp);1201		xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);1202		retried = true;1203		goto retry;1204	}1205	if (error) {1206		xfs_trans_cancel(tp);1207		return error;1208	}1209 1210	*tpp = tp;1211	return 0;1212}1213 1214/*1215 * Allocate an transaction, lock and join the inode to it, and reserve quota1216 * in preparation for inode attribute changes that include uid, gid, or prid1217 * changes.1218 *1219 * The caller must ensure that the on-disk dquots attached to this inode have1220 * already been allocated and initialized.  The ILOCK will be dropped when the1221 * transaction is committed or cancelled.1222 */1223int1224xfs_trans_alloc_ichange(1225	struct xfs_inode	*ip,1226	struct xfs_dquot	*new_udqp,1227	struct xfs_dquot	*new_gdqp,1228	struct xfs_dquot	*new_pdqp,1229	bool			force,1230	struct xfs_trans	**tpp)1231{1232	struct xfs_trans	*tp;1233	struct xfs_mount	*mp = ip->i_mount;1234	struct xfs_dquot	*udqp;1235	struct xfs_dquot	*gdqp;1236	struct xfs_dquot	*pdqp;1237	bool			retried = false;1238	int			error;1239 1240retry:1241	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);1242	if (error)1243		return error;1244 1245	xfs_ilock(ip, XFS_ILOCK_EXCL);1246	xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);1247 1248	error = xfs_qm_dqattach_locked(ip, false);1249	if (error) {1250		/* Caller should have allocated the dquots! */1251		ASSERT(error != -ENOENT);1252		goto out_cancel;1253	}1254 1255	/*1256	 * For each quota type, skip quota reservations if the inode's dquots1257	 * now match the ones that came from the caller, or the caller didn't1258	 * pass one in.  The inode's dquots can change if we drop the ILOCK to1259	 * perform a blockgc scan, so we must preserve the caller's arguments.1260	 */1261	udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL;1262	gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL;1263	pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL;1264	if (udqp || gdqp || pdqp) {1265		unsigned int	qflags = XFS_QMOPT_RES_REGBLKS;1266 1267		if (force)1268			qflags |= XFS_QMOPT_FORCE_RES;1269 1270		/*1271		 * Reserve enough quota to handle blocks on disk and reserved1272		 * for a delayed allocation.  We'll actually transfer the1273		 * delalloc reservation between dquots at chown time, even1274		 * though that part is only semi-transactional.1275		 */1276		error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp,1277				pdqp, ip->i_nblocks + ip->i_delayed_blks,1278				1, qflags);1279		if ((error == -EDQUOT || error == -ENOSPC) && !retried) {1280			xfs_trans_cancel(tp);1281			xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0);1282			retried = true;1283			goto retry;1284		}1285		if (error)1286			goto out_cancel;1287	}1288 1289	*tpp = tp;1290	return 0;1291 1292out_cancel:1293	xfs_trans_cancel(tp);1294	return error;1295}1296 1297/*1298 * Allocate an transaction, lock and join the directory and child inodes to it,1299 * and reserve quota for a directory update.  If there isn't sufficient space,1300 * @dblocks will be set to zero for a reservationless directory update and1301 * @nospace_error will be set to a negative errno describing the space1302 * constraint we hit.1303 *1304 * The caller must ensure that the on-disk dquots attached to this inode have1305 * already been allocated and initialized.  The ILOCKs will be dropped when the1306 * transaction is committed or cancelled.1307 *1308 * Caller is responsible for unlocking the inodes manually upon return1309 */1310int1311xfs_trans_alloc_dir(1312	struct xfs_inode	*dp,1313	struct xfs_trans_res	*resv,1314	struct xfs_inode	*ip,1315	unsigned int		*dblocks,1316	struct xfs_trans	**tpp,1317	int			*nospace_error)1318{1319	struct xfs_trans	*tp;1320	struct xfs_mount	*mp = ip->i_mount;1321	unsigned int		resblks;1322	bool			retried = false;1323	int			error;1324 1325retry:1326	*nospace_error = 0;1327	resblks = *dblocks;1328	error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);1329	if (error == -ENOSPC) {1330		*nospace_error = error;1331		resblks = 0;1332		error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp);1333	}1334	if (error)1335		return error;1336 1337	xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);1338 1339	xfs_trans_ijoin(tp, dp, 0);1340	xfs_trans_ijoin(tp, ip, 0);1341 1342	error = xfs_qm_dqattach_locked(dp, false);1343	if (error) {1344		/* Caller should have allocated the dquots! */1345		ASSERT(error != -ENOENT);1346		goto out_cancel;1347	}1348 1349	error = xfs_qm_dqattach_locked(ip, false);1350	if (error) {1351		/* Caller should have allocated the dquots! */1352		ASSERT(error != -ENOENT);1353		goto out_cancel;1354	}1355 1356	if (resblks == 0)1357		goto done;1358 1359	error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false);1360	if (error == -EDQUOT || error == -ENOSPC) {1361		if (!retried) {1362			xfs_trans_cancel(tp);1363			xfs_iunlock(dp, XFS_ILOCK_EXCL);1364			if (dp != ip)1365				xfs_iunlock(ip, XFS_ILOCK_EXCL);1366			xfs_blockgc_free_quota(dp, 0);1367			retried = true;1368			goto retry;1369		}1370 1371		*nospace_error = error;1372		resblks = 0;1373		error = 0;1374	}1375	if (error)1376		goto out_cancel;1377 1378done:1379	*tpp = tp;1380	*dblocks = resblks;1381	return 0;1382 1383out_cancel:1384	xfs_trans_cancel(tp);1385	return error;1386}1387