Linux Audio

Check our new training course

Loading...
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2002,2005 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_bit.h"
 
  13#include "xfs_mount.h"
  14#include "xfs_inode.h"
  15#include "xfs_btree.h"
  16#include "xfs_ialloc.h"
  17#include "xfs_ialloc_btree.h"
  18#include "xfs_alloc.h"
  19#include "xfs_errortag.h"
  20#include "xfs_error.h"
  21#include "xfs_bmap.h"
  22#include "xfs_trans.h"
  23#include "xfs_buf_item.h"
  24#include "xfs_icreate_item.h"
  25#include "xfs_icache.h"
  26#include "xfs_trace.h"
  27#include "xfs_log.h"
  28#include "xfs_rmap.h"
  29#include "xfs_ag.h"
  30#include "xfs_health.h"
  31
  32/*
  33 * Lookup a record by ino in the btree given by cur.
  34 */
  35int					/* error */
  36xfs_inobt_lookup(
  37	struct xfs_btree_cur	*cur,	/* btree cursor */
  38	xfs_agino_t		ino,	/* starting inode of chunk */
  39	xfs_lookup_t		dir,	/* <=, >=, == */
  40	int			*stat)	/* success/failure */
  41{
  42	cur->bc_rec.i.ir_startino = ino;
  43	cur->bc_rec.i.ir_holemask = 0;
  44	cur->bc_rec.i.ir_count = 0;
  45	cur->bc_rec.i.ir_freecount = 0;
  46	cur->bc_rec.i.ir_free = 0;
  47	return xfs_btree_lookup(cur, dir, stat);
  48}
  49
  50/*
  51 * Update the record referred to by cur to the value given.
  52 * This either works (return 0) or gets an EFSCORRUPTED error.
  53 */
  54STATIC int				/* error */
  55xfs_inobt_update(
  56	struct xfs_btree_cur	*cur,	/* btree cursor */
  57	xfs_inobt_rec_incore_t	*irec)	/* btree record */
  58{
  59	union xfs_btree_rec	rec;
  60
  61	rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
  62	if (xfs_has_sparseinodes(cur->bc_mp)) {
  63		rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
  64		rec.inobt.ir_u.sp.ir_count = irec->ir_count;
  65		rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
  66	} else {
  67		/* ir_holemask/ir_count not supported on-disk */
  68		rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
  69	}
  70	rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
  71	return xfs_btree_update(cur, &rec);
  72}
  73
  74/* Convert on-disk btree record to incore inobt record. */
  75void
  76xfs_inobt_btrec_to_irec(
  77	struct xfs_mount		*mp,
  78	const union xfs_btree_rec	*rec,
  79	struct xfs_inobt_rec_incore	*irec)
  80{
  81	irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
  82	if (xfs_has_sparseinodes(mp)) {
  83		irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
  84		irec->ir_count = rec->inobt.ir_u.sp.ir_count;
  85		irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
  86	} else {
  87		/*
  88		 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
  89		 * values for full inode chunks.
  90		 */
  91		irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
  92		irec->ir_count = XFS_INODES_PER_CHUNK;
  93		irec->ir_freecount =
  94				be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
  95	}
  96	irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
  97}
  98
  99/* Compute the freecount of an incore inode record. */
 100uint8_t
 101xfs_inobt_rec_freecount(
 102	const struct xfs_inobt_rec_incore	*irec)
 103{
 104	uint64_t				realfree = irec->ir_free;
 105
 106	if (xfs_inobt_issparse(irec->ir_holemask))
 107		realfree &= xfs_inobt_irec_to_allocmask(irec);
 108	return hweight64(realfree);
 109}
 110
 111/* Simple checks for inode records. */
 112xfs_failaddr_t
 113xfs_inobt_check_irec(
 114	struct xfs_perag			*pag,
 115	const struct xfs_inobt_rec_incore	*irec)
 116{
 117	/* Record has to be properly aligned within the AG. */
 118	if (!xfs_verify_agino(pag, irec->ir_startino))
 119		return __this_address;
 120	if (!xfs_verify_agino(pag,
 121				irec->ir_startino + XFS_INODES_PER_CHUNK - 1))
 122		return __this_address;
 123	if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
 124	    irec->ir_count > XFS_INODES_PER_CHUNK)
 125		return __this_address;
 126	if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
 127		return __this_address;
 128
 129	if (xfs_inobt_rec_freecount(irec) != irec->ir_freecount)
 130		return __this_address;
 131
 132	return NULL;
 133}
 134
 135static inline int
 136xfs_inobt_complain_bad_rec(
 137	struct xfs_btree_cur		*cur,
 138	xfs_failaddr_t			fa,
 139	const struct xfs_inobt_rec_incore *irec)
 140{
 141	struct xfs_mount		*mp = cur->bc_mp;
 142
 143	xfs_warn(mp,
 144		"%sbt record corruption in AG %d detected at %pS!",
 145		cur->bc_ops->name, cur->bc_ag.pag->pag_agno, fa);
 146	xfs_warn(mp,
 147"start inode 0x%x, count 0x%x, free 0x%x freemask 0x%llx, holemask 0x%x",
 148		irec->ir_startino, irec->ir_count, irec->ir_freecount,
 149		irec->ir_free, irec->ir_holemask);
 150	xfs_btree_mark_sick(cur);
 151	return -EFSCORRUPTED;
 152}
 153
 154/*
 155 * Get the data from the pointed-to record.
 156 */
 157int
 158xfs_inobt_get_rec(
 159	struct xfs_btree_cur		*cur,
 160	struct xfs_inobt_rec_incore	*irec,
 161	int				*stat)
 162{
 163	struct xfs_mount		*mp = cur->bc_mp;
 
 164	union xfs_btree_rec		*rec;
 165	xfs_failaddr_t			fa;
 166	int				error;
 
 167
 168	error = xfs_btree_get_rec(cur, &rec, stat);
 169	if (error || *stat == 0)
 170		return error;
 171
 172	xfs_inobt_btrec_to_irec(mp, rec, irec);
 173	fa = xfs_inobt_check_irec(cur->bc_ag.pag, irec);
 174	if (fa)
 175		return xfs_inobt_complain_bad_rec(cur, fa, irec);
 
 
 
 
 
 
 
 
 
 
 
 
 
 176
 177	return 0;
 
 
 
 
 
 
 
 
 
 
 178}
 179
 180/*
 181 * Insert a single inobt record. Cursor must already point to desired location.
 182 */
 183int
 184xfs_inobt_insert_rec(
 185	struct xfs_btree_cur	*cur,
 186	uint16_t		holemask,
 187	uint8_t			count,
 188	int32_t			freecount,
 189	xfs_inofree_t		free,
 190	int			*stat)
 191{
 192	cur->bc_rec.i.ir_holemask = holemask;
 193	cur->bc_rec.i.ir_count = count;
 194	cur->bc_rec.i.ir_freecount = freecount;
 195	cur->bc_rec.i.ir_free = free;
 196	return xfs_btree_insert(cur, stat);
 197}
 198
 199/*
 200 * Insert records describing a newly allocated inode chunk into the inobt.
 201 */
 202STATIC int
 203xfs_inobt_insert(
 204	struct xfs_perag	*pag,
 205	struct xfs_trans	*tp,
 206	struct xfs_buf		*agbp,
 207	xfs_agino_t		newino,
 208	xfs_agino_t		newlen,
 209	bool			is_finobt)
 210{
 211	struct xfs_btree_cur	*cur;
 
 
 212	xfs_agino_t		thisino;
 213	int			i;
 214	int			error;
 215
 216	if (is_finobt)
 217		cur = xfs_finobt_init_cursor(pag, tp, agbp);
 218	else
 219		cur = xfs_inobt_init_cursor(pag, tp, agbp);
 220
 221	for (thisino = newino;
 222	     thisino < newino + newlen;
 223	     thisino += XFS_INODES_PER_CHUNK) {
 224		error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
 225		if (error) {
 226			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 227			return error;
 228		}
 229		ASSERT(i == 0);
 230
 231		error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
 232					     XFS_INODES_PER_CHUNK,
 233					     XFS_INODES_PER_CHUNK,
 234					     XFS_INOBT_ALL_FREE, &i);
 235		if (error) {
 236			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 237			return error;
 238		}
 239		ASSERT(i == 1);
 240	}
 241
 242	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
 243
 244	return 0;
 245}
 246
 247/*
 248 * Verify that the number of free inodes in the AGI is correct.
 249 */
 250#ifdef DEBUG
 251static int
 252xfs_check_agi_freecount(
 253	struct xfs_btree_cur	*cur)
 
 254{
 255	if (cur->bc_nlevels == 1) {
 256		xfs_inobt_rec_incore_t rec;
 257		int		freecount = 0;
 258		int		error;
 259		int		i;
 260
 261		error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
 262		if (error)
 263			return error;
 264
 265		do {
 266			error = xfs_inobt_get_rec(cur, &rec, &i);
 267			if (error)
 268				return error;
 269
 270			if (i) {
 271				freecount += rec.ir_freecount;
 272				error = xfs_btree_increment(cur, 0, &i);
 273				if (error)
 274					return error;
 275			}
 276		} while (i == 1);
 277
 278		if (!xfs_is_shutdown(cur->bc_mp))
 279			ASSERT(freecount == cur->bc_ag.pag->pagi_freecount);
 280	}
 281	return 0;
 282}
 283#else
 284#define xfs_check_agi_freecount(cur)	0
 285#endif
 286
 287/*
 288 * Initialise a new set of inodes. When called without a transaction context
 289 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
 290 * than logging them (which in a transaction context puts them into the AIL
 291 * for writeback rather than the xfsbufd queue).
 292 */
 293int
 294xfs_ialloc_inode_init(
 295	struct xfs_mount	*mp,
 296	struct xfs_trans	*tp,
 297	struct list_head	*buffer_list,
 298	int			icount,
 299	xfs_agnumber_t		agno,
 300	xfs_agblock_t		agbno,
 301	xfs_agblock_t		length,
 302	unsigned int		gen)
 303{
 304	struct xfs_buf		*fbuf;
 305	struct xfs_dinode	*free;
 306	int			nbufs;
 307	int			version;
 308	int			i, j;
 309	xfs_daddr_t		d;
 310	xfs_ino_t		ino = 0;
 311	int			error;
 312
 313	/*
 314	 * Loop over the new block(s), filling in the inodes.  For small block
 315	 * sizes, manipulate the inodes in buffers  which are multiples of the
 316	 * blocks size.
 317	 */
 318	nbufs = length / M_IGEO(mp)->blocks_per_cluster;
 319
 320	/*
 321	 * Figure out what version number to use in the inodes we create.  If
 322	 * the superblock version has caught up to the one that supports the new
 323	 * inode format, then use the new inode version.  Otherwise use the old
 324	 * version so that old kernels will continue to be able to use the file
 325	 * system.
 326	 *
 327	 * For v3 inodes, we also need to write the inode number into the inode,
 328	 * so calculate the first inode number of the chunk here as
 329	 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
 330	 * across multiple filesystem blocks (such as a cluster) and so cannot
 331	 * be used in the cluster buffer loop below.
 332	 *
 333	 * Further, because we are writing the inode directly into the buffer
 334	 * and calculating a CRC on the entire inode, we have ot log the entire
 335	 * inode so that the entire range the CRC covers is present in the log.
 336	 * That means for v3 inode we log the entire buffer rather than just the
 337	 * inode cores.
 338	 */
 339	if (xfs_has_v3inodes(mp)) {
 340		version = 3;
 341		ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
 342
 343		/*
 344		 * log the initialisation that is about to take place as an
 345		 * logical operation. This means the transaction does not
 346		 * need to log the physical changes to the inode buffers as log
 347		 * recovery will know what initialisation is actually needed.
 348		 * Hence we only need to log the buffers as "ordered" buffers so
 349		 * they track in the AIL as if they were physically logged.
 350		 */
 351		if (tp)
 352			xfs_icreate_log(tp, agno, agbno, icount,
 353					mp->m_sb.sb_inodesize, length, gen);
 354	} else
 355		version = 2;
 356
 357	for (j = 0; j < nbufs; j++) {
 358		/*
 359		 * Get the block.
 360		 */
 361		d = XFS_AGB_TO_DADDR(mp, agno, agbno +
 362				(j * M_IGEO(mp)->blocks_per_cluster));
 363		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
 364				mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
 365				XBF_UNMAPPED, &fbuf);
 366		if (error)
 367			return error;
 368
 369		/* Initialize the inode buffers and log them appropriately. */
 370		fbuf->b_ops = &xfs_inode_buf_ops;
 371		xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
 372		for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
 373			int	ioffset = i << mp->m_sb.sb_inodelog;
 
 374
 375			free = xfs_make_iptr(mp, fbuf, i);
 376			free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
 377			free->di_version = version;
 378			free->di_gen = cpu_to_be32(gen);
 379			free->di_next_unlinked = cpu_to_be32(NULLAGINO);
 380
 381			if (version == 3) {
 382				free->di_ino = cpu_to_be64(ino);
 383				ino++;
 384				uuid_copy(&free->di_uuid,
 385					  &mp->m_sb.sb_meta_uuid);
 386				xfs_dinode_calc_crc(mp, free);
 387			} else if (tp) {
 388				/* just log the inode core */
 389				xfs_trans_log_buf(tp, fbuf, ioffset,
 390					  ioffset + XFS_DINODE_SIZE(mp) - 1);
 391			}
 392		}
 393
 394		if (tp) {
 395			/*
 396			 * Mark the buffer as an inode allocation buffer so it
 397			 * sticks in AIL at the point of this allocation
 398			 * transaction. This ensures the they are on disk before
 399			 * the tail of the log can be moved past this
 400			 * transaction (i.e. by preventing relogging from moving
 401			 * it forward in the log).
 402			 */
 403			xfs_trans_inode_alloc_buf(tp, fbuf);
 404			if (version == 3) {
 405				/*
 406				 * Mark the buffer as ordered so that they are
 407				 * not physically logged in the transaction but
 408				 * still tracked in the AIL as part of the
 409				 * transaction and pin the log appropriately.
 410				 */
 411				xfs_trans_ordered_buf(tp, fbuf);
 412			}
 413		} else {
 414			fbuf->b_flags |= XBF_DONE;
 415			xfs_buf_delwri_queue(fbuf, buffer_list);
 416			xfs_buf_relse(fbuf);
 417		}
 418	}
 419	return 0;
 420}
 421
 422/*
 423 * Align startino and allocmask for a recently allocated sparse chunk such that
 424 * they are fit for insertion (or merge) into the on-disk inode btrees.
 425 *
 426 * Background:
 427 *
 428 * When enabled, sparse inode support increases the inode alignment from cluster
 429 * size to inode chunk size. This means that the minimum range between two
 430 * non-adjacent inode records in the inobt is large enough for a full inode
 431 * record. This allows for cluster sized, cluster aligned block allocation
 432 * without need to worry about whether the resulting inode record overlaps with
 433 * another record in the tree. Without this basic rule, we would have to deal
 434 * with the consequences of overlap by potentially undoing recent allocations in
 435 * the inode allocation codepath.
 436 *
 437 * Because of this alignment rule (which is enforced on mount), there are two
 438 * inobt possibilities for newly allocated sparse chunks. One is that the
 439 * aligned inode record for the chunk covers a range of inodes not already
 440 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
 441 * other is that a record already exists at the aligned startino that considers
 442 * the newly allocated range as sparse. In the latter case, record content is
 443 * merged in hope that sparse inode chunks fill to full chunks over time.
 444 */
 445STATIC void
 446xfs_align_sparse_ino(
 447	struct xfs_mount		*mp,
 448	xfs_agino_t			*startino,
 449	uint16_t			*allocmask)
 450{
 451	xfs_agblock_t			agbno;
 452	xfs_agblock_t			mod;
 453	int				offset;
 454
 455	agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
 456	mod = agbno % mp->m_sb.sb_inoalignmt;
 457	if (!mod)
 458		return;
 459
 460	/* calculate the inode offset and align startino */
 461	offset = XFS_AGB_TO_AGINO(mp, mod);
 462	*startino -= offset;
 463
 464	/*
 465	 * Since startino has been aligned down, left shift allocmask such that
 466	 * it continues to represent the same physical inodes relative to the
 467	 * new startino.
 468	 */
 469	*allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
 470}
 471
 472/*
 473 * Determine whether the source inode record can merge into the target. Both
 474 * records must be sparse, the inode ranges must match and there must be no
 475 * allocation overlap between the records.
 476 */
 477STATIC bool
 478__xfs_inobt_can_merge(
 479	struct xfs_inobt_rec_incore	*trec,	/* tgt record */
 480	struct xfs_inobt_rec_incore	*srec)	/* src record */
 481{
 482	uint64_t			talloc;
 483	uint64_t			salloc;
 484
 485	/* records must cover the same inode range */
 486	if (trec->ir_startino != srec->ir_startino)
 487		return false;
 488
 489	/* both records must be sparse */
 490	if (!xfs_inobt_issparse(trec->ir_holemask) ||
 491	    !xfs_inobt_issparse(srec->ir_holemask))
 492		return false;
 493
 494	/* both records must track some inodes */
 495	if (!trec->ir_count || !srec->ir_count)
 496		return false;
 497
 498	/* can't exceed capacity of a full record */
 499	if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
 500		return false;
 501
 502	/* verify there is no allocation overlap */
 503	talloc = xfs_inobt_irec_to_allocmask(trec);
 504	salloc = xfs_inobt_irec_to_allocmask(srec);
 505	if (talloc & salloc)
 506		return false;
 507
 508	return true;
 509}
 510
 511/*
 512 * Merge the source inode record into the target. The caller must call
 513 * __xfs_inobt_can_merge() to ensure the merge is valid.
 514 */
 515STATIC void
 516__xfs_inobt_rec_merge(
 517	struct xfs_inobt_rec_incore	*trec,	/* target */
 518	struct xfs_inobt_rec_incore	*srec)	/* src */
 519{
 520	ASSERT(trec->ir_startino == srec->ir_startino);
 521
 522	/* combine the counts */
 523	trec->ir_count += srec->ir_count;
 524	trec->ir_freecount += srec->ir_freecount;
 525
 526	/*
 527	 * Merge the holemask and free mask. For both fields, 0 bits refer to
 528	 * allocated inodes. We combine the allocated ranges with bitwise AND.
 529	 */
 530	trec->ir_holemask &= srec->ir_holemask;
 531	trec->ir_free &= srec->ir_free;
 532}
 533
 534/*
 535 * Insert a new sparse inode chunk into the associated inode allocation btree.
 536 * The inode record for the sparse chunk is pre-aligned to a startino that
 537 * should match any pre-existing sparse inode record in the tree. This allows
 538 * sparse chunks to fill over time.
 539 *
 540 * If no preexisting record exists, the provided record is inserted.
 541 * If there is a preexisting record, the provided record is merged with the
 542 * existing record and updated in place. The merged record is returned in nrec.
 
 
 543 *
 544 * It is considered corruption if a merge is requested and not possible. Given
 545 * the sparse inode alignment constraints, this should never happen.
 546 */
 547STATIC int
 548xfs_inobt_insert_sprec(
 549	struct xfs_perag		*pag,
 550	struct xfs_trans		*tp,
 551	struct xfs_buf			*agbp,
 552	struct xfs_inobt_rec_incore	*nrec)	/* in/out: new/merged rec. */
 
 
 553{
 554	struct xfs_mount		*mp = pag->pag_mount;
 555	struct xfs_btree_cur		*cur;
 
 
 556	int				error;
 557	int				i;
 558	struct xfs_inobt_rec_incore	rec;
 559
 560	cur = xfs_inobt_init_cursor(pag, tp, agbp);
 561
 562	/* the new record is pre-aligned so we know where to look */
 563	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
 564	if (error)
 565		goto error;
 566	/* if nothing there, insert a new record and return */
 567	if (i == 0) {
 568		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
 569					     nrec->ir_count, nrec->ir_freecount,
 570					     nrec->ir_free, &i);
 571		if (error)
 572			goto error;
 573		if (XFS_IS_CORRUPT(mp, i != 1)) {
 574			xfs_btree_mark_sick(cur);
 575			error = -EFSCORRUPTED;
 576			goto error;
 577		}
 578
 579		goto out;
 580	}
 581
 582	/*
 583	 * A record exists at this startino.  Merge the records.
 
 584	 */
 585	error = xfs_inobt_get_rec(cur, &rec, &i);
 586	if (error)
 587		goto error;
 588	if (XFS_IS_CORRUPT(mp, i != 1)) {
 589		xfs_btree_mark_sick(cur);
 590		error = -EFSCORRUPTED;
 591		goto error;
 592	}
 593	if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
 594		xfs_btree_mark_sick(cur);
 595		error = -EFSCORRUPTED;
 596		goto error;
 597	}
 598
 599	/*
 600	 * This should never fail. If we have coexisting records that
 601	 * cannot merge, something is seriously wrong.
 602	 */
 603	if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
 604		xfs_btree_mark_sick(cur);
 605		error = -EFSCORRUPTED;
 606		goto error;
 607	}
 608
 609	trace_xfs_irec_merge_pre(mp, pag->pag_agno, rec.ir_startino,
 610				 rec.ir_holemask, nrec->ir_startino,
 611				 nrec->ir_holemask);
 612
 613	/* merge to nrec to output the updated record */
 614	__xfs_inobt_rec_merge(nrec, &rec);
 615
 616	trace_xfs_irec_merge_post(mp, pag->pag_agno, nrec->ir_startino,
 617				  nrec->ir_holemask);
 618
 619	error = xfs_inobt_rec_check_count(mp, nrec);
 620	if (error)
 621		goto error;
 
 622
 623	error = xfs_inobt_update(cur, nrec);
 624	if (error)
 625		goto error;
 626
 627out:
 628	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
 629	return 0;
 630error:
 631	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 632	return error;
 633}
 634
 635/*
 636 * Insert a new sparse inode chunk into the free inode btree. The inode
 637 * record for the sparse chunk is pre-aligned to a startino that should match
 638 * any pre-existing sparse inode record in the tree. This allows sparse chunks
 639 * to fill over time.
 640 *
 641 * The new record is always inserted, overwriting a pre-existing record if
 642 * there is one.
 643 */
 644STATIC int
 645xfs_finobt_insert_sprec(
 646	struct xfs_perag		*pag,
 647	struct xfs_trans		*tp,
 648	struct xfs_buf			*agbp,
 649	struct xfs_inobt_rec_incore	*nrec)	/* in/out: new rec. */
 650{
 651	struct xfs_mount		*mp = pag->pag_mount;
 652	struct xfs_btree_cur		*cur;
 653	int				error;
 654	int				i;
 655
 656	cur = xfs_finobt_init_cursor(pag, tp, agbp);
 657
 658	/* the new record is pre-aligned so we know where to look */
 659	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
 660	if (error)
 661		goto error;
 662	/* if nothing there, insert a new record and return */
 663	if (i == 0) {
 664		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
 665					     nrec->ir_count, nrec->ir_freecount,
 666					     nrec->ir_free, &i);
 667		if (error)
 668			goto error;
 669		if (XFS_IS_CORRUPT(mp, i != 1)) {
 670			xfs_btree_mark_sick(cur);
 671			error = -EFSCORRUPTED;
 672			goto error;
 673		}
 674	} else {
 675		error = xfs_inobt_update(cur, nrec);
 676		if (error)
 677			goto error;
 678	}
 679
 680	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
 681	return 0;
 682error:
 683	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 684	return error;
 685}
 686
 687
 688/*
 689 * Allocate new inodes in the allocation group specified by agbp.  Returns 0 if
 690 * inodes were allocated in this AG; -EAGAIN if there was no space in this AG so
 691 * the caller knows it can try another AG, a hard -ENOSPC when over the maximum
 692 * inode count threshold, or the usual negative error code for other errors.
 693 */
 694STATIC int
 695xfs_ialloc_ag_alloc(
 696	struct xfs_perag	*pag,
 697	struct xfs_trans	*tp,
 698	struct xfs_buf		*agbp)
 
 699{
 700	struct xfs_agi		*agi;
 701	struct xfs_alloc_arg	args;
 
 702	int			error;
 703	xfs_agino_t		newino;		/* new first inode's number */
 704	xfs_agino_t		newlen;		/* new number of inodes */
 705	int			isaligned = 0;	/* inode allocation at stripe */
 706						/* unit boundary */
 707	/* init. to full chunk */
 
 708	struct xfs_inobt_rec_incore rec;
 
 709	struct xfs_ino_geometry	*igeo = M_IGEO(tp->t_mountp);
 710	uint16_t		allocmask = (uint16_t) -1;
 711	int			do_sparse = 0;
 712
 713	memset(&args, 0, sizeof(args));
 714	args.tp = tp;
 715	args.mp = tp->t_mountp;
 716	args.fsbno = NULLFSBLOCK;
 717	args.oinfo = XFS_RMAP_OINFO_INODES;
 718	args.pag = pag;
 719
 720#ifdef DEBUG
 721	/* randomly do sparse inode allocations */
 722	if (xfs_has_sparseinodes(tp->t_mountp) &&
 723	    igeo->ialloc_min_blks < igeo->ialloc_blks)
 724		do_sparse = get_random_u32_below(2);
 725#endif
 726
 727	/*
 728	 * Locking will ensure that we don't have two callers in here
 729	 * at one time.
 730	 */
 731	newlen = igeo->ialloc_inos;
 732	if (igeo->maxicount &&
 733	    percpu_counter_read_positive(&args.mp->m_icount) + newlen >
 734							igeo->maxicount)
 735		return -ENOSPC;
 736	args.minlen = args.maxlen = igeo->ialloc_blks;
 737	/*
 738	 * First try to allocate inodes contiguous with the last-allocated
 739	 * chunk of inodes.  If the filesystem is striped, this will fill
 740	 * an entire stripe unit with inodes.
 741	 */
 742	agi = agbp->b_addr;
 743	newino = be32_to_cpu(agi->agi_newino);
 
 744	args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
 745		     igeo->ialloc_blks;
 746	if (do_sparse)
 747		goto sparse_alloc;
 748	if (likely(newino != NULLAGINO &&
 749		  (args.agbno < be32_to_cpu(agi->agi_length)))) {
 
 
 750		args.prod = 1;
 751
 752		/*
 753		 * We need to take into account alignment here to ensure that
 754		 * we don't modify the free list if we fail to have an exact
 755		 * block. If we don't have an exact match, and every oher
 756		 * attempt allocation attempt fails, we'll end up cancelling
 757		 * a dirty transaction and shutting down.
 758		 *
 759		 * For an exact allocation, alignment must be 1,
 760		 * however we need to take cluster alignment into account when
 761		 * fixing up the freelist. Use the minalignslop field to
 762		 * indicate that extra blocks might be required for alignment,
 763		 * but not to use them in the actual exact allocation.
 764		 */
 765		args.alignment = 1;
 766		args.minalignslop = igeo->cluster_align - 1;
 767
 768		/* Allow space for the inode btree to split. */
 769		args.minleft = igeo->inobt_maxlevels;
 770		error = xfs_alloc_vextent_exact_bno(&args,
 771				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
 772						args.agbno));
 773		if (error)
 774			return error;
 775
 776		/*
 777		 * This request might have dirtied the transaction if the AG can
 778		 * satisfy the request, but the exact block was not available.
 779		 * If the allocation did fail, subsequent requests will relax
 780		 * the exact agbno requirement and increase the alignment
 781		 * instead. It is critical that the total size of the request
 782		 * (len + alignment + slop) does not increase from this point
 783		 * on, so reset minalignslop to ensure it is not included in
 784		 * subsequent requests.
 785		 */
 786		args.minalignslop = 0;
 787	}
 788
 789	if (unlikely(args.fsbno == NULLFSBLOCK)) {
 790		/*
 791		 * Set the alignment for the allocation.
 792		 * If stripe alignment is turned on then align at stripe unit
 793		 * boundary.
 794		 * If the cluster size is smaller than a filesystem block
 795		 * then we're doing I/O for inodes in filesystem block size
 796		 * pieces, so don't need alignment anyway.
 797		 */
 798		isaligned = 0;
 799		if (igeo->ialloc_align) {
 800			ASSERT(!xfs_has_noalign(args.mp));
 801			args.alignment = args.mp->m_dalign;
 802			isaligned = 1;
 803		} else
 804			args.alignment = igeo->cluster_align;
 805		/*
 
 
 
 
 
 
 
 806		 * Allocate a fixed-size extent of inodes.
 807		 */
 
 808		args.prod = 1;
 809		/*
 810		 * Allow space for the inode btree to split.
 811		 */
 812		args.minleft = igeo->inobt_maxlevels;
 813		error = xfs_alloc_vextent_near_bno(&args,
 814				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
 815						be32_to_cpu(agi->agi_root)));
 816		if (error)
 817			return error;
 818	}
 819
 820	/*
 821	 * If stripe alignment is turned on, then try again with cluster
 822	 * alignment.
 823	 */
 824	if (isaligned && args.fsbno == NULLFSBLOCK) {
 
 
 
 825		args.alignment = igeo->cluster_align;
 826		error = xfs_alloc_vextent_near_bno(&args,
 827				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
 828						be32_to_cpu(agi->agi_root)));
 829		if (error)
 830			return error;
 831	}
 832
 833	/*
 834	 * Finally, try a sparse allocation if the filesystem supports it and
 835	 * the sparse allocation length is smaller than a full chunk.
 836	 */
 837	if (xfs_has_sparseinodes(args.mp) &&
 838	    igeo->ialloc_min_blks < igeo->ialloc_blks &&
 839	    args.fsbno == NULLFSBLOCK) {
 840sparse_alloc:
 
 
 
 841		args.alignment = args.mp->m_sb.sb_spino_align;
 842		args.prod = 1;
 843
 844		args.minlen = igeo->ialloc_min_blks;
 845		args.maxlen = args.minlen;
 846
 847		/*
 848		 * The inode record will be aligned to full chunk size. We must
 849		 * prevent sparse allocation from AG boundaries that result in
 850		 * invalid inode records, such as records that start at agbno 0
 851		 * or extend beyond the AG.
 852		 *
 853		 * Set min agbno to the first aligned, non-zero agbno and max to
 854		 * the last aligned agbno that is at least one full chunk from
 855		 * the end of the AG.
 856		 */
 857		args.min_agbno = args.mp->m_sb.sb_inoalignmt;
 858		args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
 859					    args.mp->m_sb.sb_inoalignmt) -
 860				 igeo->ialloc_blks;
 861
 862		error = xfs_alloc_vextent_near_bno(&args,
 863				XFS_AGB_TO_FSB(args.mp, pag->pag_agno,
 864						be32_to_cpu(agi->agi_root)));
 865		if (error)
 866			return error;
 867
 868		newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
 869		ASSERT(newlen <= XFS_INODES_PER_CHUNK);
 870		allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
 871	}
 872
 873	if (args.fsbno == NULLFSBLOCK)
 874		return -EAGAIN;
 875
 
 876	ASSERT(args.len == args.minlen);
 877
 878	/*
 879	 * Stamp and write the inode buffers.
 880	 *
 881	 * Seed the new inode cluster with a random generation number. This
 882	 * prevents short-term reuse of generation numbers if a chunk is
 883	 * freed and then immediately reallocated. We use random numbers
 884	 * rather than a linear progression to prevent the next generation
 885	 * number from being easily guessable.
 886	 */
 887	error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, pag->pag_agno,
 888			args.agbno, args.len, get_random_u32());
 889
 890	if (error)
 891		return error;
 892	/*
 893	 * Convert the results.
 894	 */
 895	newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
 896
 897	if (xfs_inobt_issparse(~allocmask)) {
 898		/*
 899		 * We've allocated a sparse chunk. Align the startino and mask.
 900		 */
 901		xfs_align_sparse_ino(args.mp, &newino, &allocmask);
 902
 903		rec.ir_startino = newino;
 904		rec.ir_holemask = ~allocmask;
 905		rec.ir_count = newlen;
 906		rec.ir_freecount = newlen;
 907		rec.ir_free = XFS_INOBT_ALL_FREE;
 908
 909		/*
 910		 * Insert the sparse record into the inobt and allow for a merge
 911		 * if necessary. If a merge does occur, rec is updated to the
 912		 * merged record.
 913		 */
 914		error = xfs_inobt_insert_sprec(pag, tp, agbp, &rec);
 
 915		if (error == -EFSCORRUPTED) {
 916			xfs_alert(args.mp,
 917	"invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
 918				  XFS_AGINO_TO_INO(args.mp, pag->pag_agno,
 919						   rec.ir_startino),
 920				  rec.ir_holemask, rec.ir_count);
 921			xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
 922		}
 923		if (error)
 924			return error;
 925
 926		/*
 927		 * We can't merge the part we've just allocated as for the inobt
 928		 * due to finobt semantics. The original record may or may not
 929		 * exist independent of whether physical inodes exist in this
 930		 * sparse chunk.
 931		 *
 932		 * We must update the finobt record based on the inobt record.
 933		 * rec contains the fully merged and up to date inobt record
 934		 * from the previous call. Set merge false to replace any
 935		 * existing record with this one.
 936		 */
 937		if (xfs_has_finobt(args.mp)) {
 938			error = xfs_finobt_insert_sprec(pag, tp, agbp, &rec);
 
 
 939			if (error)
 940				return error;
 941		}
 942	} else {
 943		/* full chunk - insert new records to both btrees */
 944		error = xfs_inobt_insert(pag, tp, agbp, newino, newlen, false);
 
 945		if (error)
 946			return error;
 947
 948		if (xfs_has_finobt(args.mp)) {
 949			error = xfs_inobt_insert(pag, tp, agbp, newino,
 950						 newlen, true);
 951			if (error)
 952				return error;
 953		}
 954	}
 955
 956	/*
 957	 * Update AGI counts and newino.
 958	 */
 959	be32_add_cpu(&agi->agi_count, newlen);
 960	be32_add_cpu(&agi->agi_freecount, newlen);
 
 961	pag->pagi_freecount += newlen;
 962	pag->pagi_count += newlen;
 963	agi->agi_newino = cpu_to_be32(newino);
 964
 965	/*
 966	 * Log allocation group header fields
 967	 */
 968	xfs_ialloc_log_agi(tp, agbp,
 969		XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
 970	/*
 971	 * Modify/log superblock values for inode count and inode free count.
 972	 */
 973	xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
 974	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
 
 975	return 0;
 976}
 977
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 978/*
 979 * Try to retrieve the next record to the left/right from the current one.
 980 */
 981STATIC int
 982xfs_ialloc_next_rec(
 983	struct xfs_btree_cur	*cur,
 984	xfs_inobt_rec_incore_t	*rec,
 985	int			*done,
 986	int			left)
 987{
 988	int                     error;
 989	int			i;
 990
 991	if (left)
 992		error = xfs_btree_decrement(cur, 0, &i);
 993	else
 994		error = xfs_btree_increment(cur, 0, &i);
 995
 996	if (error)
 997		return error;
 998	*done = !i;
 999	if (i) {
1000		error = xfs_inobt_get_rec(cur, rec, &i);
1001		if (error)
1002			return error;
1003		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1004			xfs_btree_mark_sick(cur);
1005			return -EFSCORRUPTED;
1006		}
1007	}
1008
1009	return 0;
1010}
1011
1012STATIC int
1013xfs_ialloc_get_rec(
1014	struct xfs_btree_cur	*cur,
1015	xfs_agino_t		agino,
1016	xfs_inobt_rec_incore_t	*rec,
1017	int			*done)
1018{
1019	int                     error;
1020	int			i;
1021
1022	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1023	if (error)
1024		return error;
1025	*done = !i;
1026	if (i) {
1027		error = xfs_inobt_get_rec(cur, rec, &i);
1028		if (error)
1029			return error;
1030		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1031			xfs_btree_mark_sick(cur);
1032			return -EFSCORRUPTED;
1033		}
1034	}
1035
1036	return 0;
1037}
1038
1039/*
1040 * Return the offset of the first free inode in the record. If the inode chunk
1041 * is sparsely allocated, we convert the record holemask to inode granularity
1042 * and mask off the unallocated regions from the inode free mask.
1043 */
1044STATIC int
1045xfs_inobt_first_free_inode(
1046	struct xfs_inobt_rec_incore	*rec)
1047{
1048	xfs_inofree_t			realfree;
1049
1050	/* if there are no holes, return the first available offset */
1051	if (!xfs_inobt_issparse(rec->ir_holemask))
1052		return xfs_lowbit64(rec->ir_free);
1053
1054	realfree = xfs_inobt_irec_to_allocmask(rec);
1055	realfree &= rec->ir_free;
1056
1057	return xfs_lowbit64(realfree);
1058}
1059
1060/*
1061 * Allocate an inode using the inobt-only algorithm.
1062 */
1063STATIC int
1064xfs_dialloc_ag_inobt(
1065	struct xfs_perag	*pag,
1066	struct xfs_trans	*tp,
1067	struct xfs_buf		*agbp,
1068	xfs_ino_t		parent,
1069	xfs_ino_t		*inop)
1070{
1071	struct xfs_mount	*mp = tp->t_mountp;
1072	struct xfs_agi		*agi = agbp->b_addr;
 
1073	xfs_agnumber_t		pagno = XFS_INO_TO_AGNO(mp, parent);
1074	xfs_agino_t		pagino = XFS_INO_TO_AGINO(mp, parent);
 
1075	struct xfs_btree_cur	*cur, *tcur;
1076	struct xfs_inobt_rec_incore rec, trec;
1077	xfs_ino_t		ino;
1078	int			error;
1079	int			offset;
1080	int			i, j;
1081	int			searchdistance = 10;
1082
1083	ASSERT(xfs_perag_initialised_agi(pag));
1084	ASSERT(xfs_perag_allows_inodes(pag));
1085	ASSERT(pag->pagi_freecount > 0);
1086
1087 restart_pagno:
1088	cur = xfs_inobt_init_cursor(pag, tp, agbp);
1089	/*
1090	 * If pagino is 0 (this is the root inode allocation) use newino.
1091	 * This must work because we've just allocated some.
1092	 */
1093	if (!pagino)
1094		pagino = be32_to_cpu(agi->agi_newino);
1095
1096	error = xfs_check_agi_freecount(cur);
1097	if (error)
1098		goto error0;
1099
1100	/*
1101	 * If in the same AG as the parent, try to get near the parent.
1102	 */
1103	if (pagno == pag->pag_agno) {
1104		int		doneleft;	/* done, to the left */
1105		int		doneright;	/* done, to the right */
1106
1107		error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1108		if (error)
1109			goto error0;
1110		if (XFS_IS_CORRUPT(mp, i != 1)) {
1111			xfs_btree_mark_sick(cur);
1112			error = -EFSCORRUPTED;
1113			goto error0;
1114		}
1115
1116		error = xfs_inobt_get_rec(cur, &rec, &j);
1117		if (error)
1118			goto error0;
1119		if (XFS_IS_CORRUPT(mp, j != 1)) {
1120			xfs_btree_mark_sick(cur);
1121			error = -EFSCORRUPTED;
1122			goto error0;
1123		}
1124
1125		if (rec.ir_freecount > 0) {
1126			/*
1127			 * Found a free inode in the same chunk
1128			 * as the parent, done.
1129			 */
1130			goto alloc_inode;
1131		}
1132
1133
1134		/*
1135		 * In the same AG as parent, but parent's chunk is full.
1136		 */
1137
1138		/* duplicate the cursor, search left & right simultaneously */
1139		error = xfs_btree_dup_cursor(cur, &tcur);
1140		if (error)
1141			goto error0;
1142
1143		/*
1144		 * Skip to last blocks looked up if same parent inode.
1145		 */
1146		if (pagino != NULLAGINO &&
1147		    pag->pagl_pagino == pagino &&
1148		    pag->pagl_leftrec != NULLAGINO &&
1149		    pag->pagl_rightrec != NULLAGINO) {
1150			error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1151						   &trec, &doneleft);
1152			if (error)
1153				goto error1;
1154
1155			error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1156						   &rec, &doneright);
1157			if (error)
1158				goto error1;
1159		} else {
1160			/* search left with tcur, back up 1 record */
1161			error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1162			if (error)
1163				goto error1;
1164
1165			/* search right with cur, go forward 1 record. */
1166			error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1167			if (error)
1168				goto error1;
1169		}
1170
1171		/*
1172		 * Loop until we find an inode chunk with a free inode.
1173		 */
1174		while (--searchdistance > 0 && (!doneleft || !doneright)) {
1175			int	useleft;  /* using left inode chunk this time */
1176
1177			/* figure out the closer block if both are valid. */
1178			if (!doneleft && !doneright) {
1179				useleft = pagino -
1180				 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1181				  rec.ir_startino - pagino;
1182			} else {
1183				useleft = !doneleft;
1184			}
1185
1186			/* free inodes to the left? */
1187			if (useleft && trec.ir_freecount) {
1188				xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1189				cur = tcur;
1190
1191				pag->pagl_leftrec = trec.ir_startino;
1192				pag->pagl_rightrec = rec.ir_startino;
1193				pag->pagl_pagino = pagino;
1194				rec = trec;
1195				goto alloc_inode;
1196			}
1197
1198			/* free inodes to the right? */
1199			if (!useleft && rec.ir_freecount) {
1200				xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1201
1202				pag->pagl_leftrec = trec.ir_startino;
1203				pag->pagl_rightrec = rec.ir_startino;
1204				pag->pagl_pagino = pagino;
1205				goto alloc_inode;
1206			}
1207
1208			/* get next record to check */
1209			if (useleft) {
1210				error = xfs_ialloc_next_rec(tcur, &trec,
1211								 &doneleft, 1);
1212			} else {
1213				error = xfs_ialloc_next_rec(cur, &rec,
1214								 &doneright, 0);
1215			}
1216			if (error)
1217				goto error1;
1218		}
1219
1220		if (searchdistance <= 0) {
1221			/*
1222			 * Not in range - save last search
1223			 * location and allocate a new inode
1224			 */
1225			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1226			pag->pagl_leftrec = trec.ir_startino;
1227			pag->pagl_rightrec = rec.ir_startino;
1228			pag->pagl_pagino = pagino;
1229
1230		} else {
1231			/*
1232			 * We've reached the end of the btree. because
1233			 * we are only searching a small chunk of the
1234			 * btree each search, there is obviously free
1235			 * inodes closer to the parent inode than we
1236			 * are now. restart the search again.
1237			 */
1238			pag->pagl_pagino = NULLAGINO;
1239			pag->pagl_leftrec = NULLAGINO;
1240			pag->pagl_rightrec = NULLAGINO;
1241			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1242			xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1243			goto restart_pagno;
1244		}
1245	}
1246
1247	/*
1248	 * In a different AG from the parent.
1249	 * See if the most recently allocated block has any free.
1250	 */
1251	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1252		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1253					 XFS_LOOKUP_EQ, &i);
1254		if (error)
1255			goto error0;
1256
1257		if (i == 1) {
1258			error = xfs_inobt_get_rec(cur, &rec, &j);
1259			if (error)
1260				goto error0;
1261
1262			if (j == 1 && rec.ir_freecount > 0) {
1263				/*
1264				 * The last chunk allocated in the group
1265				 * still has a free inode.
1266				 */
1267				goto alloc_inode;
1268			}
1269		}
1270	}
1271
1272	/*
1273	 * None left in the last group, search the whole AG
1274	 */
1275	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1276	if (error)
1277		goto error0;
1278	if (XFS_IS_CORRUPT(mp, i != 1)) {
1279		xfs_btree_mark_sick(cur);
1280		error = -EFSCORRUPTED;
1281		goto error0;
1282	}
1283
1284	for (;;) {
1285		error = xfs_inobt_get_rec(cur, &rec, &i);
1286		if (error)
1287			goto error0;
1288		if (XFS_IS_CORRUPT(mp, i != 1)) {
1289			xfs_btree_mark_sick(cur);
1290			error = -EFSCORRUPTED;
1291			goto error0;
1292		}
1293		if (rec.ir_freecount > 0)
1294			break;
1295		error = xfs_btree_increment(cur, 0, &i);
1296		if (error)
1297			goto error0;
1298		if (XFS_IS_CORRUPT(mp, i != 1)) {
1299			xfs_btree_mark_sick(cur);
1300			error = -EFSCORRUPTED;
1301			goto error0;
1302		}
1303	}
1304
1305alloc_inode:
1306	offset = xfs_inobt_first_free_inode(&rec);
1307	ASSERT(offset >= 0);
1308	ASSERT(offset < XFS_INODES_PER_CHUNK);
1309	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1310				   XFS_INODES_PER_CHUNK) == 0);
1311	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1312	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1313	rec.ir_freecount--;
1314	error = xfs_inobt_update(cur, &rec);
1315	if (error)
1316		goto error0;
1317	be32_add_cpu(&agi->agi_freecount, -1);
1318	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1319	pag->pagi_freecount--;
1320
1321	error = xfs_check_agi_freecount(cur);
1322	if (error)
1323		goto error0;
1324
1325	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1326	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1327	*inop = ino;
1328	return 0;
1329error1:
1330	xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1331error0:
1332	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1333	return error;
1334}
1335
1336/*
1337 * Use the free inode btree to allocate an inode based on distance from the
1338 * parent. Note that the provided cursor may be deleted and replaced.
1339 */
1340STATIC int
1341xfs_dialloc_ag_finobt_near(
1342	xfs_agino_t			pagino,
1343	struct xfs_btree_cur		**ocur,
1344	struct xfs_inobt_rec_incore	*rec)
1345{
1346	struct xfs_btree_cur		*lcur = *ocur;	/* left search cursor */
1347	struct xfs_btree_cur		*rcur;	/* right search cursor */
1348	struct xfs_inobt_rec_incore	rrec;
1349	int				error;
1350	int				i, j;
1351
1352	error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1353	if (error)
1354		return error;
1355
1356	if (i == 1) {
1357		error = xfs_inobt_get_rec(lcur, rec, &i);
1358		if (error)
1359			return error;
1360		if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1)) {
1361			xfs_btree_mark_sick(lcur);
1362			return -EFSCORRUPTED;
1363		}
1364
1365		/*
1366		 * See if we've landed in the parent inode record. The finobt
1367		 * only tracks chunks with at least one free inode, so record
1368		 * existence is enough.
1369		 */
1370		if (pagino >= rec->ir_startino &&
1371		    pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1372			return 0;
1373	}
1374
1375	error = xfs_btree_dup_cursor(lcur, &rcur);
1376	if (error)
1377		return error;
1378
1379	error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1380	if (error)
1381		goto error_rcur;
1382	if (j == 1) {
1383		error = xfs_inobt_get_rec(rcur, &rrec, &j);
1384		if (error)
1385			goto error_rcur;
1386		if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
1387			xfs_btree_mark_sick(lcur);
1388			error = -EFSCORRUPTED;
1389			goto error_rcur;
1390		}
1391	}
1392
1393	if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
1394		xfs_btree_mark_sick(lcur);
1395		error = -EFSCORRUPTED;
1396		goto error_rcur;
1397	}
1398	if (i == 1 && j == 1) {
1399		/*
1400		 * Both the left and right records are valid. Choose the closer
1401		 * inode chunk to the target.
1402		 */
1403		if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1404		    (rrec.ir_startino - pagino)) {
1405			*rec = rrec;
1406			xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1407			*ocur = rcur;
1408		} else {
1409			xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1410		}
1411	} else if (j == 1) {
1412		/* only the right record is valid */
1413		*rec = rrec;
1414		xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1415		*ocur = rcur;
1416	} else if (i == 1) {
1417		/* only the left record is valid */
1418		xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1419	}
1420
1421	return 0;
1422
1423error_rcur:
1424	xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1425	return error;
1426}
1427
1428/*
1429 * Use the free inode btree to find a free inode based on a newino hint. If
1430 * the hint is NULL, find the first free inode in the AG.
1431 */
1432STATIC int
1433xfs_dialloc_ag_finobt_newino(
1434	struct xfs_agi			*agi,
1435	struct xfs_btree_cur		*cur,
1436	struct xfs_inobt_rec_incore	*rec)
1437{
1438	int error;
1439	int i;
1440
1441	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1442		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1443					 XFS_LOOKUP_EQ, &i);
1444		if (error)
1445			return error;
1446		if (i == 1) {
1447			error = xfs_inobt_get_rec(cur, rec, &i);
1448			if (error)
1449				return error;
1450			if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1451				xfs_btree_mark_sick(cur);
1452				return -EFSCORRUPTED;
1453			}
1454			return 0;
1455		}
1456	}
1457
1458	/*
1459	 * Find the first inode available in the AG.
1460	 */
1461	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1462	if (error)
1463		return error;
1464	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1465		xfs_btree_mark_sick(cur);
1466		return -EFSCORRUPTED;
1467	}
1468
1469	error = xfs_inobt_get_rec(cur, rec, &i);
1470	if (error)
1471		return error;
1472	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1473		xfs_btree_mark_sick(cur);
1474		return -EFSCORRUPTED;
1475	}
1476
1477	return 0;
1478}
1479
1480/*
1481 * Update the inobt based on a modification made to the finobt. Also ensure that
1482 * the records from both trees are equivalent post-modification.
1483 */
1484STATIC int
1485xfs_dialloc_ag_update_inobt(
1486	struct xfs_btree_cur		*cur,	/* inobt cursor */
1487	struct xfs_inobt_rec_incore	*frec,	/* finobt record */
1488	int				offset) /* inode offset */
1489{
1490	struct xfs_inobt_rec_incore	rec;
1491	int				error;
1492	int				i;
1493
1494	error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1495	if (error)
1496		return error;
1497	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1498		xfs_btree_mark_sick(cur);
1499		return -EFSCORRUPTED;
1500	}
1501
1502	error = xfs_inobt_get_rec(cur, &rec, &i);
1503	if (error)
1504		return error;
1505	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1)) {
1506		xfs_btree_mark_sick(cur);
1507		return -EFSCORRUPTED;
1508	}
1509	ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1510				   XFS_INODES_PER_CHUNK) == 0);
1511
1512	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1513	rec.ir_freecount--;
1514
1515	if (XFS_IS_CORRUPT(cur->bc_mp,
1516			   rec.ir_free != frec->ir_free ||
1517			   rec.ir_freecount != frec->ir_freecount)) {
1518		xfs_btree_mark_sick(cur);
1519		return -EFSCORRUPTED;
1520	}
1521
1522	return xfs_inobt_update(cur, &rec);
1523}
1524
1525/*
1526 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1527 * back to the inobt search algorithm.
1528 *
1529 * The caller selected an AG for us, and made sure that free inodes are
1530 * available.
1531 */
1532static int
1533xfs_dialloc_ag(
1534	struct xfs_perag	*pag,
1535	struct xfs_trans	*tp,
1536	struct xfs_buf		*agbp,
1537	xfs_ino_t		parent,
1538	xfs_ino_t		*inop)
1539{
1540	struct xfs_mount		*mp = tp->t_mountp;
1541	struct xfs_agi			*agi = agbp->b_addr;
 
1542	xfs_agnumber_t			pagno = XFS_INO_TO_AGNO(mp, parent);
1543	xfs_agino_t			pagino = XFS_INO_TO_AGINO(mp, parent);
1544	struct xfs_btree_cur		*cur;	/* finobt cursor */
1545	struct xfs_btree_cur		*icur;	/* inobt cursor */
1546	struct xfs_inobt_rec_incore	rec;
1547	xfs_ino_t			ino;
1548	int				error;
1549	int				offset;
1550	int				i;
1551
1552	if (!xfs_has_finobt(mp))
1553		return xfs_dialloc_ag_inobt(pag, tp, agbp, parent, inop);
1554
1555	/*
1556	 * If pagino is 0 (this is the root inode allocation) use newino.
1557	 * This must work because we've just allocated some.
1558	 */
1559	if (!pagino)
1560		pagino = be32_to_cpu(agi->agi_newino);
1561
1562	cur = xfs_finobt_init_cursor(pag, tp, agbp);
1563
1564	error = xfs_check_agi_freecount(cur);
1565	if (error)
1566		goto error_cur;
1567
1568	/*
1569	 * The search algorithm depends on whether we're in the same AG as the
1570	 * parent. If so, find the closest available inode to the parent. If
1571	 * not, consider the agi hint or find the first free inode in the AG.
1572	 */
1573	if (pag->pag_agno == pagno)
1574		error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1575	else
1576		error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1577	if (error)
1578		goto error_cur;
1579
1580	offset = xfs_inobt_first_free_inode(&rec);
1581	ASSERT(offset >= 0);
1582	ASSERT(offset < XFS_INODES_PER_CHUNK);
1583	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1584				   XFS_INODES_PER_CHUNK) == 0);
1585	ino = XFS_AGINO_TO_INO(mp, pag->pag_agno, rec.ir_startino + offset);
1586
1587	/*
1588	 * Modify or remove the finobt record.
1589	 */
1590	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1591	rec.ir_freecount--;
1592	if (rec.ir_freecount)
1593		error = xfs_inobt_update(cur, &rec);
1594	else
1595		error = xfs_btree_delete(cur, &i);
1596	if (error)
1597		goto error_cur;
1598
1599	/*
1600	 * The finobt has now been updated appropriately. We haven't updated the
1601	 * agi and superblock yet, so we can create an inobt cursor and validate
1602	 * the original freecount. If all is well, make the equivalent update to
1603	 * the inobt using the finobt record and offset information.
1604	 */
1605	icur = xfs_inobt_init_cursor(pag, tp, agbp);
1606
1607	error = xfs_check_agi_freecount(icur);
1608	if (error)
1609		goto error_icur;
1610
1611	error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1612	if (error)
1613		goto error_icur;
1614
1615	/*
1616	 * Both trees have now been updated. We must update the perag and
1617	 * superblock before we can check the freecount for each btree.
1618	 */
1619	be32_add_cpu(&agi->agi_freecount, -1);
1620	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1621	pag->pagi_freecount--;
1622
1623	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1624
1625	error = xfs_check_agi_freecount(icur);
1626	if (error)
1627		goto error_icur;
1628	error = xfs_check_agi_freecount(cur);
1629	if (error)
1630		goto error_icur;
1631
1632	xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1633	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1634	*inop = ino;
1635	return 0;
1636
1637error_icur:
1638	xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1639error_cur:
1640	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1641	return error;
1642}
1643
1644static int
1645xfs_dialloc_roll(
1646	struct xfs_trans	**tpp,
1647	struct xfs_buf		*agibp)
1648{
1649	struct xfs_trans	*tp = *tpp;
1650	struct xfs_dquot_acct	*dqinfo;
1651	int			error;
1652
1653	/*
1654	 * Hold to on to the agibp across the commit so no other allocation can
1655	 * come in and take the free inodes we just allocated for our caller.
1656	 */
1657	xfs_trans_bhold(tp, agibp);
1658
1659	/*
1660	 * We want the quota changes to be associated with the next transaction,
1661	 * NOT this one. So, detach the dqinfo from this and attach it to the
1662	 * next transaction.
1663	 */
1664	dqinfo = tp->t_dqinfo;
1665	tp->t_dqinfo = NULL;
1666
1667	error = xfs_trans_roll(&tp);
1668
1669	/* Re-attach the quota info that we detached from prev trx. */
1670	tp->t_dqinfo = dqinfo;
1671
1672	/*
1673	 * Join the buffer even on commit error so that the buffer is released
1674	 * when the caller cancels the transaction and doesn't have to handle
1675	 * this error case specially.
1676	 */
1677	xfs_trans_bjoin(tp, agibp);
1678	*tpp = tp;
1679	return error;
1680}
1681
1682static bool
1683xfs_dialloc_good_ag(
1684	struct xfs_perag	*pag,
1685	struct xfs_trans	*tp,
1686	umode_t			mode,
1687	int			flags,
1688	bool			ok_alloc)
1689{
1690	struct xfs_mount	*mp = tp->t_mountp;
1691	xfs_extlen_t		ineed;
1692	xfs_extlen_t		longest = 0;
1693	int			needspace;
1694	int			error;
1695
1696	if (!pag)
1697		return false;
1698	if (!xfs_perag_allows_inodes(pag))
1699		return false;
1700
1701	if (!xfs_perag_initialised_agi(pag)) {
1702		error = xfs_ialloc_read_agi(pag, tp, NULL);
1703		if (error)
1704			return false;
1705	}
1706
1707	if (pag->pagi_freecount)
1708		return true;
1709	if (!ok_alloc)
1710		return false;
1711
1712	if (!xfs_perag_initialised_agf(pag)) {
1713		error = xfs_alloc_read_agf(pag, tp, flags, NULL);
1714		if (error)
1715			return false;
1716	}
1717
1718	/*
1719	 * Check that there is enough free space for the file plus a chunk of
1720	 * inodes if we need to allocate some. If this is the first pass across
1721	 * the AGs, take into account the potential space needed for alignment
1722	 * of inode chunks when checking the longest contiguous free space in
1723	 * the AG - this prevents us from getting ENOSPC because we have free
1724	 * space larger than ialloc_blks but alignment constraints prevent us
1725	 * from using it.
1726	 *
1727	 * If we can't find an AG with space for full alignment slack to be
1728	 * taken into account, we must be near ENOSPC in all AGs.  Hence we
1729	 * don't include alignment for the second pass and so if we fail
1730	 * allocation due to alignment issues then it is most likely a real
1731	 * ENOSPC condition.
1732	 *
1733	 * XXX(dgc): this calculation is now bogus thanks to the per-ag
1734	 * reservations that xfs_alloc_fix_freelist() now does via
1735	 * xfs_alloc_space_available(). When the AG fills up, pagf_freeblks will
1736	 * be more than large enough for the check below to succeed, but
1737	 * xfs_alloc_space_available() will fail because of the non-zero
1738	 * metadata reservation and hence we won't actually be able to allocate
1739	 * more inodes in this AG. We do soooo much unnecessary work near ENOSPC
1740	 * because of this.
1741	 */
1742	ineed = M_IGEO(mp)->ialloc_min_blks;
1743	if (flags && ineed > 1)
1744		ineed += M_IGEO(mp)->cluster_align;
1745	longest = pag->pagf_longest;
1746	if (!longest)
1747		longest = pag->pagf_flcount > 0;
1748	needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
1749
1750	if (pag->pagf_freeblks < needspace + ineed || longest < ineed)
1751		return false;
1752	return true;
1753}
1754
1755static int
1756xfs_dialloc_try_ag(
1757	struct xfs_perag	*pag,
1758	struct xfs_trans	**tpp,
1759	xfs_ino_t		parent,
1760	xfs_ino_t		*new_ino,
1761	bool			ok_alloc)
1762{
1763	struct xfs_buf		*agbp;
1764	xfs_ino_t		ino;
1765	int			error;
1766
1767	/*
1768	 * Then read in the AGI buffer and recheck with the AGI buffer
1769	 * lock held.
1770	 */
1771	error = xfs_ialloc_read_agi(pag, *tpp, &agbp);
1772	if (error)
1773		return error;
1774
1775	if (!pag->pagi_freecount) {
1776		if (!ok_alloc) {
1777			error = -EAGAIN;
1778			goto out_release;
1779		}
1780
1781		error = xfs_ialloc_ag_alloc(pag, *tpp, agbp);
1782		if (error < 0)
1783			goto out_release;
1784
1785		/*
1786		 * We successfully allocated space for an inode cluster in this
1787		 * AG.  Roll the transaction so that we can allocate one of the
1788		 * new inodes.
1789		 */
1790		ASSERT(pag->pagi_freecount > 0);
1791		error = xfs_dialloc_roll(tpp, agbp);
1792		if (error)
1793			goto out_release;
1794	}
1795
1796	/* Allocate an inode in the found AG */
1797	error = xfs_dialloc_ag(pag, *tpp, agbp, parent, &ino);
1798	if (!error)
1799		*new_ino = ino;
1800	return error;
1801
1802out_release:
1803	xfs_trans_brelse(*tpp, agbp);
1804	return error;
1805}
1806
1807/*
1808 * Allocate an on-disk inode.
 
 
 
1809 *
1810 * Mode is used to tell whether the new inode is a directory and hence where to
1811 * locate it. The on-disk inode that is allocated will be returned in @new_ino
1812 * on success, otherwise an error will be set to indicate the failure (e.g.
1813 * -ENOSPC).
 
 
 
 
 
 
 
 
 
 
1814 */
1815int
1816xfs_dialloc(
1817	struct xfs_trans	**tpp,
1818	xfs_ino_t		parent,
1819	umode_t			mode,
1820	xfs_ino_t		*new_ino)
 
1821{
1822	struct xfs_mount	*mp = (*tpp)->t_mountp;
 
1823	xfs_agnumber_t		agno;
1824	int			error = 0;
 
 
1825	xfs_agnumber_t		start_agno;
1826	struct xfs_perag	*pag;
1827	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1828	bool			ok_alloc = true;
1829	bool			low_space = false;
1830	int			flags;
1831	xfs_ino_t		ino = NULLFSINO;
 
 
 
 
 
 
 
1832
1833	/*
1834	 * Directories, symlinks, and regular files frequently allocate at least
1835	 * one block, so factor that potential expansion when we examine whether
1836	 * an AG has enough space for file creation.
1837	 */
1838	if (S_ISDIR(mode))
1839		start_agno = (atomic_inc_return(&mp->m_agirotor) - 1) %
1840				mp->m_maxagi;
1841	else {
1842		start_agno = XFS_INO_TO_AGNO(mp, parent);
1843		if (start_agno >= mp->m_maxagi)
1844			start_agno = 0;
1845	}
1846
1847	/*
1848	 * If we have already hit the ceiling of inode blocks then clear
1849	 * ok_alloc so we scan all available agi structures for a free
1850	 * inode.
1851	 *
1852	 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1853	 * which will sacrifice the preciseness but improve the performance.
1854	 */
1855	if (igeo->maxicount &&
1856	    percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1857							> igeo->maxicount) {
1858		ok_alloc = false;
1859	}
1860
1861	/*
1862	 * If we are near to ENOSPC, we want to prefer allocation from AGs that
1863	 * have free inodes in them rather than use up free space allocating new
1864	 * inode chunks. Hence we turn off allocation for the first non-blocking
1865	 * pass through the AGs if we are near ENOSPC to consume free inodes
1866	 * that we can immediately allocate, but then we allow allocation on the
1867	 * second pass if we fail to find an AG with free inodes in it.
1868	 */
1869	if (percpu_counter_read_positive(&mp->m_fdblocks) <
1870			mp->m_low_space[XFS_LOWSP_1_PCNT]) {
1871		ok_alloc = false;
1872		low_space = true;
1873	}
1874
1875	/*
1876	 * Loop until we find an allocation group that either has free inodes
1877	 * or in which we can allocate some inodes.  Iterate through the
1878	 * allocation groups upward, wrapping at the end.
1879	 */
1880	flags = XFS_ALLOC_FLAG_TRYLOCK;
1881retry:
1882	for_each_perag_wrap_at(mp, start_agno, mp->m_maxagi, agno, pag) {
1883		if (xfs_dialloc_good_ag(pag, *tpp, mode, flags, ok_alloc)) {
1884			error = xfs_dialloc_try_ag(pag, tpp, parent,
1885					&ino, ok_alloc);
1886			if (error != -EAGAIN)
1887				break;
1888			error = 0;
1889		}
1890
1891		if (xfs_is_shutdown(mp)) {
1892			error = -EFSCORRUPTED;
1893			break;
 
1894		}
1895	}
1896	if (pag)
1897		xfs_perag_rele(pag);
1898	if (error)
1899		return error;
1900	if (ino == NULLFSINO) {
1901		if (flags) {
1902			flags = 0;
1903			if (low_space)
1904				ok_alloc = true;
1905			goto retry;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1906		}
1907		return -ENOSPC;
1908	}
1909	*new_ino = ino;
1910	return 0;
 
 
 
 
 
1911}
1912
1913/*
1914 * Free the blocks of an inode chunk. We must consider that the inode chunk
1915 * might be sparse and only free the regions that are allocated as part of the
1916 * chunk.
1917 */
1918static int
1919xfs_difree_inode_chunk(
1920	struct xfs_trans		*tp,
1921	xfs_agnumber_t			agno,
1922	struct xfs_inobt_rec_incore	*rec)
1923{
1924	struct xfs_mount		*mp = tp->t_mountp;
1925	xfs_agblock_t			sagbno = XFS_AGINO_TO_AGBNO(mp,
1926							rec->ir_startino);
1927	int				startidx, endidx;
1928	int				nextbit;
1929	xfs_agblock_t			agbno;
1930	int				contigblk;
1931	DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1932
1933	if (!xfs_inobt_issparse(rec->ir_holemask)) {
1934		/* not sparse, calculate extent info directly */
1935		return xfs_free_extent_later(tp,
1936				XFS_AGB_TO_FSB(mp, agno, sagbno),
1937				M_IGEO(mp)->ialloc_blks, &XFS_RMAP_OINFO_INODES,
1938				XFS_AG_RESV_NONE, false);
1939	}
1940
1941	/* holemask is only 16-bits (fits in an unsigned long) */
1942	ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1943	holemask[0] = rec->ir_holemask;
1944
1945	/*
1946	 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1947	 * holemask and convert the start/end index of each range to an extent.
1948	 * We start with the start and end index both pointing at the first 0 in
1949	 * the mask.
1950	 */
1951	startidx = endidx = find_first_zero_bit(holemask,
1952						XFS_INOBT_HOLEMASK_BITS);
1953	nextbit = startidx + 1;
1954	while (startidx < XFS_INOBT_HOLEMASK_BITS) {
1955		int error;
1956
1957		nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1958					     nextbit);
1959		/*
1960		 * If the next zero bit is contiguous, update the end index of
1961		 * the current range and continue.
1962		 */
1963		if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1964		    nextbit == endidx + 1) {
1965			endidx = nextbit;
1966			goto next;
1967		}
1968
1969		/*
1970		 * nextbit is not contiguous with the current end index. Convert
1971		 * the current start/end to an extent and add it to the free
1972		 * list.
1973		 */
1974		agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1975				  mp->m_sb.sb_inopblock;
1976		contigblk = ((endidx - startidx + 1) *
1977			     XFS_INODES_PER_HOLEMASK_BIT) /
1978			    mp->m_sb.sb_inopblock;
1979
1980		ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1981		ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1982		error = xfs_free_extent_later(tp,
1983				XFS_AGB_TO_FSB(mp, agno, agbno), contigblk,
1984				&XFS_RMAP_OINFO_INODES, XFS_AG_RESV_NONE,
1985				false);
1986		if (error)
1987			return error;
1988
1989		/* reset range to current bit and carry on... */
1990		startidx = endidx = nextbit;
1991
1992next:
1993		nextbit++;
1994	}
1995	return 0;
1996}
1997
1998STATIC int
1999xfs_difree_inobt(
2000	struct xfs_perag		*pag,
2001	struct xfs_trans		*tp,
2002	struct xfs_buf			*agbp,
2003	xfs_agino_t			agino,
2004	struct xfs_icluster		*xic,
2005	struct xfs_inobt_rec_incore	*orec)
2006{
2007	struct xfs_mount		*mp = pag->pag_mount;
2008	struct xfs_agi			*agi = agbp->b_addr;
 
2009	struct xfs_btree_cur		*cur;
2010	struct xfs_inobt_rec_incore	rec;
2011	int				ilen;
2012	int				error;
2013	int				i;
2014	int				off;
2015
2016	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2017	ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
2018
2019	/*
2020	 * Initialize the cursor.
2021	 */
2022	cur = xfs_inobt_init_cursor(pag, tp, agbp);
2023
2024	error = xfs_check_agi_freecount(cur);
2025	if (error)
2026		goto error0;
2027
2028	/*
2029	 * Look for the entry describing this inode.
2030	 */
2031	if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
2032		xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
2033			__func__, error);
2034		goto error0;
2035	}
2036	if (XFS_IS_CORRUPT(mp, i != 1)) {
2037		xfs_btree_mark_sick(cur);
2038		error = -EFSCORRUPTED;
2039		goto error0;
2040	}
2041	error = xfs_inobt_get_rec(cur, &rec, &i);
2042	if (error) {
2043		xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
2044			__func__, error);
2045		goto error0;
2046	}
2047	if (XFS_IS_CORRUPT(mp, i != 1)) {
2048		xfs_btree_mark_sick(cur);
2049		error = -EFSCORRUPTED;
2050		goto error0;
2051	}
2052	/*
2053	 * Get the offset in the inode chunk.
2054	 */
2055	off = agino - rec.ir_startino;
2056	ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
2057	ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
2058	/*
2059	 * Mark the inode free & increment the count.
2060	 */
2061	rec.ir_free |= XFS_INOBT_MASK(off);
2062	rec.ir_freecount++;
2063
2064	/*
2065	 * When an inode chunk is free, it becomes eligible for removal. Don't
2066	 * remove the chunk if the block size is large enough for multiple inode
2067	 * chunks (that might not be free).
2068	 */
2069	if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
 
2070	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
 
 
2071		xic->deleted = true;
2072		xic->first_ino = XFS_AGINO_TO_INO(mp, pag->pag_agno,
2073				rec.ir_startino);
2074		xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
2075
2076		/*
2077		 * Remove the inode cluster from the AGI B+Tree, adjust the
2078		 * AGI and Superblock inode counts, and mark the disk space
2079		 * to be freed when the transaction is committed.
2080		 */
2081		ilen = rec.ir_freecount;
2082		be32_add_cpu(&agi->agi_count, -ilen);
2083		be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
2084		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
2085		pag->pagi_freecount -= ilen - 1;
2086		pag->pagi_count -= ilen;
2087		xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
2088		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
2089
2090		if ((error = xfs_btree_delete(cur, &i))) {
2091			xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
2092				__func__, error);
2093			goto error0;
2094		}
2095
2096		error = xfs_difree_inode_chunk(tp, pag->pag_agno, &rec);
2097		if (error)
2098			goto error0;
2099	} else {
2100		xic->deleted = false;
2101
2102		error = xfs_inobt_update(cur, &rec);
2103		if (error) {
2104			xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
2105				__func__, error);
2106			goto error0;
2107		}
2108
2109		/*
2110		 * Change the inode free counts and log the ag/sb changes.
2111		 */
2112		be32_add_cpu(&agi->agi_freecount, 1);
2113		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2114		pag->pagi_freecount++;
2115		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2116	}
2117
2118	error = xfs_check_agi_freecount(cur);
2119	if (error)
2120		goto error0;
2121
2122	*orec = rec;
2123	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2124	return 0;
2125
2126error0:
2127	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2128	return error;
2129}
2130
2131/*
2132 * Free an inode in the free inode btree.
2133 */
2134STATIC int
2135xfs_difree_finobt(
2136	struct xfs_perag		*pag,
2137	struct xfs_trans		*tp,
2138	struct xfs_buf			*agbp,
2139	xfs_agino_t			agino,
2140	struct xfs_inobt_rec_incore	*ibtrec) /* inobt record */
2141{
2142	struct xfs_mount		*mp = pag->pag_mount;
 
2143	struct xfs_btree_cur		*cur;
2144	struct xfs_inobt_rec_incore	rec;
2145	int				offset = agino - ibtrec->ir_startino;
2146	int				error;
2147	int				i;
2148
2149	cur = xfs_finobt_init_cursor(pag, tp, agbp);
2150
2151	error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2152	if (error)
2153		goto error;
2154	if (i == 0) {
2155		/*
2156		 * If the record does not exist in the finobt, we must have just
2157		 * freed an inode in a previously fully allocated chunk. If not,
2158		 * something is out of sync.
2159		 */
2160		if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
2161			xfs_btree_mark_sick(cur);
2162			error = -EFSCORRUPTED;
2163			goto error;
2164		}
2165
2166		error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2167					     ibtrec->ir_count,
2168					     ibtrec->ir_freecount,
2169					     ibtrec->ir_free, &i);
2170		if (error)
2171			goto error;
2172		ASSERT(i == 1);
2173
2174		goto out;
2175	}
2176
2177	/*
2178	 * Read and update the existing record. We could just copy the ibtrec
2179	 * across here, but that would defeat the purpose of having redundant
2180	 * metadata. By making the modifications independently, we can catch
2181	 * corruptions that we wouldn't see if we just copied from one record
2182	 * to another.
2183	 */
2184	error = xfs_inobt_get_rec(cur, &rec, &i);
2185	if (error)
2186		goto error;
2187	if (XFS_IS_CORRUPT(mp, i != 1)) {
2188		xfs_btree_mark_sick(cur);
2189		error = -EFSCORRUPTED;
2190		goto error;
2191	}
2192
2193	rec.ir_free |= XFS_INOBT_MASK(offset);
2194	rec.ir_freecount++;
2195
2196	if (XFS_IS_CORRUPT(mp,
2197			   rec.ir_free != ibtrec->ir_free ||
2198			   rec.ir_freecount != ibtrec->ir_freecount)) {
2199		xfs_btree_mark_sick(cur);
2200		error = -EFSCORRUPTED;
2201		goto error;
2202	}
2203
2204	/*
2205	 * The content of inobt records should always match between the inobt
2206	 * and finobt. The lifecycle of records in the finobt is different from
2207	 * the inobt in that the finobt only tracks records with at least one
2208	 * free inode. Hence, if all of the inodes are free and we aren't
2209	 * keeping inode chunks permanently on disk, remove the record.
2210	 * Otherwise, update the record with the new information.
2211	 *
2212	 * Note that we currently can't free chunks when the block size is large
2213	 * enough for multiple chunks. Leave the finobt record to remain in sync
2214	 * with the inobt.
2215	 */
2216	if (!xfs_has_ikeep(mp) && rec.ir_free == XFS_INOBT_ALL_FREE &&
2217	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
 
2218		error = xfs_btree_delete(cur, &i);
2219		if (error)
2220			goto error;
2221		ASSERT(i == 1);
2222	} else {
2223		error = xfs_inobt_update(cur, &rec);
2224		if (error)
2225			goto error;
2226	}
2227
2228out:
2229	error = xfs_check_agi_freecount(cur);
2230	if (error)
2231		goto error;
2232
2233	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2234	return 0;
2235
2236error:
2237	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2238	return error;
2239}
2240
2241/*
2242 * Free disk inode.  Carefully avoids touching the incore inode, all
2243 * manipulations incore are the caller's responsibility.
2244 * The on-disk inode is not changed by this operation, only the
2245 * btree (free inode mask) is changed.
2246 */
2247int
2248xfs_difree(
2249	struct xfs_trans	*tp,
2250	struct xfs_perag	*pag,
2251	xfs_ino_t		inode,
2252	struct xfs_icluster	*xic)
2253{
2254	/* REFERENCED */
2255	xfs_agblock_t		agbno;	/* block number containing inode */
2256	struct xfs_buf		*agbp;	/* buffer for allocation group header */
2257	xfs_agino_t		agino;	/* allocation group inode number */
 
2258	int			error;	/* error return value */
2259	struct xfs_mount	*mp = tp->t_mountp;
2260	struct xfs_inobt_rec_incore rec;/* btree record */
2261
 
 
2262	/*
2263	 * Break up inode number into its components.
2264	 */
2265	if (pag->pag_agno != XFS_INO_TO_AGNO(mp, inode)) {
2266		xfs_warn(mp, "%s: agno != pag->pag_agno (%d != %d).",
2267			__func__, XFS_INO_TO_AGNO(mp, inode), pag->pag_agno);
 
2268		ASSERT(0);
2269		return -EINVAL;
2270	}
2271	agino = XFS_INO_TO_AGINO(mp, inode);
2272	if (inode != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino))  {
2273		xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2274			__func__, (unsigned long long)inode,
2275			(unsigned long long)XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2276		ASSERT(0);
2277		return -EINVAL;
2278	}
2279	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2280	if (agbno >= mp->m_sb.sb_agblocks)  {
2281		xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2282			__func__, agbno, mp->m_sb.sb_agblocks);
2283		ASSERT(0);
2284		return -EINVAL;
2285	}
2286	/*
2287	 * Get the allocation group header.
2288	 */
2289	error = xfs_ialloc_read_agi(pag, tp, &agbp);
2290	if (error) {
2291		xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2292			__func__, error);
2293		return error;
2294	}
2295
2296	/*
2297	 * Fix up the inode allocation btree.
2298	 */
2299	error = xfs_difree_inobt(pag, tp, agbp, agino, xic, &rec);
2300	if (error)
2301		goto error0;
2302
2303	/*
2304	 * Fix up the free inode btree.
2305	 */
2306	if (xfs_has_finobt(mp)) {
2307		error = xfs_difree_finobt(pag, tp, agbp, agino, &rec);
2308		if (error)
2309			goto error0;
2310	}
2311
2312	return 0;
2313
2314error0:
2315	return error;
2316}
2317
2318STATIC int
2319xfs_imap_lookup(
2320	struct xfs_perag	*pag,
2321	struct xfs_trans	*tp,
 
2322	xfs_agino_t		agino,
2323	xfs_agblock_t		agbno,
2324	xfs_agblock_t		*chunk_agbno,
2325	xfs_agblock_t		*offset_agbno,
2326	int			flags)
2327{
2328	struct xfs_mount	*mp = pag->pag_mount;
2329	struct xfs_inobt_rec_incore rec;
2330	struct xfs_btree_cur	*cur;
2331	struct xfs_buf		*agbp;
2332	int			error;
2333	int			i;
2334
2335	error = xfs_ialloc_read_agi(pag, tp, &agbp);
2336	if (error) {
2337		xfs_alert(mp,
2338			"%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2339			__func__, error, pag->pag_agno);
2340		return error;
2341	}
2342
2343	/*
2344	 * Lookup the inode record for the given agino. If the record cannot be
2345	 * found, then it's an invalid inode number and we should abort. Once
2346	 * we have a record, we need to ensure it contains the inode number
2347	 * we are looking up.
2348	 */
2349	cur = xfs_inobt_init_cursor(pag, tp, agbp);
2350	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2351	if (!error) {
2352		if (i)
2353			error = xfs_inobt_get_rec(cur, &rec, &i);
2354		if (!error && i == 0)
2355			error = -EINVAL;
2356	}
2357
2358	xfs_trans_brelse(tp, agbp);
2359	xfs_btree_del_cursor(cur, error);
2360	if (error)
2361		return error;
2362
2363	/* check that the returned record contains the required inode */
2364	if (rec.ir_startino > agino ||
2365	    rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2366		return -EINVAL;
2367
2368	/* for untrusted inodes check it is allocated first */
2369	if ((flags & XFS_IGET_UNTRUSTED) &&
2370	    (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2371		return -EINVAL;
2372
2373	*chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2374	*offset_agbno = agbno - *chunk_agbno;
2375	return 0;
2376}
2377
2378/*
2379 * Return the location of the inode in imap, for mapping it into a buffer.
2380 */
2381int
2382xfs_imap(
2383	struct xfs_perag	*pag,
2384	struct xfs_trans	*tp,
2385	xfs_ino_t		ino,	/* inode to locate */
2386	struct xfs_imap		*imap,	/* location map structure */
2387	uint			flags)	/* flags for inode btree lookup */
2388{
2389	struct xfs_mount	*mp = pag->pag_mount;
2390	xfs_agblock_t		agbno;	/* block number of inode in the alloc group */
2391	xfs_agino_t		agino;	/* inode number within alloc group */
2392	xfs_agblock_t		chunk_agbno;	/* first block in inode chunk */
2393	xfs_agblock_t		cluster_agbno;	/* first block in inode cluster */
2394	int			error;	/* error code */
2395	int			offset;	/* index of inode in its buffer */
2396	xfs_agblock_t		offset_agbno;	/* blks from chunk start to inode */
2397
2398	ASSERT(ino != NULLFSINO);
2399
2400	/*
2401	 * Split up the inode number into its parts.
2402	 */
 
2403	agino = XFS_INO_TO_AGINO(mp, ino);
2404	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2405	if (agbno >= mp->m_sb.sb_agblocks ||
2406	    ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2407		error = -EINVAL;
2408#ifdef DEBUG
2409		/*
2410		 * Don't output diagnostic information for untrusted inodes
2411		 * as they can be invalid without implying corruption.
2412		 */
2413		if (flags & XFS_IGET_UNTRUSTED)
2414			return error;
 
 
 
 
 
2415		if (agbno >= mp->m_sb.sb_agblocks) {
2416			xfs_alert(mp,
2417		"%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2418				__func__, (unsigned long long)agbno,
2419				(unsigned long)mp->m_sb.sb_agblocks);
2420		}
2421		if (ino != XFS_AGINO_TO_INO(mp, pag->pag_agno, agino)) {
2422			xfs_alert(mp,
2423		"%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2424				__func__, ino,
2425				XFS_AGINO_TO_INO(mp, pag->pag_agno, agino));
2426		}
2427		xfs_stack_trace();
2428#endif /* DEBUG */
2429		return error;
2430	}
2431
2432	/*
2433	 * For bulkstat and handle lookups, we have an untrusted inode number
2434	 * that we have to verify is valid. We cannot do this just by reading
2435	 * the inode buffer as it may have been unlinked and removed leaving
2436	 * inodes in stale state on disk. Hence we have to do a btree lookup
2437	 * in all cases where an untrusted inode number is passed.
2438	 */
2439	if (flags & XFS_IGET_UNTRUSTED) {
2440		error = xfs_imap_lookup(pag, tp, agino, agbno,
2441					&chunk_agbno, &offset_agbno, flags);
2442		if (error)
2443			return error;
2444		goto out_map;
2445	}
2446
2447	/*
2448	 * If the inode cluster size is the same as the blocksize or
2449	 * smaller we get to the buffer by simple arithmetics.
2450	 */
2451	if (M_IGEO(mp)->blocks_per_cluster == 1) {
2452		offset = XFS_INO_TO_OFFSET(mp, ino);
2453		ASSERT(offset < mp->m_sb.sb_inopblock);
2454
2455		imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, agbno);
2456		imap->im_len = XFS_FSB_TO_BB(mp, 1);
2457		imap->im_boffset = (unsigned short)(offset <<
2458							mp->m_sb.sb_inodelog);
2459		return 0;
2460	}
2461
2462	/*
2463	 * If the inode chunks are aligned then use simple maths to
2464	 * find the location. Otherwise we have to do a btree
2465	 * lookup to find the location.
2466	 */
2467	if (M_IGEO(mp)->inoalign_mask) {
2468		offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2469		chunk_agbno = agbno - offset_agbno;
2470	} else {
2471		error = xfs_imap_lookup(pag, tp, agino, agbno,
2472					&chunk_agbno, &offset_agbno, flags);
2473		if (error)
2474			return error;
2475	}
2476
2477out_map:
2478	ASSERT(agbno >= chunk_agbno);
2479	cluster_agbno = chunk_agbno +
2480		((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2481		 M_IGEO(mp)->blocks_per_cluster);
2482	offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2483		XFS_INO_TO_OFFSET(mp, ino);
2484
2485	imap->im_blkno = XFS_AGB_TO_DADDR(mp, pag->pag_agno, cluster_agbno);
2486	imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2487	imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2488
2489	/*
2490	 * If the inode number maps to a block outside the bounds
2491	 * of the file system then return NULL rather than calling
2492	 * read_buf and panicing when we get an error from the
2493	 * driver.
2494	 */
2495	if ((imap->im_blkno + imap->im_len) >
2496	    XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2497		xfs_alert(mp,
2498	"%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2499			__func__, (unsigned long long) imap->im_blkno,
2500			(unsigned long long) imap->im_len,
2501			XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2502		return -EINVAL;
2503	}
2504	return 0;
2505}
2506
2507/*
2508 * Log specified fields for the ag hdr (inode section). The growth of the agi
2509 * structure over time requires that we interpret the buffer as two logical
2510 * regions delineated by the end of the unlinked list. This is due to the size
2511 * of the hash table and its location in the middle of the agi.
2512 *
2513 * For example, a request to log a field before agi_unlinked and a field after
2514 * agi_unlinked could cause us to log the entire hash table and use an excessive
2515 * amount of log space. To avoid this behavior, log the region up through
2516 * agi_unlinked in one call and the region after agi_unlinked through the end of
2517 * the structure in another.
2518 */
2519void
2520xfs_ialloc_log_agi(
2521	struct xfs_trans	*tp,
2522	struct xfs_buf		*bp,
2523	uint32_t		fields)
2524{
2525	int			first;		/* first byte number */
2526	int			last;		/* last byte number */
2527	static const short	offsets[] = {	/* field starting offsets */
2528					/* keep in sync with bit definitions */
2529		offsetof(xfs_agi_t, agi_magicnum),
2530		offsetof(xfs_agi_t, agi_versionnum),
2531		offsetof(xfs_agi_t, agi_seqno),
2532		offsetof(xfs_agi_t, agi_length),
2533		offsetof(xfs_agi_t, agi_count),
2534		offsetof(xfs_agi_t, agi_root),
2535		offsetof(xfs_agi_t, agi_level),
2536		offsetof(xfs_agi_t, agi_freecount),
2537		offsetof(xfs_agi_t, agi_newino),
2538		offsetof(xfs_agi_t, agi_dirino),
2539		offsetof(xfs_agi_t, agi_unlinked),
2540		offsetof(xfs_agi_t, agi_free_root),
2541		offsetof(xfs_agi_t, agi_free_level),
2542		offsetof(xfs_agi_t, agi_iblocks),
2543		sizeof(xfs_agi_t)
2544	};
2545#ifdef DEBUG
2546	struct xfs_agi		*agi = bp->b_addr;
2547
2548	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2549#endif
2550
2551	/*
2552	 * Compute byte offsets for the first and last fields in the first
2553	 * region and log the agi buffer. This only logs up through
2554	 * agi_unlinked.
2555	 */
2556	if (fields & XFS_AGI_ALL_BITS_R1) {
2557		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2558				  &first, &last);
2559		xfs_trans_log_buf(tp, bp, first, last);
2560	}
2561
2562	/*
2563	 * Mask off the bits in the first region and calculate the first and
2564	 * last field offsets for any bits in the second region.
2565	 */
2566	fields &= ~XFS_AGI_ALL_BITS_R1;
2567	if (fields) {
2568		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2569				  &first, &last);
2570		xfs_trans_log_buf(tp, bp, first, last);
2571	}
2572}
2573
2574static xfs_failaddr_t
2575xfs_agi_verify(
2576	struct xfs_buf		*bp)
2577{
2578	struct xfs_mount	*mp = bp->b_mount;
2579	struct xfs_agi		*agi = bp->b_addr;
2580	xfs_failaddr_t		fa;
2581	uint32_t		agi_seqno = be32_to_cpu(agi->agi_seqno);
2582	uint32_t		agi_length = be32_to_cpu(agi->agi_length);
2583	int			i;
2584
2585	if (xfs_has_crc(mp)) {
2586		if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2587			return __this_address;
2588		if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2589			return __this_address;
2590	}
2591
2592	/*
2593	 * Validate the magic number of the agi block.
2594	 */
2595	if (!xfs_verify_magic(bp, agi->agi_magicnum))
2596		return __this_address;
2597	if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2598		return __this_address;
2599
2600	fa = xfs_validate_ag_length(bp, agi_seqno, agi_length);
2601	if (fa)
2602		return fa;
2603
2604	if (be32_to_cpu(agi->agi_level) < 1 ||
2605	    be32_to_cpu(agi->agi_level) > M_IGEO(mp)->inobt_maxlevels)
2606		return __this_address;
2607
2608	if (xfs_has_finobt(mp) &&
2609	    (be32_to_cpu(agi->agi_free_level) < 1 ||
2610	     be32_to_cpu(agi->agi_free_level) > M_IGEO(mp)->inobt_maxlevels))
 
 
 
 
 
 
 
 
 
2611		return __this_address;
2612
2613	for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2614		if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2615			continue;
2616		if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2617			return __this_address;
2618	}
2619
2620	return NULL;
2621}
2622
2623static void
2624xfs_agi_read_verify(
2625	struct xfs_buf	*bp)
2626{
2627	struct xfs_mount *mp = bp->b_mount;
2628	xfs_failaddr_t	fa;
2629
2630	if (xfs_has_crc(mp) &&
2631	    !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2632		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2633	else {
2634		fa = xfs_agi_verify(bp);
2635		if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2636			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2637	}
2638}
2639
2640static void
2641xfs_agi_write_verify(
2642	struct xfs_buf	*bp)
2643{
2644	struct xfs_mount	*mp = bp->b_mount;
2645	struct xfs_buf_log_item	*bip = bp->b_log_item;
2646	struct xfs_agi		*agi = bp->b_addr;
2647	xfs_failaddr_t		fa;
2648
2649	fa = xfs_agi_verify(bp);
2650	if (fa) {
2651		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2652		return;
2653	}
2654
2655	if (!xfs_has_crc(mp))
2656		return;
2657
2658	if (bip)
2659		agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2660	xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2661}
2662
2663const struct xfs_buf_ops xfs_agi_buf_ops = {
2664	.name = "xfs_agi",
2665	.magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2666	.verify_read = xfs_agi_read_verify,
2667	.verify_write = xfs_agi_write_verify,
2668	.verify_struct = xfs_agi_verify,
2669};
2670
2671/*
2672 * Read in the allocation group header (inode allocation section)
2673 */
2674int
2675xfs_read_agi(
2676	struct xfs_perag	*pag,
2677	struct xfs_trans	*tp,
2678	struct xfs_buf		**agibpp)
 
2679{
2680	struct xfs_mount	*mp = pag->pag_mount;
2681	int			error;
2682
2683	trace_xfs_read_agi(pag->pag_mount, pag->pag_agno);
2684
 
2685	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2686			XFS_AG_DADDR(mp, pag->pag_agno, XFS_AGI_DADDR(mp)),
2687			XFS_FSS_TO_BB(mp, 1), 0, agibpp, &xfs_agi_buf_ops);
2688	if (xfs_metadata_is_sick(error))
2689		xfs_ag_mark_sick(pag, XFS_SICK_AG_AGI);
2690	if (error)
2691		return error;
2692	if (tp)
2693		xfs_trans_buf_set_type(tp, *agibpp, XFS_BLFT_AGI_BUF);
2694
2695	xfs_buf_set_ref(*agibpp, XFS_AGI_REF);
2696	return 0;
2697}
2698
2699/*
2700 * Read in the agi and initialise the per-ag data. If the caller supplies a
2701 * @agibpp, return the locked AGI buffer to them, otherwise release it.
2702 */
2703int
2704xfs_ialloc_read_agi(
2705	struct xfs_perag	*pag,
2706	struct xfs_trans	*tp,
2707	struct xfs_buf		**agibpp)
 
2708{
2709	struct xfs_buf		*agibp;
2710	struct xfs_agi		*agi;
2711	int			error;
2712
2713	trace_xfs_ialloc_read_agi(pag->pag_mount, pag->pag_agno);
2714
2715	error = xfs_read_agi(pag, tp, &agibp);
2716	if (error)
2717		return error;
2718
2719	agi = agibp->b_addr;
2720	if (!xfs_perag_initialised_agi(pag)) {
 
2721		pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2722		pag->pagi_count = be32_to_cpu(agi->agi_count);
2723		set_bit(XFS_AGSTATE_AGI_INIT, &pag->pag_opstate);
2724	}
2725
2726	/*
2727	 * It's possible for these to be out of sync if
2728	 * we are in the middle of a forced shutdown.
2729	 */
2730	ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2731		xfs_is_shutdown(pag->pag_mount));
2732	if (agibpp)
2733		*agibpp = agibp;
2734	else
2735		xfs_trans_brelse(tp, agibp);
2736	return 0;
2737}
2738
2739/* How many inodes are backed by inode clusters ondisk? */
2740STATIC int
2741xfs_ialloc_count_ondisk(
2742	struct xfs_btree_cur		*cur,
2743	xfs_agino_t			low,
2744	xfs_agino_t			high,
2745	unsigned int			*allocated)
 
2746{
2747	struct xfs_inobt_rec_incore	irec;
2748	unsigned int			ret = 0;
2749	int				has_record;
2750	int				error;
2751
2752	error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2753	if (error)
2754		return error;
 
 
 
 
2755
2756	while (has_record) {
2757		unsigned int		i, hole_idx;
 
 
 
 
 
 
 
 
 
 
 
 
2758
 
 
 
2759		error = xfs_inobt_get_rec(cur, &irec, &has_record);
2760		if (error)
2761			return error;
2762		if (irec.ir_startino > high)
2763			break;
2764
2765		for (i = 0; i < XFS_INODES_PER_CHUNK; i++) {
2766			if (irec.ir_startino + i < low)
 
 
 
2767				continue;
2768			if (irec.ir_startino + i > high)
2769				break;
2770
2771			hole_idx = i / XFS_INODES_PER_HOLEMASK_BIT;
2772			if (!(irec.ir_holemask & (1U << hole_idx)))
2773				ret++;
2774		}
2775
2776		error = xfs_btree_increment(cur, 0, &has_record);
2777		if (error)
2778			return error;
2779	}
2780
2781	*allocated = ret;
2782	return 0;
2783}
2784
2785/* Is there an inode record covering a given extent? */
2786int
2787xfs_ialloc_has_inodes_at_extent(
2788	struct xfs_btree_cur	*cur,
2789	xfs_agblock_t		bno,
2790	xfs_extlen_t		len,
2791	enum xbtree_recpacking	*outcome)
2792{
2793	xfs_agino_t		agino;
2794	xfs_agino_t		last_agino;
2795	unsigned int		allocated;
2796	int			error;
2797
2798	agino = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2799	last_agino = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2800
2801	error = xfs_ialloc_count_ondisk(cur, agino, last_agino, &allocated);
2802	if (error)
2803		return error;
2804
2805	if (allocated == 0)
2806		*outcome = XBTREE_RECPACKING_EMPTY;
2807	else if (allocated == last_agino - agino + 1)
2808		*outcome = XBTREE_RECPACKING_FULL;
2809	else
2810		*outcome = XBTREE_RECPACKING_SPARSE;
2811	return 0;
2812}
2813
2814struct xfs_ialloc_count_inodes {
2815	xfs_agino_t			count;
2816	xfs_agino_t			freecount;
2817};
2818
2819/* Record inode counts across all inobt records. */
2820STATIC int
2821xfs_ialloc_count_inodes_rec(
2822	struct xfs_btree_cur		*cur,
2823	const union xfs_btree_rec	*rec,
2824	void				*priv)
2825{
2826	struct xfs_inobt_rec_incore	irec;
2827	struct xfs_ialloc_count_inodes	*ci = priv;
2828	xfs_failaddr_t			fa;
2829
2830	xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
2831	fa = xfs_inobt_check_irec(cur->bc_ag.pag, &irec);
2832	if (fa)
2833		return xfs_inobt_complain_bad_rec(cur, fa, &irec);
2834
2835	ci->count += irec.ir_count;
2836	ci->freecount += irec.ir_freecount;
2837
2838	return 0;
2839}
2840
2841/* Count allocated and free inodes under an inobt. */
2842int
2843xfs_ialloc_count_inodes(
2844	struct xfs_btree_cur		*cur,
2845	xfs_agino_t			*count,
2846	xfs_agino_t			*freecount)
2847{
2848	struct xfs_ialloc_count_inodes	ci = {0};
2849	int				error;
2850
2851	ASSERT(xfs_btree_is_ino(cur->bc_ops));
2852	error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2853	if (error)
2854		return error;
2855
2856	*count = ci.count;
2857	*freecount = ci.freecount;
2858	return 0;
2859}
2860
2861/*
2862 * Initialize inode-related geometry information.
2863 *
2864 * Compute the inode btree min and max levels and set maxicount.
2865 *
2866 * Set the inode cluster size.  This may still be overridden by the file
2867 * system block size if it is larger than the chosen cluster size.
2868 *
2869 * For v5 filesystems, scale the cluster size with the inode size to keep a
2870 * constant ratio of inode per cluster buffer, but only if mkfs has set the
2871 * inode alignment value appropriately for larger cluster sizes.
2872 *
2873 * Then compute the inode cluster alignment information.
2874 */
2875void
2876xfs_ialloc_setup_geometry(
2877	struct xfs_mount	*mp)
2878{
2879	struct xfs_sb		*sbp = &mp->m_sb;
2880	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2881	uint64_t		icount;
2882	uint			inodes;
2883
2884	igeo->new_diflags2 = 0;
2885	if (xfs_has_bigtime(mp))
2886		igeo->new_diflags2 |= XFS_DIFLAG2_BIGTIME;
2887	if (xfs_has_large_extent_counts(mp))
2888		igeo->new_diflags2 |= XFS_DIFLAG2_NREXT64;
2889
2890	/* Compute inode btree geometry. */
2891	igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2892	igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2893	igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2894	igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2895	igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2896
2897	igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2898			sbp->sb_inopblock);
2899	igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2900
2901	if (sbp->sb_spino_align)
2902		igeo->ialloc_min_blks = sbp->sb_spino_align;
2903	else
2904		igeo->ialloc_min_blks = igeo->ialloc_blks;
2905
2906	/* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2907	inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2908	igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2909			inodes);
2910	ASSERT(igeo->inobt_maxlevels <= xfs_iallocbt_maxlevels_ondisk());
2911
2912	/*
2913	 * Set the maximum inode count for this filesystem, being careful not
2914	 * to use obviously garbage sb_inopblog/sb_inopblock values.  Regular
2915	 * users should never get here due to failing sb verification, but
2916	 * certain users (xfs_db) need to be usable even with corrupt metadata.
2917	 */
2918	if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2919		/*
2920		 * Make sure the maximum inode count is a multiple
2921		 * of the units we allocate inodes in.
2922		 */
2923		icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2924		do_div(icount, 100);
2925		do_div(icount, igeo->ialloc_blks);
2926		igeo->maxicount = XFS_FSB_TO_INO(mp,
2927				icount * igeo->ialloc_blks);
2928	} else {
2929		igeo->maxicount = 0;
2930	}
2931
2932	/*
2933	 * Compute the desired size of an inode cluster buffer size, which
2934	 * starts at 8K and (on v5 filesystems) scales up with larger inode
2935	 * sizes.
2936	 *
2937	 * Preserve the desired inode cluster size because the sparse inodes
2938	 * feature uses that desired size (not the actual size) to compute the
2939	 * sparse inode alignment.  The mount code validates this value, so we
2940	 * cannot change the behavior.
2941	 */
2942	igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2943	if (xfs_has_v3inodes(mp)) {
2944		int	new_size = igeo->inode_cluster_size_raw;
2945
2946		new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2947		if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2948			igeo->inode_cluster_size_raw = new_size;
2949	}
2950
2951	/* Calculate inode cluster ratios. */
2952	if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2953		igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2954				igeo->inode_cluster_size_raw);
2955	else
2956		igeo->blocks_per_cluster = 1;
2957	igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2958	igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2959
2960	/* Calculate inode cluster alignment. */
2961	if (xfs_has_align(mp) &&
2962	    mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2963		igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2964	else
2965		igeo->cluster_align = 1;
2966	igeo->inoalign_mask = igeo->cluster_align - 1;
2967	igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2968
2969	/*
2970	 * If we are using stripe alignment, check whether
2971	 * the stripe unit is a multiple of the inode alignment
2972	 */
2973	if (mp->m_dalign && igeo->inoalign_mask &&
2974	    !(mp->m_dalign & igeo->inoalign_mask))
2975		igeo->ialloc_align = mp->m_dalign;
2976	else
2977		igeo->ialloc_align = 0;
2978}
2979
2980/* Compute the location of the root directory inode that is laid out by mkfs. */
2981xfs_ino_t
2982xfs_ialloc_calc_rootino(
2983	struct xfs_mount	*mp,
2984	int			sunit)
2985{
2986	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2987	xfs_agblock_t		first_bno;
2988
2989	/*
2990	 * Pre-calculate the geometry of AG 0.  We know what it looks like
2991	 * because libxfs knows how to create allocation groups now.
2992	 *
2993	 * first_bno is the first block in which mkfs could possibly have
2994	 * allocated the root directory inode, once we factor in the metadata
2995	 * that mkfs formats before it.  Namely, the four AG headers...
2996	 */
2997	first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
2998
2999	/* ...the two free space btree roots... */
3000	first_bno += 2;
3001
3002	/* ...the inode btree root... */
3003	first_bno += 1;
3004
3005	/* ...the initial AGFL... */
3006	first_bno += xfs_alloc_min_freelist(mp, NULL);
3007
3008	/* ...the free inode btree root... */
3009	if (xfs_has_finobt(mp))
3010		first_bno++;
3011
3012	/* ...the reverse mapping btree root... */
3013	if (xfs_has_rmapbt(mp))
3014		first_bno++;
3015
3016	/* ...the reference count btree... */
3017	if (xfs_has_reflink(mp))
3018		first_bno++;
3019
3020	/*
3021	 * ...and the log, if it is allocated in the first allocation group.
3022	 *
3023	 * This can happen with filesystems that only have a single
3024	 * allocation group, or very odd geometries created by old mkfs
3025	 * versions on very small filesystems.
3026	 */
3027	if (xfs_ag_contains_log(mp, 0))
 
3028		 first_bno += mp->m_sb.sb_logblocks;
3029
3030	/*
3031	 * Now round first_bno up to whatever allocation alignment is given
3032	 * by the filesystem or was passed in.
3033	 */
3034	if (xfs_has_dalign(mp) && igeo->ialloc_align > 0)
3035		first_bno = roundup(first_bno, sunit);
3036	else if (xfs_has_align(mp) &&
3037			mp->m_sb.sb_inoalignmt > 1)
3038		first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
3039
3040	return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
3041}
3042
3043/*
3044 * Ensure there are not sparse inode clusters that cross the new EOAG.
3045 *
3046 * This is a no-op for non-spinode filesystems since clusters are always fully
3047 * allocated and checking the bnobt suffices.  However, a spinode filesystem
3048 * could have a record where the upper inodes are free blocks.  If those blocks
3049 * were removed from the filesystem, the inode record would extend beyond EOAG,
3050 * which will be flagged as corruption.
3051 */
3052int
3053xfs_ialloc_check_shrink(
3054	struct xfs_perag	*pag,
3055	struct xfs_trans	*tp,
3056	struct xfs_buf		*agibp,
3057	xfs_agblock_t		new_length)
3058{
3059	struct xfs_inobt_rec_incore rec;
3060	struct xfs_btree_cur	*cur;
3061	xfs_agino_t		agino;
3062	int			has;
3063	int			error;
3064
3065	if (!xfs_has_sparseinodes(pag->pag_mount))
3066		return 0;
3067
3068	cur = xfs_inobt_init_cursor(pag, tp, agibp);
3069
3070	/* Look up the inobt record that would correspond to the new EOFS. */
3071	agino = XFS_AGB_TO_AGINO(pag->pag_mount, new_length);
3072	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &has);
3073	if (error || !has)
3074		goto out;
3075
3076	error = xfs_inobt_get_rec(cur, &rec, &has);
3077	if (error)
3078		goto out;
3079
3080	if (!has) {
3081		xfs_ag_mark_sick(pag, XFS_SICK_AG_INOBT);
3082		error = -EFSCORRUPTED;
3083		goto out;
3084	}
3085
3086	/* If the record covers inodes that would be beyond EOFS, bail out. */
3087	if (rec.ir_startino + XFS_INODES_PER_CHUNK > agino) {
3088		error = -ENOSPC;
3089		goto out;
3090	}
3091out:
3092	xfs_btree_del_cursor(cur, error);
3093	return error;
3094}
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2002,2005 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_bit.h"
  13#include "xfs_sb.h"
  14#include "xfs_mount.h"
  15#include "xfs_inode.h"
  16#include "xfs_btree.h"
  17#include "xfs_ialloc.h"
  18#include "xfs_ialloc_btree.h"
  19#include "xfs_alloc.h"
  20#include "xfs_errortag.h"
  21#include "xfs_error.h"
  22#include "xfs_bmap.h"
  23#include "xfs_trans.h"
  24#include "xfs_buf_item.h"
  25#include "xfs_icreate_item.h"
  26#include "xfs_icache.h"
  27#include "xfs_trace.h"
  28#include "xfs_log.h"
  29#include "xfs_rmap.h"
 
 
  30
  31/*
  32 * Lookup a record by ino in the btree given by cur.
  33 */
  34int					/* error */
  35xfs_inobt_lookup(
  36	struct xfs_btree_cur	*cur,	/* btree cursor */
  37	xfs_agino_t		ino,	/* starting inode of chunk */
  38	xfs_lookup_t		dir,	/* <=, >=, == */
  39	int			*stat)	/* success/failure */
  40{
  41	cur->bc_rec.i.ir_startino = ino;
  42	cur->bc_rec.i.ir_holemask = 0;
  43	cur->bc_rec.i.ir_count = 0;
  44	cur->bc_rec.i.ir_freecount = 0;
  45	cur->bc_rec.i.ir_free = 0;
  46	return xfs_btree_lookup(cur, dir, stat);
  47}
  48
  49/*
  50 * Update the record referred to by cur to the value given.
  51 * This either works (return 0) or gets an EFSCORRUPTED error.
  52 */
  53STATIC int				/* error */
  54xfs_inobt_update(
  55	struct xfs_btree_cur	*cur,	/* btree cursor */
  56	xfs_inobt_rec_incore_t	*irec)	/* btree record */
  57{
  58	union xfs_btree_rec	rec;
  59
  60	rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
  61	if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
  62		rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
  63		rec.inobt.ir_u.sp.ir_count = irec->ir_count;
  64		rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
  65	} else {
  66		/* ir_holemask/ir_count not supported on-disk */
  67		rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
  68	}
  69	rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
  70	return xfs_btree_update(cur, &rec);
  71}
  72
  73/* Convert on-disk btree record to incore inobt record. */
  74void
  75xfs_inobt_btrec_to_irec(
  76	struct xfs_mount		*mp,
  77	union xfs_btree_rec		*rec,
  78	struct xfs_inobt_rec_incore	*irec)
  79{
  80	irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
  81	if (xfs_sb_version_hassparseinodes(&mp->m_sb)) {
  82		irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
  83		irec->ir_count = rec->inobt.ir_u.sp.ir_count;
  84		irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
  85	} else {
  86		/*
  87		 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
  88		 * values for full inode chunks.
  89		 */
  90		irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
  91		irec->ir_count = XFS_INODES_PER_CHUNK;
  92		irec->ir_freecount =
  93				be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
  94	}
  95	irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
  96}
  97
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  98/*
  99 * Get the data from the pointed-to record.
 100 */
 101int
 102xfs_inobt_get_rec(
 103	struct xfs_btree_cur		*cur,
 104	struct xfs_inobt_rec_incore	*irec,
 105	int				*stat)
 106{
 107	struct xfs_mount		*mp = cur->bc_mp;
 108	xfs_agnumber_t			agno = cur->bc_ag.agno;
 109	union xfs_btree_rec		*rec;
 
 110	int				error;
 111	uint64_t			realfree;
 112
 113	error = xfs_btree_get_rec(cur, &rec, stat);
 114	if (error || *stat == 0)
 115		return error;
 116
 117	xfs_inobt_btrec_to_irec(mp, rec, irec);
 118
 119	if (!xfs_verify_agino(mp, agno, irec->ir_startino))
 120		goto out_bad_rec;
 121	if (irec->ir_count < XFS_INODES_PER_HOLEMASK_BIT ||
 122	    irec->ir_count > XFS_INODES_PER_CHUNK)
 123		goto out_bad_rec;
 124	if (irec->ir_freecount > XFS_INODES_PER_CHUNK)
 125		goto out_bad_rec;
 126
 127	/* if there are no holes, return the first available offset */
 128	if (!xfs_inobt_issparse(irec->ir_holemask))
 129		realfree = irec->ir_free;
 130	else
 131		realfree = irec->ir_free & xfs_inobt_irec_to_allocmask(irec);
 132	if (hweight64(realfree) != irec->ir_freecount)
 133		goto out_bad_rec;
 134
 135	return 0;
 136
 137out_bad_rec:
 138	xfs_warn(mp,
 139		"%s Inode BTree record corruption in AG %d detected!",
 140		cur->bc_btnum == XFS_BTNUM_INO ? "Used" : "Free", agno);
 141	xfs_warn(mp,
 142"start inode 0x%x, count 0x%x, free 0x%x freemask 0x%llx, holemask 0x%x",
 143		irec->ir_startino, irec->ir_count, irec->ir_freecount,
 144		irec->ir_free, irec->ir_holemask);
 145	return -EFSCORRUPTED;
 146}
 147
 148/*
 149 * Insert a single inobt record. Cursor must already point to desired location.
 150 */
 151int
 152xfs_inobt_insert_rec(
 153	struct xfs_btree_cur	*cur,
 154	uint16_t		holemask,
 155	uint8_t			count,
 156	int32_t			freecount,
 157	xfs_inofree_t		free,
 158	int			*stat)
 159{
 160	cur->bc_rec.i.ir_holemask = holemask;
 161	cur->bc_rec.i.ir_count = count;
 162	cur->bc_rec.i.ir_freecount = freecount;
 163	cur->bc_rec.i.ir_free = free;
 164	return xfs_btree_insert(cur, stat);
 165}
 166
 167/*
 168 * Insert records describing a newly allocated inode chunk into the inobt.
 169 */
 170STATIC int
 171xfs_inobt_insert(
 172	struct xfs_mount	*mp,
 173	struct xfs_trans	*tp,
 174	struct xfs_buf		*agbp,
 175	xfs_agino_t		newino,
 176	xfs_agino_t		newlen,
 177	xfs_btnum_t		btnum)
 178{
 179	struct xfs_btree_cur	*cur;
 180	struct xfs_agi		*agi = agbp->b_addr;
 181	xfs_agnumber_t		agno = be32_to_cpu(agi->agi_seqno);
 182	xfs_agino_t		thisino;
 183	int			i;
 184	int			error;
 185
 186	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
 
 
 
 187
 188	for (thisino = newino;
 189	     thisino < newino + newlen;
 190	     thisino += XFS_INODES_PER_CHUNK) {
 191		error = xfs_inobt_lookup(cur, thisino, XFS_LOOKUP_EQ, &i);
 192		if (error) {
 193			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 194			return error;
 195		}
 196		ASSERT(i == 0);
 197
 198		error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
 199					     XFS_INODES_PER_CHUNK,
 200					     XFS_INODES_PER_CHUNK,
 201					     XFS_INOBT_ALL_FREE, &i);
 202		if (error) {
 203			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 204			return error;
 205		}
 206		ASSERT(i == 1);
 207	}
 208
 209	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
 210
 211	return 0;
 212}
 213
 214/*
 215 * Verify that the number of free inodes in the AGI is correct.
 216 */
 217#ifdef DEBUG
 218STATIC int
 219xfs_check_agi_freecount(
 220	struct xfs_btree_cur	*cur,
 221	struct xfs_agi		*agi)
 222{
 223	if (cur->bc_nlevels == 1) {
 224		xfs_inobt_rec_incore_t rec;
 225		int		freecount = 0;
 226		int		error;
 227		int		i;
 228
 229		error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
 230		if (error)
 231			return error;
 232
 233		do {
 234			error = xfs_inobt_get_rec(cur, &rec, &i);
 235			if (error)
 236				return error;
 237
 238			if (i) {
 239				freecount += rec.ir_freecount;
 240				error = xfs_btree_increment(cur, 0, &i);
 241				if (error)
 242					return error;
 243			}
 244		} while (i == 1);
 245
 246		if (!XFS_FORCED_SHUTDOWN(cur->bc_mp))
 247			ASSERT(freecount == be32_to_cpu(agi->agi_freecount));
 248	}
 249	return 0;
 250}
 251#else
 252#define xfs_check_agi_freecount(cur, agi)	0
 253#endif
 254
 255/*
 256 * Initialise a new set of inodes. When called without a transaction context
 257 * (e.g. from recovery) we initiate a delayed write of the inode buffers rather
 258 * than logging them (which in a transaction context puts them into the AIL
 259 * for writeback rather than the xfsbufd queue).
 260 */
 261int
 262xfs_ialloc_inode_init(
 263	struct xfs_mount	*mp,
 264	struct xfs_trans	*tp,
 265	struct list_head	*buffer_list,
 266	int			icount,
 267	xfs_agnumber_t		agno,
 268	xfs_agblock_t		agbno,
 269	xfs_agblock_t		length,
 270	unsigned int		gen)
 271{
 272	struct xfs_buf		*fbuf;
 273	struct xfs_dinode	*free;
 274	int			nbufs;
 275	int			version;
 276	int			i, j;
 277	xfs_daddr_t		d;
 278	xfs_ino_t		ino = 0;
 279	int			error;
 280
 281	/*
 282	 * Loop over the new block(s), filling in the inodes.  For small block
 283	 * sizes, manipulate the inodes in buffers  which are multiples of the
 284	 * blocks size.
 285	 */
 286	nbufs = length / M_IGEO(mp)->blocks_per_cluster;
 287
 288	/*
 289	 * Figure out what version number to use in the inodes we create.  If
 290	 * the superblock version has caught up to the one that supports the new
 291	 * inode format, then use the new inode version.  Otherwise use the old
 292	 * version so that old kernels will continue to be able to use the file
 293	 * system.
 294	 *
 295	 * For v3 inodes, we also need to write the inode number into the inode,
 296	 * so calculate the first inode number of the chunk here as
 297	 * XFS_AGB_TO_AGINO() only works within a filesystem block, not
 298	 * across multiple filesystem blocks (such as a cluster) and so cannot
 299	 * be used in the cluster buffer loop below.
 300	 *
 301	 * Further, because we are writing the inode directly into the buffer
 302	 * and calculating a CRC on the entire inode, we have ot log the entire
 303	 * inode so that the entire range the CRC covers is present in the log.
 304	 * That means for v3 inode we log the entire buffer rather than just the
 305	 * inode cores.
 306	 */
 307	if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
 308		version = 3;
 309		ino = XFS_AGINO_TO_INO(mp, agno, XFS_AGB_TO_AGINO(mp, agbno));
 310
 311		/*
 312		 * log the initialisation that is about to take place as an
 313		 * logical operation. This means the transaction does not
 314		 * need to log the physical changes to the inode buffers as log
 315		 * recovery will know what initialisation is actually needed.
 316		 * Hence we only need to log the buffers as "ordered" buffers so
 317		 * they track in the AIL as if they were physically logged.
 318		 */
 319		if (tp)
 320			xfs_icreate_log(tp, agno, agbno, icount,
 321					mp->m_sb.sb_inodesize, length, gen);
 322	} else
 323		version = 2;
 324
 325	for (j = 0; j < nbufs; j++) {
 326		/*
 327		 * Get the block.
 328		 */
 329		d = XFS_AGB_TO_DADDR(mp, agno, agbno +
 330				(j * M_IGEO(mp)->blocks_per_cluster));
 331		error = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
 332				mp->m_bsize * M_IGEO(mp)->blocks_per_cluster,
 333				XBF_UNMAPPED, &fbuf);
 334		if (error)
 335			return error;
 336
 337		/* Initialize the inode buffers and log them appropriately. */
 338		fbuf->b_ops = &xfs_inode_buf_ops;
 339		xfs_buf_zero(fbuf, 0, BBTOB(fbuf->b_length));
 340		for (i = 0; i < M_IGEO(mp)->inodes_per_cluster; i++) {
 341			int	ioffset = i << mp->m_sb.sb_inodelog;
 342			uint	isize = XFS_DINODE_SIZE(&mp->m_sb);
 343
 344			free = xfs_make_iptr(mp, fbuf, i);
 345			free->di_magic = cpu_to_be16(XFS_DINODE_MAGIC);
 346			free->di_version = version;
 347			free->di_gen = cpu_to_be32(gen);
 348			free->di_next_unlinked = cpu_to_be32(NULLAGINO);
 349
 350			if (version == 3) {
 351				free->di_ino = cpu_to_be64(ino);
 352				ino++;
 353				uuid_copy(&free->di_uuid,
 354					  &mp->m_sb.sb_meta_uuid);
 355				xfs_dinode_calc_crc(mp, free);
 356			} else if (tp) {
 357				/* just log the inode core */
 358				xfs_trans_log_buf(tp, fbuf, ioffset,
 359						  ioffset + isize - 1);
 360			}
 361		}
 362
 363		if (tp) {
 364			/*
 365			 * Mark the buffer as an inode allocation buffer so it
 366			 * sticks in AIL at the point of this allocation
 367			 * transaction. This ensures the they are on disk before
 368			 * the tail of the log can be moved past this
 369			 * transaction (i.e. by preventing relogging from moving
 370			 * it forward in the log).
 371			 */
 372			xfs_trans_inode_alloc_buf(tp, fbuf);
 373			if (version == 3) {
 374				/*
 375				 * Mark the buffer as ordered so that they are
 376				 * not physically logged in the transaction but
 377				 * still tracked in the AIL as part of the
 378				 * transaction and pin the log appropriately.
 379				 */
 380				xfs_trans_ordered_buf(tp, fbuf);
 381			}
 382		} else {
 383			fbuf->b_flags |= XBF_DONE;
 384			xfs_buf_delwri_queue(fbuf, buffer_list);
 385			xfs_buf_relse(fbuf);
 386		}
 387	}
 388	return 0;
 389}
 390
 391/*
 392 * Align startino and allocmask for a recently allocated sparse chunk such that
 393 * they are fit for insertion (or merge) into the on-disk inode btrees.
 394 *
 395 * Background:
 396 *
 397 * When enabled, sparse inode support increases the inode alignment from cluster
 398 * size to inode chunk size. This means that the minimum range between two
 399 * non-adjacent inode records in the inobt is large enough for a full inode
 400 * record. This allows for cluster sized, cluster aligned block allocation
 401 * without need to worry about whether the resulting inode record overlaps with
 402 * another record in the tree. Without this basic rule, we would have to deal
 403 * with the consequences of overlap by potentially undoing recent allocations in
 404 * the inode allocation codepath.
 405 *
 406 * Because of this alignment rule (which is enforced on mount), there are two
 407 * inobt possibilities for newly allocated sparse chunks. One is that the
 408 * aligned inode record for the chunk covers a range of inodes not already
 409 * covered in the inobt (i.e., it is safe to insert a new sparse record). The
 410 * other is that a record already exists at the aligned startino that considers
 411 * the newly allocated range as sparse. In the latter case, record content is
 412 * merged in hope that sparse inode chunks fill to full chunks over time.
 413 */
 414STATIC void
 415xfs_align_sparse_ino(
 416	struct xfs_mount		*mp,
 417	xfs_agino_t			*startino,
 418	uint16_t			*allocmask)
 419{
 420	xfs_agblock_t			agbno;
 421	xfs_agblock_t			mod;
 422	int				offset;
 423
 424	agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
 425	mod = agbno % mp->m_sb.sb_inoalignmt;
 426	if (!mod)
 427		return;
 428
 429	/* calculate the inode offset and align startino */
 430	offset = XFS_AGB_TO_AGINO(mp, mod);
 431	*startino -= offset;
 432
 433	/*
 434	 * Since startino has been aligned down, left shift allocmask such that
 435	 * it continues to represent the same physical inodes relative to the
 436	 * new startino.
 437	 */
 438	*allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
 439}
 440
 441/*
 442 * Determine whether the source inode record can merge into the target. Both
 443 * records must be sparse, the inode ranges must match and there must be no
 444 * allocation overlap between the records.
 445 */
 446STATIC bool
 447__xfs_inobt_can_merge(
 448	struct xfs_inobt_rec_incore	*trec,	/* tgt record */
 449	struct xfs_inobt_rec_incore	*srec)	/* src record */
 450{
 451	uint64_t			talloc;
 452	uint64_t			salloc;
 453
 454	/* records must cover the same inode range */
 455	if (trec->ir_startino != srec->ir_startino)
 456		return false;
 457
 458	/* both records must be sparse */
 459	if (!xfs_inobt_issparse(trec->ir_holemask) ||
 460	    !xfs_inobt_issparse(srec->ir_holemask))
 461		return false;
 462
 463	/* both records must track some inodes */
 464	if (!trec->ir_count || !srec->ir_count)
 465		return false;
 466
 467	/* can't exceed capacity of a full record */
 468	if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
 469		return false;
 470
 471	/* verify there is no allocation overlap */
 472	talloc = xfs_inobt_irec_to_allocmask(trec);
 473	salloc = xfs_inobt_irec_to_allocmask(srec);
 474	if (talloc & salloc)
 475		return false;
 476
 477	return true;
 478}
 479
 480/*
 481 * Merge the source inode record into the target. The caller must call
 482 * __xfs_inobt_can_merge() to ensure the merge is valid.
 483 */
 484STATIC void
 485__xfs_inobt_rec_merge(
 486	struct xfs_inobt_rec_incore	*trec,	/* target */
 487	struct xfs_inobt_rec_incore	*srec)	/* src */
 488{
 489	ASSERT(trec->ir_startino == srec->ir_startino);
 490
 491	/* combine the counts */
 492	trec->ir_count += srec->ir_count;
 493	trec->ir_freecount += srec->ir_freecount;
 494
 495	/*
 496	 * Merge the holemask and free mask. For both fields, 0 bits refer to
 497	 * allocated inodes. We combine the allocated ranges with bitwise AND.
 498	 */
 499	trec->ir_holemask &= srec->ir_holemask;
 500	trec->ir_free &= srec->ir_free;
 501}
 502
 503/*
 504 * Insert a new sparse inode chunk into the associated inode btree. The inode
 505 * record for the sparse chunk is pre-aligned to a startino that should match
 506 * any pre-existing sparse inode record in the tree. This allows sparse chunks
 507 * to fill over time.
 508 *
 509 * This function supports two modes of handling preexisting records depending on
 510 * the merge flag. If merge is true, the provided record is merged with the
 511 * existing record and updated in place. The merged record is returned in nrec.
 512 * If merge is false, an existing record is replaced with the provided record.
 513 * If no preexisting record exists, the provided record is always inserted.
 514 *
 515 * It is considered corruption if a merge is requested and not possible. Given
 516 * the sparse inode alignment constraints, this should never happen.
 517 */
 518STATIC int
 519xfs_inobt_insert_sprec(
 520	struct xfs_mount		*mp,
 521	struct xfs_trans		*tp,
 522	struct xfs_buf			*agbp,
 523	int				btnum,
 524	struct xfs_inobt_rec_incore	*nrec,	/* in/out: new/merged rec. */
 525	bool				merge)	/* merge or replace */
 526{
 
 527	struct xfs_btree_cur		*cur;
 528	struct xfs_agi			*agi = agbp->b_addr;
 529	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
 530	int				error;
 531	int				i;
 532	struct xfs_inobt_rec_incore	rec;
 533
 534	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
 535
 536	/* the new record is pre-aligned so we know where to look */
 537	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
 538	if (error)
 539		goto error;
 540	/* if nothing there, insert a new record and return */
 541	if (i == 0) {
 542		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
 543					     nrec->ir_count, nrec->ir_freecount,
 544					     nrec->ir_free, &i);
 545		if (error)
 546			goto error;
 547		if (XFS_IS_CORRUPT(mp, i != 1)) {
 
 548			error = -EFSCORRUPTED;
 549			goto error;
 550		}
 551
 552		goto out;
 553	}
 554
 555	/*
 556	 * A record exists at this startino. Merge or replace the record
 557	 * depending on what we've been asked to do.
 558	 */
 559	if (merge) {
 560		error = xfs_inobt_get_rec(cur, &rec, &i);
 561		if (error)
 562			goto error;
 563		if (XFS_IS_CORRUPT(mp, i != 1)) {
 564			error = -EFSCORRUPTED;
 565			goto error;
 566		}
 567		if (XFS_IS_CORRUPT(mp, rec.ir_startino != nrec->ir_startino)) {
 568			error = -EFSCORRUPTED;
 569			goto error;
 570		}
 
 571
 572		/*
 573		 * This should never fail. If we have coexisting records that
 574		 * cannot merge, something is seriously wrong.
 575		 */
 576		if (XFS_IS_CORRUPT(mp, !__xfs_inobt_can_merge(nrec, &rec))) {
 577			error = -EFSCORRUPTED;
 578			goto error;
 579		}
 
 580
 581		trace_xfs_irec_merge_pre(mp, agno, rec.ir_startino,
 582					 rec.ir_holemask, nrec->ir_startino,
 583					 nrec->ir_holemask);
 584
 585		/* merge to nrec to output the updated record */
 586		__xfs_inobt_rec_merge(nrec, &rec);
 587
 588		trace_xfs_irec_merge_post(mp, agno, nrec->ir_startino,
 589					  nrec->ir_holemask);
 590
 591		error = xfs_inobt_rec_check_count(mp, nrec);
 592		if (error)
 593			goto error;
 594	}
 595
 596	error = xfs_inobt_update(cur, nrec);
 597	if (error)
 598		goto error;
 599
 600out:
 601	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
 602	return 0;
 603error:
 604	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
 605	return error;
 606}
 607
 608/*
 609 * Allocate new inodes in the allocation group specified by agbp.
 610 * Return 0 for success, else error code.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 611 */
 612STATIC int
 613xfs_ialloc_ag_alloc(
 
 614	struct xfs_trans	*tp,
 615	struct xfs_buf		*agbp,
 616	int			*alloc)
 617{
 618	struct xfs_agi		*agi;
 619	struct xfs_alloc_arg	args;
 620	xfs_agnumber_t		agno;
 621	int			error;
 622	xfs_agino_t		newino;		/* new first inode's number */
 623	xfs_agino_t		newlen;		/* new number of inodes */
 624	int			isaligned = 0;	/* inode allocation at stripe */
 625						/* unit boundary */
 626	/* init. to full chunk */
 627	uint16_t		allocmask = (uint16_t) -1;
 628	struct xfs_inobt_rec_incore rec;
 629	struct xfs_perag	*pag;
 630	struct xfs_ino_geometry	*igeo = M_IGEO(tp->t_mountp);
 
 631	int			do_sparse = 0;
 632
 633	memset(&args, 0, sizeof(args));
 634	args.tp = tp;
 635	args.mp = tp->t_mountp;
 636	args.fsbno = NULLFSBLOCK;
 637	args.oinfo = XFS_RMAP_OINFO_INODES;
 
 638
 639#ifdef DEBUG
 640	/* randomly do sparse inode allocations */
 641	if (xfs_sb_version_hassparseinodes(&tp->t_mountp->m_sb) &&
 642	    igeo->ialloc_min_blks < igeo->ialloc_blks)
 643		do_sparse = prandom_u32() & 1;
 644#endif
 645
 646	/*
 647	 * Locking will ensure that we don't have two callers in here
 648	 * at one time.
 649	 */
 650	newlen = igeo->ialloc_inos;
 651	if (igeo->maxicount &&
 652	    percpu_counter_read_positive(&args.mp->m_icount) + newlen >
 653							igeo->maxicount)
 654		return -ENOSPC;
 655	args.minlen = args.maxlen = igeo->ialloc_blks;
 656	/*
 657	 * First try to allocate inodes contiguous with the last-allocated
 658	 * chunk of inodes.  If the filesystem is striped, this will fill
 659	 * an entire stripe unit with inodes.
 660	 */
 661	agi = agbp->b_addr;
 662	newino = be32_to_cpu(agi->agi_newino);
 663	agno = be32_to_cpu(agi->agi_seqno);
 664	args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
 665		     igeo->ialloc_blks;
 666	if (do_sparse)
 667		goto sparse_alloc;
 668	if (likely(newino != NULLAGINO &&
 669		  (args.agbno < be32_to_cpu(agi->agi_length)))) {
 670		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
 671		args.type = XFS_ALLOCTYPE_THIS_BNO;
 672		args.prod = 1;
 673
 674		/*
 675		 * We need to take into account alignment here to ensure that
 676		 * we don't modify the free list if we fail to have an exact
 677		 * block. If we don't have an exact match, and every oher
 678		 * attempt allocation attempt fails, we'll end up cancelling
 679		 * a dirty transaction and shutting down.
 680		 *
 681		 * For an exact allocation, alignment must be 1,
 682		 * however we need to take cluster alignment into account when
 683		 * fixing up the freelist. Use the minalignslop field to
 684		 * indicate that extra blocks might be required for alignment,
 685		 * but not to use them in the actual exact allocation.
 686		 */
 687		args.alignment = 1;
 688		args.minalignslop = igeo->cluster_align - 1;
 689
 690		/* Allow space for the inode btree to split. */
 691		args.minleft = igeo->inobt_maxlevels;
 692		if ((error = xfs_alloc_vextent(&args)))
 
 
 
 693			return error;
 694
 695		/*
 696		 * This request might have dirtied the transaction if the AG can
 697		 * satisfy the request, but the exact block was not available.
 698		 * If the allocation did fail, subsequent requests will relax
 699		 * the exact agbno requirement and increase the alignment
 700		 * instead. It is critical that the total size of the request
 701		 * (len + alignment + slop) does not increase from this point
 702		 * on, so reset minalignslop to ensure it is not included in
 703		 * subsequent requests.
 704		 */
 705		args.minalignslop = 0;
 706	}
 707
 708	if (unlikely(args.fsbno == NULLFSBLOCK)) {
 709		/*
 710		 * Set the alignment for the allocation.
 711		 * If stripe alignment is turned on then align at stripe unit
 712		 * boundary.
 713		 * If the cluster size is smaller than a filesystem block
 714		 * then we're doing I/O for inodes in filesystem block size
 715		 * pieces, so don't need alignment anyway.
 716		 */
 717		isaligned = 0;
 718		if (igeo->ialloc_align) {
 719			ASSERT(!(args.mp->m_flags & XFS_MOUNT_NOALIGN));
 720			args.alignment = args.mp->m_dalign;
 721			isaligned = 1;
 722		} else
 723			args.alignment = igeo->cluster_align;
 724		/*
 725		 * Need to figure out where to allocate the inode blocks.
 726		 * Ideally they should be spaced out through the a.g.
 727		 * For now, just allocate blocks up front.
 728		 */
 729		args.agbno = be32_to_cpu(agi->agi_root);
 730		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
 731		/*
 732		 * Allocate a fixed-size extent of inodes.
 733		 */
 734		args.type = XFS_ALLOCTYPE_NEAR_BNO;
 735		args.prod = 1;
 736		/*
 737		 * Allow space for the inode btree to split.
 738		 */
 739		args.minleft = igeo->inobt_maxlevels;
 740		if ((error = xfs_alloc_vextent(&args)))
 
 
 
 741			return error;
 742	}
 743
 744	/*
 745	 * If stripe alignment is turned on, then try again with cluster
 746	 * alignment.
 747	 */
 748	if (isaligned && args.fsbno == NULLFSBLOCK) {
 749		args.type = XFS_ALLOCTYPE_NEAR_BNO;
 750		args.agbno = be32_to_cpu(agi->agi_root);
 751		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
 752		args.alignment = igeo->cluster_align;
 753		if ((error = xfs_alloc_vextent(&args)))
 
 
 
 754			return error;
 755	}
 756
 757	/*
 758	 * Finally, try a sparse allocation if the filesystem supports it and
 759	 * the sparse allocation length is smaller than a full chunk.
 760	 */
 761	if (xfs_sb_version_hassparseinodes(&args.mp->m_sb) &&
 762	    igeo->ialloc_min_blks < igeo->ialloc_blks &&
 763	    args.fsbno == NULLFSBLOCK) {
 764sparse_alloc:
 765		args.type = XFS_ALLOCTYPE_NEAR_BNO;
 766		args.agbno = be32_to_cpu(agi->agi_root);
 767		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
 768		args.alignment = args.mp->m_sb.sb_spino_align;
 769		args.prod = 1;
 770
 771		args.minlen = igeo->ialloc_min_blks;
 772		args.maxlen = args.minlen;
 773
 774		/*
 775		 * The inode record will be aligned to full chunk size. We must
 776		 * prevent sparse allocation from AG boundaries that result in
 777		 * invalid inode records, such as records that start at agbno 0
 778		 * or extend beyond the AG.
 779		 *
 780		 * Set min agbno to the first aligned, non-zero agbno and max to
 781		 * the last aligned agbno that is at least one full chunk from
 782		 * the end of the AG.
 783		 */
 784		args.min_agbno = args.mp->m_sb.sb_inoalignmt;
 785		args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
 786					    args.mp->m_sb.sb_inoalignmt) -
 787				 igeo->ialloc_blks;
 788
 789		error = xfs_alloc_vextent(&args);
 
 
 790		if (error)
 791			return error;
 792
 793		newlen = XFS_AGB_TO_AGINO(args.mp, args.len);
 794		ASSERT(newlen <= XFS_INODES_PER_CHUNK);
 795		allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
 796	}
 797
 798	if (args.fsbno == NULLFSBLOCK) {
 799		*alloc = 0;
 800		return 0;
 801	}
 802	ASSERT(args.len == args.minlen);
 803
 804	/*
 805	 * Stamp and write the inode buffers.
 806	 *
 807	 * Seed the new inode cluster with a random generation number. This
 808	 * prevents short-term reuse of generation numbers if a chunk is
 809	 * freed and then immediately reallocated. We use random numbers
 810	 * rather than a linear progression to prevent the next generation
 811	 * number from being easily guessable.
 812	 */
 813	error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, agno,
 814			args.agbno, args.len, prandom_u32());
 815
 816	if (error)
 817		return error;
 818	/*
 819	 * Convert the results.
 820	 */
 821	newino = XFS_AGB_TO_AGINO(args.mp, args.agbno);
 822
 823	if (xfs_inobt_issparse(~allocmask)) {
 824		/*
 825		 * We've allocated a sparse chunk. Align the startino and mask.
 826		 */
 827		xfs_align_sparse_ino(args.mp, &newino, &allocmask);
 828
 829		rec.ir_startino = newino;
 830		rec.ir_holemask = ~allocmask;
 831		rec.ir_count = newlen;
 832		rec.ir_freecount = newlen;
 833		rec.ir_free = XFS_INOBT_ALL_FREE;
 834
 835		/*
 836		 * Insert the sparse record into the inobt and allow for a merge
 837		 * if necessary. If a merge does occur, rec is updated to the
 838		 * merged record.
 839		 */
 840		error = xfs_inobt_insert_sprec(args.mp, tp, agbp, XFS_BTNUM_INO,
 841					       &rec, true);
 842		if (error == -EFSCORRUPTED) {
 843			xfs_alert(args.mp,
 844	"invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
 845				  XFS_AGINO_TO_INO(args.mp, agno,
 846						   rec.ir_startino),
 847				  rec.ir_holemask, rec.ir_count);
 848			xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
 849		}
 850		if (error)
 851			return error;
 852
 853		/*
 854		 * We can't merge the part we've just allocated as for the inobt
 855		 * due to finobt semantics. The original record may or may not
 856		 * exist independent of whether physical inodes exist in this
 857		 * sparse chunk.
 858		 *
 859		 * We must update the finobt record based on the inobt record.
 860		 * rec contains the fully merged and up to date inobt record
 861		 * from the previous call. Set merge false to replace any
 862		 * existing record with this one.
 863		 */
 864		if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
 865			error = xfs_inobt_insert_sprec(args.mp, tp, agbp,
 866						       XFS_BTNUM_FINO, &rec,
 867						       false);
 868			if (error)
 869				return error;
 870		}
 871	} else {
 872		/* full chunk - insert new records to both btrees */
 873		error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
 874					 XFS_BTNUM_INO);
 875		if (error)
 876			return error;
 877
 878		if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
 879			error = xfs_inobt_insert(args.mp, tp, agbp, newino,
 880						 newlen, XFS_BTNUM_FINO);
 881			if (error)
 882				return error;
 883		}
 884	}
 885
 886	/*
 887	 * Update AGI counts and newino.
 888	 */
 889	be32_add_cpu(&agi->agi_count, newlen);
 890	be32_add_cpu(&agi->agi_freecount, newlen);
 891	pag = agbp->b_pag;
 892	pag->pagi_freecount += newlen;
 893	pag->pagi_count += newlen;
 894	agi->agi_newino = cpu_to_be32(newino);
 895
 896	/*
 897	 * Log allocation group header fields
 898	 */
 899	xfs_ialloc_log_agi(tp, agbp,
 900		XFS_AGI_COUNT | XFS_AGI_FREECOUNT | XFS_AGI_NEWINO);
 901	/*
 902	 * Modify/log superblock values for inode count and inode free count.
 903	 */
 904	xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, (long)newlen);
 905	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, (long)newlen);
 906	*alloc = 1;
 907	return 0;
 908}
 909
 910STATIC xfs_agnumber_t
 911xfs_ialloc_next_ag(
 912	xfs_mount_t	*mp)
 913{
 914	xfs_agnumber_t	agno;
 915
 916	spin_lock(&mp->m_agirotor_lock);
 917	agno = mp->m_agirotor;
 918	if (++mp->m_agirotor >= mp->m_maxagi)
 919		mp->m_agirotor = 0;
 920	spin_unlock(&mp->m_agirotor_lock);
 921
 922	return agno;
 923}
 924
 925/*
 926 * Select an allocation group to look for a free inode in, based on the parent
 927 * inode and the mode.  Return the allocation group buffer.
 928 */
 929STATIC xfs_agnumber_t
 930xfs_ialloc_ag_select(
 931	xfs_trans_t	*tp,		/* transaction pointer */
 932	xfs_ino_t	parent,		/* parent directory inode number */
 933	umode_t		mode)		/* bits set to indicate file type */
 934{
 935	xfs_agnumber_t	agcount;	/* number of ag's in the filesystem */
 936	xfs_agnumber_t	agno;		/* current ag number */
 937	int		flags;		/* alloc buffer locking flags */
 938	xfs_extlen_t	ineed;		/* blocks needed for inode allocation */
 939	xfs_extlen_t	longest = 0;	/* longest extent available */
 940	xfs_mount_t	*mp;		/* mount point structure */
 941	int		needspace;	/* file mode implies space allocated */
 942	xfs_perag_t	*pag;		/* per allocation group data */
 943	xfs_agnumber_t	pagno;		/* parent (starting) ag number */
 944	int		error;
 945
 946	/*
 947	 * Files of these types need at least one block if length > 0
 948	 * (and they won't fit in the inode, but that's hard to figure out).
 949	 */
 950	needspace = S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode);
 951	mp = tp->t_mountp;
 952	agcount = mp->m_maxagi;
 953	if (S_ISDIR(mode))
 954		pagno = xfs_ialloc_next_ag(mp);
 955	else {
 956		pagno = XFS_INO_TO_AGNO(mp, parent);
 957		if (pagno >= agcount)
 958			pagno = 0;
 959	}
 960
 961	ASSERT(pagno < agcount);
 962
 963	/*
 964	 * Loop through allocation groups, looking for one with a little
 965	 * free space in it.  Note we don't look for free inodes, exactly.
 966	 * Instead, we include whether there is a need to allocate inodes
 967	 * to mean that blocks must be allocated for them,
 968	 * if none are currently free.
 969	 */
 970	agno = pagno;
 971	flags = XFS_ALLOC_FLAG_TRYLOCK;
 972	for (;;) {
 973		pag = xfs_perag_get(mp, agno);
 974		if (!pag->pagi_inodeok) {
 975			xfs_ialloc_next_ag(mp);
 976			goto nextag;
 977		}
 978
 979		if (!pag->pagi_init) {
 980			error = xfs_ialloc_pagi_init(mp, tp, agno);
 981			if (error)
 982				goto nextag;
 983		}
 984
 985		if (pag->pagi_freecount) {
 986			xfs_perag_put(pag);
 987			return agno;
 988		}
 989
 990		if (!pag->pagf_init) {
 991			error = xfs_alloc_pagf_init(mp, tp, agno, flags);
 992			if (error)
 993				goto nextag;
 994		}
 995
 996		/*
 997		 * Check that there is enough free space for the file plus a
 998		 * chunk of inodes if we need to allocate some. If this is the
 999		 * first pass across the AGs, take into account the potential
1000		 * space needed for alignment of inode chunks when checking the
1001		 * longest contiguous free space in the AG - this prevents us
1002		 * from getting ENOSPC because we have free space larger than
1003		 * ialloc_blks but alignment constraints prevent us from using
1004		 * it.
1005		 *
1006		 * If we can't find an AG with space for full alignment slack to
1007		 * be taken into account, we must be near ENOSPC in all AGs.
1008		 * Hence we don't include alignment for the second pass and so
1009		 * if we fail allocation due to alignment issues then it is most
1010		 * likely a real ENOSPC condition.
1011		 */
1012		ineed = M_IGEO(mp)->ialloc_min_blks;
1013		if (flags && ineed > 1)
1014			ineed += M_IGEO(mp)->cluster_align;
1015		longest = pag->pagf_longest;
1016		if (!longest)
1017			longest = pag->pagf_flcount > 0;
1018
1019		if (pag->pagf_freeblks >= needspace + ineed &&
1020		    longest >= ineed) {
1021			xfs_perag_put(pag);
1022			return agno;
1023		}
1024nextag:
1025		xfs_perag_put(pag);
1026		/*
1027		 * No point in iterating over the rest, if we're shutting
1028		 * down.
1029		 */
1030		if (XFS_FORCED_SHUTDOWN(mp))
1031			return NULLAGNUMBER;
1032		agno++;
1033		if (agno >= agcount)
1034			agno = 0;
1035		if (agno == pagno) {
1036			if (flags == 0)
1037				return NULLAGNUMBER;
1038			flags = 0;
1039		}
1040	}
1041}
1042
1043/*
1044 * Try to retrieve the next record to the left/right from the current one.
1045 */
1046STATIC int
1047xfs_ialloc_next_rec(
1048	struct xfs_btree_cur	*cur,
1049	xfs_inobt_rec_incore_t	*rec,
1050	int			*done,
1051	int			left)
1052{
1053	int                     error;
1054	int			i;
1055
1056	if (left)
1057		error = xfs_btree_decrement(cur, 0, &i);
1058	else
1059		error = xfs_btree_increment(cur, 0, &i);
1060
1061	if (error)
1062		return error;
1063	*done = !i;
1064	if (i) {
1065		error = xfs_inobt_get_rec(cur, rec, &i);
1066		if (error)
1067			return error;
1068		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1069			return -EFSCORRUPTED;
 
1070	}
1071
1072	return 0;
1073}
1074
1075STATIC int
1076xfs_ialloc_get_rec(
1077	struct xfs_btree_cur	*cur,
1078	xfs_agino_t		agino,
1079	xfs_inobt_rec_incore_t	*rec,
1080	int			*done)
1081{
1082	int                     error;
1083	int			i;
1084
1085	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_EQ, &i);
1086	if (error)
1087		return error;
1088	*done = !i;
1089	if (i) {
1090		error = xfs_inobt_get_rec(cur, rec, &i);
1091		if (error)
1092			return error;
1093		if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1094			return -EFSCORRUPTED;
 
1095	}
1096
1097	return 0;
1098}
1099
1100/*
1101 * Return the offset of the first free inode in the record. If the inode chunk
1102 * is sparsely allocated, we convert the record holemask to inode granularity
1103 * and mask off the unallocated regions from the inode free mask.
1104 */
1105STATIC int
1106xfs_inobt_first_free_inode(
1107	struct xfs_inobt_rec_incore	*rec)
1108{
1109	xfs_inofree_t			realfree;
1110
1111	/* if there are no holes, return the first available offset */
1112	if (!xfs_inobt_issparse(rec->ir_holemask))
1113		return xfs_lowbit64(rec->ir_free);
1114
1115	realfree = xfs_inobt_irec_to_allocmask(rec);
1116	realfree &= rec->ir_free;
1117
1118	return xfs_lowbit64(realfree);
1119}
1120
1121/*
1122 * Allocate an inode using the inobt-only algorithm.
1123 */
1124STATIC int
1125xfs_dialloc_ag_inobt(
 
1126	struct xfs_trans	*tp,
1127	struct xfs_buf		*agbp,
1128	xfs_ino_t		parent,
1129	xfs_ino_t		*inop)
1130{
1131	struct xfs_mount	*mp = tp->t_mountp;
1132	struct xfs_agi		*agi = agbp->b_addr;
1133	xfs_agnumber_t		agno = be32_to_cpu(agi->agi_seqno);
1134	xfs_agnumber_t		pagno = XFS_INO_TO_AGNO(mp, parent);
1135	xfs_agino_t		pagino = XFS_INO_TO_AGINO(mp, parent);
1136	struct xfs_perag	*pag = agbp->b_pag;
1137	struct xfs_btree_cur	*cur, *tcur;
1138	struct xfs_inobt_rec_incore rec, trec;
1139	xfs_ino_t		ino;
1140	int			error;
1141	int			offset;
1142	int			i, j;
1143	int			searchdistance = 10;
1144
1145	ASSERT(pag->pagi_init);
1146	ASSERT(pag->pagi_inodeok);
1147	ASSERT(pag->pagi_freecount > 0);
1148
1149 restart_pagno:
1150	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1151	/*
1152	 * If pagino is 0 (this is the root inode allocation) use newino.
1153	 * This must work because we've just allocated some.
1154	 */
1155	if (!pagino)
1156		pagino = be32_to_cpu(agi->agi_newino);
1157
1158	error = xfs_check_agi_freecount(cur, agi);
1159	if (error)
1160		goto error0;
1161
1162	/*
1163	 * If in the same AG as the parent, try to get near the parent.
1164	 */
1165	if (pagno == agno) {
1166		int		doneleft;	/* done, to the left */
1167		int		doneright;	/* done, to the right */
1168
1169		error = xfs_inobt_lookup(cur, pagino, XFS_LOOKUP_LE, &i);
1170		if (error)
1171			goto error0;
1172		if (XFS_IS_CORRUPT(mp, i != 1)) {
 
1173			error = -EFSCORRUPTED;
1174			goto error0;
1175		}
1176
1177		error = xfs_inobt_get_rec(cur, &rec, &j);
1178		if (error)
1179			goto error0;
1180		if (XFS_IS_CORRUPT(mp, j != 1)) {
 
1181			error = -EFSCORRUPTED;
1182			goto error0;
1183		}
1184
1185		if (rec.ir_freecount > 0) {
1186			/*
1187			 * Found a free inode in the same chunk
1188			 * as the parent, done.
1189			 */
1190			goto alloc_inode;
1191		}
1192
1193
1194		/*
1195		 * In the same AG as parent, but parent's chunk is full.
1196		 */
1197
1198		/* duplicate the cursor, search left & right simultaneously */
1199		error = xfs_btree_dup_cursor(cur, &tcur);
1200		if (error)
1201			goto error0;
1202
1203		/*
1204		 * Skip to last blocks looked up if same parent inode.
1205		 */
1206		if (pagino != NULLAGINO &&
1207		    pag->pagl_pagino == pagino &&
1208		    pag->pagl_leftrec != NULLAGINO &&
1209		    pag->pagl_rightrec != NULLAGINO) {
1210			error = xfs_ialloc_get_rec(tcur, pag->pagl_leftrec,
1211						   &trec, &doneleft);
1212			if (error)
1213				goto error1;
1214
1215			error = xfs_ialloc_get_rec(cur, pag->pagl_rightrec,
1216						   &rec, &doneright);
1217			if (error)
1218				goto error1;
1219		} else {
1220			/* search left with tcur, back up 1 record */
1221			error = xfs_ialloc_next_rec(tcur, &trec, &doneleft, 1);
1222			if (error)
1223				goto error1;
1224
1225			/* search right with cur, go forward 1 record. */
1226			error = xfs_ialloc_next_rec(cur, &rec, &doneright, 0);
1227			if (error)
1228				goto error1;
1229		}
1230
1231		/*
1232		 * Loop until we find an inode chunk with a free inode.
1233		 */
1234		while (--searchdistance > 0 && (!doneleft || !doneright)) {
1235			int	useleft;  /* using left inode chunk this time */
1236
1237			/* figure out the closer block if both are valid. */
1238			if (!doneleft && !doneright) {
1239				useleft = pagino -
1240				 (trec.ir_startino + XFS_INODES_PER_CHUNK - 1) <
1241				  rec.ir_startino - pagino;
1242			} else {
1243				useleft = !doneleft;
1244			}
1245
1246			/* free inodes to the left? */
1247			if (useleft && trec.ir_freecount) {
1248				xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1249				cur = tcur;
1250
1251				pag->pagl_leftrec = trec.ir_startino;
1252				pag->pagl_rightrec = rec.ir_startino;
1253				pag->pagl_pagino = pagino;
1254				rec = trec;
1255				goto alloc_inode;
1256			}
1257
1258			/* free inodes to the right? */
1259			if (!useleft && rec.ir_freecount) {
1260				xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1261
1262				pag->pagl_leftrec = trec.ir_startino;
1263				pag->pagl_rightrec = rec.ir_startino;
1264				pag->pagl_pagino = pagino;
1265				goto alloc_inode;
1266			}
1267
1268			/* get next record to check */
1269			if (useleft) {
1270				error = xfs_ialloc_next_rec(tcur, &trec,
1271								 &doneleft, 1);
1272			} else {
1273				error = xfs_ialloc_next_rec(cur, &rec,
1274								 &doneright, 0);
1275			}
1276			if (error)
1277				goto error1;
1278		}
1279
1280		if (searchdistance <= 0) {
1281			/*
1282			 * Not in range - save last search
1283			 * location and allocate a new inode
1284			 */
1285			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1286			pag->pagl_leftrec = trec.ir_startino;
1287			pag->pagl_rightrec = rec.ir_startino;
1288			pag->pagl_pagino = pagino;
1289
1290		} else {
1291			/*
1292			 * We've reached the end of the btree. because
1293			 * we are only searching a small chunk of the
1294			 * btree each search, there is obviously free
1295			 * inodes closer to the parent inode than we
1296			 * are now. restart the search again.
1297			 */
1298			pag->pagl_pagino = NULLAGINO;
1299			pag->pagl_leftrec = NULLAGINO;
1300			pag->pagl_rightrec = NULLAGINO;
1301			xfs_btree_del_cursor(tcur, XFS_BTREE_NOERROR);
1302			xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1303			goto restart_pagno;
1304		}
1305	}
1306
1307	/*
1308	 * In a different AG from the parent.
1309	 * See if the most recently allocated block has any free.
1310	 */
1311	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1312		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1313					 XFS_LOOKUP_EQ, &i);
1314		if (error)
1315			goto error0;
1316
1317		if (i == 1) {
1318			error = xfs_inobt_get_rec(cur, &rec, &j);
1319			if (error)
1320				goto error0;
1321
1322			if (j == 1 && rec.ir_freecount > 0) {
1323				/*
1324				 * The last chunk allocated in the group
1325				 * still has a free inode.
1326				 */
1327				goto alloc_inode;
1328			}
1329		}
1330	}
1331
1332	/*
1333	 * None left in the last group, search the whole AG
1334	 */
1335	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1336	if (error)
1337		goto error0;
1338	if (XFS_IS_CORRUPT(mp, i != 1)) {
 
1339		error = -EFSCORRUPTED;
1340		goto error0;
1341	}
1342
1343	for (;;) {
1344		error = xfs_inobt_get_rec(cur, &rec, &i);
1345		if (error)
1346			goto error0;
1347		if (XFS_IS_CORRUPT(mp, i != 1)) {
 
1348			error = -EFSCORRUPTED;
1349			goto error0;
1350		}
1351		if (rec.ir_freecount > 0)
1352			break;
1353		error = xfs_btree_increment(cur, 0, &i);
1354		if (error)
1355			goto error0;
1356		if (XFS_IS_CORRUPT(mp, i != 1)) {
 
1357			error = -EFSCORRUPTED;
1358			goto error0;
1359		}
1360	}
1361
1362alloc_inode:
1363	offset = xfs_inobt_first_free_inode(&rec);
1364	ASSERT(offset >= 0);
1365	ASSERT(offset < XFS_INODES_PER_CHUNK);
1366	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1367				   XFS_INODES_PER_CHUNK) == 0);
1368	ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1369	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1370	rec.ir_freecount--;
1371	error = xfs_inobt_update(cur, &rec);
1372	if (error)
1373		goto error0;
1374	be32_add_cpu(&agi->agi_freecount, -1);
1375	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1376	pag->pagi_freecount--;
1377
1378	error = xfs_check_agi_freecount(cur, agi);
1379	if (error)
1380		goto error0;
1381
1382	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1383	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1384	*inop = ino;
1385	return 0;
1386error1:
1387	xfs_btree_del_cursor(tcur, XFS_BTREE_ERROR);
1388error0:
1389	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1390	return error;
1391}
1392
1393/*
1394 * Use the free inode btree to allocate an inode based on distance from the
1395 * parent. Note that the provided cursor may be deleted and replaced.
1396 */
1397STATIC int
1398xfs_dialloc_ag_finobt_near(
1399	xfs_agino_t			pagino,
1400	struct xfs_btree_cur		**ocur,
1401	struct xfs_inobt_rec_incore	*rec)
1402{
1403	struct xfs_btree_cur		*lcur = *ocur;	/* left search cursor */
1404	struct xfs_btree_cur		*rcur;	/* right search cursor */
1405	struct xfs_inobt_rec_incore	rrec;
1406	int				error;
1407	int				i, j;
1408
1409	error = xfs_inobt_lookup(lcur, pagino, XFS_LOOKUP_LE, &i);
1410	if (error)
1411		return error;
1412
1413	if (i == 1) {
1414		error = xfs_inobt_get_rec(lcur, rec, &i);
1415		if (error)
1416			return error;
1417		if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1))
 
1418			return -EFSCORRUPTED;
 
1419
1420		/*
1421		 * See if we've landed in the parent inode record. The finobt
1422		 * only tracks chunks with at least one free inode, so record
1423		 * existence is enough.
1424		 */
1425		if (pagino >= rec->ir_startino &&
1426		    pagino < (rec->ir_startino + XFS_INODES_PER_CHUNK))
1427			return 0;
1428	}
1429
1430	error = xfs_btree_dup_cursor(lcur, &rcur);
1431	if (error)
1432		return error;
1433
1434	error = xfs_inobt_lookup(rcur, pagino, XFS_LOOKUP_GE, &j);
1435	if (error)
1436		goto error_rcur;
1437	if (j == 1) {
1438		error = xfs_inobt_get_rec(rcur, &rrec, &j);
1439		if (error)
1440			goto error_rcur;
1441		if (XFS_IS_CORRUPT(lcur->bc_mp, j != 1)) {
 
1442			error = -EFSCORRUPTED;
1443			goto error_rcur;
1444		}
1445	}
1446
1447	if (XFS_IS_CORRUPT(lcur->bc_mp, i != 1 && j != 1)) {
 
1448		error = -EFSCORRUPTED;
1449		goto error_rcur;
1450	}
1451	if (i == 1 && j == 1) {
1452		/*
1453		 * Both the left and right records are valid. Choose the closer
1454		 * inode chunk to the target.
1455		 */
1456		if ((pagino - rec->ir_startino + XFS_INODES_PER_CHUNK - 1) >
1457		    (rrec.ir_startino - pagino)) {
1458			*rec = rrec;
1459			xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1460			*ocur = rcur;
1461		} else {
1462			xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1463		}
1464	} else if (j == 1) {
1465		/* only the right record is valid */
1466		*rec = rrec;
1467		xfs_btree_del_cursor(lcur, XFS_BTREE_NOERROR);
1468		*ocur = rcur;
1469	} else if (i == 1) {
1470		/* only the left record is valid */
1471		xfs_btree_del_cursor(rcur, XFS_BTREE_NOERROR);
1472	}
1473
1474	return 0;
1475
1476error_rcur:
1477	xfs_btree_del_cursor(rcur, XFS_BTREE_ERROR);
1478	return error;
1479}
1480
1481/*
1482 * Use the free inode btree to find a free inode based on a newino hint. If
1483 * the hint is NULL, find the first free inode in the AG.
1484 */
1485STATIC int
1486xfs_dialloc_ag_finobt_newino(
1487	struct xfs_agi			*agi,
1488	struct xfs_btree_cur		*cur,
1489	struct xfs_inobt_rec_incore	*rec)
1490{
1491	int error;
1492	int i;
1493
1494	if (agi->agi_newino != cpu_to_be32(NULLAGINO)) {
1495		error = xfs_inobt_lookup(cur, be32_to_cpu(agi->agi_newino),
1496					 XFS_LOOKUP_EQ, &i);
1497		if (error)
1498			return error;
1499		if (i == 1) {
1500			error = xfs_inobt_get_rec(cur, rec, &i);
1501			if (error)
1502				return error;
1503			if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1504				return -EFSCORRUPTED;
 
1505			return 0;
1506		}
1507	}
1508
1509	/*
1510	 * Find the first inode available in the AG.
1511	 */
1512	error = xfs_inobt_lookup(cur, 0, XFS_LOOKUP_GE, &i);
1513	if (error)
1514		return error;
1515	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1516		return -EFSCORRUPTED;
 
1517
1518	error = xfs_inobt_get_rec(cur, rec, &i);
1519	if (error)
1520		return error;
1521	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1522		return -EFSCORRUPTED;
 
1523
1524	return 0;
1525}
1526
1527/*
1528 * Update the inobt based on a modification made to the finobt. Also ensure that
1529 * the records from both trees are equivalent post-modification.
1530 */
1531STATIC int
1532xfs_dialloc_ag_update_inobt(
1533	struct xfs_btree_cur		*cur,	/* inobt cursor */
1534	struct xfs_inobt_rec_incore	*frec,	/* finobt record */
1535	int				offset) /* inode offset */
1536{
1537	struct xfs_inobt_rec_incore	rec;
1538	int				error;
1539	int				i;
1540
1541	error = xfs_inobt_lookup(cur, frec->ir_startino, XFS_LOOKUP_EQ, &i);
1542	if (error)
1543		return error;
1544	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1545		return -EFSCORRUPTED;
 
1546
1547	error = xfs_inobt_get_rec(cur, &rec, &i);
1548	if (error)
1549		return error;
1550	if (XFS_IS_CORRUPT(cur->bc_mp, i != 1))
 
1551		return -EFSCORRUPTED;
 
1552	ASSERT((XFS_AGINO_TO_OFFSET(cur->bc_mp, rec.ir_startino) %
1553				   XFS_INODES_PER_CHUNK) == 0);
1554
1555	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1556	rec.ir_freecount--;
1557
1558	if (XFS_IS_CORRUPT(cur->bc_mp,
1559			   rec.ir_free != frec->ir_free ||
1560			   rec.ir_freecount != frec->ir_freecount))
 
1561		return -EFSCORRUPTED;
 
1562
1563	return xfs_inobt_update(cur, &rec);
1564}
1565
1566/*
1567 * Allocate an inode using the free inode btree, if available. Otherwise, fall
1568 * back to the inobt search algorithm.
1569 *
1570 * The caller selected an AG for us, and made sure that free inodes are
1571 * available.
1572 */
1573STATIC int
1574xfs_dialloc_ag(
 
1575	struct xfs_trans	*tp,
1576	struct xfs_buf		*agbp,
1577	xfs_ino_t		parent,
1578	xfs_ino_t		*inop)
1579{
1580	struct xfs_mount		*mp = tp->t_mountp;
1581	struct xfs_agi			*agi = agbp->b_addr;
1582	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
1583	xfs_agnumber_t			pagno = XFS_INO_TO_AGNO(mp, parent);
1584	xfs_agino_t			pagino = XFS_INO_TO_AGINO(mp, parent);
1585	struct xfs_btree_cur		*cur;	/* finobt cursor */
1586	struct xfs_btree_cur		*icur;	/* inobt cursor */
1587	struct xfs_inobt_rec_incore	rec;
1588	xfs_ino_t			ino;
1589	int				error;
1590	int				offset;
1591	int				i;
1592
1593	if (!xfs_sb_version_hasfinobt(&mp->m_sb))
1594		return xfs_dialloc_ag_inobt(tp, agbp, parent, inop);
1595
1596	/*
1597	 * If pagino is 0 (this is the root inode allocation) use newino.
1598	 * This must work because we've just allocated some.
1599	 */
1600	if (!pagino)
1601		pagino = be32_to_cpu(agi->agi_newino);
1602
1603	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
1604
1605	error = xfs_check_agi_freecount(cur, agi);
1606	if (error)
1607		goto error_cur;
1608
1609	/*
1610	 * The search algorithm depends on whether we're in the same AG as the
1611	 * parent. If so, find the closest available inode to the parent. If
1612	 * not, consider the agi hint or find the first free inode in the AG.
1613	 */
1614	if (agno == pagno)
1615		error = xfs_dialloc_ag_finobt_near(pagino, &cur, &rec);
1616	else
1617		error = xfs_dialloc_ag_finobt_newino(agi, cur, &rec);
1618	if (error)
1619		goto error_cur;
1620
1621	offset = xfs_inobt_first_free_inode(&rec);
1622	ASSERT(offset >= 0);
1623	ASSERT(offset < XFS_INODES_PER_CHUNK);
1624	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
1625				   XFS_INODES_PER_CHUNK) == 0);
1626	ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino + offset);
1627
1628	/*
1629	 * Modify or remove the finobt record.
1630	 */
1631	rec.ir_free &= ~XFS_INOBT_MASK(offset);
1632	rec.ir_freecount--;
1633	if (rec.ir_freecount)
1634		error = xfs_inobt_update(cur, &rec);
1635	else
1636		error = xfs_btree_delete(cur, &i);
1637	if (error)
1638		goto error_cur;
1639
1640	/*
1641	 * The finobt has now been updated appropriately. We haven't updated the
1642	 * agi and superblock yet, so we can create an inobt cursor and validate
1643	 * the original freecount. If all is well, make the equivalent update to
1644	 * the inobt using the finobt record and offset information.
1645	 */
1646	icur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1647
1648	error = xfs_check_agi_freecount(icur, agi);
1649	if (error)
1650		goto error_icur;
1651
1652	error = xfs_dialloc_ag_update_inobt(icur, &rec, offset);
1653	if (error)
1654		goto error_icur;
1655
1656	/*
1657	 * Both trees have now been updated. We must update the perag and
1658	 * superblock before we can check the freecount for each btree.
1659	 */
1660	be32_add_cpu(&agi->agi_freecount, -1);
1661	xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
1662	agbp->b_pag->pagi_freecount--;
1663
1664	xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -1);
1665
1666	error = xfs_check_agi_freecount(icur, agi);
1667	if (error)
1668		goto error_icur;
1669	error = xfs_check_agi_freecount(cur, agi);
1670	if (error)
1671		goto error_icur;
1672
1673	xfs_btree_del_cursor(icur, XFS_BTREE_NOERROR);
1674	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
1675	*inop = ino;
1676	return 0;
1677
1678error_icur:
1679	xfs_btree_del_cursor(icur, XFS_BTREE_ERROR);
1680error_cur:
1681	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
1682	return error;
1683}
1684
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1685/*
1686 * Allocate an inode on disk.
1687 *
1688 * Mode is used to tell whether the new inode will need space, and whether it
1689 * is a directory.
1690 *
1691 * This function is designed to be called twice if it has to do an allocation
1692 * to make more free inodes.  On the first call, *IO_agbp should be set to NULL.
1693 * If an inode is available without having to performn an allocation, an inode
1694 * number is returned.  In this case, *IO_agbp is set to NULL.  If an allocation
1695 * needs to be done, xfs_dialloc returns the current AGI buffer in *IO_agbp.
1696 * The caller should then commit the current transaction, allocate a
1697 * new transaction, and call xfs_dialloc() again, passing in the previous value
1698 * of *IO_agbp.  IO_agbp should be held across the transactions. Since the AGI
1699 * buffer is locked across the two calls, the second call is guaranteed to have
1700 * a free inode available.
1701 *
1702 * Once we successfully pick an inode its number is returned and the on-disk
1703 * data structures are updated.  The inode itself is not read in, since doing so
1704 * would break ordering constraints with xfs_reclaim.
1705 */
1706int
1707xfs_dialloc(
1708	struct xfs_trans	*tp,
1709	xfs_ino_t		parent,
1710	umode_t			mode,
1711	struct xfs_buf		**IO_agbp,
1712	xfs_ino_t		*inop)
1713{
1714	struct xfs_mount	*mp = tp->t_mountp;
1715	struct xfs_buf		*agbp;
1716	xfs_agnumber_t		agno;
1717	int			error;
1718	int			ialloced;
1719	int			noroom = 0;
1720	xfs_agnumber_t		start_agno;
1721	struct xfs_perag	*pag;
1722	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
1723	int			okalloc = 1;
1724
1725	if (*IO_agbp) {
1726		/*
1727		 * If the caller passes in a pointer to the AGI buffer,
1728		 * continue where we left off before.  In this case, we
1729		 * know that the allocation group has free inodes.
1730		 */
1731		agbp = *IO_agbp;
1732		goto out_alloc;
1733	}
1734
1735	/*
1736	 * We do not have an agbp, so select an initial allocation
1737	 * group for inode allocation.
 
1738	 */
1739	start_agno = xfs_ialloc_ag_select(tp, parent, mode);
1740	if (start_agno == NULLAGNUMBER) {
1741		*inop = NULLFSINO;
1742		return 0;
 
 
 
1743	}
1744
1745	/*
1746	 * If we have already hit the ceiling of inode blocks then clear
1747	 * okalloc so we scan all available agi structures for a free
1748	 * inode.
1749	 *
1750	 * Read rough value of mp->m_icount by percpu_counter_read_positive,
1751	 * which will sacrifice the preciseness but improve the performance.
1752	 */
1753	if (igeo->maxicount &&
1754	    percpu_counter_read_positive(&mp->m_icount) + igeo->ialloc_inos
1755							> igeo->maxicount) {
1756		noroom = 1;
1757		okalloc = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
1758	}
1759
1760	/*
1761	 * Loop until we find an allocation group that either has free inodes
1762	 * or in which we can allocate some inodes.  Iterate through the
1763	 * allocation groups upward, wrapping at the end.
1764	 */
1765	agno = start_agno;
1766	for (;;) {
1767		pag = xfs_perag_get(mp, agno);
1768		if (!pag->pagi_inodeok) {
1769			xfs_ialloc_next_ag(mp);
1770			goto nextag;
 
 
 
1771		}
1772
1773		if (!pag->pagi_init) {
1774			error = xfs_ialloc_pagi_init(mp, tp, agno);
1775			if (error)
1776				goto out_error;
1777		}
1778
1779		/*
1780		 * Do a first racy fast path check if this AG is usable.
1781		 */
1782		if (!pag->pagi_freecount && !okalloc)
1783			goto nextag;
1784
1785		/*
1786		 * Then read in the AGI buffer and recheck with the AGI buffer
1787		 * lock held.
1788		 */
1789		error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
1790		if (error)
1791			goto out_error;
1792
1793		if (pag->pagi_freecount) {
1794			xfs_perag_put(pag);
1795			goto out_alloc;
1796		}
1797
1798		if (!okalloc)
1799			goto nextag_relse_buffer;
1800
1801
1802		error = xfs_ialloc_ag_alloc(tp, agbp, &ialloced);
1803		if (error) {
1804			xfs_trans_brelse(tp, agbp);
1805
1806			if (error != -ENOSPC)
1807				goto out_error;
1808
1809			xfs_perag_put(pag);
1810			*inop = NULLFSINO;
1811			return 0;
1812		}
1813
1814		if (ialloced) {
1815			/*
1816			 * We successfully allocated some inodes, return
1817			 * the current context to the caller so that it
1818			 * can commit the current transaction and call
1819			 * us again where we left off.
1820			 */
1821			ASSERT(pag->pagi_freecount > 0);
1822			xfs_perag_put(pag);
1823
1824			*IO_agbp = agbp;
1825			*inop = NULLFSINO;
1826			return 0;
1827		}
1828
1829nextag_relse_buffer:
1830		xfs_trans_brelse(tp, agbp);
1831nextag:
1832		xfs_perag_put(pag);
1833		if (++agno == mp->m_sb.sb_agcount)
1834			agno = 0;
1835		if (agno == start_agno) {
1836			*inop = NULLFSINO;
1837			return noroom ? -ENOSPC : 0;
1838		}
 
1839	}
1840
1841out_alloc:
1842	*IO_agbp = NULL;
1843	return xfs_dialloc_ag(tp, agbp, parent, inop);
1844out_error:
1845	xfs_perag_put(pag);
1846	return error;
1847}
1848
1849/*
1850 * Free the blocks of an inode chunk. We must consider that the inode chunk
1851 * might be sparse and only free the regions that are allocated as part of the
1852 * chunk.
1853 */
1854STATIC void
1855xfs_difree_inode_chunk(
1856	struct xfs_trans		*tp,
1857	xfs_agnumber_t			agno,
1858	struct xfs_inobt_rec_incore	*rec)
1859{
1860	struct xfs_mount		*mp = tp->t_mountp;
1861	xfs_agblock_t			sagbno = XFS_AGINO_TO_AGBNO(mp,
1862							rec->ir_startino);
1863	int				startidx, endidx;
1864	int				nextbit;
1865	xfs_agblock_t			agbno;
1866	int				contigblk;
1867	DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
1868
1869	if (!xfs_inobt_issparse(rec->ir_holemask)) {
1870		/* not sparse, calculate extent info directly */
1871		xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, sagbno),
1872				  M_IGEO(mp)->ialloc_blks,
1873				  &XFS_RMAP_OINFO_INODES);
1874		return;
1875	}
1876
1877	/* holemask is only 16-bits (fits in an unsigned long) */
1878	ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
1879	holemask[0] = rec->ir_holemask;
1880
1881	/*
1882	 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
1883	 * holemask and convert the start/end index of each range to an extent.
1884	 * We start with the start and end index both pointing at the first 0 in
1885	 * the mask.
1886	 */
1887	startidx = endidx = find_first_zero_bit(holemask,
1888						XFS_INOBT_HOLEMASK_BITS);
1889	nextbit = startidx + 1;
1890	while (startidx < XFS_INOBT_HOLEMASK_BITS) {
 
 
1891		nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
1892					     nextbit);
1893		/*
1894		 * If the next zero bit is contiguous, update the end index of
1895		 * the current range and continue.
1896		 */
1897		if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
1898		    nextbit == endidx + 1) {
1899			endidx = nextbit;
1900			goto next;
1901		}
1902
1903		/*
1904		 * nextbit is not contiguous with the current end index. Convert
1905		 * the current start/end to an extent and add it to the free
1906		 * list.
1907		 */
1908		agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
1909				  mp->m_sb.sb_inopblock;
1910		contigblk = ((endidx - startidx + 1) *
1911			     XFS_INODES_PER_HOLEMASK_BIT) /
1912			    mp->m_sb.sb_inopblock;
1913
1914		ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
1915		ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
1916		xfs_bmap_add_free(tp, XFS_AGB_TO_FSB(mp, agno, agbno),
1917				  contigblk, &XFS_RMAP_OINFO_INODES);
 
 
 
 
1918
1919		/* reset range to current bit and carry on... */
1920		startidx = endidx = nextbit;
1921
1922next:
1923		nextbit++;
1924	}
 
1925}
1926
1927STATIC int
1928xfs_difree_inobt(
1929	struct xfs_mount		*mp,
1930	struct xfs_trans		*tp,
1931	struct xfs_buf			*agbp,
1932	xfs_agino_t			agino,
1933	struct xfs_icluster		*xic,
1934	struct xfs_inobt_rec_incore	*orec)
1935{
 
1936	struct xfs_agi			*agi = agbp->b_addr;
1937	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
1938	struct xfs_btree_cur		*cur;
1939	struct xfs_inobt_rec_incore	rec;
1940	int				ilen;
1941	int				error;
1942	int				i;
1943	int				off;
1944
1945	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
1946	ASSERT(XFS_AGINO_TO_AGBNO(mp, agino) < be32_to_cpu(agi->agi_length));
1947
1948	/*
1949	 * Initialize the cursor.
1950	 */
1951	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
1952
1953	error = xfs_check_agi_freecount(cur, agi);
1954	if (error)
1955		goto error0;
1956
1957	/*
1958	 * Look for the entry describing this inode.
1959	 */
1960	if ((error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i))) {
1961		xfs_warn(mp, "%s: xfs_inobt_lookup() returned error %d.",
1962			__func__, error);
1963		goto error0;
1964	}
1965	if (XFS_IS_CORRUPT(mp, i != 1)) {
 
1966		error = -EFSCORRUPTED;
1967		goto error0;
1968	}
1969	error = xfs_inobt_get_rec(cur, &rec, &i);
1970	if (error) {
1971		xfs_warn(mp, "%s: xfs_inobt_get_rec() returned error %d.",
1972			__func__, error);
1973		goto error0;
1974	}
1975	if (XFS_IS_CORRUPT(mp, i != 1)) {
 
1976		error = -EFSCORRUPTED;
1977		goto error0;
1978	}
1979	/*
1980	 * Get the offset in the inode chunk.
1981	 */
1982	off = agino - rec.ir_startino;
1983	ASSERT(off >= 0 && off < XFS_INODES_PER_CHUNK);
1984	ASSERT(!(rec.ir_free & XFS_INOBT_MASK(off)));
1985	/*
1986	 * Mark the inode free & increment the count.
1987	 */
1988	rec.ir_free |= XFS_INOBT_MASK(off);
1989	rec.ir_freecount++;
1990
1991	/*
1992	 * When an inode chunk is free, it becomes eligible for removal. Don't
1993	 * remove the chunk if the block size is large enough for multiple inode
1994	 * chunks (that might not be free).
1995	 */
1996	if (!(mp->m_flags & XFS_MOUNT_IKEEP) &&
1997	    rec.ir_free == XFS_INOBT_ALL_FREE &&
1998	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
1999		struct xfs_perag	*pag = agbp->b_pag;
2000
2001		xic->deleted = true;
2002		xic->first_ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino);
 
2003		xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
2004
2005		/*
2006		 * Remove the inode cluster from the AGI B+Tree, adjust the
2007		 * AGI and Superblock inode counts, and mark the disk space
2008		 * to be freed when the transaction is committed.
2009		 */
2010		ilen = rec.ir_freecount;
2011		be32_add_cpu(&agi->agi_count, -ilen);
2012		be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
2013		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
2014		pag->pagi_freecount -= ilen - 1;
2015		pag->pagi_count -= ilen;
2016		xfs_trans_mod_sb(tp, XFS_TRANS_SB_ICOUNT, -ilen);
2017		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, -(ilen - 1));
2018
2019		if ((error = xfs_btree_delete(cur, &i))) {
2020			xfs_warn(mp, "%s: xfs_btree_delete returned error %d.",
2021				__func__, error);
2022			goto error0;
2023		}
2024
2025		xfs_difree_inode_chunk(tp, agno, &rec);
 
 
2026	} else {
2027		xic->deleted = false;
2028
2029		error = xfs_inobt_update(cur, &rec);
2030		if (error) {
2031			xfs_warn(mp, "%s: xfs_inobt_update returned error %d.",
2032				__func__, error);
2033			goto error0;
2034		}
2035
2036		/* 
2037		 * Change the inode free counts and log the ag/sb changes.
2038		 */
2039		be32_add_cpu(&agi->agi_freecount, 1);
2040		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_FREECOUNT);
2041		agbp->b_pag->pagi_freecount++;
2042		xfs_trans_mod_sb(tp, XFS_TRANS_SB_IFREE, 1);
2043	}
2044
2045	error = xfs_check_agi_freecount(cur, agi);
2046	if (error)
2047		goto error0;
2048
2049	*orec = rec;
2050	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2051	return 0;
2052
2053error0:
2054	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2055	return error;
2056}
2057
2058/*
2059 * Free an inode in the free inode btree.
2060 */
2061STATIC int
2062xfs_difree_finobt(
2063	struct xfs_mount		*mp,
2064	struct xfs_trans		*tp,
2065	struct xfs_buf			*agbp,
2066	xfs_agino_t			agino,
2067	struct xfs_inobt_rec_incore	*ibtrec) /* inobt record */
2068{
2069	struct xfs_agi			*agi = agbp->b_addr;
2070	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
2071	struct xfs_btree_cur		*cur;
2072	struct xfs_inobt_rec_incore	rec;
2073	int				offset = agino - ibtrec->ir_startino;
2074	int				error;
2075	int				i;
2076
2077	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_FINO);
2078
2079	error = xfs_inobt_lookup(cur, ibtrec->ir_startino, XFS_LOOKUP_EQ, &i);
2080	if (error)
2081		goto error;
2082	if (i == 0) {
2083		/*
2084		 * If the record does not exist in the finobt, we must have just
2085		 * freed an inode in a previously fully allocated chunk. If not,
2086		 * something is out of sync.
2087		 */
2088		if (XFS_IS_CORRUPT(mp, ibtrec->ir_freecount != 1)) {
 
2089			error = -EFSCORRUPTED;
2090			goto error;
2091		}
2092
2093		error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
2094					     ibtrec->ir_count,
2095					     ibtrec->ir_freecount,
2096					     ibtrec->ir_free, &i);
2097		if (error)
2098			goto error;
2099		ASSERT(i == 1);
2100
2101		goto out;
2102	}
2103
2104	/*
2105	 * Read and update the existing record. We could just copy the ibtrec
2106	 * across here, but that would defeat the purpose of having redundant
2107	 * metadata. By making the modifications independently, we can catch
2108	 * corruptions that we wouldn't see if we just copied from one record
2109	 * to another.
2110	 */
2111	error = xfs_inobt_get_rec(cur, &rec, &i);
2112	if (error)
2113		goto error;
2114	if (XFS_IS_CORRUPT(mp, i != 1)) {
 
2115		error = -EFSCORRUPTED;
2116		goto error;
2117	}
2118
2119	rec.ir_free |= XFS_INOBT_MASK(offset);
2120	rec.ir_freecount++;
2121
2122	if (XFS_IS_CORRUPT(mp,
2123			   rec.ir_free != ibtrec->ir_free ||
2124			   rec.ir_freecount != ibtrec->ir_freecount)) {
 
2125		error = -EFSCORRUPTED;
2126		goto error;
2127	}
2128
2129	/*
2130	 * The content of inobt records should always match between the inobt
2131	 * and finobt. The lifecycle of records in the finobt is different from
2132	 * the inobt in that the finobt only tracks records with at least one
2133	 * free inode. Hence, if all of the inodes are free and we aren't
2134	 * keeping inode chunks permanently on disk, remove the record.
2135	 * Otherwise, update the record with the new information.
2136	 *
2137	 * Note that we currently can't free chunks when the block size is large
2138	 * enough for multiple chunks. Leave the finobt record to remain in sync
2139	 * with the inobt.
2140	 */
2141	if (rec.ir_free == XFS_INOBT_ALL_FREE &&
2142	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK &&
2143	    !(mp->m_flags & XFS_MOUNT_IKEEP)) {
2144		error = xfs_btree_delete(cur, &i);
2145		if (error)
2146			goto error;
2147		ASSERT(i == 1);
2148	} else {
2149		error = xfs_inobt_update(cur, &rec);
2150		if (error)
2151			goto error;
2152	}
2153
2154out:
2155	error = xfs_check_agi_freecount(cur, agi);
2156	if (error)
2157		goto error;
2158
2159	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
2160	return 0;
2161
2162error:
2163	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
2164	return error;
2165}
2166
2167/*
2168 * Free disk inode.  Carefully avoids touching the incore inode, all
2169 * manipulations incore are the caller's responsibility.
2170 * The on-disk inode is not changed by this operation, only the
2171 * btree (free inode mask) is changed.
2172 */
2173int
2174xfs_difree(
2175	struct xfs_trans	*tp,		/* transaction pointer */
2176	xfs_ino_t		inode,		/* inode to be freed */
2177	struct xfs_icluster	*xic)	/* cluster info if deleted */
 
2178{
2179	/* REFERENCED */
2180	xfs_agblock_t		agbno;	/* block number containing inode */
2181	struct xfs_buf		*agbp;	/* buffer for allocation group header */
2182	xfs_agino_t		agino;	/* allocation group inode number */
2183	xfs_agnumber_t		agno;	/* allocation group number */
2184	int			error;	/* error return value */
2185	struct xfs_mount	*mp;	/* mount structure for filesystem */
2186	struct xfs_inobt_rec_incore rec;/* btree record */
2187
2188	mp = tp->t_mountp;
2189
2190	/*
2191	 * Break up inode number into its components.
2192	 */
2193	agno = XFS_INO_TO_AGNO(mp, inode);
2194	if (agno >= mp->m_sb.sb_agcount)  {
2195		xfs_warn(mp, "%s: agno >= mp->m_sb.sb_agcount (%d >= %d).",
2196			__func__, agno, mp->m_sb.sb_agcount);
2197		ASSERT(0);
2198		return -EINVAL;
2199	}
2200	agino = XFS_INO_TO_AGINO(mp, inode);
2201	if (inode != XFS_AGINO_TO_INO(mp, agno, agino))  {
2202		xfs_warn(mp, "%s: inode != XFS_AGINO_TO_INO() (%llu != %llu).",
2203			__func__, (unsigned long long)inode,
2204			(unsigned long long)XFS_AGINO_TO_INO(mp, agno, agino));
2205		ASSERT(0);
2206		return -EINVAL;
2207	}
2208	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2209	if (agbno >= mp->m_sb.sb_agblocks)  {
2210		xfs_warn(mp, "%s: agbno >= mp->m_sb.sb_agblocks (%d >= %d).",
2211			__func__, agbno, mp->m_sb.sb_agblocks);
2212		ASSERT(0);
2213		return -EINVAL;
2214	}
2215	/*
2216	 * Get the allocation group header.
2217	 */
2218	error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2219	if (error) {
2220		xfs_warn(mp, "%s: xfs_ialloc_read_agi() returned error %d.",
2221			__func__, error);
2222		return error;
2223	}
2224
2225	/*
2226	 * Fix up the inode allocation btree.
2227	 */
2228	error = xfs_difree_inobt(mp, tp, agbp, agino, xic, &rec);
2229	if (error)
2230		goto error0;
2231
2232	/*
2233	 * Fix up the free inode btree.
2234	 */
2235	if (xfs_sb_version_hasfinobt(&mp->m_sb)) {
2236		error = xfs_difree_finobt(mp, tp, agbp, agino, &rec);
2237		if (error)
2238			goto error0;
2239	}
2240
2241	return 0;
2242
2243error0:
2244	return error;
2245}
2246
2247STATIC int
2248xfs_imap_lookup(
2249	struct xfs_mount	*mp,
2250	struct xfs_trans	*tp,
2251	xfs_agnumber_t		agno,
2252	xfs_agino_t		agino,
2253	xfs_agblock_t		agbno,
2254	xfs_agblock_t		*chunk_agbno,
2255	xfs_agblock_t		*offset_agbno,
2256	int			flags)
2257{
 
2258	struct xfs_inobt_rec_incore rec;
2259	struct xfs_btree_cur	*cur;
2260	struct xfs_buf		*agbp;
2261	int			error;
2262	int			i;
2263
2264	error = xfs_ialloc_read_agi(mp, tp, agno, &agbp);
2265	if (error) {
2266		xfs_alert(mp,
2267			"%s: xfs_ialloc_read_agi() returned error %d, agno %d",
2268			__func__, error, agno);
2269		return error;
2270	}
2271
2272	/*
2273	 * Lookup the inode record for the given agino. If the record cannot be
2274	 * found, then it's an invalid inode number and we should abort. Once
2275	 * we have a record, we need to ensure it contains the inode number
2276	 * we are looking up.
2277	 */
2278	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, XFS_BTNUM_INO);
2279	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &i);
2280	if (!error) {
2281		if (i)
2282			error = xfs_inobt_get_rec(cur, &rec, &i);
2283		if (!error && i == 0)
2284			error = -EINVAL;
2285	}
2286
2287	xfs_trans_brelse(tp, agbp);
2288	xfs_btree_del_cursor(cur, error);
2289	if (error)
2290		return error;
2291
2292	/* check that the returned record contains the required inode */
2293	if (rec.ir_startino > agino ||
2294	    rec.ir_startino + M_IGEO(mp)->ialloc_inos <= agino)
2295		return -EINVAL;
2296
2297	/* for untrusted inodes check it is allocated first */
2298	if ((flags & XFS_IGET_UNTRUSTED) &&
2299	    (rec.ir_free & XFS_INOBT_MASK(agino - rec.ir_startino)))
2300		return -EINVAL;
2301
2302	*chunk_agbno = XFS_AGINO_TO_AGBNO(mp, rec.ir_startino);
2303	*offset_agbno = agbno - *chunk_agbno;
2304	return 0;
2305}
2306
2307/*
2308 * Return the location of the inode in imap, for mapping it into a buffer.
2309 */
2310int
2311xfs_imap(
2312	xfs_mount_t	 *mp,	/* file system mount structure */
2313	xfs_trans_t	 *tp,	/* transaction pointer */
2314	xfs_ino_t	ino,	/* inode to locate */
2315	struct xfs_imap	*imap,	/* location map structure */
2316	uint		flags)	/* flags for inode btree lookup */
2317{
2318	xfs_agblock_t	agbno;	/* block number of inode in the alloc group */
2319	xfs_agino_t	agino;	/* inode number within alloc group */
2320	xfs_agnumber_t	agno;	/* allocation group number */
2321	xfs_agblock_t	chunk_agbno;	/* first block in inode chunk */
2322	xfs_agblock_t	cluster_agbno;	/* first block in inode cluster */
2323	int		error;	/* error code */
2324	int		offset;	/* index of inode in its buffer */
2325	xfs_agblock_t	offset_agbno;	/* blks from chunk start to inode */
2326
2327	ASSERT(ino != NULLFSINO);
2328
2329	/*
2330	 * Split up the inode number into its parts.
2331	 */
2332	agno = XFS_INO_TO_AGNO(mp, ino);
2333	agino = XFS_INO_TO_AGINO(mp, ino);
2334	agbno = XFS_AGINO_TO_AGBNO(mp, agino);
2335	if (agno >= mp->m_sb.sb_agcount || agbno >= mp->m_sb.sb_agblocks ||
2336	    ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
 
2337#ifdef DEBUG
2338		/*
2339		 * Don't output diagnostic information for untrusted inodes
2340		 * as they can be invalid without implying corruption.
2341		 */
2342		if (flags & XFS_IGET_UNTRUSTED)
2343			return -EINVAL;
2344		if (agno >= mp->m_sb.sb_agcount) {
2345			xfs_alert(mp,
2346				"%s: agno (%d) >= mp->m_sb.sb_agcount (%d)",
2347				__func__, agno, mp->m_sb.sb_agcount);
2348		}
2349		if (agbno >= mp->m_sb.sb_agblocks) {
2350			xfs_alert(mp,
2351		"%s: agbno (0x%llx) >= mp->m_sb.sb_agblocks (0x%lx)",
2352				__func__, (unsigned long long)agbno,
2353				(unsigned long)mp->m_sb.sb_agblocks);
2354		}
2355		if (ino != XFS_AGINO_TO_INO(mp, agno, agino)) {
2356			xfs_alert(mp,
2357		"%s: ino (0x%llx) != XFS_AGINO_TO_INO() (0x%llx)",
2358				__func__, ino,
2359				XFS_AGINO_TO_INO(mp, agno, agino));
2360		}
2361		xfs_stack_trace();
2362#endif /* DEBUG */
2363		return -EINVAL;
2364	}
2365
2366	/*
2367	 * For bulkstat and handle lookups, we have an untrusted inode number
2368	 * that we have to verify is valid. We cannot do this just by reading
2369	 * the inode buffer as it may have been unlinked and removed leaving
2370	 * inodes in stale state on disk. Hence we have to do a btree lookup
2371	 * in all cases where an untrusted inode number is passed.
2372	 */
2373	if (flags & XFS_IGET_UNTRUSTED) {
2374		error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2375					&chunk_agbno, &offset_agbno, flags);
2376		if (error)
2377			return error;
2378		goto out_map;
2379	}
2380
2381	/*
2382	 * If the inode cluster size is the same as the blocksize or
2383	 * smaller we get to the buffer by simple arithmetics.
2384	 */
2385	if (M_IGEO(mp)->blocks_per_cluster == 1) {
2386		offset = XFS_INO_TO_OFFSET(mp, ino);
2387		ASSERT(offset < mp->m_sb.sb_inopblock);
2388
2389		imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, agbno);
2390		imap->im_len = XFS_FSB_TO_BB(mp, 1);
2391		imap->im_boffset = (unsigned short)(offset <<
2392							mp->m_sb.sb_inodelog);
2393		return 0;
2394	}
2395
2396	/*
2397	 * If the inode chunks are aligned then use simple maths to
2398	 * find the location. Otherwise we have to do a btree
2399	 * lookup to find the location.
2400	 */
2401	if (M_IGEO(mp)->inoalign_mask) {
2402		offset_agbno = agbno & M_IGEO(mp)->inoalign_mask;
2403		chunk_agbno = agbno - offset_agbno;
2404	} else {
2405		error = xfs_imap_lookup(mp, tp, agno, agino, agbno,
2406					&chunk_agbno, &offset_agbno, flags);
2407		if (error)
2408			return error;
2409	}
2410
2411out_map:
2412	ASSERT(agbno >= chunk_agbno);
2413	cluster_agbno = chunk_agbno +
2414		((offset_agbno / M_IGEO(mp)->blocks_per_cluster) *
2415		 M_IGEO(mp)->blocks_per_cluster);
2416	offset = ((agbno - cluster_agbno) * mp->m_sb.sb_inopblock) +
2417		XFS_INO_TO_OFFSET(mp, ino);
2418
2419	imap->im_blkno = XFS_AGB_TO_DADDR(mp, agno, cluster_agbno);
2420	imap->im_len = XFS_FSB_TO_BB(mp, M_IGEO(mp)->blocks_per_cluster);
2421	imap->im_boffset = (unsigned short)(offset << mp->m_sb.sb_inodelog);
2422
2423	/*
2424	 * If the inode number maps to a block outside the bounds
2425	 * of the file system then return NULL rather than calling
2426	 * read_buf and panicing when we get an error from the
2427	 * driver.
2428	 */
2429	if ((imap->im_blkno + imap->im_len) >
2430	    XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks)) {
2431		xfs_alert(mp,
2432	"%s: (im_blkno (0x%llx) + im_len (0x%llx)) > sb_dblocks (0x%llx)",
2433			__func__, (unsigned long long) imap->im_blkno,
2434			(unsigned long long) imap->im_len,
2435			XFS_FSB_TO_BB(mp, mp->m_sb.sb_dblocks));
2436		return -EINVAL;
2437	}
2438	return 0;
2439}
2440
2441/*
2442 * Log specified fields for the ag hdr (inode section). The growth of the agi
2443 * structure over time requires that we interpret the buffer as two logical
2444 * regions delineated by the end of the unlinked list. This is due to the size
2445 * of the hash table and its location in the middle of the agi.
2446 *
2447 * For example, a request to log a field before agi_unlinked and a field after
2448 * agi_unlinked could cause us to log the entire hash table and use an excessive
2449 * amount of log space. To avoid this behavior, log the region up through
2450 * agi_unlinked in one call and the region after agi_unlinked through the end of
2451 * the structure in another.
2452 */
2453void
2454xfs_ialloc_log_agi(
2455	xfs_trans_t	*tp,		/* transaction pointer */
2456	xfs_buf_t	*bp,		/* allocation group header buffer */
2457	int		fields)		/* bitmask of fields to log */
2458{
2459	int			first;		/* first byte number */
2460	int			last;		/* last byte number */
2461	static const short	offsets[] = {	/* field starting offsets */
2462					/* keep in sync with bit definitions */
2463		offsetof(xfs_agi_t, agi_magicnum),
2464		offsetof(xfs_agi_t, agi_versionnum),
2465		offsetof(xfs_agi_t, agi_seqno),
2466		offsetof(xfs_agi_t, agi_length),
2467		offsetof(xfs_agi_t, agi_count),
2468		offsetof(xfs_agi_t, agi_root),
2469		offsetof(xfs_agi_t, agi_level),
2470		offsetof(xfs_agi_t, agi_freecount),
2471		offsetof(xfs_agi_t, agi_newino),
2472		offsetof(xfs_agi_t, agi_dirino),
2473		offsetof(xfs_agi_t, agi_unlinked),
2474		offsetof(xfs_agi_t, agi_free_root),
2475		offsetof(xfs_agi_t, agi_free_level),
 
2476		sizeof(xfs_agi_t)
2477	};
2478#ifdef DEBUG
2479	struct xfs_agi		*agi = bp->b_addr;
2480
2481	ASSERT(agi->agi_magicnum == cpu_to_be32(XFS_AGI_MAGIC));
2482#endif
2483
2484	/*
2485	 * Compute byte offsets for the first and last fields in the first
2486	 * region and log the agi buffer. This only logs up through
2487	 * agi_unlinked.
2488	 */
2489	if (fields & XFS_AGI_ALL_BITS_R1) {
2490		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R1,
2491				  &first, &last);
2492		xfs_trans_log_buf(tp, bp, first, last);
2493	}
2494
2495	/*
2496	 * Mask off the bits in the first region and calculate the first and
2497	 * last field offsets for any bits in the second region.
2498	 */
2499	fields &= ~XFS_AGI_ALL_BITS_R1;
2500	if (fields) {
2501		xfs_btree_offsets(fields, offsets, XFS_AGI_NUM_BITS_R2,
2502				  &first, &last);
2503		xfs_trans_log_buf(tp, bp, first, last);
2504	}
2505}
2506
2507static xfs_failaddr_t
2508xfs_agi_verify(
2509	struct xfs_buf	*bp)
2510{
2511	struct xfs_mount *mp = bp->b_mount;
2512	struct xfs_agi	*agi = bp->b_addr;
2513	int		i;
 
 
 
2514
2515	if (xfs_sb_version_hascrc(&mp->m_sb)) {
2516		if (!uuid_equal(&agi->agi_uuid, &mp->m_sb.sb_meta_uuid))
2517			return __this_address;
2518		if (!xfs_log_check_lsn(mp, be64_to_cpu(agi->agi_lsn)))
2519			return __this_address;
2520	}
2521
2522	/*
2523	 * Validate the magic number of the agi block.
2524	 */
2525	if (!xfs_verify_magic(bp, agi->agi_magicnum))
2526		return __this_address;
2527	if (!XFS_AGI_GOOD_VERSION(be32_to_cpu(agi->agi_versionnum)))
2528		return __this_address;
2529
 
 
 
 
2530	if (be32_to_cpu(agi->agi_level) < 1 ||
2531	    be32_to_cpu(agi->agi_level) > XFS_BTREE_MAXLEVELS)
2532		return __this_address;
2533
2534	if (xfs_sb_version_hasfinobt(&mp->m_sb) &&
2535	    (be32_to_cpu(agi->agi_free_level) < 1 ||
2536	     be32_to_cpu(agi->agi_free_level) > XFS_BTREE_MAXLEVELS))
2537		return __this_address;
2538
2539	/*
2540	 * during growfs operations, the perag is not fully initialised,
2541	 * so we can't use it for any useful checking. growfs ensures we can't
2542	 * use it by using uncached buffers that don't have the perag attached
2543	 * so we can detect and avoid this problem.
2544	 */
2545	if (bp->b_pag && be32_to_cpu(agi->agi_seqno) != bp->b_pag->pag_agno)
2546		return __this_address;
2547
2548	for (i = 0; i < XFS_AGI_UNLINKED_BUCKETS; i++) {
2549		if (agi->agi_unlinked[i] == cpu_to_be32(NULLAGINO))
2550			continue;
2551		if (!xfs_verify_ino(mp, be32_to_cpu(agi->agi_unlinked[i])))
2552			return __this_address;
2553	}
2554
2555	return NULL;
2556}
2557
2558static void
2559xfs_agi_read_verify(
2560	struct xfs_buf	*bp)
2561{
2562	struct xfs_mount *mp = bp->b_mount;
2563	xfs_failaddr_t	fa;
2564
2565	if (xfs_sb_version_hascrc(&mp->m_sb) &&
2566	    !xfs_buf_verify_cksum(bp, XFS_AGI_CRC_OFF))
2567		xfs_verifier_error(bp, -EFSBADCRC, __this_address);
2568	else {
2569		fa = xfs_agi_verify(bp);
2570		if (XFS_TEST_ERROR(fa, mp, XFS_ERRTAG_IALLOC_READ_AGI))
2571			xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2572	}
2573}
2574
2575static void
2576xfs_agi_write_verify(
2577	struct xfs_buf	*bp)
2578{
2579	struct xfs_mount	*mp = bp->b_mount;
2580	struct xfs_buf_log_item	*bip = bp->b_log_item;
2581	struct xfs_agi		*agi = bp->b_addr;
2582	xfs_failaddr_t		fa;
2583
2584	fa = xfs_agi_verify(bp);
2585	if (fa) {
2586		xfs_verifier_error(bp, -EFSCORRUPTED, fa);
2587		return;
2588	}
2589
2590	if (!xfs_sb_version_hascrc(&mp->m_sb))
2591		return;
2592
2593	if (bip)
2594		agi->agi_lsn = cpu_to_be64(bip->bli_item.li_lsn);
2595	xfs_buf_update_cksum(bp, XFS_AGI_CRC_OFF);
2596}
2597
2598const struct xfs_buf_ops xfs_agi_buf_ops = {
2599	.name = "xfs_agi",
2600	.magic = { cpu_to_be32(XFS_AGI_MAGIC), cpu_to_be32(XFS_AGI_MAGIC) },
2601	.verify_read = xfs_agi_read_verify,
2602	.verify_write = xfs_agi_write_verify,
2603	.verify_struct = xfs_agi_verify,
2604};
2605
2606/*
2607 * Read in the allocation group header (inode allocation section)
2608 */
2609int
2610xfs_read_agi(
2611	struct xfs_mount	*mp,	/* file system mount structure */
2612	struct xfs_trans	*tp,	/* transaction pointer */
2613	xfs_agnumber_t		agno,	/* allocation group number */
2614	struct xfs_buf		**bpp)	/* allocation group hdr buf */
2615{
 
2616	int			error;
2617
2618	trace_xfs_read_agi(mp, agno);
2619
2620	ASSERT(agno != NULLAGNUMBER);
2621	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
2622			XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
2623			XFS_FSS_TO_BB(mp, 1), 0, bpp, &xfs_agi_buf_ops);
 
 
2624	if (error)
2625		return error;
2626	if (tp)
2627		xfs_trans_buf_set_type(tp, *bpp, XFS_BLFT_AGI_BUF);
2628
2629	xfs_buf_set_ref(*bpp, XFS_AGI_REF);
2630	return 0;
2631}
2632
 
 
 
 
2633int
2634xfs_ialloc_read_agi(
2635	struct xfs_mount	*mp,	/* file system mount structure */
2636	struct xfs_trans	*tp,	/* transaction pointer */
2637	xfs_agnumber_t		agno,	/* allocation group number */
2638	struct xfs_buf		**bpp)	/* allocation group hdr buf */
2639{
2640	struct xfs_agi		*agi;	/* allocation group header */
2641	struct xfs_perag	*pag;	/* per allocation group data */
2642	int			error;
2643
2644	trace_xfs_ialloc_read_agi(mp, agno);
2645
2646	error = xfs_read_agi(mp, tp, agno, bpp);
2647	if (error)
2648		return error;
2649
2650	agi = (*bpp)->b_addr;
2651	pag = (*bpp)->b_pag;
2652	if (!pag->pagi_init) {
2653		pag->pagi_freecount = be32_to_cpu(agi->agi_freecount);
2654		pag->pagi_count = be32_to_cpu(agi->agi_count);
2655		pag->pagi_init = 1;
2656	}
2657
2658	/*
2659	 * It's possible for these to be out of sync if
2660	 * we are in the middle of a forced shutdown.
2661	 */
2662	ASSERT(pag->pagi_freecount == be32_to_cpu(agi->agi_freecount) ||
2663		XFS_FORCED_SHUTDOWN(mp));
 
 
 
 
2664	return 0;
2665}
2666
2667/*
2668 * Read in the agi to initialise the per-ag data in the mount structure
2669 */
2670int
2671xfs_ialloc_pagi_init(
2672	xfs_mount_t	*mp,		/* file system mount structure */
2673	xfs_trans_t	*tp,		/* transaction pointer */
2674	xfs_agnumber_t	agno)		/* allocation group number */
2675{
2676	xfs_buf_t	*bp = NULL;
2677	int		error;
 
 
2678
2679	error = xfs_ialloc_read_agi(mp, tp, agno, &bp);
2680	if (error)
2681		return error;
2682	if (bp)
2683		xfs_trans_brelse(tp, bp);
2684	return 0;
2685}
2686
2687/* Is there an inode record covering a given range of inode numbers? */
2688int
2689xfs_ialloc_has_inode_record(
2690	struct xfs_btree_cur	*cur,
2691	xfs_agino_t		low,
2692	xfs_agino_t		high,
2693	bool			*exists)
2694{
2695	struct xfs_inobt_rec_incore	irec;
2696	xfs_agino_t		agino;
2697	uint16_t		holemask;
2698	int			has_record;
2699	int			i;
2700	int			error;
2701
2702	*exists = false;
2703	error = xfs_inobt_lookup(cur, low, XFS_LOOKUP_LE, &has_record);
2704	while (error == 0 && has_record) {
2705		error = xfs_inobt_get_rec(cur, &irec, &has_record);
2706		if (error || irec.ir_startino > high)
 
 
2707			break;
2708
2709		agino = irec.ir_startino;
2710		holemask = irec.ir_holemask;
2711		for (i = 0; i < XFS_INOBT_HOLEMASK_BITS; holemask >>= 1,
2712				i++, agino += XFS_INODES_PER_HOLEMASK_BIT) {
2713			if (holemask & 1)
2714				continue;
2715			if (agino + XFS_INODES_PER_HOLEMASK_BIT > low &&
2716					agino <= high) {
2717				*exists = true;
2718				return 0;
2719			}
 
2720		}
2721
2722		error = xfs_btree_increment(cur, 0, &has_record);
 
 
2723	}
2724	return error;
 
 
2725}
2726
2727/* Is there an inode record covering a given extent? */
2728int
2729xfs_ialloc_has_inodes_at_extent(
2730	struct xfs_btree_cur	*cur,
2731	xfs_agblock_t		bno,
2732	xfs_extlen_t		len,
2733	bool			*exists)
2734{
2735	xfs_agino_t		low;
2736	xfs_agino_t		high;
 
 
2737
2738	low = XFS_AGB_TO_AGINO(cur->bc_mp, bno);
2739	high = XFS_AGB_TO_AGINO(cur->bc_mp, bno + len) - 1;
2740
2741	return xfs_ialloc_has_inode_record(cur, low, high, exists);
 
 
 
 
 
 
 
 
 
 
2742}
2743
2744struct xfs_ialloc_count_inodes {
2745	xfs_agino_t			count;
2746	xfs_agino_t			freecount;
2747};
2748
2749/* Record inode counts across all inobt records. */
2750STATIC int
2751xfs_ialloc_count_inodes_rec(
2752	struct xfs_btree_cur		*cur,
2753	union xfs_btree_rec		*rec,
2754	void				*priv)
2755{
2756	struct xfs_inobt_rec_incore	irec;
2757	struct xfs_ialloc_count_inodes	*ci = priv;
 
2758
2759	xfs_inobt_btrec_to_irec(cur->bc_mp, rec, &irec);
 
 
 
 
2760	ci->count += irec.ir_count;
2761	ci->freecount += irec.ir_freecount;
2762
2763	return 0;
2764}
2765
2766/* Count allocated and free inodes under an inobt. */
2767int
2768xfs_ialloc_count_inodes(
2769	struct xfs_btree_cur		*cur,
2770	xfs_agino_t			*count,
2771	xfs_agino_t			*freecount)
2772{
2773	struct xfs_ialloc_count_inodes	ci = {0};
2774	int				error;
2775
2776	ASSERT(cur->bc_btnum == XFS_BTNUM_INO);
2777	error = xfs_btree_query_all(cur, xfs_ialloc_count_inodes_rec, &ci);
2778	if (error)
2779		return error;
2780
2781	*count = ci.count;
2782	*freecount = ci.freecount;
2783	return 0;
2784}
2785
2786/*
2787 * Initialize inode-related geometry information.
2788 *
2789 * Compute the inode btree min and max levels and set maxicount.
2790 *
2791 * Set the inode cluster size.  This may still be overridden by the file
2792 * system block size if it is larger than the chosen cluster size.
2793 *
2794 * For v5 filesystems, scale the cluster size with the inode size to keep a
2795 * constant ratio of inode per cluster buffer, but only if mkfs has set the
2796 * inode alignment value appropriately for larger cluster sizes.
2797 *
2798 * Then compute the inode cluster alignment information.
2799 */
2800void
2801xfs_ialloc_setup_geometry(
2802	struct xfs_mount	*mp)
2803{
2804	struct xfs_sb		*sbp = &mp->m_sb;
2805	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2806	uint64_t		icount;
2807	uint			inodes;
2808
 
 
 
 
 
 
2809	/* Compute inode btree geometry. */
2810	igeo->agino_log = sbp->sb_inopblog + sbp->sb_agblklog;
2811	igeo->inobt_mxr[0] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 1);
2812	igeo->inobt_mxr[1] = xfs_inobt_maxrecs(mp, sbp->sb_blocksize, 0);
2813	igeo->inobt_mnr[0] = igeo->inobt_mxr[0] / 2;
2814	igeo->inobt_mnr[1] = igeo->inobt_mxr[1] / 2;
2815
2816	igeo->ialloc_inos = max_t(uint16_t, XFS_INODES_PER_CHUNK,
2817			sbp->sb_inopblock);
2818	igeo->ialloc_blks = igeo->ialloc_inos >> sbp->sb_inopblog;
2819
2820	if (sbp->sb_spino_align)
2821		igeo->ialloc_min_blks = sbp->sb_spino_align;
2822	else
2823		igeo->ialloc_min_blks = igeo->ialloc_blks;
2824
2825	/* Compute and fill in value of m_ino_geo.inobt_maxlevels. */
2826	inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
2827	igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
2828			inodes);
 
2829
2830	/*
2831	 * Set the maximum inode count for this filesystem, being careful not
2832	 * to use obviously garbage sb_inopblog/sb_inopblock values.  Regular
2833	 * users should never get here due to failing sb verification, but
2834	 * certain users (xfs_db) need to be usable even with corrupt metadata.
2835	 */
2836	if (sbp->sb_imax_pct && igeo->ialloc_blks) {
2837		/*
2838		 * Make sure the maximum inode count is a multiple
2839		 * of the units we allocate inodes in.
2840		 */
2841		icount = sbp->sb_dblocks * sbp->sb_imax_pct;
2842		do_div(icount, 100);
2843		do_div(icount, igeo->ialloc_blks);
2844		igeo->maxicount = XFS_FSB_TO_INO(mp,
2845				icount * igeo->ialloc_blks);
2846	} else {
2847		igeo->maxicount = 0;
2848	}
2849
2850	/*
2851	 * Compute the desired size of an inode cluster buffer size, which
2852	 * starts at 8K and (on v5 filesystems) scales up with larger inode
2853	 * sizes.
2854	 *
2855	 * Preserve the desired inode cluster size because the sparse inodes
2856	 * feature uses that desired size (not the actual size) to compute the
2857	 * sparse inode alignment.  The mount code validates this value, so we
2858	 * cannot change the behavior.
2859	 */
2860	igeo->inode_cluster_size_raw = XFS_INODE_BIG_CLUSTER_SIZE;
2861	if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
2862		int	new_size = igeo->inode_cluster_size_raw;
2863
2864		new_size *= mp->m_sb.sb_inodesize / XFS_DINODE_MIN_SIZE;
2865		if (mp->m_sb.sb_inoalignmt >= XFS_B_TO_FSBT(mp, new_size))
2866			igeo->inode_cluster_size_raw = new_size;
2867	}
2868
2869	/* Calculate inode cluster ratios. */
2870	if (igeo->inode_cluster_size_raw > mp->m_sb.sb_blocksize)
2871		igeo->blocks_per_cluster = XFS_B_TO_FSBT(mp,
2872				igeo->inode_cluster_size_raw);
2873	else
2874		igeo->blocks_per_cluster = 1;
2875	igeo->inode_cluster_size = XFS_FSB_TO_B(mp, igeo->blocks_per_cluster);
2876	igeo->inodes_per_cluster = XFS_FSB_TO_INO(mp, igeo->blocks_per_cluster);
2877
2878	/* Calculate inode cluster alignment. */
2879	if (xfs_sb_version_hasalign(&mp->m_sb) &&
2880	    mp->m_sb.sb_inoalignmt >= igeo->blocks_per_cluster)
2881		igeo->cluster_align = mp->m_sb.sb_inoalignmt;
2882	else
2883		igeo->cluster_align = 1;
2884	igeo->inoalign_mask = igeo->cluster_align - 1;
2885	igeo->cluster_align_inodes = XFS_FSB_TO_INO(mp, igeo->cluster_align);
2886
2887	/*
2888	 * If we are using stripe alignment, check whether
2889	 * the stripe unit is a multiple of the inode alignment
2890	 */
2891	if (mp->m_dalign && igeo->inoalign_mask &&
2892	    !(mp->m_dalign & igeo->inoalign_mask))
2893		igeo->ialloc_align = mp->m_dalign;
2894	else
2895		igeo->ialloc_align = 0;
2896}
2897
2898/* Compute the location of the root directory inode that is laid out by mkfs. */
2899xfs_ino_t
2900xfs_ialloc_calc_rootino(
2901	struct xfs_mount	*mp,
2902	int			sunit)
2903{
2904	struct xfs_ino_geometry	*igeo = M_IGEO(mp);
2905	xfs_agblock_t		first_bno;
2906
2907	/*
2908	 * Pre-calculate the geometry of AG 0.  We know what it looks like
2909	 * because libxfs knows how to create allocation groups now.
2910	 *
2911	 * first_bno is the first block in which mkfs could possibly have
2912	 * allocated the root directory inode, once we factor in the metadata
2913	 * that mkfs formats before it.  Namely, the four AG headers...
2914	 */
2915	first_bno = howmany(4 * mp->m_sb.sb_sectsize, mp->m_sb.sb_blocksize);
2916
2917	/* ...the two free space btree roots... */
2918	first_bno += 2;
2919
2920	/* ...the inode btree root... */
2921	first_bno += 1;
2922
2923	/* ...the initial AGFL... */
2924	first_bno += xfs_alloc_min_freelist(mp, NULL);
2925
2926	/* ...the free inode btree root... */
2927	if (xfs_sb_version_hasfinobt(&mp->m_sb))
2928		first_bno++;
2929
2930	/* ...the reverse mapping btree root... */
2931	if (xfs_sb_version_hasrmapbt(&mp->m_sb))
2932		first_bno++;
2933
2934	/* ...the reference count btree... */
2935	if (xfs_sb_version_hasreflink(&mp->m_sb))
2936		first_bno++;
2937
2938	/*
2939	 * ...and the log, if it is allocated in the first allocation group.
2940	 *
2941	 * This can happen with filesystems that only have a single
2942	 * allocation group, or very odd geometries created by old mkfs
2943	 * versions on very small filesystems.
2944	 */
2945	if (mp->m_sb.sb_logstart &&
2946	    XFS_FSB_TO_AGNO(mp, mp->m_sb.sb_logstart) == 0)
2947		 first_bno += mp->m_sb.sb_logblocks;
2948
2949	/*
2950	 * Now round first_bno up to whatever allocation alignment is given
2951	 * by the filesystem or was passed in.
2952	 */
2953	if (xfs_sb_version_hasdalign(&mp->m_sb) && igeo->ialloc_align > 0)
2954		first_bno = roundup(first_bno, sunit);
2955	else if (xfs_sb_version_hasalign(&mp->m_sb) &&
2956			mp->m_sb.sb_inoalignmt > 1)
2957		first_bno = roundup(first_bno, mp->m_sb.sb_inoalignmt);
2958
2959	return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2960}