/fs/xfs/xfs_file.c

   1/*
   2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3 * All Rights Reserved.
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it would be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public License
  15 * along with this program; if not, write the Free Software Foundation,
  16 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17 */
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_bit.h"
  21#include "xfs_log.h"
  22#include "xfs_inum.h"
  23#include "xfs_sb.h"
  24#include "xfs_ag.h"
  25#include "xfs_trans.h"
  26#include "xfs_mount.h"
  27#include "xfs_bmap_btree.h"
  28#include "xfs_alloc.h"
  29#include "xfs_dinode.h"
  30#include "xfs_inode.h"
  31#include "xfs_inode_item.h"
  32#include "xfs_bmap.h"
  33#include "xfs_error.h"
  34#include "xfs_vnodeops.h"
  35#include "xfs_da_btree.h"
  36#include "xfs_ioctl.h"
  37#include "xfs_trace.h"
  38
  39#include <linux/dcache.h>
  40#include <linux/falloc.h>
  41
  42static const struct vm_operations_struct xfs_file_vm_ops;
  43
  44/*
  45 * Locking primitives for read and write IO paths to ensure we consistently use
  46 * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
  47 */
  48static inline void
  49xfs_rw_ilock(
  50	struct xfs_inode	*ip,
  51	int			type)
  52{
  53	if (type & XFS_IOLOCK_EXCL)
  54		mutex_lock(&VFS_I(ip)->i_mutex);
  55	xfs_ilock(ip, type);
  56}
  57
  58static inline void
  59xfs_rw_iunlock(
  60	struct xfs_inode	*ip,
  61	int			type)
  62{
  63	xfs_iunlock(ip, type);
  64	if (type & XFS_IOLOCK_EXCL)
  65		mutex_unlock(&VFS_I(ip)->i_mutex);
  66}
  67
  68static inline void
  69xfs_rw_ilock_demote(
  70	struct xfs_inode	*ip,
  71	int			type)
  72{
  73	xfs_ilock_demote(ip, type);
  74	if (type & XFS_IOLOCK_EXCL)
  75		mutex_unlock(&VFS_I(ip)->i_mutex);
  76}
  77
  78/*
  79 *	xfs_iozero
  80 *
  81 *	xfs_iozero clears the specified range of buffer supplied,
  82 *	and marks all the affected blocks as valid and modified.  If
  83 *	an affected block is not allocated, it will be allocated.  If
  84 *	an affected block is not completely overwritten, and is not
  85 *	valid before the operation, it will be read from disk before
  86 *	being partially zeroed.
  87 */
  88STATIC int
  89xfs_iozero(
  90	struct xfs_inode	*ip,	/* inode			*/
  91	loff_t			pos,	/* offset in file		*/
  92	size_t			count)	/* size of data to zero		*/
  93{
  94	struct page		*page;
  95	struct address_space	*mapping;
  96	int			status;
  97
  98	mapping = VFS_I(ip)->i_mapping;
  99	do {
 100		unsigned offset, bytes;
 101		void *fsdata;
 102
 103		offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 104		bytes = PAGE_CACHE_SIZE - offset;
 105		if (bytes > count)
 106			bytes = count;
 107
 108		status = pagecache_write_begin(NULL, mapping, pos, bytes,
 109					AOP_FLAG_UNINTERRUPTIBLE,
 110					&page, &fsdata);
 111		if (status)
 112			break;
 113
 114		zero_user(page, offset, bytes);
 115
 116		status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
 117					page, fsdata);
 118		WARN_ON(status <= 0); /* can't return less than zero! */
 119		pos += bytes;
 120		count -= bytes;
 121		status = 0;
 122	} while (count);
 123
 124	return (-status);
 125}
 126
 127STATIC int
 128xfs_file_fsync(
 129	struct file		*file,
 130	loff_t			start,
 131	loff_t			end,
 132	int			datasync)
 133{
 134	struct inode		*inode = file->f_mapping->host;
 135	struct xfs_inode	*ip = XFS_I(inode);
 136	struct xfs_mount	*mp = ip->i_mount;
 137	struct xfs_trans	*tp;
 138	int			error = 0;
 139	int			log_flushed = 0;
 140
 141	trace_xfs_file_fsync(ip);
 142
 143	error = filemap_write_and_wait_range(inode->i_mapping, start, end);
 144	if (error)
 145		return error;
 146
 147	if (XFS_FORCED_SHUTDOWN(mp))
 148		return -XFS_ERROR(EIO);
 149
 150	xfs_iflags_clear(ip, XFS_ITRUNCATED);
 151
 152	xfs_ilock(ip, XFS_IOLOCK_SHARED);
 153	xfs_ioend_wait(ip);
 154	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 155
 156	if (mp->m_flags & XFS_MOUNT_BARRIER) {
 157		/*
 158		 * If we have an RT and/or log subvolume we need to make sure
 159		 * to flush the write cache the device used for file data
 160		 * first.  This is to ensure newly written file data make
 161		 * it to disk before logging the new inode size in case of
 162		 * an extending write.
 163		 */
 164		if (XFS_IS_REALTIME_INODE(ip))
 165			xfs_blkdev_issue_flush(mp->m_rtdev_targp);
 166		else if (mp->m_logdev_targp != mp->m_ddev_targp)
 167			xfs_blkdev_issue_flush(mp->m_ddev_targp);
 168	}
 169
 170	/*
 171	 * We always need to make sure that the required inode state is safe on
 172	 * disk.  The inode might be clean but we still might need to force the
 173	 * log because of committed transactions that haven't hit the disk yet.
 174	 * Likewise, there could be unflushed non-transactional changes to the
 175	 * inode core that have to go to disk and this requires us to issue
 176	 * a synchronous transaction to capture these changes correctly.
 177	 *
 178	 * This code relies on the assumption that if the i_update_core field
 179	 * of the inode is clear and the inode is unpinned then it is clean
 180	 * and no action is required.
 181	 */
 182	xfs_ilock(ip, XFS_ILOCK_SHARED);
 183
 184	/*
 185	 * First check if the VFS inode is marked dirty.  All the dirtying
 186	 * of non-transactional updates no goes through mark_inode_dirty*,
 187	 * which allows us to distinguish beteeen pure timestamp updates
 188	 * and i_size updates which need to be caught for fdatasync.
 189	 * After that also theck for the dirty state in the XFS inode, which
 190	 * might gets cleared when the inode gets written out via the AIL
 191	 * or xfs_iflush_cluster.
 192	 */
 193	if (((inode->i_state & I_DIRTY_DATASYNC) ||
 194	    ((inode->i_state & I_DIRTY_SYNC) && !datasync)) &&
 195	    ip->i_update_core) {
 196		/*
 197		 * Kick off a transaction to log the inode core to get the
 198		 * updates.  The sync transaction will also force the log.
 199		 */
 200		xfs_iunlock(ip, XFS_ILOCK_SHARED);
 201		tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
 202		error = xfs_trans_reserve(tp, 0,
 203				XFS_FSYNC_TS_LOG_RES(mp), 0, 0, 0);
 204		if (error) {
 205			xfs_trans_cancel(tp, 0);
 206			return -error;
 207		}
 208		xfs_ilock(ip, XFS_ILOCK_EXCL);
 209
 210		/*
 211		 * Note - it's possible that we might have pushed ourselves out
 212		 * of the way during trans_reserve which would flush the inode.
 213		 * But there's no guarantee that the inode buffer has actually
 214		 * gone out yet (it's delwri).	Plus the buffer could be pinned
 215		 * anyway if it's part of an inode in another recent
 216		 * transaction.	 So we play it safe and fire off the
 217		 * transaction anyway.
 218		 */
 219		xfs_trans_ijoin(tp, ip);
 220		xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 221		xfs_trans_set_sync(tp);
 222		error = _xfs_trans_commit(tp, 0, &log_flushed);
 223
 224		xfs_iunlock(ip, XFS_ILOCK_EXCL);
 225	} else {
 226		/*
 227		 * Timestamps/size haven't changed since last inode flush or
 228		 * inode transaction commit.  That means either nothing got
 229		 * written or a transaction committed which caught the updates.
 230		 * If the latter happened and the transaction hasn't hit the
 231		 * disk yet, the inode will be still be pinned.  If it is,
 232		 * force the log.
 233		 */
 234		if (xfs_ipincount(ip)) {
 235			error = _xfs_log_force_lsn(mp,
 236					ip->i_itemp->ili_last_lsn,
 237					XFS_LOG_SYNC, &log_flushed);
 238		}
 239		xfs_iunlock(ip, XFS_ILOCK_SHARED);
 240	}
 241
 242	/*
 243	 * If we only have a single device, and the log force about was
 244	 * a no-op we might have to flush the data device cache here.
 245	 * This can only happen for fdatasync/O_DSYNC if we were overwriting
 246	 * an already allocated file and thus do not have any metadata to
 247	 * commit.
 248	 */
 249	if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
 250	    mp->m_logdev_targp == mp->m_ddev_targp &&
 251	    !XFS_IS_REALTIME_INODE(ip) &&
 252	    !log_flushed)
 253		xfs_blkdev_issue_flush(mp->m_ddev_targp);
 254
 255	return -error;
 256}
 257
 258STATIC ssize_t
 259xfs_file_aio_read(
 260	struct kiocb		*iocb,
 261	const struct iovec	*iovp,
 262	unsigned long		nr_segs,
 263	loff_t			pos)
 264{
 265	struct file		*file = iocb->ki_filp;
 266	struct inode		*inode = file->f_mapping->host;
 267	struct xfs_inode	*ip = XFS_I(inode);
 268	struct xfs_mount	*mp = ip->i_mount;
 269	size_t			size = 0;
 270	ssize_t			ret = 0;
 271	int			ioflags = 0;
 272	xfs_fsize_t		n;
 273	unsigned long		seg;
 274
 275	XFS_STATS_INC(xs_read_calls);
 276
 277	BUG_ON(iocb->ki_pos != pos);
 278
 279	if (unlikely(file->f_flags & O_DIRECT))
 280		ioflags |= IO_ISDIRECT;
 281	if (file->f_mode & FMODE_NOCMTIME)
 282		ioflags |= IO_INVIS;
 283
 284	/* START copy & waste from filemap.c */
 285	for (seg = 0; seg < nr_segs; seg++) {
 286		const struct iovec *iv = &iovp[seg];
 287
 288		/*
 289		 * If any segment has a negative length, or the cumulative
 290		 * length ever wraps negative then return -EINVAL.
 291		 */
 292		size += iv->iov_len;
 293		if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 294			return XFS_ERROR(-EINVAL);
 295	}
 296	/* END copy & waste from filemap.c */
 297
 298	if (unlikely(ioflags & IO_ISDIRECT)) {
 299		xfs_buftarg_t	*target =
 300			XFS_IS_REALTIME_INODE(ip) ?
 301				mp->m_rtdev_targp : mp->m_ddev_targp;
 302		if ((iocb->ki_pos & target->bt_smask) ||
 303		    (size & target->bt_smask)) {
 304			if (iocb->ki_pos == ip->i_size)
 305				return 0;
 306			return -XFS_ERROR(EINVAL);
 307		}
 308	}
 309
 310	n = XFS_MAXIOFFSET(mp) - iocb->ki_pos;
 311	if (n <= 0 || size == 0)
 312		return 0;
 313
 314	if (n < size)
 315		size = n;
 316
 317	if (XFS_FORCED_SHUTDOWN(mp))
 318		return -EIO;
 319
 320	/*
 321	 * Locking is a bit tricky here. If we take an exclusive lock
 322	 * for direct IO, we effectively serialise all new concurrent
 323	 * read IO to this file and block it behind IO that is currently in
 324	 * progress because IO in progress holds the IO lock shared. We only
 325	 * need to hold the lock exclusive to blow away the page cache, so
 326	 * only take lock exclusively if the page cache needs invalidation.
 327	 * This allows the normal direct IO case of no page cache pages to
 328	 * proceeed concurrently without serialisation.
 329	 */
 330	xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
 331	if ((ioflags & IO_ISDIRECT) && inode->i_mapping->nrpages) {
 332		xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
 333		xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
 334
 335		if (inode->i_mapping->nrpages) {
 336			ret = -xfs_flushinval_pages(ip,
 337					(iocb->ki_pos & PAGE_CACHE_MASK),
 338					-1, FI_REMAPF_LOCKED);
 339			if (ret) {
 340				xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
 341				return ret;
 342			}
 343		}
 344		xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
 345	}
 346
 347	trace_xfs_file_read(ip, size, iocb->ki_pos, ioflags);
 348
 349	ret = generic_file_aio_read(iocb, iovp, nr_segs, iocb->ki_pos);
 350	if (ret > 0)
 351		XFS_STATS_ADD(xs_read_bytes, ret);
 352
 353	xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
 354	return ret;
 355}
 356
 357STATIC ssize_t
 358xfs_file_splice_read(
 359	struct file		*infilp,
 360	loff_t			*ppos,
 361	struct pipe_inode_info	*pipe,
 362	size_t			count,
 363	unsigned int		flags)
 364{
 365	struct xfs_inode	*ip = XFS_I(infilp->f_mapping->host);
 366	int			ioflags = 0;
 367	ssize_t			ret;
 368
 369	XFS_STATS_INC(xs_read_calls);
 370
 371	if (infilp->f_mode & FMODE_NOCMTIME)
 372		ioflags |= IO_INVIS;
 373
 374	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 375		return -EIO;
 376
 377	xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
 378
 379	trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
 380
 381	ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 382	if (ret > 0)
 383		XFS_STATS_ADD(xs_read_bytes, ret);
 384
 385	xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
 386	return ret;
 387}
 388
 389STATIC void
 390xfs_aio_write_isize_update(
 391	struct inode	*inode,
 392	loff_t		*ppos,
 393	ssize_t		bytes_written)
 394{
 395	struct xfs_inode	*ip = XFS_I(inode);
 396	xfs_fsize_t		isize = i_size_read(inode);
 397
 398	if (bytes_written > 0)
 399		XFS_STATS_ADD(xs_write_bytes, bytes_written);
 400
 401	if (unlikely(bytes_written < 0 && bytes_written != -EFAULT &&
 402					*ppos > isize))
 403		*ppos = isize;
 404
 405	if (*ppos > ip->i_size) {
 406		xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
 407		if (*ppos > ip->i_size)
 408			ip->i_size = *ppos;
 409		xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
 410	}
 411}
 412
 413/*
 414 * If this was a direct or synchronous I/O that failed (such as ENOSPC) then
 415 * part of the I/O may have been written to disk before the error occurred.  In
 416 * this case the on-disk file size may have been adjusted beyond the in-memory
 417 * file size and now needs to be truncated back.
 418 */
 419STATIC void
 420xfs_aio_write_newsize_update(
 421	struct xfs_inode	*ip)
 422{
 423	if (ip->i_new_size) {
 424		xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
 425		ip->i_new_size = 0;
 426		if (ip->i_d.di_size > ip->i_size)
 427			ip->i_d.di_size = ip->i_size;
 428		xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
 429	}
 430}
 431
 432/*
 433 * xfs_file_splice_write() does not use xfs_rw_ilock() because
 434 * generic_file_splice_write() takes the i_mutex itself. This, in theory,
 435 * couuld cause lock inversions between the aio_write path and the splice path
 436 * if someone is doing concurrent splice(2) based writes and write(2) based
 437 * writes to the same inode. The only real way to fix this is to re-implement
 438 * the generic code here with correct locking orders.
 439 */
 440STATIC ssize_t
 441xfs_file_splice_write(
 442	struct pipe_inode_info	*pipe,
 443	struct file		*outfilp,
 444	loff_t			*ppos,
 445	size_t			count,
 446	unsigned int		flags)
 447{
 448	struct inode		*inode = outfilp->f_mapping->host;
 449	struct xfs_inode	*ip = XFS_I(inode);
 450	xfs_fsize_t		new_size;
 451	int			ioflags = 0;
 452	ssize_t			ret;
 453
 454	XFS_STATS_INC(xs_write_calls);
 455
 456	if (outfilp->f_mode & FMODE_NOCMTIME)
 457		ioflags |= IO_INVIS;
 458
 459	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 460		return -EIO;
 461
 462	xfs_ilock(ip, XFS_IOLOCK_EXCL);
 463
 464	new_size = *ppos + count;
 465
 466	xfs_ilock(ip, XFS_ILOCK_EXCL);
 467	if (new_size > ip->i_size)
 468		ip->i_new_size = new_size;
 469	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 470
 471	trace_xfs_file_splice_write(ip, count, *ppos, ioflags);
 472
 473	ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 474
 475	xfs_aio_write_isize_update(inode, ppos, ret);
 476	xfs_aio_write_newsize_update(ip);
 477	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 478	return ret;
 479}
 480
 481/*
 482 * This routine is called to handle zeroing any space in the last
 483 * block of the file that is beyond the EOF.  We do this since the
 484 * size is being increased without writing anything to that block
 485 * and we don't want anyone to read the garbage on the disk.
 486 */
 487STATIC int				/* error (positive) */
 488xfs_zero_last_block(
 489	xfs_inode_t	*ip,
 490	xfs_fsize_t	offset,
 491	xfs_fsize_t	isize)
 492{
 493	xfs_fileoff_t	last_fsb;
 494	xfs_mount_t	*mp = ip->i_mount;
 495	int		nimaps;
 496	int		zero_offset;
 497	int		zero_len;
 498	int		error = 0;
 499	xfs_bmbt_irec_t	imap;
 500
 501	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 502
 503	zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 504	if (zero_offset == 0) {
 505		/*
 506		 * There are no extra bytes in the last block on disk to
 507		 * zero, so return.
 508		 */
 509		return 0;
 510	}
 511
 512	last_fsb = XFS_B_TO_FSBT(mp, isize);
 513	nimaps = 1;
 514	error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
 515			  &nimaps, NULL);
 516	if (error) {
 517		return error;
 518	}
 519	ASSERT(nimaps > 0);
 520	/*
 521	 * If the block underlying isize is just a hole, then there
 522	 * is nothing to zero.
 523	 */
 524	if (imap.br_startblock == HOLESTARTBLOCK) {
 525		return 0;
 526	}
 527	/*
 528	 * Zero the part of the last block beyond the EOF, and write it
 529	 * out sync.  We need to drop the ilock while we do this so we
 530	 * don't deadlock when the buffer cache calls back to us.
 531	 */
 532	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 533
 534	zero_len = mp->m_sb.sb_blocksize - zero_offset;
 535	if (isize + zero_len > offset)
 536		zero_len = offset - isize;
 537	error = xfs_iozero(ip, isize, zero_len);
 538
 539	xfs_ilock(ip, XFS_ILOCK_EXCL);
 540	ASSERT(error >= 0);
 541	return error;
 542}
 543
 544/*
 545 * Zero any on disk space between the current EOF and the new,
 546 * larger EOF.  This handles the normal case of zeroing the remainder
 547 * of the last block in the file and the unusual case of zeroing blocks
 548 * out beyond the size of the file.  This second case only happens
 549 * with fixed size extents and when the system crashes before the inode
 550 * size was updated but after blocks were allocated.  If fill is set,
 551 * then any holes in the range are filled and zeroed.  If not, the holes
 552 * are left alone as holes.
 553 */
 554
 555int					/* error (positive) */
 556xfs_zero_eof(
 557	xfs_inode_t	*ip,
 558	xfs_off_t	offset,		/* starting I/O offset */
 559	xfs_fsize_t	isize)		/* current inode size */
 560{
 561	xfs_mount_t	*mp = ip->i_mount;
 562	xfs_fileoff_t	start_zero_fsb;
 563	xfs_fileoff_t	end_zero_fsb;
 564	xfs_fileoff_t	zero_count_fsb;
 565	xfs_fileoff_t	last_fsb;
 566	xfs_fileoff_t	zero_off;
 567	xfs_fsize_t	zero_len;
 568	int		nimaps;
 569	int		error = 0;
 570	xfs_bmbt_irec_t	imap;
 571
 572	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 573	ASSERT(offset > isize);
 574
 575	/*
 576	 * First handle zeroing the block on which isize resides.
 577	 * We only zero a part of that block so it is handled specially.
 578	 */
 579	error = xfs_zero_last_block(ip, offset, isize);
 580	if (error) {
 581		ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 582		return error;
 583	}
 584
 585	/*
 586	 * Calculate the range between the new size and the old
 587	 * where blocks needing to be zeroed may exist.  To get the
 588	 * block where the last byte in the file currently resides,
 589	 * we need to subtract one from the size and truncate back
 590	 * to a block boundary.  We subtract 1 in case the size is
 591	 * exactly on a block boundary.
 592	 */
 593	last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 594	start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 595	end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 596	ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 597	if (last_fsb == end_zero_fsb) {
 598		/*
 599		 * The size was only incremented on its last block.
 600		 * We took care of that above, so just return.
 601		 */
 602		return 0;
 603	}
 604
 605	ASSERT(start_zero_fsb <= end_zero_fsb);
 606	while (start_zero_fsb <= end_zero_fsb) {
 607		nimaps = 1;
 608		zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 609		error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
 610				  0, NULL, 0, &imap, &nimaps, NULL);
 611		if (error) {
 612			ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 613			return error;
 614		}
 615		ASSERT(nimaps > 0);
 616
 617		if (imap.br_state == XFS_EXT_UNWRITTEN ||
 618		    imap.br_startblock == HOLESTARTBLOCK) {
 619			/*
 620			 * This loop handles initializing pages that were
 621			 * partially initialized by the code below this
 622			 * loop. It basically zeroes the part of the page
 623			 * that sits on a hole and sets the page as P_HOLE
 624			 * and calls remapf if it is a mapped file.
 625			 */
 626			start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 627			ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 628			continue;
 629		}
 630
 631		/*
 632		 * There are blocks we need to zero.
 633		 * Drop the inode lock while we're doing the I/O.
 634		 * We'll still have the iolock to protect us.
 635		 */
 636		xfs_iunlock(ip, XFS_ILOCK_EXCL);
 637
 638		zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 639		zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 640
 641		if ((zero_off + zero_len) > offset)
 642			zero_len = offset - zero_off;
 643
 644		error = xfs_iozero(ip, zero_off, zero_len);
 645		if (error) {
 646			goto out_lock;
 647		}
 648
 649		start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 650		ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 651
 652		xfs_ilock(ip, XFS_ILOCK_EXCL);
 653	}
 654
 655	return 0;
 656
 657out_lock:
 658	xfs_ilock(ip, XFS_ILOCK_EXCL);
 659	ASSERT(error >= 0);
 660	return error;
 661}
 662
 663/*
 664 * Common pre-write limit and setup checks.
 665 *
 666 * Returns with iolock held according to @iolock.
 667 */
 668STATIC ssize_t
 669xfs_file_aio_write_checks(
 670	struct file		*file,
 671	loff_t			*pos,
 672	size_t			*count,
 673	int			*iolock)
 674{
 675	struct inode		*inode = file->f_mapping->host;
 676	struct xfs_inode	*ip = XFS_I(inode);
 677	xfs_fsize_t		new_size;
 678	int			error = 0;
 679
 680	xfs_rw_ilock(ip, XFS_ILOCK_EXCL);
 681	error = generic_write_checks(file, pos, count, S_ISBLK(inode->i_mode));
 682	if (error) {
 683		xfs_rw_iunlock(ip, XFS_ILOCK_EXCL | *iolock);
 684		*iolock = 0;
 685		return error;
 686	}
 687
 688	new_size = *pos + *count;
 689	if (new_size > ip->i_size)
 690		ip->i_new_size = new_size;
 691
 692	if (likely(!(file->f_mode & FMODE_NOCMTIME)))
 693		file_update_time(file);
 694
 695	/*
 696	 * If the offset is beyond the size of the file, we need to zero any
 697	 * blocks that fall between the existing EOF and the start of this
 698	 * write.
 699	 */
 700	if (*pos > ip->i_size)
 701		error = -xfs_zero_eof(ip, *pos, ip->i_size);
 702
 703	xfs_rw_iunlock(ip, XFS_ILOCK_EXCL);
 704	if (error)
 705		return error;
 706
 707	/*
 708	 * If we're writing the file then make sure to clear the setuid and
 709	 * setgid bits if the process is not being run by root.  This keeps
 710	 * people from modifying setuid and setgid binaries.
 711	 */
 712	return file_remove_suid(file);
 713
 714}
 715
 716/*
 717 * xfs_file_dio_aio_write - handle direct IO writes
 718 *
 719 * Lock the inode appropriately to prepare for and issue a direct IO write.
 720 * By separating it from the buffered write path we remove all the tricky to
 721 * follow locking changes and looping.
 722 *
 723 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
 724 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
 725 * pages are flushed out.
 726 *
 727 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
 728 * allowing them to be done in parallel with reads and other direct IO writes.
 729 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
 730 * needs to do sub-block zeroing and that requires serialisation against other
 731 * direct IOs to the same block. In this case we need to serialise the
 732 * submission of the unaligned IOs so that we don't get racing block zeroing in
 733 * the dio layer.  To avoid the problem with aio, we also need to wait for
 734 * outstanding IOs to complete so that unwritten extent conversion is completed
 735 * before we try to map the overlapping block. This is currently implemented by
 736 * hitting it with a big hammer (i.e. xfs_ioend_wait()).
 737 *
 738 * Returns with locks held indicated by @iolock and errors indicated by
 739 * negative return values.
 740 */
 741STATIC ssize_t
 742xfs_file_dio_aio_write(
 743	struct kiocb		*iocb,
 744	const struct iovec	*iovp,
 745	unsigned long		nr_segs,
 746	loff_t			pos,
 747	size_t			ocount,
 748	int			*iolock)
 749{
 750	struct file		*file = iocb->ki_filp;
 751	struct address_space	*mapping = file->f_mapping;
 752	struct inode		*inode = mapping->host;
 753	struct xfs_inode	*ip = XFS_I(inode);
 754	struct xfs_mount	*mp = ip->i_mount;
 755	ssize_t			ret = 0;
 756	size_t			count = ocount;
 757	int			unaligned_io = 0;
 758	struct xfs_buftarg	*target = XFS_IS_REALTIME_INODE(ip) ?
 759					mp->m_rtdev_targp : mp->m_ddev_targp;
 760
 761	*iolock = 0;
 762	if ((pos & target->bt_smask) || (count & target->bt_smask))
 763		return -XFS_ERROR(EINVAL);
 764
 765	if ((pos & mp->m_blockmask) || ((pos + count) & mp->m_blockmask))
 766		unaligned_io = 1;
 767
 768	if (unaligned_io || mapping->nrpages || pos > ip->i_size)
 769		*iolock = XFS_IOLOCK_EXCL;
 770	else
 771		*iolock = XFS_IOLOCK_SHARED;
 772	xfs_rw_ilock(ip, *iolock);
 773
 774	ret = xfs_file_aio_write_checks(file, &pos, &count, iolock);
 775	if (ret)
 776		return ret;
 777
 778	/*
 779	 * Recheck if there are cached pages that need invalidate after we got
 780	 * the iolock to protect against other threads adding new pages while
 781	 * we were waiting for the iolock.
 782	 */
 783	if (mapping->nrpages && *iolock == XFS_IOLOCK_SHARED) {
 784		xfs_rw_iunlock(ip, *iolock);
 785		*iolock = XFS_IOLOCK_EXCL;
 786		xfs_rw_ilock(ip, *iolock);
 787	}
 788
 789	if (mapping->nrpages) {
 790		ret = -xfs_flushinval_pages(ip, (pos & PAGE_CACHE_MASK), -1,
 791							FI_REMAPF_LOCKED);
 792		if (ret)
 793			return ret;
 794	}
 795
 796	/*
 797	 * If we are doing unaligned IO, wait for all other IO to drain,
 798	 * otherwise demote the lock if we had to flush cached pages
 799	 */
 800	if (unaligned_io)
 801		xfs_ioend_wait(ip);
 802	else if (*iolock == XFS_IOLOCK_EXCL) {
 803		xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
 804		*iolock = XFS_IOLOCK_SHARED;
 805	}
 806
 807	trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
 808	ret = generic_file_direct_write(iocb, iovp,
 809			&nr_segs, pos, &iocb->ki_pos, count, ocount);
 810
 811	/* No fallback to buffered IO on errors for XFS. */
 812	ASSERT(ret < 0 || ret == count);
 813	return ret;
 814}
 815
 816STATIC ssize_t
 817xfs_file_buffered_aio_write(
 818	struct kiocb		*iocb,
 819	const struct iovec	*iovp,
 820	unsigned long		nr_segs,
 821	loff_t			pos,
 822	size_t			ocount,
 823	int			*iolock)
 824{
 825	struct file		*file = iocb->ki_filp;
 826	struct address_space	*mapping = file->f_mapping;
 827	struct inode		*inode = mapping->host;
 828	struct xfs_inode	*ip = XFS_I(inode);
 829	ssize_t			ret;
 830	int			enospc = 0;
 831	size_t			count = ocount;
 832
 833	*iolock = XFS_IOLOCK_EXCL;
 834	xfs_rw_ilock(ip, *iolock);
 835
 836	ret = xfs_file_aio_write_checks(file, &pos, &count, iolock);
 837	if (ret)
 838		return ret;
 839
 840	/* We can write back this queue in page reclaim */
 841	current->backing_dev_info = mapping->backing_dev_info;
 842
 843write_retry:
 844	trace_xfs_file_buffered_write(ip, count, iocb->ki_pos, 0);
 845	ret = generic_file_buffered_write(iocb, iovp, nr_segs,
 846			pos, &iocb->ki_pos, count, ret);
 847	/*
 848	 * if we just got an ENOSPC, flush the inode now we aren't holding any
 849	 * page locks and retry *once*
 850	 */
 851	if (ret == -ENOSPC && !enospc) {
 852		ret = -xfs_flush_pages(ip, 0, -1, 0, FI_NONE);
 853		if (ret)
 854			return ret;
 855		enospc = 1;
 856		goto write_retry;
 857	}
 858	current->backing_dev_info = NULL;
 859	return ret;
 860}
 861
 862STATIC ssize_t
 863xfs_file_aio_write(
 864	struct kiocb		*iocb,
 865	const struct iovec	*iovp,
 866	unsigned long		nr_segs,
 867	loff_t			pos)
 868{
 869	struct file		*file = iocb->ki_filp;
 870	struct address_space	*mapping = file->f_mapping;
 871	struct inode		*inode = mapping->host;
 872	struct xfs_inode	*ip = XFS_I(inode);
 873	ssize_t			ret;
 874	int			iolock;
 875	size_t			ocount = 0;
 876
 877	XFS_STATS_INC(xs_write_calls);
 878
 879	BUG_ON(iocb->ki_pos != pos);
 880
 881	ret = generic_segment_checks(iovp, &nr_segs, &ocount, VERIFY_READ);
 882	if (ret)
 883		return ret;
 884
 885	if (ocount == 0)
 886		return 0;
 887
 888	xfs_wait_for_freeze(ip->i_mount, SB_FREEZE_WRITE);
 889
 890	if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 891		return -EIO;
 892
 893	if (unlikely(file->f_flags & O_DIRECT))
 894		ret = xfs_file_dio_aio_write(iocb, iovp, nr_segs, pos,
 895						ocount, &iolock);
 896	else
 897		ret = xfs_file_buffered_aio_write(iocb, iovp, nr_segs, pos,
 898						ocount, &iolock);
 899
 900	xfs_aio_write_isize_update(inode, &iocb->ki_pos, ret);
 901
 902	if (ret <= 0)
 903		goto out_unlock;
 904
 905	/* Handle various SYNC-type writes */
 906	if ((file->f_flags & O_DSYNC) || IS_SYNC(inode)) {
 907		loff_t end = pos + ret - 1;
 908		int error;
 909
 910		xfs_rw_iunlock(ip, iolock);
 911		error = xfs_file_fsync(file, pos, end,
 912				      (file->f_flags & __O_SYNC) ? 0 : 1);
 913		xfs_rw_ilock(ip, iolock);
 914		if (error)
 915			ret = error;
 916	}
 917
 918out_unlock:
 919	xfs_aio_write_newsize_update(ip);
 920	xfs_rw_iunlock(ip, iolock);
 921	return ret;
 922}
 923
 924STATIC long
 925xfs_file_fallocate(
 926	struct file	*file,
 927	int		mode,
 928	loff_t		offset,
 929	loff_t		len)
 930{
 931	struct inode	*inode = file->f_path.dentry->d_inode;
 932	long		error;
 933	loff_t		new_size = 0;
 934	xfs_flock64_t	bf;
 935	xfs_inode_t	*ip = XFS_I(inode);
 936	int		cmd = XFS_IOC_RESVSP;
 937	int		attr_flags = XFS_ATTR_NOLOCK;
 938
 939	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 940		return -EOPNOTSUPP;
 941
 942	bf.l_whence = 0;
 943	bf.l_start = offset;
 944	bf.l_len = len;
 945
 946	xfs_ilock(ip, XFS_IOLOCK_EXCL);
 947
 948	if (mode & FALLOC_FL_PUNCH_HOLE)
 949		cmd = XFS_IOC_UNRESVSP;
 950
 951	/* check the new inode size is valid before allocating */
 952	if (!(mode & FALLOC_FL_KEEP_SIZE) &&
 953	    offset + len > i_size_read(inode)) {
 954		new_size = offset + len;
 955		error = inode_newsize_ok(inode, new_size);
 956		if (error)
 957			goto out_unlock;
 958	}
 959
 960	if (file->f_flags & O_DSYNC)
 961		attr_flags |= XFS_ATTR_SYNC;
 962
 963	error = -xfs_change_file_space(ip, cmd, &bf, 0, attr_flags);
 964	if (error)
 965		goto out_unlock;
 966
 967	/* Change file size if needed */
 968	if (new_size) {
 969		struct iattr iattr;
 970
 971		iattr.ia_valid = ATTR_SIZE;
 972		iattr.ia_size = new_size;
 973		error = -xfs_setattr_size(ip, &iattr, XFS_ATTR_NOLOCK);
 974	}
 975
 976out_unlock:
 977	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 978	return error;
 979}
 980
 981
 982STATIC int
 983xfs_file_open(
 984	struct inode	*inode,
 985	struct file	*file)
 986{
 987	if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
 988		return -EFBIG;
 989	if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
 990		return -EIO;
 991	return 0;
 992}
 993
 994STATIC int
 995xfs_dir_open(
 996	struct inode	*inode,
 997	struct file	*file)
 998{
 999	struct xfs_inode *ip = XFS_I(inode);
1000	int		mode;
1001	int		error;
1002
1003	error = xfs_file_open(inode, file);
1004	if (error)
1005		return error;
1006
1007	/*
1008	 * If there are any blocks, read-ahead block 0 as we're almost
1009	 * certain to have the next operation be a read there.
1010	 */
1011	mode = xfs_ilock_map_shared(ip);
1012	if (ip->i_d.di_nextents > 0)
1013		xfs_da_reada_buf(NULL, ip, 0, XFS_DATA_FORK);
1014	xfs_iunlock(ip, mode);
1015	return 0;
1016}
1017
1018STATIC int
1019xfs_file_release(
1020	struct inode	*inode,
1021	struct file	*filp)
1022{
1023	return -xfs_release(XFS_I(inode));
1024}
1025
1026STATIC int
1027xfs_file_readdir(
1028	struct file	*filp,
1029	void		*dirent,
1030	filldir_t	filldir)
1031{
1032	struct inode	*inode = filp->f_path.dentry->d_inode;
1033	xfs_inode_t	*ip = XFS_I(inode);
1034	int		error;
1035	size_t		bufsize;
1036
1037	/*
1038	 * The Linux API doesn't pass down the total size of the buffer
1039	 * we read into down to the filesystem.  With the filldir concept
1040	 * it's not needed for correct information, but the XFS dir2 leaf
1041	 * code wants an estimate of the buffer size to calculate it's
1042	 * readahead window and size the buffers used for mapping to
1043	 * physical blocks.
1044	 *
1045	 * Try to give it an estimate that's good enough, maybe at some
1046	 * point we can change the ->readdir prototype to include the
1047	 * buffer size.  For now we use the current glibc buffer size.
1048	 */
1049	bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
1050
1051	error = xfs_readdir(ip, dirent, bufsize,
1052				(xfs_off_t *)&filp->f_pos, filldir);
1053	if (error)
1054		return -error;
1055	return 0;
1056}
1057
1058STATIC int
1059xfs_file_mmap(
1060	struct file	*filp,
1061	struct vm_area_struct *vma)
1062{
1063	vma->vm_ops = &xfs_file_vm_ops;
1064	vma->vm_flags |= VM_CAN_NONLINEAR;
1065
1066	file_accessed(filp);
1067	return 0;
1068}
1069
1070/*
1071 * mmap()d file has taken write protection fault and is being made
1072 * writable. We can set the page state up correctly for a writable
1073 * page, which means we can do correct delalloc accounting (ENOSPC
1074 * checking!) and unwritten extent mapping.
1075 */
1076STATIC int
1077xfs_vm_page_mkwrite(
1078	struct vm_area_struct	*vma,
1079	struct vm_fault		*vmf)
1080{
1081	return block_page_mkwrite(vma, vmf, xfs_get_blocks);
1082}
1083
1084const struct file_operations xfs_file_operations = {
1085	.llseek		= generic_file_llseek,
1086	.read		= do_sync_read,
1087	.write		= do_sync_write,
1088	.aio_read	= xfs_file_aio_read,
1089	.aio_write	= xfs_file_aio_write,
1090	.splice_read	= xfs_file_splice_read,
1091	.splice_write	= xfs_file_splice_write,
1092	.unlocked_ioctl	= xfs_file_ioctl,
1093#ifdef CONFIG_COMPAT
1094	.compat_ioctl	= xfs_file_compat_ioctl,
1095#endif
1096	.mmap		= xfs_file_mmap,
1097	.open		= xfs_file_open,
1098	.release	= xfs_file_release,
1099	.fsync		= xfs_file_fsync,
1100	.fallocate	= xfs_file_fallocate,
1101};
1102
1103const struct file_operations xfs_dir_file_operations = {
1104	.open		= xfs_dir_open,
1105	.read		= generic_read_dir,
1106	.readdir	= xfs_file_readdir,
1107	.llseek		= generic_file_llseek,
1108	.unlocked_ioctl	= xfs_file_ioctl,
1109#ifdef CONFIG_COMPAT
1110	.compat_ioctl	= xfs_file_compat_ioctl,
1111#endif
1112	.fsync		= xfs_file_fsync,
1113};
1114
1115static const struct vm_operations_struct xfs_file_vm_ops = {
1116	.fault		= filemap_fault,
1117	.page_mkwrite	= xfs_vm_page_mkwrite,
1118};