Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_mount.h"
  13#include "xfs_inode.h"
  14#include "xfs_trans.h"
  15#include "xfs_inode_item.h"
  16#include "xfs_bmap.h"
  17#include "xfs_bmap_util.h"
  18#include "xfs_dir2.h"
  19#include "xfs_dir2_priv.h"
  20#include "xfs_ioctl.h"
  21#include "xfs_trace.h"
  22#include "xfs_log.h"
  23#include "xfs_icache.h"
  24#include "xfs_pnfs.h"
  25#include "xfs_iomap.h"
  26#include "xfs_reflink.h"
 
  27
  28#include <linux/dax.h>
  29#include <linux/falloc.h>
  30#include <linux/backing-dev.h>
  31#include <linux/mman.h>
  32#include <linux/fadvise.h>
  33#include <linux/mount.h>
  34
  35static const struct vm_operations_struct xfs_file_vm_ops;
  36
  37/*
  38 * Decide if the given file range is aligned to the size of the fundamental
  39 * allocation unit for the file.
  40 */
  41static bool
  42xfs_is_falloc_aligned(
  43	struct xfs_inode	*ip,
  44	loff_t			pos,
  45	long long int		len)
  46{
  47	struct xfs_mount	*mp = ip->i_mount;
  48	uint64_t		mask;
  49
  50	if (XFS_IS_REALTIME_INODE(ip)) {
  51		if (!is_power_of_2(mp->m_sb.sb_rextsize)) {
  52			u64	rextbytes;
  53			u32	mod;
  54
  55			rextbytes = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize);
  56			div_u64_rem(pos, rextbytes, &mod);
  57			if (mod)
  58				return false;
  59			div_u64_rem(len, rextbytes, &mod);
  60			return mod == 0;
  61		}
  62		mask = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize) - 1;
  63	} else {
  64		mask = mp->m_sb.sb_blocksize - 1;
  65	}
  66
  67	return !((pos | len) & mask);
  68}
  69
  70/*
  71 * Fsync operations on directories are much simpler than on regular files,
  72 * as there is no file data to flush, and thus also no need for explicit
  73 * cache flush operations, and there are no non-transaction metadata updates
  74 * on directories either.
  75 */
  76STATIC int
  77xfs_dir_fsync(
  78	struct file		*file,
  79	loff_t			start,
  80	loff_t			end,
  81	int			datasync)
  82{
  83	struct xfs_inode	*ip = XFS_I(file->f_mapping->host);
  84
  85	trace_xfs_dir_fsync(ip);
  86	return xfs_log_force_inode(ip);
  87}
  88
  89static xfs_csn_t
  90xfs_fsync_seq(
  91	struct xfs_inode	*ip,
  92	bool			datasync)
  93{
  94	if (!xfs_ipincount(ip))
  95		return 0;
  96	if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
  97		return 0;
  98	return ip->i_itemp->ili_commit_seq;
  99}
 100
 101/*
 102 * All metadata updates are logged, which means that we just have to flush the
 103 * log up to the latest LSN that touched the inode.
 104 *
 105 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
 106 * the log force before we clear the ili_fsync_fields field. This ensures that
 107 * we don't get a racing sync operation that does not wait for the metadata to
 108 * hit the journal before returning.  If we race with clearing ili_fsync_fields,
 109 * then all that will happen is the log force will do nothing as the lsn will
 110 * already be on disk.  We can't race with setting ili_fsync_fields because that
 111 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
 112 * shared until after the ili_fsync_fields is cleared.
 113 */
 114static  int
 115xfs_fsync_flush_log(
 116	struct xfs_inode	*ip,
 117	bool			datasync,
 118	int			*log_flushed)
 119{
 120	int			error = 0;
 121	xfs_csn_t		seq;
 122
 123	xfs_ilock(ip, XFS_ILOCK_SHARED);
 124	seq = xfs_fsync_seq(ip, datasync);
 125	if (seq) {
 126		error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
 127					  log_flushed);
 128
 129		spin_lock(&ip->i_itemp->ili_lock);
 130		ip->i_itemp->ili_fsync_fields = 0;
 131		spin_unlock(&ip->i_itemp->ili_lock);
 132	}
 133	xfs_iunlock(ip, XFS_ILOCK_SHARED);
 134	return error;
 135}
 136
 137STATIC int
 138xfs_file_fsync(
 139	struct file		*file,
 140	loff_t			start,
 141	loff_t			end,
 142	int			datasync)
 143{
 144	struct xfs_inode	*ip = XFS_I(file->f_mapping->host);
 145	struct xfs_mount	*mp = ip->i_mount;
 146	int			error, err2;
 147	int			log_flushed = 0;
 148
 149	trace_xfs_file_fsync(ip);
 150
 151	error = file_write_and_wait_range(file, start, end);
 152	if (error)
 153		return error;
 154
 155	if (xfs_is_shutdown(mp))
 156		return -EIO;
 157
 158	xfs_iflags_clear(ip, XFS_ITRUNCATED);
 159
 160	/*
 161	 * If we have an RT and/or log subvolume we need to make sure to flush
 162	 * the write cache the device used for file data first.  This is to
 163	 * ensure newly written file data make it to disk before logging the new
 164	 * inode size in case of an extending write.
 165	 */
 166	if (XFS_IS_REALTIME_INODE(ip))
 167		error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
 168	else if (mp->m_logdev_targp != mp->m_ddev_targp)
 169		error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
 170
 171	/*
 172	 * Any inode that has dirty modifications in the log is pinned.  The
 173	 * racy check here for a pinned inode will not catch modifications
 174	 * that happen concurrently to the fsync call, but fsync semantics
 175	 * only require to sync previously completed I/O.
 176	 */
 177	if (xfs_ipincount(ip)) {
 178		err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
 179		if (err2 && !error)
 180			error = err2;
 181	}
 182
 183	/*
 184	 * If we only have a single device, and the log force about was
 185	 * a no-op we might have to flush the data device cache here.
 186	 * This can only happen for fdatasync/O_DSYNC if we were overwriting
 187	 * an already allocated file and thus do not have any metadata to
 188	 * commit.
 189	 */
 190	if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
 191	    mp->m_logdev_targp == mp->m_ddev_targp) {
 192		err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
 193		if (err2 && !error)
 194			error = err2;
 195	}
 196
 197	return error;
 198}
 199
 200static int
 201xfs_ilock_iocb(
 202	struct kiocb		*iocb,
 203	unsigned int		lock_mode)
 204{
 205	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 206
 207	if (iocb->ki_flags & IOCB_NOWAIT) {
 208		if (!xfs_ilock_nowait(ip, lock_mode))
 209			return -EAGAIN;
 210	} else {
 211		xfs_ilock(ip, lock_mode);
 212	}
 213
 214	return 0;
 215}
 216
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 217STATIC ssize_t
 218xfs_file_dio_read(
 219	struct kiocb		*iocb,
 220	struct iov_iter		*to)
 221{
 222	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 223	ssize_t			ret;
 224
 225	trace_xfs_file_direct_read(iocb, to);
 226
 227	if (!iov_iter_count(to))
 228		return 0; /* skip atime */
 229
 230	file_accessed(iocb->ki_filp);
 231
 232	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 233	if (ret)
 234		return ret;
 235	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
 236	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 237
 238	return ret;
 239}
 240
 241static noinline ssize_t
 242xfs_file_dax_read(
 243	struct kiocb		*iocb,
 244	struct iov_iter		*to)
 245{
 246	struct xfs_inode	*ip = XFS_I(iocb->ki_filp->f_mapping->host);
 247	ssize_t			ret = 0;
 248
 249	trace_xfs_file_dax_read(iocb, to);
 250
 251	if (!iov_iter_count(to))
 252		return 0; /* skip atime */
 253
 254	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 255	if (ret)
 256		return ret;
 257	ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
 258	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 259
 260	file_accessed(iocb->ki_filp);
 261	return ret;
 262}
 263
 264STATIC ssize_t
 265xfs_file_buffered_read(
 266	struct kiocb		*iocb,
 267	struct iov_iter		*to)
 268{
 269	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 270	ssize_t			ret;
 271
 272	trace_xfs_file_buffered_read(iocb, to);
 273
 274	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 275	if (ret)
 276		return ret;
 277	ret = generic_file_read_iter(iocb, to);
 278	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 279
 280	return ret;
 281}
 282
 283STATIC ssize_t
 284xfs_file_read_iter(
 285	struct kiocb		*iocb,
 286	struct iov_iter		*to)
 287{
 288	struct inode		*inode = file_inode(iocb->ki_filp);
 289	struct xfs_mount	*mp = XFS_I(inode)->i_mount;
 290	ssize_t			ret = 0;
 291
 292	XFS_STATS_INC(mp, xs_read_calls);
 293
 294	if (xfs_is_shutdown(mp))
 295		return -EIO;
 296
 297	if (IS_DAX(inode))
 298		ret = xfs_file_dax_read(iocb, to);
 299	else if (iocb->ki_flags & IOCB_DIRECT)
 300		ret = xfs_file_dio_read(iocb, to);
 301	else
 302		ret = xfs_file_buffered_read(iocb, to);
 303
 304	if (ret > 0)
 305		XFS_STATS_ADD(mp, xs_read_bytes, ret);
 306	return ret;
 307}
 308
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 309/*
 310 * Common pre-write limit and setup checks.
 311 *
 312 * Called with the iolocked held either shared and exclusive according to
 313 * @iolock, and returns with it held.  Might upgrade the iolock to exclusive
 314 * if called for a direct write beyond i_size.
 315 */
 316STATIC ssize_t
 317xfs_file_write_checks(
 318	struct kiocb		*iocb,
 319	struct iov_iter		*from,
 320	unsigned int		*iolock)
 321{
 322	struct file		*file = iocb->ki_filp;
 323	struct inode		*inode = file->f_mapping->host;
 324	struct xfs_inode	*ip = XFS_I(inode);
 325	ssize_t			error = 0;
 326	size_t			count = iov_iter_count(from);
 327	bool			drained_dio = false;
 328	loff_t			isize;
 329
 330restart:
 331	error = generic_write_checks(iocb, from);
 332	if (error <= 0)
 333		return error;
 334
 335	if (iocb->ki_flags & IOCB_NOWAIT) {
 336		error = break_layout(inode, false);
 337		if (error == -EWOULDBLOCK)
 338			error = -EAGAIN;
 339	} else {
 340		error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
 341	}
 342
 343	if (error)
 344		return error;
 345
 346	/*
 347	 * For changing security info in file_remove_privs() we need i_rwsem
 348	 * exclusively.
 349	 */
 350	if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
 351		xfs_iunlock(ip, *iolock);
 352		*iolock = XFS_IOLOCK_EXCL;
 353		error = xfs_ilock_iocb(iocb, *iolock);
 354		if (error) {
 355			*iolock = 0;
 356			return error;
 357		}
 358		goto restart;
 359	}
 360
 361	/*
 362	 * If the offset is beyond the size of the file, we need to zero any
 363	 * blocks that fall between the existing EOF and the start of this
 364	 * write.  If zeroing is needed and we are currently holding the iolock
 365	 * shared, we need to update it to exclusive which implies having to
 366	 * redo all checks before.
 367	 *
 368	 * We need to serialise against EOF updates that occur in IO completions
 369	 * here. We want to make sure that nobody is changing the size while we
 370	 * do this check until we have placed an IO barrier (i.e.  hold the
 371	 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.  The
 372	 * spinlock effectively forms a memory barrier once we have the
 373	 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
 374	 * hence be able to correctly determine if we need to run zeroing.
 375	 *
 376	 * We can do an unlocked check here safely as IO completion can only
 377	 * extend EOF. Truncate is locked out at this point, so the EOF can
 378	 * not move backwards, only forwards. Hence we only need to take the
 379	 * slow path and spin locks when we are at or beyond the current EOF.
 380	 */
 381	if (iocb->ki_pos <= i_size_read(inode))
 382		goto out;
 383
 384	spin_lock(&ip->i_flags_lock);
 385	isize = i_size_read(inode);
 386	if (iocb->ki_pos > isize) {
 387		spin_unlock(&ip->i_flags_lock);
 388
 389		if (iocb->ki_flags & IOCB_NOWAIT)
 390			return -EAGAIN;
 391
 392		if (!drained_dio) {
 393			if (*iolock == XFS_IOLOCK_SHARED) {
 394				xfs_iunlock(ip, *iolock);
 395				*iolock = XFS_IOLOCK_EXCL;
 396				xfs_ilock(ip, *iolock);
 397				iov_iter_reexpand(from, count);
 398			}
 399			/*
 400			 * We now have an IO submission barrier in place, but
 401			 * AIO can do EOF updates during IO completion and hence
 402			 * we now need to wait for all of them to drain. Non-AIO
 403			 * DIO will have drained before we are given the
 404			 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
 405			 * no-op.
 406			 */
 407			inode_dio_wait(inode);
 408			drained_dio = true;
 409			goto restart;
 410		}
 411
 412		trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
 413		error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
 414		if (error)
 415			return error;
 416	} else
 417		spin_unlock(&ip->i_flags_lock);
 418
 419out:
 420	return kiocb_modified(iocb);
 421}
 422
 423static int
 424xfs_dio_write_end_io(
 425	struct kiocb		*iocb,
 426	ssize_t			size,
 427	int			error,
 428	unsigned		flags)
 429{
 430	struct inode		*inode = file_inode(iocb->ki_filp);
 431	struct xfs_inode	*ip = XFS_I(inode);
 432	loff_t			offset = iocb->ki_pos;
 433	unsigned int		nofs_flag;
 434
 435	trace_xfs_end_io_direct_write(ip, offset, size);
 436
 437	if (xfs_is_shutdown(ip->i_mount))
 438		return -EIO;
 439
 440	if (error)
 441		return error;
 442	if (!size)
 443		return 0;
 444
 445	/*
 446	 * Capture amount written on completion as we can't reliably account
 447	 * for it on submission.
 448	 */
 449	XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
 450
 451	/*
 452	 * We can allocate memory here while doing writeback on behalf of
 453	 * memory reclaim.  To avoid memory allocation deadlocks set the
 454	 * task-wide nofs context for the following operations.
 455	 */
 456	nofs_flag = memalloc_nofs_save();
 457
 458	if (flags & IOMAP_DIO_COW) {
 459		error = xfs_reflink_end_cow(ip, offset, size);
 460		if (error)
 461			goto out;
 462	}
 463
 464	/*
 465	 * Unwritten conversion updates the in-core isize after extent
 466	 * conversion but before updating the on-disk size. Updating isize any
 467	 * earlier allows a racing dio read to find unwritten extents before
 468	 * they are converted.
 469	 */
 470	if (flags & IOMAP_DIO_UNWRITTEN) {
 471		error = xfs_iomap_write_unwritten(ip, offset, size, true);
 472		goto out;
 473	}
 474
 475	/*
 476	 * We need to update the in-core inode size here so that we don't end up
 477	 * with the on-disk inode size being outside the in-core inode size. We
 478	 * have no other method of updating EOF for AIO, so always do it here
 479	 * if necessary.
 480	 *
 481	 * We need to lock the test/set EOF update as we can be racing with
 482	 * other IO completions here to update the EOF. Failing to serialise
 483	 * here can result in EOF moving backwards and Bad Things Happen when
 484	 * that occurs.
 485	 *
 486	 * As IO completion only ever extends EOF, we can do an unlocked check
 487	 * here to avoid taking the spinlock. If we land within the current EOF,
 488	 * then we do not need to do an extending update at all, and we don't
 489	 * need to take the lock to check this. If we race with an update moving
 490	 * EOF, then we'll either still be beyond EOF and need to take the lock,
 491	 * or we'll be within EOF and we don't need to take it at all.
 492	 */
 493	if (offset + size <= i_size_read(inode))
 494		goto out;
 495
 496	spin_lock(&ip->i_flags_lock);
 497	if (offset + size > i_size_read(inode)) {
 498		i_size_write(inode, offset + size);
 499		spin_unlock(&ip->i_flags_lock);
 500		error = xfs_setfilesize(ip, offset, size);
 501	} else {
 502		spin_unlock(&ip->i_flags_lock);
 503	}
 504
 505out:
 506	memalloc_nofs_restore(nofs_flag);
 507	return error;
 508}
 509
 510static const struct iomap_dio_ops xfs_dio_write_ops = {
 511	.end_io		= xfs_dio_write_end_io,
 512};
 513
 514/*
 515 * Handle block aligned direct I/O writes
 516 */
 517static noinline ssize_t
 518xfs_file_dio_write_aligned(
 519	struct xfs_inode	*ip,
 520	struct kiocb		*iocb,
 521	struct iov_iter		*from)
 522{
 523	unsigned int		iolock = XFS_IOLOCK_SHARED;
 524	ssize_t			ret;
 525
 526	ret = xfs_ilock_iocb(iocb, iolock);
 527	if (ret)
 528		return ret;
 529	ret = xfs_file_write_checks(iocb, from, &iolock);
 530	if (ret)
 531		goto out_unlock;
 532
 533	/*
 534	 * We don't need to hold the IOLOCK exclusively across the IO, so demote
 535	 * the iolock back to shared if we had to take the exclusive lock in
 536	 * xfs_file_write_checks() for other reasons.
 537	 */
 538	if (iolock == XFS_IOLOCK_EXCL) {
 539		xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
 540		iolock = XFS_IOLOCK_SHARED;
 541	}
 542	trace_xfs_file_direct_write(iocb, from);
 543	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
 544			   &xfs_dio_write_ops, 0, NULL, 0);
 545out_unlock:
 546	if (iolock)
 547		xfs_iunlock(ip, iolock);
 548	return ret;
 549}
 550
 551/*
 552 * Handle block unaligned direct I/O writes
 553 *
 554 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
 555 * them to be done in parallel with reads and other direct I/O writes.  However,
 556 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
 557 * to do sub-block zeroing and that requires serialisation against other direct
 558 * I/O to the same block.  In this case we need to serialise the submission of
 559 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
 560 * In the case where sub-block zeroing is not required, we can do concurrent
 561 * sub-block dios to the same block successfully.
 562 *
 563 * Optimistically submit the I/O using the shared lock first, but use the
 564 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
 565 * if block allocation or partial block zeroing would be required.  In that case
 566 * we try again with the exclusive lock.
 567 */
 568static noinline ssize_t
 569xfs_file_dio_write_unaligned(
 570	struct xfs_inode	*ip,
 571	struct kiocb		*iocb,
 572	struct iov_iter		*from)
 573{
 574	size_t			isize = i_size_read(VFS_I(ip));
 575	size_t			count = iov_iter_count(from);
 576	unsigned int		iolock = XFS_IOLOCK_SHARED;
 577	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
 578	ssize_t			ret;
 579
 580	/*
 581	 * Extending writes need exclusivity because of the sub-block zeroing
 582	 * that the DIO code always does for partial tail blocks beyond EOF, so
 583	 * don't even bother trying the fast path in this case.
 584	 */
 585	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
 586		if (iocb->ki_flags & IOCB_NOWAIT)
 587			return -EAGAIN;
 588retry_exclusive:
 589		iolock = XFS_IOLOCK_EXCL;
 590		flags = IOMAP_DIO_FORCE_WAIT;
 591	}
 592
 593	ret = xfs_ilock_iocb(iocb, iolock);
 594	if (ret)
 595		return ret;
 596
 597	/*
 598	 * We can't properly handle unaligned direct I/O to reflink files yet,
 599	 * as we can't unshare a partial block.
 600	 */
 601	if (xfs_is_cow_inode(ip)) {
 602		trace_xfs_reflink_bounce_dio_write(iocb, from);
 603		ret = -ENOTBLK;
 604		goto out_unlock;
 605	}
 606
 607	ret = xfs_file_write_checks(iocb, from, &iolock);
 608	if (ret)
 609		goto out_unlock;
 610
 611	/*
 612	 * If we are doing exclusive unaligned I/O, this must be the only I/O
 613	 * in-flight.  Otherwise we risk data corruption due to unwritten extent
 614	 * conversions from the AIO end_io handler.  Wait for all other I/O to
 615	 * drain first.
 616	 */
 617	if (flags & IOMAP_DIO_FORCE_WAIT)
 618		inode_dio_wait(VFS_I(ip));
 619
 620	trace_xfs_file_direct_write(iocb, from);
 621	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
 622			   &xfs_dio_write_ops, flags, NULL, 0);
 623
 624	/*
 625	 * Retry unaligned I/O with exclusive blocking semantics if the DIO
 626	 * layer rejected it for mapping or locking reasons. If we are doing
 627	 * nonblocking user I/O, propagate the error.
 628	 */
 629	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
 630		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
 631		xfs_iunlock(ip, iolock);
 632		goto retry_exclusive;
 633	}
 634
 635out_unlock:
 636	if (iolock)
 637		xfs_iunlock(ip, iolock);
 638	return ret;
 639}
 640
 641static ssize_t
 642xfs_file_dio_write(
 643	struct kiocb		*iocb,
 644	struct iov_iter		*from)
 645{
 646	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 647	struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
 648	size_t			count = iov_iter_count(from);
 649
 650	/* direct I/O must be aligned to device logical sector size */
 651	if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
 652		return -EINVAL;
 653	if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
 654		return xfs_file_dio_write_unaligned(ip, iocb, from);
 655	return xfs_file_dio_write_aligned(ip, iocb, from);
 656}
 657
 658static noinline ssize_t
 659xfs_file_dax_write(
 660	struct kiocb		*iocb,
 661	struct iov_iter		*from)
 662{
 663	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 664	struct xfs_inode	*ip = XFS_I(inode);
 665	unsigned int		iolock = XFS_IOLOCK_EXCL;
 666	ssize_t			ret, error = 0;
 667	loff_t			pos;
 668
 669	ret = xfs_ilock_iocb(iocb, iolock);
 670	if (ret)
 671		return ret;
 672	ret = xfs_file_write_checks(iocb, from, &iolock);
 673	if (ret)
 674		goto out;
 675
 676	pos = iocb->ki_pos;
 677
 678	trace_xfs_file_dax_write(iocb, from);
 679	ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
 680	if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
 681		i_size_write(inode, iocb->ki_pos);
 682		error = xfs_setfilesize(ip, pos, ret);
 683	}
 684out:
 685	if (iolock)
 686		xfs_iunlock(ip, iolock);
 687	if (error)
 688		return error;
 689
 690	if (ret > 0) {
 691		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
 692
 693		/* Handle various SYNC-type writes */
 694		ret = generic_write_sync(iocb, ret);
 695	}
 696	return ret;
 697}
 698
 699STATIC ssize_t
 700xfs_file_buffered_write(
 701	struct kiocb		*iocb,
 702	struct iov_iter		*from)
 703{
 704	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 705	struct xfs_inode	*ip = XFS_I(inode);
 706	ssize_t			ret;
 707	bool			cleared_space = false;
 708	unsigned int		iolock;
 709
 710write_retry:
 711	iolock = XFS_IOLOCK_EXCL;
 712	ret = xfs_ilock_iocb(iocb, iolock);
 713	if (ret)
 714		return ret;
 715
 716	ret = xfs_file_write_checks(iocb, from, &iolock);
 717	if (ret)
 718		goto out;
 719
 720	/* We can write back this queue in page reclaim */
 721	current->backing_dev_info = inode_to_bdi(inode);
 722
 723	trace_xfs_file_buffered_write(iocb, from);
 724	ret = iomap_file_buffered_write(iocb, from,
 725			&xfs_buffered_write_iomap_ops);
 726	if (likely(ret >= 0))
 727		iocb->ki_pos += ret;
 728
 729	/*
 730	 * If we hit a space limit, try to free up some lingering preallocated
 731	 * space before returning an error. In the case of ENOSPC, first try to
 732	 * write back all dirty inodes to free up some of the excess reserved
 733	 * metadata space. This reduces the chances that the eofblocks scan
 734	 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
 735	 * also behaves as a filter to prevent too many eofblocks scans from
 736	 * running at the same time.  Use a synchronous scan to increase the
 737	 * effectiveness of the scan.
 738	 */
 739	if (ret == -EDQUOT && !cleared_space) {
 740		xfs_iunlock(ip, iolock);
 741		xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
 742		cleared_space = true;
 743		goto write_retry;
 744	} else if (ret == -ENOSPC && !cleared_space) {
 745		struct xfs_icwalk	icw = {0};
 746
 747		cleared_space = true;
 748		xfs_flush_inodes(ip->i_mount);
 749
 750		xfs_iunlock(ip, iolock);
 751		icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
 752		xfs_blockgc_free_space(ip->i_mount, &icw);
 753		goto write_retry;
 754	}
 755
 756	current->backing_dev_info = NULL;
 757out:
 758	if (iolock)
 759		xfs_iunlock(ip, iolock);
 760
 761	if (ret > 0) {
 762		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
 763		/* Handle various SYNC-type writes */
 764		ret = generic_write_sync(iocb, ret);
 765	}
 766	return ret;
 767}
 768
 769STATIC ssize_t
 770xfs_file_write_iter(
 771	struct kiocb		*iocb,
 772	struct iov_iter		*from)
 773{
 774	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 775	struct xfs_inode	*ip = XFS_I(inode);
 776	ssize_t			ret;
 777	size_t			ocount = iov_iter_count(from);
 778
 779	XFS_STATS_INC(ip->i_mount, xs_write_calls);
 780
 781	if (ocount == 0)
 782		return 0;
 783
 784	if (xfs_is_shutdown(ip->i_mount))
 785		return -EIO;
 786
 787	if (IS_DAX(inode))
 788		return xfs_file_dax_write(iocb, from);
 789
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 790	if (iocb->ki_flags & IOCB_DIRECT) {
 791		/*
 792		 * Allow a directio write to fall back to a buffered
 793		 * write *only* in the case that we're doing a reflink
 794		 * CoW.  In all other directio scenarios we do not
 795		 * allow an operation to fall back to buffered mode.
 796		 */
 797		ret = xfs_file_dio_write(iocb, from);
 798		if (ret != -ENOTBLK)
 799			return ret;
 800	}
 801
 802	return xfs_file_buffered_write(iocb, from);
 803}
 804
 805static void
 806xfs_wait_dax_page(
 807	struct inode		*inode)
 808{
 809	struct xfs_inode        *ip = XFS_I(inode);
 810
 811	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
 812	schedule();
 813	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
 
 
 
 
 
 814}
 815
 816int
 817xfs_break_dax_layouts(
 818	struct inode		*inode,
 819	bool			*retry)
 
 
 
 820{
 821	struct page		*page;
 822
 823	ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 824
 825	page = dax_layout_busy_page(inode->i_mapping);
 826	if (!page)
 827		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 828
 829	*retry = true;
 830	return ___wait_var_event(&page->_refcount,
 831			atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
 832			0, 0, xfs_wait_dax_page(inode));
 
 
 
 
 
 
 
 
 
 
 833}
 834
 835int
 836xfs_break_layouts(
 837	struct inode		*inode,
 838	uint			*iolock,
 839	enum layout_break_reason reason)
 840{
 841	bool			retry;
 
 842	int			error;
 843
 844	ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
 
 845
 846	do {
 847		retry = false;
 848		switch (reason) {
 849		case BREAK_UNMAP:
 850			error = xfs_break_dax_layouts(inode, &retry);
 851			if (error || retry)
 852				break;
 853			fallthrough;
 854		case BREAK_WRITE:
 855			error = xfs_break_leased_layouts(inode, iolock, &retry);
 856			break;
 857		default:
 858			WARN_ON_ONCE(1);
 859			error = -EINVAL;
 860		}
 861	} while (error == 0 && retry);
 862
 863	return error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 864}
 865
 866/* Does this file, inode, or mount want synchronous writes? */
 867static inline bool xfs_file_sync_writes(struct file *filp)
 
 
 
 
 
 
 
 
 
 
 
 
 868{
 869	struct xfs_inode	*ip = XFS_I(file_inode(filp));
 
 
 
 870
 871	if (xfs_has_wsync(ip->i_mount))
 872		return true;
 873	if (filp->f_flags & (__O_SYNC | O_DSYNC))
 874		return true;
 875	if (IS_SYNC(file_inode(filp)))
 876		return true;
 877
 878	return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 879}
 880
 881#define	XFS_FALLOC_FL_SUPPORTED						\
 882		(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |		\
 883		 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |	\
 884		 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
 885
 886STATIC long
 887xfs_file_fallocate(
 888	struct file		*file,
 889	int			mode,
 890	loff_t			offset,
 891	loff_t			len)
 892{
 893	struct inode		*inode = file_inode(file);
 894	struct xfs_inode	*ip = XFS_I(inode);
 895	long			error;
 896	uint			iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
 897	loff_t			new_size = 0;
 898	bool			do_file_insert = false;
 899
 900	if (!S_ISREG(inode->i_mode))
 901		return -EINVAL;
 902	if (mode & ~XFS_FALLOC_FL_SUPPORTED)
 903		return -EOPNOTSUPP;
 904
 905	xfs_ilock(ip, iolock);
 906	error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
 907	if (error)
 908		goto out_unlock;
 909
 910	/*
 911	 * Must wait for all AIO to complete before we continue as AIO can
 912	 * change the file size on completion without holding any locks we
 913	 * currently hold. We must do this first because AIO can update both
 914	 * the on disk and in memory inode sizes, and the operations that follow
 915	 * require the in-memory size to be fully up-to-date.
 916	 */
 917	inode_dio_wait(inode);
 918
 919	/*
 920	 * Now AIO and DIO has drained we flush and (if necessary) invalidate
 921	 * the cached range over the first operation we are about to run.
 922	 *
 923	 * We care about zero and collapse here because they both run a hole
 924	 * punch over the range first. Because that can zero data, and the range
 925	 * of invalidation for the shift operations is much larger, we still do
 926	 * the required flush for collapse in xfs_prepare_shift().
 927	 *
 928	 * Insert has the same range requirements as collapse, and we extend the
 929	 * file first which can zero data. Hence insert has the same
 930	 * flush/invalidate requirements as collapse and so they are both
 931	 * handled at the right time by xfs_prepare_shift().
 932	 */
 933	if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
 934		    FALLOC_FL_COLLAPSE_RANGE)) {
 935		error = xfs_flush_unmap_range(ip, offset, len);
 936		if (error)
 937			goto out_unlock;
 938	}
 939
 940	error = file_modified(file);
 941	if (error)
 942		goto out_unlock;
 943
 944	if (mode & FALLOC_FL_PUNCH_HOLE) {
 
 945		error = xfs_free_file_space(ip, offset, len);
 946		if (error)
 947			goto out_unlock;
 948	} else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
 949		if (!xfs_is_falloc_aligned(ip, offset, len)) {
 950			error = -EINVAL;
 951			goto out_unlock;
 952		}
 953
 954		/*
 955		 * There is no need to overlap collapse range with EOF,
 956		 * in which case it is effectively a truncate operation
 957		 */
 958		if (offset + len >= i_size_read(inode)) {
 959			error = -EINVAL;
 960			goto out_unlock;
 961		}
 962
 963		new_size = i_size_read(inode) - len;
 964
 965		error = xfs_collapse_file_space(ip, offset, len);
 966		if (error)
 967			goto out_unlock;
 968	} else if (mode & FALLOC_FL_INSERT_RANGE) {
 969		loff_t		isize = i_size_read(inode);
 970
 971		if (!xfs_is_falloc_aligned(ip, offset, len)) {
 972			error = -EINVAL;
 973			goto out_unlock;
 974		}
 975
 976		/*
 977		 * New inode size must not exceed ->s_maxbytes, accounting for
 978		 * possible signed overflow.
 979		 */
 980		if (inode->i_sb->s_maxbytes - isize < len) {
 981			error = -EFBIG;
 982			goto out_unlock;
 983		}
 984		new_size = isize + len;
 985
 986		/* Offset should be less than i_size */
 987		if (offset >= isize) {
 988			error = -EINVAL;
 989			goto out_unlock;
 990		}
 991		do_file_insert = true;
 992	} else {
 993		if (!(mode & FALLOC_FL_KEEP_SIZE) &&
 994		    offset + len > i_size_read(inode)) {
 995			new_size = offset + len;
 996			error = inode_newsize_ok(inode, new_size);
 997			if (error)
 998				goto out_unlock;
 999		}
1000
1001		if (mode & FALLOC_FL_ZERO_RANGE) {
1002			/*
1003			 * Punch a hole and prealloc the range.  We use a hole
1004			 * punch rather than unwritten extent conversion for two
1005			 * reasons:
1006			 *
1007			 *   1.) Hole punch handles partial block zeroing for us.
1008			 *   2.) If prealloc returns ENOSPC, the file range is
1009			 *       still zero-valued by virtue of the hole punch.
1010			 */
1011			unsigned int blksize = i_blocksize(inode);
1012
1013			trace_xfs_zero_file_space(ip);
1014
1015			error = xfs_free_file_space(ip, offset, len);
1016			if (error)
1017				goto out_unlock;
1018
1019			len = round_up(offset + len, blksize) -
1020			      round_down(offset, blksize);
1021			offset = round_down(offset, blksize);
1022		} else if (mode & FALLOC_FL_UNSHARE_RANGE) {
1023			error = xfs_reflink_unshare(ip, offset, len);
1024			if (error)
1025				goto out_unlock;
1026		} else {
1027			/*
1028			 * If always_cow mode we can't use preallocations and
1029			 * thus should not create them.
1030			 */
1031			if (xfs_is_always_cow_inode(ip)) {
1032				error = -EOPNOTSUPP;
1033				goto out_unlock;
1034			}
1035		}
1036
1037		if (!xfs_is_always_cow_inode(ip)) {
1038			error = xfs_alloc_file_space(ip, offset, len);
1039			if (error)
1040				goto out_unlock;
1041		}
1042	}
1043
1044	/* Change file size if needed */
1045	if (new_size) {
1046		struct iattr iattr;
1047
1048		iattr.ia_valid = ATTR_SIZE;
1049		iattr.ia_size = new_size;
1050		error = xfs_vn_setattr_size(file_mnt_user_ns(file),
1051					    file_dentry(file), &iattr);
1052		if (error)
1053			goto out_unlock;
1054	}
1055
1056	/*
1057	 * Perform hole insertion now that the file size has been
1058	 * updated so that if we crash during the operation we don't
1059	 * leave shifted extents past EOF and hence losing access to
1060	 * the data that is contained within them.
1061	 */
1062	if (do_file_insert) {
1063		error = xfs_insert_file_space(ip, offset, len);
1064		if (error)
1065			goto out_unlock;
1066	}
1067
1068	if (xfs_file_sync_writes(file))
1069		error = xfs_log_force_inode(ip);
1070
1071out_unlock:
1072	xfs_iunlock(ip, iolock);
1073	return error;
1074}
1075
1076STATIC int
1077xfs_file_fadvise(
1078	struct file	*file,
1079	loff_t		start,
1080	loff_t		end,
1081	int		advice)
1082{
1083	struct xfs_inode *ip = XFS_I(file_inode(file));
1084	int ret;
1085	int lockflags = 0;
1086
1087	/*
1088	 * Operations creating pages in page cache need protection from hole
1089	 * punching and similar ops
1090	 */
1091	if (advice == POSIX_FADV_WILLNEED) {
1092		lockflags = XFS_IOLOCK_SHARED;
1093		xfs_ilock(ip, lockflags);
1094	}
1095	ret = generic_fadvise(file, start, end, advice);
1096	if (lockflags)
1097		xfs_iunlock(ip, lockflags);
1098	return ret;
1099}
1100
1101STATIC loff_t
1102xfs_file_remap_range(
1103	struct file		*file_in,
1104	loff_t			pos_in,
1105	struct file		*file_out,
1106	loff_t			pos_out,
1107	loff_t			len,
1108	unsigned int		remap_flags)
1109{
1110	struct inode		*inode_in = file_inode(file_in);
1111	struct xfs_inode	*src = XFS_I(inode_in);
1112	struct inode		*inode_out = file_inode(file_out);
1113	struct xfs_inode	*dest = XFS_I(inode_out);
1114	struct xfs_mount	*mp = src->i_mount;
1115	loff_t			remapped = 0;
1116	xfs_extlen_t		cowextsize;
1117	int			ret;
1118
1119	if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1120		return -EINVAL;
1121
1122	if (!xfs_has_reflink(mp))
1123		return -EOPNOTSUPP;
1124
1125	if (xfs_is_shutdown(mp))
1126		return -EIO;
1127
1128	/* Prepare and then clone file data. */
1129	ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1130			&len, remap_flags);
1131	if (ret || len == 0)
1132		return ret;
1133
1134	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1135
1136	ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1137			&remapped);
1138	if (ret)
1139		goto out_unlock;
1140
1141	/*
1142	 * Carry the cowextsize hint from src to dest if we're sharing the
1143	 * entire source file to the entire destination file, the source file
1144	 * has a cowextsize hint, and the destination file does not.
1145	 */
1146	cowextsize = 0;
1147	if (pos_in == 0 && len == i_size_read(inode_in) &&
1148	    (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1149	    pos_out == 0 && len >= i_size_read(inode_out) &&
1150	    !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1151		cowextsize = src->i_cowextsize;
1152
1153	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1154			remap_flags);
1155	if (ret)
1156		goto out_unlock;
1157
1158	if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1159		xfs_log_force_inode(dest);
1160out_unlock:
1161	xfs_iunlock2_io_mmap(src, dest);
1162	if (ret)
1163		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
 
 
 
 
 
 
 
 
1164	return remapped > 0 ? remapped : ret;
1165}
1166
1167STATIC int
1168xfs_file_open(
1169	struct inode	*inode,
1170	struct file	*file)
1171{
1172	if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1173		return -EIO;
1174	file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC;
 
 
1175	return generic_file_open(inode, file);
1176}
1177
1178STATIC int
1179xfs_dir_open(
1180	struct inode	*inode,
1181	struct file	*file)
1182{
1183	struct xfs_inode *ip = XFS_I(inode);
1184	unsigned int	mode;
1185	int		error;
1186
1187	error = xfs_file_open(inode, file);
 
 
1188	if (error)
1189		return error;
1190
1191	/*
1192	 * If there are any blocks, read-ahead block 0 as we're almost
1193	 * certain to have the next operation be a read there.
1194	 */
1195	mode = xfs_ilock_data_map_shared(ip);
1196	if (ip->i_df.if_nextents > 0)
1197		error = xfs_dir3_data_readahead(ip, 0, 0);
1198	xfs_iunlock(ip, mode);
1199	return error;
1200}
1201
 
 
 
 
1202STATIC int
1203xfs_file_release(
1204	struct inode	*inode,
1205	struct file	*filp)
1206{
1207	return xfs_release(XFS_I(inode));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1208}
1209
1210STATIC int
1211xfs_file_readdir(
1212	struct file	*file,
1213	struct dir_context *ctx)
1214{
1215	struct inode	*inode = file_inode(file);
1216	xfs_inode_t	*ip = XFS_I(inode);
1217	size_t		bufsize;
1218
1219	/*
1220	 * The Linux API doesn't pass down the total size of the buffer
1221	 * we read into down to the filesystem.  With the filldir concept
1222	 * it's not needed for correct information, but the XFS dir2 leaf
1223	 * code wants an estimate of the buffer size to calculate it's
1224	 * readahead window and size the buffers used for mapping to
1225	 * physical blocks.
1226	 *
1227	 * Try to give it an estimate that's good enough, maybe at some
1228	 * point we can change the ->readdir prototype to include the
1229	 * buffer size.  For now we use the current glibc buffer size.
1230	 */
1231	bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1232
1233	return xfs_readdir(NULL, ip, ctx, bufsize);
1234}
1235
1236STATIC loff_t
1237xfs_file_llseek(
1238	struct file	*file,
1239	loff_t		offset,
1240	int		whence)
1241{
1242	struct inode		*inode = file->f_mapping->host;
1243
1244	if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1245		return -EIO;
1246
1247	switch (whence) {
1248	default:
1249		return generic_file_llseek(file, offset, whence);
1250	case SEEK_HOLE:
1251		offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1252		break;
1253	case SEEK_DATA:
1254		offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1255		break;
1256	}
1257
1258	if (offset < 0)
1259		return offset;
1260	return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1261}
1262
1263#ifdef CONFIG_FS_DAX
1264static inline vm_fault_t
1265xfs_dax_fault(
1266	struct vm_fault		*vmf,
1267	enum page_entry_size	pe_size,
1268	bool			write_fault,
1269	pfn_t			*pfn)
1270{
1271	return dax_iomap_fault(vmf, pe_size, pfn, NULL,
 
 
 
 
 
 
 
1272			(write_fault && !vmf->cow_page) ?
1273				&xfs_dax_write_iomap_ops :
1274				&xfs_read_iomap_ops);
 
 
 
1275}
1276#else
1277static inline vm_fault_t
1278xfs_dax_fault(
1279	struct vm_fault		*vmf,
1280	enum page_entry_size	pe_size,
1281	bool			write_fault,
1282	pfn_t			*pfn)
1283{
1284	ASSERT(0);
1285	return VM_FAULT_SIGBUS;
 
 
 
 
 
 
 
 
1286}
1287#endif
1288
1289/*
1290 * Locking for serialisation of IO during page faults. This results in a lock
1291 * ordering of:
1292 *
1293 * mmap_lock (MM)
1294 *   sb_start_pagefault(vfs, freeze)
1295 *     invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1296 *       page_lock (MM)
1297 *         i_lock (XFS - extent map serialisation)
1298 */
1299static vm_fault_t
1300__xfs_filemap_fault(
1301	struct vm_fault		*vmf,
1302	enum page_entry_size	pe_size,
1303	bool			write_fault)
1304{
1305	struct inode		*inode = file_inode(vmf->vma->vm_file);
1306	struct xfs_inode	*ip = XFS_I(inode);
 
1307	vm_fault_t		ret;
1308
1309	trace_xfs_filemap_fault(ip, pe_size, write_fault);
1310
1311	if (write_fault) {
1312		sb_start_pagefault(inode->i_sb);
1313		file_update_time(vmf->vma->vm_file);
1314	}
1315
1316	if (IS_DAX(inode)) {
1317		pfn_t pfn;
1318
1319		xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1320		ret = xfs_dax_fault(vmf, pe_size, write_fault, &pfn);
1321		if (ret & VM_FAULT_NEEDDSYNC)
1322			ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1323		xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1324	} else {
1325		if (write_fault) {
1326			xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1327			ret = iomap_page_mkwrite(vmf,
1328					&xfs_page_mkwrite_iomap_ops);
1329			xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1330		} else {
1331			ret = filemap_fault(vmf);
1332		}
1333	}
1334
1335	if (write_fault)
1336		sb_end_pagefault(inode->i_sb);
 
 
 
 
 
1337	return ret;
1338}
1339
1340static inline bool
1341xfs_is_write_fault(
1342	struct vm_fault		*vmf)
1343{
1344	return (vmf->flags & FAULT_FLAG_WRITE) &&
1345	       (vmf->vma->vm_flags & VM_SHARED);
1346}
1347
1348static vm_fault_t
1349xfs_filemap_fault(
1350	struct vm_fault		*vmf)
1351{
 
 
1352	/* DAX can shortcut the normal fault path on write faults! */
1353	return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1354			IS_DAX(file_inode(vmf->vma->vm_file)) &&
1355			xfs_is_write_fault(vmf));
 
 
 
 
 
1356}
1357
1358static vm_fault_t
1359xfs_filemap_huge_fault(
1360	struct vm_fault		*vmf,
1361	enum page_entry_size	pe_size)
1362{
1363	if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1364		return VM_FAULT_FALLBACK;
1365
1366	/* DAX can shortcut the normal fault path on write faults! */
1367	return __xfs_filemap_fault(vmf, pe_size,
1368			xfs_is_write_fault(vmf));
 
1369}
1370
1371static vm_fault_t
1372xfs_filemap_page_mkwrite(
1373	struct vm_fault		*vmf)
1374{
1375	return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1376}
1377
1378/*
1379 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1380 * on write faults. In reality, it needs to serialise against truncate and
1381 * prepare memory for writing so handle is as standard write fault.
1382 */
1383static vm_fault_t
1384xfs_filemap_pfn_mkwrite(
1385	struct vm_fault		*vmf)
1386{
1387
1388	return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1389}
1390
1391static vm_fault_t
1392xfs_filemap_map_pages(
1393	struct vm_fault		*vmf,
1394	pgoff_t			start_pgoff,
1395	pgoff_t			end_pgoff)
1396{
1397	struct inode		*inode = file_inode(vmf->vma->vm_file);
1398	vm_fault_t ret;
1399
1400	xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1401	ret = filemap_map_pages(vmf, start_pgoff, end_pgoff);
1402	xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1403	return ret;
1404}
1405
1406static const struct vm_operations_struct xfs_file_vm_ops = {
1407	.fault		= xfs_filemap_fault,
1408	.huge_fault	= xfs_filemap_huge_fault,
1409	.map_pages	= xfs_filemap_map_pages,
1410	.page_mkwrite	= xfs_filemap_page_mkwrite,
1411	.pfn_mkwrite	= xfs_filemap_pfn_mkwrite,
1412};
1413
1414STATIC int
1415xfs_file_mmap(
1416	struct file		*file,
1417	struct vm_area_struct	*vma)
1418{
1419	struct inode		*inode = file_inode(file);
1420	struct xfs_buftarg	*target = xfs_inode_buftarg(XFS_I(inode));
1421
1422	/*
1423	 * We don't support synchronous mappings for non-DAX files and
1424	 * for DAX files if underneath dax_device is not synchronous.
1425	 */
1426	if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1427		return -EOPNOTSUPP;
1428
1429	file_accessed(file);
1430	vma->vm_ops = &xfs_file_vm_ops;
1431	if (IS_DAX(inode))
1432		vma->vm_flags |= VM_HUGEPAGE;
1433	return 0;
1434}
1435
1436const struct file_operations xfs_file_operations = {
1437	.llseek		= xfs_file_llseek,
1438	.read_iter	= xfs_file_read_iter,
1439	.write_iter	= xfs_file_write_iter,
1440	.splice_read	= generic_file_splice_read,
1441	.splice_write	= iter_file_splice_write,
1442	.iopoll		= iocb_bio_iopoll,
1443	.unlocked_ioctl	= xfs_file_ioctl,
1444#ifdef CONFIG_COMPAT
1445	.compat_ioctl	= xfs_file_compat_ioctl,
1446#endif
1447	.mmap		= xfs_file_mmap,
1448	.mmap_supported_flags = MAP_SYNC,
1449	.open		= xfs_file_open,
1450	.release	= xfs_file_release,
1451	.fsync		= xfs_file_fsync,
1452	.get_unmapped_area = thp_get_unmapped_area,
1453	.fallocate	= xfs_file_fallocate,
1454	.fadvise	= xfs_file_fadvise,
1455	.remap_file_range = xfs_file_remap_range,
 
 
1456};
1457
1458const struct file_operations xfs_dir_file_operations = {
1459	.open		= xfs_dir_open,
1460	.read		= generic_read_dir,
1461	.iterate_shared	= xfs_file_readdir,
1462	.llseek		= generic_file_llseek,
1463	.unlocked_ioctl	= xfs_file_ioctl,
1464#ifdef CONFIG_COMPAT
1465	.compat_ioctl	= xfs_file_compat_ioctl,
1466#endif
1467	.fsync		= xfs_dir_fsync,
1468};
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4 * All Rights Reserved.
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_mount.h"
  13#include "xfs_inode.h"
  14#include "xfs_trans.h"
  15#include "xfs_inode_item.h"
  16#include "xfs_bmap.h"
  17#include "xfs_bmap_util.h"
  18#include "xfs_dir2.h"
  19#include "xfs_dir2_priv.h"
  20#include "xfs_ioctl.h"
  21#include "xfs_trace.h"
  22#include "xfs_log.h"
  23#include "xfs_icache.h"
  24#include "xfs_pnfs.h"
  25#include "xfs_iomap.h"
  26#include "xfs_reflink.h"
  27#include "xfs_file.h"
  28
  29#include <linux/dax.h>
  30#include <linux/falloc.h>
  31#include <linux/backing-dev.h>
  32#include <linux/mman.h>
  33#include <linux/fadvise.h>
  34#include <linux/mount.h>
  35
  36static const struct vm_operations_struct xfs_file_vm_ops;
  37
  38/*
  39 * Decide if the given file range is aligned to the size of the fundamental
  40 * allocation unit for the file.
  41 */
  42bool
  43xfs_is_falloc_aligned(
  44	struct xfs_inode	*ip,
  45	loff_t			pos,
  46	long long int		len)
  47{
  48	unsigned int		alloc_unit = xfs_inode_alloc_unitsize(ip);
 
  49
  50	if (!is_power_of_2(alloc_unit))
  51		return isaligned_64(pos, alloc_unit) &&
  52		       isaligned_64(len, alloc_unit);
 
 
 
 
 
 
 
 
 
 
 
 
 
  53
  54	return !((pos | len) & (alloc_unit - 1));
  55}
  56
  57/*
  58 * Fsync operations on directories are much simpler than on regular files,
  59 * as there is no file data to flush, and thus also no need for explicit
  60 * cache flush operations, and there are no non-transaction metadata updates
  61 * on directories either.
  62 */
  63STATIC int
  64xfs_dir_fsync(
  65	struct file		*file,
  66	loff_t			start,
  67	loff_t			end,
  68	int			datasync)
  69{
  70	struct xfs_inode	*ip = XFS_I(file->f_mapping->host);
  71
  72	trace_xfs_dir_fsync(ip);
  73	return xfs_log_force_inode(ip);
  74}
  75
  76static xfs_csn_t
  77xfs_fsync_seq(
  78	struct xfs_inode	*ip,
  79	bool			datasync)
  80{
  81	if (!xfs_ipincount(ip))
  82		return 0;
  83	if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
  84		return 0;
  85	return ip->i_itemp->ili_commit_seq;
  86}
  87
  88/*
  89 * All metadata updates are logged, which means that we just have to flush the
  90 * log up to the latest LSN that touched the inode.
  91 *
  92 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
  93 * the log force before we clear the ili_fsync_fields field. This ensures that
  94 * we don't get a racing sync operation that does not wait for the metadata to
  95 * hit the journal before returning.  If we race with clearing ili_fsync_fields,
  96 * then all that will happen is the log force will do nothing as the lsn will
  97 * already be on disk.  We can't race with setting ili_fsync_fields because that
  98 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
  99 * shared until after the ili_fsync_fields is cleared.
 100 */
 101static  int
 102xfs_fsync_flush_log(
 103	struct xfs_inode	*ip,
 104	bool			datasync,
 105	int			*log_flushed)
 106{
 107	int			error = 0;
 108	xfs_csn_t		seq;
 109
 110	xfs_ilock(ip, XFS_ILOCK_SHARED);
 111	seq = xfs_fsync_seq(ip, datasync);
 112	if (seq) {
 113		error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
 114					  log_flushed);
 115
 116		spin_lock(&ip->i_itemp->ili_lock);
 117		ip->i_itemp->ili_fsync_fields = 0;
 118		spin_unlock(&ip->i_itemp->ili_lock);
 119	}
 120	xfs_iunlock(ip, XFS_ILOCK_SHARED);
 121	return error;
 122}
 123
 124STATIC int
 125xfs_file_fsync(
 126	struct file		*file,
 127	loff_t			start,
 128	loff_t			end,
 129	int			datasync)
 130{
 131	struct xfs_inode	*ip = XFS_I(file->f_mapping->host);
 132	struct xfs_mount	*mp = ip->i_mount;
 133	int			error, err2;
 134	int			log_flushed = 0;
 135
 136	trace_xfs_file_fsync(ip);
 137
 138	error = file_write_and_wait_range(file, start, end);
 139	if (error)
 140		return error;
 141
 142	if (xfs_is_shutdown(mp))
 143		return -EIO;
 144
 145	xfs_iflags_clear(ip, XFS_ITRUNCATED);
 146
 147	/*
 148	 * If we have an RT and/or log subvolume we need to make sure to flush
 149	 * the write cache the device used for file data first.  This is to
 150	 * ensure newly written file data make it to disk before logging the new
 151	 * inode size in case of an extending write.
 152	 */
 153	if (XFS_IS_REALTIME_INODE(ip))
 154		error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
 155	else if (mp->m_logdev_targp != mp->m_ddev_targp)
 156		error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
 157
 158	/*
 159	 * Any inode that has dirty modifications in the log is pinned.  The
 160	 * racy check here for a pinned inode will not catch modifications
 161	 * that happen concurrently to the fsync call, but fsync semantics
 162	 * only require to sync previously completed I/O.
 163	 */
 164	if (xfs_ipincount(ip)) {
 165		err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
 166		if (err2 && !error)
 167			error = err2;
 168	}
 169
 170	/*
 171	 * If we only have a single device, and the log force about was
 172	 * a no-op we might have to flush the data device cache here.
 173	 * This can only happen for fdatasync/O_DSYNC if we were overwriting
 174	 * an already allocated file and thus do not have any metadata to
 175	 * commit.
 176	 */
 177	if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
 178	    mp->m_logdev_targp == mp->m_ddev_targp) {
 179		err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
 180		if (err2 && !error)
 181			error = err2;
 182	}
 183
 184	return error;
 185}
 186
 187static int
 188xfs_ilock_iocb(
 189	struct kiocb		*iocb,
 190	unsigned int		lock_mode)
 191{
 192	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 193
 194	if (iocb->ki_flags & IOCB_NOWAIT) {
 195		if (!xfs_ilock_nowait(ip, lock_mode))
 196			return -EAGAIN;
 197	} else {
 198		xfs_ilock(ip, lock_mode);
 199	}
 200
 201	return 0;
 202}
 203
 204static int
 205xfs_ilock_iocb_for_write(
 206	struct kiocb		*iocb,
 207	unsigned int		*lock_mode)
 208{
 209	ssize_t			ret;
 210	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 211
 212	ret = xfs_ilock_iocb(iocb, *lock_mode);
 213	if (ret)
 214		return ret;
 215
 216	/*
 217	 * If a reflink remap is in progress we always need to take the iolock
 218	 * exclusively to wait for it to finish.
 219	 */
 220	if (*lock_mode == XFS_IOLOCK_SHARED &&
 221	    xfs_iflags_test(ip, XFS_IREMAPPING)) {
 222		xfs_iunlock(ip, *lock_mode);
 223		*lock_mode = XFS_IOLOCK_EXCL;
 224		return xfs_ilock_iocb(iocb, *lock_mode);
 225	}
 226
 227	return 0;
 228}
 229
 230STATIC ssize_t
 231xfs_file_dio_read(
 232	struct kiocb		*iocb,
 233	struct iov_iter		*to)
 234{
 235	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 236	ssize_t			ret;
 237
 238	trace_xfs_file_direct_read(iocb, to);
 239
 240	if (!iov_iter_count(to))
 241		return 0; /* skip atime */
 242
 243	file_accessed(iocb->ki_filp);
 244
 245	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 246	if (ret)
 247		return ret;
 248	ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
 249	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 250
 251	return ret;
 252}
 253
 254static noinline ssize_t
 255xfs_file_dax_read(
 256	struct kiocb		*iocb,
 257	struct iov_iter		*to)
 258{
 259	struct xfs_inode	*ip = XFS_I(iocb->ki_filp->f_mapping->host);
 260	ssize_t			ret = 0;
 261
 262	trace_xfs_file_dax_read(iocb, to);
 263
 264	if (!iov_iter_count(to))
 265		return 0; /* skip atime */
 266
 267	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 268	if (ret)
 269		return ret;
 270	ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
 271	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 272
 273	file_accessed(iocb->ki_filp);
 274	return ret;
 275}
 276
 277STATIC ssize_t
 278xfs_file_buffered_read(
 279	struct kiocb		*iocb,
 280	struct iov_iter		*to)
 281{
 282	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 283	ssize_t			ret;
 284
 285	trace_xfs_file_buffered_read(iocb, to);
 286
 287	ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
 288	if (ret)
 289		return ret;
 290	ret = generic_file_read_iter(iocb, to);
 291	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 292
 293	return ret;
 294}
 295
 296STATIC ssize_t
 297xfs_file_read_iter(
 298	struct kiocb		*iocb,
 299	struct iov_iter		*to)
 300{
 301	struct inode		*inode = file_inode(iocb->ki_filp);
 302	struct xfs_mount	*mp = XFS_I(inode)->i_mount;
 303	ssize_t			ret = 0;
 304
 305	XFS_STATS_INC(mp, xs_read_calls);
 306
 307	if (xfs_is_shutdown(mp))
 308		return -EIO;
 309
 310	if (IS_DAX(inode))
 311		ret = xfs_file_dax_read(iocb, to);
 312	else if (iocb->ki_flags & IOCB_DIRECT)
 313		ret = xfs_file_dio_read(iocb, to);
 314	else
 315		ret = xfs_file_buffered_read(iocb, to);
 316
 317	if (ret > 0)
 318		XFS_STATS_ADD(mp, xs_read_bytes, ret);
 319	return ret;
 320}
 321
 322STATIC ssize_t
 323xfs_file_splice_read(
 324	struct file		*in,
 325	loff_t			*ppos,
 326	struct pipe_inode_info	*pipe,
 327	size_t			len,
 328	unsigned int		flags)
 329{
 330	struct inode		*inode = file_inode(in);
 331	struct xfs_inode	*ip = XFS_I(inode);
 332	struct xfs_mount	*mp = ip->i_mount;
 333	ssize_t			ret = 0;
 334
 335	XFS_STATS_INC(mp, xs_read_calls);
 336
 337	if (xfs_is_shutdown(mp))
 338		return -EIO;
 339
 340	trace_xfs_file_splice_read(ip, *ppos, len);
 341
 342	xfs_ilock(ip, XFS_IOLOCK_SHARED);
 343	ret = filemap_splice_read(in, ppos, pipe, len, flags);
 344	xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 345	if (ret > 0)
 346		XFS_STATS_ADD(mp, xs_read_bytes, ret);
 347	return ret;
 348}
 349
 350/*
 351 * Take care of zeroing post-EOF blocks when they might exist.
 352 *
 353 * Returns 0 if successfully, a negative error for a failure, or 1 if this
 354 * function dropped the iolock and reacquired it exclusively and the caller
 355 * needs to restart the write sanity checks.
 356 */
 357static ssize_t
 358xfs_file_write_zero_eof(
 359	struct kiocb		*iocb,
 360	struct iov_iter		*from,
 361	unsigned int		*iolock,
 362	size_t			count,
 363	bool			*drained_dio)
 364{
 365	struct xfs_inode	*ip = XFS_I(iocb->ki_filp->f_mapping->host);
 366	loff_t			isize;
 367	int			error;
 368
 369	/*
 370	 * We need to serialise against EOF updates that occur in IO completions
 371	 * here. We want to make sure that nobody is changing the size while
 372	 * we do this check until we have placed an IO barrier (i.e. hold
 373	 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.  The
 374	 * spinlock effectively forms a memory barrier once we have
 375	 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
 376	 * hence be able to correctly determine if we need to run zeroing.
 377	 */
 378	spin_lock(&ip->i_flags_lock);
 379	isize = i_size_read(VFS_I(ip));
 380	if (iocb->ki_pos <= isize) {
 381		spin_unlock(&ip->i_flags_lock);
 382		return 0;
 383	}
 384	spin_unlock(&ip->i_flags_lock);
 385
 386	if (iocb->ki_flags & IOCB_NOWAIT)
 387		return -EAGAIN;
 388
 389	if (!*drained_dio) {
 390		/*
 391		 * If zeroing is needed and we are currently holding the iolock
 392		 * shared, we need to update it to exclusive which implies
 393		 * having to redo all checks before.
 394		 */
 395		if (*iolock == XFS_IOLOCK_SHARED) {
 396			xfs_iunlock(ip, *iolock);
 397			*iolock = XFS_IOLOCK_EXCL;
 398			xfs_ilock(ip, *iolock);
 399			iov_iter_reexpand(from, count);
 400		}
 401
 402		/*
 403		 * We now have an IO submission barrier in place, but AIO can do
 404		 * EOF updates during IO completion and hence we now need to
 405		 * wait for all of them to drain.  Non-AIO DIO will have drained
 406		 * before we are given the XFS_IOLOCK_EXCL, and so for most
 407		 * cases this wait is a no-op.
 408		 */
 409		inode_dio_wait(VFS_I(ip));
 410		*drained_dio = true;
 411		return 1;
 412	}
 413
 414	trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
 415
 416	xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
 417	error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
 418	xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
 419
 420	return error;
 421}
 422
 423/*
 424 * Common pre-write limit and setup checks.
 425 *
 426 * Called with the iolock held either shared and exclusive according to
 427 * @iolock, and returns with it held.  Might upgrade the iolock to exclusive
 428 * if called for a direct write beyond i_size.
 429 */
 430STATIC ssize_t
 431xfs_file_write_checks(
 432	struct kiocb		*iocb,
 433	struct iov_iter		*from,
 434	unsigned int		*iolock)
 435{
 436	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 
 
 
 437	size_t			count = iov_iter_count(from);
 438	bool			drained_dio = false;
 439	ssize_t			error;
 440
 441restart:
 442	error = generic_write_checks(iocb, from);
 443	if (error <= 0)
 444		return error;
 445
 446	if (iocb->ki_flags & IOCB_NOWAIT) {
 447		error = break_layout(inode, false);
 448		if (error == -EWOULDBLOCK)
 449			error = -EAGAIN;
 450	} else {
 451		error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
 452	}
 453
 454	if (error)
 455		return error;
 456
 457	/*
 458	 * For changing security info in file_remove_privs() we need i_rwsem
 459	 * exclusively.
 460	 */
 461	if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
 462		xfs_iunlock(XFS_I(inode), *iolock);
 463		*iolock = XFS_IOLOCK_EXCL;
 464		error = xfs_ilock_iocb(iocb, *iolock);
 465		if (error) {
 466			*iolock = 0;
 467			return error;
 468		}
 469		goto restart;
 470	}
 471
 472	/*
 473	 * If the offset is beyond the size of the file, we need to zero all
 474	 * blocks that fall between the existing EOF and the start of this
 475	 * write.
 
 
 476	 *
 477	 * We can do an unlocked check for i_size here safely as I/O completion
 478	 * can only extend EOF.  Truncate is locked out at this point, so the
 479	 * EOF can not move backwards, only forwards. Hence we only need to take
 480	 * the slow path when we are at or beyond the current EOF.
 481	 */
 482	if (iocb->ki_pos > i_size_read(inode)) {
 483		error = xfs_file_write_zero_eof(iocb, from, iolock, count,
 484				&drained_dio);
 485		if (error == 1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 486			goto restart;
 
 
 
 
 487		if (error)
 488			return error;
 489	}
 
 490
 
 491	return kiocb_modified(iocb);
 492}
 493
 494static int
 495xfs_dio_write_end_io(
 496	struct kiocb		*iocb,
 497	ssize_t			size,
 498	int			error,
 499	unsigned		flags)
 500{
 501	struct inode		*inode = file_inode(iocb->ki_filp);
 502	struct xfs_inode	*ip = XFS_I(inode);
 503	loff_t			offset = iocb->ki_pos;
 504	unsigned int		nofs_flag;
 505
 506	trace_xfs_end_io_direct_write(ip, offset, size);
 507
 508	if (xfs_is_shutdown(ip->i_mount))
 509		return -EIO;
 510
 511	if (error)
 512		return error;
 513	if (!size)
 514		return 0;
 515
 516	/*
 517	 * Capture amount written on completion as we can't reliably account
 518	 * for it on submission.
 519	 */
 520	XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
 521
 522	/*
 523	 * We can allocate memory here while doing writeback on behalf of
 524	 * memory reclaim.  To avoid memory allocation deadlocks set the
 525	 * task-wide nofs context for the following operations.
 526	 */
 527	nofs_flag = memalloc_nofs_save();
 528
 529	if (flags & IOMAP_DIO_COW) {
 530		error = xfs_reflink_end_cow(ip, offset, size);
 531		if (error)
 532			goto out;
 533	}
 534
 535	/*
 536	 * Unwritten conversion updates the in-core isize after extent
 537	 * conversion but before updating the on-disk size. Updating isize any
 538	 * earlier allows a racing dio read to find unwritten extents before
 539	 * they are converted.
 540	 */
 541	if (flags & IOMAP_DIO_UNWRITTEN) {
 542		error = xfs_iomap_write_unwritten(ip, offset, size, true);
 543		goto out;
 544	}
 545
 546	/*
 547	 * We need to update the in-core inode size here so that we don't end up
 548	 * with the on-disk inode size being outside the in-core inode size. We
 549	 * have no other method of updating EOF for AIO, so always do it here
 550	 * if necessary.
 551	 *
 552	 * We need to lock the test/set EOF update as we can be racing with
 553	 * other IO completions here to update the EOF. Failing to serialise
 554	 * here can result in EOF moving backwards and Bad Things Happen when
 555	 * that occurs.
 556	 *
 557	 * As IO completion only ever extends EOF, we can do an unlocked check
 558	 * here to avoid taking the spinlock. If we land within the current EOF,
 559	 * then we do not need to do an extending update at all, and we don't
 560	 * need to take the lock to check this. If we race with an update moving
 561	 * EOF, then we'll either still be beyond EOF and need to take the lock,
 562	 * or we'll be within EOF and we don't need to take it at all.
 563	 */
 564	if (offset + size <= i_size_read(inode))
 565		goto out;
 566
 567	spin_lock(&ip->i_flags_lock);
 568	if (offset + size > i_size_read(inode)) {
 569		i_size_write(inode, offset + size);
 570		spin_unlock(&ip->i_flags_lock);
 571		error = xfs_setfilesize(ip, offset, size);
 572	} else {
 573		spin_unlock(&ip->i_flags_lock);
 574	}
 575
 576out:
 577	memalloc_nofs_restore(nofs_flag);
 578	return error;
 579}
 580
 581static const struct iomap_dio_ops xfs_dio_write_ops = {
 582	.end_io		= xfs_dio_write_end_io,
 583};
 584
 585/*
 586 * Handle block aligned direct I/O writes
 587 */
 588static noinline ssize_t
 589xfs_file_dio_write_aligned(
 590	struct xfs_inode	*ip,
 591	struct kiocb		*iocb,
 592	struct iov_iter		*from)
 593{
 594	unsigned int		iolock = XFS_IOLOCK_SHARED;
 595	ssize_t			ret;
 596
 597	ret = xfs_ilock_iocb_for_write(iocb, &iolock);
 598	if (ret)
 599		return ret;
 600	ret = xfs_file_write_checks(iocb, from, &iolock);
 601	if (ret)
 602		goto out_unlock;
 603
 604	/*
 605	 * We don't need to hold the IOLOCK exclusively across the IO, so demote
 606	 * the iolock back to shared if we had to take the exclusive lock in
 607	 * xfs_file_write_checks() for other reasons.
 608	 */
 609	if (iolock == XFS_IOLOCK_EXCL) {
 610		xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
 611		iolock = XFS_IOLOCK_SHARED;
 612	}
 613	trace_xfs_file_direct_write(iocb, from);
 614	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
 615			   &xfs_dio_write_ops, 0, NULL, 0);
 616out_unlock:
 617	if (iolock)
 618		xfs_iunlock(ip, iolock);
 619	return ret;
 620}
 621
 622/*
 623 * Handle block unaligned direct I/O writes
 624 *
 625 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
 626 * them to be done in parallel with reads and other direct I/O writes.  However,
 627 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
 628 * to do sub-block zeroing and that requires serialisation against other direct
 629 * I/O to the same block.  In this case we need to serialise the submission of
 630 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
 631 * In the case where sub-block zeroing is not required, we can do concurrent
 632 * sub-block dios to the same block successfully.
 633 *
 634 * Optimistically submit the I/O using the shared lock first, but use the
 635 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
 636 * if block allocation or partial block zeroing would be required.  In that case
 637 * we try again with the exclusive lock.
 638 */
 639static noinline ssize_t
 640xfs_file_dio_write_unaligned(
 641	struct xfs_inode	*ip,
 642	struct kiocb		*iocb,
 643	struct iov_iter		*from)
 644{
 645	size_t			isize = i_size_read(VFS_I(ip));
 646	size_t			count = iov_iter_count(from);
 647	unsigned int		iolock = XFS_IOLOCK_SHARED;
 648	unsigned int		flags = IOMAP_DIO_OVERWRITE_ONLY;
 649	ssize_t			ret;
 650
 651	/*
 652	 * Extending writes need exclusivity because of the sub-block zeroing
 653	 * that the DIO code always does for partial tail blocks beyond EOF, so
 654	 * don't even bother trying the fast path in this case.
 655	 */
 656	if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
 657		if (iocb->ki_flags & IOCB_NOWAIT)
 658			return -EAGAIN;
 659retry_exclusive:
 660		iolock = XFS_IOLOCK_EXCL;
 661		flags = IOMAP_DIO_FORCE_WAIT;
 662	}
 663
 664	ret = xfs_ilock_iocb_for_write(iocb, &iolock);
 665	if (ret)
 666		return ret;
 667
 668	/*
 669	 * We can't properly handle unaligned direct I/O to reflink files yet,
 670	 * as we can't unshare a partial block.
 671	 */
 672	if (xfs_is_cow_inode(ip)) {
 673		trace_xfs_reflink_bounce_dio_write(iocb, from);
 674		ret = -ENOTBLK;
 675		goto out_unlock;
 676	}
 677
 678	ret = xfs_file_write_checks(iocb, from, &iolock);
 679	if (ret)
 680		goto out_unlock;
 681
 682	/*
 683	 * If we are doing exclusive unaligned I/O, this must be the only I/O
 684	 * in-flight.  Otherwise we risk data corruption due to unwritten extent
 685	 * conversions from the AIO end_io handler.  Wait for all other I/O to
 686	 * drain first.
 687	 */
 688	if (flags & IOMAP_DIO_FORCE_WAIT)
 689		inode_dio_wait(VFS_I(ip));
 690
 691	trace_xfs_file_direct_write(iocb, from);
 692	ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
 693			   &xfs_dio_write_ops, flags, NULL, 0);
 694
 695	/*
 696	 * Retry unaligned I/O with exclusive blocking semantics if the DIO
 697	 * layer rejected it for mapping or locking reasons. If we are doing
 698	 * nonblocking user I/O, propagate the error.
 699	 */
 700	if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
 701		ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
 702		xfs_iunlock(ip, iolock);
 703		goto retry_exclusive;
 704	}
 705
 706out_unlock:
 707	if (iolock)
 708		xfs_iunlock(ip, iolock);
 709	return ret;
 710}
 711
 712static ssize_t
 713xfs_file_dio_write(
 714	struct kiocb		*iocb,
 715	struct iov_iter		*from)
 716{
 717	struct xfs_inode	*ip = XFS_I(file_inode(iocb->ki_filp));
 718	struct xfs_buftarg      *target = xfs_inode_buftarg(ip);
 719	size_t			count = iov_iter_count(from);
 720
 721	/* direct I/O must be aligned to device logical sector size */
 722	if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
 723		return -EINVAL;
 724	if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
 725		return xfs_file_dio_write_unaligned(ip, iocb, from);
 726	return xfs_file_dio_write_aligned(ip, iocb, from);
 727}
 728
 729static noinline ssize_t
 730xfs_file_dax_write(
 731	struct kiocb		*iocb,
 732	struct iov_iter		*from)
 733{
 734	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 735	struct xfs_inode	*ip = XFS_I(inode);
 736	unsigned int		iolock = XFS_IOLOCK_EXCL;
 737	ssize_t			ret, error = 0;
 738	loff_t			pos;
 739
 740	ret = xfs_ilock_iocb(iocb, iolock);
 741	if (ret)
 742		return ret;
 743	ret = xfs_file_write_checks(iocb, from, &iolock);
 744	if (ret)
 745		goto out;
 746
 747	pos = iocb->ki_pos;
 748
 749	trace_xfs_file_dax_write(iocb, from);
 750	ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
 751	if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
 752		i_size_write(inode, iocb->ki_pos);
 753		error = xfs_setfilesize(ip, pos, ret);
 754	}
 755out:
 756	if (iolock)
 757		xfs_iunlock(ip, iolock);
 758	if (error)
 759		return error;
 760
 761	if (ret > 0) {
 762		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
 763
 764		/* Handle various SYNC-type writes */
 765		ret = generic_write_sync(iocb, ret);
 766	}
 767	return ret;
 768}
 769
 770STATIC ssize_t
 771xfs_file_buffered_write(
 772	struct kiocb		*iocb,
 773	struct iov_iter		*from)
 774{
 775	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 776	struct xfs_inode	*ip = XFS_I(inode);
 777	ssize_t			ret;
 778	bool			cleared_space = false;
 779	unsigned int		iolock;
 780
 781write_retry:
 782	iolock = XFS_IOLOCK_EXCL;
 783	ret = xfs_ilock_iocb(iocb, iolock);
 784	if (ret)
 785		return ret;
 786
 787	ret = xfs_file_write_checks(iocb, from, &iolock);
 788	if (ret)
 789		goto out;
 790
 
 
 
 791	trace_xfs_file_buffered_write(iocb, from);
 792	ret = iomap_file_buffered_write(iocb, from,
 793			&xfs_buffered_write_iomap_ops, NULL);
 
 
 794
 795	/*
 796	 * If we hit a space limit, try to free up some lingering preallocated
 797	 * space before returning an error. In the case of ENOSPC, first try to
 798	 * write back all dirty inodes to free up some of the excess reserved
 799	 * metadata space. This reduces the chances that the eofblocks scan
 800	 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
 801	 * also behaves as a filter to prevent too many eofblocks scans from
 802	 * running at the same time.  Use a synchronous scan to increase the
 803	 * effectiveness of the scan.
 804	 */
 805	if (ret == -EDQUOT && !cleared_space) {
 806		xfs_iunlock(ip, iolock);
 807		xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
 808		cleared_space = true;
 809		goto write_retry;
 810	} else if (ret == -ENOSPC && !cleared_space) {
 811		struct xfs_icwalk	icw = {0};
 812
 813		cleared_space = true;
 814		xfs_flush_inodes(ip->i_mount);
 815
 816		xfs_iunlock(ip, iolock);
 817		icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
 818		xfs_blockgc_free_space(ip->i_mount, &icw);
 819		goto write_retry;
 820	}
 821
 
 822out:
 823	if (iolock)
 824		xfs_iunlock(ip, iolock);
 825
 826	if (ret > 0) {
 827		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
 828		/* Handle various SYNC-type writes */
 829		ret = generic_write_sync(iocb, ret);
 830	}
 831	return ret;
 832}
 833
 834STATIC ssize_t
 835xfs_file_write_iter(
 836	struct kiocb		*iocb,
 837	struct iov_iter		*from)
 838{
 839	struct inode		*inode = iocb->ki_filp->f_mapping->host;
 840	struct xfs_inode	*ip = XFS_I(inode);
 841	ssize_t			ret;
 842	size_t			ocount = iov_iter_count(from);
 843
 844	XFS_STATS_INC(ip->i_mount, xs_write_calls);
 845
 846	if (ocount == 0)
 847		return 0;
 848
 849	if (xfs_is_shutdown(ip->i_mount))
 850		return -EIO;
 851
 852	if (IS_DAX(inode))
 853		return xfs_file_dax_write(iocb, from);
 854
 855	if (iocb->ki_flags & IOCB_ATOMIC) {
 856		/*
 857		 * Currently only atomic writing of a single FS block is
 858		 * supported. It would be possible to atomic write smaller than
 859		 * a FS block, but there is no requirement to support this.
 860		 * Note that iomap also does not support this yet.
 861		 */
 862		if (ocount != ip->i_mount->m_sb.sb_blocksize)
 863			return -EINVAL;
 864		ret = generic_atomic_write_valid(iocb, from);
 865		if (ret)
 866			return ret;
 867	}
 868
 869	if (iocb->ki_flags & IOCB_DIRECT) {
 870		/*
 871		 * Allow a directio write to fall back to a buffered
 872		 * write *only* in the case that we're doing a reflink
 873		 * CoW.  In all other directio scenarios we do not
 874		 * allow an operation to fall back to buffered mode.
 875		 */
 876		ret = xfs_file_dio_write(iocb, from);
 877		if (ret != -ENOTBLK)
 878			return ret;
 879	}
 880
 881	return xfs_file_buffered_write(iocb, from);
 882}
 883
 884/* Does this file, inode, or mount want synchronous writes? */
 885static inline bool xfs_file_sync_writes(struct file *filp)
 
 886{
 887	struct xfs_inode	*ip = XFS_I(file_inode(filp));
 888
 889	if (xfs_has_wsync(ip->i_mount))
 890		return true;
 891	if (filp->f_flags & (__O_SYNC | O_DSYNC))
 892		return true;
 893	if (IS_SYNC(file_inode(filp)))
 894		return true;
 895
 896	return false;
 897}
 898
 899static int
 900xfs_falloc_newsize(
 901	struct file		*file,
 902	int			mode,
 903	loff_t			offset,
 904	loff_t			len,
 905	loff_t			*new_size)
 906{
 907	struct inode		*inode = file_inode(file);
 908
 909	if ((mode & FALLOC_FL_KEEP_SIZE) || offset + len <= i_size_read(inode))
 910		return 0;
 911	*new_size = offset + len;
 912	return inode_newsize_ok(inode, *new_size);
 913}
 914
 915static int
 916xfs_falloc_setsize(
 917	struct file		*file,
 918	loff_t			new_size)
 919{
 920	struct iattr iattr = {
 921		.ia_valid	= ATTR_SIZE,
 922		.ia_size	= new_size,
 923	};
 924
 925	if (!new_size)
 
 926		return 0;
 927	return xfs_vn_setattr_size(file_mnt_idmap(file), file_dentry(file),
 928			&iattr);
 929}
 930
 931static int
 932xfs_falloc_collapse_range(
 933	struct file		*file,
 934	loff_t			offset,
 935	loff_t			len)
 936{
 937	struct inode		*inode = file_inode(file);
 938	loff_t			new_size = i_size_read(inode) - len;
 939	int			error;
 940
 941	if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
 942		return -EINVAL;
 943
 944	/*
 945	 * There is no need to overlap collapse range with EOF, in which case it
 946	 * is effectively a truncate operation
 947	 */
 948	if (offset + len >= i_size_read(inode))
 949		return -EINVAL;
 950
 951	error = xfs_collapse_file_space(XFS_I(inode), offset, len);
 952	if (error)
 953		return error;
 954	return xfs_falloc_setsize(file, new_size);
 955}
 956
 957static int
 958xfs_falloc_insert_range(
 959	struct file		*file,
 960	loff_t			offset,
 961	loff_t			len)
 962{
 963	struct inode		*inode = file_inode(file);
 964	loff_t			isize = i_size_read(inode);
 965	int			error;
 966
 967	if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
 968		return -EINVAL;
 969
 970	/*
 971	 * New inode size must not exceed ->s_maxbytes, accounting for
 972	 * possible signed overflow.
 973	 */
 974	if (inode->i_sb->s_maxbytes - isize < len)
 975		return -EFBIG;
 
 
 
 
 
 
 
 
 
 
 976
 977	/* Offset should be less than i_size */
 978	if (offset >= isize)
 979		return -EINVAL;
 980
 981	error = xfs_falloc_setsize(file, isize + len);
 982	if (error)
 983		return error;
 984
 985	/*
 986	 * Perform hole insertion now that the file size has been updated so
 987	 * that if we crash during the operation we don't leave shifted extents
 988	 * past EOF and hence losing access to the data that is contained within
 989	 * them.
 990	 */
 991	return xfs_insert_file_space(XFS_I(inode), offset, len);
 992}
 993
 994/*
 995 * Punch a hole and prealloc the range.  We use a hole punch rather than
 996 * unwritten extent conversion for two reasons:
 997 *
 998 *   1.) Hole punch handles partial block zeroing for us.
 999 *   2.) If prealloc returns ENOSPC, the file range is still zero-valued by
1000 *	 virtue of the hole punch.
1001 */
1002static int
1003xfs_falloc_zero_range(
1004	struct file		*file,
1005	int			mode,
1006	loff_t			offset,
1007	loff_t			len)
1008{
1009	struct inode		*inode = file_inode(file);
1010	unsigned int		blksize = i_blocksize(inode);
1011	loff_t			new_size = 0;
1012	int			error;
1013
1014	trace_xfs_zero_file_space(XFS_I(inode));
 
 
 
 
 
1015
1016	error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1017	if (error)
1018		return error;
1019
1020	error = xfs_free_file_space(XFS_I(inode), offset, len);
1021	if (error)
1022		return error;
1023
1024	len = round_up(offset + len, blksize) - round_down(offset, blksize);
1025	offset = round_down(offset, blksize);
1026	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1027	if (error)
1028		return error;
1029	return xfs_falloc_setsize(file, new_size);
1030}
1031
1032static int
1033xfs_falloc_unshare_range(
1034	struct file		*file,
1035	int			mode,
1036	loff_t			offset,
1037	loff_t			len)
1038{
1039	struct inode		*inode = file_inode(file);
1040	loff_t			new_size = 0;
1041	int			error;
1042
1043	error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1044	if (error)
1045		return error;
1046
1047	error = xfs_reflink_unshare(XFS_I(inode), offset, len);
1048	if (error)
1049		return error;
1050
1051	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1052	if (error)
1053		return error;
1054	return xfs_falloc_setsize(file, new_size);
1055}
1056
1057static int
1058xfs_falloc_allocate_range(
1059	struct file		*file,
1060	int			mode,
1061	loff_t			offset,
1062	loff_t			len)
1063{
1064	struct inode		*inode = file_inode(file);
1065	loff_t			new_size = 0;
1066	int			error;
1067
1068	/*
1069	 * If always_cow mode we can't use preallocations and thus should not
1070	 * create them.
1071	 */
1072	if (xfs_is_always_cow_inode(XFS_I(inode)))
1073		return -EOPNOTSUPP;
1074
1075	error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1076	if (error)
1077		return error;
1078
1079	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1080	if (error)
1081		return error;
1082	return xfs_falloc_setsize(file, new_size);
1083}
1084
1085#define	XFS_FALLOC_FL_SUPPORTED						\
1086		(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |		\
1087		 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE |	\
1088		 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
1089
1090STATIC long
1091xfs_file_fallocate(
1092	struct file		*file,
1093	int			mode,
1094	loff_t			offset,
1095	loff_t			len)
1096{
1097	struct inode		*inode = file_inode(file);
1098	struct xfs_inode	*ip = XFS_I(inode);
1099	long			error;
1100	uint			iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
 
 
1101
1102	if (!S_ISREG(inode->i_mode))
1103		return -EINVAL;
1104	if (mode & ~XFS_FALLOC_FL_SUPPORTED)
1105		return -EOPNOTSUPP;
1106
1107	xfs_ilock(ip, iolock);
1108	error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
1109	if (error)
1110		goto out_unlock;
1111
1112	/*
1113	 * Must wait for all AIO to complete before we continue as AIO can
1114	 * change the file size on completion without holding any locks we
1115	 * currently hold. We must do this first because AIO can update both
1116	 * the on disk and in memory inode sizes, and the operations that follow
1117	 * require the in-memory size to be fully up-to-date.
1118	 */
1119	inode_dio_wait(inode);
1120
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1121	error = file_modified(file);
1122	if (error)
1123		goto out_unlock;
1124
1125	switch (mode & FALLOC_FL_MODE_MASK) {
1126	case FALLOC_FL_PUNCH_HOLE:
1127		error = xfs_free_file_space(ip, offset, len);
1128		break;
1129	case FALLOC_FL_COLLAPSE_RANGE:
1130		error = xfs_falloc_collapse_range(file, offset, len);
1131		break;
1132	case FALLOC_FL_INSERT_RANGE:
1133		error = xfs_falloc_insert_range(file, offset, len);
1134		break;
1135	case FALLOC_FL_ZERO_RANGE:
1136		error = xfs_falloc_zero_range(file, mode, offset, len);
1137		break;
1138	case FALLOC_FL_UNSHARE_RANGE:
1139		error = xfs_falloc_unshare_range(file, mode, offset, len);
1140		break;
1141	case FALLOC_FL_ALLOCATE_RANGE:
1142		error = xfs_falloc_allocate_range(file, mode, offset, len);
1143		break;
1144	default:
1145		error = -EOPNOTSUPP;
1146		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1147	}
1148
1149	if (!error && xfs_file_sync_writes(file))
1150		error = xfs_log_force_inode(ip);
1151
1152out_unlock:
1153	xfs_iunlock(ip, iolock);
1154	return error;
1155}
1156
1157STATIC int
1158xfs_file_fadvise(
1159	struct file	*file,
1160	loff_t		start,
1161	loff_t		end,
1162	int		advice)
1163{
1164	struct xfs_inode *ip = XFS_I(file_inode(file));
1165	int ret;
1166	int lockflags = 0;
1167
1168	/*
1169	 * Operations creating pages in page cache need protection from hole
1170	 * punching and similar ops
1171	 */
1172	if (advice == POSIX_FADV_WILLNEED) {
1173		lockflags = XFS_IOLOCK_SHARED;
1174		xfs_ilock(ip, lockflags);
1175	}
1176	ret = generic_fadvise(file, start, end, advice);
1177	if (lockflags)
1178		xfs_iunlock(ip, lockflags);
1179	return ret;
1180}
1181
1182STATIC loff_t
1183xfs_file_remap_range(
1184	struct file		*file_in,
1185	loff_t			pos_in,
1186	struct file		*file_out,
1187	loff_t			pos_out,
1188	loff_t			len,
1189	unsigned int		remap_flags)
1190{
1191	struct inode		*inode_in = file_inode(file_in);
1192	struct xfs_inode	*src = XFS_I(inode_in);
1193	struct inode		*inode_out = file_inode(file_out);
1194	struct xfs_inode	*dest = XFS_I(inode_out);
1195	struct xfs_mount	*mp = src->i_mount;
1196	loff_t			remapped = 0;
1197	xfs_extlen_t		cowextsize;
1198	int			ret;
1199
1200	if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1201		return -EINVAL;
1202
1203	if (!xfs_has_reflink(mp))
1204		return -EOPNOTSUPP;
1205
1206	if (xfs_is_shutdown(mp))
1207		return -EIO;
1208
1209	/* Prepare and then clone file data. */
1210	ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1211			&len, remap_flags);
1212	if (ret || len == 0)
1213		return ret;
1214
1215	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1216
1217	ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1218			&remapped);
1219	if (ret)
1220		goto out_unlock;
1221
1222	/*
1223	 * Carry the cowextsize hint from src to dest if we're sharing the
1224	 * entire source file to the entire destination file, the source file
1225	 * has a cowextsize hint, and the destination file does not.
1226	 */
1227	cowextsize = 0;
1228	if (pos_in == 0 && len == i_size_read(inode_in) &&
1229	    (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1230	    pos_out == 0 && len >= i_size_read(inode_out) &&
1231	    !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1232		cowextsize = src->i_cowextsize;
1233
1234	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1235			remap_flags);
1236	if (ret)
1237		goto out_unlock;
1238
1239	if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1240		xfs_log_force_inode(dest);
1241out_unlock:
1242	xfs_iunlock2_remapping(src, dest);
1243	if (ret)
1244		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1245	/*
1246	 * If the caller did not set CAN_SHORTEN, then it is not prepared to
1247	 * handle partial results -- either the whole remap succeeds, or we
1248	 * must say why it did not.  In this case, any error should be returned
1249	 * to the caller.
1250	 */
1251	if (ret && remapped < len && !(remap_flags & REMAP_FILE_CAN_SHORTEN))
1252		return ret;
1253	return remapped > 0 ? remapped : ret;
1254}
1255
1256STATIC int
1257xfs_file_open(
1258	struct inode	*inode,
1259	struct file	*file)
1260{
1261	if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1262		return -EIO;
1263	file->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
1264	if (xfs_inode_can_atomicwrite(XFS_I(inode)))
1265		file->f_mode |= FMODE_CAN_ATOMIC_WRITE;
1266	return generic_file_open(inode, file);
1267}
1268
1269STATIC int
1270xfs_dir_open(
1271	struct inode	*inode,
1272	struct file	*file)
1273{
1274	struct xfs_inode *ip = XFS_I(inode);
1275	unsigned int	mode;
1276	int		error;
1277
1278	if (xfs_is_shutdown(ip->i_mount))
1279		return -EIO;
1280	error = generic_file_open(inode, file);
1281	if (error)
1282		return error;
1283
1284	/*
1285	 * If there are any blocks, read-ahead block 0 as we're almost
1286	 * certain to have the next operation be a read there.
1287	 */
1288	mode = xfs_ilock_data_map_shared(ip);
1289	if (ip->i_df.if_nextents > 0)
1290		error = xfs_dir3_data_readahead(ip, 0, 0);
1291	xfs_iunlock(ip, mode);
1292	return error;
1293}
1294
1295/*
1296 * Don't bother propagating errors.  We're just doing cleanup, and the caller
1297 * ignores the return value anyway.
1298 */
1299STATIC int
1300xfs_file_release(
1301	struct inode		*inode,
1302	struct file		*file)
1303{
1304	struct xfs_inode	*ip = XFS_I(inode);
1305	struct xfs_mount	*mp = ip->i_mount;
1306
1307	/*
1308	 * If this is a read-only mount or the file system has been shut down,
1309	 * don't generate I/O.
1310	 */
1311	if (xfs_is_readonly(mp) || xfs_is_shutdown(mp))
1312		return 0;
1313
1314	/*
1315	 * If we previously truncated this file and removed old data in the
1316	 * process, we want to initiate "early" writeout on the last close.
1317	 * This is an attempt to combat the notorious NULL files problem which
1318	 * is particularly noticeable from a truncate down, buffered (re-)write
1319	 * (delalloc), followed by a crash.  What we are effectively doing here
1320	 * is significantly reducing the time window where we'd otherwise be
1321	 * exposed to that problem.
1322	 */
1323	if (xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED)) {
1324		xfs_iflags_clear(ip, XFS_EOFBLOCKS_RELEASED);
1325		if (ip->i_delayed_blks > 0)
1326			filemap_flush(inode->i_mapping);
1327	}
1328
1329	/*
1330	 * XFS aggressively preallocates post-EOF space to generate contiguous
1331	 * allocations for writers that append to the end of the file.
1332	 *
1333	 * To support workloads that close and reopen the file frequently, these
1334	 * preallocations usually persist after a close unless it is the first
1335	 * close for the inode.  This is a tradeoff to generate tightly packed
1336	 * data layouts for unpacking tarballs or similar archives that write
1337	 * one file after another without going back to it while keeping the
1338	 * preallocation for files that have recurring open/write/close cycles.
1339	 *
1340	 * This heuristic is skipped for inodes with the append-only flag as
1341	 * that flag is rather pointless for inodes written only once.
1342	 *
1343	 * There is no point in freeing blocks here for open but unlinked files
1344	 * as they will be taken care of by the inactivation path soon.
1345	 *
1346	 * When releasing a read-only context, don't flush data or trim post-EOF
1347	 * blocks.  This avoids open/read/close workloads from removing EOF
1348	 * blocks that other writers depend upon to reduce fragmentation.
1349	 *
1350	 * If we can't get the iolock just skip truncating the blocks past EOF
1351	 * because we could deadlock with the mmap_lock otherwise. We'll get
1352	 * another chance to drop them once the last reference to the inode is
1353	 * dropped, so we'll never leak blocks permanently.
1354	 */
1355	if (inode->i_nlink &&
1356	    (file->f_mode & FMODE_WRITE) &&
1357	    !(ip->i_diflags & XFS_DIFLAG_APPEND) &&
1358	    !xfs_iflags_test(ip, XFS_EOFBLOCKS_RELEASED) &&
1359	    xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1360		if (xfs_can_free_eofblocks(ip) &&
1361		    !xfs_iflags_test_and_set(ip, XFS_EOFBLOCKS_RELEASED))
1362			xfs_free_eofblocks(ip);
1363		xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1364	}
1365
1366	return 0;
1367}
1368
1369STATIC int
1370xfs_file_readdir(
1371	struct file	*file,
1372	struct dir_context *ctx)
1373{
1374	struct inode	*inode = file_inode(file);
1375	xfs_inode_t	*ip = XFS_I(inode);
1376	size_t		bufsize;
1377
1378	/*
1379	 * The Linux API doesn't pass down the total size of the buffer
1380	 * we read into down to the filesystem.  With the filldir concept
1381	 * it's not needed for correct information, but the XFS dir2 leaf
1382	 * code wants an estimate of the buffer size to calculate it's
1383	 * readahead window and size the buffers used for mapping to
1384	 * physical blocks.
1385	 *
1386	 * Try to give it an estimate that's good enough, maybe at some
1387	 * point we can change the ->readdir prototype to include the
1388	 * buffer size.  For now we use the current glibc buffer size.
1389	 */
1390	bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1391
1392	return xfs_readdir(NULL, ip, ctx, bufsize);
1393}
1394
1395STATIC loff_t
1396xfs_file_llseek(
1397	struct file	*file,
1398	loff_t		offset,
1399	int		whence)
1400{
1401	struct inode		*inode = file->f_mapping->host;
1402
1403	if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1404		return -EIO;
1405
1406	switch (whence) {
1407	default:
1408		return generic_file_llseek(file, offset, whence);
1409	case SEEK_HOLE:
1410		offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1411		break;
1412	case SEEK_DATA:
1413		offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1414		break;
1415	}
1416
1417	if (offset < 0)
1418		return offset;
1419	return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1420}
1421
 
1422static inline vm_fault_t
1423xfs_dax_fault_locked(
1424	struct vm_fault		*vmf,
1425	unsigned int		order,
1426	bool			write_fault)
 
1427{
1428	vm_fault_t		ret;
1429	pfn_t			pfn;
1430
1431	if (!IS_ENABLED(CONFIG_FS_DAX)) {
1432		ASSERT(0);
1433		return VM_FAULT_SIGBUS;
1434	}
1435	ret = dax_iomap_fault(vmf, order, &pfn, NULL,
1436			(write_fault && !vmf->cow_page) ?
1437				&xfs_dax_write_iomap_ops :
1438				&xfs_read_iomap_ops);
1439	if (ret & VM_FAULT_NEEDDSYNC)
1440		ret = dax_finish_sync_fault(vmf, order, pfn);
1441	return ret;
1442}
1443
1444static vm_fault_t
1445xfs_dax_read_fault(
1446	struct vm_fault		*vmf,
1447	unsigned int		order)
 
 
1448{
1449	struct xfs_inode	*ip = XFS_I(file_inode(vmf->vma->vm_file));
1450	vm_fault_t		ret;
1451
1452	trace_xfs_read_fault(ip, order);
1453
1454	xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1455	ret = xfs_dax_fault_locked(vmf, order, false);
1456	xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1457
1458	return ret;
1459}
 
1460
1461/*
1462 * Locking for serialisation of IO during page faults. This results in a lock
1463 * ordering of:
1464 *
1465 * mmap_lock (MM)
1466 *   sb_start_pagefault(vfs, freeze)
1467 *     invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1468 *       page_lock (MM)
1469 *         i_lock (XFS - extent map serialisation)
1470 */
1471static vm_fault_t
1472xfs_write_fault(
1473	struct vm_fault		*vmf,
1474	unsigned int		order)
 
1475{
1476	struct inode		*inode = file_inode(vmf->vma->vm_file);
1477	struct xfs_inode	*ip = XFS_I(inode);
1478	unsigned int		lock_mode = XFS_MMAPLOCK_SHARED;
1479	vm_fault_t		ret;
1480
1481	trace_xfs_write_fault(ip, order);
1482
1483	sb_start_pagefault(inode->i_sb);
1484	file_update_time(vmf->vma->vm_file);
 
 
 
 
 
1485
1486	/*
1487	 * Normally we only need the shared mmaplock, but if a reflink remap is
1488	 * in progress we take the exclusive lock to wait for the remap to
1489	 * finish before taking a write fault.
1490	 */
1491	xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1492	if (xfs_iflags_test(ip, XFS_IREMAPPING)) {
1493		xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1494		xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1495		lock_mode = XFS_MMAPLOCK_EXCL;
 
 
 
 
1496	}
1497
1498	if (IS_DAX(inode))
1499		ret = xfs_dax_fault_locked(vmf, order, true);
1500	else
1501		ret = iomap_page_mkwrite(vmf, &xfs_buffered_write_iomap_ops);
1502	xfs_iunlock(ip, lock_mode);
1503
1504	sb_end_pagefault(inode->i_sb);
1505	return ret;
1506}
1507
1508static inline bool
1509xfs_is_write_fault(
1510	struct vm_fault		*vmf)
1511{
1512	return (vmf->flags & FAULT_FLAG_WRITE) &&
1513	       (vmf->vma->vm_flags & VM_SHARED);
1514}
1515
1516static vm_fault_t
1517xfs_filemap_fault(
1518	struct vm_fault		*vmf)
1519{
1520	struct inode		*inode = file_inode(vmf->vma->vm_file);
1521
1522	/* DAX can shortcut the normal fault path on write faults! */
1523	if (IS_DAX(inode)) {
1524		if (xfs_is_write_fault(vmf))
1525			return xfs_write_fault(vmf, 0);
1526		return xfs_dax_read_fault(vmf, 0);
1527	}
1528
1529	trace_xfs_read_fault(XFS_I(inode), 0);
1530	return filemap_fault(vmf);
1531}
1532
1533static vm_fault_t
1534xfs_filemap_huge_fault(
1535	struct vm_fault		*vmf,
1536	unsigned int		order)
1537{
1538	if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1539		return VM_FAULT_FALLBACK;
1540
1541	/* DAX can shortcut the normal fault path on write faults! */
1542	if (xfs_is_write_fault(vmf))
1543		return xfs_write_fault(vmf, order);
1544	return xfs_dax_read_fault(vmf, order);
1545}
1546
1547static vm_fault_t
1548xfs_filemap_page_mkwrite(
1549	struct vm_fault		*vmf)
1550{
1551	return xfs_write_fault(vmf, 0);
1552}
1553
1554/*
1555 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1556 * on write faults. In reality, it needs to serialise against truncate and
1557 * prepare memory for writing so handle is as standard write fault.
1558 */
1559static vm_fault_t
1560xfs_filemap_pfn_mkwrite(
1561	struct vm_fault		*vmf)
1562{
1563	return xfs_write_fault(vmf, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1564}
1565
1566static const struct vm_operations_struct xfs_file_vm_ops = {
1567	.fault		= xfs_filemap_fault,
1568	.huge_fault	= xfs_filemap_huge_fault,
1569	.map_pages	= filemap_map_pages,
1570	.page_mkwrite	= xfs_filemap_page_mkwrite,
1571	.pfn_mkwrite	= xfs_filemap_pfn_mkwrite,
1572};
1573
1574STATIC int
1575xfs_file_mmap(
1576	struct file		*file,
1577	struct vm_area_struct	*vma)
1578{
1579	struct inode		*inode = file_inode(file);
1580	struct xfs_buftarg	*target = xfs_inode_buftarg(XFS_I(inode));
1581
1582	/*
1583	 * We don't support synchronous mappings for non-DAX files and
1584	 * for DAX files if underneath dax_device is not synchronous.
1585	 */
1586	if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1587		return -EOPNOTSUPP;
1588
1589	file_accessed(file);
1590	vma->vm_ops = &xfs_file_vm_ops;
1591	if (IS_DAX(inode))
1592		vm_flags_set(vma, VM_HUGEPAGE);
1593	return 0;
1594}
1595
1596const struct file_operations xfs_file_operations = {
1597	.llseek		= xfs_file_llseek,
1598	.read_iter	= xfs_file_read_iter,
1599	.write_iter	= xfs_file_write_iter,
1600	.splice_read	= xfs_file_splice_read,
1601	.splice_write	= iter_file_splice_write,
1602	.iopoll		= iocb_bio_iopoll,
1603	.unlocked_ioctl	= xfs_file_ioctl,
1604#ifdef CONFIG_COMPAT
1605	.compat_ioctl	= xfs_file_compat_ioctl,
1606#endif
1607	.mmap		= xfs_file_mmap,
 
1608	.open		= xfs_file_open,
1609	.release	= xfs_file_release,
1610	.fsync		= xfs_file_fsync,
1611	.get_unmapped_area = thp_get_unmapped_area,
1612	.fallocate	= xfs_file_fallocate,
1613	.fadvise	= xfs_file_fadvise,
1614	.remap_file_range = xfs_file_remap_range,
1615	.fop_flags	= FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
1616			  FOP_BUFFER_WASYNC | FOP_DIO_PARALLEL_WRITE,
1617};
1618
1619const struct file_operations xfs_dir_file_operations = {
1620	.open		= xfs_dir_open,
1621	.read		= generic_read_dir,
1622	.iterate_shared	= xfs_file_readdir,
1623	.llseek		= generic_file_llseek,
1624	.unlocked_ioctl	= xfs_file_ioctl,
1625#ifdef CONFIG_COMPAT
1626	.compat_ioctl	= xfs_file_compat_ioctl,
1627#endif
1628	.fsync		= xfs_dir_fsync,
1629};