tag name | atomic-file-updates_2021-10-22 (70bb87907d5633fb21a5240da403318d06cf9dee) |
tag date | 2021-10-22 16:45:22 -0700 |
tagged by | Darrick J. Wong <djwong@kernel.org> |
tagged object | commit 78c276eee7... |
xfs: atomic file updates
This series creates a new FIEXCHANGE_RANGE system call to exchange
ranges of bytes between two files atomically. This new functionality
enables data storage programs to stage and commit file updates such that
reader programs will see either the old contents or the new contents in
their entirety, with no chance of torn writes. A successful call
completion guarantees that the new contents will be seen even if the
system fails.
User programs will be able to update files atomically by opening an
O_TMPFILE, reflinking the source file to it, making whatever updates
they want to make, and exchange the relevant ranges of the temp file
with the original file. If the updates are aligned with the file block
size, a new (since v2) flag provides for exchanging only the written
areas. Callers can arrange for the update to be rejected if the
original file has been changed.
The intent behind this new userspace functionality is to enable atomic
rewrites of arbitrary parts of individual files. For years, application
programmers wanting to ensure the atomicity of a file update had to
write the changes to a new file in the same directory, fsync the new
file, rename the new file on top of the old filename, and then fsync the
directory. People get it wrong all the time, and $fs hacks abound.
Here is the proposed manual page:
IOCTL-FIEXCHANGE_RANGE(Linux Programmer's ManIOCTL-FIEXCHANGE_RANGE(2)
NAME
ioctl_fiexchange_range - exchange the contents of parts of two
files
SYNOPSIS
#include <sys/ioctl.h>
#include <linux/fiexchange.h>
int ioctl(int file2_fd, FIEXCHANGE_RANGE, struct
file_xchg_range *arg);
DESCRIPTION
Given a range of bytes in a first file file1_fd and a second
range of bytes in a second file file2_fd, this ioctl(2) ex‐
changes the contents of the two ranges.
Exchanges are atomic with regards to concurrent file opera‐
tions, so no userspace-level locks need to be taken to obtain
consistent results. Implementations must guarantee that read‐
ers see either the old contents or the new contents in their
entirety, even if the system fails.
The exchange parameters are conveyed in a structure of the fol‐
lowing form:
struct file_xchg_range {
__s64 file1_fd;
__s64 file1_offset;
__s64 file2_offset;
__s64 length;
__u64 flags;
__s64 file2_ino;
__s64 file2_mtime;
__s64 file2_ctime;
__s32 file2_mtime_nsec;
__s32 file2_ctime_nsec;
__u64 pad[6];
};
The field pad must be zero.
The fields file1_fd, file1_offset, and length define the first
range of bytes to be exchanged.
The fields file2_fd, file2_offset, and length define the second
range of bytes to be exchanged.
Both files must be from the same filesystem mount. If the two
file descriptors represent the same file, the byte ranges must
not overlap. Most disk-based filesystems require that the
starts of both ranges must be aligned to the file block size.
If this is the case, the ends of the ranges must also be so
aligned unless the FILE_XCHG_RANGE_TO_EOF flag is set.
The field flags control the behavior of the exchange operation.
FILE_XCHG_RANGE_FILE2_FRESH
Check the freshness of file2_fd after locking the
file but before exchanging the contents. The sup‐
plied file2_ino field must match file2's inode num‐
ber, and the supplied file2_mtime, file2_mtime_nsec,
file2_ctime, and file2_ctime_nsec fields must match
the modification time and change time of file2. If
they do not match, EBUSY will be returned.
FILE_XCHG_RANGE_TO_EOF
Ignore the length parameter. All bytes in file1_fd
from file1_offset to EOF are moved to file2_fd, and
file2's size is set to (file2_offset+(file1_length-
file1_offset)). Meanwhile, all bytes in file2 from
file2_offset to EOF are moved to file1 and file1's
size is set to (file1_offset+(file2_length-
file2_offset)). This option is not compatible with
FILE_XCHG_RANGE_FULL_FILES.
FILE_XCHG_RANGE_FSYNC
Ensure that all modified in-core data in both file
ranges and all metadata updates pertaining to the
exchange operation are flushed to persistent storage
before the call returns. Opening either file de‐
scriptor with O_SYNC or O_DSYNC will have the same
effect.
FILE_XCHG_RANGE_SKIP_FILE1_HOLES
Skip sub-ranges of file1_fd that are known not to
contain data. This facility can be used to imple‐
ment atomic scatter-gather writes of any complexity
for software-defined storage targets.
FILE_XCHG_RANGE_DRY_RUN
Check the parameters and the feasibility of the op‐
eration, but do not change anything.
FILE_XCHG_RANGE_COMMIT
This flag is a combination of
FILE_XCHG_RANGE_FILE2_FRESH | FILE_XCHG_RANGE_FSYNC
and can be used to commit changes to file2_fd to
persistent storage if and only if file2 has not
changed.
FILE_XCHG_RANGE_FULL_FILES
Require that file1_offset and file2_offset are zero,
and that the length field matches the lengths of
both files. If not, EDOM will be returned. This
option is not compatible with
FILE_XCHG_RANGE_TO_EOF.
FILE_XCHG_RANGE_NONATOMIC
This flag relaxes the requirement that readers see
only the old contents or the new contents in their
entirety. If the system fails before all modified
in-core data and metadata updates are persisted to
disk, the contents of both file ranges after recov‐
ery are not defined and may be a mix of both.
Do not use this flag unless the contents of both
ranges are known to be identical and there are no
other writers.
RETURN VALUE
On error, -1 is returned, and errno is set to indicate the er‐
ror.
ERRORS
Error codes can be one of, but are not limited to, the follow‐
ing:
EBADF file1_fd is not open for reading and writing or is open
for append-only writes; or file2_fd is not open for
reading and writing or is open for append-only writes.
EBUSY The inode number and timestamps supplied do not match
file2_fd and FILE_XCHG_RANGE_FILE2_FRESH was set in
flags.
EDOM The ranges do not cover the entirety of both files, and
FILE_XCHG_RANGE_FULL_FILES was set in flags.
EINVAL The parameters are not correct for these files. This
error can also appear if either file descriptor repre‐
sents a device, FIFO, or socket. Disk filesystems gen‐
erally require the offset and length arguments to be
aligned to the fundamental block sizes of both files.
EIO An I/O error occurred.
EISDIR One of the files is a directory.
ENOMEM The kernel was unable to allocate sufficient memory to
perform the operation.
ENOSPC There is not enough free space in the filesystem ex‐
change the contents safely.
EOPNOTSUPP
The filesystem does not support exchanging bytes between
the two files.
EPERM file1_fd or file2_fd are immutable.
ETXTBSY
One of the files is a swap file.
EUCLEAN
The filesystem is corrupt.
EXDEV file1_fd and file2_fd are not on the same mounted
filesystem.
CONFORMING TO
This API is Linux-specific.
USE CASES
Three use cases are imagined for this system call.
The first is a filesystem defragmenter, which copies the con‐
tents of a file into another file and wishes to exchange the
space mappings of the two files, provided that the original
file has not changed. The flags NONATOMIC and FILE2_FRESH are
recommended for this application.
The second is a data storage program that wants to commit non-
contiguous updates to a file atomically. This can be done by
creating a temporary file, calling FICLONE(2) to share the con‐
tents, and staging the updates into the temporary file. Either
of the FULL_FILES or TO_EOF flags are recommended, along with
FSYNC. Depending on the application's locking design, the
flags FILE2_FRESH or COMMIT may be applicable here. The tempo‐
rary file can be deleted or punched out afterwards.
The third is a software-defined storage host (e.g. a disk juke‐
box) which implements an atomic scatter-gather write command.
Provided the exported disk's logical block size matches the
file's allocation unit size, this can be done by creating a
temporary file and writing the data at the appropriate offsets.
Use this call with the SKIP_HOLES flag to exchange only the
blocks involved in the write command. The use of the FSYNC
flag is recommended here. The temporary file should be deleted
or punched out completely before being reused to stage another
write.
NOTES
Some filesystems may limit the amount of data or the number of
extents that can be exchanged in a single call.
SEE ALSO
ioctl(2)
Linux 2021-04-01 IOCTL-FIEXCHANGE_RANGE(2)
The reference implementation in XFS creates a new log incompat feature
and log intent items to track high level progress of swapping ranges of
two files and finish interrupted work if the system goes down. Sample
code can be found in the corresponding changes to xfs_io to exercise the
use case mentioned above.
Note that this function is /not/ the O_DIRECT atomic file writes concept
that has also been floating around for years. This RFC is constructed
entirely in software, which means that there are no limitations other
than the general filesystem limits.
As a side note, the original motivation behind the kernel functionality
is online repair of file-based metadata. The atomic file swap is
implemented as an atomic inode fork swap, which means that we can
implement online reconstruction of extended attributes and directories
by building a new one in another inode and atomically swap the contents.
Subsequent patchsets adapt the online filesystem repair code to use
atomic extent swapping. This enables repair functions to construct a
clean copy of a directory, xattr information, realtime bitmaps, and
realtime summary information in a temporary inode. If this completes
successfully, the new contents can be swapped atomically into the inode
being repaired. This is essential to avoid making corruption problems
worse if the system goes down in the middle of running repair.
-----BEGIN PGP SIGNATURE-----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=b28O
-----END PGP SIGNATURE-----