Testcase from Giel de Nijs <[email protected]>
on linux-ext4 list, ""Possible ext4 data corruption
with large files and async I/O," on 29 Jan 2010
ext4 put byte offsets in a block offset u32 container
in the endio struct, so 4g wrapped to 0 leading to
data corruption when the unwritten extent did not
get converted.
Signed-off-by: Eric Sandeen <[email protected]>
---
diff --git a/224 b/224
new file mode 100755
index 0000000..47e614c
--- /dev/null
+++ b/224
@@ -0,0 +1,68 @@
+#! /bin/bash
+# FS QA Test No. 224
+#
+# Test aio hole-filling past 2^32 bytes
+#
+# Testcase from Giel de Nijs <[email protected]>
+# on linux-ext4 list, ""Possible ext4 data corruption
+# with large files and async I/O," on 29 Jan 2010
+#
+#-----------------------------------------------------------------------
+# Copyright (c) 2010 Eric Sandeen. All Rights Reserved.
+#
+# This program is free software; you can redistribute it and/or
+# modify it under the terms of the GNU General Public License as
+# published by the Free Software Foundation.
+#
+# This program is distributed in the hope that it would be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program; if not, write the Free Software Foundation,
+# Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
+#
+#-----------------------------------------------------------------------
+#
+# creator
[email protected]
+
+seq=`basename $0`
+echo "QA output created by $seq"
+
+here=`pwd`
+tmp=/tmp/$$
+status=1 # failure is the default!
+
+_cleanup()
+{
+ rm -f $tmp.*
+}
+
+trap "_cleanup ; exit \$status" 0 1 2 3 15
+
+# get standard environment, filters and checks
+. ./common.rc
+. ./common.filter
+
+# real QA test starts here
+_supported_fs generic
+_supported_os Linux
+
+[ ! -x $here/src/aio-write ] && _notrun "aio-write not built"
+
+TESTFILE=$TEST_DIR/$seq-aio-write-testfile
+
+rm -f $TESTFILE
+$here/src/aio-write -v -s 0xab -o 6g -l 512k $TESTFILE
+$here/src/aio-write -v -s 0xcd -o 4g -l 512k $TESTFILE
+
+# Bug in ext4 caused the 4g offset to wrap and io
+# completion happened at 0, leaving extent unwritten,
+# and 0s returned
+
+$XFS_IO_PROG -F -c "pread -v 4g 512k" -c "pread -v 6g 512k" $TESTFILE | \
+ _filter_xfs_io_unique
+
+status=0 ; exit
diff --git a/224.out b/224.out
new file mode 100644
index 0000000..4bb51d5
--- /dev/null
+++ b/224.out
@@ -0,0 +1,19 @@
+QA output created by 224
+opening file /mnt/test/224-aio-write-testfile
+submitting write of 524288 bytes at offset 6442450944
+waiting for write to be finished
+got 1 events
+written 524288 bytes
+opening file /mnt/test/224-aio-write-testfile
+submitting write of 524288 bytes at offset 4294967296
+waiting for write to be finished
+got 1 events
+written 524288 bytes
+100000000: cd cd cd cd cd cd cd cd cd cd cd cd cd cd cd cd ................
+*
+read 524288/524288 bytes at offset 4294967296
+XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
+180000000: ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ab ................
+*
+read 524288/524288 bytes at offset 6442450944
+XXX Bytes, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
diff --git a/group b/group
index c66d965..46acc2b 100644
--- a/group
+++ b/group
@@ -336,4 +336,5 @@ deprecated
220 auto quota quick
221 auto metadata quick
222 auto fsr ioctl quick
+224 auto aio quick
223 auto quick
diff --git a/src/Makefile b/src/Makefile
index 619a752..2ac7041 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -42,6 +42,8 @@ LLDLIBS += $(LIBGDBM)
endif
ifeq ($(HAVE_AIO), true)
+LINUX_TARGETS += aio-write
+LLDLIBS += -laio
SUBDIRS += aio-dio-regress
endif
diff --git a/src/aio-write.c b/src/aio-write.c
new file mode 100644
index 0000000..59fbb24
--- /dev/null
+++ b/src/aio-write.c
@@ -0,0 +1,174 @@
+/*
+ * Author: Giel de Nijs, VectorWise B.V. <[email protected]>
+ *
+ * perform an AIO write to a file at a given length & offset
+ * Compile with -laio
+ *
+ * Copyright (c) 2010 Giel de Nijs, VectorWise B.V. <[email protected]>
+ * Modifications Copyright (c) Eric Sandeen <[email protected]>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2, or (at your option)
+ * any later version.
+
+ */
+
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+#define _LARGEFILE64_SOURCE
+#define _FILE_OFFSET_BITS 64
+#include <features.h>
+#include <libaio.h>
+#include <stdio.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <error.h>
+#include <errno.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <ctype.h>
+#include <limits.h>
+
+void usage(void)
+{
+ fprintf(stderr, "usage: aio_write [-v] [-s seed]"
+ " -o offset -l length filename\n");
+ exit(1);
+}
+
+/*
+ * Scale value by kilo, mega, or giga.
+ */
+loff_t scale_by_kmg(long long value, char scale)
+{
+switch (scale) {
+ case 'g':
+ case 'G':
+ value *= 1024;
+ /* fallthrough */
+ case 'm':
+ case 'M':
+ value *= 1024;
+ /* fallthrough */
+ case 'k':
+ case 'K':
+ value *= 1024;
+ break;
+ case '\0':
+ break;
+ default:
+ usage();
+ break;
+ }
+ return value;
+}
+
+int main(int argc, char ** argv)
+{
+ char filename[PATH_MAX];
+ loff_t offset = 0;
+ size_t length = 0;
+ int seed = 0xFF;
+ int queue_depth = 8;
+ int err;
+ char *buf;
+ io_context_t io_ctx;
+ struct iocb iocb;
+ struct iocb *iocblist[1];
+ struct io_event events[1];
+ int fd;
+ int c;
+ int verbose = 0;
+ extern char *optarg;
+ extern int optind, optopt, opterr;
+
+ if (argc < 4)
+ usage();
+
+ while ((c = getopt(argc, argv, "s:o:l:v")) != -1) {
+ char *endp;
+ switch (c) {
+ case 's':
+ seed = (int)strtol(optarg, &endp, 0);
+ break;
+ case 'o':
+ offset = strtol(optarg, &endp, 0);
+ offset = scale_by_kmg((long long)offset, *endp);
+ break;
+ case 'l':
+ length = strtol(optarg, &endp, 0);
+ length = scale_by_kmg((long long)length, *endp);
+ break;
+ case 'v':
+ verbose++;
+ break;
+ case '?':
+ usage();
+ break;
+ }
+ }
+
+ strncpy(filename, argv[argc-1], PATH_MAX);
+
+ /* allocate aligned memory (for direct i/o) */
+ err = posix_memalign((void **)&buf, getpagesize(), length);
+ if (err) {
+ printf("error allocating memory: %s\n", strerror(err));
+ return(err);
+ }
+ memset(buf, seed, length);
+
+ /* initialize async i/o */
+ err = io_queue_init(queue_depth, &io_ctx);
+ if (err < 0) {
+ printf("error initializing I/O queue: %s\n", strerror(-err));
+ return(-err);
+ }
+
+ /* create file */
+ if (verbose)
+ printf("opening file %s\n", filename);
+ fd = open(filename, O_DIRECT|O_RDWR|O_LARGEFILE|O_CREAT, 0666);
+ if (fd < 0) {
+ perror("error opening file");
+ return(errno);
+ }
+
+ /* write buf at offset */
+ io_prep_pwrite(&iocb, fd, buf, length, offset);
+ iocblist[0] = &iocb;
+ if (verbose)
+ printf("submitting write of %zd bytes at offset %zd\n", length, offset);
+ err = io_submit(io_ctx, 1, iocblist);
+ if (err < 0) {
+ printf("error submitting I/O requests: %s\n", strerror(-err));
+ return(-err);
+ }
+
+ if (verbose)
+ printf("waiting for write to be finished\n");
+ err = io_getevents(io_ctx, 1, 1, events, NULL);
+ if (err < 0) {
+ printf("error getting I/O events: %s\n", strerror(-err));
+ return(-err);
+ }
+ if (verbose)
+ printf("got %d events\n", err);
+ err = events[0].res;
+ if (err < 0) {
+ printf("error writing buffer: %s\n", strerror(-err));
+ return(-err);
+ }
+ if (verbose)
+ printf("written %ld bytes\n", events[0].res);
+
+ close(fd);
+ io_destroy(io_ctx);
+
+ free(buf);
+ return 0;
+}
+
On Fri, Jan 29, 2010 at 02:05:46PM -0600, Eric Sandeen wrote:
> Testcase from Giel de Nijs <[email protected]>
> on linux-ext4 list, ""Possible ext4 data corruption
> with large files and async I/O," on 29 Jan 2010
>
> ext4 put byte offsets in a block offset u32 container
> in the endio struct, so 4g wrapped to 0 leading to
> data corruption when the unwritten extent did not
> get converted.
There's various type messups in the test program that make it fail for
me on a 32-bit machine. The patch below fixes it up, but it seems like
we should rather add a variant of that code as aio_read/write commands
to xfs_io instead of adding a new test program.
Index: xfstests-dev/src/aio-write.c
===================================================================
--- xfstests-dev.orig/src/aio-write.c 2010-01-30 10:42:24.000000000 +0000
+++ xfstests-dev/src/aio-write.c 2010-01-30 10:45:30.000000000 +0000
@@ -42,7 +42,7 @@ void usage(void)
/*
* Scale value by kilo, mega, or giga.
*/
-loff_t scale_by_kmg(long long value, char scale)
+long long scale_by_kmg(long long value, char scale)
{
switch (scale) {
case 'g':
@@ -69,7 +69,7 @@ switch (scale) {
int main(int argc, char ** argv)
{
char filename[PATH_MAX];
- loff_t offset = 0;
+ long long offset = 0;
size_t length = 0;
int seed = 0xFF;
int queue_depth = 8;
@@ -95,12 +95,12 @@ int main(int argc, char ** argv)
seed = (int)strtol(optarg, &endp, 0);
break;
case 'o':
- offset = strtol(optarg, &endp, 0);
- offset = scale_by_kmg((long long)offset, *endp);
+ offset = strtoll(optarg, &endp, 0);
+ offset = scale_by_kmg(offset, *endp);
break;
case 'l':
length = strtol(optarg, &endp, 0);
- length = scale_by_kmg((long long)length, *endp);
+ length = scale_by_kmg(length, *endp);
break;
case 'v':
verbose++;
@@ -141,7 +141,8 @@ int main(int argc, char ** argv)
io_prep_pwrite(&iocb, fd, buf, length, offset);
iocblist[0] = &iocb;
if (verbose)
- printf("submitting write of %zd bytes at offset %zd\n", length, offset);
+ printf("submitting write of %zd bytes at offset %lld\n",
+ length, offset);
err = io_submit(io_ctx, 1, iocblist);
if (err < 0) {
printf("error submitting I/O requests: %s\n", strerror(-err));
Christoph Hellwig wrote:
> On Fri, Jan 29, 2010 at 02:05:46PM -0600, Eric Sandeen wrote:
>> Testcase from Giel de Nijs <[email protected]>
>> on linux-ext4 list, ""Possible ext4 data corruption
>> with large files and async I/O," on 29 Jan 2010
>>
>> ext4 put byte offsets in a block offset u32 container
>> in the endio struct, so 4g wrapped to 0 leading to
>> data corruption when the unwritten extent did not
>> get converted.
>
> There's various type messups in the test program that make it fail for
oops, thanks for checking.
> me on a 32-bit machine. The patch below fixes it up, but it seems like
> we should rather add a variant of that code as aio_read/write commands
> to xfs_io instead of adding a new test program.
ok, that's probably better - again, though, it takes at least a release
cycle before most folks can test it. But I guess that's not the end of
the world.
-Eric
> Index: xfstests-dev/src/aio-write.c
> ===================================================================
> --- xfstests-dev.orig/src/aio-write.c 2010-01-30 10:42:24.000000000 +0000
> +++ xfstests-dev/src/aio-write.c 2010-01-30 10:45:30.000000000 +0000
> @@ -42,7 +42,7 @@ void usage(void)
> /*
> * Scale value by kilo, mega, or giga.
> */
> -loff_t scale_by_kmg(long long value, char scale)
> +long long scale_by_kmg(long long value, char scale)
> {
> switch (scale) {
> case 'g':
> @@ -69,7 +69,7 @@ switch (scale) {
> int main(int argc, char ** argv)
> {
> char filename[PATH_MAX];
> - loff_t offset = 0;
> + long long offset = 0;
> size_t length = 0;
> int seed = 0xFF;
> int queue_depth = 8;
> @@ -95,12 +95,12 @@ int main(int argc, char ** argv)
> seed = (int)strtol(optarg, &endp, 0);
> break;
> case 'o':
> - offset = strtol(optarg, &endp, 0);
> - offset = scale_by_kmg((long long)offset, *endp);
> + offset = strtoll(optarg, &endp, 0);
> + offset = scale_by_kmg(offset, *endp);
> break;
> case 'l':
> length = strtol(optarg, &endp, 0);
> - length = scale_by_kmg((long long)length, *endp);
> + length = scale_by_kmg(length, *endp);
> break;
> case 'v':
> verbose++;
> @@ -141,7 +141,8 @@ int main(int argc, char ** argv)
> io_prep_pwrite(&iocb, fd, buf, length, offset);
> iocblist[0] = &iocb;
> if (verbose)
> - printf("submitting write of %zd bytes at offset %zd\n", length, offset);
> + printf("submitting write of %zd bytes at offset %lld\n",
> + length, offset);
> err = io_submit(io_ctx, 1, iocblist);
> if (err < 0) {
> printf("error submitting I/O requests: %s\n", strerror(-err));
>
On Sat, Jan 30, 2010 at 10:15:09AM -0600, Eric Sandeen wrote:
> > me on a 32-bit machine. The patch below fixes it up, but it seems like
> > we should rather add a variant of that code as aio_read/write commands
> > to xfs_io instead of adding a new test program.
>
> ok, that's probably better - again, though, it takes at least a release
> cycle before most folks can test it. But I guess that's not the end of
> the world.
Stupid question --- who uses xfs_io besides xfstests? Any chance we
could consider dropping in some version of xfs_io into xfstests, or
actually moving it into xfstests from xfsprogs if xfstests is the
exclusive user of that program? I've been trying to get more people
to use xfstests, since it would be good if more companies and more
projects were using it --- and one of the things that makes it hard is
all of the dependencies that it has. If there was some way we could
gradually make xfstests more self-contained, it would certainly be
nice.
- Ted
[email protected] wrote:
> On Sat, Jan 30, 2010 at 10:15:09AM -0600, Eric Sandeen wrote:
>>> me on a 32-bit machine. The patch below fixes it up, but it seems like
>>> we should rather add a variant of that code as aio_read/write commands
>>> to xfs_io instead of adding a new test program.
>> ok, that's probably better - again, though, it takes at least a release
>> cycle before most folks can test it. But I guess that's not the end of
>> the world.
>
> Stupid question --- who uses xfs_io besides xfstests? Any chance we
> could consider dropping in some version of xfs_io into xfstests, or
> actually moving it into xfstests from xfsprogs if xfstests is the
> exclusive user of that program? I've been trying to get more people
> to use xfstests, since it would be good if more companies and more
> projects were using it --- and one of the things that makes it hard is
> all of the dependencies that it has. If there was some way we could
> gradually make xfstests more self-contained, it would certainly be
> nice.
>
> - Ted
These are the deps that I know xfstests has, to build and to run:
BuildRequires: autoconf, libtool, xfsprogs-devel, e2fsprogs-devel
BuildRequires: libacl-devel, libattr-devel, libaio-devel
Requires: bash, xfsprogs, xfsdump, perl, acl, attr, bind-utils
Requires: bc, indent, quota
which isn't so bad...
(and tests are just skipped if xfsdump isn't there)
I'm not sure an xfsprogs dependency is so onerous; plenty has depended
on e2fsprogs through the years and we've lived with that ;) But
the lag time for xfsprogs to use released xfs_io functionality is a
bit of a bummer.
But I guess I don't have a great answer for who else uses xfs_io:
xfs_io(8) xfs_io(8)
NAME
xfs_io - debug the I/O path of an XFS filesystem
SYNOPSIS
xfs_io [ -adFfmrRstx ] [ -c cmd ] ... [ -p prog ] file
DESCRIPTION
xfs_io is a debugging tool like xfs_db(8), but is aimed
at examining the regular file I/O paths rather than the
raw XFS volume itself.
I guess it's not really advertised as a generic tool, and it's
in the sbin path...
I guess I could live with it either way - I suppose my main concern
is that xfstests is a mess to package for a distro, and I really like
easy access to xfs_io for my own use outside of xfstests.
-Eric
On Sat, Jan 30, 2010 at 12:31:26PM -0600, Eric Sandeen wrote:
> These are the deps that I know xfstests has, to build and to run:
>
> BuildRequires: autoconf, libtool, xfsprogs-devel, e2fsprogs-devel
> BuildRequires: libacl-devel, libattr-devel, libaio-devel
>
> Requires: bash, xfsprogs, xfsdump, perl, acl, attr, bind-utils
> Requires: bc, indent, quota
>
> which isn't so bad...
Doesn't seem to bad. Indent is afaik only needed for the weird 122 test
which doesn't apply to non-xfs filesystems.
> I'm not sure an xfsprogs dependency is so onerous; plenty has depended
> on e2fsprogs through the years and we've lived with that ;) But
> the lag time for xfsprogs to use released xfs_io functionality is a
> bit of a bummer.
>
> But I guess I don't have a great answer for who else uses xfs_io:
I use xfs_io in lots various local scripts. It's a really handly
tool for exercising some of the more weird I/O related syscalls.
Christoph Hellwig wrote:
> On Sat, Jan 30, 2010 at 12:31:26PM -0600, Eric Sandeen wrote:
>> These are the deps that I know xfstests has, to build and to run:
>>
>> BuildRequires: autoconf, libtool, xfsprogs-devel, e2fsprogs-devel
>> BuildRequires: libacl-devel, libattr-devel, libaio-devel
>>
>> Requires: bash, xfsprogs, xfsdump, perl, acl, attr, bind-utils
>> Requires: bc, indent, quota
>>
>> which isn't so bad...
>
> Doesn't seem to bad. Indent is afaik only needed for the weird 122 test
> which doesn't apply to non-xfs filesystems.
and FWIW, we do:
_require_command /usr/bin/indent
so it'll just not run if it's not there (the above was for an rpm
attempt I made, wishing to automatically pull in everything that
might possibly be needed.)
-Eric
>> I'm not sure an xfsprogs dependency is so onerous; plenty has depended
>> on e2fsprogs through the years and we've lived with that ;) But
>> the lag time for xfsprogs to use released xfs_io functionality is a
>> bit of a bummer.
>>
>> But I guess I don't have a great answer for who else uses xfs_io:
>
> I use xfs_io in lots various local scripts. It's a really handly
> tool for exercising some of the more weird I/O related syscalls.
>