This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
My /tmp is an ext3 filesystem, in case that matters.
Here is the output I get from running the program on a broken kernel:
Compiling test program
Making original file /tmp/writevtest.yzafRmFCOR/test.in
..checking original file's md5sum.
Running test to copy to /tmp/writevtest.yzafRmFCOR/test.out
..checking new file's md5sum.
Attempting to drop the page cache for this file...
..checking new file's md5sum again.
MD5SUM MISMATCH(/tmp/writevtest.yzafRmFCOR/test.out):
wanted 2fdd6851b32ae931637d4845c037b550
got 67e5e2d6d4435e8095335d86a3d3e993
(please CC responses to me, I'm not subscribed to this list).
Thanks,
James
On Nov 30, 2009, at 3:55 PM, James Y Knight wrote:
> This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
>
> The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
>
> So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
>
> This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
>
> My /tmp is an ext3 filesystem, in case that matters.
Further testing shows that the filesystem type *does* matter. The bug does not exhibit when the test is run on ext2, ext4, vfat, btrfs, jfs, or xfs (and probably all the others too). Only, so far as I can determine, on ext3.
Thanks,
James-
On Tue 01-12-09 12:42:45, Mike Galbraith wrote:
> On Mon, 2009-11-30 at 19:48 -0500, James Y Knight wrote:
> > On Nov 30, 2009, at 3:55 PM, James Y Knight wrote:
> >
> > > This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
> > >
> > > The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
> > >
> > > So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
> > >
> > > This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
> > >
> > > My /tmp is an ext3 filesystem, in case that matters.
> >
> > Further testing shows that the filesystem type *does* matter. The bug does not exhibit when the test is run on ext2, ext4, vfat, btrfs, jfs, or xfs (and probably all the others too). Only, so far as I can determine, on ext3.
>
> I bisected it this morning. Bisected cleanly to...
>
> 9eaaa2d5759837402ec5eee13b2a97921808c3eb is the first bad commit
> commit 9eaaa2d5759837402ec5eee13b2a97921808c3eb
> Author: Jan Kara <[email protected]>
> Date: Mon Jul 13 20:26:52 2009 +0200
>
> ext3: Fix truncation of symlinks after failed write
>
> Contents of long symlinks is written via standard write methods. So when the
> write fails, we add inode to orphan list. But symlinks don't have .truncate
> method defined so nobody properly removes them from the orphan list (both on
> disk and in memory).
>
> Fix this by calling ext3_truncate() directly instead of calling vmtruncate()
> (which is saner anyway since we don't need anything vmtruncate() does except
> from calling .truncate in these paths). We also add inode to orphan list only
> if ext3_can_truncate() is true (currently, it can be false for symlinks when
> there are no blocks allocated) - otherwise orphan list processing will complain
> and ext3_truncate() will not remove inode from on-disk orphan list.
>
> Signed-off-by: Jan Kara <[email protected]>
>
> Reverting that in 31.6 (two revert/apply cycles) cured it (which doesn't
> look right at a glance at changelog, but.. shrug). Doing the same
> in .git does not cure it, so either there's a part two, or something
> went wonky. I'll probably try to bisect part two, but would appreciate
> a verification before maybe wasting more time.
Huh, I don't see how that's connected either but OTOH it's touching write
path so it's probably some strange interaction. Anyway, I see it on my
machine as well so I'm investigating. Thanks for CCing me.
Honza
On Tue 01-12-09 12:42:45, Mike Galbraith wrote:
> On Mon, 2009-11-30 at 19:48 -0500, James Y Knight wrote:
> > On Nov 30, 2009, at 3:55 PM, James Y Knight wrote:
> >
> > > This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
> > >
> > > The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
> > >
> > > So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
> > >
> > > This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
> > >
> > > My /tmp is an ext3 filesystem, in case that matters.
> >
> > Further testing shows that the filesystem type *does* matter. The bug does not exhibit when the test is run on ext2, ext4, vfat, btrfs, jfs, or xfs (and probably all the others too). Only, so far as I can determine, on ext3.
>
> I bisected it this morning. Bisected cleanly to...
>
> 9eaaa2d5759837402ec5eee13b2a97921808c3eb is the first bad commit
OK, I've debugged it. This commit is really at fault. The problem is
following:
When using writev, the page we copy from is not paged in (while when we
use ordinary write, it is paged in). This difference might be worth
investigation on its own (as it is likely to heavily impact performance of
writev) but is irrelevant for us now - we should handle this without data
corruption anyway. Because the source page is not available, we pass 0 as
the number of copied bytes to write_end and thus ext3_write_end decides to
truncate the file to original size. This is perfectly fine. The problem is
that we do this by ext3_truncate() which just frees corresponding block but
does not unmap buffers. So we leave mapped buffers beyond i_size (they
actually never were inside i_size) but the blocks they are mapped to are
already free. The write is then retried (after mapping the page),
block_write_begin() sees the buffer is mapped (although it is beyond
i_size) and thus it does not call ext3_get_block() anymore. So as a result,
data is written to a block that is no longer allocated to the file. Bummer
- welcome filesystem corruption.
Ext4 also has this problem but delayed allocation mitigates the effect to
an error in accounting of blocks reserved for delayed allocation and thus
under normal circumstances nothing bad happens.
The question is how to solve this in the cleanest way. We can call
vmtruncate() instead of ext3_truncate() as we used to do but Nick wants to
get rid of that (that's why I originally changed the code to what it is
now). So probably we could just manually call truncate_pagecache() instead.
Nick, I think your truncate calling sequence patch set needs similar fix
for all filesystems as well.
Honza
On Tue 01-12-09 15:35:59, Jan Kara wrote:
> On Tue 01-12-09 12:42:45, Mike Galbraith wrote:
> > On Mon, 2009-11-30 at 19:48 -0500, James Y Knight wrote:
> > > On Nov 30, 2009, at 3:55 PM, James Y Knight wrote:
> > >
> > > > This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
> > > >
> > > > The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
> > > >
> > > > So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
> > > >
> > > > This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
> > > >
> > > > My /tmp is an ext3 filesystem, in case that matters.
> > >
> > > Further testing shows that the filesystem type *does* matter. The bug does not exhibit when the test is run on ext2, ext4, vfat, btrfs, jfs, or xfs (and probably all the others too). Only, so far as I can determine, on ext3.
> >
> > I bisected it this morning. Bisected cleanly to...
> >
> > 9eaaa2d5759837402ec5eee13b2a97921808c3eb is the first bad commit
> OK, I've debugged it. This commit is really at fault. The problem is
> following:
> When using writev, the page we copy from is not paged in (while when we
> use ordinary write, it is paged in). This difference might be worth
> investigation on its own (as it is likely to heavily impact performance of
> writev) but is irrelevant for us now - we should handle this without data
> corruption anyway. Because the source page is not available, we pass 0 as
> the number of copied bytes to write_end and thus ext3_write_end decides to
> truncate the file to original size. This is perfectly fine. The problem is
> that we do this by ext3_truncate() which just frees corresponding block but
> does not unmap buffers. So we leave mapped buffers beyond i_size (they
> actually never were inside i_size) but the blocks they are mapped to are
> already free. The write is then retried (after mapping the page),
> block_write_begin() sees the buffer is mapped (although it is beyond
> i_size) and thus it does not call ext3_get_block() anymore. So as a result,
> data is written to a block that is no longer allocated to the file. Bummer
> - welcome filesystem corruption.
> Ext4 also has this problem but delayed allocation mitigates the effect to
> an error in accounting of blocks reserved for delayed allocation and thus
> under normal circumstances nothing bad happens.
> The question is how to solve this in the cleanest way. We can call
> vmtruncate() instead of ext3_truncate() as we used to do but Nick wants to
> get rid of that (that's why I originally changed the code to what it is
> now). So probably we could just manually call truncate_pagecache() instead.
> Nick, I think your truncate calling sequence patch set needs similar fix
> for all filesystems as well.
The patch below fixes the issue for me...
Honza
>From 1b2ad411dd86afbfdb3c5b0f913230e9f1f0b858 Mon Sep 17 00:00:00 2001
From: Jan Kara <[email protected]>
Date: Tue, 1 Dec 2009 16:53:06 +0100
Subject: [PATCH] ext3: Fix data / filesystem corruption when write fails to copy data
When ext3_write_begin fails after allocating some blocks or
generic_perform_write fails to copy data to write, we truncate blocks already
instantiated beyond i_size. Although these blocks were never inside i_size, we
have to truncate pagecache of these blocks so that corresponding buffers get
unmapped. Otherwise subsequent __block_prepare_write (called because we are
retrying the write) will find the buffers mapped, not call ->get_block, and
thus the page will be backed by already freed blocks leading to filesystem and
data corruption.
Signed-off-by: Jan Kara <[email protected]>
---
fs/ext3/inode.c | 18 ++++++++++++++----
1 files changed, 14 insertions(+), 4 deletions(-)
diff --git a/fs/ext3/inode.c b/fs/ext3/inode.c
index 354ed3b..f9d6937 100644
--- a/fs/ext3/inode.c
+++ b/fs/ext3/inode.c
@@ -1151,6 +1151,16 @@ static int do_journal_get_write_access(handle_t *handle,
return ext3_journal_get_write_access(handle, bh);
}
+/*
+ * Truncate blocks that were not used by write. We have to truncate the
+ * pagecache as well so that corresponding buffers get properly unmapped.
+ */
+static void ext3_truncate_failed_write(struct inode *inode)
+{
+ truncate_inode_pages(inode->i_mapping, inode->i_size);
+ ext3_truncate(inode);
+}
+
static int ext3_write_begin(struct file *file, struct address_space *mapping,
loff_t pos, unsigned len, unsigned flags,
struct page **pagep, void **fsdata)
@@ -1209,7 +1219,7 @@ write_begin_failed:
unlock_page(page);
page_cache_release(page);
if (pos + len > inode->i_size)
- ext3_truncate(inode);
+ ext3_truncate_failed_write(inode);
}
if (ret == -ENOSPC && ext3_should_retry_alloc(inode->i_sb, &retries))
goto retry;
@@ -1304,7 +1314,7 @@ static int ext3_ordered_write_end(struct file *file,
page_cache_release(page);
if (pos + len > inode->i_size)
- ext3_truncate(inode);
+ ext3_truncate_failed_write(inode);
return ret ? ret : copied;
}
@@ -1330,7 +1340,7 @@ static int ext3_writeback_write_end(struct file *file,
page_cache_release(page);
if (pos + len > inode->i_size)
- ext3_truncate(inode);
+ ext3_truncate_failed_write(inode);
return ret ? ret : copied;
}
@@ -1383,7 +1393,7 @@ static int ext3_journalled_write_end(struct file *file,
page_cache_release(page);
if (pos + len > inode->i_size)
- ext3_truncate(inode);
+ ext3_truncate_failed_write(inode);
return ret ? ret : copied;
}
--
1.6.4.2
On Tue, 2009-12-01 at 17:03 +0100, Jan Kara wrote:
> The patch below fixes the issue for me...
Ditto.
-Mike
On Tue 01-12-09 15:35:59, Jan Kara wrote:
> On Tue 01-12-09 12:42:45, Mike Galbraith wrote:
> > On Mon, 2009-11-30 at 19:48 -0500, James Y Knight wrote:
> > > On Nov 30, 2009, at 3:55 PM, James Y Knight wrote:
> > >
> > > > This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
> > > >
> > > > The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
> > > >
> > > > So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
> > > >
> > > > This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
> > > >
> > > > My /tmp is an ext3 filesystem, in case that matters.
> > >
> > > Further testing shows that the filesystem type *does* matter. The bug does not exhibit when the test is run on ext2, ext4, vfat, btrfs, jfs, or xfs (and probably all the others too). Only, so far as I can determine, on ext3.
> >
> > I bisected it this morning. Bisected cleanly to...
> >
> > 9eaaa2d5759837402ec5eee13b2a97921808c3eb is the first bad commit
> OK, I've debugged it. This commit is really at fault. The problem is
> following:
> When using writev, the page we copy from is not paged in (while when we
> use ordinary write, it is paged in). This difference might be worth
> investigation on its own (as it is likely to heavily impact performance of
> writev) but is irrelevant for us now - we should handle this without data
> corruption anyway.
I've looked into why writev fails reliably the writes. The reason is that
iov_iter_fault_in_readable() faults in only the first IO buffer. Because
this is just 600 bytes big, following iov_iter_copy_from_user_atomic copies
only 600 bytes and block_write_end sets number of copied bytes to 0. Thus
we restart the write and do it one iov per iteration which succeeds. So
everything works as designed only it gets inefficient in this particular
case.
Honza
On Dec 1, 2009, at 11:03 AM, Jan Kara wrote:
> On Tue 01-12-09 15:35:59, Jan Kara wrote:
>> On Tue 01-12-09 12:42:45, Mike Galbraith wrote:
>>> I bisected it this morning. Bisected cleanly to...
>>>
>>> 9eaaa2d5759837402ec5eee13b2a97921808c3eb is the first bad commit
>> OK, I've debugged it. This commit is really at fault. The problem is
>> following:
>> When using writev, the page we copy from is not paged in (while when we
>> use ordinary write, it is paged in). This difference might be worth
>> investigation on its own (as it is likely to heavily impact performance of
>> writev) but is irrelevant for us now - we should handle this without data
>> corruption anyway. Because the source page is not available, we pass 0 as
>> the number of copied bytes to write_end and thus ext3_write_end decides to
>> truncate the file to original size. This is perfectly fine. The problem is
>> that we do this by ext3_truncate() which just frees corresponding block but
>> does not unmap buffers. So we leave mapped buffers beyond i_size (they
>> actually never were inside i_size) but the blocks they are mapped to are
>> already free. The write is then retried (after mapping the page),
>> block_write_begin() sees the buffer is mapped (although it is beyond
>> i_size) and thus it does not call ext3_get_block() anymore. So as a result,
>> data is written to a block that is no longer allocated to the file. Bummer
>> - welcome filesystem corruption.
>> Ext4 also has this problem but delayed allocation mitigates the effect to
>> an error in accounting of blocks reserved for delayed allocation and thus
>> under normal circumstances nothing bad happens.
>> The question is how to solve this in the cleanest way. We can call
>> vmtruncate() instead of ext3_truncate() as we used to do but Nick wants to
>> get rid of that (that's why I originally changed the code to what it is
>> now). So probably we could just manually call truncate_pagecache() instead.
>> Nick, I think your truncate calling sequence patch set needs similar fix
>> for all filesystems as well.
> The patch below fixes the issue for me...
Thank you! I can confirm that the patch fixes the issue in my real application as well.
James-
On Tue, Dec 01, 2009 at 03:35:59PM +0100, Jan Kara wrote:
> On Tue 01-12-09 12:42:45, Mike Galbraith wrote:
> > On Mon, 2009-11-30 at 19:48 -0500, James Y Knight wrote:
> > > On Nov 30, 2009, at 3:55 PM, James Y Knight wrote:
> > >
> > > > This test case fails in 2.6.23-2.6.25, because of the bug fixed in 864f24395c72b6a6c48d13f409f986dc71a5cf4a, and now again in at least 2.6.31 and 2.6.32pre8 because of a *different* bug. This test *does not* fail 2.6.26. I have not tested anything between 2.6.26 and 2.6.31.
> > > >
> > > > The bug in 2.6.31 is definitely not the same bug as 2.6.23's. This time, the zero'd area of the file doesn't show up immediately upon writing the file. Instead, the kernel waits to mangle the file until it has to flush the buffer to disk. *THEN* it zeros out parts of the file.
> > > >
> > > > So, after writing out the new file with writev, and checking the md5sum (which is correct), this test case asks the kernel to flush the cache for that file, and then checks the md5sum again. ONLY THEN is the file corrupted. That is, I won't hesitate to say *incredibly evil* behavior: it took me quite some time to figure out WTH was going wrong with my program before determining it was a kernel bug.
> > > >
> > > > This test case is distilled from an actual application which doesn't even intentionally use writev: it just uses C++'s ofstream class to write data to a file. Unfortunately, that class smart and uses writev under the covers. Unfortunately, I guess nobody ever tests linux writev behavior, since it's broken _so_much_of_the_time_. I really am quite astounded to see such a bad track record for such a fundamental core system call....
> > > >
> > > > My /tmp is an ext3 filesystem, in case that matters.
> > >
> > > Further testing shows that the filesystem type *does* matter. The bug does not exhibit when the test is run on ext2, ext4, vfat, btrfs, jfs, or xfs (and probably all the others too). Only, so far as I can determine, on ext3.
> >
> > I bisected it this morning. Bisected cleanly to...
> >
> > 9eaaa2d5759837402ec5eee13b2a97921808c3eb is the first bad commit
> OK, I've debugged it. This commit is really at fault. The problem is
> following:
> When using writev, the page we copy from is not paged in (while when we
> use ordinary write, it is paged in). This difference might be worth
> investigation on its own (as it is likely to heavily impact performance of
> writev) but is irrelevant for us now - we should handle this without data
> corruption anyway. Because the source page is not available, we pass 0 as
> the number of copied bytes to write_end and thus ext3_write_end decides to
> truncate the file to original size. This is perfectly fine. The problem is
> that we do this by ext3_truncate() which just frees corresponding block but
> does not unmap buffers. So we leave mapped buffers beyond i_size (they
> actually never were inside i_size) but the blocks they are mapped to are
> already free. The write is then retried (after mapping the page),
> block_write_begin() sees the buffer is mapped (although it is beyond
> i_size) and thus it does not call ext3_get_block() anymore. So as a result,
> data is written to a block that is no longer allocated to the file. Bummer
> - welcome filesystem corruption.
> Ext4 also has this problem but delayed allocation mitigates the effect to
> an error in accounting of blocks reserved for delayed allocation and thus
> under normal circumstances nothing bad happens.
> The question is how to solve this in the cleanest way. We can call
> vmtruncate() instead of ext3_truncate() as we used to do but Nick wants to
> get rid of that (that's why I originally changed the code to what it is
> now). So probably we could just manually call truncate_pagecache() instead.
> Nick, I think your truncate calling sequence patch set needs similar fix
> for all filesystems as well.
Thanks Jan, good analysis and yes I believe I will need to do a similar
fix there.
On Wed, Dec 02, 2009 at 08:04:26PM +0100, Jan Kara wrote:
> > When using writev, the page we copy from is not paged in (while when we
> > use ordinary write, it is paged in). This difference might be worth
> > investigation on its own (as it is likely to heavily impact performance of
> > writev) but is irrelevant for us now - we should handle this without data
> > corruption anyway.
> I've looked into why writev fails reliably the writes. The reason is that
> iov_iter_fault_in_readable() faults in only the first IO buffer. Because
> this is just 600 bytes big, following iov_iter_copy_from_user_atomic copies
> only 600 bytes and block_write_end sets number of copied bytes to 0. Thus
> we restart the write and do it one iov per iteration which succeeds. So
> everything works as designed only it gets inefficient in this particular
> case.
Yep, this would be right. We could actually do more prefaulting; I
think I was being a little over conservative and worried about earlier
pages being unmapped before we were able to consume them... but I
think being too worried about that case is optimizing an unusual case
that is probably performing badly anyway at the expense of more common
patterns.
Anyway, what I was doing to test this code when I wrote it was to
inject random failures into user copy functions. I guess this could
be useful to merge in the error injection framework?
Thanks,
Nick
On Thu 03-12-09 06:28:25, Nick Piggin wrote:
> On Wed, Dec 02, 2009 at 08:04:26PM +0100, Jan Kara wrote:
> > > When using writev, the page we copy from is not paged in (while when we
> > > use ordinary write, it is paged in). This difference might be worth
> > > investigation on its own (as it is likely to heavily impact performance of
> > > writev) but is irrelevant for us now - we should handle this without data
> > > corruption anyway.
> > I've looked into why writev fails reliably the writes. The reason is that
> > iov_iter_fault_in_readable() faults in only the first IO buffer. Because
> > this is just 600 bytes big, following iov_iter_copy_from_user_atomic copies
> > only 600 bytes and block_write_end sets number of copied bytes to 0. Thus
> > we restart the write and do it one iov per iteration which succeeds. So
> > everything works as designed only it gets inefficient in this particular
> > case.
> Yep, this would be right. We could actually do more prefaulting; I
> think I was being a little over conservative and worried about earlier
> pages being unmapped before we were able to consume them... but I
> think being too worried about that case is optimizing an unusual case
> that is probably performing badly anyway at the expense of more common
> patterns.
Yeah, IMHO optimal would be to fault in enough buffers so that we can
fill one page (although we may pose some upper bound on the number of pages
we are willing to fault in - like 1 MB of data or so).
> Anyway, what I was doing to test this code when I wrote it was to
> inject random failures into user copy functions. I guess this could
> be useful to merge in the error injection framework?
Yes, that would be definitely useful. This was exceptionally easy to
track down because it was easily reproducible. But otherwise this path is
almost never taken and bugs in there are hard to debug so it would get more
testing coverage. I've spent like a month debugging a bug in reiserfs
causing data corruption in this path - mainly because it took a few days to
reproduce it and I didn't know what could be possibly triggering it...
Honza