2008/12/19 Sage Weil <sage@xxxxxxxxxxxx>:
> Hi Chris-
>
> I noticed some data and metadata getting out of sync on disk, despite
> wrapping my writes with btrfs transactions. After digging into it a bit,
> it appears to be a larger problem with inode size/data getting written
> during a regular commit.
>
> I have a test program append a few bytes at a time to a few different
> files, in a loop. I let it run until I see a btrfs transaction commit
> (via a printk at the bottom of btrfs_commit_transaction). Then 'reboot -f
> -n'. After remounting, all files exist but are 0 bytes, and debug-tree
> shows a bunch of empty files. I would expect to see either the sizes when
> the commit happend (a few hundred KB in my case), or no files at all;
> there was actually no point in time when any of the files were 0 bytes.
>
> Similarly, if I do the same but wait for a few commits to happen, after
> remount the file sizes reflect the size from around the next-to-last
> commit, not the last commit.
>
> This is probably more information than you need, but my original test was
> a bit more complicated, with weirder results. Append to each file, then
> write it's size to an xattr on another file. Wrap both operations in a
> transaction. Start it up, run 'sync', then reboot -f -n. When I remount
> the size and xattr are out of sync by exactly one iteration: the xattr
> reflects the size that resulted from _two_ writes back, not the
> immediately preceeding write. If anything I would expect to see a larger
> actual size than xattr value (for example if the start/end transaction
> ioctls weren't working)...
>
> sage
>
>
>
> #include <stdio.h>
> #include <stdlib.h>
> #include <unistd.h>
> #include <stdlib.h>
> #include <sys/types.h>
> #include <sys/stat.h>
> #include <fcntl.h>
>
> int main(int argc, char **argv)
> {
> while (1) {
> int r, fd, pos, i = rand() % 10;
> char a[20];
>
> sprintf(a, "%d.log", i);
> fd = open(a, O_CREAT|O_APPEND|O_WRONLY, 0600);
> r = write(fd, "foobarfoo\n", 10);
> pos = lseek(fd, 0, SEEK_CUR);
> printf("write %s = %d, size = %d\n", a, r, pos);
> close(fd);
> }
> }
>
This is the desired behaviour of data=ordered. Btrfs transaction commit
don't flush data, and metadata wont get updated until data IO complete.
http://article.gmane.org/gmane.comp.file-systems.btrfs/869/match=new+data+ordered+code
Regards
Yan Zheng
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
