On Thu, Mar 03, 2011 at 08:51:55PM -0500, Chris Mason wrote:
> Excerpts from Zhong, Xin's message of 2011-03-02 05:58:49 -0500:
> > I downloaded openmotif and run the command as Mitch mentioned and was able to recreate the problem locally. And I managed to simplify the command into a very simple program which can capture the problem easily. See below code:
> >
> > #include <sys/types.h>
> > #include <sys/stat.h>
> > #include <fcntl.h>
> > static char a[4096*3];
> > int main()
> > {
> > int fd = open("out", O_WRONLY|O_CREAT|O_TRUNC, 0666);
> > write(fd,a+1, 4096*2);
> > exit(0);
> > }
> >
> > It seems that if we give an unaligned address to btrfs write and the buffer reside on more than 2 pages. It will trigger this bug.
> > If we give an aligned address to btrfs write, it works well no matter how many pages are given.
> >
> > I use ftrace to observe it. It seems iov_iter_fault_in_readable do not trigger pagefault handling when the address is not aligned. I do not quite understand the reason behind it. But the solution should be to process the page one by one. And that's also what generic file write routine does.
> >
> > Any suggestion are welcomed. Thanks!
>
> Great job guys. I'm using this on top of my debugging patch. It passes
> the unaligned test but I'll give it a real run tonight and look for
> other problems.
>
> (This is almost entirely untested, please don't use it quite yet)
>
> -chris
>
> diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
> index 89a6a26..6a44add 100644
> --- a/fs/btrfs/file.c
> +++ b/fs/btrfs/file.c
> @@ -1039,6 +1038,14 @@ static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
>
> copied = btrfs_copy_from_user(pos, num_pages,
> write_bytes, pages, &i);
> +
> + /*
> + * if we have trouble faulting in the pages, fall
> + * back to one page at a time
> + */
> + if (copied < write_bytes)
> + nrptrs = 1;
> +
> if (copied == 0)
> dirty_pages = 0;
> else
Btw this situation is taken care of in my write path rewrite patch, if copied ==
0 we switch to one segment at a time. Thanks,
Josef
--
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
