On Thu, Apr 11, 2013 at 01:44:51PM +0200, Jan Kara wrote:
> Writing a large file using direct IO in 16 MB chunks sometimes results
> in a pathological allocation pattern where 16 MB chunks of large free
> extent are allocated to a file in a reversed order. So extents of a file
> look for example as:
>
> ext logical physical expected length flags
> 0 0 13 4550656
> 1 4550656 188136807 4550668 12562432
> 2 17113088 200699240 200699238 622592
> 3 17735680 182046055 201321831 4096
> 4 17739776 182041959 182050150 4096
> 5 17743872 182037863 182046054 4096
> 6 17747968 182033767 182041958 4096
> 7 17752064 182029671 182037862 4096
> ...
> 6757 45400064 154381644 154389835 4096
> 6758 45404160 154377548 154385739 4096
> 6759 45408256 252951571 154381643 73728 eof
>
> This happens because XFS_ALLOCTYPE_THIS_BNO allocation fails (the last
> extent in the file cannot be further extended) so we fall back to
> XFS_ALLOCTYPE_NEAR_BNO allocation which picks end of a large free
> extent as the best place to continue the file. Since the chunk at the
> end of the free extent again cannot be further extended, this behavior
> repeats until the whole free extent is consumed in a reversed order.
>
> For data allocations this backward allocation isn't beneficial so make
> xfs_alloc_compute_diff() pick start of a free extent instead of its end
> for them. That avoids the backward allocation pattern.
>
> Based on idea by Dave Chinner <dchinner@xxxxxxxxxx>.
Can you add a reference to the previous discussion thread here?
I had to go back and read it to remind myself of how we ended up
with this solution, so I think that we need to capture that
information in this commit message somehow. A url to an archive
(such as on oss.sgi.com) is probably the simplest way to do this.
> CC: Dave Chinner <dchinner@xxxxxxxxxx>
> Signed-off-by: Jan Kara <jack@xxxxxxx>
> ---
> fs/xfs/xfs_alloc.c | 22 ++++++++++++++++------
> 1 files changed, 16 insertions(+), 6 deletions(-)
>
> BTW, I've tested With this patch applied I really cannot reproduce the
> problematic allocation pattern anymore.
>
> diff --git a/fs/xfs/xfs_alloc.c b/fs/xfs/xfs_alloc.c
> index 0ad2325..64c6247 100644
> --- a/fs/xfs/xfs_alloc.c
> +++ b/fs/xfs/xfs_alloc.c
> @@ -173,6 +173,7 @@ xfs_alloc_compute_diff(
> xfs_agblock_t wantbno, /* target starting block */
> xfs_extlen_t wantlen, /* target length */
> xfs_extlen_t alignment, /* target alignment */
> + char userdata, /* are we allocating data? */
> xfs_agblock_t freebno, /* freespace's starting block */
> xfs_extlen_t freelen, /* freespace's length */
> xfs_agblock_t *newbnop) /* result: best start block from free */
> @@ -187,7 +188,12 @@ xfs_alloc_compute_diff(
> ASSERT(freelen >= wantlen);
> freeend = freebno + freelen;
> wantend = wantbno + wantlen;
> - if (freebno >= wantbno) {
> + /*
> + * We want to allocate from the start of a free extent if it is past
> + * the desired block or if we are allocating user data and the free
> + * extent is before desired block.
> + */
I think this probably needs a little more detail as to why we we do
this for user data. i.e. to carve from the front edge of the free
extent to allow for contiguous allocation from the remaining free
space if the file grows in the short term.
> + if (freebno >= wantbno || (userdata && freeend < wantend)) {
> if ((newbno1 = roundup(freebno, alignment)) >= freeend)
> newbno1 = NULLAGBLOCK;
So this is the meat of the change. We have this:
freebno freeend
+---------------------------------+
+-----+
prev +----------+
wantbno wantend
and for user data this will now return:
freebno freeend
+---------------------------------+
+-----+
+--------+ prev +----------+
newbno1 wantbno wantend
I wondered for a minute about how alignment affected the extent
returned by taking this different branch, but I'm the behaviour is
no different compared to carving an aligned chunk from the rear of
the free extent. If the extent is short, we get the same result
whether we try to carve it from the front or rear of the free space.
OK, what if we have:
freebno freeend
+---------------------------------+
+----------+
wantbno wantend
The existing code treats that the same as wantbno > freeend case
above, so we should treat it the same and carve from the front edge.
So the (freeend < wantend) check is sane, as is "<" for the
comparison. If the watned range fits within the freespace block,
then we should still carve that from the end of the freespace extent
as that was what was wanted.
IOWs, the code change looks good, and as such:
Reviewed-by: Dave Chinner <dchinner@xxxxxxxxxx>
However, I think this probably needs to sit in the dev tree for a
little while before we release it on the world. I don't think that
pushing this for 3.10 is wise as we need a bit of time to determine
if there are unintended side effects from this change under
accelerated aging workloads first. I'd like to be conservative on
this as the allocation primitives being touched are devilishly
complex and getting this wrong will have permanent impact on
filesystems...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
