On Wed, Jan 08, 2020 at 11:36:04AM +0000, Filipe Manana wrote:
> On Tue, Jan 7, 2020 at 5:57 PM Darrick J. Wong <darrick.wong@xxxxxxxxxx> wrote:
> >
> > On Tue, Jan 07, 2020 at 04:23:15PM +0000, Filipe Manana wrote:
> > > On Mon, Dec 16, 2019 at 6:28 PM <fdmanana@xxxxxxxxxx> wrote:
> > > >
> > > > From: Filipe Manana <fdmanana@xxxxxxxx>
> > > >
> > > > We always round down, to a multiple of the filesystem's block size, the
> > > > length to deduplicate at generic_remap_check_len(). However this is only
> > > > needed if an attempt to deduplicate the last block into the middle of the
> > > > destination file is requested, since that leads into a corruption if the
> > > > length of the source file is not block size aligned. When an attempt to
> > > > deduplicate the last block into the end of the destination file is
> > > > requested, we should allow it because it is safe to do it - there's no
> > > > stale data exposure and we are prepared to compare the data ranges for
> > > > a length not aligned to the block (or page) size - in fact we even do
> > > > the data compare before adjusting the deduplication length.
> > > >
> > > > After btrfs was updated to use the generic helpers from VFS (by commit
> > > > 34a28e3d77535e ("Btrfs: use generic_remap_file_range_prep() for cloning
> > > > and deduplication")) we started to have user reports of deduplication
> > > > not reflinking the last block anymore, and whence users getting lower
> > > > deduplication scores. The main use case is deduplication of entire
> > > > files that have a size not aligned to the block size of the filesystem.
> > > >
> > > > We already allow cloning the last block to the end (and beyond) of the
> > > > destination file, so allow for deduplication as well.
> > > >
> > > > Link: https://lore.kernel.org/linux-btrfs/2019-1576167349.500456@xxxxxxxxxxxxxx/
> > > > Signed-off-by: Filipe Manana <fdmanana@xxxxxxxx>
> > >
> > > Darrick, Al, any feedback?
> >
> > Is there a fstest to check for correct operation of dedupe at or beyond
> > source and destfile EOF? Particularly if one range is /not/ at EOF?
>
> Such as what generic/158 does already?
Urk, heh. :)
> > And that an mmap read of the EOF block will see zeroes past EOF before
> > and after the dedupe operation?
>
> Can you elaborate a bit more? Why an mmap read and not a buffered or a
> direct IO read before and after deduplication?
> Is there anything special for the mmap reads on xfs, is that your
> concern? Or is the idea to deduplicate while the file is mmap'ed?
I cite mmap reads past EOF specifically because unlike buffered/direct
reads where the VFS will stop reading exactly at EOF, a memory mapping
maps in an entire memory page, and the fs is supposed to ensure that the
bytes past EOF are zeroed.
Hm now that I look at g/158 it doesn't actually verify mmap reads. I
looked around and can't really see anything that checks mmap reads
before and after a dedupe operation at EOF.
> > If I fallocate a 16k file, write 'X' into the first 5000 bytes,
> > write 'X' into the first 66,440 bytes (60k + 5000) of a second file, and
> > then try to dedupe (first file, 0-8k) with (second file, 60k-68k),
> > should that work?
>
> You haven't mentioned the size of the second file, nor if the first
> file has a size of 16K which I assume (instead of fallocate with the
> keep size flag).
Er, sorry, yes. The first file is 16,384 bytes long; the second file is
66,440 bytes.
> Anyway, I assume you actually meant to dedupe the range 0 - 5000 from
> the first file into the range 60k - 60k + 5000 of the second file, and
> that the second file has a size of 60k + 5000.
Nope, I meant to say to dedupe the range (offset: 0, length: 8192) from
the first file into the second file (offset: 61440, length: 8192). The
source range is entirely below EOF, and the dest range ends right at
EOF in the second file.
> If so, that fails with -EINVAL because the source range is not block
> size aligned, and we already have generic fstests that test attempt to
> duplication and clone non-aligned ranges that don't end at eof.
> This patch doesn't change that behaviour, it only aims to allow
> deduplication of the eof block of the source file into the eof of the
> destination file.
>
>
> >
> > I'm convinced that we could support dedupe to EOF when the ranges of the
> > two files both end at the respective file's EOF, but it's the weirder
> > corner cases that I worry about...
>
> Well, we used to do that in btrfs before migrating to the generic code.
> Since I discovered the corruption due to deduplication of the eof
> block into the middle of a file in 2018's summer, the btrfs fix
> allowed deduplication of the eof block only if the destination end
> offset matched the eof of the destination file:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=de02b9f6bb65a6a1848f346f7a3617b7a9b930c0
>
> Since then no issues were found nor users reported any problems so far.
<nod> I'm ok with that one scenario, it's the "one range ends at eof,
the other doesn't" case that I'm picking on. :)
(Another way to shut me up would be to run generic/52[12] with
TIME_FACTOR=1000 (i.e. 1 billion fsx ops) and see what comes exploding
out. :))
> Any other specific test you would like to see?
No, just that. And mmap reads. :)
--D
> Thanks.
>
> >
> > --D
> >
> > > Thanks.
> > >
> > > > ---
> > > > fs/read_write.c | 10 ++++------
> > > > 1 file changed, 4 insertions(+), 6 deletions(-)
> > > >
> > > > diff --git a/fs/read_write.c b/fs/read_write.c
> > > > index 5bbf587f5bc1..7458fccc59e1 100644
> > > > --- a/fs/read_write.c
> > > > +++ b/fs/read_write.c
> > > > @@ -1777,10 +1777,9 @@ static int remap_verify_area(struct file *file, loff_t pos, loff_t len,
> > > > * else. Assume that the offsets have already been checked for block
> > > > * alignment.
> > > > *
> > > > - * For deduplication we always scale down to the previous block because we
> > > > - * can't meaningfully compare post-EOF contents.
> > > > - *
> > > > - * For clone we only link a partial EOF block above the destination file's EOF.
> > > > + * For clone we only link a partial EOF block above or at the destination file's
> > > > + * EOF. For deduplication we accept a partial EOF block only if it ends at the
> > > > + * destination file's EOF (can not link it into the middle of a file).
> > > > *
> > > > * Shorten the request if possible.
> > > > */
> > > > @@ -1796,8 +1795,7 @@ static int generic_remap_check_len(struct inode *inode_in,
> > > > if ((*len & blkmask) == 0)
> > > > return 0;
> > > >
> > > > - if ((remap_flags & REMAP_FILE_DEDUP) ||
> > > > - pos_out + *len < i_size_read(inode_out))
> > > > + if (pos_out + *len < i_size_read(inode_out))
> > > > new_len &= ~blkmask;
> > > >
> > > > if (new_len == *len)
> > > > --
> > > > 2.11.0
> > > >
