David Howells <dhowells@xxxxxxxxxx> writes:
> Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
>> Yes. I do that kind of thing in iomap. What you're doing there looks
>> a bit like offset_in_folio(), but I don't understand the problem with
>> just checking pos against i_size directly.
>
> pos can be in the middle of a page. If i_size is at, say, the same point in
> the middle of that page and the page isn't currently in the cache, then we'll
> just clear the entire page and not read the bottom part of it (ie. the bit
> prior to the EOF).
>
> It's odd, though, that xfstests doesn't catch this.
Some more details/information:
1.) The test is a really simple piece of code that essentially writes
a incrementing number to a file on 'node a', while a 'tail -f'
is run and exited multiple times on 'node b'. We haven't had much
luck in this causing the problem when done on a single node
2.) ((pos & PAGE_MASK) >= i_size_read(inode)) ||' merely reflects the
logic that was in place prior to 1cc1699070bd. Patching
fs/ceph/addr.c with this does seem to eliminate the corruption.
(We'll test with the patch Jeff posted here shortly)
3.) After finding all of this, I went upstream to look at fs/ceph/addr.c
and discovered the move to leveraging fs/netfs/read_helper.c - we've
only just compiled against git master and replicated the corruption
there.
--
Andrew W. Elble
Infrastructure Engineer
Information and Technology Services
Rochester Institute of Technology
tel: (585)-475-2411 | aweits@xxxxxxx
PGP: BFAD 8461 4CCF DC95 DA2C B0EB 965B 082E 863E C912
