Messages in this thread | | | From | "Kirill A. Shutemov" <> | Subject | Re: [PATCH] mm, fs: avoid page allocation beyond i_size on read | Date | Thu, 22 Aug 2013 16:05:27 +0300 (EEST) |
| |
Steven Whitehouse wrote: > Hi, > > On Wed, 2013-08-21 at 13:58 -0700, Andrew Morton wrote: > > On Wed, 21 Aug 2013 17:42:12 +0100 Steven Whitehouse <swhiteho@redhat.com> wrote: > > > > > > I don't think the change is harmful. The worst case scenario is race with > > > > write or truncate, but it's valid to return EOF in this case. > > > > > > > > What scenario do you have in mind? > > > > > > > > > > 1. File open on node A > > > 2. Someone updates it on node B by extending the file > > > 3. Someone reads the file on node A beyond end of original file size, > > > but within end of new file size as updated by node B. Without the patch > > > this works, with it, it will fail. The reason being the i_size would not > > > be up to date until after readpage(s) has been called.
CC: +linux-fsdevel@
So in this case node A will see the file like it was never touched by node B. It's okay, if new i_size will eventually reach node A.
Is ->readpage() the only way to get i_size updated on node A or it will be eventually updated without it?
If it's the only way, we need add a explicit way to initiate i_size sync between nodes on read. Probably, distributed filesystems should provide own ->aio_read() which deal i_size as the filesystem need.
> > > I think this is likely to be an issue for any distributed fs using > > > do_generic_file_read(), although it would certainly affect GFS2, since > > > the locking is done at page cache level, > > > > Boy, that's rather subtle. I'm surprised that the generic filemap.c > > stuff works at all in that sort of scenario. > > > > Can we put the i_size check down in the no_cached_page block? afaict > > that will solve the problem without breaking GFS2 and is more > > efficient? > > > > Well I think is even more subtle, since it relies on ->readpages > updating the file size, even if it has failed to actually read the > required pages :-) Having said that, we do rely on ->readpages updating > the inode size elsewhere in this function, as per the block comment > immediately following the page_ok label.
That i_size recheck was invented to cover different use case: read vs. truncate race. Userspace should not see truncate-caused zeros in buffer. It's not to prevent file extending vs. read() race. This usually harmless: data is consistent.
-- Kirill A. Shutemov
| |