On Fri, 21 Mar 2014 12:56:34 +0400 Pavel Shilovsky <[email protected]> wrote:
> 2014-03-21 3:10 GMT+04:00 Jeff Layton <[email protected]>: > > On Fri, 21 Mar 2014 05:56:04 +0900 > > Tetsuo Handa <[email protected]> wrote: > > > >> Jeff Layton wrote: > >> > That's expected behavior. The kernel believes that the file is frozen in > >> > length so it returns short read() calls until the size is updated. > >> > >> The "size is updated" means "stat() detects the growth of file size", > >> doesn't it? Then, the former is expected behavior. > >> > >> > > >> > cache=loose is very much not recommended for use when you have multiple > >> > hosts accessing files on the server (or access by processes on the > >> > server itself). It only gives you "loose" cache coherency. The whole > >> > point of it is to allow the client to cache data even when the protocol > >> > says that it shouldn't. > >> > >> But why is the latter ( "read() returns non-0 when stat() detects the > >> growth > >> of file size but the data actually read is '\0'" ) is expected behavior? > >> It sounds like a bug that the client caches '\0' (data nobody has ever > >> wrote) > >> instead of '.' (data somebody wrote when the file size grew). > >> > > > > Yeah, that sounds wrong. What should happen is that the cache is > > invalidated when the size changes. It's possible there is a race in > > that code however. The locking around it is pretty sloppy... > > When fstat() get a new file size it sets > CIFS_I(inode)->invalid_mapping to true but do not revalidate the > cache. Then generic_file_aio_read() reads the wrong data. I think we > need to check if CIFS_I(inode)->invalid_mapping is true and revalidate > the cache before calling generic_file_aio_read() in > file->f_ops->aio_read(). Now cache revalidation happens in lookup/open > and mmap codepaths only for cache=loose. > > Of course, cache=loose is not recommended for this sort of work flow > and cache=strict should be used to provide a data coherency between > several machines. > Ahh right, you're quite correct that we need to revalidate the cache before doing an aio_read. We should have cifs do something like nfs_file_read() does... I also suspect that we have a problem with the invalid_mapping flag similar to the one NFS had until recently. That was fixed by commit d529ef83c355. We probably ought to do something similar for cifs. -- Jeff Layton <[email protected]> -- To unsubscribe from this list: send the line "unsubscribe linux-cifs" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
