I ran into a problem with 2.5's O_DIRECT read path,
generic_file_direct_IO usually ends up in generic_direct_IO
this does bounds checking on the I/O and then flushes any
cached data.
Once we return to generic_file_direct_IO we unconditionally
call invalidate_inode_pages2 if there are any cached pages.
If we make a non-aligned O_DIRECT read call, the end result is we
call invalidate_inode_pages2, but we do not do the filemap_fdatawrite,
filemap_fdatawait calls. End result is we throw the buffered data away.
Either the flush needs to happen before the bounds checks, or the
invalidate should only be done on a successful write. It looks
pretty hard to detect the latter case with the current structure,
we can get EINVAL from the bounds check and possibly from an
aligned, but invalid memory address being passed in.
This is worse for xfs than for other filesystems since if we do the
invalidate without the flush first, we end up with a delayed allocate
extent in memory and no data to flush into it. Various spots in the
filesystem dislike this.
I can fix it inside xfs by doing my own flush and invalidate first, but
the generic code should really be fixed.
Thoughts?
Steve
--Steve Lord voice: +1-651-683-3511 Principal Engineer, Filesystem Software email: lord@sgi.com - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/