On Mon, May 08, 2017 at 04:42:50PM -0700, Shaohua Li wrote: > I'm seeing below with lastet upstream. Commit > 5ecda13(generic_file_read_iter(): make use of iov_iter_revert()) > directly pass iter to ->direct_IO(), ->direct_io already advances count > bytes but return -EIOCBQUEUED. In this case, count is bigger than > iov_iter_count(iter). We really want to revert count - > iov_iter_count(iter) instead of the opposite.
Already discussed; the patch in my tree is that braino fix + sanity check in iov_iter_revert(), to prevent anything similar in the future. The reason it wasn't caught in tests (without KASAN, that is) is that to get really obvious breakage out of that, you need a short read from ->direct_IO() *and* fallback to cached read after it. -EIOCBQUEUED obviously excludes the latter, so the bug is almost certain to remain quiet. There is a non-zero chance of actually running into unmapped memory while doing that iov_iter_revert(), but it's fairly low. Unfortunately, since that allowed the damn thing to remain unnoticed... Will be in tonight's pull request.
