Re: kjournald() with DIO

[Andrea Arcangeli <andrea@suse.de>]

On Thu, Sep 15, 2005 at 04:28:44PM -0700, Mingming Cao wrote:
> Isn't direct-io and buffered IO being serialized by i_sem? If so, the

page faults aren't serialized by i_sem, so if a page fault happens the
vm may unmap the pte under memory pressure marking the page dirty. But
we have the guarantee that such dirty bit in the pte was set after we
were inside write() so we can drop it (or we can as well leave it), it's
undefined.

but the point is that the buffered-io fallback happens after the
invalidate, so it shouldn't matter if we drop dirty buffers or dirty
pages in the invalidate, they should be re-generated.

So overall I believe it's semantically equivalent if we drop the dirty
bits or not. Clearly not dropping them is a bit less intrusive, but
current code seems ready to drop them by calling ->invalidatepage
(destructive) instead of creating a blocking version of ->releasepage
(non destructive).