[dm-devel] [PATCH] PROBLEM: SSD access time with dm-crypt is way too high

[Christoph Hellwig]

On Mon, Jan 10, 2011 at 12:19:25PM -0500, Mikulas Patocka wrote:
> There are two approaches that I considered:
> 
> 1. make a special work for unplug. After queuing each IO's work, cancel the
> unplug work and queue it again (so that it will always be queued as a last
> entry). Unfortunatelly, canceling a work is rather slow operation so I decided
> to not use this approach.
> 
> 2. use a special pointer that points to the last IO. When the IO is finished
> and the pointer matches this IO, we know that it was the last IO and we should
> unplug. This patch implements this approach.

What about keeping a reference count of pending I/O requests in a shared
structure, with the last decrement doing the unplug?