Messages in this thread | | | Date | Mon, 16 Nov 2009 13:36:13 -0500 | From | Chris Mason <> | Subject | Re: [PATCH 0/7] Reduce GFP_ATOMIC allocation failures, candidate fix V3 |
| |
On Mon, Nov 16, 2009 at 05:44:07PM +0100, Milan Broz wrote: > On 11/13/2009 03:46 AM, Chris Mason wrote: > > On Thu, Nov 12, 2009 at 05:00:05PM -0500, Chris Mason wrote: > > > > [ ...] > > > >> > >> The punch line is that the btrfs guy thinks we can solve all of this with > >> just one more thread. If we change dm-crypt to have a thread dedicated > >> to sync IO and a thread dedicated to async IO the system should smooth > >> out. > > Please, can you cc DM maintainers with these kind of patches? dm-devel list at least. >
Well, my current patch is a hack. If I had come up with a proven theory (hopefully Mel can prove it ;), it definitely would have gone through the dm-devel lists.
> Note that the crypt requests can be already processed synchronously or asynchronously, > depending on used crypto module (async it is in the case of some hw acceleration). > > Adding another queue make the situation more complicated and because the crypt > requests can be queued in crypto layer I am not sure that this solution will help > in this situation at all. > (Try to run that with AES-NI acceleration for example.)
The problem is that async threads still imply a kind of ordering. If there's a fifo serviced by one thread or 10, the latency ramifications are very similar for a new entry on the list. We have to wait for a large portion of the low-prio items in order to service a high prio item.
With a queue dedicated to sync requests and one dedicated to async, you'll get better read latencies. Btrfs has a similar problem around the crc helper threads and it ends up solving things with two different lists (high and low prio) processed by one thread.
-chris
| |