Messages in this thread | | | Date | Wed, 23 Jul 2014 21:07:31 +0200 | From | Jens Axboe <> | Subject | Re: [PATCH v10] NVMe: Convert to blk-mq |
| |
On 2014-07-23 20:58, Matias Bjorling wrote: >>> + iod = nvme_alloc_iod(psegs, blk_rq_bytes(req), GFP_ATOMIC); >>> if (!iod) >>> + return result; >> >> So there's still a memory allocation for each request here. Any reason >> this can't be preallocated at least for reasonable sized I/O? > > Not at all. I've kept from adding optimizations in the first pass. The > patches following can implement the optimizations. Jens already has a > patch for this in his tree. It also removes GFP_ATOMIC.
That is correct, and that is the way the series should be done. I've got patches getting rid of this allocation for smaller IO:
http://git.kernel.dk/?p=linux-block.git;a=commit;h=04497c3394f3220111d4274704a1ff6bdd3ceae3
which is a noticable win for high IOPS smaller IO.
>> Seems like blk-mq would make your life easier by exporting an iterator >> that goes over each in-use request instead of the current >> blk_mq_tag_busy_iter prototype. blk_mq_timeout_check would also be able >> to make use of that, so maybe that would be a good preparatory patch? > > Yes. I'll prepare a patch and send it off to Jens.
Thanks, that will clean it up nicely. It's a bit backwards these days, since the mechanism got changed.
-- Jens Axboe
| |