Messages in this thread | | | Date | Wed, 24 Sep 2014 23:21:13 +0800 | Subject | Re: [PATCH v4 10/10] blk-mq: support per-distpatch_queue flush machinery | From | Ming Lei <> |
| |
On Wed, Sep 24, 2014 at 6:26 PM, Christoph Hellwig <hch@lst.de> wrote: >> +struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q, >> + struct blk_mq_hw_ctx *hctx, int cmd_size) > > I still think this should pass in the numa node instead of the hctx, and > allow node-local allocation for the old code as well.
We can do that, but have to pass NUMA_NO_NODE for old code since blk_init_allocated_queue() doesn't provide node information yet.
> As mentioned earlier > initializing mq_flush_lock for the !mq case is harmless.
q->mq_ops is another friend for the purpose, :-)
> > We also should document it where cleary somewhere that for devices that > have flushes enabled ->init_request can be called for more requests than > the queue depth, as drivers might allocate some sort of pool for them.
That does make sense.
Thanks,
| |