From mboxrd@z Thu Jan  1 00:00:00 1970
From: hch@lst.de (Christoph Hellwig)
Date: Tue, 9 Jul 2019 22:22:01 +0200
Subject: [PATCH rfc] nvme-tcp: support simple polling
In-Reply-To: <20190703210804.23137-1-sagi@grimberg.me>
References: <20190703210804.23137-1-sagi@grimberg.me>
Message-ID: <20190709202201.GB8405@lst.de>

On Wed, Jul 03, 2019@02:08:04PM -0700, Sagi Grimberg wrote:
> Simple polling support via socket busy_poll interface.
> Although we do not shutdown interrupts but simply hammer
> the socket poll, we can sometimes find completions faster
> than the normal interrupt driven RX path.
> 
> We add per queue nr_cqe counter that resets every time
> RX path is invoked such that .poll callback can return it
> to stay consistent with the semantics.
> 
> Signed-off-by: Sagi Grimberg <sagi at grimberg.me>
> ---
> There is quite a bit more to do in this direction such as
> asking the net device to allocate polling rings and steer our
> polling queues to them, but that would require some infrastructure
> work in the networking stack.
> 
> With this alone, even while still absorbing the net device interrupts
> I'm seeing around 10%-20% latency improvement while sacrificing lots
> of cpu cycles for it.

Looks good enough to me, especially now that we need to explicitly opt
into polling queues.