Re: [PATCH 5/5] net: macb: Fix race between HW and driver

From: David Miller <davem@davemloft.net>
To: soren.brinkmann@xilinx.com
Cc: michal.simek@xilinx.com, nicolas.ferre@atmel.com, git@xilinx.com,
	linux-kernel@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org, netdev@vger.kernel.org
Subject: Re: [PATCH 5/5] net: macb: Fix race between HW and driver
Date: Mon, 05 May 2014 17:09:59 -0400 (EDT)	[thread overview]
Message-ID: <20140505.170959.2172440110209805460.davem@davemloft.net> (raw)
In-Reply-To: <7c62baab-5f60-4f04-87f7-6b53fecc4746@CH1EHSMHS038.ehs.local>

From: Sören Brinkmann <soren.brinkmann@xilinx.com>
Date: Mon, 5 May 2014 14:05:19 -0700

> On Mon, 2014-05-05 at 04:56PM -0400, David Miller wrote:
>> From: Soren Brinkmann <soren.brinkmann@xilinx.com>
>> Date: Sun,  4 May 2014 15:43:02 -0700
>> 
>> > Under "heavy" RX load, the driver cannot handle the descriptors fast
>> > enough. In detail, when a descriptor is consumed, its used flag is
>> > cleared and once the RX budget is consumed all descriptors with a
>> > cleared used flag are prepared to receive more data. Under load though,
>> > the HW may constantly receive more data and use those descriptors with a
>> > cleared used flag before they are actually prepared for next usage.
>> > 
>> > The head and tail pointers into the RX-ring should always be valid and
>> > we can omit clearing and checking of the used flag.
>> > 
>> > Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>
>> 
>> Isn't the RX_USED bit the only thing that controls what RX entries
>> the chip will try to use?
>> 
>> I can't see how you can just remove the RX_USED bit handling
>> altogether.
>> 
>> The problem actually seems to be that in the current code we clear the
>> RX_USED bit before we actually reallocate the buffer and set it up.
>> 
>> It should be a bug to see the RX_USED bit set in gem_rx_refill(), and
>> the only reason why it can happen is exactly because you're clearing it
>> too early in gem_rx().
> 
> I don't follow. The HW uses the descriptor and the driver handles the
> received data. So, in gem_rx_refill we should actually only replace
> descriptor which have the RX_USED _set_, not? Currently the test tests
> for the opposite, since SW clears RX_USED in gem_rx. This patch just
> removes those two parts. The RX_USED is left as is (HW should have set
> it). And in gem_rx_refill we simply rely on the head and tail pointers
> to refill the used descriptors. I didn't see a reason to do the additional
> checking of the RX_USED bit.
> After the refill the RX_USED flags are of course cleared for all
> refilled descriptors.

Aha, that makes sense, I didn't notice that RX_USED is cleared as a side
effect of gem_rx_refill()'s work.

I'll apply this series, thanks for explaining.