RE: Regression v5.12-rc3: net: stmmac: re-init rx buffers when mac resume back

-----Original Message-----
From: Jon Hunter <jonathanh@nvidia.com>
Sent: 2021年3月25日 16:01
To: Joakim Zhang <redacted>
Cc: netdev@vger.kernel.org; Linux Kernel Mailing List
[off-list ref]; linux-tegra [off-list ref];
Jakub Kicinski [off-list ref]
Subject: Re: Regression v5.12-rc3: net: stmmac: re-init rx buffers when mac
resume back


On 25/03/2021 07:53, Joakim Zhang wrote:

quoted

quoted

-----Original Message-----
From: Jon Hunter <jonathanh@nvidia.com>
Sent: 2021年3月24日 20:39
To: Joakim Zhang <redacted>
Cc: netdev@vger.kernel.org; Linux Kernel Mailing List
[off-list ref]; linux-tegra
[off-list ref]; Jakub Kicinski [off-list ref]
Subject: Re: Regression v5.12-rc3: net: stmmac: re-init rx buffers
when mac resume back



On 24/03/2021 12:20, Joakim Zhang wrote:

...

quoted

Sorry for this breakage at your side.

You mean one of your boards? Does other boards with STMMAC can work

fine?

We have two devices with the STMMAC and one works OK and the other

fails.

quoted

quoted

They are different generation of device and so there could be some
architectural differences which is causing this to only be seen on one device.

It's really strange, but I also don't know what architectural differences could

affect this. Sorry.


Maybe caching somewhere? In other words, could there be any cache flushing
that we are missing here?

Have no idea, have not account into such case.

quoted

quoted

quoted

We do daily test with NFS to mount rootfs, on issue found. And I add
this

patch at the resume patch, and on error check, this should not break

suspend.

quoted

quoted

quoted

I even did the overnight stress test, there is no issue found.

Could you please do more test to see where the issue happen?

The issue occurs 100% of the time on the failing board and always on
the first resume from suspend. Is there any more debug I can enable
to track down what the problem is?

As commit messages described, the patch aims to re-init rx buffers
address, since the address is not fixed, so I only can recycle and then

re-allocate all of them. The page pool is allocated once when open the net
device.

quoted

Could you please debug if it fails at some functions, such as

page_pool_dev_alloc_pages() ?


Yes that was the first thing I tried, but no obvious failures from allocating the
pools.

Are you certain that the problem you are seeing, that is being fixed by this
change, is generic to all devices? The commit message states that 'descriptor
write back by DMA could exhibit unusual behavior', is this a known issue in the
STMMAC controller? If so does this impact all versions and what is the actual
problem?

Yes, I confirm this patch fix issue at my side. It should not be a generic, it can reproduce at one of our boards.
To be honest, I have not found the root cause, this should be a workaround, I upstream it since I think it will not affect others which don't suffer from this.

Best Regards,
Joakim Zhang

Jon

--
nvpublic