Re: issue with inflight pages from page_pool

[Lorenzo Bianconi <lorenzo@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 1603 bytes --]

> On Mon, Apr 17, 2023 at 7:53 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> >
> > Hi all,
> >
> > I am triggering an issue with a device running the page_pool allocator.
> > In particular, the device is running an iperf tcp server receiving traffic
> > from a remote client. On the driver I loaded a simple xdp program returning
> > xdp_pass. When I remove the ebpf program and destroy the pool, page_pool
> > allocator starts complaining in page_pool_release_retry() that not all the pages
> > have been returned to the allocator. In fact, the pool is not really destroyed
> > in this case.
> > Debugging the code it seems the pages are stuck softnet_data defer_list and
> > they are never freed in skb_defer_free_flush() since I do not have any more tcp
> > traffic. To prove it, I tried to set sysctl_skb_defer_max to 0 and the issue
> > does not occur.
> > I developed the poc patch below and the issue seems to be fixed:
> 
> I do not see why this would be different than having buffers sitting
> in some tcp receive
> (or out or order) queues for a few minutes ?

The main issue in my tests (and even in mt76 I think) is the pages are not returned
to the pool for a very long time (even hours) and doing so the pool is like in a
'limbo' state where it is not actually deallocated and page_pool_release_retry
continues complaining about it. I think this is because we do not have more tcp
traffic to deallocate them, but I am not so familiar with tcp codebase :)

Regards,
Lorenzo

> 
> Or buffers transferred to another socket or pipe (splice() and friends)

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]