Re: [PATCH net-next v6 1/8] net: add get_netmem/put_netmem support

[Jakub Kicinski <kuba@kernel.org>]

On Fri, 28 Feb 2025 17:29:13 -0800 Mina Almasry wrote:
> On Fri, Feb 28, 2025 at 4:38 PM Jakub Kicinski <kuba@kernel.org> wrote:
> > On Thu, 27 Feb 2025 04:12:02 +0000 Mina Almasry wrote:  
> > >  static inline void __skb_frag_ref(skb_frag_t *frag)
> > >  {
> > > -     get_page(skb_frag_page(frag));
> > > +     get_netmem(skb_frag_netmem(frag));
> > >  }  
> >
> > Silently handling types of memory the caller may not be expecting
> > always worries me.  
> 
> Sorry, I'm not following. What caller is not expecting netmem?
> Here we're making sure __skb_frag_ref() handles netmem correctly,
> i.e. we were not expecting netmem here before, and after this patch
> we'll handle it correctly.
> 
> > Why do we need this?
> >  
> 
> The MSG_ZEROCOPY TX path takes a page reference on the passed memory
> in zerocopy_fill_skb_from_iter() that kfree_skb() later drops when the
> skb is sent. We need an equivalent for netmem, which only supports pp
> refs today. This is my attempt at implementing a page_ref equivalent
> to net_iov and generic netmem.
> 
> I think __skb_frag_[un]ref is used elsewhere in the TX path too,
> tcp_mtu_probe for example calls skb_frag_ref eventually.

Any such caller must be inspected to make sure it generates
/ anticipates skbs with appropriate pp_recycle and readable settings.
It's possible that adding a set of _netmem APIs would be too much
churn, but if it's not - it'd make it clear which parts of the kernel
we have inspected.

> > In general, I'm surprised by the lack of bug reports for devmem.  
> 
> I guess we did a good job making sure we don't regress the page paths.

:)

> The lack of support in any driver that qemu will run is an issue. I
> wonder if also the fact that devmem needs some setup is also an issue.
> We need headersplit enabled, udmabuf created, netlink API bound, and
> then a connection referring to created and we don't support loopback.
> I think maybe it all may make it difficult for syzbot to repro. I've
> had it on my todo list to investigate this more.
> 
> > Can you think of any way we could expose this more to syzbot?
> > First thing that comes to mind is a simple hack in netdevsim,
> > to make it insert a netmem handle (allocated locally, not a real
> > memory provider), every N packets (controllable via debugfs).
> > Would that work?  
> 
> Yes, great idea. I don't see why it wouldn't work.
> 
> We don't expect mixing of net_iovs and pages in the same skb, but
> netdevsim could create one net_iov skb every N skbs.
> 
> I guess I'm not totally sure something is discoverable to syzbot. Is a
> netdevsim hack toggleable via a debugfs sufficient for syzbot? I'll
> investigate and ask.

Yeah, my unreliable memory is that syzbot has a mixed record discovering
problems with debugfs. If you could ask Dmitry for advice that'd be
ideal.