On Mon, Jul 07, 2025 at 11:40:48AM -0700, Stanislav Fomichev wrote:
> On 07/07, Alexander Lobakin wrote:
> > From: Stanislav Fomichev <stfomichev@xxxxxxxxx>
> > Date: Mon, 7 Jul 2025 08:06:21 -0700
> >
> > > On 07/07, Alexander Lobakin wrote:
> > >> From: Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx>
> > >> Date: Sat, 5 Jul 2025 15:55:12 +0200
> > >>
> > >>> Eryk reported an issue that I have put under Closes: tag, related to
> > >>> umem addrs being prematurely produced onto pool's completion queue.
> > >>> Let us make the skb's destructor responsible for producing all addrs
> > >>> that given skb used.
> > >>>
> > >>> Commit from fixes tag introduced the buggy behavior, it was not broken
> > >>> from day 1, but rather when xsk multi-buffer got introduced.
> > >>>
> > >>> Introduce a struct which will carry descriptor count with array of
> > >>> addresses taken from processed descriptors that will be carried via
> > >>> skb_shared_info::destructor_arg. This way we can refer to it within
> > >>> xsk_destruct_skb().
> > >>>
> > >>> To summarize, behavior is changed from:
> > >>> - produce addr to cq, increase cq's cached_prod
> > >>> - increment descriptor count and store it on
> > >>> - (xmit and rest of path...)
> > >>> skb_shared_info::destructor_arg
> > >>> - use destructor_arg on skb destructor to update global state of cq
> > >>> producer
> > >>>
> > >>> to the following:
> > >>> - increment cq's cached_prod
> > >>> - increment descriptor count, save xdp_desc::addr in custom array and
> > >>> store this custom array on skb_shared_info::destructor_arg
> > >>> - (xmit and rest of path...)
> > >>> - use destructor_arg on skb destructor to walk the array of addrs and
> > >>> write them to cq and finally update global state of cq producer
> > >>>
> > >>> Fixes: b7f72a30e9ac ("xsk: introduce wrappers and helpers for supporting multi-buffer in Tx path")
> > >>> Reported-by: Eryk Kubanski <e.kubanski@xxxxxxxxxxxxxxxxxxx>
> > >>> Closes: https://lore.kernel.org/netdev/20250530103456.53564-1-e.kubanski@xxxxxxxxxxxxxxxxxxx/
> > >>> Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@xxxxxxxxx>
> > >>> ---
> > >>> v1:
> > >>> https://lore.kernel.org/bpf/20250702101648.1942562-1-maciej.fijalkowski@xxxxxxxxx/
> > >>>
> > >>> v1->v2:
> > >>> * store addrs in array carried via destructor_arg instead having them
> > >>> stored in skb headroom; cleaner and less hacky approach;
> > >>
> > >> Might look cleaner, but what about the performance given that you're
> > >> adding a memory allocation?
> > >>
> > >> (I realize that's only for the skb mode, still)
> > >>
> > >> Yeah we anyway allocate an skb and may even copy the whole frame, just
> > >> curious.
> > >> I could recommend using skb->cb for that, but its 48 bytes would cover
> > >> only 6 addresses =\
> >
> > BTW isn't num_descs from that new structure would be the same as
> > shinfo->nr_frags + 1 (or just nr_frags for xsk_build_skb_zerocopy())?
>
> So you're saying we don't need to store it? Agreed. But storing the rest
> in cb still might be problematic with kconfig-configurable MAX_SKB_FRAGS?
Hi Stan & Olek,
no, as said in v1 drivers might linearize the skb and all frags will be
lost. This storage is needed unfortunately.
>
> > > Can we pre-allocate an array of xsk_addrs during xsk_bind (the number of
> > > xsk_addrs is bound by the tx ring size)? Then we can remove the alloc on tx
> > > and replace it with some code to manage that pool of xsk_addrs..
That would be pool-bound which makes it a shared resource so I believe
that we would repeat the problem being fixed here ;)
> >
> > Nice idea BTW.
> >
> > We could even use system per-cpu Page Pools to allocate these structs*
> > :D It wouldn't waste 1 page per one struct as PP is frag-aware and has
> > API for allocating only a small frag.
> >
> > Headroom stuff was also ok to me: we either way allocate a new skb, so
> > we could allocate it with a bit bigger headroom and put that table there
> > being sure that nobody will overwrite it (some drivers insert special
> > headers or descriptors in front of the actual skb->data).
headroom approach was causing one of bpf selftests to fail, but I didn't
check in-depth the reason. I didn't really like the check in destructor if
addr array was corrupted in v1 and I came up with v2 which seems to me a
cleaner fix.
> >
> > [*] Offtop: we could also use system PP to allocate skbs in
> > xsk_build_skb() just like it's done in xdp_build_skb_from_zc() +
> > xdp_copy_frags_from_zc() -- no way to avoid memcpy(), but the payload
> > buffers would be recycled then.
>
> Or maybe kmem_cache_alloc_node with a custom cache is good enough?
> Headroom also feels ok if we store the whole xsk_addrs struct in it.
Yep both of these approaches was something I considered, but keep in mind
it's a bugfix so I didn't want to go with something flashy. I have not
observed big performance impact but I checked only MAX_SKB_FRAGS being set
to standard value.
Would you guys be ok if I do the follow-up with possible optimization
after my vacation which would be a -next candidate?
Thanks,
MF
