On Wed, 2021-09-01 at 18:39 +0800, Yunsheng Lin wrote:
> Since tcp_tx_skb_cache is disabled by default in:
> commit 0b7d7f6b2208 ("tcp: add tcp_tx_skb_cache sysctl")
>
> Add tcp_tx_skb_cache_key checking in sk_stream_alloc_skb() to
> avoid possible branch-misses.
>
> Signed-off-by: Yunsheng Lin <[email protected]>
Note that MPTCP is currently exploiting sk->sk_tx_skb_cache. If we get
this patch goes in as-is, it will break mptcp.
One possible solution would be to let mptcp usage enable sk-
>sk_tx_skb_cache, but that has relevant side effects on plain TCP.
Another options would be re-work once again the mptcp xmit path to
avoid using sk->sk_tx_skb_cache.
Cheers,
Paolo
On Wed, Sep 1, 2021 at 3:52 AM Paolo Abeni <[email protected]> wrote:
>
> On Wed, 2021-09-01 at 18:39 +0800, Yunsheng Lin wrote:
> > Since tcp_tx_skb_cache is disabled by default in:
> > commit 0b7d7f6b2208 ("tcp: add tcp_tx_skb_cache sysctl")
> >
> > Add tcp_tx_skb_cache_key checking in sk_stream_alloc_skb() to
> > avoid possible branch-misses.
> >
> > Signed-off-by: Yunsheng Lin <[email protected]>
>
> Note that MPTCP is currently exploiting sk->sk_tx_skb_cache. If we get
> this patch goes in as-is, it will break mptcp.
>
> One possible solution would be to let mptcp usage enable sk-
> >sk_tx_skb_cache, but that has relevant side effects on plain TCP.
>
> Another options would be re-work once again the mptcp xmit path to
> avoid using sk->sk_tx_skb_cache.
>
Hmmm, I actually wrote a revert of this feature but forgot to submit
it last year.
commit c36cfbd791f62c0f7c6b32132af59dfdbe6be21b (HEAD -> listener_scale4)
Author: Eric Dumazet <[email protected]>
Date: Wed May 20 06:38:38 2020 -0700
tcp: remove sk_{tr}x_skb_cache
This reverts the following patches :
2e05fcae83c41eb2df10558338dc600dc783af47 ("tcp: fix compile error
if !CONFIG_SYSCTL")
4f661542a40217713f2cee0bb6678fbb30d9d367 ("tcp: fix zerocopy and
notsent_lowat issues")
472c2e07eef045145bc1493cc94a01c87140780a ("tcp: add one skb cache for tx")
8b27dae5a2e89a61c46c6dbc76c040c0e6d0ed4c ("tcp: add one skb cache for rx")
Having a cache of one skb (in each direction) per TCP socket is fragile,
since it can cause a significant increase of memory needs,
and not good enough for high speed flows anyway where more than one skb
is needed.
We want instead to add a generic infrastructure, with more flexible per-cpu
caches, for alien NUMA nodes.
Signed-off-by: Eric Dumazet <[email protected]>
I will update this commit to also remove the part in MPTCP.
Let's remove this feature and replace it with something less costly.
On Wed, Sep 1, 2021 at 8:06 AM Eric Dumazet <[email protected]> wrote:
>
> On Wed, Sep 1, 2021 at 3:52 AM Paolo Abeni <[email protected]> wrote:
> >
> > On Wed, 2021-09-01 at 18:39 +0800, Yunsheng Lin wrote:
> > > Since tcp_tx_skb_cache is disabled by default in:
> > > commit 0b7d7f6b2208 ("tcp: add tcp_tx_skb_cache sysctl")
> > >
> > > Add tcp_tx_skb_cache_key checking in sk_stream_alloc_skb() to
> > > avoid possible branch-misses.
> > >
> > > Signed-off-by: Yunsheng Lin <[email protected]>
> >
> > Note that MPTCP is currently exploiting sk->sk_tx_skb_cache. If we get
> > this patch goes in as-is, it will break mptcp.
> >
> > One possible solution would be to let mptcp usage enable sk-
> > >sk_tx_skb_cache, but that has relevant side effects on plain TCP.
> >
> > Another options would be re-work once again the mptcp xmit path to
> > avoid using sk->sk_tx_skb_cache.
> >
>
> Hmmm, I actually wrote a revert of this feature but forgot to submit
> it last year.
>
> commit c36cfbd791f62c0f7c6b32132af59dfdbe6be21b (HEAD -> listener_scale4)
> Author: Eric Dumazet <[email protected]>
> Date: Wed May 20 06:38:38 2020 -0700
>
> tcp: remove sk_{tr}x_skb_cache
>
> This reverts the following patches :
>
> 2e05fcae83c41eb2df10558338dc600dc783af47 ("tcp: fix compile error
> if !CONFIG_SYSCTL")
> 4f661542a40217713f2cee0bb6678fbb30d9d367 ("tcp: fix zerocopy and
> notsent_lowat issues")
> 472c2e07eef045145bc1493cc94a01c87140780a ("tcp: add one skb cache for tx")
> 8b27dae5a2e89a61c46c6dbc76c040c0e6d0ed4c ("tcp: add one skb cache for rx")
>
> Having a cache of one skb (in each direction) per TCP socket is fragile,
> since it can cause a significant increase of memory needs,
> and not good enough for high speed flows anyway where more than one skb
> is needed.
>
> We want instead to add a generic infrastructure, with more flexible per-cpu
> caches, for alien NUMA nodes.
>
> Signed-off-by: Eric Dumazet <[email protected]>
>
> I will update this commit to also remove the part in MPTCP.
>
> Let's remove this feature and replace it with something less costly.
Paolo, can you work on MPTP side, so that my revert can be then applied ?
Thanks !