From: Kuniyuki Iwashima <kuniyu@google.com>
To: Eric Dumazet <edumazet@kernel.org>
Cc: Neal Cardwell <ncardwell@google.com>,
"David S . Miller" <davem@davemloft.net>,
Jakub Kicinski <kuba@kernel.org>,
Paolo Abeni <pabeni@redhat.com>,
Matthieu Baerts <matttbe@kernel.org>,
Mat Martineau <martineau@kernel.org>,
Simon Horman <horms@kernel.org>,
Kuniyuki Iwashima <kuni1840@gmail.com>,
netdev@vger.kernel.org, mptcp@lists.linux.dev
Subject: Re: [PATCH v3 net-next 5/7] tcp: Allocate all new skb before collapsing.
Date: Thu, 19 Nov 2026 20:02:59 +0000 [thread overview]
Message-ID: <av9U9w4wGpO7HmDY@google.com> (raw)
In-Reply-To: <CAL4WiioNduj1AybypLPDs6HuW2F9Y9zw3B2ycLSogRMnfzBk9g@mail.gmail.com>
On Thu, 19 Nov 2026 08:44:54 +0100, Eric Dumazet wrote:
> Le jeu. 19 nov. 2026 à 08:17, Kuniyuki Iwashima <kuniyu@google.com> a écrit :
> >
> > At the final stage, tcp_collapse() iterates over collapsible
> > candidate skbs and repacks them to new skbs.
> >
> > In each iteration, a new skb is allocated and inserted to the
> > receive queue or a temporary list for RB tree.
> >
> > If alloc_skb() fails while splitting a large bloated skb across
> > multiple new skbs, the partially-copied skb is not removed and
> > new skbs remain in the queue as duplicates, consuming more memory.
> >
> > It is only when all candidates are properly collapsed that
> > we can save memory.
> >
> > To avoid complicated rollback, let's preallocate skbs before
> > collapsing.
> >
> > Now all skb in [start, sub_end) must be allocated in a batch,
> > so the section size is capped at TCP_COLLAPSE_MAX_TOTAL_SIZE
> > (16 skbs), except for the first single skb.
> >
> > Once the size exceeds the limit, the remaining skbs are
> > re-evaluated from restart:.
> >
> > Legitimate overlaps (retransmits with a different segmentation)
> > only involve a few skbs, so if an overlapping OOO chain reaches
> > the limit, we simply skip it.
> >
> > Note that the next skb can be skipped if its density already meets
> > tp->scaling_ratio, and each cut adds at most one remainder skb
> > (~3.4% worst-case truesize overhead on 4 KiB pages), which is a
> > small tradeoff to bound the temporary allocation.
> >
> > Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
> > ---
> > v3: Make TCP_COLLAPSE_MAX_TOTAL_SIZE hard-limit for multiple skbs
> > ---
> > net/ipv4/tcp_input.c | 41 ++++++++++++++++++++++++++++++++---------
> > 1 file changed, 32 insertions(+), 9 deletions(-)
> >
> > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> > index 27ed9ba11ec9..990e34896373 100644
> > --- a/net/ipv4/tcp_input.c
> > +++ b/net/ipv4/tcp_input.c
> > @@ -5822,6 +5822,11 @@ static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *li
> > #define TCP_COLLAPSE_SCALING_RATIO \
> > ((TCP_COLLAPSE_MAX_SIZE << TCP_RMEM_TO_WIN_SCALE) / \
> > (SKB_TRUESIZE(TCP_COLLAPSE_MAX_SIZE)))
> > +/* Cap pool preallocation per sub-range to at most 16 skbs in seq
> > + * space. A single first skb exceeding this is still collapsed,
> > + * whereas an exceeding overlapping ofo range is skipped.
> > + */
> > +#define TCP_COLLAPSE_MAX_TOTAL_SIZE (TCP_COLLAPSE_MAX_SIZE * 16)
> >
> > static bool tcp_collapse_skb_bloated(const struct sk_buff *skb,
> > u32 start, u8 scaling_ratio)
> > @@ -5862,8 +5867,9 @@ tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
> > {
> > unsigned int old_truesize, new_truesize;
> > struct sk_buff *skb = head, *n;
> > - struct sk_buff_head tmp;
> > + struct sk_buff_head tmp, pool;
> > int remainder, nskbs;
> > + struct sk_buff *nskb;
> > u8 scaling_ratio;
> > u32 sub_end;
> >
> > @@ -5871,6 +5877,7 @@ tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
> > TCP_COLLAPSE_SCALING_RATIO);
> >
> > __skb_queue_head_init(&tmp);
> > + __skb_queue_head_init(&pool);
> >
> > /* First, check that queue is collapsible and find
> > * the point where collapsing can be useful.
> > @@ -5926,8 +5933,9 @@ tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
> > }
> >
> > if (!overlap &&
> > - !tcp_collapse_skb_bloated(n, sub_end,
> > - TCP_COLLAPSE_SCALING_RATIO))
> > + (TCP_SKB_CB(n)->end_seq - start >= TCP_COLLAPSE_MAX_TOTAL_SIZE ||
> > + !tcp_collapse_skb_bloated(n, sub_end,
> > + TCP_COLLAPSE_SCALING_RATIO)))
> > break;
> >
> > sub_end = TCP_SKB_CB(n)->end_seq;
> > @@ -5935,25 +5943,38 @@ tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
> > n = tcp_skb_next(n, list);
> > }
> >
> > + if (sub_end != TCP_SKB_CB(skb)->end_seq &&
> > + sub_end - start >= TCP_COLLAPSE_MAX_TOTAL_SIZE)
> > + goto skip;
>
> This is fine for now, but note that a single 64K GRO skb overlapping
> a bloated skb is enough to reach the limit, and the bloated skb is
> then never collapsed. The range is walked again at every
> tcp_collapse_ofo_queue() call, for no gain.
>
> A possible followup would be to cut the range instead of skipping it.
>
> When @n overlaps and we do not want (or can not) merge it, we know
> that @n already contains [TCP_SKB_CB(n)->seq, sub_end), because
> skbs with end_seq <= sub_end were removed earlier in the loop.
>
> So we could do :
>
> sub_end = TCP_SKB_CB(n)->seq;
>
> and collapse [start, sub_end) only, as long as sub_end is after start.
>
> All new skbs would then have a seq smaller than TCP_SKB_CB(n)->seq,
> so the rbtree ordering is preserved, and we would not need to drag
> @n into the range.
>
> After the copy, the skbs of the range that were not fully consumed
> would have to be freed, since the rest of their payload is in @n.
>
> The same trick would also work for the unmergeable overlapping skb
> case added in patch 3.
>
> I have not tested this, and this can be done later.
Ah, this sounds good too, I'll follow up after your frag series.
Thank you !
next prev parent reply other threads:[~2026-11-19 20:03 UTC|newest]
Thread overview: 14+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-19 7:15 [PATCH v3 net-next 0/7] tcp: Prevent tcp_collapse() from bloating queue Kuniyuki Iwashima
2026-11-19 7:15 ` [PATCH v3 net-next 1/7] tcp: Refactor the first loop in tcp_collapse() Kuniyuki Iwashima
2026-11-19 7:15 ` [PATCH v3 net-next 2/7] tcp: Don't collapse high density skb Kuniyuki Iwashima
2026-11-19 7:15 ` [PATCH v3 net-next 3/7] tcp: Stop collapsing if not bloated Kuniyuki Iwashima
2026-11-19 7:39 ` Eric Dumazet
2026-11-20 14:58 ` Matthieu Baerts
2026-11-19 7:15 ` [PATCH v3 net-next 4/7] tcp: Skip collapsing if collapsed skb is less-dense Kuniyuki Iwashima
2026-11-19 7:41 ` Eric Dumazet
2026-11-19 7:15 ` [PATCH v3 net-next 5/7] tcp: Allocate all new skb before collapsing Kuniyuki Iwashima
2026-11-19 7:44 ` Eric Dumazet
2026-11-19 20:02 ` Kuniyuki Iwashima [this message]
2026-11-19 7:15 ` [PATCH v3 net-next 6/7] tcp: Batch memory accounting in tcp_collapse() Kuniyuki Iwashima
2026-11-19 7:15 ` [PATCH v3 net-next 7/7] selftest: packetdrill: Add test for tcp_collapse() Kuniyuki Iwashima
2026-11-19 7:45 ` Eric Dumazet
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=av9U9w4wGpO7HmDY@google.com \
--to=kuniyu@google.com \
--cc=davem@davemloft.net \
--cc=edumazet@kernel.org \
--cc=horms@kernel.org \
--cc=kuba@kernel.org \
--cc=kuni1840@gmail.com \
--cc=martineau@kernel.org \
--cc=matttbe@kernel.org \
--cc=mptcp@lists.linux.dev \
--cc=ncardwell@google.com \
--cc=netdev@vger.kernel.org \
--cc=pabeni@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.