Re: [PATCH net] tcp: fix negative sk_forward_alloc after MSG_ZEROCOPY append
From: Eric Dumazet
Date: Thu Nov 19 2026 - 05:06:54 EST
Le jeu. 19 nov. 2026 à 10:08, Jiayuan Chen <jiayuan.chen@xxxxxxxxx> a écrit :
>
> Our local syzbot reported the following warning:
>
> WARNING: ./include/linux/ucopysize.h:57 at tcp_sendmsg_locked+0x1e13/0x1f20
> Modules linked in:
> CPU: 3 UID: 0 PID: 377 Comm: syz.2.2248 Not tainted 7.4.0-rc2+ #398 PREEMPT
> RIP: 0010:tcp_sendmsg_locked+0x1e13/0x1f20
> Call Trace:
> <TASK>
> tcp_sendmsg+0x2c/0x50
> inet_sendmsg+0x62/0xb0
> __sys_sendto+0x2fe/0x350
> __x64_sys_sendto+0x7f/0xb0
> x64_sys_call+0x1375/0x27e0
> do_syscall_64+0xe4/0x470
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> </TASK>
>
> The problem is here:
>
> if (!skb_zcopy_pure(skb)) {
> copy = tcp_wmem_schedule(sk, copy);
> if (!copy)
> goto wait_for_space;
> }
>
> When MSG_ZEROCOPY data is appended to a skb that already holds copied
> data, only the user length is scheduled, but the pinned pages are then
> charged by their truesize:
>
> skb_zerocopy_iter_stream
> __zerocopy_sg_from_iter
> zerocopy_fill_skb_from_iter //truesize += PAGE_ALIGN(copied + start)
> sk_mem_charge(sk, truesize) //sk_forward_alloc goes negative
>
> For example, sending 8 iovecs of 1 byte each schedules 8 bytes and
> charges 32768 with 4K pages.
This part is a separate bug: these 8 bytes sit in a single page.
zerocopy_fill_skb_from_iter() adds PAGE_ALIGN(copied + start) to
skb->truesize for each iov_iter_get_pages2() call, even when the
chunk is merged into the previous frag and the page reference is
released.
With a corked socket, one copied byte, then one MSG_ZEROCOPY sendmsg()
of 1024 bytes from a single page:
1 iovec sk_wmem_queued += 4096
1024 iovecs sk_wmem_queued += 4194304 (sk_sndbuf is 2626560)
With your patch, this over-estimation would become a forced charge,
above tcp_mem and memcg limits, with an amount controlled by the user.
I have a fix [1] for this accounting (only charge the pages a frag did not
already cover), I will send it shortly. With it, both cases above
charge 4096 bytes, and sk_forward_alloc is still negative after the
append, so your patch is still needed.
Thanks.
>
> On a later sendmsg() without MSG_ZEROCOPY, if sk_wmem_schedule() fails
> (e.g. tcp_mem hard limit is reached), tcp_wmem_schedule() returns
> min(copy, sk->sk_forward_alloc), which is negative. The negative copy
> is passed to skb_copy_to_page_nocache() and triggers the warning, and
> sendmsg() fails with -EFAULT.
>
> A few options were considered:
>
> 1) Schedule the truesize up front. It is not known before the pages
> are pinned, it depends on the iovec layout.
> 2) Do not append zerocopy data to a non pure skb. This splits a small
> header and its zerocopy payload into two skbs.
> 3) Schedule the missing part right after the pages are charged. They
> are already in the skb at this point, so this can not fail.
>
> Go with 3).
Analysis looks correct, and 3) is consistent with what we did
for tcp_collapse() recently.
>
> Fixes: 849b425cd091 ("tcp: fix possible freeze in tx path under memory pressure")
> Signed-off-by: Jiayuan Chen <jiayuan.chen@xxxxxxxxx>
> ---
> net/ipv4/tcp.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index d510e982dcea..8bd377542a89 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -1359,6 +1359,8 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
> if (err < 0)
> goto do_error;
> copy = err;
> + /* Pages were charged by truesize, not by copy. */
> + sk_forced_mem_schedule(sk, 0);
Pure zerocopy skbs are not charged, so there is nothing to fix for
them. Please avoid the extra work in this case:
if (!skb_zcopy_pure(skb))
sk_forced_mem_schedule(sk, 0);
> } else if (zc == MSG_SPLICE_PAGES) {
> /* Splice in data if we can; copy if we can't. */
> if (tcp_downgrade_zcopy_pure(sk, skb))
> --
> 2.43.0
>
[1]
diff --git a/net/core/datagram.c b/net/core/datagram.c
index 489d92fe6dae8952da08bc5cb2a58f9863938fe0..8c418c3adf1c97b41ee7999146d09aacd8cf29d4
100644
--- a/net/core/datagram.c
+++ b/net/core/datagram.c
@@ -659,7 +659,6 @@ int zerocopy_fill_skb_from_iter(struct sk_buff *skb,
skb->data_len += copied;
skb->len += copied;
- skb->truesize += PAGE_ALIGN(copied + start);
head = compound_head(pages[n]);
order = compound_order(head);
@@ -680,6 +679,11 @@ int zerocopy_fill_skb_from_iter(struct sk_buff *skb,
if (head == skb_frag_page(last) &&
start == skb_frag_off(last) +
skb_frag_size(last)) {
+ /* Only account the pages this frag
+ * did not already cover.
+ */
+ skb->truesize +=
PAGE_ALIGN(start + size) -
+ PAGE_ALIGN(start);
skb_frag_size_add(last, size);
/* We combined this page, we
need to release
* a reference. Since compound
pages refcount
@@ -695,6 +699,7 @@ int zerocopy_fill_skb_from_iter(struct sk_buff *skb,
page_ref_sub(last_head, refs);
refs = 0;
}
+ skb->truesize += PAGE_SIZE;
skb_fill_page_desc_noacc(skb, frag++, head,
start, size);
}
if (refs)