* [BUG] tcp: shift-out-of-bounds in inet_csk_rto_backoff() for an orphaned zero-window probe
@ 2026-11-18 10:59 Kyuwon Cho
2026-11-18 12:12 ` Eric Dumazet
0 siblings, 1 reply; 4+ messages in thread
From: Kyuwon Cho @ 2026-11-18 10:59 UTC (permalink / raw)
To: Eric Dumazet, Neal Cardwell
Cc: Kuniyuki Iwashima, David S. Miller, Jakub Kicinski, Paolo Abeni,
Simon Horman, Cambda Zhu, netdev, linux-kernel
[-- Attachment #1: Type: text/plain, Size: 2340 bytes --]
Hi Eric, Neal,
While testing v7.4-rc3 with an LLM-guided test generator that we are
developing as a research prototype, we found that an orphaned socket
that keeps sending zero-window probes makes inet_csk_rto_backoff()
shift by 64 or more. An unprivileged user can set this up in its own
user and network namespace:
UBSAN: shift-out-of-bounds in include/net/inet_connection_sock.h:258:40
shift exponent 64 is too large for 64-bit type 'long long unsigned int'
CPU: 0 UID: 0 PID: 0 Comm: swapper/0 Not tainted 7.4.0-rc3 #1 PREEMPT(lazy)
Call Trace:
<IRQ>
tcp_write_timer_handler.cold+0x54/0xc4
tcp_write_timer+0xf3/0x210
call_timer_fn+0x15c/0x4b0
__run_timers+0x5f4/0x970
run_timer_softirq+0x141/0x240
handle_softirqs+0x22d/0x980
__irq_exit_rcu+0x153/0x1e0
irq_exit_rcu+0x9/0x20
sysvec_apic_timer_interrupt+0x6c/0x80
</IRQ>
The probe path raises icsk_backoff up to tcp_retries2, which can be up
to 255:
if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
icsk->icsk_backoff++;
6d4634d1b091 ("net: Limit logical shift left of TCP probe0 timeout")
clamps the shift in tcp_probe0_when(), but tcp_probe_timer() also calls
inet_csk_rto_backoff() for an orphan, and that one is not clamped:
if (sock_flag(sk, SOCK_DEAD)) {
unsigned int rto_max = tcp_rto_max(sk);
const bool alive = inet_csk_rto_backoff(icsk, rto_max) < rto_max;
u64 when = (u64)icsk->icsk_rto << icsk->icsk_backoff;
On x86 the shift count is masked, so from 64 on alive can come out true
again and the orphan keeps probing instead of being aborted.
The reproducer runs as an unprivileged user (uid 65534). It creates a
user and network namespace, brings up lo, and sets tcp_retries2 and
tcp_orphan_retries to 255 there. A loopback server with a small receive
buffer never reads. The client sets TCP_RTO_MAX_MS to 1000 so probes
come about once a second, fills its send queue against the zero window,
and closes the socket. On an x86_64 QEMU guest with 2 vCPUs and KASAN
and UBSAN (gcc 13.3, config attached), the report fires in every run
(4 of 4), after about two minutes of probing.
Checked against v7.4-rc3 (aebb72744764); the code is the same in
net-next (6ff4c5942a30). We found no earlier report of this call site.
The reproducer (repro.c) and the kernel config are attached.
Thanks,
Kyuwon Cho
[-- Attachment #2: repro.c --]
[-- Type: text/plain, Size: 4922 bytes --]
// Reproducer for the tcp_probe_timer() orphan branch: inet_csk_rto_backoff()
// shifts a u64 by icsk_backoff (a u8) without the clamp that tcp_probe0_when()
// applies. Reaching icsk_backoff >= 64 makes the shift undefined (UBSAN
// shift-out-of-bounds).
//
// Steps, all as an unprivileged user inside a fresh user+net namespace:
// 1. unshare(CLONE_NEWUSER|CLONE_NEWNET), bring lo up.
// 2. Raise this netns' tcp_retries2 and tcp_orphan_retries to 255, so the
// zero-window probe backoff may grow past 64 and the orphan path keeps
// probing.
// 3. Loopback server with a tiny receive buffer that never reads, so the
// advertised window goes to zero.
// 4. Client sets TCP_RTO_MAX_MS=1000 (per-socket icsk_rto_max, minimum 1 s)
// so each zero-window probe fires about once a second, fills its send
// queue behind the zero window, then close()s: the socket becomes an
// orphan (SOCK_DEAD) with unsent data, and the probe timer keeps firing.
// 5. Wait ~120 s for ~100 probe timer expirations.
#define _GNU_SOURCE
#include <arpa/inet.h>
#include <errno.h>
#include <fcntl.h>
#include <net/if.h>
#include <netinet/in.h>
#include <netinet/tcp.h>
#include <sched.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/socket.h>
#include <sys/types.h>
#include <time.h>
#include <unistd.h>
#ifndef TCP_RTO_MAX_MS
#define TCP_RTO_MAX_MS 44
#endif
#define PORT 4242
#define WAIT_SECS 120
static void die(const char *what)
{
fprintf(stderr, "repro: %s: %s\n", what, strerror(errno));
exit(1);
}
static void write_file(const char *path, const char *val)
{
int fd = open(path, O_WRONLY);
if (fd < 0)
die(path);
if (write(fd, val, strlen(val)) < 0)
die(path);
close(fd);
}
int main(void)
{
uid_t uid = getuid();
gid_t gid = getgid();
char buf[64];
struct sockaddr_in addr;
struct ifreq ifr;
int ls, cs, s, rcvbuf = 2048, sndbuf = 4096, rto_max = 1000;
int on = 1;
char chunk[1024];
long total = 0;
time_t t0;
memset(chunk, 'x', sizeof(chunk));
if (unshare(CLONE_NEWUSER | CLONE_NEWNET))
die("unshare(CLONE_NEWUSER|CLONE_NEWNET)");
write_file("/proc/self/setgroups", "deny");
snprintf(buf, sizeof(buf), "0 %u 1\n", (unsigned)gid);
write_file("/proc/self/gid_map", buf);
snprintf(buf, sizeof(buf), "0 %u 1\n", (unsigned)uid);
write_file("/proc/self/uid_map", buf);
/* Bring loopback up (no ip binary in the VM). */
s = socket(AF_INET, SOCK_DGRAM, 0);
if (s < 0)
die("socket");
memset(&ifr, 0, sizeof(ifr));
strcpy(ifr.ifr_name, "lo");
if (ioctl(s, SIOCGIFFLAGS, &ifr))
die("SIOCGIFFLAGS");
ifr.ifr_flags |= IFF_UP | IFF_RUNNING;
if (ioctl(s, SIOCSIFFLAGS, &ifr))
die("SIOCSIFFLAGS lo");
close(s);
/* Per-netns sysctls: allow backoff past 64 and keep the orphan alive. */
write_file("/proc/sys/net/ipv4/tcp_retries2", "255");
write_file("/proc/sys/net/ipv4/tcp_orphan_retries", "255");
puts("repro: netns sysctls set (tcp_retries2=255, tcp_orphan_retries=255)");
fflush(stdout);
/* Listener: small receive buffer, never read from. */
ls = socket(AF_INET, SOCK_STREAM, 0);
if (ls < 0)
die("socket listen");
setsockopt(ls, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on));
setsockopt(ls, SOL_SOCKET, SO_RCVBUF, &rcvbuf, sizeof(rcvbuf));
memset(&addr, 0, sizeof(addr));
addr.sin_family = AF_INET;
addr.sin_port = htons(PORT);
addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
if (bind(ls, (struct sockaddr *)&addr, sizeof(addr)))
die("bind");
if (listen(ls, 4))
die("listen");
/* Client: per-socket RTO max of 1 s, small send buffer. */
cs = socket(AF_INET, SOCK_STREAM, 0);
if (cs < 0)
die("socket client");
setsockopt(cs, SOL_SOCKET, SO_SNDBUF, &sndbuf, sizeof(sndbuf));
if (setsockopt(cs, IPPROTO_TCP, TCP_RTO_MAX_MS, &rto_max, sizeof(rto_max)))
die("TCP_RTO_MAX_MS");
if (connect(cs, (struct sockaddr *)&addr, sizeof(addr)))
die("connect");
{
int as = accept(ls, NULL, NULL);
if (as < 0)
die("accept");
/* Server never reads: its receive queue fills, window -> 0. */
}
/* Fill the send queue behind the zero window. */
{
int fl = fcntl(cs, F_GETFL);
fcntl(cs, F_SETFL, fl | O_NONBLOCK);
}
for (;;) {
ssize_t n = write(cs, chunk, sizeof(chunk));
if (n < 0) {
if (errno == EAGAIN)
break;
die("write");
}
total += n;
if (total > (1 << 20))
break;
}
printf("repro: queued %ld bytes behind zero window\n", total);
fflush(stdout);
usleep(500 * 1000); /* let ACKs and the zero window settle */
/* Orphan the socket with unsent data: the probe timer now runs the
* SOCK_DEAD branch of tcp_probe_timer() on every expiration. */
close(cs);
puts("repro: client closed (orphan with unsent data); waiting for probe timers");
fflush(stdout);
t0 = time(NULL);
while (time(NULL) - t0 < WAIT_SECS) {
sleep(10);
printf("repro: %ld s elapsed\n", (long)(time(NULL) - t0));
fflush(stdout);
}
puts("repro: done");
return 0;
}
[-- Attachment #3: config.gz --]
[-- Type: application/gzip, Size: 38105 bytes --]
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [BUG] tcp: shift-out-of-bounds in inet_csk_rto_backoff() for an orphaned zero-window probe
2026-11-18 10:59 [BUG] tcp: shift-out-of-bounds in inet_csk_rto_backoff() for an orphaned zero-window probe Kyuwon Cho
@ 2026-11-18 12:12 ` Eric Dumazet
2026-11-18 12:40 ` Jiayuan Chen
2026-11-18 13:08 ` Kyuwon Cho
0 siblings, 2 replies; 4+ messages in thread
From: Eric Dumazet @ 2026-11-18 12:12 UTC (permalink / raw)
To: Kyuwon Cho, Jiayuan Chen
Cc: Neal Cardwell, Kuniyuki Iwashima, David S. Miller, Jakub Kicinski,
Paolo Abeni, Simon Horman, Cambda Zhu, netdev, linux-kernel
Le mer. 18 nov. 2026 à 11:59, Kyuwon Cho <kyuwoncho18@gmail.com> a écrit :
>
> Hi Eric, Neal,
>
> While testing v7.4-rc3 with an LLM-guided test generator that we are
> developing as a research prototype, we found that an orphaned socket
> that keeps sending zero-window probes makes inet_csk_rto_backoff()
> shift by 64 or more. An unprivileged user can set this up in its own
> user and network namespace:
Thanks for the report and the reproducer.
This is a known issue, a fix is under review:
[PATCH net v2 2/2] tcp: fix shift-out-of-bounds in inet_csk_rto_backoff()
https://lore.kernel.org/netdev/20261026123244.239766-2-jiayuan.chen@linux.dev/
It limits the shift to 32 in inet_csk_rto_backoff() itself, so it
covers both callers: tcp_ld_RTO_revert(), which is where Jiayuan hit
the splat, and the orphan check in tcp_probe_timer() from your report.
> We found no earlier report of this call site.
Searching lore for the function name would have found it:
https://lore.kernel.org/netdev/?q=inet_csk_rto_backoff
Please do this before sending a report, especially for findings
coming from automated tools.
Note that the netns sysctls are not needed. tcp_retransmit_timer()
increments icsk_backoff without any limit, so TCP_USER_TIMEOUT and
TCP_RTO_MAX_MS on a socket with a silent peer are enough to get
icsk_backoff above 63, then an ICMP unreachable reaches the shift
from tcp_ld_RTO_revert().
> On x86 the shift count is masked, so from 64 on alive can come out true
> again and the orphan keeps probing instead of being aborted.
Right, and this starts before 64: with icsk_rto == 200, the u64
is already zero when icsk_backoff reaches 61.
Jiayuan, I have not seen a v3 addressing Neal's feedback.
When you send it, could you please add:
Reported-by: Kyuwon Cho <kyuwoncho18@gmail.com>
Closes: https://lore.kernel.org/netdev/179499954853.1440581.6627074959099619480@gmail.com/
and mention the zero window probe path in the changelog ?
Kyuwon, it would be nice if you could test Jiayuan's patch with your
reproducer and send a Tested-by: tag in that thread.
Thanks.
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [BUG] tcp: shift-out-of-bounds in inet_csk_rto_backoff() for an orphaned zero-window probe
2026-11-18 12:12 ` Eric Dumazet
@ 2026-11-18 12:40 ` Jiayuan Chen
2026-11-18 13:08 ` Kyuwon Cho
1 sibling, 0 replies; 4+ messages in thread
From: Jiayuan Chen @ 2026-11-18 12:40 UTC (permalink / raw)
To: Eric Dumazet, Kyuwon Cho
Cc: Neal Cardwell, Kuniyuki Iwashima, David S. Miller, Jakub Kicinski,
Paolo Abeni, Simon Horman, Cambda Zhu, netdev, linux-kernel
On 11/18/26 8:12 PM, Eric Dumazet wrote:
> Le mer. 18 nov. 2026 à 11:59, Kyuwon Cho <kyuwoncho18@gmail.com> a écrit :
>> Hi Eric, Neal,
>>
>> While testing v7.4-rc3 with an LLM-guided test generator that we are
>> developing as a research prototype, we found that an orphaned socket
>> that keeps sending zero-window probes makes inet_csk_rto_backoff()
>> shift by 64 or more. An unprivileged user can set this up in its own
>> user and network namespace:
> Thanks for the report and the reproducer.
>
> This is a known issue, a fix is under review:
>
> [PATCH net v2 2/2] tcp: fix shift-out-of-bounds in inet_csk_rto_backoff()
> https://lore.kernel.org/netdev/20261026123244.239766-2-jiayuan.chen@linux.dev/
>
> It limits the shift to 32 in inet_csk_rto_backoff() itself, so it
> covers both callers: tcp_ld_RTO_revert(), which is where Jiayuan hit
> the splat, and the orphan check in tcp_probe_timer() from your report.
>
>> We found no earlier report of this call site.
> Searching lore for the function name would have found it:
>
> https://lore.kernel.org/netdev/?q=inet_csk_rto_backoff
>
> Please do this before sending a report, especially for findings
> coming from automated tools.
>
> Note that the netns sysctls are not needed. tcp_retransmit_timer()
> increments icsk_backoff without any limit, so TCP_USER_TIMEOUT and
> TCP_RTO_MAX_MS on a socket with a silent peer are enough to get
> icsk_backoff above 63, then an ICMP unreachable reaches the shift
> from tcp_ld_RTO_revert().
>
>> On x86 the shift count is masked, so from 64 on alive can come out true
>> again and the orphan keeps probing instead of being aborted.
> Right, and this starts before 64: with icsk_rto == 200, the u64
> is already zero when icsk_backoff reaches 61.
>
> Jiayuan, I have not seen a v3 addressing Neal's feedback.
>
> When you send it, could you please add:
>
> Reported-by: Kyuwon Cho <kyuwoncho18@gmail.com>
> Closes: https://lore.kernel.org/netdev/179499954853.1440581.6627074959099619480@gmail.com/
>
> and mention the zero window probe path in the changelog ?
Sure I can do that.
>
> Kyuwon, it would be nice if you could test Jiayuan's patch with your
> reproducer and send a Tested-by: tag in that thread.
>
> Thanks.
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [BUG] tcp: shift-out-of-bounds in inet_csk_rto_backoff() for an orphaned zero-window probe
2026-11-18 12:12 ` Eric Dumazet
2026-11-18 12:40 ` Jiayuan Chen
@ 2026-11-18 13:08 ` Kyuwon Cho
1 sibling, 0 replies; 4+ messages in thread
From: Kyuwon Cho @ 2026-11-18 13:08 UTC (permalink / raw)
To: Eric Dumazet, Jiayuan Chen
Cc: Neal Cardwell, Kuniyuki Iwashima, David S. Miller, Jakub Kicinski,
Paolo Abeni, Simon Horman, Cambda Zhu, netdev, linux-kernel
On Wed, Nov 18, 2026 at 01:12:45PM +0100, Eric Dumazet wrote:
> Searching lore for the function name would have found it:
>
> https://lore.kernel.org/netdev/?q=inet_csk_rto_backoff
>
> Please do this before sending a report, especially for findings
> coming from automated tools.
You are right, sorry for the noise. We now search lore for the
functions involved before sending a report.
> Kyuwon, it would be nice if you could test Jiayuan's patch with your
> reproducer and send a Tested-by: tag in that thread.
Sure. I'll run it against Jiayuan's v3 once it is posted and send the
Tested-by in that thread.
Thanks,
Kyuwon Cho
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2026-11-18 13:08 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 10:59 [BUG] tcp: shift-out-of-bounds in inet_csk_rto_backoff() for an orphaned zero-window probe Kyuwon Cho
2026-11-18 12:12 ` Eric Dumazet
2026-11-18 12:40 ` Jiayuan Chen
2026-11-18 13:08 ` Kyuwon Cho
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox