[BUG] netdevsim: nsim_start_peer_tx_queue() indexes the peer's TX queues with an RX queue index
From: Kyuwon Cho
Date: Wed Nov 18 2026 - 22:05:48 EST
Hi Jakub,
While testing v7.4-rc3 with an LLM-guided test generator that we are
developing as a research prototype, we found that netdevsim can read
past the peer's TX queue array when it wakes the peer's queue from
NAPI poll:
BUG: KASAN: slab-out-of-bounds in nsim_poll (include/linux/netdevice.h:3834 drivers/net/netdevsim/netdev.c:60 drivers/net/netdevsim/netdev.c:409 drivers/net/netdevsim/netdev.c:418)
Read of size 8 at addr ff110000140a0fd0 by task test/118
Call Trace:
<IRQ>
nsim_poll (include/linux/netdevice.h:3834 drivers/net/netdevsim/netdev.c:60 drivers/net/netdevsim/netdev.c:409 drivers/net/netdevsim/netdev.c:418)
__napi_poll.constprop.0 (net/core/dev.c:7997)
net_rx_action (net/core/dev.c:8060 net/core/dev.c:8217)
handle_softirqs (kernel/softirq.c:669)
__irq_exit_rcu (kernel/softirq.c:703 kernel/softirq.c:542 kernel/softirq.c:786)
irq_exit_rcu (kernel/softirq.c:809)
sysvec_apic_timer_interrupt (arch/x86/kernel/apic/apic.c:1062)
</IRQ>
nsim_start_peer_tx_queue() uses the index of the RX queue being polled
as an index into the peer's TX queues:
idx = rq->napi.index;
...
peer_dev = peer_ns->netdev;
if (dev->real_num_tx_queues != peer_dev->num_rx_queues)
goto out;
txq = netdev_get_tx_queue(peer_dev, idx);
The guard compares this device's TX queue count with the peer's RX
queue count, not idx with the peer's TX queue count. The peer is read
again at poll time, so an RX queue filled by one peer can be polled
after the device was linked to a peer with fewer TX queues.
The reproducer runs as root and uses only sysfs, ethtool and packet
sockets: netdevsim devices X and R with 8 queues, P with 1 queue, R
reduced to 1 combined channel (so R's real_num_tx_queues equals P's
num_rx_queues and the guard passes), a packet socket flooding X with
many flows so that R's RX queues 0..7 get traffic, and a thread that
keeps relinking R between X and P. On an x86_64 QEMU guest with 2
vCPUs and KASAN (gcc 13.3), KASAN reports the read in our runs.
The code comes from ff2ac4df58ad ("netdevsim: implement peer queue
flow control"). Checked against v7.4-rc3 (aebb72744764);
nsim_start_peer_tx_queue() is the same in net and net-next.
A patch is attached (git format-patch against v7.4-rc3; it applies to
net and net-next):
0001-netdevsim-bound-the-peer-TX-queue-index-in-nsim_star.patch
It skips the wake-up when the index is not a TX queue of the peer. It
was written with LLM assistance; we reviewed and tested it. With it, the
reproducer no longer triggers the report.
The reproducer (repro.c), the kernel config and the patch are attached.
Thanks,
Kyuwon Cho
// Reproducer for the suspected sibling of 08f566e8f83b in netdevsim:
// nsim_start_peer_tx_queue() calls netdev_get_tx_queue(peer, rq->napi.index)
// with no check against peer->num_tx_queues.
//
// Setup (all netdevsim, init netns, syscalls/sysfs/ioctl only):
// X, R: bus device 10, 2 ports, 8 queues each (num_tx = num_rx = 8)
// P: bus device 11, 1 port, 1 queue (num_tx = num_rx = 1)
// R: ethtool -L combined 1 -> R->real_num_tx = 1 (== P->num_rx, passes the guard)
// Sender: X floods frames with many flows, so skb tx mapping m spreads over 0..7.
// nsim_forward_skb() puts them on R's rq[m % R->num_rx] = rq[m].
// Toggler: link/unlink R between X and P. The rq poll (armed 5us after a forward)
// reads R->peer at poll time. If the poll sees P, idx = rq->napi.index
// (1..7) is used as an index into P->_tx[], which has only one entry.
#define _GNU_SOURCE
#include <arpa/inet.h>
#include <dirent.h>
#include <errno.h>
#include <fcntl.h>
#include <linux/ethtool.h>
#include <linux/if_ether.h>
#include <linux/if_packet.h>
#include <linux/sockios.h>
#include <net/if.h>
#include <netinet/in.h>
#include <netinet/ip.h>
#include <netinet/udp.h>
#include <pthread.h>
#include <stdatomic.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/mount.h>
#include <sys/socket.h>
#include <time.h>
#include <unistd.h>
#define RUN_SECS 60
static int nsfd, ctl;
static unsigned int ifX, ifR, ifP;
static atomic_int stop;
static atomic_long sends, toggles, wr_ok, wr_fail;
static int wr(const char *path, const char *s)
{
int fd = open(path, O_WRONLY);
int r;
if (fd < 0)
return -1;
r = write(fd, s, strlen(s)) < 0 ? -1 : 0;
close(fd);
return r;
}
static int list_netdevs(int busid, char names[][IFNAMSIZ], int want)
{
char path[128];
struct dirent *de;
int n = 0;
DIR *d;
snprintf(path, sizeof(path), "/sys/bus/netdevsim/devices/netdevsim%d/net", busid);
d = opendir(path);
if (!d) {
perror(path);
return 0;
}
while ((de = readdir(d)) && n < want) {
if (de->d_name[0] == '.')
continue;
snprintf(names[n], IFNAMSIZ, "%s", de->d_name);
n++;
}
closedir(d);
return n;
}
static void set_up(const char *name)
{
struct ifreq ifr;
memset(&ifr, 0, sizeof(ifr));
snprintf(ifr.ifr_name, IFNAMSIZ, "%s", name);
if (ioctl(ctl, SIOCGIFFLAGS, &ifr) < 0)
perror("SIOCGIFFLAGS");
ifr.ifr_flags |= IFF_UP;
if (ioctl(ctl, SIOCSIFFLAGS, &ifr) < 0)
perror("SIOCSIFFLAGS");
}
static void set_channels(const char *name, unsigned int n)
{
struct ethtool_channels ch;
struct ifreq ifr;
memset(&ch, 0, sizeof(ch));
memset(&ifr, 0, sizeof(ifr));
ch.cmd = ETHTOOL_SCHANNELS;
ch.combined_count = n;
snprintf(ifr.ifr_name, IFNAMSIZ, "%s", name);
ifr.ifr_data = (void *)&ch;
if (ioctl(ctl, SIOCETHTOOL, &ifr) < 0)
perror("SCHANNELS");
else
printf("set %s combined=%u\n", name, n);
}
static uint32_t rng_state = 0x9e3779b9;
static uint32_t rnd(void)
{
rng_state ^= rng_state << 13;
rng_state ^= rng_state >> 17;
rng_state ^= rng_state << 5;
return rng_state;
}
static void *sender(void *arg)
{
unsigned char dst[ETH_ALEN] = { 0x02, 0, 0, 0, 0, 0x02 };
unsigned char src[ETH_ALEN] = { 0x02, 0, 0, 0, 0, 0x01 };
struct sockaddr_ll sll;
unsigned char f[60];
int fd;
(void)arg;
fd = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_IP));
if (fd < 0) {
perror("socket(AF_PACKET)");
return NULL;
}
memset(&sll, 0, sizeof(sll));
sll.sll_family = AF_PACKET;
sll.sll_protocol = htons(ETH_P_IP);
sll.sll_ifindex = ifX;
sll.sll_halen = ETH_ALEN;
memcpy(sll.sll_addr, dst, ETH_ALEN);
while (!atomic_load(&stop)) {
for (int b = 0; b < 32; b++) {
struct iphdr *ip = (struct iphdr *)(f + 14);
struct udphdr *u = (struct udphdr *)(f + 34);
memset(f, 0, sizeof(f));
memcpy(f, dst, ETH_ALEN);
memcpy(f + 6, src, ETH_ALEN);
f[12] = 0x08;
f[13] = 0x00;
ip->version = 4;
ip->ihl = 5;
ip->tot_len = htons(46);
ip->ttl = 64;
ip->protocol = IPPROTO_UDP;
ip->saddr = htonl(0x0a000001);
ip->daddr = htonl(0x0a000002);
u->source = htons(1024 + (rnd() % 60000));
u->dest = htons(9);
u->len = htons(26);
sendto(fd, f, sizeof(f), 0, (struct sockaddr *)&sll, sizeof(sll));
atomic_fetch_add(&sends, 1);
}
}
close(fd);
return NULL;
}
static void *toggler(void *arg)
{
char unlink_r[64], link_pr[64], link_xr[64];
int ok;
(void)arg;
snprintf(unlink_r, sizeof(unlink_r), "%d:%u", nsfd, ifR);
snprintf(link_pr, sizeof(link_pr), "%d:%u %d:%u", nsfd, ifP, nsfd, ifR);
snprintf(link_xr, sizeof(link_xr), "%d:%u %d:%u", nsfd, ifX, nsfd, ifR);
while (!atomic_load(&stop)) {
wr("/sys/bus/netdevsim/unlink_device", unlink_r);
ok = wr("/sys/bus/netdevsim/link_device", link_pr);
atomic_fetch_add(ok == 0 ? &wr_ok : &wr_fail, 1);
wr("/sys/bus/netdevsim/unlink_device", unlink_r);
ok = wr("/sys/bus/netdevsim/link_device", link_xr);
atomic_fetch_add(ok == 0 ? &wr_ok : &wr_fail, 1);
atomic_fetch_add(&toggles, 1);
}
return NULL;
}
int main(void)
{
char x[2][IFNAMSIZ], p[1][IFNAMSIZ];
char r_name[IFNAMSIZ], x_name[IFNAMSIZ], p_name[IFNAMSIZ];
pthread_t ts, tc;
int n;
setvbuf(stdout, NULL, _IONBF, 0);
mount("sysfs", "/sys", "sysfs", 0, NULL);
ctl = socket(AF_INET, SOCK_DGRAM, 0);
nsfd = open("/proc/self/ns/net", O_RDONLY);
if (wr("/sys/bus/netdevsim/new_device", "10 2 8") < 0)
perror("new_device 10");
if (wr("/sys/bus/netdevsim/new_device", "11 1 1") < 0)
perror("new_device 11");
n = list_netdevs(10, x, 2);
if (n < 2) {
printf("need 2 netdevs on bus 10, got %d\n", n);
return 1;
}
if (list_netdevs(11, p, 1) < 1) {
printf("need 1 netdev on bus 11\n");
return 1;
}
snprintf(x_name, sizeof(x_name), "%s", x[0]);
snprintf(r_name, sizeof(r_name), "%s", x[1]);
snprintf(p_name, sizeof(p_name), "%s", p[0]);
ifX = if_nametoindex(x_name);
ifR = if_nametoindex(r_name);
ifP = if_nametoindex(p_name);
printf("X=%s(%u) R=%s(%u) P=%s(%u)\n", x_name, ifX, r_name, ifR, p_name, ifP);
set_up(x_name);
set_up(r_name);
set_up(p_name);
set_channels(r_name, 1);
{
char link_xr[64];
snprintf(link_xr, sizeof(link_xr), "%d:%u %d:%u", nsfd, ifX, nsfd, ifR);
if (wr("/sys/bus/netdevsim/link_device", link_xr) < 0)
perror("link X-R");
else
printf("linked X-R\n");
}
pthread_create(&ts, NULL, sender, NULL);
pthread_create(&tc, NULL, toggler, NULL);
sleep(RUN_SECS);
atomic_store(&stop, 1);
pthread_join(ts, NULL);
pthread_join(tc, NULL);
printf("done: sends=%ld toggles=%ld link_ok=%ld link_fail=%ld\n",
atomic_load(&sends), atomic_load(&toggles),
atomic_load(&wr_ok), atomic_load(&wr_fail));
return 0;
}
Attachment:
config.gz
Description: application/gzip
From: Kyuwon Cho <kyuwoncho18@xxxxxxxxx>
Date: Thu, 19 Nov 2026 11:36:52 +0900
Subject: [PATCH net] netdevsim: bound the peer TX queue index in
nsim_start_peer_tx_queue()
nsim_start_peer_tx_queue() wakes the peer's TX queue with the index of
the RX queue being polled, and only compares this device's TX queue
count with the peer's RX queue count. When the device is relinked to a
peer with fewer TX queues while its RX queues still hold traffic from
the old peer, the index is past the peer's TX queue array:
BUG: KASAN: slab-out-of-bounds in nsim_poll (include/linux/netdevice.h:3834 drivers/net/netdevsim/netdev.c:60)
Read of size 8
Call Trace:
<IRQ>
nsim_poll (drivers/net/netdevsim/netdev.c:409)
__napi_poll.constprop.0 (net/core/dev.c:7997)
net_rx_action (net/core/dev.c:8217)
Skip the wake-up when the index is not a TX queue of the peer.
Fixes: ff2ac4df58ad ("netdevsim: implement peer queue flow control")
Signed-off-by: Kyuwon Cho <kyuwoncho18@xxxxxxxxx>
---
drivers/net/netdevsim/netdev.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/net/netdevsim/netdev.c b/drivers/net/netdevsim/netdev.c
index eebc02ccc4a9..6e287ae692a8 100644
--- a/drivers/net/netdevsim/netdev.c
+++ b/drivers/net/netdevsim/netdev.c
@@ -53,7 +53,8 @@ static void nsim_start_peer_tx_queue(struct net_device *dev, struct nsim_rq *rq)
/* TX device */
peer_dev = peer_ns->netdev;
- if (dev->real_num_tx_queues != peer_dev->num_rx_queues)
+ if (dev->real_num_tx_queues != peer_dev->num_rx_queues ||
+ idx >= peer_dev->real_num_tx_queues)
goto out;
txq = netdev_get_tx_queue(peer_dev, idx);
--
2.43.0