* [PATCH net v1] r8169: fix RTNL deadlock in runtime suspend
@ 2026-11-18 0:06 Stefano Munari
2026-11-21 1:50 ` netdev-bot+sashiko
0 siblings, 1 reply; 2+ messages in thread
From: Stefano Munari @ 2026-11-18 0:06 UTC (permalink / raw)
To: Heiner Kallweit
Cc: nic_swsd, Andrew Lunn, David S. Miller, Eric Dumazet,
Jakub Kicinski, Paolo Abeni, netdev, linux-kernel
rtl8169_runtime_suspend() and the paths that resume the device under
RTNL wait for each other in opposite order:
- The runtime suspend callback runs with the device in RPM_SUSPENDING
and then calls rtnl_lock(): it owns the suspend in progress and waits
for RTNL.
- rtl8169_close() and the ethtool ioctl/netlink paths run with RTNL
held and then call pm_runtime_get_sync(), which sleeps
uninterruptibly in rpm_resume() until the device leaves
RPM_SUSPENDING: they own RTNL and wait for the suspend to finish.
If the suspend starts after the other task has taken RTNL but before it
reaches pm_runtime_get_sync(), neither can make progress:
ip link set down / ethtool pm workqueue
-------------------------- ------------
rtnl_lock()
rpm_suspend()
status = RPM_SUSPENDING
rtl8169_runtime_suspend()
pm_runtime_get_sync()
rpm_resume()
wait for !RPM_SUSPENDING
rtnl_lock()
RTNL is never released, so every later RTNL user blocks behind it.
The suspend is the delayed one scheduled by rtl8169_runtime_idle():
the interface is up w/out link and runtime PM is enabled
(power/control = auto).
Blocked tasks, "ip link set down":
task:kworker/u66:4 state:D
Workqueue: pm pm_runtime_work
__mutex_lock
rtl8169_runtime_suspend+0x1a/0x70 [r8169]
pci_pm_runtime_suspend+0x78/0x230
__rpm_callback+0x56/0x1e0
rpm_callback+0x30/0x80
rpm_suspend+0xe7/0x630
pm_runtime_work+0x57/0xb0
task:python3 state:D
rpm_resume+0x1d0/0x750
__pm_runtime_resume+0x43/0x60
rtl8169_close+0x32/0x100 [r8169]
__dev_close_many+0xaf/0x1b0
__dev_change_flags+0xe2/0x1f0
netif_change_flags+0x27/0x70
do_setlink.isra.0+0x367/0xe70
rtnl_newlink+0x4c6/0xab0
Use rtnl_trylock() and return -EBUSY if RTNL is contended. The PM core
then sets the device back to RPM_ACTIVE and wakes up the waiters in
rpm_resume(), without recording a runtime PM error. In the deadlock
scenario the RTNL holder is the task holding runtime PM reference, so
the suspend is rescheduled by rtl8169_runtime_idle() when it drops its
usage count. If RTNL is held by an unrelated task, the device stays
active until the next idle check (usage count drop or link down).
This issue was found using an AI code review agent.
Tested on RTL8168gu/8111gu (XID 509), net tree 7.4.0-rc2.
Setup: interface up, cable unplugged, power/control = auto.
It was reproduced with the unpatched driver by timing
"ip link set down" to the expiry of the 10s runtime suspend timer.
It was also reproduced deterministically with a debug-only msleep()
before rtnl_lock() in rtl8169_runtime_suspend() to widen the race window
and "ethtool -i" issued during the sleep.
With the patch the timed reproducer completed 100 attempts, hitting
the race window in 12 of them, without deadlock.
The deterministic msleep() variant no longer hangs.
Kernel selftests were not run.
Fixes: abe5fc42f9ce ("r8169: use RTNL to protect critical sections")
Signed-off-by: Stefano Munari <stefanomunari.sm@gmail.com>
---
Race window (RTNL held before rtl8169_close() reaches
pm_runtime_get_sync()) on the unpatched driver, measured with kprobes
over 200 x "ip link set down":
idle: median ~175 us, 95% below 250 us, max 3.1 ms
moderate load: median ~165 us, 95% below 2.1 ms, max 6.0 ms
---
drivers/net/ethernet/realtek/r8169_main.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index f151ab681393..c60d87015eb9 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -5975,7 +5975,14 @@ static int rtl8169_runtime_suspend(struct device *device)
return 0;
}
- rtnl_lock();
+ /*
+ * Do not block on RTNL.
+ * Other callers holding RTNL may wait for suspend to complete
+ * in pm_runtime_get_sync(), which would deadlock.
+ */
+ if (!rtnl_trylock())
+ return -EBUSY;
+
__rtl8169_set_wol(tp, WAKE_PHY);
rtl8169_net_suspend(tp);
rtnl_unlock();
--
2.55.0
^ permalink raw reply [flat|nested] 2+ messages in thread* Re: [PATCH net v1] r8169: fix RTNL deadlock in runtime suspend
2026-11-18 0:06 [PATCH net v1] r8169: fix RTNL deadlock in runtime suspend Stefano Munari
@ 2026-11-21 1:50 ` netdev-bot+sashiko
0 siblings, 0 replies; 2+ messages in thread
From: netdev-bot+sashiko @ 2026-11-21 1:50 UTC (permalink / raw)
To: stefanomunari.sm
Cc: hkallweit1, nic_swsd, andrew+netdev, davem, edumazet, kuba,
pabeni, netdev, linux-kernel
Thank you for your contribution! Sashiko AI review found 1 potential
issue(s) to consider.
Critical: 0 · High: 0 · Medium: 1 · Low: 0
If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).
- [Medium] The patch makes rtl8169_runtime_suspend()
(drivers/net/ethernet/realtek/r8169_main.c) return -EBUSY whenever…
--
r8169: fix RTNL deadlock in runtime suspend
This patch replaces rtnl_lock() in rtl8169_runtime_suspend() with
rtnl_trylock(), and returns -EBUSY when RTNL is contended. The goal is to
avoid an ABBA deadlock. The runtime suspend callback keeps the device in
RPM_SUSPENDING while it waits for RTNL. Meanwhile, RTNL holders such as
rtl8169_close() and the ethtool paths wait in pm_runtime_get_sync().
> diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
> index f151ab6813932..c60d87015eb94 100644
> --- a/drivers/net/ethernet/realtek/r8169_main.c
> +++ b/drivers/net/ethernet/realtek/r8169_main.c
> @@ -5975,7 +5975,14 @@ static int rtl8169_runtime_suspend(struct device *device)
> return 0;
> }
>
> - rtnl_lock();
> + /*
> + * Do not block on RTNL.
> + * Other callers holding RTNL may wait for suspend to complete
> + * in pm_runtime_get_sync(), which would deadlock.
> + */
> + if (!rtnl_trylock())
> + return -EBUSY;
[Severity: Medium]
Once this returns -EBUSY, does anything schedule another runtime suspend
attempt?
r8169 arms the suspend from rtl8169_runtime_idle() with a plain timer
request, not with autosuspend:
drivers/net/ethernet/realtek/r8169_main.c:rtl8169_runtime_idle() {
...
if (!netif_running(tp->dev) || !netif_carrier_ok(tp->dev))
pm_schedule_suspend(device, 10000);
...
}
pm_schedule_suspend() leaves timer_autosuspends at 0. As a result,
pm_suspend_timer_fn() calls rpm_suspend() with RPM_ASYNC but without
RPM_AUTO. The driver also never calls pm_runtime_use_autosuspend() or
pm_runtime_mark_last_busy().
On -EBUSY, the fail path of rpm_suspend() only retries for autosuspend:
drivers/base/power/runtime.c:rpm_suspend() {
...
if (!dev->power.runtime_error && (rpmflags & RPM_AUTO) &&
pm_runtime_autosuspend_expiration(dev) != 0)
goto repeat;
pm_runtime_cancel_pending(dev);
...
}
So the pending request and the suspend timer are both cancelled.
The comment in pci_pm_runtime_suspend() says "-EBUSY and -EAGAIN is used
to request the runtime PM core to schedule a new suspend". The core
doesn't do that for non-autosuspend requests, though.
After that, the suspend is only re-armed if rtl8169_runtime_idle() runs
again. That happens on:
- open/close
- the LED helpers, which use pm_runtime_put_sync()
- the ethtool core's pm_runtime_put() in net/ethtool/ioctl.c and
net/ethtool/netlink.c
- rtl_mac_link_down(), which calls pm_runtime_idle()
rtl8169_get_stats64() uses pm_runtime_put_noidle(), so stats reads don't
re-arm it.
The patch targets a setup where the interface is up with no carrier and
power/control is set to auto. In that setup the link-down event has
already happened.
Suppose an unrelated task holds RTNL when the 10s timer fires. Examples
are a netlink or ethtool request for another NIC, a netdev notifier, or
an ip -s link dump. Can the NIC then stay in D0 indefinitely, without
WAKE_PHY armed? Before this patch, rtnl_lock() waited, so the suspend
always completed.
The commit message says:
If RTNL is held by an unrelated task, the device stays
active until the next idle check (usage count drop or link down).
Is a next idle check actually guaranteed in this setup?
This also looks like the "trylock with fallback to skip the work
entirely" pattern. That pattern is only acceptable when the failure path
reschedules the work.
Would one of these approaches work instead?
- Convert the driver to autosuspend, using pm_runtime_use_autosuspend()
plus pm_runtime_mark_last_busy() before returning -EBUSY, so the PM
core retries.
- Re-arm from the driver after the callback returns, for example with a
work item that calls pm_runtime_idle() or pm_schedule_suspend().
Re-arming from inside the callback won't work, because
pm_runtime_cancel_pending() in the fail path cancels it.
- Restructure the suspend path so it doesn't need RTNL.
> __rtl8169_set_wol(tp, WAKE_PHY);
> rtl8169_net_suspend(tp);
> rtnl_unlock();
--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261118000631.17771-1-stefanomunari.sm%40gmail.com
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-21 1:50 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-18 0:06 [PATCH net v1] r8169: fix RTNL deadlock in runtime suspend Stefano Munari
2026-11-21 1:50 ` netdev-bot+sashiko
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®