[PATCH] net: make unregister netdev warning timeout configurable

netdev_wait_allrefs() issues a warning if refcount does not drop to 0
after 10 seconds. While 10 second wait generally should not happen
under normal workload in normal environment, it seems to fire falsely
very often during fuzzing and/or in qemu emulation (~10x slower).
At least it's not possible to understand if it's really a false
positive or not. Automated testing generally bumps all timeouts
to very high values to avoid flake failures.
Make the timeout configurable for automated testing systems.
Lowering the timeout may also be useful for e.g. manual bisection.
The default value matches the current behavior.

Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=211877
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
---
 net/Kconfig    | 12 ++++++++++++
 net/core/dev.c |  4 +++-
 2 files changed, 15 insertions(+), 1 deletion(-)

diff --git a/net/Kconfig b/net/Kconfig
index 8cea808ad9e8d..ebb9cc00ac81d 100644
--- a/net/Kconfig
+++ b/net/Kconfig

@@ -461,6 +461,18 @@ config ETHTOOL_NETLINK
 	  netlink. It provides better extensibility and some new features,
 	  e.g. notification messages.
 
+config UNREGISTER_NETDEV_TIMEOUT
+	int "Unregister network device timeout in seconds"
+	default 10
+	range 0 3600
+	help
+	  This option controls the timeout (in seconds) used to issue
+	  a warning while waiting for a network device refcount to drop to 0
+	  during device unregistration.
+	  A lower value may be useful during bisection to detect a leaked
+	  reference faster. A larger value may be useful to prevent false
+	  warnings on slow/loaded systems.
+
 endif   # if NET
 
 # Used by archs to tell that they support BPF JIT compiler plus which flavour.

diff --git a/net/core/dev.c b/net/core/dev.c
index 0f72ff5d34ba0..ca03ee407133b 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -10405,7 +10405,9 @@ static void netdev_wait_allrefs(struct net_device *dev)
 
 		refcnt = netdev_refcnt_read(dev);
 
-		if (refcnt && time_after(jiffies, warning_time + 10 * HZ)) {
+		if (refcnt &&
+		    time_after(jiffies, warning_time +
+			       CONFIG_UNREGISTER_NETDEV_TIMEOUT * HZ)) {
 			pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
 				 dev->name, refcnt);
 			warning_time = jiffies;

base-commit: 5aa3c334a449bab24519c4967f5ac2b3304c8dcf
-- 
2.31.0.291.g576ba9dcdaf-goog

On Sat, Mar 20, 2021 at 03:28:51PM +0100, Dmitry Vyukov wrote:

netdev_wait_allrefs() issues a warning if refcount does not drop to 0
after 10 seconds. While 10 second wait generally should not happen
under normal workload in normal environment, it seems to fire falsely
very often during fuzzing and/or in qemu emulation (~10x slower).
At least it's not possible to understand if it's really a false
positive or not. Automated testing generally bumps all timeouts
to very high values to avoid flake failures.
Make the timeout configurable for automated testing systems.
Lowering the timeout may also be useful for e.g. manual bisection.
The default value matches the current behavior.

Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=211877
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
---
 net/Kconfig    | 12 ++++++++++++
 net/core/dev.c |  4 +++-
 2 files changed, 15 insertions(+), 1 deletion(-)

Our verification team would like to see this change too.

Thanks,
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>

From: Dmitry Vyukov <dvyukov@google.com>
Date: Sat, 20 Mar 2021 15:28:51 +0100

netdev_wait_allrefs() issues a warning if refcount does not drop to 0
after 10 seconds. While 10 second wait generally should not happen
under normal workload in normal environment, it seems to fire falsely
very often during fuzzing and/or in qemu emulation (~10x slower).
At least it's not possible to understand if it's really a false
positive or not. Automated testing generally bumps all timeouts
to very high values to avoid flake failures.
Make the timeout configurable for automated testing systems.
Lowering the timeout may also be useful for e.g. manual bisection.
The default value matches the current behavior.

Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=211877
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org

I'd say a sysctl knob is much better than a compile time setting for this.
That way stock kernels can be used in these testing scenerios.

Thanks.

On Mon, Mar 22, 2021 at 8:26 PM David Miller [off-list ref] wrote:

From: Dmitry Vyukov <dvyukov@google.com>
Date: Sat, 20 Mar 2021 15:28:51 +0100

quoted

netdev_wait_allrefs() issues a warning if refcount does not drop to 0
after 10 seconds. While 10 second wait generally should not happen
under normal workload in normal environment, it seems to fire falsely
very often during fuzzing and/or in qemu emulation (~10x slower).
At least it's not possible to understand if it's really a false
positive or not. Automated testing generally bumps all timeouts
to very high values to avoid flake failures.
Make the timeout configurable for automated testing systems.
Lowering the timeout may also be useful for e.g. manual bisection.
The default value matches the current behavior.

Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=211877
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org

I'd say a sysctl knob is much better than a compile time setting for this.
That way stock kernels can be used in these testing scenerios.

FTR, I've mailed v2 with a sysctl:
https://lore.kernel.org/netdev/20210323064923.2098711-1-dvyukov@google.com/T/#u