[PATCH net v2] net: core: make napi_disable more robust

There are chances that napi_disable can be called twice by NIC driver.
This could generate deadlock. For example,
the first napi_disable will spin until NAPI_STATE_SCHED is cleared
by napi_complete_done, then set it again.
When napi_disable is called the second time, it will loop infinitely
because no dev->poll will be running to clear NAPI_STATE_SCHED.

Though it is driver writer's responsibility to make sure it being
called only once, making napi_disable more robust does not hurt, not
to say it can prevent a buggy driver from crashing a system.
So, we check the napi state bit to make sure that if napi is already
disabled, we exit the call early enough to avoid spinning infinitely.

Fixes: bea3348eef27 ("[NET]: Make NAPI polling independent of struct net_device objects.")
Signed-off-by: Lijun Pan <redacted>
---
v2: justify that this patch makes napi_disable more robust.

 net/core/dev.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/net/core/dev.c b/net/core/dev.c
index 1f79b9aa9a3f..fa0aa212b7bb 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -6830,6 +6830,24 @@ EXPORT_SYMBOL(netif_napi_add);
 void napi_disable(struct napi_struct *n)
 {
 	might_sleep();
+
+	/* make sure napi_disable() runs only once,
+	 * When napi is disabled, the state bits are like:
+	 * NAPI_STATE_SCHED (set by previous napi_disable)
+	 * NAPI_STATE_NPSVC (set by previous napi_disable)
+	 * NAPI_STATE_DISABLE (cleared by previous napi_disable)
+	 * NAPI_STATE_PREFER_BUSY_POLL (cleared by previous napi_complete_done)
+	 * NAPI_STATE_MISSED (cleared by previous napi_complete_done)
+	 */
+
+	if (napi_disable_pending(n))
+		return;
+	if (test_bit(NAPI_STATE_SCHED, &n->state) &&
+	    test_bit(NAPI_STATE_NPSVC, &n->state) &&
+	    !test_bit(NAPI_STATE_MISSED, &n->state) &&
+	    !test_bit(NAPI_STATE_PREFER_BUSY_POLL, &n->state))
+		return;
+
 	set_bit(NAPI_STATE_DISABLE, &n->state);
 
 	while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))

-- 
2.23.0

On 2021/4/14 16:08, Lijun Pan wrote:

The NAPI_STATE_DISABLE is cleared at the end of napi_disable(),
and if a buggy driver/hw triggers a interrupt and driver calls
napi_schedule_irqoff(), which may set NAPI_STATE_MISSED
if NAPI_STATE_SCHED is set(in napi_schedule_prep()), the above
checking does not seem to handle it?

+
 	set_bit(NAPI_STATE_DISABLE, &n->state);
 
 	while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))

On Wed, Apr 14, 2021 at 3:45 AM Yunsheng Lin [off-list ref] wrote:

On 2021/4/14 16:08, Lijun Pan wrote:

quoted

There are chances that napi_disable can be called twice by NIC driver.
This could generate deadlock. For example,
the first napi_disable will spin until NAPI_STATE_SCHED is cleared
by napi_complete_done, then set it again.
When napi_disable is called the second time, it will loop infinitely
because no dev->poll will be running to clear NAPI_STATE_SCHED.

Though it is driver writer's responsibility to make sure it being
called only once, making napi_disable more robust does not hurt, not
to say it can prevent a buggy driver from crashing a system.
So, we check the napi state bit to make sure that if napi is already
disabled, we exit the call early enough to avoid spinning infinitely.

Fixes: bea3348eef27 ("[NET]: Make NAPI polling independent of struct net_device objects.")
Signed-off-by: Lijun Pan <redacted>
---
v2: justify that this patch makes napi_disable more robust.

 net/core/dev.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/net/core/dev.c b/net/core/dev.c
index 1f79b9aa9a3f..fa0aa212b7bb 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -6830,6 +6830,24 @@ EXPORT_SYMBOL(netif_napi_add);
 void napi_disable(struct napi_struct *n)
 {
      might_sleep();
+
+     /* make sure napi_disable() runs only once,
+      * When napi is disabled, the state bits are like:
+      * NAPI_STATE_SCHED (set by previous napi_disable)
+      * NAPI_STATE_NPSVC (set by previous napi_disable)
+      * NAPI_STATE_DISABLE (cleared by previous napi_disable)
+      * NAPI_STATE_PREFER_BUSY_POLL (cleared by previous napi_complete_done)
+      * NAPI_STATE_MISSED (cleared by previous napi_complete_done)
+      */
+
+     if (napi_disable_pending(n))
+             return;
+     if (test_bit(NAPI_STATE_SCHED, &n->state) &&
+         test_bit(NAPI_STATE_NPSVC, &n->state) &&
+         !test_bit(NAPI_STATE_MISSED, &n->state) &&
+         !test_bit(NAPI_STATE_PREFER_BUSY_POLL, &n->state))
+             return;

The NAPI_STATE_DISABLE is cleared at the end of napi_disable(),
and if a buggy driver/hw triggers a interrupt and driver calls
napi_schedule_irqoff(), which may set NAPI_STATE_MISSED
if NAPI_STATE_SCHED is set(in napi_schedule_prep()), the above
checking does not seem to handle it?

What I described in the commit message is the napi_disable() being
called from the same instance, same cpu. e.g.,
funcA {
    napi_disable();
    ...
    funcB{
        if (blah)
            napi_disable();
            ...
    }
    funcC;
}

The scenario you mentioned above seems to have napi already enabled
and scheduled, such that napi_schedule_prep() would set NAPI_STATE_MISSED.
The two scenarios are different per my understanding. Is there a way
that your scenario will finally call into my scenario?
Let me know if I understand you correctly.

Maybe testing NAPI_STATE_MISSED bit is not needed
because this bit is not that reliable.

Lijun

On Wed, 14 Apr 2021 03:08:45 -0500 Lijun Pan wrote:

There are chances that napi_disable can be called twice by NIC driver.
This could generate deadlock. For example,
the first napi_disable will spin until NAPI_STATE_SCHED is cleared
by napi_complete_done, then set it again.
When napi_disable is called the second time, it will loop infinitely
because no dev->poll will be running to clear NAPI_STATE_SCHED.

Though it is driver writer's responsibility to make sure it being
called only once, making napi_disable more robust does not hurt, not
to say it can prevent a buggy driver from crashing a system.
So, we check the napi state bit to make sure that if napi is already
disabled, we exit the call early enough to avoid spinning infinitely.

You've already been told by Eric & Dave to fix the driver instead.

Your check is _not_ correct - SCHED && NPSVC && !MISSED && !BUSY_POLL 
can well arise without disabling the NAPI.

But regardless, a driver bug should be relatively easy to identify with
task getting stuck in napi_disable(). We don't provide "protection" 
for taking spin locks or ref counts twice either. Unless you can show 
a strong use case please stop posting new versions of this patch.

On 4/15/21 1:21 AM, Jakub Kicinski wrote:

On Wed, 14 Apr 2021 03:08:45 -0500 Lijun Pan wrote:

quoted

There are chances that napi_disable can be called twice by NIC driver.
This could generate deadlock. For example,
the first napi_disable will spin until NAPI_STATE_SCHED is cleared
by napi_complete_done, then set it again.
When napi_disable is called the second time, it will loop infinitely
because no dev->poll will be running to clear NAPI_STATE_SCHED.

Though it is driver writer's responsibility to make sure it being
called only once, making napi_disable more robust does not hurt, not
to say it can prevent a buggy driver from crashing a system.
So, we check the napi state bit to make sure that if napi is already
disabled, we exit the call early enough to avoid spinning infinitely.

You've already been told by Eric & Dave to fix the driver instead.

Your check is _not_ correct - SCHED && NPSVC && !MISSED && !BUSY_POLL 
can well arise without disabling the NAPI.

But regardless, a driver bug should be relatively easy to identify with
task getting stuck in napi_disable(). We don't provide "protection" 
for taking spin locks or ref counts twice either. Unless you can show 
a strong use case please stop posting new versions of this patch.

+222

I notice this v2 does not even mention which driver has the issue.

I suspect an out-of-tree driver.

On 4/14/21 10:08 AM, Lijun Pan wrote:

There are chances that napi_disable can be called twice by NIC driver.
This could generate deadlock. For example,
the first napi_disable will spin until NAPI_STATE_SCHED is cleared
by napi_complete_done, then set it again.
When napi_disable is called the second time, it will loop infinitely
because no dev->poll will be running to clear NAPI_STATE_SCHED.

Though it is driver writer's responsibility to make sure it being
called only once, making napi_disable more robust does not hurt, not
to say it can prevent a buggy driver from crashing a system.

This is hard to digest. A buggy driver has plenty of ways to crash the system.

If you need help to fix the buggy driver, please ask for help.