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

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.