Re: e1000_netpoll() , BUG: sleeping function called from invalid context

From: Gabriel C
Date: Fri Feb 17 2017 - 18:54:15 EST

On 18.02.2017 00:44, Cong Wang wrote:
On Fri, Feb 17, 2017 at 3:38 PM, Gabriel C <nix.or.die@xxxxxxxxx> wrote:

My card seems to use the e1000e driver which is buit-in..

Anyway here an objdump -x :

Found disable_hardirq() but not disable_irq().

Are you sure the kernel warning was emitted by this binary rather than
some old one?

Yes , I use an clean build , remove any other files in /boot && /lib/modules before installing the new build kernel.

I've asked Thomas before I send this to lkml and to you and he said :


It's a real issue. netconsole calls disable_irq() which might sleep from an
interrupt and preemption disabled context.

> [85362.132801] __might_sleep+0x6b/0x80
> [85362.132803] synchronize_irq+0x33/0x90
> [85362.132805] ? __irq_put_desc_unlock+0x19/0x40
> [85362.132807] ? __disable_irq_nosync+0x4e/0x60
> [85362.132808] disable_irq+0x17/0x20
> [85362.132810] e1000_netpoll+0x3d/0x110

Though, what's weird is that the e1000_netpoll() does not longer call
disable_irq(). It calls disable_hardirq().

That got changed in commit 3111912971251 which got into Linus tree during
the 4.10 merge window. So it is in 4.10-rc8.