Re: [RFC PATCH] e1000e: Use rtnl_lock to prevent race conditions between net and pci/pm

[Alexander Duyck]

On Mon, Oct 7, 2019 at 10:12 AM David Z. Dai <zdai@xxxxxxxxxxxxxxxxxx> wrote:
>
> On Mon, 2019-10-07 at 10:02 -0700, Alexander Duyck wrote:
> > On Mon, Oct 7, 2019 at 8:51 AM David Z. Dai <zdai@xxxxxxxxxxxxxxxxxx> wrote:
> > >
> >
> > <snip>
> >
> > > We have tested on one of the test box.
> > > With this patch, it doesn't crash kernel anymore, which is good!
> > >
> > > However we see this warning message from the log file for irq number 0:
> > > [10206.317270] Trying to free already-free IRQ 0
> > >
> > > With this stack:
> > > [10206.317344] NIP [c00000000018cbf8] __free_irq+0x308/0x370
> > > [10206.317346] LR [c00000000018cbf4] __free_irq+0x304/0x370
> > > [10206.317347] Call Trace:
> > > [10206.317348] [c00000008b92b970] [c00000000018cbf4] __free_irq
> > > +0x304/0x370 (unreliable)
> > > [10206.317351] [c00000008b92ba00] [c00000000018cd84] free_irq+0x84/0xf0
> > > [10206.317358] [c00000008b92ba30] [d000000007449e60] e1000_free_irq
> > > +0x98/0xc0 [e1000e]
> > > [10206.317365] [c00000008b92ba60] [d000000007458a70] e1000e_pm_freeze
> > > +0xb8/0x100 [e1000e]
> > > [10206.317372] [c00000008b92baa0] [d000000007458b6c]
> > > e1000_io_error_detected+0x34/0x70 [e1000e]
> > > [10206.317375] [c00000008b92bad0] [c000000000040358] eeh_report_failure
> > > +0xc8/0x190
> > > [10206.317377] [c00000008b92bb20] [c00000000003eb2c] eeh_pe_dev_traverse
> > > +0x9c/0x170
> > > [10206.317379] [c00000008b92bbb0] [c000000000040d84]
> > > eeh_handle_normal_event+0xe4/0x580
> > > [10206.317382] [c00000008b92bc60] [c000000000041330] eeh_handle_event
> > > +0x30/0x340
> > > [10206.317384] [c00000008b92bd10] [c000000000041780] eeh_event_handler
> > > +0x140/0x200
> > > [10206.317386] [c00000008b92bdc0] [c0000000001397c8] kthread+0x1a8/0x1b0
> > > [10206.317389] [c00000008b92be30] [c00000000000b560]
> > > ret_from_kernel_thread+0x5c/0x7c
> > >
> > > Thanks! - David
> >
> > Hmm. I wonder if it is possibly calling the report
> > e1000_io_error_detected multiple times. If so then the secondary calls
> > to e1000_pm_freeze would cause issues.
> >
> > I will add a check so that we only down the interface and free the
> > IRQs if the interface is in the present and running state.
> >
> > I'll submit an update patch shortly.
> >
> > Thanks.
> >
> > - Alex
> It only complains about IRQ number 0 in the log.

I suspect that is because the IRQ is already freed. So there are no
other interrupts enabled and it thinks it is running in legacy IRQ
mode.

> Could you please let me know the actual place where you will add the
> check?
> I can retest it again.
>
> Thanks! - David

So I will need to add another variable called "present" to
e1000_pm_freeze, I will assign netif_device_present() to it after
taking the rtnl_lock and before I detach the interface, and will test
for it and netif_running() before calling the logic that calls
e1000_down and e1000_free_irq.

- Alex