IRQ/nvme_pci_complete_rq: NULL pointer dereference yet again

[Alex G.]

On 04/06/2018 01:04 PM, Keith Busch wrote:
> On Fri, Apr 06, 2018 at 12:46:06PM -0500, Alex G. wrote:
>> On 04/06/2018 12:16 PM, Scott Bauer wrote:
>>> You're using AER inject, right?
>>
>> No. I'm causing the errors in hardware with hot-unplug.
> 
> I think Scott's still on the right track for this particular sighting.
> The AER handler looks unsafe under changing topologies. It might need run
> under pci_lock_rescan_remove() before walking the bus to prevent races
> with the surprise removal, but it's not clear to me yet if holding that
> lock is okay to do in this context.

I think we have three mechanisms that can remove a device: nvme timeout,
Link Down interrupt, and AER.
Link Down comes 20-60ms after the link actually dies, in which time nvme
will queue IO, which can cause a flood of PCIe errors, which trigger AER
handling. I suspect there is a massive race condition somewhere, but I
don't yet have convincing evidence to prove it.

> This however does not appear to resemble your previous sightings. In your
> previous sightings, it looks like something has lost track of commands,
> and we're freeing the resources with them a second time.

I think they might be related.

Alex