Re: [PATCH 1/1] PCI/AER: prevent pcie_do_fatal_recovery from using device after it is removed

[Sinan Kaya <okaya@kernel.org>]

On 8/21/2018 11:50 AM, Sinan Kaya wrote:
> On 8/21/2018 11:29 AM, Keith Busch wrote:
>>> Hotplug driver needs to handle both physical removal as well as 
>>> intermittent
>>> link down issues though.
>> Back to your patch you linked to earlier, your proposal is to have
>> pciehp wait for DEVSTS.FED before deciding if it needs to handle the
>> DLLSC event. That might be a start, but it isn't enough since that
>> status isn't set if the downstream device reported ERR_FATAL. I think
>> you'd need to check the secondary status register for a Received System
>> Error.
> 
> Hmm, good feedback.
> 
> I was trying not to mix AER/DPC with HP but obviously I failed.
> I can add a flag to struct pci_dev like aer_pending for AER.
> 
> 1. AER ISR sets aer_pending
> 2. AER ISR issues a secondary bus reset
> 3. Hotplug driver bails out on aer_pending
> 4. AER ISR performs the recovery
> 
> It is slightly more challenging on the DPC front as HW brings down
> the link automatically and hotplug driver can observe the link down
> event before the DPC ISR. I'll have to go check if DPC interrupt is
> pending.
> 
> Let me know if this works out.

I forgot to mention that I didn't really want to rely on Secondary
Status Registers as most of the recent PCIe controller HW don't really
implement the old PCI status register bits correctly.