Re: [PATCH v7 2/5] PCI/DPC: Run recovery on device that detected the error

[Shuai Xue <xueshuai@linux.alibaba.com>]

On 2/27/26 6:47 PM, Lukas Wunner wrote:
> On Fri, Feb 27, 2026 at 04:28:59PM +0800, Shuai Xue wrote:
>> On 2/7/26 3:48 PM, Shuai Xue wrote:
>>> Regarding pci_restore_state() in slot_reset(): I see now that it does
>>> call pci_aer_clear_status(dev) (at line 1844 in pci.c), which will
>>> clear the AER Status registers. So if we walk the hierarchy after
>>> the slot_reset callbacks, the error bits accumulated during DPC will
>>> already be cleared.
>>>
>>> To avoid losing those errors, I think the walk should happen after
>>> dpc_reset_link() succeeds but *before* pcie_do_recovery() invokes the
>>> slot_reset callbacks. That way, we can capture the AER Status bits
>>> before pci_restore_state() clears them.
>>>
>>> Does that sound like the right approach, or would you prefer a
>>> different placement?
> 
> The problem is that if the hierarchy that was reset is deeper than
> one level, you first need to call pci_restore_state() on all the
> PCIe Upstream and Downstream Ports that were reset before you can
> access the Endpoints at the bottom of the hierarchy.
> 
> E.g. if DPC occurs at a Root Port with multiple nested PCIe switches
> below, the Endpoints at the "leafs" of that tree are only accessible
> once Config Space has been restored at all the PCIe switches
> in-between the Endpoints and the DPC-capable Root Port.
> 
> Hence your proposal unfortunately won't work.
> 
> I think the solution is to move pci_aer_clear_status() out of
> pci_restore_state() into the callers that actually need it.
> But that requires going through every single caller.
> I've begun doing that last week and am about 60% done.
> 
> Once pci_restore_state() no longer clears the error bits, we can
> report and clear them after the "report_slot_reset" stage (which
> is where drivers call pci_restore_state()).
> 
> I've also changed my mind and I think reporting and clearing
> the error bits *could* happen in pcie_do_recovery() even if it
> were used for EEH and s390 because those platforms may plug in
> AER-capable devices as well and so we do need to clear the bits
> regardless of the error recovery mechanism used.
> 
> Let me get back to you once I've gone through all the callers of
> pci_restore_state().  Please be patient.
> 

Sure, glad to hear you have been working on that.


Thanks.
SHuai