Re: [PATCH] PCI/AER: don't call recovery process for correctable errors

[Bjorn Helgaas]

On Wed, Oct 11, 2017 at 10:37:47AM -0400, Tyler Baicar wrote:
> On 10/2/2017 7:19 PM, Bjorn Helgaas wrote:
> >On Mon, Aug 28, 2017 at 11:09:44AM -0600, Tyler Baicar wrote:
> >>Correctable errors do not need any software intervention, so
> >>avoid calling into the software recovery process for correctable
> >>errors.
> >>
> >>Signed-off-by: Tyler Baicar <tbaicar@xxxxxxxxxxxxxx>
> >>---
> >>  drivers/pci/pcie/aer/aerdrv_core.c | 3 ++-
> >>  1 file changed, 2 insertions(+), 1 deletion(-)
> >>
> >>diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> >>index b1303b3..4765c11 100644
> >>--- a/drivers/pci/pcie/aer/aerdrv_core.c
> >>+++ b/drivers/pci/pcie/aer/aerdrv_core.c
> >>@@ -626,7 +626,8 @@ static void aer_recover_work_func(struct work_struct *work)
> >>  			continue;
> >>  		}
> >>  		cper_print_aer(pdev, entry.severity, entry.regs);
> >>-		do_recovery(pdev, entry.severity);
> >>+		if (entry.severity != AER_CORRECTABLE)
> >>+			do_recovery(pdev, entry.severity);
> >I think this is fine, and it mirrors what is done in
> >handle_error_source().
> >
> >But I want to converge the APEI path and the "native" AER path, so as
> >one tiny step in that direction, can you look into doing this test
> >once, e.g., move the test from handle_error_source() into
> >do_recovery(), where one test will handle both paths?
> 
> I've looked into this and it seems there is still going to need to
> be two versions of this check. The native AER path goes through
> handle_error_source() and for correctable errors requires the write
> to PCI_ERR_COR_STATUS, but the APEI path does not require this
> write. I could move this check to the beginning of do_recovery() so
> it returns right away for correctable errors and remove the else
> line in handle_error_source() so it always calls into do_recovery().
> That doesn't seem like a very clean solution though since then there
> are still two checks for correctable errors and now we're calling
> into do_recovery() in both cases for nothing :)

The PCI_ERR_COR_STATUS thing is part of what I see as the problem
here.  IMHO, the native AER path should collect up the log registers
(and do any acknowledgement, e.g., writing PCI_ERR_COR_STATUS)
*before* entering the common path.

In other words, the Linux code in the native part of AER should be
functionally the same as the BIOS code that implements the APEI path.

This is a bit of restructuring in the Linux AER code.  I haven't
looked enough to know how much.  If it's impractical, it's
impractical.  I thought this might be an opportunity for a tiny step
in that direction, but if it's not, I guess that's OK.

Bjorn