Re: [PATCH v4 1/1] PCI/ERR: don't clobber status after reset_link()

[Bjorn Helgaas]

On Mon, Nov 02, 2020 at 03:09:51PM +0000, Hedi Berriche wrote:
> Commit 6d2c89441571 ("PCI/ERR: Update error status after reset_link()")
> broke pcie_do_recovery(): updating status after reset_link() has the ill
> side effect of causing recovery to fail if the error status is
> PCI_ERS_RESULT_CAN_RECOVER or PCI_ERS_RESULT_NEED_RESET as the following
> code will *never* run in the case of a successful reset_link()
> 
>    177         if (status == PCI_ERS_RESULT_CAN_RECOVER) {
>    ...
>    181         }
> 
>    183         if (status == PCI_ERS_RESULT_NEED_RESET) {
>    ...
>    192         }

The line numbers are basically useless because they depend on some
particular version of the file.

> For instance in the case of PCI_ERS_RESULT_NEED_RESET we end up not
> calling ->slot_reset() (because we skip report_slot_reset()) thus
> breaking driver (re)initialisation.
> 
> Don't clobber status with the return value of reset_link(); set status
> to PCI_ERS_RESULT_RECOVERED, in case of successful link reset, if and
> only if the initial value of error status is PCI_ERS_RESULT_DISCONNECT
> or PCI_ERS_RESULT_NO_AER_DRIVER.
>
> Fixes: 6d2c89441571 ("PCI/ERR: Update error status after reset_link()")
> Signed-off-by: Hedi Berriche <hedi.berriche@xxxxxxx>
> 
> Reviewed-by: Sinan Kaya <okaya@xxxxxxxxxx>
> Cc: Russ Anderson <rja@xxxxxxx>
> Cc: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx>
> Cc: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> Cc: Ashok Raj <ashok.raj@xxxxxxxxx>
> Cc: Joerg Roedel <jroedel@xxxxxxxx>
> 
> Cc: stable@xxxxxxxxxx # v5.7+
> ---
>  drivers/pci/pcie/err.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
> index c543f419d8f9..2730826cfd8a 100644
> --- a/drivers/pci/pcie/err.c
> +++ b/drivers/pci/pcie/err.c
> @@ -165,10 +165,13 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
>  	pci_dbg(dev, "broadcast error_detected message\n");
>  	if (state == pci_channel_io_frozen) {
>  		pci_walk_bus(bus, report_frozen_detected, &status);
> -		status = reset_link(dev);
> -		if (status != PCI_ERS_RESULT_RECOVERED) {
> +		if (reset_link(dev) != PCI_ERS_RESULT_RECOVERED) {
>  			pci_warn(dev, "link reset failed\n");
>  			goto failed;
> +		} else {
> +			if (status == PCI_ERS_RESULT_DISCONNECT ||
> +			    status == PCI_ERS_RESULT_NO_AER_DRIVER)
> +				status = PCI_ERS_RESULT_RECOVERED;

This code (even before your patch) doesn't match
Documentation/PCI/pci-error-recovery.rst very well.  The code handles
pci_channel_io_frozen specially, but I don't think this is mentioned
in the doc.

The doc says we call ->error_detected() for all affected drivers.
Then we're supposed to do a slot reset if any driver returned
NEED_RESET.  But in fact, we always do a reset for the
pci_channel_io_frozen case and never do one otherwise, regardless of
what ->error_detected() returned.

The doc says DISCONNECT means "Driver ... doesn't want to recover at
all." Many drivers can return either NEED_RESET or DISCONNECT, and I
assume they expect them to be handled differently.  But I'm not sure
what DISCONNECT really means.  Do we reset the device?  Do we not
attempt recovery at all?

After your patch, if the reset_link() succeeded, we convert DISCONNECT
and NO_AER_DRIVER to RECOVERED.  IIUC, that means we do exactly the
same thing if the consensus of the ->error_detected() functions was
RECOVERED, DISCONNECT, or NO_AER_DRIVER: we call reset_link() and
continue with "status = PCI_ERS_RESULT_RECOVERED".

(I'd reverse the sense of the "if (reset_link())" to make this easier
to read)

>  		}
>  	} else {
>  		pci_walk_bus(bus, report_normal_detected, &status);
> -- 
> 2.28.0
>