diff mbox series

[v3,8/9] PCI/ERR: Clear fatal error status when pci_channel_io_frozen

Message ID 20220928105946.12469-9-chenzhuo.1@bytedance.com
State Superseded
Headers show
Series PCI/AER: Fix and optimize usage of status clearing api | expand

Commit Message

Zhuo Chen Sept. 28, 2022, 10:59 a.m. UTC
When state is pci_channel_io_frozen in pcie_do_recovery(), the
severity is fatal and fatal error status should be cleared.
So add pci_aer_clear_fatal_status().

Signed-off-by: Zhuo Chen <chenzhuo.1@bytedance.com>
---
 drivers/pci/pcie/err.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Comments

Bjorn Helgaas Dec. 6, 2022, 9:42 p.m. UTC | #1
Hi Zhuo,

On Wed, Sep 28, 2022 at 06:59:45PM +0800, Zhuo Chen wrote:
> When state is pci_channel_io_frozen in pcie_do_recovery(), the
> severity is fatal and fatal error status should be cleared.
> So add pci_aer_clear_fatal_status().
> 
> Signed-off-by: Zhuo Chen <chenzhuo.1@bytedance.com>
> ---
>  drivers/pci/pcie/err.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
> index f80b21244ef1..b46f1d36c090 100644
> --- a/drivers/pci/pcie/err.c
> +++ b/drivers/pci/pcie/err.c
> @@ -241,7 +241,10 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
>  	pci_walk_bridge(bridge, report_resume, &status);
>  
>  	pcie_clear_device_status(dev);
> -	pci_aer_clear_nonfatal_status(dev);
> +	if (state == pci_channel_io_frozen)
> +		pci_aer_clear_fatal_status(dev);
> +	else
> +		pci_aer_clear_nonfatal_status(dev);

I'm confused.  It seems like we certainly need to clear fatal errors
after they occur *somewhere*, and if we don't, surely this would be a
very obvious issue.  But you didn't mention this being a bug fix, so I
assume it's more of a cleanup.

If it *is* a bug fix, please say that and give a hint about what the
bug looks like, e.g., what sort of messages a user might see.

If it's not a bug fix, I don't understand how AER fatal errors get
cleared today.  The PCI_ERR_UNCOR_STATUS bits are sticky, so they're
not cleared by a reset.  In the current tree, these are the only
places I see that clear AER fatal errors:

  pci_init_capabilities
    pci_aer_init         # once at device enumeration
      pci_aer_clear_status
        pci_aer_raw_clear_status
          pci_write_config_dword(dev, aer + PCI_ERR_UNCOR_STATUS, status)

  aer_probe
    aer_enable_rootport  # once at Root Port enumeration
      pci_write_config_dword(pdev, aer + PCI_ERR_UNCOR_STATUS, reg32)

  dpc_process_error      # after DPC triggered
    pci_aer_clear_fatal_status
      pci_write_config_dword(dev, aer + PCI_ERR_UNCOR_STATUS, status)

  edr_handle_event       # after EDR event
    pci_aer_raw_clear_status
      pci_write_config_dword(dev, aer + PCI_ERR_UNCOR_STATUS, status)

  pci_restore_state      # after reset or PM sleep/resume
    pci_aer_clear_status
      pci_aer_raw_clear_status
        pci_write_config_dword(dev, aer + PCI_ERR_UNCOR_STATUS, status)

The only one that could clear errors after an AER error (not DPC or
EDR), would be the pci_restore_state() in the reset path.  If the
current code relies on that, I'd say that's a pretty non-obvious
dependency.

>  	pci_info(bridge, "device recovery successful\n");
>  	return status;
> -- 
> 2.30.1 (Apple Git-130)
>
diff mbox series

Patch

diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
index f80b21244ef1..b46f1d36c090 100644
--- a/drivers/pci/pcie/err.c
+++ b/drivers/pci/pcie/err.c
@@ -241,7 +241,10 @@  pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
 	pci_walk_bridge(bridge, report_resume, &status);
 
 	pcie_clear_device_status(dev);
-	pci_aer_clear_nonfatal_status(dev);
+	if (state == pci_channel_io_frozen)
+		pci_aer_clear_fatal_status(dev);
+	else
+		pci_aer_clear_nonfatal_status(dev);
 
 	pci_info(bridge, "device recovery successful\n");
 	return status;