在 2025/10/20 18:10, Lukas Wunner 写道:

On Wed, Oct 15, 2025 at 10:41:57AM +0800, Shuai Xue wrote:

quoted

+++ b/drivers/pci/pcie/err.c

@@ -253,6 +254,16 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
  			pci_warn(bridge, "subordinate device reset failed\n");
  			goto failed;
  		}
+
+		/* Link recovered, report fatal errors of RCiEP or EP */
+		if (state == pci_channel_io_frozen &&
+		    (type == PCI_EXP_TYPE_ENDPOINT || type == PCI_EXP_TYPE_RC_END)) {
+			aer_add_error_device(&info, dev);
+			info.severity = AER_FATAL;
+			if (aer_get_device_error_info(&info, 0, true))
+				aer_print_error(&info, 0);
+			pci_dev_put(dev);
+		}
  	}

Hi, Lukas,

Where is the the pci_dev_get() to balance the pci_dev_put() here?

The corresponding pci_dev_get() is called in add_error_device(). Please
refer to commit 60271ab044a5 ("PCI/AER: Take reference on error
devices") which introduced this reference counting mechanism.

It feels awkward to leak AER-specific details into pcie_do_recovery().
That function is supposed to implement the flow described in
Documentation/PCI/pci-error-recovery.rst in a platform-agnostic way
so that powerpc (EEH) and s390 could conceivably take advantage of it.

Can you find a way to avoid this, e.g. report errors after
pcie_do_recovery() has concluded?

I understand your concern about keeping pcie_do_recovery()
platform-agnostic. I explored the possibility of reporting errors after
recovery concludes, but unfortunately, this approach isn't feasible due
to the recovery sequence. The issue is that most drivers'
pci_error_handlers implement .slot_reset() which internally calls
pci_restore_state() to restore the device's configuration space and
state. This function also clears the device's AER status registers:

   .slot_reset()
     => pci_restore_state()
       => pci_aer_clear_status()

Therefore, the only window to capture and report the original error
information is between link recovery (after reset_subordinates()) and
before .slot_reset() is called. Once .slot_reset() executes, the error
status is cleared and lost forever.

I'm also worried that errors are reported *during* recovery.
I imagine this looks confusing to a user.  The logged messages
should make it clear that these are errors that occurred *earlier*
and are reported belatedly.

You raise an excellent point about potential user confusion. The current
aer_print_error() interface doesn't indicate that these are historical
errors being reported belatedly. Would it be acceptable to add a
clarifying message before calling aer_print_error()? For example:

   pci_err(dev, "Reporting error that occurred before recovery:\n");

Alternatively, if you have suggestions for a better approach to make
this timing clear to users, I'd be happy to implement them.

Thanks,

Lukas

Thanks for valuable comments.

Best Regards,
Shuai