Re: [PATCH for-linus] PCI/AER: Skip error recovery on false alarms

From
Bjorn Helgaas <helgaas@kernel.org>
Date
2026-10-08 19:17:08
Also in
amd-gfx, linux-pci
Reply to
Lukas Wunner, 2026-10-08
[+cc Thorsten]

On Thu, Oct 08, 2026 at 02:26:00PM +0200, Lukas Wunner wrote:
Alex is seeing a probe failure of the amdgpu driver after the Root Port
above an AMD Navi10 GPU has been reset.  The reset was performed to
recover from a Firmware First reported Fatal Error.

However all status registers in the Root Port's AER Extended Capability
are blank, so apparently the platform firmware raised a false alarm.

The issue is only occurring since commit eddba19b8b5f ("PCI/AER: Support
Advisory Non-Fatal Errors").  It looks like enabling Advisory Non-Fatal
Errors causes code paths to be exercised in platform firmware which
were never validated before.

Skip error recovery on false alarms, i.e. if no unmasked errors were
actually signaled.

Note that this will also skip recovery if both the Status and Mask
registers are "all ones", as would be the case for inaccessible devices.
However that seems justified because it would imply either a hot-unplug
event or a Surprise Down Error further up in the hierarchy.  Interfering
with recovery from that seems uncalled for.

Fixes: eddba19b8b5f ("PCI/AER: Support Advisory Non-Fatal Errors")
Reported-by: Alex Deucher <alexander.deucher@amd.com>
Tested-by: Alex Deucher <alexander.deucher@amd.com>
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=222095
Signed-off-by: Lukas Wunner <lukas@wunner.de>
Applied to pci/for-linus for v7.3, thanks!  I dropped Alex's Tested-by
since he hasn't tested this change by itself.

If we're confident that this also fixes the MacBookPro16,1 power-off
issue reported by Jason (it would be ideal if you could test this,
Jason), maybe it's ok to keep eddba19b8b5f plus this patch for v7.3.
If so, I'd like to add Jason's Reported-by and link.

But if we're not sure whether this fixes the MacBookPro16,1 power-off
issue, I think we'll have to revert eddba19b8b5f for v7.3, then squash
it with this fix and try again for v7.4.

As far as I'm aware, the reports are:

  https://lore.kernel.org/all/CADnq5_MO+ZNOzW+_EH+gUYZ27X_ggWJJ0XdK6m2QW53m_ThtUQ@mail.gmail.com/ (local)
    Alex's report for AMD Navi10 and Vega20

  https://lore.kernel.org/all/CABZrw2HcaOh3DHTJtxU5dXrb3Vg19scr6+qR-xe_+R7rME35_Q@mail.gmail.com (local)
    Jason's report of power-off issue on MacBookPro16,1
quoted hunk ↗ jump to hunk
---
 When applied to pci/for-linus, this will cause a conflict during the
 merge window with a commit queued on pci/aer, d4c842c3af5b ("PCI/AER:
 Fix memory leak in aer_recover_work_func() when pci_dev is missing").
 
 To resolve the conflict, change "if (pdev)" to "if (pdev && err)"
 and move the pci_dev_put() out of the if-clause (so that it gets
 called if pdev != NULL but err == 0).
 
 If this is all too complicated and/or late, I can respin on top of
 pci/aer or v7.4-rc1.
 
 drivers/pci/pcie/aer.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)
diff --git a/drivers/pci/pcie/aer.c b/drivers/pci/pcie/aer.c
index d8dcd238fda1..922a726a52a5 100644
--- a/drivers/pci/pcie/aer.c
+++ b/drivers/pci/pcie/aer.c
@@ -1360,8 +1360,10 @@ static DEFINE_KFIFO(aer_recover_ring, struct aer_recover_entry,
 
 static void aer_recover_work_func(struct work_struct *work)
 {
+	struct aer_capability_regs *regs;
 	struct aer_recover_entry entry;
 	struct pci_dev *pdev;
+	u32 err;
 
 	while (kfifo_get(&aer_recover_ring, &entry)) {
 		pdev = pci_get_domain_bus_and_slot(entry.domain, entry.bus,
@@ -1375,6 +1377,12 @@ static void aer_recover_work_func(struct work_struct *work)
 		}
 		pci_print_aer(pdev, entry.severity, entry.regs);
 
+		regs = entry.regs;
+		if (entry.severity == AER_CORRECTABLE)
+			err = regs->cor_status & ~regs->cor_mask;
+		else
+			err = regs->uncor_status & ~regs->uncor_mask;
+
 		/*
 		 * Memory for aer_capability_regs(entry.regs) is being
 		 * allocated from the ghes_estatus_pool to protect it from
@@ -1385,12 +1393,14 @@ static void aer_recover_work_func(struct work_struct *work)
 		ghes_estatus_pool_region_free((unsigned long)entry.regs,
 					    sizeof(struct aer_capability_regs));
 
-		if (entry.severity == AER_NONFATAL)
-			pcie_do_recovery(pdev, pci_channel_io_normal,
-					 aer_root_reset);
-		else if (entry.severity == AER_FATAL)
-			pcie_do_recovery(pdev, pci_channel_io_frozen,
-					 aer_root_reset);
+		if (err) {
+			if (entry.severity == AER_NONFATAL)
+				pcie_do_recovery(pdev, pci_channel_io_normal,
+						 aer_root_reset);
+			else if (entry.severity == AER_FATAL)
+				pcie_do_recovery(pdev, pci_channel_io_frozen,
+						 aer_root_reset);
+		}
 		pci_dev_put(pdev);
 	}
 }
-- 
2.53.0
  
Thread (13 messages) 13 messages, 5 authors, last reply 2026-10-12
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help