Re: [PATCH v1 1/1] PCI/ERR: Handle fatal error recovery for non-hotplug capable devices

From: Kuppuswamy, Sathyanarayanan
Date: Tue May 26 2020 - 21:31:27 EST


On 5/21/20 7:56 PM, Yicong Yang wrote:

On 2020/5/22 3:31, Kuppuswamy, Sathyanarayanan wrote:

On 5/21/20 3:58 AM, Yicong Yang wrote:
On 2020/5/21 1:04, Kuppuswamy, Sathyanarayanan wrote:

On 5/20/20 1:28 AM, Yicong Yang wrote:
On 2020/5/7 11:32, sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx wrote:
From: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx>

If there are non-hotplug capable devices connected to a given
port, then during the fatal error recovery(triggered by DPC or
AER), after calling reset_link() function, we cannot rely on
hotplug handler to detach and re-enumerate the device drivers
in the affected bus. Instead, we will have to let the error
recovery handler call report_slot_reset() for all devices in
the bus to notify about the reset operation. Although this is
only required for non hot-plug capable devices, doing it for
hotplug capable devices should not affect the functionality.

Along with above issue, this fix also applicable to following

Commit 6d2c89441571 ("PCI/ERR: Update error status after
reset_link()") added support to store status of reset_link()
call. Although this fixed the error recovery issue observed if
the initial value of error status is PCI_ERS_RESULT_DISCONNECT
or PCI_ERS_RESULT_NO_AER_DRIVER, it also discarded the status
result from report_frozen_detected. This can cause a failure to
recover if _NEED_RESET is returned by report_frozen_detected and
report_slot_reset is not invoked.

Such an event can be induced for testing purposes by reducing the
Max_Payload_Size of a PCIe bridge to less than that of a device
downstream from the bridge, and then initiating I/O through the
device, resulting in oversize transactions. In the presence of DPC,
this results in a containment event and attempted reset and recovery
via pcie_do_recovery. After 6d2c89441571 report_slot_reset is not
invoked, and the device does not recover.

[original patch is from jay.vosburgh@xxxxxxxxxxxxx]
[original patch link]
Fixes: 6d2c89441571 ("PCI/ERR: Update error status after reset_link()")
Signed-off-by: Jay Vosburgh <jay.vosburgh@xxxxxxxxxxxxx>
Signed-off-by: Kuppuswamy Sathyanarayanan <sathyanarayanan.kuppuswamy@xxxxxxxxxxxxxxx>
drivers/pci/pcie/err.c | 19 +++++++++++++++----
1 file changed, 15 insertions(+), 4 deletions(-)

diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
index 14bb8f54723e..db80e1ecb2dc 100644
--- a/drivers/pci/pcie/err.c
+++ b/drivers/pci/pcie/err.c
@@ -165,13 +165,24 @@ 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) {
+ } else {
+ pci_walk_bus(bus, report_normal_detected, &status);
+ }
+ if (status == PCI_ERS_RESULT_NEED_RESET) {
+ if (reset_link) {
+ if (reset_link(dev) != PCI_ERS_RESULT_RECOVERED)

we'll call reset_link() only if link is frozen. so it may have problem here.
you mean before this change right?
After this change, reset_link() will be called as long as status is

Yes. I think we should reset the link only if the io is blocked as before. There's
no reason to reset a normal link.
Currently, only AER and DPC driver uses pcie_do_recovery() call. So the
possible reset_link options are dpc_reset_link() and aer_root_reset().

In dpc_reset_link() case, the link is already disabled and hence we
don't need to do another reset. In case of aer_root_reset() it
uses pci_bus_error_reset() to reset the slot.

Not exactly. In pci_bus_error_reset(), we call pci_slot_reset() only if it's
hotpluggable. But we always call pci_bus_reset() to perform a secondary bus
reset for the bridge. That's what I think is unnecessary for a normal link,
and that's what reset link indicates us to do. The slot reset is introduced
in the process only to solve side effects. (c4eed62a2143, PCI/ERR: Use slot reset if available)

IIUC, pci_bus_reset() will do slot reset if its supported (hot-plug
capable slots). If its not supported then it will attempt secondary
bus reset. So secondary bus reset will be attempted only if slot
reset is not supported.

Since reported_error_detected() requests us to do reset, we will have
to attempt some kind of reset before we call ->slot_reset() right?
What is the side effect in calling secondary bus reset?

PCI_ERS_RESULT_NEED_RESET indicates that the driver
wants a platform-dependent slot reset and its ->slot_reset() method to be called then.
I don't think it's same as slot reset mentioned above, which is only for hotpluggable
What you think is the correct reset implementation ? Is it something
like this?

if (hotplug capable)

Previously, if link is normal and the driver reports PCI_ERS_RESULT_NEED_RESET,
we'll only call ->slot_reset() without slot reset in reset_link(). Maybe it's better
to perform just like before.


Furthermore, PCI_ERS_RESULT_NEED_RESET means device driver requires a slot reset rather
than a link reset, so it maybe improper to use it to judge whether a link reset is needed.
We decide whether to do a link reset only by the io state.



+ } else {
+ if (pci_bus_error_reset(dev))
+ }
+ if (status == PCI_ERS_RESULT_DISCONNECT) {
pci_warn(dev, "link reset failed\n");
goto failed;
- } else {
- pci_walk_bus(bus, report_normal_detected, &status);