Re: [PATCH V2 4/9] PCI/AER: Extend AER error handling to RCECs

From: Sean V Kelley
Date: Fri Aug 07 2020 - 20:55:24 EST

On 7 Aug 2020, at 15:53, Bjorn Helgaas wrote:

On Tue, Aug 04, 2020 at 12:40:47PM -0700, Sean V Kelley wrote:
From: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>

Currently the kernel does not handle AER errors for Root Complex integrated
End Points (RCiEPs)[0]. These devices sit on a root bus within the Root Complex
(RC). AER handling is performed by a Root Complex Event Collector (RCEC) [1]
which is a effectively a type of RCiEP on the same root bus.

For an RCEC (technically not a Bridge), error messages "received" from
associated RCiEPs must be enabled for "transmission" in order to cause a
System Error via the Root Control register or (when the Advanced Error
Reporting Capability is present) reporting via the Root Error Command
register and logging in the Root Error Status register and Error Source
Identification register.

In addition to the defined OS level handling of the reset flow for the
associated RCiEPs of an RCEC, it is possible to also have a firmware first
model. In that case there is no need to take any actions on the RCEC because
the firmware is responsible for them. This is true where APEI [2] is used
to report the AER errors via a GHES[v2] HEST entry [3] and relevant
AER CPER record [4] and Firmware First handling is in use.

I don't see anything in the patch that mentions "firmware first." Do
we need it in the commit log? After
https://git.kernel.org/linus/708b20003624 ("PCI/AER: Remove
HEST/FIRMWARE_FIRST parsing for AER ownership"), I think we no longer
know anything about firmware-first in the kernel.

I’ll let Jonathan reply here.


We effectively end up with two different types of discovery for
purposes of handling AER errors:

1) Normal bus walk - we pass the downstream port above a bus to which
the device is attached and it walks everything below that point.

2) An RCiEP with no visible association with an RCEC as there is no need to
walk devices. In that case, the flow is to just call the callbacks for the actual
device.

A new walk function, similar to pci_bus_walk is provided that takes a pci_dev
instead of a bus. If that dev corresponds to a downstream port it will walk
the subordinate bus of that downstream port. If the dev does not then it
will call the function on that device alone.

Maybe mention the new function name here?

Agree, I will mention it here.


Add "()" after function names in commit logs and comments so they
don't look like English words.

Will fix.


Wrap commit logs so they fit in 75 columns, so they don't wrap when
"git log" indents them in a default 80 column window. Yes, I know I
could use wider windows, but I'd still want *some* default so commits
don't just have random widths.

Will fix.


[0] ACPI PCI Express Base Specification 5.0-1 1.3.2.3 Root Complex Integrated
Endpoint Rules.
[1] ACPI PCI Express Base Specification 5.0-1 6.2 Error Signalling and Logging
[2] ACPI Specification 6.3 Chapter 18 ACPI Platform Error Interface (APEI)
[3] ACPI Specification 6.3 18.2.3.7 Generic Hardware Error Source
[4] UEFI Specification 2.8, N.2.7 PCI Express Error Section

Signed-off-by: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
Signed-off-by: Sean V Kelley <sean.v.kelley@xxxxxxxxx>
---
drivers/pci/pcie/err.c | 59 +++++++++++++++++++++++++++++++++---------
1 file changed, 47 insertions(+), 12 deletions(-)

diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
index c543f419d8f9..682302dfb55b 100644
--- a/drivers/pci/pcie/err.c
+++ b/drivers/pci/pcie/err.c
@@ -146,38 +146,69 @@ static int report_resume(struct pci_dev *dev, void *data)
return 0;
}

+/**
+ * pci_walk_dev_affected - walk devices potentially AER affected
+ * @dev device which may be an RCEC with associated RCiEPs,
+ * an RCiEP associated with an RCEC, or a Port.

Does this mean that if dev is an RCEC, we call the callback for the
*RCEC* itself? I would have thought we'd want to do that for the
associated *RCiEPs*?

Yes, we do. The errors can come from either an RCEC or its respective RCiEPs. Both Root Port and RCEC can report error for themselves. So both an error Root Port device and an error RCEC device can be passed here for error handling. In fact, the bit corresponding to the device number of the RCEC is always set in the RCiEP Bitmap (section 7.9.2). And an RCEC must also follow all the rules for an RCiEP (section 1.3.4).

I was wanting to do a test with an AER injection of the RCEC itself. But it looks like current aer_inject.c doesn’t support injecting an error to Root Ports or RCECs. Will need to take a look at it.


+ * @cb callback to be called for each device found
+ * @userdata arbitrary pointer to be passed to callback.
+ *
+ * If the device provided is a port, walk the subordinate bus,

This usage of "port" doesn't seem quite right. "Port" includes root
ports, switch upstream ports, switch downstream ports, *and* the
upstream ports on endpoints. The endpoint upstream ports obviously
don't have subordinate buses. We typically use "bridge" as the
generic term for something with a subordinate bus.

Okay, that makes sense. Will correct with “bridge”.


+ * including any bridged devices on buses under this bus.
+ * Call the provided callback on each device found.
+ *
+ * If the device provided has no subordinate bus, call the provided
+ * callback on the device itself.
+ */
+static void pci_walk_dev_affected(struct pci_dev *dev, int (*cb)(struct pci_dev *, void *),

I don't understand the "affected" reference in the function name.
This doesn't test anything to see whether devices are "affected".
Naming is the hardest part of programming :)

In earlier discussion, Cameron had suggested pci_walk_aer_affected(). But I thought perhaps that the focus should be on the devices. Perhaps a better description would be pci_walk_aer_devices() or something along those lines. The original incarnation was pci_walk_below_dev().

I’m open to anything, really.


+ void *userdata)
+{
+ if (dev->subordinate) {
+ pci_walk_bus(dev->subordinate, cb, userdata);
+ } else {
+ cb(dev, userdata);
+ }

Typical Linux style omits {} for single-line if/else branches.

Will fix.


+}
+
pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
pci_channel_state_t state,
pci_ers_result_t (*reset_link)(struct pci_dev *pdev))
{
pci_ers_result_t 'status = PCI_ERS_RESULT_CAN_RECOVER;
- struct pci_bus *bus;

/*
* Error recovery runs on all subordinates of the first downstream port.
* If the downstream port detected the error, it is cleared at the end.
+ * For RCiEPs we should reset just the RCiEP itself.
*/
if (!(pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT ||
- pci_pcie_type(dev) == PCI_EXP_TYPE_DOWNSTREAM))
+ pci_pcie_type(dev) == PCI_EXP_TYPE_DOWNSTREAM ||
+ pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END ||
+ pci_pcie_type(dev) == PCI_EXP_TYPE_RC_EC))
dev = dev->bus->self;
- bus = dev->subordinate;

pci_dbg(dev, "broadcast error_detected message\n");
if (state == pci_channel_io_frozen) {
- pci_walk_bus(bus, report_frozen_detected, &status);
+ pci_walk_dev_affected(dev, report_frozen_detected, &status);
+ if (pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END) {
+ pci_warn(dev, "link reset not possible for RCiEP\n");
+ status = PCI_ERS_RESULT_NONE;
+ goto failed;
+ }
+
status = reset_link(dev);

reset_link() might be misnamed. IIUC "dev" is a bridge, and the point
is really to reset any devices below "dev." Whether we do that by
resetting link, DPC trigger, secondary bus reset, FLR, etc, is sort of
immaterial. Some of those methods might be applicable for RCiEPs.

But you didn't add that name; I'm just trying to understand this
better.

Yes, that’s a confusing term with the _link attached. It’s difficult to relate to the different resets that might be applicable. I was thinking about that when looking at the callback path via the “reset_link” of the RCiEP to the RCEC for the sole purpose of clearing the Root Port Error Status. It would be worth time to spend looking at better descriptive naming/methods.


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

if (status == PCI_ERS_RESULT_CAN_RECOVER) {
status = PCI_ERS_RESULT_RECOVERED;
pci_dbg(dev, "broadcast mmio_enabled message\n");
- pci_walk_bus(bus, report_mmio_enabled, &status);
+ pci_walk_dev_affected(dev, report_mmio_enabled, &status);
}

if (status == PCI_ERS_RESULT_NEED_RESET) {
@@ -188,18 +219,22 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
*/
status = PCI_ERS_RESULT_RECOVERED;
pci_dbg(dev, "broadcast slot_reset message\n");
- pci_walk_bus(bus, report_slot_reset, &status);
+ pci_walk_dev_affected(dev, report_slot_reset, &status);
}

if (status != PCI_ERS_RESULT_RECOVERED)
goto failed;

pci_dbg(dev, "broadcast resume message\n");
- pci_walk_bus(bus, report_resume, &status);
-
- if (pcie_aer_is_native(dev))
- pcie_clear_device_status(dev);
- pci_aer_clear_nonfatal_status(dev);
+ pci_walk_dev_affected(dev, report_resume, &status);
+
+ if ((pci_pcie_type(dev) == PCI_EXP_TYPE_ROOT_PORT ||
+ pci_pcie_type(dev) == PCI_EXP_TYPE_DOWNSTREAM ||
+ pci_pcie_type(dev) == PCI_EXP_TYPE_RC_EC)) {
+ if (pcie_aer_is_native(dev))
+ pcie_clear_device_status(dev);
+ pci_aer_clear_nonfatal_status(dev);

This change (testing pci_pcie_type()) looks like it's not strictly
related to the rest of this patch and maybe should be split out into
its own patch?

This change was also based on a commit (068c29a24) in the pci/next branch. The type testing was brought over from Jonathan’s original V2, but actually, it went full circle by adding the RC_EC type, because now it was no longer a no-op. There was an original concern about the need for those to be called on the RCEC from Jonathan’s RFC.

Thoughts Jonathan?

Thanks,

Sean


+ }
pci_info(dev, "device recovery successful\n");
return status;

--
2.27.0