Re: [PATCH] PCI/AER: Enable SERR# forwarding in non ACPI flow

From: poza
Date: Thu Aug 02 2018 - 12:02:12 EST


On 2018-08-02 11:53, poza@xxxxxxxxxxxxxx wrote:
On 2018-08-01 04:17, Bjorn Helgaas wrote:
On Thu, Jul 12, 2018 at 08:15:19PM +0530, Bharat Kumar Gogada wrote:
Currently PCI_BRIDGE_CTL_SERR is being enabled only in
ACPI flow.
This bit is required for forwarding errors reported
by EP devices to upstream device.
This patch enables SERR# for Type-1 PCI device.

This does seem broken.

Figure 6-3 in PCIe r4.0, sec 6.2.6, would be a helpful reference to
include in the commit log.

Semi-related question: there are about 40 drivers that call
pci_enable_pcie_error_reporting() and
pci_disable_pcie_error_reporting(). I see that the PCI core
calls pci_enable_pcie_error_reporting() for Root Ports and Switch
Ports in this path:

aer_probe # for root ports only
aer_enable_rootport
set_downstream_devices_error_reporting
set_device_error_reporting
if (ROOT_PORT || UPSTREAM || DOWNSTREAM)
pci_enable_pcie_error_reporting
pci_walk_bus(..., set_device_error_reporting)

But the core doesn't call pci_enable_pcie_error_reporting() for
endpoints. I wonder why not. Could we? And then remove the calls
from those drivers? If PCI_EXP_AER_FLAGS should only be set if the
driver is prepared, the pci_driver.err_handler would be a good hint.
But I suspect we could do something sensible and at least report
errors even if the driver doesn't have err_handler callbacks.


what about hot-plug case ?
should not aer_init() call pci_enable_pcie_error_reporting() for all
the downstream pci_dev ?
and remove all the calls from drivers..


aer_init will be called for each device (pci_dev) while pciehp does re-enumeration.
so probable we might want to call pci_enable_pcie_error_reporting()
but that dictates the design where AER framework is taking decision to enable error reporting on behalf of drivers as well.
but thats fine I think, if drivers do not want to participate then they have to call pci_disable_pcie_error_reporting explicitly.
does this make sense ?

On MIPS Octeon, it looks like pcibios_plat_dev_init() does already set
PCI_EXP_AER_FLAGS for every device.

But this question is obviously far beyond the scope of this current
patch.

Signed-off-by: Bharat Kumar Gogada <bharat.kumar.gogada@xxxxxxxxxx>
---
drivers/pci/pcie/aer.c | 23 +++++++++++++++++++++++
1 files changed, 23 insertions(+), 0 deletions(-)

diff --git a/drivers/pci/pcie/aer.c b/drivers/pci/pcie/aer.c
index a2e8838..943e084 100644
--- a/drivers/pci/pcie/aer.c
+++ b/drivers/pci/pcie/aer.c
@@ -343,6 +343,19 @@ int pci_enable_pcie_error_reporting(struct pci_dev *dev)
if (!dev->aer_cap)
return -EIO;

+ if (!IS_ENABLED(CONFIG_ACPI) &&
+ dev->hdr_type == PCI_HEADER_TYPE_BRIDGE) {

I think this test needs to be refined a little bit. If the kernel
happens to be built with CONFIG_ACPI=y but the current platform
doesn't support ACPI, we still want to set PCI_BRIDGE_CTL_SERR,
don't we?

+ u16 control;
+
+ /*
+ * A Type-1 PCI bridge will not forward ERR_ messages coming
+ * from an endpoint if SERR# forwarding is not enabled.
+ */
+ pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &control);
+ control |= PCI_BRIDGE_CTL_SERR;
+ pci_write_config_word(dev, PCI_BRIDGE_CONTROL, control);
+ }
+
return pcie_capability_set_word(dev, PCI_EXP_DEVCTL, PCI_EXP_AER_FLAGS);
}
EXPORT_SYMBOL_GPL(pci_enable_pcie_error_reporting);
@@ -352,6 +365,16 @@ int pci_disable_pcie_error_reporting(struct pci_dev *dev)
if (pcie_aer_get_firmware_first(dev))
return -EIO;

+ if (!IS_ENABLED(CONFIG_ACPI) &&
+ dev->hdr_type == PCI_HEADER_TYPE_BRIDGE) {
+ u16 control;
+
+ /* Clear SERR Forwarding */
+ pci_read_config_word(dev, PCI_BRIDGE_CONTROL, &control);
+ control &= ~PCI_BRIDGE_CTL_SERR;
+ pci_write_config_word(dev, PCI_BRIDGE_CONTROL, control);
+ }
+
return pcie_capability_clear_word(dev, PCI_EXP_DEVCTL,
PCI_EXP_AER_FLAGS);
}
--
1.7.1