Re: [PATCH v9 12/15] PCI/RCEC: Add RCiEP's linked RCEC to AER/ERR

From: Sean V Kelley
Date: Sat Oct 17 2020 - 12:14:48 EST

Next message: Shuah Khan: "Re: [PATCH 4.9 00/16] 4.9.240-rc1 review"
Previous message: Shuah Khan: "Re: [PATCH 4.14 00/18] 4.14.202-rc1 review"
In reply to: Bjorn Helgaas: "Re: [PATCH v9 12/15] PCI/RCEC: Add RCiEP's linked RCEC to AER/ERR"
Next in thread: Sean V Kelley: "Re: [PATCH v9 12/15] PCI/RCEC: Add RCiEP's linked RCEC to AER/ERR"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On 16 Oct 2020, at 13:30, Bjorn Helgaas wrote:

[+to Jonathan]

On Thu, Oct 15, 2020 at 05:11:10PM -0700, Sean V Kelley wrote:

From: Qiuxu Zhuo <qiuxu.zhuo@xxxxxxxxx>

When attempting error recovery for an RCiEP associated with an RCEC device,
there needs to be a way to update the Root Error Status, the Uncorrectable
Error Status and the Uncorrectable Error Severity of the parent RCEC. In
some non-native cases in which there is no OS-visible device associated
with the RCiEP, there is nothing to act upon as the firmware is acting
before the OS.

Add handling for the linked RCEC in AER/ERR while taking into account
non-native cases.

Co-developed-by: Sean V Kelley <sean.v.kelley@xxxxxxxxx>
Link: https://lore.kernel.org/r/20201002184735.1229220-12-seanvk.dev@xxxxxxxxxxxxxxxx
Signed-off-by: Sean V Kelley <sean.v.kelley@xxxxxxxxx>
Signed-off-by: Qiuxu Zhuo <qiuxu.zhuo@xxxxxxxxx>
Signed-off-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
---
drivers/pci/pcie/aer.c | 53 ++++++++++++++++++++++++++++++------------
drivers/pci/pcie/err.c | 20 ++++++++--------
2 files changed, 48 insertions(+), 25 deletions(-)

diff --git a/drivers/pci/pcie/aer.c b/drivers/pci/pcie/aer.c
index 65dff5f3457a..083f69b67bfd 100644
--- a/drivers/pci/pcie/aer.c
+++ b/drivers/pci/pcie/aer.c
@@ -1357,27 +1357,50 @@ static int aer_probe(struct pcie_device *dev)
*/
static pci_ers_result_t aer_root_reset(struct pci_dev *dev)
{
- int aer = dev->aer_cap;
+ int type = pci_pcie_type(dev);
+ struct pci_dev *root;
+ int aer = 0;
+ int rc = 0;
u32 reg32;
- int rc;

+ if (pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END)

"type == PCI_EXP_TYPE_RC_END"

Right, I merged your suggested changes which added the type. Will correct.

+ /*
+ * The reset should only clear the Root Error Status
+ * of the RCEC. Only perform this for the
+ * native case, i.e., an RCEC is present.
+ */
+ root = dev->rcec;
+ else
+ root = dev;

- /* Disable Root's interrupt in response to error messages */
- pci_read_config_dword(dev, aer + PCI_ERR_ROOT_COMMAND, &reg32);
- reg32 &= ~ROOT_PORT_INTR_ON_MESG_MASK;
- pci_write_config_dword(dev, aer + PCI_ERR_ROOT_COMMAND, reg32);
+ if (root)
+ aer = dev->aer_cap;

- rc = pci_bus_error_reset(dev);
- pci_info(dev, "Root Port link has been reset\n");
+ if (aer) {
+ /* Disable Root's interrupt in response to error messages */
+ pci_read_config_dword(root, aer + PCI_ERR_ROOT_COMMAND, &reg32);
+ reg32 &= ~ROOT_PORT_INTR_ON_MESG_MASK;
+ pci_write_config_dword(root, aer + PCI_ERR_ROOT_COMMAND, reg32);

Not directly related to *this* patch, but my assumption was that in
the APEI case, the firmware should retain ownership of the AER
Capability, so the OS should not touch PCI_ERR_ROOT_COMMAND and
PCI_ERR_ROOT_STATUS.

But this code appears to ignore that ownership. Jonathan, you must
have looked at this recently for 068c29a248b6 ("PCI/ERR: Clear PCIe
Device Status errors only if OS owns AER"). Do you have any insight
about this?

- /* Clear Root Error Status */
- pci_read_config_dword(dev, aer + PCI_ERR_ROOT_STATUS, &reg32);
- pci_write_config_dword(dev, aer + PCI_ERR_ROOT_STATUS, reg32);
+ /* Clear Root Error Status */
+ pci_read_config_dword(root, aer + PCI_ERR_ROOT_STATUS, &reg32);
+ pci_write_config_dword(root, aer + PCI_ERR_ROOT_STATUS, reg32);

- /* Enable Root Port's interrupt in response to error messages */
- pci_read_config_dword(dev, aer + PCI_ERR_ROOT_COMMAND, &reg32);
- reg32 |= ROOT_PORT_INTR_ON_MESG_MASK;
- pci_write_config_dword(dev, aer + PCI_ERR_ROOT_COMMAND, reg32);
+ /* Enable Root Port's interrupt in response to error messages */
+ pci_read_config_dword(root, aer + PCI_ERR_ROOT_COMMAND, &reg32);
+ reg32 |= ROOT_PORT_INTR_ON_MESG_MASK;
+ pci_write_config_dword(root, aer + PCI_ERR_ROOT_COMMAND, reg32);
+ }
+
+ if ((type == PCI_EXP_TYPE_RC_EC) || (type == PCI_EXP_TYPE_RC_END)) {
+ if (pcie_has_flr(root)) {
+ rc = pcie_flr(root);
+ pci_info(dev, "has been reset (%d)\n", rc);
+ }
+ } else {
+ rc = pci_bus_error_reset(root);

Don't we want "dev" for both the FLR and pci_bus_error_reset()? I
think "root == dev" except when dev is an RCiEP. When dev is an
RCiEP, "root" is the RCEC (if present), and we want to reset the
RCiEP, not the RCEC.

Right, when I did the goto in the earlier incarnation, I always set root to dev at the start and in the merge it needs to be dev always except for the RC_END where RCEC exists. Will change without bringing back the goto…

+ struct pci_dev *root = dev;

…

+non_native:
+ if ((type == PCI_EXP_TYPE_RC_EC) || (type ==
PCI_EXP_TYPE_RC_END)) {
+ rc = flr_on_rc(root);
+ pci_info(dev, "has been reset (%d)\n", rc);
+ } else {
+ rc = pci_bus_error_reset(root);
+ pci_info(dev, "Root Port link has been reset (%d)\n",
rc);
+ }

+ pci_info(dev, "Root Port link has been reset (%d)\n", rc);
+ }

There are a couple changes here that I think should be split out.

Based on my theory that when firmware retains control of AER, the OS
should not touch PCI_ERR_ROOT_COMMAND and PCI_ERR_ROOT_STATUS, and any
updates to them would have to be done by firmware before we get here,
I suggested reordering this:

- clear PCI_ERR_ROOT_COMMAND ROOT_PORT_INTR_ON_MESG_MASK
- do reset
- clear PCI_ERR_ROOT_STATUS (for APEI, presumably done by firmware?)
- enable PCI_ERR_ROOT_COMMAND ROOT_PORT_INTR_ON_MESG_MASK

to this:

- clear PCI_ERR_ROOT_COMMAND ROOT_PORT_INTR_ON_MESG_MASK
- clear PCI_ERR_ROOT_STATUS
- enable PCI_ERR_ROOT_COMMAND ROOT_PORT_INTR_ON_MESG_MASK
- do reset

If my theory is correct, I think we should still reorder this, but:

- It's a significant behavior change that deserves its own patch so
we can document/bisect/revert.

- I'm not sure why we clear the PCI_ERR_ROOT_COMMAND error reporting
bits. In the new "clear COMMAND, clear STATUS, enable COMMAND"
order, it looks superfluous. There's no reason to disable error
reporting while clearing the status bits.

The current "clear, reset, enable" order suggests that the reset
might cause errors that we should ignore. I don't know whether
that's the case or not. It dates from 6c2b374d7485 ("PCI-Express
AER implemetation: AER core and aerdriver"), which doesn't
elaborate.

- Should we also test for OS ownership of AER before touching
PCI_ERR_ROOT_STATUS?

- If we remove the PCI_ERR_ROOT_COMMAND fiddling (and I tentatively
think we *should* unless we can justify it), that would also
deserve its own patch. Possibly (1) remove PCI_ERR_ROOT_COMMAND
fiddling, (2) reorder PCI_ERR_ROOT_STATUS clearing and reset, (3)
test for OS ownership of AER (?), (4) the rest of this patch.

You’ve highlighted some good questions.

I think we should remove the fiddling until we have a clearer picture and put that into its own patch.

Sean

return rc ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
}
diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
index 7883c9791562..cbc5abfe767b 100644
--- a/drivers/pci/pcie/err.c
+++ b/drivers/pci/pcie/err.c
@@ -148,10 +148,10 @@ static int report_resume(struct pci_dev *dev, void *data)

/**
* pci_walk_bridge - walk bridges potentially AER affected
- * @bridge: bridge which may be a Port, an RCEC with associated RCiEPs,
- * or an RCiEP associated with an RCEC
- * @cb: callback to be called for each device found
- * @userdata: arbitrary pointer to be passed to callback
+ * @bridge bridge which may be an RCEC with associated RCiEPs,
+ * or a Port.
+ * @cb callback to be called for each device found
+ * @userdata arbitrary pointer to be passed to callback.
*
* If the device provided is a bridge, walk the subordinate bus, including
* any bridged devices on buses under this bus. Call the provided callback
@@ -164,8 +164,14 @@ static void pci_walk_bridge(struct pci_dev *bridge,
int (*cb)(struct pci_dev *, void *),
void *userdata)
{
+ /*
+ * In a non-native case where there is no OS-visible reporting
+ * device the bridge will be NULL, i.e., no RCEC, no Downstream Port.
+ */
if (bridge->subordinate)
pci_walk_bus(bridge->subordinate, cb, userdata);
+ else if (bridge->rcec)
+ cb(bridge->rcec, userdata);
else
cb(bridge, userdata);
}
@@ -194,12 +200,6 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
pci_dbg(bridge, "broadcast error_detected message\n");
if (state == pci_channel_io_frozen) {
pci_walk_bridge(bridge, report_frozen_detected, &status);
- if (type == PCI_EXP_TYPE_RC_END) {
- pci_warn(dev, "subordinate device reset not possible for RCiEP\n");
- status = PCI_ERS_RESULT_NONE;
- goto failed;
- }
-
status = reset_subordinates(bridge);
if (status != PCI_ERS_RESULT_RECOVERED) {
pci_warn(bridge, "subordinate device reset failed\n");
--
2.28.0

Next message: Shuah Khan: "Re: [PATCH 4.9 00/16] 4.9.240-rc1 review"
Previous message: Shuah Khan: "Re: [PATCH 4.14 00/18] 4.14.202-rc1 review"
In reply to: Bjorn Helgaas: "Re: [PATCH v9 12/15] PCI/RCEC: Add RCiEP's linked RCEC to AER/ERR"
Next in thread: Sean V Kelley: "Re: [PATCH v9 12/15] PCI/RCEC: Add RCiEP's linked RCEC to AER/ERR"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]