PCI/DOE: Poll DOE Busy bit for up to 1 second in pci_doe_send_req()

During initial device probe, the PCI DOE busy bit for some CXL devices may
be left set for a longer period than expected by the current driver logic.
Despite local comments stating DOE Busy is unlikely to be detected, it
appears commonly specifically during boot when CXL devices are being
probed.

The symptom was messages like this:

  endpoint6: DOE failed -EBUSY

produced by cxl_cdat_get_length() or cxl_cdat_read_table().

This was observed on a single socket AMD platform with 2 CXL memory
expanders attached to the single socket. It was not the case that
concurrent accesses were being made, as validated by monitoring mailbox
commands on the device side.

This behavior has been observed with multiple CXL memory expanders from
different vendors - so it appears unrelated to the model.

In all observed tests, only a small period of the retry window is actually
used - typically only a handful of loop iterations.

Polling on the PCI DOE Busy Bit for (at max) one PCI DOE timeout interval
(1 second), resolves this issue cleanly.

Per PCIe r6.2 sec 6.30.3, the DOE Busy Bit being cleared does not raise an
interrupt, so polling is the best option in this scenario.

Subsequent code in doe_statemachine_work() and abort paths also wait for up
to 1 PCI DOE timeout interval, so this order of (potential) additional
delay is presumed acceptable.

Suggested-by: Lukas Wunner <lukas@wunner.de>
Link: https://lore.kernel.org/r/20241004162828.314-1-gourry@gourry.net
Signed-off-by: Gregory Price <gourry@gourry.net>
[bhelgaas: fix nits and add error message to commit log]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Lukas Wunner <lukas@wunner.de>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
This commit is contained in:
Gregory Price 2024-10-04 12:28:28 -04:00 committed by Bjorn Helgaas
parent 9852d85ec9
commit 86efc62d03

View File

@ -146,6 +146,7 @@ static int pci_doe_send_req(struct pci_doe_mb *doe_mb,
{
struct pci_dev *pdev = doe_mb->pdev;
int offset = doe_mb->cap_offset;
unsigned long timeout_jiffies;
size_t length, remainder;
u32 val;
int i;
@ -155,8 +156,19 @@ static int pci_doe_send_req(struct pci_doe_mb *doe_mb,
* someone other than Linux (e.g. firmware) is using the mailbox. Note
* it is expected that firmware and OS will negotiate access rights via
* an, as yet to be defined, method.
*
* Wait up to one PCI_DOE_TIMEOUT period to allow the prior command to
* finish. Otherwise, simply error out as unable to field the request.
*
* PCIe r6.2 sec 6.30.3 states no interrupt is raised when the DOE Busy
* bit is cleared, so polling here is our best option for the moment.
*/
timeout_jiffies = jiffies + PCI_DOE_TIMEOUT;
do {
pci_read_config_dword(pdev, offset + PCI_DOE_STATUS, &val);
} while (FIELD_GET(PCI_DOE_STATUS_BUSY, val) &&
!time_after(jiffies, timeout_jiffies));
if (FIELD_GET(PCI_DOE_STATUS_BUSY, val))
return -EBUSY;