On 10/2/2024 5:16 PM, Dan Williams wrote:
Smita Koralahalli wrote:
When PCIe AER is in FW-First, OS should process CXL Protocol errors from
CPER records.
Reuse the existing work queue cxl_cper_work registered with GHES to notify
the CXL subsystem on a Protocol error.
The defined trace events cxl_aer_uncorrectable_error and
cxl_aer_correctable_error currently trace native CXL AER errors. Reuse
them to trace FW-First Protocol Errors.
Signed-off-by: Smita Koralahalli <Smita.KoralahalliChannabasappa@xxxxxxx>
---
v2:
Removed pr_warn for serial number.
p_err -> rec/p_rec.
---
drivers/acpi/apei/ghes.c | 14 ++++++++++++++
drivers/cxl/core/pci.c | 24 ++++++++++++++++++++++++
drivers/cxl/cxlpci.h | 3 +++
drivers/cxl/pci.c | 20 ++++++++++++++++++--
include/cxl/event.h | 1 +
5 files changed, 60 insertions(+), 2 deletions(-)
diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
index 9dcf0f78458f..5082885e1f2c 100644
--- a/drivers/acpi/apei/ghes.c
+++ b/drivers/acpi/apei/ghes.c
@@ -723,6 +723,20 @@ static void cxl_cper_handle_prot_err(struct acpi_hest_generic_data *gdata)
if (cxl_cper_handle_prot_err_info(gdata, &wd.p_rec))
return;
+
+ guard(spinlock_irqsave)(&cxl_cper_work_lock);
+
+ if (!cxl_cper_work)
+ return;
+
+ wd.event_type = CXL_CPER_EVENT_PROT_ERR;
+
+ if (!kfifo_put(&cxl_cper_fifo, wd)) {
+ pr_err_ratelimited("CXL CPER kfifo overflow\n");
+ return;
+ }
+
+ schedule_work(cxl_cper_work);
The cxl_cper_work item is only for cases where the cxl_pci driver might care
about annotating an error report with driver specific details like the
impacted kernel object name, 'struct cxl_memdev', or address translation
for DPA data.
Protocol errors that are not endpoint errors should never be placed in
the cxl_cper_fifo. That is exclusively for errors that cxl_pci needs to
consume.
My expectation is that similar to aer_recover_queue for PCIe protocol
errors CXL needs to grow a cxl_recover_queue that at a minimum triggers
new trace events to dump these records to RAS daemon.
I am struggling to think what useful information cxl_pci could ever
append to a protocol error event.
What is more likely is that later when Terry adds port error handling a
CPER protocol error record could trigger a new cxl_do_recovery() to
react to CXL topology errors that might impact downstream CXL devices.
In that case the notification will come through something like a new
'struct cxl_error_handlers *' hanging off 'struct pci_driver' since
accelerator drivers are going to have distinct error handling from
generic memory expanders.
Hmm, I agree handling device and protocol errors separately. Unless
Terry has something to add here, I can start working on a minimal
version of cxl_recover_queue as of now.
Thanks
Smita
