Hi, Matthew,

On 06/14/2011 10:52 PM, Matthew Garrett wrote:

On Wed, May 25, 2011 at 02:05:38PM +0800, Huang Ying wrote:

quoted

To gain full APEI power on these machines, a special APEI _OSC needs
to be evaluated to tell firmware that Linux has full APEI support.
This patch add the APEI _OSC support.

(snip)

quoted

+	static DEFINE_MUTEX(mutex);
+	static int status = APEI_OSC_SETUP_UNKNOWN;
+	static u8 apei_uuid_str[] = "ed855e0c-6c90-47bf-a62a-26de0fc5ad5c";

This is the WHEA UUID, right? 

Yes.

quoted

+	u32 capbuf[3];
+	struct acpi_osc_context context = {
+		.uuid_str	= apei_uuid_str,
+		.rev		= 1,
+		.cap.length	= sizeof(capbuf),
+		.cap.pointer	= capbuf,
+	};
+
+	mutex_lock(&mutex);
+	if (status == APEI_OSC_SETUP_UNKNOWN) {
+		capbuf[OSC_QUERY_TYPE] = OSC_QUERY_ENABLE;
+		capbuf[OSC_SUPPORT_TYPE] = 0;
+		capbuf[OSC_CONTROL_TYPE] = 0;
+
+		if (ACPI_FAILURE(acpi_get_handle(NULL, "\\_SB", &handle))
+		    || ACPI_FAILURE(acpi_run_osc(handle, &context))) {
+			pr_err(APEI_PFX "APEI _OSC failed!\n");
+			status = APEI_OSC_SETUP_FAILED;
+		} else {
+			kfree(context.ret.pointer);
+			status = APEI_OSC_SETUP_SUCCEEDED;
+		}
+	}
+	mutex_unlock(&mutex);
+
+	return status == APEI_OSC_SETUP_SUCCEEDED ? 0 : -EIO;

So we fail if the platform doesn't implement WHEA...

quoted

+	rc = apei_osc_setup();
+	if (rc) {
+		ghes_remove(ghes_dev);
+		return rc;
+	}
+

And then tear down GHES. This seems wrong. A platform could predicate 
APEI functionality on the ACPI spec APEI indication (which we currently 
don't pass) without implementing WHEA, but with this patch we'd refuse 
to enable GHES support? We should probably try both the standard method 
and the WHEA method and only disable GHES if both fail.

You means the "APEI Support" bit for standard UUID?  Do you know which
machine uses this bit?  I can write the code, but I have no machine to
test it.

BTW, it is better for us to enable APEI firmware first mode (that is,
what is enabled by evaluating the WHEA UUID) after GHES reporting is
ready (that is, after GHES module is successfully loaded).  That is
later than current ACPI _OSC evaluation with standard UUID.  Is it
possible to evaluate _OSC with standard UUID twice?  So that we can
enable APEI firmware first mode later.

(Also, are there any other sideeffects of indicating that we support 
WHEA?)

After evaluating _OSC with this UUID, firmware will produce error record
to OS, otherwise only unknown NMI.

Best Regards,
Huang Ying