Skip to contentSkip to conversation

x86,apic: set cpu_has_apic manually for discrete apic

RFC In review Cyrill Gorcunov 1 participant 2 messages Aug 16, 2009 – Aug 19, 2009 x86

lore
Share & tools
lore
#0 lore raw

RFC 1/1 x86,apic: set cpu_has_apic manually for discrete apic

In case of discrete apic (due to hardware or firmware reason) cpuid may not retrieve cpu_has_apic bit. But we rely on this bit in SMP compiled kernels. So instead of long condition checking we set cpu_has_apic bit manually and check it only in further code flow.

Note that the case is applied to x86-32 pretty old machines only. Also verify_local_APIC is the only place where such a trick could be done. We can't do it earlier since there could be a BIOS bug and other checks which we are to pass to ensure that APIC is functional.

The patch also make lapic_shutdown, disable_IO_APIC callable which should take place even on discrete apics.

As a side effect it makes lapic sysfs device available for such machines so lets check for apic being integrated since we don't support suspend/resume for disrete apics.

---

Please review carefully. I've tested it with manually disabled cpu_has_apic and apic was set to "discrete" state as well under qemu.

Actually this "two-line" patch could be more dangerous then it looks like and need more wide range _long_ testing.

Or, perhaps, we could just fix lapic_shutdown and disable_IO_APIC instead since old machines already that rare to find out.

Anyway if someone has 486DX machine please give the patch a try.

 arch/x86/kernel/apic/apic.c |   21 ++++++++++++++++++++-
 1 file changed, 20 insertions(+), 1 deletion(-)
arch/x86/kernel/apic/apic.c +20 -1
Index: linux-2.6.git/arch/x86/kernel/apic/apic.c
=====================================================================
@@ -1004,6 +1004,25 @@ int __init verify_local_APIC(void)
 	unsigned int reg0, reg1;
 
 	/*
+	 * For discrete APIC cpu_has_apic bit may not
+	 * be retrieved via cpuid instruction. Same time
+	 * we rely on this bit (mostly for SMP compiled
+	 * kernel) and instead of checking the long condition
+	 * (as shown below) all the time we may check for
+	 * plain cpu_has_apic further in code.
+	 *
+	 * At this moment all the checks are passed and we've
+	 * already complained if there was a BIOS bug.
+	 * This place is safe for such a trick. --cvg
+	 */
+#ifdef CONFIG_X86_32
+	if (smp_found_config && !cpu_has_apic && !disable_apic) {
+		set_cpu_cap(&boot_cpu_data, X86_FEATURE_APIC);
+		pr_info("APIC: cpu_has_apic being set manually\n");
+	}
+#endif
+
+	/*
 	 * The version register is read-only in a real APIC.
 	 */
 	reg0 = apic_read(APIC_LVR);
@@ -2145,7 +2164,7 @@ static int __init init_lapic_sysfs(void)
 {
 	int error;
 
-	if (!cpu_has_apic)
+	if (!cpu_has_apic || !lapic_is_integrated())
 		return 0;
 	/* XXX: remove suspend/resume procs if !apic_pm_state.active? */
 
#1 lore raw

[Cyrill Gorcunov - Sun, Aug 16, 2009 at 11:57:36PM +0400]

|
| In case of discrete apic (due to hardware or firmware reason)
| cpuid may not retrieve cpu_has_apic bit. But we rely on this
| bit in SMP compiled kernels. So instead of long condition checking
| we set cpu_has_apic bit manually and check it only in further
| code flow.
| 
| Note that the case is applied to x86-32 pretty old machines only.
| Also verify_local_APIC is the only place where such a trick
| could be done. We can't do it earlier since there could be
| a BIOS bug and other checks which we are to pass to ensure
| that APIC is functional.
| 
| The patch also make lapic_shutdown, disable_IO_APIC callable
| which should take place even on discrete apics.
| 
| As a side effect it makes lapic sysfs device available
| for such machines so lets check for apic being integrated
| since we don't support suspend/resume for disrete apics.
| 
| Signed-off-by: Cyrill Gorcunov <[email protected]>
| ---

Please drop this RFC now as to eliminate side effects a number of quirks (MCE,MSI) would be needed. I didn't trigger problems because not all interrupt types were covered with qemu. The simpliest (and more safe) approach will be just to fix lapic_shutdown and disable_IO_APIC for discrete apic case. Will cook a patch.

-- Cyrill