Re: [bisected] Clocksource tsc unstable git

[Borislav Petkov <bp@amd64.org>]

On Tue, Nov 02, 2010 at 11:26:54AM -0400, Thomas Gleixner wrote:
> Ok, I'm waiting for Boris to get us the confirmation from HW folks.

Ok, I don't have much but it could be something. We could use the
minimum tick u16 value in the ACPI HPET table, offset 53:

Main Counter Minimum   2 53 Unit: Clock tick
Clock_tick in Periodic      The minimum clock ticks can be set without lost
Mode                        interrupts while the counter is programmed to operate in
*Note#3                     periodic mode

...

* Note #3: This field is written by BIOS and may be chipset and/or
platform dependent. This indicates the minimum value that must be used
for any counter programmed in periodic mode to avoid lost interrupts.
For any counter x that has been configured for periodic mode, the number
can be programmed in any Tx_Compare Register must be greater than P,
where P = (Minimum Period) / (Main counter period) in order to avoid
lost interrupts.

However, we don't know (yet) whether this can be used in the
non-periodic mode too. My gut feeling says yes but I wouldn't trust it.

We went and collected that value a bunch of systems and it looks like it
would need more massaging (read: capping) since some BIOSen simply write
crap in it. It ranges from

- 0x37ee on old nVidia and Intel boards (this is definitely crap, I
can't imagine a minimum ticks value of 14318 for a HPET but who knows)

- 0x1000 on a HP machine (also fishy)

- 0x10, 0x14 on current SBxxx boards

- 0x80 on newer Intel boards

and it looks like 0x80 is also the default value which we can use as a
capping value. Btw, I've been running the following on the machine in
question for a while now and no lockups so far.

Opinions?

---
 arch/x86/include/asm/hpet.h |    1 +
 arch/x86/kernel/acpi/boot.c |   12 +++++++++---
 arch/x86/kernel/hpet.c      |    3 ++-
 3 files changed, 12 insertions(+), 4 deletions(-)

diff --git a/arch/x86/include/asm/hpet.h b/arch/x86/include/asm/hpet.h
index 2c392d6..01d9480 100644
--- a/arch/x86/include/asm/hpet.h
+++ b/arch/x86/include/asm/hpet.h
@@ -68,6 +68,7 @@ extern unsigned long force_hpet_address;
 extern u8 hpet_blockid;
 extern int hpet_force_user;
 extern u8 hpet_msi_disable;
+extern u16 hpet_min_tick;
 extern int is_hpet_enabled(void);
 extern int hpet_enable(void);
 extern void hpet_disable(void);
diff --git a/arch/x86/kernel/acpi/boot.c b/arch/x86/kernel/acpi/boot.c
index 71232b9..37b93e8 100644
--- a/arch/x86/kernel/acpi/boot.c
+++ b/arch/x86/kernel/acpi/boot.c
@@ -772,9 +772,6 @@ static int __init acpi_parse_hpet(struct acpi_table_header *table)
 		hpet_address >>= 32;
 	}
 #endif
-	printk(KERN_INFO PREFIX "HPET id: %#x base: %#lx\n",
-	       hpet_tbl->id, hpet_address);
-
 	/*
 	 * Allocate and initialize the HPET firmware resource for adding into
 	 * the resource tree during the lateinit timeframe.
@@ -790,6 +787,15 @@ static int __init acpi_parse_hpet(struct acpi_table_header *table)
 	hpet_res->start = hpet_address;
 	hpet_res->end = hpet_address + (1 * 1024) - 1;
 
+	hpet_min_tick = hpet_tbl->minimum_tick;
+	if (hpet_min_tick > 0x80) {
+		printk(KERN_WARNING PREFIX "Limiting HPET read delay to 0x80.\n");
+		hpet_min_tick = 0x80;
+	}
+
+	printk(KERN_INFO PREFIX "HPET id: %#x base: %#lx min tick: %d\n",
+		hpet_tbl->id, hpet_address, hpet_min_tick);
+
 	return 0;
 }
 
diff --git a/arch/x86/kernel/hpet.c b/arch/x86/kernel/hpet.c
index e49c3b9..a0b790a 100644
--- a/arch/x86/kernel/hpet.c
+++ b/arch/x86/kernel/hpet.c
@@ -40,6 +40,7 @@ u8					hpet_msi_disable;
 static unsigned long			hpet_num_timers;
 #endif
 static void __iomem			*hpet_virt_address;
+u16					hpet_min_tick = 0x80;
 
 struct hpet_dev {
 	struct clock_event_device	evt;
@@ -408,7 +409,7 @@ static int hpet_next_event(unsigned long delta,
 	 */
 	res = (s32)(cnt - hpet_readl(HPET_COUNTER));
 
-	return res < 8 ? -ETIME : 0;
+	return res < hpet_min_tick ? -ETIME : 0;
 }
 
 static void hpet_legacy_set_mode(enum clock_event_mode mode,
-- 
1.7.3.1.50.g1e633




-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
General Managers: Alberto Bozzo, Andrew Bowd
Registration: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632