Skip to contentSkip to conversation

powerpc: Enable DEEPNAP power savings mode on 970MP

PATCH In review Olof Johansson 3 participants 4 messages Oct 5, 2006 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc: Enable DEEPNAP power savings mode on 970MP

Hi,

Without this patch, on an idle system I get:

cpu-power-0:21.638
cpu-power-1:27.102
cpu-power-2:29.343
cpu-power-3:25.784
Total: 103.8W

With this patch:

cpu-power-0:11.730
cpu-power-1:17.185
cpu-power-2:18.547
cpu-power-3:17.528
Total: 65.0W

If I lower HZ to 100, I can get it as low as:

cpu-power-0:10.938
cpu-power-1:16.021
cpu-power-2:17.245
cpu-power-3:16.145
Total: 60.2W

Another (older) Quad G5 went from 54W to 39W at HZ=250.

Coming back out of Deep Nap takes 40-70 cycles longer than coming back from just Nap (which already takes quite a while). I don't think it'll be a performance issue (interrupt latency on an idle system), but in case someone does measurements feel free to report them.

arch/powerpc/kernel/cpu_setup_ppc970.S +16 -0
Index: linux-2.6/arch/powerpc/kernel/cpu_setup_ppc970.S
===================================================================
@@ -83,6 +83,22 @@ _GLOBAL(__setup_cpu_ppc970)
 	rldimi	r0,r11,52,8		/* set NAP and DPM */
 	li	r11,0
 	rldimi	r0,r11,32,31		/* clear EN_ATTN */
+	b	load_hids		/* Jump to shared code */
+
+
+_GLOBAL(__setup_cpu_ppc970MP)
+	/* Do nothing if not running in HV mode */
+	mfmsr	r0
+	rldicl.	r0,r0,4,63
+	beqlr
+
+	mfspr	r0,SPRN_HID0
+	li	r11,0x15		/* clear DOZE and SLEEP */
+	rldimi	r0,r11,52,6		/* set DEEPNAP, NAP and DPM */
+	li	r11,0
+	rldimi	r0,r11,32,31		/* clear EN_ATTN */
+
+load_hids:
 	mtspr	SPRN_HID0,r0
 	mfspr	r0,SPRN_HID0
 	mfspr	r0,SPRN_HID0
arch/powerpc/kernel/cputable.c +2 -1
Index: linux-2.6/arch/powerpc/kernel/cputable.c
===================================================================
@@ -41,6 +41,7 @@ extern void __setup_cpu_745x(unsigned lo
 #endif /* CONFIG_PPC32 */
 #ifdef CONFIG_PPC64
 extern void __setup_cpu_ppc970(unsigned long offset, struct cpu_spec* spec);
+extern void __setup_cpu_ppc970MP(unsigned long offset, struct cpu_spec* spec);
 extern void __restore_cpu_ppc970(void);
 #endif /* CONFIG_PPC64 */
 
@@ -221,7 +222,7 @@ struct cpu_spec	cpu_specs[] = {
 		.icache_bsize		= 128,
 		.dcache_bsize		= 128,
 		.num_pmcs		= 8,
-		.cpu_setup		= __setup_cpu_ppc970,
+		.cpu_setup		= __setup_cpu_ppc970MP,
 		.cpu_restore		= __restore_cpu_ppc970,
 		.oprofile_cpu_type	= "ppc64/970",
 		.oprofile_type		= PPC_OPROFILE_POWER4,
#1 lore raw
Quoted 26 lines · Olof Johansson

Hi,

Without this patch, on an idle system I get:

cpu-power-0:21.638
cpu-power-1:27.102
cpu-power-2:29.343
cpu-power-3:25.784
Total: 103.8W

With this patch:

cpu-power-0:11.730
cpu-power-1:17.185
cpu-power-2:18.547
cpu-power-3:17.528
Total: 65.0W

If I lower HZ to 100, I can get it as low as:

cpu-power-0:10.938
cpu-power-1:16.021
cpu-power-2:17.245
cpu-power-3:16.145
Total: 60.2W

Quoted 6 lines

Another (older) Quad G5 went from 54W to 39W at HZ=250.

Coming back out of Deep Nap takes 40-70 cycles longer than coming back from just Nap (which already takes quite a while). I don't think it'll be a performance issue (interrupt latency on an idle system), but in case someone does measurements feel free to report them.

Isn't deep nap supposed to only get entered from F/4 which we never use ?

Ben confused...

Quoted 57 lines

Index: linux-2.6/arch/powerpc/kernel/cpu_setup_ppc970.S
===================================================================

arch/powerpc/kernel/cpu_setup_ppc970.S +16 -0
@@ -83,6 +83,22 @@ _GLOBAL(__setup_cpu_ppc970)
 	rldimi	r0,r11,52,8		/* set NAP and DPM */
 	li	r11,0
 	rldimi	r0,r11,32,31		/* clear EN_ATTN */
+	b	load_hids		/* Jump to shared code */
+
+
+_GLOBAL(__setup_cpu_ppc970MP)
+	/* Do nothing if not running in HV mode */
+	mfmsr	r0
+	rldicl.	r0,r0,4,63
+	beqlr
+
+	mfspr	r0,SPRN_HID0
+	li	r11,0x15		/* clear DOZE and SLEEP */
+	rldimi	r0,r11,52,6		/* set DEEPNAP, NAP and DPM */
+	li	r11,0
+	rldimi	r0,r11,32,31		/* clear EN_ATTN */
+
+load_hids:
 	mtspr	SPRN_HID0,r0
 	mfspr	r0,SPRN_HID0
 	mfspr	r0,SPRN_HID0

Index: linux-2.6/arch/powerpc/kernel/cputable.c
===================================================================

arch/powerpc/kernel/cputable.c +2 -1
@@ -41,6 +41,7 @@ extern void __setup_cpu_745x(unsigned lo
 #endif /* CONFIG_PPC32 */
 #ifdef CONFIG_PPC64
 extern void __setup_cpu_ppc970(unsigned long offset, struct cpu_spec* spec);
+extern void __setup_cpu_ppc970MP(unsigned long offset, struct cpu_spec* spec);
 extern void __restore_cpu_ppc970(void);
 #endif /* CONFIG_PPC64 */
 
@@ -221,7 +222,7 @@ struct cpu_spec	cpu_specs[] = {
 		.icache_bsize		= 128,
 		.dcache_bsize		= 128,
 		.num_pmcs		= 8,
-		.cpu_setup		= __setup_cpu_ppc970,
+		.cpu_setup		= __setup_cpu_ppc970MP,
 		.cpu_restore		= __restore_cpu_ppc970,
 		.oprofile_cpu_type	= "ppc64/970",
 		.oprofile_type		= PPC_OPROFILE_POWER4,

_______________________________________________
Linuxppc-dev mailing list
[email protected]
https://ozlabs.org/mailman/listinfo/linuxppc-dev

#2 lore raw
Quoted 4 lines · On Thu, 05 Oct 2006 15:34:24 +1000 Benjamin Herrenschmidt…

Isn't deep nap supposed to only get entered from F/4 which we never use ?

Ben confused...

Nope.

The User Manual clearly talks about going in/out of deep nap from the different powertune frequencies, see for example section 6.4.2 and 9.4.2.

-Olof

Acked-by on v1 1/1
#3 lore raw
In reply to Olof Johansson
Quoted 28 lines · Olof Johansson

Hi,

Without this patch, on an idle system I get:

cpu-power-0:21.638
cpu-power-1:27.102
cpu-power-2:29.343
cpu-power-3:25.784
Total: 103.8W

With this patch:

cpu-power-0:11.730
cpu-power-1:17.185
cpu-power-2:18.547
cpu-power-3:17.528
Total: 65.0W

If I lower HZ to 100, I can get it as low as:

cpu-power-0:10.938
cpu-power-1:16.021
cpu-power-2:17.245
cpu-power-3:16.145
Total: 60.2W

Another (older) Quad G5 went from 54W to 39W at HZ=250.

Quoted 7 lines

Coming back out of Deep Nap takes 40-70 cycles longer than coming back from just Nap (which already takes quite a while). I don't think it'll be a performance issue (interrupt latency on an idle system), but in case someone does measurements feel free to report them.

I am running DEEPNAP on my Quad since quite some time and I did not see any problems.
It saves quite a bit of power (I think it was about 20W for me when I measured it)
That's really worth it.

Signature · 1 line
Greetings Michael.