powerpc: Enable DEEPNAP power savings mode on 970MP
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
@@ -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
@@ -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,
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: 103.8WWith this patch:
cpu-power-0:11.730
cpu-power-1:17.185
cpu-power-2:18.547
cpu-power-3:17.528
Total: 65.0WIf 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
Signed-off-by Olof Johansson <[email protected]>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_HID0Index:
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
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
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: 103.8WWith this patch:
cpu-power-0:11.730
cpu-power-1:17.185
cpu-power-2:18.547
cpu-power-3:17.528
Total: 65.0WIf 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: 60.2WAnother (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.
Signed-off-by Olof Johansson <[email protected]>
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.