Skip to contentSkip to conversation

drm/i915: Use usleep_range() in wait_for()

PATCH In review [email protected] 5 participants 7 messages Mar 20, 2015 – Mar 23, 2015

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915: Use usleep_range() in wait_for()

From: Ville Syrjälä <[email protected]>

msleep() can sleep for way too long, so switch wait_for() to use usleep_range() instead. Following a totally unscientific method I just picked the range as W-2W.

This cuts the i915 init time on my BSW to almost half:
- initcall i915_init+0x0/0xa8 [i915] returned 0 after 419977 usecs
+ initcall i915_init+0x0/0xa8 [i915] returned 0 after 238419 usecs

Note that I didn't perform any other benchmarks on this so far.

---
 drivers/gpu/drm/i915/intel_drv.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +2 -2
index 6f20f3a..d2a4de0 100644
@@ -56,8 +56,8 @@
 				ret__ = -ETIMEDOUT;			\
 			break;						\
 		}							\
-		if (W && drm_can_sleep())  {				\
-			msleep(W);					\
+		if ((W) && drm_can_sleep()) {				\
+			usleep_range((W)*1000, (W)*2000);		\
 		} else {						\
 			cpu_relax();					\
 		}							\
Signature · 6 lines
2.0.5

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 1/1
#1 lore raw
Quoted 24 lines · On 03/20/2015 12:28 PM, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

msleep() can sleep for way too long, so switch wait_for() to use usleep_range() instead. Following a totally unscientific method I just picked the range as W-2W.

This cuts the i915 init time on my BSW to almost half:
- initcall i915_init+0x0/0xa8 [i915] returned 0 after 419977 usecs
+ initcall i915_init+0x0/0xa8 [i915] returned 0 after 238419 usecs

Note that I didn't perform any other benchmarks on this so far.

---
 drivers/gpu/drm/i915/intel_drv.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +0 -0
index 6f20f3a..d2a4de0 100644

drivers/gpu/drm/i915/intel_drv.h line 63

Quoted 10 lines
drivers/gpu/drm/i915/intel_drv.h +2 -2
@@ -56,8 +56,8 @@
 				ret__ = -ETIMEDOUT;			\
 			break;						\
 		}							\
-		if (W && drm_can_sleep())  {				\
-			msleep(W);					\
+		if ((W) && drm_can_sleep()) {				\
+			usleep_range((W)*1000, (W)*2000);		\
 		} else {						\
 			cpu_relax();					\
 		}							\

Nice improvement! I guess it's just because usleep_range() uses an hrtimeout, but since we have the added slop of the range it may actually be an improvement, power-wise too.

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#2 lore raw
Quoted 40 lines · Jesse Barnes
Quoted 24 lines · On 03/20/2015 12:28 PM, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

msleep() can sleep for way too long, so switch wait_for() to use usleep_range() instead. Following a totally unscientific method I just picked the range as W-2W.

This cuts the i915 init time on my BSW to almost half:
- initcall i915_init+0x0/0xa8 [i915] returned 0 after 419977 usecs
+ initcall i915_init+0x0/0xa8 [i915] returned 0 after 238419 usecs

Note that I didn't perform any other benchmarks on this so far.

---
 drivers/gpu/drm/i915/intel_drv.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_drv.h +0 -0
index 6f20f3a..d2a4de0 100644

drivers/gpu/drm/i915/intel_drv.h line 63

Quoted 10 lines
drivers/gpu/drm/i915/intel_drv.h +2 -2
@@ -56,8 +56,8 @@
 				ret__ = -ETIMEDOUT;			\
 			break;						\
 		}							\
-		if (W && drm_can_sleep())  {				\
-			msleep(W);					\
+		if ((W) && drm_can_sleep()) {				\
+			usleep_range((W)*1000, (W)*2000);		\
 		} else {						\
 			cpu_relax();					\
 		}							\

Nice improvement! I guess it's just because usleep_range() uses an hrtimeout, but since we have the added slop of the range it may actually be an improvement, power-wise too.

Quoted 1 line

Nice indeed. Queued for -next, thanks for the patch. -Daniel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 1/1
#3 lore raw
In reply to [email protected]
Quoted 6 lines · On Fri, Mar 20, 2015 at 09:28:08PM +0200, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

msleep() can sleep for way too long, so switch wait_for() to use usleep_range() instead. Following a totally unscientific method I just picked the range as W-2W.

Quoted 8 lines

This cuts the i915 init time on my BSW to almost half:
- initcall i915_init+0x0/0xa8 [i915] returned 0 after 419977 usecs
+ initcall i915_init+0x0/0xa8 [i915] returned 0 after 238419 usecs

Note that I didn't perform any other benchmarks on this so far.

Hmm, I think we can improve further with a more variable sleep. The maximum we pass to wait_for() is usually plucked from somewhere in the spec (or is just a safety factor). Either way, it is a good guide as to how to actually sleep for - if say we try to only sample 1000 times up to the maximum:

if (do_sleep && drm_can_sleep()) {
	usleep_range((MS), 10*(MS));
}

So whilst you have a situation where we clearly sleep too long between sampling (a register), it would be beneficial to start adding some debug infrastructure. Or even better if usleep_range already does it for us... -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw
Quoted 24 lines · Chris Wilson
Quoted 6 lines · On Fri, Mar 20, 2015 at 09:28:08PM +0200, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

msleep() can sleep for way too long, so switch wait_for() to use usleep_range() instead. Following a totally unscientific method I just picked the range as W-2W.

Quoted 8 lines

This cuts the i915 init time on my BSW to almost half:
- initcall i915_init+0x0/0xa8 [i915] returned 0 after 419977 usecs
+ initcall i915_init+0x0/0xa8 [i915] returned 0 after 238419 usecs

Note that I didn't perform any other benchmarks on this so far.

Hmm, I think we can improve further with a more variable sleep. The maximum we pass to wait_for() is usually plucked from somewhere in the spec (or is just a safety factor). Either way, it is a good guide as to how to actually sleep for - if say we try to only sample 1000 times up to the maximum:

Quoted 7 lines
if (do_sleep && drm_can_sleep()) {
	usleep_range((MS), 10*(MS));
}

So whilst you have a situation where we clearly sleep too long between sampling (a register), it would be beneficial to start adding some debug infrastructure. Or even better if usleep_range already does it for us...

Yeah, it would be nice to have some more information about how long we sleep typically in each case.

We could perhaps then even micro optimize the 'set knob -> poll for knob to become active' pattern by doing a preemptive sleep between the set and poll steps in the hopes that we don't have to check the condition more than once. Not sure that would be worth the effort though.

Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw
Quoted 33 lines · Ville Syrjälä
Quoted 24 lines · Chris Wilson
Quoted 6 lines · On Fri, Mar 20, 2015 at 09:28:08PM +0200, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

msleep() can sleep for way too long, so switch wait_for() to use usleep_range() instead. Following a totally unscientific method I just picked the range as W-2W.

Quoted 8 lines

This cuts the i915 init time on my BSW to almost half:
- initcall i915_init+0x0/0xa8 [i915] returned 0 after 419977 usecs
+ initcall i915_init+0x0/0xa8 [i915] returned 0 after 238419 usecs

Note that I didn't perform any other benchmarks on this so far.

Hmm, I think we can improve further with a more variable sleep. The maximum we pass to wait_for() is usually plucked from somewhere in the spec (or is just a safety factor). Either way, it is a good guide as to how to actually sleep for - if say we try to only sample 1000 times up to the maximum:

Quoted 7 lines
if (do_sleep && drm_can_sleep()) {
	usleep_range((MS), 10*(MS));
}

So whilst you have a situation where we clearly sleep too long between sampling (a register), it would be beneficial to start adding some debug infrastructure. Or even better if usleep_range already does it for us...

Quoted 7 lines

Yeah, it would be nice to have some more information about how long we sleep typically in each case.

We could perhaps then even micro optimize the 'set knob -> poll for knob to become active' pattern by doing a preemptive sleep between the set and poll steps in the hopes that we don't have to check the condition more than once. Not sure that would be worth the effort though.

Something like the below might be worth a quick shot. Maybe with 3 instead of 2, but then we need to make sure we don't drop down to 0 ever. -Daniel

drivers/gpu/drm/i915/intel_drv.h +2 -1
index d2a4de0e4f4a..b796e4c47da5 100644
@@ -58,6 +58,7 @@
 		}							\
 		if ((W) && drm_can_sleep()) {				\
 			usleep_range((W)*1000, (W)*2000);		\
+			W = ROUND_UP(W, 2);				\
 		} else {						\
 			cpu_relax();					\
 		}							\
@@ -65,7 +66,7 @@
 	ret__;								\
 })
 
-#define wait_for(COND, MS) _wait_for(COND, MS, 1)
+#define wait_for(COND, MS) _wait_for(COND, MS, ROUND_UP(COND, 2))
 #define wait_for_atomic(COND, MS) _wait_for(COND, MS, 0)
 #define wait_for_atomic_us(COND, US) _wait_for((COND), \
 					       DIV_ROUND_UP((US), 1000), 0)
Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Tested-by on v1 1/1
#6 lore raw
In reply to [email protected]

Task id: 6020 -------------------------------------Summary-------------------------------------
Platform Delta drm-intel-nightly Series Applied
PNV -2 274/274 272/274

ILK                                  303/303              303/303
SNB                                  303/303              303/303
IVB                 -2              342/342              340/342
BYT                                  287/287              287/287
HSW                                  362/362              362/362
BDW                                  308/308              308/308
-------------------------------------Detailed-------------------------------------
Platform  Test                                drm-intel-nightly          Series Applied
 PNV  igt_gem_userptr_blits_minor-sync-interruptible      DMESG_WARN(1)PASS(3)      DMESG_WARN(1)PASS(1)
*PNV  igt_gen3_render_linear_blits      FAIL(1)PASS(6)      CRASH(1)PASS(1)
 IVB  igt_gem_pwrite_pread_snooped-copy-performance      DMESG_WARN(2)PASS(3)      DMESG_WARN(1)PASS(1)
*IVB  igt_gem_storedw_batches_loop_secure-dispatch      PASS(4)      DMESG_WARN(1)PASS(1)

Note: You need to pay more attention to line start with '*'
_______________________________________________
Intel-gfx mailing list [email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx