Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 24/11/23 11:21, Boris Brezillon ha scritto:
On Fri, 24 Nov 2023 11:12:57 +0100
AngeloGioacchino Del Regno
wrote:
Il 24/11/23 10:17, AngeloGioacchino Del Regno ha scritto:
Il 23/11/23 16:40, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 16:14:12 +0100
AngeloGioacchino Del Regno
wrote:
Il 23/11/23 14:51, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 14:24:57 +0100
AngeloGioacchino Del Regno
wrote:
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
If we just need a quick fix to avoid PWRTRANS interrupts from kicking
in when we power-off the cores, I think we'd be better off dropping
GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
at [re]initialization time, and then have a separate series that fixes
the problem more generically.
But that didn't work:
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
I meant, your 'ignore-core_mask' fix + the
'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
So,
https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
+
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
...while this "full" solution worked:
https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
...so this *is* a "quick fix" already... :-)
It's a half-baked solution for the missing irq-synchronization-on-suspend
issue IMHO. I understand why you want it all in one patch that can serve
as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
even simpler diff (see below), and then fully address the
irq-synhronization-on-suspend issue in a follow-up patchset.
--->8---
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 09f5e1563ebd..6e2d7650cc2b 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
}
gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
- gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
We probably want a comment here:
/* Only enable the interrupts we care about. */
+ gpu_write(pfdev, GPU_INT_MASK,
+ GPU_IRQ_MASK_ERROR |
+ GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
+ GPU_IRQ_CLEAN_CACHES_COMPLETED);
...but if we do that, the next patch(es) will contain a partial revert of this
commit, putting back this to gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL)...
Why should we revert it? We're not processing the PWRTRANS interrupts
in the interrupt handler, those should never have been enabled in the
first place. The only reason we'd want to revert that change is if we
decide to do have interrupt-based waits in the poweron/off
implementation, which, as far as I'm aware, is not something we intend
to do any time soon.
You're right, yes. Okay, I'll push the new code soon.
Cheers!
Update: I was running some (rather fast) tests here because I ... felt like
playing
with it, basically :-)
So, I had an issue with MediaTek platforms being unable to cut power to the GPU
or
disable clocks aggressively... and after trying "this and that" I couldn't get
it
working (in runtime suspend).
Long story short - after implementing `panfrost_{job,mmu,gpu}_suspend_irq()`
(only
gpu irq, as you said, is a half solution), I can not only turn off clocks, but
even
turn off GPU power supplies entirely, bringing the power consumption of the GPU
itself during *runtime* suspend to ... zero.
Very nice!
The result of this test makes me truly happy, even though complete powercut
during
runtime suspend may not be feasible for other reasons (takes ~20ns on AVG,
MIN ~16ns, but the MAX is ~475000ns - and beware that I haven't run that for
long, I'd suspect to get up to 1-1.5ms as max time, so that's a big no).
Do you know what's taking so long? I'm disabling clks + the main power
domain in panthor (I leave the regulators enabled), but I didn't get to
measure the time it takes to enter/exit suspend. I might have to do
what you did in panfrost and have different paths for system and RPM
suspend.
That's SoC dependant... there's only one way to get runtime suspend right in
terms
of timing, and that is to select what to do there on a per-SoC basis: this is
why
some of them will take lots of time to turn off (or on!) clocks, because clock
controllers are not all equal: this is not only in relation to different vendors
(as in, rockchip vs nxp vs mediatek vs qcom vs...) but also for different parts
from the same vendor (as in, MSM8953 uses different clock controllers compared
to
SM8350 and MT6795 different compared to MT6985 and MT8195).
Some of
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
On Fri, 24 Nov 2023 11:12:57 +0100
AngeloGioacchino Del Regno
wrote:
> Il 24/11/23 10:17, AngeloGioacchino Del Regno ha scritto:
> > Il 23/11/23 16:40, Boris Brezillon ha scritto:
> >> On Thu, 23 Nov 2023 16:14:12 +0100
> >> AngeloGioacchino Del Regno
> >> wrote:
> >>
> >>> Il 23/11/23 14:51, Boris Brezillon ha scritto:
> On Thu, 23 Nov 2023 14:24:57 +0100
> AngeloGioacchino Del Regno
> wrote:
> >>>
> >>> So, while I agree that it'd be slightly more readable as a diff if
> >>> those
> >>> were two different commits I do have reasons against splitting.
> >>
> >> If we just need a quick fix to avoid PWRTRANS interrupts from kicking
> >> in when we power-off the cores, I think we'd be better off dropping
> >> GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
> >> at [re]initialization time, and then have a separate series that fixes
> >> the problem more generically.
> >
> > But that didn't work:
> > https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
> >
>
> I meant, your 'ignore-core_mask' fix + the
> 'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
>
> So,
>
> https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
> +
> https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
>
> >
> >
> > ...while this "full" solution worked:
> > https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
> >
> > https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
> >
> >
> > ...so this *is* a "quick fix" already... :-)
>
> It's a half-baked solution for the missing irq-synchronization-on-suspend
> issue IMHO. I understand why you want it all in one patch that can serve
> as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
> panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
> even simpler diff (see below), and then fully address the
> irq-synhronization-on-suspend issue in a follow-up patchset.
> --->8---
> diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> index 09f5e1563ebd..6e2d7650cc2b 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> @@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device
> *pfdev)
> }
> gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
> - gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
> >>
> >> We probably want a comment here:
> >>
> >> /* Only enable the interrupts we care about. */
> >>
> + gpu_write(pfdev, GPU_INT_MASK,
> + GPU_IRQ_MASK_ERROR |
> + GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
> + GPU_IRQ_CLEAN_CACHES_COMPLETED);
> >>>
> >>> ...but if we do that, the next patch(es) will contain a partial revert of
> >>> this
> >>> commit, putting back this to gpu_write(pfdev, GPU_INT_MASK,
> >>> GPU_IRQ_MASK_ALL)...
> >>
> >> Why should we revert it? We're not processing the PWRTRANS interrupts
> >> in the interrupt handler, those should never have been enabled in the
> >> first place. The only reason we'd want to revert that change is if we
> >> decide to do have interrupt-based waits in the poweron/off
> >> implementation, which, as far as I'm aware, is not something we intend
> >> to do any time soon.
> >>
> >
> > You're right, yes. Okay, I'll push the new code soon.
> >
> > Cheers!
> >
>
> Update: I was running some (rather fast) tests here because I ... felt like
> playing
> with it, basically :-)
>
> So, I had an issue with MediaTek platforms being unable to cut power to the
> GPU or
> disable clocks aggressively... and after trying "this and that" I couldn't
> get it
> working (in runtime suspend).
>
> Long story short - after implementing `panfrost_{job,mmu,gpu}_suspend_irq()`
> (only
> gpu irq, as you said, is a half solution), I can not only turn off clocks,
> but even
> turn off GPU power supplies entirely, bringing the power consumption of the
> GPU
> itself during *runtime* suspend to ... zero.
Very nice!
>
> The result of this test makes me truly happy, even though complete powercut
> during
> runtime suspend may not be feasible for other reasons (takes ~20ns on AVG,
> MIN ~16ns, but the MAX is ~475000ns - and beware that I haven't run that
> for
> long, I'd suspect to get up to 1-1.5ms as max time, so that's a big no).
Do you know what's taking so long? I'm disabling clks + the main power
domain in panthor (I leave the regulators enabled), but I didn't get to
measure the time it takes to enter/exit suspend. I might have to do
what you did in panfrost and have different paths for
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 24/11/23 10:17, AngeloGioacchino Del Regno ha scritto:
Il 23/11/23 16:40, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 16:14:12 +0100
AngeloGioacchino Del Regno
wrote:
Il 23/11/23 14:51, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 14:24:57 +0100
AngeloGioacchino Del Regno
wrote:
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
If we just need a quick fix to avoid PWRTRANS interrupts from kicking
in when we power-off the cores, I think we'd be better off dropping
GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
at [re]initialization time, and then have a separate series that fixes
the problem more generically.
But that didn't work:
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
I meant, your 'ignore-core_mask' fix + the
'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
So,
https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
+
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
...while this "full" solution worked:
https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
...so this *is* a "quick fix" already... :-)
It's a half-baked solution for the missing irq-synchronization-on-suspend
issue IMHO. I understand why you want it all in one patch that can serve
as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
even simpler diff (see below), and then fully address the
irq-synhronization-on-suspend issue in a follow-up patchset.
--->8---
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 09f5e1563ebd..6e2d7650cc2b 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
}
gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
- gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
We probably want a comment here:
/* Only enable the interrupts we care about. */
+ gpu_write(pfdev, GPU_INT_MASK,
+ GPU_IRQ_MASK_ERROR |
+ GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
+ GPU_IRQ_CLEAN_CACHES_COMPLETED);
...but if we do that, the next patch(es) will contain a partial revert of this
commit, putting back this to gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL)...
Why should we revert it? We're not processing the PWRTRANS interrupts
in the interrupt handler, those should never have been enabled in the
first place. The only reason we'd want to revert that change is if we
decide to do have interrupt-based waits in the poweron/off
implementation, which, as far as I'm aware, is not something we intend
to do any time soon.
You're right, yes. Okay, I'll push the new code soon.
Cheers!
Update: I was running some (rather fast) tests here because I ... felt like
playing
with it, basically :-)
So, I had an issue with MediaTek platforms being unable to cut power to the GPU
or
disable clocks aggressively... and after trying "this and that" I couldn't get
it
working (in runtime suspend).
Long story short - after implementing `panfrost_{job,mmu,gpu}_suspend_irq()`
(only
gpu irq, as you said, is a half solution), I can not only turn off clocks, but
even
turn off GPU power supplies entirely, bringing the power consumption of the GPU
itself during *runtime* suspend to ... zero.
The result of this test makes me truly happy, even though complete powercut
during
runtime suspend may not be feasible for other reasons (takes ~20ns on AVG,
MIN ~16ns, but the MAX is ~475000ns - and beware that I haven't run that for
long, I'd suspect to get up to 1-1.5ms as max time, so that's a big no).
This means that I will take a day or two and I'll push both the "simple" fix for
the Really-power-off and also some more commits to add the full irq sync.
Cheers!
Angelo
I'm not sure that it's worth changing this like that, then changing it back
right
after :-\
Anyway, if anyone else agrees with doing it and then partially revert, I have no
issues going with this one instead; what I care about ultimately is resolving
the
regression ASAP :-)
Cheers,
Angelo
/*
* All in-flight jobs should have released their cycle
@@ -425,11 +428,10 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
void panfrost_gpu_power_off(struct panfrost_device *pfdev)
{
- u64 core_mask = panfrost_get_core_mask(pfdev);
int ret;
u32 val;
- gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present &
core_mask);
+ gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
ret =
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 23/11/23 16:40, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 16:14:12 +0100
AngeloGioacchino Del Regno
wrote:
Il 23/11/23 14:51, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 14:24:57 +0100
AngeloGioacchino Del Regno
wrote:
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
If we just need a quick fix to avoid PWRTRANS interrupts from kicking
in when we power-off the cores, I think we'd be better off dropping
GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
at [re]initialization time, and then have a separate series that fixes
the problem more generically.
But that didn't work:
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
I meant, your 'ignore-core_mask' fix + the
'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
So,
https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
+
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
...while this "full" solution worked:
https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
...so this *is* a "quick fix" already... :-)
It's a half-baked solution for the missing irq-synchronization-on-suspend
issue IMHO. I understand why you want it all in one patch that can serve
as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
even simpler diff (see below), and then fully address the
irq-synhronization-on-suspend issue in a follow-up patchset.
--->8---
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 09f5e1563ebd..6e2d7650cc2b 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
}
gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
- gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
We probably want a comment here:
/* Only enable the interrupts we care about. */
+ gpu_write(pfdev, GPU_INT_MASK,
+ GPU_IRQ_MASK_ERROR |
+ GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
+ GPU_IRQ_CLEAN_CACHES_COMPLETED);
...but if we do that, the next patch(es) will contain a partial revert of this
commit, putting back this to gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL)...
Why should we revert it? We're not processing the PWRTRANS interrupts
in the interrupt handler, those should never have been enabled in the
first place. The only reason we'd want to revert that change is if we
decide to do have interrupt-based waits in the poweron/off
implementation, which, as far as I'm aware, is not something we intend
to do any time soon.
You're right, yes. Okay, I'll push the new code soon.
Cheers!
I'm not sure that it's worth changing this like that, then changing it back
right
after :-\
Anyway, if anyone else agrees with doing it and then partially revert, I have no
issues going with this one instead; what I care about ultimately is resolving
the
regression ASAP :-)
Cheers,
Angelo
/*
* All in-flight jobs should have released their cycle
@@ -425,11 +428,10 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
void panfrost_gpu_power_off(struct panfrost_device *pfdev)
{
- u64 core_mask = panfrost_get_core_mask(pfdev);
int ret;
u32 val;
- gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present & core_mask);
+ gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO,
val, !val, 1, 1000);
if (ret)
@@ -441,7 +443,7 @@ void panfrost_gpu_power_off(struct panfrost_device *pfdev)
if (ret)
dev_err(pfdev->dev, "tiler power transition timeout");
- gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present & core_mask);
+ gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present);
ret = readl_poll_timeout(pfdev->iomem + L2_PWRTRANS_LO,
val, !val, 0, 1000);
if (ret)
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
On Thu, 23 Nov 2023 16:14:12 +0100
AngeloGioacchino Del Regno
wrote:
> Il 23/11/23 14:51, Boris Brezillon ha scritto:
> > On Thu, 23 Nov 2023 14:24:57 +0100
> > AngeloGioacchino Del Regno
> > wrote:
> >
>
> So, while I agree that it'd be slightly more readable as a diff if those
> were two different commits I do have reasons against splitting.
> >>>
> >>> If we just need a quick fix to avoid PWRTRANS interrupts from kicking
> >>> in when we power-off the cores, I think we'd be better off dropping
> >>> GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
> >>> at [re]initialization time, and then have a separate series that fixes
> >>> the problem more generically.
> >>>
> >>
> >> But that didn't work:
> >> https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
> >>
> >
> > I meant, your 'ignore-core_mask' fix + the
> > 'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
> >
> > So,
> >
> > https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
> > +
> > https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
> >
> >>
> >>
> >> ...while this "full" solution worked:
> >> https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
> >>
> >> https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
> >>
> >>
> >> ...so this *is* a "quick fix" already... :-)
> >
> > It's a half-baked solution for the missing irq-synchronization-on-suspend
> > issue IMHO. I understand why you want it all in one patch that can serve
> > as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
> > panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
> > even simpler diff (see below), and then fully address the
> > irq-synhronization-on-suspend issue in a follow-up patchset.
> >
> > --->8---
> > diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> > b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> > index 09f5e1563ebd..6e2d7650cc2b 100644
> > --- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> > +++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> > @@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device
> > *pfdev)
> > }
> >
> > gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
> > - gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
We probably want a comment here:
/* Only enable the interrupts we care about. */
> > + gpu_write(pfdev, GPU_INT_MASK,
> > + GPU_IRQ_MASK_ERROR |
> > + GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
> > + GPU_IRQ_CLEAN_CACHES_COMPLETED);
> >
>
> ...but if we do that, the next patch(es) will contain a partial revert of this
> commit, putting back this to gpu_write(pfdev, GPU_INT_MASK,
> GPU_IRQ_MASK_ALL)...
Why should we revert it? We're not processing the PWRTRANS interrupts
in the interrupt handler, those should never have been enabled in the
first place. The only reason we'd want to revert that change is if we
decide to do have interrupt-based waits in the poweron/off
implementation, which, as far as I'm aware, is not something we intend
to do any time soon.
>
> I'm not sure that it's worth changing this like that, then changing it back
> right
> after :-\
>
> Anyway, if anyone else agrees with doing it and then partially revert, I have
> no
> issues going with this one instead; what I care about ultimately is resolving
> the
> regression ASAP :-)
>
> Cheers,
> Angelo
>
> > /*
> > * All in-flight jobs should have released their cycle
> > @@ -425,11 +428,10 @@ void panfrost_gpu_power_on(struct panfrost_device
> > *pfdev)
> >
> > void panfrost_gpu_power_off(struct panfrost_device *pfdev)
> > {
> > - u64 core_mask = panfrost_get_core_mask(pfdev);
> > int ret;
> > u32 val;
> >
> > - gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present &
> > core_mask);
> > + gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
> > ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO,
> > val, !val, 1, 1000);
> > if (ret)
> > @@ -441,7 +443,7 @@ void panfrost_gpu_power_off(struct panfrost_device
> > *pfdev)
> > if (ret)
> > dev_err(pfdev->dev, "tiler power transition timeout");
> >
> > - gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present &
> > core_mask);
> > + gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present);
> > ret = readl_poll_timeout(pfdev->iomem + L2_PWRTRANS_LO,
> > val, !val, 0, 1000);
> > if (ret)
> >
> >
>
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 23/11/23 14:51, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 14:24:57 +0100
AngeloGioacchino Del Regno
wrote:
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
If we just need a quick fix to avoid PWRTRANS interrupts from kicking
in when we power-off the cores, I think we'd be better off dropping
GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
at [re]initialization time, and then have a separate series that fixes
the problem more generically.
But that didn't work:
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
I meant, your 'ignore-core_mask' fix + the
'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
So,
https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
+
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
...while this "full" solution worked:
https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
...so this *is* a "quick fix" already... :-)
It's a half-baked solution for the missing irq-synchronization-on-suspend
issue IMHO. I understand why you want it all in one patch that can serve
as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
even simpler diff (see below), and then fully address the
irq-synhronization-on-suspend issue in a follow-up patchset.
--->8---
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 09f5e1563ebd..6e2d7650cc2b 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
}
gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
- gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
+ gpu_write(pfdev, GPU_INT_MASK,
+ GPU_IRQ_MASK_ERROR |
+ GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
+ GPU_IRQ_CLEAN_CACHES_COMPLETED);
...but if we do that, the next patch(es) will contain a partial revert of this
commit, putting back this to gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL)...
I'm not sure that it's worth changing this like that, then changing it back
right
after :-\
Anyway, if anyone else agrees with doing it and then partially revert, I have no
issues going with this one instead; what I care about ultimately is resolving
the
regression ASAP :-)
Cheers,
Angelo
/*
* All in-flight jobs should have released their cycle
@@ -425,11 +428,10 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
void panfrost_gpu_power_off(struct panfrost_device *pfdev)
{
- u64 core_mask = panfrost_get_core_mask(pfdev);
int ret;
u32 val;
- gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present & core_mask);
+ gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO,
val, !val, 1, 1000);
if (ret)
@@ -441,7 +443,7 @@ void panfrost_gpu_power_off(struct panfrost_device *pfdev)
if (ret)
dev_err(pfdev->dev, "tiler power transition timeout");
- gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present & core_mask);
+ gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present);
ret = readl_poll_timeout(pfdev->iomem + L2_PWRTRANS_LO,
val, !val, 0, 1000);
if (ret)
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
On Thu, 23 Nov 2023 14:24:57 +0100
AngeloGioacchino Del Regno
wrote:
> >>
> >> So, while I agree that it'd be slightly more readable as a diff if those
> >> were two different commits I do have reasons against splitting.
> >
> > If we just need a quick fix to avoid PWRTRANS interrupts from kicking
> > in when we power-off the cores, I think we'd be better off dropping
> > GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
> > at [re]initialization time, and then have a separate series that fixes
> > the problem more generically.
> >
>
> But that didn't work:
> https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
I meant, your 'ignore-core_mask' fix + the
'drop GPU_IRQ_POWER_CHANGED[_ALL] in GPU_INT_MASK' one.
So,
https://lore.kernel.org/all/4c73f67e-174c-497e-85a5-cb053ce65...@collabora.com/
+
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
>
>
> ...while this "full" solution worked:
> https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
>
> https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
>
>
> ...so this *is* a "quick fix" already... :-)
It's a half-baked solution for the missing irq-synchronization-on-suspend
issue IMHO. I understand why you want it all in one patch that can serve
as a fix for 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()"), which is why I'm suggesting to go for an
even simpler diff (see below), and then fully address the
irq-synhronization-on-suspend issue in a follow-up patchset.
--->8---
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 09f5e1563ebd..6e2d7650cc2b 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -78,7 +78,10 @@ int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
}
gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
- gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL);
+ gpu_write(pfdev, GPU_INT_MASK,
+ GPU_IRQ_MASK_ERROR |
+ GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
+ GPU_IRQ_CLEAN_CACHES_COMPLETED);
/*
* All in-flight jobs should have released their cycle
@@ -425,11 +428,10 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
void panfrost_gpu_power_off(struct panfrost_device *pfdev)
{
- u64 core_mask = panfrost_get_core_mask(pfdev);
int ret;
u32 val;
- gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present &
core_mask);
+ gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO,
val, !val, 1, 1000);
if (ret)
@@ -441,7 +443,7 @@ void panfrost_gpu_power_off(struct panfrost_device *pfdev)
if (ret)
dev_err(pfdev->dev, "tiler power transition timeout");
- gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present & core_mask);
+ gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present);
ret = readl_poll_timeout(pfdev->iomem + L2_PWRTRANS_LO,
val, !val, 0, 1000);
if (ret)
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 23/11/23 13:59, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 12:15:01 +0100
AngeloGioacchino Del Regno
wrote:
Il 23/11/23 11:35, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 10:53:20 +0100
AngeloGioacchino Del Regno
wrote:
Some SoCs may be equipped with a GPU containing two core groups
and this is exactly the case of Samsung's Exynos 5422 featuring
an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost
is partial, as this driver currently supports using only one
core group and that's reflected on all parts of it, including
the power on (and power off, previously to this patch) function.
The issue with this is that even though executing the soft reset
operation should power off all cores unconditionally, on at least
one platform we're seeing a crash that seems to be happening due
to an interrupt firing which may be because we are calling power
transition only on the first core group, leaving the second one
unchanged, or because ISR execution was pending before entering
the panfrost_gpu_power_off() function and executed after powering
off the GPU cores, or all of the above.
Finally, solve this by changing the power off flow to
1. Mask and clear all interrupts: we don't need nor want any, as
we are polling PWRTRANS anyway;
2. Call synchronize_irq() after that to make sure that any pending
ISR is executed before powering off the GPU Shaders/Tilers/L2
hence avoiding unpowered registers R/W; and
3. Ignore the core_mask and ask the GPU to poweroff both core groups
Could we split that in two patches? 1+2 in one patch, and 3 in another.
These are two orthogonal fixes IMO.
My initial idea was exactly that, but I opted for one patch doing 'em all
because a "full fix" comprises all of 1+2+3: the third one without the
first two and vice-versa may not fully resolve the issue that was seen
on the HC1 board.
Guess it depends how you see it. I'd argue that these are 2 orthogonal
bugs, and the suspend fix might be worth backporting to older versions.
Yes, but older versions are not affected by this regression because the GPU
was never turned off for real... so this commit is really fixing just the
issues that came out *because* of the "Really power off" commit, which is
not worth backporting as we already saw one regression with that and doing
such thing would be mostly dangerous.
If your suggestion was to backport 1+2 instead, I disagree: there are no
suspend instabilities in older kernel versions, so we'd be fixing something
that anyway always worked fine for years.
This is to say: the "Really power off" code should be treated like a commit
that implements a new feature, rather than fixing an old one, *only* because
there is a possibility to create regressions somehow (again, we already had
one).
At least IMO...
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
If we just need a quick fix to avoid PWRTRANS interrupts from kicking
in when we power-off the cores, I think we'd be better off dropping
GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
at [re]initialization time, and then have a separate series that fixes
the problem more generically.
But that didn't work:
https://lore.kernel.org/all/d95259b8-10cf-4ded-866c-47cbd2a44...@linaro.org/
...while this "full" solution worked:
https://lore.kernel.org/all/39e9514b-087c-42eb-8d0e-f75dc620e...@linaro.org/
https://lore.kernel.org/all/5b24cc73-23aa-4837-abb9-b6d138b46...@linaro.org/
...so this *is* a "quick fix" already... :-)
+ gpu_write(pfdev, GPU_INT_MASK, 0);
+ gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
+
+ /*
+* Make sure that we don't have pending ISRs, otherwise we'll be
+* reading and/or writing registers while the GPU is powered off
+*/
+ synchronize_irq(pfdev->irq);
Could we move that to a panfrost_gpu_suspend_irq() helper? I'm also not
sure making it part of panfrost_gpu_power_off() is a good idea. I'd
rather have this panfrost_gpu_suspend_irq() helper called from
panfrost_device_[runtime_]suspend(), along with
panfrost_{mmu,job}_suspend_irq().
Okay I will move that to a helper, but I still want to clarify:
- For JOB, we're checking if panfrost_job_is_idle() before trying
to runtime_suspend() (hence before trying to power off cores),
so implicitly no interrupt can fire I guess? Though there could
still be a pending ISR there too mmh. Brain ticking :-)
There's indeed no reason to see job interrupts if we're asked to enter
suspend, but it's mostly a matter of safety/correctness. If, as
expected, there's no pending interrupt, the write(_INT_MASK) +
synchronize_irq() should be relatively cheap.
- For MMU, we're not checking anything, but I guess that if there
is no job, the mmu can't be doing anything at all?
...but then you also gave me the doubt about that one as well.
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
On Thu, 23 Nov 2023 12:15:01 +0100
AngeloGioacchino Del Regno
wrote:
> Il 23/11/23 11:35, Boris Brezillon ha scritto:
> > On Thu, 23 Nov 2023 10:53:20 +0100
> > AngeloGioacchino Del Regno
> > wrote:
> >
> >> Some SoCs may be equipped with a GPU containing two core groups
> >> and this is exactly the case of Samsung's Exynos 5422 featuring
> >> an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost
> >> is partial, as this driver currently supports using only one
> >> core group and that's reflected on all parts of it, including
> >> the power on (and power off, previously to this patch) function.
> >>
> >> The issue with this is that even though executing the soft reset
> >> operation should power off all cores unconditionally, on at least
> >> one platform we're seeing a crash that seems to be happening due
> >> to an interrupt firing which may be because we are calling power
> >> transition only on the first core group, leaving the second one
> >> unchanged, or because ISR execution was pending before entering
> >> the panfrost_gpu_power_off() function and executed after powering
> >> off the GPU cores, or all of the above.
> >>
> >> Finally, solve this by changing the power off flow to
> >> 1. Mask and clear all interrupts: we don't need nor want any, as
> >> we are polling PWRTRANS anyway;
> >> 2. Call synchronize_irq() after that to make sure that any pending
> >> ISR is executed before powering off the GPU Shaders/Tilers/L2
> >> hence avoiding unpowered registers R/W; and
> >> 3. Ignore the core_mask and ask the GPU to poweroff both core groups
> >
> > Could we split that in two patches? 1+2 in one patch, and 3 in another.
> > These are two orthogonal fixes IMO.
> >
>
> My initial idea was exactly that, but I opted for one patch doing 'em all
> because a "full fix" comprises all of 1+2+3: the third one without the
> first two and vice-versa may not fully resolve the issue that was seen
> on the HC1 board.
Guess it depends how you see it. I'd argue that these are 2 orthogonal
bugs, and the suspend fix might be worth backporting to older versions.
>
> So, while I agree that it'd be slightly more readable as a diff if those
> were two different commits I do have reasons against splitting.
If we just need a quick fix to avoid PWRTRANS interrupts from kicking
in when we power-off the cores, I think we'd be better off dropping
GPU_IRQ_POWER_CHANGED[_ALL] from the value we write to GPU_INT_MASK
at [re]initialization time, and then have a separate series that fixes
the problem more generically.
> >> + gpu_write(pfdev, GPU_INT_MASK, 0);
> >> + gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
> >> +
> >> + /*
> >> + * Make sure that we don't have pending ISRs, otherwise we'll be
> >> + * reading and/or writing registers while the GPU is powered off
> >> + */
> >> + synchronize_irq(pfdev->irq);
> >
> > Could we move that to a panfrost_gpu_suspend_irq() helper? I'm also not
> > sure making it part of panfrost_gpu_power_off() is a good idea. I'd
> > rather have this panfrost_gpu_suspend_irq() helper called from
> > panfrost_device_[runtime_]suspend(), along with
> > panfrost_{mmu,job}_suspend_irq().
> >
>
> Okay I will move that to a helper, but I still want to clarify:
> - For JOB, we're checking if panfrost_job_is_idle() before trying
> to runtime_suspend() (hence before trying to power off cores),
> so implicitly no interrupt can fire I guess? Though there could
> still be a pending ISR there too mmh. Brain ticking :-)
There's indeed no reason to see job interrupts if we're asked to enter
suspend, but it's mostly a matter of safety/correctness. If, as
expected, there's no pending interrupt, the write(_INT_MASK) +
synchronize_irq() should be relatively cheap.
> - For MMU, we're not checking anything, but I guess that if there
> is no job, the mmu can't be doing anything at all?
> ...but then you also gave me the doubt about that one as well.
Same here, if we've properly flushed all jobs, and handled all pending
interrupts, we shouldn't end up with pending MMU irqs when we're asked
to suspend. But the extra mask+synchronize_irq() buys us extra safety.
>
> What I think that would be sensible to do is to get this commit as
> a "clear" fix for the "Really power off" one, then have one or more
> additional commit(s) without any fixes tag to improve the IRQ suspend
> with the new mmu/job irq suspend helpers.
If you need a self-contained fix to avoid power transition interrupts
messing up with suspend, then, as I suggested, it might make more sense
to drop GPU_IRQ_POWER_CHANGED[_ALL] when writing GPU_INT_MASK, which
you want anyway, to avoid being interrupted when you do power
transitions.
>
> Of course *this* commit would introduce the panfrost_gpu_suspend_irq()
> helper directly, instead of moving the logic to a helper in a later one.
>
> Any reason against? :-)
>
> >> +
> >> + /* Now it's safe to request
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 23/11/23 12:15, AngeloGioacchino Del Regno ha scritto:
Il 23/11/23 11:35, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 10:53:20 +0100
AngeloGioacchino Del Regno
wrote:
Some SoCs may be equipped with a GPU containing two core groups
and this is exactly the case of Samsung's Exynos 5422 featuring
an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost
is partial, as this driver currently supports using only one
core group and that's reflected on all parts of it, including
the power on (and power off, previously to this patch) function.
The issue with this is that even though executing the soft reset
operation should power off all cores unconditionally, on at least
one platform we're seeing a crash that seems to be happening due
to an interrupt firing which may be because we are calling power
transition only on the first core group, leaving the second one
unchanged, or because ISR execution was pending before entering
the panfrost_gpu_power_off() function and executed after powering
off the GPU cores, or all of the above.
Finally, solve this by changing the power off flow to
1. Mask and clear all interrupts: we don't need nor want any, as
we are polling PWRTRANS anyway;
2. Call synchronize_irq() after that to make sure that any pending
ISR is executed before powering off the GPU Shaders/Tilers/L2
hence avoiding unpowered registers R/W; and
3. Ignore the core_mask and ask the GPU to poweroff both core groups
Could we split that in two patches? 1+2 in one patch, and 3 in another.
These are two orthogonal fixes IMO.
My initial idea was exactly that, but I opted for one patch doing 'em all
because a "full fix" comprises all of 1+2+3: the third one without the
first two and vice-versa may not fully resolve the issue that was seen
on the HC1 board.
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
Of course it was also necessary to add a `irq` variable to `struct
panfrost_device` as we need to get that in panfrost_gpu_power_off()
for calling synchronize_irq() on it.
Fixes: 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()")
[Regression detected on Odroid HC1, Exynos 5422, Mali-T628 MP6]
Reported-by: Krzysztof Kozlowski
Signed-off-by: AngeloGioacchino Del Regno
---
drivers/gpu/drm/panfrost/panfrost_device.h | 1 +
drivers/gpu/drm/panfrost/panfrost_gpu.c | 26 +++---
2 files changed, 19 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h
b/drivers/gpu/drm/panfrost/panfrost_device.h
index 0fc558db6bfd..b4feaa99e34f 100644
--- a/drivers/gpu/drm/panfrost/panfrost_device.h
+++ b/drivers/gpu/drm/panfrost/panfrost_device.h
@@ -94,6 +94,7 @@ struct panfrost_device {
struct device *dev;
struct drm_device *ddev;
struct platform_device *pdev;
+ int irq;
I know it's the only irq being stored at the panfrost_device level, but
I think it's clearer if we explicitly prefix it with gpu_.
Makes sense, agreed.
void __iomem *iomem;
struct clk *clock;
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 1cc55fb9c45b..30b395125155 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -425,11 +425,21 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
void panfrost_gpu_power_off(struct panfrost_device *pfdev)
{
- u64 core_mask = panfrost_get_core_mask(pfdev);
int ret;
u32 val;
- gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present &
core_mask);
+ /* We are polling PWRTRANS and we don't need nor want interrupts */
I kinda agree with that, but that's not exactly why we're
masking+syncing IRQs here. If that was the only reason, the right fix
would be to modify GPU_IRQ_MASK_ALL so it doesn't include the PWRTRANS
bits.
This fix should cover more than just the power transition IRQ use case:
we want all IRQs to be disabled before the device is suspended.
Eh I should reword that, effectively, because what I wrote as comments make
sense (as in, the logic of it completes) only if you read both of them "as one".
I'll do that in the new suspend irq helper :-)
+ gpu_write(pfdev, GPU_INT_MASK, 0);
+ gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
+
+ /*
+ * Make sure that we don't have pending ISRs, otherwise we'll be
+ * reading and/or writing registers while the GPU is powered off
+ */
+ synchronize_irq(pfdev->irq);
Could we move that to a panfrost_gpu_suspend_irq() helper? I'm also not
sure making it part of panfrost_gpu_power_off() is a good idea. I'd
rather have this panfrost_gpu_suspend_irq() helper called from
panfrost_device_[runtime_]suspend(), along with
panfrost_{mmu,job}_suspend_irq().
Okay I will move that to a helper, but I still want to clarify:
- For JOB, we're checking
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 23/11/23 11:35, Boris Brezillon ha scritto:
On Thu, 23 Nov 2023 10:53:20 +0100
AngeloGioacchino Del Regno
wrote:
Some SoCs may be equipped with a GPU containing two core groups
and this is exactly the case of Samsung's Exynos 5422 featuring
an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost
is partial, as this driver currently supports using only one
core group and that's reflected on all parts of it, including
the power on (and power off, previously to this patch) function.
The issue with this is that even though executing the soft reset
operation should power off all cores unconditionally, on at least
one platform we're seeing a crash that seems to be happening due
to an interrupt firing which may be because we are calling power
transition only on the first core group, leaving the second one
unchanged, or because ISR execution was pending before entering
the panfrost_gpu_power_off() function and executed after powering
off the GPU cores, or all of the above.
Finally, solve this by changing the power off flow to
1. Mask and clear all interrupts: we don't need nor want any, as
we are polling PWRTRANS anyway;
2. Call synchronize_irq() after that to make sure that any pending
ISR is executed before powering off the GPU Shaders/Tilers/L2
hence avoiding unpowered registers R/W; and
3. Ignore the core_mask and ask the GPU to poweroff both core groups
Could we split that in two patches? 1+2 in one patch, and 3 in another.
These are two orthogonal fixes IMO.
My initial idea was exactly that, but I opted for one patch doing 'em all
because a "full fix" comprises all of 1+2+3: the third one without the
first two and vice-versa may not fully resolve the issue that was seen
on the HC1 board.
So, while I agree that it'd be slightly more readable as a diff if those
were two different commits I do have reasons against splitting.
Of course it was also necessary to add a `irq` variable to `struct
panfrost_device` as we need to get that in panfrost_gpu_power_off()
for calling synchronize_irq() on it.
Fixes: 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
panfrost_gpu_power_off()")
[Regression detected on Odroid HC1, Exynos 5422, Mali-T628 MP6]
Reported-by: Krzysztof Kozlowski
Signed-off-by: AngeloGioacchino Del Regno
---
drivers/gpu/drm/panfrost/panfrost_device.h | 1 +
drivers/gpu/drm/panfrost/panfrost_gpu.c| 26 +++---
2 files changed, 19 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h
b/drivers/gpu/drm/panfrost/panfrost_device.h
index 0fc558db6bfd..b4feaa99e34f 100644
--- a/drivers/gpu/drm/panfrost/panfrost_device.h
+++ b/drivers/gpu/drm/panfrost/panfrost_device.h
@@ -94,6 +94,7 @@ struct panfrost_device {
struct device *dev;
struct drm_device *ddev;
struct platform_device *pdev;
+ int irq;
I know it's the only irq being stored at the panfrost_device level, but
I think it's clearer if we explicitly prefix it with gpu_.
Makes sense, agreed.
void __iomem *iomem;
struct clk *clock;
diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
b/drivers/gpu/drm/panfrost/panfrost_gpu.c
index 1cc55fb9c45b..30b395125155 100644
--- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
+++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
@@ -425,11 +425,21 @@ void panfrost_gpu_power_on(struct panfrost_device *pfdev)
void panfrost_gpu_power_off(struct panfrost_device *pfdev)
{
- u64 core_mask = panfrost_get_core_mask(pfdev);
int ret;
u32 val;
- gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present & core_mask);
+ /* We are polling PWRTRANS and we don't need nor want interrupts */
I kinda agree with that, but that's not exactly why we're
masking+syncing IRQs here. If that was the only reason, the right fix
would be to modify GPU_IRQ_MASK_ALL so it doesn't include the PWRTRANS
bits.
This fix should cover more than just the power transition IRQ use case:
we want all IRQs to be disabled before the device is suspended.
Eh I should reword that, effectively, because what I wrote as comments make
sense (as in, the logic of it completes) only if you read both of them "as one".
I'll do that in the new suspend irq helper :-)
+ gpu_write(pfdev, GPU_INT_MASK, 0);
+ gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
+
+ /*
+* Make sure that we don't have pending ISRs, otherwise we'll be
+* reading and/or writing registers while the GPU is powered off
+*/
+ synchronize_irq(pfdev->irq);
Could we move that to a panfrost_gpu_suspend_irq() helper? I'm also not
sure making it part of panfrost_gpu_power_off() is a good idea. I'd
rather have this panfrost_gpu_suspend_irq() helper called from
panfrost_device_[runtime_]suspend(), along with
panfrost_{mmu,job}_suspend_irq().
Okay I will move that to a helper, but I still want to clarify:
- For JOB, we're checking if
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
Il 23/11/23 11:39, Krzysztof Kozlowski ha scritto: On 23/11/2023 10:53, AngeloGioacchino Del Regno wrote: Some SoCs may be equipped with a GPU containing two core groups and this is exactly the case of Samsung's Exynos 5422 featuring an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost is partial, as this driver currently supports using only one core group and that's reflected on all parts of it, including the power on (and power off, previously to this patch) function. The issue with this is that even though executing the soft reset operation should power off all cores unconditionally, on at least one platform we're seeing a crash that seems to be happening due to an interrupt firing which may be because we are calling power transition only on the first core group, leaving the second one unchanged, or because ISR execution was pending before entering the panfrost_gpu_power_off() function and executed after powering off the GPU cores, or all of the above. Does not apply - I tried next 20231117/21/22/23. Sorry about that, I'll send a v2 soon. Thanks, Angelo Best regards, Krzysztof ___ Kernel mailing list -- ker...@mailman.collabora.com To unsubscribe send an email to kernel-le...@mailman.collabora.com
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
On 23/11/2023 10:53, AngeloGioacchino Del Regno wrote: > Some SoCs may be equipped with a GPU containing two core groups > and this is exactly the case of Samsung's Exynos 5422 featuring > an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost > is partial, as this driver currently supports using only one > core group and that's reflected on all parts of it, including > the power on (and power off, previously to this patch) function. > > The issue with this is that even though executing the soft reset > operation should power off all cores unconditionally, on at least > one platform we're seeing a crash that seems to be happening due > to an interrupt firing which may be because we are calling power > transition only on the first core group, leaving the second one > unchanged, or because ISR execution was pending before entering > the panfrost_gpu_power_off() function and executed after powering > off the GPU cores, or all of the above. Does not apply - I tried next 20231117/21/22/23. Best regards, Krzysztof
Re: [PATCH] drm/panfrost: Ignore core_mask for poweroff and sync interrupts
On Thu, 23 Nov 2023 10:53:20 +0100
AngeloGioacchino Del Regno
wrote:
> Some SoCs may be equipped with a GPU containing two core groups
> and this is exactly the case of Samsung's Exynos 5422 featuring
> an ARM Mali-T628 MP6 GPU: the support for this GPU in Panfrost
> is partial, as this driver currently supports using only one
> core group and that's reflected on all parts of it, including
> the power on (and power off, previously to this patch) function.
>
> The issue with this is that even though executing the soft reset
> operation should power off all cores unconditionally, on at least
> one platform we're seeing a crash that seems to be happening due
> to an interrupt firing which may be because we are calling power
> transition only on the first core group, leaving the second one
> unchanged, or because ISR execution was pending before entering
> the panfrost_gpu_power_off() function and executed after powering
> off the GPU cores, or all of the above.
>
> Finally, solve this by changing the power off flow to
> 1. Mask and clear all interrupts: we don't need nor want any, as
> we are polling PWRTRANS anyway;
> 2. Call synchronize_irq() after that to make sure that any pending
> ISR is executed before powering off the GPU Shaders/Tilers/L2
> hence avoiding unpowered registers R/W; and
> 3. Ignore the core_mask and ask the GPU to poweroff both core groups
Could we split that in two patches? 1+2 in one patch, and 3 in another.
These are two orthogonal fixes IMO.
>
> Of course it was also necessary to add a `irq` variable to `struct
> panfrost_device` as we need to get that in panfrost_gpu_power_off()
> for calling synchronize_irq() on it.
>
> Fixes: 123b431f8a5c ("drm/panfrost: Really power off GPU cores in
> panfrost_gpu_power_off()")
> [Regression detected on Odroid HC1, Exynos 5422, Mali-T628 MP6]
> Reported-by: Krzysztof Kozlowski
> Signed-off-by: AngeloGioacchino Del Regno
>
> ---
> drivers/gpu/drm/panfrost/panfrost_device.h | 1 +
> drivers/gpu/drm/panfrost/panfrost_gpu.c| 26 +++---
> 2 files changed, 19 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/gpu/drm/panfrost/panfrost_device.h
> b/drivers/gpu/drm/panfrost/panfrost_device.h
> index 0fc558db6bfd..b4feaa99e34f 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_device.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_device.h
> @@ -94,6 +94,7 @@ struct panfrost_device {
> struct device *dev;
> struct drm_device *ddev;
> struct platform_device *pdev;
> + int irq;
I know it's the only irq being stored at the panfrost_device level, but
I think it's clearer if we explicitly prefix it with gpu_.
>
> void __iomem *iomem;
> struct clk *clock;
> diff --git a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> index 1cc55fb9c45b..30b395125155 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_gpu.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_gpu.c
> @@ -425,11 +425,21 @@ void panfrost_gpu_power_on(struct panfrost_device
> *pfdev)
>
> void panfrost_gpu_power_off(struct panfrost_device *pfdev)
> {
> - u64 core_mask = panfrost_get_core_mask(pfdev);
> int ret;
> u32 val;
>
> - gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present &
> core_mask);
> + /* We are polling PWRTRANS and we don't need nor want interrupts */
I kinda agree with that, but that's not exactly why we're
masking+syncing IRQs here. If that was the only reason, the right fix
would be to modify GPU_IRQ_MASK_ALL so it doesn't include the PWRTRANS
bits.
This fix should cover more than just the power transition IRQ use case:
we want all IRQs to be disabled before the device is suspended.
> + gpu_write(pfdev, GPU_INT_MASK, 0);
> + gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
> +
> + /*
> + * Make sure that we don't have pending ISRs, otherwise we'll be
> + * reading and/or writing registers while the GPU is powered off
> + */
> + synchronize_irq(pfdev->irq);
Could we move that to a panfrost_gpu_suspend_irq() helper? I'm also not
sure making it part of panfrost_gpu_power_off() is a good idea. I'd
rather have this panfrost_gpu_suspend_irq() helper called from
panfrost_device_[runtime_]suspend(), along with
panfrost_{mmu,job}_suspend_irq().
> +
> + /* Now it's safe to request poweroff for Shaders, Tilers and L2 */
It was safe before too, it's just that we probably don't want to be
interrupted, if all we do is ignore the interrupts we receive, hence
the suggestion to not use GPU_IRQ_MASK_ALL, and only enable the
IRQs we care about instead.
> + gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
> ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO,
>val, !val, 1, 1000);
> if (ret)
> @@ -441,7 +451,7 @@ void panfrost_gpu_power_off(struct panfrost_device *pfdev)
> if (ret)
>
