diff mbox series

drm/i915: Unlock the shared hwsp_gtt object after pinning

Message ID 20200903133144.740-1-mika.kuoppala@linux.intel.com (mailing list archive)
State New, archived
Headers show
Series drm/i915: Unlock the shared hwsp_gtt object after pinning | expand

Commit Message

Mika Kuoppala Sept. 3, 2020, 1:31 p.m. UTC
From: Thomas Hellström <thomas.hellstrom@intel.com>

The hwsp_gtt object is used for sub-allocation and could therefore
be shared by many contexts causing unnecessary contention during
concurrent context pinning.
However since we're currently locking it only for pinning, it remains
resident until we unpin it, and therefore it's safe to drop the
lock early, allowing for concurrent thread access.

Signed-off-by: Thomas Hellström <thomas.hellstrom@intel.com>
---
 drivers/gpu/drm/i915/gt/intel_context.c | 7 +++++++
 1 file changed, 7 insertions(+)

Comments

Chris Wilson Sept. 3, 2020, 2:01 p.m. UTC | #1
Quoting Mika Kuoppala (2020-09-03 14:31:44)
> From: Thomas Hellström <thomas.hellstrom@intel.com>
> 
> The hwsp_gtt object is used for sub-allocation and could therefore
> be shared by many contexts causing unnecessary contention during
> concurrent context pinning.
> However since we're currently locking it only for pinning, it remains
> resident until we unpin it, and therefore it's safe to drop the
> lock early, allowing for concurrent thread access.
> 
> Signed-off-by: Thomas Hellström <thomas.hellstrom@intel.com>
> ---
>  drivers/gpu/drm/i915/gt/intel_context.c | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
> index 61b05cd4c47a..65e956ba19e1 100644
> --- a/drivers/gpu/drm/i915/gt/intel_context.c
> +++ b/drivers/gpu/drm/i915/gt/intel_context.c
> @@ -271,6 +271,13 @@ int __intel_context_do_pin_ww(struct intel_context *ce,
>         i915_active_release(&ce->active);
>  err_ctx_unpin:
>         intel_context_post_unpin(ce);
> +
> +       /*
> +        * Unlock the hwsp_ggtt object since it's shared. This is fine for now
> +        * since the lock has been used for pinning only, not fencing.
> +        */
> +       i915_gem_ww_unlock_single(ce->timeline->hwsp_ggtt->obj);

The timeline is not special here, the same rules for locking/unlock can
equally be applied to all the global state. You may even go as far as
putting a local acquire context here, which would then have avoided the
cross driver pollution.

Anyway, if it works for the timeline, there is no reason to keep the
other globals locked. They are pinned and on completely different usage
cycles to the user objects. [They can't be evicted without interacting
with the HW to ensure the context has been flushed, so far there has
been no way to do so without stalling for a freshly submitted request,
hence we let them retire gracefully and kick the kernel timelines to
flush completed contexts.]
-Chris
Thomas Hellström (Intel) Sept. 3, 2020, 2:18 p.m. UTC | #2
On 9/3/20 4:01 PM, Chris Wilson wrote:
> Quoting Mika Kuoppala (2020-09-03 14:31:44)
>> From: Thomas Hellström <thomas.hellstrom@intel.com>
>>
>> The hwsp_gtt object is used for sub-allocation and could therefore
>> be shared by many contexts causing unnecessary contention during
>> concurrent context pinning.
>> However since we're currently locking it only for pinning, it remains
>> resident until we unpin it, and therefore it's safe to drop the
>> lock early, allowing for concurrent thread access.
>>
>> Signed-off-by: Thomas Hellström <thomas.hellstrom@intel.com>
>> ---
>>   drivers/gpu/drm/i915/gt/intel_context.c | 7 +++++++
>>   1 file changed, 7 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
>> index 61b05cd4c47a..65e956ba19e1 100644
>> --- a/drivers/gpu/drm/i915/gt/intel_context.c
>> +++ b/drivers/gpu/drm/i915/gt/intel_context.c
>> @@ -271,6 +271,13 @@ int __intel_context_do_pin_ww(struct intel_context *ce,
>>          i915_active_release(&ce->active);
>>   err_ctx_unpin:
>>          intel_context_post_unpin(ce);
>> +
>> +       /*
>> +        * Unlock the hwsp_ggtt object since it's shared. This is fine for now
>> +        * since the lock has been used for pinning only, not fencing.
>> +        */
>> +       i915_gem_ww_unlock_single(ce->timeline->hwsp_ggtt->obj);
> The timeline is not special here, the same rules for locking/unlock can
> equally be applied to all the global state. You may even go as far as
> putting a local acquire context here, which would then have avoided the
> cross driver pollution.
>
> Anyway, if it works for the timeline, there is no reason to keep the
> other globals locked. They are pinned and on completely different usage
> cycles to the user objects.

Agreed.  In principle everything we want to perma-pin we can unlock 
early, Keeping it under the same acquire context makes it possible to do 
this as part of a bigger ww transaction. Although initially this 
solution was discarded in favour of removing the suballocation to be 
able to ditch the perma-pinning in the future.

But is this acceptable to address the pi-isolated issue until we have 
something more thought-through in place? Or is there somethin specific 
you want me to change for a R-B?

Thanks,

Thomas
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
index 61b05cd4c47a..65e956ba19e1 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.c
+++ b/drivers/gpu/drm/i915/gt/intel_context.c
@@ -271,6 +271,13 @@  int __intel_context_do_pin_ww(struct intel_context *ce,
 	i915_active_release(&ce->active);
 err_ctx_unpin:
 	intel_context_post_unpin(ce);
+
+	/*
+	 * Unlock the hwsp_ggtt object since it's shared. This is fine for now
+	 * since the lock has been used for pinning only, not fencing.
+	 */
+	i915_gem_ww_unlock_single(ce->timeline->hwsp_ggtt->obj);
+
 	return err;
 }