On Fri, Feb 28, 2014 at 06:44:03PM +0000, Chris Wilson wrote:
> We don't want to suffer scheduling delay when turning off the GPU after
> waking it up to touch registers. Ideally, we only want to keep the GPU
> awake for the register access sequence, with a single forcewake dance on
> the first access and release immediately after the last. We set a timer
> on the first access so that we only dance once and on the next scheduler
> tick, we drop the forcewake again.
>
> This moves the cleanup routine from the common i915 workqueue to a timer
> func so that we don't anger powertop, and drop the forcewake again
> quicker.
>
> v2: Enable the deferred force_wake_put for regular register reads as
> well.
>
> Signed-off-by: Chris Wilson <ch...@chris-wilson.co.uk>
> Cc: Ben Widawsky <b...@bwidawsk.net>
> Cc: Ville Syrjälä <ville.syrj...@linux.intel.com>
> ---
> drivers/gpu/drm/i915/i915_drv.h | 2 +-
> drivers/gpu/drm/i915/intel_uncore.c | 34 +++++++++++++++-------------------
> 2 files changed, 16 insertions(+), 20 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index b22765192018..8af8e0dd3943 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -508,7 +508,7 @@ struct intel_uncore {
> unsigned fw_rendercount;
> unsigned fw_mediacount;
>
> - struct delayed_work force_wake_work;
> + struct timer_list force_wake_timer;
> };
>
> #define DEV_INFO_FOR_EACH_FLAG(func, sep) \
> diff --git a/drivers/gpu/drm/i915/intel_uncore.c
> b/drivers/gpu/drm/i915/intel_uncore.c
> index c62841404c82..8ee171178bfe 100644
> --- a/drivers/gpu/drm/i915/intel_uncore.c
> +++ b/drivers/gpu/drm/i915/intel_uncore.c
> @@ -289,10 +289,8 @@ void vlv_force_wake_put(struct drm_i915_private
> *dev_priv,
> spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
> }
>
> -static void gen6_force_wake_work(struct work_struct *work)
> +static void gen6_force_wake_timer(struct drm_i915_private *dev_priv)
> {
> - struct drm_i915_private *dev_priv =
> - container_of(work, typeof(*dev_priv),
> uncore.force_wake_work.work);
> unsigned long irqflags;
>
> spin_lock_irqsave(&dev_priv->uncore.lock, irqflags);
> @@ -405,9 +403,8 @@ void gen6_gt_force_wake_put(struct drm_i915_private
> *dev_priv, int fw_engine)
> spin_lock_irqsave(&dev_priv->uncore.lock, irqflags);
> if (--dev_priv->uncore.forcewake_count == 0) {
> dev_priv->uncore.forcewake_count++;
> - mod_delayed_work(dev_priv->wq,
> - &dev_priv->uncore.force_wake_work,
> - 1);
> + mod_timer_pinned(&dev_priv->uncore.force_wake_timer,
> + jiffies + 1);
This could expire more or less immediately, but it should be fine. We'd
just end up doing two forcewake_get()s instead of one, which should
still be better than >2 if the theory of the timer holds.
> }
> spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
>
> @@ -484,17 +481,15 @@ gen5_read##x(struct drm_i915_private *dev_priv, off_t
> reg, bool trace) { \
> static u##x \
> gen6_read##x(struct drm_i915_private *dev_priv, off_t reg, bool trace) { \
> REG_READ_HEADER(x); \
> - if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
> - if (dev_priv->uncore.forcewake_count == 0) \
> - dev_priv->uncore.funcs.force_wake_get(dev_priv, \
> - FORCEWAKE_ALL); \
> - val = __raw_i915_read##x(dev_priv, reg); \
> - if (dev_priv->uncore.forcewake_count == 0) \
> - dev_priv->uncore.funcs.force_wake_put(dev_priv, \
> - FORCEWAKE_ALL); \
> - } else { \
> - val = __raw_i915_read##x(dev_priv, reg); \
> + if (dev_priv->uncore.forcewake_count == 0 && \
> + NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
> + dev_priv->uncore.funcs.force_wake_get(dev_priv, \
> + FORCEWAKE_ALL); \
> + dev_priv->uncore.forcewake_count++; \
> + mod_timer_pinned(&dev_priv->uncore.force_wake_timer, \
> + jiffies + 1); \
> } \
> + val = __raw_i915_read##x(dev_priv, reg); \
> REG_READ_FOOTER; \
> }
>
> @@ -681,8 +676,9 @@ void intel_uncore_init(struct drm_device *dev)
> {
> struct drm_i915_private *dev_priv = dev->dev_private;
>
> - INIT_DELAYED_WORK(&dev_priv->uncore.force_wake_work,
> - gen6_force_wake_work);
> + setup_timer(&dev_priv->uncore.force_wake_timer,
> + (void (*)(unsigned long))gen6_force_wake_timer,
I'd prefer to do the required casting in gen6_force_wake_timer(). Seems
a bit less error prone since we'd at least get type checking for the
function pointer.
> + (unsigned long)dev_priv);
>
> if (IS_VALLEYVIEW(dev)) {
> dev_priv->uncore.funcs.force_wake_get = __vlv_force_wake_get;
> @@ -794,7 +790,7 @@ void intel_uncore_fini(struct drm_device *dev)
> {
> struct drm_i915_private *dev_priv = dev->dev_private;
>
> - flush_delayed_work(&dev_priv->uncore.force_wake_work);
> + del_timer_sync(&dev_priv->uncore.force_wake_timer);
This could leave force wake enabled.
>
> /* Paranoia: make sure we have disabled everything before we exit. */
> intel_uncore_sanitize(dev);
> --
> 1.9.0
--
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
