[Intel-gfx] [PATCH 6/8] drm/i915/rpm: d3cold Policy
Gupta, Anshuman
anshuman.gupta at intel.com
Thu Aug 4 06:29:31 UTC 2022
> -----Original Message-----
> From: Vivi, Rodrigo <rodrigo.vivi at intel.com>
> Sent: Thursday, August 4, 2022 2:16 AM
> To: Tangudu, Tilak <tilak.tangudu at intel.com>
> Cc: Ewins, Jon <jon.ewins at intel.com>; Belgaumkar, Vinay
> <vinay.belgaumkar at intel.com>; Roper, Matthew D
> <matthew.d.roper at intel.com>; Wilson, Chris P <chris.p.wilson at intel.com>;
> Nikula, Jani <jani.nikula at intel.com>; Gupta, saurabhg
> <saurabhg.gupta at intel.com>; Gupta, Anshuman
> <anshuman.gupta at intel.com>; Nilawar, Badal <badal.nilawar at intel.com>;
> Deak, Imre <imre.deak at intel.com>; Iddamsetty, Aravind
> <aravind.iddamsetty at intel.com>; intel-gfx at lists.freedesktop.org
> Subject: Re: [PATCH 6/8] drm/i915/rpm: d3cold Policy
>
> On Thu, Jul 21, 2022 at 03:29:53PM +0530, tilak.tangudu at intel.com wrote:
> > From: Tilak Tangudu <tilak.tangudu at intel.com>
>
> This needs to be sorted out... or the patch split or use the Co-developed-by:...
>
> >
> > Add d3cold_sr_lmem_threshold modparam to choose between d3cold-off
> > zero watt and d3hot/d3cold-VRAM Self Refresh.
> > i915 requires to evict the lmem objects to smem in order to support
> > d3cold-Off. if platform does not supports vram_sr feature then fall
> > back to d3hot by disabling d3cold to avoid the rpm suspend/resume
> > latency.
> > Extend the d3cold_sr_lmem_threshold modparam to debugfs i915_params so
> > that, it can be used by igt test.
> >
> > If gfx root port is not capable of sending PME from d3cold or doesn't
> > have _PR3 power resources then only d3hot state can be supported.
> >
> > Adding intel_pm_prepare_targeted_d3_state() to choose the correct
> > target d3 state and cache it to intel_runtime_pm structure, it can be
> > used in rpm suspend/resume callback accordingly.
> >
> > v2: lmem->avail stopped tracking lmem usage since ttm is introduced,
> > so removed lmem->avail usage in policy.
> > FIXME here, lmem usage is not added, need to be added by using query
> > functions.
> > FIXME, Forcing the policy to enter D3COLD_OFF for validation purpose.
> >
> > Cc: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > Signed-off-by: Anshuman Gupta <anshuman.gupta at intel.com>
> > Signed-off-by: Tilak Tangudu <tilak.tangudu at intel.com>
> > ---
> > drivers/gpu/drm/i915/i915_driver.c | 6 +++++
> > drivers/gpu/drm/i915/i915_params.c | 5 ++++
> > drivers/gpu/drm/i915/i915_params.h | 1 +
> > drivers/gpu/drm/i915/intel_pm.c | 35 +++++++++++++++++++++++++
> > drivers/gpu/drm/i915/intel_pm.h | 1 +
> > drivers/gpu/drm/i915/intel_runtime_pm.h | 7 +++++
> > 6 files changed, 55 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/i915/i915_driver.c
> > b/drivers/gpu/drm/i915/i915_driver.c
> > index 4c36554567fd..2b2e9563f149 100644
> > --- a/drivers/gpu/drm/i915/i915_driver.c
> > +++ b/drivers/gpu/drm/i915/i915_driver.c
> > @@ -1581,6 +1581,12 @@ static int intel_runtime_idle(struct device *kdev)
> > struct drm_i915_private *i915 = kdev_to_i915(kdev);
> > int ret = 1;
> >
> > + disable_rpm_wakeref_asserts(&i915->runtime_pm);
> > + ret = intel_pm_prepare_targeted_d3_state(i915);
> > + if (!ret)
> > + ret = 1;
>
> why?
>
> > +
> > + enable_rpm_wakeref_asserts(&i915->runtime_pm);
> > pm_runtime_mark_last_busy(kdev);
> > pm_runtime_autosuspend(kdev);
> >
> > diff --git a/drivers/gpu/drm/i915/i915_params.c
> > b/drivers/gpu/drm/i915/i915_params.c
> > index 6fc475a5db61..4603f5c2ed77 100644
> > --- a/drivers/gpu/drm/i915/i915_params.c
> > +++ b/drivers/gpu/drm/i915/i915_params.c
> > @@ -197,6 +197,11 @@ i915_param_named(enable_gvt, bool, 0400,
> > "Enable support for Intel GVT-g graphics virtualization host
> > support(default:false)"); #endif
> >
> > +i915_param_named_unsafe(d3cold_sr_lmem_threshold, int, 0600,
> > + "Enable VRAM Self refresh when size of lmem is greater to this
> threshold. "
> > + "If VRAM Self Refresh is not available then fall back to d3cold. "
> > + "It helps to optimize the suspend/resume latecy. (default: 300mb)");
> > +
> > #if CONFIG_DRM_I915_REQUEST_TIMEOUT
> > i915_param_named_unsafe(request_timeout_ms, uint, 0600,
> > "Default request/fence/batch buffer expiration
> timeout."); diff
> > --git a/drivers/gpu/drm/i915/i915_params.h
> > b/drivers/gpu/drm/i915/i915_params.h
> > index 2733cb6cfe09..1a86711038da 100644
> > --- a/drivers/gpu/drm/i915/i915_params.h
> > +++ b/drivers/gpu/drm/i915/i915_params.h
> > @@ -75,6 +75,7 @@ struct drm_printer;
> > param(unsigned int, request_timeout_ms,
> CONFIG_DRM_I915_REQUEST_TIMEOUT,
> CONFIG_DRM_I915_REQUEST_TIMEOUT ? 0600 : 0) \
> > param(unsigned int, lmem_size, 0, 0400) \
> > param(unsigned int, lmem_bar_size, 0, 0400) \
> > + param(int, d3cold_sr_lmem_threshold, 300, 0600) \
> > /* leave bools at the end to not create holes */ \
> > param(bool, enable_hangcheck, true, 0600) \
> > param(bool, load_detect_test, false, 0600) \ diff --git
> > a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > index f06babdb3a8c..20b0638ecd5c 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > @@ -8287,6 +8287,41 @@ void intel_pm_setup(struct drm_i915_private
> *dev_priv)
> > atomic_set(&dev_priv->runtime_pm.wakeref_count, 0); }
> >
> > +int intel_pm_prepare_targeted_d3_state(struct drm_i915_private *i915)
> > +{
> > + struct intel_runtime_pm *rpm = &i915->runtime_pm;
> > + struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
> > + u64 lmem_used = 0;
> > + struct pci_dev *root_pdev;
> > + int ret = 0;
> > +
> > + /* igfx will return from here */
>
> then this patch is preventing runtime_pm on igfx no?!
> or we need to return 0...
ret is initialized with 0. I will change it to return with explicit 'return 0'.
This won't block runtime PM for igfx as caller will request the autosuspend.
rpm->d3_state will be by default initialized with INTEL_D3HOT i.e. 0.
Please correct me in case I am missing something.
>
> > + root_pdev = pcie_find_root_port(pdev);
> > + if (!root_pdev)
> > + return ret;
> > +
> > + /* D3Cold requires PME capability and _PR3 power resource */
> > + if (!pci_pme_capable(root_pdev, PCI_D3cold) ||
> !pci_pr3_present(root_pdev))
> > + return ret;
>
> If this is the case we probably need to block D3cold, but not D3hot, right?
Similar to above rpm->d3_state will be by default initialized with INTEL_D3HOT i.e. 0.
And we do not need to block the d3cold here as device is already incapable of it.
Server platforms may hit this code path.
>
> > +
> > + /* FXME query the LMEM usage and fill lmem_used */
> > + /* Trigger D3COLD_OFF always to validate with all tests */
> > + if (lmem_used < i915->params.d3cold_sr_lmem_threshold * 1024 *
> 1024) {
> > + rpm->d3_state = INTEL_D3COLD_OFF;
> > + drm_dbg(&i915->drm, "Prepared for D3Cold off\n");
> > + } else {
> > + /* Disable D3Cold to reduce the eviction latency */
> > + rpm->d3_state = INTEL_D3HOT;
> > + }
> > +
> > + if (rpm->d3_state == INTEL_D3HOT)
> > + pci_d3cold_disable(root_pdev);
> > + else
> > + pci_d3cold_enable(root_pdev);
>
> why not merge these both if states?
For future D3COLD_VRAM_SR support scalability , target d3 state assignment and d3cold disabling kept separate.
Let me know it is must to merge both if /else block.
Thanks,
Anshuman Gupta.
>
> > +
> > + return ret;
> > +}
> > +
> > static struct intel_global_state *intel_dbuf_duplicate_state(struct
> > intel_global_obj *obj) {
> > struct intel_dbuf_state *dbuf_state; diff --git
> > a/drivers/gpu/drm/i915/intel_pm.h b/drivers/gpu/drm/i915/intel_pm.h
> > index 945503ae493e..7827b0c1a2f3 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.h
> > +++ b/drivers/gpu/drm/i915/intel_pm.h
> > @@ -31,6 +31,7 @@ int ilk_wm_max_level(const struct drm_i915_private
> > *dev_priv); void intel_init_pm(struct drm_i915_private *dev_priv);
> > void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv);
> > void intel_pm_setup(struct drm_i915_private *dev_priv);
> > +int intel_pm_prepare_targeted_d3_state(struct drm_i915_private
> > +*i915);
> > void g4x_wm_get_hw_state(struct drm_i915_private *dev_priv); void
> > vlv_wm_get_hw_state(struct drm_i915_private *dev_priv); void
> > ilk_wm_get_hw_state(struct drm_i915_private *dev_priv); diff --git
> > a/drivers/gpu/drm/i915/intel_runtime_pm.h
> > b/drivers/gpu/drm/i915/intel_runtime_pm.h
> > index 99418c3a934a..568559b71b70 100644
> > --- a/drivers/gpu/drm/i915/intel_runtime_pm.h
> > +++ b/drivers/gpu/drm/i915/intel_runtime_pm.h
> > @@ -22,6 +22,12 @@ enum i915_drm_suspend_mode {
> > I915_DRM_SUSPEND_HIBERNATE,
> > };
> >
> > +enum intel_gfx_d3_state {
> > + INTEL_D3HOT,
> > + INTEL_D3COLD_OFF,
> > + INTEL_D3COLD_VRAM_SR,
> > +};
> > +
> > /*
> > * This struct helps tracking the state needed for runtime PM, which puts the
> > * device in PCI D3 state. Notice that when this happens, nothing on
> > the @@ -52,6 +58,7 @@ struct intel_runtime_pm {
> > bool suspended;
> > bool irqs_enabled;
> > bool no_wakeref_tracking;
> > + enum intel_gfx_d3_state d3_state;
> >
> > #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_RUNTIME_PM)
> > /*
> > --
> > 2.25.1
> >
More information about the Intel-gfx
mailing list