[Intel-gfx] [PATCH 1/1] drm/i915: Implement workaround for CDCLK PLL disable/enable
Lisovskiy, Stanislav
stanislav.lisovskiy at intel.com
Mon Jan 9 10:07:43 UTC 2023
On Thu, Jan 05, 2023 at 03:05:40AM +0200, Srivatsa, Anusha wrote:
>
>
> > -----Original Message-----
> > From: Lisovskiy, Stanislav <stanislav.lisovskiy at intel.com>
> > Sent: Thursday, December 15, 2022 2:14 AM
> > To: Srivatsa, Anusha <anusha.srivatsa at intel.com>
> > Cc: intel-gfx at lists.freedesktop.org; Nikula, Jani <jani.nikula at intel.com>
> > Subject: Re: [Intel-gfx] [PATCH 1/1] drm/i915: Implement workaround for
> > CDCLK PLL disable/enable
> >
> > On Wed, Dec 14, 2022 at 09:15:07PM +0200, Srivatsa, Anusha wrote:
> > >
> > >
> > > > -----Original Message-----
> > > > From: Lisovskiy, Stanislav <stanislav.lisovskiy at intel.com>
> > > > Sent: Wednesday, December 14, 2022 2:31 AM
> > > > To: Srivatsa, Anusha <anusha.srivatsa at intel.com>
> > > > Cc: intel-gfx at lists.freedesktop.org; Nikula, Jani
> > > > <jani.nikula at intel.com>
> > > > Subject: Re: [Intel-gfx] [PATCH 1/1] drm/i915: Implement workaround
> > > > for CDCLK PLL disable/enable
> > > >
> > > > On Tue, Nov 29, 2022 at 09:19:40PM +0200, Srivatsa, Anusha wrote:
> > > > >
> > > > >
> > > > > > -----Original Message-----
> > > > > > From: Intel-gfx <intel-gfx-bounces at lists.freedesktop.org> On
> > > > > > Behalf Of Stanislav Lisovskiy
> > > > > > Sent: Thursday, November 24, 2022 2:36 AM
> > > > > > To: intel-gfx at lists.freedesktop.org
> > > > > > Cc: Nikula, Jani <jani.nikula at intel.com>
> > > > > > Subject: [Intel-gfx] [PATCH 1/1] drm/i915: Implement workaround
> > > > > > for CDCLK PLL disable/enable
> > > > > >
> > > > > > It was reported that we might get a hung and loss of register
> > > > > > access in some cases when CDCLK PLL is disabled and then
> > > > > > enabled, while squashing is enabled.
> > > > > > As a workaround it was proposed by HW team that SW should
> > > > > > disable squashing when CDCLK PLL is being reenabled.
> > > > > >
> > > > > > Signed-off-by: Stanislav Lisovskiy
> > > > > > <stanislav.lisovskiy at intel.com>
> > > > > > ---
> > > > > > drivers/gpu/drm/i915/display/intel_cdclk.c | 14 ++++++++++++--
> > > > > > 1 file changed, 12 insertions(+), 2 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > > > > b/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > > > > index 0c107a38f9d0..e338f288c9ac 100644
> > > > > > --- a/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > > > > +++ b/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > > > > @@ -1801,6 +1801,13 @@ static bool
> > > > > > cdclk_compute_crawl_and_squash_midpoint(struct
> > drm_i915_private
> > > > *i91
> > > > > > return true;
> > > > > > }
> > > > > >
> > > > > > +static bool pll_enable_wa_needed(struct drm_i915_private
> > > > > > +*dev_priv)
> > > > {
> > > > > > + return ((IS_DG2(dev_priv) || IS_METEORLAKE(dev_priv))
> > > > > > + && dev_priv->display.cdclk.hw.vco > 0
> > > > > > + && HAS_CDCLK_SQUASH(dev_priv));
> > > > > Redundant check. If it is MTL or DG2, then it will have
> > > > HAS_CDCLK_SQUASH set to true always. Shouldn't vco be 0 instead of >
> > 0.
> > > > The commit message says the hang can be observed when moving from
> > 0
> > > > to
> > > > > 0 vco.
> > > > >
> > > > > Anusha
> > > >
> > > > Idea was that we probably should bind more to the feature rather
> > > > than platform, I agree checking both "IS_DG2" and if platform has
> > > > squashing is redundant, because then we would have to add each new
> > > > platform manually, so I would leave HAS_CDCLK_SQUASH and then at
> > > > some point just cut off using some INTEL_GEN or other check all the
> > > > new platforms where this is fixed in HW.
> > > >
> > > > Regarding vco, the icl_cdclk_pll_update func works as follows:
> > > >
> > > > if (i915->display.cdclk.hw.vco != 0 &&
> > > > i915->display.cdclk.hw.vco != vco)
> > > > icl_cdclk_pll_disable(i915);
> > > >
> > > > if (i915->display.cdclk.hw.vco != vco)
> > > > icl_cdclk_pll_enable(i915, vco);
> > > >
> > > > 1) if vco changes from zero value(i915->display.cdclk.hw.vco) to
> > > > non-zero value(vco), that means
> > > > currently squashing is anyway disabled(if vco == 0, cdclk is set to
> > "bypass"
> > > > and squash waveform
> > > > is 0), so the W/A is not needed.
> > > >
> > > > 2) if vco changes from non-zero value in i915->display.cdclk.hw.vco
> > > > to zero value(vco), we are not
> > > > going to call icl_cdclk_pll_enable anyway so W/A is also not needed.
> > > >
> > > > 3) if vco changes from some non-zero value in
> > > > i915->display.cdclk.hw.vco to other non-zero value(vco),
> > > > which can happen if CDCLK changes, then icl_cdclk_pll_disable(hw
> > > > vco!=0 and vco!=0) and then
> > > > icl_cdclk_pll_enable would be called(hw vco is still not equal to vco)
> > > > So that disabling enabling in between is what we are interested
> > > > and where we should make sure
> > > > squashing is disabled.
> > > > BTW I have another question - is this code even correct?
> > > > Shouldn't we avoid disabling/enabling PLL if we have
> > > > squashing/crawling? As I understood the whole point of having
> > > > swuashing/crawling is avoiding swithing the PLL
> > > > on and off.
> > > >
> > > Squashing works when we don't need to change the PLL ratio. We fix the
> > PLL ratio to say 307 (this can change from platform to platform). Then
> > squashing can be used to vary frequencies below this. So we set squasher
> > to ffff it will mean highest. We can use squasher to change frequency with
> > squash waveform, max being ffff and any value lower will take lower
> > frequencies.
> > > Cdclk Crawling is used when the PLL has to be changed. Eg higher than
> > 307 then we need to update PLL. In that case we first program squashing to
> > ffff(this will take to 307) n then use crawling to go up to the desired
> > frequency.
> >
> > if (HAS_CDCLK_CRAWL(dev_priv) && dev_priv->display.cdclk.hw.vco > 0 &&
> > vco > 0 &&
> > !cdclk_pll_is_unknown(dev_priv->display.cdclk.hw.vco)) {
> > if (dev_priv->display.cdclk.hw.vco != vco)
> > adlp_cdclk_pll_crawl(dev_priv, vco);
> > } else if (DISPLAY_VER(dev_priv) >= 11) {
> > if (pll_enable_wa_needed(dev_priv))
> > dg2_cdclk_squash_program(dev_priv, 0);
> >
> > icl_cdclk_pll_update(dev_priv, vco);
> > } else
> > bxt_cdclk_pll_update(dev_priv, vco);
> >
> > I think that condition above will trigger CDCLK crawl whenever vco is not ~0,
> > CDCLK crawl is supported and both hw.vco and vco are > 0 and vco has to
> > be changed.
>
> Why?
Well... Because that is exactly what this condition does:
1) !cdclk_pll_is_unknown(dev_priv->display.cdclk.hw.vco) means hw.vco is not ~0
2) HAS_CDCLK_CRAWL(dev_priv) checks if CDCLK crawl is supported
3) dev_priv->display.cdclk.hw.vco > 0 && vco > 0 check if both hw.vco and vco are > 0
4) and if (dev_priv->display.cdclk.hw.vco != vco) checks if new vco is different from hw.vco
Summing all this exactly produces my statement above.
>
> >
> > In bxt_set_cdclk which calls _bxt_set_cdclk we calculate the midpoint
> > however I don't see how _bxt_set_cdclk is going to distinguish between
> > crawling and squashing.
>
> We are populating the mid_cdclk_config according to the desired action.
> "/* - If moving to a higher cdclk, the desired action is squashing.
> * The mid cdclk config should have the new (squash) waveform.
> * - If moving to a lower cdclk, the desired action is crawling.
> * The mid cdclk config should have the new vco.
> */"
>
> Anusha
Ok, lets imagine we are moving to a higher CDCLK, lets say from 480000 to
556800.
As you say above desired action is then squashing.
new_cdclk_config->vco will be then different than current vco, because
it is calculated as dev_priv->display.cdclk.hw.ref * table[i].ratio and
ratios for those CDCLK's are different in mtl_cdclk_table.
However waveforms for those are the same(0x0000 in mtl_cdclk_table) which
means cdclk_compute_crawl_and_squash_midpoint will return false.
Which means we want squash/crawl only. According to your comment above
we want squash only.
Then _bxt_set_cdclk gets called from bxt_set_cdclk here, without any midpoint:
} else {
_bxt_set_cdclk(dev_priv, cdclk_config, pipe);
}
In _bxt_set_cdclk we now at the condition:
if (HAS_CDCLK_CRAWL(dev_priv) && dev_priv->display.cdclk.hw.vco > 0 && vco > 0 &&
!cdclk_pll_is_unknown(dev_priv->display.cdclk.hw.vco)) {
if (dev_priv->display.cdclk.hw.vco != vco)
And _what_ now prevents it not to be always executed?
Both old and new vco are > 0, HAS_CDCLK_CRAWL is true, hw.vco is not ~0 and
also obviously dev_priv->display.cdclk.hw.vco != vco is also true.
Then we get adlp_cdclk_pll_crawl always executed here.
So I think what is happening currently is driven by that mtl_cdclk_table
rather than new cdclk being higher or lower than old cdclk.
adlp_cdclk_pll_crawl will be called whenever vco had changed, regardless
or whether we want squashing or not.
Stan
> > I can see that squash waveform will be returned as 0, if CDCLK is >= 307
> > MHz for MTL for example, or if CDCLK is equal to bypass, however the only
> > way to skip crawling here seems to either have vco == ~0(probably after hw
> > init) or vco == 0, but if vco == 0 we are going to then call
> > icl_cdclk_pll_update which might disable pll, if current hw.vco is not 0.
> >
> > So what am I missing here?
> >
> > Stan
> >
> > >
> > > Anusha
> > > > Stan
> > > >
> > > >
> > > > > > +}
> > > > > > +
> > > > > > static void _bxt_set_cdclk(struct drm_i915_private *dev_priv,
> > > > > > const struct intel_cdclk_config
> > *cdclk_config,
> > > > > > enum pipe pipe)
> > > > > > @@ -1815,9 +1822,12 @@ static void _bxt_set_cdclk(struct
> > > > > > drm_i915_private *dev_priv,
> > > > > > !cdclk_pll_is_unknown(dev_priv->display.cdclk.hw.vco)) {
> > > > > > if (dev_priv->display.cdclk.hw.vco != vco)
> > > > > > adlp_cdclk_pll_crawl(dev_priv, vco);
> > > > > > - } else if (DISPLAY_VER(dev_priv) >= 11)
> > > > > > + } else if (DISPLAY_VER(dev_priv) >= 11) {
> > > > > > + if (pll_enable_wa_needed(dev_priv))
> > > > > > + dg2_cdclk_squash_program(dev_priv, 0);
> > > > > > +
> > > > > > icl_cdclk_pll_update(dev_priv, vco);
> > > > > > - else
> > > > > > + } else
> > > > > > bxt_cdclk_pll_update(dev_priv, vco);
> > > > > >
> > > > > > waveform = cdclk_squash_waveform(dev_priv, cdclk);
> > > > > > --
> > > > > > 2.37.3
> > > > >
More information about the Intel-gfx
mailing list