diff mbox

[v2] drm/i915: Wait for PSR exit before checking for vblank evasion

Message ID 20180514204922.82073-3-tarun.vyas@intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Tarun Vyas May 14, 2018, 8:49 p.m. UTC
The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited, then
the pipe_update_start call schedules itself out to check back later.

On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915 but
lags w.r.t core kernel code, hot plugging an external display triggers
tons of "potential atomic update errors" in the dmesg, on *pipe A*. A
closer analysis reveals that we try to read the scanline 3 times and
eventually timeout, b/c PSR hasn't exited fully leading to a PIPEDSL
stuck @ 1599. This issue is not seen on upstream kernels, b/c for *some*
reason we loop inside intel_pipe_update start for ~2+ msec which in this
case is more than enough to exit PSR fully, hence an *unstuck* PIPEDSL
counter, hence no error. On the other hand, the ChromeOS kernel spends
~1.1 msec looping inside intel_pipe_update_start and hence errors out
b/c the source is still in PSR.

Regardless, we should wait for PSR exit (if PSR is supported and active
on the current pipe) before reading the PIPEDSL, b/c if we haven't
fully exited PSR, then checking for vblank evasion isn't actually
applicable.

This scenario applies to a configuration with an additional pipe,
as of now

Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
---
 drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

Comments

Chris Wilson May 14, 2018, 9:16 p.m. UTC | #1
Quoting Tarun Vyas (2018-05-14 21:49:22)
> The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited, then
> the pipe_update_start call schedules itself out to check back later.
> 
> On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915 but
> lags w.r.t core kernel code, hot plugging an external display triggers
> tons of "potential atomic update errors" in the dmesg, on *pipe A*. A
> closer analysis reveals that we try to read the scanline 3 times and
> eventually timeout, b/c PSR hasn't exited fully leading to a PIPEDSL
> stuck @ 1599. This issue is not seen on upstream kernels, b/c for *some*
> reason we loop inside intel_pipe_update start for ~2+ msec which in this
> case is more than enough to exit PSR fully, hence an *unstuck* PIPEDSL
> counter, hence no error. On the other hand, the ChromeOS kernel spends
> ~1.1 msec looping inside intel_pipe_update_start and hence errors out
> b/c the source is still in PSR.
> 
> Regardless, we should wait for PSR exit (if PSR is supported and active
> on the current pipe) before reading the PIPEDSL, b/c if we haven't
> fully exited PSR, then checking for vblank evasion isn't actually
> applicable.
> 
> This scenario applies to a configuration with an additional pipe,
> as of now
> 
> Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> ---
>  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_sprite.c b/drivers/gpu/drm/i915/intel_sprite.c
> index ee23613f9fd4..481d310e5c3b 100644
> --- a/drivers/gpu/drm/i915/intel_sprite.c
> +++ b/drivers/gpu/drm/i915/intel_sprite.c
> @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct intel_crtc_state *new_crtc_state)
>                                                       VBLANK_EVASION_TIME_US);
>         max = vblank_start - 1;
>  
> -       local_irq_disable();
> -
>         if (min <= 0 || max <= 0)
>                 return;
>  
>         if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
>                 return;
>  
> +       if(new_crtc_state->has_psr && dev_priv->psr.active)
> +               psr_wait_for_idle(dev_priv);
> +
> +       local_irq_disable();

Pop quiz, does intel_pipe_update_finish() unconditionally assume it is
called with irqs disabled?
-Chris
kernel test robot May 15, 2018, 3:16 a.m. UTC | #2
Hi Tarun,

Thank you for the patch! Yet something to improve:

[auto build test ERROR on drm-intel/for-linux-next]
[also build test ERROR on v4.17-rc5 next-20180514]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]

url:    https://github.com/0day-ci/linux/commits/Tarun-Vyas/drm-i915-Wait-for-PSR-exit-before-checking-for-vblank-evasion/20180515-103355
base:   git://anongit.freedesktop.org/drm-intel for-linux-next
config: i386-randconfig-x000-201819 (attached as .config)
compiler: gcc-7 (Debian 7.3.0-16) 7.3.0
reproduce:
        # save the attached .config to linux build tree
        make ARCH=i386 

All errors (new ones prefixed by >>):

   drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_pipe_update_start':
>> drivers/gpu/drm/i915/intel_sprite.c:117:3: error: implicit declaration of function 'psr_wait_for_idle'; did you mean 'apic_wait_icr_idle'? [-Werror=implicit-function-declaration]
      psr_wait_for_idle(dev_priv);
      ^~~~~~~~~~~~~~~~~
      apic_wait_icr_idle
   cc1: some warnings being treated as errors

vim +117 drivers/gpu/drm/i915/intel_sprite.c

    76	
    77	/**
    78	 * intel_pipe_update_start() - start update of a set of display registers
    79	 * @new_crtc_state: the new crtc state
    80	 *
    81	 * Mark the start of an update to pipe registers that should be updated
    82	 * atomically regarding vblank. If the next vblank will happens within
    83	 * the next 100 us, this function waits until the vblank passes.
    84	 *
    85	 * After a successful call to this function, interrupts will be disabled
    86	 * until a subsequent call to intel_pipe_update_end(). That is done to
    87	 * avoid random delays.
    88	 */
    89	void intel_pipe_update_start(const struct intel_crtc_state *new_crtc_state)
    90	{
    91		struct intel_crtc *crtc = to_intel_crtc(new_crtc_state->base.crtc);
    92		struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
    93		const struct drm_display_mode *adjusted_mode = &new_crtc_state->base.adjusted_mode;
    94		long timeout = msecs_to_jiffies_timeout(1);
    95		int scanline, min, max, vblank_start;
    96		wait_queue_head_t *wq = drm_crtc_vblank_waitqueue(&crtc->base);
    97		bool need_vlv_dsi_wa = (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) &&
    98			intel_crtc_has_type(new_crtc_state, INTEL_OUTPUT_DSI);
    99		DEFINE_WAIT(wait);
   100	
   101		vblank_start = adjusted_mode->crtc_vblank_start;
   102		if (adjusted_mode->flags & DRM_MODE_FLAG_INTERLACE)
   103			vblank_start = DIV_ROUND_UP(vblank_start, 2);
   104	
   105		/* FIXME needs to be calibrated sensibly */
   106		min = vblank_start - intel_usecs_to_scanlines(adjusted_mode,
   107							      VBLANK_EVASION_TIME_US);
   108		max = vblank_start - 1;
   109	
   110		if (min <= 0 || max <= 0)
   111			return;
   112	
   113		if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
   114			return;
   115	
   116		if(new_crtc_state->has_psr && dev_priv->psr.active)
 > 117			psr_wait_for_idle(dev_priv);
   118	
   119		local_irq_disable();
   120	
   121		crtc->debug.min_vbl = min;
   122		crtc->debug.max_vbl = max;
   123		trace_i915_pipe_update_start(crtc);
   124	
   125		for (;;) {
   126			/*
   127			 * prepare_to_wait() has a memory barrier, which guarantees
   128			 * other CPUs can see the task state update by the time we
   129			 * read the scanline.
   130			 */
   131			prepare_to_wait(wq, &wait, TASK_UNINTERRUPTIBLE);
   132	
   133			scanline = intel_get_crtc_scanline(crtc);
   134			if (scanline < min || scanline > max)
   135				break;
   136	
   137			if (!timeout) {
   138				DRM_ERROR("Potential atomic update failure on pipe %c\n",
   139					  pipe_name(crtc->pipe));
   140				break;
   141			}
   142	
   143			local_irq_enable();
   144	
   145			timeout = schedule_timeout(timeout);
   146	
   147			local_irq_disable();
   148		}
   149	
   150		finish_wait(wq, &wait);
   151	
   152		drm_crtc_vblank_put(&crtc->base);
   153	
   154		/*
   155		 * On VLV/CHV DSI the scanline counter would appear to
   156		 * increment approx. 1/3 of a scanline before start of vblank.
   157		 * The registers still get latched at start of vblank however.
   158		 * This means we must not write any registers on the first
   159		 * line of vblank (since not the whole line is actually in
   160		 * vblank). And unfortunately we can't use the interrupt to
   161		 * wait here since it will fire too soon. We could use the
   162		 * frame start interrupt instead since it will fire after the
   163		 * critical scanline, but that would require more changes
   164		 * in the interrupt code. So for now we'll just do the nasty
   165		 * thing and poll for the bad scanline to pass us by.
   166		 *
   167		 * FIXME figure out if BXT+ DSI suffers from this as well
   168		 */
   169		while (need_vlv_dsi_wa && scanline == vblank_start)
   170			scanline = intel_get_crtc_scanline(crtc);
   171	
   172		crtc->debug.scanline_start = scanline;
   173		crtc->debug.start_vbl_time = ktime_get();
   174		crtc->debug.start_vbl_count = intel_crtc_get_vblank_counter(crtc);
   175	
   176		trace_i915_pipe_update_vblank_evaded(crtc);
   177	}
   178	

---
0-DAY kernel test infrastructure                Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all                   Intel Corporation
Tarun Vyas May 15, 2018, 7:38 p.m. UTC | #3
On Mon, May 14, 2018 at 10:16:38PM +0100, Chris Wilson wrote:
> Quoting Tarun Vyas (2018-05-14 21:49:22)
> > The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited, then
> > the pipe_update_start call schedules itself out to check back later.
> > 
> > On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915 but
> > lags w.r.t core kernel code, hot plugging an external display triggers
> > tons of "potential atomic update errors" in the dmesg, on *pipe A*. A
> > closer analysis reveals that we try to read the scanline 3 times and
> > eventually timeout, b/c PSR hasn't exited fully leading to a PIPEDSL
> > stuck @ 1599. This issue is not seen on upstream kernels, b/c for *some*
> > reason we loop inside intel_pipe_update start for ~2+ msec which in this
> > case is more than enough to exit PSR fully, hence an *unstuck* PIPEDSL
> > counter, hence no error. On the other hand, the ChromeOS kernel spends
> > ~1.1 msec looping inside intel_pipe_update_start and hence errors out
> > b/c the source is still in PSR.
> > 
> > Regardless, we should wait for PSR exit (if PSR is supported and active
> > on the current pipe) before reading the PIPEDSL, b/c if we haven't
> > fully exited PSR, then checking for vblank evasion isn't actually
> > applicable.
> > 
> > This scenario applies to a configuration with an additional pipe,
> > as of now
> > 
> > Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> > ---
> >  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/intel_sprite.c b/drivers/gpu/drm/i915/intel_sprite.c
> > index ee23613f9fd4..481d310e5c3b 100644
> > --- a/drivers/gpu/drm/i915/intel_sprite.c
> > +++ b/drivers/gpu/drm/i915/intel_sprite.c
> > @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct intel_crtc_state *new_crtc_state)
> >                                                       VBLANK_EVASION_TIME_US);
> >         max = vblank_start - 1;
> >  
> > -       local_irq_disable();
> > -
> >         if (min <= 0 || max <= 0)
> >                 return;
> >  
> >         if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
> >                 return;
> >  
> > +       if(new_crtc_state->has_psr && dev_priv->psr.active)
> > +               psr_wait_for_idle(dev_priv);
> > +
> > +       local_irq_disable();
> 
> Pop quiz, does intel_pipe_update_finish() unconditionally assume it is
> called with irqs disabled?
> -Chris
Unless local_irq_disable() fails, intel_pipe_update_end() should always get called with irqs disabled, from what it looks like to me.
Dhinakaran Pandiyan June 19, 2018, 9:27 p.m. UTC | #4
On Mon, 2018-05-14 at 13:49 -0700, Tarun Vyas wrote:
> The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited, then
> the pipe_update_start call schedules itself out to check back later.
> 
> On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915 but
> lags w.r.t core kernel code, hot plugging an external display
> triggers
> tons of "potential atomic update errors" in the dmesg, on *pipe A*. A
> closer analysis reveals that we try to read the scanline 3 times and
> eventually timeout, b/c PSR hasn't exited fully leading to a PIPEDSL
> stuck @ 1599. This issue is not seen on upstream kernels, b/c for
> *some*
> reason we loop inside intel_pipe_update start for ~2+ msec which in
> this
> case is more than enough to exit PSR fully, hence an *unstuck*
> PIPEDSL
> counter, hence no error. On the other hand, the ChromeOS kernel
> spends
> ~1.1 msec looping inside intel_pipe_update_start and hence errors out
> b/c the source is still in PSR.
> 
> Regardless, we should wait for PSR exit (if PSR is supported and
> active
> on the current pipe) before reading the PIPEDSL, b/c if we haven't
> fully exited PSR, then checking for vblank evasion isn't actually
> applicable.
> 
> This scenario applies to a configuration with an additional pipe,
> as of now
> 
> Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> ---
>  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_sprite.c
> b/drivers/gpu/drm/i915/intel_sprite.c
> index ee23613f9fd4..481d310e5c3b 100644
> --- a/drivers/gpu/drm/i915/intel_sprite.c
> +++ b/drivers/gpu/drm/i915/intel_sprite.c
> @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct
> intel_crtc_state *new_crtc_state)
>  						      VBLANK_EVASION
> _TIME_US);
>  	max = vblank_start - 1;
>  
> -	local_irq_disable();
> -
>  	if (min <= 0 || max <= 0)
>  		return;
>  
>  	if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
>  		return;
>  
> +	if(new_crtc_state->has_psr && dev_priv->psr.active)
> +		psr_wait_for_idle(dev_priv);

How about just waiting for PSR_STATUS to idle without grabbing any
locks or checking whether PSR is active?

Status should be idle if PSR was disabled or on it's way to becoming
idle if it was enabled. Even if PSR did get enabled while we are in
pipe_update_start(), it will not be active as long as VBIs are enabled.



> +
> +	local_irq_disable();
> +
>  	crtc->debug.min_vbl = min;
>  	crtc->debug.max_vbl = max;
>  	trace_i915_pipe_update_start(crtc);
Dhinakaran Pandiyan June 19, 2018, 9:54 p.m. UTC | #5
On Tue, 2018-06-19 at 14:27 -0700, Dhinakaran Pandiyan wrote:
> On Mon, 2018-05-14 at 13:49 -0700, Tarun Vyas wrote:
> > 
> > The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited,
> > then
> > the pipe_update_start call schedules itself out to check back
> > later.
> > 
> > On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915
> > but
> > lags w.r.t core kernel code, hot plugging an external display
> > triggers
> > tons of "potential atomic update errors" in the dmesg, on *pipe A*.
> > A
> > closer analysis reveals that we try to read the scanline 3 times
> > and
> > eventually timeout, b/c PSR hasn't exited fully leading to a
> > PIPEDSL
> > stuck @ 1599. This issue is not seen on upstream kernels, b/c for
> > *some*
> > reason we loop inside intel_pipe_update start for ~2+ msec which in
> > this
> > case is more than enough to exit PSR fully, hence an *unstuck*
> > PIPEDSL
> > counter, hence no error. On the other hand, the ChromeOS kernel
> > spends
> > ~1.1 msec looping inside intel_pipe_update_start and hence errors
> > out
> > b/c the source is still in PSR.
> > 
> > Regardless, we should wait for PSR exit (if PSR is supported and
> > active
> > on the current pipe) before reading the PIPEDSL, b/c if we haven't
> > fully exited PSR, then checking for vblank evasion isn't actually
> > applicable.
> > 
> > This scenario applies to a configuration with an additional pipe,
> > as of now
> > 
> > Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> > ---
> >  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/intel_sprite.c
> > b/drivers/gpu/drm/i915/intel_sprite.c
> > index ee23613f9fd4..481d310e5c3b 100644
> > --- a/drivers/gpu/drm/i915/intel_sprite.c
> > +++ b/drivers/gpu/drm/i915/intel_sprite.c
> > @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct
> > intel_crtc_state *new_crtc_state)
> >  						      VBLANK_EVASI
> > ON
> > _TIME_US);
> >  	max = vblank_start - 1;
> >  
> > -	local_irq_disable();
> > -
> >  	if (min <= 0 || max <= 0)
> >  		return;
> >  
> >  	if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
> >  		return;
> >  
> > +	if(new_crtc_state->has_psr && dev_priv->psr.active)
> > +		psr_wait_for_idle(dev_priv);
> How about just waiting for PSR_STATUS to idle without grabbing any
> locks or checking whether PSR is active?
> 
> Status should be idle if PSR was disabled or on it's way to becoming
> idle if it was enabled. Even if PSR did get enabled while we are in
> pipe_update_start(), it will not be active as long as VBIs are
> enabled.
> 
Correct me if this was already considered, why not wait until the
scanline counter starts moving? I see we have a 
	intel_wait_for_pipe_scanline_moving(crtc) that's used when the
pipe is enabled.

-DK
Tarun Vyas June 19, 2018, 9:59 p.m. UTC | #6
On Tue, Jun 19, 2018 at 02:54:07PM -0700, Dhinakaran Pandiyan wrote:
> On Tue, 2018-06-19 at 14:27 -0700, Dhinakaran Pandiyan wrote:
> > On Mon, 2018-05-14 at 13:49 -0700, Tarun Vyas wrote:
> > > 
> > > The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited,
> > > then
> > > the pipe_update_start call schedules itself out to check back
> > > later.
> > > 
> > > On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915
> > > but
> > > lags w.r.t core kernel code, hot plugging an external display
> > > triggers
> > > tons of "potential atomic update errors" in the dmesg, on *pipe A*.
> > > A
> > > closer analysis reveals that we try to read the scanline 3 times
> > > and
> > > eventually timeout, b/c PSR hasn't exited fully leading to a
> > > PIPEDSL
> > > stuck @ 1599. This issue is not seen on upstream kernels, b/c for
> > > *some*
> > > reason we loop inside intel_pipe_update start for ~2+ msec which in
> > > this
> > > case is more than enough to exit PSR fully, hence an *unstuck*
> > > PIPEDSL
> > > counter, hence no error. On the other hand, the ChromeOS kernel
> > > spends
> > > ~1.1 msec looping inside intel_pipe_update_start and hence errors
> > > out
> > > b/c the source is still in PSR.
> > > 
> > > Regardless, we should wait for PSR exit (if PSR is supported and
> > > active
> > > on the current pipe) before reading the PIPEDSL, b/c if we haven't
> > > fully exited PSR, then checking for vblank evasion isn't actually
> > > applicable.
> > > 
> > > This scenario applies to a configuration with an additional pipe,
> > > as of now
> > > 
> > > Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> > > ---
> > >  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
> > >  1 file changed, 5 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/intel_sprite.c
> > > b/drivers/gpu/drm/i915/intel_sprite.c
> > > index ee23613f9fd4..481d310e5c3b 100644
> > > --- a/drivers/gpu/drm/i915/intel_sprite.c
> > > +++ b/drivers/gpu/drm/i915/intel_sprite.c
> > > @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct
> > > intel_crtc_state *new_crtc_state)
> > >  						      VBLANK_EVASI
> > > ON
> > > _TIME_US);
> > >  	max = vblank_start - 1;
> > >  
> > > -	local_irq_disable();
> > > -
> > >  	if (min <= 0 || max <= 0)
> > >  		return;
> > >  
> > >  	if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
> > >  		return;
> > >  
> > > +	if(new_crtc_state->has_psr && dev_priv->psr.active)
> > > +		psr_wait_for_idle(dev_priv);
> > How about just waiting for PSR_STATUS to idle without grabbing any
> > locks or checking whether PSR is active?
> > 
> > Status should be idle if PSR was disabled or on it's way to becoming
> > idle if it was enabled. Even if PSR did get enabled while we are in
> > pipe_update_start(), it will not be active as long as VBIs are
> > enabled.
> > 
Right, if we are OK with some duplication (of psr_wait_for_idle) inside intel_psr.c, then we can duplicate the PSR2 vs. PSR check that's being done in psr_wait_for_idle and then just wait without grabbing any locks, so essentially a lockless version of psr_wait_for_idle()
> Correct me if this was already considered, why not wait until the
> scanline counter starts moving? I see we have a 
> 	intel_wait_for_pipe_scanline_moving(crtc) that's used when the
> pipe is enabled.
> 
> -DK

Didn't consider this before, but, pipe_scanline_is_moving waits for a minimum of 5 msec. Are we OK with a min wait of 5 msec inside pipe_update_start ? Heuristically, waiting for PSR idle has almost always returned within < 2 msec. Occasionally it takes upto 1 full frame.
Dhinakaran Pandiyan June 19, 2018, 11:11 p.m. UTC | #7
On Tue, 2018-06-19 at 14:59 -0700, Tarun Vyas wrote:
> On Tue, Jun 19, 2018 at 02:54:07PM -0700, Dhinakaran Pandiyan wrote:
> > 
> > On Tue, 2018-06-19 at 14:27 -0700, Dhinakaran Pandiyan wrote:
> > > 
> > > On Mon, 2018-05-14 at 13:49 -0700, Tarun Vyas wrote:
> > > > 
> > > > 
> > > > The PIPEDSL freezes on PSR entry and if PSR hasn't fully
> > > > exited,
> > > > then
> > > > the pipe_update_start call schedules itself out to check back
> > > > later.
> > > > 
> > > > On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t
> > > > drm/i915
> > > > but
> > > > lags w.r.t core kernel code, hot plugging an external display
> > > > triggers
> > > > tons of "potential atomic update errors" in the dmesg, on *pipe
> > > > A*.
> > > > A
> > > > closer analysis reveals that we try to read the scanline 3
> > > > times
> > > > and
> > > > eventually timeout, b/c PSR hasn't exited fully leading to a
> > > > PIPEDSL
> > > > stuck @ 1599. This issue is not seen on upstream kernels, b/c
> > > > for
> > > > *some*
> > > > reason we loop inside intel_pipe_update start for ~2+ msec
> > > > which in
> > > > this
> > > > case is more than enough to exit PSR fully, hence an *unstuck*
> > > > PIPEDSL
> > > > counter, hence no error. On the other hand, the ChromeOS kernel
> > > > spends
> > > > ~1.1 msec looping inside intel_pipe_update_start and hence
> > > > errors
> > > > out
> > > > b/c the source is still in PSR.
> > > > 
> > > > Regardless, we should wait for PSR exit (if PSR is supported
> > > > and
> > > > active
> > > > on the current pipe) before reading the PIPEDSL, b/c if we
> > > > haven't
> > > > fully exited PSR, then checking for vblank evasion isn't
> > > > actually
> > > > applicable.
> > > > 
> > > > This scenario applies to a configuration with an additional
> > > > pipe,
> > > > as of now
> > > > 
> > > > Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> > > > ---
> > > >  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
> > > >  1 file changed, 5 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/i915/intel_sprite.c
> > > > b/drivers/gpu/drm/i915/intel_sprite.c
> > > > index ee23613f9fd4..481d310e5c3b 100644
> > > > --- a/drivers/gpu/drm/i915/intel_sprite.c
> > > > +++ b/drivers/gpu/drm/i915/intel_sprite.c
> > > > @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct
> > > > intel_crtc_state *new_crtc_state)
> > > >  						      VBLANK_E
> > > > VASI
> > > > ON
> > > > _TIME_US);
> > > >  	max = vblank_start - 1;
> > > >  
> > > > -	local_irq_disable();
> > > > -
> > > >  	if (min <= 0 || max <= 0)
> > > >  		return;
> > > >  
> > > >  	if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
> > > >  		return;
> > > >  
> > > > +	if(new_crtc_state->has_psr && dev_priv->psr.active)
> > > > +		psr_wait_for_idle(dev_priv);
> > > How about just waiting for PSR_STATUS to idle without grabbing
> > > any
> > > locks or checking whether PSR is active?
> > > 
> > > Status should be idle if PSR was disabled or on it's way to
> > > becoming
> > > idle if it was enabled. Even if PSR did get enabled while we are
> > > in
> > > pipe_update_start(), it will not be active as long as VBIs are
> > > enabled.
> > > 
> Right, if we are OK with some duplication (of psr_wait_for_idle)
> inside intel_psr.c, then we can duplicate the PSR2 vs. PSR check
> that's being done in psr_wait_for_idle and then just wait without
> grabbing any locks, so essentially a lockless version of
> psr_wait_for_idle()

Yeah, you can extract the wait into psr_wait_for_idle_locked() 

> > 
> > Correct me if this was already considered, why not wait until the
> > scanline counter starts moving? I see we have a 
> > 	intel_wait_for_pipe_scanline_moving(crtc) that's used when the
> > pipe is enabled.
> > 
> > -DK
> Didn't consider this before, but, pipe_scanline_is_moving waits for a
> minimum of 5 msec. Are we OK with a min wait of 5 msec inside
> pipe_update_start ? Heuristically, waiting for PSR idle has almost
> always returned within < 2 msec. Occasionally it takes upto 1 full
> frame.

We should be able to change that function.
Tarun Vyas June 21, 2018, 10:43 p.m. UTC | #8
On Tue, Jun 19, 2018 at 02:59:54PM -0700, Tarun Vyas wrote:
> On Tue, Jun 19, 2018 at 02:54:07PM -0700, Dhinakaran Pandiyan wrote:
> > On Tue, 2018-06-19 at 14:27 -0700, Dhinakaran Pandiyan wrote:
> > > On Mon, 2018-05-14 at 13:49 -0700, Tarun Vyas wrote:
> > > > 
> > > > The PIPEDSL freezes on PSR entry and if PSR hasn't fully exited,
> > > > then
> > > > the pipe_update_start call schedules itself out to check back
> > > > later.
> > > > 
> > > > On ChromeOS-4.4 kernel, which is fairly up-to-date w.r.t drm/i915
> > > > but
> > > > lags w.r.t core kernel code, hot plugging an external display
> > > > triggers
> > > > tons of "potential atomic update errors" in the dmesg, on *pipe A*.
> > > > A
> > > > closer analysis reveals that we try to read the scanline 3 times
> > > > and
> > > > eventually timeout, b/c PSR hasn't exited fully leading to a
> > > > PIPEDSL
> > > > stuck @ 1599. This issue is not seen on upstream kernels, b/c for
> > > > *some*
> > > > reason we loop inside intel_pipe_update start for ~2+ msec which in
> > > > this
> > > > case is more than enough to exit PSR fully, hence an *unstuck*
> > > > PIPEDSL
> > > > counter, hence no error. On the other hand, the ChromeOS kernel
> > > > spends
> > > > ~1.1 msec looping inside intel_pipe_update_start and hence errors
> > > > out
> > > > b/c the source is still in PSR.
> > > > 
> > > > Regardless, we should wait for PSR exit (if PSR is supported and
> > > > active
> > > > on the current pipe) before reading the PIPEDSL, b/c if we haven't
> > > > fully exited PSR, then checking for vblank evasion isn't actually
> > > > applicable.
> > > > 
> > > > This scenario applies to a configuration with an additional pipe,
> > > > as of now
> > > > 
> > > > Signed-off-by: Tarun Vyas <tarun.vyas@intel.com>
> > > > ---
> > > >  drivers/gpu/drm/i915/intel_sprite.c | 7 +++++--
> > > >  1 file changed, 5 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/i915/intel_sprite.c
> > > > b/drivers/gpu/drm/i915/intel_sprite.c
> > > > index ee23613f9fd4..481d310e5c3b 100644
> > > > --- a/drivers/gpu/drm/i915/intel_sprite.c
> > > > +++ b/drivers/gpu/drm/i915/intel_sprite.c
> > > > @@ -107,14 +107,17 @@ void intel_pipe_update_start(const struct
> > > > intel_crtc_state *new_crtc_state)
> > > >  						      VBLANK_EVASI
> > > > ON
> > > > _TIME_US);
> > > >  	max = vblank_start - 1;
> > > >  
> > > > -	local_irq_disable();
> > > > -
> > > >  	if (min <= 0 || max <= 0)
> > > >  		return;
> > > >  
> > > >  	if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
> > > >  		return;
> > > >  
> > > > +	if(new_crtc_state->has_psr && dev_priv->psr.active)
> > > > +		psr_wait_for_idle(dev_priv);
> > > How about just waiting for PSR_STATUS to idle without grabbing any
> > > locks or checking whether PSR is active?
> > > 
> > > Status should be idle if PSR was disabled or on it's way to becoming
> > > idle if it was enabled. Even if PSR did get enabled while we are in
> > > pipe_update_start(), it will not be active as long as VBIs are
> > > enabled.
> > > 
> Right, if we are OK with some duplication (of psr_wait_for_idle) inside intel_psr.c, then we can duplicate the PSR2 vs. PSR check that's being done in psr_wait_for_idle and then just wait without grabbing any locks, so essentially a lockless version of psr_wait_for_idle()
> > Correct me if this was already considered, why not wait until the
> > scanline counter starts moving? I see we have a 
> > 	intel_wait_for_pipe_scanline_moving(crtc) that's used when the
> > pipe is enabled.
> > 
> > -DK
> 
> Didn't consider this before, but, pipe_scanline_is_moving waits for a minimum of 5 msec. Are we OK with a min wait of 5 msec inside pipe_update_start ? Heuristically, waiting for PSR idle has almost always returned within < 2 msec. Occasionally it takes upto 1 full frame.
As per some preliminary measurements
Approach 1:
	Wait *unconditionally* (so no need to check for PSR enabled/disabled and hence no locks) for PSR_STATUS to IDLE out. 
This takes ~7msec when PSR is active and ~2 usec when PSR is inactive/disabled.

Approach 2:
	Use intel_wait_for_pipe_scanline_moving to wait for PIPEDSL to start moving after PSR exit. Currently, this ends up waiting for a minimum of 5 msec but I changed this to accept a caller defined value for the delay.
After the above changes, this approach takes ~7msec when PSR is active and ~25-40 usec with PSR disabled, b/c we still need to check for at least 10+ usec and see if PIPEDSL moved, if it did, we wait for longer, otherwise we move on.
diff mbox

Patch

diff --git a/drivers/gpu/drm/i915/intel_sprite.c b/drivers/gpu/drm/i915/intel_sprite.c
index ee23613f9fd4..481d310e5c3b 100644
--- a/drivers/gpu/drm/i915/intel_sprite.c
+++ b/drivers/gpu/drm/i915/intel_sprite.c
@@ -107,14 +107,17 @@  void intel_pipe_update_start(const struct intel_crtc_state *new_crtc_state)
 						      VBLANK_EVASION_TIME_US);
 	max = vblank_start - 1;
 
-	local_irq_disable();
-
 	if (min <= 0 || max <= 0)
 		return;
 
 	if (WARN_ON(drm_crtc_vblank_get(&crtc->base)))
 		return;
 
+	if(new_crtc_state->has_psr && dev_priv->psr.active)
+		psr_wait_for_idle(dev_priv);
+
+	local_irq_disable();
+
 	crtc->debug.min_vbl = min;
 	crtc->debug.max_vbl = max;
 	trace_i915_pipe_update_start(crtc);