Re: [PATCH] drm/vkms: Solve bug on kms_crc_cursor tests

From: Ville Syrjälä
Date: Fri Mar 15 2019 - 09:48:40 EST


On Fri, Mar 15, 2019 at 08:51:57AM -0300, Rodrigo Siqueira wrote:
> On 03/11, Ville SyrjÃlà wrote:
> > On Sun, Mar 10, 2019 at 05:35:05PM -0300, Rodrigo Siqueira wrote:
> > > On 03/01, Ville SyrjÃlà wrote:
> > > > On Fri, Mar 01, 2019 at 03:35:35PM -0300, Shayenne Moura wrote:
> > > > > Em sex, 1 de mar de 2019 Ãs 12:26, Ville SyrjÃlÃ
> > > > > <ville.syrjala@xxxxxxxxxxxxxxx> escreveu:
> > > > > >
> > > > > > On Fri, Mar 01, 2019 at 11:55:11AM -0300, Shayenne Moura wrote:
> > > > > > > Em qui, 28 de fev de 2019 Ãs 11:03, Ville SyrjÃlÃ
> > > > > > > <ville.syrjala@xxxxxxxxxxxxxxx> escreveu:
> > > > > > > >
> > > > > > > > On Thu, Feb 28, 2019 at 11:11:07AM +0100, Daniel Vetter wrote:
> > > > > > > > > On Mon, Feb 25, 2019 at 11:26:06AM -0300, Shayenne Moura wrote:
> > > > > > > > > > vkms_crc_work_handle needs the value of the actual frame to
> > > > > > > > > > schedule the workqueue that calls periodically the vblank
> > > > > > > > > > handler and the destroy state functions. However, the frame
> > > > > > > > > > value returned from vkms_vblank_simulate is updated and
> > > > > > > > > > diminished in vblank_get_timestamp because it is not in a
> > > > > > > > > > vblank interrupt, and return an inaccurate value.
> > > > > > > > > >
> > > > > > > > > > Solve this getting the actual vblank frame directly from the
> > > > > > > > > > vblank->count inside the `struct drm_crtc`, instead of using
> > > > > > > > > > the `drm_accurate_vblank_count` function.
> > > > > > > > > >
> > > > > > > > > > Signed-off-by: Shayenne Moura <shayenneluzmoura@xxxxxxxxx>
> > > > > > > > >
> > > > > > > > > Sorry for the delay, I'm a bit swamped right now :-/
> > > > > > > > >
> > > > > > > > > Debug work you're doing here is really impressive! But I have no idea
> > > > > > > > > what's going on. It doesn't look like it's just papering over a bug (like
> > > > > > > > > the in_vblank_irq check we've discussed on irc), but I also have no idea
> > > > > > > > > why it works.
> > > > > > > > >
> > > > > > > > > I'll pull in Ville, he understands this better than me.
> > > > > > > >
> > > > > > > > It's not entirely clear what we're trying to fix. From what I can see
> > > > > > > > the crc work seems to be in no way synchronized with page flips, so
> > > > > > > > I'm not sure how all this is really supposed to work.
> > > > > > > >
> > > > > > >
> > > > > > > Hi, Ville!
> > > > > > >
> > > > > > > Thank you for the review! :)
> > > > > > >
> > > > > > > I do not understand well what crc code is doing, but the issue that I found
> > > > > > > is related to the vblank timestamp and frame count.
> > > > > > >
> > > > > > > When vkms handles the crc_cursor it uses the start frame and end frame
> > > > > > > values to verify if it needs to call the function 'drm_crtc_add_crc_entry()'
> > > > > > > for each frame.
> > > > > > >
> > > > > > > However, when getting the frame count, the code is calling the function
> > > > > > > drm_update_vblank_count(dev, pipe, false) and, because of the 'false',
> > > > > > > subtracting the actual vblank timestamp (consequently, the frame count
> > > > > > > value), causing conflicts.
> > > > > >
> > > > > > The in_vblank_irq behavour looks sane to me. What are these conflicts?
> > > > > >
> > > > >
> > > > > The entire history was:
> > > > > - I sent the patch with bugfix for vblank extra frame. The patch changed
> > > > > our function vkms_get_vblank_timestamp() to look like this:
> > > > >
> > > > > bool vkms_get_vblank_timestamp(struct drm_device *dev, unsigned int pipe,
> > > > > int *max_error, ktime_t *vblank_time,
> > > > > bool in_vblank_irq)
> > > > > {
> > > > > struct vkms_device *vkmsdev = drm_device_to_vkms_device(dev);
> > > > > struct vkms_output *output = &vkmsdev->output;
> > > > >
> > > > > *vblank_time = output->vblank_hrtimer.node.expires;
> > > > >
> > > > > + if (!in_vblank_irq)
> > > > > + *vblank_time -= output->period_ns;
> > > > >
> > > > > return true;
> > > > > }
> > > > >
> > > > > - This patch solve the issue that I was looking for (extra vblank
> > > > > frames on kms_flip).
> > > > >
> > > > > - However, kms_cursor_crc tests, which passed before my patch, started to fail.
> > > > >
> > > > > - Debugging them, I found that the problem was inside of function
> > > > > `vkms_vblank_simulate()`
> > > > > when it was handling the crc_enabled (inside if (state && output->crc_enabled))
> > > > > and inside the function vkms_crc_work_handle() too.
> > > > >
> > > > > - Following the steps:
> > > > > 1. Inside vkms_vblank_simulate() we call drm_crtc_accurate_vblank_count()
> > > > > 2. In its turn, drm_crtc_accurate_vblank_count() calls the function
> > > > > drm_update_vblank_count(dev, pipe, false). /* This false is default */
> > > > > 3. Finally, the âfalseâ used in drm_update_vblank_count(), will be
> > > > > passed to vkms_get_vblank_timestamp() and the condition âif
> > > > > (!in_vblank_irq)â will be executed multiple times (we donât want it).
> > > > >
> > > > > - Inside vkms_crc, the issue is that the returned frame value change for
> > > > > every call of drm_crtc_accurate_vblank_count() because
> > > > > in_vblank_irq == false.
> > >
> > > Hi Ville,
> > >
> > > > OK. So why is it changing? AFAICS it should not change unless the
> > > > timer was moved forward in between the calls.
> > >
> > > Yes Ville, youâre right. We have to update it only when the function
> > > vkms_vblank_simulate() is invoked (timer move forward), and FWIU we do
> > > it when we call drm_crtc_handle_vblank(). However, the current
> > > implementation of vkms, has a call to drm_crtc_accurate_vblank_count()
> > > inside the vkms_vblank_simulate() which is a problem because it also
> > > update the vblank value. FWIU, this patch fixes this issue by taking the
> > > count value in the data struct instead of call
> > > drm_crtc_accurate_vblank_count() which will avoid the extra update.
> >
> > But why does that extra update change the vblank count?
>
> Hi,
>
> I think that Iâm failing to explain the issue and the solution in this
> patch, sorry for that... and I apologize in advance for the lengthy
> email.
>
> In this sense, allow me to elaborate the whole history in âchronological
> sectionsâ:
>
> 1. About the âfakeâ Vblank inside VKMS
>
> Since VKMS is a virtual driver, we do not have real Vblank, and we
> simulate it by using hrtimers. I registered the function
> vkms_vblank_simulate() as a callback inside hrtimer; in this function,
> we do some stuff that can be summarized in the sequence below:
>
> I) drm_crtc_handle_vblank()
> II) Check if crc is enabled or not
> III) Schedule the next vblank with hrtimer_forward_now()
>
> 2. Override the default behaviour for get_vblank_timestamp()
>
> As I told before, everytime that vkms_vblank_simulate() is invoked the
> function drm_crtc_handle_vblank() will be requested. For us, the
> important thing about drm_crtc_handle_vblank() is the sequence:
>
> a) drm_crtc_handle_vblank()
> b) drm_handle_vblank()
> c) drm_update_vblank_count()
> d) drm_get_last_vbltimestamp()
> e) get_vblank_timestamp()
>
> Since VKMS is a virtual driver, we cannot rely on the default
> implementation of get_vblank_timestamp(). In this context, we
> implemented our wrapper named vkms_get_vblank_timestamp(), and we have
> to return the expires time from hrtimer to the userspace. In my first
> implementation of this function, I had something like:
>
> bool vkms_get_vblank_timestamp(struct drm_device *dev, unsigned int pipe,
> int *max_error, ktime_t *vblank_time,
> bool in_vblank_irq)
> {
> [..]
> *vblank_time = output->vblank_hrtimer.node.expires;
> return true;
> }
>
> Around 80% of the kms_flip tests were passing with this code. However,
> we had some tests that break due to an extra vblank.
>
> 3. The extra vblank bug
>
> Shayenne and Daniel figured out that the extra vblank came from the
> function drm_crtc_arm_vblank_event() invoked inside the
> vkms_crtc_atomic_flush(). If we take a look inside
> drm_crtc_arm_vblank_event(), we can find:
>
> void drm_crtc_arm_vblank_event([..])
> {
> [..]
> e->sequence = drm_crtc_accurate_vblank_count(crtc) + 1;
> [..]
> }
>
> This function increment the sequence by one (there is a detailed
> explanation about it in the documentation of this function).
>
> 4. The bug fix
>
> After inspecting the code, we identified the following sequence when the
> function drm_crtc_arm_vblank_event() is invoked:
>
> a) drm_crtc_arm_vblank_event()
> b) drm_crtc_accurate_vblank_count()
> c) drm_update_vblank_count(dev, pipe, false)
>
> Noticed that drm_update_vblank_count() is invoked again, but now with
> âfalseâ in the in_vblank_irq because it does not have to report anything
> for userspace or something like that. With this in mind, remember that
> drm_update_vblank_count() will call get_vblank_timestamp(); with my
> first implementation of vkms_get_vblank_timestamp() (see Section 2 for
> recap) we will update the vblank_time with the wrong information. After
> understanding all of these details and the meaning of the parameter
> in_vblank_irq, the following fix was added:
>
> bool vkms_get_vblank_timestamp([..]) {
> [..]
> if (!in_vblank_irq)
> *vblank_time -= output->period_ns;
> [..]
> }
>
> With this validation, we can solve the problem of the extra Vblank added
> by drm_crtc_arm_vblank_event(). Afterward, all the kms_flip test started
> to pass.
>
> 5. The new bug on CRC operation
>
> As a side effect of the above change, the CRC tests start to failing.
> The question was, why we break the CRC test with that fix?
>
> We start hunting the problem, and the first thing to look is the
> function vkms_vblank_simulate(). Inside this function, we have this
> code:
>
> vkms_vblank_simulate([..]) {
> [..]
> if (state && output->crc_enabled) {
> u64 frame = drm_crtc_accurate_vblank_count(crtc);
> [..]
> }
>
> The output->crc_enabled is a module parameter that enables CRC
> computation which is required for CRC tests, but not for kms_flip.
> Additionally, notice that we save the value of the current frame because
> it is necessary to use it on vkms_crc_work_handle(). The only thing that
> we need to know about vkms_crc_work_handle() is the fact we call
> drm_crtc_add_crc_entry() which requires the frame number, and we also
> use it to check if the frame was updated or not.
>
> As I said in Section 2, drm_crtc_accurate_vblank_count() calls
> drm_update_vblank_count() with false parameter in in_vblank_irq. Due to
> the last change made at vkms_get_vblank_timestamp(), we broke the CRC
> code because now we invoke get_vblank_timestamp() multiple times. The
> consequence of this change is the excessive decrement of the timestamp
> since the following condition execute multiple time:
>
> bool vkms_get_vblank_timestamp([..]) {
> if (!in_vblank_irq)
> *vblank_time -= output->period_ns;

The full code is
*vblank_time = output->vblank_hrtimer.node.expires;
if (!in_vblank_irq)
*vblank_time -= output->period_ns;

So no matter how many times you call it you should always get the same
answer (unless the timer was moved forwared in between the calls).

>
> 6. The bug fix (this patch)
>
> Since we only want the frame number, as I described in Section 5, the
> fix is quite simple: avoid call drm_crtc_accurate_vblank_count() and
> just get the frame value from the data structure. In this patch,
> Shayenne removed the function drm_crtc_accurate_vblank_count() which
> avoid the extra decrement in the vkms_get_vblank_timestamp(). With this
> patch, everything started to work well.
>
> So... In a few words, it is a VKMS problem, not a vblank issue.
>
> Iâm not sure if the solution here is the best one, but I believe that
> the idea behind it is correct.
>
> Best Regards
>
> --
> Rodrigo Siqueira
> https://siqueira.tech
> Graduate Student
> Department of Computer Science
> University of SÃo Paulo



--
Ville SyrjÃlÃ
Intel