Message-ID: <fc5a927251d03e05326ecce8f4dafff06729c50c.camel@redhat.com>
Date:   Thu, 07 May 2020 14:57:10 -0400
From:   Lyude Paul <lyude@...hat.com>
To:     Daniel Vetter <daniel@...ll.ch>, Tejun Heo <tj@...nel.org>
Cc:     nouveau@...ts.freedesktop.org, dri-devel@...ts.freedesktop.org,
        Ville Syrjälä 
        <ville.syrjala@...ux.intel.com>,
        Maarten Lankhorst <maarten.lankhorst@...ux.intel.com>,
        Maxime Ripard <mripard@...nel.org>,
        Thomas Zimmermann <tzimmermann@...e.de>,
        David Airlie <airlied@...ux.ie>, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/9] drm/vblank: Add vblank works

Hey guys! Sorry this took me a little while to get to, but I was finally able
to sit down for a bit and do a thorough investigation on the latency issues
here to figure out if it's noise or not. I did this investigation with the
plain work_struct implementation, the original kthread implementation, and a
version of the kthread_worker based implementation (rewritten since the last
series I sent out, so it isn't as tightly integrated with kthread_worker as
before). Also note, this was a pretty big refresher on my stats intro back
from when I was in college, so hopefully I didn't make any mistakes :).

First, I should probably point out that I discovered the igt-gpu-tools tests
we were using for testing this were actually giving us false positives before.
Basically, we made the mistake of exiting the test early if we exceeded the
specified duration (by default, 2 seconds) even if we never actually tried any
LUT updates, which was causing this test to pass on my machine when it really
shouldn't have. I managed to fix this, and finally managed to actually
reproduce the test failures Ville was seeing by running the test in parallel
with `stress -c 1 -i 1 -b 1` (starts 1 CPU hog, 1 sync() hog, 1
write()/unlink() hog). I also made sure to stop any systemd timers and run igt
with ionice -c realtime nice -n -20, just to make sure igt's userspace process
was affected by the load as little as possible to reduce noise in the results.

But I quickly discovered the test still was far too noisy, as even after 500
test runs with each implementation I couldn't see any obvious difference in
performance between any of them. So I added a --benchmark option to the test,
which makes the test run for the entirity of the --duration value (regardless
of how many test failures we hit) and report the total number of frames
tested, along with how many of those frames had LUT tearing. I then used this
to benchmark each implementation and started putting all of the failure
percentages in a spreadsheet. I did this with a duration of 20 seconds, which
would amount to testing ~1200-1210 frames each time. Then, I kept increasing
my sample size by rerunning the benchmarks until it looked like the standard
deviation of my data set was starting to stabilize, e.g. it was no longer
fluctuating by more then ~.10% when I added more data.

Finally-this actually managed to give me some data that looked somewhat
sensible:

https://people.freedesktop.org/~lyudess/archive/05-07-2020/vbl-work-benchmarks.pdf

(remember-each value is the percentage of frames where LUT tearing was
detected)

The verdict appears to be that the new kthread_worker implementation I have
looks like it's ever so slightly faster on average then the original kthread
based implementation when the system is under heavy load. 

Also, you can find the branches I used here:

https://gitlab.freedesktop.org/lyudess/igt-gpu-tools/-/commits/wip/lut-tearing-2
https://gitlab.freedesktop.org/lyudess/linux/-/commits/wip/vbl-workers-vlv-v3 (kthread_worker v2)
https://gitlab.freedesktop.org/lyudess/linux/-/commits/wip/vbl-workers-vlv-kthread (kthread)
https://gitlab.freedesktop.org/lyudess/linux/-/commits/wip/vbl-workers-vlv-v2 (workqueues)

All tests were done on my single core valleyview machine, with one 1080p HDMI
display and one 1080p VGA display.

IMHO: Unless anyone has objections, this seems like enough evidence for me, so
I'll send out a respin of the nouveau-crc series with the new respin of the
kthread_worker based vblank works in a bit.

On Wed, 2020-03-18 at 14:46 +0100, Daniel Vetter wrote:
> On Tue, Mar 17, 2020 at 08:40:58PM -0400, Lyude Paul wrote:
> > From: Ville Syrjälä <ville.syrjala@...ux.intel.com>
> > 
> > Add some kind of vblank workers. The interface is similar to regular
> > delayed works, and also allows for re-scheduling.
> > 
> > Whatever hardware programming we do in the work must be fast
> > (must at least complete during the vblank, sometimes during
> > the first few scanlines of vblank), so we'll fire up a per-crtc
> > high priority thread for this.
> > 
> > [based off patches from Ville Syrjälä <ville.syrjala@...ux.intel.com>,
> > change below to signoff later]
> > 
> > Cc: Ville Syrjälä <ville.syrjala@...ux.intel.com>
> > Signed-off-by: Lyude Paul <lyude@...hat.com>
> 
> Hm not really sold on the idea that we have should reinvent our own worker
> infrastructure here. Imo a vblank_work should look like a delayed work,
> i.e. using struct work_struct as the base class, and wrapping the vblank
> thing around it (instead of the timer). That alos would allow drivers to
> schedule works on their own work queues, allowing for easier flushing and
> all that stuff.
> 
> Also if we do this I think we should try to follow the delayed work abi as
> closely as possible (e.g. INIT_VBLANK_WORK, queue_vblank_work,
> mod_vblank_work, ...). Delayed workers (whether timer or vblank) have a
> bunch of edges cases where consistently would be really great to avoid
> surprises and bugs.
> -Daniel
> 
> > ---
> >  drivers/gpu/drm/drm_vblank.c | 322 +++++++++++++++++++++++++++++++++++
> >  include/drm/drm_vblank.h     |  34 ++++
> >  2 files changed, 356 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/drm_vblank.c b/drivers/gpu/drm/drm_vblank.c
> > index da7b0b0c1090..06c796b6c381 100644
> > --- a/drivers/gpu/drm/drm_vblank.c
> > +++ b/drivers/gpu/drm/drm_vblank.c
> > @@ -25,7 +25,9 @@
> >   */
> >  
> >  #include <linux/export.h>
> > +#include <linux/kthread.h>
> >  #include <linux/moduleparam.h>
> > +#include <uapi/linux/sched/types.h>
> >  
> >  #include <drm/drm_crtc.h>
> >  #include <drm/drm_drv.h>
> > @@ -91,6 +93,7 @@
> >  static bool
> >  drm_get_last_vbltimestamp(struct drm_device *dev, unsigned int pipe,
> >  			  ktime_t *tvblank, bool in_vblank_irq);
> > +static int drm_vblank_get(struct drm_device *dev, unsigned int pipe);
> >  
> >  static unsigned int drm_timestamp_precision = 20;  /* Default to 20
> > usecs. */
> >  
> > @@ -440,6 +443,9 @@ void drm_vblank_cleanup(struct drm_device *dev)
> >  			drm_core_check_feature(dev, DRIVER_MODESET));
> >  
> >  		del_timer_sync(&vblank->disable_timer);
> > +
> > +		wake_up_all(&vblank->vblank_work.work_wait);
> > +		kthread_stop(vblank->vblank_work.thread);
> >  	}
> >  
> >  	kfree(dev->vblank);
> > @@ -447,6 +453,108 @@ void drm_vblank_cleanup(struct drm_device *dev)
> >  	dev->num_crtcs = 0;
> >  }
> >  
> > +static int vblank_work_thread(void *data)
> > +{
> > +	struct drm_vblank_crtc *vblank = data;
> > +
> > +	while (!kthread_should_stop()) {
> > +		struct drm_vblank_work *work, *next;
> > +		LIST_HEAD(list);
> > +		u64 count;
> > +		int ret;
> > +
> > +		spin_lock_irq(&vblank->dev->event_lock);
> > +
> > +		ret = wait_event_interruptible_lock_irq(vblank->queue,
> > +							kthread_should_stop()
> > ||
> > +							!list_empty(&vblank-
> > >vblank_work.work_list),
> > +							vblank->dev-
> > >event_lock);
> > +
> > +		WARN_ON(ret && !kthread_should_stop() &&
> > +			list_empty(&vblank->vblank_work.irq_list) &&
> > +			list_empty(&vblank->vblank_work.work_list));
> > +
> > +		list_for_each_entry_safe(work, next,
> > +					 &vblank->vblank_work.work_list,
> > +					 list) {
> > +			list_move_tail(&work->list, &list);
> > +			work->state = DRM_VBL_WORK_RUNNING;
> > +		}
> > +
> > +		spin_unlock_irq(&vblank->dev->event_lock);
> > +
> > +		if (list_empty(&list))
> > +			continue;
> > +
> > +		count = atomic64_read(&vblank->count);
> > +		list_for_each_entry(work, &list, list)
> > +			work->func(work, count);
> > +
> > +		spin_lock_irq(&vblank->dev->event_lock);
> > +
> > +		list_for_each_entry_safe(work, next, &list, list) {
> > +			if (work->reschedule) {
> > +				list_move_tail(&work->list,
> > +					       &vblank->vblank_work.irq_list);
> > +				drm_vblank_get(vblank->dev, vblank->pipe);
> > +				work->reschedule = false;
> > +				work->state = DRM_VBL_WORK_WAITING;
> > +			} else {
> > +				list_del_init(&work->list);
> > +				work->cancel = false;
> > +				work->state = DRM_VBL_WORK_IDLE;
> > +			}
> > +		}
> > +
> > +		spin_unlock_irq(&vblank->dev->event_lock);
> > +
> > +		wake_up_all(&vblank->vblank_work.work_wait);
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +static void vblank_work_init(struct drm_vblank_crtc *vblank)
> > +{
> > +	struct sched_param param = {
> > +		.sched_priority = MAX_RT_PRIO - 1,
> > +	};
> > +	int ret;
> > +
> > +	INIT_LIST_HEAD(&vblank->vblank_work.irq_list);
> > +	INIT_LIST_HEAD(&vblank->vblank_work.work_list);
> > +	init_waitqueue_head(&vblank->vblank_work.work_wait);
> > +
> > +	vblank->vblank_work.thread =
> > +		kthread_run(vblank_work_thread, vblank, "card %d crtc %d",
> > +			    vblank->dev->primary->index, vblank->pipe);
> > +
> > +	ret = sched_setscheduler(vblank->vblank_work.thread,
> > +				 SCHED_FIFO, &param);
> > +	WARN_ON(ret);
> > +}
> > +
> > +/**
> > + * drm_vblank_work_init - initialize a vblank work item
> > + * @work: vblank work item
> > + * @crtc: CRTC whose vblank will trigger the work execution
> > + * @func: work function to be executed
> > + *
> > + * Initialize a vblank work item for a specific crtc.
> > + */
> > +void drm_vblank_work_init(struct drm_vblank_work *work, struct drm_crtc
> > *crtc,
> > +			  void (*func)(struct drm_vblank_work *work, u64
> > count))
> > +{
> > +	struct drm_device *dev = crtc->dev;
> > +	struct drm_vblank_crtc *vblank = &dev->vblank[drm_crtc_index(crtc)];
> > +
> > +	work->vblank = vblank;
> > +	work->state = DRM_VBL_WORK_IDLE;
> > +	work->func = func;
> > +	INIT_LIST_HEAD(&work->list);
> > +}
> > +EXPORT_SYMBOL(drm_vblank_work_init);
> > +
> >  /**
> >   * drm_vblank_init - initialize vblank support
> >   * @dev: DRM device
> > @@ -481,6 +589,8 @@ int drm_vblank_init(struct drm_device *dev, unsigned
> > int num_crtcs)
> >  		init_waitqueue_head(&vblank->queue);
> >  		timer_setup(&vblank->disable_timer, vblank_disable_fn, 0);
> >  		seqlock_init(&vblank->seqlock);
> > +
> > +		vblank_work_init(vblank);
> >  	}
> >  
> >  	DRM_INFO("Supports vblank timestamp caching Rev 2 (21.10.2013).\n");
> > @@ -1825,6 +1935,22 @@ static void drm_handle_vblank_events(struct
> > drm_device *dev, unsigned int pipe)
> >  	trace_drm_vblank_event(pipe, seq, now, high_prec);
> >  }
> >  
> > +static void drm_handle_vblank_works(struct drm_vblank_crtc *vblank)
> > +{
> > +	struct drm_vblank_work *work, *next;
> > +	u64 count = atomic64_read(&vblank->count);
> > +
> > +	list_for_each_entry_safe(work, next, &vblank->vblank_work.irq_list,
> > +				 list) {
> > +		if (!vblank_passed(count, work->count))
> > +			continue;
> > +
> > +		drm_vblank_put(vblank->dev, vblank->pipe);
> > +		list_move_tail(&work->list, &vblank->vblank_work.work_list);
> > +		work->state = DRM_VBL_WORK_SCHEDULED;
> > +	}
> > +}
> > +
> >  /**
> >   * drm_handle_vblank - handle a vblank event
> >   * @dev: DRM device
> > @@ -1866,6 +1992,7 @@ bool drm_handle_vblank(struct drm_device *dev,
> > unsigned int pipe)
> >  
> >  	spin_unlock(&dev->vblank_time_lock);
> >  
> > +	drm_handle_vblank_works(vblank);
> >  	wake_up(&vblank->queue);
> >  
> >  	/* With instant-off, we defer disabling the interrupt until after
> > @@ -2076,3 +2203,198 @@ int drm_crtc_queue_sequence_ioctl(struct
> > drm_device *dev, void *data,
> >  	kfree(e);
> >  	return ret;
> >  }
> > +
> > +/**
> > + * drm_vblank_work_schedule - schedule a vblank work
> > + * @work: vblank work to schedule
> > + * @count: target vblank count
> > + * @nextonmiss: defer until the next vblank if target vblank was missed
> > + *
> > + * Schedule @work for execution once the crtc vblank count reaches
> > @count.
> > + *
> > + * If the crtc vblank count has already reached @count and @nextonmiss is
> > + * %false the work starts to execute immediately.
> > + *
> > + * If the crtc vblank count has already reached @count and @nextonmiss is
> > + * %true the work is deferred until the next vblank (as if @count has
> > been
> > + * specified as crtc vblank count + 1).
> > + *
> > + * If @work is already scheduled, this function will reschedule said work
> > + * using the new @count.
> > + *
> > + * Returns:
> > + * 0 on success, error code on failure.
> > + */
> > +int drm_vblank_work_schedule(struct drm_vblank_work *work,
> > +			     u64 count, bool nextonmiss)
> > +{
> > +	struct drm_vblank_crtc *vblank = work->vblank;
> > +	unsigned long irqflags;
> > +	u64 cur_vbl;
> > +	int ret = 0;
> > +	bool rescheduling = false;
> > +	bool passed;
> > +
> > +	spin_lock_irqsave(&vblank->dev->event_lock, irqflags);
> > +
> > +	if (work->cancel)
> > +		goto out;
> > +
> > +	if (work->state == DRM_VBL_WORK_RUNNING) {
> > +		work->reschedule = true;
> > +		work->count = count;
> > +		goto out;
> > +	} else if (work->state != DRM_VBL_WORK_IDLE) {
> > +		if (work->count == count)
> > +			goto out;
> > +		rescheduling = true;
> > +	}
> > +
> > +	if (work->state != DRM_VBL_WORK_WAITING) {
> > +		ret = drm_vblank_get(vblank->dev, vblank->pipe);
> > +		if (ret)
> > +			goto out;
> > +	}
> > +
> > +	work->count = count;
> > +
> > +	cur_vbl = atomic64_read(&vblank->count);
> > +	passed = vblank_passed(cur_vbl, count);
> > +	if (passed)
> > +		DRM_ERROR("crtc %d vblank %llu already passed (current
> > %llu)\n",
> > +			  vblank->pipe, count, cur_vbl);
> > +
> > +	if (!nextonmiss && passed) {
> > +		drm_vblank_put(vblank->dev, vblank->pipe);
> > +		if (rescheduling)
> > +			list_move_tail(&work->list,
> > +				       &vblank->vblank_work.work_list);
> > +		else
> > +			list_add_tail(&work->list,
> > +				      &vblank->vblank_work.work_list);
> > +		work->state = DRM_VBL_WORK_SCHEDULED;
> > +		wake_up_all(&vblank->queue);
> > +	} else {
> > +		if (rescheduling)
> > +			list_move_tail(&work->list,
> > +				       &vblank->vblank_work.irq_list);
> > +		else
> > +			list_add_tail(&work->list,
> > +				      &vblank->vblank_work.irq_list);
> > +		work->state = DRM_VBL_WORK_WAITING;
> > +	}
> > +
> > + out:
> > +	spin_unlock_irqrestore(&vblank->dev->event_lock, irqflags);
> > +
> > +	return ret;
> > +}
> > +EXPORT_SYMBOL(drm_vblank_work_schedule);
> > +
> > +static bool vblank_work_cancel(struct drm_vblank_work *work)
> > +{
> > +	struct drm_vblank_crtc *vblank = work->vblank;
> > +
> > +	switch (work->state) {
> > +	case DRM_VBL_WORK_RUNNING:
> > +		work->cancel = true;
> > +		work->reschedule = false;
> > +		/* fall through */
> > +	default:
> > +	case DRM_VBL_WORK_IDLE:
> > +		return false;
> > +	case DRM_VBL_WORK_WAITING:
> > +		drm_vblank_put(vblank->dev, vblank->pipe);
> > +		/* fall through */
> > +	case DRM_VBL_WORK_SCHEDULED:
> > +		list_del_init(&work->list);
> > +		work->state = DRM_VBL_WORK_IDLE;
> > +		return true;
> > +	}
> > +}
> > +
> > +/**
> > + * drm_vblank_work_cancel - cancel a vblank work
> > + * @work: vblank work to cancel
> > + *
> > + * Cancel an already scheduled vblank work.
> > + *
> > + * On return @work may still be executing, unless the return
> > + * value is %true.
> > + *
> > + * Returns:
> > + * True if the work was cancelled before it started to excute, false
> > otherwise.
> > + */
> > +bool drm_vblank_work_cancel(struct drm_vblank_work *work)
> > +{
> > +	struct drm_vblank_crtc *vblank = work->vblank;
> > +	bool cancelled;
> > +
> > +	spin_lock_irq(&vblank->dev->event_lock);
> > +
> > +	cancelled = vblank_work_cancel(work);
> > +
> > +	spin_unlock_irq(&vblank->dev->event_lock);
> > +
> > +	return cancelled;
> > +}
> > +EXPORT_SYMBOL(drm_vblank_work_cancel);
> > +
> > +/**
> > + * drm_vblank_work_cancel_sync - cancel a vblank work and wait for it to
> > finish executing
> > + * @work: vblank work to cancel
> > + *
> > + * Cancel an already scheduled vblank work and wait for its
> > + * execution to finish.
> > + *
> > + * On return @work is no longer guaraneed to be executing.
> > + *
> > + * Returns:
> > + * True if the work was cancelled before it started to excute, false
> > otherwise.
> > + */
> > +bool drm_vblank_work_cancel_sync(struct drm_vblank_work *work)
> > +{
> > +	struct drm_vblank_crtc *vblank = work->vblank;
> > +	bool cancelled;
> > +	long ret;
> > +
> > +	spin_lock_irq(&vblank->dev->event_lock);
> > +
> > +	cancelled = vblank_work_cancel(work);
> > +
> > +	ret = wait_event_lock_irq_timeout(vblank->vblank_work.work_wait,
> > +					  work->state == DRM_VBL_WORK_IDLE,
> > +					  vblank->dev->event_lock,
> > +					  10 * HZ);
> > +
> > +	spin_unlock_irq(&vblank->dev->event_lock);
> > +
> > +	WARN(!ret, "crtc %d vblank work timed out\n", vblank->pipe);
> > +
> > +	return cancelled;
> > +}
> > +EXPORT_SYMBOL(drm_vblank_work_cancel_sync);
> > +
> > +/**
> > + * drm_vblank_work_flush - wait for a scheduled vblank work to finish
> > excuting
> > + * @work: vblank work to flush
> > + *
> > + * Wait until @work has finished executing.
> > + */
> > +void drm_vblank_work_flush(struct drm_vblank_work *work)
> > +{
> > +	struct drm_vblank_crtc *vblank = work->vblank;
> > +	long ret;
> > +
> > +	spin_lock_irq(&vblank->dev->event_lock);
> > +
> > +	ret = wait_event_lock_irq_timeout(vblank->vblank_work.work_wait,
> > +					  work->state == DRM_VBL_WORK_IDLE,
> > +					  vblank->dev->event_lock,
> > +					  10 * HZ);
> > +
> > +	spin_unlock_irq(&vblank->dev->event_lock);
> > +
> > +	WARN(!ret, "crtc %d vblank work timed out\n", vblank->pipe);
> > +}
> > +EXPORT_SYMBOL(drm_vblank_work_flush);
> > diff --git a/include/drm/drm_vblank.h b/include/drm/drm_vblank.h
> > index dd9f5b9e56e4..ac9130f419af 100644
> > --- a/include/drm/drm_vblank.h
> > +++ b/include/drm/drm_vblank.h
> > @@ -203,8 +203,42 @@ struct drm_vblank_crtc {
> >  	 * disabling functions multiple times.
> >  	 */
> >  	bool enabled;
> > +
> > +	struct {
> > +		struct task_struct *thread;
> > +		struct list_head irq_list, work_list;
> > +		wait_queue_head_t work_wait;
> > +	} vblank_work;
> > +};
> > +
> > +struct drm_vblank_work {
> > +	u64 count;
> > +	struct drm_vblank_crtc *vblank;
> > +	void (*func)(struct drm_vblank_work *work, u64 count);
> > +	struct list_head list;
> > +	enum {
> > +		DRM_VBL_WORK_IDLE,
> > +		DRM_VBL_WORK_WAITING,
> > +		DRM_VBL_WORK_SCHEDULED,
> > +		DRM_VBL_WORK_RUNNING,
> > +	} state;
> > +	bool cancel : 1;
> > +	bool reschedule : 1;
> >  };
> >  
> > +int drm_vblank_work_schedule(struct drm_vblank_work *work,
> > +			     u64 count, bool nextonmiss);
> > +void drm_vblank_work_init(struct drm_vblank_work *work, struct drm_crtc
> > *crtc,
> > +			  void (*func)(struct drm_vblank_work *work, u64
> > count));
> > +bool drm_vblank_work_cancel(struct drm_vblank_work *work);
> > +bool drm_vblank_work_cancel_sync(struct drm_vblank_work *work);
> > +void drm_vblank_work_flush(struct drm_vblank_work *work);
> > +
> > +static inline bool drm_vblank_work_pending(struct drm_vblank_work *work)
> > +{
> > +	return work->state != DRM_VBL_WORK_IDLE;
> > +}
> > +
> >  int drm_vblank_init(struct drm_device *dev, unsigned int num_crtcs);
> >  bool drm_dev_has_vblank(const struct drm_device *dev);
> >  u64 drm_crtc_vblank_count(struct drm_crtc *crtc);
> > -- 
> > 2.24.1
> > 
-- 
Cheers,
	Lyude Paul (she/her)
	Associate Software Engineer at Red Hat

Powered by blists - more mailing lists