Since 4.7 kernel, we've seen the error messages like
kernel: [TTM] Buffer eviction failed
kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
on QXL when switching and accessing on VT. The culprit was the generic
deferred_io code (qxl driver switched to it since 4.7). There is a
race between the dirty clip update and the call of callback.
In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
while it kicks off the update worker outside the spinlock. Meanwhile
the update worker clears the dirty clip in the spinlock, too. Thus,
when drm_fb_helper_dirty() is called concurrently, schedule_work() is
called after the clip is cleared in the first worker call.
The fix is simply moving schedule_work() inside the spinlock.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98322
Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1003298
Fixes: eaa434defaca ('drm/fb-helper: Add fb_deferred_io support')
Signed-off-by: Takashi Iwai <[email protected]>
---
drivers/gpu/drm/drm_fb_helper.c | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
index 03414bde1f15..bae392dea2cc 100644
--- a/drivers/gpu/drm/drm_fb_helper.c
+++ b/drivers/gpu/drm/drm_fb_helper.c
@@ -861,9 +861,8 @@ static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
clip->y1 = min_t(u32, clip->y1, y);
clip->x2 = max_t(u32, clip->x2, x + width);
clip->y2 = max_t(u32, clip->y2, y + height);
- spin_unlock_irqrestore(&helper->dirty_lock, flags);
-
schedule_work(&helper->dirty_work);
+ spin_unlock_irqrestore(&helper->dirty_lock, flags);
}
/**
--
2.10.1
On Thu, 20 Oct 2016 15:20:55 +0200,
Takashi Iwai wrote:
>
> Since 4.7 kernel, we've seen the error messages like
>
> kernel: [TTM] Buffer eviction failed
> kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
> kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
>
> on QXL when switching and accessing on VT. The culprit was the generic
> deferred_io code (qxl driver switched to it since 4.7). There is a
> race between the dirty clip update and the call of callback.
>
> In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
> while it kicks off the update worker outside the spinlock. Meanwhile
> the update worker clears the dirty clip in the spinlock, too. Thus,
> when drm_fb_helper_dirty() is called concurrently, schedule_work() is
> called after the clip is cleared in the first worker call.
>
> The fix is simply moving schedule_work() inside the spinlock.
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98322
> Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1003298
> Fixes: eaa434defaca ('drm/fb-helper: Add fb_deferred_io support')
> Signed-off-by: Takashi Iwai <[email protected]>
I forgot to put Cc to stable. Feel free to add it once if it's
accepted.
thanks,
Takashi
> ---
> drivers/gpu/drm/drm_fb_helper.c | 3 +--
> 1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> index 03414bde1f15..bae392dea2cc 100644
> --- a/drivers/gpu/drm/drm_fb_helper.c
> +++ b/drivers/gpu/drm/drm_fb_helper.c
> @@ -861,9 +861,8 @@ static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
> clip->y1 = min_t(u32, clip->y1, y);
> clip->x2 = max_t(u32, clip->x2, x + width);
> clip->y2 = max_t(u32, clip->y2, y + height);
> - spin_unlock_irqrestore(&helper->dirty_lock, flags);
> -
> schedule_work(&helper->dirty_work);
> + spin_unlock_irqrestore(&helper->dirty_lock, flags);
> }
>
> /**
> --
> 2.10.1
>
On Thu, Oct 20, 2016 at 03:20:55PM +0200, Takashi Iwai wrote:
> Since 4.7 kernel, we've seen the error messages like
>
> kernel: [TTM] Buffer eviction failed
> kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
> kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
>
> on QXL when switching and accessing on VT. The culprit was the generic
> deferred_io code (qxl driver switched to it since 4.7). There is a
> race between the dirty clip update and the call of callback.
>
> In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
> while it kicks off the update worker outside the spinlock. Meanwhile
> the update worker clears the dirty clip in the spinlock, too. Thus,
> when drm_fb_helper_dirty() is called concurrently, schedule_work() is
> called after the clip is cleared in the first worker call.
Why does that matter? The first worker should have done all the
necessary work already, no?
>
> The fix is simply moving schedule_work() inside the spinlock.
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98322
> Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1003298
> Fixes: eaa434defaca ('drm/fb-helper: Add fb_deferred_io support')
> Signed-off-by: Takashi Iwai <[email protected]>
> ---
> drivers/gpu/drm/drm_fb_helper.c | 3 +--
> 1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> index 03414bde1f15..bae392dea2cc 100644
> --- a/drivers/gpu/drm/drm_fb_helper.c
> +++ b/drivers/gpu/drm/drm_fb_helper.c
> @@ -861,9 +861,8 @@ static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
> clip->y1 = min_t(u32, clip->y1, y);
> clip->x2 = max_t(u32, clip->x2, x + width);
> clip->y2 = max_t(u32, clip->y2, y + height);
> - spin_unlock_irqrestore(&helper->dirty_lock, flags);
> -
> schedule_work(&helper->dirty_work);
> + spin_unlock_irqrestore(&helper->dirty_lock, flags);
> }
>
> /**
> --
> 2.10.1
>
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
--
Ville Syrj?l?
Intel OTC
On Thu, 20 Oct 2016 15:28:14 +0200,
Ville Syrjälä wrote:
>
> On Thu, Oct 20, 2016 at 03:20:55PM +0200, Takashi Iwai wrote:
> > Since 4.7 kernel, we've seen the error messages like
> >
> > kernel: [TTM] Buffer eviction failed
> > kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
> > kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
> >
> > on QXL when switching and accessing on VT. The culprit was the generic
> > deferred_io code (qxl driver switched to it since 4.7). There is a
> > race between the dirty clip update and the call of callback.
> >
> > In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
> > while it kicks off the update worker outside the spinlock. Meanwhile
> > the update worker clears the dirty clip in the spinlock, too. Thus,
> > when drm_fb_helper_dirty() is called concurrently, schedule_work() is
> > called after the clip is cleared in the first worker call.
>
> Why does that matter? The first worker should have done all the
> necessary work already, no?
Before the first call, it clears the clip and passes the copied clip
to the callback. Then the second call will be with the cleared and
untouched clip, i.e. with x1=~0. This confuses
qxl_framebuffer_dirty().
Of course, we can filter out in the callback side by checking the
clip. It was actually my first version. But basically it's a race
and should be covered better in the caller side.
thanks,
Takashi
>
> >
> > The fix is simply moving schedule_work() inside the spinlock.
> >
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98322
> > Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1003298
> > Fixes: eaa434defaca ('drm/fb-helper: Add fb_deferred_io support')
> > Signed-off-by: Takashi Iwai <[email protected]>
> > ---
> > drivers/gpu/drm/drm_fb_helper.c | 3 +--
> > 1 file changed, 1 insertion(+), 2 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> > index 03414bde1f15..bae392dea2cc 100644
> > --- a/drivers/gpu/drm/drm_fb_helper.c
> > +++ b/drivers/gpu/drm/drm_fb_helper.c
> > @@ -861,9 +861,8 @@ static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
> > clip->y1 = min_t(u32, clip->y1, y);
> > clip->x2 = max_t(u32, clip->x2, x + width);
> > clip->y2 = max_t(u32, clip->y2, y + height);
> > - spin_unlock_irqrestore(&helper->dirty_lock, flags);
> > -
> > schedule_work(&helper->dirty_work);
> > + spin_unlock_irqrestore(&helper->dirty_lock, flags);
> > }
> >
> > /**
> > --
> > 2.10.1
> >
> > _______________________________________________
> > dri-devel mailing list
> > [email protected]
> > https://lists.freedesktop.org/mailman/listinfo/dri-devel
>
> --
> Ville Syrjälä
> Intel OTC
>
On Thu, Oct 20, 2016 at 03:36:54PM +0200, Takashi Iwai wrote:
> On Thu, 20 Oct 2016 15:28:14 +0200,
> Ville Syrj?l? wrote:
> >
> > On Thu, Oct 20, 2016 at 03:20:55PM +0200, Takashi Iwai wrote:
> > > Since 4.7 kernel, we've seen the error messages like
> > >
> > > kernel: [TTM] Buffer eviction failed
> > > kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
> > > kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
> > >
> > > on QXL when switching and accessing on VT. The culprit was the generic
> > > deferred_io code (qxl driver switched to it since 4.7). There is a
> > > race between the dirty clip update and the call of callback.
> > >
> > > In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
> > > while it kicks off the update worker outside the spinlock. Meanwhile
> > > the update worker clears the dirty clip in the spinlock, too. Thus,
> > > when drm_fb_helper_dirty() is called concurrently, schedule_work() is
> > > called after the clip is cleared in the first worker call.
> >
> > Why does that matter? The first worker should have done all the
> > necessary work already, no?
>
> Before the first call, it clears the clip and passes the copied clip
> to the callback. Then the second call will be with the cleared and
> untouched clip, i.e. with x1=~0. This confuses
> qxl_framebuffer_dirty().
>
> Of course, we can filter out in the callback side by checking the
> clip. It was actually my first version. But basically it's a race
> and should be covered better in the caller side.
The race is still there AFAICS. The worker may already be executing but
not yet in the critical section, at which point drm_fb_helper_dirty()
will expand the dirty rectangle, and schedule another work. So the first
worker will already see the expanded rectangle, and second worker will
get zilch.
I think the only good fix is to have the worker validate the dirty
rectangle before calling the driver.
>
>
> thanks,
>
> Takashi
>
> >
> > >
> > > The fix is simply moving schedule_work() inside the spinlock.
> > >
> > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98322
> > > Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1003298
> > > Fixes: eaa434defaca ('drm/fb-helper: Add fb_deferred_io support')
> > > Signed-off-by: Takashi Iwai <[email protected]>
> > > ---
> > > drivers/gpu/drm/drm_fb_helper.c | 3 +--
> > > 1 file changed, 1 insertion(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> > > index 03414bde1f15..bae392dea2cc 100644
> > > --- a/drivers/gpu/drm/drm_fb_helper.c
> > > +++ b/drivers/gpu/drm/drm_fb_helper.c
> > > @@ -861,9 +861,8 @@ static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
> > > clip->y1 = min_t(u32, clip->y1, y);
> > > clip->x2 = max_t(u32, clip->x2, x + width);
> > > clip->y2 = max_t(u32, clip->y2, y + height);
> > > - spin_unlock_irqrestore(&helper->dirty_lock, flags);
> > > -
> > > schedule_work(&helper->dirty_work);
> > > + spin_unlock_irqrestore(&helper->dirty_lock, flags);
> > > }
> > >
> > > /**
> > > --
> > > 2.10.1
> > >
> > > _______________________________________________
> > > dri-devel mailing list
> > > [email protected]
> > > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> >
> > --
> > Ville Syrj?l?
> > Intel OTC
> >
--
Ville Syrj?l?
Intel OTC
On Thu, Oct 20, 2016 at 03:36:54PM +0200, Takashi Iwai wrote:
> On Thu, 20 Oct 2016 15:28:14 +0200,
> Ville Syrj?l? wrote:
> >
> > On Thu, Oct 20, 2016 at 03:20:55PM +0200, Takashi Iwai wrote:
> > > Since 4.7 kernel, we've seen the error messages like
> > >
> > > kernel: [TTM] Buffer eviction failed
> > > kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
> > > kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
> > >
> > > on QXL when switching and accessing on VT. The culprit was the generic
> > > deferred_io code (qxl driver switched to it since 4.7). There is a
> > > race between the dirty clip update and the call of callback.
> > >
> > > In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
> > > while it kicks off the update worker outside the spinlock. Meanwhile
> > > the update worker clears the dirty clip in the spinlock, too. Thus,
> > > when drm_fb_helper_dirty() is called concurrently, schedule_work() is
> > > called after the clip is cleared in the first worker call.
> >
> > Why does that matter? The first worker should have done all the
> > necessary work already, no?
>
> Before the first call, it clears the clip and passes the copied clip
> to the callback. Then the second call will be with the cleared and
> untouched clip, i.e. with x1=~0. This confuses
> qxl_framebuffer_dirty().
>
> Of course, we can filter out in the callback side by checking the
> clip. It was actually my first version. But basically it's a race
> and should be covered better in the caller side.
Hm, I thought schedule_work also schedules the work when it's getting
processed right now. Which means if you're super unlucky you can still end
up with the work hitting an empty rectangle. I think filtering empty rects
in the worker is what we need to do instead.
Or is coffee not working right now?
-Daniel
>
>
> thanks,
>
> Takashi
>
> >
> > >
> > > The fix is simply moving schedule_work() inside the spinlock.
> > >
> > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98322
> > > Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1003298
> > > Fixes: eaa434defaca ('drm/fb-helper: Add fb_deferred_io support')
> > > Signed-off-by: Takashi Iwai <[email protected]>
> > > ---
> > > drivers/gpu/drm/drm_fb_helper.c | 3 +--
> > > 1 file changed, 1 insertion(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> > > index 03414bde1f15..bae392dea2cc 100644
> > > --- a/drivers/gpu/drm/drm_fb_helper.c
> > > +++ b/drivers/gpu/drm/drm_fb_helper.c
> > > @@ -861,9 +861,8 @@ static void drm_fb_helper_dirty(struct fb_info *info, u32 x, u32 y,
> > > clip->y1 = min_t(u32, clip->y1, y);
> > > clip->x2 = max_t(u32, clip->x2, x + width);
> > > clip->y2 = max_t(u32, clip->y2, y + height);
> > > - spin_unlock_irqrestore(&helper->dirty_lock, flags);
> > > -
> > > schedule_work(&helper->dirty_work);
> > > + spin_unlock_irqrestore(&helper->dirty_lock, flags);
> > > }
> > >
> > > /**
> > > --
> > > 2.10.1
> > >
> > > _______________________________________________
> > > dri-devel mailing list
> > > [email protected]
> > > https://lists.freedesktop.org/mailman/listinfo/dri-devel
> >
> > --
> > Ville Syrj?l?
> > Intel OTC
> >
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
On Thu, 20 Oct 2016 16:17:25 +0200,
Ville Syrjälä wrote:
>
> On Thu, Oct 20, 2016 at 03:36:54PM +0200, Takashi Iwai wrote:
> > On Thu, 20 Oct 2016 15:28:14 +0200,
> > Ville Syrjälä wrote:
> > >
> > > On Thu, Oct 20, 2016 at 03:20:55PM +0200, Takashi Iwai wrote:
> > > > Since 4.7 kernel, we've seen the error messages like
> > > >
> > > > kernel: [TTM] Buffer eviction failed
> > > > kernel: qxl 0000:00:02.0: object_init failed for (4026540032, 0x00000001)
> > > > kernel: [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
> > > >
> > > > on QXL when switching and accessing on VT. The culprit was the generic
> > > > deferred_io code (qxl driver switched to it since 4.7). There is a
> > > > race between the dirty clip update and the call of callback.
> > > >
> > > > In drm_fb_helper_dirty(), the dirty clip is updated in the spinlock,
> > > > while it kicks off the update worker outside the spinlock. Meanwhile
> > > > the update worker clears the dirty clip in the spinlock, too. Thus,
> > > > when drm_fb_helper_dirty() is called concurrently, schedule_work() is
> > > > called after the clip is cleared in the first worker call.
> > >
> > > Why does that matter? The first worker should have done all the
> > > necessary work already, no?
> >
> > Before the first call, it clears the clip and passes the copied clip
> > to the callback. Then the second call will be with the cleared and
> > untouched clip, i.e. with x1=~0. This confuses
> > qxl_framebuffer_dirty().
> >
> > Of course, we can filter out in the callback side by checking the
> > clip. It was actually my first version. But basically it's a race
> > and should be covered better in the caller side.
>
> The race is still there AFAICS. The worker may already be executing but
> not yet in the critical section, at which point drm_fb_helper_dirty()
> will expand the dirty rectangle, and schedule another work. So the first
> worker will already see the expanded rectangle, and second worker will
> get zilch.
Hrm, right, there's a slight race window there.
> I think the only good fix is to have the worker validate the dirty
> rectangle before calling the driver.
OK, let me cook it quickly. (It was actually the second version of
the patch I wrote, and I sent the third one :)
thanks,
Takashi