[PATCH] drm/vkms: Fix race condition around accessing frame number
Daniel Vetter
daniel at ffwll.ch
Mon Sep 3 07:29:15 UTC 2018
On Sat, Sep 01, 2018 at 08:53:52PM +0300, Haneen Mohammed wrote:
> On Fri, Aug 31, 2018 at 10:41:40AM +0200, Daniel Vetter wrote:
> > On Fri, Aug 24, 2018 at 02:16:34AM +0300, Haneen Mohammed wrote:
> > > crtc_state is accessed by both vblank_handle() and the ordered
> > > work_struct handle vkms_crc_work_handle() to retrieve and or update
> > > the frame number for computed CRC.
> > >
> > > Since work_struct can fail, add frame_end to account for missing frame
> > > numbers.
> > >
> > > use atomic_t witth appropriate flags for synchronization between hrtimer
> > > callback and ordered work_struct handle since spinlock can't be used
> > > with work_struct handle and mutex can't be used with hrtimer callback.
> > >
> > > This patch passes the following subtests from igt kms_pipe_crc_basic test:
> > > bad-source, read-crc-pipe-A, read-crc-pipe-A-frame-sequence,
> > > nonblocking-crc-pipe-A, nonblocking-crc-pipe-A-frame-sequence
> > >
> > > Signed-off-by: Haneen Mohammed <hamohammed.sa at gmail.com>
> >
> > So atomic_t is probably the greatest trap in the linux kernel. It sounds
> > like the right thing, but in 99% of all case you want to use it it isn't.
> > The trouble is that atomic_t is _very_ unordered, the only thing it
> > guarantees is that atomic_t transactions to the _same_ variable are
> > consistent. But anything else can be reordered at will.
> >
> > This is very confusing since the new C++ atomic standards has fully
> > ordered atomics as the default, and you expressedly need to ask for the
> > weakly ordered ones. In linux you need to sprinkle epic amounts of
> > smb_barrier* and similar things around them to make atomic_t behave like a
> > "normal" C++ atomic type.
> >
> > tldr; atomic_t is good for special refcounting needs, when the normal
> > refcount_t doesn't cut it. Not much else.
> >
> > What usually should be done:
> > - Use normal u64 (to match the vblank counter size) instead of atomic_t
> > here.
> > - Make sure all access is protect by an appropriate spinlock.
> >
>
> I see, thanks for the explanation!
>
> hm would the patch be fine if I just switch atomic_t to u64 without
> using spinlock? since that won't be possible with the work_struct.
Why can't you use the spin_lock from the worker? You need the
spin_lock_irq variant, plus a bit of pointer chasing, but should be doable
...
Cheers, Daniel
>
> Thanks,
> Haneen
>
> > > ---
> > > drivers/gpu/drm/vkms/vkms_crc.c | 33 ++++++++++++++++++++++++++++++--
> > > drivers/gpu/drm/vkms/vkms_crtc.c | 13 +++++++++++--
> > > drivers/gpu/drm/vkms/vkms_drv.h | 6 ++++--
> > > 3 files changed, 46 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/vkms/vkms_crc.c b/drivers/gpu/drm/vkms/vkms_crc.c
> > > index ed47d67cecd6..4a1ba5b7886a 100644
> > > --- a/drivers/gpu/drm/vkms/vkms_crc.c
> > > +++ b/drivers/gpu/drm/vkms/vkms_crc.c
> > > @@ -34,6 +34,15 @@ static uint32_t _vkms_get_crc(struct vkms_crc_data *crc_data)
> > > return crc;
> > > }
> > >
> > > +/**
> > > + * vkms_crc_work_handle - ordered work_struct to compute CRC
> > > + *
> > > + * @work: work_struct
> > > + *
> > > + * Work handler for computing CRCs. work_struct scheduled in
> > > + * an ordered workqueue that's periodically scheduled to run by
> > > + * _vblank_handle() and flushed at vkms_atomic_crtc_destroy_state().
> > > + */
> > > void vkms_crc_work_handle(struct work_struct *work)
> > > {
> > > struct vkms_crtc_state *crtc_state = container_of(work,
> > > @@ -45,8 +54,18 @@ void vkms_crc_work_handle(struct work_struct *work)
> > > output);
> > > struct vkms_crc_data *primary_crc = NULL;
> > > struct drm_plane *plane;
> > > -
> > > u32 crc32 = 0;
> > > + u32 frame_start, frame_end;
> > > +
> > > + frame_start = atomic_read(&crtc_state->frame_start);
> > > + frame_end = atomic_read(&crtc_state->frame_end);
> > > + /* _vblank_handle() hasn't updated frame_start yet */
> > > + if (!frame_start) {
> >
> > I think if we go with u64 we can ignore the issues for wrap-arround, since
> > that will simply never happen. But a comment would be good.
> >
> > Aside from the atomic_t issue I think this looks good.
> > -Daniel
> >
> > > + return;
> > > + } else if (frame_start == frame_end) {
> > > + atomic_set(&crtc_state->frame_start, 0);
> > > + return;
> > > + }
> > >
> > > drm_for_each_plane(plane, &vdev->drm) {
> > > struct vkms_plane_state *vplane_state;
> > > @@ -67,7 +86,17 @@ void vkms_crc_work_handle(struct work_struct *work)
> > > if (primary_crc)
> > > crc32 = _vkms_get_crc(primary_crc);
> > >
> > > - drm_crtc_add_crc_entry(crtc, true, crtc_state->n_frame, &crc32);
> > > + frame_end = drm_crtc_accurate_vblank_count(crtc);
> > > +
> > > + /* queue_work can fail to schedule crc_work; add crc for
> > > + * missing frames
> > > + */
> > > + while (frame_start <= frame_end)
> > > + drm_crtc_add_crc_entry(crtc, true, frame_start++, &crc32);
> > > +
> > > + /* to avoid using the same value again */
> > > + atomic_set(&crtc_state->frame_end, frame_end);
> > > + atomic_set(&crtc_state->frame_start, 0);
> > > }
> > >
> > > static int vkms_crc_parse_source(const char *src_name, bool *enabled)
> > > diff --git a/drivers/gpu/drm/vkms/vkms_crtc.c b/drivers/gpu/drm/vkms/vkms_crtc.c
> > > index 9d0b1a325a78..a170677acd46 100644
> > > --- a/drivers/gpu/drm/vkms/vkms_crtc.c
> > > +++ b/drivers/gpu/drm/vkms/vkms_crtc.c
> > > @@ -22,8 +22,17 @@ static void _vblank_handle(struct vkms_output *output)
> > > DRM_ERROR("vkms failure on handling vblank");
> > >
> > > if (state && output->crc_enabled) {
> > > - state->n_frame = drm_crtc_accurate_vblank_count(crtc);
> > > - queue_work(output->crc_workq, &state->crc_work);
> > > + u32 frame = drm_crtc_accurate_vblank_count(crtc);
> > > +
> > > + /* update frame_start only if a queued vkms_crc_work_handle has
> > > + * read the data
> > > + */
> > > + if (!atomic_read(&state->frame_start))
> > > + atomic_set(&state->frame_start, frame);
> > > +
> > > + ret = queue_work(output->crc_workq, &state->crc_work);
> > > + if (!ret)
> > > + DRM_WARN("failed to queue vkms_crc_work_handle");
> > > }
> > >
> > > spin_unlock(&output->lock);
> > > diff --git a/drivers/gpu/drm/vkms/vkms_drv.h b/drivers/gpu/drm/vkms/vkms_drv.h
> > > index 2017a2ccc43d..4a3956a0549e 100644
> > > --- a/drivers/gpu/drm/vkms/vkms_drv.h
> > > +++ b/drivers/gpu/drm/vkms/vkms_drv.h
> > > @@ -39,12 +39,14 @@ struct vkms_plane_state {
> > > * vkms_crtc_state - Driver specific CRTC state
> > > * @base: base CRTC state
> > > * @crc_work: work struct to compute and add CRC entries
> > > - * @n_frame: frame number for computed CRC
> > > + * @n_frame_start: start frame number for computed CRC
> > > + * @n_frame_end: end frame number for computed CRC
> > > */
> > > struct vkms_crtc_state {
> > > struct drm_crtc_state base;
> > > struct work_struct crc_work;
> > > - unsigned int n_frame;
> > > + atomic_t frame_start;
> > > + atomic_t frame_end;
> > > };
> > >
> > > struct vkms_output {
> > > --
> > > 2.17.1
> > >
> >
> > --
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > http://blog.ffwll.ch
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the dri-devel
mailing list