[RFC v2 8/8] drm/fence: add out-fences support
Gustavo Padovan
gustavo at padovan.org
Thu Apr 28 15:23:46 UTC 2016
2016-04-26 Daniel Vetter <daniel at ffwll.ch>:
> On Mon, Apr 25, 2016 at 07:33:28PM -0300, Gustavo Padovan wrote:
> > From: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
> >
> > Support DRM out-fences creating a sync_file with a fence for each crtc
> > update with the DRM_MODE_ATOMIC_OUT_FENCE flag.
> >
> > We then send an struct drm_out_fences array with the out-fences fds back in
> > the drm_atomic_ioctl() as an out arg in the out_fences_ptr field.
> >
> > struct drm_out_fences {
> > __u32 crtc_id;
> > __u32 fd;
> > };
> >
> > v2: Comment by Rob Clark:
> > - Squash commit that adds DRM_MODE_ATOMIC_OUT_FENCE flag here.
> >
> > Comment by Daniel Vetter:
> > - Add clean up code for out_fences
> >
> > Signed-off-by: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
> > ---
> > drivers/gpu/drm/drm_atomic.c | 163 +++++++++++++++++++++++++++++++++++++++++--
> > include/drm/drm_crtc.h | 10 +++
> > include/uapi/drm/drm_mode.h | 11 ++-
> > 3 files changed, 179 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c
> > index 5f9d434..06c6007 100644
> > --- a/drivers/gpu/drm/drm_atomic.c
> > +++ b/drivers/gpu/drm/drm_atomic.c
> > @@ -1566,6 +1566,133 @@ void drm_atomic_clean_old_fb(struct drm_device *dev,
> > }
> > EXPORT_SYMBOL(drm_atomic_clean_old_fb);
> >
> > +static struct drm_out_fence_state *get_out_fence(struct drm_device *dev,
> > + struct drm_atomic_state *state,
> > + uint32_t __user *out_fences_ptr,
> > + uint64_t count_out_fences,
> > + uint64_t user_data)
> > +{
> > + struct drm_crtc *crtc;
> > + struct drm_crtc_state *crtc_state;
> > + struct drm_out_fences *out_fences;
> > + struct drm_out_fence_state *fence_state;
> > + int num_fences = 0;
> > + int i, ret;
> > +
> > + if (count_out_fences > dev->mode_config.num_crtc)
> > + return ERR_PTR(-EINVAL);
> > +
> > + out_fences = kcalloc(count_out_fences, sizeof(*out_fences),
> > + GFP_KERNEL);
> > + if (!out_fences)
> > + return ERR_PTR(-ENOMEM);
>
> A bit tricky, but the above kcalloc is the only thing that catches integer
> overflows in count_out_fences. Needs a comment imo since this could be a
> security exploit if we accidentally screw it up.
The check above makes sure that count_out_fences is not bigger than
num_crtc. Don't that fix this?
>
> Also needs a testcase imo.
>
> > +
> > + fence_state = kcalloc(count_out_fences, sizeof(*fence_state),
> > + GFP_KERNEL);
> > + if (!fence_state) {
> > + kfree(out_fences);
> > + return ERR_PTR(-ENOMEM);
> > + }
> > +
> > + for (i = 0 ; i < count_out_fences ; i++)
> > + fence_state[i].fd = -1;
> > +
> > + for_each_crtc_in_state(state, crtc, crtc_state, i) {
> > + struct drm_pending_vblank_event *e;
> > + struct fence *fence;
> > + char name[32];
> > +
> > + fence = kzalloc(sizeof(*fence), GFP_KERNEL);
> > + if (!fence) {
> > + ret = -ENOMEM;
> > + goto out;
> > + }
> > +
> > + fence_init(fence, &drm_crtc_fence_ops, &crtc->fence_lock,
> > + crtc->fence_context, crtc->fence_seqno);
> > +
> > + snprintf(name, sizeof(name), "crtc-%d_%lu",
> > + drm_crtc_index(crtc), crtc->fence_seqno++);
>
> Hm ... fence_init_with_name? I'm kinda confused why we only name fences
> that are exported though, and why not all of them. Debugging fence
> deadlocks is real hard, so giving them all names might be a good idea.
>
> Anyway, seems like more room for a bit more sync_file/struct fence
> merging.
We just removed name from sync_file_create() so snprintf() is not even
necessary here anymore.
>
> > +
> > + fence_state[i].fd = get_unused_fd_flags(O_CLOEXEC);
> > + if (fence_state[i].fd < 0) {
> > + fence_put(fence);
> > + ret = fence_state[i].fd;
> > + goto out;
> > + }
> > +
> > + fence_state[i].sync_file = sync_file_create(name, fence);
> > + if(!fence_state[i].sync_file) {
> > + fence_put(fence);
> > + ret = -ENOMEM;
> > + goto out;
> > + }
> > +
> > + if (crtc_state->event) {
> > + crtc_state->event->base.fence = fence;
> > + } else {
>
> This looks a bit funny - I'd change the create event logic to create an
> event either if we have the either drm event or out-fence flag set.
Ok.
>
> > + e = create_vblank_event(dev, NULL, fence, user_data);
> > + if (!e) {
> > + ret = -ENOMEM;
> > + goto out;
> > + }
> > +
> > + crtc_state->event = e;
> > + }
> > +
> > + out_fences[num_fences].crtc_id = crtc->base.id;
> > + out_fences[num_fences].fd = fence_state[i].fd;
> > + num_fences++;
> > + }
> > +
> > + if (copy_to_user(out_fences_ptr, out_fences,
> > + num_fences * sizeof(*out_fences))) {
> > + ret = -EFAULT;
> > + goto out;
> > + }
> > +
> > + kfree(out_fences);
> > +
> > + return fence_state;
> > +
> > +out:
> > + for (i = 0 ; i < count_out_fences ; i++) {
> > + if (fence_state[i].sync_file)
> > + sync_file_put(fence_state[i].sync_file);
> > + if (fence_state[i].fd >= 0)
> > + put_unused_fd(fence_state[i].fd);
> > + }
> > +
> > + kfree(fence_state);
> > + kfree(out_fences);
> > +
> > + return ERR_PTR(ret);
> > +}
> > +
> > +static void install_out_fence(uint64_t count_out_fences,
> > + struct drm_out_fence_state *state)
> > +{
> > + int i;
> > +
> > + for (i = 0 ; i < count_out_fences ; i++) {
> > + if (state[i].sync_file)
> > + sync_file_install(state[i].sync_file, state[i].fd);
>
> Is sync_file_install anything more than fd_install? Imo a wrapper for
> just that function is overkill and just hides stuff. I'd nuke it (another
> sync_file patch though). In dma-buf we also don't wrap it, we only have a
> convenience wrapper for users who want to combine the
> get_unused_flags+fd_install in one go. And maybe even that is silly.
>
> Ok, I unlazied and it's indeed just a silly wrapper. Please nuke it.
already fixed in the sync file de-stage patches.
>
> > + }
> > +}
> > +
> > +static void release_out_fence(uint64_t count_out_fences,
> > + struct drm_out_fence_state *state)
> > +{
> > + int i;
> > +
> > + for (i = 0 ; i < count_out_fences ; i++) {
> > + if (state->sync_file)
> > + sync_file_put(state->sync_file);
> > + if (state->fd >= 0)
> > + put_unused_fd(state->fd);
> > + }
> > +}
> > +
> > int drm_mode_atomic_ioctl(struct drm_device *dev,
> > void *data, struct drm_file *file_priv)
> > {
> > @@ -1574,12 +1701,14 @@ int drm_mode_atomic_ioctl(struct drm_device *dev,
> > uint32_t __user *count_props_ptr = (uint32_t __user *)(unsigned long)(arg->count_props_ptr);
> > uint32_t __user *props_ptr = (uint32_t __user *)(unsigned long)(arg->props_ptr);
> > uint64_t __user *prop_values_ptr = (uint64_t __user *)(unsigned long)(arg->prop_values_ptr);
> > + uint32_t __user *out_fences_ptr = (uint32_t __user *)(unsigned long)(arg->out_fences_ptr);
> > unsigned int copied_objs, copied_props;
> > struct drm_atomic_state *state;
> > struct drm_modeset_acquire_ctx ctx;
> > struct drm_plane *plane;
> > struct drm_crtc *crtc;
> > struct drm_crtc_state *crtc_state;
> > + struct drm_out_fence_state *fence_state = NULL;
> > unsigned plane_mask;
> > int ret = 0;
> > unsigned int i, j;
> > @@ -1605,9 +1734,13 @@ int drm_mode_atomic_ioctl(struct drm_device *dev,
> > !dev->mode_config.async_page_flip)
> > return -EINVAL;
> >
> > + if ((arg->flags & DRM_MODE_ATOMIC_OUT_FENCE) && !arg->count_out_fences)
> > + return -EINVAL;
>
> We need testcases which check that arg->count_out_fences and
> arg->out_fences are 0 when the OUT_FENCE flag is not set.
>
> Definitely needs an igt testcase for this invalid input case. Ofc we also
> need tests that give the kernel nonsens in count_out_fences and out_fences
> with the flag set.
>
> > +
> > /* can't test and expect an event at the same time. */
> > if ((arg->flags & DRM_MODE_ATOMIC_TEST_ONLY) &&
> > - (arg->flags & DRM_MODE_PAGE_FLIP_EVENT))
> > + (arg->flags & (DRM_MODE_PAGE_FLIP_EVENT
> > + | DRM_MODE_ATOMIC_OUT_FENCE)))
>
> If you go with my suggestion above to create the event if either is set,
> maybe a DRM_MODE_ATOMIC_EVENT_MASK with both? Would read easier.
Ok.
Gustavo
More information about the dri-devel
mailing list