drm/atomic+msm: add helper to implement legacy dirtyfb

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
        if (state->fb)
                drm_framebuffer_get(state->fb);
 
+       state->dirty = false;
        state->fence = NULL;
        state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                             struct drm_file *file_priv, unsigned flags,
+                             unsigned color, struct drm_clip_rect *clips,
+                             unsigned num_clips)
+{
+       struct drm_modeset_acquire_ctx ctx;
+       struct drm_atomic_state *state;
+       struct drm_plane *plane;
+       int ret = 0;
+
+       /*
+        * When called from ioctl, we are interruptable, but not when
+        * called internally (ie. defio worker)
+        */
+       drm_modeset_acquire_init(&ctx,
+               file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+       state = drm_atomic_state_alloc(fb->dev);
+       if (!state) {
+               ret = -ENOMEM;
+               goto out;
+       }
+       state->acquire_ctx = &ctx;
+
+retry:
+       drm_for_each_plane(plane, fb->dev) {
+               struct drm_plane_state *plane_state;
+
+               if (plane->state->fb != fb)
+                       continue;
+
+               plane_state = drm_atomic_get_plane_state(state, plane);
+               if (IS_ERR(plane_state)) {
+                       ret = PTR_ERR(plane_state);
+                       goto out;
+               }
+
+               plane_state->dirty = true;
+       }
+
+       ret = drm_atomic_nonblocking_commit(state);
+
+out:
+       if (ret == -EDEADLK) {
+               drm_atomic_state_clear(state);
+               ret = drm_modeset_backoff(&ctx);
+               if (!ret)
+                       goto retry;
+       }
+
+       drm_atomic_state_put(state);
+
+       drm_modeset_drop_locks(&ctx);
+       drm_modeset_acquire_fini(&ctx);
+
+       return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +4 -1

index bf5f8c39f34d..bb55a048e98b 100644
@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
         * Figure out what fence to wait for:
         */
        for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-               if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+               bool sync_fb = new_plane_state->fb &&
+                       ((new_plane_state->fb != old_plane_state->fb) ||
+                        new_plane_state->dirty);
+               if (sync_fb) {
                        struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
                        struct msm_gem_object *msm_obj = to_msm_bo(obj);
                        struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
        .create_handle = msm_framebuffer_create_handle,
        .destroy = msm_framebuffer_destroy,
+       .dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
                                       u16 *red, u16 *green, u16 *blue,
                                       uint32_t size,
                                       struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                             struct drm_file *file_priv, unsigned flags,
+                             unsigned color, struct drm_clip_rect *clips,
+                             unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
                                                     struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -1

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
         */
        struct drm_framebuffer *fb;
 
+       /**
+        * @dirty:
+        *
+        * Flag that indicates the fb contents have changed even though the
+        * fb has not.  This is mostly a stop-gap solution until we have
+        * atomic dirty-rect(s) property.
+        */
+       bool dirty;
+
        /**
         * @fence:
         *
--

2.14.3

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the need for manual upload is indicated by a change in fb attached to the plane.

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

Quoted 76 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

Quoted 8 lines

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the
need for manual upload is indicated by a change in fb attached to the plane.

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks, because the buffer should be pinned already.

Quoted 112 lines · Daniel Vetter

Quoted 76 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support.  This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place.  But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit.  (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.)  The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things.  Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution.  Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

Quoted 8 lines

OK, let me rephrase:

 From the driver's point-of-view, in the atomic world, new content and the need for manual upload is indicated by a change in fb attached to the plane.

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Quoted 6 lines

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks, because the buffer should be pinned already.

I'm a bit confused.

Apparently we have different opinions in when an uploading driver should assume altered plane content and the need for re-upload.
vmwgfx is from what I know currently assuming that this happens only on changed fb attachment (what we call a page-flip) whereas if I understand you correctly it should happen on each atomic state commit?

If we should assume the latter, then it has odd implications, let's say you have 8 screens up, and you pan one of them on the large fb, why would you upload the contents of the other 7?

Likewise for cursors,  why would you want to upload the cursor image on each cursor move?

So in my POW, option 1) is the option that aligns with the current vmwgfx implementation and from what I can tell, what Rob has implemented in his patch.

/Thomas

Quoted 166 lines · Thomas Hellstrom

Quoted 144 lines · Daniel Vetter

Quoted 105 lines · Thomas Hellstrom

Quoted 93 lines · Daniel Vetter

Quoted 61 lines · Thomas Hellstrom

Hi,

Quoted 36 lines · Daniel Vetter

Quoted 34 lines · Rob Clark

Add an atomic helper to implement dirtyfb support.  This is needed to
support DSI command-mode panels with x11 userspace (ie. when we can't
rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different
folks working on getting
upstream kernel working on various different phones/tablets with qcom
SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place.  But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting
various CTL.FLUSH bits
with setting the CTL.START bit.  (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.)  The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things.  Hence adding an atomic dirtyfb helper.

Quoted 8 lines

Quoted 5 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the
final atomic dirty-
rect property solution.  Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full
dirty helpers already somewhere.
-Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob
on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to
get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the
dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the
damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 7 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like
you do here, or provide a content_age state member. Since we clear the dirty
flag on state copies, I guess that would be sufficient. The blob rectangles
would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the
dirty rects have always just been a hint, and that the driver was always
free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in
different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to
upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 8 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

I think I like the first approach best, although it may be tempting for
user-space apps to just set the dirty bit instead of providing the full
damage region.

Quoted 4 lines

Or I'm not following you here, because I don't quite see the difference
between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

 From the driver's point-of-view, in the atomic world, new content and the
need for manual upload is indicated by a change in fb attached to the plane.

Quoted 5 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and
the need for manual upload is identified by the dirty flag, (since the fb
stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same
way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 13 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate
front buffer rendering through the atomic ioctl:

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same
underlying buffer object.

Quoted 3 lines

Are you saying that people are already using 3) and we should keep using
that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working
on the fb, not the plane/crtc you want to upload) and lack of drm_event to
confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Quoted 8 lines

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at
making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks,
because the buffer should be pinned already.

I'm a bit confused.

Apparently we have different opinions in when an uploading driver should assume altered plane content and the need for re-upload. vmwgfx is from what I know currently assuming that this happens only on changed fb attachment (what we call a page-flip) whereas if I understand you correctly it should happen on each atomic state commit?

If we should assume the latter, then it has odd implications, let's say you have 8 screens up, and you pan one of them on the large fb, why would you upload the contents of the other 7?

Quoted 7 lines

Likewise for cursors,  why would you want to upload the cursor image on each cursor move?

So in my POW, option 1) is the option that aligns with the current vmwgfx implementation and from what I can tell, what Rob has implemented in his patch.

Hmm.

After doing some apparently well needed reading up on the code it looks like vmwgfx is actually doing a full upload on each plane state change, only those planes that actually got changed are referenced in the update. So that takes care of the panning example, assuming user-space is smart enough to leave the unchanged planes / crtcs out of the update.

However, the cursor example still holds, and IMHO we should have a better way to define content change than plane state update...

/Thomas

Quoted 3 lines

/Thomas

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 82 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the need for manual upload is indicated by a change in fb attached to the plane.

Quoted 3 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks, because the buffer should be pinned already.

So imo for a quick fix I think we need:
1) Fix drivers (or at least msm) to upload buffers for manual upload on all flips (well, all screen updates). Probably should do the fence waiting unconditionally too (and handle cursors with the new async stuff).
2) Tiny helper that remaps a dirtyfb call into a nonblocking atomic commit, which currently means we're throwing the dirty rect optimization into the wind. But that could easily be added to the drm_plane_state, without exposing it to userspace as a property just yet.

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 83 lines · Daniel Vetter

Quoted 82 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the need for manual upload is indicated by a change in fb attached to the plane.

Quoted 8 lines

Quoted 3 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Userspace could do a pageflip, but (with buffer-age extension) only re-render part of the frame. So there is a use-case for full frame pageflip with hints about what region(s) changed since previous frame.

(Of course userspace could screw that up and get different results on "command" vs "video" style display.. in that case, it gets to keep both pieces)

BR,
-R

Quoted 38 lines

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks, because the buffer should be pinned already.

So imo for a quick fix I think we need:
1) Fix drivers (or at least msm) to upload buffers for manual upload on all flips (well, all screen updates). Probably should do the fence waiting unconditionally too (and handle cursors with the new async stuff).
2) Tiny helper that remaps a dirtyfb call into a nonblocking atomic commit, which currently means we're throwing the dirty rect optimization into the wind. But that could easily be added to the drm_plane_state, without exposing it to userspace as a property just yet.

Quoted 5 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 93 lines · Rob Clark

Quoted 83 lines · Daniel Vetter

Quoted 82 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the need for manual upload is indicated by a change in fb attached to the plane.

Quoted 8 lines

Quoted 3 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 7 lines

Userspace could do a pageflip, but (with buffer-age extension) only re-render part of the frame. So there is a use-case for full frame pageflip with hints about what region(s) changed since previous frame.

(Of course userspace could screw that up and get different results on "command" vs "video" style display.. in that case, it gets to keep both pieces)

I'm not against dirty on flips with the same buffer as an optimization. Disagreement here is around whether a flip to the same buffer means:
a) nothing changed, upload nothing
b) entire buffer might have changed

I'm proposing we standardize on b) (with maybe the exception of legacy cursor ioctls because those are special) and add the dirty rects as an option to optimize this more.

Either way you'd always have to follow the implicit fence (which is what your msm hunk seems to be doing, but only for dirtyfb). -Daniel

Quoted 48 lines

BR,
-R

Quoted 38 lines

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks, because the buffer should be pinned already.

So imo for a quick fix I think we need:
1) Fix drivers (or at least msm) to upload buffers for manual upload on all flips (well, all screen updates). Probably should do the fence waiting unconditionally too (and handle cursors with the new async stuff).
2) Tiny helper that remaps a dirtyfb call into a nonblocking atomic commit, which currently means we're throwing the dirty rect optimization into the wind. But that could easily be added to the drm_plane_state, without exposing it to userspace as a property just yet.

Quoted 5 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 102 lines · Daniel Vetter

Quoted 93 lines · Rob Clark

Quoted 83 lines · Daniel Vetter

Quoted 82 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the need for manual upload is indicated by a change in fb attached to the plane.

Quoted 8 lines

Quoted 3 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 7 lines

Userspace could do a pageflip, but (with buffer-age extension) only re-render part of the frame. So there is a use-case for full frame pageflip with hints about what region(s) changed since previous frame.

(Of course userspace could screw that up and get different results on "command" vs "video" style display.. in that case, it gets to keep both pieces)

Quoted 8 lines

I'm not against dirty on flips with the same buffer as an optimization. Disagreement here is around whether a flip to the same buffer means:
a) nothing changed, upload nothing
b) entire buffer might have changed

I'm proposing we standardize on b) (with maybe the exception of legacy cursor ioctls because those are special) and add the dirty rects as an option to optimize this more.

ok, I can agree w/ flip to same buffer, if FB_ID is in the incoming property list.. I just want a way to differentiate that from "plane got added to new state for unrelated reasons".. so new_fb != old_fb isn't a good test for "should we synchronize", but new plane state isn't either.

Keeping the dirty flag (or sequence count or whatever mechanism) to indicate that an FB_ID that happened to be the same fb was part of what userspace passed in would do the trick.

BR,
-R

Quoted 57 lines

Either way you'd always have to follow the implicit fence (which is what your msm hunk seems to be doing, but only for dirtyfb). -Daniel

Quoted 48 lines

BR,
-R

Quoted 38 lines

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Also note that atomic helpers don't optimize out plane flips for same buffers. We only optimize out some of the waiting, in a failed attempt at making cursors stall less, but that's not fixed with the async plane update stuff. And we can obviously optimize out the prepare/cleanup hooks, because the buffer should be pinned already.

So imo for a quick fix I think we need:
1) Fix drivers (or at least msm) to upload buffers for manual upload on all flips (well, all screen updates). Probably should do the fence waiting unconditionally too (and handle cursors with the new async stuff).
2) Tiny helper that remaps a dirtyfb call into a nonblocking atomic commit, which currently means we're throwing the dirty rect optimization into the wind. But that could easily be added to the drm_plane_state, without exposing it to userspace as a property just yet.

Quoted 5 lines

Cheers, Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 76 lines · Thomas Hellstrom

Quoted 69 lines · Daniel Vetter

Quoted 49 lines · Thomas Hellstrom

Hi,

Quoted 30 lines · Daniel Vetter

Quoted 29 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 6 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere. -Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 4 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like you do here, or provide a content_age state member. Since we clear the dirty flag on state copies, I guess that would be sufficient. The blob rectangles would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 6 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

Quoted 8 lines

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and the
need for manual upload is indicated by a change in fb attached to the plane.

With Rob's patch here, (correct me if I'm wrong) in addition new content and the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

Quoted 132 lines · Daniel Vetter

Quoted 105 lines · Thomas Hellstrom

Quoted 88 lines · Daniel Vetter

Quoted 60 lines · Thomas Hellstrom

Hi,

Quoted 33 lines · Daniel Vetter

Quoted 30 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 8 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere.
-Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob
on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to
get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the
dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the
damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 7 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like
you do here, or provide a content_age state member. Since we clear the dirty
flag on state copies, I guess that would be sufficient. The blob rectangles
would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the
dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in
different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to
upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 7 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and
the
need for manual upload is indicated by a change in fb attached to the plane.

Quoted 4 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and
the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 8 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same
underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

Quoted 9 lines

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

FWIW, vmwgfx has always treated a dirtyfb as a pure front-buffer like rendering operation without any synchronization,
We've guaranteed that only the rects that are present are uploaded, but only xf86-video-vmware has taken advantage of this to keep
CPU- and GPU rendered content apart.
I think we've at some point run into problems with number of cliprects, (Old KDE lock screen?) and use multiple dirtyfb for the same update...

IIRC the reason for working with the fb, is that it's much easier for user-space, which doesn't have to care where planes are scanning out and why.
"Present my new content on any screen that's affected".

Quoted 7 lines

Quoted 7 lines

Quoted 7 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same
underlying buffer object.

Are you saying that people are already using 3) and we should keep

using

Quoted 8 lines

Quoted 8 lines

Quoted 1 line

that?

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event

to

Quoted 1 line

Quoted 1 line

confirm when exactly you uploaded your stuff. But imo they should be

the

Quoted 6 lines

Quoted 2 lines

same underlying operation.

FWIW, vmwgfx has always treated a dirtyfb as a pure front-buffer like rendering operation without any synchronization,
We've guaranteed that only the rects that are present are uploaded, but

only

Quoted 3 lines

xf86-video-vmware has taken advantage of this to keep
CPU- and GPU rendered content apart.
I think we've at some point run into problems with number of cliprects,

(Old

Quoted 2 lines

KDE lock screen?) and use multiple dirtyfb for the same update...

Ok, if we can hit this in practices then I think it's ok to have the limit. Just need to make sure userspace actually condenses the cliprects down to something within the limit, since with atomic flips you can't just do multiple updates - that would tear badly.

Wrt not syncing: I think general use pretty clearly says lots of userspace renders into buffers with gpus (not even necessarily your own) and then expects dirtyfb or a flip to upload all the bits. Otherwise Rob Clark wouldn't need his patch. Given that I think we need to make general semantics follow that requirement - I don't expect it'll harm vmwgfx since it doesn't render into the frontbuffer anyway?

Quoted 4 lines

IIRC the reason for working with the fb, is that it's much easier for user-space, which doesn't have to care where planes are scanning out and why.
"Present my new content on any screen that's affected".

Yeah, dirtyfb makes tons of sense for frontbuffer-rendering X, which also doesn't do per-scanout pixmaps. But for atomic flips you really want to flip on the crtc (or well plane), since otherwise with multiple planes it comes up all teared up. vmwgfx doesn't care I guess since you only have 1 primary plane, but all the SoC chips very much do.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - https://urldefense.proofpoint.com/v2/url?u=http-
3A__blog.ffwll.ch&d=DwIBaQ&c=uilaK90D4TOVoH58JNXRgQ&r=zOOG28inJK 0762SxAf-cyZdStnd2NQpRu98lJP2iYGw&m=XKgN7GPElFapBWdozPSC- 9rcfKmDvQC1QHhsFghexWc&s=SH9q5tw- UqpUBJVrr2v1mLpRo28Aau7iJ3YWlrgbPmU&e=

Quoted 132 lines · Daniel Vetter

Quoted 105 lines · Thomas Hellstrom

Quoted 88 lines · Daniel Vetter

Quoted 60 lines · Thomas Hellstrom

Hi,

Quoted 33 lines · Daniel Vetter

Quoted 30 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 8 lines

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Adding Noralf, who iirc already posted the full dirty helpers already somewhere.
-Daniel

I've asked Deepak to RFC the core changes suggested for the full dirty blob
on dri-devel. It builds on DisplayLink's suggestion, with a simple helper to
get to the desired coordinates.

One thing to perhaps discuss is how we would like to fit this with front-buffer rendering and the dirty ioctl. In the page-flip context, the
dirty rects, like egl's swapbuffer_with_damage is a hint to restrict the
damage region that can be fully ignored by the driver, new content is indicated by a new framebuffer.

Quoted 7 lines

We could do the same for frontbuffer rendering: Either set a dirty flag like
you do here, or provide a content_age state member. Since we clear the dirty
flag on state copies, I guess that would be sufficient. The blob rectangles
would then become a hint to restrict the damage region.

I'm not entirely following here - I thought for frontbuffer rendering the
dirty rects have always just been a hint, and that the driver was always free to re-upload the entire buffer to the screen.

And through a helper like Rob's proposing here (and have floated around in
different versions already) we'd essentially map a frontbuffer dirtyfb call to a fake flip with dirty rect. Manual upload drivers already need to
upload the entire screen if they get a flip, since some userspace uses that to flush out frontbuffer rendering (instead of calling dirtyfb).

So from that pov the new dirty flag is kinda not necessary imo.

Quoted 3 lines

Another approach would be to have the presence of dirty rects without framebuffer change to indicate frontbuffer rendering.

Quoted 7 lines

Quoted 3 lines

I think I like the first approach best, although it may be tempting for user-space apps to just set the dirty bit instead of providing the full damage region.

Or I'm not following you here, because I don't quite see the difference between dirtyfb and a flip.
-Daniel

OK, let me rephrase:

From the driver's point-of-view, in the atomic world, new content and
the
need for manual upload is indicated by a change in fb attached to the plane.

Quoted 4 lines

With Rob's patch here, (correct me if I'm wrong) in addition new content and
the need for manual upload is identified by the dirty flag, (since the fb stays the same and the driver thus never identifies a page-flip).

Hm, I'm not entirely sure Rob's approach is correct. Imo a pageflip (atomic or not) should result in the entire buffer getting uploaded. The dirty flag is kinda redundant, a flip with the same buffer works the same way as a dirtyfb with the entire buffer as the dirty rectangle.

Quoted 6 lines

In both these situations, clip rects can provide a hint to restrict the dirty region.

Now I was thinking about the preferred way for user-space to communicate front buffer rendering through the atomic ioctl:

Quoted 8 lines

1) Expose a dirty (or content_age property)
2) Attach a clip-rect blob property.
3) Fake a page-flip by ping-ponging two frame-buffers pointing to the same
underlying buffer object.

Are you saying that people are already using 3) and we should keep using that?

Quoted 9 lines

I'm saying they're using 3b), flip the same buffer wrapped in the same drm_framebuffer, and expect it to work.

The only advantage dirtyfb has is that it allows you to supply the optimized upload rectangles, but at the cost of a funny api (it's working on the fb, not the plane/crtc you want to upload) and lack of drm_event to confirm when exactly you uploaded your stuff. But imo they should be the same underlying operation.

FWIW, vmwgfx has always treated a dirtyfb as a pure front-buffer like rendering operation without any synchronization,

We've guaranteed that only the rects that are present are uploaded, but only xf86-video-vmware has taken advantage of this to keep
CPU- and GPU rendered content apart.
I think we've at some point run into problems with number of cliprects, (Old KDE lock screen?) and use multiple dirtyfb for the same update...

IIRC the reason for working with the fb, is that it's much easier for user-space, which doesn't have to care where planes are scanning out and why.
"Present my new content on any screen that's affected".

/Thomas

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
 	if (state->fb)
 		drm_framebuffer_get(state->fb);
 
+	state->dirty = false;
 	state->fence = NULL;
 	state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips)
+{
+	struct drm_modeset_acquire_ctx ctx;
+	struct drm_atomic_state *state;
+	struct drm_plane *plane;
+	int ret = 0;
+
+	/*
+	 * When called from ioctl, we are interruptable, but not when
+	 * called internally (ie. defio worker)
+	 */
+	drm_modeset_acquire_init(&ctx,
+		file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+	state = drm_atomic_state_alloc(fb->dev);
+	if (!state) {
+		ret = -ENOMEM;
+		goto out;
+	}
+	state->acquire_ctx = &ctx;
+
+retry:
+	drm_for_each_plane(plane, fb->dev) {
+		struct drm_plane_state *plane_state;
+
+		if (plane->state->fb != fb)
+			continue;
+
+		plane_state = drm_atomic_get_plane_state(state, plane);
+		if (IS_ERR(plane_state)) {
+			ret = PTR_ERR(plane_state);
+			goto out;
+		}
+
+		plane_state->dirty = true;
+	}
+
+	ret = drm_atomic_nonblocking_commit(state);
+
+out:
+	if (ret == -EDEADLK) {
+		drm_atomic_state_clear(state);
+		ret = drm_modeset_backoff(&ctx);
+		if (!ret)
+			goto retry;
+	}
+
+	drm_atomic_state_put(state);
+
+	drm_modeset_drop_locks(&ctx);
+	drm_modeset_acquire_fini(&ctx);
+
+	return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c +1 -0

+		if (sync_fb) {
 			struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
 			struct msm_gem_object *msm_obj = to_msm_bo(obj);
 			struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
 	.create_handle = msm_framebuffer_create_handle,
 	.destroy = msm_framebuffer_destroy,
+	.dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 				       u16 *red, u16 *green, u16 *blue,
 				       uint32_t size,
 				       struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
 						     struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
 	 */
 	struct drm_framebuffer *fb;
 
+	/**
+	 * @dirty:
+	 *
+	 * Flag that indicates the fb contents have changed even though the
+	 * fb has not.  This is mostly a stop-gap solution until we have
+	 * atomic dirty-rect(s) property.
+	 */
+	bool dirty;
+
 	/**
 	 * @fence:
 	 *

Signature · 1 line

2.14.3

Quoted 130 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
 	if (state->fb)
 		drm_framebuffer_get(state->fb);
 
+	state->dirty = false;
 	state->fence = NULL;
 	state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips)
+{
+	struct drm_modeset_acquire_ctx ctx;
+	struct drm_atomic_state *state;
+	struct drm_plane *plane;
+	int ret = 0;
+
+	/*
+	 * When called from ioctl, we are interruptable, but not when
+	 * called internally (ie. defio worker)
+	 */
+	drm_modeset_acquire_init(&ctx,
+		file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+	state = drm_atomic_state_alloc(fb->dev);
+	if (!state) {
+		ret = -ENOMEM;
+		goto out;
+	}
+	state->acquire_ctx = &ctx;
+
+retry:
+	drm_for_each_plane(plane, fb->dev) {
+		struct drm_plane_state *plane_state;
+
+		if (plane->state->fb != fb)
+			continue;
+
+		plane_state = drm_atomic_get_plane_state(state, plane);
+		if (IS_ERR(plane_state)) {
+			ret = PTR_ERR(plane_state);
+			goto out;
+		}
+
+		plane_state->dirty = true;
+	}
+
+	ret = drm_atomic_nonblocking_commit(state);
+
+out:
+	if (ret == -EDEADLK) {
+		drm_atomic_state_clear(state);
+		ret = drm_modeset_backoff(&ctx);
+		if (!ret)
+			goto retry;
+	}
+
+	drm_atomic_state_put(state);
+
+	drm_modeset_drop_locks(&ctx);
+	drm_modeset_acquire_fini(&ctx);
+
+	return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
 	 * Figure out what fence to wait for:
 	 */
 	for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-		if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+		bool sync_fb = new_plane_state->fb &&
+			((new_plane_state->fb != old_plane_state->fb) ||
+			 new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Quoted 54 lines

drivers/gpu/drm/msm/msm_atomic.c +1 -0

+		if (sync_fb) {
 			struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
 			struct msm_gem_object *msm_obj = to_msm_bo(obj);
 			struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
 	.create_handle = msm_framebuffer_create_handle,
 	.destroy = msm_framebuffer_destroy,
+	.dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 				       u16 *red, u16 *green, u16 *blue,
 				       uint32_t size,
 				       struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
 						     struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
 	 */
 	struct drm_framebuffer *fb;
 
+	/**
+	 * @dirty:
+	 *
+	 * Flag that indicates the fb contents have changed even though the
+	 * fb has not.  This is mostly a stop-gap solution until we have
+	 * atomic dirty-rect(s) property.
+	 */
+	bool dirty;
+
 	/**
 	 * @fence:
 	 *

Signature · 1 line

2.14.3

Signature · 3 lines

Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 142 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
 	if (state->fb)
 		drm_framebuffer_get(state->fb);
 
+	state->dirty = false;
 	state->fence = NULL;
 	state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips)
+{
+	struct drm_modeset_acquire_ctx ctx;
+	struct drm_atomic_state *state;
+	struct drm_plane *plane;
+	int ret = 0;
+
+	/*
+	 * When called from ioctl, we are interruptable, but not when
+	 * called internally (ie. defio worker)
+	 */
+	drm_modeset_acquire_init(&ctx,
+		file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+	state = drm_atomic_state_alloc(fb->dev);
+	if (!state) {
+		ret = -ENOMEM;
+		goto out;
+	}
+	state->acquire_ctx = &ctx;
+
+retry:
+	drm_for_each_plane(plane, fb->dev) {
+		struct drm_plane_state *plane_state;
+
+		if (plane->state->fb != fb)
+			continue;
+
+		plane_state = drm_atomic_get_plane_state(state, plane);
+		if (IS_ERR(plane_state)) {
+			ret = PTR_ERR(plane_state);
+			goto out;
+		}
+
+		plane_state->dirty = true;
+	}
+
+	ret = drm_atomic_nonblocking_commit(state);
+
+out:
+	if (ret == -EDEADLK) {
+		drm_atomic_state_clear(state);
+		ret = drm_modeset_backoff(&ctx);
+		if (!ret)
+			goto retry;
+	}
+
+	drm_atomic_state_put(state);
+
+	drm_modeset_drop_locks(&ctx);
+	drm_modeset_acquire_fini(&ctx);
+
+	return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
 	 * Figure out what fence to wait for:
 	 */
 	for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-		if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+		bool sync_fb = new_plane_state->fb &&
+			((new_plane_state->fb != old_plane_state->fb) ||
+			 new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

- most arm-soc drivers do have this "optimization" in their code, and it even managed to get into the new drm_gem_fb_prepare_fb helper (which I reviewed, or well claimed to have ... oops). Afaict it goes back to the original msm atomic code, and was then dutifully copypasted all over the place.

If folks are ok I'll do a patch series to align drivers with i915/nouveau. Well, any driver using reservation_object_get_excl_rcu + drm_atomic_set_fence_for_plane combo, since amdgpu and vmwgfx don't I have no idea what they're doing or whether they might have the same bug.

From looking at at least the various prepare_fb callbacks I don't see any other drivers doing funny stuff around implicit fences. -Daniel

Quoted 58 lines

Quoted 54 lines

drivers/gpu/drm/msm/msm_atomic.c +1 -0

+		if (sync_fb) {
 			struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
 			struct msm_gem_object *msm_obj = to_msm_bo(obj);
 			struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
 	.create_handle = msm_framebuffer_create_handle,
 	.destroy = msm_framebuffer_destroy,
+	.dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 				       u16 *red, u16 *green, u16 *blue,
 				       uint32_t size,
 				       struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
 						     struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
 	 */
 	struct drm_framebuffer *fb;
 
+	/**
+	 * @dirty:
+	 *
+	 * Flag that indicates the fb contents have changed even though the
+	 * fb has not.  This is mostly a stop-gap solution until we have
+	 * atomic dirty-rect(s) property.
+	 */
+	bool dirty;
+
 	/**
 	 * @fence:
 	 *

Signature · 1 line

2.14.3

Signature · 3 lines

Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 151 lines · Daniel Vetter

Quoted 150 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Quoted 6 lines

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

Quoted 65 lines

- most arm-soc drivers do have this "optimization" in their code, and it even managed to get into the new drm_gem_fb_prepare_fb helper (which I reviewed, or well claimed to have ... oops). Afaict it goes back to the original msm atomic code, and was then dutifully copypasted all over the place.

If folks are ok I'll do a patch series to align drivers with i915/nouveau. Well, any driver using reservation_object_get_excl_rcu + drm_atomic_set_fence_for_plane combo, since amdgpu and vmwgfx don't I have no idea what they're doing or whether they might have the same bug.

From looking at at least the various prepare_fb callbacks I don't see any other drivers doing funny stuff around implicit fences. -Daniel

Quoted 50 lines

Quoted 50 lines

drivers/gpu/drm/msm/msm_atomic.c +1 -0

+           if (sync_fb) {
                    struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
                    struct msm_gem_object *msm_obj = to_msm_bo(obj);
                    struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
    .create_handle = msm_framebuffer_create_handle,
    .destroy = msm_framebuffer_destroy,
+   .dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
                                   u16 *red, u16 *green, u16 *blue,
                                   uint32_t size,
                                   struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
                                                 struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
     */
    struct drm_framebuffer *fb;
 
+   /**
+    * @dirty:
+    *
+    * Flag that indicates the fb contents have changed even though the
+    * fb has not.  This is mostly a stop-gap solution until we have
+    * atomic dirty-rect(s) property.
+    */
+   bool dirty;
+
    /**
     * @fence:

Quoted 8 lines

Quoted 8 lines

Quoted 4 lines

include/drm/drm_plane.h +0 -1

     *
--

2.14.3

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 154 lines · Maarten Lankhorst

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 153 lines · Daniel Vetter

Quoted 142 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

Quoted 6 lines

Quoted 4 lines

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

for implicit fencing we fish out the implicit fence and stuff it in plane_state->fence..

Op 04-04-18 om 13:37 schreef Rob Clark:

Quoted 160 lines

Quoted 158 lines · Maarten Lankhorst

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 157 lines · Daniel Vetter

Quoted 142 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

Quoted 4 lines

Quoted 2 lines

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

for implicit fencing we fish out the implicit fence and stuff it in plane_state->fence..

What happens when userspace passes a fence fd to in_fence_fd?

Quoted 161 lines · Maarten Lankhorst

Op 04-04-18 om 13:37 schreef Rob Clark:

Quoted 160 lines

Quoted 158 lines · Maarten Lankhorst

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 157 lines · Daniel Vetter

Quoted 142 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

Quoted 5 lines

Quoted 4 lines

Quoted 2 lines

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

for implicit fencing we fish out the implicit fence and stuff it in plane_state->fence..

What happens when userspace passes a fence fd to in_fence_fd?

mixing implicit sync and explicit sync is undefined

Quoted 161 lines · Maarten Lankhorst

Op 04-04-18 om 13:37 schreef Rob Clark:

Quoted 160 lines

Quoted 158 lines · Maarten Lankhorst

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 157 lines · Daniel Vetter

Quoted 142 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

Quoted 5 lines

Quoted 4 lines

Quoted 2 lines

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

for implicit fencing we fish out the implicit fence and stuff it in plane_state->fence..

What happens when userspace passes a fence fd to in_fence_fd?

mixing implicit sync and explicit sync is undefined

BR,
-R
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 165 lines

Quoted 163 lines · Maarten Lankhorst

Op 04-04-18 om 13:37 schreef Rob Clark:

Quoted 162 lines

Quoted 160 lines · Maarten Lankhorst

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 151 lines · Daniel Vetter

Quoted 142 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Quoted 6 lines

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

Quoted 4 lines

Quoted 3 lines

Quoted 2 lines

for implicit fencing we fish out the implicit fence and stuff it in plane_state->fence..

What happens when userspace passes a fence fd to in_fence_fd?

mixing implicit sync and explicit sync is undefined

It's very well-defined, explicit sync wins. See the kerneldoc for drm_atomic_set_fence_for_plane. Since a pile of drivers don't use that the reality might be undefined though :-)

Quoted 167 lines · Rob Clark

Quoted 165 lines · Maarten Lankhorst

Op 04-04-18 om 13:37 schreef Rob Clark:

Quoted 164 lines

Quoted 152 lines · Maarten Lankhorst

Op 04-04-18 om 12:21 schreef Daniel Vetter:

Quoted 151 lines · Daniel Vetter

Quoted 150 lines · Rob Clark

Quoted 130 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Quoted 8 lines

Quoted 6 lines

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.

I thought plane_state->fence was used for explicit fences, so its use by drivers would interfere with it? I don't think fencing would work on msm or vc4..

for implicit fencing we fish out the implicit fence and stuff it in plane_state->fence..

Quoted 3 lines

Quoted 1 line

What happens when userspace passes a fence fd to in_fence_fd?

mixing implicit sync and explicit sync is undefined

It's very well-defined, explicit sync wins. See the kerneldoc for drm_atomic_set_fence_for_plane. Since a pile of drivers don't use that the reality might be undefined though :-)

-Daniel

Quoted 7 lines

BR,
-R
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Quoted 143 lines · Daniel Vetter

Quoted 130 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.
- most arm-soc drivers do have this "optimization" in their code, and it even managed to get into the new drm_gem_fb_prepare_fb helper (which I reviewed, or well claimed to have ... oops). Afaict it goes back to the original msm atomic code, and was then dutifully copypasted all over the place.

If folks are ok I'll do a patch series to align drivers with i915/nouveau. Well, any driver using reservation_object_get_excl_rcu + drm_atomic_set_fence_for_plane combo, since amdgpu and vmwgfx don't I have no idea what they're doing or whether they might have the same bug.

From looking at at least the various prepare_fb callbacks I don't see any other drivers doing funny stuff around implicit fences. -Daniel

Quoted 51 lines

Quoted 51 lines

drivers/gpu/drm/msm/msm_atomic.c +1 -0

+           if (sync_fb) {
                    struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
                    struct msm_gem_object *msm_obj = to_msm_bo(obj);
                    struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
    .create_handle = msm_framebuffer_create_handle,
    .destroy = msm_framebuffer_destroy,
+   .dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
                                   u16 *red, u16 *green, u16 *blue,
                                   uint32_t size,
                                   struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
                                                 struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
     */
    struct drm_framebuffer *fb;
 
+   /**
+    * @dirty:
+    *
+    * Flag that indicates the fb contents have changed even though the
+    * fb has not.  This is mostly a stop-gap solution until we have
+    * atomic dirty-rect(s) property.
+    */
+   bool dirty;
+
    /**
     * @fence:
     *

Quoted 8 lines

Quoted 3 lines

--
2.14.3

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Quoted 165 lines · Daniel Vetter

Quoted 143 lines · Daniel Vetter

Quoted 130 lines · Rob Clark

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

 drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
 drivers/gpu/drm/msm/msm_fb.c        |  1 +
 include/drm/drm_atomic_helper.h     |  4 +++
 include/drm/drm_plane.h             |  9 +++++
 5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
    if (state->fb)
            drm_framebuffer_get(state->fb);
 
+   state->dirty = false;
    state->fence = NULL;
    state->commit = NULL;
 }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
 }
 EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips)
+{
+   struct drm_modeset_acquire_ctx ctx;
+   struct drm_atomic_state *state;
+   struct drm_plane *plane;
+   int ret = 0;
+
+   /*
+    * When called from ioctl, we are interruptable, but not when
+    * called internally (ie. defio worker)
+    */
+   drm_modeset_acquire_init(&ctx,
+           file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+   state = drm_atomic_state_alloc(fb->dev);
+   if (!state) {
+           ret = -ENOMEM;
+           goto out;
+   }
+   state->acquire_ctx = &ctx;
+
+retry:
+   drm_for_each_plane(plane, fb->dev) {
+           struct drm_plane_state *plane_state;
+
+           if (plane->state->fb != fb)
+                   continue;
+
+           plane_state = drm_atomic_get_plane_state(state, plane);
+           if (IS_ERR(plane_state)) {
+                   ret = PTR_ERR(plane_state);
+                   goto out;
+           }
+
+           plane_state->dirty = true;
+   }
+
+   ret = drm_atomic_nonblocking_commit(state);
+
+out:
+   if (ret == -EDEADLK) {
+           drm_atomic_state_clear(state);
+           ret = drm_modeset_backoff(&ctx);
+           if (!ret)
+                   goto retry;
+   }
+
+   drm_atomic_state_put(state);
+
+   drm_modeset_drop_locks(&ctx);
+   drm_modeset_acquire_fini(&ctx);
+
+   return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
 /**
  * __drm_atomic_helper_private_duplicate_state - copy atomic private state
  * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +0 -0

index bf5f8c39f34d..bb55a048e98b 100644

drivers/gpu/drm/msm/msm_atomic.c int msm_atomic_commit(struct drm_device *dev, · line 206

Quoted 7 lines

drivers/gpu/drm/msm/msm_atomic.c +3 -1

@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
     * Figure out what fence to wait for:
     */
    for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-           if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+           bool sync_fb = new_plane_state->fb &&
+                   ((new_plane_state->fb != old_plane_state->fb) ||
+                    new_plane_state->dirty);

Why do you have this optimization even here? Imo flipping to the same fb should result in the fb getting fully uploaded, whether you're doing a legacy page_flip, and atomic one or just a plane update.

Quoted 8 lines

Iirc some userspace does use that as essentially a full-plane frontbuffer rendering flush already. IOW I don't think we need your plane_state->dirty, it's implied to always be true - why would userspace do a flip otherwise?

The helper itself to map dirtyfb to a nonblocking atomic commit looks reasonable, but misses a bunch of the trickery discussed with Noralf and others I think.

Ok, I've done some history digging:

- i915 and nouveau unconditionally wait for fences, even for same-fb flips.
- no idea what amdgpu and vmwgfx are doing, they're not using plane_state->fence for implicit fences.
- most arm-soc drivers do have this "optimization" in their code, and it even managed to get into the new drm_gem_fb_prepare_fb helper (which I reviewed, or well claimed to have ... oops). Afaict it goes back to the original msm atomic code, and was then dutifully copypasted all over the place.

Quoted 4 lines

If folks are ok I'll do a patch series to align drivers with i915/nouveau. Well, any driver using reservation_object_get_excl_rcu + drm_atomic_set_fence_for_plane combo, since amdgpu and vmwgfx don't I have no idea what they're doing or whether they might have the same bug.

if you do a patch series, I think do it the other way around and align to msm ;-)

I think otherwise we could end up stalling while x11 does front buffer rendering for something unrelated (maybe I was hitting that with cursor updates?)

BR,
-R

Quoted 68 lines

From looking at at least the various prepare_fb callbacks I don't see any other drivers doing funny stuff around implicit fences. -Daniel

Quoted 51 lines

Quoted 51 lines

drivers/gpu/drm/msm/msm_atomic.c +1 -0

+           if (sync_fb) {
                    struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
                    struct msm_gem_object *msm_obj = to_msm_bo(obj);
                    struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
 static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
    .create_handle = msm_framebuffer_create_handle,
    .destroy = msm_framebuffer_destroy,
+   .dirty = drm_atomic_helper_dirtyfb,
 };
 
 #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
                                   u16 *red, u16 *green, u16 *blue,
                                   uint32_t size,
                                   struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                         struct drm_file *file_priv, unsigned flags,
+                         unsigned color, struct drm_clip_rect *clips,
+                         unsigned num_clips);
 void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
                                                 struct drm_private_state *state);
 

include/drm/drm_plane.h +9 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
     */
    struct drm_framebuffer *fb;
 
+   /**
+    * @dirty:
+    *
+    * Flag that indicates the fb contents have changed even though the
+    * fb has not.  This is mostly a stop-gap solution until we have
+    * atomic dirty-rect(s) property.
+    */
+   bool dirty;
+
    /**
     * @fence:
     *

Quoted 8 lines

Quoted 3 lines

--
2.14.3

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

  drivers/gpu/drm/drm_atomic_helper.c | 66 +++++++++++++++++++++++++++++++++++++
  drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
  drivers/gpu/drm/msm/msm_fb.c        |  1 +
  include/drm/drm_atomic_helper.h     |  4 +++
  include/drm/drm_plane.h             |  9 +++++
  5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +66 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
  	if (state->fb)
  		drm_framebuffer_get(state->fb);
 
+	state->dirty = false;
  	state->fence = NULL;
  	state->commit = NULL;
  }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
  }
  EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
 
+/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips)
+{
+	struct drm_modeset_acquire_ctx ctx;
+	struct drm_atomic_state *state;
+	struct drm_plane *plane;
+	int ret = 0;
+
+	/*
+	 * When called from ioctl, we are interruptable, but not when
+	 * called internally (ie. defio worker)
+	 */
+	drm_modeset_acquire_init(&ctx,
+		file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+	state = drm_atomic_state_alloc(fb->dev);
+	if (!state) {
+		ret = -ENOMEM;
+		goto out;
+	}
+	state->acquire_ctx = &ctx;
+
+retry:
+	drm_for_each_plane(plane, fb->dev) {
+		struct drm_plane_state *plane_state;
+
+		if (plane->state->fb != fb)
+			continue;
+
+		plane_state = drm_atomic_get_plane_state(state, plane);
+		if (IS_ERR(plane_state)) {
+			ret = PTR_ERR(plane_state);
+			goto out;
+		}
+
+		plane_state->dirty = true;
+	}
+
+	ret = drm_atomic_nonblocking_commit(state);
+
+out:
+	if (ret == -EDEADLK) {
+		drm_atomic_state_clear(state);
+		ret = drm_modeset_backoff(&ctx);
+		if (!ret)
+			goto retry;
+	}
+
+	drm_atomic_state_put(state);
+
+	drm_modeset_drop_locks(&ctx);
+	drm_modeset_acquire_fini(&ctx);
+
+	return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
  /**
   * __drm_atomic_helper_private_duplicate_state - copy atomic private state
   * @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +4 -1

index bf5f8c39f34d..bb55a048e98b 100644
@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
  	 * Figure out what fence to wait for:
  	 */
  	for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-		if ((new_plane_state->fb != old_plane_state->fb) && new_plane_state->fb) {
+		bool sync_fb = new_plane_state->fb &&
+			((new_plane_state->fb != old_plane_state->fb) ||
+			 new_plane_state->dirty);
+		if (sync_fb) {
  			struct drm_gem_object *obj = msm_framebuffer_bo(new_plane_state->fb, 0);
  			struct msm_gem_object *msm_obj = to_msm_bo(obj);
  			struct dma_fence *fence = reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct drm_framebuffer *fb)
  static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
  	.create_handle = msm_framebuffer_create_handle,
  	.destroy = msm_framebuffer_destroy,
+	.dirty = drm_atomic_helper_dirtyfb,
  };
 
  #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct drm_crtc *crtc,
  				       u16 *red, u16 *green, u16 *blue,
  				       uint32_t size,
  				       struct drm_modeset_acquire_ctx *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+			      struct drm_file *file_priv, unsigned flags,
+			      unsigned color, struct drm_clip_rect *clips,
+			      unsigned num_clips);
  void __drm_atomic_helper_private_obj_duplicate_state(struct drm_private_obj *obj,
  						     struct drm_private_state *state);
 

include/drm/drm_plane.h +2 -0

index f7bf4a48b1c3..296fa22bda7a 100644
@@ -76,6 +76,15 @@ struct drm_plane_state {
  	 */
  	struct drm_framebuffer *fb;
 
+	/**
+	 * @dirty:

Den 04.04.2018 00.42, skrev Rob Clark:

Quoted 30 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

I have been waiting for someone to address this since I'm not skilled enough myself to tackle the atomic machinery. It would be be nice to do all flushing during commit since that would mean that the tinydrm drivers could be driven solely through drm_simple_display_pipe_funcs. I wouldn't have to wire through the dirty callback like I do now.

I see that you use a nonblocking commit. What happens if a new dirtyfb comes in before the previous is done?

If we could save the dirty clips, then I could use this in tinydrm. A full flush over SPI takes ~50ms so I need to shave off where I can.

This works for me in my tinydrm world:

include/drm/drm_plane.h +4 -0

index c64225274807..218cb36757fa 100644
@@ -28,6 +28,7 @@
 #include <drm/drm_mode_object.h>
 #include <drm/drm_color_mgmt.h>
 
+struct drm_clip_rect;
 struct drm_crtc;
 struct drm_printer;
 struct drm_modeset_acquire_ctx;

@@ -85,6 +86,9 @@ struct drm_plane_state {
         */
        bool dirty;
 
+       struct drm_clip_rect *dirty_clips;
+       unsigned int num_dirty_clips;
+
        /**
         * @fence:
         *

drivers/gpu/drm/drm_atomic_helper.c +2 -0

index 8066c508ab50..e00b8247b7c5 100644
@@ -3521,6 +3521,8 @@ void __drm_atomic_helper_plane_duplicate_state(struct
 drm_plane *plane,
                drm_framebuffer_get(state->fb);
 
        state->dirty = false;
+       state->dirty_clips = NULL;
+       state->num_dirty_clips = 0;
        state->fence = NULL;
        state->commit = NULL;

}

drivers/gpu/drm/drm_atomic_helper.c +14 -0

@@ -3567,6 +3569,8 @@ void __drm_atomic_helper_plane_destroy_state(struct
 drm_plane_state *state)
 
        if (state->commit)
                drm_crtc_commit_put(state->commit);
+
+       kfree(state->dirty_clips);
 }
 EXPORT_SYMBOL(__drm_atomic_helper_plane_destroy_state);

@@ -3877,8 +3881,20 @@ int drm_atomic_helper_dirtyfb(struct drm_framebuffer
 *fb,
        struct drm_modeset_acquire_ctx ctx;
        struct drm_atomic_state *state;
        struct drm_plane *plane;
+       bool fb_found = false;
        int ret = 0;
 
+       /* fbdev can flush even when we don't want it to */
+       drm_for_each_plane(plane, fb->dev) {
+               if (plane->state->fb == fb) {
+                       fb_found = true;
+                       break;
+               }
+       }
+
+       if (!fb_found)
+               return 0;
+
        /*
         * When called from ioctl, we are interruptable, but not when

* called internally (ie. defio worker)

drivers/gpu/drm/drm_atomic_helper.c +8 -0

@@ -3907,6 +3923,25 @@ int drm_atomic_helper_dirtyfb(struct drm_framebuffer
 *fb,
                }
 
                plane_state->dirty = true;
+               if (clips && num_clips)
+                       plane_state->dirty_clips = kmemdup(clips, num_clips
 * sizeof(*clips),
+ GFP_KERNEL);
+               else
+                       plane_state->dirty_clips = kzalloc(sizeof(*clips),
 GFP_KERNEL);
+
+               if (!plane_state->dirty_clips) {
+                       ret = -ENOMEM;

drivers/gpu/drm/drm_atomic_helper.c +1 -0

+               }

        }

        ret = drm_atomic_nonblocking_commit(state);

drivers/gpu/drm/tinydrm/core/tinydrm-pipe.c +5 -3

index e68b528ae64d..1a0c72c496f6 100644
@@ -121,12 +121,14 @@ void tinydrm_display_pipe_update(struct
 drm_simple_display_pipe *pipe,
                                 struct drm_plane_state *old_state)
 {
        struct tinydrm_device *tdev = pipe_to_tinydrm(pipe);
-       struct drm_framebuffer *fb = pipe->plane.state->fb;
+       struct drm_plane_state *new_state = pipe->plane.state;
+       struct drm_framebuffer *fb = new_state->fb;
        struct drm_crtc *crtc = &tdev->pipe.crtc;
 
-       if (fb && (fb != old_state->fb)) {
+       if (fb && ((fb != old_state->fb) || new_state->dirty_clips)) {
                if (tdev->fb_dirty)
-                       tdev->fb_dirty(fb, NULL, 0, 0, NULL, 0);
+                       tdev->fb_dirty(fb, NULL, 0, 0,
 new_state->dirty_clips,
+ new_state->num_dirty_clips);
        }

        if (crtc->state->event) {

drivers/gpu/drm/tinydrm/mipi-dbi.c +2 -1

index 4d1fb31a781f..49dee24dde60 100644
@@ -9,6 +9,7 @@
  * (at your option) any later version.
  */
 
+#include <drm/drm_atomic_helper.h>
 #include <drm/drm_gem_framebuffer_helper.h>
 #include <drm/tinydrm/mipi-dbi.h>
 #include <drm/tinydrm/tinydrm-helpers.h>

@@ -254,7 +257,7 @@ static int mipi_dbi_fb_dirty(struct drm_framebuffer *fb,
 static const struct drm_framebuffer_funcs mipi_dbi_fb_funcs = {
        .destroy        = drm_gem_fb_destroy,
        .create_handle  = drm_gem_fb_create_handle,
-       .dirty          = tinydrm_fb_dirty,
+       .dirty          = drm_atomic_helper_dirtyfb,
 };
 
 /**

I did some brief testing a few months back with PRIME buffers imported from vc4 and it looked like X11 sometimes did a page flip followed by a dirtyfb. This kills performance for me since I flush on both. Do you know any details on how/where X11 handles this?

Noralf.

Quoted 158 lines

  drivers/gpu/drm/drm_atomic_helper.c | 66
+++++++++++++++++++++++++++++++++++++
  drivers/gpu/drm/msm/msm_atomic.c    |  5 ++-
  drivers/gpu/drm/msm/msm_fb.c        |  1 +
  include/drm/drm_atomic_helper.h     |  4 +++
  include/drm/drm_plane.h             |  9 +++++
  5 files changed, 84 insertions(+), 1 deletion(-)

drivers/gpu/drm/drm_atomic_helper.c +64 -0

index c35654591c12..a578dc681b27 100644
@@ -3504,6 +3504,7 @@ void
 __drm_atomic_helper_plane_duplicate_state(struct drm_plane *plane,
        if (state->fb)
                drm_framebuffer_get(state->fb);
  +     state->dirty = false;
        state->fence = NULL;
        state->commit = NULL;
  }

@@ -3847,6 +3848,71 @@ int drm_atomic_helper_legacy_gamma_set(struct
 drm_crtc *crtc,
  }
  EXPORT_SYMBOL(drm_atomic_helper_legacy_gamma_set);
  +/**
+ * drm_atomic_helper_dirtyfb - helper for dirtyfb
+ *
+ * A helper to implement drm_framebuffer_funcs::dirty
+ */
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                             struct drm_file *file_priv, unsigned flags,
+                             unsigned color, struct drm_clip_rect *clips,
+                             unsigned num_clips)
+{
+       struct drm_modeset_acquire_ctx ctx;
+       struct drm_atomic_state *state;
+       struct drm_plane *plane;
+       int ret = 0;
+
+       /*
+        * When called from ioctl, we are interruptable, but not when
+        * called internally (ie. defio worker)
+        */
+       drm_modeset_acquire_init(&ctx,
+               file_priv ? DRM_MODESET_ACQUIRE_INTERRUPTIBLE : 0);
+
+       state = drm_atomic_state_alloc(fb->dev);
+       if (!state) {
+               ret = -ENOMEM;
+               goto out;
+       }
+       state->acquire_ctx = &ctx;
+
+retry:
+       drm_for_each_plane(plane, fb->dev) {
+               struct drm_plane_state *plane_state;
+
+               if (plane->state->fb != fb)
+                       continue;
+
+               plane_state = drm_atomic_get_plane_state(state, plane);
+               if (IS_ERR(plane_state)) {
+                       ret = PTR_ERR(plane_state);
+                       goto out;
+               }
+
+               plane_state->dirty = true;
+       }
+
+       ret = drm_atomic_nonblocking_commit(state);
+
+out:
+       if (ret == -EDEADLK) {
+               drm_atomic_state_clear(state);
+               ret = drm_modeset_backoff(&ctx);
+               if (!ret)
+                       goto retry;
+       }
+
+       drm_atomic_state_put(state);
+
+       drm_modeset_drop_locks(&ctx);
+       drm_modeset_acquire_fini(&ctx);
+
+       return ret;
+
+}
+EXPORT_SYMBOL(drm_atomic_helper_dirtyfb);
+
  /**
   * __drm_atomic_helper_private_duplicate_state - copy atomic private

state
* @obj: CRTC object

drivers/gpu/drm/msm/msm_atomic.c +4 -1

index bf5f8c39f34d..bb55a048e98b 100644
@@ -201,7 +201,10 @@ int msm_atomic_commit(struct drm_device *dev,
         * Figure out what fence to wait for:
         */
        for_each_oldnew_plane_in_state(state, plane, old_plane_state,
 new_plane_state, i) {
-               if ((new_plane_state->fb != old_plane_state->fb) &&
 new_plane_state->fb) {
+               bool sync_fb = new_plane_state->fb &&
+                       ((new_plane_state->fb != old_plane_state->fb) ||
+                        new_plane_state->dirty);
+               if (sync_fb) {
                        struct drm_gem_object *obj =

msm_framebuffer_bo(new_plane_state->fb, 0);
struct msm_gem_object *msm_obj = to_msm_bo(obj); struct dma_fence *fence =
reservation_object_get_excl_rcu(msm_obj->resv);

drivers/gpu/drm/msm/msm_fb.c +1 -0

index 0e0c87252ab0..a5d882a34a33 100644
@@ -62,6 +62,7 @@ static void msm_framebuffer_destroy(struct
 drm_framebuffer *fb)
  static const struct drm_framebuffer_funcs msm_framebuffer_funcs = {
        .create_handle = msm_framebuffer_create_handle,
        .destroy = msm_framebuffer_destroy,
+       .dirty = drm_atomic_helper_dirtyfb,
  };
    #ifdef CONFIG_DEBUG_FS

include/drm/drm_atomic_helper.h +4 -0

index 26aaba58d6ce..9b7a95c2643d 100644
@@ -183,6 +183,10 @@ int drm_atomic_helper_legacy_gamma_set(struct
 drm_crtc *crtc,
                                       u16 *red, u16 *green, u16 *blue,
                                       uint32_t size,
                                       struct drm_modeset_acquire_ctx
 *ctx);
+int drm_atomic_helper_dirtyfb(struct drm_framebuffer *fb,
+                             struct drm_file *file_priv, unsigned flags,
+                             unsigned color, struct drm_clip_rect *clips,
+                             unsigned num_clips);
  void __drm_atomic_helper_private_obj_duplicate_state(struct

drm_private_obj *obj,
struct
drm_private_state *state);
diff --git a/include/drm/drm_plane.h b/include/drm/drm_plane.h
index f7bf4a48b1c3..296fa22bda7a 100644

include/drm/drm_plane.h +3 -0

@@ -76,6 +76,15 @@ struct drm_plane_state {
         */
        struct drm_framebuffer *fb;
  +     /**
+        * @dirty:
+        *
+        * Flag that indicates the fb contents have changed even though

the

include/drm/drm_plane.h

Quoted 8 lines

include/drm/drm_plane.h +5 -0

+        * fb has not.  This is mostly a stop-gap solution until we have
+        * atomic dirty-rect(s) property.
+        */
+       bool dirty;
+
        /**
         * @fence:
         *

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 36 lines

Den 04.04.2018 00.42, skrev Rob Clark:

Quoted 29 lines

Add an atomic helper to implement dirtyfb support. This is needed to support DSI command-mode panels with x11 userspace (ie. when we can't rely on pageflips to trigger a flush to the panel).

To signal to the driver that the async atomic update needs to synchronize with fences, even though the fb didn't change, the drm_atomic_state::dirty flag is added.

Signed-off-by Rob Clark <[email protected]>

---
Background: there are a number of different folks working on getting upstream kernel working on various different phones/tablets with qcom SoC's.. many of them have command mode panels, so we kind of need a way to support the legacy dirtyfb ioctl for x11 support.

I know there is work on a proprer non-legacy atomic property for userspace to communicate dirty-rect(s) to the kernel, so this can be improved from triggering a full-frame flush once that is in place. But we kinda needa a stop-gap solution.

I had considered an in-driver solution for this, but things get a bit tricky if userspace ands up combining dirtyfb ioctls with page- flips, because we need to synchronize setting various CTL.FLUSH bits with setting the CTL.START bit. (ie. really all we need to do for cmd mode panels is bang CTL.START, but is this ends up racing with pageflips setting FLUSH bits, then bad things.) The easiest soln is to wrap this up as an atomic commit and rely on the worker to serialize things. Hence adding an atomic dirtyfb helper.

Quoted 4 lines

I guess at least the helper, with some small addition to translate and pass-thru the dirty rect(s) is useful to the final atomic dirty- rect property solution. Depending on how far off that is, a stop- gap solution could be useful.

Quoted 8 lines

I have been waiting for someone to address this since I'm not skilled enough myself to tackle the atomic machinery. It would be be nice to do all flushing during commit since that would mean that the tinydrm drivers could be driven solely through drm_simple_display_pipe_funcs. I wouldn't have to wire through the dirty callback like I do now.

I see that you use a nonblocking commit. What happens if a new dirtyfb comes in before the previous is done?

We could reuse the same trick we're doing in the fbdev code, of pushing the commit to a worker and doing it in a blocking fashion. Or at least wait for it to finish (can be done with the crtc_state->event stuff). That way userspace can pile us up in dirtyfb calls, but we'd do at most one per frame. More isn't really useful anyway.

Quoted 4 lines

If we could save the dirty clips, then I could use this in tinydrm. A full flush over SPI takes ~50ms so I need to shave off where I can.

This works for me in my tinydrm world:

One thing I thought you've had around somewhere is some additional tracking code, so we don't have to take all the plane mutexes in the ->dirtyfb callback. Does that exist, or was that just a figment of my imagination?