drm/msm: atomic core bits
authorRob Clark <robdclark@gmail.com>
Sat, 8 Nov 2014 18:21:06 +0000 (13:21 -0500)
committerRob Clark <robdclark@gmail.com>
Sun, 16 Nov 2014 19:27:37 +0000 (14:27 -0500)
The core parts for async commit.

Signed-off-by: Rob Clark <robdclark@gmail.com>
drivers/gpu/drm/msm/Makefile
drivers/gpu/drm/msm/msm_atomic.c [new file with mode: 0644]
drivers/gpu/drm/msm/msm_drv.c
drivers/gpu/drm/msm/msm_drv.h

index 6283dcb96af51f8bd5e70c26a507412f216996f6..415a145a9829af7c1656fadfa4b954f338fdb031 100644 (file)
@@ -30,6 +30,7 @@ msm-y := \
        mdp/mdp5/mdp5_kms.o \
        mdp/mdp5/mdp5_plane.o \
        mdp/mdp5/mdp5_smp.o \
+       msm_atomic.o \
        msm_drv.o \
        msm_fb.o \
        msm_gem.o \
diff --git a/drivers/gpu/drm/msm/msm_atomic.c b/drivers/gpu/drm/msm/msm_atomic.c
new file mode 100644 (file)
index 0000000..8cf3361
--- /dev/null
@@ -0,0 +1,163 @@
+/*
+ * Copyright (C) 2014 Red Hat
+ * Author: Rob Clark <robdclark@gmail.com>
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published by
+ * the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "msm_drv.h"
+#include "msm_kms.h"
+#include "msm_gem.h"
+
+struct msm_commit {
+       struct drm_atomic_state *state;
+       uint32_t fence;
+       struct msm_fence_cb fence_cb;
+};
+
+static void fence_cb(struct msm_fence_cb *cb);
+
+static struct msm_commit *new_commit(struct drm_atomic_state *state)
+{
+       struct msm_commit *c = kzalloc(sizeof(*c), GFP_KERNEL);
+
+       if (!c)
+               return NULL;
+
+       c->state = state;
+       /* TODO we might need a way to indicate to run the cb on a
+        * different wq so wait_for_vblanks() doesn't block retiring
+        * bo's..
+        */
+       INIT_FENCE_CB(&c->fence_cb, fence_cb);
+
+       return c;
+}
+
+/* The (potentially) asynchronous part of the commit.  At this point
+ * nothing can fail short of armageddon.
+ */
+static void complete_commit(struct msm_commit *c)
+{
+       struct drm_atomic_state *state = c->state;
+       struct drm_device *dev = state->dev;
+
+       drm_atomic_helper_commit_pre_planes(dev, state);
+
+       drm_atomic_helper_commit_planes(dev, state);
+
+       drm_atomic_helper_commit_post_planes(dev, state);
+
+       drm_atomic_helper_wait_for_vblanks(dev, state);
+
+       drm_atomic_helper_cleanup_planes(dev, state);
+
+       drm_atomic_state_free(state);
+
+       kfree(c);
+}
+
+static void fence_cb(struct msm_fence_cb *cb)
+{
+       struct msm_commit *c =
+                       container_of(cb, struct msm_commit, fence_cb);
+       complete_commit(c);
+}
+
+static void add_fb(struct msm_commit *c, struct drm_framebuffer *fb)
+{
+       struct drm_gem_object *obj = msm_framebuffer_bo(fb, 0);
+       c->fence = max(c->fence, msm_gem_fence(to_msm_bo(obj), MSM_PREP_READ));
+}
+
+
+/**
+ * drm_atomic_helper_commit - commit validated state object
+ * @dev: DRM device
+ * @state: the driver state object
+ * @async: asynchronous commit
+ *
+ * This function commits a with drm_atomic_helper_check() pre-validated state
+ * object. This can still fail when e.g. the framebuffer reservation fails. For
+ * now this doesn't implement asynchronous commits.
+ *
+ * RETURNS
+ * Zero for success or -errno.
+ */
+int msm_atomic_commit(struct drm_device *dev,
+               struct drm_atomic_state *state, bool async)
+{
+       struct msm_commit *c;
+       int nplanes = dev->mode_config.num_total_plane;
+       int i, ret;
+
+       ret = drm_atomic_helper_prepare_planes(dev, state);
+       if (ret)
+               return ret;
+
+       c = new_commit(state);
+
+       /*
+        * Figure out what fence to wait for:
+        */
+       for (i = 0; i < nplanes; i++) {
+               struct drm_plane *plane = state->planes[i];
+               struct drm_plane_state *new_state = state->plane_states[i];
+
+               if (!plane)
+                       continue;
+
+               if (plane->state->fb != new_state->fb)
+                       add_fb(c, new_state->fb);
+       }
+
+       /*
+        * This is the point of no return - everything below never fails except
+        * when the hw goes bonghits. Which means we can commit the new state on
+        * the software side now.
+        */
+
+       drm_atomic_helper_swap_state(dev, state);
+
+       /*
+        * Everything below can be run asynchronously without the need to grab
+        * any modeset locks at all under one conditions: It must be guaranteed
+        * that the asynchronous work has either been cancelled (if the driver
+        * supports it, which at least requires that the framebuffers get
+        * cleaned up with drm_atomic_helper_cleanup_planes()) or completed
+        * before the new state gets committed on the software side with
+        * drm_atomic_helper_swap_state().
+        *
+        * This scheme allows new atomic state updates to be prepared and
+        * checked in parallel to the asynchronous completion of the previous
+        * update. Which is important since compositors need to figure out the
+        * composition of the next frame right after having submitted the
+        * current layout.
+        */
+
+       if (async) {
+               msm_queue_fence_cb(dev, &c->fence_cb, c->fence);
+               return 0;
+       }
+
+       ret = msm_wait_fence_interruptable(dev, c->fence, NULL);
+       if (ret) {
+               WARN_ON(ret);  // TODO unswap state back?  or??
+               kfree(c);
+               return ret;
+       }
+
+       complete_commit(c);
+
+       return 0;
+}
index 1456b1c3e890f42a9f54ebb0be08f1d6b7651f05..d3b791b7ddef014e35b76ec43db3be1f8f576cf2 100644 (file)
@@ -29,6 +29,8 @@ static void msm_fb_output_poll_changed(struct drm_device *dev)
 static const struct drm_mode_config_funcs mode_config_funcs = {
        .fb_create = msm_framebuffer_create,
        .output_poll_changed = msm_fb_output_poll_changed,
+       .atomic_check = drm_atomic_helper_check,
+       .atomic_commit = msm_atomic_commit,
 };
 
 int msm_register_mmu(struct drm_device *dev, struct msm_mmu *mmu)
@@ -294,6 +296,8 @@ static int msm_load(struct drm_device *dev, unsigned long flags)
                goto fail;
        }
 
+       drm_mode_config_reset(dev);
+
 #ifdef CONFIG_DRM_MSM_FBDEV
        priv->fbdev = msm_fbdev_init(dev);
 #endif
index 4a7e8b8ff6492567e1679fd1059c578b2a5f0145..4b52d752bb6e7b58d3a85eb9dbbb6f2d25412533 100644 (file)
 #endif
 
 #include <drm/drmP.h>
+#include <drm/drm_atomic.h>
+#include <drm/drm_atomic_helper.h>
 #include <drm/drm_crtc_helper.h>
+#include <drm/drm_plane_helper.h>
 #include <drm/drm_fb_helper.h>
 #include <drm/msm_drm.h>
 #include <drm/drm_gem.h>
@@ -141,6 +144,9 @@ void __msm_fence_worker(struct work_struct *work);
                (_cb)->func = _func;                         \
        } while (0)
 
+int msm_atomic_commit(struct drm_device *dev,
+               struct drm_atomic_state *state, bool async);
+
 int msm_register_mmu(struct drm_device *dev, struct msm_mmu *mmu);
 
 int msm_wait_fence_interruptable(struct drm_device *dev, uint32_t fence,