drm/tegra: Atomic conversion, phase 3, step 3
authorThierry Reding <treding@nvidia.com>
Mon, 24 Nov 2014 16:41:23 +0000 (17:41 +0100)
committerThierry Reding <treding@nvidia.com>
Tue, 27 Jan 2015 09:14:55 +0000 (10:14 +0100)
Provide a custom ->atomic_commit() implementation which supports async
commits. The generic atomic page-flip helper can use this to implement
page-flipping.

Signed-off-by: Thierry Reding <treding@nvidia.com>
drivers/gpu/drm/tegra/dc.c
drivers/gpu/drm/tegra/drm.c
drivers/gpu/drm/tegra/drm.h

index aa66abb..2f9af13 100644 (file)
@@ -820,99 +820,6 @@ static int tegra_dc_add_planes(struct drm_device *drm, struct tegra_dc *dc)
        return 0;
 }
 
-static int tegra_dc_set_base(struct tegra_dc *dc, int x, int y,
-                            struct drm_framebuffer *fb)
-{
-       struct tegra_bo *bo = tegra_fb_get_plane(fb, 0);
-       unsigned int h_offset = 0, v_offset = 0;
-       struct tegra_bo_tiling tiling;
-       unsigned long value, flags;
-       unsigned int format, swap;
-       int err;
-
-       err = tegra_fb_get_tiling(fb, &tiling);
-       if (err < 0)
-               return err;
-
-       spin_lock_irqsave(&dc->lock, flags);
-
-       tegra_dc_writel(dc, WINDOW_A_SELECT, DC_CMD_DISPLAY_WINDOW_HEADER);
-
-       value = fb->offsets[0] + y * fb->pitches[0] +
-               x * fb->bits_per_pixel / 8;
-
-       tegra_dc_writel(dc, bo->paddr + value, DC_WINBUF_START_ADDR);
-       tegra_dc_writel(dc, fb->pitches[0], DC_WIN_LINE_STRIDE);
-
-       format = tegra_dc_format(fb->pixel_format, &swap);
-       tegra_dc_writel(dc, format, DC_WIN_COLOR_DEPTH);
-       tegra_dc_writel(dc, swap, DC_WIN_BYTE_SWAP);
-
-       if (dc->soc->supports_block_linear) {
-               unsigned long height = tiling.value;
-
-               switch (tiling.mode) {
-               case TEGRA_BO_TILING_MODE_PITCH:
-                       value = DC_WINBUF_SURFACE_KIND_PITCH;
-                       break;
-
-               case TEGRA_BO_TILING_MODE_TILED:
-                       value = DC_WINBUF_SURFACE_KIND_TILED;
-                       break;
-
-               case TEGRA_BO_TILING_MODE_BLOCK:
-                       value = DC_WINBUF_SURFACE_KIND_BLOCK_HEIGHT(height) |
-                               DC_WINBUF_SURFACE_KIND_BLOCK;
-                       break;
-               }
-
-               tegra_dc_writel(dc, value, DC_WINBUF_SURFACE_KIND);
-       } else {
-               switch (tiling.mode) {
-               case TEGRA_BO_TILING_MODE_PITCH:
-                       value = DC_WIN_BUFFER_ADDR_MODE_LINEAR_UV |
-                               DC_WIN_BUFFER_ADDR_MODE_LINEAR;
-                       break;
-
-               case TEGRA_BO_TILING_MODE_TILED:
-                       value = DC_WIN_BUFFER_ADDR_MODE_TILE_UV |
-                               DC_WIN_BUFFER_ADDR_MODE_TILE;
-                       break;
-
-               case TEGRA_BO_TILING_MODE_BLOCK:
-                       DRM_ERROR("hardware doesn't support block linear mode\n");
-                       spin_unlock_irqrestore(&dc->lock, flags);
-                       return -EINVAL;
-               }
-
-               tegra_dc_writel(dc, value, DC_WIN_BUFFER_ADDR_MODE);
-       }
-
-       /* make sure bottom-up buffers are properly displayed */
-       if (tegra_fb_is_bottom_up(fb)) {
-               value = tegra_dc_readl(dc, DC_WIN_WIN_OPTIONS);
-               value |= V_DIRECTION;
-               tegra_dc_writel(dc, value, DC_WIN_WIN_OPTIONS);
-
-               v_offset += fb->height - 1;
-       } else {
-               value = tegra_dc_readl(dc, DC_WIN_WIN_OPTIONS);
-               value &= ~V_DIRECTION;
-               tegra_dc_writel(dc, value, DC_WIN_WIN_OPTIONS);
-       }
-
-       tegra_dc_writel(dc, h_offset, DC_WINBUF_ADDR_H_OFFSET);
-       tegra_dc_writel(dc, v_offset, DC_WINBUF_ADDR_V_OFFSET);
-
-       value = GENERAL_ACT_REQ | WIN_A_ACT_REQ;
-       tegra_dc_writel(dc, value << 8, DC_CMD_STATE_CONTROL);
-       tegra_dc_writel(dc, value, DC_CMD_STATE_CONTROL);
-
-       spin_unlock_irqrestore(&dc->lock, flags);
-
-       return 0;
-}
-
 void tegra_dc_enable_vblank(struct tegra_dc *dc)
 {
        unsigned long value, flags;
@@ -991,30 +898,6 @@ void tegra_dc_cancel_page_flip(struct drm_crtc *crtc, struct drm_file *file)
        spin_unlock_irqrestore(&drm->event_lock, flags);
 }
 
-static int tegra_dc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
-                             struct drm_pending_vblank_event *event, uint32_t page_flip_flags)
-{
-       unsigned int pipe = drm_crtc_index(crtc);
-       struct tegra_dc *dc = to_tegra_dc(crtc);
-
-       if (dc->event)
-               return -EBUSY;
-
-       if (event) {
-               event->pipe = pipe;
-               dc->event = event;
-               drm_crtc_vblank_get(crtc);
-       }
-
-       if (crtc->primary->state)
-               drm_atomic_set_fb_for_plane(crtc->primary->state, fb);
-
-       tegra_dc_set_base(dc, 0, 0, fb);
-       crtc->primary->fb = fb;
-
-       return 0;
-}
-
 static void tegra_dc_destroy(struct drm_crtc *crtc)
 {
        drm_crtc_cleanup(crtc);
@@ -1056,7 +939,7 @@ static void tegra_crtc_atomic_destroy_state(struct drm_crtc *crtc,
 }
 
 static const struct drm_crtc_funcs tegra_crtc_funcs = {
-       .page_flip = tegra_dc_page_flip,
+       .page_flip = drm_atomic_helper_page_flip,
        .set_config = drm_atomic_helper_set_config,
        .destroy = tegra_dc_destroy,
        .reset = tegra_crtc_reset,
@@ -1326,6 +1209,16 @@ static int tegra_crtc_atomic_check(struct drm_crtc *crtc,
 
 static void tegra_crtc_atomic_begin(struct drm_crtc *crtc)
 {
+       struct tegra_dc *dc = to_tegra_dc(crtc);
+
+       if (crtc->state->event) {
+               crtc->state->event->pipe = drm_crtc_index(crtc);
+
+               WARN_ON(drm_crtc_vblank_get(crtc) != 0);
+
+               dc->event = crtc->state->event;
+               crtc->state->event = NULL;
+       }
 }
 
 static void tegra_crtc_atomic_flush(struct drm_crtc *crtc)
index d01484c..0edfb5e 100644 (file)
@@ -10,6 +10,7 @@
 #include <linux/host1x.h>
 #include <linux/iommu.h>
 
+#include <drm/drm_atomic.h>
 #include <drm/drm_atomic_helper.h>
 
 #include "drm.h"
@@ -26,13 +27,90 @@ struct tegra_drm_file {
        struct list_head contexts;
 };
 
+static void tegra_atomic_schedule(struct tegra_drm *tegra,
+                                 struct drm_atomic_state *state)
+{
+       tegra->commit.state = state;
+       schedule_work(&tegra->commit.work);
+}
+
+static void tegra_atomic_complete(struct tegra_drm *tegra,
+                                 struct drm_atomic_state *state)
+{
+       struct drm_device *drm = tegra->drm;
+
+       /*
+        * Everything below can be run asynchronously without the need to grab
+        * any modeset locks at all under one condition: 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.
+        */
+
+       drm_atomic_helper_commit_pre_planes(drm, state);
+       drm_atomic_helper_commit_planes(drm, state);
+       drm_atomic_helper_commit_post_planes(drm, state);
+
+       drm_atomic_helper_wait_for_vblanks(drm, state);
+
+       drm_atomic_helper_cleanup_planes(drm, state);
+       drm_atomic_state_free(state);
+}
+
+static void tegra_atomic_work(struct work_struct *work)
+{
+       struct tegra_drm *tegra = container_of(work, struct tegra_drm,
+                                              commit.work);
+
+       tegra_atomic_complete(tegra, tegra->commit.state);
+}
+
+static int tegra_atomic_commit(struct drm_device *drm,
+                              struct drm_atomic_state *state, bool async)
+{
+       struct tegra_drm *tegra = drm->dev_private;
+       int err;
+
+       err = drm_atomic_helper_prepare_planes(drm, state);
+       if (err)
+               return err;
+
+       /* serialize outstanding asynchronous commits */
+       mutex_lock(&tegra->commit.lock);
+       flush_work(&tegra->commit.work);
+
+       /*
+        * 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(drm, state);
+
+       if (async)
+               tegra_atomic_schedule(tegra, state);
+       else
+               tegra_atomic_complete(tegra, state);
+
+       mutex_unlock(&tegra->commit.lock);
+       return 0;
+}
+
 static const struct drm_mode_config_funcs tegra_drm_mode_funcs = {
        .fb_create = tegra_fb_create,
 #ifdef CONFIG_DRM_TEGRA_FBDEV
        .output_poll_changed = tegra_fb_output_poll_changed,
 #endif
        .atomic_check = drm_atomic_helper_check,
-       .atomic_commit = drm_atomic_helper_commit,
+       .atomic_commit = tegra_atomic_commit,
 };
 
 static int tegra_drm_load(struct drm_device *drm, unsigned long flags)
@@ -58,6 +136,10 @@ static int tegra_drm_load(struct drm_device *drm, unsigned long flags)
 
        mutex_init(&tegra->clients_lock);
        INIT_LIST_HEAD(&tegra->clients);
+
+       mutex_init(&tegra->commit.lock);
+       INIT_WORK(&tegra->commit.work, tegra_atomic_work);
+
        drm->dev_private = tegra;
        tegra->drm = drm;
 
index 3db719d..b1c7027 100644 (file)
@@ -50,6 +50,12 @@ struct tegra_drm {
 #endif
 
        unsigned int pitch_align;
+
+       struct {
+               struct drm_atomic_state *state;
+               struct work_struct work;
+               struct mutex lock;
+       } commit;
 };
 
 struct tegra_drm_client;