Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1281195 > unrolled thread

[PATCH 1/9] drm: Create a driver hook for allocating GEM object structs.

Back to article view | Back to linux.kernel

Contents

  [PATCH 1/9] drm: Create a driver hook for allocating GEM object structs. Eric Anholt <eric@anholt.net> - 2015-12-01 21:40 +0100
    [PATCH 4/9] drm/vc4: Add an API for creating GPU shaders in GEM BOs. Eric Anholt <eric@anholt.net> - 2015-12-01 21:40 +0100
      Re: [PATCH 4/9] drm/vc4: Add an API for creating GPU shaders in GEM BOs. Emil Velikov <emil.l.velikov@gmail.com> - 2015-12-02 12:40 +0100
        Re: [PATCH 4/9] drm/vc4: Add an API for creating GPU shaders in GEM BOs. Eric Anholt <eric@anholt.net> - 2015-12-02 20:30 +0100
    [PATCH 6/9] drm/vc4: Bind and initialize the V3D engine. Eric Anholt <eric@anholt.net> - 2015-12-01 21:40 +0100
    [PATCH 8/9] drm/vc4: Add support for async pageflips. Eric Anholt <eric@anholt.net> - 2015-12-01 21:40 +0100
      Re: [PATCH 8/9] drm/vc4: Add support for async pageflips. Daniel Stone <daniel@fooishbar.org> - 2015-12-02 15:10 +0100
        Re: [PATCH 8/9] drm/vc4: Add support for async pageflips. Eric Anholt <eric@anholt.net> - 2015-12-04 02:50 +0100
          Re: [PATCH 8/9] drm/vc4: Add support for async pageflips. Daniel Stone <daniel@fooishbar.org> - 2015-12-07 11:20 +0100
    [PATCH 3/9] drm/vc4: Add create and map BO ioctls. Eric Anholt <eric@anholt.net> - 2015-12-01 21:40 +0100
      Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls. Emil Velikov <emil.l.velikov@gmail.com> - 2015-12-04 15:30 +0100
        Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls. Eric Anholt <eric@anholt.net> - 2015-12-08 02:20 +0100
          Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls. Michel Dänzer <michel@daenzer.net> - 2015-12-08 04:00 +0100
            Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls. Eric Anholt <eric@anholt.net> - 2015-12-08 05:10 +0100
    [PATCH 2/9] drm/vc4: Add a BO cache. Eric Anholt <eric@anholt.net> - 2015-12-01 21:40 +0100
    Re: [PATCH 1/9] drm: Create a driver hook for allocating GEM object  structs. Daniel Vetter <daniel@ffwll.ch> - 2015-12-02 10:00 +0100
    Re: [PATCH 7/9] drm/vc4: Add support for drawing 3D frames. Emil Velikov <emil.l.velikov@gmail.com> - 2015-12-02 12:30 +0100
      Re: [PATCH 7/9] drm/vc4: Add support for drawing 3D frames. Eric Anholt <eric@anholt.net> - 2015-12-02 20:30 +0100

#1281195 — [PATCH 1/9] drm: Create a driver hook for allocating GEM object structs.

The CMA helpers had no way for a driver to extend the struct with its
own fields.  Since the CMA helpers are mostly "Allocate a
drm_gem_cma_object, then fill in a few fields", it's hard to write as
pure helpers without passing in a driver callback for the allocate
step.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/drm_gem_cma_helper.c | 10 ++++++----
 include/drm/drmP.h                   |  7 +++++++
 2 files changed, 13 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/drm_gem_cma_helper.c b/drivers/gpu/drm/drm_gem_cma_helper.c
index e109b49..0f7b00ba 100644
--- a/drivers/gpu/drm/drm_gem_cma_helper.c
+++ b/drivers/gpu/drm/drm_gem_cma_helper.c
@@ -59,11 +59,13 @@ __drm_gem_cma_create(struct drm_device *drm, size_t size)
 	struct drm_gem_object *gem_obj;
 	int ret;
 
-	cma_obj = kzalloc(sizeof(*cma_obj), GFP_KERNEL);
-	if (!cma_obj)
+	if (drm->driver->gem_create_object)
+		gem_obj = drm->driver->gem_create_object(drm, size);
+	else
+		gem_obj = kzalloc(sizeof(*cma_obj), GFP_KERNEL);
+	if (!gem_obj)
 		return ERR_PTR(-ENOMEM);
-
-	gem_obj = &cma_obj->base;
+	cma_obj = container_of(gem_obj, struct drm_gem_cma_object, base);
 
 	ret = drm_gem_object_init(drm, gem_obj, size);
 	if (ret)
diff --git a/include/drm/drmP.h b/include/drm/drmP.h
index 0b921ae..22ff162 100644
--- a/include/drm/drmP.h
+++ b/include/drm/drmP.h
@@ -580,6 +580,13 @@ struct drm_driver {
 	int (*gem_open_object) (struct drm_gem_object *, struct drm_file *);
 	void (*gem_close_object) (struct drm_gem_object *, struct drm_file *);
 
+	/**
+	 * Hook for allocating the GEM object struct, for use by core
+	 * helpers.
+	 */
+	struct drm_gem_object *(*gem_create_object)(struct drm_device *dev,
+						    size_t size);
+
 	/* prime: */
 	/* export handle -> fd (see drm_gem_prime_handle_to_fd() helper) */
 	int (*prime_handle_to_fd)(struct drm_device *dev, struct drm_file *file_priv,
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [next] | [standalone]

#1281196 — [PATCH 4/9] drm/vc4: Add an API for creating GPU shaders in GEM BOs.

Since we have no MMU, the kernel needs to validate that the submitted
shader code won't make any accesses to memory that the user doesn't
control, which involves banning some operations (general purpose DMA
writes), and tracking where we need to write out pointers for other
operations (texture sampling).  Once it's validated, we return a GEM
BO containing the shader, which doesn't allow mapping for write or
exporting to other subsystems.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/vc4/Makefile               |   3 +-
 drivers/gpu/drm/vc4/vc4_bo.c               | 140 ++++++++
 drivers/gpu/drm/vc4/vc4_drv.c              |   9 +-
 drivers/gpu/drm/vc4/vc4_drv.h              |  50 +++
 drivers/gpu/drm/vc4/vc4_qpu_defines.h      | 264 +++++++++++++++
 drivers/gpu/drm/vc4/vc4_validate_shaders.c | 513 +++++++++++++++++++++++++++++
 include/uapi/drm/vc4_drm.h                 |  25 ++
 7 files changed, 999 insertions(+), 5 deletions(-)
 create mode 100644 drivers/gpu/drm/vc4/vc4_qpu_defines.h
 create mode 100644 drivers/gpu/drm/vc4/vc4_validate_shaders.c

diff --git a/drivers/gpu/drm/vc4/Makefile b/drivers/gpu/drm/vc4/Makefile
index 32b4f9c..eb776a6 100644
--- a/drivers/gpu/drm/vc4/Makefile
+++ b/drivers/gpu/drm/vc4/Makefile
@@ -10,7 +10,8 @@ vc4-y := \
 	vc4_kms.o \
 	vc4_hdmi.o \
 	vc4_hvs.o \
-	vc4_plane.o
+	vc4_plane.o \
+	vc4_validate_shaders.o
 
 vc4-$(CONFIG_DEBUG_FS) += vc4_debugfs.o
 
diff --git a/drivers/gpu/drm/vc4/vc4_bo.c b/drivers/gpu/drm/vc4/vc4_bo.c
index 06cba26..18dfe3e 100644
--- a/drivers/gpu/drm/vc4/vc4_bo.c
+++ b/drivers/gpu/drm/vc4/vc4_bo.c
@@ -79,6 +79,12 @@ static void vc4_bo_destroy(struct vc4_bo *bo)
 	struct drm_gem_object *obj = &bo->base.base;
 	struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
 
+	if (bo->validated_shader) {
+		kfree(bo->validated_shader->texture_samples);
+		kfree(bo->validated_shader);
+		bo->validated_shader = NULL;
+	}
+
 	vc4->bo_stats.num_allocated--;
 	vc4->bo_stats.size_allocated -= obj->size;
 	drm_gem_cma_free_object(obj);
@@ -315,6 +321,12 @@ void vc4_free_object(struct drm_gem_object *gem_bo)
 		goto out;
 	}
 
+	if (bo->validated_shader) {
+		kfree(bo->validated_shader->texture_samples);
+		kfree(bo->validated_shader);
+		bo->validated_shader = NULL;
+	}
+
 	bo->free_time = jiffies;
 	list_add(&bo->size_head, cache_list);
 	list_add(&bo->unref_head, &vc4->bo_cache.time_list);
@@ -347,6 +359,78 @@ static void vc4_bo_cache_time_timer(unsigned long data)
 	schedule_work(&vc4->bo_cache.time_work);
 }
 
+struct dma_buf *
+vc4_prime_export(struct drm_device *dev, struct drm_gem_object *obj, int flags)
+{
+	struct vc4_bo *bo = to_vc4_bo(obj);
+
+	if (bo->validated_shader) {
+		DRM_ERROR("Attempting to export shader BO\n");
+		return ERR_PTR(-EINVAL);
+	}
+
+	return drm_gem_prime_export(dev, obj, flags);
+}
+
+int vc4_mmap(struct file *filp, struct vm_area_struct *vma)
+{
+	struct drm_gem_object *gem_obj;
+	struct vc4_bo *bo;
+	int ret;
+
+	ret = drm_gem_mmap(filp, vma);
+	if (ret)
+		return ret;
+
+	gem_obj = vma->vm_private_data;
+	bo = to_vc4_bo(gem_obj);
+
+	if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
+		DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
+		return -EINVAL;
+	}
+
+	/*
+	 * Clear the VM_PFNMAP flag that was set by drm_gem_mmap(), and set the
+	 * vm_pgoff (used as a fake buffer offset by DRM) to 0 as we want to map
+	 * the whole buffer.
+	 */
+	vma->vm_flags &= ~VM_PFNMAP;
+	vma->vm_pgoff = 0;
+
+	ret = dma_mmap_writecombine(bo->base.base.dev->dev, vma,
+				    bo->base.vaddr, bo->base.paddr,
+				    vma->vm_end - vma->vm_start);
+	if (ret)
+		drm_gem_vm_close(vma);
+
+	return ret;
+}
+
+int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
+{
+	struct vc4_bo *bo = to_vc4_bo(obj);
+
+	if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
+		DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
+		return -EINVAL;
+	}
+
+	return drm_gem_cma_prime_mmap(obj, vma);
+}
+
+void *vc4_prime_vmap(struct drm_gem_object *obj)
+{
+	struct vc4_bo *bo = to_vc4_bo(obj);
+
+	if (bo->validated_shader) {
+		DRM_ERROR("mmaping of shader BOs not allowed.\n");
+		return ERR_PTR(-EINVAL);
+	}
+
+	return drm_gem_cma_prime_vmap(obj);
+}
+
 int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *file_priv)
 {
@@ -387,6 +471,62 @@ int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
 	return 0;
 }
 
+int
+vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
+			   struct drm_file *file_priv)
+{
+	struct drm_vc4_create_shader_bo *args = data;
+	struct vc4_bo *bo = NULL;
+	int ret;
+
+	if (args->size == 0)
+		return -EINVAL;
+
+	if (args->size % sizeof(u64) != 0)
+		return -EINVAL;
+
+	if (args->flags != 0) {
+		DRM_INFO("Unknown flags set: 0x%08x\n", args->flags);
+		return -EINVAL;
+	}
+
+	if (args->pad != 0) {
+		DRM_INFO("Pad set: 0x%08x\n", args->pad);
+		return -EINVAL;
+	}
+
+	bo = vc4_bo_create(dev, args->size, true);
+	if (!bo)
+		return -ENOMEM;
+
+	ret = copy_from_user(bo->base.vaddr,
+			     (void __user *)(uintptr_t)args->data,
+			     args->size);
+	if (ret != 0)
+		goto fail;
+	/* Clear the rest of the memory from allocating from the BO
+	 * cache.
+	 */
+	memset(bo->base.vaddr + args->size, 0,
+	       bo->base.base.size - args->size);
+
+	bo->validated_shader = vc4_validate_shader(&bo->base);
+	if (!bo->validated_shader) {
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	/* We have to create the handle after validation, to avoid
+	 * races for users to do doing things like mmap the shader BO.
+	 */
+	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
+
+ fail:
+	drm_gem_object_unreference_unlocked(&bo->base.base);
+
+	return ret;
+}
+
 void vc4_bo_cache_init(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index 5fa4688..da4be9c 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -64,7 +64,7 @@ static const struct file_operations vc4_drm_fops = {
 	.open = drm_open,
 	.release = drm_release,
 	.unlocked_ioctl = drm_ioctl,
-	.mmap = drm_gem_cma_mmap,
+	.mmap = vc4_mmap,
 	.poll = drm_poll,
 	.read = drm_read,
 #ifdef CONFIG_COMPAT
@@ -76,6 +76,7 @@ static const struct file_operations vc4_drm_fops = {
 static const struct drm_ioctl_desc vc4_drm_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(VC4_CREATE_BO, vc4_create_bo_ioctl, 0),
 	DRM_IOCTL_DEF_DRV(VC4_MMAP_BO, vc4_mmap_bo_ioctl, 0),
+	DRM_IOCTL_DEF_DRV(VC4_CREATE_SHADER_BO, vc4_create_shader_bo_ioctl, 0),
 };
 
 static struct drm_driver vc4_drm_driver = {
@@ -102,12 +103,12 @@ static struct drm_driver vc4_drm_driver = {
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
 	.gem_prime_import = drm_gem_prime_import,
-	.gem_prime_export = drm_gem_prime_export,
+	.gem_prime_export = vc4_prime_export,
 	.gem_prime_get_sg_table	= drm_gem_cma_prime_get_sg_table,
 	.gem_prime_import_sg_table = drm_gem_cma_prime_import_sg_table,
-	.gem_prime_vmap = drm_gem_cma_prime_vmap,
+	.gem_prime_vmap = vc4_prime_vmap,
 	.gem_prime_vunmap = drm_gem_cma_prime_vunmap,
-	.gem_prime_mmap = drm_gem_cma_prime_mmap,
+	.gem_prime_mmap = vc4_prime_mmap,
 
 	.dumb_create = vc4_dumb_create,
 	.dumb_map_offset = drm_gem_cma_dumb_map_offset,
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index fddb0a0..bd77d55 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -69,6 +69,11 @@ struct vc4_bo {
 
 	/* List entry for the BO's position in vc4_dev->bo_cache.size_list */
 	struct list_head size_head;
+
+	/* Struct for shader validation state, if created by
+	 * DRM_IOCTL_VC4_CREATE_SHADER_BO.
+	 */
+	struct vc4_validated_shader_info *validated_shader;
 };
 
 static inline struct vc4_bo *
@@ -118,6 +123,42 @@ to_vc4_encoder(struct drm_encoder *encoder)
 #define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
 
 /**
+ * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
+ * setup parameters.
+ *
+ * This will be used at draw time to relocate the reference to the texture
+ * contents in p0, and validate that the offset combined with
+ * width/height/stride/etc. from p1 and p2/p3 doesn't sample outside the BO.
+ * Note that the hardware treats unprovided config parameters as 0, so not all
+ * of them need to be set up for every texure sample, and we'll store ~0 as
+ * the offset to mark the unused ones.
+ *
+ * See the VC4 3D architecture guide page 41 ("Texture and Memory Lookup Unit
+ * Setup") for definitions of the texture parameters.
+ */
+struct vc4_texture_sample_info {
+	bool is_direct;
+	uint32_t p_offset[4];
+};
+
+/**
+ * struct vc4_validated_shader_info - information about validated shaders that
+ * needs to be used from command list validation.
+ *
+ * For a given shader, each time a shader state record references it, we need
+ * to verify that the shader doesn't read more uniforms than the shader state
+ * record's uniform BO pointer can provide, and we need to apply relocations
+ * and validate the shader state record's uniforms that define the texture
+ * samples.
+ */
+struct vc4_validated_shader_info {
+	uint32_t uniforms_size;
+	uint32_t uniforms_src_size;
+	uint32_t num_texture_samples;
+	struct vc4_texture_sample_info *texture_samples;
+};
+
+/**
  * _wait_for - magic (register) wait macro
  *
  * Does the right thing for modeset paths when run under kdgb or similar atomic
@@ -157,8 +198,13 @@ struct dma_buf *vc4_prime_export(struct drm_device *dev,
 				 struct drm_gem_object *obj, int flags);
 int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *file_priv);
+int vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
+			       struct drm_file *file_priv);
 int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
 		      struct drm_file *file_priv);
+int vc4_mmap(struct file *filp, struct vm_area_struct *vma);
+int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
+void *vc4_prime_vmap(struct drm_gem_object *obj);
 void vc4_bo_cache_init(struct drm_device *dev);
 void vc4_bo_cache_destroy(struct drm_device *dev);
 int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
@@ -194,3 +240,7 @@ struct drm_plane *vc4_plane_init(struct drm_device *dev,
 				 enum drm_plane_type type);
 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
 u32 vc4_plane_dlist_size(struct drm_plane_state *state);
+
+/* vc4_validate_shader.c */
+struct vc4_validated_shader_info *
+vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
diff --git a/drivers/gpu/drm/vc4/vc4_qpu_defines.h b/drivers/gpu/drm/vc4/vc4_qpu_defines.h
new file mode 100644
index 0000000..d5c2f3c
--- /dev/null
+++ b/drivers/gpu/drm/vc4/vc4_qpu_defines.h
@@ -0,0 +1,264 @@
+/*
+ * Copyright © 2014 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#ifndef VC4_QPU_DEFINES_H
+#define VC4_QPU_DEFINES_H
+
+enum qpu_op_add {
+	QPU_A_NOP,
+	QPU_A_FADD,
+	QPU_A_FSUB,
+	QPU_A_FMIN,
+	QPU_A_FMAX,
+	QPU_A_FMINABS,
+	QPU_A_FMAXABS,
+	QPU_A_FTOI,
+	QPU_A_ITOF,
+	QPU_A_ADD = 12,
+	QPU_A_SUB,
+	QPU_A_SHR,
+	QPU_A_ASR,
+	QPU_A_ROR,
+	QPU_A_SHL,
+	QPU_A_MIN,
+	QPU_A_MAX,
+	QPU_A_AND,
+	QPU_A_OR,
+	QPU_A_XOR,
+	QPU_A_NOT,
+	QPU_A_CLZ,
+	QPU_A_V8ADDS = 30,
+	QPU_A_V8SUBS = 31,
+};
+
+enum qpu_op_mul {
+	QPU_M_NOP,
+	QPU_M_FMUL,
+	QPU_M_MUL24,
+	QPU_M_V8MULD,
+	QPU_M_V8MIN,
+	QPU_M_V8MAX,
+	QPU_M_V8ADDS,
+	QPU_M_V8SUBS,
+};
+
+enum qpu_raddr {
+	QPU_R_FRAG_PAYLOAD_ZW = 15, /* W for A file, Z for B file */
+	/* 0-31 are the plain regfile a or b fields */
+	QPU_R_UNIF = 32,
+	QPU_R_VARY = 35,
+	QPU_R_ELEM_QPU = 38,
+	QPU_R_NOP,
+	QPU_R_XY_PIXEL_COORD = 41,
+	QPU_R_MS_REV_FLAGS = 41,
+	QPU_R_VPM = 48,
+	QPU_R_VPM_LD_BUSY,
+	QPU_R_VPM_LD_WAIT,
+	QPU_R_MUTEX_ACQUIRE,
+};
+
+enum qpu_waddr {
+	/* 0-31 are the plain regfile a or b fields */
+	QPU_W_ACC0 = 32, /* aka r0 */
+	QPU_W_ACC1,
+	QPU_W_ACC2,
+	QPU_W_ACC3,
+	QPU_W_TMU_NOSWAP,
+	QPU_W_ACC5,
+	QPU_W_HOST_INT,
+	QPU_W_NOP,
+	QPU_W_UNIFORMS_ADDRESS,
+	QPU_W_QUAD_XY, /* X for regfile a, Y for regfile b */
+	QPU_W_MS_FLAGS = 42,
+	QPU_W_REV_FLAG = 42,
+	QPU_W_TLB_STENCIL_SETUP = 43,
+	QPU_W_TLB_Z,
+	QPU_W_TLB_COLOR_MS,
+	QPU_W_TLB_COLOR_ALL,
+	QPU_W_TLB_ALPHA_MASK,
+	QPU_W_VPM,
+	QPU_W_VPMVCD_SETUP, /* LD for regfile a, ST for regfile b */
+	QPU_W_VPM_ADDR, /* LD for regfile a, ST for regfile b */
+	QPU_W_MUTEX_RELEASE,
+	QPU_W_SFU_RECIP,
+	QPU_W_SFU_RECIPSQRT,
+	QPU_W_SFU_EXP,
+	QPU_W_SFU_LOG,
+	QPU_W_TMU0_S,
+	QPU_W_TMU0_T,
+	QPU_W_TMU0_R,
+	QPU_W_TMU0_B,
+	QPU_W_TMU1_S,
+	QPU_W_TMU1_T,
+	QPU_W_TMU1_R,
+	QPU_W_TMU1_B,
+};
+
+enum qpu_sig_bits {
+	QPU_SIG_SW_BREAKPOINT,
+	QPU_SIG_NONE,
+	QPU_SIG_THREAD_SWITCH,
+	QPU_SIG_PROG_END,
+	QPU_SIG_WAIT_FOR_SCOREBOARD,
+	QPU_SIG_SCOREBOARD_UNLOCK,
+	QPU_SIG_LAST_THREAD_SWITCH,
+	QPU_SIG_COVERAGE_LOAD,
+	QPU_SIG_COLOR_LOAD,
+	QPU_SIG_COLOR_LOAD_END,
+	QPU_SIG_LOAD_TMU0,
+	QPU_SIG_LOAD_TMU1,
+	QPU_SIG_ALPHA_MASK_LOAD,
+	QPU_SIG_SMALL_IMM,
+	QPU_SIG_LOAD_IMM,
+	QPU_SIG_BRANCH
+};
+
+enum qpu_mux {
+	/* hardware mux values */
+	QPU_MUX_R0,
+	QPU_MUX_R1,
+	QPU_MUX_R2,
+	QPU_MUX_R3,
+	QPU_MUX_R4,
+	QPU_MUX_R5,
+	QPU_MUX_A,
+	QPU_MUX_B,
+
+	/* non-hardware mux values */
+	QPU_MUX_IMM,
+};
+
+enum qpu_cond {
+	QPU_COND_NEVER,
+	QPU_COND_ALWAYS,
+	QPU_COND_ZS,
+	QPU_COND_ZC,
+	QPU_COND_NS,
+	QPU_COND_NC,
+	QPU_COND_CS,
+	QPU_COND_CC,
+};
+
+enum qpu_pack_mul {
+	QPU_PACK_MUL_NOP,
+	/* replicated to each 8 bits of the 32-bit dst. */
+	QPU_PACK_MUL_8888 = 3,
+	QPU_PACK_MUL_8A,
+	QPU_PACK_MUL_8B,
+	QPU_PACK_MUL_8C,
+	QPU_PACK_MUL_8D,
+};
+
+enum qpu_pack_a {
+	QPU_PACK_A_NOP,
+	/* convert to 16 bit float if float input, or to int16. */
+	QPU_PACK_A_16A,
+	QPU_PACK_A_16B,
+	/* replicated to each 8 bits of the 32-bit dst. */
+	QPU_PACK_A_8888,
+	/* Convert to 8-bit unsigned int. */
+	QPU_PACK_A_8A,
+	QPU_PACK_A_8B,
+	QPU_PACK_A_8C,
+	QPU_PACK_A_8D,
+
+	/* Saturating variants of the previous instructions. */
+	QPU_PACK_A_32_SAT, /* int-only */
+	QPU_PACK_A_16A_SAT, /* int or float */
+	QPU_PACK_A_16B_SAT,
+	QPU_PACK_A_8888_SAT,
+	QPU_PACK_A_8A_SAT,
+	QPU_PACK_A_8B_SAT,
+	QPU_PACK_A_8C_SAT,
+	QPU_PACK_A_8D_SAT,
+};
+
+enum qpu_unpack_r4 {
+	QPU_UNPACK_R4_NOP,
+	QPU_UNPACK_R4_F16A_TO_F32,
+	QPU_UNPACK_R4_F16B_TO_F32,
+	QPU_UNPACK_R4_8D_REP,
+	QPU_UNPACK_R4_8A,
+	QPU_UNPACK_R4_8B,
+	QPU_UNPACK_R4_8C,
+	QPU_UNPACK_R4_8D,
+};
+
+#define QPU_MASK(high, low) \
+	((((uint64_t)1 << ((high) - (low) + 1)) - 1) << (low))
+
+#define QPU_GET_FIELD(word, field) \
+	((uint32_t)(((word)  & field ## _MASK) >> field ## _SHIFT))
+
+#define QPU_SIG_SHIFT                   60
+#define QPU_SIG_MASK                    QPU_MASK(63, 60)
+
+#define QPU_UNPACK_SHIFT                57
+#define QPU_UNPACK_MASK                 QPU_MASK(59, 57)
+
+/**
+ * If set, the pack field means PACK_MUL or R4 packing, instead of normal
+ * regfile a packing.
+ */
+#define QPU_PM                          ((uint64_t)1 << 56)
+
+#define QPU_PACK_SHIFT                  52
+#define QPU_PACK_MASK                   QPU_MASK(55, 52)
+
+#define QPU_COND_ADD_SHIFT              49
+#define QPU_COND_ADD_MASK               QPU_MASK(51, 49)
+#define QPU_COND_MUL_SHIFT              46
+#define QPU_COND_MUL_MASK               QPU_MASK(48, 46)
+
+#define QPU_SF                          ((uint64_t)1 << 45)
+
+#define QPU_WADDR_ADD_SHIFT             38
+#define QPU_WADDR_ADD_MASK              QPU_MASK(43, 38)
+#define QPU_WADDR_MUL_SHIFT             32
+#define QPU_WADDR_MUL_MASK              QPU_MASK(37, 32)
+
+#define QPU_OP_MUL_SHIFT                29
+#define QPU_OP_MUL_MASK                 QPU_MASK(31, 29)
+
+#define QPU_RADDR_A_SHIFT               18
+#define QPU_RADDR_A_MASK                QPU_MASK(23, 18)
+#define QPU_RADDR_B_SHIFT               12
+#define QPU_RADDR_B_MASK                QPU_MASK(17, 12)
+#define QPU_SMALL_IMM_SHIFT             12
+#define QPU_SMALL_IMM_MASK              QPU_MASK(17, 12)
+
+#define QPU_ADD_A_SHIFT                 9
+#define QPU_ADD_A_MASK                  QPU_MASK(11, 9)
+#define QPU_ADD_B_SHIFT                 6
+#define QPU_ADD_B_MASK                  QPU_MASK(8, 6)
+#define QPU_MUL_A_SHIFT                 3
+#define QPU_MUL_A_MASK                  QPU_MASK(5, 3)
+#define QPU_MUL_B_SHIFT                 0
+#define QPU_MUL_B_MASK                  QPU_MASK(2, 0)
+
+#define QPU_WS                          ((uint64_t)1 << 44)
+
+#define QPU_OP_ADD_SHIFT                24
+#define QPU_OP_ADD_MASK                 QPU_MASK(28, 24)
+
+#endif /* VC4_QPU_DEFINES_H */
diff --git a/drivers/gpu/drm/vc4/vc4_validate_shaders.c b/drivers/gpu/drm/vc4/vc4_validate_shaders.c
new file mode 100644
index 0000000..f67124b
--- /dev/null
+++ b/drivers/gpu/drm/vc4/vc4_validate_shaders.c
@@ -0,0 +1,513 @@
+/*
+ * Copyright © 2014 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+/**
+ * DOC: Shader validator for VC4.
+ *
+ * The VC4 has no IOMMU between it and system memory, so a user with
+ * access to execute shaders could escalate privilege by overwriting
+ * system memory (using the VPM write address register in the
+ * general-purpose DMA mode) or reading system memory it shouldn't
+ * (reading it as a texture, or uniform data, or vertex data).
+ *
+ * This walks over a shader BO, ensuring that its accesses are
+ * appropriately bounded, and recording how many texture accesses are
+ * made and where so that we can do relocations for them in the
+ * uniform stream.
+ */
+
+#include "vc4_drv.h"
+#include "vc4_qpu_defines.h"
+
+struct vc4_shader_validation_state {
+	struct vc4_texture_sample_info tmu_setup[2];
+	int tmu_write_count[2];
+
+	/* For registers that were last written to by a MIN instruction with
+	 * one argument being a uniform, the address of the uniform.
+	 * Otherwise, ~0.
+	 *
+	 * This is used for the validation of direct address memory reads.
+	 */
+	uint32_t live_min_clamp_offsets[32 + 32 + 4];
+	bool live_max_clamp_regs[32 + 32 + 4];
+};
+
+static uint32_t
+waddr_to_live_reg_index(uint32_t waddr, bool is_b)
+{
+	if (waddr < 32) {
+		if (is_b)
+			return 32 + waddr;
+		else
+			return waddr;
+	} else if (waddr <= QPU_W_ACC3) {
+		return 64 + waddr - QPU_W_ACC0;
+	} else {
+		return ~0;
+	}
+}
+
+static uint32_t
+raddr_add_a_to_live_reg_index(uint64_t inst)
+{
+	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+	uint32_t add_a = QPU_GET_FIELD(inst, QPU_ADD_A);
+	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
+	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
+
+	if (add_a == QPU_MUX_A)
+		return raddr_a;
+	else if (add_a == QPU_MUX_B && sig != QPU_SIG_SMALL_IMM)
+		return 32 + raddr_b;
+	else if (add_a <= QPU_MUX_R3)
+		return 64 + add_a;
+	else
+		return ~0;
+}
+
+static bool
+is_tmu_submit(uint32_t waddr)
+{
+	return (waddr == QPU_W_TMU0_S ||
+		waddr == QPU_W_TMU1_S);
+}
+
+static bool
+is_tmu_write(uint32_t waddr)
+{
+	return (waddr >= QPU_W_TMU0_S &&
+		waddr <= QPU_W_TMU1_B);
+}
+
+static bool
+record_texture_sample(struct vc4_validated_shader_info *validated_shader,
+		      struct vc4_shader_validation_state *validation_state,
+		      int tmu)
+{
+	uint32_t s = validated_shader->num_texture_samples;
+	int i;
+	struct vc4_texture_sample_info *temp_samples;
+
+	temp_samples = krealloc(validated_shader->texture_samples,
+				(s + 1) * sizeof(*temp_samples),
+				GFP_KERNEL);
+	if (!temp_samples)
+		return false;
+
+	memcpy(&temp_samples[s],
+	       &validation_state->tmu_setup[tmu],
+	       sizeof(*temp_samples));
+
+	validated_shader->num_texture_samples = s + 1;
+	validated_shader->texture_samples = temp_samples;
+
+	for (i = 0; i < 4; i++)
+		validation_state->tmu_setup[tmu].p_offset[i] = ~0;
+
+	return true;
+}
+
+static bool
+check_tmu_write(uint64_t inst,
+		struct vc4_validated_shader_info *validated_shader,
+		struct vc4_shader_validation_state *validation_state,
+		bool is_mul)
+{
+	uint32_t waddr = (is_mul ?
+			  QPU_GET_FIELD(inst, QPU_WADDR_MUL) :
+			  QPU_GET_FIELD(inst, QPU_WADDR_ADD));
+	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
+	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
+	int tmu = waddr > QPU_W_TMU0_B;
+	bool submit = is_tmu_submit(waddr);
+	bool is_direct = submit && validation_state->tmu_write_count[tmu] == 0;
+	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+
+	if (is_direct) {
+		uint32_t add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
+		uint32_t clamp_reg, clamp_offset;
+
+		if (sig == QPU_SIG_SMALL_IMM) {
+			DRM_ERROR("direct TMU read used small immediate\n");
+			return false;
+		}
+
+		/* Make sure that this texture load is an add of the base
+		 * address of the UBO to a clamped offset within the UBO.
+		 */
+		if (is_mul ||
+		    QPU_GET_FIELD(inst, QPU_OP_ADD) != QPU_A_ADD) {
+			DRM_ERROR("direct TMU load wasn't an add\n");
+			return false;
+		}
+
+		/* We assert that the the clamped address is the first
+		 * argument, and the UBO base address is the second argument.
+		 * This is arbitrary, but simpler than supporting flipping the
+		 * two either way.
+		 */
+		clamp_reg = raddr_add_a_to_live_reg_index(inst);
+		if (clamp_reg == ~0) {
+			DRM_ERROR("direct TMU load wasn't clamped\n");
+			return false;
+		}
+
+		clamp_offset = validation_state->live_min_clamp_offsets[clamp_reg];
+		if (clamp_offset == ~0) {
+			DRM_ERROR("direct TMU load wasn't clamped\n");
+			return false;
+		}
+
+		/* Store the clamp value's offset in p1 (see reloc_tex() in
+		 * vc4_validate.c).
+		 */
+		validation_state->tmu_setup[tmu].p_offset[1] =
+			clamp_offset;
+
+		if (!(add_b == QPU_MUX_A && raddr_a == QPU_R_UNIF) &&
+		    !(add_b == QPU_MUX_B && raddr_b == QPU_R_UNIF)) {
+			DRM_ERROR("direct TMU load didn't add to a uniform\n");
+			return false;
+		}
+
+		validation_state->tmu_setup[tmu].is_direct = true;
+	} else {
+		if (raddr_a == QPU_R_UNIF || (sig != QPU_SIG_SMALL_IMM &&
+					      raddr_b == QPU_R_UNIF)) {
+			DRM_ERROR("uniform read in the same instruction as "
+				  "texture setup.\n");
+			return false;
+		}
+	}
+
+	if (validation_state->tmu_write_count[tmu] >= 4) {
+		DRM_ERROR("TMU%d got too many parameters before dispatch\n",
+			  tmu);
+		return false;
+	}
+	validation_state->tmu_setup[tmu].p_offset[validation_state->tmu_write_count[tmu]] =
+		validated_shader->uniforms_size;
+	validation_state->tmu_write_count[tmu]++;
+	/* Since direct uses a RADDR uniform reference, it will get counted in
+	 * check_instruction_reads()
+	 */
+	if (!is_direct)
+		validated_shader->uniforms_size += 4;
+
+	if (submit) {
+		if (!record_texture_sample(validated_shader,
+					   validation_state, tmu)) {
+			return false;
+		}
+
+		validation_state->tmu_write_count[tmu] = 0;
+	}
+
+	return true;
+}
+
+static bool
+check_reg_write(uint64_t inst,
+		struct vc4_validated_shader_info *validated_shader,
+		struct vc4_shader_validation_state *validation_state,
+		bool is_mul)
+{
+	uint32_t waddr = (is_mul ?
+			  QPU_GET_FIELD(inst, QPU_WADDR_MUL) :
+			  QPU_GET_FIELD(inst, QPU_WADDR_ADD));
+
+	switch (waddr) {
+	case QPU_W_UNIFORMS_ADDRESS:
+		/* XXX: We'll probably need to support this for reladdr, but
+		 * it's definitely a security-related one.
+		 */
+		DRM_ERROR("uniforms address load unsupported\n");
+		return false;
+
+	case QPU_W_TLB_COLOR_MS:
+	case QPU_W_TLB_COLOR_ALL:
+	case QPU_W_TLB_Z:
+		/* These only interact with the tile buffer, not main memory,
+		 * so they're safe.
+		 */
+		return true;
+
+	case QPU_W_TMU0_S:
+	case QPU_W_TMU0_T:
+	case QPU_W_TMU0_R:
+	case QPU_W_TMU0_B:
+	case QPU_W_TMU1_S:
+	case QPU_W_TMU1_T:
+	case QPU_W_TMU1_R:
+	case QPU_W_TMU1_B:
+		return check_tmu_write(inst, validated_shader, validation_state,
+				       is_mul);
+
+	case QPU_W_HOST_INT:
+	case QPU_W_TMU_NOSWAP:
+	case QPU_W_TLB_ALPHA_MASK:
+	case QPU_W_MUTEX_RELEASE:
+		/* XXX: I haven't thought about these, so don't support them
+		 * for now.
+		 */
+		DRM_ERROR("Unsupported waddr %d\n", waddr);
+		return false;
+
+	case QPU_W_VPM_ADDR:
+		DRM_ERROR("General VPM DMA unsupported\n");
+		return false;
+
+	case QPU_W_VPM:
+	case QPU_W_VPMVCD_SETUP:
+		/* We allow VPM setup in general, even including VPM DMA
+		 * configuration setup, because the (unsafe) DMA can only be
+		 * triggered by QPU_W_VPM_ADDR writes.
+		 */
+		return true;
+
+	case QPU_W_TLB_STENCIL_SETUP:
+		return true;
+	}
+
+	return true;
+}
+
+static void
+track_live_clamps(uint64_t inst,
+		  struct vc4_validated_shader_info *validated_shader,
+		  struct vc4_shader_validation_state *validation_state)
+{
+	uint32_t op_add = QPU_GET_FIELD(inst, QPU_OP_ADD);
+	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
+	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
+	uint32_t cond_add = QPU_GET_FIELD(inst, QPU_COND_ADD);
+	uint32_t add_a = QPU_GET_FIELD(inst, QPU_ADD_A);
+	uint32_t add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
+	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
+	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
+	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+	bool ws = inst & QPU_WS;
+	uint32_t lri_add_a, lri_add, lri_mul;
+	bool add_a_is_min_0;
+
+	/* Check whether OP_ADD's A argumennt comes from a live MAX(x, 0),
+	 * before we clear previous live state.
+	 */
+	lri_add_a = raddr_add_a_to_live_reg_index(inst);
+	add_a_is_min_0 = (lri_add_a != ~0 &&
+			  validation_state->live_max_clamp_regs[lri_add_a]);
+
+	/* Clear live state for registers written by our instruction. */
+	lri_add = waddr_to_live_reg_index(waddr_add, ws);
+	lri_mul = waddr_to_live_reg_index(waddr_mul, !ws);
+	if (lri_mul != ~0) {
+		validation_state->live_max_clamp_regs[lri_mul] = false;
+		validation_state->live_min_clamp_offsets[lri_mul] = ~0;
+	}
+	if (lri_add != ~0) {
+		validation_state->live_max_clamp_regs[lri_add] = false;
+		validation_state->live_min_clamp_offsets[lri_add] = ~0;
+	} else {
+		/* Nothing further to do for live tracking, since only ADDs
+		 * generate new live clamp registers.
+		 */
+		return;
+	}
+
+	/* Now, handle remaining live clamp tracking for the ADD operation. */
+
+	if (cond_add != QPU_COND_ALWAYS)
+		return;
+
+	if (op_add == QPU_A_MAX) {
+		/* Track live clamps of a value to a minimum of 0 (in either
+		 * arg).
+		 */
+		if (sig != QPU_SIG_SMALL_IMM || raddr_b != 0 ||
+		    (add_a != QPU_MUX_B && add_b != QPU_MUX_B)) {
+			return;
+		}
+
+		validation_state->live_max_clamp_regs[lri_add] = true;
+	} else if (op_add == QPU_A_MIN) {
+		/* Track live clamps of a value clamped to a minimum of 0 and
+		 * a maximum of some uniform's offset.
+		 */
+		if (!add_a_is_min_0)
+			return;
+
+		if (!(add_b == QPU_MUX_A && raddr_a == QPU_R_UNIF) &&
+		    !(add_b == QPU_MUX_B && raddr_b == QPU_R_UNIF &&
+		      sig != QPU_SIG_SMALL_IMM)) {
+			return;
+		}
+
+		validation_state->live_min_clamp_offsets[lri_add] =
+			validated_shader->uniforms_size;
+	}
+}
+
+static bool
+check_instruction_writes(uint64_t inst,
+			 struct vc4_validated_shader_info *validated_shader,
+			 struct vc4_shader_validation_state *validation_state)
+{
+	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
+	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
+	bool ok;
+
+	if (is_tmu_write(waddr_add) && is_tmu_write(waddr_mul)) {
+		DRM_ERROR("ADD and MUL both set up textures\n");
+		return false;
+	}
+
+	ok = (check_reg_write(inst, validated_shader, validation_state,
+			      false) &&
+	      check_reg_write(inst, validated_shader, validation_state,
+			      true));
+
+	track_live_clamps(inst, validated_shader, validation_state);
+
+	return ok;
+}
+
+static bool
+check_instruction_reads(uint64_t inst,
+			struct vc4_validated_shader_info *validated_shader)
+{
+	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
+	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
+	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+
+	if (raddr_a == QPU_R_UNIF ||
+	    (raddr_b == QPU_R_UNIF && sig != QPU_SIG_SMALL_IMM)) {
+		/* This can't overflow the uint32_t, because we're reading 8
+		 * bytes of instruction to increment by 4 here, so we'd
+		 * already be OOM.
+		 */
+		validated_shader->uniforms_size += 4;
+	}
+
+	return true;
+}
+
+struct vc4_validated_shader_info *
+vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
+{
+	bool found_shader_end = false;
+	int shader_end_ip = 0;
+	uint32_t ip, max_ip;
+	uint64_t *shader;
+	struct vc4_validated_shader_info *validated_shader;
+	struct vc4_shader_validation_state validation_state;
+	int i;
+
+	memset(&validation_state, 0, sizeof(validation_state));
+
+	for (i = 0; i < 8; i++)
+		validation_state.tmu_setup[i / 4].p_offset[i % 4] = ~0;
+	for (i = 0; i < ARRAY_SIZE(validation_state.live_min_clamp_offsets); i++)
+		validation_state.live_min_clamp_offsets[i] = ~0;
+
+	shader = shader_obj->vaddr;
+	max_ip = shader_obj->base.size / sizeof(uint64_t);
+
+	validated_shader = kcalloc(1, sizeof(*validated_shader), GFP_KERNEL);
+	if (!validated_shader)
+		return NULL;
+
+	for (ip = 0; ip < max_ip; ip++) {
+		uint64_t inst = shader[ip];
+		uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+
+		switch (sig) {
+		case QPU_SIG_NONE:
+		case QPU_SIG_WAIT_FOR_SCOREBOARD:
+		case QPU_SIG_SCOREBOARD_UNLOCK:
+		case QPU_SIG_COLOR_LOAD:
+		case QPU_SIG_LOAD_TMU0:
+		case QPU_SIG_LOAD_TMU1:
+		case QPU_SIG_PROG_END:
+		case QPU_SIG_SMALL_IMM:
+			if (!check_instruction_writes(inst, validated_shader,
+						      &validation_state)) {
+				DRM_ERROR("Bad write at ip %d\n", ip);
+				goto fail;
+			}
+
+			if (!check_instruction_reads(inst, validated_shader))
+				goto fail;
+
+			if (sig == QPU_SIG_PROG_END) {
+				found_shader_end = true;
+				shader_end_ip = ip;
+			}
+
+			break;
+
+		case QPU_SIG_LOAD_IMM:
+			if (!check_instruction_writes(inst, validated_shader,
+						      &validation_state)) {
+				DRM_ERROR("Bad LOAD_IMM write at ip %d\n", ip);
+				goto fail;
+			}
+			break;
+
+		default:
+			DRM_ERROR("Unsupported QPU signal %d at "
+				  "instruction %d\n", sig, ip);
+			goto fail;
+		}
+
+		/* There are two delay slots after program end is signaled
+		 * that are still executed, then we're finished.
+		 */
+		if (found_shader_end && ip == shader_end_ip + 2)
+			break;
+	}
+
+	if (ip == max_ip) {
+		DRM_ERROR("shader failed to terminate before "
+			  "shader BO end at %zd\n",
+			  shader_obj->base.size);
+		goto fail;
+	}
+
+	/* Again, no chance of integer overflow here because the worst case
+	 * scenario is 8 bytes of uniforms plus handles per 8-byte
+	 * instruction.
+	 */
+	validated_shader->uniforms_src_size =
+		(validated_shader->uniforms_size +
+		 4 * validated_shader->num_texture_samples);
+
+	return validated_shader;
+
+fail:
+	if (validated_shader) {
+		kfree(validated_shader->texture_samples);
+		kfree(validated_shader);
+	}
+	return NULL;
+}
diff --git a/include/uapi/drm/vc4_drm.h b/include/uapi/drm/vc4_drm.h
index 068aab9..0716b50 100644
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -28,9 +28,11 @@
 
 #define DRM_VC4_CREATE_BO                         0x03
 #define DRM_VC4_MMAP_BO                           0x04
+#define DRM_VC4_CREATE_SHADER_BO                  0x05
 
 #define DRM_IOCTL_VC4_CREATE_BO           DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_CREATE_BO, struct drm_vc4_create_bo)
 #define DRM_IOCTL_VC4_MMAP_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_MMAP_BO, struct drm_vc4_mmap_bo)
+#define DRM_IOCTL_VC4_CREATE_SHADER_BO    DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_CREATE_SHADER_BO, struct drm_vc4_create_shader_bo)
 
 /**
  * struct drm_vc4_create_bo - ioctl argument for creating VC4 BOs.
@@ -65,4 +67,27 @@ struct drm_vc4_mmap_bo {
 	uint64_t offset;
 };
 
+/**
+ * struct drm_vc4_create_shader_bo - ioctl argument for creating VC4
+ * shader BOs.
+ *
+ * Since allowing a shader to be overwritten while it's also being
+ * executed from would allow privlege escalation, shaders must be
+ * created using this ioctl, and they can't be mmapped later.
+ */
+struct drm_vc4_create_shader_bo {
+	/* Size of the data argument. */
+	uint32_t size;
+	/* Flags, currently must be 0. */
+	uint32_t flags;
+
+	/* Pointer to the data. */
+	uint64_t data;
+
+	/** Returned GEM handle for the BO. */
+	uint32_t handle;
+	/* Pad, must be 0. */
+	uint32_t pad;
+};
+
 #endif /* _UAPI_VC4_DRM_H_ */
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1281684 — Re: [PATCH 4/9] drm/vc4: Add an API for creating GPU shaders in GEM BOs.

Hi Eric,

On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
> Since we have no MMU, the kernel needs to validate that the submitted
> shader code won't make any accesses to memory that the user doesn't
> control, which involves banning some operations (general purpose DMA
> writes), and tracking where we need to write out pointers for other
> operations (texture sampling).  Once it's validated, we return a GEM
> BO containing the shader, which doesn't allow mapping for write or
> exporting to other subsystems.
>
> Signed-off-by: Eric Anholt <eric@anholt.net>
> ---
>  drivers/gpu/drm/vc4/Makefile               |   3 +-
>  drivers/gpu/drm/vc4/vc4_bo.c               | 140 ++++++++
>  drivers/gpu/drm/vc4/vc4_drv.c              |   9 +-
>  drivers/gpu/drm/vc4/vc4_drv.h              |  50 +++
>  drivers/gpu/drm/vc4/vc4_qpu_defines.h      | 264 +++++++++++++++

May I suggest that one 'exports' the header to something like
libdrm_vc4 ? There are patches in flight for nouveau which in similar
fashion, allow userspare to reuse existing api/abi.

-Emil
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1282288 — Re: [PATCH 4/9] drm/vc4: Add an API for creating GPU shaders in GEM BOs.

[Multipart message — attachments visible in raw view] — view raw

Emil Velikov <emil.l.velikov@gmail.com> writes:

> Hi Eric,
>
> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>> Since we have no MMU, the kernel needs to validate that the submitted
>> shader code won't make any accesses to memory that the user doesn't
>> control, which involves banning some operations (general purpose DMA
>> writes), and tracking where we need to write out pointers for other
>> operations (texture sampling).  Once it's validated, we return a GEM
>> BO containing the shader, which doesn't allow mapping for write or
>> exporting to other subsystems.
>>
>> Signed-off-by: Eric Anholt <eric@anholt.net>
>> ---
>>  drivers/gpu/drm/vc4/Makefile               |   3 +-
>>  drivers/gpu/drm/vc4/vc4_bo.c               | 140 ++++++++
>>  drivers/gpu/drm/vc4/vc4_drv.c              |   9 +-
>>  drivers/gpu/drm/vc4/vc4_drv.h              |  50 +++
>>  drivers/gpu/drm/vc4/vc4_qpu_defines.h      | 264 +++++++++++++++
>
> May I suggest that one 'exports' the header to something like
> libdrm_vc4 ? There are patches in flight for nouveau which in similar
> fashion, allow userspare to reuse existing api/abi.

I'll be putting this header in libdrm for use by Mesa and vc4-gpu-tools.
It will still have to live in the kernel as well, though.

[toc] | [prev] | [next] | [standalone]

#1281198 — [PATCH 6/9] drm/vc4: Bind and initialize the V3D engine.

This is the component of the GPU that does 3D rendering.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/vc4/Makefile      |   1 +
 drivers/gpu/drm/vc4/vc4_debugfs.c |   2 +
 drivers/gpu/drm/vc4/vc4_drv.c     |   1 +
 drivers/gpu/drm/vc4/vc4_drv.h     |  13 +++
 drivers/gpu/drm/vc4/vc4_v3d.c     | 225 ++++++++++++++++++++++++++++++++++++++
 5 files changed, 242 insertions(+)
 create mode 100644 drivers/gpu/drm/vc4/vc4_v3d.c

diff --git a/drivers/gpu/drm/vc4/Makefile b/drivers/gpu/drm/vc4/Makefile
index eb776a6..e87a6f2 100644
--- a/drivers/gpu/drm/vc4/Makefile
+++ b/drivers/gpu/drm/vc4/Makefile
@@ -11,6 +11,7 @@ vc4-y := \
 	vc4_hdmi.o \
 	vc4_hvs.o \
 	vc4_plane.o \
+	vc4_v3d.o \
 	vc4_validate_shaders.o
 
 vc4-$(CONFIG_DEBUG_FS) += vc4_debugfs.o
diff --git a/drivers/gpu/drm/vc4/vc4_debugfs.c b/drivers/gpu/drm/vc4/vc4_debugfs.c
index 6bcf96e..d76ad10 100644
--- a/drivers/gpu/drm/vc4/vc4_debugfs.c
+++ b/drivers/gpu/drm/vc4/vc4_debugfs.c
@@ -22,6 +22,8 @@ static const struct drm_info_list vc4_debugfs_list[] = {
 	{"crtc0_regs", vc4_crtc_debugfs_regs, 0, (void *)(uintptr_t)0},
 	{"crtc1_regs", vc4_crtc_debugfs_regs, 0, (void *)(uintptr_t)1},
 	{"crtc2_regs", vc4_crtc_debugfs_regs, 0, (void *)(uintptr_t)2},
+	{"v3d_ident", vc4_v3d_debugfs_ident, 0},
+	{"v3d_regs", vc4_v3d_debugfs_regs, 0},
 };
 
 #define VC4_DEBUGFS_ENTRIES ARRAY_SIZE(vc4_debugfs_list)
diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index da4be9c..db58d74 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -236,6 +236,7 @@ static struct platform_driver *const component_drivers[] = {
 	&vc4_hdmi_driver,
 	&vc4_crtc_driver,
 	&vc4_hvs_driver,
+	&vc4_v3d_driver,
 };
 
 static int vc4_platform_drm_probe(struct platform_device *pdev)
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index bd77d55..8945463 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -15,6 +15,7 @@ struct vc4_dev {
 	struct vc4_hdmi *hdmi;
 	struct vc4_hvs *hvs;
 	struct vc4_crtc *crtc[3];
+	struct vc4_v3d *v3d;
 
 	struct drm_fbdev_cma *fbdev;
 
@@ -82,6 +83,11 @@ to_vc4_bo(struct drm_gem_object *bo)
 	return (struct vc4_bo *)bo;
 }
 
+struct vc4_v3d {
+	struct platform_device *pdev;
+	void __iomem *regs;
+};
+
 struct vc4_hvs {
 	struct platform_device *pdev;
 	void __iomem *regs;
@@ -119,6 +125,8 @@ to_vc4_encoder(struct drm_encoder *encoder)
 	return container_of(encoder, struct vc4_encoder, base);
 }
 
+#define V3D_READ(offset) readl(vc4->v3d->regs + offset)
+#define V3D_WRITE(offset, val) writel(val, vc4->v3d->regs + offset)
 #define HVS_READ(offset) readl(vc4->hvs->regs + offset)
 #define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
 
@@ -241,6 +249,11 @@ struct drm_plane *vc4_plane_init(struct drm_device *dev,
 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
 u32 vc4_plane_dlist_size(struct drm_plane_state *state);
 
+/* vc4_v3d.c */
+extern struct platform_driver vc4_v3d_driver;
+int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused);
+int vc4_v3d_debugfs_regs(struct seq_file *m, void *unused);
+
 /* vc4_validate_shader.c */
 struct vc4_validated_shader_info *
 vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
diff --git a/drivers/gpu/drm/vc4/vc4_v3d.c b/drivers/gpu/drm/vc4/vc4_v3d.c
new file mode 100644
index 0000000..040ad0d
--- /dev/null
+++ b/drivers/gpu/drm/vc4/vc4_v3d.c
@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2014 The Linux Foundation. All rights reserved.
+ * Copyright (C) 2013 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 "linux/component.h"
+#include "vc4_drv.h"
+#include "vc4_regs.h"
+
+#ifdef CONFIG_DEBUG_FS
+#define REGDEF(reg) { reg, #reg }
+static const struct {
+	uint32_t reg;
+	const char *name;
+} vc4_reg_defs[] = {
+	REGDEF(V3D_IDENT0),
+	REGDEF(V3D_IDENT1),
+	REGDEF(V3D_IDENT2),
+	REGDEF(V3D_SCRATCH),
+	REGDEF(V3D_L2CACTL),
+	REGDEF(V3D_SLCACTL),
+	REGDEF(V3D_INTCTL),
+	REGDEF(V3D_INTENA),
+	REGDEF(V3D_INTDIS),
+	REGDEF(V3D_CT0CS),
+	REGDEF(V3D_CT1CS),
+	REGDEF(V3D_CT0EA),
+	REGDEF(V3D_CT1EA),
+	REGDEF(V3D_CT0CA),
+	REGDEF(V3D_CT1CA),
+	REGDEF(V3D_CT00RA0),
+	REGDEF(V3D_CT01RA0),
+	REGDEF(V3D_CT0LC),
+	REGDEF(V3D_CT1LC),
+	REGDEF(V3D_CT0PC),
+	REGDEF(V3D_CT1PC),
+	REGDEF(V3D_PCS),
+	REGDEF(V3D_BFC),
+	REGDEF(V3D_RFC),
+	REGDEF(V3D_BPCA),
+	REGDEF(V3D_BPCS),
+	REGDEF(V3D_BPOA),
+	REGDEF(V3D_BPOS),
+	REGDEF(V3D_BXCF),
+	REGDEF(V3D_SQRSV0),
+	REGDEF(V3D_SQRSV1),
+	REGDEF(V3D_SQCNTL),
+	REGDEF(V3D_SRQPC),
+	REGDEF(V3D_SRQUA),
+	REGDEF(V3D_SRQUL),
+	REGDEF(V3D_SRQCS),
+	REGDEF(V3D_VPACNTL),
+	REGDEF(V3D_VPMBASE),
+	REGDEF(V3D_PCTRC),
+	REGDEF(V3D_PCTRE),
+	REGDEF(V3D_PCTR0),
+	REGDEF(V3D_PCTRS0),
+	REGDEF(V3D_PCTR1),
+	REGDEF(V3D_PCTRS1),
+	REGDEF(V3D_PCTR2),
+	REGDEF(V3D_PCTRS2),
+	REGDEF(V3D_PCTR3),
+	REGDEF(V3D_PCTRS3),
+	REGDEF(V3D_PCTR4),
+	REGDEF(V3D_PCTRS4),
+	REGDEF(V3D_PCTR5),
+	REGDEF(V3D_PCTRS5),
+	REGDEF(V3D_PCTR6),
+	REGDEF(V3D_PCTRS6),
+	REGDEF(V3D_PCTR7),
+	REGDEF(V3D_PCTRS7),
+	REGDEF(V3D_PCTR8),
+	REGDEF(V3D_PCTRS8),
+	REGDEF(V3D_PCTR9),
+	REGDEF(V3D_PCTRS9),
+	REGDEF(V3D_PCTR10),
+	REGDEF(V3D_PCTRS10),
+	REGDEF(V3D_PCTR11),
+	REGDEF(V3D_PCTRS11),
+	REGDEF(V3D_PCTR12),
+	REGDEF(V3D_PCTRS12),
+	REGDEF(V3D_PCTR13),
+	REGDEF(V3D_PCTRS13),
+	REGDEF(V3D_PCTR14),
+	REGDEF(V3D_PCTRS14),
+	REGDEF(V3D_PCTR15),
+	REGDEF(V3D_PCTRS15),
+	REGDEF(V3D_DBGE),
+	REGDEF(V3D_FDBGO),
+	REGDEF(V3D_FDBGB),
+	REGDEF(V3D_FDBGR),
+	REGDEF(V3D_FDBGS),
+	REGDEF(V3D_ERRSTAT),
+};
+
+int vc4_v3d_debugfs_regs(struct seq_file *m, void *unused)
+{
+	struct drm_info_node *node = (struct drm_info_node *)m->private;
+	struct drm_device *dev = node->minor->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(vc4_reg_defs); i++) {
+		seq_printf(m, "%s (0x%04x): 0x%08x\n",
+			   vc4_reg_defs[i].name, vc4_reg_defs[i].reg,
+			   V3D_READ(vc4_reg_defs[i].reg));
+	}
+
+	return 0;
+}
+
+int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
+{
+	struct drm_info_node *node = (struct drm_info_node *)m->private;
+	struct drm_device *dev = node->minor->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	uint32_t ident1 = V3D_READ(V3D_IDENT1);
+	uint32_t nslc = VC4_GET_FIELD(ident1, V3D_IDENT1_NSLC);
+	uint32_t tups = VC4_GET_FIELD(ident1, V3D_IDENT1_TUPS);
+	uint32_t qups = VC4_GET_FIELD(ident1, V3D_IDENT1_QUPS);
+
+	seq_printf(m, "Revision:   %d\n",
+		   VC4_GET_FIELD(ident1, V3D_IDENT1_REV));
+	seq_printf(m, "Slices:     %d\n", nslc);
+	seq_printf(m, "TMUs:       %d\n", nslc * tups);
+	seq_printf(m, "QPUs:       %d\n", nslc * qups);
+	seq_printf(m, "Semaphores: %d\n",
+		   VC4_GET_FIELD(ident1, V3D_IDENT1_NSEM));
+
+	return 0;
+}
+#endif /* CONFIG_DEBUG_FS */
+
+static void vc4_v3d_init_hw(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	/* Take all the memory that would have been reserved for user
+	 * QPU programs, since we don't have an interface for running
+	 * them, anyway.
+	 */
+	V3D_WRITE(V3D_VPMBASE, 0);
+}
+
+static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
+{
+	struct platform_device *pdev = to_platform_device(dev);
+	struct drm_device *drm = dev_get_drvdata(master);
+	struct vc4_dev *vc4 = to_vc4_dev(drm);
+	struct vc4_v3d *v3d = NULL;
+
+	v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
+	if (!v3d)
+		return -ENOMEM;
+
+	v3d->pdev = pdev;
+
+	v3d->regs = vc4_ioremap_regs(pdev, 0);
+	if (IS_ERR(v3d->regs))
+		return PTR_ERR(v3d->regs);
+
+	vc4->v3d = v3d;
+
+	if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
+		DRM_ERROR("V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
+			  V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
+		return -EINVAL;
+	}
+
+	vc4_v3d_init_hw(drm);
+
+	return 0;
+}
+
+static void vc4_v3d_unbind(struct device *dev, struct device *master,
+			   void *data)
+{
+	struct drm_device *drm = dev_get_drvdata(master);
+	struct vc4_dev *vc4 = to_vc4_dev(drm);
+
+	vc4->v3d = NULL;
+}
+
+static const struct component_ops vc4_v3d_ops = {
+	.bind   = vc4_v3d_bind,
+	.unbind = vc4_v3d_unbind,
+};
+
+static int vc4_v3d_dev_probe(struct platform_device *pdev)
+{
+	return component_add(&pdev->dev, &vc4_v3d_ops);
+}
+
+static int vc4_v3d_dev_remove(struct platform_device *pdev)
+{
+	component_del(&pdev->dev, &vc4_v3d_ops);
+	return 0;
+}
+
+static const struct of_device_id vc4_v3d_dt_match[] = {
+	{ .compatible = "brcm,vc4-v3d" },
+	{}
+};
+
+struct platform_driver vc4_v3d_driver = {
+	.probe = vc4_v3d_dev_probe,
+	.remove = vc4_v3d_dev_remove,
+	.driver = {
+		.name = "vc4_v3d",
+		.of_match_table = vc4_v3d_dt_match,
+	},
+};
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1281199 — [PATCH 8/9] drm/vc4: Add support for async pageflips.

An async pageflip stores the modeset to be done and executes it once
the BOs are ready to be displayed.  This gets us about 3x performance
in full screen rendering with pageflipping.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/vc4/vc4_crtc.c  |  99 +++++++++++++++++++++++++-
 drivers/gpu/drm/vc4/vc4_drv.h   |  16 +++++
 drivers/gpu/drm/vc4/vc4_gem.c   |  40 +++++++++++
 drivers/gpu/drm/vc4/vc4_kms.c   | 149 +++++++++++++++++++++++++++++++++++++++-
 drivers/gpu/drm/vc4/vc4_plane.c |  40 +++++++++++
 5 files changed, 342 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/vc4/vc4_crtc.c b/drivers/gpu/drm/vc4/vc4_crtc.c
index 7a9f476..a319332 100644
--- a/drivers/gpu/drm/vc4/vc4_crtc.c
+++ b/drivers/gpu/drm/vc4/vc4_crtc.c
@@ -35,6 +35,7 @@
 #include "drm_atomic_helper.h"
 #include "drm_crtc_helper.h"
 #include "linux/clk.h"
+#include "drm_fb_cma_helper.h"
 #include "linux/component.h"
 #include "linux/of_device.h"
 #include "vc4_drv.h"
@@ -475,10 +476,106 @@ static irqreturn_t vc4_crtc_irq_handler(int irq, void *data)
 	return ret;
 }
 
+struct vc4_async_flip_state {
+	struct drm_crtc *crtc;
+	struct drm_framebuffer *fb;
+	struct drm_pending_vblank_event *event;
+
+	struct vc4_seqno_cb cb;
+};
+
+/* Called when the V3D execution for the BO being flipped to is done, so that
+ * we can actually update the plane's address to point to it.
+ */
+static void
+vc4_async_page_flip_complete(struct vc4_seqno_cb *cb)
+{
+	struct vc4_async_flip_state *flip_state =
+		container_of(cb, struct vc4_async_flip_state, cb);
+	struct drm_crtc *crtc = flip_state->crtc;
+	struct drm_device *dev = crtc->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_plane *plane = crtc->primary;
+
+	vc4_plane_async_set_fb(plane, flip_state->fb);
+	if (flip_state->event) {
+		unsigned long flags;
+
+		spin_lock_irqsave(&dev->event_lock, flags);
+		drm_crtc_send_vblank_event(crtc, flip_state->event);
+		spin_unlock_irqrestore(&dev->event_lock, flags);
+	}
+
+	drm_framebuffer_unreference(flip_state->fb);
+	kfree(flip_state);
+
+	up(&vc4->async_modeset);
+}
+
+/* Implements async (non-vblank-synced) page flips.
+ *
+ * The page flip ioctl needs to return immediately, so we grab the
+ * modeset semaphore on the pipe, and queue the address update for
+ * when V3D is done with the BO being flipped to.
+ */
+static int vc4_async_page_flip(struct drm_crtc *crtc,
+			       struct drm_framebuffer *fb,
+			       struct drm_pending_vblank_event *event,
+			       uint32_t flags)
+{
+	struct drm_device *dev = crtc->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_plane *plane = crtc->primary;
+	int ret = 0;
+	struct vc4_async_flip_state *flip_state;
+	struct drm_gem_cma_object *cma_bo = drm_fb_cma_get_gem_obj(fb, 0);
+	struct vc4_bo *bo = to_vc4_bo(&cma_bo->base);
+
+	flip_state = kzalloc(sizeof(*flip_state), GFP_KERNEL);
+	if (!flip_state)
+		return -ENOMEM;
+
+	drm_framebuffer_reference(fb);
+	flip_state->fb = fb;
+	flip_state->crtc = crtc;
+	flip_state->event = event;
+
+	/* Make sure all other async modesetes have landed. */
+	ret = down_interruptible(&vc4->async_modeset);
+	if (ret) {
+		kfree(flip_state);
+		return ret;
+	}
+
+	/* Immediately update the plane's legacy fb pointer, so that later
+	 * modeset prep sees the state that will be present when the semaphore
+	 * is released.
+	 */
+	drm_atomic_set_fb_for_plane(plane->state, fb);
+	plane->fb = fb;
+
+	vc4_queue_seqno_cb(dev, &flip_state->cb, bo->seqno,
+			   vc4_async_page_flip_complete);
+
+	/* Driver takes ownership of state on successful async commit. */
+	return 0;
+}
+
+static int vc4_page_flip(struct drm_crtc *crtc,
+			 struct drm_framebuffer *fb,
+			 struct drm_pending_vblank_event *event,
+			 uint32_t flags)
+{
+	if (flags & DRM_MODE_PAGE_FLIP_ASYNC)
+		return vc4_async_page_flip(crtc, fb, event, flags);
+	else
+		return drm_atomic_helper_page_flip(crtc, fb, event, flags);
+}
+
 static const struct drm_crtc_funcs vc4_crtc_funcs = {
 	.set_config = drm_atomic_helper_set_config,
 	.destroy = vc4_crtc_destroy,
-	.page_flip = drm_atomic_helper_page_flip,
+	.page_flip = vc4_page_flip,
 	.set_property = NULL,
 	.cursor_set = NULL, /* handled by drm_mode_cursor_universal */
 	.cursor_move = NULL, /* handled by drm_mode_cursor_universal */
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index a3dbfee..9a8ee23 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -76,6 +76,11 @@ struct vc4_dev {
 	wait_queue_head_t job_wait_queue;
 	struct work_struct job_done_work;
 
+	/* List of struct vc4_seqno_cb for callbacks to be made from a
+	 * workqueue when the given seqno is passed.
+	 */
+	struct list_head seqno_cb_list;
+
 	/* The binner overflow memory that's currently set up in
 	 * BPOA/BPOS registers.  When overflow occurs and a new one is
 	 * allocated, the previous one will be moved to
@@ -128,6 +133,12 @@ to_vc4_bo(struct drm_gem_object *bo)
 	return (struct vc4_bo *)bo;
 }
 
+struct vc4_seqno_cb {
+	struct work_struct work;
+	uint64_t seqno;
+	void (*func)(struct vc4_seqno_cb *cb);
+};
+
 struct vc4_v3d {
 	struct platform_device *pdev;
 	void __iomem *regs;
@@ -392,6 +403,9 @@ void vc4_submit_next_job(struct drm_device *dev);
 int vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno,
 		       uint64_t timeout_ns, bool interruptible);
 void vc4_job_handle_completed(struct vc4_dev *vc4);
+int vc4_queue_seqno_cb(struct drm_device *dev,
+		       struct vc4_seqno_cb *cb, uint64_t seqno,
+		       void (*func)(struct vc4_seqno_cb *cb));
 
 /* vc4_hdmi.c */
 extern struct platform_driver vc4_hdmi_driver;
@@ -417,6 +431,8 @@ struct drm_plane *vc4_plane_init(struct drm_device *dev,
 				 enum drm_plane_type type);
 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist);
 u32 vc4_plane_dlist_size(struct drm_plane_state *state);
+void vc4_plane_async_set_fb(struct drm_plane *plane,
+			    struct drm_framebuffer *fb);
 
 /* vc4_v3d.c */
 extern struct platform_driver vc4_v3d_driver;
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index ddcb018..a62ca68 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -459,6 +459,7 @@ void
 vc4_job_handle_completed(struct vc4_dev *vc4)
 {
 	unsigned long irqflags;
+	struct vc4_seqno_cb *cb, *cb_temp;
 
 	spin_lock_irqsave(&vc4->job_lock, irqflags);
 	while (!list_empty(&vc4->job_done_list)) {
@@ -471,7 +472,45 @@ vc4_job_handle_completed(struct vc4_dev *vc4)
 		vc4_complete_exec(vc4->dev, exec);
 		spin_lock_irqsave(&vc4->job_lock, irqflags);
 	}
+
+	list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
+		if (cb->seqno <= vc4->finished_seqno) {
+			list_del_init(&cb->work.entry);
+			schedule_work(&cb->work);
+		}
+	}
+
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+}
+
+static void vc4_seqno_cb_work(struct work_struct *work)
+{
+	struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
+
+	cb->func(cb);
+}
+
+int vc4_queue_seqno_cb(struct drm_device *dev,
+		       struct vc4_seqno_cb *cb, uint64_t seqno,
+		       void (*func)(struct vc4_seqno_cb *cb))
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret = 0;
+	unsigned long irqflags;
+
+	cb->func = func;
+	INIT_WORK(&cb->work, vc4_seqno_cb_work);
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	if (seqno > vc4->finished_seqno) {
+		cb->seqno = seqno;
+		list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
+	} else {
+		schedule_work(&cb->work);
+	}
 	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	return ret;
 }
 
 /* Scheduled when any job has been completed, this walks the list of
@@ -608,6 +647,7 @@ vc4_gem_init(struct drm_device *dev)
 
 	INIT_LIST_HEAD(&vc4->job_list);
 	INIT_LIST_HEAD(&vc4->job_done_list);
+	INIT_LIST_HEAD(&vc4->seqno_cb_list);
 	spin_lock_init(&vc4->job_lock);
 
 	INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
diff --git a/drivers/gpu/drm/vc4/vc4_kms.c b/drivers/gpu/drm/vc4/vc4_kms.c
index 2e5597d..f95f2df 100644
--- a/drivers/gpu/drm/vc4/vc4_kms.c
+++ b/drivers/gpu/drm/vc4/vc4_kms.c
@@ -15,6 +15,7 @@
  */
 
 #include "drm_crtc.h"
+#include "drm_atomic.h"
 #include "drm_atomic_helper.h"
 #include "drm_crtc_helper.h"
 #include "drm_plane_helper.h"
@@ -29,10 +30,152 @@ static void vc4_output_poll_changed(struct drm_device *dev)
 		drm_fbdev_cma_hotplug_event(vc4->fbdev);
 }
 
+struct vc4_commit {
+	struct drm_device *dev;
+	struct drm_atomic_state *state;
+	struct vc4_seqno_cb cb;
+};
+
+static void
+vc4_atomic_complete_commit(struct vc4_commit *c)
+{
+	struct drm_atomic_state *state = c->state;
+	struct drm_device *dev = state->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	drm_atomic_helper_commit_modeset_disables(dev, state);
+
+	drm_atomic_helper_commit_planes(dev, state, false);
+
+	drm_atomic_helper_commit_modeset_enables(dev, state);
+
+	drm_atomic_helper_wait_for_vblanks(dev, state);
+
+	drm_atomic_helper_cleanup_planes(dev, state);
+
+	drm_atomic_state_free(state);
+
+	up(&vc4->async_modeset);
+
+	kfree(c);
+}
+
+static void
+vc4_atomic_complete_commit_seqno_cb(struct vc4_seqno_cb *cb)
+{
+	struct vc4_commit *c = container_of(cb, struct vc4_commit, cb);
+
+	vc4_atomic_complete_commit(c);
+}
+
+static struct vc4_commit *commit_init(struct drm_atomic_state *state)
+{
+	struct vc4_commit *c = kzalloc(sizeof(*c), GFP_KERNEL);
+
+	if (!c)
+		return NULL;
+	c->dev = state->dev;
+	c->state = state;
+
+	return c;
+}
+
+/**
+ * vc4_atomic_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.
+ */
+static int vc4_atomic_commit(struct drm_device *dev,
+			     struct drm_atomic_state *state,
+			     bool async)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret;
+	int i;
+	uint64_t wait_seqno = 0;
+	struct vc4_commit *c;
+
+	c = commit_init(state);
+	if (!c)
+		return -ENOMEM;
+
+	/* Make sure that any outstanding modesets have finished. */
+	ret = down_interruptible(&vc4->async_modeset);
+	if (ret) {
+		kfree(c);
+		return ret;
+	}
+
+	ret = drm_atomic_helper_prepare_planes(dev, state);
+	if (ret) {
+		kfree(c);
+		up(&vc4->async_modeset);
+		return ret;
+	}
+
+	for (i = 0; i < dev->mode_config.num_total_plane; 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) && new_state->fb) {
+			struct drm_gem_cma_object *cma_bo =
+				drm_fb_cma_get_gem_obj(new_state->fb, 0);
+			struct vc4_bo *bo = to_vc4_bo(&cma_bo->base);
+
+			wait_seqno = max(bo->seqno, wait_seqno);
+		}
+	}
+
+	/*
+	 * 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 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.
+	 */
+
+	if (async) {
+		vc4_queue_seqno_cb(dev, &c->cb, wait_seqno,
+				   vc4_atomic_complete_commit_seqno_cb);
+	} else {
+		vc4_wait_for_seqno(dev, wait_seqno, ~0ull, false);
+		vc4_atomic_complete_commit(c);
+	}
+
+	return 0;
+}
+
 static const struct drm_mode_config_funcs vc4_mode_funcs = {
 	.output_poll_changed = vc4_output_poll_changed,
 	.atomic_check = drm_atomic_helper_check,
-	.atomic_commit = drm_atomic_helper_commit,
+	.atomic_commit = vc4_atomic_commit,
 	.fb_create = drm_fb_cma_create,
 };
 
@@ -41,6 +184,8 @@ int vc4_kms_load(struct drm_device *dev)
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	int ret;
 
+	sema_init(&vc4->async_modeset, 1);
+
 	ret = drm_vblank_init(dev, dev->mode_config.num_crtc);
 	if (ret < 0) {
 		dev_err(dev->dev, "failed to initialize vblank\n");
@@ -51,6 +196,8 @@ int vc4_kms_load(struct drm_device *dev)
 	dev->mode_config.max_height = 2048;
 	dev->mode_config.funcs = &vc4_mode_funcs;
 	dev->mode_config.preferred_depth = 24;
+	dev->mode_config.async_page_flip = true;
+
 	dev->vblank_disable_allowed = true;
 
 	drm_mode_config_reset(dev);
diff --git a/drivers/gpu/drm/vc4/vc4_plane.c b/drivers/gpu/drm/vc4/vc4_plane.c
index cdd8b10..db32c373 100644
--- a/drivers/gpu/drm/vc4/vc4_plane.c
+++ b/drivers/gpu/drm/vc4/vc4_plane.c
@@ -29,6 +29,14 @@ struct vc4_plane_state {
 	u32 *dlist;
 	u32 dlist_size; /* Number of dwords in allocated for the display list */
 	u32 dlist_count; /* Number of used dwords in the display list. */
+
+	/* Offset in the dlist to pointer word 0. */
+	u32 pw0_offset;
+
+	/* Offset where the plane's dlist was last stored in the
+	   hardware at vc4_crtc_atomic_flush() time.
+	*/
+	u32 *hw_dlist;
 };
 
 static inline struct vc4_plane_state *
@@ -197,6 +205,8 @@ static int vc4_plane_mode_set(struct drm_plane *plane,
 	/* Position Word 3: Context.  Written by the HVS. */
 	vc4_dlist_write(vc4_state, 0xc0c0c0c0);
 
+	vc4_state->pw0_offset = vc4_state->dlist_count;
+
 	/* Pointer Word 0: RGB / Y Pointer */
 	vc4_dlist_write(vc4_state, bo->paddr + offset);
 
@@ -248,6 +258,8 @@ u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist)
 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
 	int i;
 
+	vc4_state->hw_dlist = dlist;
+
 	/* Can't memcpy_toio() because it needs to be 32-bit writes. */
 	for (i = 0; i < vc4_state->dlist_count; i++)
 		writel(vc4_state->dlist[i], &dlist[i]);
@@ -262,6 +274,34 @@ u32 vc4_plane_dlist_size(struct drm_plane_state *state)
 	return vc4_state->dlist_count;
 }
 
+/* Updates the plane to immediately (well, once the FIFO needs
+ * refilling) scan out from at a new framebuffer.
+ */
+void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb)
+{
+	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
+	struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
+	uint32_t addr;
+
+	/* We're skipping the address adjustment for negative origin,
+	 * because this is only called on the primary plane.
+	 */
+	WARN_ON_ONCE(plane->state->crtc_x < 0 || plane->state->crtc_y < 0);
+	addr = bo->paddr + fb->offsets[0];
+
+	/* Write the new address into the hardware immediately.  The
+	 * scanout will start from this address as soon as the FIFO
+	 * needs to refill with pixels.
+	 */
+	writel(addr, &vc4_state->hw_dlist[vc4_state->pw0_offset]);
+
+	/* Also update the CPU-side dlist copy, so that any later
+	 * atomic updates that don't do a new modeset on our plane
+	 * also use our updated address.
+	 */
+	vc4_state->dlist[vc4_state->pw0_offset] = addr;
+}
+
 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
 	.prepare_fb = NULL,
 	.cleanup_fb = NULL,
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1281819 — Re: [PATCH 8/9] drm/vc4: Add support for async pageflips.

Hi,

On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
> An async pageflip stores the modeset to be done and executes it once
> the BOs are ready to be displayed.  This gets us about 3x performance
> in full screen rendering with pageflipping.

Looks good, but you're missing a preclose callback to reap dead events, a la:
https://git.collabora.com/cgit/user/daniels/linux.git/commit/?h=wip/4.4.x/rockchip-drm-fixes&id=d14f21bcd7e7a1b9ca129c411a9da9c911037965

(and parent commit to wire that through to core DRM preclose)

Cheers,
Daniel
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1283537 — Re: [PATCH 8/9] drm/vc4: Add support for async pageflips.

[Multipart message — attachments visible in raw view] — view raw

Daniel Stone <daniel@fooishbar.org> writes:

> Hi,
>
> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>> An async pageflip stores the modeset to be done and executes it once
>> the BOs are ready to be displayed.  This gets us about 3x performance
>> in full screen rendering with pageflipping.
>
> Looks good, but you're missing a preclose callback to reap dead events, a la:
> https://git.collabora.com/cgit/user/daniels/linux.git/commit/?h=wip/4.4.x/rockchip-drm-fixes&id=d14f21bcd7e7a1b9ca129c411a9da9c911037965
>
> (and parent commit to wire that through to core DRM preclose)

We've already got a preclose callback that reaps crtc->event -- see
vc4_cancel_page_flip().

[toc] | [prev] | [next] | [standalone]

#1285174 — Re: [PATCH 8/9] drm/vc4: Add support for async pageflips.

Hi,

On 4 December 2015 at 01:42, Eric Anholt <eric@anholt.net> wrote:
> Daniel Stone <daniel@fooishbar.org> writes:
>> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>>> An async pageflip stores the modeset to be done and executes it once
>>> the BOs are ready to be displayed.  This gets us about 3x performance
>>> in full screen rendering with pageflipping.
>>
>> Looks good, but you're missing a preclose callback to reap dead events, a la:
>> https://git.collabora.com/cgit/user/daniels/linux.git/commit/?h=wip/4.4.x/rockchip-drm-fixes&id=d14f21bcd7e7a1b9ca129c411a9da9c911037965
>>
>> (and parent commit to wire that through to core DRM preclose)
>
> We've already got a preclose callback that reaps crtc->event -- see
> vc4_cancel_page_flip().

Right you are. Carry on then. :)

Cheers,
Daniel
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1281200 — [PATCH 3/9] drm/vc4: Add create and map BO ioctls.

While there exist dumb APIs for creating and mapping BOs, one of the
rules is that drivers doing 3D acceleration have to provide their own
APIs for buffer allocation (besides, the pitch/height parameters of
the dumb alloc don't really make sense for a lot of 3D allocations).

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/vc4/vc4_bo.c  | 41 ++++++++++++++++++++++++++
 drivers/gpu/drm/vc4/vc4_drv.c |  3 ++
 drivers/gpu/drm/vc4/vc4_drv.h |  4 +++
 include/uapi/drm/Kbuild       |  1 +
 include/uapi/drm/vc4_drm.h    | 68 +++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 117 insertions(+)
 create mode 100644 include/uapi/drm/vc4_drm.h

diff --git a/drivers/gpu/drm/vc4/vc4_bo.c b/drivers/gpu/drm/vc4/vc4_bo.c
index 18faa5b..06cba26 100644
--- a/drivers/gpu/drm/vc4/vc4_bo.c
+++ b/drivers/gpu/drm/vc4/vc4_bo.c
@@ -19,6 +19,7 @@
  */
 
 #include "vc4_drv.h"
+#include "uapi/drm/vc4_drm.h"
 
 static void vc4_bo_stats_dump(struct vc4_dev *vc4)
 {
@@ -346,6 +347,46 @@ static void vc4_bo_cache_time_timer(unsigned long data)
 	schedule_work(&vc4->bo_cache.time_work);
 }
 
+int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
+			struct drm_file *file_priv)
+{
+	struct drm_vc4_create_bo *args = data;
+	struct vc4_bo *bo = NULL;
+	int ret;
+
+	/*
+	 * We can't allocate from the BO cache, because the BOs don't
+	 * get zeroed, and that might leak data between users.
+	 */
+	bo = vc4_bo_create(dev, args->size, false);
+	if (!bo)
+		return -ENOMEM;
+
+	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
+	drm_gem_object_unreference_unlocked(&bo->base.base);
+
+	return ret;
+}
+
+int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
+		      struct drm_file *file_priv)
+{
+	struct drm_vc4_mmap_bo *args = data;
+	struct drm_gem_object *gem_obj;
+
+	gem_obj = drm_gem_object_lookup(dev, file_priv, args->handle);
+	if (!gem_obj) {
+		DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
+		return -EINVAL;
+	}
+
+	/* The mmap offset was set up at BO allocation time. */
+	args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
+
+	drm_gem_object_unreference_unlocked(gem_obj);
+	return 0;
+}
+
 void vc4_bo_cache_init(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index da041fa..5fa4688 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -16,6 +16,7 @@
 #include <linux/platform_device.h>
 #include "drm_fb_cma_helper.h"
 
+#include "uapi/drm/vc4_drm.h"
 #include "vc4_drv.h"
 #include "vc4_regs.h"
 
@@ -73,6 +74,8 @@ static const struct file_operations vc4_drm_fops = {
 };
 
 static const struct drm_ioctl_desc vc4_drm_ioctls[] = {
+	DRM_IOCTL_DEF_DRV(VC4_CREATE_BO, vc4_create_bo_ioctl, 0),
+	DRM_IOCTL_DEF_DRV(VC4_MMAP_BO, vc4_mmap_bo_ioctl, 0),
 };
 
 static struct drm_driver vc4_drm_driver = {
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index 39a1ff5..fddb0a0 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -155,6 +155,10 @@ int vc4_dumb_create(struct drm_file *file_priv,
 		    struct drm_mode_create_dumb *args);
 struct dma_buf *vc4_prime_export(struct drm_device *dev,
 				 struct drm_gem_object *obj, int flags);
+int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
+			struct drm_file *file_priv);
+int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
+		      struct drm_file *file_priv);
 void vc4_bo_cache_init(struct drm_device *dev);
 void vc4_bo_cache_destroy(struct drm_device *dev);
 int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
diff --git a/include/uapi/drm/Kbuild b/include/uapi/drm/Kbuild
index 38d4370..974fcd5 100644
--- a/include/uapi/drm/Kbuild
+++ b/include/uapi/drm/Kbuild
@@ -17,4 +17,5 @@ header-y += tegra_drm.h
 header-y += via_drm.h
 header-y += vmwgfx_drm.h
 header-y += msm_drm.h
+header-y += vc4_drm.h
 header-y += virtgpu_drm.h
diff --git a/include/uapi/drm/vc4_drm.h b/include/uapi/drm/vc4_drm.h
new file mode 100644
index 0000000..068aab9
--- /dev/null
+++ b/include/uapi/drm/vc4_drm.h
@@ -0,0 +1,68 @@
+/*
+ * Copyright © 2014-2015 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#ifndef _UAPI_VC4_DRM_H_
+#define _UAPI_VC4_DRM_H_
+
+#include <drm/drm.h>
+
+#define DRM_VC4_CREATE_BO                         0x03
+#define DRM_VC4_MMAP_BO                           0x04
+
+#define DRM_IOCTL_VC4_CREATE_BO           DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_CREATE_BO, struct drm_vc4_create_bo)
+#define DRM_IOCTL_VC4_MMAP_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_MMAP_BO, struct drm_vc4_mmap_bo)
+
+/**
+ * struct drm_vc4_create_bo - ioctl argument for creating VC4 BOs.
+ *
+ * There are currently no values for the flags argument, but it may be
+ * used in a future extension.
+ */
+struct drm_vc4_create_bo {
+	uint32_t size;
+	uint32_t flags;
+	/** Returned GEM handle for the BO. */
+	uint32_t handle;
+	uint32_t pad;
+};
+
+/**
+ * struct drm_vc4_mmap_bo - ioctl argument for mapping VC4 BOs.
+ *
+ * This doesn't actually perform an mmap.  Instead, it returns the
+ * offset you need to use in an mmap on the DRM device node.  This
+ * means that tools like valgrind end up knowing about the mapped
+ * memory.
+ *
+ * There are currently no values for the flags argument, but it may be
+ * used in a future extension.
+ */
+struct drm_vc4_mmap_bo {
+	/** Handle for the object being mapped. */
+	uint32_t handle;
+	uint32_t flags;
+	/** offset into the drm node to use for subsequent mmap call. */
+	uint64_t offset;
+};
+
+#endif /* _UAPI_VC4_DRM_H_ */
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1283854 — Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls.

Hi Eric,

A couple of suggestions/requests on the UAPI header side

On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:

> --- /dev/null
> +++ b/include/uapi/drm/vc4_drm.h

> +#include <drm/drm.h>
> +
Can we make this a "drm.h" ?


> +struct drm_vc4_create_bo {
> +       uint32_t size;
> +       uint32_t flags;
> +       /** Returned GEM handle for the BO. */
> +       uint32_t handle;
> +       uint32_t pad;
... and not use the stdint.h types but __{s,u}* through the file.

I'd rather than not nag why we want/need those, but if you prefer I
can repeat myself/others.

Thanks
Emil
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1286137 — Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls.

[Multipart message — attachments visible in raw view] — view raw

Emil Velikov <emil.l.velikov@gmail.com> writes:

> Hi Eric,
>
> A couple of suggestions/requests on the UAPI header side
>
> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>
>> --- /dev/null
>> +++ b/include/uapi/drm/vc4_drm.h
>
>> +#include <drm/drm.h>
>> +
> Can we make this a "drm.h" ?

Nope.

include/uapi/drm/vc4_drm.h:27:17: fatal error: drm.h: No such file or directory

and none of the drivers do, either.

>> +struct drm_vc4_create_bo {
>> +       uint32_t size;
>> +       uint32_t flags;
>> +       /** Returned GEM handle for the BO. */
>> +       uint32_t handle;
>> +       uint32_t pad;
> ... and not use the stdint.h types but __{s,u}* through the file.
>
> I'd rather than not nag why we want/need those, but if you prefer I
> can repeat myself/others.

I've switched to using the special types.

[toc] | [prev] | [next] | [standalone]

#1286173 — Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls.

On 08.12.2015 10:16, Eric Anholt wrote:
> Emil Velikov <emil.l.velikov@gmail.com> writes:
> 
>> Hi Eric,
>>
>> A couple of suggestions/requests on the UAPI header side
>>
>> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>>
>>> --- /dev/null
>>> +++ b/include/uapi/drm/vc4_drm.h
>>
>>> +#include <drm/drm.h>
>>> +
>> Can we make this a "drm.h" ?
> 
> Nope.
> 
> include/uapi/drm/vc4_drm.h:27:17: fatal error: drm.h: No such file or directory

What happened to include/uapi/drm/drm.h in that tree?


> and none of the drivers do, either.

daenzer@kaveri|11:55:31> grep '#include "drm.h"' include/uapi/drm/*
include/uapi/drm/amdgpu_drm.h:#include "drm.h"
include/uapi/drm/radeon_drm.h:#include "drm.h"

Patches to convert the rest are pending.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1286194 — Re: [PATCH 3/9] drm/vc4: Add create and map BO ioctls.

[Multipart message — attachments visible in raw view] — view raw

Michel Dänzer <michel@daenzer.net> writes:

> On 08.12.2015 10:16, Eric Anholt wrote:
>> Emil Velikov <emil.l.velikov@gmail.com> writes:
>> 
>>> Hi Eric,
>>>
>>> A couple of suggestions/requests on the UAPI header side
>>>
>>> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>>>
>>>> --- /dev/null
>>>> +++ b/include/uapi/drm/vc4_drm.h
>>>
>>>> +#include <drm/drm.h>
>>>> +
>>> Can we make this a "drm.h" ?
>> 
>> Nope.
>> 
>> include/uapi/drm/vc4_drm.h:27:17: fatal error: drm.h: No such file or directory
>
> What happened to include/uapi/drm/drm.h in that tree?

Looks like it's just <drm.h> versus "drm.h" failure.  I've changed over
to "drm.h"

[toc] | [prev] | [next] | [standalone]

#1281201 — [PATCH 2/9] drm/vc4: Add a BO cache.

We need to allocate new BOs in the kernel as part of each frame, but
the CMA allocator is way too slow for that.  As an optimization, keep
track of recently-freed BOs and reuse them, with a 1 second timeout to
fully free them back to the system.

This improves 3D performance by about 15%.

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/vc4/vc4_bo.c      | 336 +++++++++++++++++++++++++++++++++++++-
 drivers/gpu/drm/vc4/vc4_debugfs.c |   1 +
 drivers/gpu/drm/vc4/vc4_drv.c     |   6 +-
 drivers/gpu/drm/vc4/vc4_drv.h     |  49 +++++-
 4 files changed, 384 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/vc4/vc4_bo.c b/drivers/gpu/drm/vc4/vc4_bo.c
index ab9f510..18faa5b 100644
--- a/drivers/gpu/drm/vc4/vc4_bo.c
+++ b/drivers/gpu/drm/vc4/vc4_bo.c
@@ -12,19 +12,229 @@
  * access to system memory with no MMU in between.  To support it, we
  * use the GEM CMA helper functions to allocate contiguous ranges of
  * physical memory for our BOs.
+ *
+ * Since the CMA allocator is very slow, we keep a cache of recently
+ * freed BOs around so that the kernel's allocation of objects for 3D
+ * rendering can return quickly.
  */
 
 #include "vc4_drv.h"
 
-struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size)
+static void vc4_bo_stats_dump(struct vc4_dev *vc4)
+{
+	DRM_INFO("num bos allocated: %d\n",
+		 vc4->bo_stats.num_allocated);
+	DRM_INFO("size bos allocated: %dkb\n",
+		 vc4->bo_stats.size_allocated / 1024);
+	DRM_INFO("num bos used: %d\n",
+		 vc4->bo_stats.num_allocated - vc4->bo_stats.num_cached);
+	DRM_INFO("size bos used: %dkb\n",
+		 (vc4->bo_stats.size_allocated -
+		  vc4->bo_stats.size_cached) / 1024);
+	DRM_INFO("num bos cached: %d\n",
+		 vc4->bo_stats.num_cached);
+	DRM_INFO("size bos cached: %dkb\n",
+		 vc4->bo_stats.size_cached / 1024);
+}
+
+#ifdef CONFIG_DEBUG_FS
+int vc4_bo_stats_debugfs(struct seq_file *m, void *unused)
+{
+	struct drm_info_node *node = (struct drm_info_node *)m->private;
+	struct drm_device *dev = node->minor->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_bo_stats stats;
+
+	/* Take a snapshot of the current stats with the lock held. */
+	mutex_lock(&vc4->bo_lock);
+	stats = vc4->bo_stats;
+	mutex_unlock(&vc4->bo_lock);
+
+	seq_printf(m, "num bos allocated: %d\n",
+		   stats.num_allocated);
+	seq_printf(m, "size bos allocated: %dkb\n",
+		   stats.size_allocated / 1024);
+	seq_printf(m, "num bos used: %d\n",
+		   stats.num_allocated - stats.num_cached);
+	seq_printf(m, "size bos used: %dkb\n",
+		   (stats.size_allocated - stats.size_cached) / 1024);
+	seq_printf(m, "num bos cached: %d\n",
+		   stats.num_cached);
+	seq_printf(m, "size bos cached: %dkb\n",
+		   stats.size_cached / 1024);
+
+	return 0;
+}
+#endif
+
+static uint32_t bo_page_index(size_t size)
+{
+	return (size / PAGE_SIZE) - 1;
+}
+
+/* Must be called with bo_lock held. */
+static void vc4_bo_destroy(struct vc4_bo *bo)
 {
+	struct drm_gem_object *obj = &bo->base.base;
+	struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
+
+	vc4->bo_stats.num_allocated--;
+	vc4->bo_stats.size_allocated -= obj->size;
+	drm_gem_cma_free_object(obj);
+}
+
+/* Must be called with bo_lock held. */
+static void vc4_bo_remove_from_cache(struct vc4_bo *bo)
+{
+	struct drm_gem_object *obj = &bo->base.base;
+	struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
+
+	vc4->bo_stats.num_cached--;
+	vc4->bo_stats.size_cached -= obj->size;
+
+	list_del(&bo->unref_head);
+	list_del(&bo->size_head);
+}
+
+static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev,
+						     size_t size)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	uint32_t page_index = bo_page_index(size);
+
+	if (vc4->bo_cache.size_list_size <= page_index) {
+		uint32_t new_size = max(vc4->bo_cache.size_list_size * 2,
+					page_index + 1);
+		struct list_head *new_list;
+		uint32_t i;
+
+		new_list = kmalloc_array(new_size, sizeof(struct list_head),
+					 GFP_KERNEL);
+		if (!new_list)
+			return NULL;
+
+		/* Rebase the old cached BO lists to their new list
+		 * head locations.
+		 */
+		for (i = 0; i < vc4->bo_cache.size_list_size; i++) {
+			struct list_head *old_list =
+				&vc4->bo_cache.size_list[i];
+
+			if (list_empty(old_list))
+				INIT_LIST_HEAD(&new_list[i]);
+			else
+				list_replace(old_list, &new_list[i]);
+		}
+		/* And initialize the brand new BO list heads. */
+		for (i = vc4->bo_cache.size_list_size; i < new_size; i++)
+			INIT_LIST_HEAD(&new_list[i]);
+
+		kfree(vc4->bo_cache.size_list);
+		vc4->bo_cache.size_list = new_list;
+		vc4->bo_cache.size_list_size = new_size;
+	}
+
+	return &vc4->bo_cache.size_list[page_index];
+}
+
+void vc4_bo_cache_purge(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	mutex_lock(&vc4->bo_lock);
+	while (!list_empty(&vc4->bo_cache.time_list)) {
+		struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
+						    struct vc4_bo, unref_head);
+		vc4_bo_remove_from_cache(bo);
+		vc4_bo_destroy(bo);
+	}
+	mutex_unlock(&vc4->bo_lock);
+}
+
+static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
+					    uint32_t size)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	uint32_t page_index = bo_page_index(size);
+	struct vc4_bo *bo = NULL;
+
+	size = roundup(size, PAGE_SIZE);
+
+	mutex_lock(&vc4->bo_lock);
+	if (page_index >= vc4->bo_cache.size_list_size)
+		goto out;
+
+	if (list_empty(&vc4->bo_cache.size_list[page_index]))
+		goto out;
+
+	bo = list_first_entry(&vc4->bo_cache.size_list[page_index],
+			      struct vc4_bo, size_head);
+	vc4_bo_remove_from_cache(bo);
+	kref_init(&bo->base.base.refcount);
+
+out:
+	mutex_unlock(&vc4->bo_lock);
+	return bo;
+}
+
+/**
+ * vc4_gem_create_object - Implementation of driver->gem_create_object.
+ *
+ * This lets the CMA helpers allocate object structs for us, and keep
+ * our BO stats correct.
+ */
+struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_bo *bo;
+
+	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
+	if (!bo)
+		return ERR_PTR(-ENOMEM);
+
+	mutex_lock(&vc4->bo_lock);
+	vc4->bo_stats.num_allocated++;
+	vc4->bo_stats.size_allocated += size;
+	mutex_unlock(&vc4->bo_lock);
+
+	return &bo->base.base;
+}
+
+struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
+			     bool from_cache)
+{
+	size_t size = roundup(unaligned_size, PAGE_SIZE);
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	struct drm_gem_cma_object *cma_obj;
 
-	cma_obj = drm_gem_cma_create(dev, size);
-	if (IS_ERR(cma_obj))
+	if (size == 0)
 		return NULL;
-	else
-		return to_vc4_bo(&cma_obj->base);
+
+	/* First, try to get a vc4_bo from the kernel BO cache. */
+	if (from_cache) {
+		struct vc4_bo *bo = vc4_bo_get_from_cache(dev, size);
+
+		if (bo)
+			return bo;
+	}
+
+	cma_obj = drm_gem_cma_create(dev, size);
+	if (IS_ERR(cma_obj)) {
+		/*
+		 * If we've run out of CMA memory, kill the cache of
+		 * CMA allocations we've got laying around and try again.
+		 */
+		vc4_bo_cache_purge(dev);
+
+		cma_obj = drm_gem_cma_create(dev, size);
+		if (IS_ERR(cma_obj)) {
+			DRM_ERROR("Failed to allocate from CMA:\n");
+			vc4_bo_stats_dump(vc4);
+			return NULL;
+		}
+	}
+
+	return to_vc4_bo(&cma_obj->base);
 }
 
 int vc4_dumb_create(struct drm_file *file_priv,
@@ -41,7 +251,7 @@ int vc4_dumb_create(struct drm_file *file_priv,
 	if (args->size < args->pitch * args->height)
 		args->size = args->pitch * args->height;
 
-	bo = vc4_bo_create(dev, roundup(args->size, PAGE_SIZE));
+	bo = vc4_bo_create(dev, args->size, false);
 	if (!bo)
 		return -ENOMEM;
 
@@ -50,3 +260,117 @@ int vc4_dumb_create(struct drm_file *file_priv,
 
 	return ret;
 }
+
+/* Must be called with bo_lock held. */
+static void vc4_bo_cache_free_old(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	unsigned long expire_time = jiffies - msecs_to_jiffies(1000);
+
+	while (!list_empty(&vc4->bo_cache.time_list)) {
+		struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
+						    struct vc4_bo, unref_head);
+		if (time_before(expire_time, bo->free_time)) {
+			mod_timer(&vc4->bo_cache.time_timer,
+				  round_jiffies_up(jiffies +
+						   msecs_to_jiffies(1000)));
+			return;
+		}
+
+		vc4_bo_remove_from_cache(bo);
+		vc4_bo_destroy(bo);
+	}
+}
+
+/* Called on the last userspace/kernel unreference of the BO.  Returns
+ * it to the BO cache if possible, otherwise frees it.
+ *
+ * Note that this is called with the struct_mutex held.
+ */
+void vc4_free_object(struct drm_gem_object *gem_bo)
+{
+	struct drm_device *dev = gem_bo->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_bo *bo = to_vc4_bo(gem_bo);
+	struct list_head *cache_list;
+
+	mutex_lock(&vc4->bo_lock);
+	/* If the object references someone else's memory, we can't cache it.
+	 */
+	if (gem_bo->import_attach) {
+		vc4_bo_destroy(bo);
+		goto out;
+	}
+
+	/* Don't cache if it was publicly named. */
+	if (gem_bo->name) {
+		vc4_bo_destroy(bo);
+		goto out;
+	}
+
+	cache_list = vc4_get_cache_list_for_size(dev, gem_bo->size);
+	if (!cache_list) {
+		vc4_bo_destroy(bo);
+		goto out;
+	}
+
+	bo->free_time = jiffies;
+	list_add(&bo->size_head, cache_list);
+	list_add(&bo->unref_head, &vc4->bo_cache.time_list);
+
+	vc4->bo_stats.num_cached++;
+	vc4->bo_stats.size_cached += gem_bo->size;
+
+	vc4_bo_cache_free_old(dev);
+
+out:
+	mutex_unlock(&vc4->bo_lock);
+}
+
+static void vc4_bo_cache_time_work(struct work_struct *work)
+{
+	struct vc4_dev *vc4 =
+		container_of(work, struct vc4_dev, bo_cache.time_work);
+	struct drm_device *dev = vc4->dev;
+
+	mutex_lock(&vc4->bo_lock);
+	vc4_bo_cache_free_old(dev);
+	mutex_unlock(&vc4->bo_lock);
+}
+
+static void vc4_bo_cache_time_timer(unsigned long data)
+{
+	struct drm_device *dev = (struct drm_device *)data;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	schedule_work(&vc4->bo_cache.time_work);
+}
+
+void vc4_bo_cache_init(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	mutex_init(&vc4->bo_lock);
+
+	INIT_LIST_HEAD(&vc4->bo_cache.time_list);
+
+	INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work);
+	setup_timer(&vc4->bo_cache.time_timer,
+		    vc4_bo_cache_time_timer,
+		    (unsigned long)dev);
+}
+
+void vc4_bo_cache_destroy(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	del_timer(&vc4->bo_cache.time_timer);
+	cancel_work_sync(&vc4->bo_cache.time_work);
+
+	vc4_bo_cache_purge(dev);
+
+	if (vc4->bo_stats.num_allocated) {
+		DRM_ERROR("Destroying BO cache while BOs still allocated:\n");
+		vc4_bo_stats_dump(vc4);
+	}
+}
diff --git a/drivers/gpu/drm/vc4/vc4_debugfs.c b/drivers/gpu/drm/vc4/vc4_debugfs.c
index 4297b0a5..6bcf96e 100644
--- a/drivers/gpu/drm/vc4/vc4_debugfs.c
+++ b/drivers/gpu/drm/vc4/vc4_debugfs.c
@@ -16,6 +16,7 @@
 #include "vc4_regs.h"
 
 static const struct drm_info_list vc4_debugfs_list[] = {
+	{"bo_stats", vc4_bo_stats_debugfs, 0},
 	{"hdmi_regs", vc4_hdmi_debugfs_regs, 0},
 	{"hvs_regs", vc4_hvs_debugfs_regs, 0},
 	{"crtc0_regs", vc4_crtc_debugfs_regs, 0, (void *)(uintptr_t)0},
diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index 6e73060..da041fa 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -92,7 +92,8 @@ static struct drm_driver vc4_drm_driver = {
 	.debugfs_cleanup = vc4_debugfs_cleanup,
 #endif
 
-	.gem_free_object = drm_gem_cma_free_object,
+	.gem_create_object = vc4_create_object,
+	.gem_free_object = vc4_free_object,
 	.gem_vm_ops = &drm_gem_cma_vm_ops,
 
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
@@ -170,6 +171,8 @@ static int vc4_drm_bind(struct device *dev)
 
 	drm_dev_set_unique(drm, dev_name(dev));
 
+	vc4_bo_cache_init(drm);
+
 	drm_mode_config_init(drm);
 	if (ret)
 		goto unref;
@@ -202,6 +205,7 @@ unbind_all:
 	component_unbind_all(dev, drm);
 unref:
 	drm_dev_unref(drm);
+	vc4_bo_cache_destroy(drm);
 	return ret;
 }
 
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index fd8319f..39a1ff5 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -17,6 +17,37 @@ struct vc4_dev {
 	struct vc4_crtc *crtc[3];
 
 	struct drm_fbdev_cma *fbdev;
+
+	/* The kernel-space BO cache.  Tracks buffers that have been
+	 * unreferenced by all other users (refcounts of 0!) but not
+	 * yet freed, so we can do cheap allocations.
+	 */
+	struct vc4_bo_cache {
+		/* Array of list heads for entries in the BO cache,
+		 * based on number of pages, so we can do O(1) lookups
+		 * in the cache when allocating.
+		 */
+		struct list_head *size_list;
+		uint32_t size_list_size;
+
+		/* List of all BOs in the cache, ordered by age, so we
+		 * can do O(1) lookups when trying to free old
+		 * buffers.
+		 */
+		struct list_head time_list;
+		struct work_struct time_work;
+		struct timer_list time_timer;
+	} bo_cache;
+
+	struct vc4_bo_stats {
+		u32 num_allocated;
+		u32 size_allocated;
+		u32 num_cached;
+		u32 size_cached;
+	} bo_stats;
+
+	/* Protects bo_cache and the BO stats. */
+	struct mutex bo_lock;
 };
 
 static inline struct vc4_dev *
@@ -27,6 +58,17 @@ to_vc4_dev(struct drm_device *dev)
 
 struct vc4_bo {
 	struct drm_gem_cma_object base;
+
+	/* List entry for the BO's position in either
+	 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
+	 */
+	struct list_head unref_head;
+
+	/* Time in jiffies when the BO was put in vc4->bo_cache. */
+	unsigned long free_time;
+
+	/* List entry for the BO's position in vc4_dev->bo_cache.size_list */
+	struct list_head size_head;
 };
 
 static inline struct vc4_bo *
@@ -104,13 +146,18 @@ to_vc4_encoder(struct drm_encoder *encoder)
 #define wait_for(COND, MS) _wait_for(COND, MS, 1)
 
 /* vc4_bo.c */
+struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
 void vc4_free_object(struct drm_gem_object *gem_obj);
-struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size);
+struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
+			     bool from_cache);
 int vc4_dumb_create(struct drm_file *file_priv,
 		    struct drm_device *dev,
 		    struct drm_mode_create_dumb *args);
 struct dma_buf *vc4_prime_export(struct drm_device *dev,
 				 struct drm_gem_object *obj, int flags);
+void vc4_bo_cache_init(struct drm_device *dev);
+void vc4_bo_cache_destroy(struct drm_device *dev);
+int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
 
 /* vc4_crtc.c */
 extern struct platform_driver vc4_crtc_driver;
-- 
2.6.2

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1281532 — Re: [PATCH 1/9] drm: Create a driver hook for allocating GEM object structs.

On Tue, Dec 01, 2015 at 12:35:50PM -0800, Eric Anholt wrote:
> The CMA helpers had no way for a driver to extend the struct with its
> own fields.  Since the CMA helpers are mostly "Allocate a
> drm_gem_cma_object, then fill in a few fields", it's hard to write as
> pure helpers without passing in a driver callback for the allocate
> step.
> 
> Signed-off-by: Eric Anholt <eric@anholt.net>

Yeah there's no other way to make this happen without ridiculous amounts
of boilerplate.

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

> ---
>  drivers/gpu/drm/drm_gem_cma_helper.c | 10 ++++++----
>  include/drm/drmP.h                   |  7 +++++++
>  2 files changed, 13 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/drm_gem_cma_helper.c b/drivers/gpu/drm/drm_gem_cma_helper.c
> index e109b49..0f7b00ba 100644
> --- a/drivers/gpu/drm/drm_gem_cma_helper.c
> +++ b/drivers/gpu/drm/drm_gem_cma_helper.c
> @@ -59,11 +59,13 @@ __drm_gem_cma_create(struct drm_device *drm, size_t size)
>  	struct drm_gem_object *gem_obj;
>  	int ret;
>  
> -	cma_obj = kzalloc(sizeof(*cma_obj), GFP_KERNEL);
> -	if (!cma_obj)
> +	if (drm->driver->gem_create_object)
> +		gem_obj = drm->driver->gem_create_object(drm, size);
> +	else
> +		gem_obj = kzalloc(sizeof(*cma_obj), GFP_KERNEL);
> +	if (!gem_obj)
>  		return ERR_PTR(-ENOMEM);
> -
> -	gem_obj = &cma_obj->base;
> +	cma_obj = container_of(gem_obj, struct drm_gem_cma_object, base);
>  
>  	ret = drm_gem_object_init(drm, gem_obj, size);
>  	if (ret)
> diff --git a/include/drm/drmP.h b/include/drm/drmP.h
> index 0b921ae..22ff162 100644
> --- a/include/drm/drmP.h
> +++ b/include/drm/drmP.h
> @@ -580,6 +580,13 @@ struct drm_driver {
>  	int (*gem_open_object) (struct drm_gem_object *, struct drm_file *);
>  	void (*gem_close_object) (struct drm_gem_object *, struct drm_file *);
>  
> +	/**
> +	 * Hook for allocating the GEM object struct, for use by core
> +	 * helpers.
> +	 */
> +	struct drm_gem_object *(*gem_create_object)(struct drm_device *dev,
> +						    size_t size);
> +
>  	/* prime: */
>  	/* export handle -> fd (see drm_gem_prime_handle_to_fd() helper) */
>  	int (*prime_handle_to_fd)(struct drm_device *dev, struct drm_file *file_priv,
> -- 
> 2.6.2
> 
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1281653 — Re: [PATCH 7/9] drm/vc4: Add support for drawing 3D frames.

Hi Eric,

On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
> The user submission is basically a pointer to a command list and a
> pointer to uniforms.  We copy those in to the kernel, validate and
> relocate them, and store the result in a GPU BO which we queue for
> execution.
>
> Signed-off-by: Eric Anholt <eric@anholt.net>
> ---
>  drivers/gpu/drm/vc4/Makefile           |   7 +
>  drivers/gpu/drm/vc4/vc4_drv.c          |  15 +-
>  drivers/gpu/drm/vc4/vc4_drv.h          | 192 +++++++
>  drivers/gpu/drm/vc4/vc4_gem.c          | 640 ++++++++++++++++++++++
>  drivers/gpu/drm/vc4/vc4_irq.c          | 210 ++++++++
>  drivers/gpu/drm/vc4/vc4_packet.h       | 399 ++++++++++++++
>  drivers/gpu/drm/vc4/vc4_render_cl.c    | 631 ++++++++++++++++++++++
>  drivers/gpu/drm/vc4/vc4_trace.h        |  63 +++
>  drivers/gpu/drm/vc4/vc4_trace_points.c |  14 +
>  drivers/gpu/drm/vc4/vc4_v3d.c          |  37 ++
>  drivers/gpu/drm/vc4/vc4_validate.c     | 955 +++++++++++++++++++++++++++++++++
>  include/uapi/drm/vc4_drm.h             | 141 +++++
>  12 files changed, 3303 insertions(+), 1 deletion(-)

Am I assuming correct that the above are the exact same ones copied in
mesa (and used only then simulator is used) ? In the short term
keeping the two in sync would be fine, although wondering if you have
plans to reshuffle things in the long term ?

Thanks
Emil
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]

#1282287 — Re: [PATCH 7/9] drm/vc4: Add support for drawing 3D frames.

[Multipart message — attachments visible in raw view] — view raw

Emil Velikov <emil.l.velikov@gmail.com> writes:

> Hi Eric,
>
> On 1 December 2015 at 20:35, Eric Anholt <eric@anholt.net> wrote:
>> The user submission is basically a pointer to a command list and a
>> pointer to uniforms.  We copy those in to the kernel, validate and
>> relocate them, and store the result in a GPU BO which we queue for
>> execution.
>>
>> Signed-off-by: Eric Anholt <eric@anholt.net>
>> ---
>>  drivers/gpu/drm/vc4/Makefile           |   7 +
>>  drivers/gpu/drm/vc4/vc4_drv.c          |  15 +-
>>  drivers/gpu/drm/vc4/vc4_drv.h          | 192 +++++++
>>  drivers/gpu/drm/vc4/vc4_gem.c          | 640 ++++++++++++++++++++++
>>  drivers/gpu/drm/vc4/vc4_irq.c          | 210 ++++++++
>>  drivers/gpu/drm/vc4/vc4_packet.h       | 399 ++++++++++++++
>>  drivers/gpu/drm/vc4/vc4_render_cl.c    | 631 ++++++++++++++++++++++
>>  drivers/gpu/drm/vc4/vc4_trace.h        |  63 +++
>>  drivers/gpu/drm/vc4/vc4_trace_points.c |  14 +
>>  drivers/gpu/drm/vc4/vc4_v3d.c          |  37 ++
>>  drivers/gpu/drm/vc4/vc4_validate.c     | 955 +++++++++++++++++++++++++++++++++
>>  include/uapi/drm/vc4_drm.h             | 141 +++++
>>  12 files changed, 3303 insertions(+), 1 deletion(-)
>
> Am I assuming correct that the above are the exact same ones copied in
> mesa (and used only then simulator is used) ? In the short term
> keeping the two in sync would be fine, although wondering if you have
> plans to reshuffle things in the long term ?

This code is not exactly synced with Mesa, but I do sync between them on
a regular basis.  The in-Mesa copy is secondary, and just a way for me
to get testing done faster (in the simulation environment with gdb) than
I can in the kernel with printfs.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web