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.

v2: Use __u32-style types.

Signed-off-by: Eric Anholt <eric@anholt.net>

6 files changed, 974 insertions, 5 deletions

diff --git a/drivers/gpu/drm/vc4/Makefile b/drivers/gpu/drm/vc4/Makefile
index 32b4f9cd8f52..eb776a6a9899 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 06cba268e17a..18dfe3ec9a62 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 5fa468864ec8..da4be9c8f70d 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 fddb0a06ed3a..bd77d5574e30 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 000000000000..d5c2f3c85ebb
--- /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 000000000000..f67124b4c534
--- /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;
+}