drm/vc4: Add support for drawing 3D frames.
authorEric Anholt <eric@anholt.net>
Mon, 30 Nov 2015 20:13:37 +0000 (12:13 -0800)
committerEric Anholt <eric@anholt.net>
Tue, 8 Dec 2015 04:05:10 +0000 (20:05 -0800)
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.

v2: Drop support for NV shader recs (not necessary for GL), simplify
    vc4_use_bo(), improve bin flush/semaphore checks, use __u32 style
    types.

Signed-off-by: Eric Anholt <eric@anholt.net>
12 files changed:
drivers/gpu/drm/vc4/Makefile
drivers/gpu/drm/vc4/vc4_drv.c
drivers/gpu/drm/vc4/vc4_drv.h
drivers/gpu/drm/vc4/vc4_gem.c [new file with mode: 0644]
drivers/gpu/drm/vc4/vc4_irq.c [new file with mode: 0644]
drivers/gpu/drm/vc4/vc4_packet.h [new file with mode: 0644]
drivers/gpu/drm/vc4/vc4_render_cl.c [new file with mode: 0644]
drivers/gpu/drm/vc4/vc4_trace.h [new file with mode: 0644]
drivers/gpu/drm/vc4/vc4_trace_points.c [new file with mode: 0644]
drivers/gpu/drm/vc4/vc4_v3d.c
drivers/gpu/drm/vc4/vc4_validate.c [new file with mode: 0644]
include/uapi/drm/vc4_drm.h

index e87a6f2f59162291c352107b6a2eea9c9a0fca8f..4c6a99f0398c03ce5d83dea1ec3f7906a817bf89 100644 (file)
@@ -8,12 +8,19 @@ vc4-y := \
        vc4_crtc.o \
        vc4_drv.o \
        vc4_kms.o \
+       vc4_gem.o \
        vc4_hdmi.o \
        vc4_hvs.o \
+       vc4_irq.o \
        vc4_plane.o \
+       vc4_render_cl.o \
+       vc4_trace_points.o \
        vc4_v3d.o \
+       vc4_validate.o \
        vc4_validate_shaders.o
 
 vc4-$(CONFIG_DEBUG_FS) += vc4_debugfs.o
 
 obj-$(CONFIG_DRM_VC4)  += vc4.o
+
+CFLAGS_vc4_trace_points.o := -I$(src)
index db58d74efe95de8cd0c6fae110182e8342d1d514..2cfee595945590e020945f11724e265135ced2d0 100644 (file)
@@ -74,6 +74,9 @@ static const struct file_operations vc4_drm_fops = {
 };
 
 static const struct drm_ioctl_desc vc4_drm_ioctls[] = {
+       DRM_IOCTL_DEF_DRV(VC4_SUBMIT_CL, vc4_submit_cl_ioctl, 0),
+       DRM_IOCTL_DEF_DRV(VC4_WAIT_SEQNO, vc4_wait_seqno_ioctl, 0),
+       DRM_IOCTL_DEF_DRV(VC4_WAIT_BO, vc4_wait_bo_ioctl, 0),
        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),
@@ -83,10 +86,16 @@ static struct drm_driver vc4_drm_driver = {
        .driver_features = (DRIVER_MODESET |
                            DRIVER_ATOMIC |
                            DRIVER_GEM |
+                           DRIVER_HAVE_IRQ |
                            DRIVER_PRIME),
        .lastclose = vc4_lastclose,
        .preclose = vc4_drm_preclose,
 
+       .irq_handler = vc4_irq,
+       .irq_preinstall = vc4_irq_preinstall,
+       .irq_postinstall = vc4_irq_postinstall,
+       .irq_uninstall = vc4_irq_uninstall,
+
        .enable_vblank = vc4_enable_vblank,
        .disable_vblank = vc4_disable_vblank,
        .get_vblank_counter = drm_vblank_count,
@@ -181,9 +190,11 @@ static int vc4_drm_bind(struct device *dev)
        if (ret)
                goto unref;
 
+       vc4_gem_init(drm);
+
        ret = component_bind_all(dev, drm);
        if (ret)
-               goto unref;
+               goto gem_destroy;
 
        ret = drm_dev_register(drm, 0);
        if (ret < 0)
@@ -207,6 +218,8 @@ unregister:
        drm_dev_unregister(drm);
 unbind_all:
        component_unbind_all(dev, drm);
+gem_destroy:
+       vc4_gem_destroy(drm);
 unref:
        drm_dev_unref(drm);
        vc4_bo_cache_destroy(drm);
index 8945463e70b648da480a334712e68b44b32949d7..0bc8c57196acae69984ed09e716f8d24b03fef23 100644 (file)
@@ -49,6 +49,48 @@ struct vc4_dev {
 
        /* Protects bo_cache and the BO stats. */
        struct mutex bo_lock;
+
+       /* Sequence number for the last job queued in job_list.
+        * Starts at 0 (no jobs emitted).
+        */
+       uint64_t emit_seqno;
+
+       /* Sequence number for the last completed job on the GPU.
+        * Starts at 0 (no jobs completed).
+        */
+       uint64_t finished_seqno;
+
+       /* List of all struct vc4_exec_info for jobs to be executed.
+        * The first job in the list is the one currently programmed
+        * into ct0ca/ct1ca for execution.
+        */
+       struct list_head job_list;
+       /* List of the finished vc4_exec_infos waiting to be freed by
+        * job_done_work.
+        */
+       struct list_head job_done_list;
+       /* Spinlock used to synchronize the job_list and seqno
+        * accesses between the IRQ handler and GEM ioctls.
+        */
+       spinlock_t job_lock;
+       wait_queue_head_t job_wait_queue;
+       struct work_struct job_done_work;
+
+       /* 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
+        * vc4->current_exec's free list.
+        */
+       struct vc4_bo *overflow_mem;
+       struct work_struct overflow_mem_work;
+
+       struct {
+               uint32_t last_ct0ca, last_ct1ca;
+               struct timer_list timer;
+               struct work_struct reset_work;
+       } hangcheck;
+
+       struct semaphore async_modeset;
 };
 
 static inline struct vc4_dev *
@@ -60,6 +102,9 @@ to_vc4_dev(struct drm_device *dev)
 struct vc4_bo {
        struct drm_gem_cma_object base;
 
+       /* seqno of the last job to render to this BO. */
+       uint64_t seqno;
+
        /* List entry for the BO's position in either
         * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
         */
@@ -130,6 +175,101 @@ to_vc4_encoder(struct drm_encoder *encoder)
 #define HVS_READ(offset) readl(vc4->hvs->regs + offset)
 #define HVS_WRITE(offset, val) writel(val, vc4->hvs->regs + offset)
 
+struct vc4_exec_info {
+       /* Sequence number for this bin/render job. */
+       uint64_t seqno;
+
+       /* Kernel-space copy of the ioctl arguments */
+       struct drm_vc4_submit_cl *args;
+
+       /* This is the array of BOs that were looked up at the start of exec.
+        * Command validation will use indices into this array.
+        */
+       struct drm_gem_cma_object **bo;
+       uint32_t bo_count;
+
+       /* Pointers for our position in vc4->job_list */
+       struct list_head head;
+
+       /* List of other BOs used in the job that need to be released
+        * once the job is complete.
+        */
+       struct list_head unref_list;
+
+       /* Current unvalidated indices into @bo loaded by the non-hardware
+        * VC4_PACKET_GEM_HANDLES.
+        */
+       uint32_t bo_index[2];
+
+       /* This is the BO where we store the validated command lists, shader
+        * records, and uniforms.
+        */
+       struct drm_gem_cma_object *exec_bo;
+
+       /**
+        * This tracks the per-shader-record state (packet 64) that
+        * determines the length of the shader record and the offset
+        * it's expected to be found at.  It gets read in from the
+        * command lists.
+        */
+       struct vc4_shader_state {
+               uint32_t addr;
+               /* Maximum vertex index referenced by any primitive using this
+                * shader state.
+                */
+               uint32_t max_index;
+       } *shader_state;
+
+       /** How many shader states the user declared they were using. */
+       uint32_t shader_state_size;
+       /** How many shader state records the validator has seen. */
+       uint32_t shader_state_count;
+
+       bool found_tile_binning_mode_config_packet;
+       bool found_start_tile_binning_packet;
+       bool found_increment_semaphore_packet;
+       bool found_flush;
+       uint8_t bin_tiles_x, bin_tiles_y;
+       struct drm_gem_cma_object *tile_bo;
+       uint32_t tile_alloc_offset;
+
+       /**
+        * Computed addresses pointing into exec_bo where we start the
+        * bin thread (ct0) and render thread (ct1).
+        */
+       uint32_t ct0ca, ct0ea;
+       uint32_t ct1ca, ct1ea;
+
+       /* Pointer to the unvalidated bin CL (if present). */
+       void *bin_u;
+
+       /* Pointers to the shader recs.  These paddr gets incremented as CL
+        * packets are relocated in validate_gl_shader_state, and the vaddrs
+        * (u and v) get incremented and size decremented as the shader recs
+        * themselves are validated.
+        */
+       void *shader_rec_u;
+       void *shader_rec_v;
+       uint32_t shader_rec_p;
+       uint32_t shader_rec_size;
+
+       /* Pointers to the uniform data.  These pointers are incremented, and
+        * size decremented, as each batch of uniforms is uploaded.
+        */
+       void *uniforms_u;
+       void *uniforms_v;
+       uint32_t uniforms_p;
+       uint32_t uniforms_size;
+};
+
+static inline struct vc4_exec_info *
+vc4_first_job(struct vc4_dev *vc4)
+{
+       if (list_empty(&vc4->job_list))
+               return NULL;
+       return list_first_entry(&vc4->job_list, struct vc4_exec_info, head);
+}
+
 /**
  * struct vc4_texture_sample_info - saves the offsets into the UBO for texture
  * setup parameters.
@@ -231,10 +371,31 @@ void vc4_debugfs_cleanup(struct drm_minor *minor);
 /* vc4_drv.c */
 void __iomem *vc4_ioremap_regs(struct platform_device *dev, int index);
 
+/* vc4_gem.c */
+void vc4_gem_init(struct drm_device *dev);
+void vc4_gem_destroy(struct drm_device *dev);
+int vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
+                       struct drm_file *file_priv);
+int vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
+                        struct drm_file *file_priv);
+int vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
+                     struct drm_file *file_priv);
+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);
+
 /* vc4_hdmi.c */
 extern struct platform_driver vc4_hdmi_driver;
 int vc4_hdmi_debugfs_regs(struct seq_file *m, void *unused);
 
+/* vc4_irq.c */
+irqreturn_t vc4_irq(int irq, void *arg);
+void vc4_irq_preinstall(struct drm_device *dev);
+int vc4_irq_postinstall(struct drm_device *dev);
+void vc4_irq_uninstall(struct drm_device *dev);
+void vc4_irq_reset(struct drm_device *dev);
+
 /* vc4_hvs.c */
 extern struct platform_driver vc4_hvs_driver;
 void vc4_hvs_dump_state(struct drm_device *dev);
@@ -253,6 +414,27 @@ u32 vc4_plane_dlist_size(struct drm_plane_state *state);
 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);
+int vc4_v3d_set_power(struct vc4_dev *vc4, bool on);
+
+/* vc4_validate.c */
+int
+vc4_validate_bin_cl(struct drm_device *dev,
+                   void *validated,
+                   void *unvalidated,
+                   struct vc4_exec_info *exec);
+
+int
+vc4_validate_shader_recs(struct drm_device *dev, struct vc4_exec_info *exec);
+
+struct drm_gem_cma_object *vc4_use_bo(struct vc4_exec_info *exec,
+                                     uint32_t hindex);
+
+int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec);
+
+bool vc4_check_tex_size(struct vc4_exec_info *exec,
+                       struct drm_gem_cma_object *fbo,
+                       uint32_t offset, uint8_t tiling_format,
+                       uint32_t width, uint32_t height, uint8_t cpp);
 
 /* vc4_validate_shader.c */
 struct vc4_validated_shader_info *
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
new file mode 100644 (file)
index 0000000..936dddf
--- /dev/null
@@ -0,0 +1,642 @@
+/*
+ * 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.
+ */
+
+#include <linux/module.h>
+#include <linux/platform_device.h>
+#include <linux/device.h>
+#include <linux/io.h>
+
+#include "uapi/drm/vc4_drm.h"
+#include "vc4_drv.h"
+#include "vc4_regs.h"
+#include "vc4_trace.h"
+
+static void
+vc4_queue_hangcheck(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       mod_timer(&vc4->hangcheck.timer,
+                 round_jiffies_up(jiffies + msecs_to_jiffies(100)));
+}
+
+static void
+vc4_reset(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       DRM_INFO("Resetting GPU.\n");
+       vc4_v3d_set_power(vc4, false);
+       vc4_v3d_set_power(vc4, true);
+
+       vc4_irq_reset(dev);
+
+       /* Rearm the hangcheck -- another job might have been waiting
+        * for our hung one to get kicked off, and vc4_irq_reset()
+        * would have started it.
+        */
+       vc4_queue_hangcheck(dev);
+}
+
+static void
+vc4_reset_work(struct work_struct *work)
+{
+       struct vc4_dev *vc4 =
+               container_of(work, struct vc4_dev, hangcheck.reset_work);
+
+       vc4_reset(vc4->dev);
+}
+
+static void
+vc4_hangcheck_elapsed(unsigned long data)
+{
+       struct drm_device *dev = (struct drm_device *)data;
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       uint32_t ct0ca, ct1ca;
+
+       /* If idle, we can stop watching for hangs. */
+       if (list_empty(&vc4->job_list))
+               return;
+
+       ct0ca = V3D_READ(V3D_CTNCA(0));
+       ct1ca = V3D_READ(V3D_CTNCA(1));
+
+       /* If we've made any progress in execution, rearm the timer
+        * and wait.
+        */
+       if (ct0ca != vc4->hangcheck.last_ct0ca ||
+           ct1ca != vc4->hangcheck.last_ct1ca) {
+               vc4->hangcheck.last_ct0ca = ct0ca;
+               vc4->hangcheck.last_ct1ca = ct1ca;
+               vc4_queue_hangcheck(dev);
+               return;
+       }
+
+       /* We've gone too long with no progress, reset.  This has to
+        * be done from a work struct, since resetting can sleep and
+        * this timer hook isn't allowed to.
+        */
+       schedule_work(&vc4->hangcheck.reset_work);
+}
+
+static void
+submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       /* Set the current and end address of the control list.
+        * Writing the end register is what starts the job.
+        */
+       V3D_WRITE(V3D_CTNCA(thread), start);
+       V3D_WRITE(V3D_CTNEA(thread), end);
+}
+
+int
+vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno, uint64_t timeout_ns,
+                  bool interruptible)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       int ret = 0;
+       unsigned long timeout_expire;
+       DEFINE_WAIT(wait);
+
+       if (vc4->finished_seqno >= seqno)
+               return 0;
+
+       if (timeout_ns == 0)
+               return -ETIME;
+
+       timeout_expire = jiffies + nsecs_to_jiffies(timeout_ns);
+
+       trace_vc4_wait_for_seqno_begin(dev, seqno, timeout_ns);
+       for (;;) {
+               prepare_to_wait(&vc4->job_wait_queue, &wait,
+                               interruptible ? TASK_INTERRUPTIBLE :
+                               TASK_UNINTERRUPTIBLE);
+
+               if (interruptible && signal_pending(current)) {
+                       ret = -ERESTARTSYS;
+                       break;
+               }
+
+               if (vc4->finished_seqno >= seqno)
+                       break;
+
+               if (timeout_ns != ~0ull) {
+                       if (time_after_eq(jiffies, timeout_expire)) {
+                               ret = -ETIME;
+                               break;
+                       }
+                       schedule_timeout(timeout_expire - jiffies);
+               } else {
+                       schedule();
+               }
+       }
+
+       finish_wait(&vc4->job_wait_queue, &wait);
+       trace_vc4_wait_for_seqno_end(dev, seqno);
+
+       if (ret && ret != -ERESTARTSYS) {
+               DRM_ERROR("timeout waiting for render thread idle\n");
+               return ret;
+       }
+
+       return 0;
+}
+
+static void
+vc4_flush_caches(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       /* Flush the GPU L2 caches.  These caches sit on top of system
+        * L3 (the 128kb or so shared with the CPU), and are
+        * non-allocating in the L3.
+        */
+       V3D_WRITE(V3D_L2CACTL,
+                 V3D_L2CACTL_L2CCLR);
+
+       V3D_WRITE(V3D_SLCACTL,
+                 VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
+                 VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
+                 VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
+                 VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
+}
+
+/* Sets the registers for the next job to be actually be executed in
+ * the hardware.
+ *
+ * The job_lock should be held during this.
+ */
+void
+vc4_submit_next_job(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       struct vc4_exec_info *exec = vc4_first_job(vc4);
+
+       if (!exec)
+               return;
+
+       vc4_flush_caches(dev);
+
+       /* Disable the binner's pre-loaded overflow memory address */
+       V3D_WRITE(V3D_BPOA, 0);
+       V3D_WRITE(V3D_BPOS, 0);
+
+       if (exec->ct0ca != exec->ct0ea)
+               submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
+       submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
+}
+
+static void
+vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
+{
+       struct vc4_bo *bo;
+       unsigned i;
+
+       for (i = 0; i < exec->bo_count; i++) {
+               bo = to_vc4_bo(&exec->bo[i]->base);
+               bo->seqno = seqno;
+       }
+
+       list_for_each_entry(bo, &exec->unref_list, unref_head) {
+               bo->seqno = seqno;
+       }
+}
+
+/* Queues a struct vc4_exec_info for execution.  If no job is
+ * currently executing, then submits it.
+ *
+ * Unlike most GPUs, our hardware only handles one command list at a
+ * time.  To queue multiple jobs at once, we'd need to edit the
+ * previous command list to have a jump to the new one at the end, and
+ * then bump the end address.  That's a change for a later date,
+ * though.
+ */
+static void
+vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       uint64_t seqno;
+       unsigned long irqflags;
+
+       spin_lock_irqsave(&vc4->job_lock, irqflags);
+
+       seqno = ++vc4->emit_seqno;
+       exec->seqno = seqno;
+       vc4_update_bo_seqnos(exec, seqno);
+
+       list_add_tail(&exec->head, &vc4->job_list);
+
+       /* If no job was executing, kick ours off.  Otherwise, it'll
+        * get started when the previous job's frame done interrupt
+        * occurs.
+        */
+       if (vc4_first_job(vc4) == exec) {
+               vc4_submit_next_job(dev);
+               vc4_queue_hangcheck(dev);
+       }
+
+       spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+}
+
+/**
+ * Looks up a bunch of GEM handles for BOs and stores the array for
+ * use in the command validator that actually writes relocated
+ * addresses pointing to them.
+ */
+static int
+vc4_cl_lookup_bos(struct drm_device *dev,
+                 struct drm_file *file_priv,
+                 struct vc4_exec_info *exec)
+{
+       struct drm_vc4_submit_cl *args = exec->args;
+       uint32_t *handles;
+       int ret = 0;
+       int i;
+
+       exec->bo_count = args->bo_handle_count;
+
+       if (!exec->bo_count) {
+               /* See comment on bo_index for why we have to check
+                * this.
+                */
+               DRM_ERROR("Rendering requires BOs to validate\n");
+               return -EINVAL;
+       }
+
+       exec->bo = kcalloc(exec->bo_count, sizeof(struct drm_gem_cma_object *),
+                          GFP_KERNEL);
+       if (!exec->bo) {
+               DRM_ERROR("Failed to allocate validated BO pointers\n");
+               return -ENOMEM;
+       }
+
+       handles = drm_malloc_ab(exec->bo_count, sizeof(uint32_t));
+       if (!handles) {
+               DRM_ERROR("Failed to allocate incoming GEM handles\n");
+               goto fail;
+       }
+
+       ret = copy_from_user(handles,
+                            (void __user *)(uintptr_t)args->bo_handles,
+                            exec->bo_count * sizeof(uint32_t));
+       if (ret) {
+               DRM_ERROR("Failed to copy in GEM handles\n");
+               goto fail;
+       }
+
+       spin_lock(&file_priv->table_lock);
+       for (i = 0; i < exec->bo_count; i++) {
+               struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
+                                                    handles[i]);
+               if (!bo) {
+                       DRM_ERROR("Failed to look up GEM BO %d: %d\n",
+                                 i, handles[i]);
+                       ret = -EINVAL;
+                       spin_unlock(&file_priv->table_lock);
+                       goto fail;
+               }
+               drm_gem_object_reference(bo);
+               exec->bo[i] = (struct drm_gem_cma_object *)bo;
+       }
+       spin_unlock(&file_priv->table_lock);
+
+fail:
+       kfree(handles);
+       return 0;
+}
+
+static int
+vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+       struct drm_vc4_submit_cl *args = exec->args;
+       void *temp = NULL;
+       void *bin;
+       int ret = 0;
+       uint32_t bin_offset = 0;
+       uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
+                                            16);
+       uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
+       uint32_t exec_size = uniforms_offset + args->uniforms_size;
+       uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
+                                         args->shader_rec_count);
+       struct vc4_bo *bo;
+
+       if (uniforms_offset < shader_rec_offset ||
+           exec_size < uniforms_offset ||
+           args->shader_rec_count >= (UINT_MAX /
+                                         sizeof(struct vc4_shader_state)) ||
+           temp_size < exec_size) {
+               DRM_ERROR("overflow in exec arguments\n");
+               goto fail;
+       }
+
+       /* Allocate space where we'll store the copied in user command lists
+        * and shader records.
+        *
+        * We don't just copy directly into the BOs because we need to
+        * read the contents back for validation, and I think the
+        * bo->vaddr is uncached access.
+        */
+       temp = kmalloc(temp_size, GFP_KERNEL);
+       if (!temp) {
+               DRM_ERROR("Failed to allocate storage for copying "
+                         "in bin/render CLs.\n");
+               ret = -ENOMEM;
+               goto fail;
+       }
+       bin = temp + bin_offset;
+       exec->shader_rec_u = temp + shader_rec_offset;
+       exec->uniforms_u = temp + uniforms_offset;
+       exec->shader_state = temp + exec_size;
+       exec->shader_state_size = args->shader_rec_count;
+
+       ret = copy_from_user(bin,
+                            (void __user *)(uintptr_t)args->bin_cl,
+                            args->bin_cl_size);
+       if (ret) {
+               DRM_ERROR("Failed to copy in bin cl\n");
+               goto fail;
+       }
+
+       ret = copy_from_user(exec->shader_rec_u,
+                            (void __user *)(uintptr_t)args->shader_rec,
+                            args->shader_rec_size);
+       if (ret) {
+               DRM_ERROR("Failed to copy in shader recs\n");
+               goto fail;
+       }
+
+       ret = copy_from_user(exec->uniforms_u,
+                            (void __user *)(uintptr_t)args->uniforms,
+                            args->uniforms_size);
+       if (ret) {
+               DRM_ERROR("Failed to copy in uniforms cl\n");
+               goto fail;
+       }
+
+       bo = vc4_bo_create(dev, exec_size, true);
+       if (!bo) {
+               DRM_ERROR("Couldn't allocate BO for binning\n");
+               ret = PTR_ERR(exec->exec_bo);
+               goto fail;
+       }
+       exec->exec_bo = &bo->base;
+
+       list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
+                     &exec->unref_list);
+
+       exec->ct0ca = exec->exec_bo->paddr + bin_offset;
+
+       exec->bin_u = bin;
+
+       exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
+       exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
+       exec->shader_rec_size = args->shader_rec_size;
+
+       exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
+       exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
+       exec->uniforms_size = args->uniforms_size;
+
+       ret = vc4_validate_bin_cl(dev,
+                                 exec->exec_bo->vaddr + bin_offset,
+                                 bin,
+                                 exec);
+       if (ret)
+               goto fail;
+
+       ret = vc4_validate_shader_recs(dev, exec);
+
+fail:
+       kfree(temp);
+       return ret;
+}
+
+static void
+vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+       unsigned i;
+
+       /* Need the struct lock for drm_gem_object_unreference(). */
+       mutex_lock(&dev->struct_mutex);
+       if (exec->bo) {
+               for (i = 0; i < exec->bo_count; i++)
+                       drm_gem_object_unreference(&exec->bo[i]->base);
+               kfree(exec->bo);
+       }
+
+       while (!list_empty(&exec->unref_list)) {
+               struct vc4_bo *bo = list_first_entry(&exec->unref_list,
+                                                    struct vc4_bo, unref_head);
+               list_del(&bo->unref_head);
+               drm_gem_object_unreference(&bo->base.base);
+       }
+       mutex_unlock(&dev->struct_mutex);
+
+       kfree(exec);
+}
+
+void
+vc4_job_handle_completed(struct vc4_dev *vc4)
+{
+       unsigned long irqflags;
+
+       spin_lock_irqsave(&vc4->job_lock, irqflags);
+       while (!list_empty(&vc4->job_done_list)) {
+               struct vc4_exec_info *exec =
+                       list_first_entry(&vc4->job_done_list,
+                                        struct vc4_exec_info, head);
+               list_del(&exec->head);
+
+               spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+               vc4_complete_exec(vc4->dev, exec);
+               spin_lock_irqsave(&vc4->job_lock, irqflags);
+       }
+       spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+}
+
+/* Scheduled when any job has been completed, this walks the list of
+ * jobs that had completed and unrefs their BOs and frees their exec
+ * structs.
+ */
+static void
+vc4_job_done_work(struct work_struct *work)
+{
+       struct vc4_dev *vc4 =
+               container_of(work, struct vc4_dev, job_done_work);
+
+       vc4_job_handle_completed(vc4);
+}
+
+static int
+vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
+                               uint64_t seqno,
+                               uint64_t *timeout_ns)
+{
+       unsigned long start = jiffies;
+       int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
+
+       if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
+               uint64_t delta = jiffies_to_nsecs(jiffies - start);
+
+               if (*timeout_ns >= delta)
+                       *timeout_ns -= delta;
+       }
+
+       return ret;
+}
+
+int
+vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
+                    struct drm_file *file_priv)
+{
+       struct drm_vc4_wait_seqno *args = data;
+
+       return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
+                                              &args->timeout_ns);
+}
+
+int
+vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
+                 struct drm_file *file_priv)
+{
+       int ret;
+       struct drm_vc4_wait_bo *args = data;
+       struct drm_gem_object *gem_obj;
+       struct vc4_bo *bo;
+
+       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;
+       }
+       bo = to_vc4_bo(gem_obj);
+
+       ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
+                                             &args->timeout_ns);
+
+       drm_gem_object_unreference_unlocked(gem_obj);
+       return ret;
+}
+
+/**
+ * Submits a command list to the VC4.
+ *
+ * This is what is called batchbuffer emitting on other hardware.
+ */
+int
+vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
+                   struct drm_file *file_priv)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       struct drm_vc4_submit_cl *args = data;
+       struct vc4_exec_info *exec;
+       int ret;
+
+       if ((args->flags & ~VC4_SUBMIT_CL_USE_CLEAR_COLOR) != 0) {
+               DRM_ERROR("Unknown flags: 0x%02x\n", args->flags);
+               return -EINVAL;
+       }
+
+       exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
+       if (!exec) {
+               DRM_ERROR("malloc failure on exec struct\n");
+               return -ENOMEM;
+       }
+
+       exec->args = args;
+       INIT_LIST_HEAD(&exec->unref_list);
+
+       ret = vc4_cl_lookup_bos(dev, file_priv, exec);
+       if (ret)
+               goto fail;
+
+       if (exec->args->bin_cl_size != 0) {
+               ret = vc4_get_bcl(dev, exec);
+               if (ret)
+                       goto fail;
+       } else {
+               exec->ct0ca = 0;
+               exec->ct0ea = 0;
+       }
+
+       ret = vc4_get_rcl(dev, exec);
+       if (ret)
+               goto fail;
+
+       /* Clear this out of the struct we'll be putting in the queue,
+        * since it's part of our stack.
+        */
+       exec->args = NULL;
+
+       vc4_queue_submit(dev, exec);
+
+       /* Return the seqno for our job. */
+       args->seqno = vc4->emit_seqno;
+
+       return 0;
+
+fail:
+       vc4_complete_exec(vc4->dev, exec);
+
+       return ret;
+}
+
+void
+vc4_gem_init(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       INIT_LIST_HEAD(&vc4->job_list);
+       INIT_LIST_HEAD(&vc4->job_done_list);
+       spin_lock_init(&vc4->job_lock);
+
+       INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
+       setup_timer(&vc4->hangcheck.timer,
+                   vc4_hangcheck_elapsed,
+                   (unsigned long)dev);
+
+       INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
+}
+
+void
+vc4_gem_destroy(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       /* Waiting for exec to finish would need to be done before
+        * unregistering V3D.
+        */
+       WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
+
+       /* V3D should already have disabled its interrupt and cleared
+        * the overflow allocation registers.  Now free the object.
+        */
+       if (vc4->overflow_mem) {
+               drm_gem_object_unreference_unlocked(&vc4->overflow_mem->base.base);
+               vc4->overflow_mem = NULL;
+       }
+
+       vc4_bo_cache_destroy(dev);
+}
diff --git a/drivers/gpu/drm/vc4/vc4_irq.c b/drivers/gpu/drm/vc4/vc4_irq.c
new file mode 100644 (file)
index 0000000..b68060e
--- /dev/null
@@ -0,0 +1,210 @@
+/*
+ * 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: Interrupt management for the V3D engine.
+ *
+ * We have an interrupt status register (V3D_INTCTL) which reports
+ * interrupts, and where writing 1 bits clears those interrupts.
+ * There are also a pair of interrupt registers
+ * (V3D_INTENA/V3D_INTDIS) where writing a 1 to their bits enables or
+ * disables that specific interrupt, and 0s written are ignored
+ * (reading either one returns the set of enabled interrupts).
+ *
+ * When we take a render frame interrupt, we need to wake the
+ * processes waiting for some frame to be done, and get the next frame
+ * submitted ASAP (so the hardware doesn't sit idle when there's work
+ * to do).
+ *
+ * When we take the binner out of memory interrupt, we need to
+ * allocate some new memory and pass it to the binner so that the
+ * current job can make progress.
+ */
+
+#include "vc4_drv.h"
+#include "vc4_regs.h"
+
+#define V3D_DRIVER_IRQS (V3D_INT_OUTOMEM | \
+                        V3D_INT_FRDONE)
+
+DECLARE_WAIT_QUEUE_HEAD(render_wait);
+
+static void
+vc4_overflow_mem_work(struct work_struct *work)
+{
+       struct vc4_dev *vc4 =
+               container_of(work, struct vc4_dev, overflow_mem_work);
+       struct drm_device *dev = vc4->dev;
+       struct vc4_bo *bo;
+
+       bo = vc4_bo_create(dev, 256 * 1024, true);
+       if (!bo) {
+               DRM_ERROR("Couldn't allocate binner overflow mem\n");
+               return;
+       }
+
+       /* If there's a job executing currently, then our previous
+        * overflow allocation is getting used in that job and we need
+        * to queue it to be released when the job is done.  But if no
+        * job is executing at all, then we can free the old overflow
+        * object direcctly.
+        *
+        * No lock necessary for this pointer since we're the only
+        * ones that update the pointer, and our workqueue won't
+        * reenter.
+        */
+       if (vc4->overflow_mem) {
+               struct vc4_exec_info *current_exec;
+               unsigned long irqflags;
+
+               spin_lock_irqsave(&vc4->job_lock, irqflags);
+               current_exec = vc4_first_job(vc4);
+               if (current_exec) {
+                       vc4->overflow_mem->seqno = vc4->finished_seqno + 1;
+                       list_add_tail(&vc4->overflow_mem->unref_head,
+                                     &current_exec->unref_list);
+                       vc4->overflow_mem = NULL;
+               }
+               spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+       }
+
+       if (vc4->overflow_mem)
+               drm_gem_object_unreference_unlocked(&vc4->overflow_mem->base.base);
+       vc4->overflow_mem = bo;
+
+       V3D_WRITE(V3D_BPOA, bo->base.paddr);
+       V3D_WRITE(V3D_BPOS, bo->base.base.size);
+       V3D_WRITE(V3D_INTCTL, V3D_INT_OUTOMEM);
+       V3D_WRITE(V3D_INTENA, V3D_INT_OUTOMEM);
+}
+
+static void
+vc4_irq_finish_job(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       struct vc4_exec_info *exec = vc4_first_job(vc4);
+
+       if (!exec)
+               return;
+
+       vc4->finished_seqno++;
+       list_move_tail(&exec->head, &vc4->job_done_list);
+       vc4_submit_next_job(dev);
+
+       wake_up_all(&vc4->job_wait_queue);
+       schedule_work(&vc4->job_done_work);
+}
+
+irqreturn_t
+vc4_irq(int irq, void *arg)
+{
+       struct drm_device *dev = arg;
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       uint32_t intctl;
+       irqreturn_t status = IRQ_NONE;
+
+       barrier();
+       intctl = V3D_READ(V3D_INTCTL);
+
+       /* Acknowledge the interrupts we're handling here. The render
+        * frame done interrupt will be cleared, while OUTOMEM will
+        * stay high until the underlying cause is cleared.
+        */
+       V3D_WRITE(V3D_INTCTL, intctl);
+
+       if (intctl & V3D_INT_OUTOMEM) {
+               /* Disable OUTOMEM until the work is done. */
+               V3D_WRITE(V3D_INTDIS, V3D_INT_OUTOMEM);
+               schedule_work(&vc4->overflow_mem_work);
+               status = IRQ_HANDLED;
+       }
+
+       if (intctl & V3D_INT_FRDONE) {
+               spin_lock(&vc4->job_lock);
+               vc4_irq_finish_job(dev);
+               spin_unlock(&vc4->job_lock);
+               status = IRQ_HANDLED;
+       }
+
+       return status;
+}
+
+void
+vc4_irq_preinstall(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       init_waitqueue_head(&vc4->job_wait_queue);
+       INIT_WORK(&vc4->overflow_mem_work, vc4_overflow_mem_work);
+
+       /* Clear any pending interrupts someone might have left around
+        * for us.
+        */
+       V3D_WRITE(V3D_INTCTL, V3D_DRIVER_IRQS);
+}
+
+int
+vc4_irq_postinstall(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       /* Enable both the render done and out of memory interrupts. */
+       V3D_WRITE(V3D_INTENA, V3D_DRIVER_IRQS);
+
+       return 0;
+}
+
+void
+vc4_irq_uninstall(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+       /* Disable sending interrupts for our driver's IRQs. */
+       V3D_WRITE(V3D_INTDIS, V3D_DRIVER_IRQS);
+
+       /* Clear any pending interrupts we might have left. */
+       V3D_WRITE(V3D_INTCTL, V3D_DRIVER_IRQS);
+
+       cancel_work_sync(&vc4->overflow_mem_work);
+}
+
+/** Reinitializes interrupt registers when a GPU reset is performed. */
+void vc4_irq_reset(struct drm_device *dev)
+{
+       struct vc4_dev *vc4 = to_vc4_dev(dev);
+       unsigned long irqflags;
+
+       /* Acknowledge any stale IRQs. */
+       V3D_WRITE(V3D_INTCTL, V3D_DRIVER_IRQS);
+
+       /*
+        * Turn all our interrupts on.  Binner out of memory is the
+        * only one we expect to trigger at this point, since we've
+        * just come from poweron and haven't supplied any overflow
+        * memory yet.
+        */
+       V3D_WRITE(V3D_INTENA, V3D_DRIVER_IRQS);
+
+       spin_lock_irqsave(&vc4->job_lock, irqflags);
+       vc4_irq_finish_job(dev);
+       spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+}
diff --git a/drivers/gpu/drm/vc4/vc4_packet.h b/drivers/gpu/drm/vc4/vc4_packet.h
new file mode 100644 (file)
index 0000000..0f31cc0
--- /dev/null
@@ -0,0 +1,399 @@
+/*
+ * 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_PACKET_H
+#define VC4_PACKET_H
+
+#include "vc4_regs.h" /* for VC4_MASK, VC4_GET_FIELD, VC4_SET_FIELD */
+
+enum vc4_packet {
+       VC4_PACKET_HALT = 0,
+       VC4_PACKET_NOP = 1,
+
+       VC4_PACKET_FLUSH = 4,
+       VC4_PACKET_FLUSH_ALL = 5,
+       VC4_PACKET_START_TILE_BINNING = 6,
+       VC4_PACKET_INCREMENT_SEMAPHORE = 7,
+       VC4_PACKET_WAIT_ON_SEMAPHORE = 8,
+
+       VC4_PACKET_BRANCH = 16,
+       VC4_PACKET_BRANCH_TO_SUB_LIST = 17,
+
+       VC4_PACKET_STORE_MS_TILE_BUFFER = 24,
+       VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF = 25,
+       VC4_PACKET_STORE_FULL_RES_TILE_BUFFER = 26,
+       VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER = 27,
+       VC4_PACKET_STORE_TILE_BUFFER_GENERAL = 28,
+       VC4_PACKET_LOAD_TILE_BUFFER_GENERAL = 29,
+
+       VC4_PACKET_GL_INDEXED_PRIMITIVE = 32,
+       VC4_PACKET_GL_ARRAY_PRIMITIVE = 33,
+
+       VC4_PACKET_COMPRESSED_PRIMITIVE = 48,
+       VC4_PACKET_CLIPPED_COMPRESSED_PRIMITIVE = 49,
+
+       VC4_PACKET_PRIMITIVE_LIST_FORMAT = 56,
+
+       VC4_PACKET_GL_SHADER_STATE = 64,
+       VC4_PACKET_NV_SHADER_STATE = 65,
+       VC4_PACKET_VG_SHADER_STATE = 66,
+
+       VC4_PACKET_CONFIGURATION_BITS = 96,
+       VC4_PACKET_FLAT_SHADE_FLAGS = 97,
+       VC4_PACKET_POINT_SIZE = 98,
+       VC4_PACKET_LINE_WIDTH = 99,
+       VC4_PACKET_RHT_X_BOUNDARY = 100,
+       VC4_PACKET_DEPTH_OFFSET = 101,
+       VC4_PACKET_CLIP_WINDOW = 102,
+       VC4_PACKET_VIEWPORT_OFFSET = 103,
+       VC4_PACKET_Z_CLIPPING = 104,
+       VC4_PACKET_CLIPPER_XY_SCALING = 105,
+       VC4_PACKET_CLIPPER_Z_SCALING = 106,
+
+       VC4_PACKET_TILE_BINNING_MODE_CONFIG = 112,
+       VC4_PACKET_TILE_RENDERING_MODE_CONFIG = 113,
+       VC4_PACKET_CLEAR_COLORS = 114,
+       VC4_PACKET_TILE_COORDINATES = 115,
+
+       /* Not an actual hardware packet -- this is what we use to put
+        * references to GEM bos in the command stream, since we need the u32
+        * int the actual address packet in order to store the offset from the
+        * start of the BO.
+        */
+       VC4_PACKET_GEM_HANDLES = 254,
+} __attribute__ ((__packed__));
+
+#define VC4_PACKET_HALT_SIZE                                           1
+#define VC4_PACKET_NOP_SIZE                                            1
+#define VC4_PACKET_FLUSH_SIZE                                          1
+#define VC4_PACKET_FLUSH_ALL_SIZE                                      1
+#define VC4_PACKET_START_TILE_BINNING_SIZE                             1
+#define VC4_PACKET_INCREMENT_SEMAPHORE_SIZE                            1
+#define VC4_PACKET_WAIT_ON_SEMAPHORE_SIZE                              1
+#define VC4_PACKET_BRANCH_SIZE                                         5
+#define VC4_PACKET_BRANCH_TO_SUB_LIST_SIZE                             5
+#define VC4_PACKET_STORE_MS_TILE_BUFFER_SIZE                           1
+#define VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF_SIZE                   1
+#define VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE                     5
+#define VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE                      5
+#define VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE                      7
+#define VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE                       7
+#define VC4_PACKET_GL_INDEXED_PRIMITIVE_SIZE                           14
+#define VC4_PACKET_GL_ARRAY_PRIMITIVE_SIZE                             10
+#define VC4_PACKET_COMPRESSED_PRIMITIVE_SIZE                           1
+#define VC4_PACKET_CLIPPED_COMPRESSED_PRIMITIVE_SIZE                   1
+#define VC4_PACKET_PRIMITIVE_LIST_FORMAT_SIZE                          2
+#define VC4_PACKET_GL_SHADER_STATE_SIZE                                        5
+#define VC4_PACKET_NV_SHADER_STATE_SIZE                                        5
+#define VC4_PACKET_VG_SHADER_STATE_SIZE                                        5
+#define VC4_PACKET_CONFIGURATION_BITS_SIZE                             4
+#define VC4_PACKET_FLAT_SHADE_FLAGS_SIZE                               5
+#define VC4_PACKET_POINT_SIZE_SIZE                                     5
+#define VC4_PACKET_LINE_WIDTH_SIZE                                     5
+#define VC4_PACKET_RHT_X_BOUNDARY_SIZE                                 3
+#define VC4_PACKET_DEPTH_OFFSET_SIZE                                   5
+#define VC4_PACKET_CLIP_WINDOW_SIZE                                    9
+#define VC4_PACKET_VIEWPORT_OFFSET_SIZE                                        5
+#define VC4_PACKET_Z_CLIPPING_SIZE                                     9
+#define VC4_PACKET_CLIPPER_XY_SCALING_SIZE                             9
+#define VC4_PACKET_CLIPPER_Z_SCALING_SIZE                              9
+#define VC4_PACKET_TILE_BINNING_MODE_CONFIG_SIZE                       16
+#define VC4_PACKET_TILE_RENDERING_MODE_CONFIG_SIZE                     11
+#define VC4_PACKET_CLEAR_COLORS_SIZE                                   14
+#define VC4_PACKET_TILE_COORDINATES_SIZE                               3
+#define VC4_PACKET_GEM_HANDLES_SIZE                                    9
+
+/* Number of multisamples supported. */
+#define VC4_MAX_SAMPLES                                                        4
+/* Size of a full resolution color or Z tile buffer load/store. */
+#define VC4_TILE_BUFFER_SIZE                   (64 * 64 * 4)
+
+/** @{
+ * Bits used by packets like VC4_PACKET_STORE_TILE_BUFFER_GENERAL and
+ * VC4_PACKET_TILE_RENDERING_MODE_CONFIG.
+*/
+#define VC4_TILING_FORMAT_LINEAR    0
+#define VC4_TILING_FORMAT_T         1
+#define VC4_TILING_FORMAT_LT        2
+/** @} */
+
+/** @{
+ *
+ * low bits of VC4_PACKET_STORE_FULL_RES_TILE_BUFFER and
+ * VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER.
+ */
+#define VC4_LOADSTORE_FULL_RES_EOF                     BIT(3)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       BIT(2)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              BIT(1)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           BIT(0)
+
+/** @{
+ *
+ * low bits of VC4_PACKET_STORE_FULL_RES_TILE_BUFFER and
+ * VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER.
+ */
+#define VC4_LOADSTORE_FULL_RES_EOF                     BIT(3)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       BIT(2)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              BIT(1)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           BIT(0)
+
+/** @{
+ *
+ * byte 2 of VC4_PACKET_STORE_TILE_BUFFER_GENERAL and
+ * VC4_PACKET_LOAD_TILE_BUFFER_GENERAL (low bits of the address)
+ */
+
+#define VC4_LOADSTORE_TILE_BUFFER_EOF                  BIT(3)
+#define VC4_LOADSTORE_TILE_BUFFER_DISABLE_FULL_VG_MASK BIT(2)
+#define VC4_LOADSTORE_TILE_BUFFER_DISABLE_FULL_ZS      BIT(1)
+#define VC4_LOADSTORE_TILE_BUFFER_DISABLE_FULL_COLOR   BIT(0)
+
+/** @} */
+
+/** @{
+ *
+ * byte 0-1 of VC4_PACKET_STORE_TILE_BUFFER_GENERAL and
+ * VC4_PACKET_LOAD_TILE_BUFFER_GENERAL
+ */
+#define VC4_STORE_TILE_BUFFER_DISABLE_VG_MASK_CLEAR BIT(15)
+#define VC4_STORE_TILE_BUFFER_DISABLE_ZS_CLEAR     BIT(14)
+#define VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR  BIT(13)
+#define VC4_STORE_TILE_BUFFER_DISABLE_SWAP         BIT(12)
+
+#define VC4_LOADSTORE_TILE_BUFFER_FORMAT_MASK      VC4_MASK(9, 8)
+#define VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT     8
+#define VC4_LOADSTORE_TILE_BUFFER_RGBA8888         0
+#define VC4_LOADSTORE_TILE_BUFFER_BGR565_DITHER    1
+#define VC4_LOADSTORE_TILE_BUFFER_BGR565           2
+/** @} */
+
+/** @{
+ *
+ * byte 0 of VC4_PACKET_STORE_TILE_BUFFER_GENERAL and
+ * VC4_PACKET_LOAD_TILE_BUFFER_GENERAL
+ */
+#define VC4_STORE_TILE_BUFFER_MODE_MASK            VC4_MASK(7, 6)
+#define VC4_STORE_TILE_BUFFER_MODE_SHIFT           6
+#define VC4_STORE_TILE_BUFFER_MODE_SAMPLE0         (0 << 6)
+#define VC4_STORE_TILE_BUFFER_MODE_DECIMATE_X4     (1 << 6)
+#define VC4_STORE_TILE_BUFFER_MODE_DECIMATE_X16    (2 << 6)
+
+/** The values of the field are VC4_TILING_FORMAT_* */
+#define VC4_LOADSTORE_TILE_BUFFER_TILING_MASK      VC4_MASK(5, 4)
+#define VC4_LOADSTORE_TILE_BUFFER_TILING_SHIFT     4
+
+#define VC4_LOADSTORE_TILE_BUFFER_BUFFER_MASK      VC4_MASK(2, 0)
+#define VC4_LOADSTORE_TILE_BUFFER_BUFFER_SHIFT     0
+#define VC4_LOADSTORE_TILE_BUFFER_NONE             0
+#define VC4_LOADSTORE_TILE_BUFFER_COLOR            1
+#define VC4_LOADSTORE_TILE_BUFFER_ZS               2
+#define VC4_LOADSTORE_TILE_BUFFER_Z                3
+#define VC4_LOADSTORE_TILE_BUFFER_VG_MASK          4
+#define VC4_LOADSTORE_TILE_BUFFER_FULL             5
+/** @} */
+
+#define VC4_INDEX_BUFFER_U8                        (0 << 4)
+#define VC4_INDEX_BUFFER_U16                       (1 << 4)
+
+/* This flag is only present in NV shader state. */
+#define VC4_SHADER_FLAG_SHADED_CLIP_COORDS         BIT(3)
+#define VC4_SHADER_FLAG_ENABLE_CLIPPING            BIT(2)
+#define VC4_SHADER_FLAG_VS_POINT_SIZE              BIT(1)
+#define VC4_SHADER_FLAG_FS_SINGLE_THREAD           BIT(0)
+
+/** @{ byte 2 of config bits. */
+#define VC4_CONFIG_BITS_EARLY_Z_UPDATE             BIT(1)
+#define VC4_CONFIG_BITS_EARLY_Z                    BIT(0)
+/** @} */
+
+/** @{ byte 1 of config bits. */
+#define VC4_CONFIG_BITS_Z_UPDATE                   BIT(7)
+/** same values in this 3-bit field as PIPE_FUNC_* */
+#define VC4_CONFIG_BITS_DEPTH_FUNC_SHIFT           4
+#define VC4_CONFIG_BITS_COVERAGE_READ_LEAVE        BIT(3)
+
+#define VC4_CONFIG_BITS_COVERAGE_UPDATE_NONZERO    (0 << 1)
+#define VC4_CONFIG_BITS_COVERAGE_UPDATE_ODD        (1 << 1)
+#define VC4_CONFIG_BITS_COVERAGE_UPDATE_OR         (2 << 1)
+#define VC4_CONFIG_BITS_COVERAGE_UPDATE_ZERO       (3 << 1)
+
+#define VC4_CONFIG_BITS_COVERAGE_PIPE_SELECT       BIT(0)
+/** @} */
+
+/** @{ byte 0 of config bits. */
+#define VC4_CONFIG_BITS_RASTERIZER_OVERSAMPLE_NONE (0 << 6)
+#define VC4_CONFIG_BITS_RASTERIZER_OVERSAMPLE_4X   (1 << 6)
+#define VC4_CONFIG_BITS_RASTERIZER_OVERSAMPLE_16X  (2 << 6)
+
+#define VC4_CONFIG_BITS_AA_POINTS_AND_LINES        BIT(4)
+#define VC4_CONFIG_BITS_ENABLE_DEPTH_OFFSET        BIT(3)
+#define VC4_CONFIG_BITS_CW_PRIMITIVES              BIT(2)
+#define VC4_CONFIG_BITS_ENABLE_PRIM_BACK           BIT(1)
+#define VC4_CONFIG_BITS_ENABLE_PRIM_FRONT          BIT(0)
+/** @} */
+
+/** @{ bits in the last u8 of VC4_PACKET_TILE_BINNING_MODE_CONFIG */
+#define VC4_BIN_CONFIG_DB_NON_MS                   BIT(7)
+
+#define VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK       VC4_MASK(6, 5)
+#define VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_SHIFT      5
+#define VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_32         0
+#define VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_64         1
+#define VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128        2
+#define VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_256        3
+
+#define VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK  VC4_MASK(4, 3)
+#define VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_SHIFT 3
+#define VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32    0
+#define VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_64    1
+#define VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_128   2
+#define VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_256   3
+
+#define VC4_BIN_CONFIG_AUTO_INIT_TSDA              BIT(2)
+#define VC4_BIN_CONFIG_TILE_BUFFER_64BIT           BIT(1)
+#define VC4_BIN_CONFIG_MS_MODE_4X                  BIT(0)
+/** @} */
+
+/** @{ bits in the last u16 of VC4_PACKET_TILE_RENDERING_MODE_CONFIG */
+#define VC4_RENDER_CONFIG_DB_NON_MS                BIT(12)
+#define VC4_RENDER_CONFIG_EARLY_Z_COVERAGE_DISABLE BIT(11)
+#define VC4_RENDER_CONFIG_EARLY_Z_DIRECTION_G      BIT(10)
+#define VC4_RENDER_CONFIG_COVERAGE_MODE            BIT(9)
+#define VC4_RENDER_CONFIG_ENABLE_VG_MASK           BIT(8)
+
+/** The values of the field are VC4_TILING_FORMAT_* */
+#define VC4_RENDER_CONFIG_MEMORY_FORMAT_MASK       VC4_MASK(7, 6)
+#define VC4_RENDER_CONFIG_MEMORY_FORMAT_SHIFT      6
+
+#define VC4_RENDER_CONFIG_DECIMATE_MODE_1X         (0 << 4)
+#define VC4_RENDER_CONFIG_DECIMATE_MODE_4X         (1 << 4)
+#define VC4_RENDER_CONFIG_DECIMATE_MODE_16X        (2 << 4)
+
+#define VC4_RENDER_CONFIG_FORMAT_MASK              VC4_MASK(3, 2)
+#define VC4_RENDER_CONFIG_FORMAT_SHIFT             2
+#define VC4_RENDER_CONFIG_FORMAT_BGR565_DITHERED   0
+#define VC4_RENDER_CONFIG_FORMAT_RGBA8888          1
+#define VC4_RENDER_CONFIG_FORMAT_BGR565            2
+
+#define VC4_RENDER_CONFIG_TILE_BUFFER_64BIT        BIT(1)
+#define VC4_RENDER_CONFIG_MS_MODE_4X               BIT(0)
+
+#define VC4_PRIMITIVE_LIST_FORMAT_16_INDEX         (1 << 4)
+#define VC4_PRIMITIVE_LIST_FORMAT_32_XY            (3 << 4)
+#define VC4_PRIMITIVE_LIST_FORMAT_TYPE_POINTS      (0 << 0)
+#define VC4_PRIMITIVE_LIST_FORMAT_TYPE_LINES       (1 << 0)
+#define VC4_PRIMITIVE_LIST_FORMAT_TYPE_TRIANGLES   (2 << 0)
+#define VC4_PRIMITIVE_LIST_FORMAT_TYPE_RHT         (3 << 0)
+
+enum vc4_texture_data_type {
+       VC4_TEXTURE_TYPE_RGBA8888 = 0,
+       VC4_TEXTURE_TYPE_RGBX8888 = 1,
+       VC4_TEXTURE_TYPE_RGBA4444 = 2,
+       VC4_TEXTURE_TYPE_RGBA5551 = 3,
+       VC4_TEXTURE_TYPE_RGB565 = 4,
+       VC4_TEXTURE_TYPE_LUMINANCE = 5,
+       VC4_TEXTURE_TYPE_ALPHA = 6,
+       VC4_TEXTURE_TYPE_LUMALPHA = 7,
+       VC4_TEXTURE_TYPE_ETC1 = 8,
+       VC4_TEXTURE_TYPE_S16F = 9,
+       VC4_TEXTURE_TYPE_S8 = 10,
+       VC4_TEXTURE_TYPE_S16 = 11,
+       VC4_TEXTURE_TYPE_BW1 = 12,
+       VC4_TEXTURE_TYPE_A4 = 13,
+       VC4_TEXTURE_TYPE_A1 = 14,
+       VC4_TEXTURE_TYPE_RGBA64 = 15,
+       VC4_TEXTURE_TYPE_RGBA32R = 16,
+       VC4_TEXTURE_TYPE_YUV422R = 17,
+};
+
+#define VC4_TEX_P0_OFFSET_MASK                     VC4_MASK(31, 12)
+#define VC4_TEX_P0_OFFSET_SHIFT                    12
+#define VC4_TEX_P0_CSWIZ_MASK                      VC4_MASK(11, 10)
+#define VC4_TEX_P0_CSWIZ_SHIFT                     10
+#define VC4_TEX_P0_CMMODE_MASK                     VC4_MASK(9, 9)
+#define VC4_TEX_P0_CMMODE_SHIFT                    9
+#define VC4_TEX_P0_FLIPY_MASK                      VC4_MASK(8, 8)
+#define VC4_TEX_P0_FLIPY_SHIFT                     8
+#define VC4_TEX_P0_TYPE_MASK                       VC4_MASK(7, 4)
+#define VC4_TEX_P0_TYPE_SHIFT                      4
+#define VC4_TEX_P0_MIPLVLS_MASK                    VC4_MASK(3, 0)
+#define VC4_TEX_P0_MIPLVLS_SHIFT                   0
+
+#define VC4_TEX_P1_TYPE4_MASK                      VC4_MASK(31, 31)
+#define VC4_TEX_P1_TYPE4_SHIFT                     31
+#define VC4_TEX_P1_HEIGHT_MASK                     VC4_MASK(30, 20)
+#define VC4_TEX_P1_HEIGHT_SHIFT                    20
+#define VC4_TEX_P1_ETCFLIP_MASK                    VC4_MASK(19, 19)
+#define VC4_TEX_P1_ETCFLIP_SHIFT                   19
+#define VC4_TEX_P1_WIDTH_MASK                      VC4_MASK(18, 8)
+#define VC4_TEX_P1_WIDTH_SHIFT                     8
+
+#define VC4_TEX_P1_MAGFILT_MASK                    VC4_MASK(7, 7)
+#define VC4_TEX_P1_MAGFILT_SHIFT                   7
+# define VC4_TEX_P1_MAGFILT_LINEAR                 0
+# define VC4_TEX_P1_MAGFILT_NEAREST                1
+
+#define VC4_TEX_P1_MINFILT_MASK                    VC4_MASK(6, 4)
+#define VC4_TEX_P1_MINFILT_SHIFT                   4
+# define VC4_TEX_P1_MINFILT_LINEAR                 0
+# define VC4_TEX_P1_MINFILT_NEAREST                1
+# define VC4_TEX_P1_MINFILT_NEAR_MIP_NEAR          2
+# define VC4_TEX_P1_MINFILT_NEAR_MIP_LIN           3
+# define VC4_TEX_P1_MINFILT_LIN_MIP_NEAR           4
+# define VC4_TEX_P1_MINFILT_LIN_MIP_LIN            5
+
+#define VC4_TEX_P1_WRAP_T_MASK                     VC4_MASK(3, 2)
+#define VC4_TEX_P1_WRAP_T_SHIFT                    2
+#define VC4_TEX_P1_WRAP_S_MASK                     VC4_MASK(1, 0)
+#define VC4_TEX_P1_WRAP_S_SHIFT                    0
+# define VC4_TEX_P1_WRAP_REPEAT                    0
+# define VC4_TEX_P1_WRAP_CLAMP                     1
+# define VC4_TEX_P1_WRAP_MIRROR                    2
+# define VC4_TEX_P1_WRAP_BORDER                    3
+
+#define VC4_TEX_P2_PTYPE_MASK                      VC4_MASK(31, 30)
+#define VC4_TEX_P2_PTYPE_SHIFT                     30
+# define VC4_TEX_P2_PTYPE_IGNORED                  0
+# define VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE          1
+# define VC4_TEX_P2_PTYPE_CHILD_IMAGE_DIMENSIONS   2
+# define VC4_TEX_P2_PTYPE_CHILD_IMAGE_OFFSETS      3
+
+/* VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE bits */
+#define VC4_TEX_P2_CMST_MASK                       VC4_MASK(29, 12)
+#define VC4_TEX_P2_CMST_SHIFT                      12
+#define VC4_TEX_P2_BSLOD_MASK                      VC4_MASK(0, 0)
+#define VC4_TEX_P2_BSLOD_SHIFT                     0
+
+/* VC4_TEX_P2_PTYPE_CHILD_IMAGE_DIMENSIONS */
+#define VC4_TEX_P2_CHEIGHT_MASK                    VC4_MASK(22, 12)
+#define VC4_TEX_P2_CHEIGHT_SHIFT                   12
+#define VC4_TEX_P2_CWIDTH_MASK                     VC4_MASK(10, 0)
+#define VC4_TEX_P2_CWIDTH_SHIFT                    0
+
+/* VC4_TEX_P2_PTYPE_CHILD_IMAGE_OFFSETS */
+#define VC4_TEX_P2_CYOFF_MASK                      VC4_MASK(22, 12)
+#define VC4_TEX_P2_CYOFF_SHIFT                     12
+#define VC4_TEX_P2_CXOFF_MASK                      VC4_MASK(10, 0)
+#define VC4_TEX_P2_CXOFF_SHIFT                     0
+
+#endif /* VC4_PACKET_H */
diff --git a/drivers/gpu/drm/vc4/vc4_render_cl.c b/drivers/gpu/drm/vc4/vc4_render_cl.c
new file mode 100644 (file)
index 0000000..8a2a312
--- /dev/null
@@ -0,0 +1,634 @@
+/*
+ * 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.
+ */
+
+/**
+ * DOC: Render command list generation
+ *
+ * In the VC4 driver, render command list generation is performed by the
+ * kernel instead of userspace.  We do this because validating a
+ * user-submitted command list is hard to get right and has high CPU overhead,
+ * while the number of valid configurations for render command lists is
+ * actually fairly low.
+ */
+
+#include "uapi/drm/vc4_drm.h"
+#include "vc4_drv.h"
+#include "vc4_packet.h"
+
+struct vc4_rcl_setup {
+       struct drm_gem_cma_object *color_read;
+       struct drm_gem_cma_object *color_write;
+       struct drm_gem_cma_object *zs_read;
+       struct drm_gem_cma_object *zs_write;
+       struct drm_gem_cma_object *msaa_color_write;
+       struct drm_gem_cma_object *msaa_zs_write;
+
+       struct drm_gem_cma_object *rcl;
+       u32 next_offset;
+};
+
+static inline void rcl_u8(struct vc4_rcl_setup *setup, u8 val)
+{
+       *(u8 *)(setup->rcl->vaddr + setup->next_offset) = val;
+       setup->next_offset += 1;
+}
+
+static inline void rcl_u16(struct vc4_rcl_setup *setup, u16 val)
+{
+       *(u16 *)(setup->rcl->vaddr + setup->next_offset) = val;
+       setup->next_offset += 2;
+}
+
+static inline void rcl_u32(struct vc4_rcl_setup *setup, u32 val)
+{
+       *(u32 *)(setup->rcl->vaddr + setup->next_offset) = val;
+       setup->next_offset += 4;
+}
+
+/*
+ * Emits a no-op STORE_TILE_BUFFER_GENERAL.
+ *
+ * If we emit a PACKET_TILE_COORDINATES, it must be followed by a store of
+ * some sort before another load is triggered.
+ */
+static void vc4_store_before_load(struct vc4_rcl_setup *setup)
+{
+       rcl_u8(setup, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
+       rcl_u16(setup,
+               VC4_SET_FIELD(VC4_LOADSTORE_TILE_BUFFER_NONE,
+                             VC4_LOADSTORE_TILE_BUFFER_BUFFER) |
+               VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR |
+               VC4_STORE_TILE_BUFFER_DISABLE_ZS_CLEAR |
+               VC4_STORE_TILE_BUFFER_DISABLE_VG_MASK_CLEAR);
+       rcl_u32(setup, 0); /* no address, since we're in None mode */
+}
+
+/*
+ * Calculates the physical address of the start of a tile in a RCL surface.
+ *
+ * Unlike the other load/store packets,
+ * VC4_PACKET_LOAD/STORE_FULL_RES_TILE_BUFFER don't look at the tile
+ * coordinates packet, and instead just store to the address given.
+ */
+static uint32_t vc4_full_res_offset(struct vc4_exec_info *exec,
+                                   struct drm_gem_cma_object *bo,
+                                   struct drm_vc4_submit_rcl_surface *surf,
+                                   uint8_t x, uint8_t y)
+{
+       return bo->paddr + surf->offset + VC4_TILE_BUFFER_SIZE *
+               (DIV_ROUND_UP(exec->args->width, 32) * y + x);
+}
+
+/*
+ * Emits a PACKET_TILE_COORDINATES if one isn't already pending.
+ *
+ * The tile coordinates packet triggers a pending load if there is one, are
+ * used for clipping during rendering, and determine where loads/stores happen
+ * relative to their base address.
+ */
+static void vc4_tile_coordinates(struct vc4_rcl_setup *setup,
+                                uint32_t x, uint32_t y)
+{
+       rcl_u8(setup, VC4_PACKET_TILE_COORDINATES);
+       rcl_u8(setup, x);
+       rcl_u8(setup, y);
+}
+
+static void emit_tile(struct vc4_exec_info *exec,
+                     struct vc4_rcl_setup *setup,
+                     uint8_t x, uint8_t y, bool first, bool last)
+{
+       struct drm_vc4_submit_cl *args = exec->args;
+       bool has_bin = args->bin_cl_size != 0;
+
+       /* Note that the load doesn't actually occur until the
+        * tile coords packet is processed, and only one load
+        * may be outstanding at a time.
+        */
+       if (setup->color_read) {
+               if (args->color_read.flags &
+                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+                       rcl_u8(setup, VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER);
+                       rcl_u32(setup,
+                               vc4_full_res_offset(exec, setup->color_read,
+                                                   &args->color_read, x, y) |
+                               VC4_LOADSTORE_FULL_RES_DISABLE_ZS);
+               } else {
+                       rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
+                       rcl_u16(setup, args->color_read.bits);
+                       rcl_u32(setup, setup->color_read->paddr +
+                               args->color_read.offset);
+               }
+       }
+
+       if (setup->zs_read) {
+               if (args->zs_read.flags &
+                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+                       rcl_u8(setup, VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER);
+                       rcl_u32(setup,
+                               vc4_full_res_offset(exec, setup->zs_read,
+                                                   &args->zs_read, x, y) |
+                               VC4_LOADSTORE_FULL_RES_DISABLE_COLOR);
+               } else {
+                       if (setup->color_read) {
+                               /* Exec previous load. */
+                               vc4_tile_coordinates(setup, x, y);
+                               vc4_store_before_load(setup);
+                       }
+
+                       rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
+                       rcl_u16(setup, args->zs_read.bits);
+                       rcl_u32(setup, setup->zs_read->paddr +
+                               args->zs_read.offset);
+               }
+       }
+
+       /* Clipping depends on tile coordinates having been
+        * emitted, so we always need one here.
+        */
+       vc4_tile_coordinates(setup, x, y);
+
+       /* Wait for the binner before jumping to the first
+        * tile's lists.
+        */
+       if (first && has_bin)
+               rcl_u8(setup, VC4_PACKET_WAIT_ON_SEMAPHORE);
+
+       if (has_bin) {
+               rcl_u8(setup, VC4_PACKET_BRANCH_TO_SUB_LIST);
+               rcl_u32(setup, (exec->tile_bo->paddr +
+                               exec->tile_alloc_offset +
+                               (y * exec->bin_tiles_x + x) * 32));
+       }
+
+       if (setup->msaa_color_write) {
+               bool last_tile_write = (!setup->msaa_zs_write &&
+                                       !setup->zs_write &&
+                                       !setup->color_write);
+               uint32_t bits = VC4_LOADSTORE_FULL_RES_DISABLE_ZS;
+
+               if (!last_tile_write)
+                       bits |= VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL;
+               else if (last)
+                       bits |= VC4_LOADSTORE_FULL_RES_EOF;
+               rcl_u8(setup, VC4_PACKET_STORE_FULL_RES_TILE_BUFFER);
+               rcl_u32(setup,
+                       vc4_full_res_offset(exec, setup->msaa_color_write,
+                                           &args->msaa_color_write, x, y) |
+                       bits);
+       }
+
+       if (setup->msaa_zs_write) {
+               bool last_tile_write = (!setup->zs_write &&
+                                       !setup->color_write);
+               uint32_t bits = VC4_LOADSTORE_FULL_RES_DISABLE_COLOR;
+
+               if (setup->msaa_color_write)
+                       vc4_tile_coordinates(setup, x, y);
+               if (!last_tile_write)
+                       bits |= VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL;
+               else if (last)
+                       bits |= VC4_LOADSTORE_FULL_RES_EOF;
+               rcl_u8(setup, VC4_PACKET_STORE_FULL_RES_TILE_BUFFER);
+               rcl_u32(setup,
+                       vc4_full_res_offset(exec, setup->msaa_zs_write,
+                                           &args->msaa_zs_write, x, y) |
+                       bits);
+       }
+
+       if (setup->zs_write) {
+               bool last_tile_write = !setup->color_write;
+
+               if (setup->msaa_color_write || setup->msaa_zs_write)
+                       vc4_tile_coordinates(setup, x, y);
+
+               rcl_u8(setup, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
+               rcl_u16(setup, args->zs_write.bits |
+                       (last_tile_write ?
+                        0 : VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR));
+               rcl_u32(setup,
+                       (setup->zs_write->paddr + args->zs_write.offset) |
+                       ((last && last_tile_write) ?
+                        VC4_LOADSTORE_TILE_BUFFER_EOF : 0));
+       }
+
+       if (setup->color_write) {
+               if (setup->msaa_color_write || setup->msaa_zs_write ||
+                   setup->zs_write) {
+                       vc4_tile_coordinates(setup, x, y);
+               }
+
+               if (last)
+                       rcl_u8(setup, VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF);
+               else
+                       rcl_u8(setup, VC4_PACKET_STORE_MS_TILE_BUFFER);
+       }
+}
+
+static int vc4_create_rcl_bo(struct drm_device *dev, struct vc4_exec_info *exec,
+                            struct vc4_rcl_setup *setup)
+{
+       struct drm_vc4_submit_cl *args = exec->args;
+       bool has_bin = args->bin_cl_size != 0;
+       uint8_t min_x_tile = args->min_x_tile;
+       uint8_t min_y_tile = args->min_y_tile;
+       uint8_t max_x_tile = args->max_x_tile;
+       uint8_t max_y_tile = args->max_y_tile;
+       uint8_t xtiles = max_x_tile - min_x_tile + 1;
+       uint8_t ytiles = max_y_tile - min_y_tile + 1;
+       uint8_t x, y;
+       uint32_t size, loop_body_size;
+
+       size = VC4_PACKET_TILE_RENDERING_MODE_CONFIG_SIZE;
+       loop_body_size = VC4_PACKET_TILE_COORDINATES_SIZE;
+
+       if (args->flags & VC4_SUBMIT_CL_USE_CLEAR_COLOR) {
+               size += VC4_PACKET_CLEAR_COLORS_SIZE +
+                       VC4_PACKET_TILE_COORDINATES_SIZE +
+                       VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
+       }
+
+       if (setup->color_read) {
+               if (args->color_read.flags &
+                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+                       loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
+               } else {
+                       loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
+               }
+       }
+       if (setup->zs_read) {
+               if (args->zs_read.flags &
+                   VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+                       loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
+               } else {
+                       if (setup->color_read &&
+                           !(args->color_read.flags &
+                             VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES)) {
+                               loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
+                               loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
+                       }
+                       loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
+               }
+       }
+
+       if (has_bin) {
+               size += VC4_PACKET_WAIT_ON_SEMAPHORE_SIZE;
+               loop_body_size += VC4_PACKET_BRANCH_TO_SUB_LIST_SIZE;
+       }
+
+       if (setup->msaa_color_write)
+               loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
+       if (setup->msaa_zs_write)
+               loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
+
+       if (setup->zs_write)
+               loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
+       if (setup->color_write)
+               loop_body_size += VC4_PACKET_STORE_MS_TILE_BUFFER_SIZE;
+
+       /* We need a VC4_PACKET_TILE_COORDINATES in between each store. */
+       loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE *
+               ((setup->msaa_color_write != NULL) +
+                (setup->msaa_zs_write != NULL) +
+                (setup->color_write != NULL) +
+                (setup->zs_write != NULL) - 1);
+
+       size += xtiles * ytiles * loop_body_size;
+
+       setup->rcl = &vc4_bo_create(dev, size, true)->base;
+       if (!setup->rcl)
+               return -ENOMEM;
+       list_add_tail(&to_vc4_bo(&setup->rcl->base)->unref_head,
+                     &exec->unref_list);
+
+       rcl_u8(setup, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
+       rcl_u32(setup,
+               (setup->color_write ? (setup->color_write->paddr +
+                                      args->color_write.offset) :
+                0));
+       rcl_u16(setup, args->width);
+       rcl_u16(setup, args->height);
+       rcl_u16(setup, args->color_write.bits);
+
+       /* The tile buffer gets cleared when the previous tile is stored.  If
+        * the clear values changed between frames, then the tile buffer has
+        * stale clear values in it, so we have to do a store in None mode (no
+        * writes) so that we trigger the tile buffer clear.
+        */
+       if (args->flags & VC4_SUBMIT_CL_USE_CLEAR_COLOR) {
+               rcl_u8(setup, VC4_PACKET_CLEAR_COLORS);
+               rcl_u32(setup, args->clear_color[0]);
+               rcl_u32(setup, args->clear_color[1]);
+               rcl_u32(setup, args->clear_z);
+               rcl_u8(setup, args->clear_s);
+
+               vc4_tile_coordinates(setup, 0, 0);
+
+               rcl_u8(setup, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
+               rcl_u16(setup, VC4_LOADSTORE_TILE_BUFFER_NONE);
+               rcl_u32(setup, 0); /* no address, since we're in None mode */
+       }
+
+       for (y = min_y_tile; y <= max_y_tile; y++) {
+               for (x = min_x_tile; x <= max_x_tile; x++) {
+                       bool first = (x == min_x_tile && y == min_y_tile);
+                       bool last = (x == max_x_tile && y == max_y_tile);
+
+                       emit_tile(exec, setup, x, y, first, last);
+               }
+       }
+
+       BUG_ON(setup->next_offset != size);
+       exec->ct1ca = setup->rcl->paddr;
+       exec->ct1ea = setup->rcl->paddr + setup->next_offset;
+
+       return 0;
+}
+
+static int vc4_full_res_bounds_check(struct vc4_exec_info *exec,
+                                    struct drm_gem_cma_object *obj,
+                                    struct drm_vc4_submit_rcl_surface *surf)
+{
+       struct drm_vc4_submit_cl *args = exec->args;
+       u32 render_tiles_stride = DIV_ROUND_UP(exec->args->width, 32);
+
+       if (surf->offset > obj->base.size) {
+               DRM_ERROR("surface offset %d > BO size %zd\n",
+                         surf->offset, obj->base.size);
+               return -EINVAL;
+       }
+
+       if ((obj->base.size - surf->offset) / VC4_TILE_BUFFER_SIZE <
+           render_tiles_stride * args->max_y_tile + args->max_x_tile) {
+               DRM_ERROR("MSAA tile %d, %d out of bounds "
+                         "(bo size %zd, offset %d).\n",
+                         args->max_x_tile, args->max_y_tile,
+                         obj->base.size,
+                         surf->offset);
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int vc4_rcl_msaa_surface_setup(struct vc4_exec_info *exec,
+                                     struct drm_gem_cma_object **obj,
+                                     struct drm_vc4_submit_rcl_surface *surf)
+{
+       if (surf->flags != 0 || surf->bits != 0) {
+               DRM_ERROR("MSAA surface had nonzero flags/bits\n");
+               return -EINVAL;
+       }
+
+       if (surf->hindex == ~0)
+               return 0;
+
+       *obj = vc4_use_bo(exec, surf->hindex);
+       if (!*obj)
+               return -EINVAL;
+
+       if (surf->offset & 0xf) {
+               DRM_ERROR("MSAA write must be 16b aligned.\n");
+               return -EINVAL;
+       }
+
+       return vc4_full_res_bounds_check(exec, *obj, surf);
+}
+
+static int vc4_rcl_surface_setup(struct vc4_exec_info *exec,
+                                struct drm_gem_cma_object **obj,
+                                struct drm_vc4_submit_rcl_surface *surf)
+{
+       uint8_t tiling = VC4_GET_FIELD(surf->bits,
+                                      VC4_LOADSTORE_TILE_BUFFER_TILING);
+       uint8_t buffer = VC4_GET_FIELD(surf->bits,
+                                      VC4_LOADSTORE_TILE_BUFFER_BUFFER);
+       uint8_t format = VC4_GET_FIELD(surf->bits,
+                                      VC4_LOADSTORE_TILE_BUFFER_FORMAT);
+       int cpp;
+       int ret;
+
+       if (surf->flags & ~VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+               DRM_ERROR("Extra flags set\n");
+               return -EINVAL;
+       }
+
+       if (surf->hindex == ~0)
+               return 0;
+
+       *obj = vc4_use_bo(exec, surf->hindex);
+       if (!*obj)
+               return -EINVAL;
+
+       if (surf->flags & VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+               if (surf == &exec->args->zs_write) {
+                       DRM_ERROR("general zs write may not be a full-res.\n");
+                       return -EINVAL;
+               }
+
+               if (surf->bits != 0) {
+                       DRM_ERROR("load/store general bits set with "
+                                 "full res load/store.\n");
+                       return -EINVAL;
+               }
+
+               ret = vc4_full_res_bounds_check(exec, *obj, surf);
+               if (!ret)
+                       return ret;
+
+               return 0;
+       }
+
+       if (surf->bits & ~(VC4_LOADSTORE_TILE_BUFFER_TILING_MASK |
+                          VC4_LOADSTORE_TILE_BUFFER_BUFFER_MASK |
+                          VC4_LOADSTORE_TILE_BUFFER_FORMAT_MASK)) {
+               DRM_ERROR("Unknown bits in load/store: 0x%04x\n",
+                         surf->bits);
+               return -EINVAL;
+       }
+
+       if (tiling > VC4_TILING_FORMAT_LT) {
+               DRM_ERROR("Bad tiling format\n");
+               return -EINVAL;
+       }
+
+       if (buffer == VC4_LOADSTORE_TILE_BUFFER_ZS) {
+               if (format != 0) {
+                       DRM_ERROR("No color format should be set for ZS\n");
+                       return -EINVAL;
+               }
+               cpp = 4;
+       } else if (buffer == VC4_LOADSTORE_TILE_BUFFER_COLOR) {
+               switch (format) {
+               case VC4_LOADSTORE_TILE_BUFFER_BGR565:
+               case VC4_LOADSTORE_TILE_BUFFER_BGR565_DITHER:
+                       cpp = 2;
+                       break;
+               case VC4_LOADSTORE_TILE_BUFFER_RGBA8888:
+                       cpp = 4;
+                       break;
+               default:
+                       DRM_ERROR("Bad tile buffer format\n");
+                       return -EINVAL;
+               }
+       } else {
+               DRM_ERROR("Bad load/store buffer %d.\n", buffer);
+               return -EINVAL;
+       }
+
+       if (surf->offset & 0xf) {
+               DRM_ERROR("load/store buffer must be 16b aligned.\n");
+               return -EINVAL;
+       }
+
+       if (!vc4_check_tex_size(exec, *obj, surf->offset, tiling,
+                               exec->args->width, exec->args->height, cpp)) {
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int
+vc4_rcl_render_config_surface_setup(struct vc4_exec_info *exec,
+                                   struct vc4_rcl_setup *setup,
+                                   struct drm_gem_cma_object **obj,
+                                   struct drm_vc4_submit_rcl_surface *surf)
+{
+       uint8_t tiling = VC4_GET_FIELD(surf->bits,
+                                      VC4_RENDER_CONFIG_MEMORY_FORMAT);
+       uint8_t format = VC4_GET_FIELD(surf->bits,
+                                      VC4_RENDER_CONFIG_FORMAT);
+       int cpp;
+
+       if (surf->flags != 0) {
+               DRM_ERROR("No flags supported on render config.\n");
+               return -EINVAL;
+       }
+
+       if (surf->bits & ~(VC4_RENDER_CONFIG_MEMORY_FORMAT_MASK |
+                          VC4_RENDER_CONFIG_FORMAT_MASK |
+                          VC4_RENDER_CONFIG_MS_MODE_4X |
+                          VC4_RENDER_CONFIG_DECIMATE_MODE_4X)) {
+               DRM_ERROR("Unknown bits in render config: 0x%04x\n",
+                         surf->bits);
+               return -EINVAL;
+       }
+
+       if (surf->hindex == ~0)
+               return 0;
+
+       *obj = vc4_use_bo(exec, surf->hindex);
+       if (!*obj)
+               return -EINVAL;
+
+       if (tiling > VC4_TILING_FORMAT_LT) {
+               DRM_ERROR("Bad tiling format\n");
+               return -EINVAL;
+       }
+
+       switch (format) {
+       case VC4_RENDER_CONFIG_FORMAT_BGR565_DITHERED:
+       case VC4_RENDER_CONFIG_FORMAT_BGR565:
+               cpp = 2;
+               break;
+       case VC4_RENDER_CONFIG_FORMAT_RGBA8888:
+               cpp = 4;
+               break;
+       default:
+               DRM_ERROR("Bad tile buffer format\n");
+               return -EINVAL;
+       }
+
+       if (!vc4_check_tex_size(exec, *obj, surf->offset, tiling,
+                               exec->args->width, exec->args->height, cpp)) {
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+int vc4_get_rcl(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+       struct vc4_rcl_setup setup = {0};
+       struct drm_vc4_submit_cl *args = exec->args;
+       bool has_bin = args->bin_cl_size != 0;
+       int ret;
+
+       if (args->min_x_tile > args->max_x_tile ||
+           args->min_y_tile > args->max_y_tile) {
+               DRM_ERROR("Bad render tile set (%d,%d)-(%d,%d)\n",
+                         args->min_x_tile, args->min_y_tile,
+                         args->max_x_tile, args->max_y_tile);
+               return -EINVAL;
+       }
+
+       if (has_bin &&
+           (args->max_x_tile > exec->bin_tiles_x ||
+            args->max_y_tile > exec->bin_tiles_y)) {
+               DRM_ERROR("Render tiles (%d,%d) outside of bin config "
+                         "(%d,%d)\n",
+                         args->max_x_tile, args->max_y_tile,
+                         exec->bin_tiles_x, exec->bin_tiles_y);
+               return -EINVAL;
+       }
+
+       ret = vc4_rcl_render_config_surface_setup(exec, &setup,
+                                                 &setup.color_write,
+                                                 &args->color_write);
+       if (ret)
+               return ret;
+
+       ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read);
+       if (ret)
+               return ret;
+
+       ret = vc4_rcl_surface_setup(exec, &setup.zs_read, &args->zs_read);
+       if (ret)
+               return ret;
+
+       ret = vc4_rcl_surface_setup(exec, &setup.zs_write, &args->zs_write);
+       if (ret)
+               return ret;
+
+       ret = vc4_rcl_msaa_surface_setup(exec, &setup.msaa_color_write,
+                                        &args->msaa_color_write);
+       if (ret)
+               return ret;
+
+       ret = vc4_rcl_msaa_surface_setup(exec, &setup.msaa_zs_write,
+                                        &args->msaa_zs_write);
+       if (ret)
+               return ret;
+
+       /* We shouldn't even have the job submitted to us if there's no
+        * surface to write out.
+        */
+       if (!setup.color_write && !setup.zs_write &&
+           !setup.msaa_color_write && !setup.msaa_zs_write) {
+               DRM_ERROR("RCL requires color or Z/S write\n");
+               return -EINVAL;
+       }
+
+       return vc4_create_rcl_bo(dev, exec, &setup);
+}
diff --git a/drivers/gpu/drm/vc4/vc4_trace.h b/drivers/gpu/drm/vc4/vc4_trace.h
new file mode 100644 (file)
index 0000000..ad7b1ea
--- /dev/null
@@ -0,0 +1,63 @@
+/*
+ * Copyright (C) 2015 Broadcom
+ *
+ * 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.
+ */
+
+#if !defined(_VC4_TRACE_H_) || defined(TRACE_HEADER_MULTI_READ)
+#define _VC4_TRACE_H_
+
+#include <linux/stringify.h>
+#include <linux/types.h>
+#include <linux/tracepoint.h>
+
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM vc4
+#define TRACE_INCLUDE_FILE vc4_trace
+
+TRACE_EVENT(vc4_wait_for_seqno_begin,
+           TP_PROTO(struct drm_device *dev, uint64_t seqno, uint64_t timeout),
+           TP_ARGS(dev, seqno, timeout),
+
+           TP_STRUCT__entry(
+                            __field(u32, dev)
+                            __field(u64, seqno)
+                            __field(u64, timeout)
+                            ),
+
+           TP_fast_assign(
+                          __entry->dev = dev->primary->index;
+                          __entry->seqno = seqno;
+                          __entry->timeout = timeout;
+                          ),
+
+           TP_printk("dev=%u, seqno=%llu, timeout=%llu",
+                     __entry->dev, __entry->seqno, __entry->timeout)
+);
+
+TRACE_EVENT(vc4_wait_for_seqno_end,
+           TP_PROTO(struct drm_device *dev, uint64_t seqno),
+           TP_ARGS(dev, seqno),
+
+           TP_STRUCT__entry(
+                            __field(u32, dev)
+                            __field(u64, seqno)
+                            ),
+
+           TP_fast_assign(
+                          __entry->dev = dev->primary->index;
+                          __entry->seqno = seqno;
+                          ),
+
+           TP_printk("dev=%u, seqno=%llu",
+                     __entry->dev, __entry->seqno)
+);
+
+#endif /* _VC4_TRACE_H_ */
+
+/* This part must be outside protection */
+#undef TRACE_INCLUDE_PATH
+#define TRACE_INCLUDE_PATH .
+#include <trace/define_trace.h>
diff --git a/drivers/gpu/drm/vc4/vc4_trace_points.c b/drivers/gpu/drm/vc4/vc4_trace_points.c
new file mode 100644 (file)
index 0000000..e6278f2
--- /dev/null
@@ -0,0 +1,14 @@
+/*
+ * Copyright (C) 2015 Broadcom
+ *
+ * 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.
+ */
+
+#include "vc4_drv.h"
+
+#ifndef __CHECKER__
+#define CREATE_TRACE_POINTS
+#include "vc4_trace.h"
+#endif
index 040ad0d8b8a11ded79ec8e4f3b33bea7bc39df39..424d515ffcdafae650c9d472a02600db7c0a8485 100644 (file)
@@ -144,6 +144,21 @@ int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
 }
 #endif /* CONFIG_DEBUG_FS */
 
+/*
+ * Asks the firmware to turn on power to the V3D engine.
+ *
+ * This may be doable with just the clocks interface, though this
+ * packet does some other register setup from the firmware, too.
+ */
+int
+vc4_v3d_set_power(struct vc4_dev *vc4, bool on)
+{
+       if (on)
+               return pm_generic_poweroff(&vc4->v3d->pdev->dev);
+       else
+               return pm_generic_resume(&vc4->v3d->pdev->dev);
+}
+
 static void vc4_v3d_init_hw(struct drm_device *dev)
 {
        struct vc4_dev *vc4 = to_vc4_dev(dev);
@@ -161,6 +176,7 @@ static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
        struct drm_device *drm = dev_get_drvdata(master);
        struct vc4_dev *vc4 = to_vc4_dev(drm);
        struct vc4_v3d *v3d = NULL;
+       int ret;
 
        v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
        if (!v3d)
@@ -180,8 +196,20 @@ static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
                return -EINVAL;
        }
 
+       /* Reset the binner overflow address/size at setup, to be sure
+        * we don't reuse an old one.
+        */
+       V3D_WRITE(V3D_BPOA, 0);
+       V3D_WRITE(V3D_BPOS, 0);
+
        vc4_v3d_init_hw(drm);
 
+       ret = drm_irq_install(drm, platform_get_irq(pdev, 0));
+       if (ret) {
+               DRM_ERROR("Failed to install IRQ handler\n");
+               return ret;
+       }
+
        return 0;
 }
 
@@ -191,6 +219,15 @@ static void vc4_v3d_unbind(struct device *dev, struct device *master,
        struct drm_device *drm = dev_get_drvdata(master);
        struct vc4_dev *vc4 = to_vc4_dev(drm);
 
+       drm_irq_uninstall(drm);
+
+       /* Disable the binner's overflow memory address, so the next
+        * driver probe (if any) doesn't try to reuse our old
+        * allocation.
+        */
+       V3D_WRITE(V3D_BPOA, 0);
+       V3D_WRITE(V3D_BPOS, 0);
+
        vc4->v3d = NULL;
 }
 
diff --git a/drivers/gpu/drm/vc4/vc4_validate.c b/drivers/gpu/drm/vc4/vc4_validate.c
new file mode 100644 (file)
index 0000000..0fb5b99
--- /dev/null
@@ -0,0 +1,900 @@
+/*
+ * 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.
+ */
+
+/**
+ * Command list validator for VC4.
+ *
+ * The VC4 has no IOMMU between it and system memory.  So, a user with
+ * access to execute command lists could escalate privilege by
+ * overwriting system memory (drawing to it as a framebuffer) or
+ * reading system memory it shouldn't (reading it as a texture, or
+ * uniform data, or vertex data).
+ *
+ * This validates command lists to ensure that all accesses are within
+ * the bounds of the GEM objects referenced.  It explicitly whitelists
+ * packets, and looks at the offsets in any address fields to make
+ * sure they're constrained within the BOs they reference.
+ *
+ * Note that because of the validation that's happening anyway, this
+ * is where GEM relocation processing happens.
+ */
+
+#include "uapi/drm/vc4_drm.h"
+#include "vc4_drv.h"
+#include "vc4_packet.h"
+
+#define VALIDATE_ARGS \
+       struct vc4_exec_info *exec,                     \
+       void *validated,                                \
+       void *untrusted
+
+/** Return the width in pixels of a 64-byte microtile. */
+static uint32_t
+utile_width(int cpp)
+{
+       switch (cpp) {
+       case 1:
+       case 2:
+               return 8;
+       case 4:
+               return 4;
+       case 8:
+               return 2;
+       default:
+               DRM_ERROR("unknown cpp: %d\n", cpp);
+               return 1;
+       }
+}
+
+/** Return the height in pixels of a 64-byte microtile. */
+static uint32_t
+utile_height(int cpp)
+{
+       switch (cpp) {
+       case 1:
+               return 8;
+       case 2:
+       case 4:
+       case 8:
+               return 4;
+       default:
+               DRM_ERROR("unknown cpp: %d\n", cpp);
+               return 1;
+       }
+}
+
+/**
+ * The texture unit decides what tiling format a particular miplevel is using
+ * this function, so we lay out our miptrees accordingly.
+ */
+static bool
+size_is_lt(uint32_t width, uint32_t height, int cpp)
+{
+       return (width <= 4 * utile_width(cpp) ||
+               height <= 4 * utile_height(cpp));
+}
+
+struct drm_gem_cma_object *
+vc4_use_bo(struct vc4_exec_info *exec, uint32_t hindex)
+{
+       struct drm_gem_cma_object *obj;
+       struct vc4_bo *bo;
+
+       if (hindex >= exec->bo_count) {
+               DRM_ERROR("BO index %d greater than BO count %d\n",
+                         hindex, exec->bo_count);
+               return NULL;
+       }
+       obj = exec->bo[hindex];
+       bo = to_vc4_bo(&obj->base);
+
+       if (bo->validated_shader) {
+               DRM_ERROR("Trying to use shader BO as something other than "
+                         "a shader\n");
+               return NULL;
+       }
+
+       return obj;
+}
+
+static struct drm_gem_cma_object *
+vc4_use_handle(struct vc4_exec_info *exec, uint32_t gem_handles_packet_index)
+{
+       return vc4_use_bo(exec, exec->bo_index[gem_handles_packet_index]);
+}
+
+static bool
+validate_bin_pos(struct vc4_exec_info *exec, void *untrusted, uint32_t pos)
+{
+       /* Note that the untrusted pointer passed to these functions is
+        * incremented past the packet byte.
+        */
+       return (untrusted - 1 == exec->bin_u + pos);
+}
+
+static uint32_t
+gl_shader_rec_size(uint32_t pointer_bits)
+{
+       uint32_t attribute_count = pointer_bits & 7;
+       bool extended = pointer_bits & 8;
+
+       if (attribute_count == 0)
+               attribute_count = 8;
+
+       if (extended)
+               return 100 + attribute_count * 4;
+       else
+               return 36 + attribute_count * 8;
+}
+
+bool
+vc4_check_tex_size(struct vc4_exec_info *exec, struct drm_gem_cma_object *fbo,
+                  uint32_t offset, uint8_t tiling_format,
+                  uint32_t width, uint32_t height, uint8_t cpp)
+{
+       uint32_t aligned_width, aligned_height, stride, size;
+       uint32_t utile_w = utile_width(cpp);
+       uint32_t utile_h = utile_height(cpp);
+
+       /* The shaded vertex format stores signed 12.4 fixed point
+        * (-2048,2047) offsets from the viewport center, so we should
+        * never have a render target larger than 4096.  The texture
+        * unit can only sample from 2048x2048, so it's even more
+        * restricted.  This lets us avoid worrying about overflow in
+        * our math.
+        */
+       if (width > 4096 || height > 4096) {
+               DRM_ERROR("Surface dimesions (%d,%d) too large", width, height);
+               return false;
+       }
+
+       switch (tiling_format) {
+       case VC4_TILING_FORMAT_LINEAR:
+               aligned_width = round_up(width, utile_w);
+               aligned_height = height;
+               break;
+       case VC4_TILING_FORMAT_T:
+               aligned_width = round_up(width, utile_w * 8);
+               aligned_height = round_up(height, utile_h * 8);
+               break;
+       case VC4_TILING_FORMAT_LT:
+               aligned_width = round_up(width, utile_w);
+               aligned_height = round_up(height, utile_h);
+               break;
+       default:
+               DRM_ERROR("buffer tiling %d unsupported\n", tiling_format);
+               return false;
+       }
+
+       stride = aligned_width * cpp;
+       size = stride * aligned_height;
+
+       if (size + offset < size ||
+           size + offset > fbo->base.size) {
+               DRM_ERROR("Overflow in %dx%d (%dx%d) fbo size (%d + %d > %zd)\n",
+                         width, height,
+                         aligned_width, aligned_height,
+                         size, offset, fbo->base.size);
+               return false;
+       }
+
+       return true;
+}
+
+static int
+validate_flush(VALIDATE_ARGS)
+{
+       if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 1)) {
+               DRM_ERROR("Bin CL must end with VC4_PACKET_FLUSH\n");
+               return -EINVAL;
+       }
+       exec->found_flush = true;
+
+       return 0;
+}
+
+static int
+validate_start_tile_binning(VALIDATE_ARGS)
+{
+       if (exec->found_start_tile_binning_packet) {
+               DRM_ERROR("Duplicate VC4_PACKET_START_TILE_BINNING\n");
+               return -EINVAL;
+       }
+       exec->found_start_tile_binning_packet = true;
+
+       if (!exec->found_tile_binning_mode_config_packet) {
+               DRM_ERROR("missing VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int
+validate_increment_semaphore(VALIDATE_ARGS)
+{
+       if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 2)) {
+               DRM_ERROR("Bin CL must end with "
+                         "VC4_PACKET_INCREMENT_SEMAPHORE\n");
+               return -EINVAL;
+       }
+       exec->found_increment_semaphore_packet = true;
+
+       return 0;
+}
+
+static int
+validate_indexed_prim_list(VALIDATE_ARGS)
+{
+       struct drm_gem_cma_object *ib;
+       uint32_t length = *(uint32_t *)(untrusted + 1);
+       uint32_t offset = *(uint32_t *)(untrusted + 5);
+       uint32_t max_index = *(uint32_t *)(untrusted + 9);
+       uint32_t index_size = (*(uint8_t *)(untrusted + 0) >> 4) ? 2 : 1;
+       struct vc4_shader_state *shader_state;
+
+       /* Check overflow condition */
+       if (exec->shader_state_count == 0) {
+               DRM_ERROR("shader state must precede primitives\n");
+               return -EINVAL;
+       }
+       shader_state = &exec->shader_state[exec->shader_state_count - 1];
+
+       if (max_index > shader_state->max_index)
+               shader_state->max_index = max_index;
+
+       ib = vc4_use_handle(exec, 0);
+       if (!ib)
+               return -EINVAL;
+
+       if (offset > ib->base.size ||
+           (ib->base.size - offset) / index_size < length) {
+               DRM_ERROR("IB access overflow (%d + %d*%d > %zd)\n",
+                         offset, length, index_size, ib->base.size);
+               return -EINVAL;
+       }
+
+       *(uint32_t *)(validated + 5) = ib->paddr + offset;
+
+       return 0;
+}
+
+static int
+validate_gl_array_primitive(VALIDATE_ARGS)
+{
+       uint32_t length = *(uint32_t *)(untrusted + 1);
+       uint32_t base_index = *(uint32_t *)(untrusted + 5);
+       uint32_t max_index;
+       struct vc4_shader_state *shader_state;
+
+       /* Check overflow condition */
+       if (exec->shader_state_count == 0) {
+               DRM_ERROR("shader state must precede primitives\n");
+               return -EINVAL;
+       }
+       shader_state = &exec->shader_state[exec->shader_state_count - 1];
+
+       if (length + base_index < length) {
+               DRM_ERROR("primitive vertex count overflow\n");
+               return -EINVAL;
+       }
+       max_index = length + base_index - 1;
+
+       if (max_index > shader_state->max_index)
+               shader_state->max_index = max_index;
+
+       return 0;
+}
+
+static int
+validate_gl_shader_state(VALIDATE_ARGS)
+{
+       uint32_t i = exec->shader_state_count++;
+
+       if (i >= exec->shader_state_size) {
+               DRM_ERROR("More requests for shader states than declared\n");
+               return -EINVAL;
+       }
+
+       exec->shader_state[i].addr = *(uint32_t *)untrusted;
+       exec->shader_state[i].max_index = 0;
+
+       if (exec->shader_state[i].addr & ~0xf) {
+               DRM_ERROR("high bits set in GL shader rec reference\n");
+               return -EINVAL;
+       }
+
+       *(uint32_t *)validated = (exec->shader_rec_p +
+                                 exec->shader_state[i].addr);
+
+       exec->shader_rec_p +=
+               roundup(gl_shader_rec_size(exec->shader_state[i].addr), 16);
+
+       return 0;
+}
+
+static int
+validate_tile_binning_config(VALIDATE_ARGS)
+{
+       struct drm_device *dev = exec->exec_bo->base.dev;
+       struct vc4_bo *tile_bo;
+       uint8_t flags;
+       uint32_t tile_state_size, tile_alloc_size;
+       uint32_t tile_count;
+
+       if (exec->found_tile_binning_mode_config_packet) {
+               DRM_ERROR("Duplicate VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
+               return -EINVAL;
+       }
+       exec->found_tile_binning_mode_config_packet = true;
+
+       exec->bin_tiles_x = *(uint8_t *)(untrusted + 12);
+       exec->bin_tiles_y = *(uint8_t *)(untrusted + 13);
+       tile_count = exec->bin_tiles_x * exec->bin_tiles_y;
+       flags = *(uint8_t *)(untrusted + 14);
+
+       if (exec->bin_tiles_x == 0 ||
+           exec->bin_tiles_y == 0) {
+               DRM_ERROR("Tile binning config of %dx%d too small\n",
+                         exec->bin_tiles_x, exec->bin_tiles_y);
+               return -EINVAL;
+       }
+
+       if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
+                    VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
+               DRM_ERROR("unsupported binning config flags 0x%02x\n", flags);
+               return -EINVAL;
+       }
+
+       /* The tile state data array is 48 bytes per tile, and we put it at
+        * the start of a BO containing both it and the tile alloc.
+        */
+       tile_state_size = 48 * tile_count;
+
+       /* Since the tile alloc array will follow us, align. */
+       exec->tile_alloc_offset = roundup(tile_state_size, 4096);
+
+       *(uint8_t *)(validated + 14) =
+               ((flags & ~(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK |
+                           VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK)) |
+                VC4_BIN_CONFIG_AUTO_INIT_TSDA |
+                VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32,
+                              VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE) |
+                VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128,
+                              VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE));
+
+       /* Initial block size. */
+       tile_alloc_size = 32 * tile_count;
+
+       /*
+        * The initial allocation gets rounded to the next 256 bytes before
+        * the hardware starts fulfilling further allocations.
+        */
+       tile_alloc_size = roundup(tile_alloc_size, 256);
+
+       /* Add space for the extra allocations.  This is what gets used first,
+        * before overflow memory.  It must have at least 4096 bytes, but we
+        * want to avoid overflow memory usage if possible.
+        */
+       tile_alloc_size += 1024 * 1024;
+
+       tile_bo = vc4_bo_create(dev, exec->tile_alloc_offset + tile_alloc_size,
+                               true);
+       exec->tile_bo = &tile_bo->base;
+       if (!exec->tile_bo)
+               return -ENOMEM;
+       list_add_tail(&tile_bo->unref_head, &exec->unref_list);
+
+       /* tile alloc address. */
+       *(uint32_t *)(validated + 0) = (exec->tile_bo->paddr +
+                                       exec->tile_alloc_offset);
+       /* tile alloc size. */
+       *(uint32_t *)(validated + 4) = tile_alloc_size;
+       /* tile state address. */
+       *(uint32_t *)(validated + 8) = exec->tile_bo->paddr;
+
+       return 0;
+}
+
+static int
+validate_gem_handles(VALIDATE_ARGS)
+{
+       memcpy(exec->bo_index, untrusted, sizeof(exec->bo_index));
+       return 0;
+}
+
+#define VC4_DEFINE_PACKET(packet, func) \
+       [packet] = { packet ## _SIZE, #packet, func }
+
+static const struct cmd_info {
+       uint16_t len;
+       const char *name;
+       int (*func)(struct vc4_exec_info *exec, void *validated,
+                   void *untrusted);
+} cmd_info[] = {
+       VC4_DEFINE_PACKET(VC4_PACKET_HALT, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_NOP, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_FLUSH, validate_flush),
+       VC4_DEFINE_PACKET(VC4_PACKET_FLUSH_ALL, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_START_TILE_BINNING,
+                         validate_start_tile_binning),
+       VC4_DEFINE_PACKET(VC4_PACKET_INCREMENT_SEMAPHORE,
+                         validate_increment_semaphore),
+
+       VC4_DEFINE_PACKET(VC4_PACKET_GL_INDEXED_PRIMITIVE,
+                         validate_indexed_prim_list),
+       VC4_DEFINE_PACKET(VC4_PACKET_GL_ARRAY_PRIMITIVE,
+                         validate_gl_array_primitive),
+
+       VC4_DEFINE_PACKET(VC4_PACKET_PRIMITIVE_LIST_FORMAT, NULL),
+
+       VC4_DEFINE_PACKET(VC4_PACKET_GL_SHADER_STATE, validate_gl_shader_state),
+
+       VC4_DEFINE_PACKET(VC4_PACKET_CONFIGURATION_BITS, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_FLAT_SHADE_FLAGS, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_POINT_SIZE, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_LINE_WIDTH, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_RHT_X_BOUNDARY, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_DEPTH_OFFSET, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_CLIP_WINDOW, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_VIEWPORT_OFFSET, NULL),
+       VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_XY_SCALING, NULL),
+       /* Note: The docs say this was also 105, but it was 106 in the
+        * initial userland code drop.
+        */
+       VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_Z_SCALING, NULL),
+
+       VC4_DEFINE_PACKET(VC4_PACKET_TILE_BINNING_MODE_CONFIG,
+                         validate_tile_binning_config),
+
+       VC4_DEFINE_PACKET(VC4_PACKET_GEM_HANDLES, validate_gem_handles),
+};
+
+int
+vc4_validate_bin_cl(struct drm_device *dev,
+                   void *validated,
+                   void *unvalidated,
+                   struct vc4_exec_info *exec)
+{
+       uint32_t len = exec->args->bin_cl_size;
+       uint32_t dst_offset = 0;
+       uint32_t src_offset = 0;
+
+       while (src_offset < len) {
+               void *dst_pkt = validated + dst_offset;
+               void *src_pkt = unvalidated + src_offset;
+               u8 cmd = *(uint8_t *)src_pkt;
+               const struct cmd_info *info;
+
+               if (cmd >= ARRAY_SIZE(cmd_info)) {
+                       DRM_ERROR("0x%08x: packet %d out of bounds\n",
+                                 src_offset, cmd);
+                       return -EINVAL;
+               }
+
+               info = &cmd_info[cmd];
+               if (!info->name) {
+                       DRM_ERROR("0x%08x: packet %d invalid\n",
+                                 src_offset, cmd);
+                       return -EINVAL;
+               }
+
+               if (src_offset + info->len > len) {
+                       DRM_ERROR("0x%08x: packet %d (%s) length 0x%08x "
+                                 "exceeds bounds (0x%08x)\n",
+                                 src_offset, cmd, info->name, info->len,
+                                 src_offset + len);
+                       return -EINVAL;
+               }
+
+               if (cmd != VC4_PACKET_GEM_HANDLES)
+                       memcpy(dst_pkt, src_pkt, info->len);
+
+               if (info->func && info->func(exec,
+                                            dst_pkt + 1,
+                                            src_pkt + 1)) {
+                       DRM_ERROR("0x%08x: packet %d (%s) failed to validate\n",
+                                 src_offset, cmd, info->name);
+                       return -EINVAL;
+               }
+
+               src_offset += info->len;
+               /* GEM handle loading doesn't produce HW packets. */
+               if (cmd != VC4_PACKET_GEM_HANDLES)
+                       dst_offset += info->len;
+
+               /* When the CL hits halt, it'll stop reading anything else. */
+               if (cmd == VC4_PACKET_HALT)
+                       break;
+       }
+
+       exec->ct0ea = exec->ct0ca + dst_offset;
+
+       if (!exec->found_start_tile_binning_packet) {
+               DRM_ERROR("Bin CL missing VC4_PACKET_START_TILE_BINNING\n");
+               return -EINVAL;
+       }
+
+       /* The bin CL must be ended with INCREMENT_SEMAPHORE and FLUSH.  The
+        * semaphore is used to trigger the render CL to start up, and the
+        * FLUSH is what caps the bin lists with
+        * VC4_PACKET_RETURN_FROM_SUB_LIST (so they jump back to the main
+        * render CL when they get called to) and actually triggers the queued
+        * semaphore increment.
+        */
+       if (!exec->found_increment_semaphore_packet || !exec->found_flush) {
+               DRM_ERROR("Bin CL missing VC4_PACKET_INCREMENT_SEMAPHORE + "
+                         "VC4_PACKET_FLUSH\n");
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static bool
+reloc_tex(struct vc4_exec_info *exec,
+         void *uniform_data_u,
+         struct vc4_texture_sample_info *sample,
+         uint32_t texture_handle_index)
+
+{
+       struct drm_gem_cma_object *tex;
+       uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]);
+       uint32_t p1 = *(uint32_t *)(uniform_data_u + sample->p_offset[1]);
+       uint32_t p2 = (sample->p_offset[2] != ~0 ?
+                      *(uint32_t *)(uniform_data_u + sample->p_offset[2]) : 0);
+       uint32_t p3 = (sample->p_offset[3] != ~0 ?
+                      *(uint32_t *)(uniform_data_u + sample->p_offset[3]) : 0);
+       uint32_t *validated_p0 = exec->uniforms_v + sample->p_offset[0];
+       uint32_t offset = p0 & VC4_TEX_P0_OFFSET_MASK;
+       uint32_t miplevels = VC4_GET_FIELD(p0, VC4_TEX_P0_MIPLVLS);
+       uint32_t width = VC4_GET_FIELD(p1, VC4_TEX_P1_WIDTH);
+       uint32_t height = VC4_GET_FIELD(p1, VC4_TEX_P1_HEIGHT);
+       uint32_t cpp, tiling_format, utile_w, utile_h;
+       uint32_t i;
+       uint32_t cube_map_stride = 0;
+       enum vc4_texture_data_type type;
+
+       tex = vc4_use_bo(exec, texture_handle_index);
+       if (!tex)
+               return false;
+
+       if (sample->is_direct) {
+               uint32_t remaining_size = tex->base.size - p0;
+
+               if (p0 > tex->base.size - 4) {
+                       DRM_ERROR("UBO offset greater than UBO size\n");
+                       goto fail;
+               }
+               if (p1 > remaining_size - 4) {
+                       DRM_ERROR("UBO clamp would allow reads "
+                                 "outside of UBO\n");
+                       goto fail;
+               }
+               *validated_p0 = tex->paddr + p0;
+               return true;
+       }
+
+       if (width == 0)
+               width = 2048;
+       if (height == 0)
+               height = 2048;
+
+       if (p0 & VC4_TEX_P0_CMMODE_MASK) {
+               if (VC4_GET_FIELD(p2, VC4_TEX_P2_PTYPE) ==
+                   VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE)
+                       cube_map_stride = p2 & VC4_TEX_P2_CMST_MASK;
+               if (VC4_GET_FIELD(p3, VC4_TEX_P2_PTYPE) ==
+                   VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) {
+                       if (cube_map_stride) {
+                               DRM_ERROR("Cube map stride set twice\n");
+                               goto fail;
+                       }
+
+                       cube_map_stride = p3 & VC4_TEX_P2_CMST_MASK;
+               }
+               if (!cube_map_stride) {
+                       DRM_ERROR("Cube map stride not set\n");
+                       goto fail;
+               }
+       }
+
+       type = (VC4_GET_FIELD(p0, VC4_TEX_P0_TYPE) |
+               (VC4_GET_FIELD(p1, VC4_TEX_P1_TYPE4) << 4));
+
+       switch (type) {
+       case VC4_TEXTURE_TYPE_RGBA8888:
+       case VC4_TEXTURE_TYPE_RGBX8888:
+       case VC4_TEXTURE_TYPE_RGBA32R:
+               cpp = 4;
+               break;
+       case VC4_TEXTURE_TYPE_RGBA4444:
+       case VC4_TEXTURE_TYPE_RGBA5551:
+       case VC4_TEXTURE_TYPE_RGB565:
+       case VC4_TEXTURE_TYPE_LUMALPHA:
+       case VC4_TEXTURE_TYPE_S16F:
+       case VC4_TEXTURE_TYPE_S16:
+               cpp = 2;
+               break;
+       case VC4_TEXTURE_TYPE_LUMINANCE:
+       case VC4_TEXTURE_TYPE_ALPHA:
+       case VC4_TEXTURE_TYPE_S8:
+               cpp = 1;
+               break;
+       case VC4_TEXTURE_TYPE_ETC1:
+       case VC4_TEXTURE_TYPE_BW1:
+       case VC4_TEXTURE_TYPE_A4:
+       case VC4_TEXTURE_TYPE_A1:
+       case VC4_TEXTURE_TYPE_RGBA64:
+       case VC4_TEXTURE_TYPE_YUV422R:
+       default:
+               DRM_ERROR("Texture format %d unsupported\n", type);
+               goto fail;
+       }
+       utile_w = utile_width(cpp);
+       utile_h = utile_height(cpp);
+
+       if (type == VC4_TEXTURE_TYPE_RGBA32R) {
+               tiling_format = VC4_TILING_FORMAT_LINEAR;
+       } else {
+               if (size_is_lt(width, height, cpp))
+                       tiling_format = VC4_TILING_FORMAT_LT;
+               else
+                       tiling_format = VC4_TILING_FORMAT_T;
+       }
+
+       if (!vc4_check_tex_size(exec, tex, offset + cube_map_stride * 5,
+                               tiling_format, width, height, cpp)) {
+               goto fail;
+       }
+
+       /* The mipmap levels are stored before the base of the texture.  Make
+        * sure there is actually space in the BO.
+        */
+       for (i = 1; i <= miplevels; i++) {
+               uint32_t level_width = max(width >> i, 1u);
+               uint32_t level_height = max(height >> i, 1u);
+               uint32_t aligned_width, aligned_height;
+               uint32_t level_size;
+
+               /* Once the levels get small enough, they drop from T to LT. */
+               if (tiling_format == VC4_TILING_FORMAT_T &&
+                   size_is_lt(level_width, level_height, cpp)) {
+                       tiling_format = VC4_TILING_FORMAT_LT;
+               }
+
+               switch (tiling_format) {
+               case VC4_TILING_FORMAT_T:
+                       aligned_width = round_up(level_width, utile_w * 8);
+                       aligned_height = round_up(level_height, utile_h * 8);
+                       break;
+               case VC4_TILING_FORMAT_LT:
+                       aligned_width = round_up(level_width, utile_w);
+                       aligned_height = round_up(level_height, utile_h);
+                       break;
+               default:
+                       aligned_width = round_up(level_width, utile_w);
+                       aligned_height = level_height;
+                       break;
+               }
+
+               level_size = aligned_width * cpp * aligned_height;
+
+               if (offset < level_size) {
+                       DRM_ERROR("Level %d (%dx%d -> %dx%d) size %db "
+                                 "overflowed buffer bounds (offset %d)\n",
+                                 i, level_width, level_height,
+                                 aligned_width, aligned_height,
+                                 level_size, offset);
+                       goto fail;
+               }
+
+               offset -= level_size;
+       }
+
+       *validated_p0 = tex->paddr + p0;
+
+       return true;
+ fail:
+       DRM_INFO("Texture p0 at %d: 0x%08x\n", sample->p_offset[0], p0);
+       DRM_INFO("Texture p1 at %d: 0x%08x\n", sample->p_offset[1], p1);
+       DRM_INFO("Texture p2 at %d: 0x%08x\n", sample->p_offset[2], p2);
+       DRM_INFO("Texture p3 at %d: 0x%08x\n", sample->p_offset[3], p3);
+       return false;
+}
+
+static int
+validate_gl_shader_rec(struct drm_device *dev,
+                      struct vc4_exec_info *exec,
+                      struct vc4_shader_state *state)
+{
+       uint32_t *src_handles;
+       void *pkt_u, *pkt_v;
+       static const uint32_t shader_reloc_offsets[] = {
+               4, /* fs */
+               16, /* vs */
+               28, /* cs */
+       };
+       uint32_t shader_reloc_count = ARRAY_SIZE(shader_reloc_offsets);
+       struct drm_gem_cma_object *bo[shader_reloc_count + 8];
+       uint32_t nr_attributes, nr_relocs, packet_size;
+       int i;
+
+       nr_attributes = state->addr & 0x7;
+       if (nr_attributes == 0)
+               nr_attributes = 8;
+       packet_size = gl_shader_rec_size(state->addr);
+
+       nr_relocs = ARRAY_SIZE(shader_reloc_offsets) + nr_attributes;
+       if (nr_relocs * 4 > exec->shader_rec_size) {
+               DRM_ERROR("overflowed shader recs reading %d handles "
+                         "from %d bytes left\n",
+                         nr_relocs, exec->shader_rec_size);
+               return -EINVAL;
+       }
+       src_handles = exec->shader_rec_u;
+       exec->shader_rec_u += nr_relocs * 4;
+       exec->shader_rec_size -= nr_relocs * 4;
+
+       if (packet_size > exec->shader_rec_size) {
+               DRM_ERROR("overflowed shader recs copying %db packet "
+                         "from %d bytes left\n",
+                         packet_size, exec->shader_rec_size);
+               return -EINVAL;
+       }
+       pkt_u = exec->shader_rec_u;
+       pkt_v = exec->shader_rec_v;
+       memcpy(pkt_v, pkt_u, packet_size);
+       exec->shader_rec_u += packet_size;
+       /* Shader recs have to be aligned to 16 bytes (due to the attribute
+        * flags being in the low bytes), so round the next validated shader
+        * rec address up.  This should be safe, since we've got so many
+        * relocations in a shader rec packet.
+        */
+       BUG_ON(roundup(packet_size, 16) - packet_size > nr_relocs * 4);
+       exec->shader_rec_v += roundup(packet_size, 16);
+       exec->shader_rec_size -= packet_size;
+
+       if (!(*(uint16_t *)pkt_u & VC4_SHADER_FLAG_FS_SINGLE_THREAD)) {
+               DRM_ERROR("Multi-threaded fragment shaders not supported.\n");
+               return -EINVAL;
+       }
+
+       for (i = 0; i < shader_reloc_count; i++) {
+               if (src_handles[i] > exec->bo_count) {
+                       DRM_ERROR("Shader handle %d too big\n", src_handles[i]);
+                       return -EINVAL;
+               }
+
+               bo[i] = exec->bo[src_handles[i]];
+               if (!bo[i])
+                       return -EINVAL;
+       }
+       for (i = shader_reloc_count; i < nr_relocs; i++) {
+               bo[i] = vc4_use_bo(exec, src_handles[i]);
+               if (!bo[i])
+                       return -EINVAL;
+       }
+
+       for (i = 0; i < shader_reloc_count; i++) {
+               struct vc4_validated_shader_info *validated_shader;
+               uint32_t o = shader_reloc_offsets[i];
+               uint32_t src_offset = *(uint32_t *)(pkt_u + o);
+               uint32_t *texture_handles_u;
+               void *uniform_data_u;
+               uint32_t tex;
+
+               *(uint32_t *)(pkt_v + o) = bo[i]->paddr + src_offset;
+
+               if (src_offset != 0) {
+                       DRM_ERROR("Shaders must be at offset 0 of "
+                                 "the BO.\n");
+                       return -EINVAL;
+               }
+
+               validated_shader = to_vc4_bo(&bo[i]->base)->validated_shader;
+               if (!validated_shader)
+                       return -EINVAL;
+
+               if (validated_shader->uniforms_src_size >
+                   exec->uniforms_size) {
+                       DRM_ERROR("Uniforms src buffer overflow\n");
+                       return -EINVAL;
+               }
+
+               texture_handles_u = exec->uniforms_u;
+               uniform_data_u = (texture_handles_u +
+                                 validated_shader->num_texture_samples);
+
+               memcpy(exec->uniforms_v, uniform_data_u,
+                      validated_shader->uniforms_size);
+
+               for (tex = 0;
+                    tex < validated_shader->num_texture_samples;
+                    tex++) {
+                       if (!reloc_tex(exec,
+                                      uniform_data_u,
+                                      &validated_shader->texture_samples[tex],
+                                      texture_handles_u[tex])) {
+                               return -EINVAL;
+                       }
+               }
+
+               *(uint32_t *)(pkt_v + o + 4) = exec->uniforms_p;
+
+               exec->uniforms_u += validated_shader->uniforms_src_size;
+               exec->uniforms_v += validated_shader->uniforms_size;
+               exec->uniforms_p += validated_shader->uniforms_size;
+       }
+
+       for (i = 0; i < nr_attributes; i++) {
+               struct drm_gem_cma_object *vbo =
+                       bo[ARRAY_SIZE(shader_reloc_offsets) + i];
+               uint32_t o = 36 + i * 8;
+               uint32_t offset = *(uint32_t *)(pkt_u + o + 0);
+               uint32_t attr_size = *(uint8_t *)(pkt_u + o + 4) + 1;
+               uint32_t stride = *(uint8_t *)(pkt_u + o + 5);
+               uint32_t max_index;
+
+               if (state->addr & 0x8)
+                       stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff;
+
+               if (vbo->base.size < offset ||
+                   vbo->base.size - offset < attr_size) {
+                       DRM_ERROR("BO offset overflow (%d + %d > %d)\n",
+                                 offset, attr_size, vbo->base.size);
+                       return -EINVAL;
+               }
+
+               if (stride != 0) {
+                       max_index = ((vbo->base.size - offset - attr_size) /
+                                    stride);
+                       if (state->max_index > max_index) {
+                               DRM_ERROR("primitives use index %d out of "
+                                         "supplied %d\n",
+                                         state->max_index, max_index);
+                               return -EINVAL;
+                       }
+               }
+
+               *(uint32_t *)(pkt_v + o) = vbo->paddr + offset;
+       }
+
+       return 0;
+}
+
+int
+vc4_validate_shader_recs(struct drm_device *dev,
+                        struct vc4_exec_info *exec)
+{
+       uint32_t i;
+       int ret = 0;
+
+       for (i = 0; i < exec->shader_state_count; i++) {
+               ret = validate_gl_shader_rec(dev, exec, &exec->shader_state[i]);
+               if (ret)
+                       return ret;
+       }
+
+       return ret;
+}
index 74de18416be9b566a6b15bb0c74407850a963de4..fe4161bc93ae03c2dd537f34ce88154bd49eab62 100644 (file)
 
 #include "drm.h"
 
+#define DRM_VC4_SUBMIT_CL                         0x00
+#define DRM_VC4_WAIT_SEQNO                        0x01
+#define DRM_VC4_WAIT_BO                           0x02
 #define DRM_VC4_CREATE_BO                         0x03
 #define DRM_VC4_MMAP_BO                           0x04
 #define DRM_VC4_CREATE_SHADER_BO                  0x05
 
+#define DRM_IOCTL_VC4_SUBMIT_CL           DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_SUBMIT_CL, struct drm_vc4_submit_cl)
+#define DRM_IOCTL_VC4_WAIT_SEQNO          DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_WAIT_SEQNO, struct drm_vc4_wait_seqno)
+#define DRM_IOCTL_VC4_WAIT_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_WAIT_BO, struct drm_vc4_wait_bo)
 #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_submit_rcl_surface {
+       __u32 hindex; /* Handle index, or ~0 if not present. */
+       __u32 offset; /* Offset to start of buffer. */
+       /*
+        * Bits for either render config (color_write) or load/store packet.
+        * Bits should all be 0 for MSAA load/stores.
+        */
+       __u16 bits;
+
+#define VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES                (1 << 0)
+       __u16 flags;
+};
+
+/**
+ * struct drm_vc4_submit_cl - ioctl argument for submitting commands to the 3D
+ * engine.
+ *
+ * Drivers typically use GPU BOs to store batchbuffers / command lists and
+ * their associated state.  However, because the VC4 lacks an MMU, we have to
+ * do validation of memory accesses by the GPU commands.  If we were to store
+ * our commands in BOs, we'd need to do uncached readback from them to do the
+ * validation process, which is too expensive.  Instead, userspace accumulates
+ * commands and associated state in plain memory, then the kernel copies the
+ * data to its own address space, and then validates and stores it in a GPU
+ * BO.
+ */
+struct drm_vc4_submit_cl {
+       /* Pointer to the binner command list.
+        *
+        * This is the first set of commands executed, which runs the
+        * coordinate shader to determine where primitives land on the screen,
+        * then writes out the state updates and draw calls necessary per tile
+        * to the tile allocation BO.
+        */
+       __u64 bin_cl;
+
+       /* Pointer to the shader records.
+        *
+        * Shader records are the structures read by the hardware that contain
+        * pointers to uniforms, shaders, and vertex attributes.  The
+        * reference to the shader record has enough information to determine
+        * how many pointers are necessary (fixed number for shaders/uniforms,
+        * and an attribute count), so those BO indices into bo_handles are
+        * just stored as __u32s before each shader record passed in.
+        */
+       __u64 shader_rec;
+
+       /* Pointer to uniform data and texture handles for the textures
+        * referenced by the shader.
+        *
+        * For each shader state record, there is a set of uniform data in the
+        * order referenced by the record (FS, VS, then CS).  Each set of
+        * uniform data has a __u32 index into bo_handles per texture
+        * sample operation, in the order the QPU_W_TMUn_S writes appear in
+        * the program.  Following the texture BO handle indices is the actual
+        * uniform data.
+        *
+        * The individual uniform state blocks don't have sizes passed in,
+        * because the kernel has to determine the sizes anyway during shader
+        * code validation.
+        */
+       __u64 uniforms;
+       __u64 bo_handles;
+
+       /* Size in bytes of the binner command list. */
+       __u32 bin_cl_size;
+       /* Size in bytes of the set of shader records. */
+       __u32 shader_rec_size;
+       /* Number of shader records.
+        *
+        * This could just be computed from the contents of shader_records and
+        * the address bits of references to them from the bin CL, but it
+        * keeps the kernel from having to resize some allocations it makes.
+        */
+       __u32 shader_rec_count;
+       /* Size in bytes of the uniform state. */
+       __u32 uniforms_size;
+
+       /* Number of BO handles passed in (size is that times 4). */
+       __u32 bo_handle_count;
+
+       /* RCL setup: */
+       __u16 width;
+       __u16 height;
+       __u8 min_x_tile;
+       __u8 min_y_tile;
+       __u8 max_x_tile;
+       __u8 max_y_tile;
+       struct drm_vc4_submit_rcl_surface color_read;
+       struct drm_vc4_submit_rcl_surface color_write;
+       struct drm_vc4_submit_rcl_surface zs_read;
+       struct drm_vc4_submit_rcl_surface zs_write;
+       struct drm_vc4_submit_rcl_surface msaa_color_write;
+       struct drm_vc4_submit_rcl_surface msaa_zs_write;
+       __u32 clear_color[2];
+       __u32 clear_z;
+       __u8 clear_s;
+
+       __u32 pad:24;
+
+#define VC4_SUBMIT_CL_USE_CLEAR_COLOR                  (1 << 0)
+       __u32 flags;
+
+       /* Returned value of the seqno of this render job (for the
+        * wait ioctl).
+        */
+       __u64 seqno;
+};
+
+/**
+ * struct drm_vc4_wait_seqno - ioctl argument for waiting for
+ * DRM_VC4_SUBMIT_CL completion using its returned seqno.
+ *
+ * timeout_ns is the timeout in nanoseconds, where "0" means "don't
+ * block, just return the status."
+ */
+struct drm_vc4_wait_seqno {
+       __u64 seqno;
+       __u64 timeout_ns;
+};
+
+/**
+ * struct drm_vc4_wait_bo - ioctl argument for waiting for
+ * completion of the last DRM_VC4_SUBMIT_CL on a BO.
+ *
+ * This is useful for cases where multiple processes might be
+ * rendering to a BO and you want to wait for all rendering to be
+ * completed.
+ */
+struct drm_vc4_wait_bo {
+       __u32 handle;
+       __u32 pad;
+       __u64 timeout_ns;
+};
+
 /**
  * struct drm_vc4_create_bo - ioctl argument for creating VC4 BOs.
  *