intel: move drm calls to exec buffers to libdrm_intel.

This avoids duplicating the effort in 3 places.  Also, added emit/wait fence
callbacks back in bufmgr_fake since we need it for non-drm 2d.  Sigh.

5 files changed, 168 insertions, 95 deletions

diff --git a/libdrm/intel/intel_bufmgr.c b/libdrm/intel/intel_bufmgr.c
index 5a9cdbe6..45b691d6 100644
--- a/libdrm/intel/intel_bufmgr.c
+++ b/libdrm/intel/intel_bufmgr.c
@@ -123,14 +123,12 @@ dri_bufmgr_destroy(dri_bufmgr *bufmgr)
    bufmgr->destroy(bufmgr);
 }
 
-void *dri_process_relocs(dri_bo *batch_buf)
-{
-   return batch_buf->bufmgr->process_relocs(batch_buf);
-}
-
-void dri_post_submit(dri_bo *batch_buf)
+int
+dri_bo_exec(dri_bo *bo, int used,
+	    drm_clip_rect_t *cliprects, int num_cliprects,
+	    int DR4)
 {
-   batch_buf->bufmgr->post_submit(batch_buf);
+   return bo->bufmgr->bo_exec(bo, used, cliprects, num_cliprects, DR4);
 }
 
 void
diff --git a/libdrm/intel/intel_bufmgr.h b/libdrm/intel/intel_bufmgr.h
index 57a9e33a..c44d596b 100644
--- a/libdrm/intel/intel_bufmgr.h
+++ b/libdrm/intel/intel_bufmgr.h
@@ -76,10 +76,9 @@ void dri_bo_wait_rendering(dri_bo *bo);
 
 void dri_bufmgr_set_debug(dri_bufmgr *bufmgr, int enable_debug);
 void dri_bufmgr_destroy(dri_bufmgr *bufmgr);
-
-void *dri_process_relocs(dri_bo *batch_buf);
-void dri_post_process_relocs(dri_bo *batch_buf);
-void dri_post_submit(dri_bo *batch_buf);
+int dri_bo_exec(dri_bo *bo, int used,
+		drm_clip_rect_t *cliprects, int num_cliprects,
+		int DR4);
 int dri_bufmgr_check_aperture_space(dri_bo **bo_array, int count);
 
 int dri_bo_emit_reloc(dri_bo *reloc_buf,
@@ -103,6 +102,16 @@ dri_bufmgr *intel_bufmgr_fake_init(int fd,
 				   volatile unsigned int *last_dispatch);
 void intel_bufmgr_fake_set_last_dispatch(dri_bufmgr *bufmgr,
 					 volatile unsigned int *last_dispatch);
+void intel_bufmgr_fake_set_exec_callback(dri_bufmgr *bufmgr,
+					 int (*exec)(dri_bo *bo,
+						     unsigned int used,
+						     void *priv),
+					 void *priv);
+void intel_bufmgr_fake_set_fence_callback(dri_bufmgr *bufmgr,
+					  unsigned int (*emit)(void *priv),
+					  void (*wait)(unsigned int fence,
+						       void *priv),
+					  void *priv);
 dri_bo *intel_bo_fake_alloc_static(dri_bufmgr *bufmgr, const char *name,
 				   unsigned long offset, unsigned long size,
 				   void *virtual);
diff --git a/libdrm/intel/intel_bufmgr_fake.c b/libdrm/intel/intel_bufmgr_fake.c
index 47629033..f9e1cd12 100644
--- a/libdrm/intel/intel_bufmgr_fake.c
+++ b/libdrm/intel/intel_bufmgr_fake.c
@@ -136,6 +136,31 @@ typedef struct _bufmgr_fake {
    unsigned need_fence:1;
    int thrashing;
 
+   /**
+    * Driver callback to emit a fence, returning the cookie.
+    *
+    * This allows the driver to hook in a replacement for the DRM usage in
+    * bufmgr_fake.
+    *
+    * Currently, this also requires that a write flush be emitted before
+    * emitting the fence, but this should change.
+    */
+   unsigned int (*fence_emit)(void *private);
+   /** Driver callback to wait for a fence cookie to have passed. */
+   void (*fence_wait)(unsigned int fence, void *private);
+   void *fence_priv;
+
+   /**
+    * Driver callback to execute a buffer.
+    *
+    * This allows the driver to hook in a replacement for the DRM usage in
+    * bufmgr_fake.
+    */
+   int (*exec)(dri_bo *bo, unsigned int used, void *priv);
+   void *exec_priv;
+
+   /** Driver-supplied argument to driver callbacks */
+   void *driver_priv;
    /* Pointer to kernel-updated sarea data for the last completed user irq */
    volatile unsigned int *last_dispatch;
 
@@ -205,12 +230,28 @@ static int FENCE_LTE( unsigned a, unsigned b )
    return 0;
 }
 
+void intel_bufmgr_fake_set_fence_callback(dri_bufmgr *bufmgr,
+					  unsigned int (*emit)(void *priv),
+					  void (*wait)(unsigned int fence,
+						       void *priv),
+					  void *priv)
+{
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr;
+
+   bufmgr_fake->fence_emit = emit;
+   bufmgr_fake->fence_wait = wait;
+   bufmgr_fake->fence_priv = priv;
+}
+
 static unsigned int
 _fence_emit_internal(dri_bufmgr_fake *bufmgr_fake)
 {
    struct drm_i915_irq_emit ie;
    int ret, seq = 1;
 
+   if (bufmgr_fake->fence_emit != NULL)
+      return bufmgr_fake->fence_emit(bufmgr_fake->fence_priv);
+
    ie.irq_seq = &seq;
    ret = drmCommandWriteRead(bufmgr_fake->fd, DRM_I915_IRQ_EMIT,
 			     &ie, sizeof(ie));
@@ -243,6 +284,11 @@ _fence_wait_internal(dri_bufmgr_fake *bufmgr_fake, unsigned int cookie)
    unsigned int last_dispatch;
    int ret;
 
+   if (bufmgr_fake->fence_wait != NULL) {
+      bufmgr_fake->fence_wait(cookie, bufmgr_fake->fence_priv);
+      return;
+   }
+
    DBG("wait 0x%08x\n", iw.irq_seq);
 
    /* The kernel implementation of IRQ_WAIT is broken for wraparound, and has
@@ -1092,38 +1138,6 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo)
    return dri_fake_bo_validate(bo);
 }
 
-static void *
-dri_fake_process_relocs(dri_bo *batch_buf)
-{
-   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)batch_buf->bufmgr;
-   dri_bo_fake *batch_fake = (dri_bo_fake *)batch_buf;
-   int ret;
-   int retry_count = 0;
-
-   bufmgr_fake->performed_rendering = 0;
-
-   dri_fake_calculate_domains(batch_buf);
-
-   batch_fake->read_domains = I915_GEM_DOMAIN_COMMAND;
-
-   /* we've ran out of RAM so blow the whole lot away and retry */
- restart:
-   ret = dri_fake_reloc_and_validate_buffer(batch_buf);
-   if (bufmgr_fake->fail == 1) {
-      if (retry_count == 0) {
-         retry_count++;
-         dri_fake_kick_all(bufmgr_fake);
-         bufmgr_fake->fail = 0;
-         goto restart;
-      } else /* dump out the memory here */
-         mmDumpMemInfo(bufmgr_fake->heap);
-   }
-
-   assert(ret == 0);
-
-   return NULL;
-}
-
 static void
 dri_bo_fake_post_submit(dri_bo *bo)
 {
@@ -1150,12 +1164,74 @@ dri_bo_fake_post_submit(dri_bo *bo)
 }
 
 
-static void
-dri_fake_post_submit(dri_bo *batch_buf)
+void intel_bufmgr_fake_set_exec_callback(dri_bufmgr *bufmgr,
+					 int (*exec)(dri_bo *bo,
+						     unsigned int used,
+						     void *priv),
+					 void *priv)
+{
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bufmgr;
+
+   bufmgr_fake->exec = exec;
+   bufmgr_fake->exec_priv = exec;
+}
+
+static int
+dri_fake_bo_exec(dri_bo *bo, int used,
+		 drm_clip_rect_t *cliprects, int num_cliprects,
+		 int DR4)
 {
-   dri_fake_fence_validated(batch_buf->bufmgr);
+   dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)bo->bufmgr;
+   dri_bo_fake *batch_fake = (dri_bo_fake *)bo;
+   struct drm_i915_batchbuffer batch;
+   int ret;
+   int retry_count = 0;
+
+   bufmgr_fake->performed_rendering = 0;
+
+   dri_fake_calculate_domains(bo);
 
-   dri_bo_fake_post_submit(batch_buf);
+   batch_fake->read_domains = I915_GEM_DOMAIN_COMMAND;
+
+   /* we've ran out of RAM so blow the whole lot away and retry */
+ restart:
+   ret = dri_fake_reloc_and_validate_buffer(bo);
+   if (bufmgr_fake->fail == 1) {
+      if (retry_count == 0) {
+         retry_count++;
+         dri_fake_kick_all(bufmgr_fake);
+         bufmgr_fake->fail = 0;
+         goto restart;
+      } else /* dump out the memory here */
+         mmDumpMemInfo(bufmgr_fake->heap);
+   }
+
+   assert(ret == 0);
+
+   if (bufmgr_fake->exec != NULL) {
+      int ret = bufmgr_fake->exec(bo, used, bufmgr_fake->exec_priv);
+      if (ret != 0)
+	 return ret;
+   } else {
+      batch.start = bo->offset;
+      batch.used = used;
+      batch.cliprects = cliprects;
+      batch.num_cliprects = num_cliprects;
+      batch.DR1 = 0;
+      batch.DR4 = DR4;
+
+      if (drmCommandWrite(bufmgr_fake->fd, DRM_I915_BATCHBUFFER, &batch,
+			  sizeof(batch))) {
+	 drmMsg("DRM_I915_BATCHBUFFER: %d\n", -errno);
+	 return -errno;
+      }
+   }
+
+   dri_fake_fence_validated(bo->bufmgr);
+
+   dri_bo_fake_post_submit(bo);
+
+   return 0;
 }
 
 /**
@@ -1264,8 +1340,7 @@ intel_bufmgr_fake_init(int fd,
    bufmgr_fake->bufmgr.bo_wait_rendering = dri_fake_bo_wait_rendering;
    bufmgr_fake->bufmgr.bo_emit_reloc = dri_fake_emit_reloc;
    bufmgr_fake->bufmgr.destroy = dri_fake_destroy;
-   bufmgr_fake->bufmgr.process_relocs = dri_fake_process_relocs;
-   bufmgr_fake->bufmgr.post_submit = dri_fake_post_submit;
+   bufmgr_fake->bufmgr.bo_exec = dri_fake_bo_exec;
    bufmgr_fake->bufmgr.check_aperture_space = dri_fake_check_aperture_space;
    bufmgr_fake->bufmgr.debug = 0;
 
diff --git a/libdrm/intel/intel_bufmgr_gem.c b/libdrm/intel/intel_bufmgr_gem.c
index 43d2d986..44152728 100644
--- a/libdrm/intel/intel_bufmgr_gem.c
+++ b/libdrm/intel/intel_bufmgr_gem.c
@@ -87,8 +87,6 @@ typedef struct _dri_bufmgr_gem {
 
     /** Array of lists of cached gem objects of power-of-two sizes */
     struct dri_gem_bo_bucket cache_bucket[INTEL_GEM_BO_BUCKETS];
-
-    struct drm_i915_gem_execbuffer exec_arg;
 } dri_bufmgr_gem;
 
 struct _dri_bo_gem {
@@ -706,27 +704,6 @@ dri_gem_bo_process_reloc(dri_bo *bo)
     }
 }
 
-static void *
-dri_gem_process_reloc(dri_bo *batch_buf)
-{
-    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *) batch_buf->bufmgr;
-
-    /* Update indices and set up the validate list. */
-    dri_gem_bo_process_reloc(batch_buf);
-
-    /* Add the batch buffer to the validation list.  There are no relocations
-     * pointing to it.
-     */
-    intel_add_validate_buffer(batch_buf);
-
-    bufmgr_gem->exec_arg.buffers_ptr = (uintptr_t)bufmgr_gem->exec_objects;
-    bufmgr_gem->exec_arg.buffer_count = bufmgr_gem->exec_count;
-    bufmgr_gem->exec_arg.batch_start_offset = 0;
-    bufmgr_gem->exec_arg.batch_len = 0;	/* written in intel_exec_ioctl */
-
-    return &bufmgr_gem->exec_arg;
-}
-
 static void
 intel_update_buffer_offsets (dri_bufmgr_gem *bufmgr_gem)
 {
@@ -746,11 +723,35 @@ intel_update_buffer_offsets (dri_bufmgr_gem *bufmgr_gem)
     }
 }
 
-static void
-dri_gem_post_submit(dri_bo *batch_buf)
+static int
+dri_gem_bo_exec(dri_bo *bo, int used,
+		drm_clip_rect_t *cliprects, int num_cliprects,
+		int DR4)
 {
-    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)batch_buf->bufmgr;
-    int i;
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+    struct drm_i915_gem_execbuffer execbuf;
+    int ret, i;
+
+    /* Update indices and set up the validate list. */
+    dri_gem_bo_process_reloc(bo);
+
+    /* Add the batch buffer to the validation list.  There are no relocations
+     * pointing to it.
+     */
+    intel_add_validate_buffer(bo);
+
+    execbuf.buffers_ptr = (uintptr_t)bufmgr_gem->exec_objects;
+    execbuf.buffer_count = bufmgr_gem->exec_count;
+    execbuf.batch_start_offset = 0;
+    execbuf.batch_len = used;
+    execbuf.cliprects_ptr = (uintptr_t)cliprects;
+    execbuf.num_cliprects = num_cliprects;
+    execbuf.DR1 = 0;
+    execbuf.DR4 = DR4;
+
+    do {
+	ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_I915_GEM_EXECBUFFER, &execbuf);
+    } while (ret == -EAGAIN);
 
     intel_update_buffer_offsets (bufmgr_gem);
 
@@ -770,6 +771,8 @@ dri_gem_post_submit(dri_bo *batch_buf)
 	bufmgr_gem->exec_bos[i] = NULL;
     }
     bufmgr_gem->exec_count = 0;
+
+    return 0;
 }
 
 static int
@@ -913,9 +916,8 @@ intel_bufmgr_gem_init(int fd, int batch_size)
     bufmgr_gem->bufmgr.bo_unpin = dri_gem_bo_unpin;
     bufmgr_gem->bufmgr.bo_set_tiling = dri_gem_bo_set_tiling;
     bufmgr_gem->bufmgr.bo_flink = dri_gem_bo_flink;
+    bufmgr_gem->bufmgr.bo_exec = dri_gem_bo_exec;
     bufmgr_gem->bufmgr.destroy = dri_bufmgr_gem_destroy;
-    bufmgr_gem->bufmgr.process_relocs = dri_gem_process_reloc;
-    bufmgr_gem->bufmgr.post_submit = dri_gem_post_submit;
     bufmgr_gem->bufmgr.debug = 0;
     bufmgr_gem->bufmgr.check_aperture_space = dri_gem_check_aperture_space;
     /* Initialize the linked lists for BO reuse cache. */
diff --git a/libdrm/intel/intel_bufmgr_priv.h b/libdrm/intel/intel_bufmgr_priv.h
index 8a5741fa..7f39bfc2 100644
--- a/libdrm/intel/intel_bufmgr_priv.h
+++ b/libdrm/intel/intel_bufmgr_priv.h
@@ -103,23 +103,6 @@ struct _dri_bufmgr {
     */
    void (*destroy)(dri_bufmgr *bufmgr);
 
-   /**
-    * Processes the relocations, either in userland or by converting the list
-    * for use in batchbuffer submission.
-    *
-    * Kernel-based implementations will return a pointer to the arguments
-    * to be handed with batchbuffer submission to the kernel.  The userland
-    * implementation performs the buffer validation and emits relocations
-    * into them the appopriate order.
-    *
-    * \param batch_buf buffer at the root of the tree of relocations
-    * \return argument to be completed and passed to the execbuffers ioctl
-    *   (if any).
-    */
-   void *(*process_relocs)(dri_bo *batch_buf);
-
-   void (*post_submit)(dri_bo *batch_buf);
-
     /**
      * Add relocation entry in reloc_buf, which will be updated with the
      * target buffer's real offset on on command submission.
@@ -140,6 +123,12 @@ struct _dri_bufmgr {
     int (*bo_emit_reloc)(dri_bo *reloc_buf,
 			 uint32_t read_domains, uint32_t write_domain,
 			 uint32_t delta, uint32_t offset, dri_bo *target);
+
+    /** Executes the command buffer pointed to by bo. */
+    int (*bo_exec)(dri_bo *bo, int used,
+		   drm_clip_rect_t *cliprects, int num_cliprects,
+		   int DR4);
+
     /**
      * Pin a buffer to the aperture and fix the offset until unpinned
      *