1 files changed, 327 insertions, 30 deletions

diff --git a/src/mesa/drivers/dri/intel/intel_buffer_objects.c b/src/mesa/drivers/dri/intel/intel_buffer_objects.c
index 60d7bb3770..c55c5c426e 100644
--- a/src/mesa/drivers/dri/intel/intel_buffer_objects.c
+++ b/src/mesa/drivers/dri/intel/intel_buffer_objects.c
@@ -28,16 +28,18 @@
 
 #include "main/imports.h"
 #include "main/mtypes.h"
+#include "main/macros.h"
 #include "main/bufferobj.h"
 
 #include "intel_context.h"
+#include "intel_blit.h"
 #include "intel_buffer_objects.h"
 #include "intel_batchbuffer.h"
 #include "intel_regions.h"
 
-static GLboolean intel_bufferobj_unmap(GLcontext * ctx,
-				       GLenum target,
-				       struct gl_buffer_object *obj);
+static GLboolean
+intel_bufferobj_unmap(GLcontext * ctx,
+                      GLenum target, struct gl_buffer_object *obj);
 
 /** Allocates a new dri_bo to store the data for the buffer object. */
 static void
@@ -105,11 +107,13 @@ intel_bufferobj_free(GLcontext * ctx, struct gl_buffer_object *obj)
    assert(intel_obj);
 
    /* Buffer objects are automatically unmapped when deleting according
-    * to the spec.
+    * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
+    * (though it does if you call glDeleteBuffers)
     */
    if (obj->Pointer)
       intel_bufferobj_unmap(ctx, 0, obj);
 
+   _mesa_free(intel_obj->sys_buffer);
    if (intel_obj->region) {
       intel_bufferobj_release_region(intel, intel_obj);
    }
@@ -126,9 +130,10 @@ intel_bufferobj_free(GLcontext * ctx, struct gl_buffer_object *obj)
  * Allocate space for and store data in a buffer object.  Any data that was
  * previously stored in the buffer object is lost.  If data is NULL,
  * memory will be allocated, but no copy will occur.
- * Called via glBufferDataARB().
+ * Called via ctx->Driver.BufferData().
+ * \return GL_TRUE for success, GL_FALSE if out of memory
  */
-static void
+static GLboolean
 intel_bufferobj_data(GLcontext * ctx,
                      GLenum target,
                      GLsizeiptrARB size,
@@ -141,11 +146,7 @@ intel_bufferobj_data(GLcontext * ctx,
    intel_obj->Base.Size = size;
    intel_obj->Base.Usage = usage;
 
-   /* Buffer objects are automatically unmapped when creating new data buffers
-    * according to the spec.
-    */
-   if (obj->Pointer)
-      intel_bufferobj_unmap(ctx, 0, obj);
+   assert(!obj->Pointer); /* Mesa should have unmapped it */
 
    if (intel_obj->region)
       intel_bufferobj_release_region(intel, intel_obj);
@@ -154,12 +155,32 @@ intel_bufferobj_data(GLcontext * ctx,
       dri_bo_unreference(intel_obj->buffer);
       intel_obj->buffer = NULL;
    }
+   _mesa_free(intel_obj->sys_buffer);
+   intel_obj->sys_buffer = NULL;
+
    if (size != 0) {
+#ifdef I915
+      /* On pre-965, stick VBOs in system memory, as we're always doing swtnl
+       * with their contents anyway.
+       */
+      if (target == GL_ARRAY_BUFFER || target == GL_ELEMENT_ARRAY_BUFFER) {
+	 intel_obj->sys_buffer = _mesa_malloc(size);
+	 if (intel_obj->sys_buffer != NULL) {
+	    if (data != NULL)
+	       memcpy(intel_obj->sys_buffer, data, size);
+	    return GL_TRUE;
+	 }
+      }
+#endif
       intel_bufferobj_alloc_buffer(intel, intel_obj);
+      if (!intel_obj->buffer)
+         return GL_FALSE;
 
       if (data != NULL)
 	 dri_bo_subdata(intel_obj->buffer, 0, size, data);
    }
+
+   return GL_TRUE;
 }
 
 
@@ -184,7 +205,10 @@ intel_bufferobj_subdata(GLcontext * ctx,
    if (intel_obj->region)
       intel_bufferobj_cow(intel, intel_obj);
 
-   dri_bo_subdata(intel_obj->buffer, offset, size, data);
+   if (intel_obj->sys_buffer)
+      memcpy((char *)intel_obj->sys_buffer + offset, data, size);
+   else
+      dri_bo_subdata(intel_obj->buffer, offset, size, data);
 }
 
 
@@ -216,11 +240,21 @@ intel_bufferobj_map(GLcontext * ctx,
 {
    struct intel_context *intel = intel_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
+   GLboolean read_only = (access == GL_READ_ONLY_ARB);
+   GLboolean write_only = (access == GL_WRITE_ONLY_ARB);
 
-   /* XXX: Translate access to flags arg below:
-    */
    assert(intel_obj);
 
+   if (intel_obj->sys_buffer) {
+      obj->Pointer = intel_obj->sys_buffer;
+      return obj->Pointer;
+   }
+
+   /* Flush any existing batchbuffer that might have written to this
+    * buffer.
+    */
+   intelFlush(ctx);
+
    if (intel_obj->region)
       intel_bufferobj_cow(intel, intel_obj);
 
@@ -229,27 +263,214 @@ intel_bufferobj_map(GLcontext * ctx,
       return NULL;
    }
 
-   dri_bo_map(intel_obj->buffer, GL_TRUE);
+   if (write_only && intel->intelScreen->kernel_exec_fencing) {
+      drm_intel_gem_bo_map_gtt(intel_obj->buffer);
+      intel_obj->mapped_gtt = GL_TRUE;
+   } else {
+      drm_intel_bo_map(intel_obj->buffer, !read_only);
+      intel_obj->mapped_gtt = GL_FALSE;
+   }
+
    obj->Pointer = intel_obj->buffer->virtual;
+   obj->Length = obj->Size;
+   obj->Offset = 0;
+
+   return obj->Pointer;
+}
+
+/**
+ * Called via glMapBufferRange().
+ *
+ * The goal of this extension is to allow apps to accumulate their rendering
+ * at the same time as they accumulate their buffer object.  Without it,
+ * you'd end up blocking on execution of rendering every time you mapped
+ * the buffer to put new data in.
+ *
+ * We support it in 3 ways: If unsynchronized, then don't bother
+ * flushing the batchbuffer before mapping the buffer, which can save blocking
+ * in many cases.  If we would still block, and they allow the whole buffer
+ * to be invalidated, then just allocate a new buffer to replace the old one.
+ * If not, and we'd block, and they allow the subrange of the buffer to be
+ * invalidated, then we can make a new little BO, let them write into that,
+ * and blit it into the real BO at unmap time.
+ */
+static void *
+intel_bufferobj_map_range(GLcontext * ctx,
+			  GLenum target, GLintptr offset, GLsizeiptr length,
+			  GLbitfield access, struct gl_buffer_object *obj)
+{
+   struct intel_context *intel = intel_context(ctx);
+   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
+
+   assert(intel_obj);
+
+   /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
+    * internally uses our functions directly.
+    */
+   obj->Offset = offset;
+   obj->Length = length;
+   obj->AccessFlags = access;
+
+   if (intel_obj->sys_buffer) {
+      obj->Pointer = intel_obj->sys_buffer + offset;
+      return obj->Pointer;
+   }
+
+   if (intel_obj->region)
+      intel_bufferobj_cow(intel, intel_obj);
+
+   /* If the mapping is synchronized with other GL operations, flush
+    * the batchbuffer so that GEM knows about the buffer access for later
+    * syncing.
+    */
+   if (!(access & GL_MAP_UNSYNCHRONIZED_BIT))
+      intelFlush(ctx);
+
+   if (intel_obj->buffer == NULL) {
+      obj->Pointer = NULL;
+      return NULL;
+   }
+
+   /* If the user doesn't care about existing buffer contents and mapping
+    * would cause us to block, then throw out the old buffer.
+    */
+   if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
+       (access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
+       drm_intel_bo_busy(intel_obj->buffer)) {
+      drm_intel_bo_unreference(intel_obj->buffer);
+      intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
+				       intel_obj->Base.Size, 64);
+   }
+
+   /* If the user is mapping a range of an active buffer object but
+    * doesn't require the current contents of that range, make a new
+    * BO, and we'll copy what they put in there out at unmap or
+    * FlushRange time.
+    */
+   if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
+       drm_intel_bo_busy(intel_obj->buffer)) {
+      if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
+	 intel_obj->range_map_buffer = _mesa_malloc(length);
+	 obj->Pointer = intel_obj->range_map_buffer;
+      } else {
+	 intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
+						      "range map",
+						      length, 64);
+	 if (!(access & GL_MAP_READ_BIT) &&
+	     intel->intelScreen->kernel_exec_fencing) {
+	    drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
+	    intel_obj->mapped_gtt = GL_TRUE;
+	 } else {
+	    drm_intel_bo_map(intel_obj->range_map_bo,
+			     (access & GL_MAP_WRITE_BIT) != 0);
+	    intel_obj->mapped_gtt = GL_FALSE;
+	 }
+	 obj->Pointer = intel_obj->range_map_bo->virtual;
+      }
+      return obj->Pointer;
+   }
+
+   if (!(access & GL_MAP_READ_BIT) &&
+       intel->intelScreen->kernel_exec_fencing) {
+      drm_intel_gem_bo_map_gtt(intel_obj->buffer);
+      intel_obj->mapped_gtt = GL_TRUE;
+   } else {
+      drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
+      intel_obj->mapped_gtt = GL_FALSE;
+   }
+
+   obj->Pointer = intel_obj->buffer->virtual + offset;
    return obj->Pointer;
 }
 
+/* Ideally we'd use a BO to avoid taking up cache space for the temporary
+ * data, but FlushMappedBufferRange may be followed by further writes to
+ * the pointer, so we would have to re-map after emitting our blit, which
+ * would defeat the point.
+ */
+static void
+intel_bufferobj_flush_mapped_range(GLcontext *ctx, GLenum target,
+				   GLintptr offset, GLsizeiptr length,
+				   struct gl_buffer_object *obj)
+{
+   struct intel_context *intel = intel_context(ctx);
+   struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
+   drm_intel_bo *temp_bo;
+
+   /* Unless we're in the range map using a temporary system buffer,
+    * there's no work to do.
+    */
+   if (intel_obj->range_map_buffer == NULL)
+      return;
+
+   temp_bo = drm_intel_bo_alloc(intel->bufmgr, "range map flush", length, 64);
+
+   drm_intel_bo_subdata(temp_bo, 0, length, intel_obj->range_map_buffer);
+
+   intel_emit_linear_blit(intel,
+			  intel_obj->buffer, obj->Offset + offset,
+			  temp_bo, 0,
+			  length);
+
+   drm_intel_bo_unreference(temp_bo);
+}
+
 
 /**
- * Called via glMapBufferARB().
+ * Called via glUnmapBuffer().
  */
 static GLboolean
 intel_bufferobj_unmap(GLcontext * ctx,
                       GLenum target, struct gl_buffer_object *obj)
 {
+   struct intel_context *intel = intel_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 
    assert(intel_obj);
-   if (intel_obj->buffer != NULL) {
-      assert(obj->Pointer);
-      dri_bo_unmap(intel_obj->buffer);
-      obj->Pointer = NULL;
+   assert(obj->Pointer);
+   if (intel_obj->sys_buffer != NULL) {
+      /* always keep the mapping around. */
+   } else if (intel_obj->range_map_buffer != NULL) {
+      /* Since we've emitted some blits to buffers that will (likely) be used
+       * in rendering operations in other cache domains in this batch, emit a
+       * flush.  Once again, we wish for a domain tracker in libdrm to cover
+       * usage inside of a batchbuffer.
+       */
+      intel_batchbuffer_emit_mi_flush(intel->batch);
+      free(intel_obj->range_map_buffer);
+      intel_obj->range_map_buffer = NULL;
+   } else if (intel_obj->range_map_bo != NULL) {
+      if (intel_obj->mapped_gtt) {
+	 drm_intel_gem_bo_unmap_gtt(intel_obj->range_map_bo);
+      } else {
+	 drm_intel_bo_unmap(intel_obj->range_map_bo);
+      }
+
+      intel_emit_linear_blit(intel,
+			     intel_obj->buffer, obj->Offset,
+			     intel_obj->range_map_bo, 0,
+			     obj->Length);
+
+      /* Since we've emitted some blits to buffers that will (likely) be used
+       * in rendering operations in other cache domains in this batch, emit a
+       * flush.  Once again, we wish for a domain tracker in libdrm to cover
+       * usage inside of a batchbuffer.
+       */
+      intel_batchbuffer_emit_mi_flush(intel->batch);
+
+      drm_intel_bo_unreference(intel_obj->range_map_bo);
+      intel_obj->range_map_bo = NULL;
+   } else if (intel_obj->buffer != NULL) {
+      if (intel_obj->mapped_gtt) {
+	 drm_intel_gem_bo_unmap_gtt(intel_obj->buffer);
+      } else {
+	 drm_intel_bo_unmap(intel_obj->buffer);
+      }
    }
+   obj->Pointer = NULL;
+   obj->Offset = 0;
+   obj->Length = 0;
+
    return GL_TRUE;
 }
 
@@ -266,19 +487,95 @@ intel_bufferobj_buffer(struct intel_context *intel,
       }
    }
 
+   if (intel_obj->buffer == NULL) {
+      void *sys_buffer = intel_obj->sys_buffer;
+
+      /* only one of buffer and sys_buffer could be non-NULL */
+      intel_bufferobj_alloc_buffer(intel, intel_obj);
+      intel_obj->sys_buffer = NULL;
+
+      intel_bufferobj_subdata(&intel->ctx,
+			      GL_ARRAY_BUFFER_ARB,
+			      0,
+			      intel_obj->Base.Size,
+			      sys_buffer,
+			      &intel_obj->Base);
+      _mesa_free(sys_buffer);
+      intel_obj->sys_buffer = NULL;
+   }
+
    return intel_obj->buffer;
 }
 
+static void
+intel_bufferobj_copy_subdata(GLcontext *ctx,
+			     struct gl_buffer_object *src,
+			     struct gl_buffer_object *dst,
+			     GLintptr read_offset, GLintptr write_offset,
+			     GLsizeiptr size)
+{
+   struct intel_context *intel = intel_context(ctx);
+   struct intel_buffer_object *intel_src = intel_buffer_object(src);
+   struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
+   drm_intel_bo *src_bo, *dst_bo;
+
+   if (size == 0)
+      return;
+
+   /* If we're in system memory, just map and memcpy. */
+   if (intel_src->sys_buffer || intel_dst->sys_buffer) {
+      /* The same buffer may be used, but note that regions copied may
+       * not overlap.
+       */
+      if (src == dst) {
+	 char *ptr = intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
+					 GL_READ_WRITE, dst);
+	 memcpy(ptr + write_offset, ptr + read_offset, size);
+	 intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
+      } else {
+	 const char *src_ptr;
+	 char *dst_ptr;
+
+	 src_ptr =  intel_bufferobj_map(ctx, GL_COPY_READ_BUFFER,
+					GL_READ_ONLY, src);
+	 dst_ptr =  intel_bufferobj_map(ctx, GL_COPY_WRITE_BUFFER,
+					GL_WRITE_ONLY, dst);
+
+	 memcpy(dst_ptr + write_offset, src_ptr + read_offset, size);
+
+	 intel_bufferobj_unmap(ctx, GL_COPY_READ_BUFFER, src);
+	 intel_bufferobj_unmap(ctx, GL_COPY_WRITE_BUFFER, dst);
+      }
+   }
+
+   /* Otherwise, we have real BOs, so blit them. */
+
+   dst_bo = intel_bufferobj_buffer(intel, intel_dst, INTEL_WRITE_PART);
+   src_bo = intel_bufferobj_buffer(intel, intel_src, INTEL_READ);
+
+   intel_emit_linear_blit(intel,
+			  dst_bo, write_offset,
+			  src_bo, read_offset, size);
+
+   /* Since we've emitted some blits to buffers that will (likely) be used
+    * in rendering operations in other cache domains in this batch, emit a
+    * flush.  Once again, we wish for a domain tracker in libdrm to cover
+    * usage inside of a batchbuffer.
+    */
+   intel_batchbuffer_emit_mi_flush(intel->batch);
+}
+
 void
-intel_bufferobj_init(struct intel_context *intel)
+intelInitBufferObjectFuncs(struct dd_function_table *functions)
 {
-   GLcontext *ctx = &intel->ctx;
-
-   ctx->Driver.NewBufferObject = intel_bufferobj_alloc;
-   ctx->Driver.DeleteBuffer = intel_bufferobj_free;
-   ctx->Driver.BufferData = intel_bufferobj_data;
-   ctx->Driver.BufferSubData = intel_bufferobj_subdata;
-   ctx->Driver.GetBufferSubData = intel_bufferobj_get_subdata;
-   ctx->Driver.MapBuffer = intel_bufferobj_map;
-   ctx->Driver.UnmapBuffer = intel_bufferobj_unmap;
+   functions->NewBufferObject = intel_bufferobj_alloc;
+   functions->DeleteBuffer = intel_bufferobj_free;
+   functions->BufferData = intel_bufferobj_data;
+   functions->BufferSubData = intel_bufferobj_subdata;
+   functions->GetBufferSubData = intel_bufferobj_get_subdata;
+   functions->MapBuffer = intel_bufferobj_map;
+   functions->MapBufferRange = intel_bufferobj_map_range;
+   functions->FlushMappedBufferRange = intel_bufferobj_flush_mapped_range;
+   functions->UnmapBuffer = intel_bufferobj_unmap;
+   functions->CopyBufferSubData = intel_bufferobj_copy_subdata;
 }