[intel] Move bufmgr back to context instead of screen, fixing glthreads.

Putting the bufmgr in the screen is not thread-safe since the emit_reloc
changes.  It also led to a significant performance hit from pthread usage
for the attempted thread-safety (up to 12% of a cpu spent on refcounting
protection in single-threaded 965).  The motivation had been to allow
multi-context bufmgr sharing in classic mode, but it wasn't worth the cost.

6 files changed, 129 insertions, 32 deletions

diff --git a/src/mesa/drivers/dri/i915/intel_context.c b/src/mesa/drivers/dri/i915/intel_context.c
index 326be9c76f..c67d906db0 100644
--- a/src/mesa/drivers/dri/i915/intel_context.c
+++ b/src/mesa/drivers/dri/i915/intel_context.c
@@ -60,6 +60,7 @@
 #include "intel_buffer_objects.h"
 #include "intel_fbo.h"
 #include "intel_decode.h"
+#include "intel_bufmgr_ttm.h"
 
 #include "drirenderbuffer.h"
 #include "vblank.h"
@@ -291,6 +292,81 @@ intelFinish(GLcontext * ctx)
    }
 }
 
+/** Driver-specific fence emit implementation for the fake memory manager. */
+static unsigned int
+intel_fence_emit(void *private)
+{
+   struct intel_context *intel = (struct intel_context *)private;
+   unsigned int fence;
+
+   /* XXX: Need to emit a flush, if we haven't already (at least with the
+    * current batchbuffer implementation, we have).
+    */
+
+   fence = intelEmitIrqLocked(intel);
+
+   return fence;
+}
+
+/** Driver-specific fence wait implementation for the fake memory manager. */
+static int
+intel_fence_wait(void *private, unsigned int cookie)
+{
+   struct intel_context *intel = (struct intel_context *)private;
+
+   intelWaitIrq(intel, cookie);
+
+   return 0;
+}
+
+static GLboolean
+intel_init_bufmgr(struct intel_context *intel)
+{
+   intelScreenPrivate *intelScreen = intel->intelScreen;
+   GLboolean ttm_disable = getenv("INTEL_NO_TTM") != NULL;
+
+   /* If we've got a new enough DDX that's initializing TTM and giving us
+    * object handles for the shared buffers, use that.
+    */
+   intel->ttm = GL_FALSE;
+   if (!ttm_disable &&
+       intel->intelScreen->driScrnPriv->ddx_version.minor >= 9 &&
+       intel->intelScreen->drmMinor >= 11 &&
+       intel->intelScreen->front.bo_handle != -1)
+   {
+      intel->bufmgr = intel_bufmgr_ttm_init(intel->driFd,
+					    DRM_FENCE_TYPE_EXE,
+					    DRM_FENCE_TYPE_EXE |
+					    DRM_I915_FENCE_TYPE_RW,
+					    BATCH_SZ);
+      if (intel->bufmgr != NULL)
+	 intel->ttm = GL_TRUE;
+   }
+   /* Otherwise, use the classic buffer manager. */
+   if (intel->bufmgr == NULL) {
+      if (ttm_disable) {
+	 fprintf(stderr, "TTM buffer manager disabled.  Using classic.\n");
+      } else {
+	 fprintf(stderr, "Failed to initialize TTM buffer manager.  "
+		 "Falling back to classic.\n");
+      }
+
+      if (intelScreen->tex.size == 0) {
+	 fprintf(stderr, "[%s:%u] Error initializing buffer manager.\n",
+		 __func__, __LINE__);
+	 return GL_FALSE;
+      }
+
+      intel->bufmgr = dri_bufmgr_fake_init(intelScreen->tex.offset,
+					   intelScreen->tex.map,
+					   intelScreen->tex.size,
+					   intel_fence_emit,
+					   intel_fence_wait,
+					   intel);
+   }
+
+   return GL_TRUE;
+}
 
 void
 intelInitDriverFunctions(struct dd_function_table *functions)
@@ -338,9 +414,22 @@ intelInitContext(struct intel_context *intel,
    intel->driScreen = sPriv;
    intel->sarea = saPriv;
 
+   /* Dri stuff */
+   intel->hHWContext = driContextPriv->hHWContext;
+   intel->driFd = sPriv->fd;
+   intel->driHwLock = (drmLock *) & sPriv->pSAREA->lock;
+
    intel->width = intelScreen->width;
    intel->height = intelScreen->height;
 
+   if (intelScreen->deviceID == PCI_CHIP_I865_G)
+      intel->maxBatchSize = 4096;
+   else
+      intel->maxBatchSize = BATCH_SZ;
+
+   if (!intel_init_bufmgr(intel))
+      return GL_FALSE;
+
    if (!lockMutexInit) {
       lockMutexInit = GL_TRUE;
       _glthread_INIT_MUTEX(lockMutex);
@@ -391,11 +480,6 @@ intelInitContext(struct intel_context *intel,
    _swrast_allow_pixel_fog(ctx, GL_FALSE);
    _swrast_allow_vertex_fog(ctx, GL_TRUE);
 
-   /* Dri stuff */
-   intel->hHWContext = driContextPriv->hHWContext;
-   intel->driFd = sPriv->fd;
-   intel->driHwLock = (drmLock *) & sPriv->pSAREA->lock;
-
    intel->hw_stipple = 1;
 
    /* XXX FBO: this doesn't seem to be used anywhere */
@@ -436,9 +520,10 @@ intelInitContext(struct intel_context *intel,
 /* 		      GL_TRUE, */
                      GL_FALSE);
 
-   if (intelScreen->ttm)
+   if (intel->ttm)
       driInitExtensions(ctx, ttm_extensions, GL_FALSE);
 
+   intel_recreate_static_regions(intel);
 
    intel->batch = intel_batchbuffer_alloc(intel);
    intel->last_swap_fence = NULL;
@@ -457,11 +542,10 @@ intelInitContext(struct intel_context *intel,
 
    intel->prim.primitive = ~0;
 
-
 #if DO_DEBUG
    INTEL_DEBUG = driParseDebugString(getenv("INTEL_DEBUG"), debug_control);
-   if (!intel->intelScreen->ttm && (INTEL_DEBUG & DEBUG_BUFMGR))
-      dri_bufmgr_fake_set_debug(intel->intelScreen->bufmgr, GL_TRUE);
+   if (!intel->ttm && (INTEL_DEBUG & DEBUG_BUFMGR))
+      dri_bufmgr_fake_set_debug(intel->bufmgr, GL_TRUE);
 #endif
 
    if (getenv("INTEL_NO_RAST")) {
@@ -507,6 +591,7 @@ intelDestroyContext(__DRIcontextPrivate * driContextPriv)
 	 intel->first_swap_fence = NULL;
       }
 
+      dri_bufmgr_destroy(intel->bufmgr);
 
       if (release_texture_heaps) {
          /* This share group is about to go away, free our private
@@ -551,21 +636,21 @@ intelMakeCurrent(__DRIcontextPrivate * driContextPriv,
 
          if (intel_fb->color_rb[0] && !intel_fb->color_rb[0]->region) {
             intel_region_reference(&intel_fb->color_rb[0]->region,
-				   intel->intelScreen->front_region);
+				   intel->front_region);
          }
          if (intel_fb->color_rb[1] && !intel_fb->color_rb[1]->region) {
             intel_region_reference(&intel_fb->color_rb[1]->region,
-				   intel->intelScreen->back_region);
+				   intel->back_region);
          }
          if (intel_fb->color_rb[2] && !intel_fb->color_rb[2]->region) {
             intel_region_reference(&intel_fb->color_rb[2]->region,
-				   intel->intelScreen->third_region);
+				   intel->third_region);
          }
          if (irbDepth && !irbDepth->region) {
-            intel_region_reference(&irbDepth->region, intel->intelScreen->depth_region);
+            intel_region_reference(&irbDepth->region, intel->depth_region);
          }
          if (irbStencil && !irbStencil->region) {
-            intel_region_reference(&irbStencil->region, intel->intelScreen->depth_region);
+            intel_region_reference(&irbStencil->region, intel->depth_region);
          }
       }
 
@@ -618,7 +703,6 @@ intelContendedLock(struct intel_context *intel, GLuint flags)
 {
    __DRIdrawablePrivate *dPriv = intel->driDrawable;
    __DRIscreenPrivate *sPriv = intel->driScreen;
-   intelScreenPrivate *intelScreen = (intelScreenPrivate *) sPriv->private;
    drmI830Sarea *sarea = intel->sarea;
 
    drmGetLock(intel->driFd, intel->hHWContext, flags);
@@ -639,9 +723,9 @@ intelContendedLock(struct intel_context *intel, GLuint flags)
     * between contexts of a single fake bufmgr, but this will at least make
     * things correct for now.
     */
-   if (!intel->intelScreen->ttm && sarea->texAge != intel->hHWContext) {
+   if (!intel->ttm && sarea->texAge != intel->hHWContext) {
       sarea->texAge = intel->hHWContext;
-      dri_bufmgr_fake_contended_lock_take(intel->intelScreen->bufmgr);
+      dri_bufmgr_fake_contended_lock_take(intel->bufmgr);
       if (INTEL_DEBUG & DEBUG_BATCH)
 	 intel_decode_context_reset();
    }
diff --git a/src/mesa/drivers/dri/i915/intel_context.h b/src/mesa/drivers/dri/i915/intel_context.h
index 993a08ef95..6f0051ed8b 100644
--- a/src/mesa/drivers/dri/i915/intel_context.h
+++ b/src/mesa/drivers/dri/i915/intel_context.h
@@ -34,6 +34,7 @@
 #include "drm.h"
 #include "mm.h"
 #include "texmem.h"
+#include "dri_bufmgr.h"
 
 #include "intel_screen.h"
 #include "intel_tex_obj.h"
@@ -142,10 +143,25 @@ struct intel_context
    GLuint Fallback;
    GLuint NewGLState;
 
+   dri_bufmgr *bufmgr;
+   unsigned int maxBatchSize;
+
+   struct intel_region *front_region;
+   struct intel_region *back_region;
+   struct intel_region *third_region;
+   struct intel_region *depth_region;
+
+   /**
+    * This value indicates that the kernel memory manager is being used
+    * instead of the fake client-side memory manager.
+    */
+   GLboolean ttm;
+
    dri_fence *last_swap_fence;
    dri_fence *first_swap_fence;
 
    struct intel_batchbuffer *batch;
+   unsigned batch_id;
    GLuint last_state_batch_id;
 
    struct
diff --git a/src/mesa/drivers/dri/i915/intel_ioctl.c b/src/mesa/drivers/dri/i915/intel_ioctl.c
index 94f7e73ecf..37704d66ec 100644
--- a/src/mesa/drivers/dri/i915/intel_ioctl.c
+++ b/src/mesa/drivers/dri/i915/intel_ioctl.c
@@ -47,15 +47,14 @@
 #define FILE_DEBUG_FLAG DEBUG_IOCTL
 
 int
-intelEmitIrqLocked(intelScreenPrivate *intelScreen)
+intelEmitIrqLocked(struct intel_context *intel)
 {
    drmI830IrqEmit ie;
    int ret, seq;
 
    ie.irq_seq = &seq;
 
-   ret = drmCommandWriteRead(intelScreen->driScrnPriv->fd,
-			     DRM_I830_IRQ_EMIT, &ie, sizeof(ie));
+   ret = drmCommandWriteRead(intel->driFd, DRM_I830_IRQ_EMIT, &ie, sizeof(ie));
    if (ret) {
       fprintf(stderr, "%s: drmI830IrqEmit: %d\n", __FUNCTION__, ret);
       exit(1);
@@ -67,7 +66,7 @@ intelEmitIrqLocked(intelScreenPrivate *intelScreen)
 }
 
 void
-intelWaitIrq(intelScreenPrivate *intelScreen, int seq)
+intelWaitIrq(struct intel_context *intel, int seq)
 {
    drm_i915_irq_wait_t iw;
    int ret;
@@ -77,8 +76,7 @@ intelWaitIrq(intelScreenPrivate *intelScreen, int seq)
    iw.irq_seq = seq;
 
    do {
-      ret = drmCommandWrite(intelScreen->driScrnPriv->fd,
-			    DRM_I830_IRQ_WAIT, &iw, sizeof(iw));
+      ret = drmCommandWrite(intel->driFd, DRM_I830_IRQ_WAIT, &iw, sizeof(iw));
    } while (ret == -EAGAIN || ret == -EINTR);
 
    if (ret) {
@@ -170,7 +168,7 @@ intel_exec_ioctl(struct intel_context *intel,
    }
 
 
-   fo = intel_ttm_fence_create_from_arg(intel->intelScreen->bufmgr, "fence buffers",
+   fo = intel_ttm_fence_create_from_arg(intel->bufmgr, "fence buffers",
 					&execbuf.fence_arg);
    if (!fo) {
       fprintf(stderr, "failed to fence handle: %08x\n", execbuf.fence_arg.handle);
diff --git a/src/mesa/drivers/dri/i915/intel_ioctl.h b/src/mesa/drivers/dri/i915/intel_ioctl.h
index 953fee9240..60bc4a3fb9 100644
--- a/src/mesa/drivers/dri/i915/intel_ioctl.h
+++ b/src/mesa/drivers/dri/i915/intel_ioctl.h
@@ -30,8 +30,8 @@
 
 #include "intel_context.h"
 
-void intelWaitIrq(intelScreenPrivate *intelScreen, int seq);
-int intelEmitIrqLocked(intelScreenPrivate *intelScreen);
+void intelWaitIrq(struct intel_context *intel, int seq);
+int intelEmitIrqLocked(struct intel_context *intel);
 
 void intel_batch_ioctl(struct intel_context *intel,
                        GLuint start_offset,
diff --git a/src/mesa/drivers/dri/i915/intel_pixel_copy.c b/src/mesa/drivers/dri/i915/intel_pixel_copy.c
index 629cdb979d..0bda2d863f 100644
--- a/src/mesa/drivers/dri/i915/intel_pixel_copy.c
+++ b/src/mesa/drivers/dri/i915/intel_pixel_copy.c
@@ -54,9 +54,8 @@ copypix_src_region(struct intel_context *intel, GLenum type)
    case GL_DEPTH:
       /* Don't think this is really possible execpt at 16bpp, when we have no stencil.
        */
-      if (intel->intelScreen->depth_region && 
-	  intel->intelScreen->depth_region->cpp == 2)
-         return intel->intelScreen->depth_region;
+      if (intel->depth_region && intel->depth_region->cpp == 2)
+         return intel->depth_region;
    case GL_STENCIL:
       /* Don't think this is really possible. 
        */
@@ -64,7 +63,7 @@ copypix_src_region(struct intel_context *intel, GLenum type)
    case GL_DEPTH_STENCIL_EXT:
       /* Does it matter whether it is stencil/depth or depth/stencil?
        */
-      return intel->intelScreen->depth_region;
+      return intel->depth_region;
    default:
       break;
    }
@@ -164,7 +163,7 @@ do_texture_copypixels(GLcontext * ctx,
 
    /* Set the 3d engine to draw into the destination region:
     */
-   intel->vtbl.meta_draw_region(intel, dst, intel->intelScreen->depth_region);
+   intel->vtbl.meta_draw_region(intel, dst, intel->depth_region);
 
    intel->vtbl.meta_import_pixel_state(intel);
 
diff --git a/src/mesa/drivers/dri/i915/intel_pixel_draw.c b/src/mesa/drivers/dri/i915/intel_pixel_draw.c
index 8349f4c748..0fea9a1d01 100644
--- a/src/mesa/drivers/dri/i915/intel_pixel_draw.c
+++ b/src/mesa/drivers/dri/i915/intel_pixel_draw.c
@@ -112,7 +112,7 @@ do_texture_drawpixels(GLcontext * ctx,
 
    /* Set the 3d engine to draw into the destination region:
     */
-   intel->vtbl.meta_draw_region(intel, dst, intel->intelScreen->depth_region);
+   intel->vtbl.meta_draw_region(intel, dst, intel->depth_region);
 
    intel->vtbl.meta_import_pixel_state(intel);