Bug 7260: mach64 texture memory mng cleanup

mach64 uses its own set of texture memory management routines which are buggy,
running a second DRI client kills the first one. This patch ports mach64 code
to the stock dri texture managment code.

1 files changed, 152 insertions, 503 deletions

diff --git a/src/mesa/drivers/dri/mach64/mach64_texmem.c b/src/mesa/drivers/dri/mach64/mach64_texmem.c
index 017fd3535d..3b7b93b984 100644
--- a/src/mesa/drivers/dri/mach64/mach64_texmem.c
+++ b/src/mesa/drivers/dri/mach64/mach64_texmem.c
@@ -49,333 +49,19 @@
  */
 void mach64DestroyTexObj( mach64ContextPtr mmesa, mach64TexObjPtr t )
 {
-#if ENABLE_PERF_BOXES
-   /* Bump the performace counter */
-   if (mmesa)
-      mmesa->c_textureSwaps++;
-#endif
-   if ( !t ) return;
-
-#if 0
-   if ( t->tObj && t->memBlock && mmesa ) {
-      /* not a placeholder, so release from global LRU if necessary */
-      int heap = t->heap;
-      drmTextureRegion *list = mmesa->sarea->tex_list[heap];
-      int log2sz = mmesa->mach64Screen->logTexGranularity[heap];
-      int start = t->memBlock->ofs >> log2sz;
-      int end = (t->memBlock->ofs + t->memBlock->size - 1) >> log2sz;
-      int i;
-
-      mmesa->lastTexAge[heap] = ++mmesa->sarea->tex_age[heap];
-
-      /* Update the global LRU */
-      for ( i = start ; i <= end ; i++ ) {
-	 /* do we own this block? */
-	 if (list[i].in_use == mmesa->hHWContext) {
-	    list[i].in_use = 0;
-	    list[i].age = mmesa->lastTexAge[heap];
-
-	    /* remove_from_list(i) */
-	    list[(GLuint)list[i].next].prev = list[i].prev;
-	    list[(GLuint)list[i].prev].next = list[i].next;
-	 }
-      }
-   }
-#endif
-
-   if ( t->memBlock ) {
-      mmFreeMem( t->memBlock );
-      t->memBlock = NULL;
-   }
-
-   if ( t->tObj ) {
-      t->tObj->DriverData = NULL;
-   }
-
-   if ( t->bound && mmesa )
-      mmesa->CurrentTexObj[t->bound-1] = NULL;
-
-   remove_from_list( t );
-   FREE( t );
-}
-
-/* Keep track of swapped out texture objects.
- */
-void mach64SwapOutTexObj( mach64ContextPtr mmesa,
-			  mach64TexObjPtr t )
-{
-#if ENABLE_PERF_BOXES
-   /* Bump the performace counter */
-   if (mmesa)
-     mmesa->c_textureSwaps++;
-#endif
-
-#if 0
-   if ( t->tObj && t->memBlock && mmesa ) {
-      /* not a placeholder, so release from global LRU if necessary */
-      int heap = t->heap;
-      drmTextureRegion *list = mmesa->sarea->tex_list[heap];
-      int log2sz = mmesa->mach64Screen->logTexGranularity[heap];
-      int start = t->memBlock->ofs >> log2sz;
-      int end = (t->memBlock->ofs + t->memBlock->size - 1) >> log2sz;
-      int i;
-
-      mmesa->lastTexAge[heap] = ++mmesa->sarea->tex_age[heap];
-
-      /* Update the global LRU */
-      for ( i = start ; i <= end ; i++ ) {
-	 /* do we own this block? */
-	 if (list[i].in_use == mmesa->hHWContext) {
-	    list[i].in_use = 0;
-	    list[i].age = mmesa->lastTexAge[heap];
-
-	    /* remove_from_list(i) */
-	    list[(GLuint)list[i].next].prev = list[i].prev;
-	    list[(GLuint)list[i].prev].next = list[i].next;
-	 }
-      }
-   }
-#endif
-
-   if ( t->memBlock ) {
-      mmFreeMem( t->memBlock );
-      t->memBlock = NULL;
-   }
-
-   t->dirty = ~0;
-   move_to_tail( &mmesa->SwappedOut, t );
-}
-
-/* Print out debugging information about texture LRU.
- */
-void mach64PrintLocalLRU( mach64ContextPtr mmesa, int heap )
-{
-   mach64TexObjPtr t;
-   int sz = 1 << (mmesa->mach64Screen->logTexGranularity[heap]);
-
-   fprintf( stderr, "\nLocal LRU, heap %d:\n", heap );
-
-   foreach( t, &mmesa->TexObjList[heap] ) {
-      if ( !t->tObj ) {
-	 fprintf( stderr, "Placeholder %d at 0x%x sz 0x%x\n",
-		  t->memBlock->ofs / sz,
-		  t->memBlock->ofs,
-		  t->memBlock->size );
-      } else {
-	 fprintf( stderr, "Texture (bound %d) at 0x%x sz 0x%x\n",
-		  t->bound,
-		  t->memBlock->ofs,
-		  t->memBlock->size );
-      }
-   }
-
-   fprintf( stderr, "\n" );
-}
-
-void mach64PrintGlobalLRU( mach64ContextPtr mmesa, int heap )
-{
-   drm_tex_region_t *list = mmesa->sarea->tex_list[heap];
-   int i, j;
-
-   fprintf( stderr, "\nGlobal LRU, heap %d list %p:\n", heap, list );
-
-   for ( i = 0, j = MACH64_NR_TEX_REGIONS ; i < MACH64_NR_TEX_REGIONS ; i++ ) {
-      fprintf( stderr, "list[%d] age %d in_use %d next %d prev %d\n",
-	       j, list[j].age, list[j].in_use, list[j].next, list[j].prev );
-      j = list[j].next;
-      if ( j == MACH64_NR_TEX_REGIONS ) break;
-   }
-
-   if ( j != MACH64_NR_TEX_REGIONS ) {
-      fprintf( stderr, "Loop detected in global LRU\n" );
-      for ( i = 0 ; i < MACH64_NR_TEX_REGIONS ; i++ ) {
-	 fprintf( stderr, "list[%d] age %d in_use %d next %d prev %d\n",
-		  i, list[i].age, list[i].in_use, list[i].next, list[i].prev );
-      }
-   }
+   unsigned   i;
 
-   fprintf( stderr, "\n" );
-}
-
-/* Reset the global texture LRU.
- */
-/* NOTE: This function is only called while holding the hardware lock */
-static void mach64ResetGlobalLRU( mach64ContextPtr mmesa, int heap )
-{
-   drm_tex_region_t *list = mmesa->sarea->tex_list[heap];
-   int sz = 1 << mmesa->mach64Screen->logTexGranularity[heap];
-   int i;
-
-   /* (Re)initialize the global circular LRU list.  The last element in
-    * the array (MACH64_NR_TEX_REGIONS) is the sentinal.  Keeping it at
-    * the end of the array allows it to be addressed rationally when
-    * looking up objects at a particular location in texture memory.
+   /* See if it was the driver's current object.
     */
-   for ( i = 0 ; (i+1) * sz <= mmesa->mach64Screen->texSize[heap] ; i++ ) {
-      list[i].prev = i-1;
-      list[i].next = i+1;
-      list[i].age = 0;
-      list[i].in_use = 0;
-   }
-
-   i--;
-   list[0].prev = MACH64_NR_TEX_REGIONS;
-   list[i].prev = i-1;
-   list[i].next = MACH64_NR_TEX_REGIONS;
-   list[MACH64_NR_TEX_REGIONS].prev = i;
-   list[MACH64_NR_TEX_REGIONS].next = 0;
-   mmesa->sarea->tex_age[heap] = 0;
-}
-
-/* Update the local and global texture LRUs.
- */
-/* NOTE: This function is only called while holding the hardware lock */
-void mach64UpdateTexLRU( mach64ContextPtr mmesa,
-			 mach64TexObjPtr t )
-{
-   int heap = t->heap;
-   drm_tex_region_t *list = mmesa->sarea->tex_list[heap];
-   int log2sz = mmesa->mach64Screen->logTexGranularity[heap];
-   int start = t->memBlock->ofs >> log2sz;
-   int end = (t->memBlock->ofs + t->memBlock->size - 1) >> log2sz;
-   int i;
-
-   mmesa->lastTexAge[heap] = ++mmesa->sarea->tex_age[heap];
-
-   if ( !t->memBlock ) {
-      fprintf( stderr, "no memblock\n\n" );
-      return;
-   }
-
-   /* Update our local LRU */
-   move_to_head( &mmesa->TexObjList[heap], t );
-
-   /* Update the global LRU */
-   for ( i = start ; i <= end ; i++ ) {
-      list[i].in_use = mmesa->hHWContext;
-      list[i].age = mmesa->lastTexAge[heap];
-
-#if 0
-      /* if this is the last region, it's not in the list */
-      if ( !(i*(1<<log2sz) > mmesa->mach64Screen->texSize[heap] ) ) {
-#endif
-	 /* remove_from_list(i) */
-	 list[(GLuint)list[i].next].prev = list[i].prev;
-	 list[(GLuint)list[i].prev].next = list[i].next;
-#if 0
-      }
-#endif
-
-      /* insert_at_head(list, i) */
-      list[i].prev = MACH64_NR_TEX_REGIONS;
-      list[i].next = list[MACH64_NR_TEX_REGIONS].next;
-      list[(GLuint)list[MACH64_NR_TEX_REGIONS].next].prev = i;
-      list[MACH64_NR_TEX_REGIONS].next = i;
-   }
-
-   if ( MACH64_DEBUG & DEBUG_VERBOSE_LRU ) {
-      mach64PrintGlobalLRU( mmesa, t->heap );
-      mach64PrintLocalLRU( mmesa, t->heap );
-   }
-}
-
-/* Update our notion of what textures have been changed since we last
- * held the lock.  This pertains to both our local textures and the
- * textures belonging to other clients.  Keep track of other client's
- * textures by pushing a placeholder texture onto the LRU list -- these
- * are denoted by (tObj == NULL).
- */
-/* NOTE: This function is only called while holding the hardware lock */
-static void mach64TexturesGone( mach64ContextPtr mmesa, int heap,
-				int offset, int size, int in_use )
-{
-   mach64TexObjPtr t, tmp;
-
-   foreach_s ( t, tmp, &mmesa->TexObjList[heap] ) {
-      if ( t->memBlock->ofs >= offset + size ||
-	   t->memBlock->ofs + t->memBlock->size <= offset )
-	 continue;
-
-      /* It overlaps - kick it out.  Need to hold onto the currently
-       * bound objects, however.
-       */
-      if ( t->bound ) {
-	 mach64SwapOutTexObj( mmesa, t );
-      } else {
-	 mach64DestroyTexObj( mmesa, t );
-      }
-   }
-
-   if ( in_use > 0 && in_use != mmesa->hHWContext ) {
-      t = (mach64TexObjPtr) CALLOC( sizeof(*t) );
-      if (!t) return;
-
-      t->memBlock = mmAllocMem( mmesa->texHeap[heap], size, 0, offset );
-      if ( !t->memBlock ) {
-	 fprintf( stderr, "Couldn't alloc placeholder sz %x ofs %x\n",
-		  (int)size, (int)offset );
-	 mmDumpMemInfo( mmesa->texHeap[heap] );
-	 return;
-      }
-      insert_at_head( &mmesa->TexObjList[heap], t );
-   }
-}
-
-/* Update our client's shared texture state.  If another client has
- * modified a region in which we have textures, then we need to figure
- * out which of our textures has been removed, and update our global
- * LRU.
- */
-void mach64AgeTextures( mach64ContextPtr mmesa, int heap )
-{
-   drm_mach64_sarea_t *sarea = mmesa->sarea;
-
-   if ( sarea->tex_age[heap] != mmesa->lastTexAge[heap] ) {
-      int sz = 1 << mmesa->mach64Screen->logTexGranularity[heap];
-      int nr = 0;
-      int idx;
-
-      /* Have to go right round from the back to ensure stuff ends up
-       * LRU in our local list...  Fix with a cursor pointer.
-       */
-      for ( idx = sarea->tex_list[heap][MACH64_NR_TEX_REGIONS].prev ;
-	    idx != MACH64_NR_TEX_REGIONS && nr < MACH64_NR_TEX_REGIONS ;
-	    idx = sarea->tex_list[heap][idx].prev, nr++ )
+   if ( mmesa != NULL )
+   {
+      for ( i = 0 ; i < mmesa->glCtx->Const.MaxTextureUnits ; i++ )
       {
-	 /* If switching texturing schemes, then the SAREA might not
-	  * have been properly cleared, so we need to reset the
-	  * global texture LRU.
-	  */
-	 if ( idx * sz > mmesa->mach64Screen->texSize[heap] ) {
-	    nr = MACH64_NR_TEX_REGIONS;
-	    break;
-	 }
-
-	 if ( sarea->tex_list[heap][idx].age > mmesa->lastTexAge[heap] ) {
-	    mach64TexturesGone( mmesa, heap, idx * sz, sz,
-				sarea->tex_list[heap][idx].in_use );
-	 }
-      }
-
-      /* If switching texturing schemes, then the SAREA might not
-       * have been properly cleared, so we need to reset the
-       * global texture LRU.
-       */
-      if ( nr == MACH64_NR_TEX_REGIONS ) {
-	 mach64TexturesGone( mmesa, heap, 0,
-			     mmesa->mach64Screen->texSize[heap], 0 );
-	 mach64ResetGlobalLRU( mmesa, heap );
+         if ( t == mmesa->CurrentTexObj[ i ] ) {
+            assert( t->base.bound & (1 << i) );
+            mmesa->CurrentTexObj[ i ] = NULL;
+         }
       }
-
-      if ( 0 ) {
-	 mach64PrintGlobalLRU( mmesa, heap );
-	 mach64PrintLocalLRU( mmesa, heap );
-      }
-
-      mmesa->dirty |= (MACH64_UPLOAD_CONTEXT |
-		       MACH64_UPLOAD_TEX0IMAGE |
-		       MACH64_UPLOAD_TEX1IMAGE);
-      mmesa->lastTexAge[heap] = sarea->tex_age[heap];
    }
 }
 
@@ -395,7 +81,7 @@ static void mach64UploadAGPSubImage( mach64ContextPtr mmesa,
    if ( ( level < 0 ) || ( level > mmesa->glCtx->Const.MaxTextureLevels ) )
      return;
 
-   image = t->tObj->Image[0][level];
+   image = t->base.tObj->Image[0][level];
    if ( !image )
       return;
 
@@ -424,14 +110,13 @@ static void mach64UploadAGPSubImage( mach64ContextPtr mmesa,
       fprintf( stderr, "mach64UploadSubImage: %d,%d of %d,%d at %d,%d\n",
 	       width, height, image->Width, image->Height, x, y );
       fprintf( stderr, "            blit ofs: 0x%07x pitch: 0x%x dwords: %d\n",
-	       (GLuint)t->offset, (GLint)width, dwords );
-      mmDumpMemInfo( mmesa->texHeap[t->heap] );
+	       (GLuint)t->bufAddr, (GLint)width, dwords );
    }
 
    assert(image->Data);
 
    {
-      CARD32 *dst = (CARD32 *)((char *)mach64Screen->agpTextures.map + t->memBlock->ofs);
+      CARD32 *dst = (CARD32 *)((char *)mach64Screen->agpTextures.map + t->base.memBlock->ofs);
       const GLubyte *src = (const GLubyte *) image->Data +
 	 (y * image->Width + x) * image->TexFormat->TexelBytes;
       const GLuint bytes = width * height * image->TexFormat->TexelBytes;
@@ -460,7 +145,7 @@ static void mach64UploadLocalSubImage( mach64ContextPtr mmesa,
    if ( ( level < 0 ) || ( level > mmesa->glCtx->Const.MaxTextureLevels ) )
       return;
 
-   image = t->tObj->Image[0][level];
+   image = t->base.tObj->Image[0][level];
    if ( !image )
       return;
 
@@ -543,7 +228,7 @@ static void mach64UploadLocalSubImage( mach64ContextPtr mmesa,
    }
 
    dwords = width * height / texelsPerDword;
-   offset = t->offset;
+   offset = t->bufAddr;
 
 #if ENABLE_PERF_BOXES
    /* Bump the performance counter */
@@ -555,7 +240,6 @@ static void mach64UploadLocalSubImage( mach64ContextPtr mmesa,
 	       width, height, image->Width, image->Height, x, y );
       fprintf( stderr, "            blit ofs: 0x%07x pitch: 0x%x dwords: %d\n",
 	       (GLuint)offset, (GLint)width, dwords );
-      mmDumpMemInfo( mmesa->texHeap[t->heap] );
    }
 
    /* Subdivide the texture if required (account for the registers added by the drm) */
@@ -594,78 +278,32 @@ static void mach64UploadLocalSubImage( mach64ContextPtr mmesa,
  */
 void mach64UploadTexImages( mach64ContextPtr mmesa, mach64TexObjPtr t )
 {
-   GLint heap;
-
    if ( MACH64_DEBUG & DEBUG_VERBOSE_API ) {
       fprintf( stderr, "%s( %p, %p )\n",
 	       __FUNCTION__, mmesa->glCtx, t );
    }
 
    assert(t);
-   assert(t->tObj);
-
-   /* Choose the heap appropriately */
-   heap = MACH64_CARD_HEAP;
+   assert(t->base.tObj);
 
-   if ( !mmesa->mach64Screen->IsPCI &&
-	t->size > mmesa->mach64Screen->texSize[heap] ) {
-      heap = MACH64_AGP_HEAP;
-   }
-
-   /* Do we need to eject LRU texture objects? */
-   if ( !t->memBlock ) {
-      t->heap = heap;
+   if ( !t->base.memBlock ) {
+      int heap;
 
-      /* Allocate a memory block on a 64-byte boundary */
-      t->memBlock = mmAllocMem( mmesa->texHeap[heap], t->size, 6, 0 );
+      /* NULL heaps are skipped */
+      heap = driAllocateTexture( mmesa->texture_heaps, MACH64_NR_TEX_HEAPS,
+				 (driTextureObject *) t );
 
-      /* Try AGP before kicking anything out of local mem */
-      if ( !mmesa->mach64Screen->IsPCI && !t->memBlock && heap == MACH64_CARD_HEAP ) {
-	 t->memBlock = mmAllocMem( mmesa->texHeap[MACH64_AGP_HEAP],
-				   t->size, 6, 0 );
-
-	 if ( t->memBlock )
-	    heap = t->heap = MACH64_AGP_HEAP;
+      if ( heap == -1 ) {
+	 fprintf( stderr, "%s: upload texture failure, sz=%d\n", __FUNCTION__,
+		  t->base.totalSize );
+	 exit(-1);
+	 return;
       }
 
-      /* Kick out textures until the requested texture fits */
-      while ( !t->memBlock ) {
-	 if ( mmesa->TexObjList[heap].prev->bound ) {
-	    fprintf( stderr,
-		     "mach64UploadTexImages: ran into bound texture\n" );
-	    return;
-	 }
-	 if ( mmesa->TexObjList[heap].prev == &mmesa->TexObjList[heap] ) {
-	    if ( mmesa->mach64Screen->IsPCI ) {
-	       fprintf( stderr, "%s: upload texture failure on "
-			"local texture heaps, sz=%d\n", __FUNCTION__,
-			t->size );
-	       return;
-	    } else if ( heap == MACH64_CARD_HEAP ) {
-	       heap = t->heap = MACH64_AGP_HEAP;
-	       continue;
-	    } else {
-	      int i;
-	       fprintf( stderr, "%s: upload texture failure on "
-			"%sAGP texture heaps, sz=%d\n", __FUNCTION__,
-			mmesa->firstTexHeap == MACH64_CARD_HEAP ? "both local and " : "",
-			t->size );
-	       for ( i = mmesa->firstTexHeap ; i < mmesa->lastTexHeap ; i++ ) {
-		  mach64PrintLocalLRU( mmesa, i );
-	          mmDumpMemInfo( mmesa->texHeap[i] );
-	       }
-	       exit(-1);
-	       return;
-	    }
-	 }
-
-	 mach64SwapOutTexObj( mmesa, mmesa->TexObjList[heap].prev );
-
-	 t->memBlock = mmAllocMem( mmesa->texHeap[heap], t->size, 6, 0 );
-      }
+      t->heap = heap;
 
       /* Set the base offset of the texture image */
-      t->offset = mmesa->mach64Screen->texOffset[heap] + t->memBlock->ofs;
+      t->bufAddr = mmesa->mach64Screen->texOffset[heap] + t->base.memBlock->ofs;
 
       /* Force loading the new state into the hardware */
       mmesa->dirty |= (MACH64_UPLOAD_SCALE_3D_CNTL |
@@ -673,142 +311,152 @@ void mach64UploadTexImages( mach64ContextPtr mmesa, mach64TexObjPtr t )
    }
 
    /* Let the world know we've used this memory recently */
-   mach64UpdateTexLRU( mmesa, t );
+   driUpdateTextureLRU( (driTextureObject *) t );
 
    /* Upload any images that are new */
-   if ( t->dirty ) {
+   if ( t->base.dirty_images[0] ) {
+      const GLint j = t->base.tObj->BaseLevel;
       if (t->heap == MACH64_AGP_HEAP) {
 	 /* Need to make sure any vertex buffers in the queue complete */
 	 mach64WaitForIdleLocked( mmesa );
-	 mach64UploadAGPSubImage( mmesa, t, t->tObj->BaseLevel, 0, 0,
-				  t->tObj->Image[0][t->tObj->BaseLevel]->Width,
-				  t->tObj->Image[0][t->tObj->BaseLevel]->Height );
+	 mach64UploadAGPSubImage( mmesa, t, j, 0, 0,
+				  t->base.tObj->Image[0][j]->Width,
+				  t->base.tObj->Image[0][j]->Height );
       } else {
-	 mach64UploadLocalSubImage( mmesa, t, t->tObj->BaseLevel, 0, 0,
-				    t->tObj->Image[0][t->tObj->BaseLevel]->Width,
-				    t->tObj->Image[0][t->tObj->BaseLevel]->Height );
+	 mach64UploadLocalSubImage( mmesa, t, j, 0, 0,
+				    t->base.tObj->Image[0][j]->Width,
+				    t->base.tObj->Image[0][j]->Height );
       }
 
       mmesa->setup.tex_cntl |= MACH64_TEX_CACHE_FLUSH;
+      t->base.dirty_images[0] = 0;
    }
 
    mmesa->dirty |= MACH64_UPLOAD_TEXTURE;
+}
+
+
+/* Allocate memory from the same texture heap `heap' for both textures
+ * `u0' and `u1'.
+ */
+static int mach64AllocateMultiTex( mach64ContextPtr mmesa,
+				   mach64TexObjPtr u0,
+				   mach64TexObjPtr u1,
+				   int heap, GLboolean alloc_u0 )
+{
+   /* Both objects should be bound */
+   assert( u0->base.bound && u1->base.bound );
+
+   if ( alloc_u0 ) {
+      /* Evict u0 from its current heap */
+      if ( u0->base.memBlock ) {
+	 assert( u0->heap != heap );
+	 driSwapOutTextureObject( (driTextureObject *) u0 );
+      }
+
+      /* Try to allocate u0 in the chosen heap */
+      u0->heap = driAllocateTexture( &mmesa->texture_heaps[heap], 1,
+				     (driTextureObject *) u0 );
+
+      if ( u0->heap == -1 ) {
+	 return -1;
+      }
+   }
+
+   /* Evict u1 from its current heap */
+   if ( u1->base.memBlock ) {
+      assert( u1->heap != heap );
+      driSwapOutTextureObject( (driTextureObject *) u1 );
+   }
+
+   /* Try to allocate u1 in the same heap as u0 */
+   u1->heap = driAllocateTexture( &mmesa->texture_heaps[heap], 1,
+				  (driTextureObject *) u1 );
+
+   if ( u1->heap == -1 ) {
+      return -1;
+   }
 
-   t->dirty = 0;
+   /* Bound objects are not evicted */
+   assert( u0->base.memBlock && u1->base.memBlock );
+   assert( u0->heap == u1->heap );
+
+   return heap;
 }
 
 /* The mach64 needs to have both primary and secondary textures in either
  * local or AGP memory, so we need a "buddy system" to make sure that allocation
  * succeeds or fails for both textures.
- * FIXME: This needs to be optimized better.
  */
 void mach64UploadMultiTexImages( mach64ContextPtr mmesa, 
 				 mach64TexObjPtr t0,
 				 mach64TexObjPtr t1 )
 {
-   GLint heap;
-
    if ( MACH64_DEBUG & DEBUG_VERBOSE_API ) {
       fprintf( stderr, "%s( %p, %p %p )\n",
 	       __FUNCTION__, mmesa->glCtx, t0, t1 );
    }
 
    assert(t0 && t1);
-   assert(t0->tObj && t1->tObj);
+   assert(t0->base.tObj && t1->base.tObj);
 
-   /* Choose the heap appropriately */
-   heap = MACH64_CARD_HEAP;
+   if ( !t0->base.memBlock || !t1->base.memBlock || t0->heap != t1->heap ) {
+      mach64TexObjPtr u0 = NULL;
+      mach64TexObjPtr u1 = NULL;
+      unsigned totalSize = t0->base.totalSize + t1->base.totalSize;
 
-   if ( !mmesa->mach64Screen->IsPCI &&
-	((t0->size + t1->size) > mmesa->mach64Screen->texSize[heap]) ) {
-      heap = MACH64_AGP_HEAP;
-   }
+      int heap, ret;
 
-   /* Do we need to eject LRU texture objects? */
-   if ( !t0->memBlock || !t1->memBlock || t0->heap != t1->heap ) {
-      /* FIXME: starting from scratch for now to keep it simple */
-      if ( t0->memBlock ) {
-	 mach64SwapOutTexObj( mmesa, t0 );
-      }
-      if ( t1->memBlock ) {
-	 mach64SwapOutTexObj( mmesa, t1 );
-      }
-      t0->heap = t1->heap = heap;
-      /* Allocate a memory block on a 64-byte boundary */
-      t0->memBlock = mmAllocMem( mmesa->texHeap[heap], t0->size, 6, 0 );
-      if ( t0->memBlock ) {
-	 t1->memBlock = mmAllocMem( mmesa->texHeap[heap], t1->size, 6, 0 );
-	 if ( !t1->memBlock ) {
-	    mmFreeMem( t0->memBlock );
-	    t0->memBlock = NULL;
-	 }
+      /* Check if one of the textures is already swapped in a heap and the
+       * other texture fits in that heap.
+       */
+      if ( t0->base.memBlock && totalSize <= t0->base.heap->size ) {
+	 u0 = t0;
+	 u1 = t1;
+      } else if ( t1->base.memBlock && totalSize <= t1->base.heap->size ) {
+	 u0 = t1;
+	 u1 = t0;
       }
-      /* Try AGP before kicking anything out of local mem */
-      if ( (!t0->memBlock || !t1->memBlock) && heap == MACH64_CARD_HEAP ) {
-	 t0->memBlock = mmAllocMem( mmesa->texHeap[MACH64_AGP_HEAP], t0->size, 6, 0 );
-	 if ( t0->memBlock ) {
-	    t1->memBlock = mmAllocMem( mmesa->texHeap[MACH64_AGP_HEAP], t1->size, 6, 0 );
-	    if ( !t1->memBlock ) {
-	       mmFreeMem( t0->memBlock );
-	       t0->memBlock = NULL;
-	    }
+
+      if ( u0 ) {
+	 heap = u0->heap;
+
+	 ret = mach64AllocateMultiTex( mmesa, u0, u1, heap, GL_FALSE );
+      } else {
+	 /* Both textures are swapped out or collocation is impossible */
+	 u0 = t0;
+	 u1 = t1;
+
+	 /* Choose the heap appropriately */
+	 heap = MACH64_CARD_HEAP;
+
+	 if ( totalSize > mmesa->texture_heaps[heap]->size ) {
+	    heap = MACH64_AGP_HEAP;
 	 }
 
-	 if ( t0->memBlock && t1->memBlock )
-	    heap = t0->heap = t1->heap = MACH64_AGP_HEAP;
+	 ret = mach64AllocateMultiTex( mmesa, u0, u1, heap, GL_TRUE );
       }
 
-      /* Kick out textures until the requested texture fits */
-      while ( !t0->memBlock || !t1->memBlock ) {
-	 if ( mmesa->TexObjList[heap].prev->bound ) {
-	    fprintf( stderr,
-		     "%s: ran into bound texture\n", __FUNCTION__ );
-	    return;
-	 }
-	 if ( mmesa->TexObjList[heap].prev == &mmesa->TexObjList[heap] ) {
-	    if ( mmesa->mach64Screen->IsPCI ) {
-	       fprintf( stderr, "%s: upload texture failure on local "
-			"texture heaps, tex0 sz=%d  tex1 sz=%d\n", __FUNCTION__, 
-			t0->size, t1->size );
-	       return;
-	    } else if ( heap == MACH64_CARD_HEAP ) {
-	       /* If only one allocation succeeded, start over again in AGP */
-	       if (t0->memBlock) {
-		  mmFreeMem( t0->memBlock );
-	          t0->memBlock = NULL;
-	       }
-	       if (t1->memBlock) {
-		  mmFreeMem( t1->memBlock );
-	          t1->memBlock = NULL;
-	       }
-	       heap = t0->heap = t1->heap = MACH64_AGP_HEAP;
-	       continue;
-	    } else {
-	      int i;
-	       fprintf( stderr, "%s: upload texture failure on %s"
-			"AGP texture heaps, tex0 sz=%d  tex1 sz=%d\n", __FUNCTION__,
-			mmesa->firstTexHeap == MACH64_CARD_HEAP ? "both local and " : "",
-			t0->size, t1->size );
-	       for ( i = mmesa->firstTexHeap ; i < mmesa->lastTexHeap ; i++ ) {
-		  mach64PrintLocalLRU( mmesa, i );
-	          mmDumpMemInfo( mmesa->texHeap[i] );
-	       }
-	       exit(-1);
-	       return;
-	    }
-	 }
+      if ( ret == -1 && heap == MACH64_CARD_HEAP ) {
+	 /* Try AGP if local memory failed */
+	 heap = MACH64_AGP_HEAP;
 
-	 mach64SwapOutTexObj( mmesa, mmesa->TexObjList[heap].prev );
-	 
-	 if (!t0->memBlock)
-	    t0->memBlock = mmAllocMem( mmesa->texHeap[heap], t0->size, 6, 0 );
-	 if (!t1->memBlock)
-	    t1->memBlock = mmAllocMem( mmesa->texHeap[heap], t1->size, 6, 0 );
+	 ret = mach64AllocateMultiTex( mmesa, u0, u1, heap, GL_TRUE );
+      }
+
+      if ( ret == -1 ) {
+	 /* FIXME:
+	  * Swap out all textures from the AGP heap and re-run allocation, this
+	  * should succeed in all cases.
+	  */
+	 fprintf( stderr, "%s: upload multi-texture failure, sz0=%d sz1=%d\n",
+		  __FUNCTION__, t0->base.totalSize, t1->base.totalSize );
+	 exit(-1);
       }
 
       /* Set the base offset of the texture image */
-      t0->offset = mmesa->mach64Screen->texOffset[heap] + t0->memBlock->ofs;
-      t1->offset = mmesa->mach64Screen->texOffset[heap] + t1->memBlock->ofs;
+      t0->bufAddr = mmesa->mach64Screen->texOffset[heap] + t0->base.memBlock->ofs;
+      t1->bufAddr = mmesa->mach64Screen->texOffset[heap] + t1->base.memBlock->ofs;
 
       /* Force loading the new state into the hardware */
       mmesa->dirty |= (MACH64_UPLOAD_SCALE_3D_CNTL |
@@ -816,42 +464,43 @@ void mach64UploadMultiTexImages( mach64ContextPtr mmesa,
    }
 
    /* Let the world know we've used this memory recently */
-   mach64UpdateTexLRU( mmesa, t0 );
-   mach64UpdateTexLRU( mmesa, t1 );
+   driUpdateTextureLRU( (driTextureObject *) t0 );
+   driUpdateTextureLRU( (driTextureObject *) t1 );
 
    /* Upload any images that are new */
-   if ( t0->dirty ) {
+   if ( t0->base.dirty_images[0] ) {
+      const GLint j0 = t0->base.tObj->BaseLevel;
       if (t0->heap == MACH64_AGP_HEAP) {
 	 /* Need to make sure any vertex buffers in the queue complete */
 	 mach64WaitForIdleLocked( mmesa );
-	 mach64UploadAGPSubImage( mmesa, t0, t0->tObj->BaseLevel, 0, 0,
-				    t0->tObj->Image[0][t0->tObj->BaseLevel]->Width,
-				    t0->tObj->Image[0][t0->tObj->BaseLevel]->Height );
+	 mach64UploadAGPSubImage( mmesa, t0, j0, 0, 0,
+				    t0->base.tObj->Image[0][j0]->Width,
+				    t0->base.tObj->Image[0][j0]->Height );
       } else {
-	 mach64UploadLocalSubImage( mmesa, t0, t0->tObj->BaseLevel, 0, 0,
-				    t0->tObj->Image[0][t0->tObj->BaseLevel]->Width,
-				    t0->tObj->Image[0][t0->tObj->BaseLevel]->Height );
+	 mach64UploadLocalSubImage( mmesa, t0, j0, 0, 0,
+				    t0->base.tObj->Image[0][j0]->Width,
+				    t0->base.tObj->Image[0][j0]->Height );
       }
       mmesa->setup.tex_cntl |= MACH64_TEX_CACHE_FLUSH;
+      t0->base.dirty_images[0] = 0;
    }
-   if ( t1->dirty ) {
+   if ( t1->base.dirty_images[0] ) {
+      const GLint j1 = t1->base.tObj->BaseLevel;
       if (t1->heap == MACH64_AGP_HEAP) {
 	 /* Need to make sure any vertex buffers in the queue complete */
 	 mach64WaitForIdleLocked( mmesa );
-	 mach64UploadAGPSubImage( mmesa, t1, t1->tObj->BaseLevel, 0, 0,
-			       t1->tObj->Image[0][t1->tObj->BaseLevel]->Width,
-			       t1->tObj->Image[0][t1->tObj->BaseLevel]->Height );
+	 mach64UploadAGPSubImage( mmesa, t1, j1, 0, 0,
+			       t1->base.tObj->Image[0][j1]->Width,
+			       t1->base.tObj->Image[0][j1]->Height );
       } else {
-	 mach64UploadLocalSubImage( mmesa, t1, t1->tObj->BaseLevel, 0, 0,
-			       t1->tObj->Image[0][t1->tObj->BaseLevel]->Width,
-			       t1->tObj->Image[0][t1->tObj->BaseLevel]->Height );
+	 mach64UploadLocalSubImage( mmesa, t1, j1, 0, 0,
+			       t1->base.tObj->Image[0][j1]->Width,
+			       t1->base.tObj->Image[0][j1]->Height );
       }
       
       mmesa->setup.tex_cntl |= MACH64_TEX_CACHE_FLUSH;
+      t1->base.dirty_images[0] = 0;
    }
 
    mmesa->dirty |= MACH64_UPLOAD_TEXTURE;
-
-   t0->dirty = 0;
-   t1->dirty = 0;
 }