util/vma: Add "nospan"

Add a way to request that allocations do not span a specified PoT
boundary.  This can be used, for ex, to ensure that allocations do
not span 4GB boundaries to work around hw/fw bugs.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20263>

src/util/vma.c

@@ -24,6 +24,7 @@
 #include <stdlib.h>
 #include <inttypes.h>
 
+#include "util/macros.h"
 #include "util/u_math.h"
 #include "util/vma.h"
 
@@ -52,6 +53,9 @@ util_vma_heap_init(struct util_vma_heap *heap,
 
    /* Default to using high addresses */
    heap->alloc_high = true;
+
+   /* Default to not having a nospan alignment */
+   heap->nospan_shift = 0;
 }
 
 void
@@ -158,6 +162,14 @@ util_vma_heap_alloc(struct util_vma_heap *heap,
 
    util_vma_heap_validate(heap);
 
+   /* The requested alignment should not be stronger than the block/nospan
+    * alignment.
+    */
+   if (heap->nospan_shift) {
+      assert(ALIGN(BITFIELD64_BIT(heap->nospan_shift), alignment) ==
+            BITFIELD64_BIT(heap->nospan_shift));
+   }
+
    if (heap->alloc_high) {
       util_vma_foreach_hole_safe(hole, heap) {
          if (size > hole->size)
@@ -171,6 +183,18 @@ util_vma_heap_alloc(struct util_vma_heap *heap,
           */
          uint64_t offset = (hole->size - size) + hole->offset;
 
+         if (heap->nospan_shift) {
+            uint64_t end = offset + size - 1;
+            if ((end >> heap->nospan_shift) != (offset >> heap->nospan_shift)) {
+               /* can we shift the offset down and still fit in the current hole? */
+               end &= ~BITFIELD64_MASK(heap->nospan_shift);
+               assert(end >= size);
+               offset -= size;
+               if (offset < hole->offset)
+                  continue;
+            }
+         }
+
          /* Align the offset.  We align down and not up because we are
           * allocating from the top of the hole and not the bottom.
           */
@@ -200,6 +224,16 @@ util_vma_heap_alloc(struct util_vma_heap *heap,
             offset += pad;
          }
 
+         if (heap->nospan_shift) {
+            uint64_t end = offset + size - 1;
+            if ((end >> heap->nospan_shift) != (offset >> heap->nospan_shift)) {
+               /* can we shift the offset up and still fit in the current hole? */
+               offset = end & ~BITFIELD64_MASK(heap->nospan_shift);
+               if ((offset + size) > (hole->offset + hole->size))
+                  continue;
+            }
+         }
+
          util_vma_hole_alloc(heap, hole, offset, size);
          util_vma_heap_validate(heap);
          return offset;
@@ -325,7 +359,7 @@ util_vma_heap_print(struct util_vma_heap *heap, FILE *fp,
 
    uint64_t total_free = 0;
    util_vma_foreach_hole(hole, heap) {
-      fprintf(fp, "%s    hole: offset = %"PRIu64" (0x%"PRIx64", "
+      fprintf(fp, "%s    hole: offset = %"PRIu64" (0x%"PRIx64"), "
               "size = %"PRIu64" (0x%"PRIx64")\n",
               tab, hole->offset, hole->offset, hole->size, hole->size);
       total_free += hole->size;

src/util/vma.h

@@ -44,6 +44,13 @@ struct util_vma_heap {
     * Default is true.
     */
    bool alloc_high;
+
+   /**
+    * If non-zero, util_vma_heap_alloc will avoid allocating regions which
+    * span (1 << nospan_shift) ranges.  For example, to avoid allocations
+    * which straddle 4GB boundaries, use nospan_shift=log2(4GB)
+    */
+   unsigned nospan_shift;
 };
 
 void util_vma_heap_init(struct util_vma_heap *heap,