[Mesa-dev] [PATCH] anv: Don't place scratch buffers above the 32-bit boundary

[Jason Ekstrand]

This fixes rendering corruptions in DOOM.  Hopefully, it will also make
Jenkins a bit more stable as we've been seeing some random failures and
GPU hangs ever since turning on 48bit.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100620
Cc: "17.1" <mesa-stable at lists.freedesktop.org>
---
 src/intel/vulkan/anv_allocator.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/src/intel/vulkan/anv_allocator.c b/src/intel/vulkan/anv_allocator.c
index 78327df..28bfac4 100644
--- a/src/intel/vulkan/anv_allocator.c
+++ b/src/intel/vulkan/anv_allocator.c
@@ -994,6 +994,22 @@ anv_scratch_pool_alloc(struct anv_device *device, struct anv_scratch_pool *pool,
 
    anv_bo_init_new(&bo->bo, device, size);
 
+   /* Even though the Scratch base pointers in 3DSTATE_*S are 64 bits, they
+    * are still relative to the general state base address.  When we emit
+    * STATE_BASE_ADDRESS, we set general state base address to 0 and the size
+    * to the maximum (1 page under 4GB).  This allows us to just place the
+    * scratch buffers anywhere we wish in the bottom 32 bits of address space
+    * and just set the scratch base pointer in 3DSTATE_*S using a relocation.
+    * However, in order to do so, we need to ensure that the kernel does not
+    * place the scratch BO above the 32-bit boundary.
+    *
+    * NOTE: Technically, it can't go "anywhere" because the top page is off
+    * limits.  However, it will never end up getting placed that high because
+    * the surface state and general state buffers will get placed first and
+    * the kernel likes to work top-down.
+    */
+   bo->bo.flags &= ~EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
    /* Set the exists last because it may be read by other threads */
    __sync_synchronize();
    bo->exists = true;
-- 
2.5.0.400.gff86faf