i965: Embrace "unlimited" GTT mmap support

From about kernel 4.9, GTT mmaps are virtually unlimited. A new
parameter, I915_PARAM_MMAP_GTT_VERSION, is added to advertise the
feature so query it and use it to avoid limiting tiled allocations to
only fit within the mappable aperture.

A couple of caveats:

 - fence support is still limited by stride to 262144 and the stride
needs to be a multiple of tile_width (as before, and same limitation as
the current 3D pipeline in hardware)

 - the max_gtt_map_object_size forcing untiled may be hiding a few bugs
in handling of large objects, though none were spotted in piglits.

See kernel commit 4cc6907501ed ("drm/i915: Add I915_PARAM_MMAP_GTT_VERSION
to advertise unlimited mmaps").

v2: Include some commentary on mmap virtual space vs CPU addressable
space.

Signed-off-by: Chris Wilson <[email protected]>
Cc: Kenneth Graunke <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
Reviewed-by: Daniel Vetter <[email protected]>

4 files changed, 53 insertions, 16 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
index d6f0d7bbcb3..13643930e38 100644
--- a/src/mesa/drivers/dri/i965/brw_context.c
+++ b/src/mesa/drivers/dri/i965/brw_context.c
@@ -1044,21 +1044,6 @@ brwCreateContext(gl_api api,
    brw->urb.max_ds_entries = devinfo->urb.max_ds_entries;
    brw->urb.max_gs_entries = devinfo->urb.max_gs_entries;
 
-   /* Estimate the size of the mappable aperture into the GTT.  There's an
-    * ioctl to get the whole GTT size, but not one to get the mappable subset.
-    * It turns out it's basically always 256MB, though some ancient hardware
-    * was smaller.
-    */
-   uint32_t gtt_size = 256 * 1024 * 1024;
-
-   /* We don't want to map two objects such that a memcpy between them would
-    * just fault one mapping in and then the other over and over forever.  So
-    * we would need to divide the GTT size by 2.  Additionally, some GTT is
-    * taken up by things like the framebuffer and the ringbuffer and such, so
-    * be more conservative.
-    */
-   brw->max_gtt_map_object_size = gtt_size / 4;
-
    if (brw->gen == 6)
       brw->urb.gs_present = false;
 
@@ -1069,6 +1054,8 @@ brwCreateContext(gl_api api,
 
    brw->predicate.state = BRW_PREDICATE_STATE_RENDER;
 
+   brw->max_gtt_map_object_size = screen->max_gtt_map_object_size;
+
    brw->use_resource_streamer = screen->has_resource_streamer &&
       (env_var_as_boolean("INTEL_USE_HW_BT", false) ||
        env_var_as_boolean("INTEL_USE_GATHER", false));
diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index e0f70007478..1a4efa3cb0c 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -853,7 +853,7 @@ struct brw_context
     */
    bool perf_debug;
 
-   uint32_t max_gtt_map_object_size;
+   uint64_t max_gtt_map_object_size;
 
    int gen;
    int gt;
diff --git a/src/mesa/drivers/dri/i965/intel_screen.c b/src/mesa/drivers/dri/i965/intel_screen.c
index 78766524e3c..cb007d7dc21 100644
--- a/src/mesa/drivers/dri/i965/intel_screen.c
+++ b/src/mesa/drivers/dri/i965/intel_screen.c
@@ -995,6 +995,17 @@ intel_get_boolean(struct intel_screen *screen, int param)
    return (intel_get_param(screen, param, &value) == 0) && value;
 }
 
+static int
+intel_get_integer(struct intel_screen *screen, int param)
+{
+   int value = -1;
+
+   if (intel_get_param(screen, param, &value) == 0)
+      return value;
+
+   return -1;
+}
+
 static void
 intelDestroyScreen(__DRIscreen * sPriv)
 {
@@ -1565,6 +1576,43 @@ __DRIconfig **intelInitScreen2(__DRIscreen *psp)
    if (INTEL_DEBUG & DEBUG_AUB)
       drm_intel_bufmgr_gem_set_aub_dump(intelScreen->bufmgr, true);
 
+#ifndef I915_PARAM_MMAP_GTT_VERSION
+#define I915_PARAM_MMAP_GTT_VERSION 40 /* XXX delete me with new libdrm */
+#endif
+   if (intel_get_integer(intelScreen, I915_PARAM_MMAP_GTT_VERSION) >= 1) {
+      /* Theorectically unlimited! At least for individual objects...
+       *
+       * Currently the entire (global) address space for all GTT maps is
+       * limited to 64bits. That is all objects on the system that are
+       * setup for GTT mmapping must fit within 64bits. An attempt to use
+       * one that exceeds the limit with fail in drm_intel_bo_map_gtt().
+       *
+       * Long before we hit that limit, we will be practically limited by
+       * that any single object must fit in physical memory (RAM). The upper
+       * limit on the CPU's address space is currently 48bits (Skylake), of
+       * which only 39bits can be physical memory. (The GPU itself also has
+       * a 48bit addressable virtual space.) We can fit over 32 million
+       * objects of the current maximum allocable size before running out
+       * of mmap space.
+       */
+      intelScreen->max_gtt_map_object_size = UINT64_MAX;
+   } else {
+      /* Estimate the size of the mappable aperture into the GTT.  There's an
+       * ioctl to get the whole GTT size, but not one to get the mappable subset.
+       * It turns out it's basically always 256MB, though some ancient hardware
+       * was smaller.
+       */
+      uint32_t gtt_size = 256 * 1024 * 1024;
+
+      /* We don't want to map two objects such that a memcpy between them would
+       * just fault one mapping in and then the other over and over forever.  So
+       * we would need to divide the GTT size by 2.  Additionally, some GTT is
+       * taken up by things like the framebuffer and the ringbuffer and such, so
+       * be more conservative.
+       */
+      intelScreen->max_gtt_map_object_size = gtt_size / 4;
+   }
+
    intelScreen->hw_has_swizzling = intel_detect_swizzling(intelScreen);
    intelScreen->hw_has_timestamp = intel_detect_timestamp(intelScreen);
 
diff --git a/src/mesa/drivers/dri/i965/intel_screen.h b/src/mesa/drivers/dri/i965/intel_screen.h
index 98f9d24901c..f62b39f7440 100644
--- a/src/mesa/drivers/dri/i965/intel_screen.h
+++ b/src/mesa/drivers/dri/i965/intel_screen.h
@@ -44,6 +44,8 @@ struct intel_screen
 
    __DRIscreen *driScrnPriv;
 
+   uint64_t max_gtt_map_object_size;
+
    bool no_hw;
 
    bool hw_has_swizzling;