drm/i915: correctly restore fences with objects attached
authorDaniel Vetter <daniel.vetter@ffwll.ch>
Wed, 17 Jul 2013 12:51:28 +0000 (14:51 +0200)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 18 Jul 2013 22:08:16 +0000 (00:08 +0200)
To avoid stalls we delay tiling changes and especially hold of
committing the new fence state for as long as possible.
Synchronization points are in the execbuf code and in our gtt fault
handler.

Unfortunately we've missed that tricky detail when adding proper fence
restore code in

commit 19b2dbde5732170a03bd82cc8bd442cf88d856f7
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Jun 12 10:15:12 2013 +0100

    drm/i915: Restore fences after resume and GPU resets

The result was that we've restored fences for objects with no tiling,
since the object<->fence link still existed after resume. Now that
wouldn't have been too bad since any subsequent access would have
fixed things up, but if we've changed from tiled to untiled real havoc
happened:

The tiling stride is stored -1 in the fence register, so a stride of 0
resulted in all 1s in the top 32bits, and so a completely bogus fence
spanning everything from the start of the object to the top of the
GTT. The tell-tale in the register dumps looks like:

                 FENCE START 2: 0x0214d001
                 FENCE END 2: 0xfffff3ff

Bit 11 isn't set since the hw doesn't store it, even when writing all
1s (at least on my snb here).

To prevent such a gaffle in the future add a sanity check for fences
with an untiled object attached in i915_gem_write_fence.

v2: Fix the WARN, spotted by Chris.

v3: Trying to reuse get_fences looked ugly and obfuscated the code.
Instead reuse update_fence and to make it really dtrt also move the
fence dirty state clearing into update_fence.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Stéphane Marchesin <marcheu@chromium.org>
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=60530
Cc: stable@vger.kernel.org (for 3.10 only)
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Matthew Garrett <matthew.garrett@nebula.com>
Tested-by: Björn Bidar <theodorstormgrade@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem.c

index 97afd2639fb63a1e240fce2c6f97f53fd9f82e98..d9e2208cfe98f2fdcb3957bf4ee13f0497e06dfb 100644 (file)
@@ -2258,7 +2258,17 @@ void i915_gem_restore_fences(struct drm_device *dev)
 
        for (i = 0; i < dev_priv->num_fence_regs; i++) {
                struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
-               i915_gem_write_fence(dev, i, reg->obj);
+
+               /*
+                * Commit delayed tiling changes if we have an object still
+                * attached to the fence, otherwise just clear the fence.
+                */
+               if (reg->obj) {
+                       i915_gem_object_update_fence(reg->obj, reg,
+                                                    reg->obj->tiling_mode);
+               } else {
+                       i915_gem_write_fence(dev, i, NULL);
+               }
        }
 }
 
@@ -2795,6 +2805,10 @@ static void i915_gem_write_fence(struct drm_device *dev, int reg,
        if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
                mb();
 
+       WARN(obj && (!obj->stride || !obj->tiling_mode),
+            "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
+            obj->stride, obj->tiling_mode);
+
        switch (INTEL_INFO(dev)->gen) {
        case 7:
        case 6:
@@ -2836,6 +2850,7 @@ static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
                fence->obj = NULL;
                list_del_init(&fence->lru_list);
        }
+       obj->fence_dirty = false;
 }
 
 static int
@@ -2965,7 +2980,6 @@ i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
                return 0;
 
        i915_gem_object_update_fence(obj, reg, enable);
-       obj->fence_dirty = false;
 
        return 0;
 }