]> git.karo-electronics.de Git - linux-beck.git/commitdiff
drm/i915/execlists: Reset RING registers upon resume
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 21 Sep 2016 13:51:08 +0000 (14:51 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 21 Sep 2016 15:57:48 +0000 (16:57 +0100)
There is a disparity in the context image saved to disk and our own
bookkeeping - that is we presume the RING_HEAD and RING_TAIL match our
stored ce->ring->tail value. However, as we emit WA_TAIL_DWORDS into the
ring but may not tell the GPU about them, the GPU may be lagging behind
our bookkeeping. Upon hibernation we do not save stolen pages, presuming
that their contents are volatile. This means that although we start
writing into the ring at tail, the GPU starts executing from its HEAD
and there may be some garbage in between and so the GPU promptly hangs
upon resume.

Testcase: igt/gem_exec_suspend/basic-S4
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96526
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20160921135108.29574-3-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/intel_lrc.c

index 251143361f31e5bf83a3d4df810d222d657f8731..39417b77bff2da19d85ed6d24190cd0e57543f58 100644 (file)
@@ -2129,30 +2129,42 @@ error_deref_obj:
 
 void intel_lr_context_resume(struct drm_i915_private *dev_priv)
 {
-       struct i915_gem_context *ctx = dev_priv->kernel_context;
        struct intel_engine_cs *engine;
+       struct i915_gem_context *ctx;
+
+       /* Because we emit WA_TAIL_DWORDS there may be a disparity
+        * between our bookkeeping in ce->ring->head and ce->ring->tail and
+        * that stored in context. As we only write new commands from
+        * ce->ring->tail onwards, everything before that is junk. If the GPU
+        * starts reading from its RING_HEAD from the context, it may try to
+        * execute that junk and die.
+        *
+        * So to avoid that we reset the context images upon resume. For
+        * simplicity, we just zero everything out.
+        */
+       list_for_each_entry(ctx, &dev_priv->context_list, link) {
+               for_each_engine(engine, dev_priv) {
+                       struct intel_context *ce = &ctx->engine[engine->id];
+                       u32 *reg;
 
-       for_each_engine(engine, dev_priv) {
-               struct intel_context *ce = &ctx->engine[engine->id];
-               void *vaddr;
-               uint32_t *reg_state;
-
-               if (!ce->state)
-                       continue;
-
-               vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB);
-               if (WARN_ON(IS_ERR(vaddr)))
-                       continue;
+                       if (!ce->state)
+                               continue;
 
-               reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
+                       reg = i915_gem_object_pin_map(ce->state->obj,
+                                                     I915_MAP_WB);
+                       if (WARN_ON(IS_ERR(reg)))
+                               continue;
 
-               reg_state[CTX_RING_HEAD+1] = 0;
-               reg_state[CTX_RING_TAIL+1] = 0;
+                       reg += LRC_STATE_PN * PAGE_SIZE / sizeof(*reg);
+                       reg[CTX_RING_HEAD+1] = 0;
+                       reg[CTX_RING_TAIL+1] = 0;
 
-               ce->state->obj->dirty = true;
-               i915_gem_object_unpin_map(ce->state->obj);
+                       ce->state->obj->dirty = true;
+                       i915_gem_object_unpin_map(ce->state->obj);
 
-               ce->ring->head = 0;
-               ce->ring->tail = 0;
+                       ce->ring->head = ce->ring->tail = 0;
+                       ce->ring->last_retired_head = -1;
+                       intel_ring_update_space(ce->ring);
+               }
        }
 }