drm/i915: Prevent machine death on Ivybridge context switching
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 13 Apr 2016 16:35:10 +0000 (17:35 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 14 Apr 2016 09:45:40 +0000 (10:45 +0100)
Two concurrent writes into the same register cacheline has the chance of
killing the machine on Ivybridge and other gen7. This includes LRI
emitted from the command parser.  The MI_SET_CONTEXT itself serves as
serialising barrier and prevents the pair of register writes in the first
packet from triggering the fault.  However, if a second switch-context
immediately occurs then we may have two adjacent blocks of LRI to the
same registers which may then trigger the hang. To counteract this we
need to insert a delay after the second register write using SRM.

This is easiest to reproduce with something like
igt/gem_ctx_switch/interruptible that triggers back-to-back context
switches (with no operations in between them in the command stream,
which requires the execbuf operation to be interrupted after the
MI_SET_CONTEXT) but can be observed sporadically elsewhere when running
interruptible igt. No reports from the wild though, so it must be of low
enough frequency that no one has correlated the random machine freezes
with i915.ko

The issue was introduced with
commit 2c550183476dfa25641309ae9a28d30feed14379 [v3.19]
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Dec 16 10:02:27 2014 +0000

    drm/i915: Disable PSMI sleep messages on all rings around context switches

Testcase: igt/gem_ctx_switch/render-interruptible #ivb
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel@ffwll.ch>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: stable@vger.kernel.org
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1460565315-7748-11-git-send-email-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_context.c

index 91028d9c626937ca520bf25753707eabc217207f..453b655e86fc3d6e78334e68549e2170c84fe765 100644 (file)
@@ -539,7 +539,7 @@ mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
 
        len = 4;
        if (INTEL_INFO(engine->dev)->gen >= 7)
-               len += 2 + (num_rings ? 4*num_rings + 2 : 0);
+               len += 2 + (num_rings ? 4*num_rings + 6 : 0);
 
        ret = intel_ring_begin(req, len);
        if (ret)
@@ -579,6 +579,7 @@ mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
        if (INTEL_INFO(engine->dev)->gen >= 7) {
                if (num_rings) {
                        struct intel_engine_cs *signaller;
+                       i915_reg_t last_reg = {}; /* keep gcc quiet */
 
                        intel_ring_emit(engine,
                                        MI_LOAD_REGISTER_IMM(num_rings));
@@ -586,11 +587,19 @@ mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
                                if (signaller == engine)
                                        continue;
 
-                               intel_ring_emit_reg(engine,
-                                                   RING_PSMI_CTL(signaller->mmio_base));
+                               last_reg = RING_PSMI_CTL(signaller->mmio_base);
+                               intel_ring_emit_reg(engine, last_reg);
                                intel_ring_emit(engine,
                                                _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
                        }
+
+                       /* Insert a delay before the next switch! */
+                       intel_ring_emit(engine,
+                                       MI_STORE_REGISTER_MEM |
+                                       MI_SRM_LRM_GLOBAL_GTT);
+                       intel_ring_emit_reg(engine, last_reg);
+                       intel_ring_emit(engine, engine->scratch.gtt_offset);
+                       intel_ring_emit(engine, MI_NOOP);
                }
                intel_ring_emit(engine, MI_ARB_ON_OFF | MI_ARB_ENABLE);
        }