drm/i915: Evict CS TLBs between batches

Running igt, I was encountering the invalid TLB bug on my 845g, despite
that it was using the CS workaround. Examining the w/a buffer in the
error state, showed that the copy from the user batch into the
workaround itself was suffering from the invalid TLB bug (the first
cacheline was broken with the first two words reversed). Time to try a
fresh approach. This extends the workaround to write into each page of
our scratch buffer in order to overflow the TLB and evict the invalid
entries. This could be refined to only do so after we update the GTT,
but for simplicity, we do it before each batch.

I suspect this supersedes our current workaround, but for safety keep
doing both.

v2: The magic number shall be 2.

This doesn't conclusively prove that it is the mythical TLB bug we've
been trying to workaround for so long, that it requires touching a number
of pages to prevent the corruption indicates to me that it is TLB
related, but the corruption (the reversed cacheline) is more subtle than
a TLB bug, where we would expect it to read the wrong page entirely.

Oh well, it prevents a reliable hang for me and so probably for others
as well.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: stable@vger.kernel.org
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>

1 files changed, 39 insertions, 27 deletions

diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index 16371a444426..2d068edd1adc 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -1363,54 +1363,66 @@ i965_dispatch_execbuffer(struct intel_engine_cs *ring,
 
 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
 #define I830_BATCH_LIMIT (256*1024)
+#define I830_TLB_ENTRIES (2)
+#define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
 static int
 i830_dispatch_execbuffer(struct intel_engine_cs *ring,
                                 u64 offset, u32 len,
                                 unsigned flags)
 {
+        u32 cs_offset = ring->scratch.gtt_offset;
         int ret;
 
-        if (flags & I915_DISPATCH_PINNED) {
-                ret = intel_ring_begin(ring, 4);
-                if (ret)
-                        return ret;
+        ret = intel_ring_begin(ring, 6);
+        if (ret)
+                return ret;
 
-                intel_ring_emit(ring, MI_BATCH_BUFFER);
-                intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
-                intel_ring_emit(ring, offset + len - 8);
-                intel_ring_emit(ring, MI_NOOP);
-                intel_ring_advance(ring);
-        } else {
-                u32 cs_offset = ring->scratch.gtt_offset;
+        /* Evict the invalid PTE TLBs */
+        intel_ring_emit(ring, COLOR_BLT_CMD | BLT_WRITE_RGBA);
+        intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096);
+        intel_ring_emit(ring, I830_TLB_ENTRIES << 16 | 4); /* load each page */
+        intel_ring_emit(ring, cs_offset);
+        intel_ring_emit(ring, 0xdeadbeef);
+        intel_ring_emit(ring, MI_NOOP);
+        intel_ring_advance(ring);
 
+        if ((flags & I915_DISPATCH_PINNED) == 0) {
                 if (len > I830_BATCH_LIMIT)
                         return -ENOSPC;
 
-                ret = intel_ring_begin(ring, 9+3);
+                ret = intel_ring_begin(ring, 6 + 2);
                 if (ret)
                         return ret;
-                /* Blit the batch (which has now all relocs applied) to the stable batch
-                 * scratch bo area (so that the CS never stumbles over its tlb
-                 * invalidation bug) ... */
-                intel_ring_emit(ring, XY_SRC_COPY_BLT_CMD |
-                                XY_SRC_COPY_BLT_WRITE_ALPHA |
-                                XY_SRC_COPY_BLT_WRITE_RGB);
-                intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_GXCOPY | 4096);
-                intel_ring_emit(ring, 0);
-                intel_ring_emit(ring, (DIV_ROUND_UP(len, 4096) << 16) | 1024);
+
+                /* Blit the batch (which has now all relocs applied) to the
+                 * stable batch scratch bo area (so that the CS never
+                 * stumbles over its tlb invalidation bug) ...
+                 */
+                intel_ring_emit(ring, SRC_COPY_BLT_CMD | BLT_WRITE_RGBA);
+                intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096);
+                intel_ring_emit(ring, DIV_ROUND_UP(len, 4096) << 16 | 1024);
                 intel_ring_emit(ring, cs_offset);
-                intel_ring_emit(ring, 0);
                 intel_ring_emit(ring, 4096);
                 intel_ring_emit(ring, offset);
+
                 intel_ring_emit(ring, MI_FLUSH);
+                intel_ring_emit(ring, MI_NOOP);
+                intel_ring_advance(ring);
 
                 /* ... and execute it. */
-                intel_ring_emit(ring, MI_BATCH_BUFFER);
-                intel_ring_emit(ring, cs_offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
-                intel_ring_emit(ring, cs_offset + len - 8);
-                intel_ring_advance(ring);
+                offset = cs_offset;
         }
 
+        ret = intel_ring_begin(ring, 4);
+        if (ret)
+                return ret;
+
+        intel_ring_emit(ring, MI_BATCH_BUFFER);
+        intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
+        intel_ring_emit(ring, offset + len - 8);
+        intel_ring_emit(ring, MI_NOOP);
+        intel_ring_advance(ring);
+
         return 0;
 }
 
@@ -2200,7 +2212,7 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
 
         /* Workaround batchbuffer to combat CS tlb bug. */
         if (HAS_BROKEN_CS_TLB(dev)) {
-                obj = i915_gem_alloc_object(dev, I830_BATCH_LIMIT);
+                obj = i915_gem_alloc_object(dev, I830_WA_SIZE);
                 if (obj == NULL) {
                         DRM_ERROR("Failed to allocate batch bo\n");
                         return -ENOMEM;