[Mesa-dev] [PATCH] i965: Drop need_workaround_flush flag; always do post-sync workaround.

[Kenneth Graunke]

The need_workaround_flush flag was an attempt to avoid doing the
post-sync non-zero workaround flush unless it was actually necessary.

Getting the tracking right is tricky; we've messed it up several times.

The cost of getting this wrong is high: almost certainly a GPU hang.
The cost of always doing the flush is low: it's just another flush,
and on Sandybridge, we're doing piles of them anyway.  And every time
we've added more post-sync non-zero workarounds, the system has become
more stable.

Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
---
 src/mesa/drivers/dri/i965/brw_context.h       |  1 -
 src/mesa/drivers/dri/i965/brw_draw.c          |  3 ---
 src/mesa/drivers/dri/i965/gen6_blorp.cpp      |  6 ------
 src/mesa/drivers/dri/i965/intel_batchbuffer.c | 12 ------------
 4 files changed, 22 deletions(-)

People are still getting daily GPU hangs...in the past, adding post-sync
workarounds has helped a lot.  Emitting more of them should be safe to do
(the post-sync workaround doesn't require any preceding PIPE_CONTROLs),
and may help.  Worth a try at least...?

diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index 8d098e6..bba7d97 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -870,7 +870,6 @@ struct intel_batchbuffer {
    drm_intel_bo *last_bo;
    /** BO for post-sync nonzero writes for gen6 workaround. */
    drm_intel_bo *workaround_bo;
-   bool need_workaround_flush;
 
    struct cached_batch_item *cached_items;
 
diff --git a/src/mesa/drivers/dri/i965/brw_draw.c b/src/mesa/drivers/dri/i965/brw_draw.c
index 39da953..e4c24e9 100644
--- a/src/mesa/drivers/dri/i965/brw_draw.c
+++ b/src/mesa/drivers/dri/i965/brw_draw.c
@@ -276,9 +276,6 @@ static void brw_emit_prim(struct brw_context *brw,
    OUT_BATCH(base_vertex_location);
    ADVANCE_BATCH();
 
-   /* Only used on Sandybridge; harmless to set elsewhere. */
-   brw->batch.need_workaround_flush = true;
-
    if (brw->always_flush_cache) {
       intel_batchbuffer_emit_mi_flush(brw);
    }
diff --git a/src/mesa/drivers/dri/i965/gen6_blorp.cpp b/src/mesa/drivers/dri/i965/gen6_blorp.cpp
index 90b9fbb..01db084 100644
--- a/src/mesa/drivers/dri/i965/gen6_blorp.cpp
+++ b/src/mesa/drivers/dri/i965/gen6_blorp.cpp
@@ -1010,9 +1010,6 @@ gen6_blorp_emit_primitive(struct brw_context *brw,
    OUT_BATCH(0);
    OUT_BATCH(0);
    ADVANCE_BATCH();
-
-   /* Only used on Sandybridge; harmless to set elsewhere. */
-   brw->batch.need_workaround_flush = true;
 }
 
 /**
@@ -1037,9 +1034,6 @@ gen6_blorp_exec(struct brw_context *brw,
 
    uint32_t prog_offset = params->get_wm_prog(brw, &prog_data);
 
-   /* Emit workaround flushes when we switch from drawing to blorping. */
-   brw->batch.need_workaround_flush = true;
-
    gen6_emit_3dstate_multisample(brw, params->dst.num_samples);
    gen6_emit_3dstate_sample_mask(brw,
                                  params->dst.num_samples > 1 ?
diff --git a/src/mesa/drivers/dri/i965/intel_batchbuffer.c b/src/mesa/drivers/dri/i965/intel_batchbuffer.c
index caec312..efa4ab6 100644
--- a/src/mesa/drivers/dri/i965/intel_batchbuffer.c
+++ b/src/mesa/drivers/dri/i965/intel_batchbuffer.c
@@ -70,8 +70,6 @@ intel_batchbuffer_init(struct brw_context *brw)
 						      4096, 4096);
    }
 
-   brw->batch.need_workaround_flush = true;
-
    if (!brw->has_llc) {
       brw->batch.cpu_map = malloc(BATCH_SZ);
       brw->batch.map = brw->batch.cpu_map;
@@ -206,11 +204,6 @@ brw_new_batch(struct brw_context *brw)
 
    brw->state.dirty.brw |= BRW_NEW_BATCH;
 
-   /* Assume that the last command before the start of our batch was a
-    * primitive, for safety.
-    */
-   brw->batch.need_workaround_flush = true;
-
    brw->state_batch_count = 0;
 
    brw->ib.type = -1;
@@ -610,17 +603,12 @@ gen7_emit_cs_stall_flush(struct brw_context *brw)
 void
 intel_emit_post_sync_nonzero_flush(struct brw_context *brw)
 {
-   if (!brw->batch.need_workaround_flush)
-      return;
-
    brw_emit_pipe_control_flush(brw,
                                PIPE_CONTROL_CS_STALL |
                                PIPE_CONTROL_STALL_AT_SCOREBOARD);
 
    brw_emit_pipe_control_write(brw, PIPE_CONTROL_WRITE_IMMEDIATE,
                                brw->batch.workaround_bo, 0, 0, 0);
-
-   brw->batch.need_workaround_flush = false;
 }
 
 /* Emit a pipelined flush to either flush render and texture cache for
-- 
1.8.5.2