intel/perf: fix performance counters availability after glFinish

Currently Linux kernel gathers performance counters at fixed
intervals (~5-10ms), so if application uses AMD_performance_monitor
extension and immediately after glFinish() asks GL driver for HW
performance counter values it might not get any data (values == 0).

Fix this by moving the "read counters from kernel" code from
"is query ready" to "get counter values" callback with a loop around
it. Unfortunately it means that the "read counters from kernel"
code can spin for up to 10ms.

Ideally kernel should gather performance counters whenever we ask
it for counter values, but for now we have deal with what we have.

Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Cc: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5788>
(cherry picked from commit 2fbab5a1b3998854203c43560a47beac4e1897b7)

6 files changed, 20 insertions, 17 deletions

diff --git a/.pick_status.json b/.pick_status.json
index 6c0a7f30020..2b12351b344 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -3424,7 +3424,7 @@
         "description": "intel/perf: fix performance counters availability after glFinish",
         "nominated": true,
         "nomination_type": 0,
-        "resolution": 0,
+        "resolution": 1,
         "master_sha": null,
         "because_sha": null
     },
diff --git a/src/gallium/drivers/iris/iris_monitor.c b/src/gallium/drivers/iris/iris_monitor.c
index 838c0891b5f..730fae27182 100644
--- a/src/gallium/drivers/iris/iris_monitor.c
+++ b/src/gallium/drivers/iris/iris_monitor.c
@@ -355,7 +355,7 @@ iris_get_monitor_result(struct pipe_context *ctx,
    assert(gen_perf_is_query_ready(perf_ctx, monitor->query, batch));
 
    unsigned bytes_written;
-   gen_perf_get_query_data(perf_ctx, monitor->query,
+   gen_perf_get_query_data(perf_ctx, monitor->query, batch,
                            monitor->result_size,
                            (unsigned*) monitor->result_buffer,
                            &bytes_written);
diff --git a/src/gallium/drivers/iris/iris_performance_query.c b/src/gallium/drivers/iris/iris_performance_query.c
index 0192635601a..73b25081119 100644
--- a/src/gallium/drivers/iris/iris_performance_query.c
+++ b/src/gallium/drivers/iris/iris_performance_query.c
@@ -214,7 +214,8 @@ iris_get_perf_query_data(struct pipe_context *pipe,
    struct gen_perf_query_object *obj = perf_query->query;
    struct gen_perf_context *perf_ctx = ice->perf_ctx;
 
-   gen_perf_get_query_data(perf_ctx, obj, data_size, data, bytes_written);
+   gen_perf_get_query_data(perf_ctx, obj, &ice->batches[IRIS_BATCH_RENDER],
+         data_size, data, bytes_written);
 }
 
 void
diff --git a/src/intel/perf/gen_perf_query.c b/src/intel/perf/gen_perf_query.c
index 57b01203a6f..b9744913b16 100644
--- a/src/intel/perf/gen_perf_query.c
+++ b/src/intel/perf/gen_perf_query.c
@@ -1061,17 +1061,6 @@ gen_perf_wait_query(struct gen_perf_context *perf_ctx,
       perf_cfg->vtbl.batchbuffer_flush(perf_ctx->ctx, __FILE__, __LINE__);
 
    perf_cfg->vtbl.bo_wait_rendering(bo);
-
-   /* Due to a race condition between the OA unit signaling report
-    * availability and the report actually being written into memory,
-    * we need to wait for all the reports to come in before we can
-    * read them.
-    */
-   if (query->queryinfo->kind == GEN_PERF_QUERY_TYPE_OA ||
-       query->queryinfo->kind == GEN_PERF_QUERY_TYPE_RAW) {
-      while (!read_oa_samples_for_query(perf_ctx, query, current_batch))
-         ;
-   }
 }
 
 bool
@@ -1087,8 +1076,8 @@ gen_perf_is_query_ready(struct gen_perf_context *perf_ctx,
       return (query->oa.results_accumulated ||
               (query->oa.bo &&
                !perf_cfg->vtbl.batch_references(current_batch, query->oa.bo) &&
-               !perf_cfg->vtbl.bo_busy(query->oa.bo) &&
-               read_oa_samples_for_query(perf_ctx, query, current_batch)));
+               !perf_cfg->vtbl.bo_busy(query->oa.bo)));
+
    case GEN_PERF_QUERY_TYPE_PIPELINE:
       return (query->pipeline_stats.bo &&
               !perf_cfg->vtbl.batch_references(current_batch, query->pipeline_stats.bo) &&
@@ -1513,6 +1502,7 @@ get_pipeline_stats_data(struct gen_perf_context *perf_ctx,
 void
 gen_perf_get_query_data(struct gen_perf_context *perf_ctx,
                         struct gen_perf_query_object *query,
+                        void *current_batch,
                         int data_size,
                         unsigned *data,
                         unsigned *bytes_written)
@@ -1524,6 +1514,17 @@ gen_perf_get_query_data(struct gen_perf_context *perf_ctx,
    case GEN_PERF_QUERY_TYPE_OA:
    case GEN_PERF_QUERY_TYPE_RAW:
       if (!query->oa.results_accumulated) {
+         /* Due to the sampling frequency of the OA buffer by the i915-perf
+          * driver, there can be a 5ms delay between the Mesa seeing the query
+          * complete and i915 making all the OA buffer reports available to us.
+          * We need to wait for all the reports to come in before we can do
+          * the post processing removing unrelated deltas.
+          * There is a i915-perf series to address this issue, but it's
+          * not been merged upstream yet.
+          */
+         while (!read_oa_samples_for_query(perf_ctx, query, current_batch))
+            ;
+
          read_gt_frequency(perf_ctx, query);
          uint32_t *begin_report = query->oa.map;
          uint32_t *end_report = query->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
diff --git a/src/intel/perf/gen_perf_query.h b/src/intel/perf/gen_perf_query.h
index a0246501f76..d064a5d0669 100644
--- a/src/intel/perf/gen_perf_query.h
+++ b/src/intel/perf/gen_perf_query.h
@@ -76,6 +76,7 @@ void gen_perf_delete_query(struct gen_perf_context *perf_ctx,
                            struct gen_perf_query_object *query);
 void gen_perf_get_query_data(struct gen_perf_context *perf_ctx,
                              struct gen_perf_query_object *query,
+                             void *current_batch,
                              int data_size,
                              unsigned *data,
                              unsigned *bytes_written);
diff --git a/src/mesa/drivers/dri/i965/brw_performance_query.c b/src/mesa/drivers/dri/i965/brw_performance_query.c
index 7037e25679a..ebe2c78f275 100644
--- a/src/mesa/drivers/dri/i965/brw_performance_query.c
+++ b/src/mesa/drivers/dri/i965/brw_performance_query.c
@@ -323,7 +323,7 @@ brw_get_perf_query_data(struct gl_context *ctx,
     */
    assert(o->Ready);
 
-   gen_perf_get_query_data(brw->perf_ctx, obj,
+   gen_perf_get_query_data(brw->perf_ctx, obj, &brw->batch,
                            data_size, data, bytes_written);
 }