5 files changed, 43 insertions, 8 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/channel_gk20a.c b/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
index 9902cb5e..db44d11b 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/channel_gk20a.c
@@ -2468,7 +2468,7 @@ int gk20a_submit_channel_gpfifo(struct channel_gk20a *c,
                         return -EINVAL;
 
                 need_sync_framework = force_need_sync_fence ||
-                        gk20a_channel_sync_needs_sync_framework(c) ||
+                        gk20a_channel_sync_needs_sync_framework(g) ||
                         (flags & NVGPU_SUBMIT_GPFIFO_FLAGS_SYNC_FENCE &&
                         (flags & NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_WAIT ||
                          flags & NVGPU_SUBMIT_GPFIFO_FLAGS_FENCE_GET));
diff --git a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c
index ddd94a2d..ed83663d 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.c
@@ -972,9 +972,7 @@ struct gk20a_channel_sync *gk20a_channel_sync_create(struct channel_gk20a *c)
         return gk20a_channel_semaphore_create(c);
 }
 
-bool gk20a_channel_sync_needs_sync_framework(struct channel_gk20a *c)
+bool gk20a_channel_sync_needs_sync_framework(struct gk20a *g)
 {
-        if (gk20a_platform_has_syncpoints(c->g))
-                return false;
-        return true;
+        return !gk20a_platform_has_syncpoints(g);
 }
diff --git a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h
index 8b1f85a1..4efd1b76 100644
--- a/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h
+++ b/drivers/gpu/nvgpu/gk20a/channel_sync_gk20a.h
@@ -99,7 +99,7 @@ struct gk20a_channel_sync {
 
 void gk20a_channel_sync_destroy(struct gk20a_channel_sync *sync);
 struct gk20a_channel_sync *gk20a_channel_sync_create(struct channel_gk20a *c);
-bool gk20a_channel_sync_needs_sync_framework(struct channel_gk20a *c);
+bool gk20a_channel_sync_needs_sync_framework(struct gk20a *g);
 
 #ifdef CONFIG_SYNC
 void gk20a_channel_cancel_pending_sema_waits(struct gk20a *g);
diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index 8ea2929f..7f110821 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -461,6 +461,25 @@ int gk20a_init_gpu_characteristics(struct gk20a *g)
         if (gk20a_platform_has_syncpoints(g))
                 gpu->flags |= NVGPU_GPU_FLAGS_HAS_SYNCPOINTS;
 
+        /*
+         * Railgating needs job tracking which prevents fast submits. They're
+         * supported otherwise, provided that the user doesn't request anything
+         * that depends on job tracking. (Here, fast means strictly no
+         * metadata, just the gpfifo contents are copied and gp_put updated).
+         */
+        if (!platform->can_railgate)
+                gpu->flags |= NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_NO_JOBTRACKING;
+
+        /*
+         * Railgating and sync framework require deferred job cleanup which
+         * prevents deterministic submits. They're supported otherwise,
+         * provided that the user doesn't request anything that depends on
+         * deferred cleanup.
+         */
+        if (!platform->can_railgate
+            && !gk20a_channel_sync_needs_sync_framework(g))
+                gpu->flags |= NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_FULL;
+
         gpu->flags |= NVGPU_GPU_FLAGS_SUPPORT_USERSPACE_MANAGED_AS;
         gpu->flags |= NVGPU_GPU_FLAGS_SUPPORT_TSG;
         gpu->flags |= NVGPU_GPU_FLAGS_SUPPORT_MAP_COMPBITS;
diff --git a/include/uapi/linux/nvgpu.h b/include/uapi/linux/nvgpu.h
index e25667cb..401722b1 100644
--- a/include/uapi/linux/nvgpu.h
+++ b/include/uapi/linux/nvgpu.h
@@ -137,6 +137,10 @@ struct nvgpu_gpu_zbc_query_table_args {
 #define NVGPU_GPU_FLAGS_SUPPORT_FECS_CTXSW_TRACE        (1ULL << 16)
 /* NVGPU_AS_IOCTL_MAP_BUFFER_COMPBITS is available */
 #define NVGPU_GPU_FLAGS_SUPPORT_MAP_COMPBITS            (1ULL << 17)
+/* Fast deterministic submits with no job tracking are supported */
+#define NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_NO_JOBTRACKING (1ULL << 18)
+/* Deterministic submits are supported even with job tracking */
+#define NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_FULL (1ULL << 19)
 
 struct nvgpu_gpu_characteristics {
         __u32 arch;
@@ -1348,8 +1352,22 @@ struct nvgpu_alloc_gpfifo_args {
 struct nvgpu_alloc_gpfifo_ex_args {
         __u32 num_entries;
         __u32 num_inflight_jobs;
-#define NVGPU_ALLOC_GPFIFO_EX_FLAGS_VPR_ENABLED         (1 << 0) /* set owner channel of this gpfifo as a vpr channel */
-#define NVGPU_ALLOC_GPFIFO_EX_FLAGS_DETERMINISTIC       (1 << 1) /* channel shall exhibit deterministic behavior in the submit path */
+/* Set owner channel of this gpfifo as a vpr channel. */
+#define NVGPU_ALLOC_GPFIFO_EX_FLAGS_VPR_ENABLED         (1 << 0)
+/*
+ * Channel shall exhibit deterministic behavior in the submit path.
+ *
+ * With this flag, any submits with in-kernel job tracking also require that
+ * num_inflight_jobs is nonzero, and additionally that
+ * NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_FULL is found in gpu
+ * characteristics.flags.
+ *
+ * Note that fast submits (with no in-kernel job tracking) are also
+ * deterministic and are supported if the characteristics flags contain
+ * NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_NO_JOBTRACKING; this flag or
+ * num_inflight_jobs are not necessary in that case.
+ */
+#define NVGPU_ALLOC_GPFIFO_EX_FLAGS_DETERMINISTIC       (1 << 1)
         __u32 flags;
         __u32 reserved[5];
 };