gpu: nvgpu: hold power ref for deterministic channels

To support deterministic channels even with platforms where railgating
is supported, have each deterministic-marked channel hold a power
reference during their lifetime, and skip taking power refs for jobs in
submit path for those.

Previously, railgating blocked deterministic submits in general because
of gk20a_busy()/gk20a_idle() calls in submit path possibly taking time
and more significantly because the gpu may need turning on which takes a
nondeterministic and long amount of time.

As an exception, gk20a_do_idle() can still block deterministic submits
until gk20a_do_unidle() is called. Add a rwsem to guard this. VPR resize
needs do_idle, which conflicts with deterministic channels' requirement
to keep the GPU on. This is documented in the ioctl header now.

Make NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_NO_JOBTRACKING always
set in the gpu characteristics now that it's supported. The only thing
left now blocking NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_FULL is
the sync framework.

Make the channel debug dump show which channels are deterministic.

Bug 200291300
Jira NVGPU-70

Change-Id: I47b6f3a8517cd6e4255f6ca2855e3dd912e4f5f3
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: http://git-master/r/1483038
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

1 files changed, 8 insertions, 11 deletions

diff --git a/drivers/gpu/nvgpu/gk20a/gk20a.c b/drivers/gpu/nvgpu/gk20a/gk20a.c
index 8624d601..13635706 100644
--- a/drivers/gpu/nvgpu/gk20a/gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/gk20a.c
@@ -456,22 +456,19 @@ int gk20a_init_gpu_characteristics(struct gk20a *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
+         * Fast submits are supported as long as 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 (!g->can_railgate)
-                gpu->flags |= NVGPU_GPU_FLAGS_SUPPORT_DETERMINISTIC_SUBMIT_NO_JOBTRACKING;
+        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.
+         * Sync framework requires deferred job cleanup, wrapping syncs in FDs,
+         * and other heavy stuff, which prevents deterministic submits. This is
+         * supported otherwise, provided that the user doesn't request anything
+         * that depends on deferred cleanup.
          */
-        if (!g->can_railgate
-            && !gk20a_channel_sync_needs_sync_framework(g))
+        if (!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;