gpu: nvgpu: wait for all prefence semas on gpu

The pre-fence wait for semaphores in the submit path has supported a
fast path for fences that have only one underlying semaphore. The fast
path just inserts the wait on this sema to the pushbuffer directly. For
other fences, the path has been using a CPU wait indirection, signaling
another semaphore when we get the CPU-side callback.

Instead of only supporting prefences with one sema, unroll all the
individual semaphores and insert waits for each to a pushbuffer, like
we've already been doing with syncpoints. Now all sema-backed syncs get
the fast path. This simplifies the logic and makes it more explicit that
only foreign fences need the CPU wait.

There is no need to hold references to the sync fence or the semas
inside: this submitted job only needs the global read-only sema mapping
that is guaranteed to stay alive while the VM of this channel stays
alive, and the job does not outlive this channel.

Jira NVGPU-43
Jira NVGPU-66
Jira NVGPU-513

Change-Id: I7cfbb510001d998a864aed8d6afd1582b9adb80d
Signed-off-by: Konsta Holtta <kholtta@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1636345
Reviewed-by: Alex Waterman <alexw@nvidia.com>
Reviewed-by: Terje Bergstrom <tbergstrom@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

0 files changed, 0 insertions, 0 deletions