gpu-paging: Allow for more than one buffer to be swapped at a timertss22-aegpu-paging

This uses a very primitive linear disk sector allocation scheme.
Sectors are only reused when userspace resets assignment to 0 with
an NVGPU_AS_IOCTL_SWAP_RESET ioctl (which invalidates all current
swap buffers).

This sector assignment scheme is sufficient for use in a TimeWall-
like system, where all allocations are assumed to be static after
after task system release. This is not suitable for a system with
dynamic allocations, unless userspace manually resets swap state
regularly (benchmarks run a reset at start for example).

Support for dynamic allocations is on the backlog.

No significant speed impact.

Benchmarks, 100 iters, after:
gpu_paging_speed, write: 186.0ms +/- 3.51
gpu_paging_speed, read: 162.7ms +/- 2.58
gpu_paging_overhead_speed, write start: 35.4ms +/- 4.47
gpu_paging_overhead_speed, write finish: 3.3ms +/- 0.18
gpu_paging_overhead_speed, read start: 69.8ms +/- 6.42
gpu_paging_overhead_speed, read finish: 43.2ms +/- 0.91

1 files changed, 1 insertions, 0 deletions

diff --git a/drivers/gpu/nvgpu/os/linux/vm.c b/drivers/gpu/nvgpu/os/linux/vm.c
index 9cd17981..a1c19a3a 100644
--- a/drivers/gpu/nvgpu/os/linux/vm.c
+++ b/drivers/gpu/nvgpu/os/linux/vm.c
@@ -269,6 +269,7 @@ int nvgpu_vm_map_linux(struct vm_gk20a *vm,
         else
                 // So we can always safely call list_del()
                 INIT_LIST_HEAD(&mapped_buffer->os_priv.nvmap_priv_entry);
+        mapped_buffer->os_priv.swap_sector = 0;
 
         *gpu_va = mapped_buffer->addr;
         return 0;