From bcf83fab33d8f1821cfcfd8c03411ea2fcf5334b Mon Sep 17 00:00:00 2001 From: Alex Waterman Date: Wed, 18 Jul 2018 12:40:24 -0700 Subject: gpu: nvgpu: Set default large page size to 64K for gm20b Since all userspace apps are using 64K pages these days it makes sense to set the default large page size to 64K. This in turn causes the PDE coverage field to be set to 64M in the GPU characteristics field. While it would therefor be possible to create a VM with a PDE coverage that's larger than 64M (128M if you set the large page size to 128K) this will make the defaults work properly. This in turn fixes a CUDA issue where CUDA tries to determine the PDE coverage (and correspondingly a minimum alignemnt) from the characteristics IOCTL. Bug 200105199 Change-Id: Iee3c213f1b81d8628571f46c7ad5e16fbfe07499 Signed-off-by: Alex Waterman Reviewed-on: https://git-master.nvidia.com/r/1781088 Reviewed-by: Terje Bergstrom GVS: Gerrit_Virtual_Submit Reviewed-by: mobile promotions Tested-by: mobile promotions --- drivers/gpu/nvgpu/gm20b/mm_gm20b.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu/nvgpu/gm20b') diff --git a/drivers/gpu/nvgpu/gm20b/mm_gm20b.c b/drivers/gpu/nvgpu/gm20b/mm_gm20b.c index deca6686..9be3180c 100644 --- a/drivers/gpu/nvgpu/gm20b/mm_gm20b.c +++ b/drivers/gpu/nvgpu/gm20b/mm_gm20b.c @@ -58,7 +58,7 @@ u32 gm20b_mm_get_big_page_sizes(void) u32 gm20b_mm_get_default_big_page_size(void) { - return SZ_128K; + return SZ_64K; } bool gm20b_mm_support_sparse(struct gk20a *g) -- cgit v1.2.2