On 2023-09-29 14:18, Rajneesh Bhardwaj wrote:
To allow bigger allocations specially on systems such as GFXIP 9.4.3
that use GTT memory for VRAM allocations, relax the limits to
maximize ROCm allocations.
Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@xxxxxxx>
If the new heuristic in patch 1 causes regressions, we can change
AMDGPU_RESERVE_MEM_LIMIT to something larger, e.g. 1.5 GB (3UL << 29).
Maybe we should start with that more conservative value to minimize the
risk on 16GB system memory systems that have proven problematic in the past.
Other than that, patches 1 and 3 are
Reviewed-by: Felix Kuehling <Felix.Kuehling@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 10 ++++++++--
1 file changed, 8 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index b5b940485059..b1c4e9c0e036 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -42,6 +42,7 @@
* changes to accumulate
*/
#define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
+#define AMDGPU_RESERVE_MEM_LIMIT (1UL << 30)
/*
* Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
@@ -115,11 +116,16 @@ void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
return;
si_meminfo(&si);
- mem = si.freeram - si.freehigh;
+ mem = si.totalram - si.totalhigh;
mem *= si.mem_unit;
spin_lock_init(&kfd_mem_limit.mem_limit_lock);
- kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
+ kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6);
+ if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT)
+ kfd_mem_limit.max_system_mem_limit >>= 1;
+ else
+ kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
+
kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
(kfd_mem_limit.max_system_mem_limit >> 20),
