On 8/31/2025 9:27 PM, Yifan Zhang wrote:
Current mem limit check leaks some GTT memory (reserved_for_pt
reserved_for_ras + adev->vram_pin_size) for small APUs.
Since carveout VRAM is tunable on APUs, there are three case
regarding the carveout VRAM size relative to GTT:
1. 0 < carveout < gtt
apu_prefer_gtt = true, is_app_apu = false
2. carveout > gtt / 2
apu_prefer_gtt = false, is_app_apu = false
3. 0 = carveout
apu_prefer_gtt = true, is_app_apu = true
It doesn't make sense to check below limitation in case 1
(default case, small carveout) because the values in the below
expression are mixed with carveout and gtt.
adev->kfd.vram_used[xcp_id] + vram_needed >
vram_size - reserved_for_pt - reserved_for_ras -
atomic64_read(&adev->vram_pin_size)
gtt: kfd.vram_used, vram_needed, vram_size
carveout: reserved_for_pt, reserved_for_ras, adev->vram_pin_size
In case 1, vram allocation will go to gtt domain, skip vram check
since ttm_mem_limit check already cover this allocation.
v2: simplify judgement logic (Mario)
You can strip the changelog from the commit message.
Signed-off-by: Yifan Zhang <yifan1.zh...@amd.com>
Reviewed-by: Mario Limonciello <mario.limoncie...@amd.com>
---
.../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 44 ++++++++++++++-----
1 file changed, 32 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index d478acb4568a..c3b34a410375 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -213,19 +213,35 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device
*adev,
spin_lock(&kfd_mem_limit.mem_limit_lock);
if (kfd_mem_limit.system_mem_used + system_mem_needed >
- kfd_mem_limit.max_system_mem_limit)
+ kfd_mem_limit.max_system_mem_limit) {
pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
+ if (!no_system_mem_limit) {
+ ret = -ENOMEM;
+ goto release;
+ }
+ }
- if ((kfd_mem_limit.system_mem_used + system_mem_needed >
- kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
- (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
- kfd_mem_limit.max_ttm_mem_limit) ||
- (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
- vram_size - reserved_for_pt - reserved_for_ras -
atomic64_read(&adev->vram_pin_size))) {
+ if (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
+ kfd_mem_limit.max_ttm_mem_limit) {
ret = -ENOMEM;
goto release;
}
+ /*if is_app_apu is false and apu_prefer_gtt is true, it is an APU with
+ * carve out < gtt. In that case, VRAM allocation will go to gtt
domain, skip
+ * VRAM check since ttm_mem_limit check already cover this allocation
+ */
+
+ if (adev && xcp_id >= 0 && (!adev->apu_prefer_gtt ||
adev->gmc.is_app_apu)) {
+ uint64_t vram_available =
+ vram_size - reserved_for_pt - reserved_for_ras -
+ atomic64_read(&adev->vram_pin_size);
+ if (adev->kfd.vram_used[xcp_id] + vram_needed > vram_available)
{
+ ret = -ENOMEM;
+ goto release;
+ }
+ }
+
/* Update memory accounting by decreasing available system
* memory, TTM memory and GPU memory as computed above
*/
@@ -1627,11 +1643,15 @@ size_t amdgpu_amdkfd_get_available_memory(struct
amdgpu_device *adev,
uint64_t vram_available, system_mem_available, ttm_mem_available;
spin_lock(&kfd_mem_limit.mem_limit_lock);
- vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
- - adev->kfd.vram_used_aligned[xcp_id]
- - atomic64_read(&adev->vram_pin_size)
- - reserved_for_pt
- - reserved_for_ras;
+ if (adev->apu_prefer_gtt && !adev->gmc.is_app_apu)
+ vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
+ - adev->kfd.vram_used_aligned[xcp_id];
+ else
+ vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
+ - adev->kfd.vram_used_aligned[xcp_id]
+ - atomic64_read(&adev->vram_pin_size)
+ - reserved_for_pt
+ - reserved_for_ras;
if (adev->apu_prefer_gtt) {
system_mem_available = no_system_mem_limit ?
