[PATCH 1/3] drm/amdgpu: Rework KFD memory max limits

[PATCH 1/3] drm/amdgpu: Rework KFD memory max limits

[Rajneesh Bhardwaj]

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@amd.com>
---
 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),
-- 
2.34.1

[PATCH 2/3] drm/amdgpu: Initialize acpi mem ranges after TTM

[Rajneesh Bhardwaj]

Move ttm init before acpi mem range init so we can use ttm_pages_limit
to override vram size for GFXIP 9.4.3. The vram size override change
will be introduced in a future commit.

Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
index 268ee533e7c1..005ea719d2fd 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
@@ -2190,17 +2190,17 @@ static int gmc_v9_0_sw_init(void *handle)
 
 	amdgpu_gmc_get_vbios_allocations(adev);
 
+	/* Memory manager */
+	r = amdgpu_bo_init(adev);
+	if (r)
+		return r;
+
 	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) {
 		r = gmc_v9_0_init_mem_ranges(adev);
 		if (r)
 			return r;
 	}
 
-	/* Memory manager */
-	r = amdgpu_bo_init(adev);
-	if (r)
-		return r;
-
 	r = gmc_v9_0_gart_init(adev);
 	if (r)
 		return r;
-- 
2.34.1

[PATCH 3/3] drm/amdgpu: Use ttm_pages_limit to override vram reporting

[Rajneesh Bhardwaj]

On GFXIP9.4.3 APU, allow the memory reporting as per the ttm pages
limit in NPS1 mode.

Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
index 005ea719d2fd..fadc4d2ed071 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
@@ -25,6 +25,7 @@
 #include <linux/pci.h>
 
 #include <drm/drm_cache.h>
+#include <drm/ttm/ttm_tt.h>
 
 #include "amdgpu.h"
 #include "gmc_v9_0.h"
@@ -1896,7 +1897,8 @@ static void
 gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev,
 			      struct amdgpu_mem_partition_info *mem_ranges)
 {
-	int num_ranges = 0, ret, mem_groups;
+	int num_ranges = 0, ret, num_nodes;
+	uint64_t node_size_ttm_override = 0;
 	struct amdgpu_numa_info numa_info;
 	int node_ids[MAX_MEM_RANGES];
 	int num_xcc, xcc_id;
@@ -1904,7 +1906,8 @@ gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev,
 
 	num_xcc = NUM_XCC(adev->gfx.xcc_mask);
 	xcc_mask = (1U << num_xcc) - 1;
-	mem_groups = hweight32(adev->aid_mask);
+	num_nodes = num_online_nodes();
+	node_size_ttm_override = (ttm_tt_pages_limit() << PAGE_SHIFT) / num_nodes;
 
 	for_each_inst(xcc_id, xcc_mask)	{
 		ret = amdgpu_acpi_get_mem_info(adev, xcc_id, &numa_info);
@@ -1912,7 +1915,6 @@ gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev,
 			continue;
 
 		if (numa_info.nid == NUMA_NO_NODE) {
-			mem_ranges[0].size = numa_info.size;
 			mem_ranges[0].numa.node = numa_info.nid;
 			num_ranges = 1;
 			break;
@@ -1930,11 +1932,16 @@ gmc_v9_0_init_acpi_mem_ranges(struct amdgpu_device *adev,
 
 	adev->gmc.num_mem_partitions = num_ranges;
 
-	/* If there is only partition, don't use entire size */
-	if (adev->gmc.num_mem_partitions == 1) {
-		mem_ranges[0].size = mem_ranges[0].size * (mem_groups - 1);
-		do_div(mem_ranges[0].size, mem_groups);
-	}
+	/* In NPS1 mode, we should restrict the vram reporting
+	 * tied to the ttm_pages_limit which is 1/2 of the
+	 * system memory. For other partition modes, the HBM is
+	 * uniformly divided already per numa node reported. If
+	 * user wants to go beyond the default ttm limit and
+	 * maximize the ROCm allocations, they can go up to max
+	 * ttm and sysmem limits.
+	 */
+	if (adev->gmc.num_mem_partitions == 1)
+		mem_ranges[0].size = node_size_ttm_override;
 }
 
 static void
-- 
2.34.1

Re: [PATCH 1/3] drm/amdgpu: Rework KFD memory max limits

[Felix Kuehling]

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@amd.com>

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@amd.com>


> ---
>   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),

Re: [PATCH 2/3] drm/amdgpu: Initialize acpi mem ranges after TTM

[Bhardwaj, Rajneesh]

I found an issue with this patch, that leads to performance drop. This 
leads to incorrectly initialize numa pools on a multi node system. I am 
working on the fix and will send another change set.

On 9/29/2023 2:18 PM, Rajneesh Bhardwaj wrote:
> Move ttm init before acpi mem range init so we can use ttm_pages_limit
> to override vram size for GFXIP 9.4.3. The vram size override change
> will be introduced in a future commit.
>
> Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
> Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 10 +++++-----
>   1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> index 268ee533e7c1..005ea719d2fd 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> @@ -2190,17 +2190,17 @@ static int gmc_v9_0_sw_init(void *handle)
>   
>   	amdgpu_gmc_get_vbios_allocations(adev);
>   
> +	/* Memory manager */
> +	r = amdgpu_bo_init(adev);
> +	if (r)
> +		return r;
> +
>   	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) {
>   		r = gmc_v9_0_init_mem_ranges(adev);
>   		if (r)
>   			return r;
>   	}
>   
> -	/* Memory manager */
> -	r = amdgpu_bo_init(adev);
> -	if (r)
> -		return r;
> -
>   	r = gmc_v9_0_gart_init(adev);
>   	if (r)
>   		return r;