---
v6:
- rebase v5 on patch 1/2, which adds kfd.vram_pinned (Mukul Joshi)
v5:
- add the kfd.vram_pinned[] counter. v4 read it in get_available_memory()
but it does not exist in the tree (Mukul Joshi)
- track pins per XCP in vram_pin_size_xcp[]; pins with no partition go to
vram_pin_size_common (Philip Yang, Lijo Lazar)
- pass the VRAM counter into the shared helper: the gate uses vram_used[],
the ioctl vram_used_aligned[]
- WARN_ONCE if a KFD BO reaches the VRAM pin path with xcp < 0 (Tony Zhou)
v4:
- use get_available_memory() in reserve_mem_limit()
- move ALIGN_DOWN from get_available_memory() to the ioctl
- drop v3 extra call in amdgpu_dma_buf.c
- update kfd.vram_pinned from amdgpu_bo_pin() via a KFD helper
v3:
- update kfd.vram_pinned from dma-buf pin/vmap and gpuvm_pin_bo
on first pin / last unpin in amdgpu_dma_buf.c
v2:
- keep vram_pin_size; add kfd.vram_pinned in reserve_mem_limit (Mukul Joshi)
- leave get_available_memory unchanged
- drop former 2/2 pin/unpin changes in amdgpu_object.c
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 11 ++-
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 20 +++-
.../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 94 +++++++++++--------
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 33 +++++++
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 12 ++-
5 files changed, 127 insertions(+), 43 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 79b69d74eb2e..04151eab3b3f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -839,8 +839,17 @@ struct amdgpu_device {
struct mutex mn_lock;
DECLARE_HASHTABLE(mn_hash, 7);
- /* tracking pinned memory */
+ /* tracking pinned memory
+ *
+ * vram_pin_size = sum(vram_pin_size_xcp[]) + vram_pin_size_common
+ *
+ * vram_pin_size total pin size in device
+ * vram_pin_size_xcp[i] pins owned by partition i (bo->xcp_id = i, i
>= 0)
+ * vram_pin_size_common pins not owned by any partition, bo->xcp_id =
-1
+ */
atomic64_t vram_pin_size;
+ atomic64_t vram_pin_size_xcp[MAX_XCP];
+ atomic64_t vram_pin_size_common;
atomic64_t visible_pin_size;
atomic64_t gart_pin_size;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 056dd7375d24..f462b40220eb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -42,6 +42,12 @@
extern uint64_t amdgpu_amdkfd_total_mem_size;
+/*
+ * Align VRAM availability to 2MB to avoid fragmentation caused by 4K
+ * allocations in the tail 2MB BO chunk.
+ */
+#define VRAM_AVAILABLITY_ALIGN (1 << 21)
+
enum TLB_FLUSH_TYPE {
TLB_FLUSH_LEGACY = 0,
TLB_FLUSH_LIGHTWEIGHT,
@@ -106,7 +112,7 @@ struct amdgpu_kfd_dev {
struct kfd_dev *dev;
int64_t vram_used[MAX_XCP];
uint64_t vram_used_aligned[MAX_XCP];
- atomic64_t vram_pinned;
+ atomic64_t vram_pinned[MAX_XCP];
bool init_complete;
struct work_struct reset_work;
@@ -391,6 +397,8 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device
*adev,
* Allows KFD to release its resources associated with the GEM object.
*/
void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo);
+void amdgpu_amdkfd_account_vram_pin(struct amdgpu_bo *bo, int xcp_id);
+void amdgpu_amdkfd_account_vram_unpin(struct amdgpu_bo *bo, int
+xcp_id);
void amdgpu_amdkfd_reserve_system_mem(uint64_t size); #else static inline
@@ -408,6 +416,16 @@ static inline void amdgpu_amdkfd_release_notify(struct
amdgpu_bo *bo) { }
+
+static inline
+void amdgpu_amdkfd_account_vram_pin(struct amdgpu_bo *bo, int xcp_id) {
+}
+
+static inline
+void amdgpu_amdkfd_account_vram_unpin(struct amdgpu_bo *bo, int xcp_id)
+{ }
#endif
#if IS_ENABLED(CONFIG_HSA_AMD_SVM)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 5a4072792a57..34cd791d826e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -46,12 +46,6 @@
#define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
#define AMDGPU_RESERVE_MEM_LIMIT (3UL << 29)
-/*
- * Align VRAM availability to 2MB to avoid fragmentation caused by 4K
allocations in the tail 2MB
- * BO chunk
- */
-#define VRAM_AVAILABLITY_ALIGN (1 << 21)
-
/* Impose limit on how much memory KFD can use */ static struct {
uint64_t max_system_mem_limit;
@@ -73,6 +67,8 @@ static const char * const domain_bit_to_string[] = { #define
domain_string(domain) domain_bit_to_string[ffs(domain)-1]
static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
+static size_t __get_available_memory(struct amdgpu_device *adev,
+ uint8_t xcp_id, int64_t vram_used);
static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
struct kgd_mem *mem)
@@ -169,13 +165,8 @@ void amdgpu_amdkfd_reserve_system_mem(uint64_t size) int
amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
uint64_t size, u32 alloc_flag, int8_t xcp_id) {
- uint64_t reserved_for_pt =
- ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
- struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
- uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0);
size_t system_mem_needed, ttm_mem_needed, vram_needed;
int ret = 0;
- uint64_t vram_size = 0;
system_mem_needed = 0;
ttm_mem_needed = 0;
@@ -196,7 +187,6 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device
*adev,
if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
return -EINVAL;
- vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
if (adev->apu_prefer_gtt) {
system_mem_needed = size;
ttm_mem_needed = size;
@@ -232,14 +222,11 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device
*adev,
* 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;
- }
+ if (adev && xcp_id >= 0 && (!adev->apu_prefer_gtt || adev->gmc.is_app_apu)
&&
+ vram_needed > __get_available_memory(adev, xcp_id,
+ adev->kfd.vram_used[xcp_id])) {
+ ret = -ENOMEM;
+ goto release;
}
/* Update memory accounting by decreasing available system @@ -318,6
+305,30 @@ void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
kfree(bo->kfd_bo);
}
+void amdgpu_amdkfd_account_vram_pin(struct amdgpu_bo *bo, int xcp_id) {
+ if (!bo->kfd_bo)
+ return;
+
+ if (WARN_ONCE(xcp_id < 0, "KFD BO pinned in VRAM with no partition"))
+ return;
+
+ atomic64_add(amdgpu_bo_size(bo),
+ &amdgpu_ttm_adev(bo->tbo.bdev)->kfd.vram_pinned[xcp_id]);
+}
+
+void amdgpu_amdkfd_account_vram_unpin(struct amdgpu_bo *bo, int xcp_id)
+{
+ if (!bo->kfd_bo)
+ return;
+
+ if (WARN_ONCE(xcp_id < 0, "KFD BO unpinned in VRAM with no partition"))
+ return;
+
+ atomic64_sub(amdgpu_bo_size(bo),
+ &amdgpu_ttm_adev(bo->tbo.bdev)->kfd.vram_pinned[xcp_id]);
+}
+
/**
* create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
* about USERPTR or DOOREBELL or MMIO BO.
@@ -1537,10 +1548,6 @@ static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo
*bo, u32 domain)
amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
- if (!ret && bo->tbo.resource->mem_type == TTM_PL_VRAM)
- atomic64_add(amdgpu_bo_size(bo),
- &amdgpu_ttm_adev(bo->tbo.bdev)->kfd.vram_pinned);
-
out:
amdgpu_bo_unreserve(bo);
return ret;
@@ -1564,10 +1571,6 @@ static void amdgpu_amdkfd_gpuvm_unpin_bo(struct
amdgpu_bo *bo)
amdgpu_bo_unpin(bo);
- if (bo->tbo.resource->mem_type == TTM_PL_VRAM)
- atomic64_sub(amdgpu_bo_size(bo),
- &amdgpu_ttm_adev(bo->tbo.bdev)->kfd.vram_pinned);
-
amdgpu_bo_unreserve(bo);
}
@@ -1671,8 +1674,9 @@ int amdgpu_amdkfd_criu_resume(void *p)
return ret;
}
-size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
- uint8_t xcp_id)
+/* Caller must hold kfd_mem_limit.mem_limit_lock. */ static size_t
+__get_available_memory(struct amdgpu_device *adev,
+ uint8_t xcp_id, int64_t vram_used)
{
uint64_t reserved_for_pt =
ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
@@ -1681,15 +1685,17 @@ size_t amdgpu_amdkfd_get_available_memory(struct
amdgpu_device *adev,
ssize_t available;
uint64_t vram_available, system_mem_available, ttm_mem_available;
- spin_lock(&kfd_mem_limit.mem_limit_lock);
+ lockdep_assert_held(&kfd_mem_limit.mem_limit_lock);
+
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];
+ - vram_used;
else
vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
- - adev->kfd.vram_used_aligned[xcp_id]
- - atomic64_read(&adev->vram_pin_size)
- + atomic64_read(&adev->kfd.vram_pinned)
+ - vram_used
+ - atomic64_read(&adev->vram_pin_size_xcp[xcp_id])
+ - atomic64_read(&adev->vram_pin_size_common)
+ + atomic64_read(&adev->kfd.vram_pinned[xcp_id])
- reserved_for_pt
- reserved_for_ras;
@@ -1704,19 +1710,29 @@ size_t amdgpu_amdkfd_get_available_memory(struct
amdgpu_device *adev,
available = min3(system_mem_available, ttm_mem_available,
vram_available);
- available = ALIGN_DOWN(available, PAGE_SIZE);
} else {
- available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
+ available = vram_available;
}
- spin_unlock(&kfd_mem_limit.mem_limit_lock);
-
if (available < 0)
available = 0;
return available;
}
+size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
+ uint8_t xcp_id)
+{
+ size_t available;
+
+ spin_lock(&kfd_mem_limit.mem_limit_lock);
+ available = __get_available_memory(adev, xcp_id,
+ adev->kfd.vram_used_aligned[xcp_id]);
+ spin_unlock(&kfd_mem_limit.mem_limit_lock);
+
+ return available;
+}
+
int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
struct amdgpu_device *adev, uint64_t va, uint64_t size,
void *drm_priv, struct kgd_mem **mem, diff --git
a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
index fc6d3fa62d79..dc0ab5bc467f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
@@ -923,6 +923,19 @@ void amdgpu_bo_unref(struct amdgpu_bo **bo)
*bo = NULL;
}
+/* Map a BO onto vram_pin_size_xcp[]. bo->xcp_id < 0 means "any partition":
+ * such pins are not owned by a single partition and go to
+ * vram_pin_size_common, which every partition subtracts.
+ */
+static int amdgpu_bo_get_xcp(struct amdgpu_bo *bo) {
+ s8 xcp = bo->xcp_id;
+
+ if (xcp >= 0 && xcp < MAX_XCP)
+ return xcp;
+ return -1;
+}
+
/**
* amdgpu_bo_pin - pin an &amdgpu_bo buffer object
* @bo: &amdgpu_bo buffer object to be pinned @@ -1003,7 +1016,17 @@ int
amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
ttm_bo_pin(&bo->tbo);
if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
+ int xcp = amdgpu_bo_get_xcp(bo);
+
atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
+ if (xcp >= 0) {
+ atomic64_add(amdgpu_bo_size(bo),
+ &adev->vram_pin_size_xcp[xcp]);
+ } else {
+ atomic64_add(amdgpu_bo_size(bo),
+ &adev->vram_pin_size_common);
+ }
+ amdgpu_amdkfd_account_vram_pin(bo, xcp);
atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
&adev->visible_pin_size);
} else if (bo->tbo.resource->mem_type == TTM_PL_TT) { @@ -1036,7
+1059,17 @@ void amdgpu_bo_unpin(struct amdgpu_bo *bo)
dma_buf_unpin(bo->tbo.base.import_attach);
if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
+ int xcp = amdgpu_bo_get_xcp(bo);
+
atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
+ if (xcp >= 0) {
+ atomic64_sub(amdgpu_bo_size(bo),
+ &adev->vram_pin_size_xcp[xcp]);
+ } else {
+ atomic64_sub(amdgpu_bo_size(bo),
+ &adev->vram_pin_size_common);
+ }
+ amdgpu_amdkfd_account_vram_unpin(bo, xcp);
atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
&adev->visible_pin_size);
} else if (bo->tbo.resource->mem_type == TTM_PL_TT) { diff --git
a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
index 7fcfc150a7fc..2295f9bcc5e4 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
@@ -1069,11 +1069,19 @@ static int kfd_ioctl_get_available_memory(struct file
*filep, {
struct kfd_ioctl_get_available_memory_args *args = data;
struct kfd_process_device *pdd = kfd_lock_pdd_by_id(p, args->gpu_id);
+ struct amdgpu_device *adev;
+ size_t available;
if (!pdd)
return -EINVAL;
- args->available = amdgpu_amdkfd_get_available_memory(pdd->dev->adev,
- pdd->dev->node_id);
+
+ adev = pdd->dev->adev;
+ available = amdgpu_amdkfd_get_available_memory(adev, pdd->dev->node_id);
+ if (adev->apu_prefer_gtt)
+ args->available = ALIGN_DOWN(available, PAGE_SIZE);
+ else
+ args->available = ALIGN_DOWN(available, VRAM_AVAILABLITY_ALIGN);
+
kfd_unlock_pdd(pdd);
return 0;
}
--
2.43.0
