* [PATCH 2/3] drm/amdkfd: Avoid thrashing of stack and heap
2021-11-01 12:55 [PATCH 1/3] drm/amdkfd: Fix SVM_ATTR_PREFERRED_LOC Felix Kuehling
@ 2021-11-01 12:55 ` Felix Kuehling
2021-11-01 12:55 ` [PATCH 3/3] drm/amdkfd: Handle incomplete migration to system memory Felix Kuehling
2021-11-01 15:11 ` [PATCH 1/3] drm/amdkfd: Fix SVM_ATTR_PREFERRED_LOC philip yang
2 siblings, 0 replies; 4+ messages in thread
From: Felix Kuehling @ 2021-11-01 12:55 UTC (permalink / raw)
To: amd-gfx; +Cc: philip.yang, Philip Yang
Stack and heap pages tend to be shared by many small allocations.
Concurrent access by CPU and GPU is therefore likely, which can lead to
thrashing. Avoid this by setting the preferred location to system memory.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Philip Yang <yangp@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 22 +++++++++++++++++-----
1 file changed, 17 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 8d37ee755b34..854735dbb96e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -2310,7 +2310,8 @@ svm_range_best_restore_location(struct svm_range *prange,
static int
svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
- unsigned long *start, unsigned long *last)
+ unsigned long *start, unsigned long *last,
+ bool *is_heap_stack)
{
struct vm_area_struct *vma;
struct interval_tree_node *node;
@@ -2321,6 +2322,12 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
pr_debug("VMA does not exist in address [0x%llx]\n", addr);
return -EFAULT;
}
+
+ *is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
+ vma->vm_end >= vma->vm_mm->start_brk) ||
+ (vma->vm_start <= vma->vm_mm->start_stack &&
+ vma->vm_end >= vma->vm_mm->start_stack);
+
start_limit = max(vma->vm_start >> PAGE_SHIFT,
(unsigned long)ALIGN_DOWN(addr, 2UL << 8));
end_limit = min(vma->vm_end >> PAGE_SHIFT,
@@ -2350,9 +2357,9 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
*start = start_limit;
*last = end_limit - 1;
- pr_debug("vma start: 0x%lx start: 0x%lx vma end: 0x%lx last: 0x%lx\n",
- vma->vm_start >> PAGE_SHIFT, *start,
- vma->vm_end >> PAGE_SHIFT, *last);
+ pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
+ vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
+ *start, *last, *is_heap_stack);
return 0;
}
@@ -2417,11 +2424,13 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
struct svm_range *prange = NULL;
unsigned long start, last;
uint32_t gpuid, gpuidx;
+ bool is_heap_stack;
uint64_t bo_s = 0;
uint64_t bo_l = 0;
int r;
- if (svm_range_get_range_boundaries(p, addr, &start, &last))
+ if (svm_range_get_range_boundaries(p, addr, &start, &last,
+ &is_heap_stack))
return NULL;
r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
@@ -2448,6 +2457,9 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
return NULL;
}
+ if (is_heap_stack)
+ prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
+
svm_range_add_to_svms(prange);
svm_range_add_notifier_locked(mm, prange);
--
2.32.0
^ permalink raw reply related [flat|nested] 4+ messages in thread* [PATCH 3/3] drm/amdkfd: Handle incomplete migration to system memory
2021-11-01 12:55 [PATCH 1/3] drm/amdkfd: Fix SVM_ATTR_PREFERRED_LOC Felix Kuehling
2021-11-01 12:55 ` [PATCH 2/3] drm/amdkfd: Avoid thrashing of stack and heap Felix Kuehling
@ 2021-11-01 12:55 ` Felix Kuehling
2021-11-01 15:11 ` [PATCH 1/3] drm/amdkfd: Fix SVM_ATTR_PREFERRED_LOC philip yang
2 siblings, 0 replies; 4+ messages in thread
From: Felix Kuehling @ 2021-11-01 12:55 UTC (permalink / raw)
To: amd-gfx; +Cc: philip.yang
If some pages fail to migrate to system memory, don't update
prange->actual_loc = 0. This prevents endless CPU page faults after
partial migration failures due to contested page locks.
Migration to RAM must be complete during migrations from VRAM to VRAM and
during evictions. Implement retry and fail if the migration to RAM fails.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 45 +++++++++++++++++-------
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 8 ++++-
2 files changed, 40 insertions(+), 13 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index 6d8634e40b3b..cc1525095937 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
@@ -281,6 +281,19 @@ static unsigned long svm_migrate_successful_pages(struct migrate_vma *migrate)
return cpages;
}
+static unsigned long svm_migrate_unsuccessful_pages(struct migrate_vma *migrate)
+{
+ unsigned long upages = 0;
+ unsigned long i;
+
+ for (i = 0; i < migrate->npages; i++) {
+ if (migrate->src[i] & MIGRATE_PFN_VALID &&
+ !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
+ upages++;
+ }
+ return upages;
+}
+
static int
svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
struct migrate_vma *migrate, struct dma_fence **mfence,
@@ -634,10 +647,11 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
struct vm_area_struct *vma, uint64_t start, uint64_t end)
{
uint64_t npages = (end - start) >> PAGE_SHIFT;
+ unsigned long upages = npages;
+ unsigned long cpages = 0;
struct kfd_process_device *pdd;
struct dma_fence *mfence = NULL;
struct migrate_vma migrate;
- unsigned long cpages = 0;
dma_addr_t *scratch;
size_t size;
void *buf;
@@ -671,6 +685,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
if (!cpages) {
pr_debug("failed collect migrate device pages [0x%lx 0x%lx]\n",
prange->start, prange->last);
+ upages = svm_migrate_unsuccessful_pages(&migrate);
goto out_free;
}
if (cpages != npages)
@@ -683,8 +698,9 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
scratch, npages);
migrate_vma_pages(&migrate);
- pr_debug("successful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
- svm_migrate_successful_pages(&migrate), cpages, migrate.npages);
+ upages = svm_migrate_unsuccessful_pages(&migrate);
+ pr_debug("unsuccessful/cpages/npages 0x%lx/0x%lx/0x%lx\n",
+ upages, cpages, migrate.npages);
svm_migrate_copy_done(adev, mfence);
migrate_vma_finalize(&migrate);
@@ -698,9 +714,9 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
if (pdd)
WRITE_ONCE(pdd->page_out, pdd->page_out + cpages);
- return cpages;
+ return upages;
}
- return r;
+ return r ? r : upages;
}
/**
@@ -720,7 +736,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
unsigned long addr;
unsigned long start;
unsigned long end;
- unsigned long cpages = 0;
+ unsigned long upages = 0;
long r = 0;
if (!prange->actual_loc) {
@@ -756,12 +772,12 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)
pr_debug("failed %ld to migrate\n", r);
break;
} else {
- cpages += r;
+ upages += r;
}
addr = next;
}
- if (cpages) {
+ if (!upages) {
svm_range_vram_node_free(prange);
prange->actual_loc = 0;
}
@@ -784,7 +800,7 @@ static int
svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
struct mm_struct *mm)
{
- int r;
+ int r, retries = 3;
/*
* TODO: for both devices with PCIe large bar or on same xgmi hive, skip
@@ -793,9 +809,14 @@ svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,
pr_debug("from gpu 0x%x to gpu 0x%x\n", prange->actual_loc, best_loc);
- r = svm_migrate_vram_to_ram(prange, mm);
- if (r)
- return r;
+ do {
+ r = svm_migrate_vram_to_ram(prange, mm);
+ if (r)
+ return r;
+ } while (prange->actual_loc && --retries);
+
+ if (prange->actual_loc)
+ return -EDEADLK;
return svm_migrate_ram_to_vram(prange, best_loc, mm);
}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 854735dbb96e..e6897e931057 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -3083,6 +3083,8 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work)
struct svm_range *prange =
list_first_entry(&svm_bo->range_list,
struct svm_range, svm_bo_list);
+ int retries = 3;
+
list_del_init(&prange->svm_bo_list);
spin_unlock(&svm_bo->list_lock);
@@ -3090,7 +3092,11 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work)
prange->start, prange->last);
mutex_lock(&prange->migrate_mutex);
- svm_migrate_vram_to_ram(prange, svm_bo->eviction_fence->mm);
+ do {
+ svm_migrate_vram_to_ram(prange,
+ svm_bo->eviction_fence->mm);
+ } while (prange->actual_loc && --retries);
+ WARN(prange->actual_loc, "Migration failed during eviction");
mutex_lock(&prange->lock);
prange->svm_bo = NULL;
--
2.32.0
^ permalink raw reply related [flat|nested] 4+ messages in thread