[PATCH 44/44] drm/amdkfd: Nested locking and invalidation of child ranges

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



This allows validation of child ranges, so the GPU page fault handler
can be more light-weight.

Signed-off-by: Felix Kuehling <Felix.Kuehling@xxxxxxx>
---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c |  8 +++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 40 +++++++++++++++++-------
 2 files changed, 37 insertions(+), 11 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
index 4d79d69d8aac..cc8bf6438383 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
@@ -826,12 +826,18 @@ static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
 	}
 
 	mutex_lock(&parent->migrate_mutex);
+	if (prange != parent)
+		mutex_lock_nested(&prange->migrate_mutex, 1);
 
 	if (!prange->actual_loc)
 		goto out_unlock_prange;
 
 	svm_range_lock(parent);
+	if (prange != parent)
+		mutex_lock_nested(&prange->lock, 1);
 	r = svm_range_split_by_granularity(p, mm, addr, parent, prange);
+	if (prange != parent)
+		mutex_unlock(&prange->lock);
 	svm_range_unlock(parent);
 	if (r) {
 		pr_debug("failed %d to split range by granularity\n", r);
@@ -852,6 +858,8 @@ static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf)
 	schedule_deferred_list_work(&p->svms);
 
 out_unlock_prange:
+	if (prange != parent)
+		mutex_unlock(&prange->migrate_mutex);
 	mutex_unlock(&parent->migrate_mutex);
 out:
 	mutex_unlock(&p->svms.lock);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 0a6e28f0dcaf..bc1a9e9ba722 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1281,10 +1281,6 @@ void svm_range_unreserve_bos(struct svm_validate_context *ctx)
  * serialize concurrent migrations or validations of the same range, the
  * prange->migrate_mutex must be held.
  *
- * The range must be in the inverval tree and have an MMU notifier to catch
- * concurrent invalidations of the virtual address range. This means it cannot
- * be a child range.
- *
  * For VRAM ranges, the SVM BO must be allocated and valid (protected by its
  * eviction fence.
  *
@@ -1568,10 +1564,24 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
 		schedule_delayed_work(&svms->restore_work,
 			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
 	} else {
+		struct svm_range *pchild;
+		unsigned long s, l;
+
 		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",
 			 prange->svms, start, last);
 		svm_range_lock(prange);
-		svm_range_unmap_from_gpus(prange, start, last);
+		list_for_each_entry(pchild, &prange->child_list, child_list) {
+			mutex_lock_nested(&pchild->lock, 1);
+			s = max(start, pchild->start);
+			l = min(last, pchild->last);
+			if (l >= s)
+				svm_range_unmap_from_gpus(pchild, s, l);
+			mutex_unlock(&pchild->lock);
+		}
+		s = max(start, prange->start);
+		l = min(last, prange->last);
+		if (l >= s)
+			svm_range_unmap_from_gpus(prange, s, l);
 		svm_range_unlock(prange);
 	}
 
@@ -1927,6 +1937,7 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
 	struct svm_range_list *svms;
 	struct svm_range *pchild;
 	struct kfd_process *p;
+	unsigned long s, l;
 	bool unmap_parent;
 
 	p = kfd_lookup_process_by_mm(mm);
@@ -1937,14 +1948,23 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
 		 prange, prange->start, prange->last, start, last);
 
-	svm_range_unmap_from_gpus(prange, start, last);
-
 	svm_range_lock(prange);
 
 	unmap_parent = start <= prange->start && last >= prange->last;
 
-	list_for_each_entry(pchild, &prange->child_list, child_list)
+	list_for_each_entry(pchild, &prange->child_list, child_list) {
+		mutex_lock_nested(&pchild->lock, 1);
+		s = max(start, pchild->start);
+		l = min(last, pchild->last);
+		if (l >= s)
+			svm_range_unmap_from_gpus(pchild, s, l);
 		svm_range_unmap_split(mm, prange, pchild, start, last);
+		mutex_unlock(&pchild->lock);
+	}
+	s = max(start, prange->start);
+	l = min(last, prange->last);
+	if (l >= s)
+		svm_range_unmap_from_gpus(prange, s, l);
 	svm_range_unmap_split(mm, prange, prange, start, last);
 
 	svm_range_unlock(prange);
@@ -2142,12 +2162,10 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 		goto out;
 	}
 
-	svm_range_list_lock_and_flush_work(svms, mm);
+	mmap_read_lock(mm);
 	mutex_lock(&svms->lock);
 	prange = svm_range_from_addr(svms, addr, NULL);
 
-	mmap_write_downgrade(mm);
-
 	if (!prange) {
 		pr_debug("failed to find prange svms 0x%p address [0x%llx]\n",
 			 svms, addr);
-- 
2.31.0

_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx



[Index of Archives]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux