[PATCH v4 0/2] Defer vm_bo cleanup in GPUVM with DRM_GPUVM_IMMEDIATE_MODE

[PATCH v4 0/2] Defer vm_bo cleanup in GPUVM with DRM_GPUVM_IMMEDIATE_MODE

[Alice Ryhl]

There are two main ways that GPUVM might be used:

* staged mode, where VM_BIND ioctls update the GPUVM immediately so that
  the GPUVM reflects the state of the VM *including* staged changes that
  are not yet applied to the GPU's virtual address space.
* immediate mode, where the GPUVM state is updated during run_job(),
  i.e., in the DMA fence signalling critical path, to ensure that the
  GPUVM and the GPU's virtual address space has the same state at all
  times.

Currently, only Panthor uses GPUVM in immediate mode, but the Rust
drivers Tyr and Nova will also use GPUVM in immediate mode, so it is
worth to support both staged and immediate mode well in GPUVM. To use
immediate mode, we must manage the vm_bos and vas during the fence
signalling critical path.

The first part of that work was the introduction of a fence signalling
safe mutex for the GEMs GPUVA list in commit e7fa80e2932c ("drm_gem: add
mutex to drm_gem_object.gpuva").

This is series the second part of that work: Dropping a vm_bo object in
the fence signalling critical path is problematic for two reasons:

* When using DRM_GPUVM_RESV_PROTECTED, you cannot remove the vm_bo from
  the extobj/evicted lists during the fence signalling path.
* Dropping a vm_bo could lead to the GEM object getting destroyed.
  The requirement that GEM object cleanup is fence signalling safe is
  dubious and likely to be violated in practice.

Panthor already has its own custom implementation of postponing vm_bo
cleanup. Take inspiration from that by moving the logic into GPUVM, and
adjust Panthor to use the new GPUVM logic.

Signed-off-by: Alice Ryhl <aliceryhl@google.com>
---
Changes in v4:
- Go back to using kref_put_mutex().
- Change terminology of deferral methods to 'zombie' and improve their
  docs.
- Link to v3: https://lore.kernel.org/r/20251001-vmbo-defer-v3-0-a3fe6b6ae185@google.com

Changes in v3:
- Unpin in panthor on drm_gpuvm_bo_create() failure.
- Use llist for bo_defer list.
- Rename drm_gpuvm_bo_is_dead() to drm_gpuvm_bo_is_zombie().
- Rename drm_gpuvm_bo_defer() to drm_gpuvm_bo_defer_free().
- Link to v2: https://lore.kernel.org/r/20250909-vmbo-defer-v2-0-9835d7349089@google.com

Changes in v2:
- Fix missing kfree in Panthor.
- Rework mutex_lock() calls to be less confusing.
- Add note about resv lock in drm_gpuvm_bo_is_dead() docs.
- Link to v1: https://lore.kernel.org/r/20250905-vmbo-defer-v1-0-7ae1a382b674@google.com

---
Alice Ryhl (2):
      drm/gpuvm: add deferred vm_bo cleanup
      panthor: use drm_gpuva_unlink_defer()

 drivers/gpu/drm/drm_gpuvm.c           | 191 ++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/panthor/panthor_mmu.c | 110 ++++----------------
 include/drm/drm_gpuvm.h               |  16 +++
 3 files changed, 226 insertions(+), 91 deletions(-)
---
base-commit: b2ec5ca9d5c2c019e2316f7ba447596d1dcd8fde
change-id: 20250905-vmbo-defer-3faf90d821f5

Best regards,
-- 
Alice Ryhl <aliceryhl@google.com>

[PATCH v4 1/2] drm/gpuvm: add deferred vm_bo cleanup

[Alice Ryhl]

When using GPUVM in immediate mode, it is necessary to call
drm_gpuvm_unlink() from the fence signalling critical path. However,
unlink may call drm_gpuvm_bo_put(), which causes some challenges:

1. drm_gpuvm_bo_put() often requires you to take resv locks, which you
   can't do from the fence signalling critical path.
2. drm_gpuvm_bo_put() calls drm_gem_object_put(), which is often going
   to be unsafe to call from the fence signalling critical path.

To solve these issues, add a deferred version of drm_gpuvm_unlink() that
adds the vm_bo to a deferred cleanup list, and then clean it up later.

The new methods take the GEMs GPUVA lock internally rather than letting
the caller do it because it also needs to perform an operation after
releasing the mutex again. This is to prevent freeing the GEM while
holding the mutex (more info as comments in the patch). This means that
the new methods can only be used with DRM_GPUVM_IMMEDIATE_MODE.

Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Acked-by: Danilo Krummrich <dakr@kernel.org>
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
---
 drivers/gpu/drm/drm_gpuvm.c | 191 ++++++++++++++++++++++++++++++++++++++++++++
 include/drm/drm_gpuvm.h     |  16 ++++
 2 files changed, 207 insertions(+)

diff --git a/drivers/gpu/drm/drm_gpuvm.c b/drivers/gpu/drm/drm_gpuvm.c
index a52e95555549a16c062168253477035679d4775b..c36f1b8b68d2435628bbdbd03a3c8d3a3f0c123f 100644
--- a/drivers/gpu/drm/drm_gpuvm.c
+++ b/drivers/gpu/drm/drm_gpuvm.c
@@ -876,6 +876,31 @@ __drm_gpuvm_bo_list_add(struct drm_gpuvm *gpuvm, spinlock_t *lock,
 	cond_spin_unlock(lock, !!lock);
 }
 
+/**
+ * drm_gpuvm_bo_is_zombie() - check whether this vm_bo is scheduled for cleanup
+ * @vm_bo: the &drm_gpuvm_bo
+ *
+ * When a vm_bo is scheduled for cleanup using the bo_defer list, it is not
+ * immediately removed from the evict and extobj lists. Therefore, anyone
+ * iterating these lists should skip entries that are being destroyed.
+ *
+ * Checking the refcount without incrementing it is okay as long as the lock
+ * protecting the evict/extobj list is held for as long as you are using the
+ * vm_bo, because even if the refcount hits zero while you are using it, freeing
+ * the vm_bo requires taking the list's lock.
+ *
+ * Zombie entries can be observed on the evict and extobj lists regardless of
+ * whether DRM_GPUVM_RESV_PROTECTED is used, but they remain on the lists for a
+ * longer time when the resv lock is used because we can't take the resv lock
+ * during run_job() in immediate mode, meaning that they need to remain on the
+ * lists until drm_gpuvm_bo_deferred_cleanup() is called.
+ */
+static bool
+drm_gpuvm_bo_is_zombie(struct drm_gpuvm_bo *vm_bo)
+{
+	return !kref_read(&vm_bo->kref);
+}
+
 /**
  * drm_gpuvm_bo_list_add() - insert a vm_bo into the given list
  * @__vm_bo: the &drm_gpuvm_bo
@@ -1081,6 +1106,8 @@ drm_gpuvm_init(struct drm_gpuvm *gpuvm, const char *name,
 	INIT_LIST_HEAD(&gpuvm->evict.list);
 	spin_lock_init(&gpuvm->evict.lock);
 
+	init_llist_head(&gpuvm->bo_defer);
+
 	kref_init(&gpuvm->kref);
 
 	gpuvm->name = name ? name : "unknown";
@@ -1122,6 +1149,8 @@ drm_gpuvm_fini(struct drm_gpuvm *gpuvm)
 		 "Extobj list should be empty.\n");
 	drm_WARN(gpuvm->drm, !list_empty(&gpuvm->evict.list),
 		 "Evict list should be empty.\n");
+	drm_WARN(gpuvm->drm, !llist_empty(&gpuvm->bo_defer),
+		 "VM BO cleanup list should be empty.\n");
 
 	drm_gem_object_put(gpuvm->r_obj);
 }
@@ -1217,6 +1246,9 @@ drm_gpuvm_prepare_objects_locked(struct drm_gpuvm *gpuvm,
 
 	drm_gpuvm_resv_assert_held(gpuvm);
 	list_for_each_entry(vm_bo, &gpuvm->extobj.list, list.entry.extobj) {
+		if (drm_gpuvm_bo_is_zombie(vm_bo))
+			continue;
+
 		ret = exec_prepare_obj(exec, vm_bo->obj, num_fences);
 		if (ret)
 			break;
@@ -1460,6 +1492,9 @@ drm_gpuvm_validate_locked(struct drm_gpuvm *gpuvm, struct drm_exec *exec)
 
 	list_for_each_entry_safe(vm_bo, next, &gpuvm->evict.list,
 				 list.entry.evict) {
+		if (drm_gpuvm_bo_is_zombie(vm_bo))
+			continue;
+
 		ret = ops->vm_bo_validate(vm_bo, exec);
 		if (ret)
 			break;
@@ -1560,6 +1595,7 @@ drm_gpuvm_bo_create(struct drm_gpuvm *gpuvm,
 
 	INIT_LIST_HEAD(&vm_bo->list.entry.extobj);
 	INIT_LIST_HEAD(&vm_bo->list.entry.evict);
+	init_llist_node(&vm_bo->list.entry.bo_defer);
 
 	return vm_bo;
 }
@@ -1621,6 +1657,127 @@ drm_gpuvm_bo_put(struct drm_gpuvm_bo *vm_bo)
 }
 EXPORT_SYMBOL_GPL(drm_gpuvm_bo_put);
 
+/*
+ * drm_gpuvm_bo_into_zombie() - called when the vm_bo becomes a zombie due to
+ * deferred cleanup
+ *
+ * If deferred cleanup is used, then this must be called right after the vm_bo
+ * refcount drops to zero. Must be called with GEM mutex held. After releasing
+ * the GEM mutex, drm_gpuvm_bo_defer_zombie_cleanup() must be called.
+ */
+static void
+drm_gpuvm_bo_into_zombie(struct kref *kref)
+{
+	struct drm_gpuvm_bo *vm_bo = container_of(kref, struct drm_gpuvm_bo,
+						  kref);
+
+	if (!drm_gpuvm_resv_protected(vm_bo->vm)) {
+		drm_gpuvm_bo_list_del(vm_bo, extobj, true);
+		drm_gpuvm_bo_list_del(vm_bo, evict, true);
+	}
+
+	list_del(&vm_bo->list.entry.gem);
+}
+
+/*
+ * drm_gpuvm_bo_defer_zombie_cleanup() - adds a new zombie vm_bo to the
+ * bo_defer list
+ *
+ * Called after drm_gpuvm_bo_into_zombie(). GEM mutex must not be held.
+ *
+ * It's important that the GEM stays alive for the duration in which we hold
+ * the mutex, but the instant we add the vm_bo to bo_defer, another thread
+ * might call drm_gpuvm_bo_deferred_cleanup() and put the GEM. Therefore, to
+ * avoid kfreeing a mutex we are holding, the GEM mutex must be released
+ * *before* calling this function.
+ */
+static void
+drm_gpuvm_bo_defer_zombie_cleanup(struct drm_gpuvm_bo *vm_bo)
+{
+	llist_add(&vm_bo->list.entry.bo_defer, &vm_bo->vm->bo_defer);
+}
+
+static void
+drm_gpuvm_bo_defer_free(struct kref *kref)
+{
+	struct drm_gpuvm_bo *vm_bo = container_of(kref, struct drm_gpuvm_bo,
+						  kref);
+
+	drm_gpuvm_bo_into_zombie(kref);
+	mutex_unlock(&vm_bo->obj->gpuva.lock);
+	drm_gpuvm_bo_defer_zombie_cleanup(vm_bo);
+}
+
+/**
+ * drm_gpuvm_bo_put_deferred() - drop a struct drm_gpuvm_bo reference with
+ * deferred cleanup
+ * @vm_bo: the &drm_gpuvm_bo to release the reference of
+ *
+ * This releases a reference to @vm_bo.
+ *
+ * This might take and release the GEMs GPUVA lock. You should call
+ * drm_gpuvm_bo_deferred_cleanup() later to complete the cleanup process.
+ *
+ * Returns: true if vm_bo is being destroyed, false otherwise.
+ */
+bool
+drm_gpuvm_bo_put_deferred(struct drm_gpuvm_bo *vm_bo)
+{
+	if (!vm_bo)
+		return false;
+
+	drm_WARN_ON(vm_bo->vm->drm, !drm_gpuvm_immediate_mode(vm_bo->vm));
+
+	return !!kref_put_mutex(&vm_bo->kref,
+				drm_gpuvm_bo_defer_free,
+				&vm_bo->obj->gpuva.lock);
+}
+EXPORT_SYMBOL_GPL(drm_gpuvm_bo_put_deferred);
+
+/**
+ * drm_gpuvm_bo_deferred_cleanup() - clean up BOs in the deferred list
+ * deferred cleanup
+ * @gpuvm: the VM to clean up
+ *
+ * Cleans up &drm_gpuvm_bo instances in the deferred cleanup list.
+ */
+void
+drm_gpuvm_bo_deferred_cleanup(struct drm_gpuvm *gpuvm)
+{
+	const struct drm_gpuvm_ops *ops = gpuvm->ops;
+	struct drm_gpuvm_bo *vm_bo;
+	struct drm_gem_object *obj;
+	struct llist_node *bo_defer;
+
+	bo_defer = llist_del_all(&gpuvm->bo_defer);
+	if (!bo_defer)
+		return;
+
+	if (drm_gpuvm_resv_protected(gpuvm)) {
+		dma_resv_lock(drm_gpuvm_resv(gpuvm), NULL);
+		llist_for_each_entry(vm_bo, bo_defer, list.entry.bo_defer) {
+			drm_gpuvm_bo_list_del(vm_bo, extobj, false);
+			drm_gpuvm_bo_list_del(vm_bo, evict, false);
+		}
+		dma_resv_unlock(drm_gpuvm_resv(gpuvm));
+	}
+
+	while (bo_defer) {
+		vm_bo = llist_entry(bo_defer,
+			struct drm_gpuvm_bo, list.entry.bo_defer);
+		bo_defer = bo_defer->next;
+		obj = vm_bo->obj;
+		if (ops && ops->vm_bo_free)
+			ops->vm_bo_free(vm_bo);
+		else
+			kfree(vm_bo);
+
+		drm_gpuvm_put(gpuvm);
+		drm_gem_object_put(obj);
+	}
+}
+EXPORT_SYMBOL_GPL(drm_gpuvm_bo_deferred_cleanup);
+
 static struct drm_gpuvm_bo *
 __drm_gpuvm_bo_find(struct drm_gpuvm *gpuvm,
 		    struct drm_gem_object *obj)
@@ -1948,6 +2105,40 @@ drm_gpuva_unlink(struct drm_gpuva *va)
 }
 EXPORT_SYMBOL_GPL(drm_gpuva_unlink);
 
+/**
+ * drm_gpuva_unlink_defer() - unlink a &drm_gpuva with deferred vm_bo cleanup
+ * @va: the &drm_gpuva to unlink
+ *
+ * Similar to drm_gpuva_unlink(), but uses drm_gpuvm_bo_put_deferred() and takes
+ * the lock for the caller.
+ */
+void
+drm_gpuva_unlink_defer(struct drm_gpuva *va)
+{
+	struct drm_gem_object *obj = va->gem.obj;
+	struct drm_gpuvm_bo *vm_bo = va->vm_bo;
+	bool should_defer_bo;
+
+	if (unlikely(!obj))
+		return;
+
+	drm_WARN_ON(vm_bo->vm->drm, !drm_gpuvm_immediate_mode(vm_bo->vm));
+
+	mutex_lock(&obj->gpuva.lock);
+	list_del_init(&va->gem.entry);
+
+	/*
+	 * This is drm_gpuvm_bo_put_deferred() except we already hold the mutex.
+	 */
+	should_defer_bo = kref_put(&vm_bo->kref, drm_gpuvm_bo_into_zombie);
+	mutex_unlock(&obj->gpuva.lock);
+	if (should_defer_bo)
+		drm_gpuvm_bo_defer_zombie_cleanup(vm_bo);
+
+	va->vm_bo = NULL;
+}
+EXPORT_SYMBOL_GPL(drm_gpuva_unlink_defer);
+
 /**
  * drm_gpuva_find_first() - find the first &drm_gpuva in the given range
  * @gpuvm: the &drm_gpuvm to search in
diff --git a/include/drm/drm_gpuvm.h b/include/drm/drm_gpuvm.h
index 8890ded1d90752a2acbb564f697aa5ab03b5d052..81cc7672cf2d5362c637abfa2a75471e5274ed08 100644
--- a/include/drm/drm_gpuvm.h
+++ b/include/drm/drm_gpuvm.h
@@ -27,6 +27,7 @@
 
 #include <linux/dma-resv.h>
 #include <linux/list.h>
+#include <linux/llist.h>
 #include <linux/rbtree.h>
 #include <linux/types.h>
 
@@ -152,6 +153,7 @@ void drm_gpuva_remove(struct drm_gpuva *va);
 
 void drm_gpuva_link(struct drm_gpuva *va, struct drm_gpuvm_bo *vm_bo);
 void drm_gpuva_unlink(struct drm_gpuva *va);
+void drm_gpuva_unlink_defer(struct drm_gpuva *va);
 
 struct drm_gpuva *drm_gpuva_find(struct drm_gpuvm *gpuvm,
 				 u64 addr, u64 range);
@@ -331,6 +333,11 @@ struct drm_gpuvm {
 		 */
 		spinlock_t lock;
 	} evict;
+
+	/**
+	 * @bo_defer: structure holding vm_bos that need to be destroyed
+	 */
+	struct llist_head bo_defer;
 };
 
 void drm_gpuvm_init(struct drm_gpuvm *gpuvm, const char *name,
@@ -714,6 +721,12 @@ struct drm_gpuvm_bo {
 			 * &drm_gpuvms evict list.
 			 */
 			struct list_head evict;
+
+			/**
+			 * @list.entry.bo_defer: List entry to attach to
+			 * the &drm_gpuvms bo_defer list.
+			 */
+			struct llist_node bo_defer;
 		} entry;
 	} list;
 };
@@ -746,6 +759,9 @@ drm_gpuvm_bo_get(struct drm_gpuvm_bo *vm_bo)
 
 bool drm_gpuvm_bo_put(struct drm_gpuvm_bo *vm_bo);
 
+bool drm_gpuvm_bo_put_deferred(struct drm_gpuvm_bo *vm_bo);
+void drm_gpuvm_bo_deferred_cleanup(struct drm_gpuvm *gpuvm);
+
 struct drm_gpuvm_bo *
 drm_gpuvm_bo_find(struct drm_gpuvm *gpuvm,
 		  struct drm_gem_object *obj);

-- 
2.51.0.618.g983fd99d29-goog

[PATCH v4 2/2] panthor: use drm_gpuva_unlink_defer()

[Alice Ryhl]

Instead of manually deferring cleanup of vm_bos, use the new GPUVM
infrastructure for doing so.

To avoid manual management of vm_bo refcounts, the panthor_vma_link()
and panthor_vma_unlink() methods are changed to get and put a vm_bo
refcount on the vm_bo. This simplifies the code a lot. I preserved the
behavior where panthor_gpuva_sm_step_map() drops the refcount right away
rather than letting panthor_vm_cleanup_op_ctx() do it later.

Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
---
 drivers/gpu/drm/panthor/panthor_mmu.c | 110 ++++++----------------------------
 1 file changed, 19 insertions(+), 91 deletions(-)

diff --git a/drivers/gpu/drm/panthor/panthor_mmu.c b/drivers/gpu/drm/panthor/panthor_mmu.c
index 6dec4354e3789d17c5a87fc8de3bc86764b804bc..9f5f4ddf291024121f3fd5644f2fdeba354fa67c 100644
--- a/drivers/gpu/drm/panthor/panthor_mmu.c
+++ b/drivers/gpu/drm/panthor/panthor_mmu.c
@@ -181,20 +181,6 @@ struct panthor_vm_op_ctx {
 		u64 range;
 	} va;
 
-	/**
-	 * @returned_vmas: List of panthor_vma objects returned after a VM operation.
-	 *
-	 * For unmap operations, this will contain all VMAs that were covered by the
-	 * specified VA range.
-	 *
-	 * For map operations, this will contain all VMAs that previously mapped to
-	 * the specified VA range.
-	 *
-	 * Those VMAs, and the resources they point to will be released as part of
-	 * the op_ctx cleanup operation.
-	 */
-	struct list_head returned_vmas;
-
 	/** @map: Fields specific to a map operation. */
 	struct {
 		/** @map.vm_bo: Buffer object to map. */
@@ -1081,47 +1067,18 @@ void panthor_vm_free_va(struct panthor_vm *vm, struct drm_mm_node *va_node)
 	mutex_unlock(&vm->mm_lock);
 }
 
-static void panthor_vm_bo_put(struct drm_gpuvm_bo *vm_bo)
+static void panthor_vm_bo_free(struct drm_gpuvm_bo *vm_bo)
 {
 	struct panthor_gem_object *bo = to_panthor_bo(vm_bo->obj);
-	struct drm_gpuvm *vm = vm_bo->vm;
-	bool unpin;
-
-	/* We must retain the GEM before calling drm_gpuvm_bo_put(),
-	 * otherwise the mutex might be destroyed while we hold it.
-	 * Same goes for the VM, since we take the VM resv lock.
-	 */
-	drm_gem_object_get(&bo->base.base);
-	drm_gpuvm_get(vm);
-
-	/* We take the resv lock to protect against concurrent accesses to the
-	 * gpuvm evicted/extobj lists that are modified in
-	 * drm_gpuvm_bo_destroy(), which is called if drm_gpuvm_bo_put()
-	 * releases sthe last vm_bo reference.
-	 * We take the BO GPUVA list lock to protect the vm_bo removal from the
-	 * GEM vm_bo list.
-	 */
-	dma_resv_lock(drm_gpuvm_resv(vm), NULL);
-	mutex_lock(&bo->base.base.gpuva.lock);
-	unpin = drm_gpuvm_bo_put(vm_bo);
-	mutex_unlock(&bo->base.base.gpuva.lock);
-	dma_resv_unlock(drm_gpuvm_resv(vm));
 
-	/* If the vm_bo object was destroyed, release the pin reference that
-	 * was hold by this object.
-	 */
-	if (unpin && !drm_gem_is_imported(&bo->base.base))
+	if (!drm_gem_is_imported(&bo->base.base))
 		drm_gem_shmem_unpin(&bo->base);
-
-	drm_gpuvm_put(vm);
-	drm_gem_object_put(&bo->base.base);
+	kfree(vm_bo);
 }
 
 static void panthor_vm_cleanup_op_ctx(struct panthor_vm_op_ctx *op_ctx,
 				      struct panthor_vm *vm)
 {
-	struct panthor_vma *vma, *tmp_vma;
-
 	u32 remaining_pt_count = op_ctx->rsvd_page_tables.count -
 				 op_ctx->rsvd_page_tables.ptr;
 
@@ -1134,16 +1091,12 @@ static void panthor_vm_cleanup_op_ctx(struct panthor_vm_op_ctx *op_ctx,
 	kfree(op_ctx->rsvd_page_tables.pages);
 
 	if (op_ctx->map.vm_bo)
-		panthor_vm_bo_put(op_ctx->map.vm_bo);
+		drm_gpuvm_bo_put_deferred(op_ctx->map.vm_bo);
 
 	for (u32 i = 0; i < ARRAY_SIZE(op_ctx->preallocated_vmas); i++)
 		kfree(op_ctx->preallocated_vmas[i]);
 
-	list_for_each_entry_safe(vma, tmp_vma, &op_ctx->returned_vmas, node) {
-		list_del(&vma->node);
-		panthor_vm_bo_put(vma->base.vm_bo);
-		kfree(vma);
-	}
+	drm_gpuvm_bo_deferred_cleanup(&vm->base);
 }
 
 static struct panthor_vma *
@@ -1232,7 +1185,6 @@ static int panthor_vm_prepare_map_op_ctx(struct panthor_vm_op_ctx *op_ctx,
 		return -EINVAL;
 
 	memset(op_ctx, 0, sizeof(*op_ctx));
-	INIT_LIST_HEAD(&op_ctx->returned_vmas);
 	op_ctx->flags = flags;
 	op_ctx->va.range = size;
 	op_ctx->va.addr = va;
@@ -1243,7 +1195,9 @@ static int panthor_vm_prepare_map_op_ctx(struct panthor_vm_op_ctx *op_ctx,
 
 	if (!drm_gem_is_imported(&bo->base.base)) {
 		/* Pre-reserve the BO pages, so the map operation doesn't have to
-		 * allocate.
+		 * allocate. This pin is dropped in panthor_vm_bo_free(), so
+		 * once we have successfully called drm_gpuvm_bo_create(),
+		 * GPUVM will take care of dropping the pin for us.
 		 */
 		ret = drm_gem_shmem_pin(&bo->base);
 		if (ret)
@@ -1282,16 +1236,6 @@ static int panthor_vm_prepare_map_op_ctx(struct panthor_vm_op_ctx *op_ctx,
 	mutex_unlock(&bo->base.base.gpuva.lock);
 	dma_resv_unlock(panthor_vm_resv(vm));
 
-	/* If the a vm_bo for this <VM,BO> combination exists, it already
-	 * retains a pin ref, and we can release the one we took earlier.
-	 *
-	 * If our pre-allocated vm_bo is picked, it now retains the pin ref,
-	 * which will be released in panthor_vm_bo_put().
-	 */
-	if (preallocated_vm_bo != op_ctx->map.vm_bo &&
-	    !drm_gem_is_imported(&bo->base.base))
-		drm_gem_shmem_unpin(&bo->base);
-
 	op_ctx->map.bo_offset = offset;
 
 	/* L1, L2 and L3 page tables.
@@ -1339,7 +1283,6 @@ static int panthor_vm_prepare_unmap_op_ctx(struct panthor_vm_op_ctx *op_ctx,
 	int ret;
 
 	memset(op_ctx, 0, sizeof(*op_ctx));
-	INIT_LIST_HEAD(&op_ctx->returned_vmas);
 	op_ctx->va.range = size;
 	op_ctx->va.addr = va;
 	op_ctx->flags = DRM_PANTHOR_VM_BIND_OP_TYPE_UNMAP;
@@ -1387,7 +1330,6 @@ static void panthor_vm_prepare_sync_only_op_ctx(struct panthor_vm_op_ctx *op_ctx
 						struct panthor_vm *vm)
 {
 	memset(op_ctx, 0, sizeof(*op_ctx));
-	INIT_LIST_HEAD(&op_ctx->returned_vmas);
 	op_ctx->flags = DRM_PANTHOR_VM_BIND_OP_TYPE_SYNC_ONLY;
 }
 
@@ -2033,26 +1975,13 @@ static void panthor_vma_link(struct panthor_vm *vm,
 
 	mutex_lock(&bo->base.base.gpuva.lock);
 	drm_gpuva_link(&vma->base, vm_bo);
-	drm_WARN_ON(&vm->ptdev->base, drm_gpuvm_bo_put(vm_bo));
 	mutex_unlock(&bo->base.base.gpuva.lock);
 }
 
-static void panthor_vma_unlink(struct panthor_vm *vm,
-			       struct panthor_vma *vma)
+static void panthor_vma_unlink(struct panthor_vma *vma)
 {
-	struct panthor_gem_object *bo = to_panthor_bo(vma->base.gem.obj);
-	struct drm_gpuvm_bo *vm_bo = drm_gpuvm_bo_get(vma->base.vm_bo);
-
-	mutex_lock(&bo->base.base.gpuva.lock);
-	drm_gpuva_unlink(&vma->base);
-	mutex_unlock(&bo->base.base.gpuva.lock);
-
-	/* drm_gpuva_unlink() release the vm_bo, but we manually retained it
-	 * when entering this function, so we can implement deferred VMA
-	 * destruction. Re-assign it here.
-	 */
-	vma->base.vm_bo = vm_bo;
-	list_add_tail(&vma->node, &vm->op_ctx->returned_vmas);
+	drm_gpuva_unlink_defer(&vma->base);
+	kfree(vma);
 }
 
 static void panthor_vma_init(struct panthor_vma *vma, u32 flags)
@@ -2084,12 +2013,12 @@ static int panthor_gpuva_sm_step_map(struct drm_gpuva_op *op, void *priv)
 	if (ret)
 		return ret;
 
-	/* Ref owned by the mapping now, clear the obj field so we don't release the
-	 * pinning/obj ref behind GPUVA's back.
-	 */
 	drm_gpuva_map(&vm->base, &vma->base, &op->map);
 	panthor_vma_link(vm, vma, op_ctx->map.vm_bo);
+
+	drm_gpuvm_bo_put_deferred(op_ctx->map.vm_bo);
 	op_ctx->map.vm_bo = NULL;
+
 	return 0;
 }
 
@@ -2128,16 +2057,14 @@ static int panthor_gpuva_sm_step_remap(struct drm_gpuva_op *op,
 		 * owned by the old mapping which will be released when this
 		 * mapping is destroyed, we need to grab a ref here.
 		 */
-		panthor_vma_link(vm, prev_vma,
-				 drm_gpuvm_bo_get(op->remap.unmap->va->vm_bo));
+		panthor_vma_link(vm, prev_vma, op->remap.unmap->va->vm_bo);
 	}
 
 	if (next_vma) {
-		panthor_vma_link(vm, next_vma,
-				 drm_gpuvm_bo_get(op->remap.unmap->va->vm_bo));
+		panthor_vma_link(vm, next_vma, op->remap.unmap->va->vm_bo);
 	}
 
-	panthor_vma_unlink(vm, unmap_vma);
+	panthor_vma_unlink(unmap_vma);
 	return 0;
 }
 
@@ -2154,12 +2081,13 @@ static int panthor_gpuva_sm_step_unmap(struct drm_gpuva_op *op,
 		return ret;
 
 	drm_gpuva_unmap(&op->unmap);
-	panthor_vma_unlink(vm, unmap_vma);
+	panthor_vma_unlink(unmap_vma);
 	return 0;
 }
 
 static const struct drm_gpuvm_ops panthor_gpuvm_ops = {
 	.vm_free = panthor_vm_free,
+	.vm_bo_free = panthor_vm_bo_free,
 	.sm_step_map = panthor_gpuva_sm_step_map,
 	.sm_step_remap = panthor_gpuva_sm_step_remap,
 	.sm_step_unmap = panthor_gpuva_sm_step_unmap,

-- 
2.51.0.618.g983fd99d29-goog

Re: [PATCH v4 0/2] Defer vm_bo cleanup in GPUVM with DRM_GPUVM_IMMEDIATE_MODE

[Alice Ryhl]

On Mon, Oct 06, 2025 at 12:05:54PM +0000, Alice Ryhl wrote:
> There are two main ways that GPUVM might be used:
> 
> * staged mode, where VM_BIND ioctls update the GPUVM immediately so that
>   the GPUVM reflects the state of the VM *including* staged changes that
>   are not yet applied to the GPU's virtual address space.
> * immediate mode, where the GPUVM state is updated during run_job(),
>   i.e., in the DMA fence signalling critical path, to ensure that the
>   GPUVM and the GPU's virtual address space has the same state at all
>   times.
> 
> Currently, only Panthor uses GPUVM in immediate mode, but the Rust
> drivers Tyr and Nova will also use GPUVM in immediate mode, so it is
> worth to support both staged and immediate mode well in GPUVM. To use
> immediate mode, we must manage the vm_bos and vas during the fence
> signalling critical path.
> 
> The first part of that work was the introduction of a fence signalling
> safe mutex for the GEMs GPUVA list in commit e7fa80e2932c ("drm_gem: add
> mutex to drm_gem_object.gpuva").
> 
> This is series the second part of that work: Dropping a vm_bo object in
> the fence signalling critical path is problematic for two reasons:
> 
> * When using DRM_GPUVM_RESV_PROTECTED, you cannot remove the vm_bo from
>   the extobj/evicted lists during the fence signalling path.
> * Dropping a vm_bo could lead to the GEM object getting destroyed.
>   The requirement that GEM object cleanup is fence signalling safe is
>   dubious and likely to be violated in practice.
> 
> Panthor already has its own custom implementation of postponing vm_bo
> cleanup. Take inspiration from that by moving the logic into GPUVM, and
> adjust Panthor to use the new GPUVM logic.
> 
> Signed-off-by: Alice Ryhl <aliceryhl@google.com>

Pushed to drm-rust-next.

I added this to drm-rust instead of drm-misc as a prerequisite for
another patch. If merge conflicts show up, we can do a backmerge.