On Fri, Oct 11, 2024 at 06:29:13PM +0800, yaozhenguo wrote:
During the hot-unplugging of vhost-user-net type network cards,
the vhost_user_cleanup function may add the same rcu node to
the rcu linked list. The function call in this case is as follows:
vhost_user_cleanup
->vhost_user_host_notifier_remove
->call_rcu(n, vhost_user_host_notifier_free, rcu);
->g_free_rcu(n, rcu);
When this happens, QEMU will abort in try_dequeue:
if (head == &dummy && qatomic_mb_read(&tail) == &dummy.next) {
abort();
}
backtrace is as follows:
0 __pthread_kill_implementation () at /usr/lib64/libc.so.6
1 raise () at /usr/lib64/libc.so.6
2 abort () at /usr/lib64/libc.so.6
3 try_dequeue () at ../util/rcu.c:235
4 call_rcu_thread (0) at ../util/rcu.c:288
5 qemu_thread_start (0) at ../util/qemu-thread-posix.c:541
6 start_thread () at /usr/lib64/libc.so.6
7 clone3 () at /usr/lib64/libc.so.6
The reason for the abort is that adding two identical nodes to
the rcu linked list will cause the rcu linked list to become a ring,
but when the dummy node is added after the two identical nodes,
the ring is opened. But only one node is added to list with
rcu_call_count added twice. This will cause rcu try_dequeue abort.
This happens when n->addr != 0. In some scenarios, this does happen.
For example, this situation will occur when using a 32-queue DPU
vhost-user-net type network card for hot-unplug testing, because
VhostUserHostNotifier->addr will be cleared during the processing of
VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG. However,it is asynchronous,
so we cannot guarantee that VhostUserHostNotifier->addr is zero in
vhost_user_cleanup. Therefore, it is necessary to merge g_free_rcu
and vhost_user_host_notifier_free into one rcu node.
Fixes: 503e355465 ("virtio/vhost-user: dynamically assign
VhostUserHostNotifiers")
Signed-off-by: yaozhenguo <yaozhen...@jd.com>
---
V1->V2: add n->addr check in vhost_user_get_vring_base and
vhost_user_backend_handle_vring_host_notifier to prevent submit same node to rcu list.
V2->V3: 1. change "free" to "destroy"
2. move "!n->addr && !destroy" checking to
vhost_user_host_notifier_remove
3. move "!n" checking to vhost_user_host_notifier_remove
V3->V4 fix checkpatch error
---
hw/virtio/vhost-user.c | 46 +++++++++++++++++++---------------
include/hw/virtio/vhost-user.h | 1 +
2 files changed, 27 insertions(+), 20 deletions(-)
LGTM!
Reviewed-by: Stefano Garzarella <sgarz...@redhat.com>
diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
index 00561daa06..d1b0893b4d 100644
--- a/hw/virtio/vhost-user.c
+++ b/hw/virtio/vhost-user.c
@@ -1185,9 +1185,16 @@ static int vhost_user_set_vring_num(struct vhost_dev
*dev,
static void vhost_user_host_notifier_free(VhostUserHostNotifier *n)
{
- assert(n && n->unmap_addr);
- munmap(n->unmap_addr, qemu_real_host_page_size());
- n->unmap_addr = NULL;
+ if (n->unmap_addr) {
+ munmap(n->unmap_addr, qemu_real_host_page_size());
+ n->unmap_addr = NULL;
+ }
+ if (n->destroy) {
+ memory_region_transaction_begin();
+ object_unparent(OBJECT(&n->mr));
+ memory_region_transaction_commit();
+ g_free(n);
+ }
}
/*
@@ -1195,17 +1202,28 @@ static void
vhost_user_host_notifier_free(VhostUserHostNotifier *n)
* under rcu.
*/
static void vhost_user_host_notifier_remove(VhostUserHostNotifier *n,
- VirtIODevice *vdev)
+ VirtIODevice *vdev, bool destroy)
{
+ /*
+ * if destroy == false and n->addr == NULL, we have nothing to do.
+ * so, just return.
+ */
+ if (!n || (!destroy && !n->addr)) {
+ return;
+ }
+
if (n->addr) {
if (vdev) {
+ memory_region_transaction_begin();
virtio_queue_set_host_notifier_mr(vdev, n->idx, &n->mr, false);
+ memory_region_transaction_commit();
}
assert(!n->unmap_addr);
n->unmap_addr = n->addr;
n->addr = NULL;
- call_rcu(n, vhost_user_host_notifier_free, rcu);
}
+ n->destroy = destroy;
+ call_rcu(n, vhost_user_host_notifier_free, rcu);
}
static int vhost_user_set_vring_base(struct vhost_dev *dev,
@@ -1279,9 +1297,7 @@ static int vhost_user_get_vring_base(struct vhost_dev
*dev,
struct vhost_user *u = dev->opaque;
VhostUserHostNotifier *n = fetch_notifier(u->user, ring->index);
- if (n) {
- vhost_user_host_notifier_remove(n, dev->vdev);
- }
+ vhost_user_host_notifier_remove(n, dev->vdev, false);
ret = vhost_user_write(dev, &msg, NULL, 0);
if (ret < 0) {
@@ -1562,7 +1578,7 @@ static int
vhost_user_backend_handle_vring_host_notifier(struct vhost_dev *dev,
* new mapped address.
*/
n = fetch_or_create_notifier(user, queue_idx);
- vhost_user_host_notifier_remove(n, vdev);
+ vhost_user_host_notifier_remove(n, vdev, false);
if (area->u64 & VHOST_USER_VRING_NOFD_MASK) {
return 0;
@@ -2736,15 +2752,7 @@ static int vhost_user_set_inflight_fd(struct vhost_dev
*dev,
static void vhost_user_state_destroy(gpointer data)
{
VhostUserHostNotifier *n = (VhostUserHostNotifier *) data;
- if (n) {
- vhost_user_host_notifier_remove(n, NULL);
- object_unparent(OBJECT(&n->mr));
- /*
- * We can't free until vhost_user_host_notifier_remove has
- * done it's thing so schedule the free with RCU.
- */
- g_free_rcu(n, rcu);
- }
+ vhost_user_host_notifier_remove(n, NULL, true);
}
bool vhost_user_init(VhostUserState *user, CharBackend *chr, Error **errp)
@@ -2765,9 +2773,7 @@ void vhost_user_cleanup(VhostUserState *user)
if (!user->chr) {
return;
}
- memory_region_transaction_begin();
user->notifiers = (GPtrArray *) g_ptr_array_free(user->notifiers, true);
- memory_region_transaction_commit();
user->chr = NULL;
}
diff --git a/include/hw/virtio/vhost-user.h b/include/hw/virtio/vhost-user.h
index 324cd8663a..9a3f238b43 100644
--- a/include/hw/virtio/vhost-user.h
+++ b/include/hw/virtio/vhost-user.h
@@ -54,6 +54,7 @@ typedef struct VhostUserHostNotifier {
void *addr;
void *unmap_addr;
int idx;
+ bool destroy;
} VhostUserHostNotifier;
/**
--
2.41.0
