hw/virtio/vhost-user.c | 34 ++++++++++++++++++---------------- include/hw/virtio/vhost-user.h | 1 + 2 files changed, 19 insertions(+), 16 deletions(-)
During the process of hot-unplug in vhost-user-net NIC, vhost_user_cleanup
may add same rcu node to rcu list. Function calls are 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
Reason for the abort is that adding two identical nodes to the rcu list will
cause it becomes a ring. After dummy node is added to the tail of the list in
try_dequeue, the ring is opened. But lead to a situation that only one node is
added to list and rcu_call_count is added twice. This will cause try_dequeue
abort.
This issue happens when n->addr != 0 in vhost_user_host_notifier_remove. It can
happens in a hotplug stress test with a 32queue vhost-user-net type NIC.
Because n->addr is set in VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG function.
during device hotplug process and it is cleared in vhost_dev_stop during device
hot-unplug. Since VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG is a message sent
by DPDK to qemu, it is asynchronous. So, there is no guaranteed order between
the two processes of setting n->addr and clearing n->addr. If setting n->addr
in hotplug occurs after clearing n->addr in hotunplug, the issue will occur.
So, 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 <yaozhenguo@jd.com>
---
hw/virtio/vhost-user.c | 34 ++++++++++++++++++----------------
include/hw/virtio/vhost-user.h | 1 +
2 files changed, 19 insertions(+), 16 deletions(-)
diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
index 00561da..b2da3cf 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->need_free) {
+ memory_region_transaction_begin();
+ object_unparent(OBJECT(&n->mr));
+ memory_region_transaction_commit();
+ g_free(n);
+ }
}
/*
@@ -1195,17 +1202,20 @@ 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 free)
{
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->need_free = free;
+ call_rcu(n, vhost_user_host_notifier_free, rcu);
}
static int vhost_user_set_vring_base(struct vhost_dev *dev,
@@ -1280,7 +1290,7 @@ static int vhost_user_get_vring_base(struct vhost_dev *dev,
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);
@@ -1562,7 +1572,7 @@ 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;
@@ -2737,13 +2747,7 @@ 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);
}
}
@@ -2765,9 +2769,7 @@ 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 324cd86..a171f29 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 need_free;
} VhostUserHostNotifier;
/**
--
1.8.3.1yaozhenguo <yaozhenguo1@gmail.com> 于2024年9月13日周五 16:08写道：
>
> During the process of hot-unplug in vhost-user-net NIC, vhost_user_cleanup
> may add same rcu node to rcu list. Function calls are 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
>
> Reason for the abort is that adding two identical nodes to the rcu list will
> cause it becomes a ring. After dummy node is added to the tail of the list in
> try_dequeue, the ring is opened. But lead to a situation that only one node is
> added to list and rcu_call_count is added twice. This will cause try_dequeue
> abort.
>
> This issue happens when n->addr != 0 in vhost_user_host_notifier_remove. It can
> happens in a hotplug stress test with a 32queue vhost-user-net type NIC.
> Because n->addr is set in VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG function.
> during device hotplug process and it is cleared in vhost_dev_stop during device
> hot-unplug. Since VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG is a message sent
> by DPDK to qemu, it is asynchronous. So, there is no guaranteed order between
> the two processes of setting n->addr and clearing n->addr. If setting n->addr
> in hotplug occurs after clearing n->addr in hotunplug, the issue will occur.
> So, 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 <yaozhenguo@jd.com>
> ---
> hw/virtio/vhost-user.c | 34 ++++++++++++++++++----------------
> include/hw/virtio/vhost-user.h | 1 +
> 2 files changed, 19 insertions(+), 16 deletions(-)
>
> diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
> index 00561da..b2da3cf 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->need_free) {
> + memory_region_transaction_begin();
> + object_unparent(OBJECT(&n->mr));
> + memory_region_transaction_commit();
> + g_free(n);
> + }
> }
>
> /*
> @@ -1195,17 +1202,20 @@ 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 free)
> {
> 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->need_free = free;
> + call_rcu(n, vhost_user_host_notifier_free, rcu);
> }
This brings up another problem. When the device is initialized,
problems may occur when the execution times of
vhost_user_get_vring_base and
vhost_user_backend_handle_vring_host_notifier are relatively close,
and the same node is submitted to the rcu list. This happened during
our regression testing. Therefore, modifications must be made to avoid
this problem.
vhost_user_get_vring_base
vhost_user_host_notifier_remove(n, dev->vdev, false);
vhost_user_backend_handle_vring_host_notifier
vhost_user_host_notifier_remove(n, dev->vdev, false);
>
> static int vhost_user_set_vring_base(struct vhost_dev *dev,
> @@ -1280,7 +1290,7 @@ static int vhost_user_get_vring_base(struct vhost_dev *dev,
>
> 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);
> @@ -1562,7 +1572,7 @@ 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;
> @@ -2737,13 +2747,7 @@ 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);
> }
> }
>
> @@ -2765,9 +2769,7 @@ 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 324cd86..a171f29 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 need_free;
> } VhostUserHostNotifier;
>
> /**
> --
> 1.8.3.1
>
© 2016 - 2024 Red Hat, Inc.