hw/virtio/vhost-user.c | 67 ++++++++++++++++++++++++++++++++++++++---- 1 file changed, 61 insertions(+), 6 deletions(-)
fix: QEMU deadlock with dpdk-vdpa
QEMU start vhost-user with modern net and blk, backend use dpdk-vdpa process,
after live migration, dest QEMU deadlock with dpdk-vdpa
- QEMU sends VHOST_USER_SET_VRING_KICK to dpdk-vdpa net
- QEMU does not need to wait for a response to this message
- QEMU then sends VHOST_USER_SET_MEM_TABLE to dpdk-vdpa blk
- QEMU needs to wait reply in this message
- when dpdk-vdpa recv VHOST_USER_SET_VRING_KICK, it will send VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG to QEMU
- dpdk-vdpa needs to wait for a response to this message
- since QEMU vhost_user_read and backend channel are synchronous in the same thread
- QEMU will deadlock with dpdk-vdpa
Signed-off-by: Rma Ma <rma.ma@jaguarmicro.com>
---
hw/virtio/vhost-user.c | 67 ++++++++++++++++++++++++++++++++++++++----
1 file changed, 61 insertions(+), 6 deletions(-)
diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
index c4e0cbd702..1f6b3a5a63 100644
--- a/hw/virtio/vhost-user.c
+++ b/hw/virtio/vhost-user.c
@@ -274,6 +274,17 @@ struct scrub_regions {
int fd_idx;
};
+struct backend_thread {
+ QemuThread thread;
+ QemuSemaphore init_done_sem;
+ int thread_id;
+ GMainContext *ctxt;
+ GMainLoop *loop;
+};
+
+static struct backend_thread *backend_th;
+static bool backend_thread_run;
+
static bool ioeventfd_enabled(void)
{
return !kvm_enabled() || kvm_eventfds_enabled();
@@ -1696,7 +1707,8 @@ fdcleanup:
return rc;
}
-static int vhost_setup_backend_channel(struct vhost_dev *dev)
+static int vhost_setup_backend_channel(struct vhost_dev *dev,
+ GMainContext *ctxt)
{
VhostUserMsg msg = {
.hdr.request = VHOST_USER_SET_BACKEND_REQ_FD,
@@ -1728,7 +1740,7 @@ static int vhost_setup_backend_channel(struct vhost_dev *dev)
u->backend_ioc = ioc;
u->backend_src = qio_channel_add_watch_source(u->backend_ioc,
G_IO_IN | G_IO_HUP,
- backend_read, dev, NULL, NULL);
+ backend_read, dev, NULL, ctxt);
if (reply_supported) {
msg.hdr.flags |= VHOST_USER_NEED_REPLY_MASK;
@@ -1981,6 +1993,42 @@ static int vhost_user_postcopy_notifier(NotifierWithReturn *notifier,
return 0;
}
+static void *vhost_backend_channel_worker(void *opaque)
+{
+ struct backend_thread *backend_th = opaque;
+
+ rcu_register_thread();
+
+ backend_th->ctxt = g_main_context_new();
+ backend_th->loop = g_main_loop_new(backend_th->ctxt, false);
+ backend_th->thread_id = qemu_get_thread_id();
+
+ qemu_sem_post(&backend_th->init_done_sem);
+
+ g_main_loop_run(backend_th->loop);
+
+ g_main_loop_unref(backend_th->loop);
+ g_main_context_unref(backend_th->ctxt);
+ g_free(backend_th);
+ rcu_unregister_thread();
+ return NULL;
+}
+
+static void vhost_backend_thread_init(void)
+{
+ backend_th = g_malloc0(sizeof(struct backend_thread));
+ backend_th->thread_id = -1;
+ qemu_sem_init(&backend_th->init_done_sem, 0);
+
+ qemu_thread_create(&backend_th->thread, "backend-channel-worker",
+ vhost_backend_channel_worker, backend_th,
+ QEMU_THREAD_DETACHED);
+
+ while (backend_th->thread_id == -1) {
+ qemu_sem_wait(&backend_th->init_done_sem);
+ }
+}
+
static int vhost_user_backend_init(struct vhost_dev *dev, void *opaque,
Error **errp)
{
@@ -2108,10 +2156,17 @@ static int vhost_user_backend_init(struct vhost_dev *dev, void *opaque,
}
if (dev->vq_index == 0) {
- err = vhost_setup_backend_channel(dev);
- if (err < 0) {
- error_setg_errno(errp, EPROTO, "vhost_backend_init failed");
- return -EPROTO;
+ if (!backend_thread_run) {
+ vhost_backend_thread_init();
+ backend_thread_run = true;
+ }
+
+ if (backend_thread_run && backend_th) {
+ err = vhost_setup_backend_channel(dev, backend_th->ctxt);
+ if (err < 0) {
+ error_setg_errno(errp, EPROTO, "vhost_backend_init failed");
+ return -EPROTO;
+ }
}
}
--
2.17.1On Thu, Jun 29, 2023 at 02:54:15PM +0800, Rma Ma wrote:
> fix: QEMU deadlock with dpdk-vdpa
>
> QEMU start vhost-user with modern net and blk, backend use dpdk-vdpa process,
> after live migration, dest QEMU deadlock with dpdk-vdpa
>
> - QEMU sends VHOST_USER_SET_VRING_KICK to dpdk-vdpa net
> - QEMU does not need to wait for a response to this message
> - QEMU then sends VHOST_USER_SET_MEM_TABLE to dpdk-vdpa blk
> - QEMU needs to wait reply in this message
> - when dpdk-vdpa recv VHOST_USER_SET_VRING_KICK, it will send VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG to QEMU
> - dpdk-vdpa needs to wait for a response to this message
> - since QEMU vhost_user_read and backend channel are synchronous in the same thread
> - QEMU will deadlock with dpdk-vdpa
>
> Signed-off-by: Rma Ma <rma.ma@jaguarmicro.com>
Hmm this will need some thought. I'm concerned that this
is making what was previously a simple synchronous
code, multithreaded.
Also I feel dpdk needs to fix this too.
Let's document in the doc that when backend is allowed to send messages
(VHOST_USER_PROTOCOL_F_BACKEND_REQ) , neither side is allowed to block
processing incoming messsages while waiting for response.
> ---
> hw/virtio/vhost-user.c | 67 ++++++++++++++++++++++++++++++++++++++----
> 1 file changed, 61 insertions(+), 6 deletions(-)
>
> diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
> index c4e0cbd702..1f6b3a5a63 100644
> --- a/hw/virtio/vhost-user.c
> +++ b/hw/virtio/vhost-user.c
> @@ -274,6 +274,17 @@ struct scrub_regions {
> int fd_idx;
> };
>
> +struct backend_thread {
> + QemuThread thread;
> + QemuSemaphore init_done_sem;
> + int thread_id;
> + GMainContext *ctxt;
> + GMainLoop *loop;
> +};
> +
> +static struct backend_thread *backend_th;
> +static bool backend_thread_run;
> +
> static bool ioeventfd_enabled(void)
> {
> return !kvm_enabled() || kvm_eventfds_enabled();
> @@ -1696,7 +1707,8 @@ fdcleanup:
> return rc;
> }
>
> -static int vhost_setup_backend_channel(struct vhost_dev *dev)
> +static int vhost_setup_backend_channel(struct vhost_dev *dev,
> + GMainContext *ctxt)
> {
> VhostUserMsg msg = {
> .hdr.request = VHOST_USER_SET_BACKEND_REQ_FD,
> @@ -1728,7 +1740,7 @@ static int vhost_setup_backend_channel(struct vhost_dev *dev)
> u->backend_ioc = ioc;
> u->backend_src = qio_channel_add_watch_source(u->backend_ioc,
> G_IO_IN | G_IO_HUP,
> - backend_read, dev, NULL, NULL);
> + backend_read, dev, NULL, ctxt);
>
> if (reply_supported) {
> msg.hdr.flags |= VHOST_USER_NEED_REPLY_MASK;
> @@ -1981,6 +1993,42 @@ static int vhost_user_postcopy_notifier(NotifierWithReturn *notifier,
> return 0;
> }
>
> +static void *vhost_backend_channel_worker(void *opaque)
> +{
> + struct backend_thread *backend_th = opaque;
> +
> + rcu_register_thread();
> +
> + backend_th->ctxt = g_main_context_new();
> + backend_th->loop = g_main_loop_new(backend_th->ctxt, false);
> + backend_th->thread_id = qemu_get_thread_id();
> +
> + qemu_sem_post(&backend_th->init_done_sem);
> +
> + g_main_loop_run(backend_th->loop);
> +
> + g_main_loop_unref(backend_th->loop);
> + g_main_context_unref(backend_th->ctxt);
> + g_free(backend_th);
> + rcu_unregister_thread();
> + return NULL;
> +}
> +
> +static void vhost_backend_thread_init(void)
> +{
> + backend_th = g_malloc0(sizeof(struct backend_thread));
> + backend_th->thread_id = -1;
> + qemu_sem_init(&backend_th->init_done_sem, 0);
> +
> + qemu_thread_create(&backend_th->thread, "backend-channel-worker",
> + vhost_backend_channel_worker, backend_th,
> + QEMU_THREAD_DETACHED);
> +
> + while (backend_th->thread_id == -1) {
> + qemu_sem_wait(&backend_th->init_done_sem);
> + }
> +}
> +
> static int vhost_user_backend_init(struct vhost_dev *dev, void *opaque,
> Error **errp)
> {
> @@ -2108,10 +2156,17 @@ static int vhost_user_backend_init(struct vhost_dev *dev, void *opaque,
> }
>
> if (dev->vq_index == 0) {
> - err = vhost_setup_backend_channel(dev);
> - if (err < 0) {
> - error_setg_errno(errp, EPROTO, "vhost_backend_init failed");
> - return -EPROTO;
> + if (!backend_thread_run) {
> + vhost_backend_thread_init();
> + backend_thread_run = true;
> + }
> +
> + if (backend_thread_run && backend_th) {
> + err = vhost_setup_backend_channel(dev, backend_th->ctxt);
> + if (err < 0) {
> + error_setg_errno(errp, EPROTO, "vhost_backend_init failed");
> + return -EPROTO;
> + }
> }
> }
>
> --
> 2.17.1
When the vm is created，it will process vhost_dev_start in kvm_vcpu_thread_fn，
and backend channel in main thread，it seems that there is no multithreaded synchronous problem
after live migration, it will process vhost_dev_start in os_host_main_loop_wait,
and backend channel is also in main thread, it will risk of potential deadlock
so i think backend channel in separate thread can fix deadlock
does it really exist multithreaded synchronous problem while backend channel in separate thread?
________________________________
发件人: Michael S. Tsirkin <mst@redhat.com>
发送时间: 2023年6月29日 15:13
收件人: Rma Ma <rma.ma@jaguarmicro.com>
抄送: qemu-devel <qemu-devel@nongnu.org>
主题: Re: [PATCH] vhost-user: use new thread to loop backend channel.
On Thu, Jun 29, 2023 at 02:54:15PM +0800, Rma Ma wrote:
> fix: QEMU deadlock with dpdk-vdpa
>
> QEMU start vhost-user with modern net and blk, backend use dpdk-vdpa process,
> after live migration, dest QEMU deadlock with dpdk-vdpa
>
> - QEMU sends VHOST_USER_SET_VRING_KICK to dpdk-vdpa net
> - QEMU does not need to wait for a response to this message
> - QEMU then sends VHOST_USER_SET_MEM_TABLE to dpdk-vdpa blk
> - QEMU needs to wait reply in this message
> - when dpdk-vdpa recv VHOST_USER_SET_VRING_KICK, it will send VHOST_USER_BACKEND_VRING_HOST_NOTIFIER_MSG to QEMU
> - dpdk-vdpa needs to wait for a response to this message
> - since QEMU vhost_user_read and backend channel are synchronous in the same thread
> - QEMU will deadlock with dpdk-vdpa
>
> Signed-off-by: Rma Ma <rma.ma@jaguarmicro.com>
Hmm this will need some thought. I'm concerned that this
is making what was previously a simple synchronous
code, multithreaded.
Also I feel dpdk needs to fix this too.
Let's document in the doc that when backend is allowed to send messages
(VHOST_USER_PROTOCOL_F_BACKEND_REQ) , neither side is allowed to block
processing incoming messsages while waiting for response.
> ---
> hw/virtio/vhost-user.c | 67 ++++++++++++++++++++++++++++++++++++++----
> 1 file changed, 61 insertions(+), 6 deletions(-)
>
> diff --git a/hw/virtio/vhost-user.c b/hw/virtio/vhost-user.c
> index c4e0cbd702..1f6b3a5a63 100644
> --- a/hw/virtio/vhost-user.c
> +++ b/hw/virtio/vhost-user.c
> @@ -274,6 +274,17 @@ struct scrub_regions {
> int fd_idx;
> };
>
> +struct backend_thread {
> + QemuThread thread;
> + QemuSemaphore init_done_sem;
> + int thread_id;
> + GMainContext *ctxt;
> + GMainLoop *loop;
> +};
> +
> +static struct backend_thread *backend_th;
> +static bool backend_thread_run;
> +
> static bool ioeventfd_enabled(void)
> {
> return !kvm_enabled() || kvm_eventfds_enabled();
> @@ -1696,7 +1707,8 @@ fdcleanup:
> return rc;
> }
>
> -static int vhost_setup_backend_channel(struct vhost_dev *dev)
> +static int vhost_setup_backend_channel(struct vhost_dev *dev,
> + GMainContext *ctxt)
> {
> VhostUserMsg msg = {
> .hdr.request = VHOST_USER_SET_BACKEND_REQ_FD,
> @@ -1728,7 +1740,7 @@ static int vhost_setup_backend_channel(struct vhost_dev *dev)
> u->backend_ioc = ioc;
> u->backend_src = qio_channel_add_watch_source(u->backend_ioc,
> G_IO_IN | G_IO_HUP,
> - backend_read, dev, NULL, NULL);
> + backend_read, dev, NULL, ctxt);
>
> if (reply_supported) {
> msg.hdr.flags |= VHOST_USER_NEED_REPLY_MASK;
> @@ -1981,6 +1993,42 @@ static int vhost_user_postcopy_notifier(NotifierWithReturn *notifier,
> return 0;
> }
>
> +static void *vhost_backend_channel_worker(void *opaque)
> +{
> + struct backend_thread *backend_th = opaque;
> +
> + rcu_register_thread();
> +
> + backend_th->ctxt = g_main_context_new();
> + backend_th->loop = g_main_loop_new(backend_th->ctxt, false);
> + backend_th->thread_id = qemu_get_thread_id();
> +
> + qemu_sem_post(&backend_th->init_done_sem);
> +
> + g_main_loop_run(backend_th->loop);
> +
> + g_main_loop_unref(backend_th->loop);
> + g_main_context_unref(backend_th->ctxt);
> + g_free(backend_th);
> + rcu_unregister_thread();
> + return NULL;
> +}
> +
> +static void vhost_backend_thread_init(void)
> +{
> + backend_th = g_malloc0(sizeof(struct backend_thread));
> + backend_th->thread_id = -1;
> + qemu_sem_init(&backend_th->init_done_sem, 0);
> +
> + qemu_thread_create(&backend_th->thread, "backend-channel-worker",
> + vhost_backend_channel_worker, backend_th,
> + QEMU_THREAD_DETACHED);
> +
> + while (backend_th->thread_id == -1) {
> + qemu_sem_wait(&backend_th->init_done_sem);
> + }
> +}
> +
> static int vhost_user_backend_init(struct vhost_dev *dev, void *opaque,
> Error **errp)
> {
> @@ -2108,10 +2156,17 @@ static int vhost_user_backend_init(struct vhost_dev *dev, void *opaque,
> }
>
> if (dev->vq_index == 0) {
> - err = vhost_setup_backend_channel(dev);
> - if (err < 0) {
> - error_setg_errno(errp, EPROTO, "vhost_backend_init failed");
> - return -EPROTO;
> + if (!backend_thread_run) {
> + vhost_backend_thread_init();
> + backend_thread_run = true;
> + }
> +
> + if (backend_thread_run && backend_th) {
> + err = vhost_setup_backend_channel(dev, backend_th->ctxt);
> + if (err < 0) {
> + error_setg_errno(errp, EPROTO, "vhost_backend_init failed");
> + return -EPROTO;
> + }
> }
> }
>
> --
> 2.17.1
© 2016 - 2024 Red Hat, Inc.