Re: [PATCH v9 2/4] vhost: Reintroduce kthread mode support in vhost
On Mon, Apr 21, 2025 at 11:39:14AM +0800, Jason Wang wrote:
> On Mon, Apr 21, 2025 at 10:45 AM Cindy Lu wrote:
> >
> > This patch reintroduces kthread mode support in vhost,
> > It also introduces struct vhost_worker_ops to abstract
> > worker create/stop/wakeup operations.
> >
> > * Bring back the original vhost_worker() implementation,
> > and renamed to vhost_run_work_kthread_list().
> >
> > * Add cgroup support for the kthread
> >
> > * Introduce struct vhost_worker_ops:
> > - Encapsulates create / stop / wake‑up callbacks.
> > - vhost_worker_create() selects the proper ops according to
> > inherit_owner.
> >
> > This partially reverts or improves upon:
> > commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
> > commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
> >
> > Signed-off-by: Cindy Lu
> > ---
> > drivers/vhost/vhost.c | 188 ++
> > drivers/vhost/vhost.h | 12 +++
> > 2 files changed, 182 insertions(+), 18 deletions(-)
> >
> > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > index 250dc43f1786..be97028a8baf 100644
> > --- a/drivers/vhost/vhost.c
> > +++ b/drivers/vhost/vhost.c
> > @@ -22,6 +22,7 @@
> > #include
> > #include
> > #include
> > +#include
> > #include
> > #include
> > #include
> > @@ -242,7 +243,7 @@ static void vhost_worker_queue(struct vhost_worker
> > *worker,
> > * test_and_set_bit() implies a memory barrier.
> > */
> > llist_add(&work->node, &worker->work_list);
> > - vhost_task_wake(worker->vtsk);
> > + worker->ops->wakeup(worker);
> > }
> > }
> >
> > @@ -388,6 +389,44 @@ static void vhost_vq_reset(struct vhost_dev *dev,
> > __vhost_vq_meta_reset(vq);
> > }
> >
> > +static int vhost_run_work_kthread_list(void *data)
> > +{
> > + struct vhost_worker *worker = data;
> > + struct vhost_work *work, *work_next;
> > + struct vhost_dev *dev = worker->dev;
> > + struct llist_node *node;
> > +
> > + kthread_use_mm(dev->mm);
> > +
> > + for (;;) {
> > + /* mb paired w/ kthread_stop */
> > + set_current_state(TASK_INTERRUPTIBLE);
> > +
> > + if (kthread_should_stop()) {
> > + __set_current_state(TASK_RUNNING);
> > + break;
> > + }
> > + node = llist_del_all(&worker->work_list);
> > + if (!node)
> > + schedule();
> > +
> > + node = llist_reverse_order(node);
> > + /* make sure flag is seen after deletion */
> > + smp_wmb();
> > + llist_for_each_entry_safe(work, work_next, node, node) {
> > + clear_bit(VHOST_WORK_QUEUED, &work->flags);
> > + __set_current_state(TASK_RUNNING);
> > + kcov_remote_start_common(worker->kcov_handle);
> > + work->fn(work);
> > + kcov_remote_stop();
> > + cond_resched();
> > + }
> > + }
> > + kthread_unuse_mm(dev->mm);
> > +
> > + return 0;
> > +}
> > +
> > static bool vhost_run_work_list(void *data)
> > {
> > struct vhost_worker *worker = data;
> > @@ -582,6 +621,46 @@ long vhost_dev_check_owner(struct vhost_dev *dev)
> > }
> > EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
> >
> > +struct vhost_attach_cgroups_struct {
> > + struct vhost_work work;
> > + struct task_struct *owner;
> > + int ret;
> > +};
> > +
> > +static void vhost_attach_cgroups_work(struct vhost_work *work)
> > +{
> > + struct vhost_attach_cgroups_struct *s;
> > +
> > + s = container_of(work, struct vhost_attach_cgroups_struct, work);
> > + s->ret = cgroup_attach_task_all(s->owner, current);
> > +}
> > +
> > +static int vhost_attach_task_to_cgroups(struct vhost_worker *worker)
> > +{
> > + struct vhost_attach_cgroups_struct attach;
> > + int saved_cnt;
> > +
> > + attach.owner = current;
> > +
> > + vhost_work_init(&attach.work, vhost_attach_cgroups_work);
> > + vhost_worker_queue(worker, &attach.work);
> > +
> > + mutex_lock(&worker->mutex);
> > +
> > + /*
> > +* Bypass attachment_cnt check in __vhost_worker_flush:
> > +* Temporarily change it to INT_MAX to bypass the check
> > +*/
> > + saved_cnt = worker->attachment_cnt;
> > + worker->attachment_cnt = INT_MAX;
> > + __vhost_worker_flush(worker);
> > + worker->attachment_cnt = saved_cnt;
>
> I wonder if it's easier to re-introduce the flush that was used before
> vhost kthread to avoid the tricks here. We can have flush ops for
> example.
>
> Thanks
Nah we do not need ops, __vhost_worker_flush is just an internal
function. Refactor it so we can call the part without the check.
--
MST
Re: [PATCH v9 2/4] vhost: Reintroduce kthread mode support in vhost
On Mon, Apr 21, 2025 at 10:44:08AM +0800, Cindy Lu wrote:
> This patch reintroduces kthread mode support in vhost,
> It also introduces struct vhost_worker_ops to abstract
> worker create/stop/wakeup operations.
>
> * Bring back the original vhost_worker() implementation,
> and renamed to vhost_run_work_kthread_list().
>
> * Add cgroup support for the kthread
>
> * Introduce struct vhost_worker_ops:
> - Encapsulates create / stop / wake‑up callbacks.
> - vhost_worker_create() selects the proper ops according to
> inherit_owner.
>
> This partially reverts or improves upon:
> commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
> commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
>
> Signed-off-by: Cindy Lu
> ---
> drivers/vhost/vhost.c | 188 ++
> drivers/vhost/vhost.h | 12 +++
> 2 files changed, 182 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 250dc43f1786..be97028a8baf 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -22,6 +22,7 @@
> #include
> #include
> #include
> +#include
> #include
> #include
> #include
> @@ -242,7 +243,7 @@ static void vhost_worker_queue(struct vhost_worker
> *worker,
>* test_and_set_bit() implies a memory barrier.
>*/
> llist_add(&work->node, &worker->work_list);
> - vhost_task_wake(worker->vtsk);
> + worker->ops->wakeup(worker);
> }
> }
>
> @@ -388,6 +389,44 @@ static void vhost_vq_reset(struct vhost_dev *dev,
> __vhost_vq_meta_reset(vq);
> }
>
> +static int vhost_run_work_kthread_list(void *data)
> +{
> + struct vhost_worker *worker = data;
> + struct vhost_work *work, *work_next;
> + struct vhost_dev *dev = worker->dev;
> + struct llist_node *node;
> +
> + kthread_use_mm(dev->mm);
> +
> + for (;;) {
> + /* mb paired w/ kthread_stop */
> + set_current_state(TASK_INTERRUPTIBLE);
> +
> + if (kthread_should_stop()) {
> + __set_current_state(TASK_RUNNING);
> + break;
> + }
> + node = llist_del_all(&worker->work_list);
> + if (!node)
> + schedule();
> +
> + node = llist_reverse_order(node);
> + /* make sure flag is seen after deletion */
> + smp_wmb();
> + llist_for_each_entry_safe(work, work_next, node, node) {
> + clear_bit(VHOST_WORK_QUEUED, &work->flags);
> + __set_current_state(TASK_RUNNING);
> + kcov_remote_start_common(worker->kcov_handle);
> + work->fn(work);
> + kcov_remote_stop();
> + cond_resched();
> + }
> + }
> + kthread_unuse_mm(dev->mm);
> +
> + return 0;
> +}
> +
> static bool vhost_run_work_list(void *data)
> {
> struct vhost_worker *worker = data;
> @@ -582,6 +621,46 @@ long vhost_dev_check_owner(struct vhost_dev *dev)
> }
> EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
>
> +struct vhost_attach_cgroups_struct {
> + struct vhost_work work;
> + struct task_struct *owner;
> + int ret;
> +};
> +
> +static void vhost_attach_cgroups_work(struct vhost_work *work)
> +{
> + struct vhost_attach_cgroups_struct *s;
> +
> + s = container_of(work, struct vhost_attach_cgroups_struct, work);
> + s->ret = cgroup_attach_task_all(s->owner, current);
> +}
> +
> +static int vhost_attach_task_to_cgroups(struct vhost_worker *worker)
> +{
> + struct vhost_attach_cgroups_struct attach;
> + int saved_cnt;
> +
> + attach.owner = current;
> +
> + vhost_work_init(&attach.work, vhost_attach_cgroups_work);
> + vhost_worker_queue(worker, &attach.work);
> +
> + mutex_lock(&worker->mutex);
> +
> + /*
> + * Bypass attachment_cnt check in __vhost_worker_flush:
> + * Temporarily change it to INT_MAX to bypass the check
> + */
> + saved_cnt = worker->attachment_cnt;
> + worker->attachment_cnt = INT_MAX;
> + __vhost_worker_flush(worker);
> + worker->attachment_cnt = saved_cnt;
You mean this one?
if (!worker->attachment_cnt || worker->killed)
return;
Just introduce a variant of __vhost_worker_flush that
skips this check.
E.g.
Rename __vhost_worker_flush -> _vhost_worker_flush.
then rework:
static void _vhost_worker_flush(struct vhost_worker *worker)
{
struct vhost_flush_struct flush;
if (!worker->attachment_cnt || worker->killed)
return;
__vhost_worker_flush(worker);
}
> +
> + mutex_unlock(&worker->mutex);
> +
> + return attach.ret;
> +}
> +
> /* Caller should have device mutex */
> bool vhost_dev_has_owner(struct vhost_dev *dev)
> {
> @@ -627,7 +706,7 @@ static void vhost_worker_d
Re: [PATCH v9 2/4] vhost: Reintroduce kthread mode support in vhost
On Mon, Apr 21, 2025 at 10:45 AM Cindy Lu wrote:
>
> This patch reintroduces kthread mode support in vhost,
> It also introduces struct vhost_worker_ops to abstract
> worker create/stop/wakeup operations.
>
> * Bring back the original vhost_worker() implementation,
> and renamed to vhost_run_work_kthread_list().
>
> * Add cgroup support for the kthread
>
> * Introduce struct vhost_worker_ops:
> - Encapsulates create / stop / wake‑up callbacks.
> - vhost_worker_create() selects the proper ops according to
> inherit_owner.
>
> This partially reverts or improves upon:
> commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
> commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
>
> Signed-off-by: Cindy Lu
> ---
> drivers/vhost/vhost.c | 188 ++
> drivers/vhost/vhost.h | 12 +++
> 2 files changed, 182 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 250dc43f1786..be97028a8baf 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -22,6 +22,7 @@
> #include
> #include
> #include
> +#include
> #include
> #include
> #include
> @@ -242,7 +243,7 @@ static void vhost_worker_queue(struct vhost_worker
> *worker,
> * test_and_set_bit() implies a memory barrier.
> */
> llist_add(&work->node, &worker->work_list);
> - vhost_task_wake(worker->vtsk);
> + worker->ops->wakeup(worker);
> }
> }
>
> @@ -388,6 +389,44 @@ static void vhost_vq_reset(struct vhost_dev *dev,
> __vhost_vq_meta_reset(vq);
> }
>
> +static int vhost_run_work_kthread_list(void *data)
> +{
> + struct vhost_worker *worker = data;
> + struct vhost_work *work, *work_next;
> + struct vhost_dev *dev = worker->dev;
> + struct llist_node *node;
> +
> + kthread_use_mm(dev->mm);
> +
> + for (;;) {
> + /* mb paired w/ kthread_stop */
> + set_current_state(TASK_INTERRUPTIBLE);
> +
> + if (kthread_should_stop()) {
> + __set_current_state(TASK_RUNNING);
> + break;
> + }
> + node = llist_del_all(&worker->work_list);
> + if (!node)
> + schedule();
> +
> + node = llist_reverse_order(node);
> + /* make sure flag is seen after deletion */
> + smp_wmb();
> + llist_for_each_entry_safe(work, work_next, node, node) {
> + clear_bit(VHOST_WORK_QUEUED, &work->flags);
> + __set_current_state(TASK_RUNNING);
> + kcov_remote_start_common(worker->kcov_handle);
> + work->fn(work);
> + kcov_remote_stop();
> + cond_resched();
> + }
> + }
> + kthread_unuse_mm(dev->mm);
> +
> + return 0;
> +}
> +
> static bool vhost_run_work_list(void *data)
> {
> struct vhost_worker *worker = data;
> @@ -582,6 +621,46 @@ long vhost_dev_check_owner(struct vhost_dev *dev)
> }
> EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
>
> +struct vhost_attach_cgroups_struct {
> + struct vhost_work work;
> + struct task_struct *owner;
> + int ret;
> +};
> +
> +static void vhost_attach_cgroups_work(struct vhost_work *work)
> +{
> + struct vhost_attach_cgroups_struct *s;
> +
> + s = container_of(work, struct vhost_attach_cgroups_struct, work);
> + s->ret = cgroup_attach_task_all(s->owner, current);
> +}
> +
> +static int vhost_attach_task_to_cgroups(struct vhost_worker *worker)
> +{
> + struct vhost_attach_cgroups_struct attach;
> + int saved_cnt;
> +
> + attach.owner = current;
> +
> + vhost_work_init(&attach.work, vhost_attach_cgroups_work);
> + vhost_worker_queue(worker, &attach.work);
> +
> + mutex_lock(&worker->mutex);
> +
> + /*
> +* Bypass attachment_cnt check in __vhost_worker_flush:
> +* Temporarily change it to INT_MAX to bypass the check
> +*/
> + saved_cnt = worker->attachment_cnt;
> + worker->attachment_cnt = INT_MAX;
> + __vhost_worker_flush(worker);
> + worker->attachment_cnt = saved_cnt;
I wonder if it's easier to re-introduce the flush that was used before
vhost kthread to avoid the tricks here. We can have flush ops for
example.
Thanks
[PATCH v9 2/4] vhost: Reintroduce kthread mode support in vhost
This patch reintroduces kthread mode support in vhost,
It also introduces struct vhost_worker_ops to abstract
worker create/stop/wakeup operations.
* Bring back the original vhost_worker() implementation,
and renamed to vhost_run_work_kthread_list().
* Add cgroup support for the kthread
* Introduce struct vhost_worker_ops:
- Encapsulates create / stop / wake‑up callbacks.
- vhost_worker_create() selects the proper ops according to
inherit_owner.
This partially reverts or improves upon:
commit 6e890c5d5021 ("vhost: use vhost_tasks for worker threads")
commit 1cdaafa1b8b4 ("vhost: replace single worker pointer with xarray")
Signed-off-by: Cindy Lu
---
drivers/vhost/vhost.c | 188 ++
drivers/vhost/vhost.h | 12 +++
2 files changed, 182 insertions(+), 18 deletions(-)
diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index 250dc43f1786..be97028a8baf 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -22,6 +22,7 @@
#include
#include
#include
+#include
#include
#include
#include
@@ -242,7 +243,7 @@ static void vhost_worker_queue(struct vhost_worker *worker,
* test_and_set_bit() implies a memory barrier.
*/
llist_add(&work->node, &worker->work_list);
- vhost_task_wake(worker->vtsk);
+ worker->ops->wakeup(worker);
}
}
@@ -388,6 +389,44 @@ static void vhost_vq_reset(struct vhost_dev *dev,
__vhost_vq_meta_reset(vq);
}
+static int vhost_run_work_kthread_list(void *data)
+{
+ struct vhost_worker *worker = data;
+ struct vhost_work *work, *work_next;
+ struct vhost_dev *dev = worker->dev;
+ struct llist_node *node;
+
+ kthread_use_mm(dev->mm);
+
+ for (;;) {
+ /* mb paired w/ kthread_stop */
+ set_current_state(TASK_INTERRUPTIBLE);
+
+ if (kthread_should_stop()) {
+ __set_current_state(TASK_RUNNING);
+ break;
+ }
+ node = llist_del_all(&worker->work_list);
+ if (!node)
+ schedule();
+
+ node = llist_reverse_order(node);
+ /* make sure flag is seen after deletion */
+ smp_wmb();
+ llist_for_each_entry_safe(work, work_next, node, node) {
+ clear_bit(VHOST_WORK_QUEUED, &work->flags);
+ __set_current_state(TASK_RUNNING);
+ kcov_remote_start_common(worker->kcov_handle);
+ work->fn(work);
+ kcov_remote_stop();
+ cond_resched();
+ }
+ }
+ kthread_unuse_mm(dev->mm);
+
+ return 0;
+}
+
static bool vhost_run_work_list(void *data)
{
struct vhost_worker *worker = data;
@@ -582,6 +621,46 @@ long vhost_dev_check_owner(struct vhost_dev *dev)
}
EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
+struct vhost_attach_cgroups_struct {
+ struct vhost_work work;
+ struct task_struct *owner;
+ int ret;
+};
+
+static void vhost_attach_cgroups_work(struct vhost_work *work)
+{
+ struct vhost_attach_cgroups_struct *s;
+
+ s = container_of(work, struct vhost_attach_cgroups_struct, work);
+ s->ret = cgroup_attach_task_all(s->owner, current);
+}
+
+static int vhost_attach_task_to_cgroups(struct vhost_worker *worker)
+{
+ struct vhost_attach_cgroups_struct attach;
+ int saved_cnt;
+
+ attach.owner = current;
+
+ vhost_work_init(&attach.work, vhost_attach_cgroups_work);
+ vhost_worker_queue(worker, &attach.work);
+
+ mutex_lock(&worker->mutex);
+
+ /*
+* Bypass attachment_cnt check in __vhost_worker_flush:
+* Temporarily change it to INT_MAX to bypass the check
+*/
+ saved_cnt = worker->attachment_cnt;
+ worker->attachment_cnt = INT_MAX;
+ __vhost_worker_flush(worker);
+ worker->attachment_cnt = saved_cnt;
+
+ mutex_unlock(&worker->mutex);
+
+ return attach.ret;
+}
+
/* Caller should have device mutex */
bool vhost_dev_has_owner(struct vhost_dev *dev)
{
@@ -627,7 +706,7 @@ static void vhost_worker_destroy(struct vhost_dev *dev,
WARN_ON(!llist_empty(&worker->work_list));
xa_erase(&dev->worker_xa, worker->id);
- vhost_task_stop(worker->vtsk);
+ worker->ops->stop(worker);
kfree(worker);
}
@@ -650,42 +729,115 @@ static void vhost_workers_free(struct vhost_dev *dev)
xa_destroy(&dev->worker_xa);
}
+static void vhost_task_wakeup(struct vhost_worker *worker)
+{
+ return vhost_task_wake(worker->vtsk);
+}
+
+static void vhost_kthread_wakeup(struct vhost_worker *worker)
+{
+ wake_up_process(worker->kthread_task);
+}
+
+static void vhost_task_do_stop(struct vhost_worker *worker)
+{
+ return vhost_task_stop(worker->vtsk);
+}
+
+static void vhost_kthread_do_stop
