On Sun, Jun 1, 2025 at 6:49 PM Michael S. Tsirkin [off-list ref] wrote:

On Sat, May 31, 2025 at 05:57:28PM +0800, Cindy Lu wrote:

quoted

This patch introduces a new UAPI that allows the vhost device to select
in kthread mode. Userspace applications can utilize IOCTL
VHOST_FORK_FROM_OWNER to select between task and kthread modes, which
must be invoked before IOCTL VHOST_SET_OWNER, as the worker will be
created during that call.

The VHOST_NEW_WORKER requires the inherit_owner setting to be true, and
a check has been added to ensure proper configuration.

Additionally, a new KConfig option, CONFIG_VHOST_ENABLE_FORK_OWNER_IOCTL,
is introduced to control the availability of the IOCTL
VHOST_FORK_FROM_OWNER. When CONFIG_VHOST_ENABLE_FORK_OWNER_IOCTL is set
to n, the IOCTL is disabled, and any attempt to use it will result in a
failure.

Signed-off-by: Cindy Lu <redacted>

I propose renaming
CONFIG_VHOST_ENABLE_FORK_OWNER_IOCTL
to
CONFIG_VHOST_ENABLE_FORK_OWNER_CONTROL
and it should also control the presence of the module parameter
and a get ioctl (more on which below).

Otherwise we can get a situation where task mode is disabled and
there is no way for userspace to override or check.

sure, will do
Thanks
cindy

quoted

---
 drivers/vhost/Kconfig      | 13 +++++++++++++
 drivers/vhost/vhost.c      | 30 +++++++++++++++++++++++++++++-
 include/uapi/linux/vhost.h | 16 ++++++++++++++++
 3 files changed, 58 insertions(+), 1 deletion(-)

diff --git a/drivers/vhost/Kconfig b/drivers/vhost/Kconfig
index 020d4fbb947c..300e474b60fd 100644
--- a/drivers/vhost/Kconfig
+++ b/drivers/vhost/Kconfig

@@ -96,3 +96,16 @@ config VHOST_CROSS_ENDIAN_LEGACY
        If unsure, say "N".

 endif
+
+config VHOST_ENABLE_FORK_OWNER_IOCTL
+     bool "Enable IOCTL VHOST_FORK_FROM_OWNER"
+     default n
+     help
+       This option enables the IOCTL VHOST_FORK_FROM_OWNER, allowing
+       userspace applications to modify the thread mode for vhost devices.
+
+       By default, `CONFIG_VHOST_ENABLE_FORK_OWNER_IOCTL` is set to `n`,
+       which disables the IOCTL. When enabled (y), the IOCTL allows
+       users to set the mode as needed.
+
+       If unsure, say "N".

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index 2d2909be1bb2..cfa60dc438f9 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c

@@ -1022,6 +1022,13 @@ long vhost_worker_ioctl(struct vhost_dev *dev, unsigned int ioctl,
      switch (ioctl) {
      /* dev worker ioctls */
      case VHOST_NEW_WORKER:
+             /*
+              * vhost_tasks will account for worker threads under the parent's
+              * NPROC value but kthreads do not. To avoid userspace overflowing
+              * the system with worker threads inherit_owner must be true.
+              */
+             if (!dev->inherit_owner)
+                     return -EFAULT;
              ret = vhost_new_worker(dev, &state);
              if (!ret && copy_to_user(argp, &state, sizeof(state)))
                      ret = -EFAULT;

@@ -1138,7 +1145,7 @@ void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
      int i;

      vhost_dev_cleanup(dev);
-
+     dev->inherit_owner = inherit_owner_default;
      dev->umem = umem;
      /* We don't need VQ locks below since vhost_dev_cleanup makes sure
       * VQs aren't running.

@@ -2292,6 +2299,27 @@ long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
              goto done;
      }

+#ifdef CONFIG_VHOST_ENABLE_FORK_OWNER_IOCTL
+     if (ioctl == VHOST_FORK_FROM_OWNER) {
+             u8 inherit_owner;
+             /*inherit_owner can only be modified before owner is set*/
+             if (vhost_dev_has_owner(d)) {
+                     r = -EBUSY;
+                     goto done;
+             }
+             if (copy_from_user(&inherit_owner, argp, sizeof(u8))) {
+                     r = -EFAULT;
+                     goto done;
+             }
+             if (inherit_owner > 1) {
+                     r = -EINVAL;
+                     goto done;
+             }
+             d->inherit_owner = (bool)inherit_owner;
+             r = 0;
+             goto done;
+     }
+#endif
      /* You must be the owner to do anything else */
      r = vhost_dev_check_owner(d);
      if (r)

diff --git a/include/uapi/linux/vhost.h b/include/uapi/linux/vhost.h
index d4b3e2ae1314..d2692c7ef450 100644
--- a/include/uapi/linux/vhost.h
+++ b/include/uapi/linux/vhost.h

@@ -235,4 +235,20 @@
  */
 #define VHOST_VDPA_GET_VRING_SIZE    _IOWR(VHOST_VIRTIO, 0x82,       \
                                            struct vhost_vring_state)
+
+/**
+ * VHOST_FORK_FROM_OWNER - Set the inherit_owner flag for the vhost device,
+ * This ioctl must called before VHOST_SET_OWNER.
+ *
+ * @param inherit_owner: An 8-bit value that determines the vhost thread mode
+ *
+ * When inherit_owner is set to 1(default value):
+ *   - Vhost will create tasks similar to processes forked from the owner,
+ *     inheriting all of the owner's attributes.
+ *
+ * When inherit_owner is set to 0:
+ *   - Vhost will create tasks as kernel thread.
+ */
+#define VHOST_FORK_FROM_OWNER _IOW(VHOST_VIRTIO, 0x83, __u8)

Given default now depends on the module parameter, we should
have both GET and SET ioctls. All controlled by the kconfig knob.

Sure, will do
Thanks
cindy

quoted

+
 #endif
--
2.45.0