| Message ID | 20210525180600.6349-3-michael.christie@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/9] vhost: move worker thread fields to new struct | expand |
On Tue, May 25, 2021 at 01:05:53PM -0500, Mike Christie wrote: > The next patch will add new ioctls that allows userspace to create workers > and bind them to devs and vqs after VHOST_SET_OWNER. To support older > tools, newer tools that want to go wild with worker threads, and newer > tools that want the old/default behavior this patch moves the default > worker setup to the kick setup. > > After the first vq's kick/poll setup is done we could start to get works > queued, so this is the point when we must have a worker setup. If we are > using older tools or the newer tools just want the default single vhost > thread per dev behavior then it will automatically be done here. If the > tools are using the newer ioctls that have already created the workers > then we also detect that here and do nothing. > > Signed-off-by: Mike Christie <michael.christie@oracle.com> > --- > drivers/vhost/vhost.c | 23 ++++++++++++++++------- > 1 file changed, 16 insertions(+), 7 deletions(-) Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c index 0cd19b1a832e..a291cde95c43 100644 --- a/drivers/vhost/vhost.c +++ b/drivers/vhost/vhost.c @@ -629,6 +629,15 @@ static int vhost_worker_create(struct vhost_dev *dev) return ret; } +/* Caller must have device mutex */ +static int vhost_worker_try_create_def(struct vhost_dev *dev) +{ + if (!dev->use_worker || dev->worker) + return 0; + + return vhost_worker_create(dev); +} + /* Caller should have device mutex */ long vhost_dev_set_owner(struct vhost_dev *dev) { @@ -643,11 +652,6 @@ long vhost_dev_set_owner(struct vhost_dev *dev) vhost_attach_mm(dev); dev->kcov_handle = kcov_common_handle(); - if (dev->use_worker) { - err = vhost_worker_create(dev); - if (err) - goto err_worker; - } err = vhost_dev_alloc_iovecs(dev); if (err) @@ -655,8 +659,6 @@ long vhost_dev_set_owner(struct vhost_dev *dev) return 0; err_iovecs: - vhost_worker_free(dev); -err_worker: vhost_detach_mm(dev); dev->kcov_handle = 0; err_mm: @@ -1665,6 +1667,13 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg r = -EFAULT; break; } + + if (f.fd != VHOST_FILE_UNBIND) { + r = vhost_worker_try_create_def(d); + if (r) + break; + } + eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd); if (IS_ERR(eventfp)) { r = PTR_ERR(eventfp);
The next patch will add new ioctls that allows userspace to create workers and bind them to devs and vqs after VHOST_SET_OWNER. To support older tools, newer tools that want to go wild with worker threads, and newer tools that want the old/default behavior this patch moves the default worker setup to the kick setup. After the first vq's kick/poll setup is done we could start to get works queued, so this is the point when we must have a worker setup. If we are using older tools or the newer tools just want the default single vhost thread per dev behavior then it will automatically be done here. If the tools are using the newer ioctls that have already created the workers then we also detect that here and do nothing. Signed-off-by: Mike Christie <michael.christie@oracle.com> --- drivers/vhost/vhost.c | 23 ++++++++++++++++------- 1 file changed, 16 insertions(+), 7 deletions(-)