Jason Wang
2021-Oct-26 05:37 UTC
[PATCH V3 11/11] vhost: allow userspace to create workers
? 2021/10/22 ??1:19, Mike Christie ??:> This patch allows userspace to create workers and bind them to vqs. You > can have N workers per dev and also share N workers with M vqs. > > Signed-off-by: Mike Christie <michael.christie at oracle.com>A question, who is the best one to determine the binding? Is it the VMM (Qemu etc) or the management stack? If the latter, it looks to me it's better to expose this via sysfs?> --- > drivers/vhost/vhost.c | 99 ++++++++++++++++++++++++++++---- > drivers/vhost/vhost.h | 2 +- > include/uapi/linux/vhost.h | 11 ++++ > include/uapi/linux/vhost_types.h | 12 ++++ > 4 files changed, 112 insertions(+), 12 deletions(-) > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c > index 04f43a6445e1..c86e88d7f35c 100644 > --- a/drivers/vhost/vhost.c > +++ b/drivers/vhost/vhost.c > @@ -493,7 +493,6 @@ void vhost_dev_init(struct vhost_dev *dev, > dev->umem = NULL; > dev->iotlb = NULL; > dev->mm = NULL; > - dev->worker = NULL; > dev->iov_limit = iov_limit; > dev->weight = weight; > dev->byte_weight = byte_weight; > @@ -576,20 +575,40 @@ static void vhost_worker_stop(struct vhost_worker *worker) > wait_for_completion(worker->exit_done); > } > > -static void vhost_worker_free(struct vhost_dev *dev) > -{ > - struct vhost_worker *worker = dev->worker; > > +static void vhost_worker_put(struct vhost_worker *worker) > +{ > if (!worker) > return; > > - dev->worker = NULL; > + if (!refcount_dec_and_test(&worker->refcount)) > + return; > + > WARN_ON(!llist_empty(&worker->work_list)); > vhost_worker_stop(worker); > kfree(worker); > } > > -static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev) > +static void vhost_vq_clear_worker(struct vhost_virtqueue *vq) > +{ > + if (vq->worker) > + vhost_worker_put(vq->worker); > + vq->worker = NULL; > +} > + > +static void vhost_workers_free(struct vhost_dev *dev) > +{ > + int i; > + > + if (!dev->use_worker) > + return; > + > + for (i = 0; i < dev->nvqs; i++) > + vhost_vq_clear_worker(dev->vqs[i]); > +} > + > +static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev, > + int init_vq_map_count) > { > struct vhost_worker *worker; > struct task_struct *task; > @@ -598,9 +617,9 @@ static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev) > if (!worker) > return NULL; > > - dev->worker = worker; > worker->kcov_handle = kcov_common_handle(); > init_llist_head(&worker->work_list); > + refcount_set(&worker->refcount, init_vq_map_count); > > /* > * vhost used to use the kthread API which ignores all signals by > @@ -617,10 +636,58 @@ static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev) > > free_worker: > kfree(worker); > - dev->worker = NULL; > return NULL; > } > > +static struct vhost_worker *vhost_worker_find(struct vhost_dev *dev, pid_t pid) > +{ > + struct vhost_worker *worker = NULL; > + int i; > + > + for (i = 0; i < dev->nvqs; i++) { > + if (dev->vqs[i]->worker->task->pid != pid) > + continue; > + > + worker = dev->vqs[i]->worker; > + break; > + } > + > + return worker; > +} > + > +/* Caller must have device mutex */ > +static int vhost_vq_setup_worker(struct vhost_virtqueue *vq, > + struct vhost_vring_worker *info) > +{ > + struct vhost_dev *dev = vq->dev; > + struct vhost_worker *worker; > + > + if (!dev->use_worker) > + return -EINVAL; > + > + /* We don't support setting a worker on an active vq */ > + if (vq->private_data) > + return -EBUSY;Is it valuable to allow the worker switching on active vq?> + > + if (info->pid == VHOST_VRING_NEW_WORKER) { > + worker = vhost_worker_create(dev, 1); > + if (!worker) > + return -ENOMEM; > + > + info->pid = worker->task->pid; > + } else { > + worker = vhost_worker_find(dev, info->pid); > + if (!worker) > + return -ENODEV; > + > + refcount_inc(&worker->refcount); > + } > + > + vhost_vq_clear_worker(vq); > + vq->worker = worker; > + return 0; > +} > + > /* Caller should have device mutex */ > long vhost_dev_set_owner(struct vhost_dev *dev) > { > @@ -636,7 +703,7 @@ long vhost_dev_set_owner(struct vhost_dev *dev) > vhost_attach_mm(dev); > > if (dev->use_worker) { > - worker = vhost_worker_create(dev); > + worker = vhost_worker_create(dev, dev->nvqs); > if (!worker) > goto err_worker; > > @@ -650,7 +717,7 @@ long vhost_dev_set_owner(struct vhost_dev *dev) > > return 0; > err_iovecs: > - vhost_worker_free(dev); > + vhost_workers_free(dev); > err_worker: > vhost_detach_mm(dev); > err_mm: > @@ -742,7 +809,7 @@ void vhost_dev_cleanup(struct vhost_dev *dev) > dev->iotlb = NULL; > vhost_clear_msg(dev); > wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM); > - vhost_worker_free(dev); > + vhost_workers_free(dev); > vhost_detach_mm(dev); > } > EXPORT_SYMBOL_GPL(vhost_dev_cleanup); > @@ -1612,6 +1679,7 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg > struct eventfd_ctx *ctx = NULL; > u32 __user *idxp = argp; > struct vhost_virtqueue *vq; > + struct vhost_vring_worker w; > struct vhost_vring_state s; > struct vhost_vring_file f; > u32 idx; > @@ -1719,6 +1787,15 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg > if (copy_to_user(argp, &s, sizeof(s))) > r = -EFAULT; > break; > + case VHOST_SET_VRING_WORKER: > + if (copy_from_user(&w, argp, sizeof(w))) { > + r = -EFAULT; > + break; > + } > + r = vhost_vq_setup_worker(vq, &w); > + if (!r && copy_to_user(argp, &w, sizeof(w))) > + r = -EFAULT; > + break; > default: > r = -ENOIOCTLCMD; > } > diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h > index 33c63b24187a..0911d1a9bd3b 100644 > --- a/drivers/vhost/vhost.h > +++ b/drivers/vhost/vhost.h > @@ -35,6 +35,7 @@ struct vhost_worker { > struct llist_head work_list; > u64 kcov_handle; > unsigned long flags; > + refcount_t refcount; > }; > > /* Poll a file (eventfd or socket) */ > @@ -160,7 +161,6 @@ struct vhost_dev { > struct vhost_virtqueue **vqs; > int nvqs; > struct eventfd_ctx *log_ctx; > - struct vhost_worker *worker; > struct vhost_iotlb *umem; > struct vhost_iotlb *iotlb; > spinlock_t iotlb_lock; > diff --git a/include/uapi/linux/vhost.h b/include/uapi/linux/vhost.h > index c998860d7bbc..e5c0669430e5 100644 > --- a/include/uapi/linux/vhost.h > +++ b/include/uapi/linux/vhost.h > @@ -70,6 +70,17 @@ > #define VHOST_VRING_BIG_ENDIAN 1 > #define VHOST_SET_VRING_ENDIAN _IOW(VHOST_VIRTIO, 0x13, struct vhost_vring_state) > #define VHOST_GET_VRING_ENDIAN _IOW(VHOST_VIRTIO, 0x14, struct vhost_vring_state) > +/* By default, a device gets one vhost_worker created during VHOST_SET_OWNER > + * that its virtqueues share. This allows userspace to create a vhost_worker > + * and map a virtqueue to it or map a virtqueue to an existing worker. > + * > + * If pid > 0 and it matches an existing vhost_worker thread it will be bound > + * to the vq. If pid is VHOST_VRING_NEW_WORKER, then a new worker will be > + * created and bound to the vq. > + * > + * This must be called after VHOST_SET_OWNER and before the vq is active. > + */ > +#define VHOST_SET_VRING_WORKER _IOWR(VHOST_VIRTIO, 0x15, struct vhost_vring_worker) > > /* The following ioctls use eventfd file descriptors to signal and poll > * for events. */ > diff --git a/include/uapi/linux/vhost_types.h b/include/uapi/linux/vhost_types.h > index f7f6a3a28977..af654e3cef0e 100644 > --- a/include/uapi/linux/vhost_types.h > +++ b/include/uapi/linux/vhost_types.h > @@ -47,6 +47,18 @@ struct vhost_vring_addr { > __u64 log_guest_addr; > }; > > +#define VHOST_VRING_NEW_WORKER -1Do we need VHOST_VRING_FREE_WORKER? And I wonder if using dedicated ioctls are better: VHOST_VRING_NEW/FREE_WORKER VHOST_VRING_ATTACH_WORKER etc. Thanks> + > +struct vhost_vring_worker { > + unsigned int index; > + /* > + * The pid of the vhost worker that the vq will be bound to. If > + * pid is VHOST_VRING_NEW_WORKER a new worker will be created and its > + * pid will be returned in pid. > + */ > + __kernel_pid_t pid; > +}; > + > /* no alignment requirement */ > struct vhost_iotlb_msg { > __u64 iova;
Michael S. Tsirkin
2021-Oct-26 13:09 UTC
[PATCH V3 11/11] vhost: allow userspace to create workers
On Tue, Oct 26, 2021 at 01:37:14PM +0800, Jason Wang wrote:> > ? 2021/10/22 ??1:19, Mike Christie ??: > > This patch allows userspace to create workers and bind them to vqs. You > > can have N workers per dev and also share N workers with M vqs. > > > > Signed-off-by: Mike Christie <michael.christie at oracle.com> > > > A question, who is the best one to determine the binding? Is it the VMM > (Qemu etc) or the management stack? If the latter, it looks to me it's > better to expose this via sysfs?I think it's a bit much to expect this from management.> > > --- > > drivers/vhost/vhost.c | 99 ++++++++++++++++++++++++++++---- > > drivers/vhost/vhost.h | 2 +- > > include/uapi/linux/vhost.h | 11 ++++ > > include/uapi/linux/vhost_types.h | 12 ++++ > > 4 files changed, 112 insertions(+), 12 deletions(-) > > > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c > > index 04f43a6445e1..c86e88d7f35c 100644 > > --- a/drivers/vhost/vhost.c > > +++ b/drivers/vhost/vhost.c > > @@ -493,7 +493,6 @@ void vhost_dev_init(struct vhost_dev *dev, > > dev->umem = NULL; > > dev->iotlb = NULL; > > dev->mm = NULL; > > - dev->worker = NULL; > > dev->iov_limit = iov_limit; > > dev->weight = weight; > > dev->byte_weight = byte_weight; > > @@ -576,20 +575,40 @@ static void vhost_worker_stop(struct vhost_worker *worker) > > wait_for_completion(worker->exit_done); > > } > > -static void vhost_worker_free(struct vhost_dev *dev) > > -{ > > - struct vhost_worker *worker = dev->worker; > > +static void vhost_worker_put(struct vhost_worker *worker) > > +{ > > if (!worker) > > return; > > - dev->worker = NULL; > > + if (!refcount_dec_and_test(&worker->refcount)) > > + return; > > + > > WARN_ON(!llist_empty(&worker->work_list)); > > vhost_worker_stop(worker); > > kfree(worker); > > } > > -static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev) > > +static void vhost_vq_clear_worker(struct vhost_virtqueue *vq) > > +{ > > + if (vq->worker) > > + vhost_worker_put(vq->worker); > > + vq->worker = NULL; > > +} > > + > > +static void vhost_workers_free(struct vhost_dev *dev) > > +{ > > + int i; > > + > > + if (!dev->use_worker) > > + return; > > + > > + for (i = 0; i < dev->nvqs; i++) > > + vhost_vq_clear_worker(dev->vqs[i]); > > +} > > + > > +static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev, > > + int init_vq_map_count) > > { > > struct vhost_worker *worker; > > struct task_struct *task; > > @@ -598,9 +617,9 @@ static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev) > > if (!worker) > > return NULL; > > - dev->worker = worker; > > worker->kcov_handle = kcov_common_handle(); > > init_llist_head(&worker->work_list); > > + refcount_set(&worker->refcount, init_vq_map_count); > > /* > > * vhost used to use the kthread API which ignores all signals by > > @@ -617,10 +636,58 @@ static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev) > > free_worker: > > kfree(worker); > > - dev->worker = NULL; > > return NULL; > > } > > +static struct vhost_worker *vhost_worker_find(struct vhost_dev *dev, pid_t pid) > > +{ > > + struct vhost_worker *worker = NULL; > > + int i; > > + > > + for (i = 0; i < dev->nvqs; i++) { > > + if (dev->vqs[i]->worker->task->pid != pid) > > + continue; > > + > > + worker = dev->vqs[i]->worker; > > + break; > > + } > > + > > + return worker; > > +} > > + > > +/* Caller must have device mutex */ > > +static int vhost_vq_setup_worker(struct vhost_virtqueue *vq, > > + struct vhost_vring_worker *info) > > +{ > > + struct vhost_dev *dev = vq->dev; > > + struct vhost_worker *worker; > > + > > + if (!dev->use_worker) > > + return -EINVAL; > > + > > + /* We don't support setting a worker on an active vq */ > > + if (vq->private_data) > > + return -EBUSY; > > > Is it valuable to allow the worker switching on active vq? > > > > + > > + if (info->pid == VHOST_VRING_NEW_WORKER) { > > + worker = vhost_worker_create(dev, 1); > > + if (!worker) > > + return -ENOMEM; > > + > > + info->pid = worker->task->pid; > > + } else { > > + worker = vhost_worker_find(dev, info->pid); > > + if (!worker) > > + return -ENODEV; > > + > > + refcount_inc(&worker->refcount); > > + } > > + > > + vhost_vq_clear_worker(vq); > > + vq->worker = worker; > > + return 0; > > +} > > + > > /* Caller should have device mutex */ > > long vhost_dev_set_owner(struct vhost_dev *dev) > > { > > @@ -636,7 +703,7 @@ long vhost_dev_set_owner(struct vhost_dev *dev) > > vhost_attach_mm(dev); > > if (dev->use_worker) { > > - worker = vhost_worker_create(dev); > > + worker = vhost_worker_create(dev, dev->nvqs); > > if (!worker) > > goto err_worker; > > @@ -650,7 +717,7 @@ long vhost_dev_set_owner(struct vhost_dev *dev) > > return 0; > > err_iovecs: > > - vhost_worker_free(dev); > > + vhost_workers_free(dev); > > err_worker: > > vhost_detach_mm(dev); > > err_mm: > > @@ -742,7 +809,7 @@ void vhost_dev_cleanup(struct vhost_dev *dev) > > dev->iotlb = NULL; > > vhost_clear_msg(dev); > > wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM); > > - vhost_worker_free(dev); > > + vhost_workers_free(dev); > > vhost_detach_mm(dev); > > } > > EXPORT_SYMBOL_GPL(vhost_dev_cleanup); > > @@ -1612,6 +1679,7 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg > > struct eventfd_ctx *ctx = NULL; > > u32 __user *idxp = argp; > > struct vhost_virtqueue *vq; > > + struct vhost_vring_worker w; > > struct vhost_vring_state s; > > struct vhost_vring_file f; > > u32 idx; > > @@ -1719,6 +1787,15 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg > > if (copy_to_user(argp, &s, sizeof(s))) > > r = -EFAULT; > > break; > > + case VHOST_SET_VRING_WORKER: > > + if (copy_from_user(&w, argp, sizeof(w))) { > > + r = -EFAULT; > > + break; > > + } > > + r = vhost_vq_setup_worker(vq, &w); > > + if (!r && copy_to_user(argp, &w, sizeof(w))) > > + r = -EFAULT; > > + break; > > default: > > r = -ENOIOCTLCMD; > > } > > diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h > > index 33c63b24187a..0911d1a9bd3b 100644 > > --- a/drivers/vhost/vhost.h > > +++ b/drivers/vhost/vhost.h > > @@ -35,6 +35,7 @@ struct vhost_worker { > > struct llist_head work_list; > > u64 kcov_handle; > > unsigned long flags; > > + refcount_t refcount; > > }; > > /* Poll a file (eventfd or socket) */ > > @@ -160,7 +161,6 @@ struct vhost_dev { > > struct vhost_virtqueue **vqs; > > int nvqs; > > struct eventfd_ctx *log_ctx; > > - struct vhost_worker *worker; > > struct vhost_iotlb *umem; > > struct vhost_iotlb *iotlb; > > spinlock_t iotlb_lock; > > diff --git a/include/uapi/linux/vhost.h b/include/uapi/linux/vhost.h > > index c998860d7bbc..e5c0669430e5 100644 > > --- a/include/uapi/linux/vhost.h > > +++ b/include/uapi/linux/vhost.h > > @@ -70,6 +70,17 @@ > > #define VHOST_VRING_BIG_ENDIAN 1 > > #define VHOST_SET_VRING_ENDIAN _IOW(VHOST_VIRTIO, 0x13, struct vhost_vring_state) > > #define VHOST_GET_VRING_ENDIAN _IOW(VHOST_VIRTIO, 0x14, struct vhost_vring_state) > > +/* By default, a device gets one vhost_worker created during VHOST_SET_OWNER > > + * that its virtqueues share. This allows userspace to create a vhost_worker > > + * and map a virtqueue to it or map a virtqueue to an existing worker. > > + * > > + * If pid > 0 and it matches an existing vhost_worker thread it will be bound > > + * to the vq. If pid is VHOST_VRING_NEW_WORKER, then a new worker will be > > + * created and bound to the vq. > > + * > > + * This must be called after VHOST_SET_OWNER and before the vq is active. > > + */ > > +#define VHOST_SET_VRING_WORKER _IOWR(VHOST_VIRTIO, 0x15, struct vhost_vring_worker) > > /* The following ioctls use eventfd file descriptors to signal and poll > > * for events. */ > > diff --git a/include/uapi/linux/vhost_types.h b/include/uapi/linux/vhost_types.h > > index f7f6a3a28977..af654e3cef0e 100644 > > --- a/include/uapi/linux/vhost_types.h > > +++ b/include/uapi/linux/vhost_types.h > > @@ -47,6 +47,18 @@ struct vhost_vring_addr { > > __u64 log_guest_addr; > > }; > > +#define VHOST_VRING_NEW_WORKER -1 > > > Do we need VHOST_VRING_FREE_WORKER? And I wonder if using dedicated ioctls > are better: > > VHOST_VRING_NEW/FREE_WORKER > VHOST_VRING_ATTACH_WORKER > > etc. > > Thanks > > > > + > > +struct vhost_vring_worker { > > + unsigned int index; > > + /* > > + * The pid of the vhost worker that the vq will be bound to. If > > + * pid is VHOST_VRING_NEW_WORKER a new worker will be created and its > > + * pid will be returned in pid. > > + */ > > + __kernel_pid_t pid; > > +}; > > + > > /* no alignment requirement */ > > struct vhost_iotlb_msg { > > __u64 iova;
Stefan Hajnoczi
2021-Oct-26 15:44 UTC
[PATCH V3 11/11] vhost: allow userspace to create workers
On Tue, Oct 26, 2021 at 01:37:14PM +0800, Jason Wang wrote:> > ? 2021/10/22 ??1:19, Mike Christie ??: > > This patch allows userspace to create workers and bind them to vqs. You > > can have N workers per dev and also share N workers with M vqs. > > > > Signed-off-by: Mike Christie <michael.christie at oracle.com> > > > A question, who is the best one to determine the binding? Is it the VMM > (Qemu etc) or the management stack? If the latter, it looks to me it's > better to expose this via sysfs?A few options that let the management stack control vhost worker CPU affinity: 1. The management tool opens the vhost device node, calls ioctl(VHOST_SET_VRING_WORKER), sets up CPU affinity, and then passes the fd to the VMM. In this case the VMM is still able to call the ioctl, which may be undesirable from an attack surface perspective. 2. The VMM calls ioctl(VHOST_SET_VRING_WORKER) itself and the management tool queries the vq:worker details from the VMM (e.g. a new QEMU QMP query-vhost-workers command similar to query-iothreads). The management tool can then control CPU affinity on the vhost worker threads. (This is how CPU affinity works in QEMU and libvirt today.) 3. The sysfs approach you suggested. Does sysfs export vq-0/, vq-1/, etc directories with a "worker" attribute? Do we need to define a point when the VMM has set up vqs and the management stack is able to query them? Vhost devices currently pre-allocate the maximum number of vqs and I'm not sure how to determine the number of vqs that will actually be used? One advantage of this is that access to the vq:worker mapping can be limited to the management stack and the VMM cannot access it. But it seems a little tricky because the vhost model today doesn't use sysfs or define a lifecycle where the management stack can configure devices. Stefan -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 488 bytes Desc: not available URL: <http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20211026/23ff1c32/attachment.sig>
michael.christie at oracle.com
2021-Oct-26 16:49 UTC
[PATCH V3 11/11] vhost: allow userspace to create workers
On 10/26/21 12:37 AM, Jason Wang wrote:> > ? 2021/10/22 ??1:19, Mike Christie ??: >> This patch allows userspace to create workers and bind them to vqs. You >> can have N workers per dev and also share N workers with M vqs. >> >> Signed-off-by: Mike Christie <michael.christie at oracle.com> > > > A question, who is the best one to determine the binding? Is it the VMM (Qemu etc) or the management stack? If the latter, it looks to me it's better to expose this via sysfs?I thought it would be where you have management app settings, then the management app talks to the qemu control interface like it does when it adds new devices on the fly. A problem with the management app doing it is to handle the RLIMIT_NPROC review comment, this patchset: https://lore.kernel.org/all/20211007214448.6282-1-michael.christie at oracle.com/ basically has the kernel do a clone() from the caller's context. So adding a worker is like doing the VHOST_SET_OWNER ioctl where it still has to be done from a process you can inherit values like the mm, cgroups, and now RLIMITs.>> diff --git a/include/uapi/linux/vhost_types.h b/include/uapi/linux/vhost_types.h >> index f7f6a3a28977..af654e3cef0e 100644 >> --- a/include/uapi/linux/vhost_types.h >> +++ b/include/uapi/linux/vhost_types.h >> @@ -47,6 +47,18 @@ struct vhost_vring_addr { >> ????? __u64 log_guest_addr; >> ? }; >> ? +#define VHOST_VRING_NEW_WORKER -1 > > > Do we need VHOST_VRING_FREE_WORKER? And I wonder if using dedicated ioctls are better: > > VHOST_VRING_NEW/FREE_WORKER > VHOST_VRING_ATTACH_WORKERWe didn't need a free worker, because the kernel handles it for userspace. I tried to make it easy for userspace because in some cases it may not be able to do syscalls like close on the device. For example if qemu crashes or for vhost-scsi we don't do an explicit close during VM shutdown. So we start off with the default worker thread that's used by all vqs like we do today. Userspace can then override it by creating a new worker. That also unbinds/ detaches the existing worker and does a put on the workers refcount. We also do a put on the worker when we stop using it during device shutdown/closure/release. When the worker's refcount goes to zero the kernel deletes it. I think separating the calls could be helpful though.