This series introduces a framework that makes it possible to implement
software-emulated vDPA devices in userspace. And to make it simple, the
emulated vDPA device's control path is handled in the kernel and only the
data path is implemented in the userspace.
Since the emuldated vDPA device's control path is handled in the kernel,
a message mechnism is introduced to make userspace be aware of the data
path related changes. Userspace can use read()/write() to receive/reply
the control messages.
In the data path, the core is mapping dma buffer into VDUSE daemon's
address space, which can be implemented in different ways depending on
the vdpa bus to which the vDPA device is attached.
In virtio-vdpa case, we implements a MMU-based on-chip IOMMU driver with
bounce-buffering mechanism to achieve that. And in vhost-vdpa case, the dma
buffer is reside in a userspace memory region which can be shared to the
VDUSE userspace processs via transferring the shmfd.
The details and our user case is shown below:
------------------------ ------------------------- ----------------------------------------------
| Container | | QEMU(VM) | | VDUSE daemon |
| --------- | | ------------------- | | ------------------------- ---------------- |
| |dev/vdx| | | |/dev/vhost-vdpa-x| | | | vDPA device emulation | | block driver | |
------------+----------- -----------+------------ -------------+----------------------+---------
| | | |
| | | |
------------+---------------------------+----------------------------+----------------------+---------
| | block device | | vhost device | | vduse driver | | TCP/IP | |
| -------+-------- --------+-------- -------+-------- -----+---- |
| | | | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio-blk driver | | vhost-vdpa driver | | vdpa device | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio bus | | | |
| --------+----+----------- | | | |
| | | | | |
| ----------+---------- | | | |
| | virtio-blk device | | | | |
| ----------+---------- | | | |
| | | | | |
| -----------+----------- | | | |
| | virtio-vdpa driver | | | | |
| -----------+----------- | | | |
| | | | vdpa bus | |
| -----------+----------------------+---------------------------+------------ | |
| ---+--- |
-----------------------------------------------------------------------------------------| NIC |------
---+---
|
---------+---------
| Remote Storages |
-------------------
We make use of it to implement a block device connecting to
our distributed storage, which can be used both in containers and
VMs. Thus, we can have an unified technology stack in this two cases.
To test it with null-blk:
$ qemu-storage-daemon \
--chardev socket,id=charmonitor,path=/tmp/qmp.sock,server,nowait \
--monitor chardev=charmonitor \
--blockdev driver=host_device,cache.direct=on,aio=native,filename=/dev/nullb0,node-name=disk0 \
--export type=vduse-blk,id=test,node-name=disk0,writable=on,name=vduse-null,num-queues=16,queue-size=128
The qemu-storage-daemon can be found at https://github.com/bytedance/qemu/tree/vduse
To make the userspace VDUSE processes such as qemu-storage-daemon able to
be run by an unprivileged user. We did some works on virtio driver to avoid
trusting device, including:
- validating the used length:
* https://lore.kernel.org/lkml/20210531135852.113-1-xieyongji@bytedance.com/
* https://lore.kernel.org/lkml/20210525125622.1203-1-xieyongji@bytedance.com/
- validating the device config:
* https://lore.kernel.org/lkml/20210615104810.151-1-xieyongji@bytedance.com/
- validating the device response:
* https://lore.kernel.org/lkml/20210615105218.214-1-xieyongji@bytedance.com/
Since I'm not sure if I missing something during auditing, especially on some
virtio device drivers that I'm not familiar with, we limit the supported device
type to virtio block device currently. The support for other device types can be
added after the security issue of corresponding device driver is clarified or
fixed in the future.
Future work:
- Improve performance
- Userspace library (find a way to reuse device emulation code in qemu/rust-vmm)
- Support more device types
V7 to V8:
- Rebased to newest kernel tree
- Rework VDUSE driver to handle the device's control path in kernel
- Limit the supported device type to virtio block device
- Export free_iova_fast()
- Remove the virtio-blk and virtio-scsi patches (will send them alone)
- Remove all module parameters
- Use the same MAJOR for both control device and VDUSE devices
- Avoid eventfd cleanup in vduse_dev_release()
V6 to V7:
- Export alloc_iova_fast()
- Add get_config_size() callback
- Add some patches to avoid trusting virtio devices
- Add limited device emulation
- Add some documents
- Use workqueue to inject config irq
- Add parameter on vq irq injecting
- Rename vduse_domain_get_mapping_page() to vduse_domain_get_coherent_page()
- Add WARN_ON() to catch message failure
- Add some padding/reserved fields to uAPI structure
- Fix some bugs
- Rebase to vhost.git
V5 to V6:
- Export receive_fd() instead of __receive_fd()
- Factor out the unmapping logic of pa and va separatedly
- Remove the logic of bounce page allocation in page fault handler
- Use PAGE_SIZE as IOVA allocation granule
- Add EPOLLOUT support
- Enable setting API version in userspace
- Fix some bugs
V4 to V5:
- Remove the patch for irq binding
- Use a single IOTLB for all types of mapping
- Factor out vhost_vdpa_pa_map()
- Add some sample codes in document
- Use receice_fd_user() to pass file descriptor
- Fix some bugs
V3 to V4:
- Rebase to vhost.git
- Split some patches
- Add some documents
- Use ioctl to inject interrupt rather than eventfd
- Enable config interrupt support
- Support binding irq to the specified cpu
- Add two module parameter to limit bounce/iova size
- Create char device rather than anon inode per vduse
- Reuse vhost IOTLB for iova domain
- Rework the message mechnism in control path
V2 to V3:
- Rework the MMU-based IOMMU driver
- Use the iova domain as iova allocator instead of genpool
- Support transferring vma->vm_file in vhost-vdpa
- Add SVA support in vhost-vdpa
- Remove the patches on bounce pages reclaim
V1 to V2:
- Add vhost-vdpa support
- Add some documents
- Based on the vdpa management tool
- Introduce a workqueue for irq injection
- Replace interval tree with array map to store the iova_map
Xie Yongji (10):
iova: Export alloc_iova_fast() and free_iova_fast();
file: Export receive_fd() to modules
eventfd: Increase the recursion depth of eventfd_signal()
vhost-iotlb: Add an opaque pointer for vhost IOTLB
vdpa: Add an opaque pointer for vdpa_config_ops.dma_map()
vdpa: factor out vhost_vdpa_pa_map() and vhost_vdpa_pa_unmap()
vdpa: Support transferring virtual addressing during DMA mapping
vduse: Implement an MMU-based IOMMU driver
vduse: Introduce VDUSE - vDPA Device in Userspace
Documentation: Add documentation for VDUSE
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 +++
drivers/iommu/iova.c | 2 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/ifcvf/ifcvf_main.c | 2 +-
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
drivers/vdpa/vdpa.c | 9 +-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 8 +-
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/iova_domain.c | 545 ++++++++
drivers/vdpa/vdpa_user/iova_domain.h | 73 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
drivers/vdpa/virtio_pci/vp_vdpa.c | 2 +-
drivers/vhost/iotlb.c | 20 +-
drivers/vhost/vdpa.c | 148 +-
fs/eventfd.c | 2 +-
fs/file.c | 6 +
include/linux/eventfd.h | 5 +-
include/linux/file.h | 7 +-
include/linux/vdpa.h | 21 +-
include/linux/vhost_iotlb.h | 3 +
include/uapi/linux/vduse.h | 143 ++
24 files changed, 2641 insertions(+), 50 deletions(-)
create mode 100644 Documentation/userspace-api/vduse.rst
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.c
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.h
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
--
2.11.0
Export alloc_iova_fast() and free_iova_fast() so that
some modules can use it to improve iova allocation efficiency.
Signed-off-by: Xie Yongji <redacted>
---
drivers/iommu/iova.c | 2 ++
1 file changed, 2 insertions(+)
Export receive_fd() so that some modules can use
it to pass file descriptor between processes without
missing any security stuffs.
Signed-off-by: Xie Yongji <redacted>
---
fs/file.c | 6 ++++++
include/linux/file.h | 7 +++----
2 files changed, 9 insertions(+), 4 deletions(-)
Increase the recursion depth of eventfd_signal() to 1. This
is the maximum recursion depth we have found so far, which
can be triggered with the following call chain:
kvm_io_bus_write [kvm]
--> ioeventfd_write [kvm]
--> eventfd_signal [eventfd]
--> vhost_poll_wakeup [vhost]
--> vduse_vdpa_kick_vq [vduse]
--> eventfd_signal [eventfd]
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
---
fs/eventfd.c | 2 +-
include/linux/eventfd.h | 5 ++++-
2 files changed, 5 insertions(+), 2 deletions(-)
Add an opaque pointer for vhost IOTLB. And introduce
vhost_iotlb_add_range_ctx() to accept it.
Suggested-by: Jason Wang <redacted>
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
---
drivers/vhost/iotlb.c | 20 ++++++++++++++++----
include/linux/vhost_iotlb.h | 3 +++
2 files changed, 19 insertions(+), 4 deletions(-)
The upcoming patch is going to support VA mapping/unmapping.
So let's factor out the logic of PA mapping/unmapping firstly
to make the code more readable.
Suggested-by: Jason Wang <redacted>
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
---
drivers/vhost/vdpa.c | 53 +++++++++++++++++++++++++++++++++-------------------
1 file changed, 34 insertions(+), 19 deletions(-)
This patch introduces an attribute for vDPA device to indicate
whether virtual address can be used. If vDPA device driver set
it, vhost-vdpa bus driver will not pin user page and transfer
userspace virtual address instead of physical address during
DMA mapping. And corresponding vma->vm_file and offset will be
also passed as an opaque pointer.
Suggested-by: Jason Wang <redacted>
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
---
drivers/vdpa/ifcvf/ifcvf_main.c | 2 +-
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
drivers/vdpa/vdpa.c | 9 +++-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 2 +-
drivers/vdpa/virtio_pci/vp_vdpa.c | 2 +-
drivers/vhost/vdpa.c | 99 ++++++++++++++++++++++++++++++++++-----
include/linux/vdpa.h | 19 ++++++--
7 files changed, 116 insertions(+), 19 deletions(-)
@@ -91,6 +93,10 @@ struct vdpa_device *__vdpa_alloc_device(struct device *parent,if(!!config->dma_map!=!!config->dma_unmap)gotoerr;+/* It should only work for the device that use on-chip IOMMU */+if(use_va&&!(config->dma_map||config->set_map))+gotoerr;+err=-ENOMEM;vdev=kzalloc(size,GFP_KERNEL);if(!vdev)
This implements an MMU-based IOMMU driver to support mapping
kernel dma buffer into userspace. The basic idea behind it is
treating MMU (VA->PA) as IOMMU (IOVA->PA). The driver will set
up MMU mapping instead of IOMMU mapping for the DMA transfer so
that the userspace process is able to use its virtual address to
access the dma buffer in kernel.
And to avoid security issue, a bounce-buffering mechanism is
introduced to prevent userspace accessing the original buffer
directly.
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
---
drivers/vdpa/vdpa_user/iova_domain.c | 545 +++++++++++++++++++++++++++++++++++
drivers/vdpa/vdpa_user/iova_domain.h | 73 +++++
2 files changed, 618 insertions(+)
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.c
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.h
This VDUSE driver enables implementing vDPA devices in userspace.
The vDPA device's control path is handled in kernel and the data
path is handled in userspace.
A message mechnism is used by VDUSE driver to forward some control
messages such as starting/stopping datapath to userspace. Userspace
can use read()/write() to receive/reply those control messages.
And some ioctls are introduced to help userspace to implement the
data path. VDUSE_IOTLB_GET_FD ioctl can be used to get the file
descriptors referring to vDPA device's iova regions. Then userspace
can use mmap() to access those iova regions. VDUSE_DEV_GET_FEATURES
and VDUSE_VQ_GET_INFO ioctls are used to get the negotiated features
and metadata of virtqueues. VDUSE_INJECT_VQ_IRQ and VDUSE_VQ_SETUP_KICKFD
ioctls can be used to inject interrupt and setup the kickfd for
virtqueues. VDUSE_DEV_UPDATE_CONFIG ioctl is used to update the
configuration space and inject a config interrupt.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
include/uapi/linux/vduse.h | 143 ++
6 files changed, 1613 insertions(+)
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
@@ -0,0 +1,1453 @@+// SPDX-License-Identifier: GPL-2.0-only+/*+*VDUSE:vDPADeviceinUserspace+*+*Copyright(C)2020-2021BytedanceInc.and/oritsaffiliates.Allrightsreserved.+*+*Author:XieYongji<xieyongji@bytedance.com>+*+*/++#include<linux/init.h>+#include<linux/module.h>+#include<linux/cdev.h>+#include<linux/device.h>+#include<linux/eventfd.h>+#include<linux/slab.h>+#include<linux/wait.h>+#include<linux/dma-map-ops.h>+#include<linux/poll.h>+#include<linux/file.h>+#include<linux/uio.h>+#include<linux/vdpa.h>+#include<linux/nospec.h>+#include<uapi/linux/vduse.h>+#include<uapi/linux/vdpa.h>+#include<uapi/linux/virtio_config.h>+#include<uapi/linux/virtio_ids.h>+#include<uapi/linux/virtio_blk.h>+#include<linux/mod_devicetable.h>++#include"iova_domain.h"++#define DRV_AUTHOR "Yongji Xie <xieyongji@bytedance.com>"+#define DRV_DESC "vDPA Device in Userspace"+#define DRV_LICENSE "GPL v2"++#define VDUSE_DEV_MAX (1U << MINORBITS)+#define VDUSE_MAX_BOUNCE_SIZE (64 * 1024 * 1024)+#define VDUSE_IOVA_SIZE (128 * 1024 * 1024)+#define VDUSE_REQUEST_TIMEOUT 30++structvduse_virtqueue{+u16index;+u32num;+u32avail_idx;+u64desc_addr;+u64driver_addr;+u64device_addr;+boolready;+boolkicked;+spinlock_tkick_lock;+spinlock_tirq_lock;+structeventfd_ctx*kickfd;+structvdpa_callbackcb;+structwork_structinject;+};++structvduse_dev;++structvduse_vdpa{+structvdpa_devicevdpa;+structvduse_dev*dev;+};++structvduse_dev{+structvduse_vdpa*vdev;+structdevice*dev;+structvduse_virtqueue*vqs;+structvduse_iova_domain*domain;+char*name;+structmutexlock;+spinlock_tmsg_lock;+u64msg_unique;+wait_queue_head_twaitq;+structlist_headsend_list;+structlist_headrecv_list;+structvdpa_callbackconfig_cb;+structwork_structinject;+spinlock_tirq_lock;+intminor;+boolconnected;+boolstarted;+u64api_version;+u64user_features;+u64features;+u32device_id;+u32vendor_id;+u32generation;+u32config_size;+void*config;+u8status;+u16vq_size_max;+u32vq_num;+u32vq_align;+};++structvduse_dev_msg{+structvduse_dev_requestreq;+structvduse_dev_responseresp;+structlist_headlist;+wait_queue_head_twaitq;+boolcompleted;+};++structvduse_control{+u64api_version;+};++staticDEFINE_MUTEX(vduse_lock);+staticDEFINE_IDR(vduse_idr);++staticdev_tvduse_major;+staticstructclass*vduse_class;+staticstructcdevvduse_ctrl_cdev;+staticstructcdevvduse_cdev;+staticstructworkqueue_struct*vduse_irq_wq;++staticu32allowed_device_id[]={+VIRTIO_ID_BLOCK,+};++staticinlinestructvduse_dev*vdpa_to_vduse(structvdpa_device*vdpa)+{+structvduse_vdpa*vdev=container_of(vdpa,structvduse_vdpa,vdpa);++returnvdev->dev;+}++staticinlinestructvduse_dev*dev_to_vduse(structdevice*dev)+{+structvdpa_device*vdpa=dev_to_vdpa(dev);++returnvdpa_to_vduse(vdpa);+}++staticstructvduse_dev_msg*vduse_find_msg(structlist_head*head,+uint32_trequest_id)+{+structvduse_dev_msg*msg;++list_for_each_entry(msg,head,list){+if(msg->req.request_id==request_id){+list_del(&msg->list);+returnmsg;+}+}++returnNULL;+}++staticstructvduse_dev_msg*vduse_dequeue_msg(structlist_head*head)+{+structvduse_dev_msg*msg=NULL;++if(!list_empty(head)){+msg=list_first_entry(head,structvduse_dev_msg,list);+list_del(&msg->list);+}++returnmsg;+}++staticvoidvduse_enqueue_msg(structlist_head*head,+structvduse_dev_msg*msg)+{+list_add_tail(&msg->list,head);+}++staticintvduse_dev_msg_send(structvduse_dev*dev,+structvduse_dev_msg*msg,boolno_reply)+{+init_waitqueue_head(&msg->waitq);+spin_lock(&dev->msg_lock);+msg->req.request_id=dev->msg_unique++;+vduse_enqueue_msg(&dev->send_list,msg);+wake_up(&dev->waitq);+spin_unlock(&dev->msg_lock);+if(no_reply)+return0;++wait_event_killable_timeout(msg->waitq,msg->completed,+VDUSE_REQUEST_TIMEOUT*HZ);+spin_lock(&dev->msg_lock);+if(!msg->completed){+list_del(&msg->list);+msg->resp.result=VDUSE_REQ_RESULT_FAILED;+}+spin_unlock(&dev->msg_lock);++return(msg->resp.result==VDUSE_REQ_RESULT_OK)?0:-EIO;+}++staticvoidvduse_dev_msg_cleanup(structvduse_dev*dev)+{+structvduse_dev_msg*msg;++spin_lock(&dev->msg_lock);+while((msg=vduse_dequeue_msg(&dev->send_list))){+if(msg->req.flags&VDUSE_REQ_FLAGS_NO_REPLY)+kfree(msg);+else+vduse_enqueue_msg(&dev->recv_list,msg);+}+while((msg=vduse_dequeue_msg(&dev->recv_list))){+msg->resp.result=VDUSE_REQ_RESULT_FAILED;+msg->completed=1;+wake_up(&msg->waitq);+}+spin_unlock(&dev->msg_lock);+}++staticvoidvduse_dev_start_dataplane(structvduse_dev*dev)+{+structvduse_dev_msg*msg=kzalloc(sizeof(*msg),+GFP_KERNEL|__GFP_NOFAIL);++msg->req.type=VDUSE_START_DATAPLANE;+msg->req.flags|=VDUSE_REQ_FLAGS_NO_REPLY;+vduse_dev_msg_send(dev,msg,true);+}++staticvoidvduse_dev_stop_dataplane(structvduse_dev*dev)+{+structvduse_dev_msg*msg=kzalloc(sizeof(*msg),+GFP_KERNEL|__GFP_NOFAIL);++msg->req.type=VDUSE_STOP_DATAPLANE;+msg->req.flags|=VDUSE_REQ_FLAGS_NO_REPLY;+vduse_dev_msg_send(dev,msg,true);+}++staticintvduse_dev_get_vq_state(structvduse_dev*dev,+structvduse_virtqueue*vq,+structvdpa_vq_state*state)+{+structvduse_dev_msgmsg={0};+intret;++msg.req.type=VDUSE_GET_VQ_STATE;+msg.req.vq_state.index=vq->index;++ret=vduse_dev_msg_send(dev,&msg,false);+if(ret)+returnret;++state->avail_index=msg.resp.vq_state.avail_idx;+return0;+}++staticintvduse_dev_update_iotlb(structvduse_dev*dev,+u64start,u64last)+{+structvduse_dev_msgmsg={0};++if(last<start)+return-EINVAL;++msg.req.type=VDUSE_UPDATE_IOTLB;+msg.req.iova.start=start;+msg.req.iova.last=last;++returnvduse_dev_msg_send(dev,&msg,false);+}++staticssize_tvduse_dev_read_iter(structkiocb*iocb,structiov_iter*to)+{+structfile*file=iocb->ki_filp;+structvduse_dev*dev=file->private_data;+structvduse_dev_msg*msg;+intsize=sizeof(structvduse_dev_request);+ssize_tret;++if(iov_iter_count(to)<size)+return-EINVAL;++spin_lock(&dev->msg_lock);+while(1){+msg=vduse_dequeue_msg(&dev->send_list);+if(msg)+break;++ret=-EAGAIN;+if(file->f_flags&O_NONBLOCK)+gotounlock;++spin_unlock(&dev->msg_lock);+ret=wait_event_interruptible_exclusive(dev->waitq,+!list_empty(&dev->send_list));+if(ret)+returnret;++spin_lock(&dev->msg_lock);+}+spin_unlock(&dev->msg_lock);+ret=copy_to_iter(&msg->req,size,to);+spin_lock(&dev->msg_lock);+if(ret!=size){+ret=-EFAULT;+vduse_enqueue_msg(&dev->send_list,msg);+gotounlock;+}+if(msg->req.flags&VDUSE_REQ_FLAGS_NO_REPLY)+kfree(msg);+else+vduse_enqueue_msg(&dev->recv_list,msg);+unlock:+spin_unlock(&dev->msg_lock);++returnret;+}++staticssize_tvduse_dev_write_iter(structkiocb*iocb,structiov_iter*from)+{+structfile*file=iocb->ki_filp;+structvduse_dev*dev=file->private_data;+structvduse_dev_responseresp;+structvduse_dev_msg*msg;+size_tret;++ret=copy_from_iter(&resp,sizeof(resp),from);+if(ret!=sizeof(resp))+return-EINVAL;++spin_lock(&dev->msg_lock);+msg=vduse_find_msg(&dev->recv_list,resp.request_id);+if(!msg){+ret=-ENOENT;+gotounlock;+}++memcpy(&msg->resp,&resp,sizeof(resp));+msg->completed=1;+wake_up(&msg->waitq);+unlock:+spin_unlock(&dev->msg_lock);++returnret;+}++static__poll_tvduse_dev_poll(structfile*file,poll_table*wait)+{+structvduse_dev*dev=file->private_data;+__poll_tmask=0;++poll_wait(file,&dev->waitq,wait);++if(!list_empty(&dev->send_list))+mask|=EPOLLIN|EPOLLRDNORM;+if(!list_empty(&dev->recv_list))+mask|=EPOLLOUT|EPOLLWRNORM;++returnmask;+}++staticvoidvduse_dev_reset(structvduse_dev*dev)+{+inti;+structvduse_iova_domain*domain=dev->domain;++/* The coherent mappings are handled in vduse_dev_free_coherent() */+if(domain->bounce_map)+vduse_domain_reset_bounce_map(domain);++dev->features=0;+dev->generation++;+spin_lock(&dev->irq_lock);+dev->config_cb.callback=NULL;+dev->config_cb.private=NULL;+spin_unlock(&dev->irq_lock);++for(i=0;i<dev->vq_num;i++){+structvduse_virtqueue*vq=&dev->vqs[i];++vq->ready=false;+vq->desc_addr=0;+vq->driver_addr=0;+vq->device_addr=0;+vq->avail_idx=0;+vq->num=0;++spin_lock(&vq->kick_lock);+vq->kicked=false;+if(vq->kickfd)+eventfd_ctx_put(vq->kickfd);+vq->kickfd=NULL;+spin_unlock(&vq->kick_lock);++spin_lock(&vq->irq_lock);+vq->cb.callback=NULL;+vq->cb.private=NULL;+spin_unlock(&vq->irq_lock);+}+}++staticintvduse_vdpa_set_vq_address(structvdpa_device*vdpa,u16idx,+u64desc_area,u64driver_area,+u64device_area)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++vq->desc_addr=desc_area;+vq->driver_addr=driver_area;+vq->device_addr=device_area;++return0;+}++staticvoidvduse_vdpa_kick_vq(structvdpa_device*vdpa,u16idx)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++spin_lock(&vq->kick_lock);+if(!vq->ready)+gotounlock;++if(vq->kickfd)+eventfd_signal(vq->kickfd,1);+else+vq->kicked=true;+unlock:+spin_unlock(&vq->kick_lock);+}++staticvoidvduse_vdpa_set_vq_cb(structvdpa_device*vdpa,u16idx,+structvdpa_callback*cb)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++spin_lock(&vq->irq_lock);+vq->cb.callback=cb->callback;+vq->cb.private=cb->private;+spin_unlock(&vq->irq_lock);+}++staticvoidvduse_vdpa_set_vq_num(structvdpa_device*vdpa,u16idx,u32num)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++vq->num=num;+}++staticvoidvduse_vdpa_set_vq_ready(structvdpa_device*vdpa,+u16idx,boolready)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++vq->ready=ready;+}++staticboolvduse_vdpa_get_vq_ready(structvdpa_device*vdpa,u16idx)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++returnvq->ready;+}++staticintvduse_vdpa_set_vq_state(structvdpa_device*vdpa,u16idx,+conststructvdpa_vq_state*state)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++vq->avail_idx=state->avail_index;+return0;+}++staticintvduse_vdpa_get_vq_state(structvdpa_device*vdpa,u16idx,+structvdpa_vq_state*state)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+structvduse_virtqueue*vq=&dev->vqs[idx];++returnvduse_dev_get_vq_state(dev,vq,state);+}++staticu32vduse_vdpa_get_vq_align(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->vq_align;+}++staticu64vduse_vdpa_get_features(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->user_features;+}++staticintvduse_vdpa_set_features(structvdpa_device*vdpa,u64features)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++dev->features=features;+return0;+}++staticvoidvduse_vdpa_set_config_cb(structvdpa_device*vdpa,+structvdpa_callback*cb)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++spin_lock(&dev->irq_lock);+dev->config_cb.callback=cb->callback;+dev->config_cb.private=cb->private;+spin_unlock(&dev->irq_lock);+}++staticu16vduse_vdpa_get_vq_num_max(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->vq_size_max;+}++staticu32vduse_vdpa_get_device_id(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->device_id;+}++staticu32vduse_vdpa_get_vendor_id(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->vendor_id;+}++staticu8vduse_vdpa_get_status(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->status;+}++staticvoidvduse_vdpa_set_status(structvdpa_device*vdpa,u8status)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+boolstarted=!!(status&VIRTIO_CONFIG_S_DRIVER_OK);++dev->status=status;++if(dev->started==started)+return;++dev->started=started;+if(dev->started){+vduse_dev_start_dataplane(dev);+}else{+vduse_dev_reset(dev);+vduse_dev_stop_dataplane(dev);+}+}++staticsize_tvduse_vdpa_get_config_size(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->config_size;+}++staticvoidvduse_vdpa_get_config(structvdpa_device*vdpa,unsignedintoffset,+void*buf,unsignedintlen)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++memcpy(buf,dev->config+offset,len);+}++staticvoidvduse_vdpa_set_config(structvdpa_device*vdpa,unsignedintoffset,+constvoid*buf,unsignedintlen)+{+/* Now we only support read-only configuration space */+}++staticu32vduse_vdpa_get_generation(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++returndev->generation;+}++staticintvduse_vdpa_set_map(structvdpa_device*vdpa,+structvhost_iotlb*iotlb)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);+intret;++ret=vduse_domain_set_map(dev->domain,iotlb);+if(ret)+returnret;++ret=vduse_dev_update_iotlb(dev,0ULL,ULLONG_MAX);+if(ret){+vduse_domain_clear_map(dev->domain,iotlb);+returnret;+}++return0;+}++staticvoidvduse_vdpa_free(structvdpa_device*vdpa)+{+structvduse_dev*dev=vdpa_to_vduse(vdpa);++dev->vdev=NULL;+}++staticconststructvdpa_config_opsvduse_vdpa_config_ops={+.set_vq_address=vduse_vdpa_set_vq_address,+.kick_vq=vduse_vdpa_kick_vq,+.set_vq_cb=vduse_vdpa_set_vq_cb,+.set_vq_num=vduse_vdpa_set_vq_num,+.set_vq_ready=vduse_vdpa_set_vq_ready,+.get_vq_ready=vduse_vdpa_get_vq_ready,+.set_vq_state=vduse_vdpa_set_vq_state,+.get_vq_state=vduse_vdpa_get_vq_state,+.get_vq_align=vduse_vdpa_get_vq_align,+.get_features=vduse_vdpa_get_features,+.set_features=vduse_vdpa_set_features,+.set_config_cb=vduse_vdpa_set_config_cb,+.get_vq_num_max=vduse_vdpa_get_vq_num_max,+.get_device_id=vduse_vdpa_get_device_id,+.get_vendor_id=vduse_vdpa_get_vendor_id,+.get_status=vduse_vdpa_get_status,+.set_status=vduse_vdpa_set_status,+.get_config_size=vduse_vdpa_get_config_size,+.get_config=vduse_vdpa_get_config,+.set_config=vduse_vdpa_set_config,+.get_generation=vduse_vdpa_get_generation,+.set_map=vduse_vdpa_set_map,+.free=vduse_vdpa_free,+};++staticdma_addr_tvduse_dev_map_page(structdevice*dev,structpage*page,+unsignedlongoffset,size_tsize,+enumdma_data_directiondir,+unsignedlongattrs)+{+structvduse_dev*vdev=dev_to_vduse(dev);+structvduse_iova_domain*domain=vdev->domain;++returnvduse_domain_map_page(domain,page,offset,size,dir,attrs);+}++staticvoidvduse_dev_unmap_page(structdevice*dev,dma_addr_tdma_addr,+size_tsize,enumdma_data_directiondir,+unsignedlongattrs)+{+structvduse_dev*vdev=dev_to_vduse(dev);+structvduse_iova_domain*domain=vdev->domain;++returnvduse_domain_unmap_page(domain,dma_addr,size,dir,attrs);+}++staticvoid*vduse_dev_alloc_coherent(structdevice*dev,size_tsize,+dma_addr_t*dma_addr,gfp_tflag,+unsignedlongattrs)+{+structvduse_dev*vdev=dev_to_vduse(dev);+structvduse_iova_domain*domain=vdev->domain;+unsignedlongiova;+void*addr;++*dma_addr=DMA_MAPPING_ERROR;+addr=vduse_domain_alloc_coherent(domain,size,+(dma_addr_t*)&iova,flag,attrs);+if(!addr)+returnNULL;++*dma_addr=(dma_addr_t)iova;++returnaddr;+}++staticvoidvduse_dev_free_coherent(structdevice*dev,size_tsize,+void*vaddr,dma_addr_tdma_addr,+unsignedlongattrs)+{+structvduse_dev*vdev=dev_to_vduse(dev);+structvduse_iova_domain*domain=vdev->domain;++vduse_domain_free_coherent(domain,size,vaddr,dma_addr,attrs);+}++staticsize_tvduse_dev_max_mapping_size(structdevice*dev)+{+structvduse_dev*vdev=dev_to_vduse(dev);+structvduse_iova_domain*domain=vdev->domain;++returndomain->bounce_size;+}++staticconststructdma_map_opsvduse_dev_dma_ops={+.map_page=vduse_dev_map_page,+.unmap_page=vduse_dev_unmap_page,+.alloc=vduse_dev_alloc_coherent,+.free=vduse_dev_free_coherent,+.max_mapping_size=vduse_dev_max_mapping_size,+};++staticunsignedintperm_to_file_flags(u8perm)+{+unsignedintflags=0;++switch(perm){+caseVDUSE_ACCESS_WO:+flags|=O_WRONLY;+break;+caseVDUSE_ACCESS_RO:+flags|=O_RDONLY;+break;+caseVDUSE_ACCESS_RW:+flags|=O_RDWR;+break;+default:+WARN(1,"invalidate vhost IOTLB permission\n");+break;+}++returnflags;+}++staticintvduse_kickfd_setup(structvduse_dev*dev,+structvduse_vq_eventfd*eventfd)+{+structeventfd_ctx*ctx=NULL;+structvduse_virtqueue*vq;+u32index;++if(eventfd->index>=dev->vq_num)+return-EINVAL;++index=array_index_nospec(eventfd->index,dev->vq_num);+vq=&dev->vqs[index];+if(eventfd->fd>=0){+ctx=eventfd_ctx_fdget(eventfd->fd);+if(IS_ERR(ctx))+returnPTR_ERR(ctx);+}elseif(eventfd->fd!=VDUSE_EVENTFD_DEASSIGN)+return0;++spin_lock(&vq->kick_lock);+if(vq->kickfd)+eventfd_ctx_put(vq->kickfd);+vq->kickfd=ctx;+if(vq->ready&&vq->kicked&&vq->kickfd){+eventfd_signal(vq->kickfd,1);+vq->kicked=false;+}+spin_unlock(&vq->kick_lock);++return0;+}++staticvoidvduse_dev_irq_inject(structwork_struct*work)+{+structvduse_dev*dev=container_of(work,structvduse_dev,inject);++spin_lock_irq(&dev->irq_lock);+if(dev->config_cb.callback)+dev->config_cb.callback(dev->config_cb.private);+spin_unlock_irq(&dev->irq_lock);+}++staticvoidvduse_vq_irq_inject(structwork_struct*work)+{+structvduse_virtqueue*vq=container_of(work,+structvduse_virtqueue,inject);++spin_lock_irq(&vq->irq_lock);+if(vq->ready&&vq->cb.callback)+vq->cb.callback(vq->cb.private);+spin_unlock_irq(&vq->irq_lock);+}++staticlongvduse_dev_ioctl(structfile*file,unsignedintcmd,+unsignedlongarg)+{+structvduse_dev*dev=file->private_data;+void__user*argp=(void__user*)arg;+intret;++switch(cmd){+caseVDUSE_IOTLB_GET_FD:{+structvduse_iotlb_entryentry;+structvhost_iotlb_map*map;+structvdpa_map_file*map_file;+structvduse_iova_domain*domain=dev->domain;+structfile*f=NULL;++ret=-EFAULT;+if(copy_from_user(&entry,argp,sizeof(entry)))+break;++ret=-EINVAL;+if(entry.start>entry.last)+break;++spin_lock(&domain->iotlb_lock);+map=vhost_iotlb_itree_first(domain->iotlb,+entry.start,entry.last);+if(map){+map_file=(structvdpa_map_file*)map->opaque;+f=get_file(map_file->file);+entry.offset=map_file->offset;+entry.start=map->start;+entry.last=map->last;+entry.perm=map->perm;+}+spin_unlock(&domain->iotlb_lock);+ret=-EINVAL;+if(!f)+break;++ret=-EFAULT;+if(copy_to_user(argp,&entry,sizeof(entry))){+fput(f);+break;+}+ret=receive_fd(f,perm_to_file_flags(entry.perm));+fput(f);+break;+}+caseVDUSE_DEV_GET_FEATURES:+ret=put_user(dev->features,(u64__user*)argp);+break;+caseVDUSE_DEV_UPDATE_CONFIG:{+structvduse_config_updateconfig;+unsignedlongsize=offsetof(structvduse_config_update,+buffer);++ret=-EFAULT;+if(copy_from_user(&config,argp,size))+break;++ret=-EINVAL;+if(config.length==0||+config.length>dev->config_size-config.offset)+break;++ret=-EFAULT;+if(copy_from_user(dev->config+config.offset,argp+size,+config.length))+break;++ret=0;+queue_work(vduse_irq_wq,&dev->inject);+break;+}+caseVDUSE_VQ_GET_INFO:{+structvduse_vq_infovq_info;+u32vq_index;++ret=-EFAULT;+if(copy_from_user(&vq_info,argp,sizeof(vq_info)))+break;++ret=-EINVAL;+if(vq_info.index>=dev->vq_num)+break;++vq_index=array_index_nospec(vq_info.index,dev->vq_num);+vq_info.desc_addr=dev->vqs[vq_index].desc_addr;+vq_info.driver_addr=dev->vqs[vq_index].driver_addr;+vq_info.device_addr=dev->vqs[vq_index].device_addr;+vq_info.num=dev->vqs[vq_index].num;+vq_info.avail_idx=dev->vqs[vq_index].avail_idx;+vq_info.ready=dev->vqs[vq_index].ready;++ret=-EFAULT;+if(copy_to_user(argp,&vq_info,sizeof(vq_info)))+break;++ret=0;+break;+}+caseVDUSE_VQ_SETUP_KICKFD:{+structvduse_vq_eventfdeventfd;++ret=-EFAULT;+if(copy_from_user(&eventfd,argp,sizeof(eventfd)))+break;++ret=vduse_kickfd_setup(dev,&eventfd);+break;+}+caseVDUSE_VQ_INJECT_IRQ:{+u32vq_index;++ret=-EFAULT;+if(get_user(vq_index,(u32__user*)argp))+break;++ret=-EINVAL;+if(vq_index>=dev->vq_num)+break;++ret=0;+vq_index=array_index_nospec(vq_index,dev->vq_num);+queue_work(vduse_irq_wq,&dev->vqs[vq_index].inject);+break;+}+default:+ret=-ENOIOCTLCMD;+break;+}++returnret;+}++staticintvduse_dev_release(structinode*inode,structfile*file)+{+structvduse_dev*dev=file->private_data;++spin_lock(&dev->msg_lock);+/* Make sure the inflight messages can processed after reconncection */+list_splice_init(&dev->recv_list,&dev->send_list);+spin_unlock(&dev->msg_lock);+dev->connected=false;++return0;+}++staticstructvduse_dev*vduse_dev_get_from_minor(intminor)+{+structvduse_dev*dev;++mutex_lock(&vduse_lock);+dev=idr_find(&vduse_idr,minor);+mutex_unlock(&vduse_lock);++returndev;+}++staticintvduse_dev_open(structinode*inode,structfile*file)+{+intret;+structvduse_dev*dev=vduse_dev_get_from_minor(iminor(inode));++if(!dev)+return-ENODEV;++ret=-EBUSY;+mutex_lock(&dev->lock);+if(dev->connected)+gotounlock;++ret=0;+dev->connected=true;+file->private_data=dev;+unlock:+mutex_unlock(&dev->lock);++returnret;+}++staticconststructfile_operationsvduse_dev_fops={+.owner=THIS_MODULE,+.open=vduse_dev_open,+.release=vduse_dev_release,+.read_iter=vduse_dev_read_iter,+.write_iter=vduse_dev_write_iter,+.poll=vduse_dev_poll,+.unlocked_ioctl=vduse_dev_ioctl,+.compat_ioctl=compat_ptr_ioctl,+.llseek=noop_llseek,+};++staticstructvduse_dev*vduse_dev_create(void)+{+structvduse_dev*dev=kzalloc(sizeof(*dev),GFP_KERNEL);++if(!dev)+returnNULL;++mutex_init(&dev->lock);+spin_lock_init(&dev->msg_lock);+INIT_LIST_HEAD(&dev->send_list);+INIT_LIST_HEAD(&dev->recv_list);+spin_lock_init(&dev->irq_lock);++INIT_WORK(&dev->inject,vduse_dev_irq_inject);+init_waitqueue_head(&dev->waitq);++returndev;+}++staticvoidvduse_dev_destroy(structvduse_dev*dev)+{+kfree(dev);+}++staticstructvduse_dev*vduse_find_dev(constchar*name)+{+structvduse_dev*dev;+intid;++idr_for_each_entry(&vduse_idr,dev,id)+if(!strcmp(dev->name,name))+returndev;++returnNULL;+}++staticintvduse_destroy_dev(char*name)+{+structvduse_dev*dev=vduse_find_dev(name);++if(!dev)+return-EINVAL;++mutex_lock(&dev->lock);+if(dev->vdev||dev->connected){+mutex_unlock(&dev->lock);+return-EBUSY;+}+dev->connected=true;+mutex_unlock(&dev->lock);++vduse_dev_msg_cleanup(dev);+device_destroy(vduse_class,MKDEV(MAJOR(vduse_major),dev->minor));+idr_remove(&vduse_idr,dev->minor);+kvfree(dev->config);+kfree(dev->vqs);+vduse_domain_destroy(dev->domain);+kfree(dev->name);+vduse_dev_destroy(dev);+module_put(THIS_MODULE);++return0;+}++staticbooldevice_is_allowed(u32device_id)+{+inti;++for(i=0;i<ARRAY_SIZE(allowed_device_id);i++)+if(allowed_device_id[i]==device_id)+returntrue;++returnfalse;+}++staticboolfeatures_is_valid(u64features)+{+if(!(features&(1ULL<<VIRTIO_F_ACCESS_PLATFORM)))+returnfalse;++/* Now we only support read-only configuration space */+if(features&(1ULL<<VIRTIO_BLK_F_CONFIG_WCE))+returnfalse;++returntrue;+}++staticboolvduse_validate_config(structvduse_dev_config*config)+{+if(config->bounce_size>VDUSE_MAX_BOUNCE_SIZE)+returnfalse;++if(config->vq_align>PAGE_SIZE)+returnfalse;++if(config->config_size>PAGE_SIZE)+returnfalse;++if(!device_is_allowed(config->device_id))+returnfalse;++if(!features_is_valid(config->features))+returnfalse;++returntrue;+}++staticintvduse_create_dev(structvduse_dev_config*config,+void*config_buf,u64api_version)+{+inti,ret;+structvduse_dev*dev;++ret=-EEXIST;+if(vduse_find_dev(config->name))+gotoerr;++ret=-ENOMEM;+dev=vduse_dev_create();+if(!dev)+gotoerr;++dev->api_version=api_version;+dev->user_features=config->features;+dev->device_id=config->device_id;+dev->vendor_id=config->vendor_id;+dev->name=kstrdup(config->name,GFP_KERNEL);+if(!dev->name)+gotoerr_str;++dev->domain=vduse_domain_create(VDUSE_IOVA_SIZE-1,+config->bounce_size);+if(!dev->domain)+gotoerr_domain;++dev->config=config_buf;+dev->config_size=config->config_size;+dev->vq_align=config->vq_align;+dev->vq_size_max=config->vq_size_max;+dev->vq_num=config->vq_num;+dev->vqs=kcalloc(dev->vq_num,sizeof(*dev->vqs),GFP_KERNEL);+if(!dev->vqs)+gotoerr_vqs;++for(i=0;i<dev->vq_num;i++){+dev->vqs[i].index=i;+INIT_WORK(&dev->vqs[i].inject,vduse_vq_irq_inject);+spin_lock_init(&dev->vqs[i].kick_lock);+spin_lock_init(&dev->vqs[i].irq_lock);+}++ret=idr_alloc(&vduse_idr,dev,1,VDUSE_DEV_MAX,GFP_KERNEL);+if(ret<0)+gotoerr_idr;++dev->minor=ret;+dev->dev=device_create(vduse_class,NULL,+MKDEV(MAJOR(vduse_major),dev->minor),+NULL,"%s",config->name);+if(IS_ERR(dev->dev)){+ret=PTR_ERR(dev->dev);+gotoerr_dev;+}+__module_get(THIS_MODULE);++return0;+err_dev:+idr_remove(&vduse_idr,dev->minor);+err_idr:+kfree(dev->vqs);+err_vqs:+vduse_domain_destroy(dev->domain);+err_domain:+kfree(dev->name);+err_str:+vduse_dev_destroy(dev);+err:+kvfree(config_buf);+returnret;+}++staticlongvduse_ioctl(structfile*file,unsignedintcmd,+unsignedlongarg)+{+intret;+void__user*argp=(void__user*)arg;+structvduse_control*control=file->private_data;++mutex_lock(&vduse_lock);+switch(cmd){+caseVDUSE_GET_API_VERSION:+ret=put_user(control->api_version,(u64__user*)argp);+break;+caseVDUSE_SET_API_VERSION:{+u64api_version;++ret=-EFAULT;+if(get_user(api_version,(u64__user*)argp))+break;++ret=-EINVAL;+if(api_version>VDUSE_API_VERSION)+break;++ret=0;+control->api_version=api_version;+break;+}+caseVDUSE_CREATE_DEV:{+structvduse_dev_configconfig;+unsignedlongsize=offsetof(structvduse_dev_config,config);+void*buf;++ret=-EFAULT;+if(copy_from_user(&config,argp,size))+break;++ret=-EINVAL;+if(vduse_validate_config(&config)==false)+break;++buf=vmemdup_user(argp+size,config.config_size);+if(IS_ERR(buf)){+ret=PTR_ERR(buf);+break;+}+ret=vduse_create_dev(&config,buf,control->api_version);+break;+}+caseVDUSE_DESTROY_DEV:{+charname[VDUSE_NAME_MAX];++ret=-EFAULT;+if(copy_from_user(name,argp,VDUSE_NAME_MAX))+break;++ret=vduse_destroy_dev(name);+break;+}+default:+ret=-EINVAL;+break;+}+mutex_unlock(&vduse_lock);++returnret;+}++staticintvduse_release(structinode*inode,structfile*file)+{+structvduse_control*control=file->private_data;++kfree(control);+return0;+}++staticintvduse_open(structinode*inode,structfile*file)+{+structvduse_control*control;++control=kmalloc(sizeof(structvduse_control),GFP_KERNEL);+if(!control)+return-ENOMEM;++control->api_version=VDUSE_API_VERSION;+file->private_data=control;++return0;+}++staticconststructfile_operationsvduse_ctrl_fops={+.owner=THIS_MODULE,+.open=vduse_open,+.release=vduse_release,+.unlocked_ioctl=vduse_ioctl,+.compat_ioctl=compat_ptr_ioctl,+.llseek=noop_llseek,+};++staticchar*vduse_devnode(structdevice*dev,umode_t*mode)+{+returnkasprintf(GFP_KERNEL,"vduse/%s",dev_name(dev));+}++staticvoidvduse_mgmtdev_release(structdevice*dev)+{+}++staticstructdevicevduse_mgmtdev={+.init_name="vduse",+.release=vduse_mgmtdev_release,+};++staticstructvdpa_mgmt_devmgmt_dev;++staticintvduse_dev_init_vdpa(structvduse_dev*dev,constchar*name)+{+structvduse_vdpa*vdev;+intret;++if(dev->vdev)+return-EEXIST;++vdev=vdpa_alloc_device(structvduse_vdpa,vdpa,dev->dev,+&vduse_vdpa_config_ops,name,true);+if(!vdev)+return-ENOMEM;++dev->vdev=vdev;+vdev->dev=dev;+vdev->vdpa.dev.dma_mask=&vdev->vdpa.dev.coherent_dma_mask;+ret=dma_set_mask_and_coherent(&vdev->vdpa.dev,DMA_BIT_MASK(64));+if(ret){+put_device(&vdev->vdpa.dev);+returnret;+}+set_dma_ops(&vdev->vdpa.dev,&vduse_dev_dma_ops);+vdev->vdpa.dma_dev=&vdev->vdpa.dev;+vdev->vdpa.mdev=&mgmt_dev;++return0;+}++staticintvdpa_dev_add(structvdpa_mgmt_dev*mdev,constchar*name)+{+structvduse_dev*dev;+intret;++mutex_lock(&vduse_lock);+dev=vduse_find_dev(name);+if(!dev){+mutex_unlock(&vduse_lock);+return-EINVAL;+}+ret=vduse_dev_init_vdpa(dev,name);+mutex_unlock(&vduse_lock);+if(ret)+returnret;++ret=_vdpa_register_device(&dev->vdev->vdpa,dev->vq_num);+if(ret){+put_device(&dev->vdev->vdpa.dev);+returnret;+}++return0;+}++staticvoidvdpa_dev_del(structvdpa_mgmt_dev*mdev,structvdpa_device*dev)+{+_vdpa_unregister_device(dev);+}++staticconststructvdpa_mgmtdev_opsvdpa_dev_mgmtdev_ops={+.dev_add=vdpa_dev_add,+.dev_del=vdpa_dev_del,+};++staticstructvirtio_device_idid_table[]={+{VIRTIO_ID_BLOCK,VIRTIO_DEV_ANY_ID},+{0},+};++staticstructvdpa_mgmt_devmgmt_dev={+.device=&vduse_mgmtdev,+.id_table=id_table,+.ops=&vdpa_dev_mgmtdev_ops,+};++staticintvduse_mgmtdev_init(void)+{+intret;++ret=device_register(&vduse_mgmtdev);+if(ret)+returnret;++ret=vdpa_mgmtdev_register(&mgmt_dev);+if(ret)+gotoerr;++return0;+err:+device_unregister(&vduse_mgmtdev);+returnret;+}++staticvoidvduse_mgmtdev_exit(void)+{+vdpa_mgmtdev_unregister(&mgmt_dev);+device_unregister(&vduse_mgmtdev);+}++staticintvduse_init(void)+{+intret;+structdevice*dev;++vduse_class=class_create(THIS_MODULE,"vduse");+if(IS_ERR(vduse_class))+returnPTR_ERR(vduse_class);++vduse_class->devnode=vduse_devnode;++ret=alloc_chrdev_region(&vduse_major,0,VDUSE_DEV_MAX,"vduse");+if(ret)+gotoerr_chardev_region;++/* /dev/vduse/control */+cdev_init(&vduse_ctrl_cdev,&vduse_ctrl_fops);+vduse_ctrl_cdev.owner=THIS_MODULE;+ret=cdev_add(&vduse_ctrl_cdev,vduse_major,1);+if(ret)+gotoerr_ctrl_cdev;++dev=device_create(vduse_class,NULL,vduse_major,NULL,"control");+if(IS_ERR(dev)){+ret=PTR_ERR(dev);+gotoerr_device;+}++/* /dev/vduse/$DEVICE */+cdev_init(&vduse_cdev,&vduse_dev_fops);+vduse_cdev.owner=THIS_MODULE;+ret=cdev_add(&vduse_cdev,MKDEV(MAJOR(vduse_major),1),+VDUSE_DEV_MAX-1);+if(ret)+gotoerr_cdev;++vduse_irq_wq=alloc_workqueue("vduse-irq",+WQ_HIGHPRI|WQ_SYSFS|WQ_UNBOUND,0);+if(!vduse_irq_wq)+gotoerr_wq;++ret=vduse_domain_init();+if(ret)+gotoerr_domain;++ret=vduse_mgmtdev_init();+if(ret)+gotoerr_mgmtdev;++return0;+err_mgmtdev:+vduse_domain_exit();+err_domain:+destroy_workqueue(vduse_irq_wq);+err_wq:+cdev_del(&vduse_cdev);+err_cdev:+device_destroy(vduse_class,vduse_major);+err_device:+cdev_del(&vduse_ctrl_cdev);+err_ctrl_cdev:+unregister_chrdev_region(vduse_major,VDUSE_DEV_MAX);+err_chardev_region:+class_destroy(vduse_class);+returnret;+}+module_init(vduse_init);++staticvoidvduse_exit(void)+{+vduse_mgmtdev_exit();+vduse_domain_exit();+destroy_workqueue(vduse_irq_wq);+cdev_del(&vduse_cdev);+device_destroy(vduse_class,vduse_major);+cdev_del(&vduse_ctrl_cdev);+unregister_chrdev_region(vduse_major,VDUSE_DEV_MAX);+class_destroy(vduse_class);+}+module_exit(vduse_exit);++MODULE_LICENSE(DRV_LICENSE);+MODULE_AUTHOR(DRV_AUTHOR);+MODULE_DESCRIPTION(DRV_DESC);
@@ -0,0 +1,143 @@+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */+#ifndef _UAPI_VDUSE_H_+#define _UAPI_VDUSE_H_++#include<linux/types.h>++#define VDUSE_API_VERSION 0++#define VDUSE_NAME_MAX 256++/* the control messages definition for read/write */++enumvduse_req_type{+/* Get the state for virtqueue from userspace */+VDUSE_GET_VQ_STATE,+/* Notify userspace to start the dataplane, no reply */+VDUSE_START_DATAPLANE,+/* Notify userspace to stop the dataplane, no reply */+VDUSE_STOP_DATAPLANE,+/* Notify userspace to update the memory mapping in device IOTLB */+VDUSE_UPDATE_IOTLB,+};++structvduse_vq_state{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+};++structvduse_iova_range{+__u64start;/* start of the IOVA range */+__u64last;/* end of the IOVA range */+};++structvduse_dev_request{+__u32type;/* request type */+__u32request_id;/* request id */+#define VDUSE_REQ_FLAGS_NO_REPLY (1 << 0) /* No need to reply */+__u32flags;/* request flags */+__u32reserved;/* for future use */+union{+structvduse_vq_statevq_state;/* virtqueue state */+structvduse_iova_rangeiova;/* iova range for updating */+__u32padding[16];/* padding */+};+};++structvduse_dev_response{+__u32request_id;/* corresponding request id */+#define VDUSE_REQ_RESULT_OK 0x00+#define VDUSE_REQ_RESULT_FAILED 0x01+__u32result;/* the result of request */+__u32reserved[2];/* for future use */+union{+structvduse_vq_statevq_state;/* virtqueue state */+__u32padding[16];/* padding */+};+};++/* ioctls */++structvduse_dev_config{+charname[VDUSE_NAME_MAX];/* vduse device name */+__u32vendor_id;/* virtio vendor id */+__u32device_id;/* virtio device id */+__u64features;/* device features */+__u64bounce_size;/* bounce buffer size for iommu */+__u16vq_size_max;/* the max size of virtqueue */+__u16padding;/* padding */+__u32vq_num;/* the number of virtqueues */+__u32vq_align;/* the allocation alignment of virtqueue's metadata */+__u32config_size;/* the size of the configuration space */+__u32reserved[15];/* for future use */+__u8config[0];/* the buffer of the configuration space */+};++structvduse_iotlb_entry{+__u64offset;/* the mmap offset on fd */+__u64start;/* start of the IOVA range */+__u64last;/* last of the IOVA range */+#define VDUSE_ACCESS_RO 0x1+#define VDUSE_ACCESS_WO 0x2+#define VDUSE_ACCESS_RW 0x3+__u8perm;/* access permission of this range */+};++structvduse_config_update{+__u32offset;/* offset from the beginning of configuration space */+__u32length;/* the length to write to configuration space */+__u8buffer[0];/* buffer used to write from */+};++structvduse_vq_info{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+__u64desc_addr;/* address of desc area */+__u64driver_addr;/* address of driver area */+__u64device_addr;/* address of device area */+__u32num;/* the size of virtqueue */+__u8ready;/* ready status of virtqueue */+};++structvduse_vq_eventfd{+__u32index;/* virtqueue index */+#define VDUSE_EVENTFD_DEASSIGN -1+intfd;/* eventfd, -1 means de-assigning the eventfd */+};++#define VDUSE_BASE 0x81++/* Get the version of VDUSE API. This is used for future extension */+#define VDUSE_GET_API_VERSION _IOR(VDUSE_BASE, 0x00, __u64)++/* Set the version of VDUSE API. */+#define VDUSE_SET_API_VERSION _IOW(VDUSE_BASE, 0x01, __u64)++/* Create a vduse device which is represented by a char device (/dev/vduse/<name>) */+#define VDUSE_CREATE_DEV _IOW(VDUSE_BASE, 0x02, struct vduse_dev_config)++/* Destroy a vduse device. Make sure there are no references to the char device */+#define VDUSE_DESTROY_DEV _IOW(VDUSE_BASE, 0x03, char[VDUSE_NAME_MAX])++/*+*Getafiledescriptorforthefirstoverlappediovaregion,+*-EINVALmeanstheiovaregiondoesn'texist.+*/+#define VDUSE_IOTLB_GET_FD _IOWR(VDUSE_BASE, 0x04, struct vduse_iotlb_entry)++/* Get the negotiated features */+#define VDUSE_DEV_GET_FEATURES _IOR(VDUSE_BASE, 0x05, __u64)++/* Update the configuration space */+#define VDUSE_DEV_UPDATE_CONFIG _IOW(VDUSE_BASE, 0x06, struct vduse_config_update)++/* Get the specified virtqueue's information */+#define VDUSE_VQ_GET_INFO _IOWR(VDUSE_BASE, 0x07, struct vduse_vq_info)++/* Setup an eventfd to receive kick for virtqueue */+#define VDUSE_VQ_SETUP_KICKFD _IOW(VDUSE_BASE, 0x08, struct vduse_vq_eventfd)++/* Inject an interrupt for specific virtqueue */+#define VDUSE_VQ_INJECT_IRQ _IOW(VDUSE_BASE, 0x09, __u32)++#endif /* _UAPI_VDUSE_H_ */
VDUSE (vDPA Device in Userspace) is a framework to support
implementing software-emulated vDPA devices in userspace. This
document is intended to clarify the VDUSE design and usage.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 ++++++++++++++++++++++++++++++++++
2 files changed, 223 insertions(+)
create mode 100644 Documentation/userspace-api/vduse.rst
@@ -0,0 +1,222 @@+==================================+VDUSE - "vDPA Device in Userspace"+==================================++vDPA (virtio data path acceleration) device is a device that uses a+datapath which complies with the virtio specifications with vendor+specific control path. vDPA devices can be both physically located on+the hardware or emulated by software. VDUSE is a framework that makes it+possible to implement software-emulated vDPA devices in userspace. And+to make it simple, the emulated vDPA device's control path is handled in+the kernel and only the data path is implemented in the userspace.++Note that only virtio block device is supported by VDUSE framework now,+which can reduce security risks when the userspace process that implements+the data path is run by an unprivileged user. The Support for other device+types can be added after the security issue is clarified or fixed in the future.++Start/Stop VDUSE devices+------------------------++VDUSE devices are started as follows:++1. Create a new VDUSE instance with ioctl(VDUSE_CREATE_DEV) on+ /dev/vduse/control.++2. Begin processing VDUSE messages from /dev/vduse/$NAME. The first+ messages will arrive while attaching the VDUSE instance to vDPA bus.++3. Send the VDPA_CMD_DEV_NEW netlink message to attach the VDUSE+ instance to vDPA bus.++VDUSE devices are stopped as follows:++1. Send the VDPA_CMD_DEV_DEL netlink message to detach the VDUSE+ instance from vDPA bus.++2. Close the file descriptor referring to /dev/vduse/$NAME++3. Destroy the VDUSE instance with ioctl(VDUSE_DESTROY_DEV) on+ /dev/vduse/control++The netlink messages metioned above can be sent via vdpa tool in iproute2+or use the below sample codes:++..code-block:: c++ static int netlink_add_vduse(const char *name, enum vdpa_command cmd)+ {+ struct nl_sock *nlsock;+ struct nl_msg *msg;+ int famid;++ nlsock = nl_socket_alloc();+ if (!nlsock)+ return -ENOMEM;++ if (genl_connect(nlsock))+ goto free_sock;++ famid = genl_ctrl_resolve(nlsock, VDPA_GENL_NAME);+ if (famid < 0)+ goto close_sock;++ msg = nlmsg_alloc();+ if (!msg)+ goto close_sock;++ if (!genlmsg_put(msg, NL_AUTO_PORT, NL_AUTO_SEQ, famid, 0, 0, cmd, 0))+ goto nla_put_failure;++ NLA_PUT_STRING(msg, VDPA_ATTR_DEV_NAME, name);+ if (cmd == VDPA_CMD_DEV_NEW)+ NLA_PUT_STRING(msg, VDPA_ATTR_MGMTDEV_DEV_NAME, "vduse");++ if (nl_send_sync(nlsock, msg))+ goto close_sock;++ nl_close(nlsock);+ nl_socket_free(nlsock);++ return 0;+ nla_put_failure:+ nlmsg_free(msg);+ close_sock:+ nl_close(nlsock);+ free_sock:+ nl_socket_free(nlsock);+ return -1;+ }++How VDUSE works+---------------++Since the emuldated vDPA device's control path is handled in the kernel,+a message-based communication protocol and few types of control messages+are introduced by VDUSE framework to make userspace be aware of the data+path related changes:++- VDUSE_GET_VQ_STATE: Get the state for virtqueue from userspace++- VDUSE_START_DATAPLANE: Notify userspace to start the dataplane++- VDUSE_STOP_DATAPLANE: Notify userspace to stop the dataplane++- VDUSE_UPDATE_IOTLB: Notify userspace to update the memory mapping in device IOTLB++Userspace needs to read()/write() on /dev/vduse/$NAME to receive/reply+those control messages from/to VDUSE kernel module as follows:++..code-block:: c++ static int vduse_message_handler(int dev_fd)+ {+ int len;+ struct vduse_dev_request req;+ struct vduse_dev_response resp;++ len = read(dev_fd, &req, sizeof(req));+ if (len != sizeof(req))+ return -1;++ resp.request_id = req.request_id;++ switch (req.type) {++ /* handle different types of message */++ }++ if (req.flags & VDUSE_REQ_FLAGS_NO_REPLY)+ return 0;++ len = write(dev_fd, &resp, sizeof(resp));+ if (len != sizeof(resp))+ return -1;++ return 0;+ }++After VDUSE_START_DATAPLANE messages is received, userspace should start the+dataplane processing with the help of some ioctls on /dev/vduse/$NAME:++- VDUSE_IOTLB_GET_FD: get the file descriptor to the first overlapped iova region.+ Userspace can access this iova region by passing fd and corresponding size, offset,+ perm to mmap(). For example:++..code-block:: c++ static int perm_to_prot(uint8_t perm)+ {+ int prot = 0;++ switch (perm) {+ case VDUSE_ACCESS_WO:+ prot |= PROT_WRITE;+ break;+ case VDUSE_ACCESS_RO:+ prot |= PROT_READ;+ break;+ case VDUSE_ACCESS_RW:+ prot |= PROT_READ | PROT_WRITE;+ break;+ }++ return prot;+ }++ static void *iova_to_va(int dev_fd, uint64_t iova, uint64_t *len)+ {+ int fd;+ void *addr;+ size_t size;+ struct vduse_iotlb_entry entry;++ entry.start = iova;+ entry.last = iova + 1;+ fd = ioctl(dev_fd, VDUSE_IOTLB_GET_FD, &entry);+ if (fd < 0)+ return NULL;++ size = entry.last - entry.start + 1;+ *len = entry.last - iova + 1;+ addr = mmap(0, size, perm_to_prot(entry.perm), MAP_SHARED,+ fd, entry.offset);+ close(fd);+ if (addr == MAP_FAILED)+ return NULL;++ /* do something to cache this iova region */++ return addr + iova - entry.start;+ }++- VDUSE_DEV_GET_FEATURES: Get the negotiated features++- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt++- VDUSE_VQ_GET_INFO: Get the specified virtqueue's metadata++- VDUSE_VQ_SETUP_KICKFD: set the kickfd for virtqueue, this eventfd is used+ by VDUSE kernel module to notify userspace to consume the vring.++- VDUSE_INJECT_VQ_IRQ: inject an interrupt for specific virtqueue++MMU-based IOMMU Driver+----------------------++VDUSE framework implements an MMU-based on-chip IOMMU driver to support+mapping the kernel DMA buffer into the userspace iova region dynamically.+This is mainly designed for virtio-vdpa case (kernel virtio drivers).++The basic idea behind this driver is treating MMU (VA->PA) as IOMMU (IOVA->PA).+The driver will set up MMU mapping instead of IOMMU mapping for the DMA transfer+so that the userspace process is able to use its virtual address to access+the DMA buffer in kernel.++And to avoid security issue, a bounce-buffering mechanism is introduced to+prevent userspace accessing the original buffer directly which may contain other+kernel data. During the mapping, unmapping, the driver will copy the data from+the original buffer to the bounce buffer and back, depending on the direction of+the transfer. And the bounce-buffer addresses will be mapped into the user address+space instead of the original one.
Increase the recursion depth of eventfd_signal() to 1. This
is the maximum recursion depth we have found so far, which
can be triggered with the following call chain:
kvm_io_bus_write [kvm]
--> ioeventfd_write [kvm]
--> eventfd_signal [eventfd]
--> vhost_poll_wakeup [vhost]
--> vduse_vdpa_kick_vq [vduse]
--> eventfd_signal [eventfd]
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
On Thu, Jun 17, 2021 at 4:34 PM He Zhe [off-list ref] wrote:
On 6/15/21 10:13 PM, Xie Yongji wrote:
quoted
Increase the recursion depth of eventfd_signal() to 1. This
is the maximum recursion depth we have found so far, which
can be triggered with the following call chain:
kvm_io_bus_write [kvm]
--> ioeventfd_write [kvm]
--> eventfd_signal [eventfd]
--> vhost_poll_wakeup [vhost]
--> vduse_vdpa_kick_vq [vduse]
--> eventfd_signal [eventfd]
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
count is just count. How deep is acceptable should be put
where eventfd_signal_count is called.
The return value of this function is boolean rather than integer.
Please see the comments in eventfd_signal():
"then it should check eventfd_signal_count() before calling this
function. If it returns true, the eventfd_signal() call should be
deferred to a safe context."
Thanks,
Yongji
On Thu, Jun 17, 2021 at 4:34 PM He Zhe [off-list ref] wrote:
quoted
On 6/15/21 10:13 PM, Xie Yongji wrote:
quoted
Increase the recursion depth of eventfd_signal() to 1. This
is the maximum recursion depth we have found so far, which
can be triggered with the following call chain:
kvm_io_bus_write [kvm]
--> ioeventfd_write [kvm]
--> eventfd_signal [eventfd]
--> vhost_poll_wakeup [vhost]
--> vduse_vdpa_kick_vq [vduse]
--> eventfd_signal [eventfd]
Signed-off-by: Xie Yongji <redacted>
Acked-by: Jason Wang <redacted>
OK, so it seems to be a fix for the RT system if my understanding is
correct? Any reason why it's not merged? I'm happy to rebase my series
on your patch if you'd like to repost it.
It works for both mainline and RT kernel. The folks just reproduced in their RT
environments.
This patch somehow hasn't got maintainer's reply, so not merged yet.
And OK, I'll resend the patch.
count is just count. How deep is acceptable should be put
where eventfd_signal_count is called.
The return value of this function is boolean rather than integer.
Please see the comments in eventfd_signal():
"then it should check eventfd_signal_count() before calling this
function. If it returns true, the eventfd_signal() call should be
deferred to a safe context."
OK. Now that the maintainer comments as such we can use it accordingly,
though I still got the feeling that the function name and the type of the return
value don't match.
Thanks,
Zhe
From: Jason Wang <hidden> Date: 2021-06-21 09:14:14
在 2021/6/15 下午10:13, Xie Yongji 写道:
quoted hunk
This VDUSE driver enables implementing vDPA devices in userspace.
The vDPA device's control path is handled in kernel and the data
path is handled in userspace.
A message mechnism is used by VDUSE driver to forward some control
messages such as starting/stopping datapath to userspace. Userspace
can use read()/write() to receive/reply those control messages.
And some ioctls are introduced to help userspace to implement the
data path. VDUSE_IOTLB_GET_FD ioctl can be used to get the file
descriptors referring to vDPA device's iova regions. Then userspace
can use mmap() to access those iova regions. VDUSE_DEV_GET_FEATURES
and VDUSE_VQ_GET_INFO ioctls are used to get the negotiated features
and metadata of virtqueues. VDUSE_INJECT_VQ_IRQ and VDUSE_VQ_SETUP_KICKFD
ioctls can be used to inject interrupt and setup the kickfd for
virtqueues. VDUSE_DEV_UPDATE_CONFIG ioctl is used to update the
configuration space and inject a config interrupt.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
include/uapi/linux/vduse.h | 143 ++
6 files changed, 1613 insertions(+)
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
It looks to me the only user for no_reply=true is the dataplane start. I
wonder no_reply is really needed consider we have switched to use
wait_event_killable_timeout().
In another way, no_reply is false for vq state synchronization and IOTLB
updating. I wonder if we can simply use no_reply = true for them.
I wonder if no_reply work for the case of vhost-vdpa. For virtio-vDPA,
we have bouncing buffers so it's harmless if usersapce dataplane keeps
performing read/write. For vhost-vDPA we don't have such stuffs.
@@ -0,0 +1,143 @@+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */+#ifndef _UAPI_VDUSE_H_+#define _UAPI_VDUSE_H_++#include<linux/types.h>++#define VDUSE_API_VERSION 0++#define VDUSE_NAME_MAX 256++/* the control messages definition for read/write */++enumvduse_req_type{+/* Get the state for virtqueue from userspace */+VDUSE_GET_VQ_STATE,+/* Notify userspace to start the dataplane, no reply */+VDUSE_START_DATAPLANE,+/* Notify userspace to stop the dataplane, no reply */+VDUSE_STOP_DATAPLANE,+/* Notify userspace to update the memory mapping in device IOTLB */+VDUSE_UPDATE_IOTLB,+};++structvduse_vq_state{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+};
This needs some tweaks to support packed virtqueue.
+
+struct vduse_iova_range {
+ __u64 start; /* start of the IOVA range */
+ __u64 last; /* end of the IOVA range */
+};
+
+struct vduse_dev_request {
+ __u32 type; /* request type */
+ __u32 request_id; /* request id */
+#define VDUSE_REQ_FLAGS_NO_REPLY (1 << 0) /* No need to reply */
+ __u32 flags; /* request flags */
+ __u32 reserved; /* for future use */
+ union {
+ struct vduse_vq_state vq_state; /* virtqueue state */
+ struct vduse_iova_range iova; /* iova range for updating */
+ __u32 padding[16]; /* padding */
+ };
+};
+
+struct vduse_dev_response {
+ __u32 request_id; /* corresponding request id */
+#define VDUSE_REQ_RESULT_OK 0x00
+#define VDUSE_REQ_RESULT_FAILED 0x01
+ __u32 result; /* the result of request */
+ __u32 reserved[2]; /* for future use */
+ union {
+ struct vduse_vq_state vq_state; /* virtqueue state */
+ __u32 padding[16]; /* padding */
+ };
+};
+
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
+ __u16 padding; /* padding */
+ __u32 vq_num; /* the number of virtqueues */
+ __u32 vq_align; /* the allocation alignment of virtqueue's metadata */
+ __u32 config_size; /* the size of the configuration space */
+ __u32 reserved[15]; /* for future use */
+ __u8 config[0]; /* the buffer of the configuration space */
+};
+
+struct vduse_iotlb_entry {
+ __u64 offset; /* the mmap offset on fd */
+ __u64 start; /* start of the IOVA range */
+ __u64 last; /* last of the IOVA range */
+#define VDUSE_ACCESS_RO 0x1
+#define VDUSE_ACCESS_WO 0x2
+#define VDUSE_ACCESS_RW 0x3
+ __u8 perm; /* access permission of this range */
+};
+
+struct vduse_config_update {
+ __u32 offset; /* offset from the beginning of configuration space */
+ __u32 length; /* the length to write to configuration space */
+ __u8 buffer[0]; /* buffer used to write from */
+};
+
+struct vduse_vq_info {
+ __u32 index; /* virtqueue index */
+ __u32 avail_idx; /* virtqueue state (last_avail_idx) */
+ __u64 desc_addr; /* address of desc area */
+ __u64 driver_addr; /* address of driver area */
+ __u64 device_addr; /* address of device area */
+ __u32 num; /* the size of virtqueue */
+ __u8 ready; /* ready status of virtqueue */
+};
+
+struct vduse_vq_eventfd {
+ __u32 index; /* virtqueue index */
+#define VDUSE_EVENTFD_DEASSIGN -1
+ int fd; /* eventfd, -1 means de-assigning the eventfd */
+};
+
+#define VDUSE_BASE 0x81
+
+/* Get the version of VDUSE API. This is used for future extension */
+#define VDUSE_GET_API_VERSION _IOR(VDUSE_BASE, 0x00, __u64)
+
+/* Set the version of VDUSE API. */
+#define VDUSE_SET_API_VERSION _IOW(VDUSE_BASE, 0x01, __u64)
+
+/* Create a vduse device which is represented by a char device (/dev/vduse/<name>) */
+#define VDUSE_CREATE_DEV _IOW(VDUSE_BASE, 0x02, struct vduse_dev_config)
+
+/* Destroy a vduse device. Make sure there are no references to the char device */
+#define VDUSE_DESTROY_DEV _IOW(VDUSE_BASE, 0x03, char[VDUSE_NAME_MAX])
+
+/*
+ * Get a file descriptor for the first overlapped iova region,
+ * -EINVAL means the iova region doesn't exist.
+ */
+#define VDUSE_IOTLB_GET_FD _IOWR(VDUSE_BASE, 0x04, struct vduse_iotlb_entry)
+
+/* Get the negotiated features */
+#define VDUSE_DEV_GET_FEATURES _IOR(VDUSE_BASE, 0x05, __u64)
+
+/* Update the configuration space */
+#define VDUSE_DEV_UPDATE_CONFIG _IOW(VDUSE_BASE, 0x06, struct vduse_config_update)
+
+/* Get the specified virtqueue's information */
+#define VDUSE_VQ_GET_INFO _IOWR(VDUSE_BASE, 0x07, struct vduse_vq_info)
+
+/* Setup an eventfd to receive kick for virtqueue */
+#define VDUSE_VQ_SETUP_KICKFD _IOW(VDUSE_BASE, 0x08, struct vduse_vq_eventfd)
+
+/* Inject an interrupt for specific virtqueue */
+#define VDUSE_VQ_INJECT_IRQ _IOW(VDUSE_BASE, 0x09, __u32)
+
+#endif /* _UAPI_VDUSE_H_ */
On Mon, Jun 21, 2021 at 5:14 PM Jason Wang [off-list ref] wrote:
在 2021/6/15 下午10:13, Xie Yongji 写道:
quoted
This VDUSE driver enables implementing vDPA devices in userspace.
The vDPA device's control path is handled in kernel and the data
path is handled in userspace.
A message mechnism is used by VDUSE driver to forward some control
messages such as starting/stopping datapath to userspace. Userspace
can use read()/write() to receive/reply those control messages.
And some ioctls are introduced to help userspace to implement the
data path. VDUSE_IOTLB_GET_FD ioctl can be used to get the file
descriptors referring to vDPA device's iova regions. Then userspace
can use mmap() to access those iova regions. VDUSE_DEV_GET_FEATURES
and VDUSE_VQ_GET_INFO ioctls are used to get the negotiated features
and metadata of virtqueues. VDUSE_INJECT_VQ_IRQ and VDUSE_VQ_SETUP_KICKFD
ioctls can be used to inject interrupt and setup the kickfd for
virtqueues. VDUSE_DEV_UPDATE_CONFIG ioctl is used to update the
configuration space and inject a config interrupt.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
include/uapi/linux/vduse.h | 143 ++
6 files changed, 1613 insertions(+)
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
It looks to me the only user for no_reply=true is the dataplane start. I
wonder no_reply is really needed consider we have switched to use
wait_event_killable_timeout().
Do we need to handle the error in this case if we remove the no_reply
flag. Print a warning message?
In another way, no_reply is false for vq state synchronization and IOTLB
updating. I wonder if we can simply use no_reply = true for them.
Looks like we can't, e.g. we need to get a reply from userspace for vq state.
I wonder if no_reply work for the case of vhost-vdpa. For virtio-vDPA,
we have bouncing buffers so it's harmless if usersapce dataplane keeps
performing read/write. For vhost-vDPA we don't have such stuffs.
OK. So it still needs to be synchronized here. If so, how to handle
the error? Looks like printing a warning message should be enough.
I wonder if it's better to separate config interrupt out of config
update or we need document this.
I have documented it in the docs. Looks like a config update should be
always followed by a config interrupt. I didn't find a case that uses
them separately.
quoted
+ break;
+ }
+ case VDUSE_VQ_GET_INFO: {
Do we need to limit this only when DRIVER_OK is set?
@@ -0,0 +1,143 @@+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */+#ifndef _UAPI_VDUSE_H_+#define _UAPI_VDUSE_H_++#include<linux/types.h>++#define VDUSE_API_VERSION 0++#define VDUSE_NAME_MAX 256++/* the control messages definition for read/write */++enumvduse_req_type{+/* Get the state for virtqueue from userspace */+VDUSE_GET_VQ_STATE,+/* Notify userspace to start the dataplane, no reply */+VDUSE_START_DATAPLANE,+/* Notify userspace to stop the dataplane, no reply */+VDUSE_STOP_DATAPLANE,+/* Notify userspace to update the memory mapping in device IOTLB */+VDUSE_UPDATE_IOTLB,+};++structvduse_vq_state{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+};
This needs some tweaks to support packed virtqueue.
From: Jason Wang <hidden> Date: 2021-06-22 05:07:04
在 2021/6/21 下午6:41, Yongji Xie 写道:
On Mon, Jun 21, 2021 at 5:14 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/15 下午10:13, Xie Yongji 写道:
quoted
This VDUSE driver enables implementing vDPA devices in userspace.
The vDPA device's control path is handled in kernel and the data
path is handled in userspace.
A message mechnism is used by VDUSE driver to forward some control
messages such as starting/stopping datapath to userspace. Userspace
can use read()/write() to receive/reply those control messages.
And some ioctls are introduced to help userspace to implement the
data path. VDUSE_IOTLB_GET_FD ioctl can be used to get the file
descriptors referring to vDPA device's iova regions. Then userspace
can use mmap() to access those iova regions. VDUSE_DEV_GET_FEATURES
and VDUSE_VQ_GET_INFO ioctls are used to get the negotiated features
and metadata of virtqueues. VDUSE_INJECT_VQ_IRQ and VDUSE_VQ_SETUP_KICKFD
ioctls can be used to inject interrupt and setup the kickfd for
virtqueues. VDUSE_DEV_UPDATE_CONFIG ioctl is used to update the
configuration space and inject a config interrupt.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
include/uapi/linux/vduse.h | 143 ++
6 files changed, 1613 insertions(+)
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
It looks to me the only user for no_reply=true is the dataplane start. I
wonder no_reply is really needed consider we have switched to use
wait_event_killable_timeout().
Do we need to handle the error in this case if we remove the no_reply
flag. Print a warning message?
See below.
quoted
In another way, no_reply is false for vq state synchronization and IOTLB
updating. I wonder if we can simply use no_reply = true for them.
Looks like we can't, e.g. we need to get a reply from userspace for vq state.
I wonder if no_reply work for the case of vhost-vdpa. For virtio-vDPA,
we have bouncing buffers so it's harmless if usersapce dataplane keeps
performing read/write. For vhost-vDPA we don't have such stuffs.
OK. So it still needs to be synchronized here. If so, how to handle
the error? Looks like printing a warning message should be enough.
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I wonder if it's better to separate config interrupt out of config
update or we need document this.
I have documented it in the docs. Looks like a config update should be
always followed by a config interrupt. I didn't find a case that uses
them separately.
The uAPI doesn't prevent us from the following scenario:
update_config(mac[0], ..);
update_config(max[1], ..);
So it looks to me it's better to separate the config interrupt from the
config updating.
quoted
quoted
+ break;
+ }
+ case VDUSE_VQ_GET_INFO: {
Do we need to limit this only when DRIVER_OK is set?
Any reason to add this limitation?
Otherwise the vq is not fully initialized, e.g the desc_addr might not
be correct.
@@ -0,0 +1,143 @@+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */+#ifndef _UAPI_VDUSE_H_+#define _UAPI_VDUSE_H_++#include<linux/types.h>++#define VDUSE_API_VERSION 0++#define VDUSE_NAME_MAX 256++/* the control messages definition for read/write */++enumvduse_req_type{+/* Get the state for virtqueue from userspace */+VDUSE_GET_VQ_STATE,+/* Notify userspace to start the dataplane, no reply */+VDUSE_START_DATAPLANE,+/* Notify userspace to stop the dataplane, no reply */+VDUSE_STOP_DATAPLANE,+/* Notify userspace to update the memory mapping in device IOTLB */+VDUSE_UPDATE_IOTLB,+};++structvduse_vq_state{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+};
This needs some tweaks to support packed virtqueue.
.On Tue, Jun 22, 2021 at 1:07 PM Jason Wang [off-list ref] wrote:
在 2021/6/21 下午6:41, Yongji Xie 写道:
quoted
On Mon, Jun 21, 2021 at 5:14 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/15 下午10:13, Xie Yongji 写道:
quoted
This VDUSE driver enables implementing vDPA devices in userspace.
The vDPA device's control path is handled in kernel and the data
path is handled in userspace.
A message mechnism is used by VDUSE driver to forward some control
messages such as starting/stopping datapath to userspace. Userspace
can use read()/write() to receive/reply those control messages.
And some ioctls are introduced to help userspace to implement the
data path. VDUSE_IOTLB_GET_FD ioctl can be used to get the file
descriptors referring to vDPA device's iova regions. Then userspace
can use mmap() to access those iova regions. VDUSE_DEV_GET_FEATURES
and VDUSE_VQ_GET_INFO ioctls are used to get the negotiated features
and metadata of virtqueues. VDUSE_INJECT_VQ_IRQ and VDUSE_VQ_SETUP_KICKFD
ioctls can be used to inject interrupt and setup the kickfd for
virtqueues. VDUSE_DEV_UPDATE_CONFIG ioctl is used to update the
configuration space and inject a config interrupt.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
include/uapi/linux/vduse.h | 143 ++
6 files changed, 1613 insertions(+)
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
It looks to me the only user for no_reply=true is the dataplane start. I
wonder no_reply is really needed consider we have switched to use
wait_event_killable_timeout().
Do we need to handle the error in this case if we remove the no_reply
flag. Print a warning message?
See below.
quoted
quoted
In another way, no_reply is false for vq state synchronization and IOTLB
updating. I wonder if we can simply use no_reply = true for them.
Looks like we can't, e.g. we need to get a reply from userspace for vq state.
I wonder if no_reply work for the case of vhost-vdpa. For virtio-vDPA,
we have bouncing buffers so it's harmless if usersapce dataplane keeps
performing read/write. For vhost-vDPA we don't have such stuffs.
OK. So it still needs to be synchronized here. If so, how to handle
the error? Looks like printing a warning message should be enough.
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
I wonder if it's better to separate config interrupt out of config
update or we need document this.
I have documented it in the docs. Looks like a config update should be
always followed by a config interrupt. I didn't find a case that uses
them separately.
The uAPI doesn't prevent us from the following scenario:
update_config(mac[0], ..);
update_config(max[1], ..);
So it looks to me it's better to separate the config interrupt from the
config updating.
Fine.
quoted
quoted
quoted
+ break;
+ }
+ case VDUSE_VQ_GET_INFO: {
Do we need to limit this only when DRIVER_OK is set?
Any reason to add this limitation?
Otherwise the vq is not fully initialized, e.g the desc_addr might not
be correct.
The vq_info->ready can be used to tell userspace whether the vq is
initialized or not.
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-22 07:50:15
在 2021/6/22 下午3:22, Yongji Xie 写道:
quoted
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
You're right. So let's make set_status() can fail first, then propagate
its failure via VHOST_VDPA_SET_STATUS.
I wonder if it's better to separate config interrupt out of config
update or we need document this.
I have documented it in the docs. Looks like a config update should be
always followed by a config interrupt. I didn't find a case that uses
them separately.
The uAPI doesn't prevent us from the following scenario:
update_config(mac[0], ..);
update_config(max[1], ..);
So it looks to me it's better to separate the config interrupt from the
config updating.
Fine.
quoted
quoted
quoted
quoted
+ break;
+ }
+ case VDUSE_VQ_GET_INFO: {
Do we need to limit this only when DRIVER_OK is set?
Any reason to add this limitation?
Otherwise the vq is not fully initialized, e.g the desc_addr might not
be correct.
The vq_info->ready can be used to tell userspace whether the vq is
initialized or not.
On Tue, Jun 22, 2021 at 3:50 PM Jason Wang [off-list ref] wrote:
在 2021/6/22 下午3:22, Yongji Xie 写道:
quoted
quoted
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
You're right. So let's make set_status() can fail first, then propagate
its failure via VHOST_VDPA_SET_STATUS.
OK. So we only need to propagate the failure in the vhost-vdpa case, right?
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-23 03:31:14
在 2021/6/22 下午4:14, Yongji Xie 写道:
On Tue, Jun 22, 2021 at 3:50 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/22 下午3:22, Yongji Xie 写道:
quoted
quoted
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
You're right. So let's make set_status() can fail first, then propagate
its failure via VHOST_VDPA_SET_STATUS.
OK. So we only need to propagate the failure in the vhost-vdpa case, right?
I think not, we need to deal with the reset for virtio as well:
E.g in register_virtio_devices(), we have:
/* We always start by resetting the device, in case a previous
* driver messed it up. This also tests that code path a
little. */
dev->config->reset(dev);
We probably need to make reset can fail and then fail the
register_virtio_device() as well.
Thanks
On Wed, Jun 23, 2021 at 11:31 AM Jason Wang [off-list ref] wrote:
在 2021/6/22 下午4:14, Yongji Xie 写道:
quoted
On Tue, Jun 22, 2021 at 3:50 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/22 下午3:22, Yongji Xie 写道:
quoted
quoted
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
You're right. So let's make set_status() can fail first, then propagate
its failure via VHOST_VDPA_SET_STATUS.
OK. So we only need to propagate the failure in the vhost-vdpa case, right?
I think not, we need to deal with the reset for virtio as well:
E.g in register_virtio_devices(), we have:
/* We always start by resetting the device, in case a previous
* driver messed it up. This also tests that code path a
little. */
dev->config->reset(dev);
We probably need to make reset can fail and then fail the
register_virtio_device() as well.
From: Jason Wang <hidden> Date: 2021-06-24 03:35:07
在 2021/6/23 下午1:50, Yongji Xie 写道:
On Wed, Jun 23, 2021 at 11:31 AM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/22 下午4:14, Yongji Xie 写道:
quoted
On Tue, Jun 22, 2021 at 3:50 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/22 下午3:22, Yongji Xie 写道:
quoted
quoted
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
You're right. So let's make set_status() can fail first, then propagate
its failure via VHOST_VDPA_SET_STATUS.
OK. So we only need to propagate the failure in the vhost-vdpa case, right?
I think not, we need to deal with the reset for virtio as well:
E.g in register_virtio_devices(), we have:
/* We always start by resetting the device, in case a previous
* driver messed it up. This also tests that code path a
little. */
dev->config->reset(dev);
We probably need to make reset can fail and then fail the
register_virtio_device() as well.
My understanding is DRIVER_OK is not something that needs to be validated:
"
DRIVER_OK (4)
Indicates that the driver is set up and ready to drive the device.
"
Since the spec doesn't require to re-read the and check if DRIVER_OK is
set in 3.1.1 Driver Requirements: Device Initialization.
It's more about "telling the device that driver is ready."
But we don have some status bit that requires the synchronization with
the device.
1) FEATURES_OK, spec requires to re-read the status bit to check whether
or it it was set by the device:
"
Re-read device status to ensure the FEATURES_OK bit is still set:
otherwise, the device does not support our subset of features and the
device is unusable.
"
This is useful for some device which can only support a subset of the
features. E.g a device that can only work for packed virtqueue. This
means the current design of set_features won't work, we need either:
1a) relay the set_features request to userspace
or
1b) introduce a mandated_device_features during device creation and
validate the driver features during the set_features(), and don't set
FEATURES_OK if they don't match.
2) Some transports (PCI) requires to re-read the status to ensure the
synchronization.
"
After writing 0 to device_status, the driver MUST wait for a read of
device_status to return 0 before reinitializing the device.
"
So we need to deal with both FEATURES_OK and reset, but probably not
DRIVER_OK.
Thanks
On Thu, Jun 24, 2021 at 11:35 AM Jason Wang [off-list ref] wrote:
在 2021/6/23 下午1:50, Yongji Xie 写道:
quoted
On Wed, Jun 23, 2021 at 11:31 AM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/22 下午4:14, Yongji Xie 写道:
quoted
On Tue, Jun 22, 2021 at 3:50 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/22 下午3:22, Yongji Xie 写道:
quoted
quoted
We need fix a way to propagate the error to the userspace.
E.g if we want to stop the deivce, we will delay the status reset until
we get respose from the userspace?
I didn't get how to delay the status reset. And should it be a DoS
that we want to fix if the userspace doesn't give a response forever?
You're right. So let's make set_status() can fail first, then propagate
its failure via VHOST_VDPA_SET_STATUS.
OK. So we only need to propagate the failure in the vhost-vdpa case, right?
I think not, we need to deal with the reset for virtio as well:
E.g in register_virtio_devices(), we have:
/* We always start by resetting the device, in case a previous
* driver messed it up. This also tests that code path a
little. */
dev->config->reset(dev);
We probably need to make reset can fail and then fail the
register_virtio_device() as well.
My understanding is DRIVER_OK is not something that needs to be validated:
"
DRIVER_OK (4)
Indicates that the driver is set up and ready to drive the device.
"
Since the spec doesn't require to re-read the and check if DRIVER_OK is
set in 3.1.1 Driver Requirements: Device Initialization.
It's more about "telling the device that driver is ready."
But we don have some status bit that requires the synchronization with
the device.
1) FEATURES_OK, spec requires to re-read the status bit to check whether
or it it was set by the device:
"
Re-read device status to ensure the FEATURES_OK bit is still set:
otherwise, the device does not support our subset of features and the
device is unusable.
"
This is useful for some device which can only support a subset of the
features. E.g a device that can only work for packed virtqueue. This
means the current design of set_features won't work, we need either:
1a) relay the set_features request to userspace
or
1b) introduce a mandated_device_features during device creation and
validate the driver features during the set_features(), and don't set
FEATURES_OK if they don't match.
2) Some transports (PCI) requires to re-read the status to ensure the
synchronization.
"
After writing 0 to device_status, the driver MUST wait for a read of
device_status to return 0 before reinitializing the device.
"
So we need to deal with both FEATURES_OK and reset, but probably not
DRIVER_OK.
OK, I see. Thanks for the explanation. One more question is how about
clearing the corresponding status bit in get_status() rather than
making set_status() fail. Since the spec recommends this way for
validation which is done in virtio_dev_remove() and
virtio_finalize_features().
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-24 08:14:16
在 2021/6/24 下午12:46, Yongji Xie 写道:
quoted
So we need to deal with both FEATURES_OK and reset, but probably not
DRIVER_OK.
OK, I see. Thanks for the explanation. One more question is how about
clearing the corresponding status bit in get_status() rather than
making set_status() fail. Since the spec recommends this way for
validation which is done in virtio_dev_remove() and
virtio_finalize_features().
Thanks,
Yongji
I think you can. Or it would be even better that we just don't set the
bit during set_status().
I just realize that in vdpa_reset() we had:
static inline void vdpa_reset(struct vdpa_device *vdev)
{
const struct vdpa_config_ops *ops = vdev->config;
vdev->features_valid = false;
ops->set_status(vdev, 0);
}
We probably need to add the synchronization here. E.g re-read with a
timeout.
Thanks
On Thu, Jun 24, 2021 at 4:14 PM Jason Wang [off-list ref] wrote:
在 2021/6/24 下午12:46, Yongji Xie 写道:
quoted
quoted
So we need to deal with both FEATURES_OK and reset, but probably not
DRIVER_OK.
OK, I see. Thanks for the explanation. One more question is how about
clearing the corresponding status bit in get_status() rather than
making set_status() fail. Since the spec recommends this way for
validation which is done in virtio_dev_remove() and
virtio_finalize_features().
Thanks,
Yongji
I think you can. Or it would be even better that we just don't set the
bit during set_status().
Yes, that's what I mean.
I just realize that in vdpa_reset() we had:
static inline void vdpa_reset(struct vdpa_device *vdev)
{
const struct vdpa_config_ops *ops = vdev->config;
vdev->features_valid = false;
ops->set_status(vdev, 0);
}
We probably need to add the synchronization here. E.g re-read with a
timeout.
Looks like the timeout is already in set_status(). Do we really need a
duplicated one here? And how to handle failure? Adding a return value
to virtio_config_ops->reset() and passing the error to the upper
layer?
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-25 03:09:04
在 2021/6/24 下午5:16, Yongji Xie 写道:
On Thu, Jun 24, 2021 at 4:14 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/24 下午12:46, Yongji Xie 写道:
quoted
quoted
So we need to deal with both FEATURES_OK and reset, but probably not
DRIVER_OK.
OK, I see. Thanks for the explanation. One more question is how about
clearing the corresponding status bit in get_status() rather than
making set_status() fail. Since the spec recommends this way for
validation which is done in virtio_dev_remove() and
virtio_finalize_features().
Thanks,
Yongji
I think you can. Or it would be even better that we just don't set the
bit during set_status().
Yes, that's what I mean.
quoted
I just realize that in vdpa_reset() we had:
static inline void vdpa_reset(struct vdpa_device *vdev)
{
const struct vdpa_config_ops *ops = vdev->config;
vdev->features_valid = false;
ops->set_status(vdev, 0);
}
We probably need to add the synchronization here. E.g re-read with a
timeout.
Looks like the timeout is already in set_status().
Do you mean the VDUSE's implementation?
Do we really need a
duplicated one here?
1) this is the timeout at the vDPA layer instead of the VDUSE layer.
2) it really depends on what's the meaning of the timeout for set_status
of VDUSE.
Do we want:
2a) for set_status(): relay the message to userspace and wait for the
userspace to quiescence the datapath
or
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
And how to handle failure? Adding a return value
to virtio_config_ops->reset() and passing the error to the upper
layer?
On Fri, Jun 25, 2021 at 11:09 AM Jason Wang [off-list ref] wrote:
在 2021/6/24 下午5:16, Yongji Xie 写道:
quoted
On Thu, Jun 24, 2021 at 4:14 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/24 下午12:46, Yongji Xie 写道:
quoted
quoted
So we need to deal with both FEATURES_OK and reset, but probably not
DRIVER_OK.
OK, I see. Thanks for the explanation. One more question is how about
clearing the corresponding status bit in get_status() rather than
making set_status() fail. Since the spec recommends this way for
validation which is done in virtio_dev_remove() and
virtio_finalize_features().
Thanks,
Yongji
I think you can. Or it would be even better that we just don't set the
bit during set_status().
Yes, that's what I mean.
quoted
I just realize that in vdpa_reset() we had:
static inline void vdpa_reset(struct vdpa_device *vdev)
{
const struct vdpa_config_ops *ops = vdev->config;
vdev->features_valid = false;
ops->set_status(vdev, 0);
}
We probably need to add the synchronization here. E.g re-read with a
timeout.
Looks like the timeout is already in set_status().
Do you mean the VDUSE's implementation?
Yes.
quoted
Do we really need a
duplicated one here?
1) this is the timeout at the vDPA layer instead of the VDUSE layer.
OK, I get it.
2) it really depends on what's the meaning of the timeout for set_status
of VDUSE.
Do we want:
2a) for set_status(): relay the message to userspace and wait for the
userspace to quiescence the datapath
or
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
Looks good to me. And I think we can use the reply of the message to
update the status instead of introducing a new command.
quoted
And how to handle failure? Adding a return value
to virtio_config_ops->reset() and passing the error to the upper
layer?
From: Liu Xiaodong <hidden> Date: 2021-06-28 02:55:41
On Tue, Jun 15, 2021 at 10:13:21PM +0800, Xie Yongji wrote:
This series introduces a framework that makes it possible to implement
software-emulated vDPA devices in userspace. And to make it simple, the
emulated vDPA device's control path is handled in the kernel and only the
data path is implemented in the userspace.
Since the emuldated vDPA device's control path is handled in the kernel,
a message mechnism is introduced to make userspace be aware of the data
path related changes. Userspace can use read()/write() to receive/reply
the control messages.
In the data path, the core is mapping dma buffer into VDUSE daemon's
address space, which can be implemented in different ways depending on
the vdpa bus to which the vDPA device is attached.
In virtio-vdpa case, we implements a MMU-based on-chip IOMMU driver with
bounce-buffering mechanism to achieve that. And in vhost-vdpa case, the dma
buffer is reside in a userspace memory region which can be shared to the
VDUSE userspace processs via transferring the shmfd.
The details and our user case is shown below:
------------------------ ------------------------- ----------------------------------------------
| Container | | QEMU(VM) | | VDUSE daemon |
| --------- | | ------------------- | | ------------------------- ---------------- |
| |dev/vdx| | | |/dev/vhost-vdpa-x| | | | vDPA device emulation | | block driver | |
------------+----------- -----------+------------ -------------+----------------------+---------
| | | |
| | | |
------------+---------------------------+----------------------------+----------------------+---------
| | block device | | vhost device | | vduse driver | | TCP/IP | |
| -------+-------- --------+-------- -------+-------- -----+---- |
| | | | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio-blk driver | | vhost-vdpa driver | | vdpa device | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio bus | | | |
| --------+----+----------- | | | |
| | | | | |
| ----------+---------- | | | |
| | virtio-blk device | | | | |
| ----------+---------- | | | |
| | | | | |
| -----------+----------- | | | |
| | virtio-vdpa driver | | | | |
| -----------+----------- | | | |
| | | | vdpa bus | |
| -----------+----------------------+---------------------------+------------ | |
| ---+--- |
-----------------------------------------------------------------------------------------| NIC |------
---+---
|
---------+---------
| Remote Storages |
-------------------
We make use of it to implement a block device connecting to
our distributed storage, which can be used both in containers and
VMs. Thus, we can have an unified technology stack in this two cases.
To test it with null-blk:
$ qemu-storage-daemon \
--chardev socket,id=charmonitor,path=/tmp/qmp.sock,server,nowait \
--monitor chardev=charmonitor \
--blockdev driver=host_device,cache.direct=on,aio=native,filename=/dev/nullb0,node-name=disk0 \
--export type=vduse-blk,id=test,node-name=disk0,writable=on,name=vduse-null,num-queues=16,queue-size=128
The qemu-storage-daemon can be found at https://github.com/bytedance/qemu/tree/vduse
To make the userspace VDUSE processes such as qemu-storage-daemon able to
be run by an unprivileged user. We did some works on virtio driver to avoid
trusting device, including:
- validating the used length:
* https://lore.kernel.org/lkml/20210531135852.113-1-xieyongji@bytedance.com/
* https://lore.kernel.org/lkml/20210525125622.1203-1-xieyongji@bytedance.com/
- validating the device config:
* https://lore.kernel.org/lkml/20210615104810.151-1-xieyongji@bytedance.com/
- validating the device response:
* https://lore.kernel.org/lkml/20210615105218.214-1-xieyongji@bytedance.com/
Since I'm not sure if I missing something during auditing, especially on some
virtio device drivers that I'm not familiar with, we limit the supported device
type to virtio block device currently. The support for other device types can be
added after the security issue of corresponding device driver is clarified or
fixed in the future.
Future work:
- Improve performance
- Userspace library (find a way to reuse device emulation code in qemu/rust-vmm)
- Support more device types
V7 to V8:
- Rebased to newest kernel tree
- Rework VDUSE driver to handle the device's control path in kernel
- Limit the supported device type to virtio block device
- Export free_iova_fast()
- Remove the virtio-blk and virtio-scsi patches (will send them alone)
- Remove all module parameters
- Use the same MAJOR for both control device and VDUSE devices
- Avoid eventfd cleanup in vduse_dev_release()
V6 to V7:
- Export alloc_iova_fast()
- Add get_config_size() callback
- Add some patches to avoid trusting virtio devices
- Add limited device emulation
- Add some documents
- Use workqueue to inject config irq
- Add parameter on vq irq injecting
- Rename vduse_domain_get_mapping_page() to vduse_domain_get_coherent_page()
- Add WARN_ON() to catch message failure
- Add some padding/reserved fields to uAPI structure
- Fix some bugs
- Rebase to vhost.git
V5 to V6:
- Export receive_fd() instead of __receive_fd()
- Factor out the unmapping logic of pa and va separatedly
- Remove the logic of bounce page allocation in page fault handler
- Use PAGE_SIZE as IOVA allocation granule
- Add EPOLLOUT support
- Enable setting API version in userspace
- Fix some bugs
V4 to V5:
- Remove the patch for irq binding
- Use a single IOTLB for all types of mapping
- Factor out vhost_vdpa_pa_map()
- Add some sample codes in document
- Use receice_fd_user() to pass file descriptor
- Fix some bugs
V3 to V4:
- Rebase to vhost.git
- Split some patches
- Add some documents
- Use ioctl to inject interrupt rather than eventfd
- Enable config interrupt support
- Support binding irq to the specified cpu
- Add two module parameter to limit bounce/iova size
- Create char device rather than anon inode per vduse
- Reuse vhost IOTLB for iova domain
- Rework the message mechnism in control path
V2 to V3:
- Rework the MMU-based IOMMU driver
- Use the iova domain as iova allocator instead of genpool
- Support transferring vma->vm_file in vhost-vdpa
- Add SVA support in vhost-vdpa
- Remove the patches on bounce pages reclaim
V1 to V2:
- Add vhost-vdpa support
- Add some documents
- Based on the vdpa management tool
- Introduce a workqueue for irq injection
- Replace interval tree with array map to store the iova_map
Xie Yongji (10):
iova: Export alloc_iova_fast() and free_iova_fast();
file: Export receive_fd() to modules
eventfd: Increase the recursion depth of eventfd_signal()
vhost-iotlb: Add an opaque pointer for vhost IOTLB
vdpa: Add an opaque pointer for vdpa_config_ops.dma_map()
vdpa: factor out vhost_vdpa_pa_map() and vhost_vdpa_pa_unmap()
vdpa: Support transferring virtual addressing during DMA mapping
vduse: Implement an MMU-based IOMMU driver
vduse: Introduce VDUSE - vDPA Device in Userspace
Documentation: Add documentation for VDUSE
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 +++
drivers/iommu/iova.c | 2 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/ifcvf/ifcvf_main.c | 2 +-
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
drivers/vdpa/vdpa.c | 9 +-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 8 +-
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/iova_domain.c | 545 ++++++++
drivers/vdpa/vdpa_user/iova_domain.h | 73 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
drivers/vdpa/virtio_pci/vp_vdpa.c | 2 +-
drivers/vhost/iotlb.c | 20 +-
drivers/vhost/vdpa.c | 148 +-
fs/eventfd.c | 2 +-
fs/file.c | 6 +
include/linux/eventfd.h | 5 +-
include/linux/file.h | 7 +-
include/linux/vdpa.h | 21 +-
include/linux/vhost_iotlb.h | 3 +
include/uapi/linux/vduse.h | 143 ++
24 files changed, 2641 insertions(+), 50 deletions(-)
create mode 100644 Documentation/userspace-api/vduse.rst
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.c
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.h
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
--
2.11.0
Hi, Yongji
Great work! your method is really wise that implements a software IOMMU
so that data path gets processed by userspace application efficiently.
Sorry, I've just realized your work and patches.
I was working on a similar thing aiming to get vhost-user-blk device
from SPDK vhost-target to be exported as local host kernel block device.
It's diagram is like this:
-----------------------------
------------------------ | ----------------- | ---------------------------------------
| <RunC Container> | <<<<<<<<| Shared-Memory |>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> |
| --------- | v | ----------------- | | v |
| |dev/vdx| | v | <virtio-local-agent> | | <Vhost-user Target> v |
------------+----------- v | ------------------------ | | --------------------------v------ |
| v | |/dev/virtio-local-ctrl| | | | unix socket | |block driver | |
| v ------------+---------------- --------+--------------------v---------
| v | | v
------------+----------------v--------------+----------------------------+--------------------v--------|
| | block device | v | Misc device | | v |
| -------+-------- v --------+------- | v |
| | v | | v |
| ----------+---------- v | | v |
| | virtio-blk driver | v | | v |
| ----------+---------- v | | v |
| | virtio bus v | | v |
| --------+---+------- v | | v |
| | v | | v |
| | v | | v |
| ----------+---------- v ---------+----------- | v |
| | virtio-blk device |--<----| virtio-local driver |----------------< v |
| ----------+---------- ----------+----------- v |
| ---------+--------|
-------------------------------------------------------------------------------------| RNIC |--| PCIe |-
----+--- | NVMe |
| --------
---------+---------
| Remote Storages |
-------------------
I just draft out an initial proof version. When seeing your RFC mail,
I'm thinking that SPDK target may depends on your work, so I could
directly drop mine.
But after a glance of the RFC patches, seems it is not so easy or
efficient to get vduse leveraged by SPDK.
(Please correct me, if I get wrong understanding on vduse. :) )
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer,
then SPDK could also be a potential user of vduse.
It would be better if SPDK vhost-target could leverage the datapath of
vduse directly and efficiently. Even the control path is vdpa based,
we may work out one daemon as agent to bridge SPDK vhost-target with vduse.
Then users who already deployed SPDK vhost-target, can smoothly run
some agent daemon without code modification on SPDK vhost-target itself.
(It is only better-to-have for SPDK vhost-target app, not mandatory for SPDK) :)
At least, some small barrier is there that blocked a vhost-target use vduse
datapath efficiently:
- Current IO completion irq of vduse is IOCTL based. If add one option
to get it eventfd based, then vhost-target can directly notify IO
completion via negotiated eventfd.
Thanks
From Xiaodong
From: Jason Wang <hidden> Date: 2021-06-28 04:35:39
在 2021/6/28 下午6:33, Liu Xiaodong 写道:
On Tue, Jun 15, 2021 at 10:13:21PM +0800, Xie Yongji wrote:
quoted
This series introduces a framework that makes it possible to implement
software-emulated vDPA devices in userspace. And to make it simple, the
emulated vDPA device's control path is handled in the kernel and only the
data path is implemented in the userspace.
Since the emuldated vDPA device's control path is handled in the kernel,
a message mechnism is introduced to make userspace be aware of the data
path related changes. Userspace can use read()/write() to receive/reply
the control messages.
In the data path, the core is mapping dma buffer into VDUSE daemon's
address space, which can be implemented in different ways depending on
the vdpa bus to which the vDPA device is attached.
In virtio-vdpa case, we implements a MMU-based on-chip IOMMU driver with
bounce-buffering mechanism to achieve that. And in vhost-vdpa case, the dma
buffer is reside in a userspace memory region which can be shared to the
VDUSE userspace processs via transferring the shmfd.
The details and our user case is shown below:
------------------------ ------------------------- ----------------------------------------------
| Container | | QEMU(VM) | | VDUSE daemon |
| --------- | | ------------------- | | ------------------------- ---------------- |
| |dev/vdx| | | |/dev/vhost-vdpa-x| | | | vDPA device emulation | | block driver | |
------------+----------- -----------+------------ -------------+----------------------+---------
| | | |
| | | |
------------+---------------------------+----------------------------+----------------------+---------
| | block device | | vhost device | | vduse driver | | TCP/IP | |
| -------+-------- --------+-------- -------+-------- -----+---- |
| | | | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio-blk driver | | vhost-vdpa driver | | vdpa device | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio bus | | | |
| --------+----+----------- | | | |
| | | | | |
| ----------+---------- | | | |
| | virtio-blk device | | | | |
| ----------+---------- | | | |
| | | | | |
| -----------+----------- | | | |
| | virtio-vdpa driver | | | | |
| -----------+----------- | | | |
| | | | vdpa bus | |
| -----------+----------------------+---------------------------+------------ | |
| ---+--- |
-----------------------------------------------------------------------------------------| NIC |------
---+---
|
---------+---------
| Remote Storages |
-------------------
We make use of it to implement a block device connecting to
our distributed storage, which can be used both in containers and
VMs. Thus, we can have an unified technology stack in this two cases.
To test it with null-blk:
$ qemu-storage-daemon \
--chardev socket,id=charmonitor,path=/tmp/qmp.sock,server,nowait \
--monitor chardev=charmonitor \
--blockdev driver=host_device,cache.direct=on,aio=native,filename=/dev/nullb0,node-name=disk0 \
--export type=vduse-blk,id=test,node-name=disk0,writable=on,name=vduse-null,num-queues=16,queue-size=128
The qemu-storage-daemon can be found at https://github.com/bytedance/qemu/tree/vduse
To make the userspace VDUSE processes such as qemu-storage-daemon able to
be run by an unprivileged user. We did some works on virtio driver to avoid
trusting device, including:
- validating the used length:
* https://lore.kernel.org/lkml/20210531135852.113-1-xieyongji@bytedance.com/
* https://lore.kernel.org/lkml/20210525125622.1203-1-xieyongji@bytedance.com/
- validating the device config:
* https://lore.kernel.org/lkml/20210615104810.151-1-xieyongji@bytedance.com/
- validating the device response:
* https://lore.kernel.org/lkml/20210615105218.214-1-xieyongji@bytedance.com/
Since I'm not sure if I missing something during auditing, especially on some
virtio device drivers that I'm not familiar with, we limit the supported device
type to virtio block device currently. The support for other device types can be
added after the security issue of corresponding device driver is clarified or
fixed in the future.
Future work:
- Improve performance
- Userspace library (find a way to reuse device emulation code in qemu/rust-vmm)
- Support more device types
V7 to V8:
- Rebased to newest kernel tree
- Rework VDUSE driver to handle the device's control path in kernel
- Limit the supported device type to virtio block device
- Export free_iova_fast()
- Remove the virtio-blk and virtio-scsi patches (will send them alone)
- Remove all module parameters
- Use the same MAJOR for both control device and VDUSE devices
- Avoid eventfd cleanup in vduse_dev_release()
V6 to V7:
- Export alloc_iova_fast()
- Add get_config_size() callback
- Add some patches to avoid trusting virtio devices
- Add limited device emulation
- Add some documents
- Use workqueue to inject config irq
- Add parameter on vq irq injecting
- Rename vduse_domain_get_mapping_page() to vduse_domain_get_coherent_page()
- Add WARN_ON() to catch message failure
- Add some padding/reserved fields to uAPI structure
- Fix some bugs
- Rebase to vhost.git
V5 to V6:
- Export receive_fd() instead of __receive_fd()
- Factor out the unmapping logic of pa and va separatedly
- Remove the logic of bounce page allocation in page fault handler
- Use PAGE_SIZE as IOVA allocation granule
- Add EPOLLOUT support
- Enable setting API version in userspace
- Fix some bugs
V4 to V5:
- Remove the patch for irq binding
- Use a single IOTLB for all types of mapping
- Factor out vhost_vdpa_pa_map()
- Add some sample codes in document
- Use receice_fd_user() to pass file descriptor
- Fix some bugs
V3 to V4:
- Rebase to vhost.git
- Split some patches
- Add some documents
- Use ioctl to inject interrupt rather than eventfd
- Enable config interrupt support
- Support binding irq to the specified cpu
- Add two module parameter to limit bounce/iova size
- Create char device rather than anon inode per vduse
- Reuse vhost IOTLB for iova domain
- Rework the message mechnism in control path
V2 to V3:
- Rework the MMU-based IOMMU driver
- Use the iova domain as iova allocator instead of genpool
- Support transferring vma->vm_file in vhost-vdpa
- Add SVA support in vhost-vdpa
- Remove the patches on bounce pages reclaim
V1 to V2:
- Add vhost-vdpa support
- Add some documents
- Based on the vdpa management tool
- Introduce a workqueue for irq injection
- Replace interval tree with array map to store the iova_map
Xie Yongji (10):
iova: Export alloc_iova_fast() and free_iova_fast();
file: Export receive_fd() to modules
eventfd: Increase the recursion depth of eventfd_signal()
vhost-iotlb: Add an opaque pointer for vhost IOTLB
vdpa: Add an opaque pointer for vdpa_config_ops.dma_map()
vdpa: factor out vhost_vdpa_pa_map() and vhost_vdpa_pa_unmap()
vdpa: Support transferring virtual addressing during DMA mapping
vduse: Implement an MMU-based IOMMU driver
vduse: Introduce VDUSE - vDPA Device in Userspace
Documentation: Add documentation for VDUSE
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 +++
drivers/iommu/iova.c | 2 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/ifcvf/ifcvf_main.c | 2 +-
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
drivers/vdpa/vdpa.c | 9 +-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 8 +-
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/iova_domain.c | 545 ++++++++
drivers/vdpa/vdpa_user/iova_domain.h | 73 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
drivers/vdpa/virtio_pci/vp_vdpa.c | 2 +-
drivers/vhost/iotlb.c | 20 +-
drivers/vhost/vdpa.c | 148 +-
fs/eventfd.c | 2 +-
fs/file.c | 6 +
include/linux/eventfd.h | 5 +-
include/linux/file.h | 7 +-
include/linux/vdpa.h | 21 +-
include/linux/vhost_iotlb.h | 3 +
include/uapi/linux/vduse.h | 143 ++
24 files changed, 2641 insertions(+), 50 deletions(-)
create mode 100644 Documentation/userspace-api/vduse.rst
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.c
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.h
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
--
2.11.0
Hi, Yongji
Great work! your method is really wise that implements a software IOMMU
so that data path gets processed by userspace application efficiently.
Sorry, I've just realized your work and patches.
I was working on a similar thing aiming to get vhost-user-blk device
from SPDK vhost-target to be exported as local host kernel block device.
It's diagram is like this:
-----------------------------
------------------------ | ----------------- | ---------------------------------------
| <RunC Container> | <<<<<<<<| Shared-Memory |>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> |
| --------- | v | ----------------- | | v |
| |dev/vdx| | v | <virtio-local-agent> | | <Vhost-user Target> v |
------------+----------- v | ------------------------ | | --------------------------v------ |
| v | |/dev/virtio-local-ctrl| | | | unix socket | |block driver | |
| v ------------+---------------- --------+--------------------v---------
| v | | v
------------+----------------v--------------+----------------------------+--------------------v--------|
| | block device | v | Misc device | | v |
| -------+-------- v --------+------- | v |
| | v | | v |
| ----------+---------- v | | v |
| | virtio-blk driver | v | | v |
| ----------+---------- v | | v |
| | virtio bus v | | v |
| --------+---+------- v | | v |
| | v | | v |
| | v | | v |
| ----------+---------- v ---------+----------- | v |
| | virtio-blk device |--<----| virtio-local driver |----------------< v |
| ----------+---------- ----------+----------- v |
| ---------+--------|
-------------------------------------------------------------------------------------| RNIC |--| PCIe |-
----+--- | NVMe |
| --------
---------+---------
| Remote Storages |
-------------------
I just draft out an initial proof version. When seeing your RFC mail,
I'm thinking that SPDK target may depends on your work, so I could
directly drop mine.
But after a glance of the RFC patches, seems it is not so easy or
efficient to get vduse leveraged by SPDK.
(Please correct me, if I get wrong understanding on vduse. :) )
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer,
then SPDK could also be a potential user of vduse.
Several issues:
- VDUSE needs to limit the total size of the bounce buffers (64M if I
was not wrong). Does it work for SPDK?
- VDUSE can use hugepages but I'm not sure we can mandate hugepages (or
we need introduce new flags for supporting this)
Thanks
It would be better if SPDK vhost-target could leverage the datapath of
vduse directly and efficiently. Even the control path is vdpa based,
we may work out one daemon as agent to bridge SPDK vhost-target with vduse.
Then users who already deployed SPDK vhost-target, can smoothly run
some agent daemon without code modification on SPDK vhost-target itself.
(It is only better-to-have for SPDK vhost-target app, not mandatory for SPDK) :)
At least, some small barrier is there that blocked a vhost-target use vduse
datapath efficiently:
- Current IO completion irq of vduse is IOCTL based. If add one option
to get it eventfd based, then vhost-target can directly notify IO
completion via negotiated eventfd.
Thanks
From Xiaodong
From: Jason Wang <hidden> Date: 2021-06-28 04:40:34
在 2021/6/25 下午12:19, Yongji Xie 写道:
quoted
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
Looks good to me. And I think we can use the reply of the message to
update the status instead of introducing a new command.
Just notice this part in virtio_finalize_features():
virtio_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
status = dev->config->get_status(dev);
if (!(status & VIRTIO_CONFIG_S_FEATURES_OK)) {
So we no reply doesn't work for FEATURES_OK.
So my understanding is:
1) We must not use noreply for set_status()
2) We can use noreply for get_status(), but it requires a new ioctl to
update the status.
So it looks to me we need synchronize for both get_status() and
set_status().
Thanks
On Tue, Jun 15, 2021 at 10:13:21PM +0800, Xie Yongji wrote:
quoted
This series introduces a framework that makes it possible to
implement software-emulated vDPA devices in userspace. And to make it
simple, the emulated vDPA device's control path is handled in the
kernel and only the data path is implemented in the userspace.
Since the emuldated vDPA device's control path is handled in the
kernel, a message mechnism is introduced to make userspace be aware
of the data path related changes. Userspace can use read()/write() to
receive/reply the control messages.
In the data path, the core is mapping dma buffer into VDUSE daemon's
address space, which can be implemented in different ways depending
on the vdpa bus to which the vDPA device is attached.
In virtio-vdpa case, we implements a MMU-based on-chip IOMMU driver
with bounce-buffering mechanism to achieve that. And in vhost-vdpa
case, the dma buffer is reside in a userspace memory region which can
be shared to the VDUSE userspace processs via transferring the shmfd.
The details and our user case is shown below:
------------------------ ------------------------- ------------------------------------------
| ---+--- |
-----------------------------------------------------------------------------------------| NIC
|------
quoted
quoted
---+---
|
---------+---------
| Remote Storages |
-------------------
We make use of it to implement a block device connecting to our
distributed storage, which can be used both in containers and VMs.
Thus, we can have an unified technology stack in this two cases.
To test it with null-blk:
$ qemu-storage-daemon \
--chardev socket,id=charmonitor,path=/tmp/qmp.sock,server,nowait \
--monitor chardev=charmonitor \
--blockdev
--export
type=vduse-blk,id=test,node-name=disk0,writable=on,name=vduse-null,nu
m-queues=16,queue-size=128
The qemu-storage-daemon can be found at
https://github.com/bytedance/qemu/tree/vduse
To make the userspace VDUSE processes such as qemu-storage-daemon
able to be run by an unprivileged user. We did some works on virtio
driver to avoid trusting device, including:
- validating the used length:
* https://lore.kernel.org/lkml/20210531135852.113-1-
- Support more device types
V7 to V8:
- Rebased to newest kernel tree
- Rework VDUSE driver to handle the device's control path in kernel
- Limit the supported device type to virtio block device
- Export free_iova_fast()
- Remove the virtio-blk and virtio-scsi patches (will send them
alone)
- Remove all module parameters
- Use the same MAJOR for both control device and VDUSE devices
- Avoid eventfd cleanup in vduse_dev_release()
V6 to V7:
- Export alloc_iova_fast()
- Add get_config_size() callback
- Add some patches to avoid trusting virtio devices
- Add limited device emulation
- Add some documents
- Use workqueue to inject config irq
- Add parameter on vq irq injecting
- Rename vduse_domain_get_mapping_page() to
vduse_domain_get_coherent_page()
- Add WARN_ON() to catch message failure
- Add some padding/reserved fields to uAPI structure
- Fix some bugs
- Rebase to vhost.git
V5 to V6:
- Export receive_fd() instead of __receive_fd()
- Factor out the unmapping logic of pa and va separatedly
- Remove the logic of bounce page allocation in page fault handler
- Use PAGE_SIZE as IOVA allocation granule
- Add EPOLLOUT support
- Enable setting API version in userspace
- Fix some bugs
V4 to V5:
- Remove the patch for irq binding
- Use a single IOTLB for all types of mapping
- Factor out vhost_vdpa_pa_map()
- Add some sample codes in document
- Use receice_fd_user() to pass file descriptor
- Fix some bugs
V3 to V4:
- Rebase to vhost.git
- Split some patches
- Add some documents
- Use ioctl to inject interrupt rather than eventfd
- Enable config interrupt support
- Support binding irq to the specified cpu
- Add two module parameter to limit bounce/iova size
- Create char device rather than anon inode per vduse
- Reuse vhost IOTLB for iova domain
- Rework the message mechnism in control path
V2 to V3:
- Rework the MMU-based IOMMU driver
- Use the iova domain as iova allocator instead of genpool
- Support transferring vma->vm_file in vhost-vdpa
- Add SVA support in vhost-vdpa
- Remove the patches on bounce pages reclaim
V1 to V2:
- Add vhost-vdpa support
- Add some documents
- Based on the vdpa management tool
- Introduce a workqueue for irq injection
- Replace interval tree with array map to store the iova_map
Xie Yongji (10):
iova: Export alloc_iova_fast() and free_iova_fast();
file: Export receive_fd() to modules
eventfd: Increase the recursion depth of eventfd_signal()
vhost-iotlb: Add an opaque pointer for vhost IOTLB
vdpa: Add an opaque pointer for vdpa_config_ops.dma_map()
vdpa: factor out vhost_vdpa_pa_map() and vhost_vdpa_pa_unmap()
vdpa: Support transferring virtual addressing during DMA mapping
vduse: Implement an MMU-based IOMMU driver
vduse: Introduce VDUSE - vDPA Device in Userspace
Documentation: Add documentation for VDUSE
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 +++
drivers/iommu/iova.c | 2 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/ifcvf/ifcvf_main.c | 2 +-
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
drivers/vdpa/vdpa.c | 9 +-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 8 +-
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/iova_domain.c | 545 ++++++++
drivers/vdpa/vdpa_user/iova_domain.h | 73 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453
Hi, Yongji
Great work! your method is really wise that implements a software
IOMMU so that data path gets processed by userspace application efficiently.
Sorry, I've just realized your work and patches.
I was working on a similar thing aiming to get vhost-user-blk device
from SPDK vhost-target to be exported as local host kernel block device.
It's diagram is like this:
-----------------------------
------------------------ | ----------------- | --------------------------------------
-
quoted
| <RunC Container> | <<<<<<<<| Shared-Memory
|>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> |
quoted
| --------- | v | ----------------- | | v |
| |dev/vdx| | v | <virtio-local-agent> | | <Vhost-user Target>
v |
quoted
------------+----------- v | ------------------------ | | --------------------------v-----
| v ------------+---------------- --------+--------------------v---------
| v | | v
------------+----------------v--------------+----------------------------+--------------------
v--------|
quoted
| | block device | v | Misc device | | v |
| -------+-------- v --------+------- | v |
| | v | | v |
| ----------+---------- v | | v |
| | virtio-blk driver | v | | v |
| ----------+---------- v | | v |
| | virtio bus v | | v |
| --------+---+------- v | | v |
| | v | | v |
| | v | | v |
| ----------+---------- v ---------+----------- | v |
| | virtio-blk device |--<----| virtio-local driver |----------------< v
----+--- | NVMe |
| --------
---------+---------
| Remote Storages |
-------------------
I just draft out an initial proof version. When seeing your RFC mail,
I'm thinking that SPDK target may depends on your work, so I could
directly drop mine.
But after a glance of the RFC patches, seems it is not so easy or
efficient to get vduse leveraged by SPDK.
(Please correct me, if I get wrong understanding on vduse. :) )
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer, then SPDK
could also be a potential user of vduse.
Several issues:
- VDUSE needs to limit the total size of the bounce buffers (64M if I was not
wrong). Does it work for SPDK?
Yes, Jason. It is enough and works for SPDK.
Since it's a kind of bounce buffer mainly for in-flight IO, so limited size like
64MB is enough.
- VDUSE can use hugepages but I'm not sure we can mandate hugepages (or we
need introduce new flags for supporting this)
Same with your worry, I'm afraid too that it is a hard for a kernel module
to directly preallocate hugepage internal.
What I tried is that:
1. A simple agent daemon (represents for one device) `preallocates` and maps
dozens of 2MB hugepages (like 64MB) for one device.
2. The daemon passes its mapping addr&len and hugepage fd to kernel
module through created IOCTL.
3. Kernel module remaps the hugepages inside kernel.
4. Vhost user target gets and maps hugepage fd from kernel module
in vhost-user msg through Unix Domain Socket cmsg.
Then kernel module and target map on the same hugepage based
bounce buffer for in-flight IO.
If there is one option in VDUSE to map userspace preallocated memory, then
VDUSE should be able to mandate it even it is hugepage based.
Thanks
quoted
It would be better if SPDK vhost-target could leverage the datapath of
vduse directly and efficiently. Even the control path is vdpa based,
we may work out one daemon as agent to bridge SPDK vhost-target with vduse.
Then users who already deployed SPDK vhost-target, can smoothly run
some agent daemon without code modification on SPDK vhost-target itself.
(It is only better-to-have for SPDK vhost-target app, not mandatory
for SPDK) :) At least, some small barrier is there that blocked a
vhost-target use vduse datapath efficiently:
- Current IO completion irq of vduse is IOCTL based. If add one option
to get it eventfd based, then vhost-target can directly notify IO
completion via negotiated eventfd.
Thanks
From Xiaodong
On Mon, 28 Jun 2021 at 10:55, Liu Xiaodong [off-list ref] wrote:
On Tue, Jun 15, 2021 at 10:13:21PM +0800, Xie Yongji wrote:
quoted
This series introduces a framework that makes it possible to implement
software-emulated vDPA devices in userspace. And to make it simple, the
emulated vDPA device's control path is handled in the kernel and only the
data path is implemented in the userspace.
Since the emuldated vDPA device's control path is handled in the kernel,
a message mechnism is introduced to make userspace be aware of the data
path related changes. Userspace can use read()/write() to receive/reply
the control messages.
In the data path, the core is mapping dma buffer into VDUSE daemon's
address space, which can be implemented in different ways depending on
the vdpa bus to which the vDPA device is attached.
In virtio-vdpa case, we implements a MMU-based on-chip IOMMU driver with
bounce-buffering mechanism to achieve that. And in vhost-vdpa case, the dma
buffer is reside in a userspace memory region which can be shared to the
VDUSE userspace processs via transferring the shmfd.
The details and our user case is shown below:
------------------------ ------------------------- ----------------------------------------------
| Container | | QEMU(VM) | | VDUSE daemon |
| --------- | | ------------------- | | ------------------------- ---------------- |
| |dev/vdx| | | |/dev/vhost-vdpa-x| | | | vDPA device emulation | | block driver | |
------------+----------- -----------+------------ -------------+----------------------+---------
| | | |
| | | |
------------+---------------------------+----------------------------+----------------------+---------
| | block device | | vhost device | | vduse driver | | TCP/IP | |
| -------+-------- --------+-------- -------+-------- -----+---- |
| | | | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio-blk driver | | vhost-vdpa driver | | vdpa device | | |
| ----------+---------- ----------+----------- -------+------- | |
| | virtio bus | | | |
| --------+----+----------- | | | |
| | | | | |
| ----------+---------- | | | |
| | virtio-blk device | | | | |
| ----------+---------- | | | |
| | | | | |
| -----------+----------- | | | |
| | virtio-vdpa driver | | | | |
| -----------+----------- | | | |
| | | | vdpa bus | |
| -----------+----------------------+---------------------------+------------ | |
| ---+--- |
-----------------------------------------------------------------------------------------| NIC |------
---+---
|
---------+---------
| Remote Storages |
-------------------
We make use of it to implement a block device connecting to
our distributed storage, which can be used both in containers and
VMs. Thus, we can have an unified technology stack in this two cases.
To test it with null-blk:
$ qemu-storage-daemon \
--chardev socket,id=charmonitor,path=/tmp/qmp.sock,server,nowait \
--monitor chardev=charmonitor \
--blockdev driver=host_device,cache.direct=on,aio=native,filename=/dev/nullb0,node-name=disk0 \
--export type=vduse-blk,id=test,node-name=disk0,writable=on,name=vduse-null,num-queues=16,queue-size=128
The qemu-storage-daemon can be found at https://github.com/bytedance/qemu/tree/vduse
To make the userspace VDUSE processes such as qemu-storage-daemon able to
be run by an unprivileged user. We did some works on virtio driver to avoid
trusting device, including:
- validating the used length:
* https://lore.kernel.org/lkml/20210531135852.113-1-xieyongji@bytedance.com/
* https://lore.kernel.org/lkml/20210525125622.1203-1-xieyongji@bytedance.com/
- validating the device config:
* https://lore.kernel.org/lkml/20210615104810.151-1-xieyongji@bytedance.com/
- validating the device response:
* https://lore.kernel.org/lkml/20210615105218.214-1-xieyongji@bytedance.com/
Since I'm not sure if I missing something during auditing, especially on some
virtio device drivers that I'm not familiar with, we limit the supported device
type to virtio block device currently. The support for other device types can be
added after the security issue of corresponding device driver is clarified or
fixed in the future.
Future work:
- Improve performance
- Userspace library (find a way to reuse device emulation code in qemu/rust-vmm)
- Support more device types
V7 to V8:
- Rebased to newest kernel tree
- Rework VDUSE driver to handle the device's control path in kernel
- Limit the supported device type to virtio block device
- Export free_iova_fast()
- Remove the virtio-blk and virtio-scsi patches (will send them alone)
- Remove all module parameters
- Use the same MAJOR for both control device and VDUSE devices
- Avoid eventfd cleanup in vduse_dev_release()
V6 to V7:
- Export alloc_iova_fast()
- Add get_config_size() callback
- Add some patches to avoid trusting virtio devices
- Add limited device emulation
- Add some documents
- Use workqueue to inject config irq
- Add parameter on vq irq injecting
- Rename vduse_domain_get_mapping_page() to vduse_domain_get_coherent_page()
- Add WARN_ON() to catch message failure
- Add some padding/reserved fields to uAPI structure
- Fix some bugs
- Rebase to vhost.git
V5 to V6:
- Export receive_fd() instead of __receive_fd()
- Factor out the unmapping logic of pa and va separatedly
- Remove the logic of bounce page allocation in page fault handler
- Use PAGE_SIZE as IOVA allocation granule
- Add EPOLLOUT support
- Enable setting API version in userspace
- Fix some bugs
V4 to V5:
- Remove the patch for irq binding
- Use a single IOTLB for all types of mapping
- Factor out vhost_vdpa_pa_map()
- Add some sample codes in document
- Use receice_fd_user() to pass file descriptor
- Fix some bugs
V3 to V4:
- Rebase to vhost.git
- Split some patches
- Add some documents
- Use ioctl to inject interrupt rather than eventfd
- Enable config interrupt support
- Support binding irq to the specified cpu
- Add two module parameter to limit bounce/iova size
- Create char device rather than anon inode per vduse
- Reuse vhost IOTLB for iova domain
- Rework the message mechnism in control path
V2 to V3:
- Rework the MMU-based IOMMU driver
- Use the iova domain as iova allocator instead of genpool
- Support transferring vma->vm_file in vhost-vdpa
- Add SVA support in vhost-vdpa
- Remove the patches on bounce pages reclaim
V1 to V2:
- Add vhost-vdpa support
- Add some documents
- Based on the vdpa management tool
- Introduce a workqueue for irq injection
- Replace interval tree with array map to store the iova_map
Xie Yongji (10):
iova: Export alloc_iova_fast() and free_iova_fast();
file: Export receive_fd() to modules
eventfd: Increase the recursion depth of eventfd_signal()
vhost-iotlb: Add an opaque pointer for vhost IOTLB
vdpa: Add an opaque pointer for vdpa_config_ops.dma_map()
vdpa: factor out vhost_vdpa_pa_map() and vhost_vdpa_pa_unmap()
vdpa: Support transferring virtual addressing during DMA mapping
vduse: Implement an MMU-based IOMMU driver
vduse: Introduce VDUSE - vDPA Device in Userspace
Documentation: Add documentation for VDUSE
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/ioctl/ioctl-number.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 +++
drivers/iommu/iova.c | 2 +
drivers/vdpa/Kconfig | 10 +
drivers/vdpa/Makefile | 1 +
drivers/vdpa/ifcvf/ifcvf_main.c | 2 +-
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
drivers/vdpa/vdpa.c | 9 +-
drivers/vdpa/vdpa_sim/vdpa_sim.c | 8 +-
drivers/vdpa/vdpa_user/Makefile | 5 +
drivers/vdpa/vdpa_user/iova_domain.c | 545 ++++++++
drivers/vdpa/vdpa_user/iova_domain.h | 73 +
drivers/vdpa/vdpa_user/vduse_dev.c | 1453 ++++++++++++++++++++
drivers/vdpa/virtio_pci/vp_vdpa.c | 2 +-
drivers/vhost/iotlb.c | 20 +-
drivers/vhost/vdpa.c | 148 +-
fs/eventfd.c | 2 +-
fs/file.c | 6 +
include/linux/eventfd.h | 5 +-
include/linux/file.h | 7 +-
include/linux/vdpa.h | 21 +-
include/linux/vhost_iotlb.h | 3 +
include/uapi/linux/vduse.h | 143 ++
24 files changed, 2641 insertions(+), 50 deletions(-)
create mode 100644 Documentation/userspace-api/vduse.rst
create mode 100644 drivers/vdpa/vdpa_user/Makefile
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.c
create mode 100644 drivers/vdpa/vdpa_user/iova_domain.h
create mode 100644 drivers/vdpa/vdpa_user/vduse_dev.c
create mode 100644 include/uapi/linux/vduse.h
--
2.11.0
Hi, Yongji
Great work! your method is really wise that implements a software IOMMU
so that data path gets processed by userspace application efficiently.
Sorry, I've just realized your work and patches.
I was working on a similar thing aiming to get vhost-user-blk device
from SPDK vhost-target to be exported as local host kernel block device.
It's diagram is like this:
-----------------------------
------------------------ | ----------------- | ---------------------------------------
| <RunC Container> | <<<<<<<<| Shared-Memory |>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> |
| --------- | v | ----------------- | | v |
| |dev/vdx| | v | <virtio-local-agent> | | <Vhost-user Target> v |
------------+----------- v | ------------------------ | | --------------------------v------ |
| v | |/dev/virtio-local-ctrl| | | | unix socket | |block driver | |
| v ------------+---------------- --------+--------------------v---------
| v | | v
------------+----------------v--------------+----------------------------+--------------------v--------|
| | block device | v | Misc device | | v |
| -------+-------- v --------+------- | v |
| | v | | v |
| ----------+---------- v | | v |
| | virtio-blk driver | v | | v |
| ----------+---------- v | | v |
| | virtio bus v | | v |
| --------+---+------- v | | v |
| | v | | v |
| | v | | v |
| ----------+---------- v ---------+----------- | v |
| | virtio-blk device |--<----| virtio-local driver |----------------< v |
| ----------+---------- ----------+----------- v |
| ---------+--------|
-------------------------------------------------------------------------------------| RNIC |--| PCIe |-
----+--- | NVMe |
| --------
---------+---------
| Remote Storages |
-------------------
Oh, yes, this design is similar to VDUSE.
I just draft out. an initial proof version. When seeing your RFC mail,
I'm thinking that SPDK target may depends on your work, so I could
directly drop mine.
Great to hear that! I think we can extend VDUSE to meet your needs.
But I prefer to do that after this initial version merged.
But after a glance of the RFC patches, seems it is not so easy or
efficient to get vduse leveraged by SPDK.
(Please correct me, if I get wrong understanding on vduse. :) )
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer,
then SPDK could also be a potential user of vduse.
I think we can support registering user space memory for bounce-buffer
use like XDP does. But this needs to pin the pages, so I didn't
consider it in this initial version.
It would be better if SPDK vhost-target could leverage the datapath of
vduse directly and efficiently. Even the control path is vdpa based,
we may work out one daemon as agent to bridge SPDK vhost-target with vduse.
Then users who already deployed SPDK vhost-target, can smoothly run
some agent daemon without code modification on SPDK vhost-target itself.
That's a good idea!
(It is only better-to-have for SPDK vhost-target app, not mandatory for SPDK) :)
At least, some small barrier is there that blocked a vhost-target use vduse
datapath efficiently:
- Current IO completion irq of vduse is IOCTL based. If add one option
to get it eventfd based, then vhost-target can directly notify IO
completion via negotiated eventfd.
Make sense. Actually we did use the eventfd mechanism for this purpose
in the old version. But using ioctl would be simple, so we choose it
in this initial version.
Thanks,
Yongji
@@ -0,0 +1,143 @@+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */+#ifndef _UAPI_VDUSE_H_+#define _UAPI_VDUSE_H_++#include<linux/types.h>++#define VDUSE_API_VERSION 0++#define VDUSE_NAME_MAX 256++/* the control messages definition for read/write */++enumvduse_req_type{+/* Get the state for virtqueue from userspace */+VDUSE_GET_VQ_STATE,+/* Notify userspace to start the dataplane, no reply */+VDUSE_START_DATAPLANE,+/* Notify userspace to stop the dataplane, no reply */+VDUSE_STOP_DATAPLANE,+/* Notify userspace to update the memory mapping in device IOTLB */+VDUSE_UPDATE_IOTLB,+};++structvduse_vq_state{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+};++structvduse_iova_range{+__u64start;/* start of the IOVA range */+__u64last;/* end of the IOVA range */
Please clarify whether this describes a closed range [start, last] or an
open range [start, last).
+};
+
+struct vduse_dev_request {
+ __u32 type; /* request type */
+ __u32 request_id; /* request id */
+#define VDUSE_REQ_FLAGS_NO_REPLY (1 << 0) /* No need to reply */
+ __u32 flags; /* request flags */
+ __u32 reserved; /* for future use */
+ union {
+ struct vduse_vq_state vq_state; /* virtqueue state */
+ struct vduse_iova_range iova; /* iova range for updating */
+ __u32 padding[16]; /* padding */
+ };
+};
+
+struct vduse_dev_response {
+ __u32 request_id; /* corresponding request id */
+#define VDUSE_REQ_RESULT_OK 0x00
+#define VDUSE_REQ_RESULT_FAILED 0x01
+ __u32 result; /* the result of request */
+ __u32 reserved[2]; /* for future use */
+ union {
+ struct vduse_vq_state vq_state; /* virtqueue state */
+ __u32 padding[16]; /* padding */
+ };
+};
+
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
This constant seems to impose the constraint that all virtqueues have
the same maximum size. Is this really necessary?
+ __u16 padding; /* padding */
+ __u32 vq_num; /* the number of virtqueues */
+ __u32 vq_align; /* the allocation alignment of virtqueue's metadata */
I'm not sure what this is?
+ __u32 config_size; /* the size of the configuration space */
+ __u32 reserved[15]; /* for future use */
+ __u8 config[0]; /* the buffer of the configuration space */
+};
+
+struct vduse_iotlb_entry {
+ __u64 offset; /* the mmap offset on fd */
+ __u64 start; /* start of the IOVA range */
+ __u64 last; /* last of the IOVA range */
Same here, please specify whether this is an open range or a closed
range.
+#define VDUSE_ACCESS_RO 0x1
+#define VDUSE_ACCESS_WO 0x2
+#define VDUSE_ACCESS_RW 0x3
+ __u8 perm; /* access permission of this range */
+};
+
+struct vduse_config_update {
+ __u32 offset; /* offset from the beginning of configuration space */
+ __u32 length; /* the length to write to configuration space */
+ __u8 buffer[0]; /* buffer used to write from */
+};
+
+struct vduse_vq_info {
+ __u32 index; /* virtqueue index */
+ __u32 avail_idx; /* virtqueue state (last_avail_idx) */
+ __u64 desc_addr; /* address of desc area */
+ __u64 driver_addr; /* address of driver area */
+ __u64 device_addr; /* address of device area */
+ __u32 num; /* the size of virtqueue */
+ __u8 ready; /* ready status of virtqueue */
+};
+
+struct vduse_vq_eventfd {
+ __u32 index; /* virtqueue index */
+#define VDUSE_EVENTFD_DEASSIGN -1
+ int fd; /* eventfd, -1 means de-assigning the eventfd */
+};
+
+#define VDUSE_BASE 0x81
+
+/* Get the version of VDUSE API. This is used for future extension */
+#define VDUSE_GET_API_VERSION _IOR(VDUSE_BASE, 0x00, __u64)
+
+/* Set the version of VDUSE API. */
+#define VDUSE_SET_API_VERSION _IOW(VDUSE_BASE, 0x01, __u64)
+
+/* Create a vduse device which is represented by a char device (/dev/vduse/<name>) */
+#define VDUSE_CREATE_DEV _IOW(VDUSE_BASE, 0x02, struct vduse_dev_config)
+
+/* Destroy a vduse device. Make sure there are no references to the char device */
+#define VDUSE_DESTROY_DEV _IOW(VDUSE_BASE, 0x03, char[VDUSE_NAME_MAX])
+
+/*
+ * Get a file descriptor for the first overlapped iova region,
+ * -EINVAL means the iova region doesn't exist.
+ */
+#define VDUSE_IOTLB_GET_FD _IOWR(VDUSE_BASE, 0x04, struct vduse_iotlb_entry)
+
+/* Get the negotiated features */
+#define VDUSE_DEV_GET_FEATURES _IOR(VDUSE_BASE, 0x05, __u64)
+
+/* Update the configuration space */
+#define VDUSE_DEV_UPDATE_CONFIG _IOW(VDUSE_BASE, 0x06, struct vduse_config_update)
+
+/* Get the specified virtqueue's information */
+#define VDUSE_VQ_GET_INFO _IOWR(VDUSE_BASE, 0x07, struct vduse_vq_info)
+
+/* Setup an eventfd to receive kick for virtqueue */
+#define VDUSE_VQ_SETUP_KICKFD _IOW(VDUSE_BASE, 0x08, struct vduse_vq_eventfd)
+
+/* Inject an interrupt for specific virtqueue */
+#define VDUSE_VQ_INJECT_IRQ _IOW(VDUSE_BASE, 0x09, __u32)
There is not enough documentation to use this header file. For example,
which ioctls are used with /dev/vduse and which are used with
/dev/vduse/<name>?
Please document that ioctl API fully. It will not only help userspace
developers but also define what is part of the interface and what is an
implementation detail that can change in the future.
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-06-28 13:02:25
On Tue, Jun 15, 2021 at 10:13:31PM +0800, Xie Yongji wrote:
quoted hunk
VDUSE (vDPA Device in Userspace) is a framework to support
implementing software-emulated vDPA devices in userspace. This
document is intended to clarify the VDUSE design and usage.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 ++++++++++++++++++++++++++++++++++
2 files changed, 223 insertions(+)
create mode 100644 Documentation/userspace-api/vduse.rst
@@ -0,0 +1,222 @@+==================================+VDUSE - "vDPA Device in Userspace"+==================================++vDPA (virtio data path acceleration) device is a device that uses a+datapath which complies with the virtio specifications with vendor+specific control path. vDPA devices can be both physically located on+the hardware or emulated by software. VDUSE is a framework that makes it+possible to implement software-emulated vDPA devices in userspace. And+to make it simple, the emulated vDPA device's control path is handled in+the kernel and only the data path is implemented in the userspace.++Note that only virtio block device is supported by VDUSE framework now,+which can reduce security risks when the userspace process that implements+the data path is run by an unprivileged user. The Support for other device+types can be added after the security issue is clarified or fixed in the future.++Start/Stop VDUSE devices+------------------------++VDUSE devices are started as follows:++1. Create a new VDUSE instance with ioctl(VDUSE_CREATE_DEV) on+ /dev/vduse/control.++2. Begin processing VDUSE messages from /dev/vduse/$NAME. The first+ messages will arrive while attaching the VDUSE instance to vDPA bus.++3. Send the VDPA_CMD_DEV_NEW netlink message to attach the VDUSE+ instance to vDPA bus.++VDUSE devices are stopped as follows:++1. Send the VDPA_CMD_DEV_DEL netlink message to detach the VDUSE+ instance from vDPA bus.++2. Close the file descriptor referring to /dev/vduse/$NAME++3. Destroy the VDUSE instance with ioctl(VDUSE_DESTROY_DEV) on+ /dev/vduse/control++The netlink messages metioned above can be sent via vdpa tool in iproute2+or use the below sample codes:++..code-block:: c++ static int netlink_add_vduse(const char *name, enum vdpa_command cmd)+ {+ struct nl_sock *nlsock;+ struct nl_msg *msg;+ int famid;++ nlsock = nl_socket_alloc();+ if (!nlsock)+ return -ENOMEM;++ if (genl_connect(nlsock))+ goto free_sock;++ famid = genl_ctrl_resolve(nlsock, VDPA_GENL_NAME);+ if (famid < 0)+ goto close_sock;++ msg = nlmsg_alloc();+ if (!msg)+ goto close_sock;++ if (!genlmsg_put(msg, NL_AUTO_PORT, NL_AUTO_SEQ, famid, 0, 0, cmd, 0))+ goto nla_put_failure;++ NLA_PUT_STRING(msg, VDPA_ATTR_DEV_NAME, name);+ if (cmd == VDPA_CMD_DEV_NEW)+ NLA_PUT_STRING(msg, VDPA_ATTR_MGMTDEV_DEV_NAME, "vduse");++ if (nl_send_sync(nlsock, msg))+ goto close_sock;++ nl_close(nlsock);+ nl_socket_free(nlsock);++ return 0;+ nla_put_failure:+ nlmsg_free(msg);+ close_sock:+ nl_close(nlsock);+ free_sock:+ nl_socket_free(nlsock);+ return -1;+ }++How VDUSE works+---------------++Since the emuldated vDPA device's control path is handled in the kernel,
s/emuldated/emulated/
+a message-based communication protocol and few types of control messages
+are introduced by VDUSE framework to make userspace be aware of the data
+path related changes:
+
+- VDUSE_GET_VQ_STATE: Get the state for virtqueue from userspace
+
+- VDUSE_START_DATAPLANE: Notify userspace to start the dataplane
+
+- VDUSE_STOP_DATAPLANE: Notify userspace to stop the dataplane
+
+- VDUSE_UPDATE_IOTLB: Notify userspace to update the memory mapping in device IOTLB
+
+Userspace needs to read()/write() on /dev/vduse/$NAME to receive/reply
+those control messages from/to VDUSE kernel module as follows:
+
+.. code-block:: c
+
+ static int vduse_message_handler(int dev_fd)
+ {
+ int len;
+ struct vduse_dev_request req;
+ struct vduse_dev_response resp;
+
+ len = read(dev_fd, &req, sizeof(req));
+ if (len != sizeof(req))
+ return -1;
+
+ resp.request_id = req.request_id;
+
+ switch (req.type) {
+
+ /* handle different types of message */
+
+ }
+
+ if (req.flags & VDUSE_REQ_FLAGS_NO_REPLY)
+ return 0;
+
+ len = write(dev_fd, &resp, sizeof(resp));
+ if (len != sizeof(resp))
+ return -1;
+
+ return 0;
+ }
+
+After VDUSE_START_DATAPLANE messages is received, userspace should start the
+dataplane processing with the help of some ioctls on /dev/vduse/$NAME:
+
+- VDUSE_IOTLB_GET_FD: get the file descriptor to the first overlapped iova region.
+ Userspace can access this iova region by passing fd and corresponding size, offset,
+ perm to mmap(). For example:
+
+.. code-block:: c
+
+ static int perm_to_prot(uint8_t perm)
+ {
+ int prot = 0;
+
+ switch (perm) {
+ case VDUSE_ACCESS_WO:
+ prot |= PROT_WRITE;
+ break;
+ case VDUSE_ACCESS_RO:
+ prot |= PROT_READ;
+ break;
+ case VDUSE_ACCESS_RW:
+ prot |= PROT_READ | PROT_WRITE;
+ break;
+ }
+
+ return prot;
+ }
+
+ static void *iova_to_va(int dev_fd, uint64_t iova, uint64_t *len)
+ {
+ int fd;
+ void *addr;
+ size_t size;
+ struct vduse_iotlb_entry entry;
+
+ entry.start = iova;
+ entry.last = iova + 1;
Why +1?
I expected the request to include *len so that VDUSE can create a bounce
buffer for the full iova range, if necessary.
How is userspace expected to manage iotlb mmaps? When should munmap(2)
be called?
Should userspace expect VDUSE_IOTLB_GET_FD to return a full chunk of
guest RAM (e.g. multiple gigabytes) that can be cached permanently or
will it return just enough pages to cover [start, last)?
+
+ return addr + iova - entry.start;
+ }
+
+- VDUSE_DEV_GET_FEATURES: Get the negotiated features
Are these VIRTIO feature bits? Please explain how feature negotiation
works. There must be a way for userspace to report the device's
supported feature bits to the kernel.
+- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt
Does this mean the contents of the configuration space are cached by
VDUSE? The downside is that the userspace code cannot generate the
contents on demand. Most devices doin't need to generate the contents
on demand, so I think this is okay but I had expected a different
interface:
kernel->userspace VDUSE_DEV_GET_CONFIG
userspace->kernel VDUSE_DEV_INJECT_CONFIG_IRQ
I think you can leave it the way it is, but I wanted to mention this in
case someone thinks it's important to support generating the contents of
the configuration space on demand.
+- VDUSE_VQ_GET_INFO: Get the specified virtqueue's metadata
+
+- VDUSE_VQ_SETUP_KICKFD: set the kickfd for virtqueue, this eventfd is used
+ by VDUSE kernel module to notify userspace to consume the vring.
+
+- VDUSE_INJECT_VQ_IRQ: inject an interrupt for specific virtqueue
This information is useful but it's not enough to be able to implement a
userspace device. Please provide more developer documentation or at
least refer to uapi header files, published documents, etc that contain
the details.
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-06-28 13:02:40
On Tue, Jun 15, 2021 at 10:13:21PM +0800, Xie Yongji wrote:
This series introduces a framework that makes it possible to implement
software-emulated vDPA devices in userspace. And to make it simple, the
emulated vDPA device's control path is handled in the kernel and only the
data path is implemented in the userspace.
This looks interesting. Unfortunately I don't have enough time to do a
full review, but I looked at the documentation and uapi header file to
give feedback on the userspace ABI.
Stefan
On Mon, Jun 28, 2021 at 12:40 PM Jason Wang [off-list ref] wrote:
在 2021/6/25 下午12:19, Yongji Xie 写道:
quoted
quoted
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
Looks good to me. And I think we can use the reply of the message to
update the status instead of introducing a new command.
Just notice this part in virtio_finalize_features():
virtio_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
status = dev->config->get_status(dev);
if (!(status & VIRTIO_CONFIG_S_FEATURES_OK)) {
So we no reply doesn't work for FEATURES_OK.
So my understanding is:
1) We must not use noreply for set_status()
2) We can use noreply for get_status(), but it requires a new ioctl to
update the status.
So it looks to me we need synchronize for both get_status() and
set_status().
We should not send messages to userspace in the FEATURES_OK case. So
the synchronization is not necessary.
Thanks,
Yongji
@@ -0,0 +1,143 @@+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */+#ifndef _UAPI_VDUSE_H_+#define _UAPI_VDUSE_H_++#include<linux/types.h>++#define VDUSE_API_VERSION 0++#define VDUSE_NAME_MAX 256++/* the control messages definition for read/write */++enumvduse_req_type{+/* Get the state for virtqueue from userspace */+VDUSE_GET_VQ_STATE,+/* Notify userspace to start the dataplane, no reply */+VDUSE_START_DATAPLANE,+/* Notify userspace to stop the dataplane, no reply */+VDUSE_STOP_DATAPLANE,+/* Notify userspace to update the memory mapping in device IOTLB */+VDUSE_UPDATE_IOTLB,+};++structvduse_vq_state{+__u32index;/* virtqueue index */+__u32avail_idx;/* virtqueue state (last_avail_idx) */+};++structvduse_iova_range{+__u64start;/* start of the IOVA range */+__u64last;/* end of the IOVA range */
Please clarify whether this describes a closed range [start, last] or an
open range [start, last).
OK.
quoted
+};
+
+struct vduse_dev_request {
+ __u32 type; /* request type */
+ __u32 request_id; /* request id */
+#define VDUSE_REQ_FLAGS_NO_REPLY (1 << 0) /* No need to reply */
+ __u32 flags; /* request flags */
+ __u32 reserved; /* for future use */
+ union {
+ struct vduse_vq_state vq_state; /* virtqueue state */
+ struct vduse_iova_range iova; /* iova range for updating */
+ __u32 padding[16]; /* padding */
+ };
+};
+
+struct vduse_dev_response {
+ __u32 request_id; /* corresponding request id */
+#define VDUSE_REQ_RESULT_OK 0x00
+#define VDUSE_REQ_RESULT_FAILED 0x01
+ __u32 result; /* the result of request */
+ __u32 reserved[2]; /* for future use */
+ union {
+ struct vduse_vq_state vq_state; /* virtqueue state */
+ __u32 padding[16]; /* padding */
+ };
+};
+
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
Good point! But it looks like virtio-vdpa/virtio-pci doesn't support
that now. All virtqueues have the same maximum size.
This constant seems to impose the constraint that all virtqueues have
the same maximum size. Is this really necessary?
This will be used by vring_create_virtqueue(). We need to specify the
maximum queue size supported by the device.
quoted
+ __u16 padding; /* padding */
+ __u32 vq_num; /* the number of virtqueues */
+ __u32 vq_align; /* the allocation alignment of virtqueue's metadata */
I'm not sure what this is?
This will be used by vring_create_virtqueue() too.
quoted
+ __u32 config_size; /* the size of the configuration space */
+ __u32 reserved[15]; /* for future use */
+ __u8 config[0]; /* the buffer of the configuration space */
+};
+
+struct vduse_iotlb_entry {
+ __u64 offset; /* the mmap offset on fd */
+ __u64 start; /* start of the IOVA range */
+ __u64 last; /* last of the IOVA range */
Same here, please specify whether this is an open range or a closed
range.
Sure.
quoted
+#define VDUSE_ACCESS_RO 0x1
+#define VDUSE_ACCESS_WO 0x2
+#define VDUSE_ACCESS_RW 0x3
+ __u8 perm; /* access permission of this range */
+};
+
+struct vduse_config_update {
+ __u32 offset; /* offset from the beginning of configuration space */
+ __u32 length; /* the length to write to configuration space */
+ __u8 buffer[0]; /* buffer used to write from */
+};
+
+struct vduse_vq_info {
+ __u32 index; /* virtqueue index */
+ __u32 avail_idx; /* virtqueue state (last_avail_idx) */
+ __u64 desc_addr; /* address of desc area */
+ __u64 driver_addr; /* address of driver area */
+ __u64 device_addr; /* address of device area */
+ __u32 num; /* the size of virtqueue */
+ __u8 ready; /* ready status of virtqueue */
+};
+
+struct vduse_vq_eventfd {
+ __u32 index; /* virtqueue index */
+#define VDUSE_EVENTFD_DEASSIGN -1
+ int fd; /* eventfd, -1 means de-assigning the eventfd */
+};
+
+#define VDUSE_BASE 0x81
+
+/* Get the version of VDUSE API. This is used for future extension */
+#define VDUSE_GET_API_VERSION _IOR(VDUSE_BASE, 0x00, __u64)
+
+/* Set the version of VDUSE API. */
+#define VDUSE_SET_API_VERSION _IOW(VDUSE_BASE, 0x01, __u64)
+
+/* Create a vduse device which is represented by a char device (/dev/vduse/<name>) */
+#define VDUSE_CREATE_DEV _IOW(VDUSE_BASE, 0x02, struct vduse_dev_config)
+
+/* Destroy a vduse device. Make sure there are no references to the char device */
+#define VDUSE_DESTROY_DEV _IOW(VDUSE_BASE, 0x03, char[VDUSE_NAME_MAX])
+
+/*
+ * Get a file descriptor for the first overlapped iova region,
+ * -EINVAL means the iova region doesn't exist.
+ */
+#define VDUSE_IOTLB_GET_FD _IOWR(VDUSE_BASE, 0x04, struct vduse_iotlb_entry)
+
+/* Get the negotiated features */
+#define VDUSE_DEV_GET_FEATURES _IOR(VDUSE_BASE, 0x05, __u64)
+
+/* Update the configuration space */
+#define VDUSE_DEV_UPDATE_CONFIG _IOW(VDUSE_BASE, 0x06, struct vduse_config_update)
+
+/* Get the specified virtqueue's information */
+#define VDUSE_VQ_GET_INFO _IOWR(VDUSE_BASE, 0x07, struct vduse_vq_info)
+
+/* Setup an eventfd to receive kick for virtqueue */
+#define VDUSE_VQ_SETUP_KICKFD _IOW(VDUSE_BASE, 0x08, struct vduse_vq_eventfd)
+
+/* Inject an interrupt for specific virtqueue */
+#define VDUSE_VQ_INJECT_IRQ _IOW(VDUSE_BASE, 0x09, __u32)
There is not enough documentation to use this header file. For example,
which ioctls are used with /dev/vduse and which are used with
/dev/vduse/<name>?
Please document that ioctl API fully. It will not only help userspace
developers but also define what is part of the interface and what is an
implementation detail that can change in the future.
OK, I will try to add more details.
Thanks,
Yongji
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
On Tue, Jun 15, 2021 at 10:13:21PM +0800, Xie Yongji wrote:
quoted
This series introduces a framework that makes it possible to implement
software-emulated vDPA devices in userspace. And to make it simple, the
emulated vDPA device's control path is handled in the kernel and only the
data path is implemented in the userspace.
This looks interesting. Unfortunately I don't have enough time to do a
full review, but I looked at the documentation and uapi header file to
give feedback on the userspace ABI.
OK. Thanks for your comments. It's helpful!
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-29 03:29:52
在 2021/6/29 上午10:26, Yongji Xie 写道:
On Mon, Jun 28, 2021 at 12:40 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/25 下午12:19, Yongji Xie 写道:
quoted
quoted
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
Looks good to me. And I think we can use the reply of the message to
update the status instead of introducing a new command.
Just notice this part in virtio_finalize_features():
virtio_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
status = dev->config->get_status(dev);
if (!(status & VIRTIO_CONFIG_S_FEATURES_OK)) {
So we no reply doesn't work for FEATURES_OK.
So my understanding is:
1) We must not use noreply for set_status()
2) We can use noreply for get_status(), but it requires a new ioctl to
update the status.
So it looks to me we need synchronize for both get_status() and
set_status().
We should not send messages to userspace in the FEATURES_OK case. So
the synchronization is not necessary.
As discussed previously, there could be a device that mandates some
features (VIRTIO_F_RING_PACKED). So it can choose to not accept
FEATURES_OK is packed virtqueue is not negotiated.
In this case we need to relay the message to userspace.
Thanks
On Tue, Jun 29, 2021 at 11:29 AM Jason Wang [off-list ref] wrote:
在 2021/6/29 上午10:26, Yongji Xie 写道:
quoted
On Mon, Jun 28, 2021 at 12:40 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/25 下午12:19, Yongji Xie 写道:
quoted
quoted
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
Looks good to me. And I think we can use the reply of the message to
update the status instead of introducing a new command.
Just notice this part in virtio_finalize_features():
virtio_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
status = dev->config->get_status(dev);
if (!(status & VIRTIO_CONFIG_S_FEATURES_OK)) {
So we no reply doesn't work for FEATURES_OK.
So my understanding is:
1) We must not use noreply for set_status()
2) We can use noreply for get_status(), but it requires a new ioctl to
update the status.
So it looks to me we need synchronize for both get_status() and
set_status().
We should not send messages to userspace in the FEATURES_OK case. So
the synchronization is not necessary.
As discussed previously, there could be a device that mandates some
features (VIRTIO_F_RING_PACKED). So it can choose to not accept
FEATURES_OK is packed virtqueue is not negotiated.
In this case we need to relay the message to userspace.
OK, I see. If so, I prefer to only use noreply for set_status(). We do
not set the status bit if the message is failed. In this way, we don't
need to change lots of virtio core codes to handle the failure of
set_status()/get_status().
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-29 04:03:52
在 2021/6/29 上午11:56, Yongji Xie 写道:
On Tue, Jun 29, 2021 at 11:29 AM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/29 上午10:26, Yongji Xie 写道:
quoted
On Mon, Jun 28, 2021 at 12:40 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/25 下午12:19, Yongji Xie 写道:
quoted
quoted
2b) for set_status(): simply relay the message to userspace, reply is no
needed. Userspace will use a command to update the status when the
datapath is stop. The the status could be fetched via get_stats().
2b looks more spec complaint.
Looks good to me. And I think we can use the reply of the message to
update the status instead of introducing a new command.
Just notice this part in virtio_finalize_features():
virtio_add_status(dev, VIRTIO_CONFIG_S_FEATURES_OK);
status = dev->config->get_status(dev);
if (!(status & VIRTIO_CONFIG_S_FEATURES_OK)) {
So we no reply doesn't work for FEATURES_OK.
So my understanding is:
1) We must not use noreply for set_status()
2) We can use noreply for get_status(), but it requires a new ioctl to
update the status.
So it looks to me we need synchronize for both get_status() and
set_status().
We should not send messages to userspace in the FEATURES_OK case. So
the synchronization is not necessary.
As discussed previously, there could be a device that mandates some
features (VIRTIO_F_RING_PACKED). So it can choose to not accept
FEATURES_OK is packed virtqueue is not negotiated.
In this case we need to relay the message to userspace.
OK, I see. If so, I prefer to only use noreply for set_status(). We do
not set the status bit if the message is failed. In this way, we don't
need to change lots of virtio core codes to handle the failure of
set_status()/get_status().
From: Jason Wang <hidden> Date: 2021-06-29 04:11:19
在 2021/6/28 下午1:54, Liu, Xiaodong 写道:
quoted
Several issues:
- VDUSE needs to limit the total size of the bounce buffers (64M if I was not
wrong). Does it work for SPDK?
Yes, Jason. It is enough and works for SPDK.
Since it's a kind of bounce buffer mainly for in-flight IO, so limited size like
64MB is enough.
Ok.
quoted
- VDUSE can use hugepages but I'm not sure we can mandate hugepages (or we
need introduce new flags for supporting this)
Same with your worry, I'm afraid too that it is a hard for a kernel module
to directly preallocate hugepage internal.
What I tried is that:
1. A simple agent daemon (represents for one device) `preallocates` and maps
dozens of 2MB hugepages (like 64MB) for one device.
2. The daemon passes its mapping addr&len and hugepage fd to kernel
module through created IOCTL.
3. Kernel module remaps the hugepages inside kernel.
Such model should work, but the main "issue" is that it introduce
overheads in the case of vhost-vDPA.
Note that in the case of vhost-vDPA, we don't use bounce buffer, the
userspace pages were shared directly.
And since DMA is not done per page, it prevents us from using tricks
like vm_insert_page() in those cases.
4. Vhost user target gets and maps hugepage fd from kernel module
in vhost-user msg through Unix Domain Socket cmsg.
Then kernel module and target map on the same hugepage based
bounce buffer for in-flight IO.
If there is one option in VDUSE to map userspace preallocated memory, then
VDUSE should be able to mandate it even it is hugepage based.
As above, this requires some kind of re-design since VDUSE depends on
the model of mmap(MAP_SHARED) instead of umem registering.
Thanks
From: Jason Wang <hidden> Date: 2021-06-29 04:13:13
在 2021/6/28 下午6:32, Yongji Xie 写道:
quoted
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer,
then SPDK could also be a potential user of vduse.
I think we can support registering user space memory for bounce-buffer
use like XDP does. But this needs to pin the pages, so I didn't
consider it in this initial version.
Note that userspace should be unaware of the existence of the bounce buffer.
So we need to think carefully of mmap() vs umem registering.
Thanks
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
On Tue, Jun 15, 2021 at 10:13:31PM +0800, Xie Yongji wrote:
quoted
VDUSE (vDPA Device in Userspace) is a framework to support
implementing software-emulated vDPA devices in userspace. This
document is intended to clarify the VDUSE design and usage.
Signed-off-by: Xie Yongji <redacted>
---
Documentation/userspace-api/index.rst | 1 +
Documentation/userspace-api/vduse.rst | 222 ++++++++++++++++++++++++++++++++++
2 files changed, 223 insertions(+)
create mode 100644 Documentation/userspace-api/vduse.rst
@@ -0,0 +1,222 @@+==================================+VDUSE - "vDPA Device in Userspace"+==================================++vDPA (virtio data path acceleration) device is a device that uses a+datapath which complies with the virtio specifications with vendor+specific control path. vDPA devices can be both physically located on+the hardware or emulated by software. VDUSE is a framework that makes it+possible to implement software-emulated vDPA devices in userspace. And+to make it simple, the emulated vDPA device's control path is handled in+the kernel and only the data path is implemented in the userspace.++Note that only virtio block device is supported by VDUSE framework now,+which can reduce security risks when the userspace process that implements+the data path is run by an unprivileged user. The Support for other device+types can be added after the security issue is clarified or fixed in the future.++Start/Stop VDUSE devices+------------------------++VDUSE devices are started as follows:++1. Create a new VDUSE instance with ioctl(VDUSE_CREATE_DEV) on+ /dev/vduse/control.++2. Begin processing VDUSE messages from /dev/vduse/$NAME. The first+ messages will arrive while attaching the VDUSE instance to vDPA bus.++3. Send the VDPA_CMD_DEV_NEW netlink message to attach the VDUSE+ instance to vDPA bus.++VDUSE devices are stopped as follows:++1. Send the VDPA_CMD_DEV_DEL netlink message to detach the VDUSE+ instance from vDPA bus.++2. Close the file descriptor referring to /dev/vduse/$NAME++3. Destroy the VDUSE instance with ioctl(VDUSE_DESTROY_DEV) on+ /dev/vduse/control++The netlink messages metioned above can be sent via vdpa tool in iproute2+or use the below sample codes:++..code-block:: c++ static int netlink_add_vduse(const char *name, enum vdpa_command cmd)+ {+ struct nl_sock *nlsock;+ struct nl_msg *msg;+ int famid;++ nlsock = nl_socket_alloc();+ if (!nlsock)+ return -ENOMEM;++ if (genl_connect(nlsock))+ goto free_sock;++ famid = genl_ctrl_resolve(nlsock, VDPA_GENL_NAME);+ if (famid < 0)+ goto close_sock;++ msg = nlmsg_alloc();+ if (!msg)+ goto close_sock;++ if (!genlmsg_put(msg, NL_AUTO_PORT, NL_AUTO_SEQ, famid, 0, 0, cmd, 0))+ goto nla_put_failure;++ NLA_PUT_STRING(msg, VDPA_ATTR_DEV_NAME, name);+ if (cmd == VDPA_CMD_DEV_NEW)+ NLA_PUT_STRING(msg, VDPA_ATTR_MGMTDEV_DEV_NAME, "vduse");++ if (nl_send_sync(nlsock, msg))+ goto close_sock;++ nl_close(nlsock);+ nl_socket_free(nlsock);++ return 0;+ nla_put_failure:+ nlmsg_free(msg);+ close_sock:+ nl_close(nlsock);+ free_sock:+ nl_socket_free(nlsock);+ return -1;+ }++How VDUSE works+---------------++Since the emuldated vDPA device's control path is handled in the kernel,
s/emuldated/emulated/
Will fix it.
quoted
+a message-based communication protocol and few types of control messages
+are introduced by VDUSE framework to make userspace be aware of the data
+path related changes:
+
+- VDUSE_GET_VQ_STATE: Get the state for virtqueue from userspace
+
+- VDUSE_START_DATAPLANE: Notify userspace to start the dataplane
+
+- VDUSE_STOP_DATAPLANE: Notify userspace to stop the dataplane
+
+- VDUSE_UPDATE_IOTLB: Notify userspace to update the memory mapping in device IOTLB
+
+Userspace needs to read()/write() on /dev/vduse/$NAME to receive/reply
+those control messages from/to VDUSE kernel module as follows:
+
+.. code-block:: c
+
+ static int vduse_message_handler(int dev_fd)
+ {
+ int len;
+ struct vduse_dev_request req;
+ struct vduse_dev_response resp;
+
+ len = read(dev_fd, &req, sizeof(req));
+ if (len != sizeof(req))
+ return -1;
+
+ resp.request_id = req.request_id;
+
+ switch (req.type) {
+
+ /* handle different types of message */
+
+ }
+
+ if (req.flags & VDUSE_REQ_FLAGS_NO_REPLY)
+ return 0;
+
+ len = write(dev_fd, &resp, sizeof(resp));
+ if (len != sizeof(resp))
+ return -1;
+
+ return 0;
+ }
+
+After VDUSE_START_DATAPLANE messages is received, userspace should start the
+dataplane processing with the help of some ioctls on /dev/vduse/$NAME:
+
+- VDUSE_IOTLB_GET_FD: get the file descriptor to the first overlapped iova region.
+ Userspace can access this iova region by passing fd and corresponding size, offset,
+ perm to mmap(). For example:
+
+.. code-block:: c
+
+ static int perm_to_prot(uint8_t perm)
+ {
+ int prot = 0;
+
+ switch (perm) {
+ case VDUSE_ACCESS_WO:
+ prot |= PROT_WRITE;
+ break;
+ case VDUSE_ACCESS_RO:
+ prot |= PROT_READ;
+ break;
+ case VDUSE_ACCESS_RW:
+ prot |= PROT_READ | PROT_WRITE;
+ break;
+ }
+
+ return prot;
+ }
+
+ static void *iova_to_va(int dev_fd, uint64_t iova, uint64_t *len)
+ {
+ int fd;
+ void *addr;
+ size_t size;
+ struct vduse_iotlb_entry entry;
+
+ entry.start = iova;
+ entry.last = iova + 1;
Why +1?
I expected the request to include *len so that VDUSE can create a bounce
buffer for the full iova range, if necessary.
The function is used to translate iova to va. And the *len is not
specified by the caller. Instead, it's used to tell the caller the
length of the contiguous iova region from the specified iova. And the
ioctl VDUSE_IOTLB_GET_FD will get the file descriptor to the first
overlapped iova region. So using iova + 1 should be enough here.
How is userspace expected to manage iotlb mmaps? When should munmap(2)
be called?
The simple way is using a list to store the iotlb mappings. And we
should call the munmap(2) for the old mappings when VDUSE_UPDATE_IOTLB
or VDUSE_STOP_DATAPLANE message is received.
Should userspace expect VDUSE_IOTLB_GET_FD to return a full chunk of
guest RAM (e.g. multiple gigabytes) that can be cached permanently or
will it return just enough pages to cover [start, last)?
It should return one iotlb mapping that covers [start, last). In
vhost-vdpa cases, it might be a full chunk of guest RAM. In
virtio-vdpa cases, it might be the whole bounce buffer or one coherent
mapping (produced by dma_alloc_coherent()).
quoted
+
+ return addr + iova - entry.start;
+ }
+
+- VDUSE_DEV_GET_FEATURES: Get the negotiated features
Are these VIRTIO feature bits? Please explain how feature negotiation
works. There must be a way for userspace to report the device's
supported feature bits to the kernel.
Yes, these are VIRTIO feature bits. Userspace will specify the
device's supported feature bits when creating a new VDUSE device with
ioctl(VDUSE_CREATE_DEV).
quoted
+- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt
Does this mean the contents of the configuration space are cached by
VDUSE?
Yes, but the kernel will also store the same contents.
The downside is that the userspace code cannot generate the
contents on demand. Most devices doin't need to generate the contents
on demand, so I think this is okay but I had expected a different
interface:
kernel->userspace VDUSE_DEV_GET_CONFIG
userspace->kernel VDUSE_DEV_INJECT_CONFIG_IRQ
The problem is how to handle the failure of VDUSE_DEV_GET_CONFIG. We
will need lots of modification of virtio codes to support that. So to
make it simple, we choose this way:
userspace -> kernel VDUSE_DEV_SET_CONFIG
userspace -> kernel VDUSE_DEV_INJECT_CONFIG_IRQ
I think you can leave it the way it is, but I wanted to mention this in
case someone thinks it's important to support generating the contents of
the configuration space on demand.
Sorry, I didn't get you here. Can't VDUSE_DEV_SET_CONFIG and
VDUSE_DEV_INJECT_CONFIG_IRQ achieve that?
quoted
+- VDUSE_VQ_GET_INFO: Get the specified virtqueue's metadata
+
+- VDUSE_VQ_SETUP_KICKFD: set the kickfd for virtqueue, this eventfd is used
+ by VDUSE kernel module to notify userspace to consume the vring.
+
+- VDUSE_INJECT_VQ_IRQ: inject an interrupt for specific virtqueue
This information is useful but it's not enough to be able to implement a
userspace device. Please provide more developer documentation or at
least refer to uapi header files, published documents, etc that contain
the details.
OK, I will try to add more details.
Thanks,
Yongji
On Tue, Jun 29, 2021 at 12:13 PM Jason Wang [off-list ref] wrote:
在 2021/6/28 下午6:32, Yongji Xie 写道:
quoted
quoted
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer,
then SPDK could also be a potential user of vduse.
I think we can support registering user space memory for bounce-buffer
use like XDP does. But this needs to pin the pages, so I didn't
consider it in this initial version.
Note that userspace should be unaware of the existence of the bounce buffer.
If so, it might be hard to use umem. Because we can't use umem for
coherent mapping which needs physical address contiguous space.
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-06-29 07:34:07
在 2021/6/29 下午2:40, Yongji Xie 写道:
On Tue, Jun 29, 2021 at 12:13 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/6/28 下午6:32, Yongji Xie 写道:
quoted
quoted
The large barrier is bounce-buffer mapping: SPDK requires hugepages
for NVMe over PCIe and RDMA, so take some preallcoated hugepages to
map as bounce buffer is necessary. Or it's hard to avoid an extra
memcpy from bounce-buffer to hugepage.
If you can add an option to map hugepages as bounce-buffer,
then SPDK could also be a potential user of vduse.
I think we can support registering user space memory for bounce-buffer
use like XDP does. But this needs to pin the pages, so I didn't
consider it in this initial version.
Note that userspace should be unaware of the existence of the bounce buffer.
If so, it might be hard to use umem. Because we can't use umem for
coherent mapping which needs physical address contiguous space.
Thanks,
Yongji
We probably can use umem for memory other than the virtqueue (still via
mmap()).
Thanks
Several issues:
- VDUSE needs to limit the total size of the bounce buffers (64M if I was not
wrong). Does it work for SPDK?
Yes, Jason. It is enough and works for SPDK.
Since it's a kind of bounce buffer mainly for in-flight IO, so limited size like
64MB is enough.
Ok.
quoted
quoted
- VDUSE can use hugepages but I'm not sure we can mandate hugepages (or
we
quoted
quoted
need introduce new flags for supporting this)
Same with your worry, I'm afraid too that it is a hard for a kernel module
to directly preallocate hugepage internal.
What I tried is that:
1. A simple agent daemon (represents for one device) `preallocates` and maps
dozens of 2MB hugepages (like 64MB) for one device.
2. The daemon passes its mapping addr&len and hugepage fd to kernel
module through created IOCTL.
3. Kernel module remaps the hugepages inside kernel.
Such model should work, but the main "issue" is that it introduce
overheads in the case of vhost-vDPA.
Note that in the case of vhost-vDPA, we don't use bounce buffer, the
userspace pages were shared directly.
And since DMA is not done per page, it prevents us from using tricks
like vm_insert_page() in those cases.
Yes, really, it's a problem to handle vhost-vDPA case.
But there are already several solutions to get VM served, like vhost-user,
vfio-user, so at least for SPDK, it won't serve VM through VDUSE. If a user
still want to do that, then the user should tolerate Introduced overhead.
In other words, software backend like SPDK, will appreciate the virtio
datapath of VDUSE to serve local host instead of VM. That's why I also drafted
a "virtio-local" to bridge vhost-user target and local host kernel virtio-blk.
quoted
4. Vhost user target gets and maps hugepage fd from kernel module
in vhost-user msg through Unix Domain Socket cmsg.
Then kernel module and target map on the same hugepage based
bounce buffer for in-flight IO.
If there is one option in VDUSE to map userspace preallocated memory, then
VDUSE should be able to mandate it even it is hugepage based.
As above, this requires some kind of re-design since VDUSE depends on
the model of mmap(MAP_SHARED) instead of umem registering.
Got it, Jason, this may be hard for current version of VDUSE.
Maybe we can consider these options after VDUSE merged later.
Since if VDUSE datapath could be directly leveraged by vhost-user target,
its value will be propagated immediately.
Several issues:
- VDUSE needs to limit the total size of the bounce buffers (64M if I was not
wrong). Does it work for SPDK?
Yes, Jason. It is enough and works for SPDK.
Since it's a kind of bounce buffer mainly for in-flight IO, so limited size like
64MB is enough.
Ok.
quoted
quoted
- VDUSE can use hugepages but I'm not sure we can mandate hugepages (or
we
quoted
quoted
need introduce new flags for supporting this)
Same with your worry, I'm afraid too that it is a hard for a kernel module
to directly preallocate hugepage internal.
What I tried is that:
1. A simple agent daemon (represents for one device) `preallocates` and maps
dozens of 2MB hugepages (like 64MB) for one device.
2. The daemon passes its mapping addr&len and hugepage fd to kernel
module through created IOCTL.
3. Kernel module remaps the hugepages inside kernel.
Such model should work, but the main "issue" is that it introduce
overheads in the case of vhost-vDPA.
Note that in the case of vhost-vDPA, we don't use bounce buffer, the
userspace pages were shared directly.
And since DMA is not done per page, it prevents us from using tricks
like vm_insert_page() in those cases.
Yes, really, it's a problem to handle vhost-vDPA case.
But there are already several solutions to get VM served, like vhost-user,
vfio-user, so at least for SPDK, it won't serve VM through VDUSE. If a user
still want to do that, then the user should tolerate Introduced overhead.
In other words, software backend like SPDK, will appreciate the virtio
datapath of VDUSE to serve local host instead of VM. That's why I also drafted
a "virtio-local" to bridge vhost-user target and local host kernel virtio-blk.
quoted
quoted
4. Vhost user target gets and maps hugepage fd from kernel module
in vhost-user msg through Unix Domain Socket cmsg.
Then kernel module and target map on the same hugepage based
bounce buffer for in-flight IO.
If there is one option in VDUSE to map userspace preallocated memory, then
VDUSE should be able to mandate it even it is hugepage based.
As above, this requires some kind of re-design since VDUSE depends on
the model of mmap(MAP_SHARED) instead of umem registering.
Got it, Jason, this may be hard for current version of VDUSE.
Maybe we can consider these options after VDUSE merged later.
Since if VDUSE datapath could be directly leveraged by vhost-user target,
its value will be propagated immediately.
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-06-30 09:51:56
On Tue, Jun 29, 2021 at 10:59:51AM +0800, Yongji Xie wrote:
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 15, 2021 at 10:13:30PM +0800, Xie Yongji wrote:
quoted
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
Good point! But it looks like virtio-vdpa/virtio-pci doesn't support
that now. All virtqueues have the same maximum size.
I see struct vpda_config_ops only supports a per-device max vq size:
u16 (*get_vq_num_max)(struct vdpa_device *vdev);
virtio-pci supports per-virtqueue sizes because the struct
virtio_pci_common_cfg->queue_size register is per-queue (controlled by
queue_select).
I guess this is a question for Jason: will vdpa will keep this limitation?
If yes, then VDUSE can stick to it too without running into problems in
the future.
quoted
quoted
+ __u16 padding; /* padding */
+ __u32 vq_num; /* the number of virtqueues */
+ __u32 vq_align; /* the allocation alignment of virtqueue's metadata */
I'm not sure what this is?
This will be used by vring_create_virtqueue() too.
If there is no official definition for the meaning of this value then
"/* same as vring_create_virtqueue()'s vring_align parameter */" would
be clearer. That way the reader knows what to research in order to
understand how this field works.
I don't remember but maybe it was used to support vrings when the
host/guest have non-4KB page sizes. I wonder if anyone has an official
definition for this value?
Why +1?
I expected the request to include *len so that VDUSE can create a bounce
buffer for the full iova range, if necessary.
The function is used to translate iova to va. And the *len is not
specified by the caller. Instead, it's used to tell the caller the
length of the contiguous iova region from the specified iova. And the
ioctl VDUSE_IOTLB_GET_FD will get the file descriptor to the first
overlapped iova region. So using iova + 1 should be enough here.
Does the entry.last field have any purpose with VDUSE_IOTLB_GET_FD? I
wonder why userspace needs to assign a value at all if it's always +1.
How is userspace expected to manage iotlb mmaps? When should munmap(2)
be called?
The simple way is using a list to store the iotlb mappings. And we
should call the munmap(2) for the old mappings when VDUSE_UPDATE_IOTLB
or VDUSE_STOP_DATAPLANE message is received.
Thanks for explaining. It would be helpful to have a description of
IOTLB operation in this document.
quoted
Should userspace expect VDUSE_IOTLB_GET_FD to return a full chunk of
guest RAM (e.g. multiple gigabytes) that can be cached permanently or
will it return just enough pages to cover [start, last)?
It should return one iotlb mapping that covers [start, last). In
vhost-vdpa cases, it might be a full chunk of guest RAM. In
virtio-vdpa cases, it might be the whole bounce buffer or one coherent
mapping (produced by dma_alloc_coherent()).
Great, thanks. Adding something about this to the documentation would
help others implementing VDUSE devices or libraries.
quoted
quoted
+
+ return addr + iova - entry.start;
+ }
+
+- VDUSE_DEV_GET_FEATURES: Get the negotiated features
Are these VIRTIO feature bits? Please explain how feature negotiation
works. There must be a way for userspace to report the device's
supported feature bits to the kernel.
Yes, these are VIRTIO feature bits. Userspace will specify the
device's supported feature bits when creating a new VDUSE device with
ioctl(VDUSE_CREATE_DEV).
Can the VDUSE device influence feature bit negotiation? For example, if
the VDUSE virtio-blk device does not implement discard/write-zeroes, how
does QEMU or the guest find out about this?
quoted
quoted
+- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt
Does this mean the contents of the configuration space are cached by
VDUSE?
Yes, but the kernel will also store the same contents.
quoted
The downside is that the userspace code cannot generate the
contents on demand. Most devices doin't need to generate the contents
on demand, so I think this is okay but I had expected a different
interface:
kernel->userspace VDUSE_DEV_GET_CONFIG
userspace->kernel VDUSE_DEV_INJECT_CONFIG_IRQ
The problem is how to handle the failure of VDUSE_DEV_GET_CONFIG. We
will need lots of modification of virtio codes to support that. So to
make it simple, we choose this way:
userspace -> kernel VDUSE_DEV_SET_CONFIG
userspace -> kernel VDUSE_DEV_INJECT_CONFIG_IRQ
quoted
I think you can leave it the way it is, but I wanted to mention this in
case someone thinks it's important to support generating the contents of
the configuration space on demand.
Sorry, I didn't get you here. Can't VDUSE_DEV_SET_CONFIG and
VDUSE_DEV_INJECT_CONFIG_IRQ achieve that?
If the contents of the configuration space change continuously, then the
VDUSE_DEV_SET_CONFIG approach is inefficient and might have race
conditions. For example, imagine a device where the driver can read a
timer from the configuration space. I think the VIRTIO device model
allows that although I'm not aware of any devices that do something like
it today. The problem is that VDUSE_DEV_SET_CONFIG would have to be
called frequently to keep the timer value updated even though the guest
driver probably isn't accessing it.
What's worse is that there might be race conditions where other
driver->device operations are supposed to update the configuration space
but VDUSE_DEV_SET_CONFIG means that the VDUSE kernel code is caching an
outdated copy.
Again, I don't think it's a problem for existing devices in the VIRTIO
specification. But I'm not 100% sure and future devices might require
what I've described, so the VDUSE_DEV_SET_CONFIG interface could become
a problem.
Stefan
On Wed, Jun 30, 2021 at 5:51 PM Stefan Hajnoczi [off-list ref] wrote:
On Tue, Jun 29, 2021 at 10:59:51AM +0800, Yongji Xie wrote:
quoted
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 15, 2021 at 10:13:30PM +0800, Xie Yongji wrote:
quoted
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
Good point! But it looks like virtio-vdpa/virtio-pci doesn't support
that now. All virtqueues have the same maximum size.
I see struct vpda_config_ops only supports a per-device max vq size:
u16 (*get_vq_num_max)(struct vdpa_device *vdev);
virtio-pci supports per-virtqueue sizes because the struct
virtio_pci_common_cfg->queue_size register is per-queue (controlled by
queue_select).
Oh, yes. I miss queue_select.
I guess this is a question for Jason: will vdpa will keep this limitation?
If yes, then VDUSE can stick to it too without running into problems in
the future.
quoted
quoted
quoted
+ __u16 padding; /* padding */
+ __u32 vq_num; /* the number of virtqueues */
+ __u32 vq_align; /* the allocation alignment of virtqueue's metadata */
I'm not sure what this is?
This will be used by vring_create_virtqueue() too.
If there is no official definition for the meaning of this value then
"/* same as vring_create_virtqueue()'s vring_align parameter */" would
be clearer. That way the reader knows what to research in order to
understand how this field works.
OK.
I don't remember but maybe it was used to support vrings when the
host/guest have non-4KB page sizes. I wonder if anyone has an official
definition for this value?
Not sure. Maybe we might need some alignment which is less than
PAGE_SIZE sometimes.
Thanks,
Yongji
From: Jason Wang <hidden> Date: 2021-07-01 07:56:01
在 2021/7/1 下午2:50, Yongji Xie 写道:
On Wed, Jun 30, 2021 at 5:51 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 29, 2021 at 10:59:51AM +0800, Yongji Xie wrote:
quoted
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 15, 2021 at 10:13:30PM +0800, Xie Yongji wrote:
quoted
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
Good point! But it looks like virtio-vdpa/virtio-pci doesn't support
that now. All virtqueues have the same maximum size.
I see struct vpda_config_ops only supports a per-device max vq size:
u16 (*get_vq_num_max)(struct vdpa_device *vdev);
virtio-pci supports per-virtqueue sizes because the struct
virtio_pci_common_cfg->queue_size register is per-queue (controlled by
queue_select).
Oh, yes. I miss queue_select.
quoted
I guess this is a question for Jason: will vdpa will keep this limitation?
If yes, then VDUSE can stick to it too without running into problems in
the future.
I think it's better to extend the get_vq_num_max() per virtqueue.
Currently, vDPA assumes the parent to have a global max size. This seems
to work on most of the parents but not vp-vDPA (which could be backed by
QEMU, in that case cvq's size is smaller).
Fortunately, we haven't enabled had cvq support in the userspace now.
I can post the fixes.
quoted
quoted
quoted
quoted
+ __u16 padding; /* padding */
+ __u32 vq_num; /* the number of virtqueues */
+ __u32 vq_align; /* the allocation alignment of virtqueue's metadata */
I'm not sure what this is?
This will be used by vring_create_virtqueue() too.
If there is no official definition for the meaning of this value then
"/* same as vring_create_virtqueue()'s vring_align parameter */" would
be clearer. That way the reader knows what to research in order to
understand how this field works.
OK.
quoted
I don't remember but maybe it was used to support vrings when the
host/guest have non-4KB page sizes. I wonder if anyone has an official
definition for this value?
Not sure. Maybe we might need some alignment which is less than
PAGE_SIZE sometimes.
So I see CCW always use 4096, but I'm not sure whether or not it's
smaller than PAGE_SIZE.
Thanks
Why +1?
I expected the request to include *len so that VDUSE can create a bounce
buffer for the full iova range, if necessary.
The function is used to translate iova to va. And the *len is not
specified by the caller. Instead, it's used to tell the caller the
length of the contiguous iova region from the specified iova. And the
ioctl VDUSE_IOTLB_GET_FD will get the file descriptor to the first
overlapped iova region. So using iova + 1 should be enough here.
Does the entry.last field have any purpose with VDUSE_IOTLB_GET_FD? I
wonder why userspace needs to assign a value at all if it's always +1.
If we need to get some iova regions in the specified range, we need
the entry.last field. For example, we can use [0, ULONG_MAX] to get
the first overlapped iova region which might be [4096, 8192]. But in
this function, we don't use VDUSE_IOTLB_GET_FD like this. We need to
get the iova region including the specified iova.
How is userspace expected to manage iotlb mmaps? When should munmap(2)
be called?
The simple way is using a list to store the iotlb mappings. And we
should call the munmap(2) for the old mappings when VDUSE_UPDATE_IOTLB
or VDUSE_STOP_DATAPLANE message is received.
Thanks for explaining. It would be helpful to have a description of
IOTLB operation in this document.
Sure.
quoted
quoted
Should userspace expect VDUSE_IOTLB_GET_FD to return a full chunk of
guest RAM (e.g. multiple gigabytes) that can be cached permanently or
will it return just enough pages to cover [start, last)?
It should return one iotlb mapping that covers [start, last). In
vhost-vdpa cases, it might be a full chunk of guest RAM. In
virtio-vdpa cases, it might be the whole bounce buffer or one coherent
mapping (produced by dma_alloc_coherent()).
Great, thanks. Adding something about this to the documentation would
help others implementing VDUSE devices or libraries.
OK.
quoted
quoted
quoted
+
+ return addr + iova - entry.start;
+ }
+
+- VDUSE_DEV_GET_FEATURES: Get the negotiated features
Are these VIRTIO feature bits? Please explain how feature negotiation
works. There must be a way for userspace to report the device's
supported feature bits to the kernel.
Yes, these are VIRTIO feature bits. Userspace will specify the
device's supported feature bits when creating a new VDUSE device with
ioctl(VDUSE_CREATE_DEV).
Can the VDUSE device influence feature bit negotiation? For example, if
the VDUSE virtio-blk device does not implement discard/write-zeroes, how
does QEMU or the guest find out about this?
There is a "features" field in struct vduse_dev_config which is used
to do feature negotiation.
quoted
quoted
quoted
+- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt
Does this mean the contents of the configuration space are cached by
VDUSE?
Yes, but the kernel will also store the same contents.
quoted
The downside is that the userspace code cannot generate the
contents on demand. Most devices doin't need to generate the contents
on demand, so I think this is okay but I had expected a different
interface:
kernel->userspace VDUSE_DEV_GET_CONFIG
userspace->kernel VDUSE_DEV_INJECT_CONFIG_IRQ
The problem is how to handle the failure of VDUSE_DEV_GET_CONFIG. We
will need lots of modification of virtio codes to support that. So to
make it simple, we choose this way:
userspace -> kernel VDUSE_DEV_SET_CONFIG
userspace -> kernel VDUSE_DEV_INJECT_CONFIG_IRQ
quoted
I think you can leave it the way it is, but I wanted to mention this in
case someone thinks it's important to support generating the contents of
the configuration space on demand.
Sorry, I didn't get you here. Can't VDUSE_DEV_SET_CONFIG and
VDUSE_DEV_INJECT_CONFIG_IRQ achieve that?
If the contents of the configuration space change continuously, then the
VDUSE_DEV_SET_CONFIG approach is inefficient and might have race
conditions. For example, imagine a device where the driver can read a
timer from the configuration space. I think the VIRTIO device model
allows that although I'm not aware of any devices that do something like
it today. The problem is that VDUSE_DEV_SET_CONFIG would have to be
called frequently to keep the timer value updated even though the guest
driver probably isn't accessing it.
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
What's worse is that there might be race conditions where other
driver->device operations are supposed to update the configuration space
but VDUSE_DEV_SET_CONFIG means that the VDUSE kernel code is caching an
outdated copy.
I'm not sure. Should the device and driver be able to access the same
fields concurrently?
Again, I don't think it's a problem for existing devices in the VIRTIO
specification. But I'm not 100% sure and future devices might require
what I've described, so the VDUSE_DEV_SET_CONFIG interface could become
a problem.
If so, maybe a new interface can be added at that time. The
VDUSE_DEV_GET_CONFIG might be better, but I still did not find a good
way for failure handling.
Thanks,
Yongji
On Thu, Jul 1, 2021 at 3:55 PM Jason Wang [off-list ref] wrote:
在 2021/7/1 下午2:50, Yongji Xie 写道:
quoted
On Wed, Jun 30, 2021 at 5:51 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 29, 2021 at 10:59:51AM +0800, Yongji Xie wrote:
quoted
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 15, 2021 at 10:13:30PM +0800, Xie Yongji wrote:
quoted
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
Good point! But it looks like virtio-vdpa/virtio-pci doesn't support
that now. All virtqueues have the same maximum size.
I see struct vpda_config_ops only supports a per-device max vq size:
u16 (*get_vq_num_max)(struct vdpa_device *vdev);
virtio-pci supports per-virtqueue sizes because the struct
virtio_pci_common_cfg->queue_size register is per-queue (controlled by
queue_select).
Oh, yes. I miss queue_select.
quoted
I guess this is a question for Jason: will vdpa will keep this limitation?
If yes, then VDUSE can stick to it too without running into problems in
the future.
I think it's better to extend the get_vq_num_max() per virtqueue.
Currently, vDPA assumes the parent to have a global max size. This seems
to work on most of the parents but not vp-vDPA (which could be backed by
QEMU, in that case cvq's size is smaller).
Fortunately, we haven't enabled had cvq support in the userspace now.
I can post the fixes.
OK. If so, it looks like we need to support the per-vq configuration.
I wonder if it's better to use something like: VDUSE_CREATE_DEVICE ->
VDUSE_SETUP_VQ -> VDUSE_SETUP_VQ -> ... -> VDUSE_ENABLE_DEVICE to do
initialization rather than only use VDUSE_CREATE_DEVICE.
Thanks,
Yongji
Why +1?
I expected the request to include *len so that VDUSE can create a bounce
buffer for the full iova range, if necessary.
The function is used to translate iova to va. And the *len is not
specified by the caller. Instead, it's used to tell the caller the
length of the contiguous iova region from the specified iova. And the
ioctl VDUSE_IOTLB_GET_FD will get the file descriptor to the first
overlapped iova region. So using iova + 1 should be enough here.
Does the entry.last field have any purpose with VDUSE_IOTLB_GET_FD? I
wonder why userspace needs to assign a value at all if it's always +1.
If we need to get some iova regions in the specified range, we need
the entry.last field. For example, we can use [0, ULONG_MAX] to get
the first overlapped iova region which might be [4096, 8192]. But in
this function, we don't use VDUSE_IOTLB_GET_FD like this. We need to
get the iova region including the specified iova.
I see, thanks for explaining!
quoted
quoted
quoted
quoted
+ return addr + iova - entry.start;
+ }
+
+- VDUSE_DEV_GET_FEATURES: Get the negotiated features
Are these VIRTIO feature bits? Please explain how feature negotiation
works. There must be a way for userspace to report the device's
supported feature bits to the kernel.
Yes, these are VIRTIO feature bits. Userspace will specify the
device's supported feature bits when creating a new VDUSE device with
ioctl(VDUSE_CREATE_DEV).
Can the VDUSE device influence feature bit negotiation? For example, if
the VDUSE virtio-blk device does not implement discard/write-zeroes, how
does QEMU or the guest find out about this?
There is a "features" field in struct vduse_dev_config which is used
to do feature negotiation.
This approach is more restrictive than required by the VIRTIO
specification:
"The device SHOULD accept any valid subset of features the driver
accepts, otherwise it MUST fail to set the FEATURES_OK device status
bit when the driver writes it."
https://docs.oasis-open.org/virtio/virtio/v1.1/cs01/virtio-v1.1-cs01.html#x1-130002
The spec allows a device to reject certain subsets of features. For
example, if feature B depends on feature A and can only be enabled when
feature A is also enabled.
From your description I think VDUSE would accept feature B without
feature A since the device implementation has no opportunity to fail
negotiation with custom logic.
Ideally VDUSE would send a SET_FEATURES message to userspace, allowing
the device implementation full flexibility in which subsets of features
to accept.
This is a corner case. Many or maybe even all existing VIRTIO devices
don't need this flexibility, but I want to point out this limitation in
the VDUSE interface because it may cause issues in the future.
quoted
quoted
quoted
quoted
+- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt
Does this mean the contents of the configuration space are cached by
VDUSE?
Yes, but the kernel will also store the same contents.
quoted
The downside is that the userspace code cannot generate the
contents on demand. Most devices doin't need to generate the contents
on demand, so I think this is okay but I had expected a different
interface:
kernel->userspace VDUSE_DEV_GET_CONFIG
userspace->kernel VDUSE_DEV_INJECT_CONFIG_IRQ
The problem is how to handle the failure of VDUSE_DEV_GET_CONFIG. We
will need lots of modification of virtio codes to support that. So to
make it simple, we choose this way:
userspace -> kernel VDUSE_DEV_SET_CONFIG
userspace -> kernel VDUSE_DEV_INJECT_CONFIG_IRQ
quoted
I think you can leave it the way it is, but I wanted to mention this in
case someone thinks it's important to support generating the contents of
the configuration space on demand.
Sorry, I didn't get you here. Can't VDUSE_DEV_SET_CONFIG and
VDUSE_DEV_INJECT_CONFIG_IRQ achieve that?
If the contents of the configuration space change continuously, then the
VDUSE_DEV_SET_CONFIG approach is inefficient and might have race
conditions. For example, imagine a device where the driver can read a
timer from the configuration space. I think the VIRTIO device model
allows that although I'm not aware of any devices that do something like
it today. The problem is that VDUSE_DEV_SET_CONFIG would have to be
called frequently to keep the timer value updated even though the guest
driver probably isn't accessing it.
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
quoted
What's worse is that there might be race conditions where other
driver->device operations are supposed to update the configuration space
but VDUSE_DEV_SET_CONFIG means that the VDUSE kernel code is caching an
outdated copy.
I'm not sure. Should the device and driver be able to access the same
fields concurrently?
Again, I don't think it's a problem for existing devices in the VIRTIO
specification. But I'm not 100% sure and future devices might require
what I've described, so the VDUSE_DEV_SET_CONFIG interface could become
a problem.
If so, maybe a new interface can be added at that time. The
VDUSE_DEV_GET_CONFIG might be better, but I still did not find a good
way for failure handling.
I'm not aware of the details of why the current approach was necessary,
so I don't have any concrete suggestions. Sorry!
Stefan
From: Jason Wang <hidden> Date: 2021-07-02 03:25:30
在 2021/7/1 下午6:26, Yongji Xie 写道:
On Thu, Jul 1, 2021 at 3:55 PM Jason Wang [off-list ref] wrote:
quoted
在 2021/7/1 下午2:50, Yongji Xie 写道:
quoted
On Wed, Jun 30, 2021 at 5:51 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 29, 2021 at 10:59:51AM +0800, Yongji Xie wrote:
quoted
On Mon, Jun 28, 2021 at 9:02 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jun 15, 2021 at 10:13:30PM +0800, Xie Yongji wrote:
quoted
+/* ioctls */
+
+struct vduse_dev_config {
+ char name[VDUSE_NAME_MAX]; /* vduse device name */
+ __u32 vendor_id; /* virtio vendor id */
+ __u32 device_id; /* virtio device id */
+ __u64 features; /* device features */
+ __u64 bounce_size; /* bounce buffer size for iommu */
+ __u16 vq_size_max; /* the max size of virtqueue */
The VIRTIO specification allows per-virtqueue sizes. A device can have
two virtqueues, where the first one allows up to 1024 descriptors and
the second one allows only 128 descriptors, for example.
Good point! But it looks like virtio-vdpa/virtio-pci doesn't support
that now. All virtqueues have the same maximum size.
I see struct vpda_config_ops only supports a per-device max vq size:
u16 (*get_vq_num_max)(struct vdpa_device *vdev);
virtio-pci supports per-virtqueue sizes because the struct
virtio_pci_common_cfg->queue_size register is per-queue (controlled by
queue_select).
Oh, yes. I miss queue_select.
quoted
I guess this is a question for Jason: will vdpa will keep this limitation?
If yes, then VDUSE can stick to it too without running into problems in
the future.
I think it's better to extend the get_vq_num_max() per virtqueue.
Currently, vDPA assumes the parent to have a global max size. This seems
to work on most of the parents but not vp-vDPA (which could be backed by
QEMU, in that case cvq's size is smaller).
Fortunately, we haven't enabled had cvq support in the userspace now.
I can post the fixes.
OK. If so, it looks like we need to support the per-vq configuration.
I wonder if it's better to use something like: VDUSE_CREATE_DEVICE ->
VDUSE_SETUP_VQ -> VDUSE_SETUP_VQ -> ... -> VDUSE_ENABLE_DEVICE to do
initialization rather than only use VDUSE_CREATE_DEVICE.
Why +1?
I expected the request to include *len so that VDUSE can create a bounce
buffer for the full iova range, if necessary.
The function is used to translate iova to va. And the *len is not
specified by the caller. Instead, it's used to tell the caller the
length of the contiguous iova region from the specified iova. And the
ioctl VDUSE_IOTLB_GET_FD will get the file descriptor to the first
overlapped iova region. So using iova + 1 should be enough here.
Does the entry.last field have any purpose with VDUSE_IOTLB_GET_FD? I
wonder why userspace needs to assign a value at all if it's always +1.
If we need to get some iova regions in the specified range, we need
the entry.last field. For example, we can use [0, ULONG_MAX] to get
the first overlapped iova region which might be [4096, 8192]. But in
this function, we don't use VDUSE_IOTLB_GET_FD like this. We need to
get the iova region including the specified iova.
I see, thanks for explaining!
quoted
quoted
quoted
quoted
quoted
+ return addr + iova - entry.start;
+ }
+
+- VDUSE_DEV_GET_FEATURES: Get the negotiated features
Are these VIRTIO feature bits? Please explain how feature negotiation
works. There must be a way for userspace to report the device's
supported feature bits to the kernel.
Yes, these are VIRTIO feature bits. Userspace will specify the
device's supported feature bits when creating a new VDUSE device with
ioctl(VDUSE_CREATE_DEV).
Can the VDUSE device influence feature bit negotiation? For example, if
the VDUSE virtio-blk device does not implement discard/write-zeroes, how
does QEMU or the guest find out about this?
There is a "features" field in struct vduse_dev_config which is used
to do feature negotiation.
This approach is more restrictive than required by the VIRTIO
specification:
"The device SHOULD accept any valid subset of features the driver
accepts, otherwise it MUST fail to set the FEATURES_OK device status
bit when the driver writes it."
https://docs.oasis-open.org/virtio/virtio/v1.1/cs01/virtio-v1.1-cs01.html#x1-130002
The spec allows a device to reject certain subsets of features. For
example, if feature B depends on feature A and can only be enabled when
feature A is also enabled.
From your description I think VDUSE would accept feature B without
feature A since the device implementation has no opportunity to fail
negotiation with custom logic.
Yes, we discussed it [1] before. So I'd like to re-introduce
SET_STATUS messages so that the userspace can fail feature negotiation
during setting FEATURES_OK status bit.
[1] https://lkml.org/lkml/2021/6/28/1587
Ideally VDUSE would send a SET_FEATURES message to userspace, allowing
the device implementation full flexibility in which subsets of features
to accept.
This is a corner case. Many or maybe even all existing VIRTIO devices
don't need this flexibility, but I want to point out this limitation in
the VDUSE interface because it may cause issues in the future.
quoted
quoted
quoted
quoted
quoted
+- VDUSE_DEV_UPDATE_CONFIG: Update the configuration space and inject a config interrupt
Does this mean the contents of the configuration space are cached by
VDUSE?
Yes, but the kernel will also store the same contents.
quoted
The downside is that the userspace code cannot generate the
contents on demand. Most devices doin't need to generate the contents
on demand, so I think this is okay but I had expected a different
interface:
kernel->userspace VDUSE_DEV_GET_CONFIG
userspace->kernel VDUSE_DEV_INJECT_CONFIG_IRQ
The problem is how to handle the failure of VDUSE_DEV_GET_CONFIG. We
will need lots of modification of virtio codes to support that. So to
make it simple, we choose this way:
userspace -> kernel VDUSE_DEV_SET_CONFIG
userspace -> kernel VDUSE_DEV_INJECT_CONFIG_IRQ
quoted
I think you can leave it the way it is, but I wanted to mention this in
case someone thinks it's important to support generating the contents of
the configuration space on demand.
Sorry, I didn't get you here. Can't VDUSE_DEV_SET_CONFIG and
VDUSE_DEV_INJECT_CONFIG_IRQ achieve that?
If the contents of the configuration space change continuously, then the
VDUSE_DEV_SET_CONFIG approach is inefficient and might have race
conditions. For example, imagine a device where the driver can read a
timer from the configuration space. I think the VIRTIO device model
allows that although I'm not aware of any devices that do something like
it today. The problem is that VDUSE_DEV_SET_CONFIG would have to be
called frequently to keep the timer value updated even though the guest
driver probably isn't accessing it.
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
quoted
quoted
What's worse is that there might be race conditions where other
driver->device operations are supposed to update the configuration space
but VDUSE_DEV_SET_CONFIG means that the VDUSE kernel code is caching an
outdated copy.
I'm not sure. Should the device and driver be able to access the same
fields concurrently?
From: Jason Wang <hidden> Date: 2021-07-05 03:36:36
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space
is not suitable for this kind of error propagating. And it would be very
hard to implement such kind of semantic in some transports. Virtqueue
should be much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
I'd like to stick to the current assumption thich get_config won't fail.
That is to say,
1) maintain a config in the kernel, make sure the config space read can
always succeed
2) introduce an ioctl for the vduse usersapce to update the config space.
3) we can synchronize with the vduse userspace during set_config
Does this work?
Thanks
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-07-05 12:50:15
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
I'd like to stick to the current assumption thich get_config won't fail.
That is to say,
1) maintain a config in the kernel, make sure the config space read can
always succeed
2) introduce an ioctl for the vduse usersapce to update the config space.
3) we can synchronize with the vduse userspace during set_config
Does this work?
I noticed that caching is also allowed by the vhost-user protocol
messages (QEMU's docs/interop/vhost-user.rst), but the device doesn't
know whether or not caching is in effect. The interface you outlined
above requires caching.
Is there a reason why the host kernel vDPA code needs to cache the
configuration space?
Here are the vhost-user protocol messages:
Virtio device config space
^^^^^^^^^^^^^^^^^^^^^^^^^^
+--------+------+-------+---------+
| offset | size | flags | payload |
+--------+------+-------+---------+
:offset: a 32-bit offset of virtio device's configuration space
:size: a 32-bit configuration space access size in bytes
:flags: a 32-bit value:
- 0: Vhost master messages used for writeable fields
- 1: Vhost master messages used for live migration
:payload: Size bytes array holding the contents of the virtio
device's configuration space
...
``VHOST_USER_GET_CONFIG``
:id: 24
:equivalent ioctl: N/A
:master payload: virtio device config space
:slave payload: virtio device config space
When ``VHOST_USER_PROTOCOL_F_CONFIG`` is negotiated, this message is
submitted by the vhost-user master to fetch the contents of the
virtio device configuration space, vhost-user slave's payload size
MUST match master's request, vhost-user slave uses zero length of
payload to indicate an error to vhost-user master. The vhost-user
master may cache the contents to avoid repeated
``VHOST_USER_GET_CONFIG`` calls.
``VHOST_USER_SET_CONFIG``
:id: 25
:equivalent ioctl: N/A
:master payload: virtio device config space
:slave payload: N/A
When ``VHOST_USER_PROTOCOL_F_CONFIG`` is negotiated, this message is
submitted by the vhost-user master when the Guest changes the virtio
device configuration space and also can be used for live migration
on the destination host. The vhost-user slave must check the flags
field, and slaves MUST NOT accept SET_CONFIG for read-only
configuration space fields unless the live migration bit is set.
Stefan
From: Jason Wang <hidden> Date: 2021-07-06 02:34:48
在 2021/7/5 下午8:49, Stefan Hajnoczi 写道:
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
quoted
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
quoted
I'd like to stick to the current assumption thich get_config won't fail.
That is to say,
1) maintain a config in the kernel, make sure the config space read can
always succeed
2) introduce an ioctl for the vduse usersapce to update the config space.
3) we can synchronize with the vduse userspace during set_config
Does this work?
I noticed that caching is also allowed by the vhost-user protocol
messages (QEMU's docs/interop/vhost-user.rst), but the device doesn't
know whether or not caching is in effect. The interface you outlined
above requires caching.
Is there a reason why the host kernel vDPA code needs to cache the
configuration space?
Because:
1) Kernel can not wait forever in get_config(), this is the major
difference with vhost-user.
2) Stick to the current assumption that virtio_cread() should always
succeed.
Thanks
Here are the vhost-user protocol messages:
Virtio device config space
^^^^^^^^^^^^^^^^^^^^^^^^^^
+--------+------+-------+---------+
| offset | size | flags | payload |
+--------+------+-------+---------+
:offset: a 32-bit offset of virtio device's configuration space
:size: a 32-bit configuration space access size in bytes
:flags: a 32-bit value:
- 0: Vhost master messages used for writeable fields
- 1: Vhost master messages used for live migration
:payload: Size bytes array holding the contents of the virtio
device's configuration space
...
``VHOST_USER_GET_CONFIG``
:id: 24
:equivalent ioctl: N/A
:master payload: virtio device config space
:slave payload: virtio device config space
When ``VHOST_USER_PROTOCOL_F_CONFIG`` is negotiated, this message is
submitted by the vhost-user master to fetch the contents of the
virtio device configuration space, vhost-user slave's payload size
MUST match master's request, vhost-user slave uses zero length of
payload to indicate an error to vhost-user master. The vhost-user
master may cache the contents to avoid repeated
``VHOST_USER_GET_CONFIG`` calls.
``VHOST_USER_SET_CONFIG``
:id: 25
:equivalent ioctl: N/A
:master payload: virtio device config space
:slave payload: N/A
When ``VHOST_USER_PROTOCOL_F_CONFIG`` is negotiated, this message is
submitted by the vhost-user master when the Guest changes the virtio
device configuration space and also can be used for live migration
on the destination host. The vhost-user slave must check the flags
field, and slaves MUST NOT accept SET_CONFIG for read-only
configuration space fields unless the live migration bit is set.
Stefan
On Mon, Jul 5, 2021 at 8:50 PM Stefan Hajnoczi [off-list ref] wrote:
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
quoted
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
We add a timeout and return error in case userspace never replies to
the message.
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
I noticed that virtio_config_read*() only returns -1 when we access a
invalid field. But in the VDUSE case, VDUSE_DEV_GET_CONFIG might fail
when we access a valid field. Not sure if it's ok to silently ignore
this kind of error.
Thanks,
Yongji
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-07-06 10:15:04
On Tue, Jul 06, 2021 at 10:34:33AM +0800, Jason Wang wrote:
在 2021/7/5 下午8:49, Stefan Hajnoczi 写道:
quoted
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
quoted
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
quoted
I'd like to stick to the current assumption thich get_config won't fail.
That is to say,
1) maintain a config in the kernel, make sure the config space read can
always succeed
2) introduce an ioctl for the vduse usersapce to update the config space.
3) we can synchronize with the vduse userspace during set_config
Does this work?
I noticed that caching is also allowed by the vhost-user protocol
messages (QEMU's docs/interop/vhost-user.rst), but the device doesn't
know whether or not caching is in effect. The interface you outlined
above requires caching.
Is there a reason why the host kernel vDPA code needs to cache the
configuration space?
Because:
1) Kernel can not wait forever in get_config(), this is the major difference
with vhost-user.
virtio_cread() can sleep:
#define virtio_cread(vdev, structname, member, ptr) \
do { \
typeof(((structname*)0)->member) virtio_cread_v; \
\
might_sleep(); \
^^^^^^^^^^^^^^
Which code path cannot sleep?
2) Stick to the current assumption that virtio_cread() should always
succeed.
That can be done by reading -1 (like QEMU does) when the read fails.
Stefan
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-07-06 10:23:01
On Tue, Jul 06, 2021 at 11:04:18AM +0800, Yongji Xie wrote:
On Mon, Jul 5, 2021 at 8:50 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
quoted
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
We add a timeout and return error in case userspace never replies to
the message.
quoted
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
I noticed that virtio_config_read*() only returns -1 when we access a
invalid field. But in the VDUSE case, VDUSE_DEV_GET_CONFIG might fail
when we access a valid field. Not sure if it's ok to silently ignore
this kind of error.
That's a good point but it's a general VIRTIO issue. Any device
implementation (QEMU userspace, hardware vDPA, etc) can fail, so the
VIRTIO specification needs to provide a way for the driver to detect
this.
If userspace violates the contract then VDUSE needs to mark the device
broken. QEMU's device emulation does something similar with the
vdev->broken flag.
The VIRTIO Device Status field DEVICE_NEEDS_RESET bit can be set by
vDPA/VDUSE to indicate that the device is not operational and must be
reset.
The driver code may still process the -1 value read from the
configuration space. Hopefully this isn't a problem. There is currently
no VIRTIO interface besides DEVICE_NEEDS_RESET to indicate configuration
space access failures. On the other hand, drivers need to handle
malicious devices so they should be able to cope with the -1 value
anyway.
Stefan
On Tue, Jul 6, 2021 at 6:22 PM Stefan Hajnoczi [off-list ref] wrote:
On Tue, Jul 06, 2021 at 11:04:18AM +0800, Yongji Xie wrote:
quoted
On Mon, Jul 5, 2021 at 8:50 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
quoted
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
We add a timeout and return error in case userspace never replies to
the message.
quoted
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
I noticed that virtio_config_read*() only returns -1 when we access a
invalid field. But in the VDUSE case, VDUSE_DEV_GET_CONFIG might fail
when we access a valid field. Not sure if it's ok to silently ignore
this kind of error.
That's a good point but it's a general VIRTIO issue. Any device
implementation (QEMU userspace, hardware vDPA, etc) can fail, so the
VIRTIO specification needs to provide a way for the driver to detect
this.
If userspace violates the contract then VDUSE needs to mark the device
broken. QEMU's device emulation does something similar with the
vdev->broken flag.
The VIRTIO Device Status field DEVICE_NEEDS_RESET bit can be set by
vDPA/VDUSE to indicate that the device is not operational and must be
reset.
It might be a solution. But DEVICE_NEEDS_RESET is not implemented
currently. So I'm thinking whether it's ok to add a check of
DEVICE_NEEDS_RESET status bit in probe function of virtio device
driver (e.g. virtio-blk driver). Then VDUSE can make use of it to fail
device initailization when configuration space access failed.
Thanks,
Yongji
From: Stefan Hajnoczi <stefanha@redhat.com> Date: 2021-07-08 09:07:27
On Wed, Jul 07, 2021 at 05:09:13PM +0800, Yongji Xie wrote:
On Tue, Jul 6, 2021 at 6:22 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Tue, Jul 06, 2021 at 11:04:18AM +0800, Yongji Xie wrote:
quoted
On Mon, Jul 5, 2021 at 8:50 PM Stefan Hajnoczi [off-list ref] wrote:
quoted
On Mon, Jul 05, 2021 at 11:36:15AM +0800, Jason Wang wrote:
quoted
在 2021/7/4 下午5:49, Yongji Xie 写道:
quoted
quoted
quoted
OK, I get you now. Since the VIRTIO specification says "Device
configuration space is generally used for rarely-changing or
initialization-time parameters". I assume the VDUSE_DEV_SET_CONFIG
ioctl should not be called frequently.
The spec uses MUST and other terms to define the precise requirements.
Here the language (especially the word "generally") is weaker and means
there may be exceptions.
Another type of access that doesn't work with the VDUSE_DEV_SET_CONFIG
approach is reads that have side-effects. For example, imagine a field
containing an error code if the device encounters a problem unrelated to
a specific virtqueue request. Reading from this field resets the error
code to 0, saving the driver an extra configuration space write access
and possibly race conditions. It isn't possible to implement those
semantics suing VDUSE_DEV_SET_CONFIG. It's another corner case, but it
makes me think that the interface does not allow full VIRTIO semantics.
Note that though you're correct, my understanding is that config space is
not suitable for this kind of error propagating. And it would be very hard
to implement such kind of semantic in some transports. Virtqueue should be
much better. As Yong Ji quoted, the config space is used for
"rarely-changing or intialization-time parameters".
quoted
Agreed. I will use VDUSE_DEV_GET_CONFIG in the next version. And to
handle the message failure, I'm going to add a return value to
virtio_config_ops.get() and virtio_cread_* API so that the error can
be propagated to the virtio device driver. Then the virtio-blk device
driver can be modified to handle that.
Jason and Stefan, what do you think of this way?
Why does VDUSE_DEV_GET_CONFIG need to support an error return value?
We add a timeout and return error in case userspace never replies to
the message.
quoted
The VIRTIO spec provides no way for the device to report errors from
config space accesses.
The QEMU virtio-pci implementation returns -1 from invalid
virtio_config_read*() and silently discards virtio_config_write*()
accesses.
VDUSE can take the same approach with
VDUSE_DEV_GET_CONFIG/VDUSE_DEV_SET_CONFIG.
I noticed that virtio_config_read*() only returns -1 when we access a
invalid field. But in the VDUSE case, VDUSE_DEV_GET_CONFIG might fail
when we access a valid field. Not sure if it's ok to silently ignore
this kind of error.
That's a good point but it's a general VIRTIO issue. Any device
implementation (QEMU userspace, hardware vDPA, etc) can fail, so the
VIRTIO specification needs to provide a way for the driver to detect
this.
If userspace violates the contract then VDUSE needs to mark the device
broken. QEMU's device emulation does something similar with the
vdev->broken flag.
The VIRTIO Device Status field DEVICE_NEEDS_RESET bit can be set by
vDPA/VDUSE to indicate that the device is not operational and must be
reset.
It might be a solution. But DEVICE_NEEDS_RESET is not implemented
currently. So I'm thinking whether it's ok to add a check of
DEVICE_NEEDS_RESET status bit in probe function of virtio device
driver (e.g. virtio-blk driver). Then VDUSE can make use of it to fail
device initailization when configuration space access failed.