Re: [PATCH v8 00/10] Introduce VDUSE - vDPA Device in Userspace
From: Liu Xiaodong
Date: Sun Jun 27 2021 - 22:55:40 EST
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@xxxxxxxxxxxxx/
> * https://lore.kernel.org/lkml/20210525125622.1203-1-xieyongji@xxxxxxxxxxxxx/
>
> - validating the device config:
>
> * https://lore.kernel.org/lkml/20210615104810.151-1-xieyongji@xxxxxxxxxxxxx/
>
> - validating the device response:
>
> * https://lore.kernel.org/lkml/20210615105218.214-1-xieyongji@xxxxxxxxxxxxx/
>
> 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