[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240626133528.GE2494510@nvidia.com>
Date: Wed, 26 Jun 2024 10:35:28 -0300
From: Jason Gunthorpe <jgg@...dia.com>
To: Yan Zhao <yan.y.zhao@...el.com>
Cc: kvm@...r.kernel.org, linux-kernel@...r.kernel.org,
alex.williamson@...hat.com, kevin.tian@...el.com, peterx@...hat.com,
ajones@...tanamicro.com
Subject: Re: [PATCH] vfio: Reuse file f_inode as vfio device inode
On Mon, Jun 17, 2024 at 05:53:32PM +0800, Yan Zhao wrote:
> Reuse file f_inode as vfio device inode and associate pseudo path file
> directly to inode allocated in vfio fs.
>
> Currently, vfio device is opened via 2 ways:
> 1) via cdev open
> vfio device is opened with a cdev device with file f_inode and address
> space associated with a cdev inode;
> 2) via VFIO_GROUP_GET_DEVICE_FD ioctl
> vfio device is opened via a pseudo path file with file f_inode and
> address space associated with an inode in anon_inode_fs.
>
> In commit b7c5e64fecfa ("vfio: Create vfio_fs_type with inode per device"),
> an inode in vfio fs is allocated for each vfio device. However, this inode
> in vfio fs is only used to assign its address space to that of a file
> associated with another cdev inode or an inode in anon_inode_fs.
>
> This patch
> - reuses cdev device inode as the vfio device inode when it's opened via
> cdev way;
> - allocates an inode in vfio fs, associate it to the pseudo path file,
> and save it as the vfio device inode when the vfio device is opened via
> VFIO_GROUP_GET_DEVICE_FD ioctl.
>
> File address space will then point automatically to the address space of
> the vfio device inode. Tools like unmap_mapping_range() can then zap all
> vmas associated with the vfio device.
>
> Signed-off-by: Yan Zhao <yan.y.zhao@...el.com>
> ---
> drivers/vfio/device_cdev.c | 9 ++++---
> drivers/vfio/group.c | 21 ++--------------
> drivers/vfio/vfio.h | 2 ++
> drivers/vfio/vfio_main.c | 49 +++++++++++++++++++++++++++-----------
> 4 files changed, 43 insertions(+), 38 deletions(-)
>
> diff --git a/drivers/vfio/device_cdev.c b/drivers/vfio/device_cdev.c
> index bb1817bd4ff3..a4eec8e88f5c 100644
> --- a/drivers/vfio/device_cdev.c
> +++ b/drivers/vfio/device_cdev.c
> @@ -40,12 +40,11 @@ int vfio_device_fops_cdev_open(struct inode *inode, struct file *filep)
> filep->private_data = df;
>
> /*
> - * Use the pseudo fs inode on the device to link all mmaps
> - * to the same address space, allowing us to unmap all vmas
> - * associated to this device using unmap_mapping_range().
> + * mmaps are linked to the address space of the inode of device cdev.
> + * Save the inode of device cdev in device->inode to allow
> + * unmap_mapping_range() to unmap all vmas.
> */
> - filep->f_mapping = device->inode->i_mapping;
> -
> + device->inode = inode;
This doesn't seem right.. There is only one device but multiple file
can be opened on that device.
We expect every open file to have a unique inode otherwise the
unmap_mapping_range() will not function properly.
Jason
Powered by blists - more mailing lists