Re: [PATCH v6 5/7] vfio: Enable cdev noiommu mode under iommufd

[Jacob Pan <jacob.pan@linux.microsoft.com>]

Hi Yi,

On Mon, 25 May 2026 14:29:31 +0800
Yi Liu <yi.l.liu@intel.com> wrote:

> Hi Jacob,
> 
> On 5/24/26 06:01, Jacob Pan wrote:
> > Hi Yi,
> > 
> > On Fri, 22 May 2026 17:19:41 +0800
> > Yi Liu <yi.l.liu@intel.com> wrote:
> >   
> >> On 5/22/26 06:11, Jacob Pan wrote:  
> >>> Now that devices under noiommu mode can bind with IOMMUFD and
> >>> perform IOAS operations, lift restrictions on cdev from VFIO side.
> >>> Use cases are documented in Documentation/driver-api/vfio.rst
> >>>
> >>> Signed-off-by: Jacob Pan <jacob.pan@linux.microsoft.com>
> >>> ---
> >>> v6:
> >>>     - Revert back to unified VFIO_NOIOMMU Kconfig for both cdev
> >>> and group. Use Kconfig dependency to restrict usages and avoid
> >>> null group checks. (Alex & Yi)
> >>>     - Add CAP_SYS_RAWIO checks for cdev open to maintain security
> >>> parity with the group noiommu path. (Alex)
> >>> v5:
> >>>      - Add Kconfig VFIO_CDEV_NOIOMMU to select IOMMUFD_NOIOMMU
> >>>        and its dependencies
> >>>      - Add comment to explain vfio_noiommu conditional definition
> >>> (Alex)
> >>>      - Removed early return for group noiommu in bind/unbind
> >>>      - Use consistent wording referring to VFIO noiommu mode
> >>> (Kevin)
> >>>      - Update unsafe_noiommu Kconfig help text (Kevin)
> >>>      - Change dev_warn to dev_info for noiommu enabling msg
> >>> (Kevin) v4:
> >>>      - Remove early return in iommufd_bind for noiommu (Alex)
> >>> v3:
> >>>      - Consolidate into fewer patches
> >>> v2:
> >>>      - removed unnecessary device->noiommu set in
> >>>        iommufd_vfio_compat_ioas_get_id()
> >>> Signed-off-by: Jacob Pan <jacob.pan@linux.microsoft.com>
> >>> ---
> >>>    drivers/vfio/Kconfig       |  8 +++++---
> >>>    drivers/vfio/device_cdev.c |  3 +++
> >>>    drivers/vfio/iommufd.c     |  6 +++---
> >>>    drivers/vfio/vfio.h        | 20 +++++++++++++-------
> >>>    drivers/vfio/vfio_main.c   | 23 +++++++++++++++++++----
> >>>    include/linux/vfio.h       |  1 +
> >>>    6 files changed, 44 insertions(+), 17 deletions(-)
> >>>
> >>> diff --git a/drivers/vfio/Kconfig b/drivers/vfio/Kconfig
> >>> index ceae52fd7586..d3d8fef2855c 100644
> >>> --- a/drivers/vfio/Kconfig
> >>> +++ b/drivers/vfio/Kconfig
> >>> @@ -22,8 +22,7 @@ config VFIO_DEVICE_CDEV
> >>>    	  The VFIO device cdev is another way for userspace to
> >>> get device access. Userspace gets device fd by opening device cdev
> >>> under /dev/vfio/devices/vfioX, and then bind the device fd with an
> >>> iommufd
> >>> -	  to set up secure DMA context for device access.  This
> >>> interface does
> >>> -	  not support noiommu.
> >>> +	  to set up secure DMA context for device access.  
> >>
> >> if noiommu, it's unsafe DMA. :)  
> > yes, here I just want to remove "This interface does not support
> > noiommu.".  
> >>  
> >>>    	  If you don't know what to do here, say N.
> >>>    
> >>> @@ -62,7 +61,10 @@ endif
> >>>    
> >>>    config VFIO_NOIOMMU
> >>>    	bool "VFIO No-IOMMU support"
> >>> -	depends on VFIO_GROUP
> >>> +	depends on VFIO_GROUP || VFIO_DEVICE_CDEV
> >>> +	depends on !VFIO_GROUP || VFIO_CONTAINER ||
> >>> IOMMUFD_VFIO_CONTAINER
> >>> +	depends on !VFIO_DEVICE_CDEV || !GENERIC_ATOMIC64
> >>> +	select IOMMUFD_NOIOMMU if VFIO_DEVICE_CDEV
> >>>    	help
> >>>    	  VFIO is built on the ability to isolate devices using
> >>> the IOMMU. Only with an IOMMU can userspace access to DMA capable
> >>> devices be diff --git a/drivers/vfio/device_cdev.c
> >>> b/drivers/vfio/device_cdev.c index 54abf312cf04..4e2c1e4fc1f8
> >>> 100644 --- a/drivers/vfio/device_cdev.c
> >>> +++ b/drivers/vfio/device_cdev.c
> >>> @@ -27,6 +27,9 @@ int vfio_device_fops_cdev_open(struct inode
> >>> *inode, struct file *filep) struct vfio_device_file *df;
> >>>    	int ret;
> >>>    
> >>> +	if (device->noiommu && !capable(CAP_SYS_RAWIO))
> >>> +		return -EPERM;
> >>> +
> >>>    	/* Paired with the put in vfio_device_fops_release() */
> >>>    	if (!vfio_device_try_get_registration(device))
> >>>    		return -ENODEV;
> >>> diff --git a/drivers/vfio/iommufd.c b/drivers/vfio/iommufd.c
> >>> index a38d262c6028..d4f2e2a0f2f3 100644
> >>> --- a/drivers/vfio/iommufd.c
> >>> +++ b/drivers/vfio/iommufd.c
> >>> @@ -25,8 +25,8 @@ int vfio_df_iommufd_bind(struct vfio_device_file
> >>> *df)
> >>>    	lockdep_assert_held(&vdev->dev_set->lock);
> >>>    
> >>> -	/* Returns 0 to permit device opening under noiommu mode
> >>> */
> >>> -	if (vfio_device_is_noiommu(vdev))
> >>> +	/* Group noiommu via iommufd compat needs no device
> >>> binding */
> >>> +	if (df->group && vfio_device_is_noiommu(vdev))  
> >>
> >> seems like vfio_device_is_noiommu() implies group path, then no
> >> need to use df->group.
> >>  
> > df->group is needed because only the legacy VFIO
> > group/iommufd-compat noiommu path should skip real iommufd device
> > binding.  
> 
> got it. It should be opened via group path. Otherwise, it should go
> ahead to bind iommufd. BTW. the noiommu check in
> vfio_iommufd_compat_attach_ioas() is skipped. I know this helper is
> only for the group path, so the df->group is not added, can we add
> a note for it?
> 
Sure, I will add a comment:

diff --git a/drivers/vfio/iommufd.c b/drivers/vfio/iommufd.c
index d4f2e2a0f2f3..e9893d34d07b 100644
--- a/drivers/vfio/iommufd.c
+++ b/drivers/vfio/iommufd.c
@@ -40,7 +40,11 @@ int vfio_iommufd_compat_attach_ioas(struct vfio_device *vdev,
 
        lockdep_assert_held(&vdev->dev_set->lock);
 
-       /* compat noiommu does not need to do ioas attach */
+       /*
+        * Compat noiommu does not need to do ioas attach. This helper is
+        * only called from the legacy group/iommufd compat path, so no
+        * explicit df->group check is needed.
+        */

> > For df->group == NULL, the fd is a VFIO cdev fd. That path uses
> > VFIO_DEVICE_BIND_IOMMUFD and later VFIO_DEVICE_ATTACH_IOMMUFD_PT.
> > Even in noiommu cdev mode, bind must still call:
> > 
> >   vdev->ops->bind_iommufd(vdev, ictx, &df->devid);
> > 
> > so vdev->iommufd_device can get initialized. If the check were only:
> > 
> >   if (vfio_device_is_noiommu(vdev))
> >          return 0;
> > then cdev noiommu bind would falsely “succeed” without setting
> > vdev->iommufd_device. Later VFIO_DEVICE_ATTACH_IOMMUFD_PT calls
> > vfio_iommufd_physical_attach_ioas(), hits:
> > 
> >   if (WARN_ON(!vdev->iommufd_device))
> >          return -EINVAL;
> > 
> > In the noiommu test, you will get:
> >    185.870670] ------------[ cut here ]------------
> > [  185.871952] WARNING: drivers/vfio/iommufd.c:157 at
> > vfio_iommufd_physical_attach_ioas+0x3f/0x50, CPU#0:
> > vfio-noiommu-pc/157[  185.875010] Modules linked in:[  185.875882]
> > CPU: 0 UID: 0 PID: 157 Comm: vfio-noiommu-pc Tainted: G     U  W
> > 7.1.0-rc1+ #20 PREEMPT[  185.878637] Tainted: [U]=USER, [W]=WARN[
> > 185.879711] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS
> > rel-1.17.0-0-gb52ca86e094d-prebuilt.qemu.org 04/01/2014[
> > 185.882913] RIP: 0010:vfio_iommufd_physical_attach_ioas+0x3f/0x50[
> > 185.884624] Code: 89 f2 31 f6 f6 83 50 04 00 00 01 75 16 e8 d9 aa
> > c6 ff 85 c0 75 07 80 8b 50 04 00 00 01 5b c3 cc cc cc cc e8 43 ab
> > c6 ff eb e8 <0f> 0b b80[  185.889701] RSP: 0018:ffa000000062fd88
> > EFLAGS: 00010246[ 185.891161] RAX: ffffffff81f59ee0 RBX:
> > ff1100010c43b800 RCX: 0000000000000000[  185.893141] RDX:
> > ff1100010c708040 RSI: ffa000000062fda0 RDI: 0000000000000000[
> > 185.895127] RBP: ff1100010c43b800 R08: ff1100010c7c12b0 R09:
> > 0000000000000000[ 185.897119] R10: 0000000000000000 R11:
> > 0000000000000000 R12: 00007ffec4c2f720[  185.899102] R13:
> > ffa000000062fda0 R14: ff11000103bd40d0 R15: ff1100010c43b800[
> > 185.901075] FS: 0000000028d69380(0000) GS:ff110004e4a8d000(0000)
> > knlGS:0000000000000000[  185.903284] CS:  0010 DS: 0000 ES: 0000
> > CR0: 0000000080050033[  185.904888] CR2: 0000000028d73988 CR3:
> > 0000000103507002 CR4: 0000000000f73ef0[  185.906853] PKRU: 55555554[
> > 185.907636] Call Trace:[  185.908373]  <TASK>[  185.908932]
> > vfio_df_ioctl_attach_pt+0xc7/0x170[  185.910085]
> > vfio_device_fops_unl_ioctl+0x49b/0xa50[  185.911322]  ?
> > file_tty_write.isra.0+0x202/0x320[  185.912507]
> > __x64_sys_ioctl+0x425/0xa30[  185.913502]  do_syscall_64+0x5e/0xf80[
> > 185.914444]  ? irqentry_exit+0x3b/0x5e0[  185.915414]
> > entry_SYSCALL_64_after_hwframe+0x76/0x7e[  185.916701] RIP:
> > 0033:0x434a4d[  185.917498] Code: 04 25 28 00 00 00 48 89 45 c8 31
> > c0 48 8d 45 10 c7 45 b0 10 00 00 00 48 89 45 b8 48 8d 45 d0 48 89
> > 45 c0 b8 10 00 00 00 0f 05 <89> c2 3d0[  185.922052] RSP:
> > 002b:00007ffec4c2f6b0 EFLAGS: 00000246 ORIG_RAX: 0000000000000010[
> > 185.923785] RAX: ffffffffffffffda RBX: 0000000000000004 RCX:
> > 0000000000434a4d[ 185.925398] RDX: 00007ffec4c2f720 RSI:
> > 0000000000003b77 RDI: 0000000000000004[  185.927007] RBP:
> > 00007ffec4c2f700 R08: 0000000000000064 R09: 0000000000000000[
> > 185.928611] R10: 0000000000000000 R11: 0000000000000246 R12:
> > 00007ffec4c30918[ 185.930211] R13: 00007ffec4c30940 R14:
> > 00000000004cf868 R15: 0000000000000001[  185.931758]  </TASK>[
> > 185.932258] ---[ end trace 0000000000000000 ]---Failed to attach pt
> > to device 
> >> static inline bool vfio_device_is_noiommu(struct vfio_device *vdev)
> >> {
> >>       return IS_ENABLED(CONFIG_VFIO_NOIOMMU) &&
> >>              vdev->group->type == VFIO_NO_IOMMU;
> >> }
> >>  
> >>>    		return 0;
> >>>    
> >>>    	return vdev->ops->bind_iommufd(vdev, ictx, &df->devid);
> >>> @@ -58,7 +58,7 @@ void vfio_df_iommufd_unbind(struct
> >>> vfio_device_file *df)
> >>>    	lockdep_assert_held(&vdev->dev_set->lock);
> >>>    
> >>> -	if (vfio_device_is_noiommu(vdev))
> >>> +	if (df->group && vfio_device_is_noiommu(vdev))
> >>>    		return;
> >>>    
> >>>    	if (vdev->ops->unbind_iommufd)
> >>> diff --git a/drivers/vfio/vfio.h b/drivers/vfio/vfio.h
> >>> index e4b72e79b7e3..6f0a2dfc8a00 100644
> >>> --- a/drivers/vfio/vfio.h
> >>> +++ b/drivers/vfio/vfio.h
> >>> @@ -358,19 +358,13 @@ void vfio_init_device_cdev(struct
> >>> vfio_device *device);
> >>>    static inline int vfio_device_add(struct vfio_device *device)
> >>>    {
> >>> -	/* cdev does not support noiommu device */
> >>> -	if (vfio_device_is_noiommu(device))
> >>> -		return device_add(&device->device);
> >>>    	vfio_init_device_cdev(device);
> >>>    	return cdev_device_add(&device->cdev, &device->device);
> >>>    }
> >>>    
> >>>    static inline void vfio_device_del(struct vfio_device *device)
> >>>    {
> >>> -	if (vfio_device_is_noiommu(device))
> >>> -		device_del(&device->device);
> >>> -	else
> >>> -		cdev_device_del(&device->cdev, &device->device);
> >>> +	cdev_device_del(&device->cdev, &device->device);
> >>>    }
> >>>    
> >>>    int vfio_device_fops_cdev_open(struct inode *inode, struct file
> >>> *filep); @@ -420,6 +414,18 @@ static inline void
> >>> vfio_cdev_cleanup(void) }
> >>>    #endif /* CONFIG_VFIO_DEVICE_CDEV */
> >>>    
> >>> +#if IS_ENABLED(CONFIG_VFIO_NOIOMMU)
> >>> +static inline bool vfio_device_is_cdev_noiommu(struct vfio_device
> >>> *vdev) +{
> >>> +	return vdev->noiommu;
> >>> +}
> >>> +#else
> >>> +static inline bool vfio_device_is_cdev_noiommu(struct vfio_device
> >>> *vdev) +{
> >>> +	return false;
> >>> +}
> >>> +#endif
> >>> +
> >>>    #if IS_ENABLED(CONFIG_VFIO_VIRQFD)
> >>>    int __init vfio_virqfd_init(void);
> >>>    void vfio_virqfd_exit(void);
> >>> diff --git a/drivers/vfio/vfio_main.c b/drivers/vfio/vfio_main.c
> >>> index 6222376ab6ab..84381c500623 100644
> >>> --- a/drivers/vfio/vfio_main.c
> >>> +++ b/drivers/vfio/vfio_main.c
> >>> @@ -321,6 +321,20 @@ static int vfio_init_device(struct
> >>> vfio_device *device, struct device *dev, return ret;
> >>>    }
> >>>    
> >>> +static int vfio_device_set_noiommu_and_name(struct vfio_device
> >>> *device) +{
> >>> +	if (IS_ENABLED(CONFIG_VFIO_DEVICE_CDEV) && vfio_noiommu
> >>> && !device->dev->iommu) {
> >>> +		device->noiommu = true;
> >>> +		add_taint(TAINT_USER, LOCKDEP_STILL_OK);
> >>> +		dev_warn(device->dev,
> >>> +			 "Adding kernel taint for vfio-noiommu
> >>> cdev on device\n");
> >>> +	}
> >>> +
> >>> +	/* Just to be safe, expose to user explicitly noiommu
> >>> cdev node */
> >>> +	return dev_set_name(&device->device, "%svfio%d",
> >>> +		     device->noiommu ? "noiommu-" : "",
> >>> device->index); +}
> >>> +
> >>>    static int __vfio_register_dev(struct vfio_device *device,
> >>>    			       enum vfio_group_type type)
> >>>    {
> >>> @@ -340,20 +354,21 @@ static int __vfio_register_dev(struct
> >>> vfio_device *device, if (!device->dev_set)
> >>>    		vfio_assign_device_set(device, device);
> >>>    
> >>> -	ret = dev_set_name(&device->device, "vfio%d",
> >>> device->index);
> >>> +	ret = vfio_device_set_group(device, type);
> >>>    	if (ret)
> >>>    		return ret;
> >>>    
> >>> -	ret = vfio_device_set_group(device, type);
> >>> +	ret = vfio_device_set_noiommu_and_name(device);  
> >>
> >> the order of dev_set_name and vfio_device_set_group() are swapped,
> >> any special reason?  
> > The ordering was intentional in an earlier version where the cdev
> > noiommu check depended on device->group. With the current check
> > using !device->dev->iommu, the ordering is no longer strictly
> > required for that test.
> >   
> > I kept vfio_device_set_group() first because the rest of
> > registration already treats group setup as the first VFIO state to
> > unwind, and this lets the existing err_out path handle failures
> > after group assignment, including dev_set_name(). I can restore the
> > old order if you prefer, since it is not functionally required
> > anymore.  
> 
> I think it's better to keep the original order if no functional
> requirement anymore.
will do,

> 
> >>>    	if (ret)
> >>> -		return ret;
> >>> +		goto err_out;
> >>>    
> >>>    	/*
> >>>    	 * VFIO always sets IOMMU_CACHE because we offer no way
> >>> for userspace to
> >>>    	 * restore cache coherency. It has to be checked here
> >>> because it is only
> >>>    	 * valid for cases where we are using iommu groups.
> >>>    	 */
> >>> -	if (type == VFIO_IOMMU && !vfio_device_is_noiommu(device)
> >>> &&
> >>> +	if (type == VFIO_IOMMU &&
> >>> !(vfio_device_is_noiommu(device) ||
> >>> +
> >>> vfio_device_is_cdev_noiommu(device)) &&  
> >>
> >> now, the group path and cdev path have their own is_noiommu helper,
> >> can the two helpers be consolidated?
> >>  
> > They could be consolidated mechanically, but I feel they are
> > checking different things it is more clear to keep them separate?  
> 
> IMHO. They are actually checking if the device is noiommu. I found the
> current usage of vfio_device_is_noiommu(). #1 is totally specific for
> group path. #2, #3 and #4 are for the common path to identify the
> noiommu of group. It also implies the info of the open path (group
> path?). #7 and #8 is going to be dropped. And 9 is totally for
> checking noiommu attribute. So I'm wondering if using !dev->iommu is
> a good choice. Should be able to cover both group and cdev path. Let
> me know if this is not workable.
I'm not sure I follow -- do you mean using !dev->iommu as the common
no-IOMMU check?
 
I don't think that should be the helper condition directly. !dev->iommu
is the low-level device state, but VFIO no-IOMMU is a VFIO mode. For
example, VFIO_EMULATED_IOMMU devices may also not have dev->iommu, but
they should not be treated as VFIO_NO_IOMMU.

What I can do is consolidate the helpers around the VFIO state instead:
legacy group no-IOMMU is represented by group->type, while the cdev
path uses vdev->noiommu

static inline bool vfio_device_is_noiommu(struct vfio_device *vdev)
{
#if IS_ENABLED(CONFIG_VFIO_GROUP)
	if (vdev->group && vdev->group->type == VFIO_NO_IOMMU)
		return true;
#endif

	return IS_ENABLED(CONFIG_IOMMUFD_NOIOMMU) && vdev->noiommu;
}

> 
>     #   line  filename / context / line
>     1    193  drivers/vfio/group.c <<vfio_df_group_open>>
>               if (df->iommufd && vfio_device_is_noiommu(device) && 
> device->open_count == 0) {
>     2     29  drivers/vfio/iommufd.c <<vfio_df_iommufd_bind>>
>               if (vfio_device_is_noiommu(vdev))
>     3     44  drivers/vfio/iommufd.c
> <<vfio_iommufd_compat_attach_ioas>> if (vfio_device_is_noiommu(vdev))
>     4     61  drivers/vfio/iommufd.c <<vfio_df_iommufd_unbind>>
>               if (vfio_device_is_noiommu(vdev))
>     5    115  drivers/vfio/vfio.h <<vfio_device_is_noiommu>>
>               static inline bool vfio_device_is_noiommu(struct 
> vfio_device *vdev)
>     6    191  drivers/vfio/vfio.h <<vfio_device_is_noiommu>>
>               static inline bool vfio_device_is_noiommu(struct 
> vfio_device *vdev)
>     7    362  drivers/vfio/vfio.h <<vfio_device_add>>
>               if (vfio_device_is_noiommu(device))
>     8    370  drivers/vfio/vfio.h <<vfio_device_del>>
>               if (vfio_device_is_noiommu(device))
>     9    356  drivers/vfio/vfio_main.c <<__vfio_register_dev>>
>               if (type == VFIO_IOMMU &&
> !vfio_device_is_noiommu(device) &&
> 
> >>>    	    !device_iommu_capable(device->dev,
> >>> IOMMU_CAP_CACHE_COHERENCY)) { ret = -EINVAL;
> >>>    		goto err_out;
> >>> diff --git a/include/linux/vfio.h b/include/linux/vfio.h
> >>> index 31b826efba00..45f08986359e 100644
> >>> --- a/include/linux/vfio.h
> >>> +++ b/include/linux/vfio.h
> >>> @@ -74,6 +74,7 @@ struct vfio_device {
> >>>    	u8 iommufd_attached:1;
> >>>    #endif
> >>>    	u8 cdev_opened:1;
> >>> +	u8 noiommu:1;
> >>>    	/*
> >>>    	 * debug_root is a static property of the vfio_device
> >>>    	 * which must be set prior to registering the
> >>> vfio_device.  
> >   
> 
> Regards,
> Yi Liu