Re: [PATCH v9 4/4] vfio-pci/zdev: Add VFIO FMB device features

[Omar Elghoul <oelghoul@linux.ibm.com>]

On 10/7/26 10:07 AM, sashiko-bot@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
> - [High] vfio_pci: NULL pointer deref in fmb_read on fmb allocation failure
> - [High] vfio_pci: stack buffer overflow in fmb_read via unbounded fmb_length
> --
> 
> commit 759143700fb3898fd78661e48f994eff036f8cf5
> Author: Omar Elghoul <oelghoul@linux.ibm.com>
> 
> vfio-pci/zdev: Add VFIO FMB device features
> 
> This commit introduces new VFIO features for zPCI devices to provide FMB
> passthrough to userspace. It allows enabling or disabling the FMB using a SET
> feature, and reading the latest FMB using a GET feature.
> 
>> diff --git a/drivers/vfio/pci/vfio_pci_zdev.c b/drivers/vfio/pci/vfio_pci_zdev.c
>> index f47f36314a1cc..1a66f455545ab 100644
>> --- a/drivers/vfio/pci/vfio_pci_zdev.c
>> +++ b/drivers/vfio/pci/vfio_pci_zdev.c
> [ ... ]
>> +int vfio_pci_zdev_feature_fmb_enable(struct vfio_pci_core_device *vdev, u32 flags,
>> +				     void __user *arg, size_t argsz)
>> +{
> [ ... ]
>> +	if (fmb_enable.enabled) {
>> +		zdev->fmb_requested = true;
>> +		return zpci_fmb_reenable_device(zdev);
> 
> [Severity: High]
> If zpci_fmb_reenable_device() fails (for example, if kmem_cache_zalloc()
> returns -ENOMEM), does zdev->fmb_requested remain set to true?
> 
> Since the failure doesn't revert the flag to false, could this leave the
> device state inconsistent for future FMB reads?

This flag is intentionally kept true and is documented in patch 2/4 to
track the desired FMB state, which is not necessarily the actual FMB
enablement state in firmware. In hindsight, the liveness check should've
been zdev->fmb_requested && zdev->fmb rather than just one or the other.
The "inconsistent" device state isn't an issue either because the FMB
sample includes a timestamp, so a consumer would know when it's reading
stale data.

> 
> [Severity: High]
> Is it intentional to bypass the zdev->fmb_length bounds check here?
> 
> The standard enablement path typically validates that zdev->fmb_length fits
> within the structure boundaries. By calling zpci_fmb_reenable_device()
> directly, it appears a hardware-supplied length greater than the size of
> struct zpci_fmb could be accepted and saved.

In practice this shouldn't be an issue, but it would've been a good idea
to move the sizeof(*zdev->fmb) < zdev->fmb_length check into
zpci_fmb_reenable_device() as well to prevent this issue from existing
even in theory.

> 
>> +	}
>> +
>> +	return zpci_fmb_disable_device(zdev);
>> +}
>> +
>> +int vfio_pci_zdev_feature_fmb_read(struct vfio_pci_core_device *vdev, u32 flags,
>> +				   void __user *arg, size_t argsz)
>> +{
>> +	struct zpci_dev *zdev;
>> +	struct vfio_device_feature_zpci_fmb_read fmb_read;
>> +	struct zpci_fmb fmb_bounce;
> [ ... ]
>> +	scoped_guard(mutex, &zdev->fmb_lock) {
>> +		if (!zdev->fmb_requested)
>> +			return -ENOMSG;
>> +		memcpy(&fmb_bounce, zdev->fmb, zdev->fmb_length);
> 
> [Severity: High]
> If a previous call to vfio_pci_zdev_feature_fmb_enable() failed during
> allocation but left zdev->fmb_requested = true, this check will pass.
> 
> Will this result in a NULL pointer dereference when accessing zdev->fmb in
> this memcpy()?

Admittedly I did miss this one, sorry about that. I will address it in
the next iteration.

> 
> [Severity: High]
> Can zdev->fmb_length be larger than sizeof(struct zpci_fmb)?
> 
> If an oversized fmb_length is reported by the hardware or firmware and was not
> validated during enablement, could copying up to that length into the
> fixed-size local stack variable fmb_bounce result in a stack buffer overflow?

See above

> 
>> +	}
>> +
>> +	if (copy_to_user(u64_to_user_ptr(fmb_read.data), &fmb_bounce, zdev->fmb_length))
>> +		return -EFAULT;
>> +	return 0;
>> +}
>