On Wed, Mar 04, 2026 at 01:53:42AM -0500, Demi Marie Obenour wrote:
> I noticed potentially missing input sanitization in dma_buf_set_name(),
> which is reachable from DMA_BUF_SET_NAME.  This allows inserting a name
> containing a newline, which is then used to construct the contents of
> /proc/PID/task/TID/fdinfo/FD.  This could confuse userspace programs
> that access this data, possibly tricking them into thinking a file
> descriptor is of a different type than it actually is.
> 
> Other code might have similar bugs.  For instance, there is code that
> uses a sysfs path, a driver name, or a device name from /dev.  It is
> possible to sanitize the first, and the second and third should come
> from trusted sources within the kernel itself.  The last area where
> I found a potential problem is BPF.  I don't know if this can happen.
> 
> I think this should be fixed by either sanitizing data on write
> (by limiting the allowed characters in dma_buf_set_name()), on read
> (by using one of the formats that escapes special characters), or both.
> 
> Is there a better way to identify that a file descriptor is of
> a particular type, such as an eventfd?  fdinfo is subject to

The problem is that most of the anonymous inodes share a single
anonymous inode so any uapi that returns information based inode->i_op
is not going to be usable.

> bugs of this type, which might happen again.  readlink() reports
> "anon_inode:[eventfd]" and S_IFMT reports a mode of 0, but but my

That is definitely uapi by now. We've tried to change S_IFMT and it
breaks lsfd and other tools so we can't reasonably change it. In fact,
pidfds pretend to be anon_inode even though they're not simply because
some tools parse that out.

> reading of the kernel source code is that neither is intended to be
> stable uAPI.  Is there a better interface that can be used?

Reply via email to