Re: [for-linus][PATCH 1/3] eventfs: Have the inodes all for files and directories all be the same

[Christian Brauner]

On Mon, Jan 22, 2024 at 02:44:43PM -0500, Steven Rostedt wrote:
> On Mon, 22 Jan 2024 10:19:12 -0800
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Mon, 22 Jan 2024 at 09:39, Linus Torvalds
> > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > Actually, why not juist add an inode number to your data structures,
> > > at least for directories? And just do a static increment on it as they
> > > get registered?
> > >
> > > That avoids the whole issue with possibly leaking kernel address data.  
> > 
> > The 'nlink = 1' thing doesn't seem to make 'find' any happier for this
> > case, sadly.
> > 
> > But the inode number in the 'struct eventfs_inode' looks trivial. And
> > doesn't even grow that structure on 64-bit architectures at least,
> > because the struct is already 64-bit aligned, and had only one 32-bit
> > entry at the end.
> > 
> > On 32-bit architectures the structure size grows, but I'm not sure the
> > allocation size grows. Our kmalloc() is quantized at odd numbers.
> > 
> > IOW, this trivial patch seems to be much safer than worrying about
> > some pointer exposure.
> 
> I originally wanted to avoid the addition of the 4 bytes, but your comment
> about it not making a difference on 64bit due to alignment makes sense.

Non-unique inode numbers aren't exactly great for userspace and there
are a surprisingly small yet large enough number of tools that trip over
this in various ways. So if that can be avoided then great.

But luckily no one is probably going to tar up tracefs. ;)