Re: [RFC bpf-next fanotify 2/5] samples/fanotify: Add a sample fanotify fastpath handler

[Song Liu <songliubraving@meta.com>]

Hi Jeff, 

> On Oct 30, 2024, at 5:23 PM, Jeff Layton <jlayton@kernel.org> wrote:

[...]

>> If the subtree is all in the same file system, we can attach fanotify to 
>> the whole file system, and then use some dget_parent() and follow_up() 
>> to walk up the directory tree in the fastpath handler. However, if there
>> are other mount points in the subtree, we will need more logic to handle
>> these mount points. 
>> 
> 
> My 2 cents...
> 
> I'd just confine it to a single vfsmount. If you want to monitor in
> several submounts, then you need to add new fanotify watches.
> 
> Alternately, maybe there is some way to designate that an entire
> vfsmount is a child of a watched (or ignored) directory?
> 
>> @Christian, I would like to know your thoughts on this (walking up the 
>> directory tree in fanotify fastpath handler). It can be expensive for 
>> very very deep subtree. 
>> 
> 
> I'm not Christian, but I'll make the case for it. It's basically a
> bunch of pointer chasing. That's probably not "cheap", but if you can
> do it under RCU it might not be too awful. It might still suck with
> really deep paths, but this is a sample module. It's not expected that
> everyone will want to use it anyway.

Thanks for the suggestion! I will try to do it under RCU. 

> 
>> How should we pass in the subtree? I guess we can just use full path in
>> a string as the argument.
>> 
> 
> I'd stay away from string parsing. How about this instead?
> 
> Allow a process to open a directory fd, and then hand that fd to an
> fanotify ioctl that says that you want to ignore everything that has
> that directory as an ancestor. Or, maybe make it so that you only watch
> dentries that have that directory as an ancestor? I'm not sure what
> makes the most sense.

Yes, directory fd is another option. Currently, the "attach to group"
function only takes a string as input. I guess it makes sense to allow
taking a fd, or maybe we should allow any random format (pass in a 
pointer to a structure. Let me give it a try. 

Thanks again!

Song

> 
> Then, when you get a dentry-based event, you just walk the d_parent
> pointers back up to the root of the vfsmount. If one of them matches
> the dentry in your fd then you're done. If you hit the root of the
> vfsmount, then you're also done (and know that it's not a child of that
> dentry).
> -- 
> Jeff Layton <jlayton@kernel.org>