Re: [RFC 2/2] fanotify: emit FAN_MODIFY_DIR on filesystem changes

From: Amir Goldstein
Date: Tue Mar 14 2017 - 11:36:01 EST


On Tue, Mar 14, 2017 at 4:58 PM, Filip ÅtÄdronskà <r.lkml@xxxxxxxxxx> wrote:
> Hi,
>
> On Tue, Mar 14, 2017 at 01:18:01PM +0200, Amir Goldstein wrote:
>> I claim that fanotify filters event by mount not because it
>> was a requirement, but because it was an implementation challenge
>> to do otherwise.
>>
>> And I claim that what mount watchers are really interested in is
>> "all the changes that happen in the file system in the area
>> that is visible to me through this mount point".
>>
>> In other words, an indexer needs to know if files were modified\
>> create/deleted if that indexer sits in container host namespace
>> regardless if those files were modified from within a container
>> namespace.
>>
>> It's not a matter of security/isolation. It's a matter of functionality.
>> I agree that for some event (e.g. permission events) it is possible
>> to argue both ways (i.e. that the namespace context should be used
>> as a filter for events).
>> But for the new proposed events (FS_MODIFY_DIR), I really don't
>> see the point in isolation by mount/namespace.
>
> there are basically two classes of uses for a fantotify-like
> interface:
>
> (1) Keeping an up-to-date representation of the file system.
> For this, superblock watches are clearly what you want.
>
> * You are interested to know the current state of the
> filesystem so you need to know about every change,
> regardless of where it came from.
> * As I mentioned earlier, in case of remote, ditributed
> and virtual filesystems, the change might come from
> within the filesystem itself (if the protocol supports
> reporting such changes). This can probably be
> implemented only with superblock-scoped watches because
> the change is fundamentally not related to any mount.
> * Some filesystems might also support change journalling
> and it might be concievable to extend the API in the
> future to report "past" events (for example by passing
> sequence number of last seen event or similar).
> * The argument about containers escaping change notification
> you mentioned earlier.
>
> All those factors speak greatly in favour of superblock
> watches.
>
> (2) Tracking filesystem *activity*. Now you are not building
> an image of current filesystem state but rather a log of
> what happened. Perhaps you are also interested in who
> (user/process/...) did what. Permission events also fit
> mostly in this category.
>
> For those it *might* make sense to have mount-scoped
> watches, for example if you want to monitor only one
> container or a subset of processes.
>
> We both concentrate on the first but we shouldn't forget about
> the second, which was one of the original motivations for
> fanotify.
>
> Thus I conclude that it might be desirable to implement
> mount-scoped filename events in the long run. Even though
> I agree that the sb-scoped events are more important because
> they cover more use cases and you can do additional filtering
> (e.g. by pid) if deemed necessary.
>
> This would require:
>
> (a) Sprinkling the callers of vfs_* with fanotify calls
> as I did, or
> (b) Creating wrapper functions like vfs_path_unlink & co.
> that would make the necessary fanotify call (and probably
> tell the lower function not to generate another
> notification), as I suggested earlier.
> (c) Give the vfs_* functions an *optional* vfsmount argument.
>
> In the end I probably find (c) the most elegant but this
> can be discussed later, even after your changes are merged.
>

Agreed. That is an independent question.
Thanks for the thorough summary.

Amir.