Re: [RFC][PATCH] fanotify: allow setting FAN_CREATE in mount mark mask

[Amir Goldstein <amir73il@gmail.com>]

> > This implementation is a compromise for not having clear user mount
> > context in all places that call for an event.
> > For every person you find that thinks it is intuitive to get an event on /B
> > for touch C/bla, you will find another person that thinks it is not intuitive
>
> And I think here we disagree. The technical implementation currently
> requires this since the two mounts are both clearly marked and the first
> mount creates objects by going through the other mount and they don't
> have a private mount. All I was saying is that the current patchset
> can't handle this case and asked whether we are ok with that and if not
> what we do to fix it.
> My proposal two or three mails ago and then picked up by you is: make
> them both use a private clone mount which is - as I said in earlier
> mails - the correct solution anyway and falls in line with overlayfs
> too.
>

As long as we agree on the solution ;-)

> > to get an event. I think we are way beyond the stage with mount
> > namespaces where intuition alone suffice.
> >
> > w.r.t consistent, you gave a few examples and I suggested how IMO
> > they should be fixed to behave consistently.
> > If you have other examples of alleged inconsistencies, please list them.
>
> It feels like I somehow upset you with this.

You do not upset me.

I just didn't find a better way to address "consistent and intuitive" concern
without asking for more concrete examples, after we eliminated the
ecryptfs example, which we already agreed(?), is a non issue.

My claim about "intuitive" is that there is a limit to how intuitive
this could be.
I do not see myself explaining in the man page why FAN_DELETE_SELF
cannot be requested for a mount mark. It's just too low level.

So the best I can do is document the events that are available to inode/sb
mark and not available to mount mark.

Currently, the fanotify_mark.2 man page reads:
"...The events which require that filesystem objects are identified by
file handles,
 such as FAN_CREATE, FAN_ATTRIB,  FAN_MOVE,  and FAN_DELETE_SELF,
 cannot  be provided as a mask when flags contains FAN_MARK_MOUNT..."

I will change that to:
"...The events FAN_ATTRIB,  FAN_MOVE,  and FAN_DELETE_SELF,
 cannot  be provided as a mask when flags contain FAN_MARK_MOUNT..."

Without providing a rationale to the list of forbidden events.

BTW, there is an undocumented fact about FAN_MODIFY -
This event is allowed in a mount mark mask, but it only reports the events
generated by fsnotify_modify() on file writes. It does not report to a mount
mark, the FAN_MODIFY event generated by fsnotify_change() from
truncate() and utimensat(), because of the missing mount context.

So yeh, I do understand where the "inconsistent" feeling is coming from... ;-)

[...]
> > So I would like to know that we really have all the pieces needed for
> > a useful solution, before proposing the fanotify patches.
>
> Sure, if you think that you have your branch in the shape that you want
> to. So far it has been evolving quite rapidly as you said yourself. :)
> I can probably test this soon early next week seems most likely since I
> need to find a timeslot to actually do the work you're asking. Hope that
> works.
>

No plans to make any more changes to those branches and no rush
as to when to post tham. This is not v4.13-rc1 material anyway.

Thanks,
Amir.