Re: [RFC][PATCH] fanotify: introduce filesystem view mark

On Sat, May 15, 2021 at 05:28:27PM +0300, Amir Goldstein wrote:

On Fri, May 14, 2021 at 4:56 PM Christian Brauner
[off-list ref] wrote:

quoted

On Thu, May 13, 2021 at 12:55:26PM +0200, Jan Kara wrote:

quoted

On Wed 12-05-21 17:26:25, Christian Brauner wrote:

quoted

On Mon, May 10, 2021 at 02:37:59PM +0300, Amir Goldstein wrote:

quoted

On Mon, May 10, 2021 at 1:13 PM Jan Kara [off-list ref] wrote:

quoted

OK, so this feature would effectively allow sb-wide watching of events that
are generated from within the container (or its descendants). That sounds
useful. Just one question: If there's some part of a filesystem, that is
accesible by multiple containers (and thus multiple namespaces), or if
there's some change done to the filesystem say by container management SW,
then event for this change won't be visible inside the container (despite
that the fs change itself will be visible).

That is correct.
FYI, a privileged user can already mount an overlayfs in order to indirectly
open and write to a file.

Because overlayfs opens the underlying file FMODE_NONOTIFY this will
hide OPEN/ACCESS/MODIFY/CLOSE events also for inode/sb marks.
Since 459c7c565ac3 ("ovl: unprivieged mounts"), so can unprivileged users.

I wonder if that is a problem that we need to fix...

quoted

This is kind of a similar
problem to the one we had with mount marks and why sb marks were created.
So aren't we just repeating the mistake with mount marks? Because it seems
to me that more often than not, applications are interested in getting
notification when what they can actually access within the fs has changed
(and this is what they actually get with the inode marks) and they don't
care that much where the change came from... Do you have some idea how
frequent are such cross-ns filesystem changes?

The use case surely exist, the question is whether this use case will be
handled by a single idmapped userns or multiple userns.

You see, we simplified the discussion to an idmapped mount that uses
the same userns and the userns the container processes are associated
with, but in fact, container A can use userns A container B userns B and they
can both access a shared idmapped mount mapped with userns AB.

I think at this point in time, there are only ideas about how the shared data
case would be managed, but Christian should know better than me.

I think there are two major immediate container use-cases right now that
are already actively used:
1. idmapped rootfs
A container manager wants to avoid recursively chowning the rootfs or
image for a container. To this end an idmapped mount is created. The
idmapped mount can either share the same userns as the container itself
or a separate userns can be used. What people use depends on their
concept of a container.
For example, systemd has merged support for idmapping a containers
rootfs in [1]. The systemd approach to containers never puts the
container itself in control of most things including most of its mounts.
That is very much the approach of having it be a rather tightly managed
system. Specifically, this means that systemd currently uses a separate
userns to idmap.
In contrast other container managers usually treat the container as a
mostly separate system and put it in charge of all its mounts. This
means the userns used for the idmapped mount will be the same as the
container runs in (see [2]).

OK, thanks for explanation. So to make fanotify idmap-filtered marks work
for systemd-style containers we would indeed need what Amir proposed -
i.e., the container manager intercepts fanotify_mark calls and decides
which namespace to setup the mark in as there's no sufficient priviledge
within the container to do that AFAIU.

Yes, that's how that would work.

quoted

quoted

2. data sharing among containers or among the host and containers etc.
The most common use-case is to share data from the host with the
container such as a download folder or the Linux folder on ChromeOS.
Most container managers will simly re-use the container's userns for
that too. More complex cases arise where data is shared between
containers with different idmappings then often a separate userns will
have to be used.

OK, but if say on ChromeOS you copy something to the Linux folder by app A
(say file manager) and containerized app B (say browser) watches that mount

For ChromeOS it is currently somewhat simple since they currently only
allow a single container by default. So everytime you start an app in
the container it's the same app so they all write to the Linux Files
folder through the same container. (I'm glossing over a range of details
but that's not really relevant to the general spirit of the example.).

quoted

for changes with idmap-filtered mark, then it won't see notification for
those changes because A presumably runs in a different namespace than B, am
I imagining this right? So mark which filters events based on namespace of
the originating process won't be usable for such usecase AFAICT.

Idmap filtered marks won't cover that use-case as envisioned now. Though
I'm not sure they really need to as the semantics are related to mount
marks.

We really need to refer to those as filesystem marks. They are definitely
NOT mount marks. We are trying to design a better API that will not share
as many flaws with mount marks...

quoted

A mount mark would allow you to receive events based on the
originating mount. If two mounts A and B are separate but expose the
same files you wouldn't see events caused by B if you're watching A.
Similarly you would only see events from mounts that have been delegated
to you through the idmapped userns. I find this acceptable especially if
clearly documented.

The way I see it, we should delegate all the decisions over to userspace,
but I agree that the current "simple" proposal may not provide a good
enough answer to the case of a subtree that is shared with the host.

I was focussed on what happens if you set an idmapped filtered mark for
a container for a set of files that is exposed to another container via
another idmapped mount. And it seemed to me that it was ok if the
container A doesn't see events from container B.

You seem to be looking at this from the host's perspective right now
which is interesting as well.

IMO, it should be a container manager decision whether changes done by
the host are:
a) Not visible to containerized application

Yes, that seems ok.

b) Watched in host via recursive inode watches
c) Watched in host by filesystem mark filtered in userspace
d) Watched in host by an "noop" idmapped mount in host, through
     which all relevant apps in host access the shared folder

So b)-d) are concerned with the host getting notifcations for changes
done from any container that uses a given set of files possibly through
different mounts.

We can later provide the option of "subtree filtered filesystem mark"
which can be choice (e). It will incur performance overhead on the system
that is higher than option (d) but lower than option (c).

In the end, it all depends on the individual use case.

The way forward as I see it is:
1. Need to write a proposal where FAN_MARK_IDMAPPED is the
    first step towards a wider API that also includes subtree marks -
    whether we end up implementing subtree mark or not
2. Need to make sure that setting up N idmapped marks
    does not have O(N) performance overhead on the system
3. Need to make sure that the idmapped marks proposal is deemed
    desired by concrete container manager project and use cases

If there are no objections to this roadmap, I will prepare the
proposal on my next context switch.

Yes, sounds good.

Christian