Re: [PATCH] mm: mmap_lock: fix use-after-free race and css ref leak in tracepoints

[Axel Rasmussen <axelrasmussen@google.com>]

On Tue, Dec 1, 2020 at 10:42 AM Shakeel Butt <shakeelb@google.com> wrote:
>
> On Tue, Dec 1, 2020 at 9:56 AM Greg Thelen <gthelen@google.com> wrote:
> >
> > Axel Rasmussen <axelrasmussen@google.com> wrote:
> >
> > > On Mon, Nov 30, 2020 at 5:34 PM Shakeel Butt <shakeelb@google.com> wrote:
> > >>
> > >> On Mon, Nov 30, 2020 at 3:43 PM Axel Rasmussen <axelrasmussen@google.com> wrote:
> > >> >
> > >> > syzbot reported[1] a use-after-free introduced in 0f818c4bc1f3. The bug
> > >> > is that an ongoing trace event might race with the tracepoint being
> > >> > disabled (and therefore the _unreg() callback being called). Consider
> > >> > this ordering:
> > >> >
> > >> > T1: trace event fires, get_mm_memcg_path() is called
> > >> > T1: get_memcg_path_buf() returns a buffer pointer
> > >> > T2: trace_mmap_lock_unreg() is called, buffers are freed
> > >> > T1: cgroup_path() is called with the now-freed buffer
> > >>
> > >> Any reason to use the cgroup_path instead of the cgroup_ino? There are
> > >> other examples of trace points using cgroup_ino and no need to
> > >> allocate buffers. Also cgroup namespace might complicate the path
> > >> usage.
> > >
> > > Hmm, so in general I would love to use a numeric identifier instead of a string.
> > >
> > > I did some reading, and it looks like the cgroup_ino() mainly has to
> > > do with writeback, instead of being just a general identifier?
> > > https://www.kernel.org/doc/Documentation/cgroup-v2.txt
>
> I think you are confusing cgroup inodes with real filesystem inodes in that doc.
>
> > >
> > > There is cgroup_id() which I think is almost what I'd want, but there
> > > are a couple problems with it:
> > >
> > > - I don't know of a way for userspace to translate IDs -> paths, to
> > > make them human readable?
> >
> > The id => name map can be built from user space with a tree walk.
> > Example:
> >
> > $ find /sys/fs/cgroup/memory -type d -printf '%i %P\n'                                                                          # ~ [main]
> > 20387 init.scope
> > 31 system.slice
> >
> > > - Also I think the ID implementation we use for this is "dense",
> > > meaning if a cgroup is removed, its ID is likely to be quickly reused.
> > >
>
> The ID for cgroup nodes (underlying it is kernfs) are allocated from
> idr_alloc_cyclic() which gives new ID after the last allocated ID and
> wrap after around INT_MAX IDs. So, likeliness of repetition is very
> low. Also the file_handle returned by name_to_handle_at() for cgroupfs
> returns the inode ID which gives confidence to the claim of low chance
> of ID reusing.

Ah, for some reason I remembered it using idr_alloc(), but you're
right, it does use cyclical IDs. Even so, tracepoints which expose
these IDs would still be difficult to use I think. Say we're trying to
collect a histogram of lock latencies over the course of some test
we're running. At the end, we want to produce some kind of
human-readable report.

cgroups may come and go throughout the test. Even if we never re-use
IDs, in order to be able to map all of them to human-readable paths,
it seems like we'd need some background process to poll the
/sys/fs/cgroup/memory directory tree as Greg described, keeping track
of the ID<->path mapping. This seems expensive, and even if we poll
relatively frequently we might still miss short-lived cgroups.

Trying to aggregate such statistics across physical machines, or
reboots of the same machine, is further complicated. The machine(s)
may be running the same application, which runs in a container with
the same path, but it'll end up with different IDs. So we'd have to
collect the ID<->path mapping from each, and then try to match up the
names for aggregation.