Re: [PATCH] pasta: Don't try to watch namespaces in procfs with inotify, use timer instead

[Stefano Brivio <sbrivio@redhat.com>]

On Mon, 19 Feb 2024 13:35:48 +1100
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Sat, Feb 17, 2024 at 02:34:57PM +0100, Stefano Brivio wrote:
> > We watch network namespace entries to detect when we should quit
> > (unless --no-netns-quit is passed), and these might stored in a tmpfs
> > typically mounted at /run/user/UID or /var/run/user/UID, or found in
> > procfs at /proc/PID/ns/.
> > 
> > Currently, we try to use inotify for any possible location of those
> > entries, but inotify, of course, doesn't work on pseudo-filesystems
> > (see inotify(7)).
> > 
> > The man page reflects this: the description of --no-netns-quit
> > implies that we won't quit anyway if the namespace is not "bound to
> > the filesystem".
> > 
> > Well, we won't quit, but, since commit 9e0dbc894813 ("More
> > deterministic detection of whether argument is a PID, PATH or NAME"),
> > we try. And, indeed, this is harmless, as the caveat from that
> > commit message states.
> > 
> > Now, it turns out that Buildah, a tool to create container images,
> > sharing its codebase with Podman, passes a procfs entry to pasta, and
> > expects pasta to exit once the network namespace is not needed
> > anymore, that is, once the original Buildah process terminates.
> > 
> > Get this to work by using the timer fallback mechanism if the
> > namespace name is passed as a path belonging to a pseudo-filesystem.
> > This is expected to be procfs, but I covered sysfs and devpts
> > pseudo-filesystems as well, because nothing actually prevents
> > creating this kind of directory structure and links there.
> > 
> > Note that statfs(), according to some versions of man pages, was
> > apparently "deprecated" by the LSB. My reasoning for using it is
> > essentially this:
> >   https://lore.kernel.org/linux-man/f54kudgblgk643u32tb6at4cd3kkzha6hslahv24szs4raroaz@ogivjbfdaqtb/t/#u
> > 
> > ...that is, there was no such thing as an LSB deprecation, and
> > anyway there's no other way to get the filesystem type.  
> 
> Huh, weird.
> 
> > Also note that, while it might sound more robust to detect the
> > filesystem type using fstatfs() on the file descriptor
> > (c->pasta_netns_fd), the reported filesystem type for it is nsfs, no
> > matter what path was given to pasta. If we use the parent directory,
> > we'll typically have either tmpfs or procfs reported.  
> 
> We could, however, move the opening of the parent directory earlier
> and use fstatfs() on that instead of statfs() on the path.

Done in v3.

> > The timer, however, still uses the file descriptor of the parent
> > directory later, as it has no access to the filesystem, and that
> > avoids possible races if the PID is recycled: if the original process
> > terminates, the handle we have on /proc/PID/ns still refers to it,
> > not to any other process with the same PID.  
> 
> True, but it's not obvious to me how that's relevant to this patch.
> 
> > We could have used pidfd_open() to get a handle on the parent
> > process. But it's not guaranteed that the parent process is actually
> > the one associated to the network namespace we operate on, and if we
> > get a PID file descriptor for a PID (parent or not) or path that was
> > given on our command line, this inherently causes a race condition as
> > that PID might have been recycled by the time we call pidfd_open().  
> 
> Again, not really sure of the relevance of this to the patch at hand.

Right, sorry, in both paragraphs I took the context of the Podman issue
for granted. Changed in v3.

> > Even assuming the process we want to watch is the parent process, and
> > a race-free usage of pidfd_open() would have been possible, I'm not
> > very enthusiastic about enabling yet another system call in the
> > seccomp profile just for this, while openat() is needed anyway.
> > 
> > Update the man page to reflect that, even if the target network
> > namespace is passed as a procfs path or a PID, we'll now quit when
> > the procfs entry is gone.
> > 
> > Reported-by: Paul Holzinger <pholzing@redhat.com>
> > Link: https://github.com/containers/podman/pull/21563#issuecomment-1948200214
> > Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
> > ---
> >  passt.1 |  8 ++++++--
> >  pasta.c | 15 +++++++++++++--
> >  2 files changed, 19 insertions(+), 4 deletions(-)
> > 
> > diff --git a/passt.1 b/passt.1
> > index dc2d719..de6e3bf 100644
> > --- a/passt.1
> > +++ b/passt.1
> > @@ -550,8 +550,12 @@ without \-\-userns.
> >  
> >  .TP
> >  .BR \-\-no-netns-quit
> > -If the target network namespace is bound to the filesystem (that is, if PATH or
> > -NAME are given as target), do not exit once the network namespace is deleted.
> > +If the target network namespace is bound to the filesystem, do not exit once
> > +that path is deleted.
> > +
> > +If the target network namespace is represented by a procfs entry, do not exit
> > +once that entry is removed from procfs (representing the fact that a process
> > +with the given PID terminated).  
> 
> Can't you now simplify this all to "Do not exit when the target
> network namespace ends"?

That was my first attempt, but as I was writing something of that sort,
I realised it's actually false: as long as pasta is using the network
namespace, the namespace exists and networking is still up and running.

One could delete the entry in tmpfs, or the original process could
terminate, join the network namespace after that, and keep using it.

So I guess it's better to clarify what are our conditions for
terminating (and, most likely, terminating the target namespace as a
consequence).

> >  .TP
> >  .BR \-\-config-net
> > diff --git a/pasta.c b/pasta.c
> > index 01d1511..4110917 100644
> > --- a/pasta.c
> > +++ b/pasta.c
> > @@ -33,6 +33,7 @@
> >  #include <sys/timerfd.h>
> >  #include <sys/types.h>
> >  #include <sys/stat.h>
> > +#include <sys/statfs.h>
> >  #include <fcntl.h>
> >  #include <sys/wait.h>
> >  #include <signal.h>
> > @@ -41,6 +42,7 @@
> >  #include <netinet/in.h>
> >  #include <net/ethernet.h>
> >  #include <sys/syscall.h>
> > +#include <linux/magic.h>
> >  
> >  #include "util.h"
> >  #include "passt.h"
> > @@ -390,12 +392,18 @@ void pasta_netns_quit_init(const struct ctx *c)
> >  	union epoll_ref ref = { .type = EPOLL_TYPE_NSQUIT_INOTIFY };
> >  	struct epoll_event ev = { .events = EPOLLIN };
> >  	int flags = O_NONBLOCK | O_CLOEXEC;
> > -	int fd;
> > +	struct statfs s = { 0 };  
> 
> I don't think you need to initialise this, since you should only be
> reading it in the case that statfs() succeeds.

There's no functional need, but I don't want to expose our stack memory
to the kernel.

Even though the kernel is generally able to control the process, I feel
like it's nice to avoid it, and I think it's consistent with most of
our system call usage.

In a number of cases, valgrind might complain about uninitialised
bytes, too.

> > +	bool try_inotify = true;
> > +	int fd = -1;
> >  
> >  	if (c->mode != MODE_PASTA || c->no_netns_quit || !*c->netns_base)
> >  		return;
> >  
> > -	if ((fd = inotify_init1(flags)) < 0)
> > +	if (statfs(c->netns_dir, &s)     || s.f_type == DEVPTS_SUPER_MAGIC ||
> > +	    s.f_type == PROC_SUPER_MAGIC || s.f_type == SYSFS_MAGIC)
> > +		try_inotify = false;
> > +
> > +	if (try_inotify && (fd = inotify_init1(flags)) < 0)
> >  		warn("inotify_init1(): %s, use a timer", strerror(errno));
> >  
> >  	if (fd >= 0 && inotify_add_watch(fd, c->netns_dir, IN_DELETE) < 0) {
> > @@ -463,6 +471,9 @@ void pasta_netns_quit_timer_handler(struct ctx *c, union epoll_ref ref)
> >  
> >  	fd = openat(ref.nsdir_fd, c->netns_base, O_PATH | O_CLOEXEC);
> >  	if (fd < 0) {
> > +		if (errno == EACCES)	/* Expected for existing procfs entry */
> > +			return;  
> 
> This seems like an unrelated fix.

It's actually related because, before this patch, procfs entries aren't
handled by the timer, and with tmpfs entries you don't get EACCES.

> > +
> >  		info("Namespace %s is gone, exiting", c->netns_base);
> >  		exit(EXIT_SUCCESS);
> >  	}  

-- 
Stefano