On Mon, Mar 25, 2019 at 10:36 AM Joel Fernandes [off-list ref] wrote:

On Mon, Mar 25, 2019 at 09:48:43AM -0700, Daniel Colascione wrote:

quoted

On Mon, Mar 25, 2019 at 9:21 AM Christian Brauner [off-list ref] wrote:

quoted

The pidctl() syscalls builds on, extends, and improves translate_pid() [4].
I quote Konstantins original patchset first that has already been acked and
picked up by Eric before and whose functionality is preserved in this
syscall. Multiple people have asked when this patchset will be sent in
for merging (cf. [1], [2]). It has recently been revived by Nagarathnam
Muthusamy from Oracle [3].

The intention of the original translate_pid() syscall was twofold:
1. Provide translation of pids between pid namespaces
2. Provide implicit pid namespace introspection

Both functionalities are preserved. The latter task has been improved
upon though. In the original version of the pachset passing pid as 1
would allow to deterimine the relationship between the pid namespaces.
This is inherhently racy. If pid 1 inside a pid namespace has died it
would report false negatives. For example, if pid 1 inside of the target
pid namespace already died, it would report that the target pid
namespace cannot be reached from the source pid namespace because it
couldn't find the pid inside of the target pid namespace and thus
falsely report to the user that the two pid namespaces are not related.
This problem is simple to avoid. In the new version we simply walk the
list of ancestors and check whether the namespace are related to each
other. By doing it this way we can reliably report what the relationship
between two pid namespace file descriptors looks like.

Additionally, this syscall has been extended to allow the retrieval of
pidfds independent of procfs. These pidfds can e.g. be used with the new
pidfd_send_signal() syscall we recently merged. The ability to retrieve
pidfds independent of procfs had already been requested in the
pidfd_send_signal patchset by e.g. Andrew [4] and later again by Alexey
[5]. A use-case where a kernel is compiled without procfs but where
pidfds are still useful has been outlined by Andy in [6]. Regular
anon-inode based file descriptors are used that stash a reference to
struct pid in file->private_data and drop that reference on close.

With this translate_pid() has three closely related but still distinct
functionalities. To clarify the semantics and to make it easier for
userspace to use the syscall it has:
- gained a command argument and three commands clearly reflecting the
  distinct functionalities (PIDCMD_QUERY_PID, PIDCMD_QUERY_PIDNS,
  PIDCMD_GET_PIDFD).
- been renamed to pidctl()

[snip]

quoted

Also, I'm still confused about how metadata access is supposed to work
for these procfs-less pidfs. If I use PIDCMD_GET_PIDFD on a process,
You snipped out a portion of a previous email in which I asked about
your thoughts on this question. With the PIDCMD_GET_PIDFD command in
place, we have two different kinds of file descriptors for processes,
one derived from procfs and one that's independent. The former works
with openat(2). The latter does not. To be very specific; if I'm
writing a function that accepts a pidfd and I get a pidfd that comes
from PIDCMD_GET_PIDFD, how am I supposed to get the equivalent of
smaps or oom_score_adj or statm for the named process in a race-free
manner?

This is true, that such usecase will not be supportable.  But the advantage
on the other hand, is that suchs "pidfd" can be made pollable or readable in
the future. Potentially allowing us to return exit status without a new
syscall (?). And we can add IOCTLs to the pidfd descriptor which we cannot do
with proc.

I don't like the idea of having one kind of pollfd be pollable and
another not. Such an interface would be confusing for users. If, as
you suggest below, we instead make the procfs-less FD the only thing
we call a "pidfd", then this proposal becomes less confusing and more
viable.

But.. one thing we could do for Daniel usecase is if a /proc/pid directory fd
can be translated into a "pidfd" using another syscall or even a node, like
/proc/pid/handle or something. I think this is what Christian suggested in
the previous threads.

/proc/pid/handle, if I understand correctly, "translates" a
procfs-based pidfd to a non-pidfd one. The needed interface would have
to perform the opposite translation, providing a procfs directory for
a given pidfd.

And also for the translation the other way, add a syscall or modify
translate_fd or something, to covert a anon_inode pidfd into a /proc/pid
directory fd. Then the user is welcomed to do openat(2) on _that_ directory fd.
Then we modify pidfd_send_signal to only send signals to pure pidfd fds, not
to /proc/pid directory fds.

This approach would work, but there's one subtlety to take into
account: which procfs? You'd want to take, as inputs, 1) the procfs
root you want, and 2) the pidfd, this way you could return to the user
a directory FD in the right procfs.

Should we work on patches for these? Please let us know if this idea makes
sense and thanks a lot for adding us to the review as well.

I would strongly prefer that we not merge pidctl (or whatever it is)
without a story for metadata access, be it your suggestion or
something else.