Message-ID: <CAKOZuesw9JK6DLhFTxN74rH5PDqGn7qbCQT4aaP1qtahSoRZ8w@mail.gmail.com>
Date:   Fri, 19 Apr 2019 16:29:54 -0700
From:   Daniel Colascione <dancol@...gle.com>
To:     Christian Brauner <christian@...uner.io>
Cc:     Joel Fernandes <joel@...lfernandes.org>,
        Jann Horn <jannh@...gle.com>, Oleg Nesterov <oleg@...hat.com>,
        Florian Weimer <fweimer@...hat.com>,
        kernel list <linux-kernel@...r.kernel.org>,
        Andy Lutomirski <luto@...capital.net>,
        Steven Rostedt <rostedt@...dmis.org>,
        Suren Baghdasaryan <surenb@...gle.com>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        Alexey Dobriyan <adobriyan@...il.com>,
        Al Viro <viro@...iv.linux.org.uk>,
        Andrei Vagin <avagin@...il.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Arnd Bergmann <arnd@...db.de>,
        "Eric W. Biederman" <ebiederm@...ssion.com>,
        Kees Cook <keescook@...omium.org>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        "open list:KERNEL SELFTEST FRAMEWORK" 
        <linux-kselftest@...r.kernel.org>, Michal Hocko <mhocko@...e.com>,
        Nadav Amit <namit@...are.com>, Serge Hallyn <serge@...lyn.com>,
        Shuah Khan <shuah@...nel.org>,
        Stephen Rothwell <sfr@...b.auug.org.au>,
        Taehee Yoo <ap420073@...il.com>, Tejun Heo <tj@...nel.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        kernel-team <kernel-team@...roid.com>,
        Tycho Andersen <tycho@...ho.ws>
Subject: Re: [PATCH RFC 1/2] Add polling support to pidfd

On Fri, Apr 19, 2019 at 4:02 PM Christian Brauner <christian@...uner.io> wrote:
>
> On Sat, Apr 20, 2019 at 12:35 AM Daniel Colascione <dancol@...gle.com> wrote:
> >
> > On Fri, Apr 19, 2019 at 2:48 PM Christian Brauner <christian@...uner.io> wrote:
> > >
> > > On Fri, Apr 19, 2019 at 11:21 PM Daniel Colascione <dancol@...gle.com> wrote:
> > > >
> > > > On Fri, Apr 19, 2019 at 1:57 PM Christian Brauner <christian@...uner.io> wrote:
> > > > >
> > > > > On Fri, Apr 19, 2019 at 10:34 PM Daniel Colascione <dancol@...gle.com> wrote:
> > > > > >
> > > > > > On Fri, Apr 19, 2019 at 12:49 PM Joel Fernandes <joel@...lfernandes.org> wrote:
> > > > > > >
> > > > > > > On Fri, Apr 19, 2019 at 09:18:59PM +0200, Christian Brauner wrote:
> > > > > > > > On Fri, Apr 19, 2019 at 03:02:47PM -0400, Joel Fernandes wrote:
> > > > > > > > > On Thu, Apr 18, 2019 at 07:26:44PM +0200, Christian Brauner wrote:
> > > > > > > > > > On April 18, 2019 7:23:38 PM GMT+02:00, Jann Horn <jannh@...gle.com> wrote:
> > > > > > > > > > >On Wed, Apr 17, 2019 at 3:09 PM Oleg Nesterov <oleg@...hat.com> wrote:
> > > > > > > > > > >> On 04/16, Joel Fernandes wrote:
> > > > > > > > > > >> > On Tue, Apr 16, 2019 at 02:04:31PM +0200, Oleg Nesterov wrote:
> > > > > > > > > > >> > >
> > > > > > > > > > >> > > Could you explain when it should return POLLIN? When the whole
> > > > > > > > > > >process exits?
> > > > > > > > > > >> >
> > > > > > > > > > >> > It returns POLLIN when the task is dead or doesn't exist anymore,
> > > > > > > > > > >or when it
> > > > > > > > > > >> > is in a zombie state and there's no other thread in the thread
> > > > > > > > > > >group.
> > > > > > > > > > >>
> > > > > > > > > > >> IOW, when the whole thread group exits, so it can't be used to
> > > > > > > > > > >monitor sub-threads.
> > > > > > > > > > >>
> > > > > > > > > > >> just in case... speaking of this patch it doesn't modify
> > > > > > > > > > >proc_tid_base_operations,
> > > > > > > > > > >> so you can't poll("/proc/sub-thread-tid") anyway, but iiuc you are
> > > > > > > > > > >going to use
> > > > > > > > > > >> the anonymous file returned by CLONE_PIDFD ?
> > > > > > > > > > >
> > > > > > > > > > >I don't think procfs works that way. /proc/sub-thread-tid has
> > > > > > > > > > >proc_tgid_base_operations despite not being a thread group leader.
> > > > > > > > > > >(Yes, that's kinda weird.) AFAICS the WARN_ON_ONCE() in this code can
> > > > > > > > > > >be hit trivially, and then the code will misbehave.
> > > > > > > > > > >
> > > > > > > > > > >@Joel: I think you'll have to either rewrite this to explicitly bail
> > > > > > > > > > >out if you're dealing with a thread group leader, or make the code
> > > > > > > > > > >work for threads, too.
> > > > > > > > > >
> > > > > > > > > > The latter case probably being preferred if this API is supposed to be
> > > > > > > > > > useable for thread management in userspace.
> > > > > > > > >
> > > > > > > > > At the moment, we are not planning to use this for sub-thread management. I
> > > > > > > > > am reworking this patch to only work on clone(2) pidfds which makes the above
> > > > > > > >
> > > > > > > > Indeed and agreed.
> > > > > > > >
> > > > > > > > > discussion about /proc a bit unnecessary I think. Per the latest CLONE_PIDFD
> > > > > > > > > patches, CLONE_THREAD with pidfd is not supported.
> > > > > > > >
> > > > > > > > Yes. We have no one asking for it right now and we can easily add this
> > > > > > > > later.
> > > > > > > >
> > > > > > > > Admittedly I haven't gotten around to reviewing the patches here yet
> > > > > > > > completely. But one thing about using POLLIN. FreeBSD is using POLLHUP
> > > > > > > > on process exit which I think is nice as well. How about returning
> > > > > > > > POLLIN | POLLHUP on process exit?
> > > > > > > > We already do things like this. For example, when you proxy between
> > > > > > > > ttys. If the process that you're reading data from has exited and closed
> > > > > > > > it's end you still can't usually simply exit because it might have still
> > > > > > > > buffered data that you want to read.  The way one can deal with this
> > > > > > > > from  userspace is that you can observe a (POLLHUP | POLLIN) event and
> > > > > > > > you keep on reading until you only observe a POLLHUP without a POLLIN
> > > > > > > > event at which point you know you have read
> > > > > > > > all data.
> > > > > > > > I like the semantics for pidfds as well as it would indicate:
> > > > > > > > - POLLHUP -> process has exited
> > > > > > > > - POLLIN  -> information can be read
> > > > > > >
> > > > > > > Actually I think a bit different about this, in my opinion the pidfd should
> > > > > > > always be readable (we would store the exit status somewhere in the future
> > > > > > > which would be readable, even after task_struct is dead). So I was thinking
> > > > > > > we always return EPOLLIN.  If process has not exited, then it blocks.
> > > > > >
> > > > > > ITYM that a pidfd polls as readable *once a task exits* and stays
> > > > > > readable forever. Before a task exit, a poll on a pidfd should *not*
> > > > > > yield POLLIN and reading that pidfd should *not* complete immediately.
> > > > > > There's no way that, having observed POLLIN on a pidfd, you should
> > > > > > ever then *not* see POLLIN on that pidfd in the future --- it's a
> > > > > > one-way transition from not-ready-to-get-exit-status to
> > > > > > ready-to-get-exit-status.
> > > > >
> > > > > What do you consider interesting state transitions? A listener on a pidfd
> > > > > in epoll_wait() might be interested if the process execs for example.
> > > > > That's a very valid use-case for e.g. systemd.
> > > >
> > > > Sure, but systemd is specialized.
> > >
> > > So is Android and we're not designing an interface for Android but for
> > > all of userspace.
> >
> > Nothing in my post is Android-specific. Waiting for non-child
> > processes is something that lots of people want to do, which is why
> > patches to enable it have been getting posted every few years for many
> > years (e.g., Andy's from 2011). I, too, want to make an API for all
> > over userspace. Don't attribute to me arguments that I'm not actually
> > making.
> >
> > > I hope this is clear. Service managers are quite important and systemd
> > > is the largest one
> > > and they can make good use of this feature.
> >
> > Service managers already have the tools they need to do their job. The
>
> No they don't. Even if they quite often have kludges and run into a lot
> of problems. That's why there's interest in these features as well.

Yes, and these facilities should have a powerful toolkit that they can
use to do their job in the right way. This toolkit will probably
involve pretty powerful kinds of process monitoring. I don't see a
reason to gate the ability to wait for process death via pidfd on that
toolkit. Please don't interpret my position as saying that the service
monitor usecase is unimportant or not worth adding to Linux.

> > kind of monitoring you're talking about is a niche case and an
> > improved API for this niche --- which amounts to a rethought ptrace
> > --- can wait for a future date, when it can be done right. Nothing in
> > the model I'm advocating precludes adding an event stream API in the
> > future. I don't think we should gate the ability to wait for process
> > exit via pidfd on pidfds providing an entire ptrace replacement
> > facility.
> >
> > > > There are two broad classes of programs that care about process exit
> > > > status: 1) those that just want to do something and wait for it to
> > > > complete, and 2) programs that want to perform detailed monitoring of
> > > > processes and intervention in their state. #1 is overwhelmingly more
> > > > common. The basic pidfd feature should take care of case #1 only, as
> > > > wait*() in file descriptor form. I definitely don't think we should be
> > > > complicating the interface and making it more error-prone (see below)
> > > > for the sake of that rare program that cares about non-exit
> > > > notification conditions. You're proposing a complicated combination of
> > > > poll bit flags that most users (the ones who just wait to wait for
> > > > processes) don't care about and that risk making the facility hard to
> > > > use with existing event loops, which generally recognize readability
> > > > and writability as the only properties that are worth monitoring.
> > >
> > > That whole pargraph is about dismissing a range of valid use-cases based on
> > > assumptions such as "way more common" and
> >
> > It really ought not to be controversial to say that process managers
> > make up a small fraction of the programs that wait for child
> > processes.
>
> Well, daemons tend to do those things do. System managers and container
> managers are just an example of a whole class. Even if you just consider system
> managers like openrc, systemd you have gotten yourself quite a large userbase.

When I said "niche", I didn't mean "unimportant". I meant
"specialized", as in the people who write these sorts of programs are
willing to dive in to low-level operational details and get things
right. I also think the community of people *writing* programs like
systemd is relatively small.

> > > even argues that service managers are special cases and therefore not
> > > really worth considering. I would like to be more open to other use cases.
> >
> > It's not my position that service managers are "not worth considering"
> > and you know that, so I'd appreciate your not attributing to me views
> > hat I don't hold. I *am* saying that an event-based process-monitoring
>
> It very much sounded like it. Calling them a "niche" case didn't help
> given that they run quite a lot of workloads everywhere.
>
> > API is out of scope and that it should be separate work: the
> > overwhelmingly majority of process manipulation (say, in libraries
> > wanting private helper processes, which is something I thought we all
> > agreed would be beneficial to support) is waiting for exit.
> >
> > > > > We can't use EPOLLIN for that too otherwise you'd need to to waitid(_WNOHANG)
> > > > > to check whether an exit status can be read which is not nice and then you
> > > > > multiplex different meanings on the same bit.
> > > > > I would prefer if the exit status can only be read from the parent which is
> > > > > clean and the least complicated semantics, i.e. Linus waitid() idea.
> > > >
> > > > Exit status information should be *at least* as broadly available
> > > > through pidfds as it is through the last field of /proc/pid/stat
> > > > today, and probably more broadly. I've been saying for six months now
> > > > that we need to talk about *who* should have access to exit status
> > > > information. We haven't had that conversation yet. My preference is to
> > >
> > > > just make exit status information globally available, as FreeBSD seems
> > > > to do. I think it would be broadly useful for something like pkill to
> > >
> > > From the pdfork() FreeBSD manpage:
> > > "poll(2) and select(2) allow waiting for process state transitions;
> > > currently only POLLHUP is defined, and will be raised when the process dies.
> > > Process state transitions can also be monitored using kqueue(2) filter
> > > EVFILT_PROCDESC; currently only NOTE_EXIT is implemented."
> >
> > I don't understand what you're trying to demonstrate by quoting that passage.
>
> FreeBSD obviously has thought about being able to observe
> more than just NOTE_EXIT in the future.

Yes. Did I say anywhere that we should never be able to observe execs
and forks? I think that what FreeBSD got *right* is making process
exit status broadly available. What I think they got wrong is the
mixing of exit information with other EVFILT_PROCDESC messages.

> > > > wait for processes to exit and to retrieve their exit information.
> > > >
> > > > Speaking of pkill: AIUI, in your current patch set, one can get a
> > > > pidfd *only* via clone. Joel indicated that he believes poll(2)
> > > > shouldn't be supported on procfs pidfds. Is that your thinking as
> > > > well? If that's the case, then we're in a state where non-parents
> > >
> > > Yes, it is.
> >
> > If reading process status information from a pidfd is destructive,
> > it's dangerous to share pidfds between processes. If reading
> > information *isn't* destructive, how are you supposed to use poll(2)
> > to wait for the next transition? Is poll destructive? If you can only
> > make a new pidfd via clone, you can't get two separate event streams
> > for two different users. Sharing a single pidfd via dup or SCM_RIGHTS
> > becomes dangerous, because if reading status is destructive, only one
> > reader can observe each event. Your proposed edge-triggered design
> > makes pidfds significantly less useful, because in your design, it's
> > unsafe to share a single pidfd open file description *and* there's no
> > way to create a new pidfd open file description for an existing
> > process.
> >
> > I think we should make an API for all of userspace and not just for
> > container managers and systemd.
>
> I mean,  you can go and try making arguments based on syntactical
> rearrangements of things I said but I'm going to pass.

I'd prefer if we focused on substantive technical issues instead of
unproductive comments on syntax. I'd prefer to talk about the
technical concerns I raised regarding an edge-triggered event-queue
design making pidfd file descriptor sharing dangerous.

> My point simply was: There are more users that would be interested
> in observing more state transitions in the future.
> Your argument made it sound like they are not worth considering.
> I disagree.

I believe you've misunderstood me. I've never said that this use case
is "not worth considering". I would appreciate it if you wouldn't
claim that I believe these use cases aren't worth considering. My
point is that these uses would be better served through a dedicated
process event monitoring facility, one that could replace ptrace. I
would be thrilled by something like that. The point I'm making, to be
very clear, is *NOT* that process monitoring is "not worth
considering", but that process monitoring is subtle and complicated
enough that it ought to be considered as a standalone project,
independent of pidfds proper and of the very simple and effective
pidfd system that Joel has proposed in his patch series.

> > > > can't wait for process exit, and providing this facility is an
> > > > important goal of the whole project.
> > >
> > > That's your goal.
> >
> > I thought we all agreed on that months ago that it's reasonable to
> > allow processes to wait for non-child processes to exit. Now, out of
>
> Uhm, I can't remember being privy to that agreement but the threads get
> so long that maybe I forgot what I wrote?
>
> > the blue, you're saying that 1) actually, we want a rich API for all
> > kinds of things that aren't process exit, because systemd, and 2)
>
> - I'm not saying we have to. It just makes it more flexible and is something
> we can at least consider.

I've spent a few emails and a lot of thought considering the idea. If
I weren't considering it, I wouldn't have thought through the
implications of destructive event reads above. My position, *after due
consideration*, is that we're better off making the pidfd poll()
system a level-triggered exit-only signal and defer more sophisticated
monitoring to a separate facility. I'm not saying that we shouldn't be
able to monitor processes. I'm saying that there are important
technical and API design reasons why the initial version of the pidfd
wait system should be simple and support only one thing. The model
you've proposed has important technical disadvantages, and you haven't
addressed these technical disadvantages.

> - systemd is an example of another *huge* user of this api. That neither implies
> this api is "because systemd" it simply makes it worth that we
> consider this use-case.
>
> > actually, non-parents shouldn't be able to wait for process death. I
>
> I'm sorry, who has agreed that a non-parent should be able to wait for
> process death?

Aleksa Sarai, for starters. See [1]. In any case, I don't see where
your sudden opposition to the idea is coming from. I've been saying
over and over that it's important that we allow processes to wait for
the death of non-children. Why are you just objecting now? And on what
basis are you objecting? Why *shouldn't* we be able to wait for
process death in a general way? Both FreeBSD and Windows can do it.

> I know you proposed that but has anyone ever substantially supported this?
> I'm happy if you can gather the necessary support for this but I just
> haven't seen that yet.

I recommend re-reading the whole exithand thread, since we covered a
lot of the ground that we're also covering here. In any case, there's
a legitimate technical reason why we'd want to wait for non-child
death, and I would appreciate it if you didn't just summarily dismiss
it as "[my] goal". Let's talk about what's best for users, not use
that kind of unproductive rhetoric.

[1] https://lore.kernel.org/lkml/20181101070036.l24c2p432ohuwmqf@yavin/

Powered by blists - more mailing lists