[Shepherd] Use of ‘waitpid’,_ ‘system*’, etc. in service code can cause deadlocks

DoneSubmitted by Ludovic Courtès.
Details
3 participants
  • Ludovic Courtès
  • Mathieu Othacehe
  • Maxime Devos
Owner
unassigned
Severity
important
Merged with
L
L
Ludovic Courtès wrote on 20 Jul 23:39 +0200
[Shepherd] Use of ‘waitpid’, ‘system* ’, etc. in service code can cause deadlocks
(address . bug-guix@gnu.org)
8735evpipv.fsf@inria.fr
Hi!

We’ve just had a bad experience with the nginx service on berlin, where
‘herd restart nginx’ would cause shepherd to get stuck forever in
‘waitpid’ on the process that was supposed to start nginx.

The details are unclear, but one thing is clear is that using ‘waitpid’
(either directly or indirectly with ‘system*’, which is what
‘nginx-service-type’ does) is not great:

1. In the best case, shepherd (as of 0.9.1) is stuck while ‘system*’
is in ‘waitpid’ waiting for child process completion (“stuck” as
in: doesn’t do anything, not even answering ‘herd’ requests or
inetd connections.)

2. I don’t think that can happen with ‘system*’ (because it’s in C),
but generally speaking, there’s a possibility that shepherd’s event
loop will handle child process termination before some other
user-made ‘waitpid’ call does.

Anyway, that’s a bad situation.

So I can think of several ways to address it:

1. Change the nginx service ‘stop’ method to just
(make-kill-destructor), which should work just as well as invoking
“nginx -s stop”.

2. Have Shepherd provide a replacement for ‘system*’.

Thoughts?

Ludo’.
L
L
Ludovic Courtès wrote on 20 Jul 23:43 +0200
control message for bug #56674
(address . control@debbugs.gnu.org)
87y1wno3yx.fsf@gnu.org
severity 56674 important
quit
M
M
Maxime Devos wrote on 21 Jul 01:48 +0200
Re: bug#56674: [Shepherd] Use of ‘waitpi d’, ‘system*’, etc. in service code can ca use deadlocks
c4045c06-2024-b49e-cee9-88dafd3612e6@telenet.be
On 20-07-2022 23:39, Ludovic Courtès wrote:
Toggle quote (29 lines)
> Hi!
>
> We’ve just had a bad experience with the nginx service on berlin, where
> ‘herd restart nginx’ would cause shepherd to get stuck forever in
> ‘waitpid’ on the process that was supposed to start nginx.
>
> The details are unclear, but one thing is clear is that using ‘waitpid’
> (either directly or indirectly with ‘system*’, which is what
> ‘nginx-service-type’ does) is not great:
>
> 1. In the best case, shepherd (as of 0.9.1) is stuck while ‘system*’
> is in ‘waitpid’ waiting for child process completion (“stuck” as
> in: doesn’t do anything, not even answering ‘herd’ requests or
> inetd connections.)
>
> 2. I don’t think that can happen with ‘system*’ (because it’s in C),
> but generally speaking, there’s a possibility that shepherd’s event
> loop will handle child process termination before some other
> user-made ‘waitpid’ call does.
>
> Anyway, that’s a bad situation.
>
> So I can think of several ways to address it:
>
> 1. Change the nginx service ‘stop’ method to just
> (make-kill-destructor), which should work just as well as invoking
> “nginx -s stop”.
>
> 2. Have Shepherd provide a replacement for ‘system*’.
Why Shepherd and not guile fibers? Is this a Shepherd-specific problem?
Toggle quote (2 lines)
>
> Thoughts?
3. Make waitpid (or a variant that does what we need) interact well with
guile-fibers, like how 'accept' is doesn't inhibit switching to another
fiber. There some Linux API with signal handlers or pid fds or such that
might be useful here, though I don't recall the name. Presumably
something similar can be done for the Hurd, though some C glue may be
needed to access the right Hurd APIs if the signal handler API isn't
portable.
Alternatively:
4. Do the waitpid in a separate thread (needs work-around for the
multi-threaded fork problem, probably C things? Or modifying Guile and
maybe glibc to avoid async-unsafe things or make more things async-safe
or whatever the appropriate ...-safe is here.)
If not a Guile Fibers interaction problem, then the asynchronous signal
handler API might still be useful.
Greetings,
Maxime
Attachment: OpenPGP_signature
L
L
Ludovic Courtès wrote on 21 Jul 17:39 +0200
Re: bug#56674: [Shepherd] Use of ‘waitpid’, ‘system*’, etc. in service code can cause deadlocks
(name . Maxime Devos)(address . maximedevos@telenet.be)(address . 56674@debbugs.gnu.org)
87fsiujwzo.fsf@gnu.org
Maxime Devos <maximedevos@telenet.be> skribis:

Toggle quote (2 lines)
> Why Shepherd and not guile fibers? Is this a Shepherd-specific problem?

Blocking calls are a problem for Fibers in general, and ‘waitpid’ is no
exception.

The problem here is Shepherd-specific in the sense that we’re more
likely to use ‘system*’ and ‘waitpid’ in this context. It’s also
Shepherd-specific because shepherd already runs an event loop that
tracks signal FDs and will thus “see” SIGCHLD events.

Toggle quote (8 lines)
> 3. Make waitpid (or a variant that does what we need) interact well
> with guile-fibers, like how 'accept' is doesn't inhibit switching to
> another fiber. There some Linux API with signal handlers or pid fds or
> such that might be useful here, though I don't recall the
> name. Presumably something similar can be done for the Hurd, though
> some C glue may be needed to access the right Hurd APIs if the signal
> handler API isn't portable.

Yes, that’s roughly what I had in mind when I mentioned providing a
replacement for ‘system*’ (but you’re right, it’s a replacement for
‘waitpid’ at its core).

Toggle quote (7 lines)
> Alternatively:
>
> 4. Do the waitpid in a separate thread (needs work-around for the
> multi-threaded fork problem, probably C things? Or modifying Guile and
> maybe glibc to avoid async-unsafe things or make more things
> async-safe or whatever the appropriate ...-safe is here.)

For shepherd, multithreading is not an option due to the semantics of
fork in the presence of threads.

Ludo’.
M
M
Maxime Devos wrote on 13 Aug 16:59 +0200
Re: bug#56674: [Shepherd] Use of ‘waitpi d’, ‘system*’, etc. in service code can ca use deadlocks
(name . Ludovic Courtès)(address . ludo@gnu.org)(address . 56674@debbugs.gnu.org)
f13570c6-95bc-548b-5084-8f727efe5881@telenet.be
On 21-07-2022 17:39, Ludovic Courtès wrote:
Toggle quote (8 lines)
>> Alternatively:
>>
>> 4. Do the waitpid in a separate thread (needs work-around for the
>> multi-threaded fork problem, probably C things? Or modifying Guile and
>> maybe glibc to avoid async-unsafe things or make more things
>> async-safe or whatever the appropriate ...-safe is here.)
> For shepherd, multithreading is not an option due to the semantics of
> fork in the presence of threads.
From what I've read, multi-threaded fork is safe as long as you do an
exec 'immediately' afterwards, without doing things like taking locks or
allocating memory with malloc in-between the fork and exec. I don't
think it's possible to do that in Guile code, but that's what the C
things are for.
Greetings,
Maxime.
Attachment: file
Attachment: OpenPGP_signature
M
M
Mathieu Othacehe wrote on 12 Nov 09:36 +0100
control message for bug #58926
(address . control@debbugs.gnu.org)
87leog603u.fsf@meije.mail-host-address-is-not-set
merge 58926 56674
quit
L
L
Ludovic Courtès wrote on 14 Nov 00:16 +0100
Re: bug#56674: [Shepherd] Use of ‘waitpid’, ‘system*’, etc. in service code can cause deadlocks
(address . 56674@debbugs.gnu.org)
87o7tabg2x.fsf@gnu.org
Hi,

Ludovic Courtès <ludo@gnu.org> skribis:

Toggle quote (20 lines)
> 1. In the best case, shepherd (as of 0.9.1) is stuck while ‘system*’
> is in ‘waitpid’ waiting for child process completion (“stuck” as
> in: doesn’t do anything, not even answering ‘herd’ requests or
> inetd connections.)
>
> 2. I don’t think that can happen with ‘system*’ (because it’s in C),
> but generally speaking, there’s a possibility that shepherd’s event
> loop will handle child process termination before some other
> user-made ‘waitpid’ call does.
>
> Anyway, that’s a bad situation.
>
> So I can think of several ways to address it:
>
> 1. Change the nginx service ‘stop’ method to just
> (make-kill-destructor), which should work just as well as invoking
> “nginx -s stop”.
>
> 2. Have Shepherd provide a replacement for ‘system*’.

These fresh Shepherd commits install a non-blocking ‘system*’ replacement:

975b0aa service: Provide a non-blocking replacement of 'system*'.
039c7a8 service: Spawn a fiber responsible for process monitoring.

We’ll have to do more testing and probably go for a 0.9.3 release soon.

Protip: you can test the latest shepherd with:

Toggle snippet (18 lines)
(operating-system
;; …
(essential-services
(modify-services (operating-system-default-essential-services
this-operating-system)
(shepherd-root-service-type
config =>
(shepherd-configuration
(shepherd (package
(inherit shepherd-0.9)
(version "0.9.3pre")
(source (git-checkout
(url "https://git.savannah.gnu.org/git/shepherd.git")))
(native-inputs
(modify-inputs (package-native-inputs shepherd-0.9)
(append autoconf automake help2man texinfo gnu-gettext))))))))))

Full example attached.

Ludo’.
;; This is an operating system configuration template
;; for a "bare bones" setup, with no X11 display server.

(use-modules (gnu) (guix) (guix git))
(use-service-modules networking ssh web vpn shepherd)
(use-package-modules linux screen ssh
admin autotools gettext man texinfo)

(operating-system
(host-name "komputilo")
(timezone "Europe/Berlin")
(locale "en_US.utf8")

;; Boot in "legacy" BIOS mode, assuming /dev/sdX is the
;; target hard disk, and "my-root" is the label of the target
;; root file system.
(bootloader (bootloader-configuration
(bootloader grub-bootloader)
(targets '("/dev/sdX"))))
;; It's fitting to support the equally bare bones ‘-nographic’
;; QEMU option, which also nicely sidesteps forcing QWERTY.
(kernel-arguments (list "console=ttyS0,115200"))
(file-systems (cons (file-system
(device (file-system-label "my-root"))
(mount-point "/")
(type "ext4"))
%base-file-systems))

;; This is where user accounts are specified. The "root"
;; account is implicit, and is initially created with the
;; empty password.
(users (cons (user-account
(name "alice")
(comment "Bob's sister")
(group "users")

;; Adding the account to the "wheel" group
;; makes it a sudoer. Adding it to "audio"
;; and "video" allows the user to play sound
;; and access the webcam.
(supplementary-groups '("wheel"
"audio" "video")))
%base-user-accounts))

;; Globally-installed packages.
(packages (append (list screen strace) %base-packages))

(essential-services
(modify-services (operating-system-default-essential-services
this-operating-system)
(shepherd-root-service-type
config =>
(shepherd-configuration
(shepherd (package
(inherit shepherd-0.9)
(version "0.9.3pre")
(source (git-checkout
(native-inputs
(modify-inputs (package-native-inputs shepherd-0.9)
(append autoconf automake help2man texinfo gnu-gettext)))))))))

;; Add services to the baseline: a DHCP client and
;; an SSH server.
(services (append (list (service dhcp-client-service-type)
(service nginx-service-type
(nginx-configuration
(server-blocks
(list (nginx-server-configuration
(listen '("80"))
(server-name '("www.example.org"))
(root "/srv/whatever"))))))
(service wireguard-service-type
(wireguard-configuration
(addresses (list "10.0.0.2/24"))
(dns '("10.0.0.50")))) ;does not exit


(service openssh-service-type
(openssh-configuration
(openssh openssh-sans-x)
(port-number 2222))))
%base-services)))
L
L
Ludovic Courtès wrote on 14 Nov 17:32 +0100
Re: bug#58926: Shepherd becomes unresponsive after an interrupt
(address . 56674@debbugs.gnu.org)
87wn7xo5ss.fsf_-_@gnu.org
Hello!

Ludovic Courtès <ludo@gnu.org> skribis:

Toggle quote (7 lines)
> These fresh Shepherd commits install a non-blocking ‘system*’ replacement:
>
> 975b0aa service: Provide a non-blocking replacement of 'system*'.
> 039c7a8 service: Spawn a fiber responsible for process monitoring.
>
> We’ll have to do more testing and probably go for a 0.9.3 release soon.

Shepherd commit ada88074f0ab7551fd0f3dce8bf06de971382e79 passes my
tests. It definitely solves the wireguard example and similar things
(uses of ‘system*’ in service constructors/destructors); I can’t tell
for sure about nginx because I haven’t been able to reproduce it in a
VM. I’m interested in ways to reproduce it.

It does look like we could go with 0.9.3 real soon now.

Ludo’.
?
Your comment

Commenting via the web interface is currently disabled.

To comment on this conversation send email to 56674@debbugs.gnu.org