bug#41948: Shepherd deadlocks

[Ludovic Courtès]

Hi!

Mathieu Othacehe <othacehe@gnu.org> skribis:

> Those two finalizer threads share the same pipe. When we try to
> stop the finalizer thread in Shepherd, right before forking a new
> process, we send a '\1' byte to the finalizer pipe.
>
> 1     write(6, "\1", 1 <unfinished ...>
>
>
> which is received by (line 183597): 
>
> 253   <... read resumed>"\1", 1)        = 1
>
> the marionette finalizer thread. Then, we pthread_join the Shepherd
> finalizer thread, which never stops! Quite unfortunate.

While working on a fix for this issue (finalizer pipe shared between
parent and child process), I found the ‘sleep_pipe’ of the main thread
is also shared between the parent and its child.

Attached is a reproducer.  It prints something like this before hanging:

--8<---------------cut here---------------start------------->8---
$ guile ~/src/guile-debugging/signal-pipe.scm
parent: 25947
child: 25953
alarm in parent!
alarm in parent!
alarm in parent!

[...]

alarm in parent!
alarm in parent!
child woken up!
--8<---------------cut here---------------end--------------->8---

“child woken up” means that it’s the child process that won the race
reading on the sleep pipe (from ‘scm_std_select’).

The parent process then hangs because, in ‘scm_std_select’, it did:

  1. select(1), which returned due to available data on ‘wakeup_fd’;

  2. ‘full_read (wakeup_fd, &dummy, 1)’ gets stuck forever in read(2)
     because the child process read that byte in the meantime so
     there’s nothing left to read.

Here’s the sequence:

--8<---------------cut here---------------start------------->8---
25947 select(4, [3], NULL, NULL, {tv_sec=0, tv_usec=100000}) = 0 (Timeout)
25947 getpid()                          = 25947
25947 kill(25947, SIGALRM)              = 0
25947 --- SIGALRM {si_signo=SIGALRM, si_code=SI_USER, si_pid=25947, 
si_uid=1000} ---
25947 write(8, "\16", 1)                = 1
25947 rt_sigreturn({mask=[]})           = 0
25952 <... read resumed>"\16", 1)       = 1
25947 rt_sigprocmask(SIG_BLOCK, NULL,  <unfinished ...>
25952 write(4, "\0", 1 <unfinished ...>
25947 <... rt_sigprocmask resumed>[], 8) = 0
25953 <... select resumed>)             = 1 (in [3], left {tv_sec=0, 
tv_usec=346370})
25952 <... write resumed>)              = 1
25947 select(4, [3], NULL, NULL, {tv_sec=0, tv_usec=100000} <unfinished ...>
25953 read(3,  <unfinished ...>
25952 rt_sigprocmask(SIG_BLOCK, NULL,  <unfinished ...>
25947 <... select resumed>)             = 1 (in [3], left {tv_sec=0, 
tv_usec=99999})
25953 <... read resumed>"\0", 1)        = 1
25947 read(3,  <unfinished ...>
25952 <... rt_sigprocmask resumed>~[KILL STOP PWR RTMIN RT_1], 8) = 0
25953 write(1, "child woken up!\n", 16 <unfinished ...>
25952 read(7,  <unfinished ...>
--8<---------------cut here---------------end--------------->8---

Notice how “\16” (= SIGALRM) is written by the parent’s main thread and
read by the child’s main thread.

Thoughts?

Ludo’.

;; https://issues.guix.gnu.org/41948

(use-modules (ice-9 match))

(setvbuf (current-output-port) 'line)
(sigaction SIGCHLD pk)                            ;start signal thread

(match (primitive-fork)
  (0
   (format #t "child: ~a~%" (getpid))
   (let loop ()
     (unless (zero? (usleep 500000))
       ;; If this happens, it means the select(2) call in 'scm_std_select'
       ;; returned because one of our file descriptors had input data
       ;; available (which shouldn't happen).
       (format #t "child woken up!~%"))
     (loop)))
  (pid
   (format #t "parent: ~a~%" (getpid))
   (sigaction SIGALRM (lambda _
                        (format #t "alarm in parent!~%")))
   (let loop ()
     (kill (getpid) SIGALRM)
     (usleep 100000)
     (loop))))