PATCH: add -q option to kill for sigqueue

PATCH: add -q option to kill for sigqueue

[Oliver Kiddle]

The following adds support for a -q option to kill which takes a numeric
argument. It then uses sigqueue(2) to send the signal with the provided
number which a signal handler can pick up as si_value.sival_int in the
siginfo_t parameter. The util-linux kill also supports a similar option.
sigqueue and a range of "realtime" signals are specified in POSIX and
well supported. As far as I can tell, the RT signals are intended for
user-defined purposes like SIGUSR1 and USR2. The value can be an int or
a pointer but in practice, sending a pointer is fairly useless given
that signals are usually sent to a different process.

We could potentially make the signal value and other details such as
si_pid available to trap handlers. Perhaps using namespaced variables -
e.g. .signal.pid and .signal.value? However, I don't really understand
enough about how trap is implemented to know how (or whether) that can
be done safely.

It also ought to be possible to make SIGRTMIN through to SIGRTMAX
available. util-linux's kill allows those to be specified as RT<n>,
RTMIN+<n> or RTMAX-<n>. Would that need us to muck around with
signames2.awk or is leaving them out of $signals / kill -l and using
#ifdef RTMIN sufficient?

Oliver

diff --git a/Src/jobs.c b/Src/jobs.c
index a3b9f667a..4e9767ee4 100644
--- a/Src/jobs.c
+++ b/Src/jobs.c
@@ -2677,9 +2677,35 @@ bin_kill(char *nam, char **argv, UNUSED(Options ops), UNUSED(int func))
 {
     int sig = SIGTERM;
     int returnval = 0;
+#ifdef HAVE_SIGQUEUE
+    union sigval sigqueue_info;
+#endif
+    int use_sigqueue = 0, got_sig = 0;
+
+    while (*argv && **argv == '-') {
+	if (!use_sigqueue && (*argv)[1] == 'q' && (*argv)[2] == '\0') {
+	    char *endp;
+
+	    if (!*++argv) {
+		zwarnnam(nam, "-q: argument expected");
+		return 1;
+	    }
+#ifdef HAVE_SIGQUEUE
+	    sigqueue_info.sival_int =
+#endif
+		    zstrtol(*argv, &endp, 10);
+	    if (*endp) {
+		zwarnnam(nam, "invalid number: %s", *argv);
+		return 1;
+	    }
+	    use_sigqueue = 1;
+	    argv++;
+	    continue;
+	}
+	if (got_sig)
+	    break;
 
     /* check for, and interpret, a signal specifier */
-    if (*argv && **argv == '-') {
 	if (idigit((*argv)[1])) {
 	    char *endp;
 	    /* signal specified by number */
@@ -2796,6 +2822,7 @@ bin_kill(char *nam, char **argv, UNUSED(Options ops), UNUSED(int func))
 	    }
 	}
 	argv++;
+	got_sig = 1;
     }
 
     /* Discard the standard "-" and "--" option breaks */
@@ -2844,7 +2871,12 @@ bin_kill(char *nam, char **argv, UNUSED(Options ops), UNUSED(int func))
 	    returnval++;
 	} else {
 	    int pid = atoi(*argv);
-	    if (kill(pid, sig) == -1) {
+	    if (
+#ifdef HAVE_SIGQUEUE
+		use_sigqueue ? sigqueue(pid, sig, sigqueue_info) :
+#endif
+		kill(pid, sig) == -1)
+	    {
 		zwarnnam("kill", "kill %s failed: %e", *argv, errno);
 		returnval++;
 	    } 
diff --git a/configure.ac b/configure.ac
index c5263035e..9cb6e032b 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1299,6 +1299,7 @@ AC_CHECK_FUNCS(strftime strptime mktime timelocal \
 	       mkfifo _mktemp mkstemp \
 	       waitpid wait3 \
 	       sigaction sigblock sighold sigrelse sigsetmask sigprocmask \
+	       sigqueue \
 	       killpg setpgid setpgrp tcsetpgrp tcgetattr nice \
 	       gethostname gethostbyname2 getipnodebyname \
 	       inet_aton inet_pton inet_ntop \

Re: PATCH: add -q option to kill for sigqueue

[Oliver Kiddle]

On 25 Nov, I wrote:
> The following adds support for a -q option to kill which takes a numeric
> argument. It then uses sigqueue(2) to send the signal with the provided

And with the following addition, the documentation and completion also
cover the option. Seems it isn't a system call on Linux so the man page
there is sigqueue(3).

Oliver

diff --git a/Completion/Zsh/Command/_kill b/Completion/Zsh/Command/_kill
index b9dfde3f0..084cf42c8 100644
--- a/Completion/Zsh/Command/_kill
+++ b/Completion/Zsh/Command/_kill
@@ -5,6 +5,7 @@ typeset -A opt_args
 
 _arguments -C \
   '(-s -l 1)-n[specify signal number]:signal number' \
+  '(-l)-q[send the specified integer with the signal using sigqueue]:value' \
   '(-n -l 1)-s[specify signal name]:signal:_signals -s' \
   '(-n -s)-l[list signal names or numbers of specified signals]:*:signal:_signals' \
   '(-n -s -l)1::signal:_signals -p -s' \
diff --git a/Doc/Zsh/builtins.yo b/Doc/Zsh/builtins.yo
index 8f310f6cf..81694225d 100644
--- a/Doc/Zsh/builtins.yo
+++ b/Doc/Zsh/builtins.yo
@@ -1142,7 +1142,7 @@ findex(kill)
 cindex(killing jobs)
 cindex(jobs, killing)
 xitem(tt(kill) [ tt(-s) var(signal_name) | tt(-n) var(signal_number) | \
-tt(-)var(sig) ] var(job) ...)
+tt(-)var(sig) ] [ tt(-q) var(value) ] var(job) ...)
 item(tt(kill) tt(-l) [ var(sig) ... ])(
 Sends either tt(SIGTERM) or the specified signal to the given
 jobs or processes.
@@ -1169,6 +1169,9 @@ tt(kill -IO) and tt(kill -POLL) have the same effect.
 
 Many systems will allow process IDs to be negative to kill a process
 group or zero to kill the current process group.
+
+The tt(-q) option allows an integer value to be sent with the signal
+on systems that support tt(sigqueue+LPAR()RPAR()).
 )
 findex(let)
 item(tt(let) var(arg) ...)(

Re: PATCH: add -q option to kill for sigqueue

[Bart Schaefer]

Missed the questions in this before ...

On Fri, Nov 24, 2023 at 3:46 PM Oliver Kiddle <opk@zsh.org> wrote:
>
> We could potentially make the signal value and other details such as
> si_pid available to trap handlers. Perhaps using namespaced variables -
> e.g. .signal.pid and .signal.value?

The signal is first picked up by signals.c:zhandler(), and thereafter
only the signal number is passed down the call chain to
signals.c:dotrapargs() which arranges to call every trap function with
its signal number as $1.  So you probably need to wrangle si_pid etc.
in zhandler(), but you have to be very careful about memory management
(ideally, don't do any).

> It also ought to be possible to make SIGRTMIN through to SIGRTMAX
> available. util-linux's kill allows those to be specified as RT<n>,
> RTMIN+<n> or RTMAX-<n>. Would that need us to muck around with
> signames2.awk or is leaving them out of $signals / kill -l and using
> #ifdef RTMIN sufficient?

No opinion on whether they should be left out, but it does appear that
signames2.awk would have to be tweaked, yes.