ash: fix ash-signals/signal8 testcase failure

function                                             old     new   delta
killcmd                                              109     224    +115
kill_main                                            882     910     +28
changepath                                           194     195      +1
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 3/0 up/down: 144/0)             Total: 144 bytes

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>

5 files changed, 130 insertions, 12 deletions

diff --git a/shell/ash.c b/shell/ash.c
index cccd6dd79..98d2c7c29 100644
--- a/shell/ash.c
+++ b/shell/ash.c
@@ -3783,18 +3783,51 @@ setjobctl(int on)
 static int FAST_FUNC
 killcmd(int argc, char **argv)
 {
-        int i = 1;
         if (argv[1] && strcmp(argv[1], "-l") != 0) {
+                int i = 1;
                 do {
                         if (argv[i][0] == '%') {
-                                struct job *jp = getjob(argv[i], 0);
-                                unsigned pid = jp->ps[0].ps_pid;
-                                /* Enough space for ' -NNN<nul>' */
-                                argv[i] = alloca(sizeof(int)*3 + 3);
-                                /* kill_main has matching code to expect
-                                 * leading space. Needed to not confuse
-                                 * negative pids with "kill -SIGNAL_NO" syntax */
-                                sprintf(argv[i], " -%u", pid);
+                                /*
+                                 * "kill %N" - job kill
+                                 * Converting to pgrp / pid kill
+                                 */
+                                struct job *jp;
+                                char *dst;
+                                int j, n;
+
+                                jp = getjob(argv[i], 0);
+                                /*
+                                 * In jobs started under job control, we signal
+                                 * entire process group by kill -PGRP_ID.
+                                 * This happens, f.e., in interactive shell.
+                                 *
+                                 * Otherwise, we signal each child via
+                                 * kill PID1 PID2 PID3.
+                                 * Testcases:
+                                 * sh -c 'sleep 1|sleep 1 & kill %1'
+                                 * sh -c 'true|sleep 2 & sleep 1; kill %1'
+                                 * sh -c 'true|sleep 1 & sleep 2; kill %1'
+                                 */
+                                n = jp->nprocs; /* can't be 0 (I hope) */
+                                if (jp->jobctl)
+                                        n = 1;
+                                dst = alloca(n * sizeof(int)*4);
+                                argv[i] = dst;
+                                for (j = 0; j < n; j++) {
+                                        struct procstat *ps = &jp->ps[j];
+                                        /* Skip non-running and not-stopped members
+                                         * (i.e. dead members) of the job
+                                         */
+                                        if (ps->ps_status != -1 && !WIFSTOPPED(ps->ps_status))
+                                                continue;
+                                        /*
+                                         * kill_main has matching code to expect
+                                         * leading space. Needed to not confuse
+                                         * negative pids with "kill -SIGNAL_NO" syntax
+                                         */
+                                        dst += sprintf(dst, jp->jobctl ? " -%u" : " %u", (int)ps->ps_pid);
+                                }
+                                *dst = '\0';
                         }
                 } while (argv[++i]);
         }
@@ -4227,8 +4260,9 @@ waitcmd(int argc UNUSED_PARAM, char **argv)
                                         break;
                                 job = job->prev_job;
                         }
-                } else
+                } else {
                         job = getjob(*argv, 0);
+                }
                 /* loop until process terminated or stopped */
                 while (job->state == JOBRUNNING)
                         blocking_wait_with_raise_on_sig();
@@ -4724,7 +4758,7 @@ forkchild(struct job *jp, union node *n, int mode)
 #if JOBS
         /* do job control only in root shell */
         doing_jobctl = 0;
-        if (mode != FORK_NOJOB && jp->jobctl && !oldlvl) {
+        if (mode != FORK_NOJOB && jp->jobctl && oldlvl == 0) {
                 pid_t pgrp;
 
                 if (jp->nprocs == 0)
@@ -4750,7 +4784,7 @@ forkchild(struct job *jp, union node *n, int mode)
                                 ash_msg_and_raise_error("can't open '%s'", bb_dev_null);
                 }
         }
-        if (!oldlvl) {
+        if (oldlvl == 0) {
                 if (iflag) { /* why if iflag only? */
                         setsignal(SIGINT);
                         setsignal(SIGTERM);
diff --git a/shell/ash_test/ash-signals/sigint1.right b/shell/ash_test/ash-signals/sigint1.right
new file mode 100644
index 000000000..a9094b056
--- /dev/null
+++ b/shell/ash_test/ash-signals/sigint1.right
@@ -0,0 +1 @@
+Sending SIGINT to main shell PID
diff --git a/shell/ash_test/ash-signals/sigint1.tests b/shell/ash_test/ash-signals/sigint1.tests
new file mode 100755
index 000000000..20e45d940
--- /dev/null
+++ b/shell/ash_test/ash-signals/sigint1.tests
@@ -0,0 +1,41 @@
+# What should happen if non-interactive shell gets SIGINT?
+
+(sleep 1; echo Sending SIGINT to main shell PID; exec kill -INT $$) &
+
+# We create a child which exits with 0 even on SIGINT
+# (This is truly necessary only if SIGINT is generated by ^C,
+# in this testcase even bare "sleep 2" would do because
+# we don't send SIGINT _to_ the_ child_...)
+$THIS_SH -c 'trap "exit 0" SIGINT; sleep 2'
+
+# In one second, we (main shell) get SIGINT here.
+# The question is whether we should, or should not, exit.
+
+# bash will not stop here. It will execute next command(s).
+
+# The rationale for this is described here:
+# http://www.cons.org/cracauer/sigint.html
+#
+# Basically, bash will not exit on SIGINT immediately if it waits
+# for a child. It will wait for the child to exit.
+# If child exits NOT by dying on SIGINT, then bash will not exit.
+#
+# The idea is that the following script:
+# | emacs file.txt
+# | more cmds
+# User may use ^C to interrupt editor's ops like search. But then
+# emacs exits normally. User expects that script doesn't stop.
+#
+# This is a nice idea, but detecting "did process really exit
+# with SIGINT?" is racy. Consider:
+# | bash -c 'while true; do /bin/true; done'
+# When ^C is pressed while bash waits for /bin/true to exit,
+# it may happen that /bin/true exits with exitcode 0 before
+# ^C is delivered to it as SIGINT. bash will see SIGINT, then
+# it will see that child exited with 0, and bash will NOT EXIT.
+
+# Therefore we do not implement bash behavior.
+# I'd say that emacs need to put itself into a separate pgrp
+# to isolate shell from getting stray SIGINTs from ^C.
+
+echo Next command after SIGINT was executed
diff --git a/shell/hush_test/hush-misc/sigint1.right b/shell/hush_test/hush-misc/sigint1.right
new file mode 100644
index 000000000..a9094b056
--- /dev/null
+++ b/shell/hush_test/hush-misc/sigint1.right
@@ -0,0 +1 @@
+Sending SIGINT to main shell PID
diff --git a/shell/hush_test/hush-misc/sigint1.tests b/shell/hush_test/hush-misc/sigint1.tests
new file mode 100755
index 000000000..20e45d940
--- /dev/null
+++ b/shell/hush_test/hush-misc/sigint1.tests
@@ -0,0 +1,41 @@
+# What should happen if non-interactive shell gets SIGINT?
+
+(sleep 1; echo Sending SIGINT to main shell PID; exec kill -INT $$) &
+
+# We create a child which exits with 0 even on SIGINT
+# (This is truly necessary only if SIGINT is generated by ^C,
+# in this testcase even bare "sleep 2" would do because
+# we don't send SIGINT _to_ the_ child_...)
+$THIS_SH -c 'trap "exit 0" SIGINT; sleep 2'
+
+# In one second, we (main shell) get SIGINT here.
+# The question is whether we should, or should not, exit.
+
+# bash will not stop here. It will execute next command(s).
+
+# The rationale for this is described here:
+# http://www.cons.org/cracauer/sigint.html
+#
+# Basically, bash will not exit on SIGINT immediately if it waits
+# for a child. It will wait for the child to exit.
+# If child exits NOT by dying on SIGINT, then bash will not exit.
+#
+# The idea is that the following script:
+# | emacs file.txt
+# | more cmds
+# User may use ^C to interrupt editor's ops like search. But then
+# emacs exits normally. User expects that script doesn't stop.
+#
+# This is a nice idea, but detecting "did process really exit
+# with SIGINT?" is racy. Consider:
+# | bash -c 'while true; do /bin/true; done'
+# When ^C is pressed while bash waits for /bin/true to exit,
+# it may happen that /bin/true exits with exitcode 0 before
+# ^C is delivered to it as SIGINT. bash will see SIGINT, then
+# it will see that child exited with 0, and bash will NOT EXIT.
+
+# Therefore we do not implement bash behavior.
+# I'd say that emacs need to put itself into a separate pgrp
+# to isolate shell from getting stray SIGINTs from ^C.
+
+echo Next command after SIGINT was executed