From 49bfaaa65107a9d79fbb1276e44fd4e3c98b0e9c Mon Sep 17 00:00:00 2001

From: Lennart Poettering <lennart@poettering.net>

Date: Fri, 7 Feb 2014 11:58:25 +0100

Subject: [PATCH] core: allow PIDs to be watched by two units at the same time

In some cases it is interesting to map a PID to two units at the same

time. For example, when a user logs in via a getty, which is reexeced to

/sbin/login that binary will be explicitly referenced as main pid of the

getty service, as well as implicitly referenced as part of the session

scope.

Conflicts:

	src/core/manager.c

	src/core/manager.h

---

 src/core/manager.c | 201 ++++++++++++++++++++++++++++++-----------------------

 src/core/manager.h |   9 ++-

 src/core/unit.c    |  28 ++++++--

 3 files changed, 142 insertions(+), 96 deletions(-)

diff --git a/src/core/manager.c b/src/core/manager.c

index db5094f..2829c95 100644

--- a/src/core/manager.c

+++ b/src/core/manager.c

@@ -525,7 +525,10 @@ int manager_new(SystemdRunningAs running_as, bool reexecuting, Manager **_m) {

         if (!(m->jobs = hashmap_new(trivial_hash_func, trivial_compare_func)))

                 goto fail;

-        if (!(m->watch_pids = hashmap_new(trivial_hash_func, trivial_compare_func)))

+        if (!(m->watch_pids1 = hashmap_new(trivial_hash_func, trivial_compare_func)))

+                goto fail;

+

+        if (!(m->watch_pids2 = hashmap_new(trivial_hash_func, trivial_compare_func)))

                 goto fail;

         m->cgroup_unit = hashmap_new(string_hash_func, string_compare_func);

@@ -740,7 +743,8 @@ void manager_free(Manager *m) {

         hashmap_free(m->units);

         hashmap_free(m->jobs);

-        hashmap_free(m->watch_pids);

+        hashmap_free(m->watch_pids1);

+        hashmap_free(m->watch_pids2);

         hashmap_free(m->watch_bus);

         if (m->epoll_fd >= 0)

@@ -1247,6 +1251,26 @@ unsigned manager_dispatch_dbus_queue(Manager *m) {

         return n;

 }

+static void manager_invoke_notify_message(Manager *m, Unit *u, pid_t pid, char *buf, size_t n) {

+        _cleanup_strv_free_ char **tags = NULL;

+

+        assert(m);

+        assert(u);

+        assert(buf);

+        assert(n > 0);

+

+        tags = strv_split(buf, "\n\r");

+        if (!tags) {

+                log_oom();

+                return;

+        }

+

+        log_debug_unit(u->id, "Got notification message for unit %s", u->id);

+

+        if (UNIT_VTABLE(u)->notify_message)

+                UNIT_VTABLE(u)->notify_message(u, pid, tags);

+}

+

 static int manager_process_notify_fd(Manager *m) {

         ssize_t n;

@@ -1259,6 +1283,8 @@ static int manager_process_notify_fd(Manager *m) {

                         .iov_len = sizeof(buf)-1,

                 };

+                bool found = false;

+

                 union {

                         struct cmsghdr cmsghdr;

                         uint8_t buf[CMSG_SPACE(sizeof(struct ucred))];

@@ -1272,7 +1298,6 @@ static int manager_process_notify_fd(Manager *m) {

                 };

                 struct ucred *ucred;

                 Unit *u;

-                _cleanup_strv_free_ char **tags = NULL;

                 n = recvmsg(m->notify_watch.fd, &msghdr, MSG_DONTWAIT);

                 if (n <= 0) {

@@ -1295,105 +1320,105 @@ static int manager_process_notify_fd(Manager *m) {

                 ucred = (struct ucred*) CMSG_DATA(&control.cmsghdr);

-                u = hashmap_get(m->watch_pids, LONG_TO_PTR(ucred->pid));

-                if (!u) {

-                        u = manager_get_unit_by_pid(m, ucred->pid);

-                        if (!u) {

-                                log_warning("Cannot find unit for notify message of PID %lu.", (unsigned long) ucred->pid);

-                                continue;

-                        }

-                }

-

                 assert((size_t) n < sizeof(buf));

                 buf[n] = 0;

-                tags = strv_split(buf, "\n\r");

-                if (!tags)

-                        return log_oom();

-

-                log_debug_unit(u->id, "Got notification message for unit %s", u->id);

-

-                if (UNIT_VTABLE(u)->notify_message)

-                        UNIT_VTABLE(u)->notify_message(u, ucred->pid, tags);

-        }

-

-        return 0;

-}

-

-static int manager_dispatch_sigchld(Manager *m) {

-        assert(m);

-

-        for (;;) {

-                siginfo_t si = {};

-                Unit *u;

-                int r;

-

-                /* First we call waitd() for a PID and do not reap the

-                 * zombie. That way we can still access /proc/$PID for

-                 * it while it is a zombie. */

-                if (waitid(P_ALL, 0, &si, WEXITED|WNOHANG|WNOWAIT) < 0) {

-

-                        if (errno == ECHILD)

-                                break;

-                        if (errno == EINTR)

-                                continue;

-

-                        return -errno;

+                u = manager_get_unit_by_pid(m, ucred->pid);

+                if (u) {

+                        manager_invoke_notify_message(m, u, ucred->pid, buf, n);

+                        found = true;

                 }

-                if (si.si_pid <= 0)

-                        break;

-

-                if (si.si_code == CLD_EXITED || si.si_code == CLD_KILLED || si.si_code == CLD_DUMPED) {

-                        _cleanup_free_ char *name = NULL;

-

-                        get_process_comm(si.si_pid, &name);

-                        log_debug("Got SIGCHLD for process %lu (%s)", (unsigned long) si.si_pid, strna(name));

+                u = hashmap_get(m->watch_pids1, LONG_TO_PTR(ucred->pid));

+                if (u) {

+                        manager_invoke_notify_message(m, u, ucred->pid, buf, n);

+                        found = true;

                 }

-                /* Let's flush any message the dying child might still

-                 * have queued for us. This ensures that the process

-                 * still exists in /proc so that we can figure out

-                 * which cgroup and hence unit it belongs to. */

-                r = manager_process_notify_fd(m);

-                if (r < 0)

-                        return r;

-

-                /* And now figure out the unit this belongs to */

-                u = hashmap_get(m->watch_pids, LONG_TO_PTR(si.si_pid));

-                if (!u)

-                        u = manager_get_unit_by_pid(m, si.si_pid);

-

-                /* And now, we actually reap the zombie. */

-                if (waitid(P_PID, si.si_pid, &si, WEXITED) < 0) {

-                        if (errno == EINTR)

-                                continue;

-

-                        return -errno;

+                u = hashmap_get(m->watch_pids2, LONG_TO_PTR(ucred->pid));

+                if (u) {

+                        manager_invoke_notify_message(m, u, ucred->pid, buf, n);

+                        found = true;

                 }

-                if (si.si_code != CLD_EXITED && si.si_code != CLD_KILLED && si.si_code != CLD_DUMPED)

-                        continue;

+                if (!found)

+                        log_warning("Cannot find unit for notify message of PID %lu.",(long unsigned) ucred->pid);

+        }

-                log_debug("Child %lu died (code=%s, status=%i/%s)",

-                          (long unsigned) si.si_pid,

-                          sigchld_code_to_string(si.si_code),

-                          si.si_status,

-                          strna(si.si_code == CLD_EXITED

-                                ? exit_status_to_string(si.si_status, EXIT_STATUS_FULL)

-                                : signal_to_string(si.si_status)));

+        return 0;

+}

-                if (!u)

-                        continue;

+static void invoke_sigchld_event(Manager *m, Unit *u, siginfo_t *si) {

+        assert(m);

+        assert(u);

+        assert(si);

-                log_debug_unit(u->id,

-                               "Child %lu belongs to %s", (long unsigned) si.si_pid, u->id);

+        log_debug_unit(u->id, "Child %lu belongs to %s",(long unsigned) si->si_pid, u->id);

-                unit_unwatch_pid(u, si.si_pid);

-                UNIT_VTABLE(u)->sigchld_event(u, si.si_pid, si.si_code, si.si_status);

-        }

+        unit_unwatch_pid(u, si->si_pid);

+        UNIT_VTABLE(u)->sigchld_event(u, si->si_pid, si->si_code, si->si_status);

+}

-        return 0;

+static int manager_dispatch_sigchld(Manager *m) {

+    assert(m);

+

+    for (;;) {

+            siginfo_t si = {};

+

+            /* First we call waitd() for a PID and do not reap the

+             * zombie. That way we can still access /proc/$PID for

+             * it while it is a zombie. */

+            if (waitid(P_ALL, 0, &si, WEXITED|WNOHANG|WNOWAIT) < 0) {

+

+                    if (errno == ECHILD)

+                            break;

+

+                    if (errno == EINTR)

+                            continue;

+

+                    return -errno;

+            }

+

+            if (si.si_pid <= 0)

+                    break;

+

+            if (si.si_code == CLD_EXITED || si.si_code == CLD_KILLED || si.si_code == CLD_DUMPED) {

+                    _cleanup_free_ char *name = NULL;

+                    Unit *u;

+

+                    get_process_comm(si.si_pid, &name);

+

+                    log_debug("Child %lu (%s) died (code=%s, status=%i/%s)",

+                              (long unsigned) si.si_pid, strna(name),

+                              sigchld_code_to_string(si.si_code),

+                              si.si_status,

+                              strna(si.si_code == CLD_EXITED

+                                    ? exit_status_to_string(si.si_status, EXIT_STATUS_FULL)

+                                    : signal_to_string(si.si_status)));

+

+                    /* And now figure out the unit this belongs

+                     * to, it might be multiple... */

+                    u = manager_get_unit_by_pid(m, si.si_pid);

+                    if (u)

+                            invoke_sigchld_event(m, u, &si);

+                    u = hashmap_get(m->watch_pids1, LONG_TO_PTR(si.si_pid));

+                    if (u)

+                            invoke_sigchld_event(m, u, &si);

+                    u = hashmap_get(m->watch_pids2, LONG_TO_PTR(si.si_pid));

+                    if (u)

+                            invoke_sigchld_event(m, u, &si);

+            }

+

+            /* And now, we actually reap the zombie. */

+            if (waitid(P_PID, si.si_pid, &si, WEXITED) < 0) {

+                    if (errno == EINTR)

+                            continue;

+

+                    return -errno;

+            }

+    }

+

+    return 0;

 }

 static int manager_start_target(Manager *m, const char *name, JobMode mode) {

diff --git a/src/core/manager.h b/src/core/manager.h

index ee42c5e..0133ea5 100644

--- a/src/core/manager.h

+++ b/src/core/manager.h

@@ -125,7 +125,14 @@ struct Manager {

         /* Units that should be realized */

         LIST_HEAD(Unit, cgroup_queue);

-        Hashmap *watch_pids;  /* pid => Unit object n:1 */

+        /* We use two hash tables here, since the same PID might be

+         * watched by two different units: once the unit that forked

+         * it off, and possibly a different unit to which it was

+         * joined as cgroup member. Since we know that it is either

+         * one or two units for each PID we just use to hashmaps

+         * here. */

+        Hashmap *watch_pids1;  /* pid => Unit object n:1 */

+        Hashmap *watch_pids2;  /* pid => Unit object n:1 */

         char *notify_socket;

diff --git a/src/core/unit.c b/src/core/unit.c

index 0332094..a510eb2 100644

--- a/src/core/unit.c

+++ b/src/core/unit.c

@@ -1665,16 +1665,27 @@ int unit_watch_pid(Unit *u, pid_t pid) {

         assert(u);

         assert(pid >= 1);

+        /* Watch a specific PID. We only support one or two units

+         * watching each PID for now, not more. */

+

+        r = hashmap_ensure_allocated(&u->manager->watch_pids1, trivial_hash_func, trivial_compare_func);

+        if (r < 0)

+                return r;

+

         r = set_ensure_allocated(&u->pids, trivial_hash_func, trivial_compare_func);

         if (r < 0)

                 return r;

-        /* Watch a specific PID. We only support one unit watching

-         * each PID for now. */

+        r = hashmap_put(u->manager->watch_pids1, LONG_TO_PTR(pid), u);

+        if (r == -EEXIST) {

+                r = hashmap_ensure_allocated(&u->manager->watch_pids2, trivial_hash_func, trivial_compare_func);

+                if (r < 0)

+                        return r;

-        r = set_put(u->pids, LONG_TO_PTR(pid));

+                r = hashmap_put(u->manager->watch_pids2, LONG_TO_PTR(pid), u);

+        }

-        q = hashmap_put(u->manager->watch_pids, LONG_TO_PTR(pid), u);

+        q = set_put(u->pids, LONG_TO_PTR(pid));

         if (q < 0)

                 return q;

@@ -1685,7 +1696,8 @@ void unit_unwatch_pid(Unit *u, pid_t pid) {

         assert(u);

         assert(pid >= 1);

-        hashmap_remove_value(u->manager->watch_pids, LONG_TO_PTR(pid), u);

+        hashmap_remove_value(u->manager->watch_pids1, LONG_TO_PTR(pid), u);

+        hashmap_remove_value(u->manager->watch_pids2, LONG_TO_PTR(pid), u);

         set_remove(u->pids, LONG_TO_PTR(pid));

 }

@@ -1758,8 +1770,10 @@ void unit_unwatch_all_pids(Unit *u) {

         assert(u);

-        SET_FOREACH(e, u->pids, i)

-                hashmap_remove_value(u->manager->watch_pids, e, u);

+        SET_FOREACH(e, u->pids, i) {

+                hashmap_remove_value(u->manager->watch_pids1, e, u);

+                hashmap_remove_value(u->manager->watch_pids2, e, u);

+        }

         set_free(u->pids);

         u->pids = NULL;