Bug 1814551 - Try waiting for child processes directly when the fork server is in use. r=nika

When the fork server is enabled, not all IPC child processes are children of
the fork server; currently, process types other than content processes
are still spawned directly.  This means that we need to `waitid` or
`waitpid` them when they exit in order to not leak zombie processes.

Specifically, we can just try to `waitid` the process, and then if that
fails with `ECHILD` we can assume it was a fork server child and fall
back to the previous `kill(pid, 0)` workaround.  This patch does that,
but only if the fork server is active; otherwise we maintain the current
behavior of only waiting for child processes directly.

Differential Revision: https://phabricator.services.mozilla.com/D168756

ipc/chromium/src/base/process_util_posix.cc

@@ -30,6 +30,7 @@
 #include "base/dir_reader_posix.h"
 
 #include "mozilla/UniquePtr.h"
+#include "mozilla/Unused.h"
 // For PR_DuplicateEnvironment:
 #include "prenv.h"
 #include "prmem.h"
@@ -196,20 +197,25 @@ void CloseSuperfluousFds(void* aCtx, bool (*aShouldPreserve)(void*, int)) {
 }
 
 bool IsProcessDead(ProcessHandle handle, bool blocking) {
+  auto handleForkServer = [handle]() -> mozilla::Maybe<bool> {
 #ifdef MOZ_ENABLE_FORKSERVER
-  if (mozilla::ipc::ForkServiceChild::Get()) {
-    // We only know if a process exists, but not if it has crashed.
-    //
-    // Since content processes are not direct children of the chrome
-    // process any more, it is impossible to use |waitpid()| to wait for
-    // them.
-    const int r = kill(handle, 0);
-    // FIXME: for unexpected errors we should probably log a warning
-    // and return true, so that the caller doesn't loop / hang /
-    // try to kill the process.  (Bug 1658072 will rewrite this code.)
-    return r < 0 && errno == ESRCH;
-  }
+    if (errno == ECHILD && mozilla::ipc::ForkServiceChild::Get()) {
+      // We only know if a process exists, but not if it has crashed.
+      //
+      // Since content processes are not direct children of the chrome
+      // process any more, it is impossible to use |waitpid()| to wait for
+      // them.
+      const int r = kill(handle, 0);
+      // FIXME: for unexpected errors we should probably log a warning
+      // and return true, so that the caller doesn't loop / hang /
+      // try to kill the process.  (Bug 1658072 will rewrite this code.)
+      return mozilla::Some(r < 0 && errno == ESRCH);
+    }
+#else
+    mozilla::Unused << handle;
 #endif
+    return mozilla::Nothing();
+  };
 
 #ifdef HAVE_WAITID
 
@@ -222,6 +228,10 @@ bool IsProcessDead(ProcessHandle handle, bool blocking) {
   const int wflags = WEXITED | WNOWAIT | (blocking ? 0 : WNOHANG);
   int result = HANDLE_EINTR(waitid(P_PID, handle, &si, wflags));
   if (result == -1) {
+    if (auto forkServerReturn = handleForkServer()) {
+      return *forkServerReturn;
+    }
+
     // This shouldn't happen, but sometimes it does.  The error is
     // probably ECHILD and the reason is probably that a pid was
     // waited on again after a previous wait reclaimed its zombie.
@@ -287,6 +297,10 @@ bool IsProcessDead(ProcessHandle handle, bool blocking) {
   int status;
   const int result = waitpid(handle, &status, blocking ? 0 : WNOHANG);
   if (result == -1) {
+    if (auto forkServerReturn = handleForkServer()) {
+      return *forkServerReturn;
+    }
+
     CHROMIUM_LOG(ERROR) << "waitpid failed pid:" << handle
                         << " errno:" << errno;
     return true;