We sometimes erroneously detected processes in other sessions

as belonging to our session. Consequently, we would terminate
these processes when our session shut down. We are now more
conservative about deleting processes when a session ends.

BUG=n/a
TEST=Start two instances of the same chroot (e.g. precise32) in two different shell windows. Terminate one of them. Notice that the other one is still running.
NOTRY=true

Review URL: https://chromiumcodereview.appspot.com/20010003

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@213049 0039d316-1c4b-4281-b951-d872f2087c98

1 files changed, 40 insertions, 4 deletions

diff --git a/build/install-chroot.sh b/build/install-chroot.sh
index 155fc79..f9e165ec 100755
--- a/build/install-chroot.sh
+++ b/build/install-chroot.sh
@@ -497,26 +497,62 @@ while [ "$#" -ne 0 ]; do
   esac
 done
 
+# Start a new chroot session and keep track of the session id. We inject this
+# id into all processes that run inside the chroot. Unless they go out of their
+# way to clear their environment, we can then later identify our child and
+# grand-child processes by scanning their environment.
 session="$(schroot -c "${chroot}" -b)"
+export CHROOT_SESSION_ID="${session}"
 
 if [ $# -eq 0 ]; then
+  # Run an interactive shell session
   schroot -c "${session}" -r -p
 else
+  # Run a command inside of the chroot environment
   p="$1"; shift
   schroot -c "${session}" -r -p "$p" -- "$@"
 fi
 rc=$?
 
+# Compute the inode of the root directory inside of the chroot environment.
 i=$(schroot -c "${session}" -r -p ls -- -id /proc/self/root/. |
      awk '{ print $1 }') 2>/dev/null
+other_pids=
 while [ -n "$i" ]; do
-  pids=$(ls -id1 /proc/*/root/. 2>/dev/null |
+  # Identify processes by the inode number of their root directory. Then
+  # remove all processes that we know belong to other sessions. We use
+  # "sort | uniq -u" to do what amounts to a "set substraction operation".
+  pids=$({ ls -id1 /proc/*/root/. 2>/dev/null |
          sed -e 's,^[^0-9]*'$i'.*/\([1-9][0-9]*\)/.*$,\1,
                  t
-                 d') >/dev/null 2>&1
-  [ -z "$pids" ] && break
-  kill -9 $pids
+                 d';
+         echo "${other_pids}";
+         echo "${other_pids}"; } | sort | uniq -u) >/dev/null 2>&1
+  # Kill all processes that are still left running in the session. This is
+  # typically an assortment of daemon processes that were started
+  # automatically. They result in us being unable to tear down the session
+  # cleanly.
+  [ -z "${pids}" ] && break
+  for j in $pids; do
+    # Unfortunately, the way that schroot sets up sessions has the
+    # side-effect of being unable to tell one session apart from another.
+    # This can result in us attempting to kill processes in other sessions.
+    # We make a best-effort to avoid doing so.
+    k="$( ( xargs -0 -n1 </proc/$j/environ ) 2>/dev/null |
+         sed 's/^CHROOT_SESSION_ID=/x/;t1;d;:1;q')"
+    if [ -n "${k}" -a "${k#x}" != "${session}" ]; then
+      other_pids="${other_pids}
+${j}"
+      continue
+    fi
+    kill -9 $pids
+  done
 done
+# End the chroot session. This should clean up all temporary files. But if we
+# earlier failed to terminate all (daemon) processes inside of the session,
+# deleting the session could fail. When that happens, the user has to manually
+# clean up the stale files by invoking us with "--clean" after having killed
+# all running processes.
 schroot -c "${session}" -e
 exit $rc
 EOF