check: run _check_filesystems in an OOM-happy subshell
authorDarrick J. Wong <djwong@kernel.org>
Wed, 7 Jul 2021 00:21:34 +0000 (17:21 -0700)
committerEryu Guan <guaneryu@gmail.com>
Sun, 18 Jul 2021 14:32:29 +0000 (22:32 +0800)
While running fstests one night, I observed that fstests stopped
abruptly because ./check ran _check_filesystems to run xfs_repair.
In turn, repair (which inherited oom_score_adj=-1000 from ./check)
consumed so much memory that the OOM killer ran around killing other
daemons, rendering the system nonfunctional.

This is silly -- we set an OOM score adjustment of -1000 on the
./check process so that the test framework itself wouldn't get
OOM-killed, because that aborts the entire run.  Everything else is
fair game for that, including subprocesses started by
_check_filesystems.

Therefore, adapt _check_filesystems (and its children) to run in a
subshell with a much higher oom score adjustment.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Eryu Guan <guaneryu@gmail.com>
Signed-off-by: Eryu Guan <guaneryu@gmail.com>
check

diff --git a/check b/check
index de8104d..bb7e030 100755 (executable)
--- a/check
+++ b/check
@@ -525,17 +525,20 @@ _summary()
 
 _check_filesystems()
 {
+       local ret=0
+
        if [ -f ${RESULT_DIR}/require_test ]; then
-               _check_test_fs || err=true
+               _check_test_fs || ret=1
                rm -f ${RESULT_DIR}/require_test*
        else
                _test_unmount 2> /dev/null
        fi
        if [ -f ${RESULT_DIR}/require_scratch ]; then
-               _check_scratch_fs || err=true
+               _check_scratch_fs || ret=1
                rm -f ${RESULT_DIR}/require_scratch*
        fi
        _scratch_unmount 2> /dev/null
+       return $ret
 }
 
 _expunge_test()
@@ -558,11 +561,15 @@ test $? -eq 77 && HAVE_SYSTEMD_SCOPES=yes
 
 # Make the check script unattractive to the OOM killer...
 OOM_SCORE_ADJ="/proc/self/oom_score_adj"
-test -w ${OOM_SCORE_ADJ} && echo -1000 > ${OOM_SCORE_ADJ}
+function _adjust_oom_score() {
+       test -w "${OOM_SCORE_ADJ}" && echo "$1" > "${OOM_SCORE_ADJ}"
+}
+_adjust_oom_score -1000
 
 # ...and make the tests themselves somewhat more attractive to it, so that if
 # the system runs out of memory it'll be the test that gets killed and not the
-# test framework.
+# test framework.  The test is run in a separate process without any of our
+# functions, so we open-code adjusting the OOM score.
 #
 # If systemd is available, run the entire test script in a scope so that we can
 # kill all subprocesses of the test if it fails to clean up after itself.  This
@@ -875,9 +882,12 @@ function run_section()
                        rm -f ${RESULT_DIR}/require_scratch*
                        err=true
                else
-                       # the test apparently passed, so check for corruption
-                       # and log messages that shouldn't be there.
-                       _check_filesystems
+                       # The test apparently passed, so check for corruption
+                       # and log messages that shouldn't be there.  Run the
+                       # checking tools from a subshell with adjusted OOM
+                       # score so that the OOM killer will target them instead
+                       # of the check script itself.
+                       (_adjust_oom_score 250; _check_filesystems) || err=true
                        _check_dmesg || err=true
                fi