From mboxrd@z Thu Jan 1 00:00:00 1970 From: mwilck@suse.com Subject: [PATCH 02/23] multipathd: Fix liburcu memory leak Date: Thu, 24 Sep 2020 15:40:33 +0200 Message-ID: <20200924134054.14632-3-mwilck@suse.com> References: <20200924134054.14632-1-mwilck@suse.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20200924134054.14632-1-mwilck@suse.com> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: dm-devel-bounces@redhat.com Errors-To: dm-devel-bounces@redhat.com To: Christophe Varoqui , Benjamin Marzinski Cc: lixiaokeng@huawei.com, dm-devel@redhat.com, Martin Wilck List-Id: dm-devel.ids From: Martin Wilck Fix this leak in multipathd, reported by valgrind, that messes up multipathd's otherwise clean leak report: ==23823== 336 bytes in 1 blocks are possibly lost in loss record 3 of 3 ==23823== at 0x483AB65: calloc (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so) ==23823== by 0x4012F16: _dl_allocate_tls (in /lib64/ld-2.31.so) ==23823== by 0x493BB8E: pthread_create@@GLIBC_2.2.5 (in /lib64/libpthread-2.31.so) ==23823== by 0x492A9A9: call_rcu_data_init (urcu-call-rcu-impl.h:437) ==23823== by 0x492AD2F: UnknownInlinedFun (urcu-call-rcu-impl.h:492) ==23823== by 0x492AD2F: create_call_rcu_data_memb (urcu-call-rcu-impl.h:504) ==23823== by 0x1164E3: child.constprop.0.isra.0 (main.c:2915) ==23823== by 0x10F50C: main (main.c:3335) ==23823== ==23823== LEAK SUMMARY: ==23823== definitely lost: 0 bytes in 0 blocks ==23823== indirectly lost: 0 bytes in 0 blocks ==23823== possibly lost: 336 bytes in 1 blocks The problem is caused by using liburcu's default RCU call handler, which liburcu refuses to stop/join. See comments in the code. Signed-off-by: Martin Wilck --- multipathd/main.c | 45 ++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 44 insertions(+), 1 deletion(-) diff --git a/multipathd/main.c b/multipathd/main.c index c5c374b..be1b5ae 100644 --- a/multipathd/main.c +++ b/multipathd/main.c @@ -2889,6 +2889,47 @@ set_oom_adj (void) condlog(0, "couldn't adjust oom score"); } +/* + * Use a non-default call_rcu_data for child(). + * + * We do this to avoid a memory leak from liburcu. + * liburcu never frees the default rcu handler (see comments on + * call_rcu_data_free() in urcu-call-rcu-impl.h), its thread + * can't be joined with pthread_join(), leaving a memory leak. + * + * Therefore we create our own, which can be destroyed and joined. + */ +static struct call_rcu_data *setup_rcu(void) +{ + struct call_rcu_data *crdp; + + rcu_init(); + rcu_register_thread(); + crdp = create_call_rcu_data(0UL, -1); + if (crdp != NULL) + set_thread_call_rcu_data(crdp); + return crdp; +} + +static void cleanup_rcu(int dummy __attribute__((unused)), void *arg) +{ + struct call_rcu_data *crdp = arg; + pthread_t rcu_thread; + + /* Wait for any pending RCU calls */ + rcu_barrier(); + if (crdp != NULL) { + rcu_thread = get_call_rcu_thread(crdp); + /* detach this thread from the RCU thread */ + set_thread_call_rcu_data(NULL); + synchronize_rcu(); + /* tell RCU thread to exit */ + call_rcu_data_free(crdp); + pthread_join(rcu_thread, NULL); + } + rcu_unregister_thread(); +} + static int child (__attribute__((unused)) void *param) { @@ -2903,10 +2944,12 @@ child (__attribute__((unused)) void *param) char *envp; int queue_without_daemon; enum daemon_status state; + struct call_rcu_data *crdp; mlockall(MCL_CURRENT | MCL_FUTURE); signal_init(); - rcu_init(); + crdp = setup_rcu(); + on_exit(cleanup_rcu, crdp); setup_thread_attr(&misc_attr, 64 * 1024, 0); setup_thread_attr(&uevent_attr, DEFAULT_UEVENT_STACKSIZE * 1024, 0); -- 2.28.0