[dm-devel] [PATCH v3 02/29] multipathd: Fix liburcu memory leak

[Benjamin Marzinski]

On Wed, Dec 16, 2020 at 07:16:41PM +0100, mwilck at suse.com wrote:
> From: Martin Wilck <mwilck at suse.com>
> 
> 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.
> 
Reviewed-by: Benjamin Marzinski <bmarzins at redhat.com>
> Signed-off-by: Martin Wilck <mwilck at suse.com>
> ---
>  multipathd/main.c | 45 ++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 44 insertions(+), 1 deletion(-)
> 
> diff --git a/multipathd/main.c b/multipathd/main.c
> index c5c374b..ce14bb6 100644
> --- a/multipathd/main.c
> +++ b/multipathd/main.c
> @@ -2889,6 +2889,48 @@ 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 struct call_rcu_data *mp_rcu_data;
> +
> +static void cleanup_rcu(void)
> +{
> +	pthread_t rcu_thread;
> +
> +	/* Wait for any pending RCU calls */
> +	rcu_barrier();
> +	if (mp_rcu_data != NULL) {
> +		rcu_thread = get_call_rcu_thread(mp_rcu_data);
> +		/* detach this thread from the RCU thread */
> +		set_thread_call_rcu_data(NULL);
> +		synchronize_rcu();
> +		/* tell RCU thread to exit */
> +		call_rcu_data_free(mp_rcu_data);
> +		pthread_join(rcu_thread, NULL);
> +	}
> +	rcu_unregister_thread();
> +}
> +
>  static int
>  child (__attribute__((unused)) void *param)
>  {
> @@ -2906,7 +2948,8 @@ child (__attribute__((unused)) void *param)
>  
>  	mlockall(MCL_CURRENT | MCL_FUTURE);
>  	signal_init();
> -	rcu_init();
> +	mp_rcu_data = setup_rcu();
> +	atexit(cleanup_rcu);
>  
>  	setup_thread_attr(&misc_attr, 64 * 1024, 0);
>  	setup_thread_attr(&uevent_attr, DEFAULT_UEVENT_STACKSIZE * 1024, 0);
> -- 
> 2.29.0