[PATCH v4] lib/eal: fix segfaults due to thread exit order

[Bruce Richardson]

On Wed, Jun 15, 2022 at 02:01:54PM +0800, zhichaox.zeng at intel.com wrote:
> From: Zhichao Zeng <zhichaox.zeng at intel.com>
> 
> The eal-intr-thread is not closed before memory cleanup in the
> process of exiting. There is a small probability that when the
> eal-intr-thread is about to use some pointers, the memory were
> just cleaned, which cause the segment fault error caught by ASan.
> 
> This patch close the eal-intr-thread before memory cleanup when
> exiting to avoid segment fault. And add some atomic operations
> to avoid executing rte_eal_cleanup in the child process spawned
> by fork() in some test cases, e.g. debug_autotest of dpdk-test.
> 
> Cc: stable at dpdk.org
> 

Hi,

some comments inline below.

/Bruce

> ---
> v2:
> add the same API for FreeBSD
> ---
> v3:
> fix rte_eal_cleanup crash in debug_autotest
> ---
> v4:
> shorten the prompt message and optimize the commit log
> 

Please put these updates below the cutline after the sign-offs, i.e.
immediately before the diffstat.

> Suggested-by: David Marchand <david.marchand at redhat.com>
> Signed-off-by: Zhichao Zeng <zhichaox.zeng at intel.com>
> ---
>  lib/eal/common/eal_private.h     |  7 +++++++
>  lib/eal/freebsd/eal.c            | 21 ++++++++++++++++++++-
>  lib/eal/freebsd/eal_interrupts.c | 12 ++++++++++++
>  lib/eal/linux/eal.c              | 20 +++++++++++++++++++-
>  lib/eal/linux/eal_interrupts.c   | 12 ++++++++++++
>  5 files changed, 70 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/eal/common/eal_private.h b/lib/eal/common/eal_private.h
> index 44d14241f0..7adf41b7d7 100644
> --- a/lib/eal/common/eal_private.h
> +++ b/lib/eal/common/eal_private.h
> @@ -152,6 +152,13 @@ int rte_eal_tailqs_init(void);
>   */
>  int rte_eal_intr_init(void);
>  
> +/**
> + * Destroy interrupt handling thread.
> + *
> + * This function is private to EAL.
> + */
> +void rte_eal_intr_destroy(void);
> +
>  /**
>   * Close the default log stream
>   *
> diff --git a/lib/eal/freebsd/eal.c b/lib/eal/freebsd/eal.c
> index a6b20960f2..4882f27abd 100644
> --- a/lib/eal/freebsd/eal.c
> +++ b/lib/eal/freebsd/eal.c
> @@ -72,6 +72,8 @@ struct lcore_config lcore_config[RTE_MAX_LCORE];
>  /* used by rte_rdtsc() */
>  int rte_cycles_vmware_tsc_map;
>  
> +/* used to judge the running status of the eal */
> +static uint32_t run_once;
>  

I don't like just moving this variable from the eal_init function. When in
eal_init the name "run_once" made sense as it tracked how often the EAL
init function was run. However, now as a global variable the name
"run_once" no longer makes sense.

Two suggestions:
1. Keep run_once in EAL init as-is, and use a different variable or value
   to indicate that DPDK is initialized for cleanup.
2. Move the variable as you have here, just rename it to a more meaningful
   name.


>  int
>  eal_clean_runtime_dir(void)
> @@ -574,12 +576,22 @@ static void rte_eal_init_alert(const char *msg)
>  	RTE_LOG(ERR, EAL, "%s\n", msg);
>  }
>  
> +static void warn_parent(void)
> +{
> +	RTE_LOG(WARNING, EAL, "DPDK won't work in the child process\n");
> +}

I wonder if this contains enough information. Can we identify briefly what
parts will or won't work, or if we just want to deny everything, can we
give a brief reason why?

> +
> +static void scratch_child(void)
> +{
> +	/* Scratch run_once so that a call to rte_eal_cleanup won't crash... */
> +	__atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
> +}
> +

I think the name of this function needs improvement. I'm not sure that
"scratch" is the best term to use. Something like "clear_eal_flag" is
probably better.

>  /* Launch threads, called at application init(). */
>  int
>  rte_eal_init(int argc, char **argv)
>  {
>  	int i, fctret, ret;
> -	static uint32_t run_once;
>  	uint32_t has_run = 0;
>  	char cpuset[RTE_CPU_AFFINITY_STR_LEN];
>  	char thread_name[RTE_MAX_THREAD_NAME_LEN];
> @@ -883,6 +895,8 @@ rte_eal_init(int argc, char **argv)
>  
>  	eal_mcfg_complete();
>  
> +	pthread_atfork(NULL, warn_parent, scratch_child);
> +
>  	return fctret;
>  }
>  
> @@ -891,8 +905,13 @@ rte_eal_cleanup(void)
>  {
>  	struct internal_config *internal_conf =
>  		eal_get_internal_configuration();
> +
> +	if (__atomic_load_n(&run_once, __ATOMIC_RELAXED) == 0)
> +		return 0;
> +
>  	rte_service_finalize();
>  	rte_mp_channel_cleanup();
> +	rte_eal_intr_destroy();
>  	/* after this point, any DPDK pointers will become dangling */
>  	rte_eal_memory_detach();
>  	rte_eal_alarm_cleanup();
<snip for brevity>