Re: [dpdk-dev] [PATCH v3] eal: fix race in ctrl thread creation

[Honnappa Nagarahalli]

<snip>

> 
> The creation of control threads used a pthread barrier for synchronization.
> This patch fixes a race condition where the pthread barrier could get
> destroyed while one of the threads has not yet returned from the
> pthread_barrier_wait function, which could result in undefined behaviour.
> The barrier has been completely removed in favour of a reference count on
> the control thread parameters struct.
> 
> Fixes: 3a0d465d4c53 ("eal: fix use-after-free on control thread creation")
> Cc: jianfeng....@intel.com
> Cc: sta...@dpdk.org
> 
> Signed-off-by: Luc Pelletier <lucp.at.w...@gmail.com>
> ---
> 
> Hi Olivier,
> Hi Honnappa,
> 
> Thanks for your input Honnappa. I've made the changes to completely
> remove the barrier. However, I didn't move the call to pthread_setaffinity_np
I looked at the rte_ctrl_thread_create API definition. I am not sure if we have 
much leavy here as the API definition bakes in calling pthread_setaffinity_np.
With the barrier, the control thread did not execute its function till it was 
moved to the assigned core. If we remove the barrier and not move the 
pthread_setaffinity_np to ctrl_thread_init, the behavior might change.

One more comment below on how we are handling the setaffinity failure.

> to the control thread; I think we want to report the result of that function 
> to
> the caller of rte_ctrl_thread_create and doing so from ctrl_thread_init would
> be a lot trickier.
> 
> Olivier, what do you think of these changes?
> 
>  lib/librte_eal/common/eal_common_thread.c | 35 ++++++++---------------
>  1 file changed, 12 insertions(+), 23 deletions(-)
> 
> diff --git a/lib/librte_eal/common/eal_common_thread.c
> b/lib/librte_eal/common/eal_common_thread.c
> index 73a055902..2421066f9 100644
> --- a/lib/librte_eal/common/eal_common_thread.c
> +++ b/lib/librte_eal/common/eal_common_thread.c
> @@ -169,12 +169,11 @@ __rte_thread_uninit(void)  struct
> rte_thread_ctrl_params {
>       void *(*start_routine)(void *);
>       void *arg;
> -     pthread_barrier_t configured;
> +     unsigned int refcnt;
>  };
> 
>  static void *ctrl_thread_init(void *arg)  {
> -     int ret;
>       struct internal_config *internal_conf =
>               eal_get_internal_configuration();
>       rte_cpuset_t *cpuset = &internal_conf->ctrl_cpuset; @@ -184,11
> +183,8 @@ static void *ctrl_thread_init(void *arg)
> 
>       __rte_thread_init(rte_lcore_id(), cpuset);
> 
> -     ret = pthread_barrier_wait(&params->configured);
> -     if (ret == PTHREAD_BARRIER_SERIAL_THREAD) {
> -             pthread_barrier_destroy(&params->configured);
> +     if (__atomic_sub_fetch(&params->refcnt, 1, __ATOMIC_ACQ_REL) ==
> 0)
>               free(params);
> -     }
> 
>       return start_routine(routine_arg);
>  }
> @@ -210,14 +206,11 @@ rte_ctrl_thread_create(pthread_t *thread, const
> char *name,
> 
>       params->start_routine = start_routine;
>       params->arg = arg;
> -
> -     pthread_barrier_init(&params->configured, NULL, 2);
> +     params->refcnt = 2;
> 
>       ret = pthread_create(thread, attr, ctrl_thread_init, (void *)params);
> -     if (ret != 0) {
> -             free(params);
> -             return -ret;
> -     }
> +     if (ret != 0)
> +             goto fail;
> 
>       if (name != NULL) {
>               ret = rte_thread_setname(*thread, name); @@ -228,24
> +221,20 @@ rte_ctrl_thread_create(pthread_t *thread, const char *name,
> 
>       ret = pthread_setaffinity_np(*thread, sizeof(*cpuset), cpuset);
>       if (ret)
> -             goto fail;
> +             goto fail_cancel;
> 
> -     ret = pthread_barrier_wait(&params->configured);
> -     if (ret == PTHREAD_BARRIER_SERIAL_THREAD) {
> -             pthread_barrier_destroy(&params->configured);
> +     if (__atomic_sub_fetch(&params->refcnt, 1, __ATOMIC_ACQ_REL) ==
> 0)
>               free(params);
> -     }
> 
>       return 0;
> 
> -fail:
> -     if (PTHREAD_BARRIER_SERIAL_THREAD ==
> -         pthread_barrier_wait(&params->configured)) {
> -             pthread_barrier_destroy(&params->configured);
> -             free(params);
> -     }
> +fail_cancel:
>       pthread_cancel(*thread);
For the control thread to be able to respond to pthread_cancel, it needs to 
call one of the library functions that is considered a cancellation point. I do 
not see such requirements in the API definition.


>       pthread_join(*thread, NULL);
> +
> +fail:
> +     free(params);
> +
>       return -ret;
>  }
> 
> --
> 2.25.1