[dpdk-dev] [PATCH 1/2] eal/hotplug: allow monitor to be setup by multiple places

[Luca Boccassi]

On Mon, 2020-11-30 at 23:56 -0800, Long Li wrote:
> From: Stephen Hemminger <stephen at networkplumber.org>
> 
> In some cases, a device or infrastructure may want to enable hotplug
> but application may also try and start hotplug as well. Therefore
> change the monitor_started from a boolean into a reference count.
> 
> Signed-off-by: Stephen Hemminger <stephen at networkplumber.org>
> Signed-off-by: Long Li <longli at microsoft.com>
> ---
>  lib/librte_eal/linux/eal_dev.c | 15 +++++++--------
>  1 file changed, 7 insertions(+), 8 deletions(-)
> 
> diff --git a/lib/librte_eal/linux/eal_dev.c b/lib/librte_eal/linux/eal_dev.c
> index 5c0e752b2d..5fa679989e 100644
> --- a/lib/librte_eal/linux/eal_dev.c
> +++ b/lib/librte_eal/linux/eal_dev.c
> @@ -23,8 +23,11 @@
>  
>  #include "eal_private.h"
>  
> -static struct rte_intr_handle intr_handle = {.fd = -1 };
> -static bool monitor_started;
> +static struct rte_intr_handle intr_handle = {
> +	.type = RTE_INTR_HANDLE_DEV_EVENT,
> +	.fd = -1,
> +};
> +static uint32_t monitor_refcount;
>  static bool hotplug_handle;
>  
>  #define EAL_UEV_MSG_LEN 4096
> @@ -300,7 +303,7 @@ rte_dev_event_monitor_start(void)
>  {
>  	int ret;
>  
> -	if (monitor_started)
> +	if (__atomic_fetch_add(&monitor_refcount, 1, __ATOMIC_RELAXED))
>  		return 0;

If dev_uev_socket_fd_create or rte_intr_callback_register fail, you'll
have incremented the refcount but there won't be anything registered,
and calls to rte_dev_event_monitor_start will do nothing. Will that be
a problem?

In other words, it seems to me the semantics is changing from call
rte_dev_event_monitor_start until it succeeds, to call
rte_dev_event_monitor_start, and if it fails call
rte_dev_event_monitor_stop and then rte_dev_event_monitor_start again

>  	ret = dev_uev_socket_fd_create();
> @@ -309,7 +312,6 @@ rte_dev_event_monitor_start(void)
>  		return -1;
>  	}
>  
> -	intr_handle.type = RTE_INTR_HANDLE_DEV_EVENT;
>  	ret = rte_intr_callback_register(&intr_handle, dev_uev_handler, NULL);
>  
>  	if (ret) {
> @@ -317,8 +319,6 @@ rte_dev_event_monitor_start(void)
>  		return -1;
>  	}
>  
> -	monitor_started = true;
> -
>  	return 0;
>  }
>  
> @@ -327,7 +327,7 @@ rte_dev_event_monitor_stop(void)
>  {
>  	int ret;
>  
> -	if (!monitor_started)
> +	if (__atomic_sub_fetch(&monitor_refcount, 1, __ATOMIC_RELAXED))
>  		return 0;

Same question in reverse - if rte_intr_callback_unregister fails, the
refcount will have been decreased anyway, so nothing will close the
file handle, right?

>  	ret = rte_intr_callback_unregister(&intr_handle, dev_uev_handler,
> @@ -339,7 +339,6 @@ rte_dev_event_monitor_stop(void)
>  
>  	close(intr_handle.fd);
>  	intr_handle.fd = -1;
> -	monitor_started = false;
>  
>  	return 0;
>  }