[PATCH v6] net/ice: fix ice dcf control thread crash

[Zhang, Qi Z]

> -----Original Message-----
> From: Ye, MingjinX <mingjinx.ye at intel.com>
> Sent: Wednesday, March 22, 2023 1:56 PM
> To: dev at dpdk.org
> Cc: Yang, Qiming <qiming.yang at intel.com>; stable at dpdk.org; Zhou, YidingX
> <yidingx.zhou at intel.com>; Ye, MingjinX <mingjinx.ye at intel.com>; Zhang,
> Ke1X <ke1x.zhang at intel.com>; Zhang, Qi Z <qi.z.zhang at intel.com>
> Subject: [PATCH v6] net/ice: fix ice dcf control thread crash
> 
> The control thread accesses the hardware resources after the resources were
> released, which results in a segment error.
> 
> The 'ice-reset' threads are detached, so thread resources cannot be
> reclaimed by `pthread_join` calls.
> 
> This commit synchronizes the number of "ice-reset" threads by adding a
> variable ("vsi_update_thread_num") to the "struct ice_dcf_hw" and
> performing an atomic operation on this variable. When releasing HW
> resources, we wait for the number of "ice-reset" threads to be reduced to 0
> before releasing the resources.
> 
> Fixes: c7e1a1a3bfeb ("net/ice: refactor DCF VLAN handling")
> Fixes: 7564d5509611 ("net/ice: add DCF hardware initialization")
> Fixes: 0b02c9519432 ("net/ice: handle PF initialization by DCF")
> Cc: stable at dpdk.org
> 
> Signed-off-by: Ke Zhang <ke1x.zhang at intel.com>
> Signed-off-by: Mingjin Ye <mingjinx.ye at intel.com>
> ---
> v2: add pthread_exit() for windows
> ---
> v3: Optimization. It is unsafe for a thread to forcibly exit, which will cause the
> spin lock to not be released correctly
> ---
> v4: Safely wait for all event threads to end
> ---
> v5: Spinlock moved to struct ice_dcf_hw
> ---
> v6: Spinlock changed to atomic
> ---
>  drivers/net/ice/ice_dcf.c        | 9 +++++++++
>  drivers/net/ice/ice_dcf.h        | 2 ++
>  drivers/net/ice/ice_dcf_parent.c | 6 ++++++
>  3 files changed, 17 insertions(+)
> 
> diff --git a/drivers/net/ice/ice_dcf.c b/drivers/net/ice/ice_dcf.c index
> 1c3d22ae0f..adf2cf2cb6 100644
> --- a/drivers/net/ice/ice_dcf.c
> +++ b/drivers/net/ice/ice_dcf.c
> @@ -32,6 +32,8 @@
>  #define ICE_DCF_ARQ_MAX_RETRIES 200
>  #define ICE_DCF_ARQ_CHECK_TIME  2   /* msecs */
> 
> +#define ICE_DCF_CHECK_INTERVAL  100   /* 100ms */
> +
>  #define ICE_DCF_VF_RES_BUF_SZ	\
>  	(sizeof(struct virtchnl_vf_resource) +	\
>  		IAVF_MAX_VF_VSI * sizeof(struct virtchnl_vsi_resource)) @@
> -639,6 +641,8 @@ ice_dcf_init_hw(struct rte_eth_dev *eth_dev, struct
> ice_dcf_hw *hw)
>  	rte_spinlock_init(&hw->vc_cmd_queue_lock);
>  	TAILQ_INIT(&hw->vc_cmd_queue);
> 
> +	__atomic_store_n(&hw->vsi_update_thread_num, 0,
> __ATOMIC_RELAXED);
> +
>  	hw->arq_buf = rte_zmalloc("arq_buf", ICE_DCF_AQ_BUF_SZ, 0);
>  	if (hw->arq_buf == NULL) {
>  		PMD_INIT_LOG(ERR, "unable to allocate AdminQ buffer
> memory"); @@ -760,6 +764,11 @@ ice_dcf_uninit_hw(struct rte_eth_dev
> *eth_dev, struct ice_dcf_hw *hw)
>  	rte_intr_callback_unregister(intr_handle,
>  				     ice_dcf_dev_interrupt_handler, hw);
> 
> +	/* Wait for all `ice-thread` threads to exit. */
> +	while (__atomic_load_n(&hw->vsi_update_thread_num,
> +		__ATOMIC_ACQUIRE) != 0)
> +		rte_delay_ms(ICE_DCF_CHECK_INTERVAL);
> +
>  	ice_dcf_mode_disable(hw);
>  	iavf_shutdown_adminq(&hw->avf);
> 
> diff --git a/drivers/net/ice/ice_dcf.h b/drivers/net/ice/ice_dcf.h index
> 7f42ebabe9..7becf6d187 100644
> --- a/drivers/net/ice/ice_dcf.h
> +++ b/drivers/net/ice/ice_dcf.h
> @@ -105,6 +105,8 @@ struct ice_dcf_hw {
>  	void (*vc_event_msg_cb)(struct ice_dcf_hw *dcf_hw,
>  				uint8_t *msg, uint16_t msglen);
> 
> +	int vsi_update_thread_num;
> +
>  	uint8_t *arq_buf;
> 
>  	uint16_t num_vfs;
> diff --git a/drivers/net/ice/ice_dcf_parent.c
> b/drivers/net/ice/ice_dcf_parent.c
> index 01e390ddda..6ee0fbcd2b 100644
> --- a/drivers/net/ice/ice_dcf_parent.c
> +++ b/drivers/net/ice/ice_dcf_parent.c
> @@ -154,6 +154,9 @@ ice_dcf_vsi_update_service_handler(void *param)
> 
>  	free(param);
> 
> +	__atomic_fetch_sub(&hw->vsi_update_thread_num, 1,
> +		__ATOMIC_RELEASE);
> +
>  	return NULL;
>  }
> 
> @@ -183,6 +186,9 @@ start_vsi_reset_thread(struct ice_dcf_hw *dcf_hw,
> bool vfr, uint16_t vf_id)
>  		PMD_DRV_LOG(ERR, "Failed to start the thread for reset
> handling");
>  		free(param);
>  	}
> +
> +	__atomic_fetch_add(&dcf_hw->vsi_update_thread_num, 1,
> +		__ATOMIC_RELAXED);

Is this correct? you increment the counter after you start the new thread with rte_ctrl_thread_create, 
It is possible, __atomic_fetch_sub in ice_dcf_vsi_update_service_handler will be executed before __atomic_fetch_add?

Why not moving __atomic_fetch_add to the service_handler thread? So the order of "add" and "sub" can be guaranteed ?

>  }
> 
>  static uint32_t
> --
> 2.25.1