Re: [PATCH net-next v6] net: af_packet: Use hrtimer to do the retire operation

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Xin Zhao wrote:
> In a system with high real-time requirements, the timeout mechanism of
> ordinary timers with jiffies granularity is insufficient to meet the
> demands for real-time performance. Meanwhile, the optimization of CPU
> usage with af_packet is quite significant. Use hrtimer instead of timer
> to help compensate for the shortcomings in real-time performance.
> In HZ=100 or HZ=250 system, the update of TP_STATUS_USER is not real-time
> enough, with fluctuations reaching over 8ms (on a system with HZ=250).
> This is unacceptable in some high real-time systems that require timely
> processing of network packets. By replacing it with hrtimer, if a timeout
> of 2ms is set, the update of TP_STATUS_USER can be stabilized to within
> 3 ms.
> 
> Signed-off-by: Xin Zhao <jackzxcui1989@163.com>
> 
> ---
> Changes in v6:
> - Use hrtimer_is_queued instead to check whether it is within the callback function.
>   So do not need to add 'bool callback' parameter to _prb_refresh_rx_retire_blk_timer
>   as suggested by Willem de Bruijn;
> - Do not need local_irq_save and local_irq_restore to protect the race of the timer
>   callback running in softirq context or the open_block from tpacket_rcv in process
>   context
>   as suggested by Willem de Bruijn;
> 
> Changes in v5:
> - Remove the unnecessary comments at the top of the _prb_refresh_rx_retire_blk_timer,
>   branch is self-explanatory enough
>   as suggested by Willem de Bruijn;
> - Indentation of _prb_refresh_rx_retire_blk_timer, align with first argument on
>   previous line
>   as suggested by Willem de Bruijn;
> - Do not call hrtimer_start within the hrtimer callback
>   as suggested by Willem de Bruijn
>   So add 'bool callback' parameter to _prb_refresh_rx_retire_blk_timer to indicate
>   whether it is within the callback function. Use hrtimer_forward_now instead of
>   hrtimer_start when it is in the callback function and is doing prb_open_block.
> - Link to v5: https://lore.kernel.org/all/20250819091447.1199980-1-jackzxcui1989@163.com/
> 
> Changes in v4:
> - Add 'bool start' to distinguish whether the call to _prb_refresh_rx_retire_blk_timer
>   is for prb_open_block. When it is for prb_open_block, execute hrtimer_start to
>   (re)start the hrtimer; otherwise, use hrtimer_forward_now to set the expiration
>   time as it is more commonly used compared to hrtimer_set_expires.
>   as suggested by Willem de Bruijn;
> - Delete the comments to explain why hrtimer_set_expires(not hrtimer_forward_now)
>   is used, as we do not use hrtimer_set_expires any more;
> - Link to v4: https://lore.kernel.org/all/20250818050233.155344-1-jackzxcui1989@163.com/
> 
> Changes in v3:
> - return HRTIMER_NORESTART when pkc->delete_blk_timer is true
>   as suggested by Willem de Bruijn;
> - Drop the retire_blk_tov field of tpacket_kbdq_core, add interval_ktime instead
>   as suggested by Willem de Bruijn;
> - Add comments to explain why hrtimer_set_expires(not hrtimer_forward_now) is used in
>   _prb_refresh_rx_retire_blk_timer
>   as suggested by Willem de Bruijn;
> - Link to v3: https://lore.kernel.org/all/20250816170130.3969354-1-jackzxcui1989@163.com/
> 
> Changes in v2:
> - Drop the tov_in_msecs field of tpacket_kbdq_core added by the patch
>   as suggested by Willem de Bruijn;
> - Link to v2: https://lore.kernel.org/all/20250815044141.1374446-1-jackzxcui1989@163.com/
> 
> Changes in v1:
> - Do not add another config for the current changes
>   as suggested by Eric Dumazet;
> - Mention the beneficial cases 'HZ=100 or HZ=250' in the changelog
>   as suggested by Eric Dumazet;
> - Add some performance details to the changelog
>   as suggested by Ferenc Fejes;
> - Delete the 'pkc->tov_in_msecs == 0' bounds check which is not necessary
>   as suggested by Willem de Bruijn;
> - Use hrtimer_set_expires instead of hrtimer_start_range_ns when retire timer needs update
>   as suggested by Willem de Bruijn. Start the hrtimer in prb_setup_retire_blk_timer;
> - Just return HRTIMER_RESTART directly as all cases return the same value
>   as suggested by Willem de Bruijn;
> - Link to v1: https://lore.kernel.org/all/20250813165201.1492779-1-jackzxcui1989@163.com/
> - Link to v0: https://lore.kernel.org/all/20250806055210.1530081-1-jackzxcui1989@163.com/
> ---
>  net/packet/af_packet.c | 40 +++++++++++++++++++++++-----------------
>  net/packet/diag.c      |  2 +-
>  net/packet/internal.h  |  5 ++---
>  3 files changed, 26 insertions(+), 21 deletions(-)
> 
> diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> index a7017d7f0..9b13939a6 100644
> --- a/net/packet/af_packet.c
> +++ b/net/packet/af_packet.c
> @@ -203,8 +203,8 @@ static void prb_retire_current_block(struct tpacket_kbdq_core *,
>  static int prb_queue_frozen(struct tpacket_kbdq_core *);
>  static void prb_open_block(struct tpacket_kbdq_core *,
>  		struct tpacket_block_desc *);
> -static void prb_retire_rx_blk_timer_expired(struct timer_list *);
> -static void _prb_refresh_rx_retire_blk_timer(struct tpacket_kbdq_core *);
> +static enum hrtimer_restart prb_retire_rx_blk_timer_expired(struct hrtimer *);
> +static void _prb_refresh_rx_retire_blk_timer(struct tpacket_kbdq_core *, bool);
>  static void prb_fill_rxhash(struct tpacket_kbdq_core *, struct tpacket3_hdr *);
>  static void prb_clear_rxhash(struct tpacket_kbdq_core *,
>  		struct tpacket3_hdr *);
> @@ -581,7 +581,7 @@ static __be16 vlan_get_protocol_dgram(const struct sk_buff *skb)
>  
>  static void prb_del_retire_blk_timer(struct tpacket_kbdq_core *pkc)
>  {
> -	timer_delete_sync(&pkc->retire_blk_timer);
> +	hrtimer_cancel(&pkc->retire_blk_timer);
>  }
>  
>  static void prb_shutdown_retire_blk_timer(struct packet_sock *po,
> @@ -603,9 +603,8 @@ static void prb_setup_retire_blk_timer(struct packet_sock *po)
>  	struct tpacket_kbdq_core *pkc;
>  
>  	pkc = GET_PBDQC_FROM_RB(&po->rx_ring);
> -	timer_setup(&pkc->retire_blk_timer, prb_retire_rx_blk_timer_expired,
> -		    0);
> -	pkc->retire_blk_timer.expires = jiffies;
> +	hrtimer_setup(&pkc->retire_blk_timer, prb_retire_rx_blk_timer_expired,
> +		      CLOCK_MONOTONIC, HRTIMER_MODE_REL_SOFT);
>  }
>  
>  static int prb_calc_retire_blk_tmo(struct packet_sock *po,
> @@ -672,11 +671,10 @@ static void init_prb_bdqc(struct packet_sock *po,
>  	p1->last_kactive_blk_num = 0;
>  	po->stats.stats3.tp_freeze_q_cnt = 0;
>  	if (req_u->req3.tp_retire_blk_tov)
> -		p1->retire_blk_tov = req_u->req3.tp_retire_blk_tov;
> +		p1->interval_ktime = ms_to_ktime(req_u->req3.tp_retire_blk_tov);
>  	else
> -		p1->retire_blk_tov = prb_calc_retire_blk_tmo(po,
> -						req_u->req3.tp_block_size);
> -	p1->tov_in_jiffies = msecs_to_jiffies(p1->retire_blk_tov);
> +		p1->interval_ktime = ms_to_ktime(prb_calc_retire_blk_tmo(po,
> +						req_u->req3.tp_block_size));
>  	p1->blk_sizeof_priv = req_u->req3.tp_sizeof_priv;
>  	rwlock_init(&p1->blk_fill_in_prog_lock);
>  
> @@ -689,10 +687,14 @@ static void init_prb_bdqc(struct packet_sock *po,
>  /*  Do NOT update the last_blk_num first.
>   *  Assumes sk_buff_head lock is held.
>   */
> -static void _prb_refresh_rx_retire_blk_timer(struct tpacket_kbdq_core *pkc)
> +static void _prb_refresh_rx_retire_blk_timer(struct tpacket_kbdq_core *pkc,
> +					     bool start)
>  {
> -	mod_timer(&pkc->retire_blk_timer,
> -			jiffies + pkc->tov_in_jiffies);
> +	if (start && !hrtimer_is_queued(&pkc->retire_blk_timer))
> +		hrtimer_start(&pkc->retire_blk_timer, pkc->interval_ktime,
> +			      HRTIMER_MODE_REL_SOFT);
> +	else
> +		hrtimer_forward_now(&pkc->retire_blk_timer, pkc->interval_ktime);

Is the hrtimer still queued when prb_retire_rx_blk_timer_expired
fires? Based on the existence of hrtimer_forward_now, I assume so. But
have not checked yet. If so, hrtimer_is_queued alone suffices to
detect the other callstack from tpacket_rcv where hrtimer_start is
needed. No need for bool start?

>  	pkc->last_kactive_blk_num = pkc->kactive_blk_num;
>  }