netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH net-next v10 0/2] net: af_packet: optimize retire operation
@ 2025-08-31 10:08 Xin Zhao
  2025-08-31 10:08 ` [PATCH net-next v10 1/2] net: af_packet: remove last_kactive_blk_num field Xin Zhao
  2025-08-31 10:08 ` [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation Xin Zhao
  0 siblings, 2 replies; 23+ messages in thread
From: Xin Zhao @ 2025-08-31 10:08 UTC (permalink / raw)
  To: willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel, Xin Zhao

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.

---
Changes in v10:
- kactive_blk_num (K) is incremented on block close. last_kactive_blk_num (L)
  is set to match K on block open and each timer. So the only time that they
  differ is if a block is closed in tpacket_rcv and no new block could be
  opened. So the origin check L==K in timer callback only skip the case 'no
  new block to open'. If we remove L==K check, it will make prb_curr_blk_in_use
  check earlier, which will not cause any side effect
  as suggested by Willem de Bruijn.
- Submit a precursor patch that removes last_kactive_blk_num
  as suggested by Willem de Bruijn.

Changes in v9:
- Remove the function prb_setup_retire_blk_timer and move hrtimer setup and start
  logic into function init_prb_bdqc
  as suggested by Willem de Bruijn.
- Always update last_kactive_blk_num before hrtimer callback return as the origin
  logic does, as suggested by Willem de Bruijn.
  In tpacket_rcv, it may call prb_close_block but do not call prb_open_block in
  prb_dispatch_next_block, leading to inconsistency between last_kactive_blk_num
  and kactive_blk_num. In hrtimer callback, we should update last_kactive_blk_num
  in this case.
- Remove 'refresh_timer:' label which is not needed while I change goto logic to
  if-else implementation.
- Link to v9: https://lore.kernel.org/netdev/20250828155127.3076551-1-jackzxcui1989@163.com/

Changes in v8:
- Delete delete_blk_timer field, as suggested by Willem de Bruijn,
  hrtimer_cancel will check and wait until the timer callback return and ensure
  enter enter callback again;
- Simplify the logic related to setting timeout, as suggestd by Willem de Bruijn.
  Currently timer callback just restarts itself unconditionally, so delete the
 'out:' label, do not forward hrtimer in prb_open_block, call hrtimer_forward_now
  directly and always return HRTIMER_RESTART. The only special case is when
  prb_open_block is called from tpacket_rcv. That would set the timeout further
  into the future than the already queued timer. An earlier timeout is not
  problematic. No need to add complexity to avoid that.
- Link to v8: https://lore.kernel.org/all/20250827150131.2193485-1-jackzxcui1989@163.com/

Changes in v7:
- Only update the hrtimer expire time within the hrtimer callback.
  When the callback return, without sk_buff_head lock protection, __run_hrtimer will
  enqueue the timer if return HRTIMER_RESTART. Setting the hrtimer expires while
  enqueuing a timer may cause chaos in the hrtimer red-black tree.
  The setting expire time is monotonic, so if we do not update the expire time to the
  retire_blk_timer when it is not in callback, it will not cause problem if we skip
  the timeout event and update it when find out that expire_ktime is bigger than the
  expire time of retire_blk_timer.
- Use hrtimer_set_expires instead of hrtimer_forward_now.
  The end time for retiring each block is not fixed because when network packets are
  received quickly, blocks are retired rapidly, and the new block retire time needs
  to be recalculated. However, hrtimer_forward_now increments the previous timeout
  by an interval, which is not correct.
- The expire time is monotonic, so if we do not update the expire time to the
  retire_blk_timer when it is not in callback, it will not cause problem if we skip
  the timeout event and update it when find out that expire_ktime is bigger than the
  expire time of retire_blk_timer.
- Adding the 'bool callback' parameter back is intended to more accurately determine
  whether we are inside the hrtimer callback when executing
  _prb_refresh_rx_retire_blk_timer. This ensures that we only update the hrtimer's
  timeout value within the hrtimer callback.
- Link to v7: https://lore.kernel.org/all/20250822132051.266787-1-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;
- Link to v6: https://lore.kernel.org/all/20250820092925.2115372-1-jackzxcui1989@163.com/

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/

Xin Zhao (2):
  net: af_packet: remove last_kactive_blk_num field
  net: af_packet: Use hrtimer to do the retire operation

 net/packet/af_packet.c | 107 +++++++++++------------------------------
 net/packet/diag.c      |   2 +-
 net/packet/internal.h  |  12 +----
 3 files changed, 32 insertions(+), 89 deletions(-)

-- 
2.34.1


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-01  2:27 Xin Zhao
  2025-09-01 13:35 ` Willem de Bruijn
  0 siblings, 1 reply; 23+ messages in thread
From: Xin Zhao @ 2025-09-01  2:27 UTC (permalink / raw)
  To: willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Sun, 2025-08-31 at 21:21 -0400, Willem wrote:

> > -		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));
> 
> req_u is not aligned with the line above.

I have some questions regarding the alignment here. According to the alignment requirements,
req_u should be aligned below the po variable. However, if it is aligned below po, the line
will become very long, which may affect readability. In this special case, can I align it to
prb_calc_retire_blk_tmo instead, or should I continue to align it to the po variable?


What should I do next?
Should I change the alignment, and resend PATCH with the reviewed information of version 10?

Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-01 14:16 Xin Zhao
  0 siblings, 0 replies; 23+ messages in thread
From: Xin Zhao @ 2025-09-01 14:16 UTC (permalink / raw)
  To: willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Mon, 2025-09-01 at 09:35 -0400, Willem wrote:

> > On Sun, 2025-08-31 at 21:21 -0400, Willem wrote:
> > 
> > > > -		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));
> > > 
> > > req_u is not aligned with the line above.
> > 
> > I have some questions regarding the alignment here. According to the alignment requirements,
> > req_u should be aligned below the po variable. However, if it is aligned below po, the line
> > will become very long, which may affect readability. In this special case, can I align it to
> > prb_calc_retire_blk_tmo instead, or should I continue to align it to the po variable?
> 
> The (minor) issue here is with the second req_u. Which is one space
> off from the argument above. See checkpath.
> 
> In general, the line length and break rules are documented in the
> kernel coding style page, which checkpatch follows.
> 
> > 
> > What should I do next?
> > Should I change the alignment, and resend PATCH with the reviewed information of version 10?
> 
> I did not think this one space was worth resending, so I added my
> Reviewed-by. Others may disagree, but so far no other opinions.

Okay, I will not resend the patch if there are no other opinions.


Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-03 16:17 Xin Zhao
  2025-09-04  2:50 ` Jason Xing
  0 siblings, 1 reply; 23+ messages in thread
From: Xin Zhao @ 2025-09-03 16:17 UTC (permalink / raw)
  To: kerneljasonxing, willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Tue, Sep 2, 2025 at 23:43 +0800 Jason Xing <kerneljasonxing@gmail.com> wrote:

> >         p1->max_frame_len = p1->kblk_size - BLK_PLUS_PRIV(p1->blk_sizeof_priv);
> >         prb_init_ft_ops(p1, req_u);
> > -       prb_setup_retire_blk_timer(po);
> > +       hrtimer_setup(&p1->retire_blk_timer, prb_retire_rx_blk_timer_expired,
> > +                     CLOCK_MONOTONIC, HRTIMER_MODE_REL_SOFT);
> > +       hrtimer_start(&p1->retire_blk_timer, p1->interval_ktime,
> > +                     HRTIMER_MODE_REL_SOFT);
> 
> You expect to see it start at the setsockopt phase? Even if it's far
> from the real use of recv at the moment.
> 
> >         prb_open_block(p1, pbd);
> >  }

Before applying this patch, init_prb_bdqc also start the timer by mod_timer:

init_prb_bdqc
  prb_open_block
    _prb_refresh_rx_retire_blk_timer
      mod_timer

So the current timer's start time is almost the same as it was before applying
the patch.


> > @@ -917,7 +873,6 @@ static void prb_open_block(struct tpacket_kbdq_core *pkc1,
> >         pkc1->pkblk_end = pkc1->pkblk_start + pkc1->kblk_size;
> >
> >         prb_thaw_queue(pkc1);
> > -       _prb_refresh_rx_retire_blk_timer(pkc1);
> 
> Could you say more on why you remove this here and only reset/update
> the expiry time in the timer handler? Probably I missed something
> appearing in the previous long discussion.
> 
> >
> >         smp_wmb();
> >  }

In the description of [PATCH net-next v10 0/2] net: af_packet: optimize retire operation:

Changes in v7:
  When the callback return, without sk_buff_head lock protection, __run_hrtimer will
  enqueue the timer if return HRTIMER_RESTART. Setting the hrtimer expires while
  enqueuing a timer may cause chaos in the hrtimer red-black tree.

Neither hrtimer_set_expires nor hrtimer_forward_now is allowed when the hrtimer has
already been enqueued. Therefore, the only place where the hrtimer timeout can be set is
within the callback, at which point the hrtimer is in a non-enqueued state and can have
its timeout set.


Changes in v8:
  Simplify the logic related to setting timeout, as suggestd by Willem de Bruijn.
  Currently timer callback just restarts itself unconditionally, so delete the
 'out:' label, do not forward hrtimer in prb_open_block, call hrtimer_forward_now
  directly and always return HRTIMER_RESTART. The only special case is when
  prb_open_block is called from tpacket_rcv. That would set the timeout further
  into the future than the already queued timer. An earlier timeout is not
  problematic. No need to add complexity to avoid that.

This paragraph explains that if the block's retire timeout is not adjusted within
the timer callback, it will only result in an earlier-than-expected retire timeout,
which is not problematic. Therefore, it is unnecessary to increase the logical complexity
to ensure block retire timeout occurs as expected each time.


> The whole structure needs a new organization?
> 
> Before:
>         /* size: 152, cachelines: 3, members: 22 */
>         /* sum members: 144, holes: 2, sum holes: 8 */
>         /* paddings: 1, sum paddings: 4 */
>         /* last cacheline: 24 bytes */
> After:
>         /* size: 176, cachelines: 3, members: 19 */
>         /* sum members: 163, holes: 4, sum holes: 13 */
>         /* paddings: 1, sum paddings: 4 */
>         /* forced alignments: 1, forced holes: 1, sum forced holes: 6 */
>         /* last cacheline: 48 bytes */

What about the following organization:?

/* kbdq - kernel block descriptor queue */
struct tpacket_kbdq_core {
	struct pgv	*pkbdq;
	unsigned int	feature_req_word;
	unsigned int	hdrlen;
	unsigned short	kactive_blk_num;
	unsigned short	blk_sizeof_priv;
	unsigned char	reset_pending_on_curr_blk;

	char		*pkblk_start;
	char		*pkblk_end;
	int		kblk_size;
	unsigned int	max_frame_len;
	unsigned int	knum_blocks;
	char		*prev;
	char		*nxt_offset;

	unsigned short  version;
	
	uint64_t	knxt_seq_num;
	struct sk_buff	*skb;

	rwlock_t	blk_fill_in_prog_lock;

	/* timer to retire an outstanding block */
	struct hrtimer  retire_blk_timer;

	/* Default is set to 8ms */
#define DEFAULT_PRB_RETIRE_TOV	(8)

	ktime_t		interval_ktime;
};



Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-03 17:07 Xin Zhao
  2025-09-04  3:26 ` Jason Xing
  0 siblings, 1 reply; 23+ messages in thread
From: Xin Zhao @ 2025-09-03 17:07 UTC (permalink / raw)
  To: kerneljasonxing, willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Wed, Sep 3, 2025 at 00:42 +0800 Jason Xing <kerneljasonxing@gmail.com> wrote:

> One more review from my side is that as to the removal of
> delete_blk_timer, I'm afraid it deserves a clarification in the commit
> message.
> 
> > > -       spin_unlock_bh(&rb_queue->lock);
> > > -
> > > -       prb_del_retire_blk_timer(pkc);
> > > -}
> > > -

In the description of [PATCH net-next v10 0/2] net: af_packet: optimize retire operation:

Changes in v8:
- Delete delete_blk_timer field, as suggested by Willem de Bruijn,
  hrtimer_cancel will check and wait until the timer callback return and ensure
  enter enter callback again;

I will also emphasize the removal of delete_blk_timer in the commit message for this 2/2
commit. The updated commit message for the 2/2 patch is as follows：

Changes in v8:
- Simplify the logic related to setting timeout.
- Delete delete_blk_timer field, hrtimer_cancel will check and wait until
  the timer callback return.


> I gradually understand your thought behind this modification. You're
> trying to move the timer operation out of prb_open_block() and then
> spread the timer operation into each caller.
> 
> You probably miss the following call trace:
> packet_current_rx_frame() -> __packet_lookup_frame_in_block() ->
> prb_open_block() -> _prb_refresh_rx_retire_blk_timer()
> ?
> 
> May I ask why bother introducing so many changes like this instead of
> leaving it as-is?




Consider the following timing sequence:
timer   cpu0 (softirq context, hrtimer timeout)                cpu1 (process context)
0       hrtimer_run_softirq
1         __hrtimer_run_queues
2           __run_hrtimer
3             prb_retire_rx_blk_timer_expired
4               spin_lock(&po->sk.sk_receive_queue.lock);
5               _prb_refresh_rx_retire_blk_timer
6                 hrtimer_forward_now
7               spin_unlock(&po->sk.sk_receive_queue.lock)
8             raw_spin_lock_irq(&cpu_base->lock);              tpacket_rcv
9             enqueue_hrtimer                                    spin_lock(&sk->sk_receive_queue.lock);
10                                                               packet_current_rx_frame
11                                                                 __packet_lookup_frame_in_block
12            finish enqueue_hrtimer                                 prb_open_block
13                                                                     _prb_refresh_rx_retire_blk_timer
14                                                                       hrtimer_is_queued(&pkc->retire_blk_timer) == true
15                                                                       hrtimer_forward_now
16                                                                         WARN_ON
On cpu0 in the timing sequence above, enqueue_hrtimer is not protected by sk_receive_queue.lock,
while the hrtimer_forward_now is not protected by raw_spin_lock_irq(&cpu_base->lock).

In my previous email, I provided an explanation. As a supplement, I would
like to reiterate a paragraph from my earlier response to Willem.
The point is that when the hrtimer is in the enqueued state, you cannot
call interfaces like hrtimer_forward_now. The kernel has a WARN_ON check
in hrtimer_forward_now for this reason. Similarly, you also cannot call
interfaces like hrtimer_set_expires. The kernel does not include a WARN_ON
check in hrtimer_set_expires to avoid increasing the code size, as
hrtimer_set_expires is an inline function.


Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-04 14:59 Xin Zhao
  2025-09-05  0:09 ` Jason Xing
  0 siblings, 1 reply; 23+ messages in thread
From: Xin Zhao @ 2025-09-04 14:59 UTC (permalink / raw)
  To: kerneljasonxing, willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Thu, Sep 4, 2025 at 10:50 +0800 Jason Xing <kerneljasonxing@gmail.com> wrote:

> > In the description of [PATCH net-next v10 0/2] net: af_packet: optimize retire operation:
> >
> > Changes in v7:
> >   When the callback return, without sk_buff_head lock protection, __run_hrtimer will
> >   enqueue the timer if return HRTIMER_RESTART. Setting the hrtimer expires while
> >   enqueuing a timer may cause chaos in the hrtimer red-black tree.
> >
> > Neither hrtimer_set_expires nor hrtimer_forward_now is allowed when the hrtimer has
> > already been enqueued. Therefore, the only place where the hrtimer timeout can be set is
> > within the callback, at which point the hrtimer is in a non-enqueued state and can have
> > its timeout set.
> 
> Can we use hrtimer_is_queued() instead? Please see tcp_pacing_check()
> as an example. But considering your following explanation, I think
> it's okay now.

Okay， let's keep the current logic as it is.



> > /* kbdq - kernel block descriptor queue */
> > struct tpacket_kbdq_core {
> >         struct pgv      *pkbdq;
> >         unsigned int    feature_req_word;
> >         unsigned int    hdrlen;
> >         unsigned short  kactive_blk_num;
> >         unsigned short  blk_sizeof_priv;
> >         unsigned char   reset_pending_on_curr_blk;
> >
> >         char            *pkblk_start;
> >         char            *pkblk_end;
> >         int             kblk_size;
> >         unsigned int    max_frame_len;
> >         unsigned int    knum_blocks;
> >         char            *prev;
> >         char            *nxt_offset;
> >
> >         unsigned short  version;
> >
> >         uint64_t        knxt_seq_num;
> >         struct sk_buff  *skb;
> >
> >         rwlock_t        blk_fill_in_prog_lock;
> >
> >         /* timer to retire an outstanding block */
> >         struct hrtimer  retire_blk_timer;
> >
> >         /* Default is set to 8ms */
> > #define DEFAULT_PRB_RETIRE_TOV  (8)
> >
> >         ktime_t         interval_ktime;
> > };
> 
> Could you share the result after running 'pahole --hex -C
> tpacket_kbdq_core vmlinux'?

I change the struct tpacket_kbdq_core as following:

/* kbdq - kernel block descriptor queue */
struct tpacket_kbdq_core {
	struct pgv	*pkbdq;
	unsigned int	feature_req_word;
	unsigned int	hdrlen;
	unsigned char	reset_pending_on_curr_blk;
	unsigned short	kactive_blk_num;
	unsigned short	blk_sizeof_priv;

	unsigned short  version;

	char		*pkblk_start;
	char		*pkblk_end;
	int		kblk_size;
	unsigned int	max_frame_len;
	unsigned int	knum_blocks;
	char		*prev;
	char		*nxt_offset;

	uint64_t	knxt_seq_num;
	struct sk_buff	*skb;

	rwlock_t	blk_fill_in_prog_lock;

	/* timer to retire an outstanding block */
	struct hrtimer  retire_blk_timer;

	/* Default is set to 8ms */
#define DEFAULT_PRB_RETIRE_TOV	(8)

	ktime_t		interval_ktime;
};


pahole --hex -C tpacket_kbdq_core vmlinux

running results:

struct tpacket_kbdq_core {
        struct pgv *               pkbdq;                /*     0   0x8 */
        unsigned int               feature_req_word;     /*   0x8   0x4 */
        unsigned int               hdrlen;               /*   0xc   0x4 */
        unsigned char              reset_pending_on_curr_blk; /*  0x10   0x1 */

        /* XXX 1 byte hole, try to pack */

        short unsigned int         kactive_blk_num;      /*  0x12   0x2 */
        short unsigned int         blk_sizeof_priv;      /*  0x14   0x2 */
        short unsigned int         version;              /*  0x16   0x2 */
        char *                     pkblk_start;          /*  0x18   0x8 */
        char *                     pkblk_end;            /*  0x20   0x8 */
        int                        kblk_size;            /*  0x28   0x4 */
        unsigned int               max_frame_len;        /*  0x2c   0x4 */
        unsigned int               knum_blocks;          /*  0x30   0x4 */

        /* XXX 4 bytes hole, try to pack */

        char *                     prev;                 /*  0x38   0x8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        char *                     nxt_offset;           /*  0x40   0x8 */
        uint64_t                   knxt_seq_num;         /*  0x48   0x8 */
        struct sk_buff *           skb;                  /*  0x50   0x8 */
        rwlock_t                   blk_fill_in_prog_lock; /*  0x58  0x30 */
        /* --- cacheline 2 boundary (128 bytes) was 8 bytes ago --- */
        struct hrtimer             retire_blk_timer __attribute__((__aligned__(8))); /*  0x88  0x40 */

        /* XXX last struct has 4 bytes of padding */

        /* --- cacheline 3 boundary (192 bytes) was 8 bytes ago --- */
        ktime_t                    interval_ktime;       /*  0xc8   0x8 */

        /* size: 208, cachelines: 4, members: 19 */
        /* sum members: 203, holes: 2, sum holes: 5 */
        /* paddings: 1, sum paddings: 4 */
        /* forced alignments: 1 */
        /* last cacheline: 16 bytes */
} __attribute__((__aligned__(8)));


Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-05  4:00 Xin Zhao
  2025-09-05  6:03 ` Jason Xing
  0 siblings, 1 reply; 23+ messages in thread
From: Xin Zhao @ 2025-09-05  4:00 UTC (permalink / raw)
  To: kerneljasonxing, willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Thu, Sep 4, 2025 at 11:26 +0800 Jason Xing <kerneljasonxing@gmail.com> wrote:

> > In the description of [PATCH net-next v10 0/2] net: af_packet: optimize retire operation:
> >
> > Changes in v8:
> > - Delete delete_blk_timer field, as suggested by Willem de Bruijn,
> >   hrtimer_cancel will check and wait until the timer callback return and ensure
> >   enter enter callback again;
> 
> I see the reason now :)
> 
> Please know that the history changes through versions will finally be
> removed, only the official message that will be kept in the git. So
> this kind of change, I think, should be clarified officially since
> you're removing a structure member. Adding more descriptions will be
> helpful to readers in the future. Thank you.

I will add some more information to the commit message of this 2/2 PATCH.



> > Consider the following timing sequence:
> > timer   cpu0 (softirq context, hrtimer timeout)                cpu1 (process context)
> > 0       hrtimer_run_softirq
> > 1         __hrtimer_run_queues
> > 2           __run_hrtimer
> > 3             prb_retire_rx_blk_timer_expired
> > 4               spin_lock(&po->sk.sk_receive_queue.lock);
> > 5               _prb_refresh_rx_retire_blk_timer
> > 6                 hrtimer_forward_now
> > 7               spin_unlock(&po->sk.sk_receive_queue.lock)
> > 8             raw_spin_lock_irq(&cpu_base->lock);              tpacket_rcv
> > 9             enqueue_hrtimer                                    spin_lock(&sk->sk_receive_queue.lock);
> > 10                                                               packet_current_rx_frame
> > 11                                                                 __packet_lookup_frame_in_block
> > 12            finish enqueue_hrtimer                                 prb_open_block
> > 13                                                                     _prb_refresh_rx_retire_blk_timer
> > 14                                                                       hrtimer_is_queued(&pkc->retire_blk_timer) == true
> > 15                                                                       hrtimer_forward_now
> > 16                                                                         WARN_ON
> > On cpu0 in the timing sequence above, enqueue_hrtimer is not protected by sk_receive_queue.lock,
> > while the hrtimer_forward_now is not protected by raw_spin_lock_irq(&cpu_base->lock).
> >
> > In my previous email, I provided an explanation. As a supplement, I would
> > like to reiterate a paragraph from my earlier response to Willem.
> > The point is that when the hrtimer is in the enqueued state, you cannot
> 
> How about tring hrtimer_is_queued() beforehand?
> 
> IIUC, with this patch applied, we will lose the opportunity to refresh
> the timer when the lookup function (in the above path I mentioned)
> gets called compared to before. If the packet socket tries to look up
> a new block and it doesn't update its expiry time, the timer will soon
> wake up. Does it sound unreasonable?


I actually pointed out the issue with the timeout setting in a previous email:
https://lore.kernel.org/netdev/20250826030328.878001-1-jackzxcui1989@163.com/.

Regarding the method you mentioned, using hrtimer_is_queued to assist in judgment, I had
discussed this extensively with Willem in previous emails, and the conclusion was that
it is not feasible. The reason is that in our scenario, the hrtimer always returns
HRTIMER_RESTART, unlike the places you pointed out, such as tcp_pacing_check, where the
corresponding hrtimer callbacks all return HRTIMER_NORESTART. Since our scenario returns
HRTIMER_RESTART, this can lead to many troublesome issues. The fundamental reason is that
if HRTIMER_RESTART is returned, the hrtimer module will enqueue the hrtimer after the
callback returns, which leads to exiting the protection of our sk_receive_queue lock.

Returning to the functionality here, if we really want to update the hrtimer's timeout
outside of the timer callback, there are two key points to note:

1. Accurately knowing whether the current context is a timer callback or tpacket_rcv.
2. How to update the hrtimer's timeout in a non-timer callback scenario.

To start with the first point, it has already been explained in previous emails that
executing hrtimer_forward outside of a timer callback is not allowed. Therefore, we
must accurately determine whether we are in a timer callback; only in that context can
we use the hrtimer_forward function to update.
In the original code, since the same _prb_refresh_rx_retire_blk_timer function was called,
distinguishing between contexts required code restructuring. Now that this patch removes
the _prb_refresh_rx_retire_blk_timer function, achieving this accurate distinction is not
too difficult.
The key issue is the second point. If we are not inside the hrtimer's callback, we cannot
use hrtimer_forward to update the timeout. So what other interface can we use? You might
suggest using hrtimer_start, but fundamentally, hrtimer_start cannot be called if it has
already been started previously. Therefore, wouldn’t you need to add hrtimer_cancel to
confirm that the hrtimer has been canceled? Once hrtimer_cancel is added, there will also
be scenarios where it is restarted, which means we need to consider the concurrent
scenario when the socket exits and also calls hrtimer_cancel. This might require adding
logic for that concurrency scenario, and you might even need to reintroduce the
delete_blk_timer variable to indicate whether the packet_release operation has been
triggered so that the hrtimer does not restart in the tpacket_rcv scenario.

In fact, in a previous v7 version, I proposed a change that I personally thought was
quite good, which can be seen here:
https://lore.kernel.org/netdev/20250822132051.266787-1-jackzxcui1989@163.com/. However,
this change introduced an additional variable and more logic. Willem also pointed out
that the added complexity to avoid a non-problematic issue was unnecessary.

As mentioned in Changes in v8:
  The only special case is when prb_open_block is called from tpacket_rcv.
  That would set the timeout further into the future than the already queued
  timer. An earlier timeout is not problematic. No need to add complexity to
  avoid that.

It is not problematic, as Willem point it out in
https://lore.kernel.org/netdev/willemdebruijn.kernel.2d7599ee951fd@gmail.com/


In the end:

So, if you agree with the current changes in v10 and do not wish to add the timeout
setting under tpacket_rcv, that’s fine.
If you do not agree, then the only alternative I can think of is to use a combination
of hrtimer_cancel and hrtimer_start under prb_open_block, and we would also need to
reintroduce the delete_blk_timer variable to help determine whether the hrtimer was
canceled due to the packet_release behavior. If we really want to make this change,
it does add quite a bit of logic, and we would also need Willem's agreement.


Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
* Re: [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation
@ 2025-09-05 14:47 Xin Zhao
  2025-09-05 16:16 ` Willem de Bruijn
  0 siblings, 1 reply; 23+ messages in thread
From: Xin Zhao @ 2025-09-05 14:47 UTC (permalink / raw)
  To: kerneljasonxing, willemdebruijn.kernel, edumazet, ferenc
  Cc: davem, kuba, pabeni, horms, netdev, linux-kernel

On Fri, Sep 5, 2025 at 14:45 +0800 Jason Xing <kerneljasonxing@gmail.com> wrote:

> BTW, I have to emphasize that after this patch, the hrtimer will run
> periodically and unconditionally. As far as I know, it's not possible
> to run hundreds and thousands packet sockets in production, so it
> might not be a huge problem. Or else, numerous timers are likely to
> cause spikes/jitters, especially when timeout is very small (which can
> be 1ms timeout for HZ=1000 system). It would be great if you state the
> possible side effects in the next version.

The original logic actually involves an unconditional restart in the timer's
callback. You might be suggesting that if packets come in particularly fast,
the original logic would reset the timeout when opening a new block in
tpacket_rcv, so the timeout does not expire immediately. However, if packets
arrive very quickly, it will also lead to frequent timeout resets, which can
waste CPU resources.
I will emphasize in the comments that the current hrtimer expiration logic
is unconditional and periodic.


Thanks
Xin Zhao


^ permalink raw reply	[flat|nested] 23+ messages in thread
end of thread, other threads:[~2025-09-05 16:16 UTC | newest]

Thread overview: 23+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-08-31 10:08 [PATCH net-next v10 0/2] net: af_packet: optimize retire operation Xin Zhao
2025-08-31 10:08 ` [PATCH net-next v10 1/2] net: af_packet: remove last_kactive_blk_num field Xin Zhao
2025-09-01  1:18   ` Willem de Bruijn
2025-09-02 14:04   ` Jason Xing
2025-08-31 10:08 ` [PATCH net-next v10 2/2] net: af_packet: Use hrtimer to do the retire operation Xin Zhao
2025-09-01  1:21   ` Willem de Bruijn
2025-09-02 15:43   ` Jason Xing
2025-09-02 16:42     ` Jason Xing
  -- strict thread matches above, loose matches on Subject: below --
2025-09-01  2:27 Xin Zhao
2025-09-01 13:35 ` Willem de Bruijn
2025-09-01 14:16 Xin Zhao
2025-09-03 16:17 Xin Zhao
2025-09-04  2:50 ` Jason Xing
2025-09-03 17:07 Xin Zhao
2025-09-04  3:26 ` Jason Xing
2025-09-04 14:59 Xin Zhao
2025-09-05  0:09 ` Jason Xing
2025-09-05  4:00 Xin Zhao
2025-09-05  6:03 ` Jason Xing
2025-09-05  6:45   ` Jason Xing
2025-09-05 16:16     ` Willem de Bruijn
2025-09-05 14:47 Xin Zhao
2025-09-05 16:16 ` Willem de Bruijn

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).