Message-ID: <aYHlhKiSh8YrXjT1@thinkpad>
Date: Tue, 3 Feb 2026 13:09:40 +0100
From: Felix Maurer <fmaurer@...hat.com>
To: Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Cc: netdev@...r.kernel.org, davem@...emloft.net, edumazet@...gle.com,
	kuba@...nel.org, pabeni@...hat.com, horms@...nel.org,
	jkarrenpalo@...il.com, tglx@...utronix.de, mingo@...nel.org,
	allison.henderson@...cle.com, petrm@...dia.com, antonio@...nvpn.net
Subject: Re: [PATCH net-next v2 5/9] selftests: hsr: Add tests for more link
 faults with PRP

On Mon, Feb 02, 2026 at 05:45:50PM +0100, Sebastian Andrzej Siewior wrote:
> On 2026-02-02 12:30:44 [+0100], Felix Maurer wrote:
> > On Thu, Jan 29, 2026 at 02:32:17PM +0100, Sebastian Andrzej Siewior wrote:
> > > On 2026-01-22 15:57:00 [+0100], Felix Maurer wrote:
> > > > Add tests where one link has different rates of packet loss or reorders
> > > > packets. PRP should still be able to recover from these link faults and
> > > > show no packet loss.  However, it is acceptable to receive some level of
> > > > duplicate packets. This matches the current specification (IEC
> > > > 62439-3:2021) of the duplicate discard algorithm that requires it to be
> > > > "designed such that it never rejects a legitimate frame, while occasional
> > > > acceptance of a duplicate can be tolerated." The rate of acceptable
> > > > duplicates in this test is intentionally high (10%) to make the test
> > > > stable, the values I observed in the worst test cases (20% loss) are around
> > > > 5% duplicates.
> > >
> > > Do you know why we have the duplicates? It is because of the high
> > > sending rate at which point the blocks are dropped and we can't
> > > recognise the duplicate?
> >
> > I could not give a definitive answer on this last week, I had to dig a
> > bit and think through this. The reason for the duplicates is not a high
> > sending rate leading to dropped blocks. Rather, it's the low sending
> > rate leading to expiring blocks: in the test, the ping interval is set
> > to 10ms. As the blocks expire based on the timestamp of the first
> > received sequence number, the 400ms entry expiry gets hit every ca. 40
> > frames, at which we basically clear all received sequence number in the
> > block by creating a new block. This happens on both nodes (for ping
> > requests and replies) ca. 10 times during the test, leading to about 20
> > duplicate packets, which is the ca. 5% duplicates I observed in multiple
> > runs of this test.
> >
> > This means the duplicate rate is highly dependent on the traffic pattern
> > and the test manages to hit exactly this situation. I'll add the
> > explanation to the message in v3. If similar traffic patterns are hit in
> > the real world too often, we could test different block sizes. I just
> > did this and the results are as expected:
> > - 64 bit blocks: the duplicate rate is about the same because the block
> >   still expires once.
> > - 32 bit blocks: the duplicate rate drops to around 4% because,
> >   including the netem delay in the test, this is roughly the edge
> >   between expiring a block and already hitting a new one. But if we keep
> >   the bitmap as it is, we waste 50% of memory.
> > - 16 bit blocks: duplicate rate is zero because we use new blocks way
> >   before the expiry time. But this wastes about 75% memory for each
> >   block and the overhead of the timestamp in each block gets quite
> >   significant.
> >
> > I'd suggest to keep the 128 bit blocks to have a chance to detect
> > duplicates at higher data rates for some delay difference (up to about
> > 5.5ms with 1Gbit/s). 64 bit blocks would halve the maximum delay
> > difference, but probably still strike a good balance between memory
> > overhead and duplicate rate. Do you prefer one over the other?
>
> We had false positives with the old algorithm so having now some is not
> the end of the world. 128bit blocks are fine I guess. The question is
> how far do we want to tune it and everything comes with a price.
>
> If I update bitmap every time there is a valid packet:
>
> diff --git a/net/hsr/hsr_framereg.c b/net/hsr/hsr_framereg.c
> index 08f91f10a9347..bb9b13940c35b 100644
> --- a/net/hsr/hsr_framereg.c
> +++ b/net/hsr/hsr_framereg.c
> @@ -572,6 +572,7 @@ static int hsr_check_duplicate(struct hsr_frame_info *frame,
>  	if (test_and_set_bit(seq_bit, block->seq_nrs[port_type]))
>  		goto out_seen;
>
> +	block->time = jiffies;
>  out_new:
>  	spin_unlock_bh(&node->seq_out_lock);
>  	return 0;
>
> then I the test
> | TEST: test_cut_link - HSRv1                                         [ OK ]
> | INFO: Wait for node table entries to be merged.
> | INFO: Running ping node1-NUb1aJ -> 100.64.0.2
> | INFO: 400 packets transmitted, 400 received, +22 duplicates, 0% packet loss, time 6113ms
>
> turns into
>
> | TEST: test_cut_link - HSRv1                                         [ OK ]
> | INFO: Wait for node table entries to be merged.
> | INFO: Running ping node1-6i4xNe -> 100.64.0.2
> | INFO: 400 packets transmitted, 400 received, 0% packet loss, time 6067ms
>
> which could be considered as an improvement.

Thank you for testing this! That's indeed an improvement.

> On the other hand the live
> time of the block gets pushed past the HSR_ENTRY_FORGET_TIME so
> the first entry survives longer theoretically 127 * (400-1) = ~50 seconds.
> In the reboot case the box should be quiet for 500ms at which point the
> entry gets removed anyway so the longer "hold time" shouldn't matter.
>
> It sounds like a good idea but I don't want to rush anything ;)

I agree, the reboot case is safe because of the 500ms quiet period. But
I kinda fear a situation where a node sends a lot of packets (making
it's sequence number wrap within the 400ms), but only a small fraction
of that to us, so that we don't clear blocks because of a full buffer.
That could make a block live pretty long, when we hit it multiple times,
which could in turn lead to valid, new frames being dropped. I'd say
it's a somewhat artificial scenario, but not impossible in reality
(especially considering that I'd expect a good chunk of the traffic in
HSR and PRP networks being periodic, which may lead to reliably hitting
the same blocks over and over).

Thanks,
   Felix

Powered by blists - more mailing lists