Re: [PATCH net-next v9 2/3] net: airoha: fix ETS QoS stats counter underflow and cross-channel corruption

Lorenzo Bianconi <[email protected]>
Newsgroups org.infradead.lists.linux-mediatek,org.infradead.lists.linux-arm-kernel,org.kernel.vger.netdev
Message-ID <amByVU4m8YuCphO2@lore-desk>
> airoha_qdma_get_tx_ets_stats() has two bugs:
> - The hardware counters read via airoha_qdma_rr() are 32-bit values
>   but are stored in u64 locals and subtracted from u64 baselines. When
>   a 32-bit hardware counter wraps around, the subtraction produces a
>   large underflow value passed to _bstats_update().
> - The baseline counters (cpu_tx_packets, fwd_tx_packets) are stored as
>   single per-device fields, but airoha_qdma_get_tx_ets_stats() is
>   called with different channel values (0-3). Each call reads a
>   different channel's hardware counter but overwrites the same
>   baseline, corrupting the delta computation for other channels.
> 
> Fix both by:
> - Narrowing the counter locals and baselines to u32 so that 32-bit
>   unsigned subtraction handles wrap-around naturally.
> - Grouping the baselines into a per-channel qos_stats array so each
>   channel tracks its own previous counter value independently.
> - Splitting the delta addition into two statements so the first u32
>   delta is widened to u64 on assignment and the second is added in
>   u64 arithmetic, preventing overflow when both deltas are large.
> 
> Fixes: 20bf7d07c956 ("net: airoha: Add sched ETS offload support")
> Reviewed-by: Simon Horman <[email protected]>
> Reviewed-by: Alexander Lobakin <[email protected]>
> Signed-off-by: Lorenzo Bianconi <[email protected]>
> ---
>  drivers/net/ethernet/airoha/airoha_eth.c | 18 +++++++++++-------
>  drivers/net/ethernet/airoha/airoha_eth.h |  7 ++++---
>  2 files changed, 15 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/net/ethernet/airoha/airoha_eth.c b/drivers/net/ethernet/airoha/airoha_eth.c
> index 41c1a0ffbdd8..aaf2a4717d12 100644
> --- a/drivers/net/ethernet/airoha/airoha_eth.c
> +++ b/drivers/net/ethernet/airoha/airoha_eth.c
> @@ -2482,16 +2482,20 @@ static int airoha_qdma_get_tx_ets_stats(struct net_device *netdev, int channel,
>  {
>  	struct airoha_gdm_dev *dev = netdev_priv(netdev);
>  	struct airoha_qdma *qdma = dev->qdma;
> +	u32 cpu_tx_packets, fwd_tx_packets;
> +	u64 tx_packets;
>  
> -	u64 cpu_tx_packets = airoha_qdma_rr(qdma, REG_CNTR_VAL(channel << 1));
> -	u64 fwd_tx_packets = airoha_qdma_rr(qdma,
> -					    REG_CNTR_VAL((channel << 1) + 1));
> -	u64 tx_packets = (cpu_tx_packets - dev->cpu_tx_packets) +
> -			 (fwd_tx_packets - dev->fwd_tx_packets);
> +	cpu_tx_packets = airoha_qdma_rr(qdma, REG_CNTR_VAL(channel << 1));
> +	fwd_tx_packets = airoha_qdma_rr(qdma,
> +					REG_CNTR_VAL((channel << 1) + 1));
> +	tx_packets = (u32)(cpu_tx_packets -
> +			   dev->qos_stats[channel].cpu_tx_packets);
> +	tx_packets += (u32)(fwd_tx_packets -
> +			    dev->qos_stats[channel].fwd_tx_packets);
>  
>  	_bstats_update(opt->stats.bstats, 0, tx_packets);
> -	dev->cpu_tx_packets = cpu_tx_packets;
> -	dev->fwd_tx_packets = fwd_tx_packets;
> +	dev->qos_stats[channel].cpu_tx_packets = cpu_tx_packets;
> +	dev->qos_stats[channel].fwd_tx_packets = fwd_tx_packets;

commenting on sashiko's report:
https://sashiko.dev/#/patchset/20260721-airoha-ethtool-priv_flags-v9-0-9c15d8b71a56%40kernel.org
- This is a pre-existing issue, but does calling _bstats_update() here from
  the netlink dump path (process context) without holding qdisc_lock() or
  disabling bottom-halves risk a seqcount deadlock?
  On 32-bit platforms, _bstats_update() uses a seqcount_t write lock via
  u64_stats_update_begin(). If this path is preempted by a softirq on the same
  CPU that updates the same shared opt->stats.bstats (e.g., during a software
  packet dequeue), will the softirq spin forever attempting to acquire the
  same seqcount lock?
  - As pointed out by sashiko, this issue has not been introduced by this
    patch. Moreover, since airoha_eth is available just on 64bit system, I
    guess this is not a real issue.

Regards,
Lorenzo

>  
>  	return 0;
>  }
> diff --git a/drivers/net/ethernet/airoha/airoha_eth.h b/drivers/net/ethernet/airoha/airoha_eth.h
> index bf1c249255bd..bf44be9f0954 100644
> --- a/drivers/net/ethernet/airoha/airoha_eth.h
> +++ b/drivers/net/ethernet/airoha/airoha_eth.h
> @@ -553,9 +553,10 @@ struct airoha_gdm_dev {
>  	struct airoha_eth *eth;
>  
>  	DECLARE_BITMAP(qos_sq_bmap, AIROHA_NUM_QOS_CHANNELS);
> -	/* qos stats counters */
> -	u64 cpu_tx_packets;
> -	u64 fwd_tx_packets;
> +	struct {
> +		u32 cpu_tx_packets;
> +		u32 fwd_tx_packets;
> +	} qos_stats[AIROHA_NUM_QOS_CHANNELS];
>  
>  	u32 flags;
>  	int nbq;
> 
> -- 
> 2.55.0
>
signature.asc (application/pgp-signature, 228 B)
-----BEGIN PGP SIGNATURE-----

iHUEABYKAB0WIQTquNwa3Txd3rGGn7Y6cBh0uS2trAUCamByVQAKCRA6cBh0uS2t
rEeCAP9oauJEsmnW2DxM567A3ateWe95l4O9RA67X8/ppn06vwEAhEQQijbawDmV
M+opV8dtSOGZsMLViinpM3cZiPy5sA0=
=d7yO
-----END PGP SIGNATURE-----
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.