> REG_FE_GDM_MIB_CLEAR after every read creates a race window where
> packets arriving between read and clear are lost from statistics.
> 
> Switch to a delta-based approach instead:
> 
> - 64-bit H+L registers (ok pkts/bytes, E64..L1023): read absolute
>   hardware total directly; use a local variable and max(new, old)
>   clamping to prevent intermediate visibility and torn-read regression.
> 
> - 32-bit registers (drops, bc, mc, errors, runt, long): accumulate
>   (u32)(curr - prev) into a 64-bit software counter; unsigned
>   subtraction handles wrap-around transparently.
> 
> - tx/rx_len[0] ([0,64] bucket): combines RUNT_CNT (32-bit, delta via
>   tx_runt/rx_runt) and E64_CNT (64-bit, absolute) into a single
>   assignment using a local accumulator to avoid double-counting.
> 
> Clear MIB counters once in airoha_fe_init() to establish a clean
> baseline, preventing spurious stats from pre-driver activity (kexec,
> driver rebind, warm reboot).
> 
> Merge airoha_dev_get_hw_stats() into airoha_update_hw_stats() and
> move stats_lock inside. Plain spin_lock() is correct: the function
> is only called from ndo_get_stats64() in process context. Each dev
> refreshes only its own MIB counters; sibling devs on a shared GDM3/4
> port are polled when their own netdev is queried.
> 
> Fixes: 8f4695fb67b2 ("net: airoha: better handle MIBs for GDM ports with multiple devs attached")
> Signed-off-by: Aniket Negi <aniket.negi03@gmail.com>
> Acked-by: Lorenzo Bianconi <lorenzo@kernel.org>
> ---
> Changes in v4:
> - Add max(new, old) clamping for 64-bit H+L register pairs to ensure
>   monotonically non-decreasing stats despite torn reads between H and L
> - Use local variable (tmp) for all 64-bit H+L computations to prevent
>   lockless readers from seeing intermediate values during piecewise write
> - Add one-shot MIB counter clear in airoha_fe_init() to establish a
>   clean baseline (kexec, driver rebind, warm reboot)
> - Document sibling dev polling design in commit message
> 
> Changes in v3:
> - Link to V2: https://lore.kernel.org/20260701173941.314795-1-aniket.negi03@gmail.com/
> - Add Acked-by tag from Lorenzo
> - Rename from tx_runt_cnt to tx_runt, tx_long_cnt to tx_long,
>   tx_runt_accum64 to tx_runt64
> - Rename from rx_runt_cnt to rx_runt, rx_long_cnt to rx_long,
>   rx_runt_accum64 to rx_runt64
> - Condense the marked comments in V2, remove new line after comment
> 
> Changes in v2:
> - Store _CNT_L register reads in val before adding to stats, improving
>   readability (suggested by Lorenzo Bianconi)
> - Fix double-counting bug in the RUNT+E64 combined bucket: previously
>   "+=" for E64 re-added the full absolute counter each poll; now a
>   dedicated tx_runt_accum64/rx_runt_accum64 accumulator holds the
>   running RUNT delta, and tx_len[0] is assigned (not accumulated) each
>   poll as runt_accum64 + E64_abs
> - Replace 7-element tx_len[]/rx_len[] shadow arrays in mib_prev with
>   focused tx_runt_cnt/tx_long_cnt and rx_runt_cnt/rx_long_cnt fields;
>   only RUNT and LONG are 32-bit and need wrap-around tracking
> - Rename inner struct hw_prev_stats to mib_prev; rename accumulator
>   fields to tx_runt_accum64/rx_runt_accum64 for clarity
> - Fix comment alignment in mib_prev struct block
> - Rename airoha_dev_get_hw_stats() to airoha_update_hw_stats() and
>   move the port spin_lock inside, removing the separate wrapper
> 
>  drivers/net/ethernet/airoha/airoha_eth.c | 183 +++++++++++++++--------
>  drivers/net/ethernet/airoha/airoha_eth.h |  27 ++++
>  2 files changed, 144 insertions(+), 66 deletions(-)
> 
> diff --git a/drivers/net/ethernet/airoha/airoha_eth.c b/drivers/net/ethernet/airoha/airoha_eth.c
> index 59001fd4b6f7..2d032cc6dfca 100644
> --- a/drivers/net/ethernet/airoha/airoha_eth.c
> +++ b/drivers/net/ethernet/airoha/airoha_eth.c
> @@ -493,6 +493,8 @@ static void airoha_fe_crsn_qsel_init(struct airoha_eth *eth)
>  
>  static int airoha_fe_init(struct airoha_eth *eth)
>  {
> +	int i;
> +
>  	airoha_fe_maccr_init(eth);
>  
>  	/* PSE IQ reserve */
> @@ -586,6 +588,14 @@ static int airoha_fe_init(struct airoha_eth *eth)
>  	/* enable 1:N vlan action, init vlan table */
>  	airoha_fe_set(eth, REG_MC_VLAN_EN, MC_VLAN_EN_MASK);
>  
> +	/* Clear MIB counters to establish clean baseline for delta tracking.
> +	 * This prevents spurious statistics from pre-driver activity (e.g.,
> +	 * kexec, driver rebind, warm reboot) on first poll.
> +	 */
> +	for (i = AIROHA_GDM1_IDX; i <= AIROHA_GDM4_IDX; i++)
> +		airoha_fe_set(eth, REG_FE_GDM_MIB_CLEAR(i),
> +			      FE_GDM_MIB_RX_CLEAR_MASK | FE_GDM_MIB_TX_CLEAR_MASK);

This configuration is not needed since we reset FE at module load.

> +
>  	return airoha_fe_mc_vlan_clear(eth);
>  }
>  
> @@ -1686,11 +1696,14 @@ static void airoha_qdma_stop_napi(struct airoha_qdma *qdma)
>  	}
>  }
>  
> -static void airoha_dev_get_hw_stats(struct airoha_gdm_dev *dev)
> +static void airoha_update_hw_stats(struct airoha_gdm_dev *dev)
>  {
>  	struct airoha_gdm_port *port = dev->port;
>  	struct airoha_eth *eth = dev->eth;
>  	u32 val, i = 0;
> +	u64 tmp, prev;

nit: tmp is not so meaningful, maybe better something like data?

> +
> +	spin_lock(&port->stats_lock);
>  
>  	/* Read relevant MIB for GDM with multiple port attached */
>  	if (port->id == AIROHA_GDM3_IDX || port->id == AIROHA_GDM4_IDX)
> @@ -1701,152 +1714,190 @@ static void airoha_dev_get_hw_stats(struct airoha_gdm_dev *dev)
>  
>  	u64_stats_update_begin(&dev->stats.syncp);
>  
> -	/* TX */
> +	/* TX - 64-bit H+L registers: hw accumulates the total, read directly.
> +	 * Use local variable to prevent readers from seeing intermediate values.
> +	 * Clamp to prevent regression from torn reads between H and L.
> +	 */
> +	prev = dev->stats.tx_ok_pkts;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_PKT_CNT_H(port->id));
> -	dev->stats.tx_ok_pkts += ((u64)val << 32);
> +	tmp = ((u64)val << 32);

nit: data = (u64)val << 32;

>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_PKT_CNT_L(port->id));
> -	dev->stats.tx_ok_pkts += val;
> +	tmp |= val;

nit: I would prefer "+" instead of "|"

> +	dev->stats.tx_ok_pkts = max(tmp, prev);

nit: please drop prev and just do:
	dev->stats.tx_ok_pkts = max(data, dev->stats.tx_ok_pkts);

please redo it for all the occurrences.

Regards,
Lorenzo

>  
> +	prev = dev->stats.tx_ok_bytes;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_BYTE_CNT_H(port->id));
> -	dev->stats.tx_ok_bytes += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_BYTE_CNT_L(port->id));
> -	dev->stats.tx_ok_bytes += val;
> +	tmp |= val;
> +	dev->stats.tx_ok_bytes = max(tmp, prev);
>  
> +	/* TX - 32-bit registers: accumulate delta to handle wrap-around. */
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_DROP_CNT(port->id));
> -	dev->stats.tx_drops += val;
> +	dev->stats.tx_drops += (u32)(val - dev->stats.mib_prev.tx_drops);
> +	dev->stats.mib_prev.tx_drops = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_BC_CNT(port->id));
> -	dev->stats.tx_broadcast += val;
> +	dev->stats.tx_broadcast += (u32)(val - dev->stats.mib_prev.tx_broadcast);
> +	dev->stats.mib_prev.tx_broadcast = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_MC_CNT(port->id));
> -	dev->stats.tx_multicast += val;
> +	dev->stats.tx_multicast += (u32)(val - dev->stats.mib_prev.tx_multicast);
> +	dev->stats.mib_prev.tx_multicast = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_RUNT_CNT(port->id));
> -	dev->stats.tx_len[i] += val;
> +	dev->stats.mib_prev.tx_runt64 +=
> +		(u32)(val - dev->stats.mib_prev.tx_runt);
> +	dev->stats.mib_prev.tx_runt = val;
>  
> +	/* tx_len[0]: RUNT (32-bit, delta) + E64 (64-bit, absolute). */
> +	tmp = dev->stats.mib_prev.tx_runt64;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_E64_CNT_H(port->id));
> -	dev->stats.tx_len[i] += ((u64)val << 32);
> +	tmp += ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_E64_CNT_L(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	tmp += val;
> +	dev->stats.tx_len[i++] = tmp;
>  
> +	prev = dev->stats.tx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L64_CNT_H(port->id));
> -	dev->stats.tx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L64_CNT_L(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.tx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.tx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L127_CNT_H(port->id));
> -	dev->stats.tx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L127_CNT_L(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.tx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.tx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L255_CNT_H(port->id));
> -	dev->stats.tx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L255_CNT_L(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.tx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.tx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L511_CNT_H(port->id));
> -	dev->stats.tx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L511_CNT_L(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.tx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.tx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L1023_CNT_H(port->id));
> -	dev->stats.tx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_L1023_CNT_L(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.tx_len[i++] = max(tmp, prev);
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_TX_ETH_LONG_CNT(port->id));
> -	dev->stats.tx_len[i++] += val;
> +	dev->stats.tx_len[i++] += (u32)(val - dev->stats.mib_prev.tx_long);
> +	dev->stats.mib_prev.tx_long = val;
>  
>  	/* RX */
> +	prev = dev->stats.rx_ok_pkts;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_OK_PKT_CNT_H(port->id));
> -	dev->stats.rx_ok_pkts += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_OK_PKT_CNT_L(port->id));
> -	dev->stats.rx_ok_pkts += val;
> +	tmp |= val;
> +	dev->stats.rx_ok_pkts = max(tmp, prev);
>  
> +	prev = dev->stats.rx_ok_bytes;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_OK_BYTE_CNT_H(port->id));
> -	dev->stats.rx_ok_bytes += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_OK_BYTE_CNT_L(port->id));
> -	dev->stats.rx_ok_bytes += val;
> +	tmp |= val;
> +	dev->stats.rx_ok_bytes = max(tmp, prev);
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_DROP_CNT(port->id));
> -	dev->stats.rx_drops += val;
> +	dev->stats.rx_drops += (u32)(val - dev->stats.mib_prev.rx_drops);
> +	dev->stats.mib_prev.rx_drops = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_BC_CNT(port->id));
> -	dev->stats.rx_broadcast += val;
> +	dev->stats.rx_broadcast += (u32)(val - dev->stats.mib_prev.rx_broadcast);
> +	dev->stats.mib_prev.rx_broadcast = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_MC_CNT(port->id));
> -	dev->stats.rx_multicast += val;
> +	dev->stats.rx_multicast += (u32)(val - dev->stats.mib_prev.rx_multicast);
> +	dev->stats.mib_prev.rx_multicast = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ERROR_DROP_CNT(port->id));
> -	dev->stats.rx_errors += val;
> +	dev->stats.rx_errors += (u32)(val - dev->stats.mib_prev.rx_errors);
> +	dev->stats.mib_prev.rx_errors = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_CRC_ERR_CNT(port->id));
> -	dev->stats.rx_crc_error += val;
> +	dev->stats.rx_crc_error += (u32)(val - dev->stats.mib_prev.rx_crc_error);
> +	dev->stats.mib_prev.rx_crc_error = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_OVERFLOW_DROP_CNT(port->id));
> -	dev->stats.rx_over_errors += val;
> +	dev->stats.rx_over_errors += (u32)(val - dev->stats.mib_prev.rx_over_errors);
> +	dev->stats.mib_prev.rx_over_errors = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_FRAG_CNT(port->id));
> -	dev->stats.rx_fragment += val;
> +	dev->stats.rx_fragment += (u32)(val - dev->stats.mib_prev.rx_fragment);
> +	dev->stats.mib_prev.rx_fragment = val;
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_JABBER_CNT(port->id));
> -	dev->stats.rx_jabber += val;
> +	dev->stats.rx_jabber += (u32)(val - dev->stats.mib_prev.rx_jabber);
> +	dev->stats.mib_prev.rx_jabber = val;
>  
>  	i = 0;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_RUNT_CNT(port->id));
> -	dev->stats.rx_len[i] += val;
> +	dev->stats.mib_prev.rx_runt64 +=
> +		(u32)(val - dev->stats.mib_prev.rx_runt);
> +	dev->stats.mib_prev.rx_runt = val;
>  
> +	/* rx_len[0]: RUNT (32-bit, delta) + E64 (64-bit, absolute). */
> +	tmp = dev->stats.mib_prev.rx_runt64;
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_E64_CNT_H(port->id));
> -	dev->stats.rx_len[i] += ((u64)val << 32);
> +	tmp += ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_E64_CNT_L(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	tmp += val;
> +	dev->stats.rx_len[i++] = tmp;
>  
> +	prev = dev->stats.rx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L64_CNT_H(port->id));
> -	dev->stats.rx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L64_CNT_L(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.rx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.rx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L127_CNT_H(port->id));
> -	dev->stats.rx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L127_CNT_L(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.rx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.rx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L255_CNT_H(port->id));
> -	dev->stats.rx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L255_CNT_L(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.rx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.rx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L511_CNT_H(port->id));
> -	dev->stats.rx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L511_CNT_L(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.rx_len[i++] = max(tmp, prev);
>  
> +	prev = dev->stats.rx_len[i];
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L1023_CNT_H(port->id));
> -	dev->stats.rx_len[i] += ((u64)val << 32);
> +	tmp = ((u64)val << 32);
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_L1023_CNT_L(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	tmp |= val;
> +	dev->stats.rx_len[i++] = max(tmp, prev);
>  
>  	val = airoha_fe_rr(eth, REG_FE_GDM_RX_ETH_LONG_CNT(port->id));
> -	dev->stats.rx_len[i++] += val;
> +	dev->stats.rx_len[i] += (u32)(val - dev->stats.mib_prev.rx_long);
> +	dev->stats.mib_prev.rx_long = val;
>  
>  	u64_stats_update_end(&dev->stats.syncp);
> -}
> -
> -static void airoha_update_hw_stats(struct airoha_gdm_dev *dev)
> -{
> -	struct airoha_gdm_port *port = dev->port;
> -	int i;
> -
> -	spin_lock(&port->stats_lock);
> -
> -	for (i = 0; i < ARRAY_SIZE(port->devs); i++) {
> -		if (port->devs[i])
> -			airoha_dev_get_hw_stats(port->devs[i]);
> -	}
> -
> -	/* Reset MIB counters */
> -	airoha_fe_set(dev->eth, REG_FE_GDM_MIB_CLEAR(port->id),
> -		      FE_GDM_MIB_RX_CLEAR_MASK | FE_GDM_MIB_TX_CLEAR_MASK);
>  
>  	spin_unlock(&port->stats_lock);
>  }
> diff --git a/drivers/net/ethernet/airoha/airoha_eth.h b/drivers/net/ethernet/airoha/airoha_eth.h
> index f6d01a8e8da1..fe934f9ffe8a 100644
> --- a/drivers/net/ethernet/airoha/airoha_eth.h
> +++ b/drivers/net/ethernet/airoha/airoha_eth.h
> @@ -245,6 +245,33 @@ struct airoha_hw_stats {
>  	u64 rx_fragment;
>  	u64 rx_jabber;
>  	u64 rx_len[7];
> +
> +	struct {
> +		/* Previous HW register values for 32-bit counter delta
> +		 * tracking. Storing the last seen value and accumulating
> +		 * (u32)(curr - prev) into the 64-bit software counter
> +		 * handles wrap-around transparently via unsigned arithmetic.
> +		 * tx_runt64/rx_runt64 hold the running sum of runt deltas.
> +		 * These fields are never reported to userspace.
> +		 */
> +		u32 tx_drops;
> +		u32 tx_broadcast;
> +		u32 tx_multicast;
> +		u32 tx_runt;
> +		u32 tx_long;
> +		u64 tx_runt64;
> +		u32 rx_drops;
> +		u32 rx_broadcast;
> +		u32 rx_multicast;
> +		u32 rx_errors;
> +		u32 rx_crc_error;
> +		u32 rx_over_errors;
> +		u32 rx_fragment;
> +		u32 rx_jabber;
> +		u32 rx_runt;
> +		u32 rx_long;
> +		u64 rx_runt64;
> +	} mib_prev;
>  };
>  
>  enum {
> 
> base-commit: a225f8c20712713406ae47024b8df42deacddd4a
> -- 
> 2.43.0
>