On Mon, Jan 26, 2026 at 6:35 AM David Laight
[off-list ref] wrote:

On Fri, 23 Jan 2026 02:52:51 +0800
David Yang [off-list ref] wrote:

quoted

On 64bit arches, struct u64_stats_sync is empty and provides no help
against load/store tearing. Convert to u64_stats_t to ensure atomic
operations.

Signed-off-by: David Yang <mmyangfl@gmail.com>
---
 drivers/net/ethernet/airoha/airoha_eth.c | 136 +++++++++++------------
 drivers/net/ethernet/airoha/airoha_eth.h |  34 +++---
 2 files changed, 85 insertions(+), 85 deletions(-)

diff --git a/drivers/net/ethernet/airoha/airoha_eth.c b/drivers/net/ethernet/airoha/airoha_eth.c
index 62bcbbbe2a95..6ed220e5a094 100644
--- a/drivers/net/ethernet/airoha/airoha_eth.c
+++ b/drivers/net/ethernet/airoha/airoha_eth.c

@@ -1472,131 +1472,131 @@ static void airoha_update_hw_stats(struct airoha_gdm_port *port)

      /* TX */
      val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_PKT_CNT_H(port->id));
-     port->stats.tx_ok_pkts += ((u64)val << 32);
+     u64_stats_add(&port->stats.tx_ok_pkts, (u64)val << 32);
      val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_PKT_CNT_L(port->id));
-     port->stats.tx_ok_pkts += val;
+     u64_stats_add(&port->stats.tx_ok_pkts, val);

Wouldn't that be better written as:
        u64 val = airoha_fe_rr(eth, REG_FE_GDM_TX_OK_PKT_CNT_H(port->id));
        val = val << 32 + airoha_fe_rr(eth, REG_FE_GDM_TX_OK_PKT_CNT_L(port->id));
        port->stats.tx_ok_pkts += val;
(Assuming there something has stopped the hardware increment the register
between the two accesses, and there is an associated atomic zero.)

Otherwise you are generating 'tearing' on 64bit systems by adding the high
and low halves separately - regardless of how the stats are read.
I think that works for 32bit as well.

Making the code completely unreadable with 'special' types and 'special'
copy routines really doesn't seem worth while.
The compiler won't generate code that does 'data tearing' for aligned word
accesses, and even if it did nothing would really break.

Quoting 316580b69d0a ("u64_stats: provide u64_stats_t type"):

  On 64bit arches, struct u64_stats_sync is empty and provides no help
against load/store tearing.

  Using READ_ONCE()/WRITE_ONCE() would be needed.

The most you want is a memcpy_in_words() function that guarantees to use
'word' size tranfsers for aligned buffers - and is probably an alias for
normal memcpy() on all current systems.

See https://lore.kernel.org/r/20260120092137.2161162-1-mmyangfl@gmail.com (local)

And if you want a memcpy, you need a buffer. If you want to fold the
buffer, that's basically WRITE_ONCE().

OTOH data tearing can be seen for 64 bit adds on 32bit systems.

It is worth nothing that on 32bits the 'packet count' and 'byte count'
increments get seen as a pair, this doesn't happen on 64bit - one happens
before the other.

        David