On Wed, 12 Nov 2025 12:28:01 +0000
Mark Rutland [off-list ref] wrote:

On Wed, Nov 12, 2025 at 09:58:46AM +0800, Chenghai Huang wrote:

quoted

From: Weili Qian <qianweili@huawei.com>

Starting from ARMv8.4, stp and ldp instructions become atomic.  

That's not true for accesses to Device memory types.

Per ARM DDI 0487, L.b, section B2.2.1.1 ("Changes to single-copy atomicity in
Armv8.4"):

  If FEAT_LSE2 is implemented, LDP, LDNP, and STP instructions that load
  or store two 64-bit registers are single-copy atomic when all of the
  following conditions are true:
  • The overall memory access is aligned to 16 bytes.
  • Accesses are to Inner Write-Back, Outer Write-Back Normal cacheable memory.

IIUC when used for Device memory types, those can be split, and a part
of the access could be replayed multiple times (e.g. due to an
intetrupt).

That can't be right.
IO accesses can reference hardware FIFO so must only happen once.
(Or is 'Device memory' something different from 'Device register'?
I'm also not sure that the bus cycles could get split by an interrupt,
that would require a mid-instruction interrupt - very unlikely.
Interleaving is most likely to come from another cpu.

More interesting would be whether the instructions generate a single
PCIe TLP? (perhaps even only most of the time.)
PCIe reads are high latency, anything that can be done to increase the
size of the TLP improves PIO throughput massively.

	David

I don't think we can add this generally. It is not atomic, and not
generally safe.

Mark.

...