Thanks for the review. > Can this lead to a deadlock on 32-bit systems?
The described odd-seqcount deadlock requires a 32-bit MANA configuration, which is not supported by either the v6 base or the rebased v7. MICROSOFT_MANA is restricted to X86_64 or little-endian ARM64. The upstream CDX transport present in the v7 base also explicitly requires 64BIT. The common GDMA core is selected by those supported transports; it does not provide a separate 32-bit MANA device path. On 64-bit, include/linux/u64_stats_sync.h has no seqcount in struct u64_stats_sync. Its update helpers are empty and fetch_retry always returns false. A reader cannot enter the odd-sequence retry loop described in the report, even if the folding task is preempted. The MTU change does not change these architecture constraints or the 64-bit helper implementation. The fold remains under RTNL and after quiescing the retiring RX drain-stat writers. I have therefore not added preemption/IRQ disabling for this report in v7. This conclusion concerns the reported seqcount deadlock, not a guarantee for a hypothetical future 32-bit port or for unrelated concurrent-writer statistics bugs. Thanks, Wei
