On Fri, 31 Jul 2026 16:53:45 +0200 Morten Brørup <[email protected]> wrote:
> +TO: x86 maintainers, ThunderX maintainers > > > From: Stephen Hemminger [mailto:[email protected]] > > Sent: Wednesday, 29 July 2026 19.54 > > > > The generic and C11 lock-free stack implementations differ only in > > memory ordering. The generic version uses a full barrier where its > > own comments state an acquire fence is sufficient, and seq_cst for > > all length counter operations. > > > > Only x86 and ThunderX still used the generic version. On x86 the > > switch removes a locked add per CAS attempt in push and pop; TSO > > provides the acquire semantics. On ThunderX the pop fence weakens > > from dmb ish to dmb ishld and the push fence goes away. Unlike the > > ring, no platform selected the generic stack for measured > > performance reasons. > > > > Remove it and use the C11 implementation everywhere. > > The lack of measured performance difference documentation is not a valid > reason to remove the generic version! > > It would be reasonable to assume that x86 (and ThunderX) use the generic > version for non-insignificant performance reasons. > > If there is no performance difference, I agree with this patch. Otherwise not. > This could be verified by providing the missing measurements. > Surprisingly, the performance of the C11 version is better than the old generic version that had smp_mb. That is because C11 code generates no locked prefixes. Gets speedup of upto 60%. Between main (with rte_smp_mb) and the unified C11 version on the 32-core x86 machine: Test main (n=9) unified C11 (n=9) delta single push/pop 46.62 ±0.30 33.41 ±0.10 -28% empty pop 1.47 ±0.01 0.98 ±0.01 -33% 1 lcore, bulk 8 9.06 ±0.05 8.20 ±0.08 -10% 1 lcore, bulk 32 6.09 ±0.02 6.15 ±0.03 +1% 2 HT, bulk 8 42.05 ±0.31 39.24 ±0.52 -7% 2 HT, bulk 32 11.92 ±0.13 11.89 ±0.10 0 2 cores, bulk 8 78.90 ±0.60 72.96 ±1.11 -7% 2 cores, bulk 32 20.74 ±1.56 7.70 ±0.13 -63% 32 cores, bulk 8 6126 ±72 6121 ±89 0 32 cores, bulk 32 1953.9 ±2.9 1984.6 ±13.3 +1.6%
