On Sun, 7 Dec 2025 at 11:29, Jon Perryman <[email protected]> wrote:
> On Sat, 6 Dec 2025 22:39:51 -0500, Tony Harminc <[email protected]> wrote: > > >On Sat, 6 Dec 2025 at 15:53, Binyamin Dissen wrote: > >> Does the cache get flushed if, for example, SAC 512 is issued when in > AR mode? > > > >For SAC, the PofO says "a serialization and checkpoint-synchronization > >function *is *performed before the operation begins and again after the > >operation is completed" > > Hardware serialize (cpu) and checkpoint doesn't cause cache to be flushed > otherwise 32MB L2 cache wouldn't be useful. Instead, checkpoint commits > cache changes to storage. I suspect CPU serialization is required for > checkpoint to guarantee multiple cores aren't committing the same cache at > the same time. Think of this in terms of the "compare & swap" instruction. > Checkpoint is caused by various instructions (PC, SVC, ...), interrupts, > checkpoint interval and more. > > >I don't see that they can fast-path (i.e. skip) these in the > "SAC-to-current" case, > >because the lack of the checkpoint-synchronization might in theory be > >detectable by the program. > > Cache checkpoint is a hardware feature. Why would a program need to detect > it? It's not a question of why a program would need to detect it (in the normal course it would be extremely rare); it's a matter of whether the program could, in theory, detect optimizations done by the hardware. The PofO guaranties that most such optimizations cannot be detected by a program running on the same CPU. So any optimization that can be detected would be in violation of the PofO, and IBM is highly unlikely to do that. Although IBM has been documenting this kind of stuff for decades, it became sharply important - rather than just academic niceness - as this is at the root of integrity exposures like the Spectre/Meltdown vulnerabilities. Before you get all carried away on this topic, let me be quite clear that I don't think Binyamin's queries relate to security/integrity issues. Rather, my answer just points out that IBM is unlikely to ship functionality on a machine that violates the PofO, unless they insert a "model dependant" exception that allows for it. Tony H. ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO IBM-MAIN
