Hi Boris and Tony, On 1/9/2023 1:39 PM, Borislav Petkov wrote: > On Mon, Jan 09, 2023 at 09:25:32PM +0000, Luck, Tony wrote: >> It feels like the old "rule" was "make it visible in /proc/cpuid" unless there was some >> good reason NOT to do it. But that has resulted in the "flags" line getting ridiculously >> long and hard for humans to read (141 fields with 926 bytes on my Skylake, >> more on more modern CPUs). > > Yap, imagine every possible CPUID bit were in there... > >> I don't know if we'd break anything if we dropped: >> >> cat_l3 cdp_l3 mba cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local >> >> from /proc/cpuinfo. > > I wouldn't mind if we remove them from cpuinfo, frankly. I am afraid that I am not aware of all the resctrl user space apps and tools being used. I did take a quick look at intel-cmt-cat that has an active user base and from what I can tell it uses /proc/cpuinfo to learn some capabilities: Example of looking for "cqm" (although "cqm" is not in Tony's list): https://github.com/intel/intel-cmt-cat/blob/master/lib/os_cap.c#L420 Example of looking for "cdp_l3": https://github.com/intel/intel-cmt-cat/blob/master/lib/os_cap.c#L520 Example of looking for "cdp_l2": https://github.com/intel/intel-cmt-cat/blob/master/lib/os_cap.c#L564 > >> Perhaps the "rule" should be written in Documentation/{somewhere}? > > This started documenting it: > > Documentation/x86/cpuinfo.rst > We could make a rule that no more resctrl related features are added to cpuinfo but I am hesitant to remove the ones that are already there. Reinette
