On 2024-06-30 17:23:36 [-0400], Kent Overstreet wrote: > On Sun, Jun 30, 2024 at 11:11:42PM GMT, Sebastian Andrzej Siewior wrote: > > On 2024-06-28 16:20:27 [-0400], Kent Overstreet wrote: > > > > Kent, you said you didn't want it where it currently is. Fine. You said > > > > you want it at the front next to `flags'. This isn't going to work since > > > > there is no space left. You didn't make another suggestion or say how to > > > > make room. > > > > > > It doesn't need to be on the exact same cacheline, just as near as you > > > can get it. > > > > the first possible thing would be somewhere after the scheduler. > > However, what difference does it make if it s two cache lines later or > > more? I don't understand the requirement "closer". > > take advantage of CPU prefetching; CPUs will bring in more than just the > cacheline you touched because 64 bytes is small and it's cheap to fetch > from the same DRAM bank while it's open. Looking at the layout: | unsigned int flags; /* 44 4 */ | unsigned int ptrace; /* 48 4 */ | int on_cpu; /* 52 4 */ | struct __call_single_node wake_entry; /* 56 16 */ | /* --- cacheline 1 boundary (64 bytes) was 8 bytes ago --- */ … Starting with sched … | struct sched_statistics stats __attribute__((__aligned__(64))); /* 704 256 */ | | /* XXX last struct has 32 bytes of padding */ sched end, earliest spot imho | /* --- cacheline 15 boundary (960 bytes) --- */ | unsigned int btrace_seq; /* 960 4 */ If I add this before `btrace_seq' right after `stats' then it will be 14 caches lines later or 912 bytes after. How big is this prefetch going to be? Sebastian
