now that I'm slowly catching up… On 2021-07-02 22:25:05 [+0200], Vlastimil Babka wrote: > > - perf_5.10 stat -r 10 hackbench -g200 -s 4096 -l500 > > Old: > > | 464.967,20 msec task-clock # 27,220 CPUs utilized ( +- 0,16% ) > > New: > > | 422.865,71 msec task-clock # 4,782 CPUs utilized ( +- 0,34% ) > > The series shouldn't significantly change the memory allocator > interaction, though. > Seems there's less cycles, but more time elapsed, thus more sleeping - > is it locks becoming mutexes on RT? yes, most likely since the !RT parts are mostly unchanged. > My second guess - list_lock remains spinlock with my series, thus RT > mutex, but the current RT tree converts it to raw_spinlock. I'd hope > leaving that one as non-raw spinlock would still be much better for RT > goals, even if hackbench (which is AFAIK very slab intensive) throughput > regresses - hopefully not that much. Yes, the list_lock seems to be the case. I picked your slub-local-lock-v3r0 and changed the list_lock (+slab_lock()) to use raw_spinlock_t and disable interrupts and CPUs utilisation went to ~23CPUs (plus a bunch of warnings which probably made it a little slower again). The difference between a sleeping lock (spinlock_t) and a mutex is that we attempt not to preempt a task that acquired a spinlock_t even if it is running for some time and the scheduler would preempt it (like it would do if the task had a mutex acquired. These are the "lazy preempt" bits in the RT patch). By making the list_lock a raw_spinlock_t a lot of IRQ-flags dancing needs to be done as the page-allocator must be entered with enabled interrupts. And then there is the possibility that you may need to free some memory even if you allocate memory which requires some extra steps on RT due to the IRQ-off part. All this vanishes by keeping list_lock a spinlock_t. The kernel-build test on /dev/shm remained unchanged so that is good. Unless there is a real-world use-case, that gets worse, I don't mind keeping the spinlock_t here. I haven't seen tglx complaining so far. Sebastian
