On Thu, Jun 23, 2022 at 3:05 PM Huacai Chen <chenhuacai@xxxxxxxxxx> wrote: > On Thu, Jun 23, 2022 at 4:26 PM Arnd Bergmann <arnd@xxxxxxxx> wrote: > > > > On Thu, Jun 23, 2022 at 9:56 AM Huacai Chen <chenhuacai@xxxxxxxxxx> wrote: > > > On Thu, Jun 23, 2022 at 1:45 PM Guo Ren <guoren@xxxxxxxxxx> wrote: > > > > > > > > On Thu, Jun 23, 2022 at 12:46 PM Huacai Chen <chenhuacai@xxxxxxxxxxx> wrote: > > > > > > > > > > On NUMA system, the performance of qspinlock is better than generic > > > > > spinlock. Below is the UnixBench test results on a 8 nodes (4 cores > > > > > per node, 32 cores in total) machine. > > > > You are still missing an explanation here about why this is safe to > > do. Is there are > > architectural guarantee for forward progress, or do you rely on > > specific microarchitectural > > behavior? > In my understanding, "guarantee for forward progress" means to avoid > many ll/sc happening at the same time and no one succeeds. > LoongArch uses "exclusive access (with timeout) of ll" to avoid > simultaneous ll (it also blocks other memory load/store on the same > address), and uses "random delay of sc" to avoid simultaneous sc > (introduced in CPUCFG3, bit 3 and bit 4 [1]). This mechanism can > guarantee forward progress in practice. > > [1] https://loongson.github.io/LoongArch-Documentation/LoongArch-Vol1-EN.html#_cpucfg If there is an architected feature bit for the delay, does that mean that there is a chance of CPUs getting released that set this to zero? In that case, you probably need a boot-time check for this feature bit to refuse booting a kernel with qspinlock enabled when it has more than one active CPU but does not support the random backoff, and you need to make the choice user-visible, so users are able to configure their kernels using the ticket spinlock. The ticket lock may also be the best choice for smaller configurations such as a single-socket 3A5000 with four cores and no NUMA. Arnd
