Catalin Marinas <catalin.marinas@xxxxxxx> writes: > On Thu, Sep 11, 2025 at 02:58:22PM -0700, Ankur Arora wrote: >> >> Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> writes: >> >> > On Thu, 11 Sept 2025 at 16:32, Catalin Marinas <catalin.marinas@xxxxxxx> wrote: >> >> >> >> On Wed, Sep 10, 2025 at 08:46:55PM -0700, Ankur Arora wrote: >> >> > Switch out the conditional load inerfaces used by rqspinlock >> >> > to smp_cond_read_acquire_timeout(). >> >> > This interface handles the timeout check explicitly and does any >> >> > necessary amortization, so use check_timeout() directly. >> >> >> >> It's worth mentioning that the default smp_cond_load_acquire_timeout() >> >> implementation (without hardware support) only spins 200 times instead >> >> of 16K times in the rqspinlock code. That's probably fine but it would >> >> be good to have confirmation from Kumar or Alexei. >> >> >> > >> > This looks good, but I would still redefine the spin count from 200 to >> > 16k for rqspinlock.c, especially because we need to keep >> > RES_CHECK_TIMEOUT around which still uses 16k spins to amortize >> > check_timeout. >> >> By my count that amounts to ~100us per check_timeout() on x86 >> systems I've tested with cpu_relax(). Which seems quite reasonable. >> >> 16k also seems safer on CPUs where cpu_relax() is basically a NOP. > > Does this spin count work for poll_idle()? I don't remember where the > 200 value came from. Just reusing the value of POLL_IDLE_RELAX_COUNT which is is defined as 200. For the poll_idle() case I don't think the value of 200 makes sense for all architectures, so they'll need to redefine it (before defining ARCH_HAS_OPTIMIZED_POLL which gates poll_idle().) -- ankur
