On 2020-12-09, Sergey Senozhatsky <[email protected]> wrote: >> Sigh, atomic64_read() uses a spin lock in the generic implementation >> that is used on some architectures. > > Oh... So on those archs prb is not lockless in fact, it actually > takes the spin_lock each time we read the descriptor state? > > desc_read() > atomic_long_read(state_var) > atomic64_read() > raw_spin_lock_irqsave(lock, flags) > << NMI panic >> > > Am I missing something?
For the state variable we chose atomic_long_t instead of atomic64_t for this reason. atomic_long_t operations are available atomically on all architectures. However, for clear_seq we need 64-bit (even on 32-bit machines). The seqcount_latch is an excellent solution here since clear_seq does not require lockless writers. John Ogness
