On 4/25/21 5:49 PM, Ming Lei wrote: > On Sun, Apr 25, 2021 at 01:42:59PM -0700, Bart Van Assche wrote: >> Using cmpxchg() on set->tags[] is only safe if all other set->tags[] >> accesses are changed into WRITE_ONCE() or READ_ONCE(). > > Why? > > Semantic of cmpxchg() is to modify value pointed by the address if its > old value is same with passed 'rq'. That is exactly what we need. > > writting 'void *' is always atomic. if someone has touched > '->rqs[tag]', cmpxchg() won't modify the value. WRITE_ONCE() supports data types that have the same size as char, short, int, long and long long. That includes void *. If writes to these data types would always be atomic then we wouldn't need the WRITE_ONCE() macro. The explanation at the top of the rwonce.h header file is as follows: "Prevent the compiler from merging or refetching reads or writes. [ ... ] Ensuring that the compiler does not fold, spindle, or otherwise mutilate accesses that either do not require ordering or that interact with an explicit memory barrier or atomic instruction that provides the required ordering." Bart.
