On Mon, Nov 11, 2019 at 12:47 PM Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Mon, Nov 11, 2019 at 12:43 PM Linus Torvalds > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > > > Yeah, maybe we could have some model for marking "this is statistics, > > doesn't need to be exact". > > Side note: that marking MUST NOT be "READ_ONCE + WRITE_ONCE", because > that makes gcc create horrible code, and only makes the race worse. > > At least with a regular add, it might stay as a single r-m-w > instruction on architectures that have that, and makes the quality of > the statistics slightly better (no preemption etc). > > So that's an excellent example of where changing code to use > WRITE_ONCE actually makes the code objectively worse in practice - > even if it might be the same in theory. Yes, I believe that was the rationale of the ADD_ONCE() thing I mentioned earlier. I do not believe we have a solution right now ? We have similar non atomic increments in some virtual network drivers doing "dev->stats.tx_errors++;" in their error path.
