On Tue, Jul 15, 2025 at 11:37:05AM +0200, Yann Ylavic wrote: > On Mon, Jul 14, 2025 at 9:48 PM Nam Cao <namcao@xxxxxxxxxxxxx> wrote: > > > > And my lesson is that lockless is hard. I still have no clue what is the > > bug in this patch. > > Maybe this is related: > https://lore.kernel.org/all/20250704180804.3598503-1-shakeel.butt@xxxxxxxxx/ > ? Maybe, you would need to ask Xi to validate it, because I cannot reproduce the issue. But I have abandoned this patch, even if we figure out the problem, sorry. The patch was a mistake, it makes the code much more complicated. Even myself got confused when I looked at it again after a few weeks. The performance numbers were surely impressive, but it is probably not that important. I don't think the benchmark reflects real workload anyway. I'm going to simply switch the rwlock to spinlock instead. The numbers are not as nice as before, but the code is simpler. The new patch is obvious to look at. I'm running tests right now. If no issue appears then I'm going to post it. Best regards, Nam
