On 3/28/24 15:38, Christoph Hellwig wrote: > On Thu, Mar 28, 2024 at 03:33:13PM +0900, Damien Le Moal wrote: >> Ha. OK. I did not see that one. But that means that the plug kfree() can then >> block the caller. Given that the last ref drop may happen from BIO completion >> context (when the last write to a zone making the zone full complete), I do not >> think we can use this function... > > Ah, damn. So yes, we probably still need the rcu head. We can kill > the gendisk pointer, though. Or just stick with the existing version > and don't bother with the micro-optimization, at which point the > mempool might actually be the simpler implementation? I am all for not micro-optimizing the free path right now. I am not so sure about the mempool being simpler... And I do see some improvements in perf for SMR HDDs with the free list. Could be noise though but it feels a little more solid perf-wise. I have not seen any benefit for faster devices with the free list though... If you prefer the mempool, I can go back to using it though, not a big deal. For other micro-optimizations worth looking at later would be to try out the new low latency workqueues for the plug BIO work. -- Damien Le Moal Western Digital Research
