Re: [PATCH] io-uring: futex: cache io_futex_data than kfree

[Sidong Yang <sidong.yang@furiosa.ai>]

On Fri, Jan 24, 2025 at 06:37:13AM -0700, Jens Axboe wrote:
> On 1/23/25 6:15 PM, Sidong Yang wrote:
> > On Thu, Jan 23, 2025 at 06:36:06AM -0700, Jens Axboe wrote:
> > 
> > Hi, Jens.
> > Thanks for review!
> > 
> >> On 1/23/25 3:50 AM, Sidong Yang wrote:
> >>> If futex_wait_setup() fails in io_futex_wait(), Old code just releases
> >>> io_futex_data. This patch tries to cache io_futex_data before kfree.
> >>
> >> It's not that the patch is incorrect, but:
> >>
> >> 1) This is an error path, surely we do not care about caching in
> >>    that case. If it's often hit, then the application would be buggy.
> >>
> >> 2) If you're going to add an io_free_ifd() helper, then at least use it
> >>    for the normal case too that still open-codes it.
> > 
> > Agreed, So this patch could be make it buggy. You can drop this patch. I'll
> > find another task to work on. 
> 
> It won't make it buggy, it's just a bit questionnable if it's worth
> doing. And if it is, then it should have io_free_ifd() being used in the
> other place that puts to the cache as well, to make it complete.

I found that io_free_ifd() could be used in io_futex_complete(). It needs same
routine that try to cache or kfree. Is it good to make v2 includes changing it?

> 
> -- 
> Jens Axboe