On 05/02/2020 22:07, Pavel Begunkov wrote:
> This is the first series of shaving some overhead for wq-offloading.
> The 1st removes extra allocations, and the 3rd req->refs abusing.

Rechecked a couple of assumptions, this patchset is messed up.
Drop it for now.

> There are plenty of opportunities to leak memory similarly to
> the way mentioned in [PATCH 1/3], and I'm working a generic fix,
> as I need it to close holes in waiting splice(2) patches.
> 
> The untold idea behind [PATCH 3/3] is to get rid referencing even
> further. As submission ref always pin request, there is no need
> in the second (i.e. completion) ref. Even more, With a bit of
> retossing, we can get rid of req->refs at all by using non-atomic
> ref under @compl_lock, which usually will be bundled fill_event().
> I'll play with it soon. Any ideas or concerns regarding it?
> 
> Regarding [PATCH 3/3], is there better way to do it for io_poll_add()?
> 
> 
> Pavel Begunkov (3):
>   io_uring: pass sqe for link head
>   io_uring: deduce force_nonblock in io_issue_sqe()
>   io_uring: pass submission ref to async
> 
>  fs/io_uring.c | 60 +++++++++++++++++++++++++++++----------------------
>  1 file changed, 34 insertions(+), 26 deletions(-)
> 

-- 
Pavel Begunkov