On 2/10/23 10:13 PM, Ming Lei wrote: > On Fri, Feb 10, 2023 at 02:54:29PM -0700, Jens Axboe wrote: >> On 2/10/23 8:32 AM, Ming Lei wrote: >>> Hello, >>> >>> Add two OPs which buffer is retrieved via kernel splice for supporting >>> fuse/ublk zero copy. >>> >>> The 1st patch enhances direct pipe & splice for moving pages in kernel, >>> so that the two added OPs won't be misused, and avoid potential security >>> hole. >>> >>> The 2nd patch allows splice_direct_to_actor() caller to ignore signal >>> if the actor won't block and can be done in bound time. >>> >>> The 3rd patch add the two OPs. >>> >>> The 4th patch implements ublk's ->splice_read() for supporting >>> zero copy. >>> >>> ublksrv(userspace): >>> >>> https://github.com/ming1/ubdsrv/commits/io_uring_splice_buf >>> >>> So far, only loop/null target implements zero copy in above branch: >>> >>> ublk add -t loop -f $file -z >>> ublk add -t none -z >>> >>> Basic FS/IO function is verified, mount/kernel building & fio >>> works fine, and big chunk IO(BS: 64k/512k) performance gets improved >>> obviously. >> >> Do you have any performance numbers? > > Simple test on ublk-loop over image in btrfs shows the improvement is > 100% ~ 200%. That is pretty tasty... >> Also curious on liburing regression >> tests, would be nice to see as it helps with review. > > It isn't easy since it requires ublk device so far, it looks like > read to/write from one device buffer. It can't be tested without ublk itself? Surely the two new added ops can have separate test cases? -- Jens Axboe
