Hi, On 2019/11/12 下午6:48, Amir Goldstein wrote: > On Tue, Nov 12, 2019 at 11:15 AM Miklos Szeredi <miklos@xxxxxxxxxx> wrote: >> >> On Tue, Nov 12, 2019 at 8:53 AM Jiufei Xue <jiufei.xue@xxxxxxxxxxxxxxxxx> wrote: >>> >>> Hi Miklos, >>> >>> A performance regression is observed since linux v4.19 when we do aio >>> test using fio with iodepth 128 on overlayfs. And we found that queue >>> depth of the device is always 1 which is unexpected. >>> >>> After investigation, it is found that commit 16914e6fc7 >>> (“ovl: add ovl_read_iter()”) and commit 2a92e07edc >>> (“ovl: add ovl_write_iter()”) use do_iter_readv_writev() to submit >>> requests to real filesystem. Async IOs are converted to sync IOs here >>> and cause performance regression. >>> >>> I wondered that is this a design flaw or supposed to be. >> >> It's not theoretically difficult to fix. The challenge is to do it >> without too much complexity or code duplication. >> >> Maybe best would be to introduce VFS helpers specially for stacked >> operation such as: >> >> ssize_t vfs_read_iter_on_file(struct file *file, struct kiocb >> *orig_iocb, struct iov_iter *iter); >> ssize_t vfs_write_iter_on_file(struct file *file, struct kiocb >> *orig_iocb, struct iov_iter *iter); >> >> Implementation-wise I'm quite sure we need to allocate a new kiocb and >> initialize it from the old one, adding our own completion callback. >> > > Isn't it "just" a matter of implementing ovl-aops and > generic_file_read/write_iter() will have done the aio properly? > > I don't remember at what state we left the ovl-aops experiment [1] > IIRC, it passed most xfstests, but had some more corner cases to > cover. > > Jiufei, > > If you are interested, you can try out the experimental code [2] to > see how it plays with aio, although since readpages/writepages > are not implemented, overall performance may not be better > (or even worse). > Thanks for your reply. I have checked the experimental code [2] and found that ovl_direct_IO() also converted async IOs to sync. So I don't think it can solve the regression. Thanks, Jiufei. > Thanks, > Amir. > > [1] https://marc.info/?l=linux-unionfs&m=154995908004146&w=2 > [2] https://github.com/amir73il/linux/commits/ovl-aops-wip >
