On 27/05/2020 19:41, Xiaoguang Wang wrote: >> >> See a comment for the [2/3], can spark some ideas. >> >> It's a bit messy and makes it more difficult to keep in mind -- all that extra >> state (i.e. initialised or not) + caring whether func was already set. IMHO, the >> nop-test do not really justifies extra complexity, unless the whole stuff is >> pretty and clear. Can you benchmark something more realistic? at least >> reads/writes to null_blk (completion_nsec=0). > Indeed for this patch set, I also don't expect any obvious performance improvement, > just think current codes are not good, so try to improve it. I'm sure, we'll figure out something good in the process! There are shaky places where io_uring can use having a more orderly workflow. > I will send a v2 version later, in which I'll use null_blk to evaluate performance, > please have a check. > -- Pavel Begunkov
