On 5/29/20 2:58 AM, Pavel Begunkov wrote: > On 28/05/2020 12:15, Xiaoguang Wang wrote: >> If requests can be submitted and completed inline, we don't need to >> initialize whole io_wq_work in io_init_req(), which is an expensive >> operation, add a new 'REQ_F_WORK_INITIALIZED' to control whether >> io_wq_work is initialized. > > It looks nicer. Especially if you'd add a helper as Jens supposed. > > The other thing, even though I hate treating a part of the fields differently > from others, I don't like ->creds tossing either. > > Did you consider trying using only ->work.creds without adding req->creds? like > in the untested incremental below. init_io_work() there is misleading, should be > somehow played around better. I had that thought too when reading the patchset, would be nice _not_ to have to add a new creds field. -- Jens Axboe