On 9/4/19 2:43 PM, Jens Axboe wrote: > On 9/4/19 2:10 PM, Jens Axboe wrote: >> On 9/3/19 5:49 AM, Jeff Moyer wrote: >>> Jens Axboe <axboe@xxxxxxxxx> writes: >>> >>>> On 8/29/19 10:38 AM, Jeff Moyer wrote: >>>>> Hi, >>>>> >>>>> sqthread_poll requires registered files in order to work. That's not >>>>> implemented in fio, and I don't see how one would go about it (but I'm >>>>> not all that familiar with fio's architecture). We basically need all >>>>> files to be opened before we can register them. It looks to me as >>>>> though the files are opened on demand, after the io engine is >>>>> initialized. >>>>> >>>>> Any suggestions? >>>> >>>> Leftover from previously, when that requirement wasn't there... Using >>>> it from fio would require a little bit of footwork, but it's not >>>> impossible. >>> >>> OK, if you can point me in the right direction, I'll take a stab at it. >> >> I think if we impose the restriction that the total number of files used >> must fit in a registered file set, it isn't too hard. Just have the >> engine open the files when it starts, and provide a new ->open_file() >> (and ->close_file()) in the engine ops. >> >> I'll take a stab at it. > > Here's a quick 10 min hack to get this going, if you want to run with > this, that'd be great. What is still needed: > > 1) Test it... I just ran a basic test to verify it does what I think it > should, that's all. > 2) Keep the standalone option, but make sqthread_poll imply this option > as well. > 3) Probably add a check whether o->nr_files fits in a set. Or maybe it's > fine not to, the system call will complain anyway. > 4) Ensure it works with things like loops= that restart things. > 5) Probably various things I didn't think of yet. I ran a few more tests and it seems fine, so I fixed up a bit of error handling and committed it. -- Jens Axboe
