Am 22.09.2017 um 04:30 hat Fam Zheng geschrieben: > On Thu, 09/21 18:39, Manos Pitsidianakis wrote: > > On Thu, Sep 21, 2017 at 09:29:43PM +0800, Fam Zheng wrote: > > > On Thu, 09/21 16:17, Manos Pitsidianakis wrote: > > It might imply to someone that there's an assert(drv->bdrv_co_drain_begin && > > drv->bdrv_co_drain_end) somewhere unless you state they don't have to be > > implemented at the same time. How about we be completely explicit: > > > > bdrv_co_drain_begin is called if implemented in the beggining of a > > drain operation to drain and stop any internal sources of requests in > > the driver. > > bdrv_co_drain_end is called if implemented at the end of the drain. > > > > They should be used by the driver to e.g. manage scheduled I/O > > requests, or toggle an internal state. After the end of the drain new > > requests will continue normally. > > > > I hope this is easier for a reader to understand! > > I don't like the inconsistent semantics of when the drained section > ends, if we allow drivers to implement bdrv_co_drain_begin but omit > bdrv_co_drained_end. Currently the point where the section ends is, > as said in the comment, when next I/O callback is invoked. Now we are > adding the explicit ".bdrv_co_drain_end" into the fomular, if we still > keep the previous convention, the interface contract is just mixed of > two things for no good reason. I don't think it's technically > necessary.
We don't keep the convention with the next I/O callback. We just allow drivers to omit an empty implementation of either callback, which seems to be a very sensible default to me. > Let's just add the assert: > > assert(!!drv->bdrv_co_drain_begin == !!bdrv_co_drain_end); > > in bdrv_drain_invoke, add a comment here I'm not in favour of this, but if we do want to have it, wouldn't bdrv_register() be a much better place for the assertion? > then add an empty .bdrv_co_drain_end in qed. If you need empty functions anywhere, it's a sign that we have a bad default behaviour. Kevin