On Mon, Mar 11, 2019 at 09:06:18PM +0000, Al Viro wrote: > On Mon, Mar 11, 2019 at 08:58:31PM +0100, Christoph Hellwig wrote: > > Where do we put the second iocb reference in case we return from > > vfs_poll without ever being woken? > > Depends. If mask is non-zero (i.e. vfs_poll() has returned something > we care about) and it has never been woken, we steal it and drop the > reference ourselves. If it is zero and we see that ->poll() has tried > to put it on two queues, we steal it (again, assuming it's not on > waitqueue and _can_ be stolen) and return -EINVAL. In that case > __io_submit_one() (or, by the end of the series, io_submit_one()) > will call iocb_destroy(). And in the normal waiting case (nothing > interesting reported and no errors) it will end up on the list of > cancellables. Then it either will get completed by later wakeup, which > will drop the reference, or it will get eventually cancelled, which will > hit the same aio_poll_complete_work() and drop the reference... Ok, seems like the logic is sane. I was missing how the actual mask logic worked in aio_poll(). > > Also it seems like the complete code would still benefit from a little > > helper, something like: > > Umm... Not sure I like the name (something like aio_poll_done() seems > to be better), but other than that - no problem. I don't care about the name. Feel free to change it to whatever suits you.
