Hi Pete, On Thu, 13 Dec 2007 13:59:16 -0800, Pete Zaitcev <zaitcev@xxxxxxxxxx> wrote: > > > > - end_that_request_first(rq, uptodate, rq->hard_nr_sectors); > > > > - end_that_request_last(rq, uptodate); > > > > + if (__blk_end_request(rq, error, blk_rq_bytes(rq))) > > > > + BUG(); > > > > > > My understanding was, blk_end_request() is the same thing, only > > > takes the queue lock. But then, should I refactor ub so that it > > > calls __blk_end_request if request function ends with an error > > > and blk_end_request if the end-of-IO even is processed? > > > > I'm using __blk_end_request() here and I think it's sufficient, because: > > o end_that_request_last() must be called with the queue lock held > > o ub_end_rq() calls end_that_request_last() without taking > > the queue lock in itself. > > So the queue lock must have been taken outside ub_end_rq(). > > > > But, if ub is calling end_that_request_last() without the queue lock, > > it is a bug in the original code and we should use blk_end_request() > > to fix that. > > So, I have to rewrite ub to split the paths after all, right? > Let's do this then: I'll wait until your patch gets to Linus and > then update it with the split. The reason is, I need the whole > enchilada applied and I don't want to bother tracking iterations > and all the little segments (of which you already have 30). > Until then, ub will have a race by using your original small patch. No. Are you doubting that the current ub code has the problem, aren't you? My patch shouldn't introduce a NEW problem to ub. I have investigated all code paths which call ub_end_rq() in ub.c, and confirmed that ub_end_rq() is always called with the queue lock held. (sc->lock is registered as a queue lock.) So there is no such race in the current ub code. You don't need to rewrite ub. Thanks, Kiyoshi Ueda - To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
