On Wednesday, 14 March 2007 04:17, Nick Piggin wrote: > Rafael J. Wysocki wrote: > > On Tuesday, 13 March 2007 11:31, Nick Piggin wrote: > > > >>Rafael J. Wysocki wrote: > >> > >>>On Tuesday, 13 March 2007 10:23, Nick Piggin wrote: > >>> > >> > >>>>I wouldn't say that. You're creating an interface here that is going to be > >>>>used outside swsusp. Users of that interface may not need locking now, but > >>>>that could cause problems down the line. > >>> > >>> > >>>I think we can add the locking when it's necessary. For now, IMHO, it could be > >>>confusing to someone who doesn't know the locking is not needed. > >> > >>I don't know why it would confuse them. We just define the API to > >>guarantee the correct locking, and that means the locking _is_ needed. > > > > > > Even if there are no users that actually need the locking and probably never > > will be? > > Probably is the keyword. > > Why would you *not* make this a sane API? Surely performance isn't the > reason? Nor complexity. > > > For now, register_nosave_region() is to be called by architecture > > initialization code _only_ and there's no reason whatsoever why any > > architecture would need to call it concurrently from many places. > > > > > >>You don't have to care what the callers are doing. That's the beauty > >>of a sane API. > > > > > > Well, I don't think adding unneded infrastructure is a good thing. > > > But defining good APIs is a very good thing. And with my good API, the > lock is not unneeded. > > >>>>But that's because you even use mark_nosave_pages in your implementation. > >>>>Mine uses the nosave regions directly. > >>> > >>> > >>>Well, I think we need two bits per page anyway, to mark free pages and > >>>pages allocated by swsusp, so using the nosave regions directly won't save us > >>>much. > >> > >>Well I think it is a cleaner though. > > > > > > This is a matter of opinion, too ... > > > Well, as I'm not volunteering to maintain swsusp, if your opinion is that > your way is cleaner, I can't argue ;) So long as it stops wasting those page > flags then I'm happy. > > However the register_nosave API really should use locking, I think. There > is absolutely no downside, AFAIKS. Hm, I can add the lock, but currently register_nosave_region() calls alloc_bootmem_low() directly, which means it won't even work outside the early initialization code which is single-threaded. For this reason I think that the locking can be added later, when register_nosave_region() is extended so that it can be called from other contexts too. Greetings, Rafael _______________________________________________ linux-pm mailing list linux-pm@xxxxxxxxxxxxxx https://lists.osdl.org/mailman/listinfo/linux-pm
