On Fri, Aug 11, 2023 at 10:34:52AM +0200, Richard Weinberger wrote: > ----- Ursprüngliche Mail ----- > > Von: "Christoph Hellwig" <hch@xxxxxxxxxxxxx> > > On Thu, Aug 10, 2023 at 09:54:46PM +0200, Richard Weinberger wrote: > >> > But why add the bounce buffering first if you can avoid it from the > >> > very beginning by just using kmap_local instead of adding a new > >> > caller for the deprecate kmap_atomic? > >> > >> Because I want this fix also in all stable trees. kmap_local() is rather > >> new. When back porting patch 1/7, bounce buffers and kmap_atomic() > >> are needed anyway. > >> By doing this in patch 1/7 I avoid backport troubles and keep the > >> delta between upstream and stable trees minimal. > > > > Just use plain kmap for the historic backports. > > Hm, yes. For UBIblock kmap should be too expensive. ? kmap is a no-op for !highmem. And it will always be way cheaper than bounce buffering for the case where you are actually fed highmem pages.
