On Wed, May 25, 2022 at 11:34:16AM +0200, Sebastian Andrzej Siewior wrote: > On 2022-04-29 08:59:26 [-0700], Ira Weiny wrote: > > I think some discussion needs to happen around this API. > > > > Highmem has little use. I don't think anyone disagrees with Linus there. > > (Although I think there are still a few users out there.) > > arm32 is still built and they have sometimes 1 - 2 GiB of memory. Yep :-) I was thinking of arm when I said this. > > > kmap may be a poor name for an API without the highmem functionality. But > > perhaps not. One could interpret it to mean simply getting the kernel mapping > > of the page rather than creating one. After all that is what 64bit has done > > all along. > > > > This interpretation helps when you consider features which attempt to layer the > > direct map with additional protections like PKS.[1] Those protections mean > > that a simple page_address() is insufficient to access the direct map. > > > > As far as calling kmap() and kmap_atomic() deprecated I'm ok with that if the > > community is. > > > > The current kmap() call sites need work and Fabio's work on auditing them is > > extremely helpful. That said, if we officially deprecate kmap_atomic() then > > those sites could be added to the list for rework. > > Maybe I oversee something obvious but there is no problem with removing > kmap_atomic*() and keeping only kmap_local*() around, is there? No there is not. But some kmap_atomic() sites may have to open code the preempt_disable() while others may not. I have not done a full audit of the kmap_atomic() sites but I suspect most don't really need the preempt_disable() but many may need to. I just don't know. Regardless marking it deprecated can stop the growth of kmap_atomic() calls. > I never intended to deprecated kmap(), only kmap_atomic*() in favour of > kmap_local*(). Ok. But I do want to see kmap() use removed. The PKS code can't work with kmap() and in general we are seeing more and more restrictions on the direct map which may or may not be compatible with kmap(). What I presented at LSFmm was to turn the kmap* interfaces into a more generic 'get me a temp kernel address' interface instead of a highmem interface. Any user who needs a long term address will need something other than kmap(). To that end there was some discussion on making vmap() more efficient or other alternatives. First we need to focus on reducing the kmap() call sites. This documentation change, making kmap() deprecated, will help ensure the kernel does not grow more of them. Ira > > > Ira > > Sebastian
