On Mon, 2014-09-15 at 18:22 -0700, Andy Lutomirski wrote: > On Mon, Sep 15, 2014 at 2:19 PM, Toshi Kani <toshi.kani@xxxxxx> wrote: > > On Wed, 2014-09-10 at 15:34 -0600, Toshi Kani wrote: > >> On Wed, 2014-09-10 at 13:29 -0700, Andy Lutomirski wrote: > >> > On Wed, Sep 10, 2014 at 1:12 PM, Toshi Kani <toshi.kani@xxxxxx> wrote: > >> > > On Wed, 2014-09-10 at 11:30 -0700, Andy Lutomirski wrote: > >> > >> On Wed, Sep 10, 2014 at 9:51 AM, Toshi Kani <toshi.kani@xxxxxx> wrote: > >> > >> > +Drivers may map the entire NV-DIMM range with ioremap_cache and then change > >> > >> > +a specific range to wt with set_memory_wt. > >> > >> > >> > >> That's mighty specific :) > >> > > > >> > > How about below? > >> > > > >> > > Drivers may use set_memory_wt to set WT type for cached reserve ranges. > >> > > >> > Do they have to be cached? > >> > >> Yes, set_memory_xyz only supports WB->type->WB transition. > >> > >> > How about: > >> > > >> > Drivers may call set_memory_wt on ioremapped ranges. In this case, > >> > there is no need to change the memory type back before calling > >> > iounmap. > >> > > >> > (Or only on cached ioremapped ranges if that is, in fact, the case.) > >> > >> Sounds good. Yes, I will use cashed ioremapped ranges. > > > > Well, testing "no need to change the memory type back before calling > > iounmap" turns out to be a good test case. I realized that > > set_memory_xyz only works properly for RAM. There are two problems for > > using this interface for ioremapped ranges. > > > > 1) set_memory_xyz calls reserve_memtype() with __pa(addr). However, > > __pa() translates the addr into a fake physical address when it is an > > ioremapped address. > > > > 2) reserve_memtype() does not work for set_memory_xyz. For RAM, the WB > > state is managed untracked. Hence, WB->new->WB is not considered as a > > conflict. For ioremapped ranges, WB is tracked in the same way as other > > cache types. Hence, WB->new is considered as a conflict. > > > > In my previous testing, 2) was undetected since 1) led using a fake > > physical address which was not tracked for WB. This made ioremapped > > ranges worked just like RAM. :-( > > > > Anyway, 1) can be fixed by using slow_virt_to_phys() instead of __pa(). > > set_memory_xyz is already slow, but this makes it even slower, though. > > > > For 2), WB has to be continuously tracked in order to detect aliasing, > > ex. ioremap_cache and ioremap to a same address. So, I think > > reserve_memtype() needs the following changes: > > - Add a new arg to see if an operation is to create a new mapping or to > > change cache attribute. > > - Track overlapping maps so that cache type change to an overlapping > > range can be detected and failed. > > > > This level of changes requires a separate set of patches if we pursue to > > support ioremapped ranges. So, I am considering to take one of the two > > options below. > > > > A) Drop the patch for set_memory_wt. > > > > B) Keep the patch for set_memory_wt, but document that it fails with > > -EINVAL and its use is for RAM only. > > > > I vote A. I see no great reason to add code that can't be used. Once > someone needs this ability, they can add it :) Agreed. I will drop the patch for now. Since _PGMT_WB does not seem to be used for tracking WB, we might be able to use this bit for WT. But I need to look at the code more carefully for sure. > It's too bad that ioremap is called ioremap and not iomap. Otherwise > the natural solution would be to add a different function call > ioremap_wt that's like set_memory_wt but for ioremap ranges. Calling > it ioreremap_wt sounds kind of disgusting :) :) Thanks, -Toshi -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>