Messages in this thread | | | From | Arnd Bergmann <> | Subject | Re: [PATCH] ARM: use default ioremap alignment for SMP or LPAE | Date | Thu, 22 Jan 2015 12:03 +0100 |
| |
On Thursday 22 January 2015 10:04:41 Catalin Marinas wrote: > On Thu, Jan 22, 2015 at 07:17:55AM +0000, Sergey Dyasly wrote: > > 16MB alignment for ioremap mappings was added by commit a069c896d0d6 ("[ARM] > > 3705/1: add supersection support to ioremap()") in order to support supersection > > mappings. But __arm_ioremap_pfn_caller uses section and supersection mappings > > only in !SMP && !LPAE case. There is no need for such big alignment if either > > SMP or LPAE is enabled. > [...] > > diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h > > index 184def0..c3ef139 100644 > > --- a/arch/arm/include/asm/memory.h > > +++ b/arch/arm/include/asm/memory.h > > @@ -78,10 +78,12 @@ > > */ > > #define XIP_VIRT_ADDR(physaddr) (MODULES_VADDR + ((physaddr) & 0x000fffff)) > > > > +#if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE) > > /* > > * Allow 16MB-aligned ioremap pages > > */ > > #define IOREMAP_MAX_ORDER 24 > > +#endif > > Actually, I think we could make this depend only on CONFIG_IO_36. That's > the only scenario where we get the supersections matter, and maybe make > CONFIG_IO_36 dependent on !SMP or !ARM_LPAE.
Good point, I assumed this was just a performance optimization, but it is in fact required for dynamic high mappings on XSC3.
> My assumption is that we > don't support single zImage with CPU_XSC3 enabled (but I haven't > followed the latest developments here).
I have said in the past that I do not expect any of the xscale or strongarm based platforms to be used with a single zImage kernel. This has changed slightly given the work that Robert Jarzmik and others are doing on mach-pxa with DT conversion. I still don't think it's likely that they will move on to multiplatform, but I no longer think it's impossible.
Fortunately, PXA also does not have any high physical address mappings that I can see. The only platform we have that has these (see 'git grep ioremap_pfn') is mach-iop13xx, and I am still seeing this in the 'never going to be multiplatform' category.
Unrelated to this question however is whether we want to keep supersection mappings as a performance optimization to save TLBs. It seems useful to me, but not critical.
Arnd
| |