Question on virtual memory layout: lowmem with memory hole

[linux@arm.linux.org.uk (Russell King - ARM Linux)]

On Fri, Aug 26, 2011 at 12:06:20AM +0530, Pedanekar, Hemant wrote:
> Russell King - ARM Linux wrote on Thursday, August 25, 2011 3:53 PM:
> 
> > On Thu, Aug 25, 2011 at 09:35:07AM +0530, Pedanekar, Hemant wrote:
> >> E.g., on OMAP3 with mem=32M at 0x80000000 mem=8M at 0x87800000
> >> 
> >> (CASE 1)
> >>  Memory: 32MB 8MB = 40MB total
> >>  Memory: 28408k/28408k available, 12552k reserved, 0K highmem
> >>  Virtual kernel memory layout:
> >>      vector  : 0xffff0000 - 0xffff1000   (   4 kB)
> >>      fixmap  : 0xfff00000 - 0xfffe0000   ( 896 kB)
> >>      DMA     : 0xffc00000 - 0xffe00000   (   2 MB)
> >>      vmalloc : 0xc8800000 - 0xf8000000   ( 760 MB)
> >>      lowmem  : 0xc0000000 - 0xc8000000   ( 128 MB)
> >>      modules : 0xbf000000 - 0xc0000000   (  16 MB)
> >>        .text : 0xc0008000 - 0xc05ac2c8   (5777 kB)
> >>        .init : 0xc05ad000 - 0xc05f8780   ( 302 kB)
> >>        .data : 0xc05fa000 - 0xc06838d0   ( 551 kB)
> >>         .bss : 0xc06838f4 - 0xc0bd8a14   (5461 kB)
> >> 
> >> Is this expected? 88MB space between two 'mem's seems to have lost.
> > 
> > Yes.  The memory layout lines gives an overview of the virtual memory
> > address space _regions_. 
> > 
> > What it's saying is that the virtual addresses from 0xc0000000 - 0xc8000000
> > are used for lowmem.  That may not be fully populated, but that's what
> > the address range is reserved for.
> 
> So, larger the hole, more address space will be unusable - is that correct?

Correct.

> >> This also
> >> means vmalloc space is lower compared to when a single mem=40M is passed.
> > 
> > Huh.  Either your maths is wrong or...
> > 
> > Here's case 1:
> >>      vmalloc : 0xc8800000 - 0xf8000000   ( 760 MB) And case 2:
> >>      vmalloc : 0xc3000000 - 0xf8000000   ( 848 MB)
> > 
> > Looks to me like case 1, vmalloc space is _higher_ not _lower_.  That's
> > expected because you told the kernel it had more memory in case 1.
> 
> Sorry, my mistake - I actually meant "vmalloc space is _smaller_ compared to
> when a single mem=40M is passed" though the actual physical RAM available is
> same in both the cases.

Again, that's expected.

We require a 1:1 relationship between virtual and physical addresses for
efficiency - having non-linear translation means we'd need a lookup table,
and as these translations are heavily used, that will impact performance.