Re: [RFC] virtual memmap for sparsemem [1/2] arch independent part

From: KAMEZAWA Hiroyuki <>
Date: 2006-10-20 11:18:57
On Thu, 19 Oct 2006 09:39:55 -0700 (PDT)
Christoph Lameter <> wrote:

> On Thu, 19 Oct 2006, KAMEZAWA Hiroyuki wrote:
> > For make patches simple, pfn_valid() uses sparsemem's logic. 
> Hmm... pfn_valid is much less costly if you use ia64's scheme.
> You can simply probe without having to walk tables.
Yes. but it seems to need per-arch implementation (in page fault handler).
like this (from ia64)
         * If fault is in region 5 and we are in the kernel, we may already
         * have the mmap_sem (pfn_valid macro is called during mmap). There
         * is no vma for region 5 addr's anyway, so skip getting the semaphore
         * and go directly to the exception handling code.

        if ((REGION_NUMBER(address) == 5) && !user_mode(regs))
                goto bad_area_no_up;

Maybe extra optimization patch can be discussed after this generic code is settled.

> > This patch maps sparsemem's *sparse* memmap into contiguous virtual address range
> > starting from virt_memmap_start.
> Could you make that a static address instead of a variable? Also we 
> already have vmem_map (ia64 specific) and mem_map. The logic here is the 
> same as FLATMEM. Why not use the definitions for FLATMEM?
It depends on how #ifdef looks. Here, I just wanted to throw this stuff into
SPARSEMEM subsystem.

> > * memmap is allocated per SECTION_SIZE, so there will be some of RESERVED pages.
> > * no holes in MAX_ORDER range. so HOLE_IN_ZONE=n here.
> Good. Had a patch here to do the same but I do not have time to get to 
> it. Surely wish that this will become the default config and that we can 
> get rid of at least some of the memory models.
> > +extern struct page *virt_memmap_start;
> extern struct page[] would be better performance wise. Use the definitions 
> for FLATMEM?
Okay. will make it as array. or some constant value.

> > +		if (pte_none(*pte))
> > +			set_pte(pte, pfn_pte(__pa(map) >> PAGE_SHIFT, PAGE_KERNEL));
> Would it be possible to add support for larger page sizes? On x86_64 we 
> probably would like to use 2MB pages and it may be good to have 
> configurable page size on ia64.
> The virtual memmap has the potential of becoming the default for x86_64 
> and many other platforms that already map memory. There is no performance 
> difference between FLATMEM and this virtual memmap approach if there are 
> already mappings in play.
We can add per-arch patches afterwards.


To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
the body of a message to
More majordomo info at
Received on Fri Oct 20 11:15:06 2006

This archive was generated by hypermail 2.1.8 : 2006-10-20 11:15:17 EST