Add alloc_pages_exact() and free_pages_exact()
Subsystems (3)
PATCH 1/1 Add alloc_pages_exact() and free_pages_exact()
alloc_pages_exact() is similar to alloc_pages(), except that it allocates the minimum number of pages to fulfill the request. This is useful if you want to allocate a very large buffer that is slightly larger than an even power-of-two number of pages. In that case, alloc_pages() will waste a lot of memory.
---
I have a video driver that wants to allocate a 5MB buffer. alloc_pages() will waste 3MB of physically-contiguous memory. Therefore, I would like to see alloc_pages_exact() added to 2.6.27.
Please note that I am not a Linux VM expert. I wrote these functions based on guidance from Andi Kleen. I have no familiarity with NUMA, so I don't know how to handle that. Any and all suggestions are welcome.
include/linux/gfp.h | 3 ++ mm/page_alloc.c | 53 +++++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 56 insertions(+), 0 deletions(-)
include/linux/gfp.h +3 -0
@@ -215,6 +215,9 @@ extern struct page *alloc_page_vma(gfp_t gfp_mask,extern unsigned long __get_free_pages (gfp_t gfp_mask,unsigned int order);extern unsigned long get_zeroed_page (gfp_t gfp_mask); +void *alloc_pages_exact (size_t size,gfp_t gfp_mask); +void free_pages_exact (void *virt,size_t size); +#define __get_free_page (gfp_mask) \ __get_free_pages((gfp_mask),0)
mm/page_alloc.c +53 -0
59 lines not on this page · raw message
Signature · 1 line
1.5.5
Quoted 9 lines · Timur Tabi
mm/page_alloc.c +9
-0+void *alloc_pages_exact (size_t size,gfp_t gfp_mask) +{ +unsigned int order =get_order (size); +unsigned long addr; + + addr =__get_free_pages (gfp_mask, order); +if (addr) { +unsigned long alloc_end = addr + (PAGE_SIZE << order); +unsigned long used = addr +PAGE_ALIGN (size);
mm/page_alloc.c void free_pages(unsigned long addr, unsigned int order) · line 1745
Quoted 10 lines
mm/page_alloc.c +10
-0+ +split_page (virt_to_page (addr), order); +while (used < alloc_end) { +free_page (used); + used +=PAGE_SIZE ; + } + } + +return (void *)addr; +}
Hi Timur,
This looks like a really good idea. It looks pretty good to me, no functional problems. My brain had a really hard time parsing that code for some reason, though. Could just be a lack of coffee.
I think the thing that confused me was trying to figure out if 'alloc_end' was the end of what we *did* allocate from __get_free_pages() or if it was the *goal* allocation end.
'used' also seemed like a slightly strange variable name because it points to the memory which is about to be freed and ends up *unused*.
I'll offer this up just in case you like it better. For me, it is easier to parse, and should do the exact same thing. I also think it's slightly nicer to do the arithmetic on 'struct page *' rather than vaddrs in 'unsigned long'. It is _slightly_ cheaper not having to do a virt_to_page() on each free_page() call. The same would go for the free side as well.
All of the 'struct page *' arithmetic is OK since it is all done inside one MAX_ORDER area.
void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
{
unsigned int order = get_order(size);
void *alloc;
struct page *surplus_start;
struct page *surplus_end;
struct page *page;
size = PAGE_ALIGN(size);
alloc = (void *)__get_free_pages(gfp_mask, order);
if (!alloc)
return NULL;
/* Turn the big allocation into a bunch of single pages */
split_page(virt_to_page(alloc), order);
surplus_start = virt_to_page(alloc + size);
surplus_end = surplus_start + (1 << order);
for (page = surplus_start; page < surplus_end; page++)
__free_page(page);
return alloc;
}-- Dave
Quoted 2 lines · Dave Hansen
This looks like a really good idea. It looks pretty good to me, no functional problems.
Thanks.
Quoted 6 lines
I think the thing that confused me was trying to figure out if '
alloc_end' was the end of what we *did* allocate from__get_free_pages()or if it was the *goal* allocation end.'used' also seemed like a slightly strange variable name because it points to the memory which is about to be freed and ends up *unused*.
This function is taken almost verbatim from alloc_large_system_hash(). I figured if the terminology was good for that function, it's good for mine.
Quoted 6 lines
I'll offer this up just in case you like it better. For me, it is easier to parse, and should do the exact same thing. I also think it's slightly nicer to do the arithmetic on 'struct page *' rather than vaddrs in 'unsigned long'. It is _slightly_ cheaper not having to do a
virt_to_page()on eachfree_page()call. The same would go for the free side as well.
It does seem to be an improvement, although AKPM just accepted this patch. I'd hate to bother him with a replacement patch for something so minor.
Signature · 2 lines
Timur Tabi Linux kernel developer at Freescale