The vmap allocator is used to, among other things, allocate per-cpu
vmap blocks, where each vmap block is naturally aligned to its own
size. Obviously, leaving a guard page after each vmap area forbids
packing vmap blocks efficiently and can make the kernel run out of
possible vmap blocks long before overall vmap space is exhausted.
The new interface to map a user-supplied page array into linear
vmalloc space (vm_map_ram) insists on allocating from a vmap block
(instead of falling back to a custom area) when the area size is below
a certain threshold. With heavy users of this interface (e.g. XFS)
and limited vmalloc space on 32-bit, vmap block exhaustion is a real
problem.
Remove the guard page from the core vmap allocator. vmalloc and the
old vmap interface enforce a guard page on their own at a higher
level.
Signed-off-by: Johannes Weiner <[email protected]>
Cc: Nick Piggin <[email protected]>
Cc: Dave Chinner <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: Hugh Dickins <[email protected]>
Cc: Christoph Hellwig <[email protected]>
---
mm/vmalloc.c | 6 +++---
1 files changed, 3 insertions(+), 3 deletions(-)
Note that without this patch, we had accidental guard pages after
those vm_map_ram areas that happened to be at the end of a vmap block,
but not between every area. This patch removes this accidental guard
page only.
If we want guard pages after every vm_map_ram area, this should be
done separately. And just like with vmalloc and the old interface on
a different level, not in the core allocator.
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index cbd9f9f..5d8666b 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -307,7 +307,7 @@ nocache:
/* find starting point for our search */
if (free_vmap_cache) {
first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
- addr = ALIGN(first->va_end + PAGE_SIZE, align);
+ addr = ALIGN(first->va_end, align);
if (addr < vstart)
goto nocache;
if (addr + size - 1 < addr)
@@ -338,10 +338,10 @@ nocache:
}
/* from the starting point, walk areas until a suitable hole is found */
- while (addr + size >= first->va_start && addr + size <= vend) {
+ while (addr + size > first->va_start && addr + size <= vend) {
if (addr + cached_hole_size < first->va_start)
cached_hole_size = first->va_start - addr;
- addr = ALIGN(first->va_end + PAGE_SIZE, align);
+ addr = ALIGN(first->va_end, align);
if (addr + size - 1 < addr)
goto overflow;
--
1.7.4
On Thu, Apr 14, 2011 at 05:14:41PM -0400, Johannes Weiner wrote:
> The vmap allocator is used to, among other things, allocate per-cpu
> vmap blocks, where each vmap block is naturally aligned to its own
> size. Obviously, leaving a guard page after each vmap area forbids
> packing vmap blocks efficiently and can make the kernel run out of
> possible vmap blocks long before overall vmap space is exhausted.
>
> The new interface to map a user-supplied page array into linear
> vmalloc space (vm_map_ram) insists on allocating from a vmap block
> (instead of falling back to a custom area) when the area size is below
> a certain threshold. With heavy users of this interface (e.g. XFS)
> and limited vmalloc space on 32-bit, vmap block exhaustion is a real
> problem.
>
> Remove the guard page from the core vmap allocator. vmalloc and the
> old vmap interface enforce a guard page on their own at a higher
> level.
>
> Signed-off-by: Johannes Weiner <[email protected]>
> Cc: Nick Piggin <[email protected]>
> Cc: Dave Chinner <[email protected]>
> Cc: Mel Gorman <[email protected]>
> Cc: Hugh Dickins <[email protected]>
> Cc: Christoph Hellwig <[email protected]>
If necessary, the guard page could be reintroduced as a debugging-only
option (CONFIG_DEBUG_PAGEALLOC?). Otherwise it seems reasonable.
Acked-by: Mel Gorman <[email protected]>
--
Mel Gorman
SUSE Labs