Using __get_free_pages() has some drawbacks :
1) Not NUMA aware
2) 2^X page granularity :
On arches with big PAGE_SIZE, we waste some ram for each cpu.
(We currently use 1024 pointers, that is at most 8192 bytes, but
PAGE_SIZE can be 65536 for example : With say 64 possible cpus,
thats about 56*64 Kbytes that are wasted)
Using kmalloc_node() can help to solve these two points.
Signed-off-by: Eric Dumazet <[EMAIL PROTECTED]>
net/core/flow.c | 15 +++++----------
1 files changed, 5 insertions(+), 10 deletions(-)
diff --git a/net/core/flow.c b/net/core/flow.c
index a618f89..bcbbe2d 100644
--- a/net/core/flow.c
+++ b/net/core/flow.c
@@ -311,18 +311,13 @@ void flow_cache_flush(void)
static void __devinit flow_cache_cpu_prepare(int cpu)
{
struct tasklet_struct *tasklet;
- unsigned long order;
- for (order = 0;
- (PAGE_SIZE << order) <
- (sizeof(struct flow_cache_entry *)*flow_hash_size);
- order++)
- /* NOTHING */;
-
- flow_table(cpu) = (struct flow_cache_entry **)
- __get_free_pages(GFP_KERNEL|__GFP_ZERO, order);
+ flow_table(cpu) = kmalloc_node(
+ sizeof(struct flow_cache_entry *) * flow_hash_size,
+ GFP_KERNEL | __GFP_ZERO,
+ cpu_to_node(cpu));
if (!flow_table(cpu))
- panic("NET: failed to allocate flow cache order %lu\n", order);
+ panic("NET: failed to allocate flow cache for cpu %d\n", cpu);
flow_hash_rnd_recalc(cpu) = 1;
flow_count(cpu) = 0;
