On 05/25/2018 05:51 PM, Christopher Lameter wrote:

On Thu, 24 May 2018, Vlastimil Babka wrote:

quoted

diff --git a/include/linux/slab.h b/include/linux/slab.h
index 9ebe659bd4a5..5bff0571b360 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h

@@ -296,11 +296,16 @@ static inline void __check_heap_object(const void *ptr, unsigned long n,
                                (KMALLOC_MIN_SIZE) : 16)

 #ifndef CONFIG_SLOB
-extern struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
+extern struct kmem_cache *kmalloc_caches[2][KMALLOC_SHIFT_HIGH + 1];
 #ifdef CONFIG_ZONE_DMA
 extern struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 #endif

In the existing code we used a different array name for the DMA caches.
This is a similar situation.

I would suggest to use

kmalloc_reclaimable_caches[]

or make it consistent by folding the DMA caches into the array too (but
then note the issues below).

quoted

@@ -536,12 +541,13 @@ static __always_inline void *kmalloc(size_t size, gfp_t flags)
 #ifndef CONFIG_SLOB
 		if (!(flags & GFP_DMA)) {
 			unsigned int index = kmalloc_index(size);
+			unsigned int recl = kmalloc_reclaimable(flags);

This is a hotpath reserved for regular allocations. The reclaimable slabs
need to be handled like the DMA slabs.  So check for GFP_DMA plus the
reclaimable flags.

Yeah I thought that by doing reclaimable via array index manipulation
and not a branch, there would be no noticeable overhead. And GFP_DMA
should go away eventually. I will see if I can convert GFP_DMA to
another index, and completely remove the branch quoted above.

quoted

@@ -588,12 +594,13 @@ static __always_inline void *kmalloc_node(size_t size, gfp_t flags, int node)
 	if (__builtin_constant_p(size) &&
 		size <= KMALLOC_MAX_CACHE_SIZE && !(flags & GFP_DMA)) {
 		unsigned int i = kmalloc_index(size);
+		unsigned int recl = kmalloc_reclaimable(flags);

Same situation here and additional times below.