1 files changed, 103 insertions, 5 deletions

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index bbcfb1c7f59e..1d9719e72f9d 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -134,8 +134,66 @@ static void skb_under_panic(struct sk_buff *skb, unsigned int sz, void *addr)
 #define NAPI_SKB_CACHE_BULK	16
 #define NAPI_SKB_CACHE_HALF	(NAPI_SKB_CACHE_SIZE / 2)
 
+#if PAGE_SIZE == SZ_4K
+
+#define NAPI_HAS_SMALL_PAGE_FRAG	1
+#define NAPI_SMALL_PAGE_PFMEMALLOC(nc)	((nc).pfmemalloc)
+
+/* specialized page frag allocator using a single order 0 page
+ * and slicing it into 1K sized fragment. Constrained to systems
+ * with a very limited amount of 1K fragments fitting a single
+ * page - to avoid excessive truesize underestimation
+ */
+
+struct page_frag_1k {
+	void *va;
+	u16 offset;
+	bool pfmemalloc;
+};
+
+static void *page_frag_alloc_1k(struct page_frag_1k *nc, gfp_t gfp)
+{
+	struct page *page;
+	int offset;
+
+	offset = nc->offset - SZ_1K;
+	if (likely(offset >= 0))
+		goto use_frag;
+
+	page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
+	if (!page)
+		return NULL;
+
+	nc->va = page_address(page);
+	nc->pfmemalloc = page_is_pfmemalloc(page);
+	offset = PAGE_SIZE - SZ_1K;
+	page_ref_add(page, offset / SZ_1K);
+
+use_frag:
+	nc->offset = offset;
+	return nc->va + offset;
+}
+#else
+
+/* the small page is actually unused in this build; add dummy helpers
+ * to please the compiler and avoid later preprocessor's conditionals
+ */
+#define NAPI_HAS_SMALL_PAGE_FRAG	0
+#define NAPI_SMALL_PAGE_PFMEMALLOC(nc)	false
+
+struct page_frag_1k {
+};
+
+static void *page_frag_alloc_1k(struct page_frag_1k *nc, gfp_t gfp_mask)
+{
+	return NULL;
+}
+
+#endif
+
 struct napi_alloc_cache {
 	struct page_frag_cache page;
+	struct page_frag_1k page_small;
 	unsigned int skb_count;
 	void *skb_cache[NAPI_SKB_CACHE_SIZE];
 };
@@ -143,6 +201,23 @@ struct napi_alloc_cache {
 static DEFINE_PER_CPU(struct page_frag_cache, netdev_alloc_cache);
 static DEFINE_PER_CPU(struct napi_alloc_cache, napi_alloc_cache);
 
+/* Double check that napi_get_frags() allocates skbs with
+ * skb->head being backed by slab, not a page fragment.
+ * This is to make sure bug fixed in 3226b158e67c
+ * ("net: avoid 32 x truesize under-estimation for tiny skbs")
+ * does not accidentally come back.
+ */
+void napi_get_frags_check(struct napi_struct *napi)
+{
+	struct sk_buff *skb;
+
+	local_bh_disable();
+	skb = napi_get_frags(napi);
+	WARN_ON_ONCE(!NAPI_HAS_SMALL_PAGE_FRAG && skb && skb->head_frag);
+	napi_free_frags(napi);
+	local_bh_enable();
+}
+
 void *__napi_alloc_frag_align(unsigned int fragsz, unsigned int align_mask)
 {
 	struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
@@ -561,6 +636,7 @@ struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
 {
 	struct napi_alloc_cache *nc;
 	struct sk_buff *skb;
+	bool pfmemalloc;
 	void *data;
 
 	DEBUG_NET_WARN_ON_ONCE(!in_softirq());
@@ -568,8 +644,10 @@ struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
 
 	/* If requested length is either too small or too big,
 	 * we use kmalloc() for skb->head allocation.
+	 * When the small frag allocator is available, prefer it over kmalloc
+	 * for small fragments
 	 */
-	if (len <= SKB_WITH_OVERHEAD(1024) ||
+	if ((!NAPI_HAS_SMALL_PAGE_FRAG && len <= SKB_WITH_OVERHEAD(1024)) ||
 	    len > SKB_WITH_OVERHEAD(PAGE_SIZE) ||
 	    (gfp_mask & (__GFP_DIRECT_RECLAIM | GFP_DMA))) {
 		skb = __alloc_skb(len, gfp_mask, SKB_ALLOC_RX | SKB_ALLOC_NAPI,
@@ -580,13 +658,33 @@ struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
 	}
 
 	nc = this_cpu_ptr(&napi_alloc_cache);
-	len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-	len = SKB_DATA_ALIGN(len);
 
 	if (sk_memalloc_socks())
 		gfp_mask |= __GFP_MEMALLOC;
 
-	data = page_frag_alloc(&nc->page, len, gfp_mask);
+	if (NAPI_HAS_SMALL_PAGE_FRAG && len <= SKB_WITH_OVERHEAD(1024)) {
+		/* we are artificially inflating the allocation size, but
+		 * that is not as bad as it may look like, as:
+		 * - 'len' less than GRO_MAX_HEAD makes little sense
+		 * - On most systems, larger 'len' values lead to fragment
+		 *   size above 512 bytes
+		 * - kmalloc would use the kmalloc-1k slab for such values
+		 * - Builds with smaller GRO_MAX_HEAD will very likely do
+		 *   little networking, as that implies no WiFi and no
+		 *   tunnels support, and 32 bits arches.
+		 */
+		len = SZ_1K;
+
+		data = page_frag_alloc_1k(&nc->page_small, gfp_mask);
+		pfmemalloc = NAPI_SMALL_PAGE_PFMEMALLOC(nc->page_small);
+	} else {
+		len += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+		len = SKB_DATA_ALIGN(len);
+
+		data = page_frag_alloc(&nc->page, len, gfp_mask);
+		pfmemalloc = nc->page.pfmemalloc;
+	}
+
 	if (unlikely(!data))
 		return NULL;
 
@@ -596,7 +694,7 @@ struct sk_buff *__napi_alloc_skb(struct napi_struct *napi, unsigned int len,
 		return NULL;
 	}
 
-	if (nc->page.pfmemalloc)
+	if (pfmemalloc)
 		skb->pfmemalloc = 1;
 	skb->head_frag = 1;