]> git.baikalelectronics.ru Git - kernel.git/commitdiff
skbuff: introduce {,__}napi_build_skb() which reuses NAPI cache heads
authorAlexander Lobakin <alobakin@pm.me>
Sat, 13 Feb 2021 14:12:25 +0000 (14:12 +0000)
committerDavid S. Miller <davem@davemloft.net>
Sat, 13 Feb 2021 22:32:04 +0000 (14:32 -0800)
Instead of just bulk-flushing skbuff_heads queued up through
napi_consume_skb() or __kfree_skb_defer(), try to reuse them
on allocation path.
If the cache is empty on allocation, bulk-allocate the first
16 elements, which is more efficient than per-skb allocation.
If the cache is full on freeing, bulk-wipe the second half of
the cache (32 elements).
This also includes custom KASAN poisoning/unpoisoning to be
double sure there are no use-after-free cases.

To not change current behaviour, introduce a new function,
napi_build_skb(), to optionally use a new approach later
in drivers.

Note on selected bulk size, 16:
 - this equals to XDP_BULK_QUEUE_SIZE, DEV_MAP_BULK_SIZE
   and especially VETH_XDP_BATCH, which is also used to
   bulk-allocate skbuff_heads and was tested on powerful
   setups;
 - this also showed the best performance in the actual
   test series (from the array of {8, 16, 32}).

Suggested-by: Edward Cree <ecree.xilinx@gmail.com> # Divide on two halves
Suggested-by: Eric Dumazet <edumazet@google.com> # KASAN poisoning
Cc: Dmitry Vyukov <dvyukov@google.com> # Help with KASAN
Cc: Paolo Abeni <pabeni@redhat.com> # Reduced batch size
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/skbuff.h
net/core/skbuff.c

index 0e07072960983c800755dfc3e21880423b1548ae..906122eac82a6e2054be2c3c768c969929358aed 100644 (file)
@@ -1087,6 +1087,8 @@ struct sk_buff *build_skb(void *data, unsigned int frag_size);
 struct sk_buff *build_skb_around(struct sk_buff *skb,
                                 void *data, unsigned int frag_size);
 
+struct sk_buff *napi_build_skb(void *data, unsigned int frag_size);
+
 /**
  * alloc_skb - allocate a network buffer
  * @size: size to allocate
index 860a9d4f752f678fda3fa5a062337bfe99c19eb2..9e1a8ded4accb6d20f2b4373cee644aaaaeeb4e6 100644 (file)
@@ -120,6 +120,8 @@ static void skb_under_panic(struct sk_buff *skb, unsigned int sz, void *addr)
 }
 
 #define NAPI_SKB_CACHE_SIZE    64
+#define NAPI_SKB_CACHE_BULK    16
+#define NAPI_SKB_CACHE_HALF    (NAPI_SKB_CACHE_SIZE / 2)
 
 struct napi_alloc_cache {
        struct page_frag_cache page;
@@ -164,6 +166,25 @@ void *__netdev_alloc_frag_align(unsigned int fragsz, unsigned int align_mask)
 }
 EXPORT_SYMBOL(__netdev_alloc_frag_align);
 
+static struct sk_buff *napi_skb_cache_get(void)
+{
+       struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
+       struct sk_buff *skb;
+
+       if (unlikely(!nc->skb_count))
+               nc->skb_count = kmem_cache_alloc_bulk(skbuff_head_cache,
+                                                     GFP_ATOMIC,
+                                                     NAPI_SKB_CACHE_BULK,
+                                                     nc->skb_cache);
+       if (unlikely(!nc->skb_count))
+               return NULL;
+
+       skb = nc->skb_cache[--nc->skb_count];
+       kasan_unpoison_object_data(skbuff_head_cache, skb);
+
+       return skb;
+}
+
 /* Caller must provide SKB that is memset cleared */
 static void __build_skb_around(struct sk_buff *skb, void *data,
                               unsigned int frag_size)
@@ -265,6 +286,53 @@ struct sk_buff *build_skb_around(struct sk_buff *skb,
 }
 EXPORT_SYMBOL(build_skb_around);
 
+/**
+ * __napi_build_skb - build a network buffer
+ * @data: data buffer provided by caller
+ * @frag_size: size of data, or 0 if head was kmalloced
+ *
+ * Version of __build_skb() that uses NAPI percpu caches to obtain
+ * skbuff_head instead of inplace allocation.
+ *
+ * Returns a new &sk_buff on success, %NULL on allocation failure.
+ */
+static struct sk_buff *__napi_build_skb(void *data, unsigned int frag_size)
+{
+       struct sk_buff *skb;
+
+       skb = napi_skb_cache_get();
+       if (unlikely(!skb))
+               return NULL;
+
+       memset(skb, 0, offsetof(struct sk_buff, tail));
+       __build_skb_around(skb, data, frag_size);
+
+       return skb;
+}
+
+/**
+ * napi_build_skb - build a network buffer
+ * @data: data buffer provided by caller
+ * @frag_size: size of data, or 0 if head was kmalloced
+ *
+ * Version of __napi_build_skb() that takes care of skb->head_frag
+ * and skb->pfmemalloc when the data is a page or page fragment.
+ *
+ * Returns a new &sk_buff on success, %NULL on allocation failure.
+ */
+struct sk_buff *napi_build_skb(void *data, unsigned int frag_size)
+{
+       struct sk_buff *skb = __napi_build_skb(data, frag_size);
+
+       if (likely(skb) && frag_size) {
+               skb->head_frag = 1;
+               skb_propagate_pfmemalloc(virt_to_head_page(data), skb);
+       }
+
+       return skb;
+}
+EXPORT_SYMBOL(napi_build_skb);
+
 /*
  * kmalloc_reserve is a wrapper around kmalloc_node_track_caller that tells
  * the caller if emergency pfmemalloc reserves are being used. If it is and
@@ -838,31 +906,31 @@ void __consume_stateless_skb(struct sk_buff *skb)
        kfree_skbmem(skb);
 }
 
-static inline void _kfree_skb_defer(struct sk_buff *skb)
+static void napi_skb_cache_put(struct sk_buff *skb)
 {
        struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
+       u32 i;
 
        /* drop skb->head and call any destructors for packet */
        skb_release_all(skb);
 
-       /* record skb to CPU local list */
+       kasan_poison_object_data(skbuff_head_cache, skb);
        nc->skb_cache[nc->skb_count++] = skb;
 
-#ifdef CONFIG_SLUB
-       /* SLUB writes into objects when freeing */
-       prefetchw(skb);
-#endif
-
-       /* flush skb_cache if it is filled */
        if (unlikely(nc->skb_count == NAPI_SKB_CACHE_SIZE)) {
-               kmem_cache_free_bulk(skbuff_head_cache, NAPI_SKB_CACHE_SIZE,
-                                    nc->skb_cache);
-               nc->skb_count = 0;
+               for (i = NAPI_SKB_CACHE_HALF; i < NAPI_SKB_CACHE_SIZE; i++)
+                       kasan_unpoison_object_data(skbuff_head_cache,
+                                                  nc->skb_cache[i]);
+
+               kmem_cache_free_bulk(skbuff_head_cache, NAPI_SKB_CACHE_HALF,
+                                    nc->skb_cache + NAPI_SKB_CACHE_HALF);
+               nc->skb_count = NAPI_SKB_CACHE_HALF;
        }
 }
+
 void __kfree_skb_defer(struct sk_buff *skb)
 {
-       _kfree_skb_defer(skb);
+       napi_skb_cache_put(skb);
 }
 
 void napi_consume_skb(struct sk_buff *skb, int budget)
@@ -887,7 +955,7 @@ void napi_consume_skb(struct sk_buff *skb, int budget)
                return;
        }
 
-       _kfree_skb_defer(skb);
+       napi_skb_cache_put(skb);
 }
 EXPORT_SYMBOL(napi_consume_skb);