bpf: cpumap do bulk allocation of SKBs
authorJesper Dangaard Brouer <brouer@redhat.com>
Fri, 12 Apr 2019 15:07:43 +0000 (17:07 +0200)
committerAlexei Starovoitov <ast@kernel.org>
Thu, 18 Apr 2019 02:09:25 +0000 (19:09 -0700)
As cpumap now batch consume xdp_frame's from the ptr_ring, it knows how many
SKBs it need to allocate. Thus, lets bulk allocate these SKBs via
kmem_cache_alloc_bulk() API, and use the previously introduced function
build_skb_around().

Notice that the flag __GFP_ZERO asks the slab/slub allocator to clear the
memory for us. This does clear a larger area than needed, but my micro
benchmarks on Intel CPUs show that this is slightly faster due to being a
cacheline aligned area is cleared for the SKBs. (For SLUB allocator, there
is a future optimization potential, because SKBs will with high probability
originate from same page. If we can find/identify continuous memory areas
then the Intel CPU memset rep stos will have a real performance gain.)

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
kernel/bpf/cpumap.c

index 430103e182a04d780754d209f89ddc79ae717476..732d6ced3987059b9608b86305f00bd813f609d8 100644 (file)
@@ -160,12 +160,12 @@ static void cpu_map_kthread_stop(struct work_struct *work)
 }
 
 static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
-                                        struct xdp_frame *xdpf)
+                                        struct xdp_frame *xdpf,
+                                        struct sk_buff *skb)
 {
        unsigned int hard_start_headroom;
        unsigned int frame_size;
        void *pkt_data_start;
-       struct sk_buff *skb;
 
        /* Part of headroom was reserved to xdpf */
        hard_start_headroom = sizeof(struct xdp_frame) +  xdpf->headroom;
@@ -191,8 +191,8 @@ static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
                SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 
        pkt_data_start = xdpf->data - hard_start_headroom;
-       skb = build_skb(pkt_data_start, frame_size);
-       if (!skb)
+       skb = build_skb_around(skb, pkt_data_start, frame_size);
+       if (unlikely(!skb))
                return NULL;
 
        skb_reserve(skb, hard_start_headroom);
@@ -256,7 +256,9 @@ static int cpu_map_kthread_run(void *data)
        while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
                unsigned int drops = 0, sched = 0;
                void *frames[CPUMAP_BATCH];
-               int i, n;
+               void *skbs[CPUMAP_BATCH];
+               gfp_t gfp = __GFP_ZERO | GFP_ATOMIC;
+               int i, n, m;
 
                /* Release CPU reschedule checks */
                if (__ptr_ring_empty(rcpu->queue)) {
@@ -278,14 +280,20 @@ static int cpu_map_kthread_run(void *data)
                 * consume side valid as no-resize allowed of queue.
                 */
                n = ptr_ring_consume_batched(rcpu->queue, frames, CPUMAP_BATCH);
+               m = kmem_cache_alloc_bulk(skbuff_head_cache, gfp, n, skbs);
+               if (unlikely(m == 0)) {
+                       for (i = 0; i < n; i++)
+                               skbs[i] = NULL; /* effect: xdp_return_frame */
+                       drops = n;
+               }
 
                local_bh_disable();
                for (i = 0; i < n; i++) {
                        struct xdp_frame *xdpf = frames[i];
-                       struct sk_buff *skb;
+                       struct sk_buff *skb = skbs[i];
                        int ret;
 
-                       skb = cpu_map_build_skb(rcpu, xdpf);
+                       skb = cpu_map_build_skb(rcpu, xdpf, skb);
                        if (!skb) {
                                xdp_return_frame(xdpf);
                                continue;