[net-next PATCH 0/5] New bpf cpumap type for XDP_REDIRECT

[net-next PATCH 0/5] New bpf cpumap type for XDP_REDIRECT

[Jesper Dangaard Brouer]

Introducing a new way to redirect XDP frames.  Notice how no driver
changes are necessary given the design of XDP_REDIRECT.

This redirect map type is called 'cpumap', as it allows redirection
XDP frames to remote CPUs.  The remote CPU will do the SKB allocation
and start the network stack invocation on that CPU.

This is a scalability and isolation mechanism, that allow separating
the early driver network XDP layer, from the rest of the netstack, and
assigning dedicated CPUs for this stage.  The sysadm control/configure
the RX-CPU to NIC-RX queue (as usual) via procfs smp_affinity and how
many queues are configured via ethtool --set-channels.  Benchmarks
show that a single CPU can handle approx 11Mpps.  Thus, only assigning
two NIC RX-queues (and two CPUs) is sufficient for handling 10Gbit/s
wirespeed smallest packet 14.88Mpps.  Reducing the number of queues
have the advantage that more packets being "bulk" available per hard
interrupt[1].

[1] https://www.netdevconf.org/2.1/papers/BusyPollingNextGen.pdf

Use-cases:

1. End-host based pre-filtering for DDoS mitigation.  This is fast
   enough to allow software to see and filter all packets wirespeed.
   Thus, no packets getting silently dropped by hardware.

2. Given NIC HW unevenly distributes packets across RX queue, this
   mechanism can be used for redistribution load across CPUs.  This
   usually happens when HW is unaware of a new protocol.  This
   resembles RPS (Receive Packet Steering), just faster, but with more
   responsibility placed on the BPF program for correct steering.

3. Auto-scaling or power saving via only activating the appropriate
   number of remote CPUs for handling the current load.  The cpumap
   tracepoints can function as a feedback loop for this purpose.

Patchset based on net-next at:
 commit 14a0d032f4ec ("Merge branch 'mlxsw-pass-gact'")

---

Jesper Dangaard Brouer (5):
      bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP
      bpf: XDP_REDIRECT enable use of cpumap
      bpf: cpumap xdp_buff to skb conversion and allocation
      bpf: cpumap add tracepoints
      samples/bpf: add cpumap sample program xdp_redirect_cpu


 include/linux/bpf.h                 |    7 
 include/linux/bpf_types.h           |    1 
 include/trace/events/xdp.h          |   80 ++++
 include/uapi/linux/bpf.h            |    1 
 kernel/bpf/Makefile                 |    1 
 kernel/bpf/cpumap.c                 |  671 +++++++++++++++++++++++++++++++++++
 kernel/bpf/syscall.c                |    8 
 kernel/bpf/verifier.c               |    3 
 net/core/filter.c                   |   65 +++
 samples/bpf/Makefile                |    4 
 samples/bpf/xdp_redirect_cpu_kern.c |  640 +++++++++++++++++++++++++++++++++
 samples/bpf/xdp_redirect_cpu_user.c |  639 +++++++++++++++++++++++++++++++++
 tools/include/uapi/linux/bpf.h      |    1 
 13 files changed, 2109 insertions(+), 12 deletions(-)
 create mode 100644 kernel/bpf/cpumap.c
 create mode 100644 samples/bpf/xdp_redirect_cpu_kern.c
 create mode 100644 samples/bpf/xdp_redirect_cpu_user.c

[net-next PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Jesper Dangaard Brouer]

The 'cpumap' is primary used as a backend map for XDP BPF helper
call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.

This patch implement the main part of the map.  It is not connected to
the XDP redirect system yet, and no SKB allocation are done yet.

The main concern in this patch is to ensure the datapath can run
without any locking.  This adds complexity to the setup and tear-down
procedure, which assumptions are extra carefully documented in the
code comments.

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/linux/bpf_types.h      |    1 
 include/uapi/linux/bpf.h       |    1 
 kernel/bpf/Makefile            |    1 
 kernel/bpf/cpumap.c            |  547 ++++++++++++++++++++++++++++++++++++++++
 kernel/bpf/syscall.c           |    8 +
 tools/include/uapi/linux/bpf.h |    1 
 6 files changed, 558 insertions(+), 1 deletion(-)
 create mode 100644 kernel/bpf/cpumap.c

diff --git a/include/linux/bpf_types.h b/include/linux/bpf_types.h
index 6f1a567667b8..814c1081a4a9 100644
--- a/include/linux/bpf_types.h
+++ b/include/linux/bpf_types.h
@@ -41,4 +41,5 @@ BPF_MAP_TYPE(BPF_MAP_TYPE_DEVMAP, dev_map_ops)
 #ifdef CONFIG_STREAM_PARSER
 BPF_MAP_TYPE(BPF_MAP_TYPE_SOCKMAP, sock_map_ops)
 #endif
+BPF_MAP_TYPE(BPF_MAP_TYPE_CPUMAP, cpu_map_ops)
 #endif
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index e43491ac4823..f14e15702533 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -111,6 +111,7 @@ enum bpf_map_type {
 	BPF_MAP_TYPE_HASH_OF_MAPS,
 	BPF_MAP_TYPE_DEVMAP,
 	BPF_MAP_TYPE_SOCKMAP,
+	BPF_MAP_TYPE_CPUMAP,
 };
 
 enum bpf_prog_type {
diff --git a/kernel/bpf/Makefile b/kernel/bpf/Makefile
index 897daa005b23..dba0bd33a43c 100644
--- a/kernel/bpf/Makefile
+++ b/kernel/bpf/Makefile
@@ -4,6 +4,7 @@ obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o tnum.o
 obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o lpm_trie.o map_in_map.o
 ifeq ($(CONFIG_NET),y)
 obj-$(CONFIG_BPF_SYSCALL) += devmap.o
+obj-$(CONFIG_BPF_SYSCALL) += cpumap.o
 ifeq ($(CONFIG_STREAM_PARSER),y)
 obj-$(CONFIG_BPF_SYSCALL) += sockmap.o
 endif
diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
new file mode 100644
index 000000000000..f0948af82e65
--- /dev/null
+++ b/kernel/bpf/cpumap.c
@@ -0,0 +1,547 @@
+/* bpf/cpumap.c
+ *
+ * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
+ * Released under terms in GPL version 2.  See COPYING.
+ */
+
+/* The 'cpumap' is primary used as a backend map for XDP BPF helper
+ * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
+ *
+ * Unlike devmap which redirect XDP frames out another NIC device,
+ * this map type redirect raw XDP frames to another CPU.  The remote
+ * CPU will do SKB-allocation and call the normal network stack.
+ *
+ * This is a scalability and isolation mechanism, that allow
+ * separating the early driver network XDP layer, from the rest of the
+ * netstack, and assigning dedicated CPUs for this stage.  This
+ * basically allows for 10G wirespeed pre-filtering via bpf.
+ */
+#include <linux/bpf.h>
+#include <linux/filter.h>
+#include <linux/ptr_ring.h>
+
+#include <linux/sched.h>
+#include <linux/workqueue.h>
+#include <linux/kthread.h>
+
+/*
+ * General idea: XDP packets getting XDP redirected to another CPU,
+ * will maximum be stored/queued for one driver ->poll() call.  It is
+ * guaranteed that setting flush bit and flush operation happen on
+ * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
+ * which queue in bpf_cpu_map_entry contains packets.
+ */
+
+#define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
+struct xdp_bulk_queue {
+	void *q[CPU_MAP_BULK_SIZE];
+	unsigned int count;
+};
+
+/* Struct for every remote "destination" CPU in map */
+struct bpf_cpu_map_entry {
+	u32 cpu;    /* kthread CPU and map index */
+	int map_id; /* Back reference to map */
+	u32 qsize;  /* Redundant queue size for map lookup */
+
+	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
+	struct xdp_bulk_queue __percpu *bulkq;
+
+	/* Queue with potential multi-producers, and single-consumer kthread */
+	struct ptr_ring *queue;
+	struct task_struct *kthread;
+	struct work_struct kthread_stop_wq;
+
+	atomic_t refcnt; /* Control when this struct can be free'ed */
+	struct rcu_head rcu;
+};
+
+struct bpf_cpu_map {
+	struct bpf_map map;
+	/* Below members specific for map type */
+	struct bpf_cpu_map_entry **cpu_map;
+	unsigned long __percpu *flush_needed;
+};
+
+static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
+			     struct xdp_bulk_queue *bq);
+
+static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
+{
+	return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
+}
+
+static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
+{
+	struct bpf_cpu_map *cmap;
+	u64 cost;
+	int err;
+
+	/* check sanity of attributes */
+	if (attr->max_entries == 0 || attr->key_size != 4 ||
+	    attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
+		return ERR_PTR(-EINVAL);
+
+	cmap = kzalloc(sizeof(*cmap), GFP_USER);
+	if (!cmap)
+		return ERR_PTR(-ENOMEM);
+
+	/* mandatory map attributes */
+	cmap->map.map_type = attr->map_type;
+	cmap->map.key_size = attr->key_size;
+	cmap->map.value_size = attr->value_size;
+	cmap->map.max_entries = attr->max_entries;
+	cmap->map.map_flags = attr->map_flags;
+	cmap->map.numa_node = bpf_map_attr_numa_node(attr);
+
+	/* make sure page count doesn't overflow */
+	cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
+	cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
+	if (cost >= U32_MAX - PAGE_SIZE)
+		goto free_cmap;
+	cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
+
+	/* if map size is larger than memlock limit, reject it early */
+	err = bpf_map_precharge_memlock(cmap->map.pages);
+	if (err)
+		goto free_cmap;
+
+	/* A per cpu bitfield with a bit per possible CPU in map  */
+	cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
+					    __alignof__(unsigned long));
+	if (!cmap->flush_needed)
+		goto free_cmap;
+
+	/* Alloc array for possible remote "destination" CPUs */
+	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
+					   sizeof(struct bpf_cpu_map_entry *),
+					   cmap->map.numa_node);
+	if (!cmap->cpu_map)
+		goto free_cmap;
+
+	return &cmap->map;
+free_cmap:
+	free_percpu(cmap->flush_needed);
+	kfree(cmap);
+	return ERR_PTR(-ENOMEM);
+}
+
+void __cpu_map_queue_destructor(void *ptr)
+{
+	/* For now, just catch this as an error */
+	if (!ptr)
+		return;
+	pr_err("ERROR: %s() cpu_map queue was not empty\n", __func__);
+	page_frag_free(ptr);
+}
+
+static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
+{
+	if (atomic_dec_and_test(&rcpu->refcnt)) {
+		/* The queue should be empty at this point */
+		ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor);
+		kfree(rcpu->queue);
+		kfree(rcpu);
+	}
+}
+
+static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
+{
+	atomic_inc(&rcpu->refcnt);
+}
+
+/* called from workqueue, to workaround syscall using preempt_disable */
+static void cpu_map_kthread_stop(struct work_struct *work)
+{
+	struct bpf_cpu_map_entry *rcpu;
+
+	rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
+	synchronize_rcu(); /* wait for flush in __cpu_map_entry_free() */
+	kthread_stop(rcpu->kthread); /* calls put_cpu_map_entry */
+}
+
+static int cpu_map_kthread_run(void *data)
+{
+	struct bpf_cpu_map_entry *rcpu = data;
+
+	set_current_state(TASK_INTERRUPTIBLE);
+	while (!kthread_should_stop()) {
+		struct xdp_pkt *xdp_pkt;
+
+		schedule();
+		/* Do work */
+		while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
+			/* For now just "refcnt-free" */
+			page_frag_free(xdp_pkt);
+		}
+		__set_current_state(TASK_INTERRUPTIBLE);
+	}
+	put_cpu_map_entry(rcpu);
+
+	__set_current_state(TASK_RUNNING);
+	return 0;
+}
+
+struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id)
+{
+	gfp_t gfp = GFP_ATOMIC|__GFP_NOWARN;
+	struct bpf_cpu_map_entry *rcpu;
+	int numa, err;
+
+	/* Have map->numa_node, but choose node of redirect target CPU */
+	numa = cpu_to_node(cpu);
+
+	rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
+	if (!rcpu)
+		return NULL;
+
+	/* Alloc percpu bulkq */
+	rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
+					 sizeof(void *), gfp);
+	if (!rcpu->bulkq)
+		goto fail;
+
+	/* Alloc queue */
+	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
+	if (!rcpu->queue)
+		goto fail;
+
+	err = ptr_ring_init(rcpu->queue, qsize, gfp);
+	if (err)
+		goto fail;
+	rcpu->qsize = qsize;
+
+	/* Setup kthread */
+	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
+					       "cpumap/%d/map:%d", cpu, map_id);
+	if (IS_ERR(rcpu->kthread))
+		goto fail;
+
+	/* Make sure kthread runs on a single CPU */
+	kthread_bind(rcpu->kthread, cpu);
+	wake_up_process(rcpu->kthread);
+
+	get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
+	get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
+
+	return rcpu;
+
+fail:   /* Hint: free API detect NULL values */
+	free_percpu(rcpu->bulkq);
+	kfree(rcpu->queue);
+	kfree(rcpu);
+	return NULL;
+}
+
+void __cpu_map_entry_free(struct rcu_head *rcu)
+{
+	struct bpf_cpu_map_entry *rcpu;
+	int cpu;
+
+	/* This cpu_map_entry have been disconnected from map and one
+	 * RCU graze-period have elapsed.  Thus, XDP cannot queue any
+	 * new packets and cannot change/set flush_needed that can
+	 * find this entry.
+	 */
+	rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
+
+	/* Flush remaining packets in percpu bulkq */
+	for_each_online_cpu(cpu) {
+		struct xdp_bulk_queue *bq = per_cpu_ptr(rcpu->bulkq, cpu);
+
+		/* No concurrent bq_enqueue can run at this point */
+		bq_flush_to_queue(rcpu, bq);
+	}
+	free_percpu(rcpu->bulkq);
+	/* Cannot kthread_stop() here, last put free rcpu resources */
+	put_cpu_map_entry(rcpu);
+}
+
+/*
+ * After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
+ * ensure any driver rcu critical sections have completed, but this
+ * does not guarantee a flush has happened yet. Because driver side
+ * rcu_read_lock/unlock only protects the running XDP program.  The
+ * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
+ * pending flush op doesn't fail.
+ *
+ * The bpf_cpu_map_entry is still used by the kthread, and there can
+ * still be pending packets (in queue and percpu bulkq).  A refcnt
+ * makes sure to last user (kthread_stop vs. call_rcu) free memory
+ * resources.
+ *
+ * The rcu callback __cpu_map_entry_free flush remaining packets in
+ * percpu bulkq to queue.  Due to caller map_delete_elem() disable
+ * preemption, cannot call kthread_stop() to make sure queue is empty.
+ * Instead a work_queue is started for stopping kthread,
+ * cpu_map_kthread_stop, which waits for an RCU graze period before
+ * stopping kthread, emptying the queue.
+ */
+void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
+			     u32 key_cpu, struct bpf_cpu_map_entry *rcpu)
+{
+	struct bpf_cpu_map_entry *old_rcpu;
+
+	old_rcpu = xchg(&cmap->cpu_map[key_cpu], rcpu);
+	if (old_rcpu) {
+		call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
+		INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
+		schedule_work(&old_rcpu->kthread_stop_wq);
+	}
+}
+
+int cpu_map_delete_elem(struct bpf_map *map, void *key)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	u32 key_cpu = *(u32 *)key;
+
+	if (key_cpu >= map->max_entries)
+		return -EINVAL;
+
+	/* notice caller map_delete_elem() use preempt_disable() */
+	__cpu_map_entry_replace(cmap, key_cpu, NULL);
+	return 0;
+}
+
+int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
+				u64 map_flags)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	struct bpf_cpu_map_entry *rcpu;
+
+	/* Array index key correspond to CPU number */
+	u32 key_cpu = *(u32 *)key;
+	/* Q: The value could be the queue size? */
+	u32 qsize = *(u32 *)value;
+
+	if (unlikely(map_flags > BPF_EXIST))
+		return -EINVAL;
+	if (unlikely(key_cpu >= cmap->map.max_entries))
+		return -E2BIG;
+	if (unlikely(map_flags == BPF_NOEXIST))
+		return -EEXIST;
+	if (unlikely(qsize > 16384)) /* sanity limit on qsize */
+		return -EOVERFLOW;
+
+	if (qsize == 0) {
+		rcpu = NULL; /* Same as deleting */
+	} else {
+		/* Updating qsize cause re-allocation of bpf_cpu_map_entry */
+		rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
+		if (!rcpu)
+			return -ENOMEM;
+	}
+	rcu_read_lock();
+	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
+	rcu_read_unlock();
+	return 0;
+}
+
+void cpu_map_free(struct bpf_map *map)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	int cpu;
+	u32 i;
+
+	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
+	 * so the bpf programs (can be more than one that used this map) were
+	 * disconnected from events. Wait for outstanding critical sections in
+	 * these programs to complete. The rcu critical section only guarantees
+	 * no further "XDP/bpf-side" reads against bpf_cpu_map->cpu_map.
+	 * It does __not__ ensure pending flush operations (if any) are
+	 * complete.
+	 */
+	synchronize_rcu();
+
+	/* To ensure all pending flush operations have completed wait for flush
+	 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
+	 * Because the above synchronize_rcu() ensures the map is disconnected
+	 * from the program we can assume no new bits will be set.
+	 */
+	for_each_online_cpu(cpu) {
+		unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu);
+
+		while (!bitmap_empty(bitmap, cmap->map.max_entries))
+			cond_resched();
+	}
+
+	/* For cpu_map the remote CPUs can still be using the entries
+	 * (struct bpf_cpu_map_entry).
+	 */
+	for (i = 0; i < cmap->map.max_entries; i++) {
+		struct bpf_cpu_map_entry *rcpu;
+
+		rcpu = READ_ONCE(cmap->cpu_map[i]);
+		if (!rcpu)
+			continue;
+
+		/* bq flush and cleanup happens after RCU graze-period */
+		__cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
+	}
+	free_percpu(cmap->flush_needed);
+	bpf_map_area_free(cmap->cpu_map);
+	kfree(cmap);
+}
+
+struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	struct bpf_cpu_map_entry *rcpu;
+
+	if (key >= map->max_entries)
+		return NULL;
+
+	rcpu = READ_ONCE(cmap->cpu_map[key]);
+	return rcpu;
+}
+
+static void *cpu_map_lookup_elem(struct bpf_map *map, void *key)
+{
+	struct bpf_cpu_map_entry *rcpu =
+		__cpu_map_lookup_elem(map, *(u32 *)key);
+
+	return rcpu ? &rcpu->qsize : NULL;
+}
+
+static int cpu_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	u32 index = key ? *(u32 *)key : U32_MAX;
+	u32 *next = next_key;
+
+	if (index >= cmap->map.max_entries) {
+		*next = 0;
+		return 0;
+	}
+
+	if (index == cmap->map.max_entries - 1)
+		return -ENOENT;
+	*next = index + 1;
+	return 0;
+}
+
+const struct bpf_map_ops cpu_map_ops = {
+	.map_alloc		= cpu_map_alloc,
+	.map_free		= cpu_map_free,
+	.map_delete_elem	= cpu_map_delete_elem,
+	.map_update_elem	= cpu_map_update_elem,
+	.map_lookup_elem	= cpu_map_lookup_elem,
+	.map_get_next_key	= cpu_map_get_next_key,
+};
+
+
+static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
+			     struct xdp_bulk_queue *bq)
+{
+	struct ptr_ring *q;
+	int i;
+
+	if (unlikely(!bq->count))
+		return 0;
+
+	q = rcpu->queue;
+	spin_lock(&q->producer_lock);
+
+	for (i = 0; i < bq->count; i++) {
+		void *xdp_pkt = bq->q[i];
+		int err;
+
+		err = __ptr_ring_produce(q, xdp_pkt);
+		if (err) {
+			/* Free xdp_pkt */
+			page_frag_free(xdp_pkt);
+		}
+	}
+	bq->count = 0;
+	spin_unlock(&q->producer_lock);
+
+	return 0;
+}
+
+/* Notice: Will change in later patch */
+struct xdp_pkt {
+	void *data;
+	u16 len;
+	u16 headroom;
+};
+
+/* Runs under RCU-read-side, plus in softirq under NAPI protection.
+ * Thus, safe percpu variable access.
+ */
+static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt)
+{
+	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
+
+	if (unlikely(bq->count == CPU_MAP_BULK_SIZE)) {
+		bq_flush_to_queue(rcpu, bq);
+	}
+	/* Notice, xdp_buff/page MUST be queued here, long enough for
+	 * driver to code invoking us to finished, due to driver
+	 * (e.g. ixgbe) recycle tricks based on page-refcnt.
+	 *
+	 * Thus, incoming xdp_pkt is always queued here (else we race
+	 * with another CPU on page-refcnt and remaining driver code).
+	 * Queue time is very short, as driver will invoke flush
+	 * operation, when completing napi->poll call.
+	 */
+	bq->q[bq->count++] = xdp_pkt;
+	return 0;
+}
+
+int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp)
+{
+	struct xdp_pkt *xdp_pkt;
+	int headroom;
+
+	/* Convert xdp_buff to xdp_pkt */
+	headroom = xdp->data - xdp->data_hard_start;
+	if (headroom < sizeof(*xdp_pkt))
+		return -EOVERFLOW;
+	xdp_pkt = xdp->data_hard_start;
+	xdp_pkt->data = xdp->data;
+	xdp_pkt->len  = xdp->data_end - xdp->data;
+	xdp_pkt->headroom = headroom;
+	/* For now this is just used as a void pointer to data_hard_start */
+
+	bq_enqueue(rcpu, xdp_pkt);
+	return 0;
+}
+
+void __cpu_map_insert_ctx(struct bpf_map *map, u32 bit)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
+
+	__set_bit(bit, bitmap);
+}
+
+void __cpu_map_flush(struct bpf_map *map)
+{
+	struct bpf_cpu_map *cmap = container_of(map, struct bpf_cpu_map, map);
+	unsigned long *bitmap = this_cpu_ptr(cmap->flush_needed);
+	u32 bit;
+
+	/* The napi->poll softirq makes sure __cpu_map_insert_ctx()
+	 * and __cpu_map_flush() happen on same CPU. Thus, the percpu
+	 * bitmap indicate which percpu bulkq have packets.
+	 */
+	for_each_set_bit(bit, bitmap, map->max_entries) {
+		struct bpf_cpu_map_entry *rcpu = READ_ONCE(cmap->cpu_map[bit]);
+		struct xdp_bulk_queue *bq;
+
+		/* This is possible if entry is removed by user space
+		 * between xdp redirect and flush op.
+		 */
+		if (unlikely(!rcpu))
+			continue;
+
+		__clear_bit(bit, bitmap);
+
+		/* Flush all frames in bulkq to real queue */
+		bq = this_cpu_ptr(rcpu->bulkq);
+		bq_flush_to_queue(rcpu, bq);
+
+		/* If already running, costs spin_lock_irqsave + smb_mb */
+		wake_up_process(rcpu->kthread);
+	}
+}
diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
index 25d074920a00..68fe3f51e1a0 100644
--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c
@@ -562,6 +562,12 @@ static int map_update_elem(union bpf_attr *attr)
 	if (copy_from_user(value, uvalue, value_size) != 0)
 		goto free_value;
 
+	/* Need to create a kthread, thus must support schedule */
+	if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
+		err = map->ops->map_update_elem(map, key, value, attr->flags);
+		goto out;
+	}
+
 	/* must increment bpf_prog_active to avoid kprobe+bpf triggering from
 	 * inside bpf map update or delete otherwise deadlocks are possible
 	 */
@@ -592,7 +598,7 @@ static int map_update_elem(union bpf_attr *attr)
 	}
 	__this_cpu_dec(bpf_prog_active);
 	preempt_enable();
-
+out:
 	if (!err)
 		trace_bpf_map_update_elem(map, ufd, key, value);
 free_value:
diff --git a/tools/include/uapi/linux/bpf.h b/tools/include/uapi/linux/bpf.h
index e43491ac4823..f14e15702533 100644
--- a/tools/include/uapi/linux/bpf.h
+++ b/tools/include/uapi/linux/bpf.h
@@ -111,6 +111,7 @@ enum bpf_map_type {
 	BPF_MAP_TYPE_HASH_OF_MAPS,
 	BPF_MAP_TYPE_DEVMAP,
 	BPF_MAP_TYPE_SOCKMAP,
+	BPF_MAP_TYPE_CPUMAP,
 };
 
 enum bpf_prog_type {

Re: [net-next PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Alexei Starovoitov]

On Thu, Sep 28, 2017 at 02:57:08PM +0200, Jesper Dangaard Brouer wrote:
> The 'cpumap' is primary used as a backend map for XDP BPF helper
> call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
> 
> This patch implement the main part of the map.  It is not connected to
> the XDP redirect system yet, and no SKB allocation are done yet.
> 
> The main concern in this patch is to ensure the datapath can run
> without any locking.  This adds complexity to the setup and tear-down
> procedure, which assumptions are extra carefully documented in the
> code comments.
> 
> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> ---
>  include/linux/bpf_types.h      |    1 
>  include/uapi/linux/bpf.h       |    1 
>  kernel/bpf/Makefile            |    1 
>  kernel/bpf/cpumap.c            |  547 ++++++++++++++++++++++++++++++++++++++++
>  kernel/bpf/syscall.c           |    8 +
>  tools/include/uapi/linux/bpf.h |    1 
>  6 files changed, 558 insertions(+), 1 deletion(-)
>  create mode 100644 kernel/bpf/cpumap.c
> 
> diff --git a/include/linux/bpf_types.h b/include/linux/bpf_types.h
> index 6f1a567667b8..814c1081a4a9 100644
> --- a/include/linux/bpf_types.h
> +++ b/include/linux/bpf_types.h
> @@ -41,4 +41,5 @@ BPF_MAP_TYPE(BPF_MAP_TYPE_DEVMAP, dev_map_ops)
>  #ifdef CONFIG_STREAM_PARSER
>  BPF_MAP_TYPE(BPF_MAP_TYPE_SOCKMAP, sock_map_ops)
>  #endif
> +BPF_MAP_TYPE(BPF_MAP_TYPE_CPUMAP, cpu_map_ops)
>  #endif
> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> index e43491ac4823..f14e15702533 100644
> --- a/include/uapi/linux/bpf.h
> +++ b/include/uapi/linux/bpf.h
> @@ -111,6 +111,7 @@ enum bpf_map_type {
>  	BPF_MAP_TYPE_HASH_OF_MAPS,
>  	BPF_MAP_TYPE_DEVMAP,
>  	BPF_MAP_TYPE_SOCKMAP,
> +	BPF_MAP_TYPE_CPUMAP,
>  };
>  
>  enum bpf_prog_type {
> diff --git a/kernel/bpf/Makefile b/kernel/bpf/Makefile
> index 897daa005b23..dba0bd33a43c 100644
> --- a/kernel/bpf/Makefile
> +++ b/kernel/bpf/Makefile
> @@ -4,6 +4,7 @@ obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o tnum.o
>  obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o lpm_trie.o map_in_map.o
>  ifeq ($(CONFIG_NET),y)
>  obj-$(CONFIG_BPF_SYSCALL) += devmap.o
> +obj-$(CONFIG_BPF_SYSCALL) += cpumap.o
>  ifeq ($(CONFIG_STREAM_PARSER),y)
>  obj-$(CONFIG_BPF_SYSCALL) += sockmap.o
>  endif
> diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
> new file mode 100644
> index 000000000000..f0948af82e65
> --- /dev/null
> +++ b/kernel/bpf/cpumap.c
> @@ -0,0 +1,547 @@
> +/* bpf/cpumap.c
> + *
> + * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
> + * Released under terms in GPL version 2.  See COPYING.
> + */
> +
> +/* The 'cpumap' is primary used as a backend map for XDP BPF helper
> + * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
> + *
> + * Unlike devmap which redirect XDP frames out another NIC device,
> + * this map type redirect raw XDP frames to another CPU.  The remote
> + * CPU will do SKB-allocation and call the normal network stack.
> + *
> + * This is a scalability and isolation mechanism, that allow
> + * separating the early driver network XDP layer, from the rest of the
> + * netstack, and assigning dedicated CPUs for this stage.  This
> + * basically allows for 10G wirespeed pre-filtering via bpf.
> + */
> +#include <linux/bpf.h>
> +#include <linux/filter.h>
> +#include <linux/ptr_ring.h>
> +
> +#include <linux/sched.h>
> +#include <linux/workqueue.h>
> +#include <linux/kthread.h>
> +
> +/*
> + * General idea: XDP packets getting XDP redirected to another CPU,
> + * will maximum be stored/queued for one driver ->poll() call.  It is
> + * guaranteed that setting flush bit and flush operation happen on
> + * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
> + * which queue in bpf_cpu_map_entry contains packets.
> + */
> +
> +#define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
> +struct xdp_bulk_queue {
> +	void *q[CPU_MAP_BULK_SIZE];
> +	unsigned int count;
> +};
> +
> +/* Struct for every remote "destination" CPU in map */
> +struct bpf_cpu_map_entry {
> +	u32 cpu;    /* kthread CPU and map index */
> +	int map_id; /* Back reference to map */
> +	u32 qsize;  /* Redundant queue size for map lookup */
> +
> +	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
> +	struct xdp_bulk_queue __percpu *bulkq;
> +
> +	/* Queue with potential multi-producers, and single-consumer kthread */
> +	struct ptr_ring *queue;
> +	struct task_struct *kthread;
> +	struct work_struct kthread_stop_wq;
> +
> +	atomic_t refcnt; /* Control when this struct can be free'ed */
> +	struct rcu_head rcu;
> +};
> +
> +struct bpf_cpu_map {
> +	struct bpf_map map;
> +	/* Below members specific for map type */
> +	struct bpf_cpu_map_entry **cpu_map;
> +	unsigned long __percpu *flush_needed;
> +};
> +
> +static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
> +			     struct xdp_bulk_queue *bq);
> +
> +static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
> +{
> +	return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
> +}
> +
> +static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
> +{
> +	struct bpf_cpu_map *cmap;
> +	u64 cost;
> +	int err;
> +
> +	/* check sanity of attributes */
> +	if (attr->max_entries == 0 || attr->key_size != 4 ||
> +	    attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
> +		return ERR_PTR(-EINVAL);
> +
> +	cmap = kzalloc(sizeof(*cmap), GFP_USER);
> +	if (!cmap)
> +		return ERR_PTR(-ENOMEM);
> +
> +	/* mandatory map attributes */
> +	cmap->map.map_type = attr->map_type;
> +	cmap->map.key_size = attr->key_size;
> +	cmap->map.value_size = attr->value_size;
> +	cmap->map.max_entries = attr->max_entries;
> +	cmap->map.map_flags = attr->map_flags;
> +	cmap->map.numa_node = bpf_map_attr_numa_node(attr);
> +
> +	/* make sure page count doesn't overflow */
> +	cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
> +	cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
> +	if (cost >= U32_MAX - PAGE_SIZE)
> +		goto free_cmap;
> +	cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
> +
> +	/* if map size is larger than memlock limit, reject it early */
> +	err = bpf_map_precharge_memlock(cmap->map.pages);
> +	if (err)
> +		goto free_cmap;
> +
> +	/* A per cpu bitfield with a bit per possible CPU in map  */
> +	cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
> +					    __alignof__(unsigned long));
> +	if (!cmap->flush_needed)
> +		goto free_cmap;
> +
> +	/* Alloc array for possible remote "destination" CPUs */
> +	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
> +					   sizeof(struct bpf_cpu_map_entry *),
> +					   cmap->map.numa_node);
> +	if (!cmap->cpu_map)
> +		goto free_cmap;
> +
> +	return &cmap->map;
> +free_cmap:
> +	free_percpu(cmap->flush_needed);
> +	kfree(cmap);
> +	return ERR_PTR(-ENOMEM);
> +}
> +
> +void __cpu_map_queue_destructor(void *ptr)
> +{
> +	/* For now, just catch this as an error */
> +	if (!ptr)
> +		return;
> +	pr_err("ERROR: %s() cpu_map queue was not empty\n", __func__);
> +	page_frag_free(ptr);
> +}
> +
> +static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
> +{
> +	if (atomic_dec_and_test(&rcpu->refcnt)) {
> +		/* The queue should be empty at this point */
> +		ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor);
> +		kfree(rcpu->queue);
> +		kfree(rcpu);
> +	}
> +}
> +
> +static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
> +{
> +	atomic_inc(&rcpu->refcnt);
> +}
> +
> +/* called from workqueue, to workaround syscall using preempt_disable */
> +static void cpu_map_kthread_stop(struct work_struct *work)
> +{
> +	struct bpf_cpu_map_entry *rcpu;
> +
> +	rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
> +	synchronize_rcu(); /* wait for flush in __cpu_map_entry_free() */
> +	kthread_stop(rcpu->kthread); /* calls put_cpu_map_entry */
> +}
> +
> +static int cpu_map_kthread_run(void *data)
> +{
> +	struct bpf_cpu_map_entry *rcpu = data;
> +
> +	set_current_state(TASK_INTERRUPTIBLE);
> +	while (!kthread_should_stop()) {
> +		struct xdp_pkt *xdp_pkt;
> +
> +		schedule();
> +		/* Do work */
> +		while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
> +			/* For now just "refcnt-free" */
> +			page_frag_free(xdp_pkt);
> +		}
> +		__set_current_state(TASK_INTERRUPTIBLE);
> +	}
> +	put_cpu_map_entry(rcpu);
> +
> +	__set_current_state(TASK_RUNNING);
> +	return 0;
> +}
> +
> +struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id)
> +{
> +	gfp_t gfp = GFP_ATOMIC|__GFP_NOWARN;
> +	struct bpf_cpu_map_entry *rcpu;
> +	int numa, err;
> +
> +	/* Have map->numa_node, but choose node of redirect target CPU */
> +	numa = cpu_to_node(cpu);
> +
> +	rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
> +	if (!rcpu)
> +		return NULL;
> +
> +	/* Alloc percpu bulkq */
> +	rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
> +					 sizeof(void *), gfp);
> +	if (!rcpu->bulkq)
> +		goto fail;
> +
> +	/* Alloc queue */
> +	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
> +	if (!rcpu->queue)
> +		goto fail;
> +
> +	err = ptr_ring_init(rcpu->queue, qsize, gfp);
> +	if (err)
> +		goto fail;
> +	rcpu->qsize = qsize;
> +
> +	/* Setup kthread */
> +	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
> +					       "cpumap/%d/map:%d", cpu, map_id);
> +	if (IS_ERR(rcpu->kthread))
> +		goto fail;
> +
> +	/* Make sure kthread runs on a single CPU */
> +	kthread_bind(rcpu->kthread, cpu);

is there a check that max_entries <= num_possible_cpu ? I couldn't find it.
otherwise it will be binding to impossible cpu?

> +	wake_up_process(rcpu->kthread);

In general the whole thing looks like 'threaded NAPI' that Hannes was
proposing some time back. I liked it back then and I like it now.
I don't remember what were the objections back then.
Something scheduler related?
Adding Hannes.

Still curious about the questions I asked in the other thread
on what's causing it to be so much better than RPS

Re: [net-next PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Hannes Frederic Sowa]

[adding Paolo, Eric]

Alexei Starovoitov <alexei.starovoitov@gmail.com> writes:

> On Thu, Sep 28, 2017 at 02:57:08PM +0200, Jesper Dangaard Brouer wrote:

[...]

>> +	wake_up_process(rcpu->kthread);
>
> In general the whole thing looks like 'threaded NAPI' that Hannes was
> proposing some time back. I liked it back then and I like it now.
> I don't remember what were the objections back then.
> Something scheduler related?
> Adding Hannes.

Yes.

The main objection from Eric at that time was that user space now starts
to compete with the threaded NAPI threads depending on process
priorities, which are under control of user space. Softirq always runs
first to end. Networking could starve because a process with higher
priority is runnable. At that time Eric found a way to fix the
particular problem, which resulted in commit 4cd13c21b207e80d. Pinning
and other control is also possible from user space, causing more complex
tuning set ups and problems will be harder to debug.

In particular after Eric's patch threaded NAPI proofed itself to be not
useful anymore, because his patch successfully deferred work to the
ksoftirqd more reliable thus allowing the UDP rx queue to get drained by
user space.

> Still curious about the questions I asked in the other thread
> on what's causing it to be so much better than RPS

My guess is that RPS uses expensive IPI to notify the remote
softirq. The batching size on RPS depends on how many packets could get
worked on during one softirq invocation on the source CPU until we wake
up remote CPU(s!), if they are not constantly running.

Re: [net-next PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Paolo Abeni]

On Fri, 2017-09-29 at 09:56 +0200, Hannes Frederic Sowa wrote:
> [adding Paolo, Eric]
> 
> Alexei Starovoitov <alexei.starovoitov@gmail.com> writes:
> 
> > On Thu, Sep 28, 2017 at 02:57:08PM +0200, Jesper Dangaard Brouer wrote:
> 
> [...]
> 
> > > +	wake_up_process(rcpu->kthread);
> > 
> > In general the whole thing looks like 'threaded NAPI' that Hannes was
> > proposing some time back. I liked it back then and I like it now.
> > I don't remember what were the objections back then.
> > Something scheduler related?
> > Adding Hannes.

Beyond the added scheduling complexity, the threaded NAPI
implementation proposed some time ago also possibly introduced OoO
packet delivery, because the NAPI threads were left unbound to any CPU.

Cheers,

Paolo

Re: [net-next PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Hannes Frederic Sowa]

Paolo Abeni <pabeni@redhat.com> writes:

> On Fri, 2017-09-29 at 09:56 +0200, Hannes Frederic Sowa wrote:
>> [adding Paolo, Eric]
>> 
>> Alexei Starovoitov <alexei.starovoitov@gmail.com> writes:
>> 
>> > On Thu, Sep 28, 2017 at 02:57:08PM +0200, Jesper Dangaard Brouer wrote:
>> 
>> [...]
>> 
>> > > +	wake_up_process(rcpu->kthread);
>> > 
>> > In general the whole thing looks like 'threaded NAPI' that Hannes was
>> > proposing some time back. I liked it back then and I like it now.
>> > I don't remember what were the objections back then.
>> > Something scheduler related?
>> > Adding Hannes.
>
> Beyond the added scheduling complexity, the threaded NAPI
> implementation proposed some time ago also possibly introduced OoO
> packet delivery, because the NAPI threads were left unbound to any CPU.

Right, yes, but that can be resolved. The problem was just in that
particular patch.

Re: [net-next PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Jesper Dangaard Brouer]

On Thu, 28 Sep 2017 20:21:47 -0700
Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:

> On Thu, Sep 28, 2017 at 02:57:08PM +0200, Jesper Dangaard Brouer wrote:
> > The 'cpumap' is primary used as a backend map for XDP BPF helper
> > call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
> > 
> > This patch implement the main part of the map.  It is not connected to
> > the XDP redirect system yet, and no SKB allocation are done yet.
> > 
> > The main concern in this patch is to ensure the datapath can run
> > without any locking.  This adds complexity to the setup and tear-down
> > procedure, which assumptions are extra carefully documented in the
> > code comments.
> > 
> > Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > ---
> >  include/linux/bpf_types.h      |    1 
> >  include/uapi/linux/bpf.h       |    1 
> >  kernel/bpf/Makefile            |    1 
> >  kernel/bpf/cpumap.c            |  547 ++++++++++++++++++++++++++++++++++++++++
> >  kernel/bpf/syscall.c           |    8 +
> >  tools/include/uapi/linux/bpf.h |    1 
> >  6 files changed, 558 insertions(+), 1 deletion(-)
> >  create mode 100644 kernel/bpf/cpumap.c
> > 
> > diff --git a/include/linux/bpf_types.h b/include/linux/bpf_types.h
> > index 6f1a567667b8..814c1081a4a9 100644
> > --- a/include/linux/bpf_types.h
> > +++ b/include/linux/bpf_types.h
> > @@ -41,4 +41,5 @@ BPF_MAP_TYPE(BPF_MAP_TYPE_DEVMAP, dev_map_ops)
> >  #ifdef CONFIG_STREAM_PARSER
> >  BPF_MAP_TYPE(BPF_MAP_TYPE_SOCKMAP, sock_map_ops)
> >  #endif
> > +BPF_MAP_TYPE(BPF_MAP_TYPE_CPUMAP, cpu_map_ops)
> >  #endif
> > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > index e43491ac4823..f14e15702533 100644
> > --- a/include/uapi/linux/bpf.h
> > +++ b/include/uapi/linux/bpf.h
> > @@ -111,6 +111,7 @@ enum bpf_map_type {
> >  	BPF_MAP_TYPE_HASH_OF_MAPS,
> >  	BPF_MAP_TYPE_DEVMAP,
> >  	BPF_MAP_TYPE_SOCKMAP,
> > +	BPF_MAP_TYPE_CPUMAP,
> >  };
> >  
> >  enum bpf_prog_type {
> > diff --git a/kernel/bpf/Makefile b/kernel/bpf/Makefile
> > index 897daa005b23..dba0bd33a43c 100644
> > --- a/kernel/bpf/Makefile
> > +++ b/kernel/bpf/Makefile
> > @@ -4,6 +4,7 @@ obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o tnum.o
> >  obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o lpm_trie.o map_in_map.o
> >  ifeq ($(CONFIG_NET),y)
> >  obj-$(CONFIG_BPF_SYSCALL) += devmap.o
> > +obj-$(CONFIG_BPF_SYSCALL) += cpumap.o
> >  ifeq ($(CONFIG_STREAM_PARSER),y)
> >  obj-$(CONFIG_BPF_SYSCALL) += sockmap.o
> >  endif
> > diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
> > new file mode 100644
> > index 000000000000..f0948af82e65
> > --- /dev/null
> > +++ b/kernel/bpf/cpumap.c
> > @@ -0,0 +1,547 @@
> > +/* bpf/cpumap.c
> > + *
> > + * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
> > + * Released under terms in GPL version 2.  See COPYING.
> > + */
> > +
> > +/* The 'cpumap' is primary used as a backend map for XDP BPF helper
> > + * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
> > + *
> > + * Unlike devmap which redirect XDP frames out another NIC device,
> > + * this map type redirect raw XDP frames to another CPU.  The remote
> > + * CPU will do SKB-allocation and call the normal network stack.
> > + *
> > + * This is a scalability and isolation mechanism, that allow
> > + * separating the early driver network XDP layer, from the rest of the
> > + * netstack, and assigning dedicated CPUs for this stage.  This
> > + * basically allows for 10G wirespeed pre-filtering via bpf.
> > + */
> > +#include <linux/bpf.h>
> > +#include <linux/filter.h>
> > +#include <linux/ptr_ring.h>
> > +
> > +#include <linux/sched.h>
> > +#include <linux/workqueue.h>
> > +#include <linux/kthread.h>
> > +
> > +/*
> > + * General idea: XDP packets getting XDP redirected to another CPU,
> > + * will maximum be stored/queued for one driver ->poll() call.  It is
> > + * guaranteed that setting flush bit and flush operation happen on
> > + * same CPU.  Thus, cpu_map_flush operation can deduct via this_cpu_ptr()
> > + * which queue in bpf_cpu_map_entry contains packets.
> > + */
> > +
> > +#define CPU_MAP_BULK_SIZE 8  /* 8 == one cacheline on 64-bit archs */
> > +struct xdp_bulk_queue {
> > +	void *q[CPU_MAP_BULK_SIZE];
> > +	unsigned int count;
> > +};
> > +
> > +/* Struct for every remote "destination" CPU in map */
> > +struct bpf_cpu_map_entry {
> > +	u32 cpu;    /* kthread CPU and map index */
> > +	int map_id; /* Back reference to map */
> > +	u32 qsize;  /* Redundant queue size for map lookup */
> > +
> > +	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
> > +	struct xdp_bulk_queue __percpu *bulkq;
> > +
> > +	/* Queue with potential multi-producers, and single-consumer kthread */
> > +	struct ptr_ring *queue;
> > +	struct task_struct *kthread;
> > +	struct work_struct kthread_stop_wq;
> > +
> > +	atomic_t refcnt; /* Control when this struct can be free'ed */
> > +	struct rcu_head rcu;
> > +};
> > +
> > +struct bpf_cpu_map {
> > +	struct bpf_map map;
> > +	/* Below members specific for map type */
> > +	struct bpf_cpu_map_entry **cpu_map;
> > +	unsigned long __percpu *flush_needed;
> > +};
> > +
> > +static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
> > +			     struct xdp_bulk_queue *bq);
> > +
> > +static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
> > +{
> > +	return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
> > +}
> > +
> > +static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
> > +{
> > +	struct bpf_cpu_map *cmap;
> > +	u64 cost;
> > +	int err;
> > +
> > +	/* check sanity of attributes */
> > +	if (attr->max_entries == 0 || attr->key_size != 4 ||
> > +	    attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
> > +		return ERR_PTR(-EINVAL);
> > +
> > +	cmap = kzalloc(sizeof(*cmap), GFP_USER);
> > +	if (!cmap)
> > +		return ERR_PTR(-ENOMEM);
> > +
> > +	/* mandatory map attributes */
> > +	cmap->map.map_type = attr->map_type;
> > +	cmap->map.key_size = attr->key_size;
> > +	cmap->map.value_size = attr->value_size;
> > +	cmap->map.max_entries = attr->max_entries;
> > +	cmap->map.map_flags = attr->map_flags;
> > +	cmap->map.numa_node = bpf_map_attr_numa_node(attr);
> > +
> > +	/* make sure page count doesn't overflow */
> > +	cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
> > +	cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
> > +	if (cost >= U32_MAX - PAGE_SIZE)
> > +		goto free_cmap;
> > +	cmap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
> > +
> > +	/* if map size is larger than memlock limit, reject it early */
> > +	err = bpf_map_precharge_memlock(cmap->map.pages);
> > +	if (err)
> > +		goto free_cmap;
> > +
> > +	/* A per cpu bitfield with a bit per possible CPU in map  */
> > +	cmap->flush_needed = __alloc_percpu(cpu_map_bitmap_size(attr),
> > +					    __alignof__(unsigned long));
> > +	if (!cmap->flush_needed)
> > +		goto free_cmap;
> > +
> > +	/* Alloc array for possible remote "destination" CPUs */
> > +	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
> > +					   sizeof(struct bpf_cpu_map_entry *),
> > +					   cmap->map.numa_node);
> > +	if (!cmap->cpu_map)
> > +		goto free_cmap;
> > +
> > +	return &cmap->map;
> > +free_cmap:
> > +	free_percpu(cmap->flush_needed);
> > +	kfree(cmap);
> > +	return ERR_PTR(-ENOMEM);
> > +}
> > +
> > +void __cpu_map_queue_destructor(void *ptr)
> > +{
> > +	/* For now, just catch this as an error */
> > +	if (!ptr)
> > +		return;
> > +	pr_err("ERROR: %s() cpu_map queue was not empty\n", __func__);
> > +	page_frag_free(ptr);
> > +}
> > +
> > +static void put_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
> > +{
> > +	if (atomic_dec_and_test(&rcpu->refcnt)) {
> > +		/* The queue should be empty at this point */
> > +		ptr_ring_cleanup(rcpu->queue, __cpu_map_queue_destructor);
> > +		kfree(rcpu->queue);
> > +		kfree(rcpu);
> > +	}
> > +}
> > +
> > +static void get_cpu_map_entry(struct bpf_cpu_map_entry *rcpu)
> > +{
> > +	atomic_inc(&rcpu->refcnt);
> > +}
> > +
> > +/* called from workqueue, to workaround syscall using preempt_disable */
> > +static void cpu_map_kthread_stop(struct work_struct *work)
> > +{
> > +	struct bpf_cpu_map_entry *rcpu;
> > +
> > +	rcpu = container_of(work, struct bpf_cpu_map_entry, kthread_stop_wq);
> > +	synchronize_rcu(); /* wait for flush in __cpu_map_entry_free() */
> > +	kthread_stop(rcpu->kthread); /* calls put_cpu_map_entry */
> > +}
> > +
> > +static int cpu_map_kthread_run(void *data)
> > +{
> > +	struct bpf_cpu_map_entry *rcpu = data;
> > +
> > +	set_current_state(TASK_INTERRUPTIBLE);
> > +	while (!kthread_should_stop()) {
> > +		struct xdp_pkt *xdp_pkt;
> > +
> > +		schedule();
> > +		/* Do work */
> > +		while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
> > +			/* For now just "refcnt-free" */
> > +			page_frag_free(xdp_pkt);
> > +		}
> > +		__set_current_state(TASK_INTERRUPTIBLE);
> > +	}
> > +	put_cpu_map_entry(rcpu);
> > +
> > +	__set_current_state(TASK_RUNNING);
> > +	return 0;
> > +}
> > +
> > +struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id)
> > +{
> > +	gfp_t gfp = GFP_ATOMIC|__GFP_NOWARN;
> > +	struct bpf_cpu_map_entry *rcpu;
> > +	int numa, err;
> > +
> > +	/* Have map->numa_node, but choose node of redirect target CPU */
> > +	numa = cpu_to_node(cpu);
> > +
> > +	rcpu = kzalloc_node(sizeof(*rcpu), gfp, numa);
> > +	if (!rcpu)
> > +		return NULL;
> > +
> > +	/* Alloc percpu bulkq */
> > +	rcpu->bulkq = __alloc_percpu_gfp(sizeof(*rcpu->bulkq),
> > +					 sizeof(void *), gfp);
> > +	if (!rcpu->bulkq)
> > +		goto fail;
> > +
> > +	/* Alloc queue */
> > +	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
> > +	if (!rcpu->queue)
> > +		goto fail;
> > +
> > +	err = ptr_ring_init(rcpu->queue, qsize, gfp);
> > +	if (err)
> > +		goto fail;
> > +	rcpu->qsize = qsize;
> > +
> > +	/* Setup kthread */
> > +	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
> > +					       "cpumap/%d/map:%d", cpu, map_id);
> > +	if (IS_ERR(rcpu->kthread))
> > +		goto fail;
> > +
> > +	/* Make sure kthread runs on a single CPU */
> > +	kthread_bind(rcpu->kthread, cpu);  
> 
> is there a check that max_entries <= num_possible_cpu ? I couldn't
> find it. otherwise it will be binding to impossible cpu?

Good point! -- I'll find an appropriate place to add such a limit.


> > +	wake_up_process(rcpu->kthread);  
> 
> In general the whole thing looks like 'threaded NAPI' that Hannes was
> proposing some time back. I liked it back then and I like it now.
> I don't remember what were the objections back then.
> Something scheduler related?
> Adding Hannes.

It is related to the threaded NAPI' idea[1], and I did choose kthreads
because this was used by this patch[1].
(Link to Hannes & Paolo's patch:[1] http://patchwork.ozlabs.org/patch/620657/)

It's less-intrusive, as it's only activated specifically when activating
bpf+XDP+cpumap.  Plus, it's not taking over the calling of napi->poll,
it is "just" making to "cost" of calling napi->poll significantly
smaller, as it moves invoking the network stack to another kthread. And
the choice is done on a per packet level (you don't get more
flexibility than that).

> Still curious about the questions I asked in the other thread
> on what's causing it to be so much better than RPS

Answered in that thread.  It is simply that the RPS-RX CPU have to do
too much work (like memory allocations).  Plus it uses more expensive
IPI calls, where I use wake_up_process() which doesn't do a IPI if it
can see that the remote thread is already running.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

[net-next PATCH 2/5] bpf: XDP_REDIRECT enable use of cpumap

[Jesper Dangaard Brouer]

This patch connects cpumap to the xdp_do_redirect_map infrastructure.

Still no SKB allocation are done yet.  The XDP frames are transferred
to the other CPU, but they are simply refcnt decremented on the remote
CPU.  This served as a good benchmark for measuring the overhead of
remote refcnt decrement.  If driver page recycle cache is not
efficient then this, exposes a bottleneck in the page allocator.

A shout-out to MST's ptr_ring, which is the secret behind is being so
efficient to transfer memory pointers between CPUs, without constantly
bouncing cache-lines between CPUs.

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/linux/bpf.h        |    7 +++++
 include/trace/events/xdp.h |   10 +++++--
 kernel/bpf/cpumap.c        |    5 ++-
 kernel/bpf/verifier.c      |    3 +-
 net/core/filter.c          |   65 +++++++++++++++++++++++++++++++++++++++-----
 5 files changed, 77 insertions(+), 13 deletions(-)

diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index 2b672c50f160..7f70b03e7426 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -317,6 +317,13 @@ struct net_device  *__dev_map_lookup_elem(struct bpf_map *map, u32 key);
 void __dev_map_insert_ctx(struct bpf_map *map, u32 index);
 void __dev_map_flush(struct bpf_map *map);
 
+struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key);
+void __cpu_map_insert_ctx(struct bpf_map *map, u32 index);
+void __cpu_map_flush(struct bpf_map *map);
+struct xdp_buff;
+int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
+		    struct net_device *dev_rx);
+
 /* Return map's numa specified by userspace */
 static inline int bpf_map_attr_numa_node(const union bpf_attr *attr)
 {
diff --git a/include/trace/events/xdp.h b/include/trace/events/xdp.h
index 4e16c43fba10..eb2ece96c1a2 100644
--- a/include/trace/events/xdp.h
+++ b/include/trace/events/xdp.h
@@ -136,12 +136,18 @@ DEFINE_EVENT_PRINT(xdp_redirect_template, xdp_redirect_map_err,
 		  __entry->map_id, __entry->map_index)
 );
 
+#define devmap_ifindex(fwd, map)				\
+	(!fwd ? 0 :						\
+	 (!map ? 0 :						\
+	  ((map->map_type == BPF_MAP_TYPE_DEVMAP) ?		\
+	   ((struct net_device *)fwd)->ifindex : 0)))
+
 #define _trace_xdp_redirect_map(dev, xdp, fwd, map, idx)		\
-	 trace_xdp_redirect_map(dev, xdp, fwd ? fwd->ifindex : 0,	\
+	 trace_xdp_redirect_map(dev, xdp, devmap_ifindex(fwd, map),	\
 				0, map, idx)
 
 #define _trace_xdp_redirect_map_err(dev, xdp, fwd, map, idx, err)	\
-	 trace_xdp_redirect_map_err(dev, xdp, fwd ? fwd->ifindex : 0,	\
+	 trace_xdp_redirect_map_err(dev, xdp, devmap_ifindex(fwd, map),	\
 				    err, map, idx)
 
 #endif /* _TRACE_XDP_H */
diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index f0948af82e65..ce2490ad860d 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -488,7 +488,8 @@ static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt)
 	return 0;
 }
 
-int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp)
+int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
+		    struct net_device *dev_rx)
 {
 	struct xdp_pkt *xdp_pkt;
 	int headroom;
@@ -500,7 +501,7 @@ int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp)
 	xdp_pkt = xdp->data_hard_start;
 	xdp_pkt->data = xdp->data;
 	xdp_pkt->len  = xdp->data_end - xdp->data;
-	xdp_pkt->headroom = headroom;
+	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
 	/* For now this is just used as a void pointer to data_hard_start */
 
 	bq_enqueue(rcpu, xdp_pkt);
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index f849eca36052..a712c7431c2d 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -1589,7 +1589,8 @@ static int check_map_func_compatibility(struct bpf_map *map, int func_id)
 			goto error;
 		break;
 	case BPF_FUNC_redirect_map:
-		if (map->map_type != BPF_MAP_TYPE_DEVMAP)
+		if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
+		    map->map_type != BPF_MAP_TYPE_CPUMAP)
 			goto error;
 		break;
 	case BPF_FUNC_sk_redirect_map:
diff --git a/net/core/filter.c b/net/core/filter.c
index 9b6e7e84aafd..37fe9e631ee4 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -2521,10 +2521,37 @@ static int __bpf_tx_xdp(struct net_device *dev,
 	err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
 	if (err)
 		return err;
-	if (map)
+	dev->netdev_ops->ndo_xdp_flush(dev);
+	return 0;
+}
+
+static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd,
+			    struct bpf_map *map,
+			    struct xdp_buff *xdp,
+			    u32 index)
+{
+	int err;
+
+	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
+		struct net_device *dev = fwd;
+
+		if (!dev->netdev_ops->ndo_xdp_xmit) {
+			return -EOPNOTSUPP;
+		}
+
+		err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
+		if (err)
+			return err;
 		__dev_map_insert_ctx(map, index);
-	else
-		dev->netdev_ops->ndo_xdp_flush(dev);
+
+	} else if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
+		struct bpf_cpu_map_entry *rcpu = fwd;
+
+		err = cpu_map_enqueue(rcpu, xdp, dev_rx);
+		if (err)
+			return err;
+		__cpu_map_insert_ctx(map, index);
+	}
 	return 0;
 }
 
@@ -2534,11 +2561,33 @@ void xdp_do_flush_map(void)
 	struct bpf_map *map = ri->map_to_flush;
 
 	ri->map_to_flush = NULL;
-	if (map)
-		__dev_map_flush(map);
+	if (map) {
+		switch (map->map_type) {
+		case BPF_MAP_TYPE_DEVMAP:
+			__dev_map_flush(map);
+			break;
+		case BPF_MAP_TYPE_CPUMAP:
+			__cpu_map_flush(map);
+			break;
+		default:
+			break;
+		}
+	}
 }
 EXPORT_SYMBOL_GPL(xdp_do_flush_map);
 
+static void *__xdp_map_lookup_elem(struct bpf_map *map, u32 index)
+{
+	switch (map->map_type) {
+	case BPF_MAP_TYPE_DEVMAP:
+		return __dev_map_lookup_elem(map, index);
+	case BPF_MAP_TYPE_CPUMAP:
+		return __cpu_map_lookup_elem(map, index);
+	default:
+		return NULL;
+	}
+}
+
 static inline bool xdp_map_invalid(const struct bpf_prog *xdp_prog,
 				   unsigned long aux)
 {
@@ -2551,8 +2600,8 @@ static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
 	struct redirect_info *ri = this_cpu_ptr(&redirect_info);
 	unsigned long map_owner = ri->map_owner;
 	struct bpf_map *map = ri->map;
-	struct net_device *fwd = NULL;
 	u32 index = ri->ifindex;
+	void *fwd = NULL;
 	int err;
 
 	ri->ifindex = 0;
@@ -2565,7 +2614,7 @@ static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
 		goto err;
 	}
 
-	fwd = __dev_map_lookup_elem(map, index);
+	fwd = __xdp_map_lookup_elem(map, index);
 	if (!fwd) {
 		err = -EINVAL;
 		goto err;
@@ -2573,7 +2622,7 @@ static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
 	if (ri->map_to_flush && ri->map_to_flush != map)
 		xdp_do_flush_map();
 
-	err = __bpf_tx_xdp(fwd, map, xdp, index);
+	err = __bpf_tx_xdp_map(dev, fwd, map, xdp, index);
 	if (unlikely(err))
 		goto err;
 

[net-next PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Jesper Dangaard Brouer]

This patch makes cpumap functional, by adding SKB allocation and
invoking the network stack on the dequeuing CPU.

For constructing the SKB on the remote CPU, the xdp_buff in converted
into a struct xdp_pkt, and it mapped into the top headroom of the
packet, to avoid allocating separate mem.  For now, struct xdp_pkt is
just a cpumap internal data structure, with info carried between
enqueue to dequeue.

If a driver doesn't have enough headroom it is simply dropped, with
return code -EOVERFLOW.  This will be picked up the xdp tracepoint
infrastructure, to allow users to catch this.

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 kernel/bpf/cpumap.c |  153 ++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 132 insertions(+), 21 deletions(-)

diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index ce2490ad860d..352cc071c9cc 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -24,6 +24,9 @@
 #include <linux/workqueue.h>
 #include <linux/kthread.h>
 
+#include <linux/netdevice.h>   /* netif_receive_skb */
+#include <linux/etherdevice.h> /* eth_type_trans */
+
 /*
  * General idea: XDP packets getting XDP redirected to another CPU,
  * will maximum be stored/queued for one driver ->poll() call.  It is
@@ -160,20 +163,139 @@ static void cpu_map_kthread_stop(struct work_struct *work)
 	kthread_stop(rcpu->kthread); /* calls put_cpu_map_entry */
 }
 
+/* For now, xdp_pkt is a cpumap internal data structure, with info
+ * carried between enqueue to dequeue. It is mapped into the top
+ * headroom of the packet, to avoid allocating separate mem.
+ */
+struct xdp_pkt {
+	void *data;
+	u16 len;
+	u16 headroom;
+	struct net_device *dev_rx;
+};
+
+/* Convert xdp_buff to xdp_pkt */
+static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
+{
+	struct xdp_pkt *xdp_pkt;
+	int headroom;
+
+	/* Assure headroom is available for storing info */
+	headroom = xdp->data - xdp->data_hard_start;
+	if (headroom < sizeof(*xdp_pkt))
+		return NULL;
+
+	/* Store info in top of packet */
+	xdp_pkt = xdp->data_hard_start;
+
+	xdp_pkt->data = xdp->data;
+	xdp_pkt->len  = xdp->data_end - xdp->data;
+	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
+
+	return xdp_pkt;
+}
+
+static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
+					 struct xdp_pkt *xdp_pkt)
+{
+	unsigned int frame_size;
+	void *pkt_data_start;
+	struct sk_buff *skb;
+
+	/* build_skb need to place skb_shared_info after SKB end, and
+	 * also want to know the memory "truesize".  Thus, need to
+	 * know the memory frame size backing xdp_buff.
+	 *
+	 * XDP was designed to have PAGE_SIZE frames, but this
+	 * assumption is not longer true with ixgbe and i40e.  It
+	 * would be preferred to set frame_size to 2048 or 4096
+	 * depending on the driver.
+	 *   frame_size = 2048;
+	 *   frame_len  = frame_size - sizeof(*xdp_pkt);
+	 *
+	 * Instead, with info avail, skb_shared_info in placed after
+	 * packet len.  This, unfortunately fakes the truesize.
+	 * Another disadvantage of this approach, the skb_shared_info
+	 * is not at a fixed memory location, with mixed length
+	 * packets, which is bad for cache-line hotness.
+	 */
+	frame_size = SKB_DATA_ALIGN(xdp_pkt->len) + xdp_pkt->headroom +
+		SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+
+	pkt_data_start = xdp_pkt->data - xdp_pkt->headroom;
+	skb = build_skb(pkt_data_start, frame_size);
+	if (!skb)
+		return NULL;
+
+	skb_reserve(skb, xdp_pkt->headroom);
+	__skb_put(skb, xdp_pkt->len);
+
+	/* Essential SKB info: protocol and skb->dev */
+	skb->protocol = eth_type_trans(skb, xdp_pkt->dev_rx);
+
+	/* Optional SKB info, currently missing:
+	 * - HW checksum info		(skb->ip_summed)
+	 * - HW RX hash			(skb_set_hash)
+	 * - RX ring dev queue index	(skb_record_rx_queue)
+	 */
+
+	return skb;
+}
+
 static int cpu_map_kthread_run(void *data)
 {
+	const unsigned long busy_poll_jiffies = usecs_to_jiffies(2000);
+	unsigned long time_limit = jiffies + busy_poll_jiffies;
 	struct bpf_cpu_map_entry *rcpu = data;
+	unsigned int empty_cnt = 0;
 
 	set_current_state(TASK_INTERRUPTIBLE);
 	while (!kthread_should_stop()) {
+		unsigned int processed = 0, drops = 0;
 		struct xdp_pkt *xdp_pkt;
 
-		schedule();
-		/* Do work */
-		while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
-			/* For now just "refcnt-free" */
-			page_frag_free(xdp_pkt);
+		/* Release CPU reschedule checks */
+		if ((time_after_eq(jiffies, time_limit) || empty_cnt > 25) &&
+		    __ptr_ring_empty(rcpu->queue)) {
+			empty_cnt++;
+			schedule();
+			time_limit = jiffies + busy_poll_jiffies;
+			WARN_ON(smp_processor_id() != rcpu->cpu);
+		} else {
+			cond_resched();
 		}
+
+		/* Process packets in rcpu->queue */
+		local_bh_disable();
+		/*
+		 * The bpf_cpu_map_entry is single consumer, with this
+		 * kthread CPU pinned. Lockless access to ptr_ring
+		 * consume side valid as no-resize allowed of queue.
+		 */
+		while ((xdp_pkt = __ptr_ring_consume(rcpu->queue))) {
+			struct sk_buff *skb;
+			int ret;
+
+			/* Allow busy polling again */
+			empty_cnt = 0;
+
+			skb = cpu_map_build_skb(rcpu, xdp_pkt);
+			if (!skb) {
+				page_frag_free(xdp_pkt);
+				continue;
+			}
+
+			/* Inject into network stack */
+			ret = netif_receive_skb(skb);
+			if (ret == NET_RX_DROP)
+				drops++;
+
+			/* Limit BH-disable period */
+			if (++processed == 8)
+				break;
+		}
+		local_bh_enable();
+
 		__set_current_state(TASK_INTERRUPTIBLE);
 	}
 	put_cpu_map_entry(rcpu);
@@ -458,13 +580,6 @@ static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
 	return 0;
 }
 
-/* Notice: Will change in later patch */
-struct xdp_pkt {
-	void *data;
-	u16 len;
-	u16 headroom;
-};
-
 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
  * Thus, safe percpu variable access.
  */
@@ -492,17 +607,13 @@ int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
 		    struct net_device *dev_rx)
 {
 	struct xdp_pkt *xdp_pkt;
-	int headroom;
 
-	/* Convert xdp_buff to xdp_pkt */
-	headroom = xdp->data - xdp->data_hard_start;
-	if (headroom < sizeof(*xdp_pkt))
+	xdp_pkt = convert_to_xdp_pkt(xdp);
+	if (!xdp_pkt)
 		return -EOVERFLOW;
-	xdp_pkt = xdp->data_hard_start;
-	xdp_pkt->data = xdp->data;
-	xdp_pkt->len  = xdp->data_end - xdp->data;
-	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
-	/* For now this is just used as a void pointer to data_hard_start */
+
+	/* Info needed when constructing SKB on remote CPU */
+	xdp_pkt->dev_rx = dev_rx;
 
 	bq_enqueue(rcpu, xdp_pkt);
 	return 0;

Re: [net-next PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Daniel Borkmann]

On 09/28/2017 02:57 PM, Jesper Dangaard Brouer wrote:
[...]
> +/* Convert xdp_buff to xdp_pkt */
> +static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
> +{
> +	struct xdp_pkt *xdp_pkt;
> +	int headroom;
> +
> +	/* Assure headroom is available for storing info */
> +	headroom = xdp->data - xdp->data_hard_start;
> +	if (headroom < sizeof(*xdp_pkt))
> +		return NULL;
> +
> +	/* Store info in top of packet */
> +	xdp_pkt = xdp->data_hard_start;

(You'd also need to handle data_meta here if set, and for below
cpu_map_build_skb(), e.g. headroom is data_meta-data_hard_start.)

> +	xdp_pkt->data = xdp->data;
> +	xdp_pkt->len  = xdp->data_end - xdp->data;
> +	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
> +
> +	return xdp_pkt;
> +}
> +
> +static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
> +					 struct xdp_pkt *xdp_pkt)
> +{
> +	unsigned int frame_size;
> +	void *pkt_data_start;
> +	struct sk_buff *skb;
> +
> +	/* build_skb need to place skb_shared_info after SKB end, and
> +	 * also want to know the memory "truesize".  Thus, need to
[...]
>   static int cpu_map_kthread_run(void *data)
>   {
> +	const unsigned long busy_poll_jiffies = usecs_to_jiffies(2000);
> +	unsigned long time_limit = jiffies + busy_poll_jiffies;
>   	struct bpf_cpu_map_entry *rcpu = data;
> +	unsigned int empty_cnt = 0;
>
>   	set_current_state(TASK_INTERRUPTIBLE);
>   	while (!kthread_should_stop()) {
> +		unsigned int processed = 0, drops = 0;
>   		struct xdp_pkt *xdp_pkt;
>
> -		schedule();
> -		/* Do work */
> -		while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
> -			/* For now just "refcnt-free" */
> -			page_frag_free(xdp_pkt);
> +		/* Release CPU reschedule checks */
> +		if ((time_after_eq(jiffies, time_limit) || empty_cnt > 25) &&
> +		    __ptr_ring_empty(rcpu->queue)) {
> +			empty_cnt++;
> +			schedule();
> +			time_limit = jiffies + busy_poll_jiffies;
> +			WARN_ON(smp_processor_id() != rcpu->cpu);
> +		} else {
> +			cond_resched();
>   		}
> +
> +		/* Process packets in rcpu->queue */
> +		local_bh_disable();
> +		/*
> +		 * The bpf_cpu_map_entry is single consumer, with this
> +		 * kthread CPU pinned. Lockless access to ptr_ring
> +		 * consume side valid as no-resize allowed of queue.
> +		 */
> +		while ((xdp_pkt = __ptr_ring_consume(rcpu->queue))) {
> +			struct sk_buff *skb;
> +			int ret;
> +
> +			/* Allow busy polling again */
> +			empty_cnt = 0;
> +
> +			skb = cpu_map_build_skb(rcpu, xdp_pkt);
> +			if (!skb) {
> +				page_frag_free(xdp_pkt);
> +				continue;
> +			}
> +
> +			/* Inject into network stack */
> +			ret = netif_receive_skb(skb);

Have you looked into whether it's feasible to reuse GRO
engine here as well?

> +			if (ret == NET_RX_DROP)
> +				drops++;
> +
> +			/* Limit BH-disable period */
> +			if (++processed == 8)
> +				break;
> +		}
> +		local_bh_enable();
> +
>   		__set_current_state(TASK_INTERRUPTIBLE);
>   	}
>   	put_cpu_map_entry(rcpu);
[...]

Re: [net-next PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Jesper Dangaard Brouer]

On Fri, 29 Sep 2017 01:21:01 +0200
Daniel Borkmann <daniel@iogearbox.net> wrote:

> On 09/28/2017 02:57 PM, Jesper Dangaard Brouer wrote:
> [...]
> > +/* Convert xdp_buff to xdp_pkt */
> > +static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
> > +{
> > +	struct xdp_pkt *xdp_pkt;
> > +	int headroom;
> > +
> > +	/* Assure headroom is available for storing info */
> > +	headroom = xdp->data - xdp->data_hard_start;
> > +	if (headroom < sizeof(*xdp_pkt))
> > +		return NULL;
> > +
> > +	/* Store info in top of packet */
> > +	xdp_pkt = xdp->data_hard_start;  
> 
> (You'd also need to handle data_meta here if set, and for below
> cpu_map_build_skb(), e.g. headroom is data_meta-data_hard_start.)

I'll look into this.  The data_meta patchset was in-flight while I
rebased this.

> > +	xdp_pkt->data = xdp->data;
> > +	xdp_pkt->len  = xdp->data_end - xdp->data;
> > +	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
> > +
> > +	return xdp_pkt;
> > +}
> > +
> > +static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
> > +					 struct xdp_pkt *xdp_pkt)
> > +{
> > +	unsigned int frame_size;
> > +	void *pkt_data_start;
> > +	struct sk_buff *skb;
> > +
> > +	/* build_skb need to place skb_shared_info after SKB end, and
> > +	 * also want to know the memory "truesize".  Thus, need to  
> [...]
> >   static int cpu_map_kthread_run(void *data)
> >   {
> > +	const unsigned long busy_poll_jiffies = usecs_to_jiffies(2000);
> > +	unsigned long time_limit = jiffies + busy_poll_jiffies;
> >   	struct bpf_cpu_map_entry *rcpu = data;
> > +	unsigned int empty_cnt = 0;
> >
> >   	set_current_state(TASK_INTERRUPTIBLE);
> >   	while (!kthread_should_stop()) {
> > +		unsigned int processed = 0, drops = 0;
> >   		struct xdp_pkt *xdp_pkt;
> >
> > -		schedule();
> > -		/* Do work */
> > -		while ((xdp_pkt = ptr_ring_consume(rcpu->queue))) {
> > -			/* For now just "refcnt-free" */
> > -			page_frag_free(xdp_pkt);
> > +		/* Release CPU reschedule checks */
> > +		if ((time_after_eq(jiffies, time_limit) || empty_cnt > 25) &&
> > +		    __ptr_ring_empty(rcpu->queue)) {
> > +			empty_cnt++;
> > +			schedule();
> > +			time_limit = jiffies + busy_poll_jiffies;
> > +			WARN_ON(smp_processor_id() != rcpu->cpu);
> > +		} else {
> > +			cond_resched();
> >   		}
> > +
> > +		/* Process packets in rcpu->queue */
> > +		local_bh_disable();
> > +		/*
> > +		 * The bpf_cpu_map_entry is single consumer, with this
> > +		 * kthread CPU pinned. Lockless access to ptr_ring
> > +		 * consume side valid as no-resize allowed of queue.
> > +		 */
> > +		while ((xdp_pkt = __ptr_ring_consume(rcpu->queue))) {
> > +			struct sk_buff *skb;
> > +			int ret;
> > +
> > +			/* Allow busy polling again */
> > +			empty_cnt = 0;
> > +
> > +			skb = cpu_map_build_skb(rcpu, xdp_pkt);
> > +			if (!skb) {
> > +				page_frag_free(xdp_pkt);
> > +				continue;
> > +			}
> > +
> > +			/* Inject into network stack */
> > +			ret = netif_receive_skb(skb);  
> 
> Have you looked into whether it's feasible to reuse GRO
> engine here as well?

This is the first step. I'll work on adding the GRO-engine later. And
it should be feasible.  There are plenty of optimizations in this area
that can do done later ;-)

> 
> > +			if (ret == NET_RX_DROP)
> > +				drops++;
> > +
> > +			/* Limit BH-disable period */
> > +			if (++processed == 8)
> > +				break;
> > +		}
> > +		local_bh_enable();
> > +
> >   		__set_current_state(TASK_INTERRUPTIBLE);
> >   	}
> >   	put_cpu_map_entry(rcpu);  
> [...]



-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

Re: [net-next PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Jason Wang]

On 2017年09月28日 20:57, Jesper Dangaard Brouer wrote:
> +};
> +
> +/* Convert xdp_buff to xdp_pkt */
> +static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
> +{
> +	struct xdp_pkt *xdp_pkt;
> +	int headroom;
> +
> +	/* Assure headroom is available for storing info */
> +	headroom = xdp->data - xdp->data_hard_start;
> +	if (headroom < sizeof(*xdp_pkt))
> +		return NULL;

Hi Jesper:

Do you consider this as a trick or a long term solution? Is it better to 
store XDP in a circular buffer? (I'm asking since I meet similar issue 
when doing xdp_xmit for tun).

> +
> +	/* Store info in top of packet */
> +	xdp_pkt = xdp->data_hard_start;
> +
> +	xdp_pkt->data = xdp->data;
> +	xdp_pkt->len  = xdp->data_end - xdp->data;
> +	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
> +

Is wmb() needed here?

> +	return xdp_pkt;
> +}

Thanks

Re: [net-next PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Jesper Dangaard Brouer]

On Fri, 29 Sep 2017 17:49:23 +0800
Jason Wang <jasowang@redhat.com> wrote:

> On 2017年09月28日 20:57, Jesper Dangaard Brouer wrote:
> > +};
> > +
> > +/* Convert xdp_buff to xdp_pkt */
> > +static struct xdp_pkt *convert_to_xdp_pkt(struct xdp_buff *xdp)
> > +{
> > +	struct xdp_pkt *xdp_pkt;
> > +	int headroom;
> > +
> > +	/* Assure headroom is available for storing info */
> > +	headroom = xdp->data - xdp->data_hard_start;
> > +	if (headroom < sizeof(*xdp_pkt))
> > +		return NULL;  
> 
> Hi Jesper:
> 
> Do you consider this as a trick or a long term solution? Is it better to 
> store XDP in a circular buffer? (I'm asking since I meet similar issue 
> when doing xdp_xmit for tun).

(The way you ask the question is slightly ambiguous, but I hope I understand.)

IMHO the best solution to allow queueing of XDP packets is to create a
meta-data structure, with the needed info.  For performance reasons, we
don't want to allocate a new memory area for this.  Thus, we simply use
the available headroom in the page that the packet is stored into.
Notice that DPDK also use the first cache-line of the packet data, for
its packet meta-data structure. (This is not a performance problem.
I've done several PoC benchmarks, before choosing to do this)

For now, this "trick" is local to the cpumap, and thus not exposed as
any API.  Thus we can evolve and change the contents easily.  But I
would in time, like to see this generalized. When/if more places need
to queue XDP packets, this header meta-data format should be
standardized.


Pipe-dreaming: Taking this to the extreme... if I could get away with
it, I would actually like to store the (232 bytes) SKB meta-data header
inside headroom too.  That would eliminate any real SKB memory alloc.


> > +
> > +	/* Store info in top of packet */
> > +	xdp_pkt = xdp->data_hard_start;
> > +
> > +	xdp_pkt->data = xdp->data;
> > +	xdp_pkt->len  = xdp->data_end - xdp->data;
> > +	xdp_pkt->headroom = headroom - sizeof(*xdp_pkt);
> > +  
> 
> Is wmb() needed here?

No. This xdp_pkt is queued into a into a ptr_ring, which have a
spin_lock on enqueue, and any atomic operation works as a full memory
barrirer mb().

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

[net-next PATCH 4/5] bpf: cpumap add tracepoints

[Jesper Dangaard Brouer]

This adds two tracepoint to the cpumap.  One for the enqueue side
trace_xdp_cpumap_enqueue() and one for the kthread dequeue side
trace_xdp_cpumap_kthread().

To mitigate the tracepoint overhead, these are invoked during the
enqueue/dequeue bulking phases, thus amortizing the cost.

The obvious use-cases are for debugging and monitoring.  The
non-intuitive use-case is using these as a feedback loop to know the
system load.  One can imagine auto-scaling by reducing, adding or
activating more worker CPUs on demand.

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/trace/events/xdp.h |   70 ++++++++++++++++++++++++++++++++++++++++++++
 kernel/bpf/cpumap.c        |   18 +++++++++--
 2 files changed, 85 insertions(+), 3 deletions(-)

diff --git a/include/trace/events/xdp.h b/include/trace/events/xdp.h
index eb2ece96c1a2..bc48c13892c4 100644
--- a/include/trace/events/xdp.h
+++ b/include/trace/events/xdp.h
@@ -150,6 +150,76 @@ DEFINE_EVENT_PRINT(xdp_redirect_template, xdp_redirect_map_err,
 	 trace_xdp_redirect_map_err(dev, xdp, devmap_ifindex(fwd, map),	\
 				    err, map, idx)
 
+TRACE_EVENT(xdp_cpumap_kthread,
+
+	TP_PROTO(int map_id, unsigned int processed,  unsigned int drops,
+		 int time_limit),
+
+	TP_ARGS(map_id, processed, drops, time_limit),
+
+	TP_STRUCT__entry(
+		__field(int, map_id)
+		__field(u32, act)
+		__field(int, cpu)
+		__field(unsigned int, drops)
+		__field(unsigned int, processed)
+		__field(int, time_limit)
+	),
+
+	TP_fast_assign(
+		__entry->map_id		= map_id;
+		__entry->act		= XDP_REDIRECT;
+		__entry->cpu		= smp_processor_id();
+		__entry->drops		= drops;
+		__entry->processed	= processed;
+		__entry->time_limit	= time_limit;
+	),
+
+	TP_printk("kthread"
+		  " cpu=%d map_id=%d action=%s"
+		  " processed=%u drops=%u"
+		  " time_limit=%d",
+		  __entry->cpu, __entry->map_id,
+		  __print_symbolic(__entry->act, __XDP_ACT_SYM_TAB),
+		  __entry->processed, __entry->drops,
+		  __entry->time_limit)
+);
+
+TRACE_EVENT(xdp_cpumap_enqueue,
+
+	TP_PROTO(int map_id, unsigned int processed,  unsigned int drops,
+		 int to_cpu),
+
+	TP_ARGS(map_id, processed, drops, to_cpu),
+
+	TP_STRUCT__entry(
+		__field(int, map_id)
+		__field(u32, act)
+		__field(int, cpu)
+		__field(unsigned int, drops)
+		__field(unsigned int, processed)
+		__field(int, to_cpu)
+	),
+
+	TP_fast_assign(
+		__entry->map_id		= map_id;
+		__entry->act		= XDP_REDIRECT;
+		__entry->cpu		= smp_processor_id();
+		__entry->drops		= drops;
+		__entry->processed	= processed;
+		__entry->to_cpu		= to_cpu;
+	),
+
+	TP_printk("enqueue"
+		  " cpu=%d map_id=%d action=%s"
+		  " processed=%u drops=%u"
+		  " to_cpu=%d",
+		  __entry->cpu, __entry->map_id,
+		  __print_symbolic(__entry->act, __XDP_ACT_SYM_TAB),
+		  __entry->processed, __entry->drops,
+		  __entry->to_cpu)
+);
+
 #endif /* _TRACE_XDP_H */
 
 #include <trace/define_trace.h>
diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index 352cc071c9cc..3b0288e4e998 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -23,6 +23,7 @@
 #include <linux/sched.h>
 #include <linux/workqueue.h>
 #include <linux/kthread.h>
+#include <trace/events/xdp.h>
 
 #include <linux/netdevice.h>   /* netif_receive_skb */
 #include <linux/etherdevice.h> /* eth_type_trans */
@@ -294,6 +295,9 @@ static int cpu_map_kthread_run(void *data)
 			if (++processed == 8)
 				break;
 		}
+		/* Feedback loop via tracepoint */
+		trace_xdp_cpumap_kthread(rcpu->map_id, processed, drops,
+					 time_after_eq(jiffies, time_limit));
 		local_bh_enable();
 
 		__set_current_state(TASK_INTERRUPTIBLE);
@@ -331,7 +335,10 @@ struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id)
 	err = ptr_ring_init(rcpu->queue, qsize, gfp);
 	if (err)
 		goto fail;
-	rcpu->qsize = qsize;
+
+	rcpu->cpu    = cpu;
+	rcpu->map_id = map_id;
+	rcpu->qsize  = qsize;
 
 	/* Setup kthread */
 	rcpu->kthread = kthread_create_on_node(cpu_map_kthread_run, rcpu, numa,
@@ -555,6 +562,8 @@ const struct bpf_map_ops cpu_map_ops = {
 static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
 			     struct xdp_bulk_queue *bq)
 {
+	unsigned int processed = 0, drops = 0;
+	const int to_cpu = rcpu->cpu;
 	struct ptr_ring *q;
 	int i;
 
@@ -570,13 +579,16 @@ static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
 
 		err = __ptr_ring_produce(q, xdp_pkt);
 		if (err) {
-			/* Free xdp_pkt */
-			page_frag_free(xdp_pkt);
+			drops++;
+			page_frag_free(xdp_pkt); /* Free xdp_pkt */
 		}
+		processed++;
 	}
 	bq->count = 0;
 	spin_unlock(&q->producer_lock);
 
+	/* Feedback loop via tracepoints */
+	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 	return 0;
 }
 

[net-next PATCH 5/5] samples/bpf: add cpumap sample program xdp_redirect_cpu

[Jesper Dangaard Brouer]

This sample program show how to use cpumap and the associated
tracepoints.

It provides command line stats, which shows how the XDP-RX process,
cpumap-enqueue and cpumap kthread dequeue is cooperating on a per CPU
basis.  It also utilize the xdp_exception and xdp_redirect_err
transpoints to allow users quickly to identify setup issues.

One issue with ixgbe driver is that the driver reset the link when
loading XDP.  This reset the procfs smp_affinity settings.  Thus,
after loading the program, these must be reconfigured.  The easiest
workaround it to reduce the RX-queue to e.g. two via:

 # ethtool --set-channels ixgbe1 combined 2

And then add CPUs above 0 and 1, like:

 # xdp_redirect_cpu --dev ixgbe1 --prog 2 --cpu 2 --cpu 3 --cpu 4

Another issue with ixgbe is that the page recycle mechanism is tied to
the RX-ring size.  And the default setting of 512 elements is too
small.  This is the same issue with regular devmap XDP_REDIRECT.
To overcome this I've been using 1024 rx-ring size:

 # ethtool -G ixgbe1 rx 1024 tx 1024

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 samples/bpf/Makefile                |    4 
 samples/bpf/xdp_redirect_cpu_kern.c |  640 +++++++++++++++++++++++++++++++++++
 samples/bpf/xdp_redirect_cpu_user.c |  639 +++++++++++++++++++++++++++++++++++
 3 files changed, 1283 insertions(+)
 create mode 100644 samples/bpf/xdp_redirect_cpu_kern.c
 create mode 100644 samples/bpf/xdp_redirect_cpu_user.c

diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile
index ebc2ad69b62c..52c4dab2c153 100644
--- a/samples/bpf/Makefile
+++ b/samples/bpf/Makefile
@@ -39,6 +39,7 @@ hostprogs-y += per_socket_stats_example
 hostprogs-y += load_sock_ops
 hostprogs-y += xdp_redirect
 hostprogs-y += xdp_redirect_map
+hostprogs-y += xdp_redirect_cpu
 hostprogs-y += xdp_monitor
 hostprogs-y += syscall_tp
 
@@ -84,6 +85,7 @@ test_map_in_map-objs := bpf_load.o $(LIBBPF) test_map_in_map_user.o
 per_socket_stats_example-objs := $(LIBBPF) cookie_uid_helper_example.o
 xdp_redirect-objs := bpf_load.o $(LIBBPF) xdp_redirect_user.o
 xdp_redirect_map-objs := bpf_load.o $(LIBBPF) xdp_redirect_map_user.o
+xdp_redirect_cpu-objs := bpf_load.o $(LIBBPF) xdp_redirect_cpu_user.o
 xdp_monitor-objs := bpf_load.o $(LIBBPF) xdp_monitor_user.o
 syscall_tp-objs := bpf_load.o $(LIBBPF) syscall_tp_user.o
 
@@ -129,6 +131,7 @@ always += tcp_iw_kern.o
 always += tcp_clamp_kern.o
 always += xdp_redirect_kern.o
 always += xdp_redirect_map_kern.o
+always += xdp_redirect_cpu_kern.o
 always += xdp_monitor_kern.o
 always += syscall_tp_kern.o
 
@@ -169,6 +172,7 @@ HOSTLOADLIBES_xdp_tx_iptunnel += -lelf
 HOSTLOADLIBES_test_map_in_map += -lelf
 HOSTLOADLIBES_xdp_redirect += -lelf
 HOSTLOADLIBES_xdp_redirect_map += -lelf
+HOSTLOADLIBES_xdp_redirect_cpu += -lelf
 HOSTLOADLIBES_xdp_monitor += -lelf
 HOSTLOADLIBES_syscall_tp += -lelf
 
diff --git a/samples/bpf/xdp_redirect_cpu_kern.c b/samples/bpf/xdp_redirect_cpu_kern.c
new file mode 100644
index 000000000000..7403e7841a88
--- /dev/null
+++ b/samples/bpf/xdp_redirect_cpu_kern.c
@@ -0,0 +1,640 @@
+/*  XDP redirect to CPUs via cpumap (BPF_MAP_TYPE_CPUMAP)
+ *
+ *  GPLv2, Copyright(c) 2017 Jesper Dangaard Brouer, Red Hat, Inc.
+ */
+#include <uapi/linux/if_ether.h>
+#include <uapi/linux/if_packet.h>
+#include <uapi/linux/if_vlan.h>
+#include <uapi/linux/ip.h>
+#include <uapi/linux/ipv6.h>
+#include <uapi/linux/in.h>
+#include <uapi/linux/tcp.h>
+#include <uapi/linux/udp.h>
+
+#include <uapi/linux/bpf.h>
+#include "bpf_helpers.h"
+
+#define MAX_CPUS 12 /* WARNING - sync with _user.c */
+
+/* Special map type that can XDP_REDIRECT frames to another CPU */
+struct bpf_map_def SEC("maps") cpu_map = {
+	.type		= BPF_MAP_TYPE_CPUMAP,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(u32),
+	.max_entries	= MAX_CPUS,
+};
+
+/* Common stats data record to keep userspace more simple */
+struct datarec {
+	__u64 processed;
+	__u64 dropped;
+	__u64 issue;
+};
+
+/* Count RX packets, as XDP bpf_prog doesn't get direct TX-success
+ * feedback.  Redirect TX errors can be caught via a tracepoint.
+ */
+struct bpf_map_def SEC("maps") rx_cnt = {
+	.type		= BPF_MAP_TYPE_PERCPU_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(struct datarec),
+	.max_entries	= 1,
+};
+
+/* Used by trace point */
+struct bpf_map_def SEC("maps") redirect_err_cnt = {
+	.type		= BPF_MAP_TYPE_PERCPU_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(struct datarec),
+	.max_entries	= 2,
+	/* TODO: have entries for all possible errno's */
+};
+
+/* Used by trace point */
+struct bpf_map_def SEC("maps") cpumap_enqueue_cnt = {
+	.type		= BPF_MAP_TYPE_PERCPU_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(struct datarec),
+	.max_entries	= MAX_CPUS,
+};
+
+/* Used by trace point */
+struct bpf_map_def SEC("maps") cpumap_kthread_cnt = {
+	.type		= BPF_MAP_TYPE_PERCPU_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(struct datarec),
+	.max_entries	= 1,
+};
+
+/* Set of maps controlling available CPU, and for iterating through
+ * selectable redirect CPUs.
+ */
+struct bpf_map_def SEC("maps") cpus_available = {
+	.type		= BPF_MAP_TYPE_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(u32),
+	.max_entries	= MAX_CPUS,
+};
+struct bpf_map_def SEC("maps") cpus_count = {
+	.type		= BPF_MAP_TYPE_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(u32),
+	.max_entries	= 1,
+};
+struct bpf_map_def SEC("maps") cpus_iterator = {
+	.type		= BPF_MAP_TYPE_PERCPU_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(u32),
+	.max_entries	= 1,
+};
+
+/* Used by trace point */
+struct bpf_map_def SEC("maps") exception_cnt = {
+	.type		= BPF_MAP_TYPE_PERCPU_ARRAY,
+	.key_size	= sizeof(u32),
+	.value_size	= sizeof(struct datarec),
+	.max_entries	= 1,
+};
+
+/* Helper parse functions */
+
+/* Parse Ethernet layer 2, extract network layer 3 offset and protocol
+ *
+ * Returns false on error and non-supported ether-type
+ */
+struct vlan_hdr {
+	__be16 h_vlan_TCI;
+	__be16 h_vlan_encapsulated_proto;
+};
+
+static __always_inline
+bool parse_eth(struct ethhdr *eth, void *data_end,
+	       u16 *eth_proto, u64 *l3_offset)
+{
+	u16 eth_type;
+	u64 offset;
+
+	offset = sizeof(*eth);
+	if ((void *)eth + offset > data_end)
+		return false;
+
+	eth_type = eth->h_proto;
+
+	/* Skip non 802.3 Ethertypes */
+	if (unlikely(ntohs(eth_type) < ETH_P_802_3_MIN))
+		return false;
+
+	/* Handle VLAN tagged packet */
+	if (eth_type == htons(ETH_P_8021Q) || eth_type == htons(ETH_P_8021AD)) {
+		struct vlan_hdr *vlan_hdr;
+
+		vlan_hdr = (void *)eth + offset;
+		offset += sizeof(*vlan_hdr);
+		if ((void *)eth + offset > data_end)
+			return false;
+		eth_type = vlan_hdr->h_vlan_encapsulated_proto;
+	}
+	/* TODO: Handle double VLAN tagged packet */
+
+	*eth_proto = ntohs(eth_type);
+	*l3_offset = offset;
+	return true;
+}
+
+static __always_inline
+u16 get_dest_port_ipv4_udp(struct xdp_md *ctx, u64 nh_off)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+        struct iphdr *iph = data + nh_off;
+	struct udphdr *udph;
+	u16 dport;
+
+        if (iph + 1 > data_end)
+                return 0;
+	if (!(iph->protocol == IPPROTO_UDP))
+		return 0;
+
+	udph = (void *)(iph + 1);
+	if (udph + 1 > data_end)
+		return 0;
+
+	dport = ntohs(udph->dest);
+	return dport;
+}
+
+static __always_inline
+int get_proto_ipv4(struct xdp_md *ctx, u64 nh_off)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+        struct iphdr *iph = data + nh_off;
+
+        if (iph + 1 > data_end)
+                return 0;
+        return iph->protocol;
+}
+
+static __always_inline
+int get_proto_ipv6(struct xdp_md *ctx, u64 nh_off)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+        struct ipv6hdr *ip6h = data + nh_off;
+
+        if (ip6h + 1 > data_end)
+                return 0;
+        return ip6h->nexthdr;
+}
+
+SEC("xdp_cpu_map0")
+int  xdp_prognum0_no_touch(struct xdp_md *ctx)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+	struct datarec* rec;
+	u32 cpu_dest;
+	u32 key = 0;
+
+	/* Only use first entry in cpus_available */
+	u32 *cpu_selected;
+	cpu_selected = bpf_map_lookup_elem(&cpus_available, &key);
+	if (!cpu_selected)
+		return XDP_ABORTED;
+	cpu_dest = *cpu_selected;
+
+	/* Count RX packet in map */
+	rec = bpf_map_lookup_elem(&rx_cnt, &key);
+	if (rec)
+		rec->processed++;
+
+	return bpf_redirect_map(&cpu_map, cpu_dest, 0);
+}
+
+SEC("xdp_cpu_map1_touch_data")
+int  xdp_prognum1_touch_data(struct xdp_md *ctx)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+	struct ethhdr *eth = data;
+	volatile u16 eth_type;
+	struct datarec* rec;
+	u32 cpu_dest;
+	u32 key = 0;
+
+	/* Only use first entry in cpus_available */
+	u32 *cpu_selected;
+	cpu_selected = bpf_map_lookup_elem(&cpus_available, &key);
+	if (!cpu_selected)
+		return XDP_ABORTED;
+	cpu_dest = *cpu_selected;
+
+	/* Validate packet length is minimum Eth header size */
+	if (eth + 1 > data_end) {
+		return XDP_ABORTED;
+	}
+
+	/* Count RX packet in map */
+	rec = bpf_map_lookup_elem(&rx_cnt, &key);
+	if (!rec)
+		return XDP_ABORTED;
+	rec->processed++;
+
+	/* Read packet data, and use it (drop non 802.3 Ethertypes) */
+	eth_type = eth->h_proto;
+	if (ntohs(eth_type) < ETH_P_802_3_MIN) {
+		rec->dropped++;
+		return XDP_DROP;
+	}
+
+	return bpf_redirect_map(&cpu_map, cpu_dest, 0);
+}
+
+SEC("xdp_cpu_map2_round_robin")
+int  xdp_prognum2_round_robin(struct xdp_md *ctx)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+	struct ethhdr *eth = data;
+	struct datarec* rec;
+	u32 cpu_dest;
+	u32 *cpu_lookup;
+	u32 key0 = 0;
+
+	u32 *cpu_selected;
+	u32 *cpu_iterator;
+	u32 *cpu_max;
+	u32 cpu_idx;
+
+	cpu_max = bpf_map_lookup_elem(&cpus_count, &key0);
+	if (!cpu_max)
+		return XDP_ABORTED;
+
+	cpu_iterator = bpf_map_lookup_elem(&cpus_iterator, &key0);
+	if (!cpu_iterator)
+		return XDP_ABORTED;
+	cpu_idx = *cpu_iterator;
+
+	*cpu_iterator += 1;
+	if (*cpu_iterator == *cpu_max)
+		*cpu_iterator = 0;
+
+	cpu_selected = bpf_map_lookup_elem(&cpus_available, &cpu_idx);
+	if (!cpu_selected)
+		return XDP_ABORTED;
+	cpu_dest = *cpu_selected;
+
+	/* Count RX packet in map */
+	rec = bpf_map_lookup_elem(&rx_cnt, &key0);
+	if (!rec)
+		return XDP_ABORTED;
+	rec->processed++;
+
+	/* Check cpu_dest is valid */
+	cpu_lookup = bpf_map_lookup_elem(&cpu_map, &cpu_dest);
+	if (!cpu_lookup) {
+		rec->issue++;
+		return XDP_DROP;
+	}
+
+	if (cpu_dest >= MAX_CPUS )
+		return XDP_ABORTED;
+
+	return bpf_redirect_map(&cpu_map, cpu_dest, 0);
+}
+
+SEC("xdp_cpu_map3_proto_separate")
+int  xdp_prognum3_proto_separate(struct xdp_md *ctx)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+	struct ethhdr *eth = data;
+	u8 ip_proto = IPPROTO_UDP;
+	struct datarec* rec;
+	u16 eth_proto = 0;
+	u64 l3_offset = 0;
+	u32 cpu_dest = 0;
+	u32 cpu_idx = 0;
+	u32 *cpu_lookup;
+	u32 key = 0;
+
+	/* Count RX packet in map */
+	rec = bpf_map_lookup_elem(&rx_cnt, &key);
+	if (!rec)
+		return XDP_ABORTED;
+	rec->processed++;
+
+	if (!(parse_eth(eth, data_end, &eth_proto, &l3_offset))) {
+		return XDP_PASS; /* Just skip */
+	}
+
+	/* Extract L4 protocol */
+	switch (eth_proto) {
+	case ETH_P_IP:
+		ip_proto = get_proto_ipv4(ctx, l3_offset);
+		break;
+	case ETH_P_IPV6:
+		ip_proto = get_proto_ipv6(ctx, l3_offset);
+		break;
+	case ETH_P_ARP:
+		cpu_idx = 0; /* ARP packet handled on separate CPU */
+		break;
+	default:
+		cpu_idx = 0;
+	}
+
+	/* Choose CPU based on L4 protocol */
+	switch (ip_proto) {
+	case IPPROTO_ICMP:
+	case IPPROTO_ICMPV6:
+		cpu_idx = 2;
+		break;
+	case IPPROTO_TCP:
+		cpu_idx = 0;
+		break;
+	case IPPROTO_UDP:
+		cpu_idx = 1;
+		break;
+	default:
+		cpu_idx = 0;
+	}
+
+	cpu_lookup = bpf_map_lookup_elem(&cpus_available, &cpu_idx);
+	if (!cpu_lookup)
+		return XDP_ABORTED;
+	cpu_dest = *cpu_lookup;
+
+	if (cpu_dest >= MAX_CPUS )
+		return XDP_ABORTED;
+
+	/* Check cpu_dest is valid */
+	cpu_lookup = bpf_map_lookup_elem(&cpu_map, &cpu_dest);
+	if (!cpu_lookup) {
+		rec->issue++;
+		return XDP_DROP;
+	}
+
+	return bpf_redirect_map(&cpu_map, cpu_dest, 0);
+}
+
+SEC("xdp_cpu_map4_ddos_filter_pktgen")
+int  xdp_prognum4_ddos_filter_pktgen(struct xdp_md *ctx)
+{
+	void *data_end = (void *)(long)ctx->data_end;
+	void *data     = (void *)(long)ctx->data;
+	struct ethhdr *eth = data;
+	u8 ip_proto = IPPROTO_UDP;
+	struct datarec* rec;
+	u16 eth_proto = 0;
+	u64 l3_offset = 0;
+	u32 cpu_dest = 0;
+	u32 cpu_idx = 0;
+	u16 dest_port;
+	u32 *cpu_lookup;
+	u32 key = 0;
+
+	/* Count RX packet in map */
+	rec = bpf_map_lookup_elem(&rx_cnt, &key);
+	if (!rec)
+		return XDP_ABORTED;
+	rec->processed++;
+
+	if (!(parse_eth(eth, data_end, &eth_proto, &l3_offset))) {
+		return XDP_PASS; /* Just skip */
+	}
+
+	/* Extract L4 protocol */
+	switch (eth_proto) {
+	case ETH_P_IP:
+		ip_proto = get_proto_ipv4(ctx, l3_offset);
+		break;
+	case ETH_P_IPV6:
+		ip_proto = get_proto_ipv6(ctx, l3_offset);
+		break;
+	case ETH_P_ARP:
+		cpu_idx = 0; /* ARP packet handled on separate CPU */
+		break;
+	default:
+		cpu_idx = 0;
+	}
+
+	/* Choose CPU based on L4 protocol */
+	switch (ip_proto) {
+	case IPPROTO_ICMP:
+	case IPPROTO_ICMPV6:
+		cpu_idx = 2;
+		break;
+	case IPPROTO_TCP:
+		cpu_idx = 0;
+		break;
+	case IPPROTO_UDP:
+		cpu_idx = 1;
+		/* DDoS filter UDP port 9 (pktgen) */
+		dest_port = get_dest_port_ipv4_udp(ctx, l3_offset);
+		if (dest_port == 9) {
+			if (rec)
+				rec->dropped++;
+			return XDP_DROP;
+		}
+		break;
+	default:
+		cpu_idx = 0;
+	}
+
+	cpu_lookup = bpf_map_lookup_elem(&cpus_available, &cpu_idx);
+	if (!cpu_lookup)
+		return XDP_ABORTED;
+	cpu_dest = *cpu_lookup;
+
+	if (cpu_dest >= MAX_CPUS )
+		return XDP_ABORTED;
+
+	/* Check cpu_dest is valid */
+	cpu_lookup = bpf_map_lookup_elem(&cpu_map, &cpu_dest);
+	if (!cpu_lookup) {
+		rec->issue++;
+		return XDP_DROP;
+	}
+
+	if (cpu_dest >= MAX_CPUS )
+		return XDP_ABORTED;
+
+	return bpf_redirect_map(&cpu_map, cpu_dest, 0);
+}
+
+
+char _license[] SEC("license") = "GPL";
+
+/*** Trace point code ***/
+
+/* Tracepoint format: /sys/kernel/debug/tracing/events/xdp/xdp_redirect/format
+ * Code in:                kernel/include/trace/events/xdp.h
+ */
+struct xdp_redirect_ctx {
+	unsigned short common_type;	//	offset:0;  size:2; signed:0;
+	unsigned char common_flags;	//	offset:2;  size:1; signed:0;
+	unsigned char common_preempt_count;//	offset:3;  size:1; signed:0;
+	int common_pid;			//	offset:4;  size:4; signed:1;
+
+	int prog_id;			//	offset:8;  size:4; signed:1;
+	u32 act;			//	offset:12  size:4; signed:0;
+	int ifindex;			//	offset:16  size:4; signed:1;
+	int err;			//	offset:20  size:4; signed:1;
+	int to_ifindex;			//	offset:24  size:4; signed:1;
+	u32 map_id;			//	offset:28  size:4; signed:0;
+	int map_index;			//	offset:32  size:4; signed:1;
+};					//	offset:36
+
+enum {
+	XDP_REDIRECT_SUCCESS = 0,
+	XDP_REDIRECT_ERROR = 1
+};
+
+static __always_inline
+int xdp_redirect_collect_stat(struct xdp_redirect_ctx *ctx)
+{
+	u32 key = XDP_REDIRECT_ERROR;
+	struct datarec *rec;
+	int err = ctx->err;
+
+	if (!err)
+		key = XDP_REDIRECT_SUCCESS;
+
+	rec = bpf_map_lookup_elem(&redirect_err_cnt, &key);
+	if (!rec)
+		return 0;
+	rec->dropped += 1;
+
+	return 0; /* Indicate event was filtered (no further processing)*/
+	/*
+	 * Returning 1 here would allow e.g. a perf-record tracepoint
+	 * to see and record these events, but it doesn't work well
+	 * in-practice as stopping perf-record also unload this
+	 * bpf_prog.  Plus, there is additional overhead of doing so.
+	 */
+}
+
+SEC("tracepoint/xdp/xdp_redirect_err")
+int trace_xdp_redirect_err(struct xdp_redirect_ctx *ctx)
+{
+	return xdp_redirect_collect_stat(ctx);
+}
+
+
+SEC("tracepoint/xdp/xdp_redirect_map_err")
+int trace_xdp_redirect_map_err(struct xdp_redirect_ctx *ctx)
+{
+	return xdp_redirect_collect_stat(ctx);
+}
+
+/* Tracepoint format: /sys/kernel/debug/tracing/events/xdp/xdp_exception/format
+ * Code in:                kernel/include/trace/events/xdp.h
+ */
+struct xdp_exception_ctx {
+	unsigned short common_type;	//	offset:0;  size:2; signed:0;
+	unsigned char common_flags;	//	offset:2;  size:1; signed:0;
+	unsigned char common_preempt_count;//	offset:3;  size:1; signed:0;
+	int common_pid;			//	offset:4;  size:4; signed:1;
+
+	int prog_id;			//	offset:8;  size:4; signed:1;
+	u32 act;			//	offset:12; size:4; signed:0;
+	int ifindex;			//	offset:16; size:4; signed:1;
+};
+
+SEC("tracepoint/xdp/xdp_exception")
+int trace_xdp_exception(struct xdp_exception_ctx *ctx)
+{
+	struct datarec *rec;
+	u32 key = 0;
+
+	rec = bpf_map_lookup_elem(&exception_cnt, &key);
+	if (!rec)
+		return 1;
+	rec->dropped += 1;
+
+	return 0;
+}
+
+/* Tracepoint: /sys/kernel/debug/tracing/events/xdp/xdp_cpumap_enqueue/format
+ * Code in:         kernel/include/trace/events/xdp.h
+ */
+struct cpumap_enqueue_ctx {
+	unsigned short common_type;	//	offset:0;  size:2; signed:0;
+	unsigned char common_flags;	//	offset:2;  size:1; signed:0;
+	unsigned char common_preempt_count;//	offset:3;  size:1; signed:0;
+	int common_pid;			//	offset:4;  size:4; signed:1;
+
+	int map_id;			//	offset:8;  size:4; signed:1;
+	u32 act;			//	offset:12; size:4; signed:0;
+	int cpu;			//	offset:16; size:4; signed:1;
+	unsigned int drops;		//	offset:20; size:4; signed:0;
+	unsigned int processed;		//	offset:24; size:4; signed:0;
+	int to_cpu;			//	offset:28; size:4; signed:1;
+};
+
+SEC("tracepoint/xdp/xdp_cpumap_enqueue")
+int trace_xdp_cpumap_enqueue(struct cpumap_enqueue_ctx *ctx)
+{
+	u32 to_cpu = ctx->to_cpu;
+	struct datarec *rec;
+
+	if (to_cpu >= MAX_CPUS)
+		return 1;
+
+	rec = bpf_map_lookup_elem(&cpumap_enqueue_cnt, &to_cpu);
+	if (!rec)
+		return 0;
+	rec->processed += ctx->processed;
+	rec->dropped   += ctx->drops;
+
+	/* Detect misconfig. Redirect to "same" CPU, makes no sense
+	 * and indicate user of cpumap have not done proper IRQ RXq
+	 * setup.
+	 */
+	if (ctx->cpu == ctx->to_cpu)
+		rec->issue += ctx->processed;
+
+	/* Keep seperate map for feedback loop */
+	// have map that boolean mark drops, and RX side can clean
+	// this, indicating it have got the notification. TODO, should
+	// this also contain a (k)timestamp.
+
+	return 0;
+}
+
+/* Tracepoint: /sys/kernel/debug/tracing/events/xdp/xdp_cpumap_kthread/format
+ * Code in:         kernel/include/trace/events/xdp.h
+ */
+struct cpumap_kthread_ctx {
+	unsigned short common_type;	//	offset:0;  size:2; signed:0;
+	unsigned char common_flags;	//	offset:2;  size:1; signed:0;
+	unsigned char common_preempt_count;//	offset:3;  size:1; signed:0;
+	int common_pid;			//	offset:4;  size:4; signed:1;
+
+	int map_id;			//	offset:8;  size:4; signed:1;
+	u32 act;			//	offset:12; size:4; signed:0;
+	int cpu;			//	offset:16; size:4; signed:1;
+	unsigned int drops;		//	offset:20; size:4; signed:0;
+	unsigned int processed;		//	offset:24; size:4; signed:0;
+	int time_limit;			//	offset:28; size:4; signed:1;
+};
+
+SEC("tracepoint/xdp/xdp_cpumap_kthread")
+int trace_xdp_cpumap_kthread(struct cpumap_kthread_ctx *ctx)
+{
+	struct datarec *rec;
+	u32 key = 0;
+
+	rec = bpf_map_lookup_elem(&cpumap_kthread_cnt, &key);
+	if (!rec)
+		return 0;
+	rec->processed += ctx->processed;
+	rec->dropped   += ctx->drops;
+
+	/* Detect when time limit was exceeded, but queue was not-empty */
+	if (ctx->processed > 0 && ctx->time_limit)
+		rec->issue++;
+
+	return 0;
+}
+
diff --git a/samples/bpf/xdp_redirect_cpu_user.c b/samples/bpf/xdp_redirect_cpu_user.c
new file mode 100644
index 000000000000..c2c971ab7078
--- /dev/null
+++ b/samples/bpf/xdp_redirect_cpu_user.c
@@ -0,0 +1,639 @@
+/* GPLv2 Copyright(c) 2017 Jesper Dangaard Brouer, Red Hat, Inc.
+ */
+static const char *__doc__=
+ " XDP redirect with a CPU-map type \"BPF_MAP_TYPE_CPUMAP\"";
+
+#include <errno.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <locale.h>
+#include <sys/resource.h>
+#include <getopt.h>
+#include <net/if.h>
+#include <time.h>
+
+#include <arpa/inet.h>
+#include <linux/if_link.h>
+
+#define MAX_CPUS 12 /* WARNING - sync with _kern.c */
+
+/* How many xdp_progs are defined in _kern.c */
+#define MAX_PROG 5
+
+/* Wanted to get rid of bpf_load.h and fake-"libbpf.h" (and instead
+ * use bpf/libbpf.h), but cannot as (currently) needed for XDP
+ * attaching to a device via set_link_xdp_fd()
+ */
+#include "libbpf.h"
+#include "bpf_load.h"
+
+#include "bpf_util.h"
+
+static int ifindex = -1;
+static char ifname_buf[IF_NAMESIZE];
+static char *ifname = NULL;
+static __u32 xdp_flags = 0;
+
+/* Exit return codes */
+#define EXIT_OK			0
+#define EXIT_FAIL		1
+#define EXIT_FAIL_OPTION	2
+#define EXIT_FAIL_XDP		3
+#define EXIT_FAIL_BPF		4
+#define EXIT_FAIL_MEM		5
+
+static const struct option long_options[] = {
+	{"help",	no_argument,		NULL, 'h' },
+	{"dev",		required_argument,	NULL, 'd' },
+	{"skb-mode", 	no_argument,		NULL, 'S' },
+	{"debug",	no_argument,		NULL, 'D' },
+	{"sec", 	required_argument,	NULL, 's' },
+	{"prognum", 	required_argument,	NULL, 'p' },
+	{"qsize", 	required_argument,	NULL, 'q' },
+	{"cpu", 	required_argument,	NULL, 'c' },
+	{"no-separators",no_argument,		NULL, 'z' },
+	{0, 0, NULL,  0 }
+};
+
+static void int_exit(int sig)
+{
+	fprintf(stderr,
+		"Interrupted: Removing XDP program on ifindex:%d device:%s\n",
+		ifindex, ifname);
+	if (ifindex > -1)
+		set_link_xdp_fd(ifindex, -1, xdp_flags);
+	exit(EXIT_OK);
+}
+
+static void usage(char *argv[])
+{
+	int i;
+	printf("\nDOCUMENTATION:\n%s\n", __doc__);
+	printf("\n");
+	printf(" Usage: %s (options-see-below)\n",
+	       argv[0]);
+	printf(" Listing options:\n");
+	for (i = 0; long_options[i].name != 0; i++) {
+		printf(" --%-12s", long_options[i].name);
+		if (long_options[i].flag != NULL)
+			printf(" flag (internal value:%d)",
+			       *long_options[i].flag);
+		else
+			printf(" short-option: -%c",
+			       long_options[i].val);
+		printf("\n");
+	}
+	printf("\n");
+}
+
+/* gettime returns the current time of day in nanoseconds.
+ * Cost: clock_gettime (ns) => 26ns (CLOCK_MONOTONIC)
+ *       clock_gettime (ns) =>  9ns (CLOCK_MONOTONIC_COARSE)
+ */
+#define NANOSEC_PER_SEC 1000000000 /* 10^9 */
+static __u64 gettime(void)
+{
+	struct timespec t;
+	int res;
+
+	res = clock_gettime(CLOCK_MONOTONIC, &t);
+	if (res < 0) {
+		fprintf(stderr, "Error with gettimeofday! (%i)\n", res);
+		exit(EXIT_FAIL);
+	}
+	return (__u64) t.tv_sec * NANOSEC_PER_SEC + t.tv_nsec;
+}
+
+/* Common stats data record shared with _kern.c */
+struct datarec {
+	__u64 processed;
+	__u64 dropped;
+	__u64 issue;
+};
+struct record {
+	__u64 timestamp;
+	struct datarec total;
+	struct datarec *cpu;
+};
+struct stats_record {
+	struct record rx_cnt;
+	struct record redir_err;
+	struct record kthread;
+	struct record exception;
+	struct record enq[MAX_CPUS];
+};
+
+static bool map_collect_percpu(int fd, __u32 key, struct record* rec)
+{
+	/* For percpu maps, userspace gets a value per possible CPU */
+	unsigned int nr_cpus = bpf_num_possible_cpus();
+	struct datarec values[nr_cpus];
+	__u64 sum_processed = 0;
+	__u64 sum_dropped = 0;
+	__u64 sum_issue = 0;
+	int i;
+
+	if ((bpf_map_lookup_elem(fd, &key, values)) != 0) {
+		fprintf(stderr,
+			"ERR: bpf_map_lookup_elem failed key:0x%X\n", key);
+		return false;
+	}
+	/* Get time as close as possible to reading map contents */
+	rec->timestamp = gettime();
+
+	/* Record and sum values from each CPU */
+	for (i = 0; i < nr_cpus; i++) {
+		rec->cpu[i].processed = values[i].processed;
+		sum_processed        += values[i].processed;
+		rec->cpu[i].dropped = values[i].dropped;
+		sum_dropped        += values[i].dropped;
+		rec->cpu[i].issue = values[i].issue;
+		sum_issue        += values[i].issue;
+	}
+	rec->total.processed = sum_processed;
+	rec->total.dropped   = sum_dropped;
+	rec->total.issue     = sum_issue;
+	return true;
+}
+
+static struct datarec *alloc_record_per_cpu(void)
+{
+	unsigned int nr_cpus = bpf_num_possible_cpus();
+	struct datarec *array;
+	size_t size;
+
+	size = sizeof(struct datarec) * nr_cpus;
+	array = malloc(size);
+	memset(array, 0, size);
+	if (!array) {
+		fprintf(stderr, "Mem alloc error (nr_cpus:%u)\n", nr_cpus);
+		exit(EXIT_FAIL_MEM);
+	}
+	return array;
+}
+
+static struct stats_record* alloc_stats_record(void)
+{
+	struct stats_record* rec;
+	int i;
+
+	rec = malloc(sizeof(*rec));
+	memset(rec, 0, sizeof(*rec));
+	if (!rec) {
+		fprintf(stderr, "Mem alloc error\n");
+		exit(EXIT_FAIL_MEM);
+	}
+	rec->rx_cnt.cpu    = alloc_record_per_cpu();
+	rec->redir_err.cpu = alloc_record_per_cpu();
+	rec->kthread.cpu   = alloc_record_per_cpu();
+	rec->exception.cpu = alloc_record_per_cpu();
+	for (i = 0; i < MAX_CPUS; i++)
+		rec->enq[i].cpu = alloc_record_per_cpu();
+
+	return rec;
+}
+
+static void free_stats_record(struct stats_record* r)
+{
+	int i;
+
+	for (i = 0; i < MAX_CPUS; i++)
+		free(r->enq[i].cpu);
+	free(r->exception.cpu);
+	free(r->kthread.cpu);
+	free(r->redir_err.cpu);
+	free(r->rx_cnt.cpu);
+	free(r);
+}
+
+static double calc_period(struct record *r, struct record *p)
+{
+	double period_ = 0;
+	__u64 period  = 0;
+
+	period = r->timestamp - p->timestamp;
+	if (period > 0) {
+		period_ = ((double) period / NANOSEC_PER_SEC);
+	}
+	return period_;
+}
+
+static __u64 calc_pps(struct datarec *r, struct datarec *p, double period_)
+{
+	__u64 packets = 0;
+	__u64 pps = 0;
+
+	if (period_ > 0) {
+		packets = r->processed - p->processed;
+		pps = packets / period_;
+	}
+	return pps;
+}
+
+static __u64 calc_drop_pps(struct datarec *r, struct datarec *p, double period_)
+{
+	__u64 packets = 0;
+	__u64 pps = 0;
+
+	if (period_ > 0) {
+		packets = r->dropped - p->dropped;
+		pps = packets / period_;
+	}
+	return pps;
+}
+
+static __u64 calc_errs_pps(struct datarec *r,
+			    struct datarec *p, double period_)
+{
+	__u64 packets = 0;
+	__u64 pps = 0;
+
+	if (period_ > 0) {
+		packets = r->issue - p->issue;
+		pps = packets / period_;
+	}
+	return pps;
+}
+
+static void stats_print(struct stats_record *stats_rec,
+			struct stats_record *stats_prev,
+			int prog_num)
+{
+	unsigned int nr_cpus = bpf_num_possible_cpus();
+	double pps = 0, drop = 0, err = 0;
+	struct record *rec, *prev;
+	int to_cpu;
+	double t;
+	int i;
+
+	/* Header */
+	printf("Running XDP/eBPF prog_num:%d\n", prog_num);
+	printf("%-15s %-7s %-14s %-11s %-9s\n",
+	       "XDP-cpumap", "CPU:to", "pps", "drop-pps", "extra-info");
+
+	/* XDP rx_cnt */
+	{
+		char * fmt_rx = "%-15s %-7d %'-14.0f %'-11.0f %'-10.0f %s\n";
+		char * fm2_rx = "%-15s %-7s %'-14.0f %'-11.0f\n";
+		char *errstr = "";
+
+		rec  = &stats_rec->rx_cnt;
+		prev = &stats_prev->rx_cnt;
+		t = calc_period(rec, prev);
+		for (i = 0; i < nr_cpus; i++) {
+			struct datarec *r = &rec->cpu[i];
+			struct datarec *p = &prev->cpu[i];
+			pps = calc_pps(r, p, t);
+			drop = calc_drop_pps(r, p, t);
+			err  = calc_errs_pps(r, p, t);
+			if (err > 0)
+				errstr = "cpu-dest/err";
+			if (pps > 0)
+				printf(fmt_rx, "XDP-RX",
+				       i, pps, drop, err, errstr);
+		}
+		pps  = calc_pps(&rec->total, &prev->total, t);
+		drop = calc_drop_pps(&rec->total, &prev->total, t);
+		err  = calc_errs_pps(&rec->total, &prev->total, t);
+		printf(fm2_rx, "XDP-RX", "total", pps, drop);
+	}
+
+	/* cpumap enqueue stats */
+	for (to_cpu = 0; to_cpu < MAX_CPUS; to_cpu++) {
+		char *fmt="%-15s %3d:%-3d %'-14.0f %'-11.0f %'-10.0f %s\n";
+		char *fm2="%-15s %3s:%-3d %'-14.0f %'-11.0f %'-10.0f %s\n";
+		char *errstr = "";
+
+		rec  =  &stats_rec->enq[to_cpu];
+		prev = &stats_prev->enq[to_cpu];
+		t = calc_period(rec, prev);
+		for (i = 0; i < nr_cpus; i++) {
+			struct datarec *r = &rec->cpu[i];
+			struct datarec *p = &prev->cpu[i];
+			pps  = calc_pps(r, p, t);
+			drop = calc_drop_pps(r, p, t);
+			err  = calc_errs_pps(r, p, t);
+			if (err > 0)
+				errstr = "same-cpu/pps";
+			if (pps > 0)
+				printf(fmt, "cpumap-enqueue",
+				       i, to_cpu, pps, drop, err, errstr);
+		}
+		pps = calc_pps(&rec->total, &prev->total, t);
+		if (pps > 0) {
+			drop = calc_drop_pps(&rec->total, &prev->total, t);
+			err  = calc_errs_pps(&rec->total, &prev->total, t);
+			printf(fm2, "cpumap-enqueue",
+			       "sum", to_cpu, pps, drop, err, errstr);
+		}
+	}
+
+	/* cpumap kthread stats */
+	{
+		char *fmt_k = "%-15s %-7d %'-14.0f %'-11.0f %-10.0f %s\n";
+		char *fm2_k = "%-15s %-7s %'-14.0f %'-11.0f %-10.0f %s\n";
+		char *errstr = "";
+		rec  = &stats_rec->kthread;
+		prev = &stats_prev->kthread;
+		t = calc_period(rec, prev);
+		for (i = 0; i < nr_cpus; i++) {
+			struct datarec *r = &rec->cpu[i];
+			struct datarec *p = &prev->cpu[i];
+			pps  = calc_pps(r, p, t);
+			drop = calc_drop_pps(r, p, t);
+			err  = calc_errs_pps(r, p, t);
+			if (err > 0)
+				errstr = "time_exceed";
+			if (pps > 0)
+				printf(fmt_k, "cpumap_kthread",
+				       i, pps, drop, err, errstr);
+		}
+		pps = calc_pps(&rec->total, &prev->total, t);
+		drop = calc_drop_pps(&rec->total, &prev->total, t);
+		printf(fm2_k, "cpumap_kthread", "total", pps, drop);
+	}
+
+	/* XDP redirect err tracepoints (very unlikely) */
+	{
+		char *fmt_err = "%-15s %-7d %'-14.0f %'-11.0f\n";
+		char *fm2_err = "%-15s %-7s %'-14.0f %'-11.0f\n";
+		rec  = &stats_rec->redir_err;
+		prev = &stats_prev->redir_err;
+		t = calc_period(rec, prev);
+		for (i = 0; i < nr_cpus; i++) {
+			struct datarec *r = &rec->cpu[i];
+			struct datarec *p = &prev->cpu[i];
+			pps  = calc_pps(r, p, t);
+			drop = calc_drop_pps(r, p, t);
+			if (pps > 0)
+				printf(fmt_err, "redirect_err", i, pps, drop);
+		}
+		pps = calc_pps(&rec->total, &prev->total, t);
+		drop = calc_drop_pps(&rec->total, &prev->total, t);
+		printf(fm2_err, "redirect_err", "total", pps, drop);
+	}
+
+	/* XDP general exception tracepoints */
+	{
+		char *fmt_err = "%-15s %-7d %'-14.0f %'-11.0f\n";
+		char *fm2_err = "%-15s %-7s %'-14.0f %'-11.0f\n";
+		rec  = &stats_rec->exception;
+		prev = &stats_prev->exception;
+		t = calc_period(rec, prev);
+		for (i = 0; i < nr_cpus; i++) {
+			struct datarec *r = &rec->cpu[i];
+			struct datarec *p = &prev->cpu[i];
+			pps  = calc_pps(r, p, t);
+			drop = calc_drop_pps(r, p, t);
+			if (pps > 0)
+				printf(fmt_err, "xdp_exception", i, pps, drop);
+		}
+		pps = calc_pps(&rec->total, &prev->total, t);
+		drop = calc_drop_pps(&rec->total, &prev->total, t);
+		printf(fm2_err, "xdp_exception", "total", pps, drop);
+	}
+
+	printf("\n");
+	fflush(stdout);
+}
+
+static void stats_collect(struct stats_record *rec)
+{
+	int fd, i;
+
+	fd = map_fd[1]; /* map: rx_cnt */
+	map_collect_percpu(fd, 0, &rec->rx_cnt);
+
+	fd = map_fd[2]; /* map: redirect_err_cnt */
+	map_collect_percpu(fd, 1, &rec->redir_err);
+
+	fd = map_fd[3]; /* map: cpumap_enqueue_cnt */
+	for (i = 0; i < MAX_CPUS; i++) {
+		map_collect_percpu(fd, i, &rec->enq[i]);
+	}
+
+	fd = map_fd[4]; /* map: cpumap_kthread_cnt */
+	map_collect_percpu(fd, 0, &rec->kthread);
+
+	fd = map_fd[8]; /* map: exception_cnt */
+	map_collect_percpu(fd, 0, &rec->exception);
+}
+
+
+/* Pointer swap trick */
+static inline void swap(struct stats_record **a, struct stats_record **b)
+{
+	struct stats_record *tmp;
+
+	tmp = *a;
+	*a = *b;
+	*b = tmp;
+}
+
+static void stats_poll(int interval, bool use_separators, int prog_num)
+{
+	struct stats_record *record, *prev;
+
+	record = alloc_stats_record();
+	prev   = alloc_stats_record();
+	stats_collect(record);
+
+	/* Trick to pretty printf with thousands separators use %' */
+	if (use_separators)
+		setlocale(LC_NUMERIC, "en_US");
+
+	while (1) {
+		swap(&prev, &record);
+		stats_collect(record);
+		stats_print(record, prev, prog_num);
+		sleep(interval);
+	}
+
+	free_stats_record(record);
+	free_stats_record(prev);
+}
+
+static int create_cpu_entry(__u32 cpu, __u32 queue_size,
+			    __u32 avail_idx, bool new)
+{
+	__u32 curr_cpus_count;
+	__u32 key = 0;
+	int ret;
+
+	/* Add a CPU entry to cpumap, as this allocate a cpu entry in
+	 * the kernel for the cpu.
+	 */
+	ret = bpf_map_update_elem(map_fd[0], &cpu, &queue_size, 0);
+	if (ret) {
+		fprintf(stderr, "Create CPU entry failed\n");
+		exit(EXIT_FAIL_BPF);
+	}
+
+	/* Inform bpf_prog's that a new CPU is available to select
+	 * from via some control maps.
+	 */
+	/* map_fd[5] = cpus_available */
+	ret = bpf_map_update_elem(map_fd[5], &avail_idx, &cpu, 0);
+	if (ret) {
+		fprintf(stderr, "Add to avail CPUs failed\n");
+		exit(EXIT_FAIL_BPF);
+	}
+
+	/* When not replacing/updating existing entry, bump the count */
+	/* map_fd[6] = cpus_count */
+	if (new) {
+		ret = bpf_map_lookup_elem(map_fd[6], &key, &curr_cpus_count);
+		if (ret) {
+			fprintf(stderr, "Failed reading curr cpus_count \n");
+			exit(EXIT_FAIL_BPF);
+		}
+		curr_cpus_count++;
+		ret = bpf_map_update_elem(map_fd[6], &key, &curr_cpus_count, 0);
+		if (ret) {
+			fprintf(stderr, "Failed write curr cpus_count \n");
+			exit(EXIT_FAIL_BPF);
+		}
+	}
+	/* map_fd[7] = cpus_iterator */
+	printf("%s CPU:%u as idx:%u cpus_count:%u\n",
+	       new ? "Add-new":"Replace", cpu, avail_idx, curr_cpus_count);
+
+	return 0;
+}
+
+int main(int argc, char **argv)
+{
+	struct rlimit r = {10 * 1024*1024, RLIM_INFINITY};
+	bool use_separators = true;
+	char filename[256];
+	bool debug = false;
+	int added_cpus = 0;
+	int longindex = 0;
+	int interval = 2;
+	int prog_num = 0;
+	int add_cpu = -1;
+	__u32 qsize;
+	int opt;
+
+	/* Notice: choosing he queue size is very important with the
+	 * ixgbe driver, because it's driver page recycling trick is
+	 * dependend on pages being returned quickly.  The number of
+	 * out-standing packets in the system must be less-than 2x
+	 * RX-ring size.
+	 */
+	qsize = 128+64;
+
+	snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
+
+	if (setrlimit(RLIMIT_MEMLOCK, &r)) {
+		perror("setrlimit(RLIMIT_MEMLOCK)");
+		return 1;
+	}
+
+	if (load_bpf_file(filename)) {
+		fprintf(stderr, "ERR in load_bpf_file(): %s", bpf_log_buf);
+		return EXIT_FAIL;
+	}
+
+	if (!prog_fd[0]) {
+		fprintf(stderr, "ERR: load_bpf_file: %s\n", strerror(errno));
+		return EXIT_FAIL;
+	}
+
+	/* Parse commands line args */
+	while ((opt = getopt_long(argc, argv, "hSd:",
+				  long_options, &longindex)) != -1) {
+		switch (opt) {
+		case 'd':
+			if (strlen(optarg) >= IF_NAMESIZE) {
+				fprintf(stderr, "ERR: --dev name too long\n");
+				goto error;
+			}
+			ifname = (char *)&ifname_buf;
+			strncpy(ifname, optarg, IF_NAMESIZE);
+			ifindex = if_nametoindex(ifname);
+			if (ifindex == 0) {
+				fprintf(stderr,
+					"ERR: --dev name unknown err(%d):%s\n",
+					errno, strerror(errno));
+				goto error;
+			}
+			break;
+		case 's':
+			interval = atoi(optarg);
+			break;
+		case 'S':
+			xdp_flags |= XDP_FLAGS_SKB_MODE;
+			break;
+		case 'D':
+			debug = true;
+			break;
+		case 'z':
+			use_separators = false;
+			break;
+		case 'p':
+			/* Selecting eBPF prog to load */
+			prog_num = atoi(optarg);
+			if (prog_num < 0 || prog_num >= MAX_PROG) {
+				fprintf(stderr,
+					"--prognum too large err(%d):%s\n",
+					errno, strerror(errno));
+				goto error;
+			}
+			break;
+		case 'c':
+			/* Add multiple CPUs */
+			add_cpu = strtoul(optarg, NULL, 0);
+			if (add_cpu > MAX_CPUS) {
+				fprintf(stderr,
+				"--cpu nr too large for cpumap err(%d):%s\n",
+					errno, strerror(errno));
+				goto error;
+			}
+			create_cpu_entry(add_cpu, qsize, added_cpus, true);
+			added_cpus++;
+			break;
+		case 'q':
+			qsize = atoi(optarg);
+			break;
+		case 'h':
+		error:
+		default:
+			usage(argv);
+			return EXIT_FAIL_OPTION;
+		}
+	}
+	/* Required option */
+	if (ifindex == -1) {
+		fprintf(stderr, "ERR: required option --dev missing\n");
+		usage(argv);
+		return EXIT_FAIL_OPTION;
+	}
+	/* Required option */
+	if (add_cpu == -1) {
+		fprintf(stderr, "ERR: required option --cpu missing\n");
+		fprintf(stderr, " Specify multiple --cpu option to add more\n");
+		usage(argv);
+		return EXIT_FAIL_OPTION;
+	}
+
+	/* Remove XDP program when program is interrupted */
+	signal(SIGINT, int_exit);
+
+	if (set_link_xdp_fd(ifindex, prog_fd[prog_num], xdp_flags) < 0) {
+		fprintf(stderr, "link set xdp fd failed\n");
+		return EXIT_FAIL_XDP;
+	}
+
+	if (debug) {
+		printf("Debug-mode reading trace pipe (fix #define DEBUG)\n");
+		read_trace_pipe();
+	}
+
+	stats_poll(interval, use_separators, prog_num);
+	return EXIT_OK;
+}

Re: [net-next PATCH 0/5] New bpf cpumap type for XDP_REDIRECT

[Daniel Borkmann]

On 09/28/2017 02:57 PM, Jesper Dangaard Brouer wrote:
> Introducing a new way to redirect XDP frames.  Notice how no driver
> changes are necessary given the design of XDP_REDIRECT.
>
> This redirect map type is called 'cpumap', as it allows redirection
> XDP frames to remote CPUs.  The remote CPU will do the SKB allocation
> and start the network stack invocation on that CPU.
>
> This is a scalability and isolation mechanism, that allow separating
> the early driver network XDP layer, from the rest of the netstack, and
> assigning dedicated CPUs for this stage.  The sysadm control/configure
> the RX-CPU to NIC-RX queue (as usual) via procfs smp_affinity and how
> many queues are configured via ethtool --set-channels.  Benchmarks
> show that a single CPU can handle approx 11Mpps.  Thus, only assigning
> two NIC RX-queues (and two CPUs) is sufficient for handling 10Gbit/s
> wirespeed smallest packet 14.88Mpps.  Reducing the number of queues
> have the advantage that more packets being "bulk" available per hard
> interrupt[1].
>
> [1] https://www.netdevconf.org/2.1/papers/BusyPollingNextGen.pdf
>
> Use-cases:
>
> 1. End-host based pre-filtering for DDoS mitigation.  This is fast
>     enough to allow software to see and filter all packets wirespeed.
>     Thus, no packets getting silently dropped by hardware.
>
> 2. Given NIC HW unevenly distributes packets across RX queue, this
>     mechanism can be used for redistribution load across CPUs.  This
>     usually happens when HW is unaware of a new protocol.  This
>     resembles RPS (Receive Packet Steering), just faster, but with more
>     responsibility placed on the BPF program for correct steering.
>
> 3. Auto-scaling or power saving via only activating the appropriate
>     number of remote CPUs for handling the current load.  The cpumap
>     tracepoints can function as a feedback loop for this purpose.

Interesting work, thanks! Still digesting the code a bit. I think
it pretty much goes into the direction that Eric describes in his
netdev paper quoted above; not on a generic level though but specific
to XDP at least; theoretically XDP could just run transparently on
the CPU doing the filtering, and raw buffers are handed to remote
CPU with similar batching, but it would need some different config
interface at minimum.

Shouldn't we take the CPU(s) running XDP on the RX queues out from
the normal process scheduler, so that we have a guarantee that user
space or unrelated kernel tasks cannot interfere with them anymore,
and we could then turn them into busy polling eventually (e.g. as
long as XDP is running there and once off could put them back into
normal scheduling domain transparently)?

What about RPS/RFS in the sense that once you punt them to remote
CPU, could we reuse application locality information so they'd end
up on the right CPU in the first place (w/o backlog detour), or is
the intent to rather disable it and have some own orchestration
with relation to the CPU map?

Cheers,
Daniel

Re: [net-next PATCH 0/5] New bpf cpumap type for XDP_REDIRECT

[Jesper Dangaard Brouer]

On Fri, 29 Sep 2017 00:45:40 +0200
Daniel Borkmann <daniel@iogearbox.net> wrote:

> On 09/28/2017 02:57 PM, Jesper Dangaard Brouer wrote:
> > Introducing a new way to redirect XDP frames.  Notice how no driver
> > changes are necessary given the design of XDP_REDIRECT.
> >
> > This redirect map type is called 'cpumap', as it allows redirection
> > XDP frames to remote CPUs.  The remote CPU will do the SKB allocation
> > and start the network stack invocation on that CPU.
> >
> > This is a scalability and isolation mechanism, that allow separating
> > the early driver network XDP layer, from the rest of the netstack, and
> > assigning dedicated CPUs for this stage.  The sysadm control/configure
> > the RX-CPU to NIC-RX queue (as usual) via procfs smp_affinity and how
> > many queues are configured via ethtool --set-channels.  Benchmarks
> > show that a single CPU can handle approx 11Mpps.  Thus, only assigning
> > two NIC RX-queues (and two CPUs) is sufficient for handling 10Gbit/s
> > wirespeed smallest packet 14.88Mpps.  Reducing the number of queues
> > have the advantage that more packets being "bulk" available per hard
> > interrupt[1].
> >
> > [1] https://www.netdevconf.org/2.1/papers/BusyPollingNextGen.pdf
> >
> > Use-cases:
> >
> > 1. End-host based pre-filtering for DDoS mitigation.  This is fast
> >     enough to allow software to see and filter all packets wirespeed.
> >     Thus, no packets getting silently dropped by hardware.
> >
> > 2. Given NIC HW unevenly distributes packets across RX queue, this
> >     mechanism can be used for redistribution load across CPUs.  This
> >     usually happens when HW is unaware of a new protocol.  This
> >     resembles RPS (Receive Packet Steering), just faster, but with more
> >     responsibility placed on the BPF program for correct steering.
> >
> > 3. Auto-scaling or power saving via only activating the appropriate
> >     number of remote CPUs for handling the current load.  The cpumap
> >     tracepoints can function as a feedback loop for this purpose.  
> 
> Interesting work, thanks! Still digesting the code a bit. I think
> it pretty much goes into the direction that Eric describes in his
> netdev paper quoted above; not on a generic level though but specific
> to XDP at least; theoretically XDP could just run transparently on
> the CPU doing the filtering, and raw buffers are handed to remote
> CPU with similar batching, but it would need some different config
> interface at minimum.

Good that you noticed this is (implicit) implementing RX bulking, which
is where much of the performance gain originates from.

It is true, I am inspired by Eric's paper (I love it). Do notice that
this is not blocking or interfering with Erics/others continued work in
this area.  This implementation just show that the section "break the
pipe!" idea works very well for XDP. 

More on config knobs below.
 
> Shouldn't we take the CPU(s) running XDP on the RX queues out from
> the normal process scheduler, so that we have a guarantee that user
> space or unrelated kernel tasks cannot interfere with them anymore,
> and we could then turn them into busy polling eventually (e.g. as
> long as XDP is running there and once off could put them back into
> normal scheduling domain transparently)?

We should be careful not to invent networking config knobs that belongs
to other parts of the kernel, like the scheduler.  We already have
ability to control where IRQ's land via procfs smp_affinity.  And if
you want to avoid CPU isolation, we can use the boot cmdline
"isolcpus" (hint like DPDK recommend/use for zero-loss configs).  It is
the userspace tool (or sysadm) loading the XDP program, who is
responsible for having configures the CPU smp_affinity alignment.

Making NAPI busy-poll is out of scope for this patchset. Someone
should work on this separately.  It would just help/improve this kind
of scheme.

I actually think it would be more relevant to add/put the "remote" CPUs
in the 'cpumap' into a separate scheduler group.  To implement stuff
like auto-scaling and power-saving.


> What about RPS/RFS in the sense that once you punt them to remote
> CPU, could we reuse application locality information so they'd end
> up on the right CPU in the first place (w/o backlog detour), or is
> the intent to rather disable it and have some own orchestration
> with relation to the CPU map?

An advanced bpf orchestration could basically implement what you
describe, combined with a userspace side tool that taskset/pin
applications.  To know when a task can move between CPUs, you use the
tracepoints to see when the CPU queue is empty (hint, time_limit=true
and processed=0).

For now, I'm not targeting such advanced use-cases.  My main target is
a customer that have double tagged VLANS, and ixgbe cannot RSS
distribute these, thus they all end-up on queue 0.  And as I
demonstrated (in another email) RPS is too slow to fix this.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer