[net-next V8 PATCH 0/5] New bpf cpumap type for XDP_REDIRECT

[net-next V8 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.

In V7, a --stress-mode was implemented for the samples program, which
between each stats update, adds + removes CPUs from the map
concurrently with traffic.  I did find and fix some concurrency issues
in the tear-down path, details in patch desc.  The stress test have
now been running for 15 hours without any issues, while being
bombarded with 11.6 Mpps via pktgen_sample04_many_flows.sh.

See individual patches for patchset-version changes.

---

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                 |   31 +-
 include/linux/bpf_types.h           |    1 
 include/linux/netdevice.h           |    1 
 include/trace/events/xdp.h          |   80 ++++
 include/uapi/linux/bpf.h            |    1 
 kernel/bpf/Makefile                 |    1 
 kernel/bpf/cpumap.c                 |  702 +++++++++++++++++++++++++++++++++++
 kernel/bpf/syscall.c                |    8 
 kernel/bpf/verifier.c               |    8 
 net/core/dev.c                      |   27 +
 net/core/filter.c                   |  140 ++++++-
 samples/bpf/Makefile                |    4 
 samples/bpf/xdp_redirect_cpu_kern.c |  609 ++++++++++++++++++++++++++++++
 samples/bpf/xdp_redirect_cpu_user.c |  697 +++++++++++++++++++++++++++++++++++
 tools/include/uapi/linux/bpf.h      |    1 
 15 files changed, 2277 insertions(+), 34 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 V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Jesper Dangaard Brouer]

The 'cpumap' is primarily 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.

V2:
 - make sure array isn't larger than NR_CPUS
 - make sure CPUs added is a valid possible CPU

V3: fix nitpicks from Jakub Kicinski <kubakici@wp.pl>

V5:
 - Restrict map allocation to root / CAP_SYS_ADMIN
 - WARN_ON_ONCE if queue is not empty on tear-down
 - Return -EPERM on memlock limit instead of -ENOMEM
 - Error code in __cpu_map_entry_alloc() also handle ptr_ring_cleanup()
 - Moved cpu_map_enqueue() to next patch

V6: all notice by Daniel Borkmann
 - Fix err return code in cpu_map_alloc() introduced in V5
 - Move cpu_possible() check after max_entries boundary check
 - Forbid usage initially in check_map_func_compatibility()

V7:
 - Fix alloc error path spotted by Daniel Borkmann
 - Did stress test adding+removing CPUs from the map concurrently
 - Fixed refcnt issue on cpu_map_entry, kthread started too soon
 - Make sure packets are flushed during tear-down, involved use of
   rcu_barrier() and kthread_run only exit after queue is empty
 - Fix alloc error path in __cpu_map_entry_alloc() for ptr_ring

V8:
 - Nitpicking comments and gramma by Edward Cree
 - Fix missing semi-colon introduced in V7 due to rebasing
 - Move struct bpf_cpu_map_entry members cpu+map_id to tracepoint patch

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            |  560 ++++++++++++++++++++++++++++++++++++++++
 kernel/bpf/syscall.c           |    8 +
 kernel/bpf/verifier.c          |    5 
 tools/include/uapi/linux/bpf.h |    1 
 7 files changed, 576 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 6db9e1d679cd..4303fb6c3817 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -112,6 +112,7 @@ 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 53fb09f92e3f..e597daae6120 100644
--- a/kernel/bpf/Makefile
+++ b/kernel/bpf/Makefile
@@ -5,6 +5,7 @@ obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list
 obj-$(CONFIG_BPF_SYSCALL) += disasm.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..e1e25ddba038
--- /dev/null
+++ b/kernel/bpf/cpumap.c
@@ -0,0 +1,560 @@
+/* bpf/cpumap.c
+ *
+ * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
+ * Released under terms in GPL version 2.  See COPYING.
+ */
+
+/* The 'cpumap' is primarily used as a backend map for XDP BPF helper
+ * call bpf_redirect_map() and XDP_REDIRECT action, like 'devmap'.
+ *
+ * Unlike devmap which redirects XDP frames out another NIC device,
+ * this map type redirects 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>
+#include <linux/capability.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 qsize;  /* Queue size placeholder 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;
+	int err = -ENOMEM;
+	u64 cost;
+	int ret;
+
+	if (!capable(CAP_SYS_ADMIN))
+		return ERR_PTR(-EPERM);
+
+	/* 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);
+
+	/* Pre-limit array size based on NR_CPUS, not final CPU check */
+	if (cmap->map.max_entries > NR_CPUS) {
+		err = -E2BIG;
+		goto free_cmap;
+	}
+
+	/* 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;
+
+	/* Notice returns -EPERM on if map size is larger than memlock limit */
+	ret = bpf_map_precharge_memlock(cmap->map.pages);
+	if (ret) {
+		err = ret;
+		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_percpu;
+
+	return &cmap->map;
+free_percpu:
+	free_percpu(cmap->flush_needed);
+free_cmap:
+	kfree(cmap);
+	return ERR_PTR(err);
+}
+
+void __cpu_map_queue_destructor(void *ptr)
+{
+	/* The tear-down procedure should have made sure that queue is
+	 * empty.  See __cpu_map_entry_replace() and work-queue
+	 * invoked cpu_map_kthread_stop(). Catch any broken behaviour
+	 * gracefully and warn once.
+	 */
+	if (WARN_ON_ONCE(ptr))
+		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);
+
+	/* Wait for flush in __cpu_map_entry_free(), via full RCU barrier,
+	 * as it waits until all in-flight call_rcu() callbacks complete.
+	 */
+	rcu_barrier();
+
+	/* kthread_stop will wake_up_process and wait for it to complete */
+	kthread_stop(rcpu->kthread);
+}
+
+static int cpu_map_kthread_run(void *data)
+{
+	struct bpf_cpu_map_entry *rcpu = data;
+
+	set_current_state(TASK_INTERRUPTIBLE);
+
+	/* When kthread gives stop order, then rcpu have been disconnected
+	 * from map, thus no new packets can enter. Remaining in-flight
+	 * per CPU stored packets are flushed to this queue.  Wait honoring
+	 * kthread_stop signal until queue is empty.
+	 */
+	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
+		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);
+	}
+	__set_current_state(TASK_RUNNING);
+
+	put_cpu_map_entry(rcpu);
+	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 free_rcu;
+
+	/* Alloc queue */
+	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
+	if (!rcpu->queue)
+		goto free_bulkq;
+
+	err = ptr_ring_init(rcpu->queue, qsize, gfp);
+	if (err)
+		goto free_queue;
+
+	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 free_ptr_ring;
+
+	get_cpu_map_entry(rcpu); /* 1-refcnt for being in cmap->cpu_map[] */
+	get_cpu_map_entry(rcpu); /* 1-refcnt for kthread */
+
+	/* Make sure kthread runs on a single CPU */
+	kthread_bind(rcpu->kthread, cpu);
+	wake_up_process(rcpu->kthread);
+
+	return rcpu;
+
+free_ptr_ring:
+	ptr_ring_cleanup(rcpu->queue, NULL);
+free_queue:
+	kfree(rcpu->queue);
+free_bulkq:
+	free_percpu(rcpu->bulkq);
+free_rcu:
+	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;
+	/* Value is 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;
+
+	/* Make sure CPU is a valid possible cpu */
+	if (!cpu_possible(key_cpu))
+		return -ENODEV;
+
+	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.
+ */
+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;
+}
+
+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 d124e702e040..54fba06942f5 100644
--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c
@@ -592,6 +592,12 @@ 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
 	 */
@@ -622,7 +628,7 @@ 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/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 2cdbcc4f8f6b..d7368e8f37d6 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -1413,6 +1413,11 @@ static int check_map_func_compatibility(struct bpf_verifier_env *env,
 		if (func_id != BPF_FUNC_redirect_map)
 			goto error;
 		break;
+	/* Restrict bpf side of cpumap, open when use-cases appear */
+	case BPF_MAP_TYPE_CPUMAP:
+		if (func_id != BPF_FUNC_redirect_map)
+			goto error;
+		break;
 	case BPF_MAP_TYPE_ARRAY_OF_MAPS:
 	case BPF_MAP_TYPE_HASH_OF_MAPS:
 		if (func_id != BPF_FUNC_map_lookup_elem)
diff --git a/tools/include/uapi/linux/bpf.h b/tools/include/uapi/linux/bpf.h
index fb4fb81ce5b0..fa93033dc521 100644
--- a/tools/include/uapi/linux/bpf.h
+++ b/tools/include/uapi/linux/bpf.h
@@ -112,6 +112,7 @@ 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 V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Alexei Starovoitov]

On Mon, Oct 16, 2017 at 12:19:28PM +0200, Jesper Dangaard Brouer wrote:
> The 'cpumap' is primarily 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.
> 
> V2:
>  - make sure array isn't larger than NR_CPUS
>  - make sure CPUs added is a valid possible CPU
> 
> V3: fix nitpicks from Jakub Kicinski <kubakici@wp.pl>
> 
> V5:
>  - Restrict map allocation to root / CAP_SYS_ADMIN
>  - WARN_ON_ONCE if queue is not empty on tear-down
>  - Return -EPERM on memlock limit instead of -ENOMEM
>  - Error code in __cpu_map_entry_alloc() also handle ptr_ring_cleanup()
>  - Moved cpu_map_enqueue() to next patch
> 
> V6: all notice by Daniel Borkmann
>  - Fix err return code in cpu_map_alloc() introduced in V5
>  - Move cpu_possible() check after max_entries boundary check
>  - Forbid usage initially in check_map_func_compatibility()
> 
> V7:
>  - Fix alloc error path spotted by Daniel Borkmann
>  - Did stress test adding+removing CPUs from the map concurrently
>  - Fixed refcnt issue on cpu_map_entry, kthread started too soon
>  - Make sure packets are flushed during tear-down, involved use of
>    rcu_barrier() and kthread_run only exit after queue is empty
>  - Fix alloc error path in __cpu_map_entry_alloc() for ptr_ring
> 
> V8:
>  - Nitpicking comments and gramma by Edward Cree
>  - Fix missing semi-colon introduced in V7 due to rebasing
>  - Move struct bpf_cpu_map_entry members cpu+map_id to tracepoint patch
> 
> 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            |  560 ++++++++++++++++++++++++++++++++++++++++
>  kernel/bpf/syscall.c           |    8 +
>  kernel/bpf/verifier.c          |    5 
>  tools/include/uapi/linux/bpf.h |    1 
>  7 files changed, 576 insertions(+), 1 deletion(-)
>  create mode 100644 kernel/bpf/cpumap.c

Looks good to me
I like the idea of running networking stack from kthread
and hope adding GRO won't change the api. 
Acked-by: Alexei Starovoitov <ast@kernel.org>

Re: [net-next V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Jesper Dangaard Brouer]

On Mon, 16 Oct 2017 14:49:53 -0700
Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:

> On Mon, Oct 16, 2017 at 12:19:28PM +0200, Jesper Dangaard Brouer wrote:
> > The 'cpumap' is primarily 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.
> > 
> > V2:
> >  - make sure array isn't larger than NR_CPUS
> >  - make sure CPUs added is a valid possible CPU
> > 
> > V3: fix nitpicks from Jakub Kicinski <kubakici@wp.pl>
> > 
> > V5:
> >  - Restrict map allocation to root / CAP_SYS_ADMIN
> >  - WARN_ON_ONCE if queue is not empty on tear-down
> >  - Return -EPERM on memlock limit instead of -ENOMEM
> >  - Error code in __cpu_map_entry_alloc() also handle ptr_ring_cleanup()
> >  - Moved cpu_map_enqueue() to next patch
> > 
> > V6: all notice by Daniel Borkmann
> >  - Fix err return code in cpu_map_alloc() introduced in V5
> >  - Move cpu_possible() check after max_entries boundary check
> >  - Forbid usage initially in check_map_func_compatibility()
> > 
> > V7:
> >  - Fix alloc error path spotted by Daniel Borkmann
> >  - Did stress test adding+removing CPUs from the map concurrently
> >  - Fixed refcnt issue on cpu_map_entry, kthread started too soon
> >  - Make sure packets are flushed during tear-down, involved use of
> >    rcu_barrier() and kthread_run only exit after queue is empty
> >  - Fix alloc error path in __cpu_map_entry_alloc() for ptr_ring
> > 
> > V8:
> >  - Nitpicking comments and gramma by Edward Cree
> >  - Fix missing semi-colon introduced in V7 due to rebasing
> >  - Move struct bpf_cpu_map_entry members cpu+map_id to tracepoint patch
> > 
> > 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            |  560 ++++++++++++++++++++++++++++++++++++++++
> >  kernel/bpf/syscall.c           |    8 +
> >  kernel/bpf/verifier.c          |    5 
> >  tools/include/uapi/linux/bpf.h |    1 
> >  7 files changed, 576 insertions(+), 1 deletion(-)
> >  create mode 100644 kernel/bpf/cpumap.c  
> 
> Looks good to me
> I like the idea of running networking stack from kthread
> and hope adding GRO won't change the api. 
> Acked-by: Alexei Starovoitov <ast@kernel.org>

Thanks!

I think adding GRO is still safe API-wise after this patchset.  I
imagine that the GRO API will be tied to how we implement/expose the
RX-hash to the eBPF program.  Thus, the API will be the eBPF prog can
change the RX-hash, to influence the GRO aggregation/partial-sort. If
the map need to behave differently for GRO then we have the map_flags
to adjust this behavior (but I assume this would not be needed).

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

Re: [net-next V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Daniel Borkmann]

On 10/17/2017 12:47 PM, Jesper Dangaard Brouer wrote:
> On Mon, 16 Oct 2017 14:49:53 -0700
> Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
>
>> On Mon, Oct 16, 2017 at 12:19:28PM +0200, Jesper Dangaard Brouer wrote:
>>> The 'cpumap' is primarily 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.
>>>
>>> V2:
>>>   - make sure array isn't larger than NR_CPUS
>>>   - make sure CPUs added is a valid possible CPU
>>>
>>> V3: fix nitpicks from Jakub Kicinski <kubakici@wp.pl>
>>>
>>> V5:
>>>   - Restrict map allocation to root / CAP_SYS_ADMIN
>>>   - WARN_ON_ONCE if queue is not empty on tear-down
>>>   - Return -EPERM on memlock limit instead of -ENOMEM
>>>   - Error code in __cpu_map_entry_alloc() also handle ptr_ring_cleanup()
>>>   - Moved cpu_map_enqueue() to next patch
>>>
>>> V6: all notice by Daniel Borkmann
>>>   - Fix err return code in cpu_map_alloc() introduced in V5
>>>   - Move cpu_possible() check after max_entries boundary check
>>>   - Forbid usage initially in check_map_func_compatibility()
>>>
>>> V7:
>>>   - Fix alloc error path spotted by Daniel Borkmann
>>>   - Did stress test adding+removing CPUs from the map concurrently
>>>   - Fixed refcnt issue on cpu_map_entry, kthread started too soon
>>>   - Make sure packets are flushed during tear-down, involved use of
>>>     rcu_barrier() and kthread_run only exit after queue is empty
>>>   - Fix alloc error path in __cpu_map_entry_alloc() for ptr_ring
>>>
>>> V8:
>>>   - Nitpicking comments and gramma by Edward Cree
>>>   - Fix missing semi-colon introduced in V7 due to rebasing
>>>   - Move struct bpf_cpu_map_entry members cpu+map_id to tracepoint patch
>>>
>>> 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            |  560 ++++++++++++++++++++++++++++++++++++++++
>>>   kernel/bpf/syscall.c           |    8 +
>>>   kernel/bpf/verifier.c          |    5
>>>   tools/include/uapi/linux/bpf.h |    1
>>>   7 files changed, 576 insertions(+), 1 deletion(-)
>>>   create mode 100644 kernel/bpf/cpumap.c
>>
>> Looks good to me
>> I like the idea of running networking stack from kthread
>> and hope adding GRO won't change the api.
>> Acked-by: Alexei Starovoitov <ast@kernel.org>
>
> Thanks!
>
> I think adding GRO is still safe API-wise after this patchset.  I
> imagine that the GRO API will be tied to how we implement/expose the
> RX-hash to the eBPF program.  Thus, the API will be the eBPF prog can
> change the RX-hash, to influence the GRO aggregation/partial-sort. If
> the map need to behave differently for GRO then we have the map_flags
> to adjust this behavior (but I assume this would not be needed).

+1, this should happen transparent to the user. But would it mean that
the cpumap threads get a fake napi_struct for napi_gro_receive() or
would it require larger refactoring/splitting of gro engine internals?
Would be good to have a clearer picture on that before uapi freezes.

Thanks,
Daniel

Re: [net-next V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Yann Ylavic]

On Mon, Oct 16, 2017 at 12:19 PM, Jesper Dangaard Brouer
<brouer@redhat.com> wrote:
> +
> +       /* Notice returns -EPERM on if map size is larger than memlock limit */
> +       ret = bpf_map_precharge_memlock(cmap->map.pages);
> +       if (ret) {
> +               err = ret;
> +               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)

ret = -ENOMEM; ?

> +               goto free_percpu;
> +
> +       return &cmap->map;
> +free_percpu:
> +       free_percpu(cmap->flush_needed);
> +free_cmap:
> +       kfree(cmap);
> +       return ERR_PTR(err);
> +}


Regards,
Yann.

Re: [net-next V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Jesper Dangaard Brouer]

On Mon, 16 Oct 2017 12:19:28 +0200 Jesper Dangaard Brouer <brouer@redhat.com> wrote:

> +static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
> +{
> +	struct bpf_cpu_map *cmap;
> +	int err = -ENOMEM;

Notice err is set to -ENOMEM.

> +	u64 cost;
> +	int ret;
> +
> +	if (!capable(CAP_SYS_ADMIN))
> +		return ERR_PTR(-EPERM);
> +
> +	/* 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);
> +
> +	/* Pre-limit array size based on NR_CPUS, not final CPU check */
> +	if (cmap->map.max_entries > NR_CPUS) {
> +		err = -E2BIG;
> +		goto free_cmap;
> +	}
> +
> +	/* 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;
> +
> +	/* Notice returns -EPERM on if map size is larger than memlock limit */

[... continued below ...]

On Wed, 18 Oct 2017 09:45:59 +0200 Yann Ylavic <ylavic.dev@gmail.com> wrote:

> On Mon, Oct 16, 2017 at 12:19 PM, Jesper Dangaard Brouer
> <brouer@redhat.com> wrote:
> > +
> > +       /* Notice returns -EPERM on if map size is larger than memlock limit */
> > +       ret = bpf_map_precharge_memlock(cmap->map.pages);
> > +       if (ret) {
> > +               err = ret;
> > +               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)  
> 
> ret = -ENOMEM; ?

Did you notice that "err" already is = -ENOMEM at this point?


> > +               goto free_percpu;
> > +
> > +       return &cmap->map;
> > +free_percpu:
> > +       free_percpu(cmap->flush_needed);
> > +free_cmap:
> > +       kfree(cmap);
> > +       return ERR_PTR(err);
> > +}  
> 
> 
> Regards,
> Yann.



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

Re: [net-next V8 PATCH 1/5] bpf: introduce new bpf cpu map type BPF_MAP_TYPE_CPUMAP

[Yann Ylavic]

On Wed, Oct 18, 2017 at 10:38 AM, Jesper Dangaard Brouer
<brouer@redhat.com> wrote:
>
> On Wed, 18 Oct 2017 09:45:59 +0200 Yann Ylavic <ylavic.dev@gmail.com> wrote:
>
>> On Mon, Oct 16, 2017 at 12:19 PM, Jesper Dangaard Brouer
>> <brouer@redhat.com> wrote:
>> > +
>> > +       /* Notice returns -EPERM on if map size is larger than memlock limit */
>> > +       ret = bpf_map_precharge_memlock(cmap->map.pages);
>> > +       if (ret) {
>> > +               err = ret;
>> > +               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)
>>
>> ret = -ENOMEM; ?
>
> Did you notice that "err" already is = -ENOMEM at this point?

My bad, I was confused by the use of "ret"...

>
>
>> > +               goto free_percpu;
>> > +
>> > +       return &cmap->map;
>> > +free_percpu:
>> > +       free_percpu(cmap->flush_needed);
>> > +free_cmap:
>> > +       kfree(cmap);
>> > +       return ERR_PTR(err);

.. while actually "err" is returned.
Sorry for wasting your time.

>> > +}
>>
>>
>> Regards,
>> Yann.

[net-next V8 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.

V3: Handle !CONFIG_BPF_SYSCALL pointed out by kbuild test robot.

V4: Make Generic-XDP aware of cpumap type, but don't allow redirect yet,
 as implementation require a separate upstream discussion.

V5:
 - Fix a maybe-uninitialized pointed out by kbuild test robot.
 - Restrict bpf-prog side access to cpumap, open when use-cases appear
 - Implement cpu_map_enqueue() as a more simple void pointer enqueue

V6:
 - Allow cpumap type for usage in helper bpf_redirect_map,
   general bpf-prog side restriction moved to earlier patch.

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/linux/bpf.h        |   31 +++++++++-
 include/trace/events/xdp.h |   10 +++
 kernel/bpf/cpumap.c        |   22 +++++++
 kernel/bpf/verifier.c      |    3 +
 net/core/filter.c          |  140 +++++++++++++++++++++++++++++++++++---------
 5 files changed, 172 insertions(+), 34 deletions(-)

diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index 4373125de1f3..6d4dd844828a 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -355,6 +355,13 @@ 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)
 {
@@ -362,7 +369,7 @@ static inline int bpf_map_attr_numa_node(const union bpf_attr *attr)
 		attr->numa_node : NUMA_NO_NODE;
 }
 
-#else
+#else /* !CONFIG_BPF_SYSCALL */
 static inline struct bpf_prog *bpf_prog_get(u32 ufd)
 {
 	return ERR_PTR(-EOPNOTSUPP);
@@ -425,6 +432,28 @@ static inline void __dev_map_insert_ctx(struct bpf_map *map, u32 index)
 static inline void __dev_map_flush(struct bpf_map *map)
 {
 }
+
+static inline
+struct bpf_cpu_map_entry *__cpu_map_lookup_elem(struct bpf_map *map, u32 key)
+{
+	return NULL;
+}
+
+static inline void __cpu_map_insert_ctx(struct bpf_map *map, u32 index)
+{
+}
+
+static inline void __cpu_map_flush(struct bpf_map *map)
+{
+}
+
+struct xdp_buff;
+static inline int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu,
+				  struct xdp_buff *xdp,
+				  struct net_device *dev_rx)
+{
+	return 0;
+}
 #endif /* CONFIG_BPF_SYSCALL */
 
 #if defined(CONFIG_STREAM_PARSER) && defined(CONFIG_BPF_SYSCALL)
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 e1e25ddba038..768da6a2c265 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -500,7 +500,7 @@ struct xdp_pkt {
 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
  * Thus, safe percpu variable access.
  */
-int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_pkt *xdp_pkt)
+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);
 
@@ -520,6 +520,26 @@ 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,
+		    struct net_device *dev_rx)
+{
+	struct xdp_pkt *xdp_pkt;
+	int headroom;
+
+	/* For now this is just used as a void pointer to data_hard_start.
+	 * Followup patch will generalize this.
+	 */
+	xdp_pkt = xdp->data_hard_start;
+
+	/* Fake writing into xdp_pkt->data to measure overhead */
+	headroom = xdp->data - xdp->data_hard_start;
+	if (headroom < sizeof(*xdp_pkt))
+		xdp_pkt->data = xdp->data;
+
+	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);
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index d7368e8f37d6..748484287e5e 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -1455,7 +1455,8 @@ static int check_map_func_compatibility(struct bpf_verifier_env *env,
 			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 140fa9f9c0f4..4d88e0665c41 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -2526,10 +2526,36 @@ 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;
 }
 
@@ -2539,11 +2565,33 @@ 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)
 {
@@ -2556,8 +2604,8 @@ 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;
@@ -2570,7 +2618,7 @@ 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;
@@ -2578,7 +2626,7 @@ 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;
 
@@ -2620,54 +2668,88 @@ int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
 }
 EXPORT_SYMBOL_GPL(xdp_do_redirect);
 
-int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
-			    struct bpf_prog *xdp_prog)
+static int __xdp_generic_ok_fwd_dev(struct sk_buff *skb, struct net_device *fwd)
+{
+	unsigned int len;
+
+	if (unlikely(!(fwd->flags & IFF_UP)))
+		return -ENETDOWN;
+
+	len = fwd->mtu + fwd->hard_header_len + VLAN_HLEN;
+	if (skb->len > len)
+		return -EMSGSIZE;
+
+	return 0;
+}
+
+int xdp_do_generic_redirect_map(struct net_device *dev, struct sk_buff *skb,
+				struct bpf_prog *xdp_prog)
 {
 	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;
-	unsigned int len;
 	int err = 0;
 
 	ri->ifindex = 0;
 	ri->map = NULL;
 	ri->map_owner = 0;
 
-	if (map) {
-		if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
-			err = -EFAULT;
-			map = NULL;
-			goto err;
-		}
-		fwd = __dev_map_lookup_elem(map, index);
-	} else {
-		fwd = dev_get_by_index_rcu(dev_net(dev), index);
+	if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
+		err = -EFAULT;
+		map = NULL;
+		goto err;
 	}
+	fwd = __xdp_map_lookup_elem(map, index);
 	if (unlikely(!fwd)) {
 		err = -EINVAL;
 		goto err;
 	}
 
-	if (unlikely(!(fwd->flags & IFF_UP))) {
-		err = -ENETDOWN;
+	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
+		if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
+			goto err;
+		skb->dev = fwd;
+	} else {
+		/* TODO: Handle BPF_MAP_TYPE_CPUMAP */
+		err = -EBADRQC;
 		goto err;
 	}
 
-	len = fwd->mtu + fwd->hard_header_len + VLAN_HLEN;
-	if (skb->len > len) {
-		err = -EMSGSIZE;
+	_trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
+	return 0;
+err:
+	_trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
+	return err;
+}
+
+int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
+			    struct bpf_prog *xdp_prog)
+{
+	struct redirect_info *ri = this_cpu_ptr(&redirect_info);
+	u32 index = ri->ifindex;
+	struct net_device *fwd;
+	int err = 0;
+
+	if (ri->map)
+		return xdp_do_generic_redirect_map(dev, skb, xdp_prog);
+
+	ri->ifindex = 0;
+	fwd = dev_get_by_index_rcu(dev_net(dev), index);
+	if (unlikely(!fwd)) {
+		err = -EINVAL;
 		goto err;
 	}
 
+	if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
+		goto err;
+
 	skb->dev = fwd;
-	map ? _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index)
-		: _trace_xdp_redirect(dev, xdp_prog, index);
+	_trace_xdp_redirect(dev, xdp_prog, index);
 	return 0;
 err:
-	map ? _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err)
-		: _trace_xdp_redirect_err(dev, xdp_prog, index, err);
+	_trace_xdp_redirect_err(dev, xdp_prog, index, err);
 	return err;
 }
 EXPORT_SYMBOL_GPL(xdp_do_generic_redirect);

[net-next V8 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.

V2: take into account xdp->data_meta

V4:
 - Drop busypoll tricks, keeping it more simple.
 - Skip RPS and Generic-XDP-recursive-reinjection, suggested by Alexei

V5: correct RCU read protection around __netif_receive_skb_core.

V6: Setting TASK_RUNNING vs TASK_INTERRUPTIBLE based on talk with Rik van Riel

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/linux/netdevice.h |    1 
 kernel/bpf/cpumap.c       |  152 ++++++++++++++++++++++++++++++++++++++-------
 net/core/dev.c            |   27 ++++++++
 3 files changed, 158 insertions(+), 22 deletions(-)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 31bb3010c69b..bf014afcb914 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -3260,6 +3260,7 @@ int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb);
 int netif_rx(struct sk_buff *skb);
 int netif_rx_ni(struct sk_buff *skb);
 int netif_receive_skb(struct sk_buff *skb);
+int netif_receive_skb_core(struct sk_buff *skb);
 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb);
 void napi_gro_flush(struct napi_struct *napi, bool flush_old);
 struct sk_buff *napi_get_frags(struct napi_struct *napi);
diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index 768da6a2c265..ee7adf4352dd 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -25,6 +25,9 @@
 #include <linux/kthread.h>
 #include <linux/capability.h>
 
+#include <linux/netdevice.h>   /* netif_receive_skb_core */
+#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
  * guaranteed that setting flush bit and flush operation happen on
@@ -179,6 +182,92 @@ static void cpu_map_kthread_stop(struct work_struct *work)
 	kthread_stop(rcpu->kthread);
 }
 
+/* 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;
+	u16 metasize;
+	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 metasize;
+	int headroom;
+
+	/* Assure headroom is available for storing info */
+	headroom = xdp->data - xdp->data_hard_start;
+	metasize = xdp->data - xdp->data_meta;
+	metasize = metasize > 0 ? metasize : 0;
+	if ((headroom - metasize) < 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);
+	xdp_pkt->metasize = metasize;
+
+	return xdp_pkt;
+}
+
+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);
+	if (xdp_pkt->metasize)
+		skb_metadata_set(skb, xdp_pkt->metasize);
+
+	/* 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)
 {
 	struct bpf_cpu_map_entry *rcpu = data;
@@ -191,15 +280,45 @@ static int cpu_map_kthread_run(void *data)
 	 * kthread_stop signal until queue is empty.
 	 */
 	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
+		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 (__ptr_ring_empty(rcpu->queue)) {
+			__set_current_state(TASK_INTERRUPTIBLE);
+			schedule();
+		} else {
+			cond_resched();
+		}
+		__set_current_state(TASK_RUNNING);
+
+		/* 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;
+
+			skb = cpu_map_build_skb(rcpu, xdp_pkt);
+			if (!skb) {
+				page_frag_free(xdp_pkt);
+				continue;
+			}
+
+			/* Inject into network stack */
+			ret = netif_receive_skb_core(skb);
+			if (ret == NET_RX_DROP)
+				drops++;
+
+			/* Limit BH-disable period */
+			if (++processed == 8)
+				break;
 		}
-		__set_current_state(TASK_INTERRUPTIBLE);
+		local_bh_enable(); /* resched point, may call do_softirq() */
 	}
 	__set_current_state(TASK_RUNNING);
 
@@ -490,13 +609,6 @@ 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.
  */
@@ -524,17 +636,13 @@ 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;
 
-	/* For now this is just used as a void pointer to data_hard_start.
-	 * Followup patch will generalize this.
-	 */
-	xdp_pkt = xdp->data_hard_start;
+	xdp_pkt = convert_to_xdp_pkt(xdp);
+	if (!xdp_pkt)
+		return -EOVERFLOW;
 
-	/* Fake writing into xdp_pkt->data to measure overhead */
-	headroom = xdp->data - xdp->data_hard_start;
-	if (headroom < sizeof(*xdp_pkt))
-		xdp_pkt->data = xdp->data;
+	/* Info needed when constructing SKB on remote CPU */
+	xdp_pkt->dev_rx = dev_rx;
 
 	bq_enqueue(rcpu, xdp_pkt);
 	return 0;
diff --git a/net/core/dev.c b/net/core/dev.c
index fcddccb6be41..36bb68f3a2c7 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4491,6 +4491,33 @@ static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
 	return ret;
 }
 
+/**
+ *	netif_receive_skb_core - special purpose version of netif_receive_skb
+ *	@skb: buffer to process
+ *
+ *	More direct receive version of netif_receive_skb().  It should
+ *	only be used by callers that have a need to skip RPS and Generic XDP.
+ *	Caller must also take care of handling if (page_is_)pfmemalloc.
+ *
+ *	This function may only be called from softirq context and interrupts
+ *	should be enabled.
+ *
+ *	Return values (usually ignored):
+ *	NET_RX_SUCCESS: no congestion
+ *	NET_RX_DROP: packet was dropped
+ */
+int netif_receive_skb_core(struct sk_buff *skb)
+{
+	int ret;
+
+	rcu_read_lock();
+	ret = __netif_receive_skb_core(skb, false);
+	rcu_read_unlock();
+
+	return ret;
+}
+EXPORT_SYMBOL(netif_receive_skb_core);
+
 static int __netif_receive_skb(struct sk_buff *skb)
 {
 	int ret;

Re: [net-next V8 PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Michael S. Tsirkin]

On Mon, Oct 16, 2017 at 12:19:39PM +0200, Jesper Dangaard Brouer wrote:
> @@ -191,15 +280,45 @@ static int cpu_map_kthread_run(void *data)
>  	 * kthread_stop signal until queue is empty.
>  	 */
>  	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
> +		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 (__ptr_ring_empty(rcpu->queue)) {


I suspect this is racy: if ring becomes non empty here and
you wake the task, next line will put it to sleep.
I think you want to reverse the order:

			__set_current_state(TASK_INTERRUPTIBLE);

	then check __ptr_ring_empty.

I note using the __ version means you can not resize the ring.
Hope you do not need to.

> +			__set_current_state(TASK_INTERRUPTIBLE);
> +			schedule();
> +		} else {
> +			cond_resched();
> +		}
> +		__set_current_state(TASK_RUNNING);
> +
> +		/* 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;
> +
> +			skb = cpu_map_build_skb(rcpu, xdp_pkt);
> +			if (!skb) {
> +				page_frag_free(xdp_pkt);
> +				continue;
> +			}
> +
> +			/* Inject into network stack */
> +			ret = netif_receive_skb_core(skb);
> +			if (ret == NET_RX_DROP)
> +				drops++;
> +
> +			/* Limit BH-disable period */
> +			if (++processed == 8)
> +				break;
>  		}
> -		__set_current_state(TASK_INTERRUPTIBLE);
> +		local_bh_enable(); /* resched point, may call do_softirq() */
>  	}
>  	__set_current_state(TASK_RUNNING);
>  
> @@ -490,13 +609,6 @@ 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.
>   */
> @@ -524,17 +636,13 @@ 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;
>  
> -	/* For now this is just used as a void pointer to data_hard_start.
> -	 * Followup patch will generalize this.
> -	 */
> -	xdp_pkt = xdp->data_hard_start;
> +	xdp_pkt = convert_to_xdp_pkt(xdp);
> +	if (!xdp_pkt)
> +		return -EOVERFLOW;
>  
> -	/* Fake writing into xdp_pkt->data to measure overhead */
> -	headroom = xdp->data - xdp->data_hard_start;
> -	if (headroom < sizeof(*xdp_pkt))
> -		xdp_pkt->data = xdp->data;
> +	/* Info needed when constructing SKB on remote CPU */
> +	xdp_pkt->dev_rx = dev_rx;
>  
>  	bq_enqueue(rcpu, xdp_pkt);
>  	return 0;
> diff --git a/net/core/dev.c b/net/core/dev.c
> index fcddccb6be41..36bb68f3a2c7 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -4491,6 +4491,33 @@ static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
>  	return ret;
>  }
>  
> +/**
> + *	netif_receive_skb_core - special purpose version of netif_receive_skb
> + *	@skb: buffer to process
> + *
> + *	More direct receive version of netif_receive_skb().  It should
> + *	only be used by callers that have a need to skip RPS and Generic XDP.
> + *	Caller must also take care of handling if (page_is_)pfmemalloc.
> + *
> + *	This function may only be called from softirq context and interrupts
> + *	should be enabled.
> + *
> + *	Return values (usually ignored):
> + *	NET_RX_SUCCESS: no congestion
> + *	NET_RX_DROP: packet was dropped
> + */
> +int netif_receive_skb_core(struct sk_buff *skb)
> +{
> +	int ret;
> +
> +	rcu_read_lock();
> +	ret = __netif_receive_skb_core(skb, false);
> +	rcu_read_unlock();
> +
> +	return ret;
> +}
> +EXPORT_SYMBOL(netif_receive_skb_core);
> +
>  static int __netif_receive_skb(struct sk_buff *skb)
>  {
>  	int ret;

Re: [net-next V8 PATCH 3/5] bpf: cpumap xdp_buff to skb conversion and allocation

[Jesper Dangaard Brouer]

On Wed, 18 Oct 2017 17:12:09 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Mon, Oct 16, 2017 at 12:19:39PM +0200, Jesper Dangaard Brouer wrote:
> > @@ -191,15 +280,45 @@ static int cpu_map_kthread_run(void *data)
> >  	 * kthread_stop signal until queue is empty.
> >  	 */
> >  	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
> > +		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 (__ptr_ring_empty(rcpu->queue)) {  
> 
> 
> I suspect this is racy: if ring becomes non empty here and
> you wake the task, next line will put it to sleep.
> I think you want to reverse the order:
> 
> 			__set_current_state(TASK_INTERRUPTIBLE);
> 
> 	then check __ptr_ring_empty.

I'll look into this.

The window will be minimal, as __cpu_map_flush() after the last packets
enqueue will call wake_up_process(rcpu->kthread).  But I guess there is
still small race possible.  Worst case, a packet could be stuck in the
queue until a new packet arrive.  Thanks for spotting this.


> I note using the __ version means you can not resize the ring.
> Hope you do not need to.

Resize is not supported.  If user change the queue size, a new ptr_ring
and kthread is created, and logic assured the old ptr_ring and kthread
flush packets appropriately (this is tested with the --stress-mode).

 
> > +			__set_current_state(TASK_INTERRUPTIBLE);
> > +			schedule();
> > +		} else {
> > +			cond_resched();
> > +		}
> > +		__set_current_state(TASK_RUNNING);
> > +
> > +		/* 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;
> > +
> > +			skb = cpu_map_build_skb(rcpu, xdp_pkt);
> > +			if (!skb) {
> > +				page_frag_free(xdp_pkt);
> > +				continue;
> > +			}
> > +
> > +			/* Inject into network stack */
> > +			ret = netif_receive_skb_core(skb);
> > +			if (ret == NET_RX_DROP)
> > +				drops++;
> > +
> > +			/* Limit BH-disable period */
> > +			if (++processed == 8)
> > +				break;
> >  		}
> > -		__set_current_state(TASK_INTERRUPTIBLE);
> > +		local_bh_enable(); /* resched point, may call do_softirq() */
> >  	}
> >  	__set_current_state(TASK_RUNNING);
> >


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

[net-next V8 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.

V4: tracepoint remove time_limit info, instead add sched info

V8: intro struct bpf_cpu_map_entry members cpu+map_id in this patch

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/trace/events/xdp.h |   70 ++++++++++++++++++++++++++++++++++++++++++++
 kernel/bpf/cpumap.c        |   24 ++++++++++++---
 2 files changed, 89 insertions(+), 5 deletions(-)

diff --git a/include/trace/events/xdp.h b/include/trace/events/xdp.h
index eb2ece96c1a2..0c8dec61987e 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 sched),
+
+	TP_ARGS(map_id, processed, drops, sched),
+
+	TP_STRUCT__entry(
+		__field(int, map_id)
+		__field(u32, act)
+		__field(int, cpu)
+		__field(unsigned int, drops)
+		__field(unsigned int, processed)
+		__field(int, sched)
+	),
+
+	TP_fast_assign(
+		__entry->map_id		= map_id;
+		__entry->act		= XDP_REDIRECT;
+		__entry->cpu		= smp_processor_id();
+		__entry->drops		= drops;
+		__entry->processed	= processed;
+		__entry->sched	= sched;
+	),
+
+	TP_printk("kthread"
+		  " cpu=%d map_id=%d action=%s"
+		  " processed=%u drops=%u"
+		  " sched=%d",
+		  __entry->cpu, __entry->map_id,
+		  __print_symbolic(__entry->act, __XDP_ACT_SYM_TAB),
+		  __entry->processed, __entry->drops,
+		  __entry->sched)
+);
+
+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 ee7adf4352dd..b4358d84ddf1 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -24,6 +24,7 @@
 #include <linux/workqueue.h>
 #include <linux/kthread.h>
 #include <linux/capability.h>
+#include <trace/events/xdp.h>
 
 #include <linux/netdevice.h>   /* netif_receive_skb_core */
 #include <linux/etherdevice.h> /* eth_type_trans */
@@ -43,6 +44,8 @@ struct xdp_bulk_queue {
 
 /* 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;  /* Queue size placeholder for map lookup */
 
 	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
@@ -280,15 +283,16 @@ static int cpu_map_kthread_run(void *data)
 	 * kthread_stop signal until queue is empty.
 	 */
 	while (!kthread_should_stop() || !__ptr_ring_empty(rcpu->queue)) {
-		unsigned int processed = 0, drops = 0;
+		unsigned int processed = 0, drops = 0, sched = 0;
 		struct xdp_pkt *xdp_pkt;
 
 		/* Release CPU reschedule checks */
 		if (__ptr_ring_empty(rcpu->queue)) {
 			__set_current_state(TASK_INTERRUPTIBLE);
 			schedule();
+			sched = 1;
 		} else {
-			cond_resched();
+			sched = cond_resched();
 		}
 		__set_current_state(TASK_RUNNING);
 
@@ -318,6 +322,9 @@ 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, sched);
+
 		local_bh_enable(); /* resched point, may call do_softirq() */
 	}
 	__set_current_state(TASK_RUNNING);
@@ -354,7 +361,9 @@ struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu, int map_id)
 	if (err)
 		goto free_queue;
 
-	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,
@@ -584,6 +593,8 @@ 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;
 
@@ -599,13 +610,16 @@ 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 V8 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

V3:
 - whitespace cleanups
 - bpf tracepoint cannot access top part of struct

V4:
 - report on kthread sched events, according to tracepoint change
 - report average bulk enqueue size

V5:
 - bpf_map_lookup_elem on cpumap not allowed from bpf_prog
   use separate map to mark CPUs not available

V6:
 - correct kthread sched summary output

V7:
 - Added a --stress-mode for concurrently changing underlying cpumap

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 samples/bpf/Makefile                |    4 
 samples/bpf/xdp_redirect_cpu_kern.c |  609 +++++++++++++++++++++++++++++++
 samples/bpf/xdp_redirect_cpu_user.c |  697 +++++++++++++++++++++++++++++++++++
 3 files changed, 1310 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..303e9e7161f3
--- /dev/null
+++ b/samples/bpf/xdp_redirect_cpu_kern.c
@@ -0,0 +1,609 @@
+/*  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_selected;
+	u32 cpu_dest;
+	u32 key = 0;
+
+	/* Only use first entry in cpus_available */
+	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)
+		return XDP_ABORTED;
+	rec->processed++;
+
+	if (cpu_dest >= MAX_CPUS) {
+		rec->issue++;
+		return XDP_ABORTED;
+	}
+
+	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;
+	struct datarec *rec;
+	u32 *cpu_selected;
+	u32 cpu_dest;
+	u16 eth_type;
+	u32 key = 0;
+
+	/* Only use first entry in cpus_available */
+	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;
+	}
+
+	if (cpu_dest >= MAX_CPUS) {
+		rec->issue++;
+		return XDP_ABORTED;
+	}
+
+	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++;
+
+	if (cpu_dest >= MAX_CPUS) {
+		rec->issue++;
+		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) {
+		rec->issue++;
+		return XDP_ABORTED;
+	}
+
+	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) {
+		rec->issue++;
+		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 {
+	u64 __pad;	// First 8 bytes are not accessible by bpf code
+	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 {
+	u64 __pad;	// First 8 bytes are not accessible by bpf code
+	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 {
+	u64 __pad;		// First 8 bytes are not accessible by bpf code
+	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;
+
+	/* Record bulk events, then userspace can calc average bulk size */
+	if (ctx->processed > 0)
+		rec->issue += 1;
+
+	/* Inception: It's possible to detect overload situations, via
+	 * this tracepoint.  This can be used for creating a feedback
+	 * loop to XDP, which can take appropriate actions to mitigate
+	 * this overload situation.
+	 */
+	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 {
+	u64 __pad;		// First 8 bytes are not accessible by bpf code
+	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 sched;		//	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;
+
+	/* Count times kthread yielded CPU via schedule call */
+	if (ctx->sched)
+		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..35fec9fecb57
--- /dev/null
+++ b/samples/bpf/xdp_redirect_cpu_user.c
@@ -0,0 +1,697 @@
+/* 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;
+
+static __u32 xdp_flags;
+
+/* 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' },
+	{"stress-mode", no_argument,		NULL, 'x' },
+	{"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 %'-10.2f %s\n";
+		char *fm2 = "%-15s %3s:%-3d %'-14.0f %'-11.0f %'-10.2f %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 = "bulk-average";
+				err = pps / err; /* calc average bulk size */
+			}
+			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);
+			if (err > 0) {
+				errstr = "bulk-average";
+				err = pps / err; /* calc average bulk size */
+			}
+			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 *e_str = "";
+
+		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)
+				e_str = "sched";
+			if (pps > 0)
+				printf(fmt_k, "cpumap_kthread",
+				       i, pps, drop, err, e_str);
+		}
+		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);
+		if (err > 0)
+			e_str = "sched-sum";
+		printf(fm2_k, "cpumap_kthread", "total", pps, drop, err, e_str);
+	}
+
+	/* 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 int create_cpu_entry(__u32 cpu, __u32 queue_size,
+			    __u32 avail_idx, bool new)
+{
+	__u32 curr_cpus_count = 0;
+	__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 (err:%d)\n", ret);
+		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 */
+	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);
+	}
+	if (new) {
+		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 queue_size:%d (total cpus_count:%u)\n",
+	       new ? "Add-new":"Replace", cpu, avail_idx,
+	       queue_size, curr_cpus_count);
+
+	return 0;
+}
+
+/* CPUs are zero-indexed. Thus, add a special sentinel default value
+ * in map cpus_available to mark CPU index'es not configured
+ */
+static void mark_cpus_unavailable(void)
+{
+	__u32 invalid_cpu = MAX_CPUS;
+	int ret, i;
+
+	for (i = 0; i < MAX_CPUS; i++) {
+		/* map_fd[5] = cpus_available */
+		ret = bpf_map_update_elem(map_fd[5], &i, &invalid_cpu, 0);
+		if (ret) {
+			fprintf(stderr, "Failed marking CPU unavailable\n");
+			exit(EXIT_FAIL_BPF);
+		}
+	}
+}
+
+/* Stress cpumap management code by concurrently changing underlying cpumap */
+static void stress_cpumap(void)
+{
+	/* Changing qsize will cause kernel to free and alloc a new
+	 * bpf_cpu_map_entry, with an associated/complicated tear-down
+	 * procedure.
+	 */
+	create_cpu_entry(1,  1024, 0, false);
+	create_cpu_entry(1,   128, 0, false);
+	create_cpu_entry(1, 16000, 0, false);
+}
+
+static void stats_poll(int interval, bool use_separators, int prog_num,
+		       bool stress_mode)
+{
+	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);
+		if (stress_mode)
+			stress_cpumap();
+	}
+
+	free_stats_record(record);
+	free_stats_record(prev);
+}
+
+int main(int argc, char **argv)
+{
+	struct rlimit r = {10 * 1024 * 1024, RLIM_INFINITY};
+	bool use_separators = true;
+	bool stress_mode = false;
+	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;
+	}
+
+	mark_cpus_unavailable();
+
+	/* 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 'x':
+			stress_mode = 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, stress_mode);
+	return EXIT_OK;
+}

Re: [net-next V8 PATCH 0/5] New bpf cpumap type for XDP_REDIRECT

[David Miller]

From: Jesper Dangaard Brouer <brouer@redhat.com>
Date: Mon, 16 Oct 2017 12:19:23 +0200

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

Series applied, thanks Jesper.