[PATCH bpf-next v6 2/5] devmap/cpumap: Use flush list instead of bitmap

[PATCH bpf-next v6 2/5] devmap/cpumap: Use flush list instead of bitmap

[Toke Høiland-Jørgensen]

From: Toke Høiland-Jørgensen <toke@redhat.com>

The socket map uses a linked list instead of a bitmap to keep track of
which entries to flush. Do the same for devmap and cpumap, as this means we
don't have to care about the map index when enqueueing things into the
map (and so we can cache the map lookup).

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Acked-by: Andrii Nakryiko <andriin@fb.com>
---
 kernel/bpf/cpumap.c |  105 +++++++++++++++++++++++---------------------------
 kernel/bpf/devmap.c |  107 ++++++++++++++++++++++-----------------------------
 net/core/filter.c   |    2 -
 3 files changed, 95 insertions(+), 119 deletions(-)

diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index 8dff08768087..ef49e17ae47c 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -32,14 +32,19 @@
 
 /* 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
+ * guaranteed that queueing the frame and the 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 bpf_cpu_map_entry;
+struct bpf_cpu_map;
+
 struct xdp_bulk_queue {
 	void *q[CPU_MAP_BULK_SIZE];
+	struct list_head flush_node;
+	struct bpf_cpu_map_entry *obj;
 	unsigned int count;
 };
 
@@ -52,6 +57,8 @@ struct bpf_cpu_map_entry {
 	/* XDP can run multiple RX-ring queues, need __percpu enqueue store */
 	struct xdp_bulk_queue __percpu *bulkq;
 
+	struct bpf_cpu_map *cmap;
+
 	/* Queue with potential multi-producers, and single-consumer kthread */
 	struct ptr_ring *queue;
 	struct task_struct *kthread;
@@ -65,23 +72,17 @@ 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;
+	struct list_head __percpu *flush_list;
 };
 
-static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
-			     struct xdp_bulk_queue *bq, bool in_napi_ctx);
-
-static u64 cpu_map_bitmap_size(const union bpf_attr *attr)
-{
-	return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
-}
+static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx);
 
 static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
 {
 	struct bpf_cpu_map *cmap;
 	int err = -ENOMEM;
+	int ret, cpu;
 	u64 cost;
-	int ret;
 
 	if (!capable(CAP_SYS_ADMIN))
 		return ERR_PTR(-EPERM);
@@ -105,7 +106,7 @@ static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
 
 	/* make sure page count doesn't overflow */
 	cost = (u64) cmap->map.max_entries * sizeof(struct bpf_cpu_map_entry *);
-	cost += cpu_map_bitmap_size(attr) * num_possible_cpus();
+	cost += sizeof(struct list_head) * num_possible_cpus();
 
 	/* Notice returns -EPERM on if map size is larger than memlock limit */
 	ret = bpf_map_charge_init(&cmap->map.memory, cost);
@@ -114,12 +115,13 @@ static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
 		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)
+	cmap->flush_list = alloc_percpu(struct list_head);
+	if (!cmap->flush_list)
 		goto free_charge;
 
+	for_each_possible_cpu(cpu)
+		INIT_LIST_HEAD(per_cpu_ptr(cmap->flush_list, cpu));
+
 	/* Alloc array for possible remote "destination" CPUs */
 	cmap->cpu_map = bpf_map_area_alloc(cmap->map.max_entries *
 					   sizeof(struct bpf_cpu_map_entry *),
@@ -129,7 +131,7 @@ static struct bpf_map *cpu_map_alloc(union bpf_attr *attr)
 
 	return &cmap->map;
 free_percpu:
-	free_percpu(cmap->flush_needed);
+	free_percpu(cmap->flush_list);
 free_charge:
 	bpf_map_charge_finish(&cmap->map.memory);
 free_cmap:
@@ -334,7 +336,8 @@ static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
 {
 	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
 	struct bpf_cpu_map_entry *rcpu;
-	int numa, err;
+	struct xdp_bulk_queue *bq;
+	int numa, err, i;
 
 	/* Have map->numa_node, but choose node of redirect target CPU */
 	numa = cpu_to_node(cpu);
@@ -349,6 +352,11 @@ static struct bpf_cpu_map_entry *__cpu_map_entry_alloc(u32 qsize, u32 cpu,
 	if (!rcpu->bulkq)
 		goto free_rcu;
 
+	for_each_possible_cpu(i) {
+		bq = per_cpu_ptr(rcpu->bulkq, i);
+		bq->obj = rcpu;
+	}
+
 	/* Alloc queue */
 	rcpu->queue = kzalloc_node(sizeof(*rcpu->queue), gfp, numa);
 	if (!rcpu->queue)
@@ -405,7 +413,7 @@ static void __cpu_map_entry_free(struct rcu_head *rcu)
 		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, false);
+		bq_flush_to_queue(bq, false);
 	}
 	free_percpu(rcpu->bulkq);
 	/* Cannot kthread_stop() here, last put free rcpu resources */
@@ -488,6 +496,7 @@ static int cpu_map_update_elem(struct bpf_map *map, void *key, void *value,
 		rcpu = __cpu_map_entry_alloc(qsize, key_cpu, map->id);
 		if (!rcpu)
 			return -ENOMEM;
+		rcpu->cmap = cmap;
 	}
 	rcu_read_lock();
 	__cpu_map_entry_replace(cmap, key_cpu, rcpu);
@@ -514,14 +523,14 @@ static void cpu_map_free(struct bpf_map *map)
 	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.
+	 * list be empty on _all_ cpus. Because the above synchronize_rcu()
+	 * ensures the map is disconnected from the program we can assume no new
+	 * items will be added to the list.
 	 */
 	for_each_online_cpu(cpu) {
-		unsigned long *bitmap = per_cpu_ptr(cmap->flush_needed, cpu);
+		struct list_head *flush_list = per_cpu_ptr(cmap->flush_list, cpu);
 
-		while (!bitmap_empty(bitmap, cmap->map.max_entries))
+		while (!list_empty(flush_list))
 			cond_resched();
 	}
 
@@ -538,7 +547,7 @@ static void cpu_map_free(struct bpf_map *map)
 		/* bq flush and cleanup happens after RCU graze-period */
 		__cpu_map_entry_replace(cmap, i, NULL); /* call_rcu */
 	}
-	free_percpu(cmap->flush_needed);
+	free_percpu(cmap->flush_list);
 	bpf_map_area_free(cmap->cpu_map);
 	kfree(cmap);
 }
@@ -590,9 +599,9 @@ const struct bpf_map_ops cpu_map_ops = {
 	.map_check_btf		= map_check_no_btf,
 };
 
-static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
-			     struct xdp_bulk_queue *bq, bool in_napi_ctx)
+static int bq_flush_to_queue(struct xdp_bulk_queue *bq, bool in_napi_ctx)
 {
+	struct bpf_cpu_map_entry *rcpu = bq->obj;
 	unsigned int processed = 0, drops = 0;
 	const int to_cpu = rcpu->cpu;
 	struct ptr_ring *q;
@@ -621,6 +630,8 @@ static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
 	bq->count = 0;
 	spin_unlock(&q->producer_lock);
 
+	__list_del_clearprev(&bq->flush_node);
+
 	/* Feedback loop via tracepoints */
 	trace_xdp_cpumap_enqueue(rcpu->map_id, processed, drops, to_cpu);
 	return 0;
@@ -631,10 +642,11 @@ static int bq_flush_to_queue(struct bpf_cpu_map_entry *rcpu,
  */
 static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
 {
+	struct list_head *flush_list = this_cpu_ptr(rcpu->cmap->flush_list);
 	struct xdp_bulk_queue *bq = this_cpu_ptr(rcpu->bulkq);
 
 	if (unlikely(bq->count == CPU_MAP_BULK_SIZE))
-		bq_flush_to_queue(rcpu, bq, true);
+		bq_flush_to_queue(bq, true);
 
 	/* Notice, xdp_buff/page MUST be queued here, long enough for
 	 * driver to code invoking us to finished, due to driver
@@ -646,6 +658,10 @@ static int bq_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_frame *xdpf)
 	 * operation, when completing napi->poll call.
 	 */
 	bq->q[bq->count++] = xdpf;
+
+	if (!bq->flush_node.prev)
+		list_add(&bq->flush_node, flush_list);
+
 	return 0;
 }
 
@@ -665,41 +681,16 @@ int cpu_map_enqueue(struct bpf_cpu_map_entry *rcpu, struct xdp_buff *xdp,
 	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);
+	struct list_head *flush_list = this_cpu_ptr(cmap->flush_list);
+	struct xdp_bulk_queue *bq, *tmp;
 
-		/* Flush all frames in bulkq to real queue */
-		bq = this_cpu_ptr(rcpu->bulkq);
-		bq_flush_to_queue(rcpu, bq, true);
+	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
+		bq_flush_to_queue(bq, true);
 
 		/* If already running, costs spin_lock_irqsave + smb_mb */
-		wake_up_process(rcpu->kthread);
+		wake_up_process(bq->obj->kthread);
 	}
 }
diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c
index 40e86a7e0ef0..a4dddc867cbf 100644
--- a/kernel/bpf/devmap.c
+++ b/kernel/bpf/devmap.c
@@ -17,9 +17,8 @@
  * datapath always has a valid copy. However, the datapath does a "flush"
  * operation that pushes any pending packets in the driver outside the RCU
  * critical section. Each bpf_dtab_netdev tracks these pending operations using
- * an atomic per-cpu bitmap. The bpf_dtab_netdev object will not be destroyed
- * until all bits are cleared indicating outstanding flush operations have
- * completed.
+ * a per-cpu flush list. The bpf_dtab_netdev object will not be destroyed  until
+ * this list is empty, indicating outstanding flush operations have completed.
  *
  * BPF syscalls may race with BPF program calls on any of the update, delete
  * or lookup operations. As noted above the xchg() operation also keep the
@@ -48,9 +47,13 @@
 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
 
 #define DEV_MAP_BULK_SIZE 16
+struct bpf_dtab_netdev;
+
 struct xdp_bulk_queue {
 	struct xdp_frame *q[DEV_MAP_BULK_SIZE];
+	struct list_head flush_node;
 	struct net_device *dev_rx;
+	struct bpf_dtab_netdev *obj;
 	unsigned int count;
 };
 
@@ -65,23 +68,18 @@ struct bpf_dtab_netdev {
 struct bpf_dtab {
 	struct bpf_map map;
 	struct bpf_dtab_netdev **netdev_map;
-	unsigned long __percpu *flush_needed;
+	struct list_head __percpu *flush_list;
 	struct list_head list;
 };
 
 static DEFINE_SPINLOCK(dev_map_lock);
 static LIST_HEAD(dev_map_list);
 
-static u64 dev_map_bitmap_size(const union bpf_attr *attr)
-{
-	return BITS_TO_LONGS((u64) attr->max_entries) * sizeof(unsigned long);
-}
-
 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 {
 	struct bpf_dtab *dtab;
+	int err, cpu;
 	u64 cost;
-	int err;
 
 	if (!capable(CAP_NET_ADMIN))
 		return ERR_PTR(-EPERM);
@@ -99,7 +97,7 @@ static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 
 	/* make sure page count doesn't overflow */
 	cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
-	cost += dev_map_bitmap_size(attr) * num_possible_cpus();
+	cost += sizeof(struct list_head) * num_possible_cpus();
 
 	/* if map size is larger than memlock limit, reject it */
 	err = bpf_map_charge_init(&dtab->map.memory, cost);
@@ -108,28 +106,30 @@ static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 
 	err = -ENOMEM;
 
-	/* A per cpu bitfield with a bit per possible net device */
-	dtab->flush_needed = __alloc_percpu_gfp(dev_map_bitmap_size(attr),
-						__alignof__(unsigned long),
-						GFP_KERNEL | __GFP_NOWARN);
-	if (!dtab->flush_needed)
+	dtab->flush_list = alloc_percpu(struct list_head);
+	if (!dtab->flush_list)
 		goto free_charge;
 
+	for_each_possible_cpu(cpu)
+		INIT_LIST_HEAD(per_cpu_ptr(dtab->flush_list, cpu));
+
 	dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
 					      sizeof(struct bpf_dtab_netdev *),
 					      dtab->map.numa_node);
 	if (!dtab->netdev_map)
-		goto free_charge;
+		goto free_percpu;
 
 	spin_lock(&dev_map_lock);
 	list_add_tail_rcu(&dtab->list, &dev_map_list);
 	spin_unlock(&dev_map_lock);
 
 	return &dtab->map;
+
+free_percpu:
+	free_percpu(dtab->flush_list);
 free_charge:
 	bpf_map_charge_finish(&dtab->map.memory);
 free_dtab:
-	free_percpu(dtab->flush_needed);
 	kfree(dtab);
 	return ERR_PTR(err);
 }
@@ -158,14 +158,14 @@ static void dev_map_free(struct bpf_map *map)
 	rcu_barrier();
 
 	/* To ensure all pending flush operations have completed wait for flush
-	 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
+	 * list to empty 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.
+	 * from the program we can assume no new items will be added.
 	 */
 	for_each_online_cpu(cpu) {
-		unsigned long *bitmap = per_cpu_ptr(dtab->flush_needed, cpu);
+		struct list_head *flush_list = per_cpu_ptr(dtab->flush_list, cpu);
 
-		while (!bitmap_empty(bitmap, dtab->map.max_entries))
+		while (!list_empty(flush_list))
 			cond_resched();
 	}
 
@@ -181,7 +181,7 @@ static void dev_map_free(struct bpf_map *map)
 		kfree(dev);
 	}
 
-	free_percpu(dtab->flush_needed);
+	free_percpu(dtab->flush_list);
 	bpf_map_area_free(dtab->netdev_map);
 	kfree(dtab);
 }
@@ -203,18 +203,10 @@ static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 	return 0;
 }
 
-void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
-{
-	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
-	unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
-
-	__set_bit(bit, bitmap);
-}
-
-static int bq_xmit_all(struct bpf_dtab_netdev *obj,
-		       struct xdp_bulk_queue *bq, u32 flags,
+static int bq_xmit_all(struct xdp_bulk_queue *bq, u32 flags,
 		       bool in_napi_ctx)
 {
+	struct bpf_dtab_netdev *obj = bq->obj;
 	struct net_device *dev = obj->dev;
 	int sent = 0, drops = 0, err = 0;
 	int i;
@@ -241,6 +233,7 @@ static int bq_xmit_all(struct bpf_dtab_netdev *obj,
 	trace_xdp_devmap_xmit(&obj->dtab->map, obj->bit,
 			      sent, drops, bq->dev_rx, dev, err);
 	bq->dev_rx = NULL;
+	__list_del_clearprev(&bq->flush_node);
 	return 0;
 error:
 	/* If ndo_xdp_xmit fails with an errno, no frames have been
@@ -263,31 +256,18 @@ static int bq_xmit_all(struct bpf_dtab_netdev *obj,
  * from the driver before returning from its napi->poll() routine. The poll()
  * routine is called either from busy_poll context or net_rx_action signaled
  * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
- * net device can be torn down. On devmap tear down we ensure the ctx bitmap
- * is zeroed before completing to ensure all flush operations have completed.
+ * net device can be torn down. On devmap tear down we ensure the flush list
+ * is empty before completing to ensure all flush operations have completed.
  */
 void __dev_map_flush(struct bpf_map *map)
 {
 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
-	unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
-	u32 bit;
+	struct list_head *flush_list = this_cpu_ptr(dtab->flush_list);
+	struct xdp_bulk_queue *bq, *tmp;
 
 	rcu_read_lock();
-	for_each_set_bit(bit, bitmap, map->max_entries) {
-		struct bpf_dtab_netdev *dev = READ_ONCE(dtab->netdev_map[bit]);
-		struct xdp_bulk_queue *bq;
-
-		/* This is possible if the dev entry is removed by user space
-		 * between xdp redirect and flush op.
-		 */
-		if (unlikely(!dev))
-			continue;
-
-		bq = this_cpu_ptr(dev->bulkq);
-		bq_xmit_all(dev, bq, XDP_XMIT_FLUSH, true);
-
-		__clear_bit(bit, bitmap);
-	}
+	list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
+		bq_xmit_all(bq, XDP_XMIT_FLUSH, true);
 	rcu_read_unlock();
 }
 
@@ -314,10 +294,11 @@ static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
 		      struct net_device *dev_rx)
 
 {
+	struct list_head *flush_list = this_cpu_ptr(obj->dtab->flush_list);
 	struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
 
 	if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
-		bq_xmit_all(obj, bq, 0, true);
+		bq_xmit_all(bq, 0, true);
 
 	/* Ingress dev_rx will be the same for all xdp_frame's in
 	 * bulk_queue, because bq stored per-CPU and must be flushed
@@ -327,6 +308,10 @@ static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
 		bq->dev_rx = dev_rx;
 
 	bq->q[bq->count++] = xdpf;
+
+	if (!bq->flush_node.prev)
+		list_add(&bq->flush_node, flush_list);
+
 	return 0;
 }
 
@@ -377,17 +362,12 @@ static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
 {
 	if (dev->dev->netdev_ops->ndo_xdp_xmit) {
 		struct xdp_bulk_queue *bq;
-		unsigned long *bitmap;
-
 		int cpu;
 
 		rcu_read_lock();
 		for_each_online_cpu(cpu) {
-			bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
-			__clear_bit(dev->bit, bitmap);
-
 			bq = per_cpu_ptr(dev->bulkq, cpu);
-			bq_xmit_all(dev, bq, XDP_XMIT_FLUSH, false);
+			bq_xmit_all(bq, XDP_XMIT_FLUSH, false);
 		}
 		rcu_read_unlock();
 	}
@@ -434,8 +414,10 @@ static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 	struct net *net = current->nsproxy->net_ns;
 	gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
 	struct bpf_dtab_netdev *dev, *old_dev;
-	u32 i = *(u32 *)key;
 	u32 ifindex = *(u32 *)value;
+	struct xdp_bulk_queue *bq;
+	u32 i = *(u32 *)key;
+	int cpu;
 
 	if (unlikely(map_flags > BPF_EXIST))
 		return -EINVAL;
@@ -458,6 +440,11 @@ static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 			return -ENOMEM;
 		}
 
+		for_each_possible_cpu(cpu) {
+			bq = per_cpu_ptr(dev->bulkq, cpu);
+			bq->obj = dev;
+		}
+
 		dev->dev = dev_get_by_index(net, ifindex);
 		if (!dev->dev) {
 			free_percpu(dev->bulkq);
diff --git a/net/core/filter.c b/net/core/filter.c
index 2014d76e0d2a..183bf4d8e301 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -3523,7 +3523,6 @@ static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd,
 		err = dev_map_enqueue(dst, xdp, dev_rx);
 		if (unlikely(err))
 			return err;
-		__dev_map_insert_ctx(map, index);
 		break;
 	}
 	case BPF_MAP_TYPE_CPUMAP: {
@@ -3532,7 +3531,6 @@ static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd,
 		err = cpu_map_enqueue(rcpu, xdp, dev_rx);
 		if (unlikely(err))
 			return err;
-		__cpu_map_insert_ctx(map, index);
 		break;
 	}
 	case BPF_MAP_TYPE_XSKMAP: {

[PATCH bpf-next v6 0/5] xdp: Allow lookup into devmaps before redirect

[Toke Høiland-Jørgensen]

When using the bpf_redirect_map() helper to redirect packets from XDP, the eBPF
program cannot currently know whether the redirect will succeed, which makes it
impossible to gracefully handle errors. To properly fix this will probably
require deeper changes to the way TX resources are allocated, but one thing that
is fairly straight forward to fix is to allow lookups into devmaps, so programs
can at least know when a redirect is *guaranteed* to fail because there is no
entry in the map. Currently, programs work around this by keeping a shadow map
of another type which indicates whether a map index is valid.

This series contains two changes that are complementary ways to fix this issue:

- Moving the map lookup into the bpf_redirect_map() helper (and caching the
  result), so the helper can return an error if no value is found in the map.
  This includes a refactoring of the devmap and cpumap code to not care about
  the index on enqueue.

- Allowing regular lookups into devmaps from eBPF programs, using the read-only
  flag to make sure they don't change the values.

The performance impact of the series is negligible, in the sense that I cannot
measure it because the variance between test runs is higher than the difference
pre/post series.

Changelog:

v6:
  - Factor out list handling in maps to a helper in list.h (new patch 1)
  - Rename variables in struct bpf_redirect_info (new patch 3 + patch 4)
  - Explain why we are clearing out the map in the info struct on lookup failure
  - Remove unneeded check for forwarding target in tracepoint macro

v5:
  - Rebase on latest bpf-next.
  - Update documentation for bpf_redirect_map() with the new meaning of flags.

v4:
  - Fix a few nits from Andrii
  - Lose the #defines in bpf.h and just compare the flags argument directly to
    XDP_TX in bpf_xdp_redirect_map().

v3:
  - Adopt Jonathan's idea of using the lower two bits of the flag value as the
    return code.
  - Always do the lookup, and cache the result for use in xdp_do_redirect(); to
    achieve this, refactor the devmap and cpumap code to get rid the bitmap for
    selecting which devices to flush.

v2:
  - For patch 1, make it clear that the change works for any map type.
  - For patch 2, just use the new BPF_F_RDONLY_PROG flag to make the return
    value read-only.

---

Toke Høiland-Jørgensen (5):
      xskmap: Move non-standard list manipulation to helper
      devmap/cpumap: Use flush list instead of bitmap
      devmap: Rename ifindex member in bpf_redirect_info
      bpf_xdp_redirect_map: Perform map lookup in eBPF helper
      devmap: Allow map lookups from eBPF


 include/linux/filter.h     |    3 +
 include/linux/list.h       |   14 ++++++
 include/trace/events/xdp.h |    5 +-
 include/uapi/linux/bpf.h   |    7 ++-
 kernel/bpf/cpumap.c        |  105 +++++++++++++++++++----------------------
 kernel/bpf/devmap.c        |  112 ++++++++++++++++++++------------------------
 kernel/bpf/verifier.c      |    7 +--
 kernel/bpf/xskmap.c        |    3 -
 net/core/filter.c          |   60 ++++++++++++------------
 9 files changed, 157 insertions(+), 159 deletions(-)

[PATCH bpf-next v6 4/5] bpf_xdp_redirect_map: Perform map lookup in eBPF helper

[Toke Høiland-Jørgensen]

From: Toke Høiland-Jørgensen <toke@redhat.com>

The bpf_redirect_map() helper used by XDP programs doesn't return any
indication of whether it can successfully redirect to the map index it was
given. Instead, BPF programs have to track this themselves, leading to
programs using duplicate maps to track which entries are populated in the
devmap.

This patch fixes this by moving the map lookup into the bpf_redirect_map()
helper, which makes it possible to return failure to the eBPF program. The
lower bits of the flags argument is used as the return code, which means
that existing users who pass a '0' flag argument will get XDP_ABORTED.

With this, a BPF program can check the return code from the helper call and
react by, for instance, substituting a different redirect. This works for
any type of map used for redirect.

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Acked-by: Andrii Nakryiko <andriin@fb.com>
---
 include/linux/filter.h     |    1 +
 include/trace/events/xdp.h |    5 ++---
 include/uapi/linux/bpf.h   |    7 +++++--
 net/core/filter.c          |   32 ++++++++++++++++++--------------
 4 files changed, 26 insertions(+), 19 deletions(-)

diff --git a/include/linux/filter.h b/include/linux/filter.h
index af8601340dae..f2a03f232386 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h
@@ -580,6 +580,7 @@ struct bpf_skb_data_end {
 struct bpf_redirect_info {
 	u32 flags;
 	u32 tgt_index;
+	void *tgt_value;
 	struct bpf_map *map;
 	struct bpf_map *map_to_flush;
 	u32 kern_flags;
diff --git a/include/trace/events/xdp.h b/include/trace/events/xdp.h
index bb5e380e2ef3..393f3d7ca819 100644
--- a/include/trace/events/xdp.h
+++ b/include/trace/events/xdp.h
@@ -146,9 +146,8 @@ struct _bpf_dtab_netdev {
 #endif /* __DEVMAP_OBJ_TYPE */
 
 #define devmap_ifindex(fwd, map)				\
-	(!fwd ? 0 :						\
-	 ((map->map_type == BPF_MAP_TYPE_DEVMAP) ?		\
-	  ((struct _bpf_dtab_netdev *)fwd)->dev->ifindex : 0))
+	((map->map_type == BPF_MAP_TYPE_DEVMAP) ?		\
+	  ((struct _bpf_dtab_netdev *)fwd)->dev->ifindex : 0)
 
 #define _trace_xdp_redirect_map(dev, xdp, fwd, map, idx)		\
 	 trace_xdp_redirect_map(dev, xdp, devmap_ifindex(fwd, map),	\
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index b077507efa3f..59f3fe61b2b0 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -1568,8 +1568,11 @@ union bpf_attr {
  * 		but this is only implemented for native XDP (with driver
  * 		support) as of this writing).
  *
- * 		All values for *flags* are reserved for future usage, and must
- * 		be left at zero.
+ * 		The lower two bits of *flags* are used as the return code if
+ * 		the map lookup fails. This is so that the return value can be
+ * 		one of the XDP program return codes up to XDP_TX, as chosen by
+ * 		the caller. Any higher bits in the *flags* argument must be
+ * 		unset.
  *
  * 		When used to redirect packets to net devices, this helper
  * 		provides a high performance increase over **bpf_redirect**\ ().
diff --git a/net/core/filter.c b/net/core/filter.c
index 1d7f5f7f32cf..f9fcc8168e1c 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -3605,17 +3605,13 @@ static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
 			       struct bpf_redirect_info *ri)
 {
 	u32 index = ri->tgt_index;
-	void *fwd = NULL;
+	void *fwd = ri->tgt_value;
 	int err;
 
 	ri->tgt_index = 0;
+	ri->tgt_value = NULL;
 	WRITE_ONCE(ri->map, NULL);
 
-	fwd = __xdp_map_lookup_elem(map, index);
-	if (unlikely(!fwd)) {
-		err = -EINVAL;
-		goto err;
-	}
 	if (ri->map_to_flush && unlikely(ri->map_to_flush != map))
 		xdp_do_flush_map();
 
@@ -3652,18 +3648,13 @@ static int xdp_do_generic_redirect_map(struct net_device *dev,
 {
 	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
 	u32 index = ri->tgt_index;
-	void *fwd = NULL;
+	void *fwd = ri->tgt_value;
 	int err = 0;
 
 	ri->tgt_index = 0;
+	ri->tgt_value = NULL;
 	WRITE_ONCE(ri->map, NULL);
 
-	fwd = __xdp_map_lookup_elem(map, index);
-	if (unlikely(!fwd)) {
-		err = -EINVAL;
-		goto err;
-	}
-
 	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 		struct bpf_dtab_netdev *dst = fwd;
 
@@ -3732,6 +3723,7 @@ BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
 
 	ri->flags = flags;
 	ri->tgt_index = ifindex;
+	ri->tgt_value = NULL;
 	WRITE_ONCE(ri->map, NULL);
 
 	return XDP_REDIRECT;
@@ -3750,9 +3742,21 @@ BPF_CALL_3(bpf_xdp_redirect_map, struct bpf_map *, map, u32, ifindex,
 {
 	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
 
-	if (unlikely(flags))
+	/* Lower bits of the flags are used as return code on lookup failure */
+	if (unlikely(flags > XDP_TX))
 		return XDP_ABORTED;
 
+	ri->tgt_value = __xdp_map_lookup_elem(map, ifindex);
+	if (unlikely(!ri->tgt_value)) {
+		/* If the lookup fails we want to clear out the state in the
+		 * redirect_info struct completely, so that if an eBPF program
+		 * performs multiple lookups, the last one always takes
+		 * precedence.
+		 */
+		WRITE_ONCE(ri->map, NULL);
+		return flags;
+	}
+
 	ri->flags = flags;
 	ri->tgt_index = ifindex;
 	WRITE_ONCE(ri->map, map);

[PATCH bpf-next v6 3/5] devmap: Rename ifindex member in bpf_redirect_info

[Toke Høiland-Jørgensen]

From: Toke Høiland-Jørgensen <toke@redhat.com>

The bpf_redirect_info struct has an 'ifindex' member which was named back
when the redirects could only target egress interfaces. Now that we can
also redirect to sockets and CPUs, this is a bit misleading, so rename the
member to tgt_index.

Reorder the struct members so we can have 'tgt_index' and 'tgt_value' next
to each other in a subsequent patch.

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
---
 include/linux/filter.h |    2 +-
 net/core/filter.c      |   26 +++++++++++++-------------
 2 files changed, 14 insertions(+), 14 deletions(-)

diff --git a/include/linux/filter.h b/include/linux/filter.h
index 43b45d6db36d..af8601340dae 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h
@@ -578,8 +578,8 @@ struct bpf_skb_data_end {
 };
 
 struct bpf_redirect_info {
-	u32 ifindex;
 	u32 flags;
+	u32 tgt_index;
 	struct bpf_map *map;
 	struct bpf_map *map_to_flush;
 	u32 kern_flags;
diff --git a/net/core/filter.c b/net/core/filter.c
index 183bf4d8e301..1d7f5f7f32cf 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -2158,8 +2158,8 @@ BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
 	if (unlikely(flags & ~(BPF_F_INGRESS)))
 		return TC_ACT_SHOT;
 
-	ri->ifindex = ifindex;
 	ri->flags = flags;
+	ri->tgt_index = ifindex;
 
 	return TC_ACT_REDIRECT;
 }
@@ -2169,8 +2169,8 @@ int skb_do_redirect(struct sk_buff *skb)
 	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
 	struct net_device *dev;
 
-	dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
-	ri->ifindex = 0;
+	dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->tgt_index);
+	ri->tgt_index = 0;
 	if (unlikely(!dev)) {
 		kfree_skb(skb);
 		return -EINVAL;
@@ -3488,11 +3488,11 @@ xdp_do_redirect_slow(struct net_device *dev, struct xdp_buff *xdp,
 		     struct bpf_prog *xdp_prog, struct bpf_redirect_info *ri)
 {
 	struct net_device *fwd;
-	u32 index = ri->ifindex;
+	u32 index = ri->tgt_index;
 	int err;
 
 	fwd = dev_get_by_index_rcu(dev_net(dev), index);
-	ri->ifindex = 0;
+	ri->tgt_index = 0;
 	if (unlikely(!fwd)) {
 		err = -EINVAL;
 		goto err;
@@ -3604,11 +3604,11 @@ static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
 			       struct bpf_prog *xdp_prog, struct bpf_map *map,
 			       struct bpf_redirect_info *ri)
 {
-	u32 index = ri->ifindex;
+	u32 index = ri->tgt_index;
 	void *fwd = NULL;
 	int err;
 
-	ri->ifindex = 0;
+	ri->tgt_index = 0;
 	WRITE_ONCE(ri->map, NULL);
 
 	fwd = __xdp_map_lookup_elem(map, index);
@@ -3651,11 +3651,11 @@ static int xdp_do_generic_redirect_map(struct net_device *dev,
 				       struct bpf_map *map)
 {
 	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
-	u32 index = ri->ifindex;
+	u32 index = ri->tgt_index;
 	void *fwd = NULL;
 	int err = 0;
 
-	ri->ifindex = 0;
+	ri->tgt_index = 0;
 	WRITE_ONCE(ri->map, NULL);
 
 	fwd = __xdp_map_lookup_elem(map, index);
@@ -3695,14 +3695,14 @@ int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
 {
 	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
 	struct bpf_map *map = READ_ONCE(ri->map);
-	u32 index = ri->ifindex;
+	u32 index = ri->tgt_index;
 	struct net_device *fwd;
 	int err = 0;
 
 	if (map)
 		return xdp_do_generic_redirect_map(dev, skb, xdp, xdp_prog,
 						   map);
-	ri->ifindex = 0;
+	ri->tgt_index = 0;
 	fwd = dev_get_by_index_rcu(dev_net(dev), index);
 	if (unlikely(!fwd)) {
 		err = -EINVAL;
@@ -3730,8 +3730,8 @@ BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
 	if (unlikely(flags))
 		return XDP_ABORTED;
 
-	ri->ifindex = ifindex;
 	ri->flags = flags;
+	ri->tgt_index = ifindex;
 	WRITE_ONCE(ri->map, NULL);
 
 	return XDP_REDIRECT;
@@ -3753,8 +3753,8 @@ BPF_CALL_3(bpf_xdp_redirect_map, struct bpf_map *, map, u32, ifindex,
 	if (unlikely(flags))
 		return XDP_ABORTED;
 
-	ri->ifindex = ifindex;
 	ri->flags = flags;
+	ri->tgt_index = ifindex;
 	WRITE_ONCE(ri->map, map);
 
 	return XDP_REDIRECT;

[PATCH bpf-next v6 1/5] xskmap: Move non-standard list manipulation to helper

[Toke Høiland-Jørgensen]

From: Toke Høiland-Jørgensen <toke@redhat.com>

Add a helper in list.h for the non-standard way of clearing a list that is
used in xskmap. This makes it easier to reuse it in the other map types,
and also makes sure this usage is not forgotten in any list refactorings in
the future.

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
---
 include/linux/list.h |   14 ++++++++++++++
 kernel/bpf/xskmap.c  |    3 +--
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/include/linux/list.h b/include/linux/list.h
index e951228db4b2..85c92555e31f 100644
--- a/include/linux/list.h
+++ b/include/linux/list.h
@@ -106,6 +106,20 @@ static inline void __list_del(struct list_head * prev, struct list_head * next)
 	WRITE_ONCE(prev->next, next);
 }
 
+/*
+ * Delete a list entry and clear the 'prev' pointer.
+ *
+ * This is a special-purpose list clearing method used in the networking code
+ * for lists allocated as per-cpu, where we don't want to incur the extra
+ * WRITE_ONCE() overhead of a regular list_del_init(). The code that uses this
+ * needs to check the node 'prev' pointer instead of calling list_empty().
+ */
+static inline void __list_del_clearprev(struct list_head *entry)
+{
+	__list_del(entry->prev, entry->next);
+	entry->prev = NULL;
+}
+
 /**
  * list_del - deletes entry from list.
  * @entry: the element to delete from the list.
diff --git a/kernel/bpf/xskmap.c b/kernel/bpf/xskmap.c
index ef7338cebd18..9bb96ace9fa1 100644
--- a/kernel/bpf/xskmap.c
+++ b/kernel/bpf/xskmap.c
@@ -145,8 +145,7 @@ void __xsk_map_flush(struct bpf_map *map)
 
 	list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
 		xsk_flush(xs);
-		__list_del(xs->flush_node.prev, xs->flush_node.next);
-		xs->flush_node.prev = NULL;
+		__list_del_clearprev(&xs->flush_node);
 	}
 }
 

[PATCH bpf-next v6 5/5] devmap: Allow map lookups from eBPF

[Toke Høiland-Jørgensen]

From: Toke Høiland-Jørgensen <toke@redhat.com>

We don't currently allow lookups into a devmap from eBPF, because the map
lookup returns a pointer directly to the dev->ifindex, which shouldn't be
modifiable from eBPF.

However, being able to do lookups in devmaps is useful to know (e.g.)
whether forwarding to a specific interface is enabled. Currently, programs
work around this by keeping a shadow map of another type which indicates
whether a map index is valid.

Since we now have a flag to make maps read-only from the eBPF side, we can
simply lift the lookup restriction if we make sure this flag is always set.

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Acked-by: Jonathan Lemon <jonathan.lemon@gmail.com>
Acked-by: Andrii Nakryiko <andriin@fb.com>
---
 kernel/bpf/devmap.c   |    5 +++++
 kernel/bpf/verifier.c |    7 ++-----
 2 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c
index a4dddc867cbf..d83cf8ccc872 100644
--- a/kernel/bpf/devmap.c
+++ b/kernel/bpf/devmap.c
@@ -89,6 +89,11 @@ static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 	    attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
 		return ERR_PTR(-EINVAL);
 
+	/* Lookup returns a pointer straight to dev->ifindex, so make sure the
+	 * verifier prevents writes from the BPF side
+	 */
+	attr->map_flags |= BPF_F_RDONLY_PROG;
+
 	dtab = kzalloc(sizeof(*dtab), GFP_USER);
 	if (!dtab)
 		return ERR_PTR(-ENOMEM);
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 0e079b2298f8..51c327ee7d3f 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -3407,12 +3407,9 @@ static int check_map_func_compatibility(struct bpf_verifier_env *env,
 		if (func_id != BPF_FUNC_get_local_storage)
 			goto error;
 		break;
-	/* devmap returns a pointer to a live net_device ifindex that we cannot
-	 * allow to be modified from bpf side. So do not allow lookup elements
-	 * for now.
-	 */
 	case BPF_MAP_TYPE_DEVMAP:
-		if (func_id != BPF_FUNC_redirect_map)
+		if (func_id != BPF_FUNC_redirect_map &&
+		    func_id != BPF_FUNC_map_lookup_elem)
 			goto error;
 		break;
 	/* Restrict bpf side of cpumap and xskmap, open when use-cases

Re: [PATCH bpf-next v6 0/5] xdp: Allow lookup into devmaps before redirect

[Daniel Borkmann]

On 06/28/2019 11:12 AM, Toke Høiland-Jørgensen wrote:
> When using the bpf_redirect_map() helper to redirect packets from XDP, the eBPF
> program cannot currently know whether the redirect will succeed, which makes it
> impossible to gracefully handle errors. To properly fix this will probably
> require deeper changes to the way TX resources are allocated, but one thing that
> is fairly straight forward to fix is to allow lookups into devmaps, so programs
> can at least know when a redirect is *guaranteed* to fail because there is no
> entry in the map. Currently, programs work around this by keeping a shadow map
> of another type which indicates whether a map index is valid.
> 
> This series contains two changes that are complementary ways to fix this issue:
> 
> - Moving the map lookup into the bpf_redirect_map() helper (and caching the
>   result), so the helper can return an error if no value is found in the map.
>   This includes a refactoring of the devmap and cpumap code to not care about
>   the index on enqueue.
> 
> - Allowing regular lookups into devmaps from eBPF programs, using the read-only
>   flag to make sure they don't change the values.
> 
> The performance impact of the series is negligible, in the sense that I cannot
> measure it because the variance between test runs is higher than the difference
> pre/post series.
> 
> Changelog:
> 
> v6:
>   - Factor out list handling in maps to a helper in list.h (new patch 1)
>   - Rename variables in struct bpf_redirect_info (new patch 3 + patch 4)
>   - Explain why we are clearing out the map in the info struct on lookup failure
>   - Remove unneeded check for forwarding target in tracepoint macro
> 
> v5:
>   - Rebase on latest bpf-next.
>   - Update documentation for bpf_redirect_map() with the new meaning of flags.
> 
> v4:
>   - Fix a few nits from Andrii
>   - Lose the #defines in bpf.h and just compare the flags argument directly to
>     XDP_TX in bpf_xdp_redirect_map().
> 
> v3:
>   - Adopt Jonathan's idea of using the lower two bits of the flag value as the
>     return code.
>   - Always do the lookup, and cache the result for use in xdp_do_redirect(); to
>     achieve this, refactor the devmap and cpumap code to get rid the bitmap for
>     selecting which devices to flush.
> 
> v2:
>   - For patch 1, make it clear that the change works for any map type.
>   - For patch 2, just use the new BPF_F_RDONLY_PROG flag to make the return
>     value read-only.
> 
> ---
> 
> Toke Høiland-Jørgensen (5):
>       xskmap: Move non-standard list manipulation to helper
>       devmap/cpumap: Use flush list instead of bitmap
>       devmap: Rename ifindex member in bpf_redirect_info
>       bpf_xdp_redirect_map: Perform map lookup in eBPF helper
>       devmap: Allow map lookups from eBPF
> 
> 
>  include/linux/filter.h     |    3 +
>  include/linux/list.h       |   14 ++++++
>  include/trace/events/xdp.h |    5 +-
>  include/uapi/linux/bpf.h   |    7 ++-
>  kernel/bpf/cpumap.c        |  105 +++++++++++++++++++----------------------
>  kernel/bpf/devmap.c        |  112 ++++++++++++++++++++------------------------
>  kernel/bpf/verifier.c      |    7 +--
>  kernel/bpf/xskmap.c        |    3 -
>  net/core/filter.c          |   60 ++++++++++++------------
>  9 files changed, 157 insertions(+), 159 deletions(-)
> 

Applied, thanks!