The XDP redirect process is two staged:
- bpf_prog_run_xdp() is invoked to run a eBPF program which inspects the
  packet and makes decisions. While doing that, the per-CPU variable
  bpf_redirect_info is used.

- Afterwards xdp_do_redirect() is invoked and accesses bpf_redirect_info
  and it may also access other per-CPU variables like xskmap_flush_list.

At the very end of the NAPI callback, xdp_do_flush() is invoked which
does not access bpf_redirect_info but will touch the individual per-CPU
lists.

The per-CPU variables are only used in the NAPI callback hence disabling
bottom halves is the only protection mechanism. Users from preemptible
context (like cpu_map_kthread_run()) explicitly disable bottom halves
for protections reasons.
Without locking in local_bh_disable() on PREEMPT_RT this data structure
requires explicit locking.

PREEMPT_RT has forced-threaded interrupts enabled and every
NAPI-callback runs in a thread. If each thread has its own data
structure then locking can be avoided.

Create a struct bpf_net_context which contains struct bpf_redirect_info.
Define the variable on stack, use bpf_net_ctx_set() to save a pointer to
it. Use the __free() annotation to automatically reset the pointer once
function returns.
The bpf_net_ctx_set() may nest. For instance a function can be used from
within NET_RX_SOFTIRQ/ net_rx_action which uses bpf_net_ctx_set() and
NET_TX_SOFTIRQ which does not. Therefore only the first invocations
updates the pointer.
Use bpf_net_ctx_get_ri() as a wrapper to retrieve the current struct
bpf_redirect_info.

On PREEMPT_RT the pointer to bpf_net_context is saved task's
task_struct. On non-PREEMPT_RT builds the pointer saved in a per-CPU
variable (which is always NODE-local memory). Using always the
bpf_net_context approach has the advantage that there is almost zero
differences between PREEMPT_RT and non-PREEMPT_RT builds.

Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Eduard Zingerman <eddyz87@gmail.com>
Cc: Hao Luo <haoluo@google.com>
Cc: Jesper Dangaard Brouer <hawk@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: John Fastabend <john.fastabend@gmail.com>
Cc: KP Singh <kpsingh@kernel.org>
Cc: Martin KaFai Lau <martin.lau@linux.dev>
Cc: Song Liu <song@kernel.org>
Cc: Stanislav Fomichev <sdf@google.com>
Cc: Toke Høiland-Jørgensen <toke@redhat.com>
Cc: Yonghong Song <yonghong.song@linux.dev>
Cc: bpf@vger.kernel.org
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 include/linux/filter.h | 42 ++++++++++++++++++++++++++++++++-----
 include/linux/sched.h  |  3 +++
 kernel/bpf/cpumap.c    |  3 +++
 kernel/fork.c          |  1 +
 net/bpf/test_run.c     | 11 +++++++++-
 net/core/dev.c         | 19 ++++++++++++++++-
 net/core/filter.c      | 47 +++++++++++++++++++-----------------------
 net/core/lwt_bpf.c     |  3 +++
 8 files changed, 96 insertions(+), 33 deletions(-)

diff --git a/include/linux/filter.h b/include/linux/filter.h
index b02aea291b7e8..a0d0ea356f925 100644
--- a/include/linux/filter.h
+++ b/include/linux/filter.h
@@ -744,7 +744,39 @@ struct bpf_redirect_info {
 	struct bpf_nh_params nh;
 };
 
-DECLARE_PER_CPU(struct bpf_redirect_info, bpf_redirect_info);
+struct bpf_net_context {
+	struct bpf_redirect_info ri;
+};
+
+static inline struct bpf_net_context *bpf_net_ctx_set(struct bpf_net_context *bpf_net_ctx)
+{
+	struct task_struct *tsk = current;
+
+	if (tsk->bpf_net_context != NULL)
+		return NULL;
+	tsk->bpf_net_context = bpf_net_ctx;
+	return bpf_net_ctx;
+}
+
+static inline void bpf_net_ctx_clear(struct bpf_net_context *bpf_net_ctx)
+{
+	if (bpf_net_ctx)
+		current->bpf_net_context = NULL;
+}
+
+static inline struct bpf_net_context *bpf_net_ctx_get(void)
+{
+	return current->bpf_net_context;
+}
+
+static inline struct bpf_redirect_info *bpf_net_ctx_get_ri(void)
+{
+	struct bpf_net_context *bpf_net_ctx = bpf_net_ctx_get();
+
+	return &bpf_net_ctx->ri;
+}
+
+DEFINE_FREE(bpf_net_ctx_clear, struct bpf_net_context *, bpf_net_ctx_clear(_T));
 
 /* flags for bpf_redirect_info kern_flags */
 #define BPF_RI_F_RF_NO_DIRECT	BIT(0)	/* no napi_direct on return_frame */
@@ -1022,21 +1054,21 @@ void bpf_clear_redirect_map(struct bpf_map *map);
 
 static inline bool xdp_return_frame_no_direct(void)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	return ri->kern_flags & BPF_RI_F_RF_NO_DIRECT;
 }
 
 static inline void xdp_set_return_frame_no_direct(void)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	ri->kern_flags |= BPF_RI_F_RF_NO_DIRECT;
 }
 
 static inline void xdp_clear_return_frame_no_direct(void)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	ri->kern_flags &= ~BPF_RI_F_RF_NO_DIRECT;
 }
@@ -1592,7 +1624,7 @@ static __always_inline long __bpf_xdp_redirect_map(struct bpf_map *map, u64 inde
 						   u64 flags, const u64 flag_mask,
 						   void *lookup_elem(struct bpf_map *map, u32 key))
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	const u64 action_mask = XDP_ABORTED | XDP_DROP | XDP_PASS | XDP_TX;
 
 	/* Lower bits of the flags are used as return code on lookup failure */
diff --git a/include/linux/sched.h b/include/linux/sched.h
index a9b0ca72db55f..dfa1843ab2916 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -53,6 +53,7 @@ struct bio_list;
 struct blk_plug;
 struct bpf_local_storage;
 struct bpf_run_ctx;
+struct bpf_net_context;
 struct capture_control;
 struct cfs_rq;
 struct fs_struct;
@@ -1508,6 +1509,8 @@ struct task_struct {
 	/* Used for BPF run context */
 	struct bpf_run_ctx		*bpf_ctx;
 #endif
+	/* Used by BPF for per-TASK xdp storage */
+	struct bpf_net_context		*bpf_net_context;
 
 #ifdef CONFIG_GCC_PLUGIN_STACKLEAK
 	unsigned long			lowest_stack;
diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index a8e34416e960f..66974bd027109 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -240,12 +240,14 @@ static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
 				int xdp_n, struct xdp_cpumap_stats *stats,
 				struct list_head *list)
 {
+	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
 	int nframes;
 
 	if (!rcpu->prog)
 		return xdp_n;
 
 	rcu_read_lock_bh();
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 
 	nframes = cpu_map_bpf_prog_run_xdp(rcpu, frames, xdp_n, stats);
 
@@ -255,6 +257,7 @@ static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
 	if (unlikely(!list_empty(list)))
 		cpu_map_bpf_prog_run_skb(rcpu, list, stats);
 
+	bpf_net_ctx_clear(bpf_net_ctx);
 	rcu_read_unlock_bh(); /* resched point, may call do_softirq() */
 
 	return nframes;
diff --git a/kernel/fork.c b/kernel/fork.c
index 99076dbe27d83..f314bdd7e6108 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -2355,6 +2355,7 @@ __latent_entropy struct task_struct *copy_process(
 	RCU_INIT_POINTER(p->bpf_storage, NULL);
 	p->bpf_ctx = NULL;
 #endif
+	p->bpf_net_context =  NULL;
 
 	/* Perform scheduler related setup. Assign this task to a CPU. */
 	retval = sched_fork(clone_flags, p);
diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
index f6aad4ed2ab2f..600cc8e428c1a 100644
--- a/net/bpf/test_run.c
+++ b/net/bpf/test_run.c
@@ -283,9 +283,10 @@ static int xdp_recv_frames(struct xdp_frame **frames, int nframes,
 static int xdp_test_run_batch(struct xdp_test_data *xdp, struct bpf_prog *prog,
 			      u32 repeat)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
 	int err = 0, act, ret, i, nframes = 0, batch_sz;
 	struct xdp_frame **frames = xdp->frames;
+	struct bpf_redirect_info *ri;
 	struct xdp_page_head *head;
 	struct xdp_frame *frm;
 	bool redirect = false;
@@ -295,6 +296,8 @@ static int xdp_test_run_batch(struct xdp_test_data *xdp, struct bpf_prog *prog,
 	batch_sz = min_t(u32, repeat, xdp->batch_size);
 
 	local_bh_disable();
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
+	ri = bpf_net_ctx_get_ri();
 	xdp_set_return_frame_no_direct();
 
 	for (i = 0; i < batch_sz; i++) {
@@ -359,6 +362,7 @@ static int xdp_test_run_batch(struct xdp_test_data *xdp, struct bpf_prog *prog,
 	}
 
 	xdp_clear_return_frame_no_direct();
+	bpf_net_ctx_clear(bpf_net_ctx);
 	local_bh_enable();
 	return err;
 }
@@ -394,6 +398,7 @@ static int bpf_test_run_xdp_live(struct bpf_prog *prog, struct xdp_buff *ctx,
 static int bpf_test_run(struct bpf_prog *prog, void *ctx, u32 repeat,
 			u32 *retval, u32 *time, bool xdp)
 {
+	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
 	struct bpf_prog_array_item item = {.prog = prog};
 	struct bpf_run_ctx *old_ctx;
 	struct bpf_cg_run_ctx run_ctx;
@@ -419,10 +424,14 @@ static int bpf_test_run(struct bpf_prog *prog, void *ctx, u32 repeat,
 	do {
 		run_ctx.prog_item = &item;
 		local_bh_disable();
+		bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
+
 		if (xdp)
 			*retval = bpf_prog_run_xdp(prog, ctx);
 		else
 			*retval = bpf_prog_run(prog, ctx);
+
+		bpf_net_ctx_clear(bpf_net_ctx);
 		local_bh_enable();
 	} while (bpf_test_timer_continue(&t, 1, repeat, &ret, time));
 	bpf_reset_run_ctx(old_ctx);
diff --git a/net/core/dev.c b/net/core/dev.c
index 2c3f86c8cd176..3c0f5474eae78 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4030,11 +4030,15 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
 		   struct net_device *orig_dev, bool *another)
 {
 	struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress);
+	struct bpf_net_context *bpf_net_ctx __free(bpf_net_ctx_clear) = NULL;
 	enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS;
+	struct bpf_net_context __bpf_net_ctx;
 	int sch_ret;
 
 	if (!entry)
 		return skb;
+
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 	if (*pt_prev) {
 		*ret = deliver_skb(skb, *pt_prev, orig_dev);
 		*pt_prev = NULL;
@@ -4085,13 +4089,17 @@ sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
 static __always_inline struct sk_buff *
 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
 {
+	struct bpf_net_context *bpf_net_ctx __free(bpf_net_ctx_clear) = NULL;
 	struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress);
 	enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS;
+	struct bpf_net_context __bpf_net_ctx;
 	int sch_ret;
 
 	if (!entry)
 		return skb;
 
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
+
 	/* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was
 	 * already set by the caller.
 	 */
@@ -6356,13 +6364,15 @@ static void __napi_busy_loop(unsigned int napi_id,
 		      bool (*loop_end)(void *, unsigned long),
 		      void *loop_end_arg, unsigned flags, u16 budget)
 {
+	struct bpf_net_context *bpf_net_ctx __free(bpf_net_ctx_clear) = NULL;
 	unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
 	int (*napi_poll)(struct napi_struct *napi, int budget);
+	struct bpf_net_context __bpf_net_ctx;
 	void *have_poll_lock = NULL;
 	struct napi_struct *napi;
 
 	WARN_ON_ONCE(!rcu_read_lock_held());
-
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 restart:
 	napi_poll = NULL;
 
@@ -6833,6 +6843,7 @@ static int napi_thread_wait(struct napi_struct *napi)
 
 static void napi_threaded_poll_loop(struct napi_struct *napi)
 {
+	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
 	struct softnet_data *sd;
 	unsigned long last_qs = jiffies;
 
@@ -6841,6 +6852,8 @@ static void napi_threaded_poll_loop(struct napi_struct *napi)
 		void *have;
 
 		local_bh_disable();
+		bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
+
 		sd = this_cpu_ptr(&softnet_data);
 		sd->in_napi_threaded_poll = true;
 
@@ -6856,6 +6869,7 @@ static void napi_threaded_poll_loop(struct napi_struct *napi)
 			net_rps_action_and_irq_enable(sd);
 		}
 		skb_defer_free_flush(sd);
+		bpf_net_ctx_clear(bpf_net_ctx);
 		local_bh_enable();
 
 		if (!repoll)
@@ -6878,13 +6892,16 @@ static int napi_threaded_poll(void *data)
 
 static __latent_entropy void net_rx_action(struct softirq_action *h)
 {
+	struct bpf_net_context *bpf_net_ctx __free(bpf_net_ctx_clear) = NULL;
 	struct softnet_data *sd = this_cpu_ptr(&softnet_data);
 	unsigned long time_limit = jiffies +
 		usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs));
 	int budget = READ_ONCE(net_hotdata.netdev_budget);
+	struct bpf_net_context __bpf_net_ctx;
 	LIST_HEAD(list);
 	LIST_HEAD(repoll);
 
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 start:
 	sd->in_net_rx_action = true;
 	local_irq_disable();
diff --git a/net/core/filter.c b/net/core/filter.c
index d6cf1a63c3f43..0a3cc71505b43 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c
@@ -2475,9 +2475,6 @@ static const struct bpf_func_proto bpf_clone_redirect_proto = {
 	.arg3_type      = ARG_ANYTHING,
 };
 
-DEFINE_PER_CPU(struct bpf_redirect_info, bpf_redirect_info);
-EXPORT_PER_CPU_SYMBOL_GPL(bpf_redirect_info);
-
 static struct net_device *skb_get_peer_dev(struct net_device *dev)
 {
 	const struct net_device_ops *ops = dev->netdev_ops;
@@ -2490,7 +2487,7 @@ static struct net_device *skb_get_peer_dev(struct net_device *dev)
 
 int skb_do_redirect(struct sk_buff *skb)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	struct net *net = dev_net(skb->dev);
 	struct net_device *dev;
 	u32 flags = ri->flags;
@@ -2523,7 +2520,7 @@ int skb_do_redirect(struct sk_buff *skb)
 
 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	if (unlikely(flags & (~(BPF_F_INGRESS) | BPF_F_REDIRECT_INTERNAL)))
 		return TC_ACT_SHOT;
@@ -2544,7 +2541,7 @@ static const struct bpf_func_proto bpf_redirect_proto = {
 
 BPF_CALL_2(bpf_redirect_peer, u32, ifindex, u64, flags)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	if (unlikely(flags))
 		return TC_ACT_SHOT;
@@ -2566,7 +2563,7 @@ static const struct bpf_func_proto bpf_redirect_peer_proto = {
 BPF_CALL_4(bpf_redirect_neigh, u32, ifindex, struct bpf_redir_neigh *, params,
 	   int, plen, u64, flags)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	if (unlikely((plen && plen < sizeof(*params)) || flags))
 		return TC_ACT_SHOT;
@@ -4294,19 +4291,17 @@ void xdp_do_check_flushed(struct napi_struct *napi)
 
 void bpf_clear_redirect_map(struct bpf_map *map)
 {
-	struct bpf_redirect_info *ri;
-	int cpu;
-
-	for_each_possible_cpu(cpu) {
-		ri = per_cpu_ptr(&bpf_redirect_info, cpu);
-		/* Avoid polluting remote cacheline due to writes if
-		 * not needed. Once we pass this test, we need the
-		 * cmpxchg() to make sure it hasn't been changed in
-		 * the meantime by remote CPU.
-		 */
-		if (unlikely(READ_ONCE(ri->map) == map))
-			cmpxchg(&ri->map, map, NULL);
-	}
+	/* ri->map is assigned in __bpf_xdp_redirect_map() from within a eBPF
+	 * program/ during NAPI callback. It is used during
+	 * xdp_do_generic_redirect_map()/ __xdp_do_redirect_frame() from the
+	 * redirect callback afterwards. ri->map is cleared after usage.
+	 * The path has no explicit RCU read section but the local_bh_disable()
+	 * is also a RCU read section which makes the complete softirq callback
+	 * RCU protected. This in turn makes ri->map RCU protected and it is
+	 * sufficient to wait a grace period to ensure that no "ri->map == map"
+	 * exists. dev_map_free() removes the map from the list and then
+	 * invokes synchronize_rcu() after calling this function.
+	 */
 }
 
 DEFINE_STATIC_KEY_FALSE(bpf_master_redirect_enabled_key);
@@ -4314,8 +4309,8 @@ EXPORT_SYMBOL_GPL(bpf_master_redirect_enabled_key);
 
 u32 xdp_master_redirect(struct xdp_buff *xdp)
 {
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	struct net_device *master, *slave;
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
 
 	master = netdev_master_upper_dev_get_rcu(xdp->rxq->dev);
 	slave = master->netdev_ops->ndo_xdp_get_xmit_slave(master, xdp);
@@ -4432,7 +4427,7 @@ static __always_inline int __xdp_do_redirect_frame(struct bpf_redirect_info *ri,
 int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
 		    struct bpf_prog *xdp_prog)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	enum bpf_map_type map_type = ri->map_type;
 
 	if (map_type == BPF_MAP_TYPE_XSKMAP)
@@ -4446,7 +4441,7 @@ EXPORT_SYMBOL_GPL(xdp_do_redirect);
 int xdp_do_redirect_frame(struct net_device *dev, struct xdp_buff *xdp,
 			  struct xdp_frame *xdpf, struct bpf_prog *xdp_prog)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	enum bpf_map_type map_type = ri->map_type;
 
 	if (map_type == BPF_MAP_TYPE_XSKMAP)
@@ -4463,7 +4458,7 @@ static int xdp_do_generic_redirect_map(struct net_device *dev,
 				       enum bpf_map_type map_type, u32 map_id,
 				       u32 flags)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	struct bpf_map *map;
 	int err;
 
@@ -4517,7 +4512,7 @@ static int xdp_do_generic_redirect_map(struct net_device *dev,
 int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
 			    struct xdp_buff *xdp, struct bpf_prog *xdp_prog)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 	enum bpf_map_type map_type = ri->map_type;
 	void *fwd = ri->tgt_value;
 	u32 map_id = ri->map_id;
@@ -4553,7 +4548,7 @@ int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
 
 BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
 {
-	struct bpf_redirect_info *ri = this_cpu_ptr(&bpf_redirect_info);
+	struct bpf_redirect_info *ri = bpf_net_ctx_get_ri();
 
 	if (unlikely(flags))
 		return XDP_ABORTED;
diff --git a/net/core/lwt_bpf.c b/net/core/lwt_bpf.c
index a94943681e5aa..afb05f58b64c5 100644
--- a/net/core/lwt_bpf.c
+++ b/net/core/lwt_bpf.c
@@ -38,12 +38,14 @@ static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
 static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
 		       struct dst_entry *dst, bool can_redirect)
 {
+	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
 	int ret;
 
 	/* Disabling BH is needed to protect per-CPU bpf_redirect_info between
 	 * BPF prog and skb_do_redirect().
 	 */
 	local_bh_disable();
+	bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx);
 	bpf_compute_data_pointers(skb);
 	ret = bpf_prog_run_save_cb(lwt->prog, skb);
 
@@ -76,6 +78,7 @@ static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
 		break;
 	}
 
+	bpf_net_ctx_clear(bpf_net_ctx);
 	local_bh_enable();
 
 	return ret;
-- 
2.45.1

Sebastian Andrzej Siewior <bigeasy@linutronix.de> writes:

> The XDP redirect process is two staged:
> - bpf_prog_run_xdp() is invoked to run a eBPF program which inspects the
>   packet and makes decisions. While doing that, the per-CPU variable
>   bpf_redirect_info is used.
>
> - Afterwards xdp_do_redirect() is invoked and accesses bpf_redirect_info
>   and it may also access other per-CPU variables like xskmap_flush_list.
>
> At the very end of the NAPI callback, xdp_do_flush() is invoked which
> does not access bpf_redirect_info but will touch the individual per-CPU
> lists.
>
> The per-CPU variables are only used in the NAPI callback hence disabling
> bottom halves is the only protection mechanism. Users from preemptible
> context (like cpu_map_kthread_run()) explicitly disable bottom halves
> for protections reasons.
> Without locking in local_bh_disable() on PREEMPT_RT this data structure
> requires explicit locking.
>
> PREEMPT_RT has forced-threaded interrupts enabled and every
> NAPI-callback runs in a thread. If each thread has its own data
> structure then locking can be avoided.
>
> Create a struct bpf_net_context which contains struct bpf_redirect_info.
> Define the variable on stack, use bpf_net_ctx_set() to save a pointer to
> it. Use the __free() annotation to automatically reset the pointer once
> function returns.
> The bpf_net_ctx_set() may nest. For instance a function can be used from
> within NET_RX_SOFTIRQ/ net_rx_action which uses bpf_net_ctx_set() and
> NET_TX_SOFTIRQ which does not. Therefore only the first invocations
> updates the pointer.
> Use bpf_net_ctx_get_ri() as a wrapper to retrieve the current struct
> bpf_redirect_info.
>
> On PREEMPT_RT the pointer to bpf_net_context is saved task's
> task_struct. On non-PREEMPT_RT builds the pointer saved in a per-CPU
> variable (which is always NODE-local memory). Using always the
> bpf_net_context approach has the advantage that there is almost zero
> differences between PREEMPT_RT and non-PREEMPT_RT builds.
>
> Cc: Alexei Starovoitov <ast@kernel.org>
> Cc: Andrii Nakryiko <andrii@kernel.org>
> Cc: Eduard Zingerman <eddyz87@gmail.com>
> Cc: Hao Luo <haoluo@google.com>
> Cc: Jesper Dangaard Brouer <hawk@kernel.org>
> Cc: Jiri Olsa <jolsa@kernel.org>
> Cc: John Fastabend <john.fastabend@gmail.com>
> Cc: KP Singh <kpsingh@kernel.org>
> Cc: Martin KaFai Lau <martin.lau@linux.dev>
> Cc: Song Liu <song@kernel.org>
> Cc: Stanislav Fomichev <sdf@google.com>
> Cc: Toke Høiland-Jørgensen <toke@redhat.com>
> Cc: Yonghong Song <yonghong.song@linux.dev>
> Cc: bpf@vger.kernel.org
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

[...]
> @@ -240,12 +240,14 @@ static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
>  				int xdp_n, struct xdp_cpumap_stats *stats,
>  				struct list_head *list)
>  {
> +	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
>  	int nframes;

I think we need to zero-initialise all the context objects we allocate
on the stack.

The reason being that an XDP program can return XDP_REDIRECT without
calling any of the redirect helpers first; which will lead to
xdp_do_redirect() being called without any of the fields in struct
bpf_redirect_info having being set. This can lead to a crash if the
values happen to be the wrong value; and if we're not initialising the
stack space used by this struct, we have no guarantees about what value
they will end up with.

We fixed a similar bug relatively recently, see:

5bcf0dcbf906 ("xdp: use flags field to disambiguate broadcast redirect")

>  void bpf_clear_redirect_map(struct bpf_map *map)
>  {
> -	struct bpf_redirect_info *ri;
> -	int cpu;
> -
> -	for_each_possible_cpu(cpu) {
> -		ri = per_cpu_ptr(&bpf_redirect_info, cpu);
> -		/* Avoid polluting remote cacheline due to writes if
> -		 * not needed. Once we pass this test, we need the
> -		 * cmpxchg() to make sure it hasn't been changed in
> -		 * the meantime by remote CPU.
> -		 */
> -		if (unlikely(READ_ONCE(ri->map) == map))
> -			cmpxchg(&ri->map, map, NULL);
> -	}
> +	/* ri->map is assigned in __bpf_xdp_redirect_map() from within a eBPF
> +	 * program/ during NAPI callback. It is used during
> +	 * xdp_do_generic_redirect_map()/ __xdp_do_redirect_frame() from the
> +	 * redirect callback afterwards. ri->map is cleared after usage.
> +	 * The path has no explicit RCU read section but the local_bh_disable()
> +	 * is also a RCU read section which makes the complete softirq callback
> +	 * RCU protected. This in turn makes ri->map RCU protected and it is
> +	 * sufficient to wait a grace period to ensure that no "ri->map == map"
> +	 * exists. dev_map_free() removes the map from the list and then
> +	 * invokes synchronize_rcu() after calling this function.
> +	 */
>  }

With the zeroing of the stack variable mentioned above, I agree that
this is not needed anymore, but I think we should just get rid of the
function entirely and put a comment in devmap.c instead of the call to
the (now empty) function.

-Toke

On 2024-05-30 00:09:21 [+0200], Toke Høiland-Jørgensen wrote:
> [...]
> > @@ -240,12 +240,14 @@ static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
> >  				int xdp_n, struct xdp_cpumap_stats *stats,
> >  				struct list_head *list)
> >  {
> > +	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
> >  	int nframes;
> 
> I think we need to zero-initialise all the context objects we allocate
> on the stack.
> 
> The reason being that an XDP program can return XDP_REDIRECT without
> calling any of the redirect helpers first; which will lead to
> xdp_do_redirect() being called without any of the fields in struct
> bpf_redirect_info having being set. This can lead to a crash if the
> values happen to be the wrong value; and if we're not initialising the
> stack space used by this struct, we have no guarantees about what value
> they will end up with.

Okay, I can do that.

> >  void bpf_clear_redirect_map(struct bpf_map *map)
> >  {
> > -	struct bpf_redirect_info *ri;
> > -	int cpu;
> > -
> > -	for_each_possible_cpu(cpu) {
> > -		ri = per_cpu_ptr(&bpf_redirect_info, cpu);
> > -		/* Avoid polluting remote cacheline due to writes if
> > -		 * not needed. Once we pass this test, we need the
> > -		 * cmpxchg() to make sure it hasn't been changed in
> > -		 * the meantime by remote CPU.
> > -		 */
> > -		if (unlikely(READ_ONCE(ri->map) == map))
> > -			cmpxchg(&ri->map, map, NULL);
> > -	}
> > +	/* ri->map is assigned in __bpf_xdp_redirect_map() from within a eBPF
> > +	 * program/ during NAPI callback. It is used during
> > +	 * xdp_do_generic_redirect_map()/ __xdp_do_redirect_frame() from the
> > +	 * redirect callback afterwards. ri->map is cleared after usage.
> > +	 * The path has no explicit RCU read section but the local_bh_disable()
> > +	 * is also a RCU read section which makes the complete softirq callback
> > +	 * RCU protected. This in turn makes ri->map RCU protected and it is
> > +	 * sufficient to wait a grace period to ensure that no "ri->map == map"
> > +	 * exists. dev_map_free() removes the map from the list and then
> > +	 * invokes synchronize_rcu() after calling this function.
> > +	 */
> >  }
> 
> With the zeroing of the stack variable mentioned above, I agree that
> this is not needed anymore, but I think we should just get rid of the
> function entirely and put a comment in devmap.c instead of the call to
> the (now empty) function.

I wasn't entirely sure if my reasoning is valid. In that case…

> -Toke

Sebastian

On 31/05/2024 12.38, Sebastian Andrzej Siewior wrote:
> On 2024-05-30 00:09:21 [+0200], Toke Høiland-Jørgensen wrote:
>> [...]
>>> @@ -240,12 +240,14 @@ static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
>>>   				int xdp_n, struct xdp_cpumap_stats *stats,
>>>   				struct list_head *list)
>>>   {
>>> +	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
>>>   	int nframes;
 >>
>> I think we need to zero-initialise all the context objects we allocate
>> on the stack.
>>
> Okay, I can do that.

Hmm, but how will this affect performance?

--Jesper

On 2024-06-05 12:28:08 [+0200], Jesper Dangaard Brouer wrote:
> 
> Hmm, but how will this affect performance?

As I wrote in the changelog for v4, I haven't notice a difference. I
tried to move bpf_net_ctx_set() from cpu_map_bpf_prog_run() to
cpu_map_kthread_run() to have this assignment only once and I didn't see
a difference/ I couldn't tell the two kernels apart.

This is what I have been using for testing

| xdp-bench redirect-cpu --cpu 3 --remote-action drop eth1 -e

in case I was changing the wrong part…

> --Jesper

Sebastian

On 05/06/2024 12.41, Sebastian Andrzej Siewior wrote:
> On 2024-06-05 12:28:08 [+0200], Jesper Dangaard Brouer wrote:
>>
>> Hmm, but how will this affect performance?
> 
> As I wrote in the changelog for v4, I haven't notice a difference. I
> tried to move bpf_net_ctx_set() from cpu_map_bpf_prog_run() to
> cpu_map_kthread_run() to have this assignment only once and I didn't see
> a difference/ I couldn't tell the two kernels apart.
> 

This would be my preferred solution.
See below, your benchmark wasn't testing/measuring this changed code on
remote CPU running kthread.

> This is what I have been using for testing
> 
> | xdp-bench redirect-cpu --cpu 3 --remote-action drop eth1 -e
> 
> in case I was changing the wrong part…

As we saw earlier (with your hardware setup) this test is benchmarking
the RX-NAPI XDP-redirect code.  As the cpumap "remote" CPUs kthread had
idle cycles.

The extra clearing bpf_net_ctx_set() for each packet in the kthread on
the remote CPU will not change the benchmark numbers (as it have idle
cycles).

Looking closer at kernel code + your patch, I see that this clearing
isn't done for each packet, but per bulk (up-to CPUMAP_BATCH 8).  Given
that, I'm more okay with this change.

--Jesper

Jesper Dangaard Brouer <hawk@kernel.org> writes:

> On 31/05/2024 12.38, Sebastian Andrzej Siewior wrote:
>> On 2024-05-30 00:09:21 [+0200], Toke Høiland-Jørgensen wrote:
>>> [...]
>>>> @@ -240,12 +240,14 @@ static int cpu_map_bpf_prog_run(struct bpf_cpu_map_entry *rcpu, void **frames,
>>>>   				int xdp_n, struct xdp_cpumap_stats *stats,
>>>>   				struct list_head *list)
>>>>   {
>>>> +	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
>>>>   	int nframes;
>  >>
>>> I think we need to zero-initialise all the context objects we allocate
>>> on the stack.
>>>
>> Okay, I can do that.
>
> Hmm, but how will this affect performance?

My hunch would be that this would be in a cache line we're touching
anyway, so it won't make much difference? But better measure, I suppose :)

-Toke

On Wed, May 29, 2024 at 9:29 AM Sebastian Andrzej Siewior
<bigeasy@linutronix.de> wrote:
>
> The XDP redirect process is two staged:
> - bpf_prog_run_xdp() is invoked to run a eBPF program which inspects the
>   packet and makes decisions. While doing that, the per-CPU variable
>   bpf_redirect_info is used.
>
> - Afterwards xdp_do_redirect() is invoked and accesses bpf_redirect_info
>   and it may also access other per-CPU variables like xskmap_flush_list.
>
> At the very end of the NAPI callback, xdp_do_flush() is invoked which
> does not access bpf_redirect_info but will touch the individual per-CPU
> lists.
>
> The per-CPU variables are only used in the NAPI callback hence disabling
> bottom halves is the only protection mechanism. Users from preemptible
> context (like cpu_map_kthread_run()) explicitly disable bottom halves
> for protections reasons.
> Without locking in local_bh_disable() on PREEMPT_RT this data structure
> requires explicit locking.
>
> PREEMPT_RT has forced-threaded interrupts enabled and every
> NAPI-callback runs in a thread. If each thread has its own data
> structure then locking can be avoided.
>
> Create a struct bpf_net_context which contains struct bpf_redirect_info.
> Define the variable on stack, use bpf_net_ctx_set() to save a pointer to
> it. Use the __free() annotation to automatically reset the pointer once
> function returns.
> The bpf_net_ctx_set() may nest. For instance a function can be used from
> within NET_RX_SOFTIRQ/ net_rx_action which uses bpf_net_ctx_set() and
> NET_TX_SOFTIRQ which does not. Therefore only the first invocations
> updates the pointer.
> Use bpf_net_ctx_get_ri() as a wrapper to retrieve the current struct
> bpf_redirect_info.
>
> On PREEMPT_RT the pointer to bpf_net_context is saved task's
> task_struct. On non-PREEMPT_RT builds the pointer saved in a per-CPU
> variable (which is always NODE-local memory). Using always the
> bpf_net_context approach has the advantage that there is almost zero
> differences between PREEMPT_RT and non-PREEMPT_RT builds.
>
> Cc: Alexei Starovoitov <ast@kernel.org>
> Cc: Andrii Nakryiko <andrii@kernel.org>
> Cc: Eduard Zingerman <eddyz87@gmail.com>
> Cc: Hao Luo <haoluo@google.com>
> Cc: Jesper Dangaard Brouer <hawk@kernel.org>
> Cc: Jiri Olsa <jolsa@kernel.org>
> Cc: John Fastabend <john.fastabend@gmail.com>
> Cc: KP Singh <kpsingh@kernel.org>
> Cc: Martin KaFai Lau <martin.lau@linux.dev>
> Cc: Song Liu <song@kernel.org>
> Cc: Stanislav Fomichev <sdf@google.com>
> Cc: Toke Høiland-Jørgensen <toke@redhat.com>
> Cc: Yonghong Song <yonghong.song@linux.dev>
> Cc: bpf@vger.kernel.org
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

lgtm
Acked-by: Alexei Starovoitov <ast@kernel.org>