Skip to contentSkip to conversation

Various BPF helper improvements

PATCH Superseded Daniel Borkmann 7 participants 28 messages Sep 24, 2020 – Sep 25, 2020

Subsystems (8)

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Daniel Borkmann

author maintainer
#0 lore raw

cover Various BPF helper improvements

This series adds two BPF helpers, that is, one for retrieving the classid of an skb and another one to redirect via the neigh subsystem, and improves also the cookie helpers by removing the atomic counter. I've also added the bpf_tail_call_static() helper to the libbpf API that we've been using in Cilium for a while now, and last but not least the series adds a few selftests. For details, please check individual patches, thanks!

Daniel Borkmann (6):
bpf: add classid helper only based on skb->sk
bpf, net: rework cookie generator as per-cpu one
bpf: add redirect_neigh helper as redirect drop-in bpf, libbpf: add bpf_tail_call_static helper for bpf programs bpf, selftests: use bpf_tail_call_static where appropriate bpf, selftests: add redirect_neigh selftest

 include/linux/cookie.h                        |  41 +++
 include/linux/skbuff.h                        |   5 +
 include/uapi/linux/bpf.h                      |  24 ++
 net/core/filter.c                             | 277 +++++++++++++++++-
 net/core/net_namespace.c                      |   5 +-
 net/core/sock_diag.c                          |   7 +-
 tools/include/uapi/linux/bpf.h                |  24 ++
 tools/lib/bpf/bpf_helpers.h                   |  32 ++
 tools/testing/selftests/bpf/progs/bpf_flow.c  |  12 +-
 tools/testing/selftests/bpf/progs/tailcall1.c |  28 +-
 tools/testing/selftests/bpf/progs/tailcall2.c |  14 +-
 tools/testing/selftests/bpf/progs/tailcall3.c |   4 +-
 .../selftests/bpf/progs/tailcall_bpf2bpf1.c   |   4 +-
 .../selftests/bpf/progs/tailcall_bpf2bpf2.c   |   6 +-
 .../selftests/bpf/progs/tailcall_bpf2bpf3.c   |   6 +-
 .../selftests/bpf/progs/tailcall_bpf2bpf4.c   |   6 +-
 .../selftests/bpf/progs/test_tc_neigh.c       | 144 +++++++++
 tools/testing/selftests/bpf/test_tc_neigh.sh  | 168 +++++++++++
 18 files changed, 749 insertions(+), 58 deletions(-)
 create mode 100644 include/linux/cookie.h
 create mode 100644 tools/testing/selftests/bpf/progs/test_tc_neigh.c
 create mode 100755 tools/testing/selftests/bpf/test_tc_neigh.sh
Signature · 1 line
2.21.0
author maintainer
#1 lore raw

PATCH 1/6 bpf: add classid helper only based on skb->sk

Similarly to 5a52ae4e32a6 ("bpf: Allow to retrieve cgroup v1 classid from v2 hooks"), add a helper to retrieve cgroup v1 classid solely based on the skb->sk, so it can be used as key as part of BPF map lookups out of tc from host ns, in particular given the skb->sk is retained these days when crossing net ns thanks to 9c4c325252c5 ("skbuff: preserve sock reference when scrubbing the skb."). This is similar to bpf_skb_cgroup_id() which implements the same for v2. Kubernetes ecosystem is still operating on v1 however, hence net_cls needs to be used there until this can be dropped in with the v2 helper of bpf_skb_cgroup_id().

---
 include/uapi/linux/bpf.h       | 10 ++++++++++
 net/core/filter.c              | 21 +++++++++++++++++++++
 tools/include/uapi/linux/bpf.h | 10 ++++++++++
 3 files changed, 41 insertions(+)
include/uapi/linux/bpf.h +10 -0
index a22812561064..48ecf246d047 100644
@@ -3586,6 +3586,15 @@ union bpf_attr {
  * 		the data in *dst*. This is a wrapper of **copy_from_user**\ ().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u64 bpf_skb_cgroup_classid(struct sk_buff *skb)
+ * 	Description
+ * 		See **bpf_get_cgroup_classid**\ () for the main description.
+ * 		This helper differs from **bpf_get_cgroup_classid**\ () in that
+ * 		the cgroup v1 net_cls class is retrieved only from the *skb*'s
+ * 		associated socket instead of the current process.
+ * 	Return
+ * 		The id is returned or 0 in case the id could not be retrieved.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3737,6 +3746,7 @@ union bpf_attr {
 	FN(inode_storage_delete),	\
 	FN(d_path),			\
 	FN(copy_from_user),		\
+	FN(skb_cgroup_classid),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
net/core/filter.c +21 -0
index 6014e5f40c58..0f913755bcba 100644
@@ -2704,6 +2704,23 @@ static const struct bpf_func_proto bpf_get_cgroup_classid_curr_proto = {
 	.gpl_only	= false,
 	.ret_type	= RET_INTEGER,
 };
+
+BPF_CALL_1(bpf_skb_cgroup_classid, const struct sk_buff *, skb)
+{
+	struct sock *sk = skb_to_full_sk(skb);
+
+	if (!sk || !sk_fullsock(sk))
+		return 0;
+
+	return sock_cgroup_classid(&sk->sk_cgrp_data);
+}
+
+static const struct bpf_func_proto bpf_skb_cgroup_classid_proto = {
+	.func		= bpf_skb_cgroup_classid,
+	.gpl_only	= false,
+	.ret_type	= RET_INTEGER,
+	.arg1_type	= ARG_PTR_TO_CTX,
+};
 #endif
 
 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
@@ -6770,6 +6787,10 @@ tc_cls_act_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 	case BPF_FUNC_skb_get_xfrm_state:
 		return &bpf_skb_get_xfrm_state_proto;
 #endif
+#ifdef CONFIG_CGROUP_NET_CLASSID
+	case BPF_FUNC_skb_cgroup_classid:
+		return &bpf_skb_cgroup_classid_proto;
+#endif
 #ifdef CONFIG_SOCK_CGROUP_DATA
 	case BPF_FUNC_skb_cgroup_id:
 		return &bpf_skb_cgroup_id_proto;
tools/include/uapi/linux/bpf.h +10 -0
index a22812561064..48ecf246d047 100644
@@ -3586,6 +3586,15 @@ union bpf_attr {
  * 		the data in *dst*. This is a wrapper of **copy_from_user**\ ().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u64 bpf_skb_cgroup_classid(struct sk_buff *skb)
+ * 	Description
+ * 		See **bpf_get_cgroup_classid**\ () for the main description.
+ * 		This helper differs from **bpf_get_cgroup_classid**\ () in that
+ * 		the cgroup v1 net_cls class is retrieved only from the *skb*'s
+ * 		associated socket instead of the current process.
+ * 	Return
+ * 		The id is returned or 0 in case the id could not be retrieved.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3737,6 +3746,7 @@ union bpf_attr {
 	FN(inode_storage_delete),	\
 	FN(d_path),			\
 	FN(copy_from_user),		\
+	FN(skb_cgroup_classid),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
Signature · 1 line
2.21.0
maintainer
#2 lore raw

Re: [PATCH bpf-next 1/6] bpf: add classid helper only based on skb->sk

Quoted 11 lines · On Thu, 24 Sep 2020 20:21:22 +0200 Daniel Borkmann wrote:

Similarly to 5a52ae4e32a6 ("bpf: Allow to retrieve cgroup v1 classid from v2 hooks"), add a helper to retrieve cgroup v1 classid solely based on the skb->sk, so it can be used as key as part of BPF map lookups out of tc from host ns, in particular given the skb->sk is retained these days when crossing net ns thanks to 9c4c325252c5 ("skbuff: preserve sock reference when scrubbing the skb."). This is similar to bpf_skb_cgroup_id() which implements the same for v2. Kubernetes ecosystem is still operating on v1 however, hence net_cls needs to be used there until this can be dropped in with the v2 helper of bpf_skb_cgroup_id().

Quoted 1 line

FWIW lot's of whitespace warnings from checkpatch --strict about comments having spaces before tabs here.

author maintainer
#3 lore raw

Re: [PATCH bpf-next 1/6] bpf: add classid helper only based on skb->sk

Quoted 12 lines · Jakub Kicinski
Quoted 11 lines · On Thu, 24 Sep 2020 20:21:22 +0200 Daniel Borkmann wrote:

Similarly to 5a52ae4e32a6 ("bpf: Allow to retrieve cgroup v1 classid from v2 hooks"), add a helper to retrieve cgroup v1 classid solely based on the skb->sk, so it can be used as key as part of BPF map lookups out of tc from host ns, in particular given the skb->sk is retained these days when crossing net ns thanks to 9c4c325252c5 ("skbuff: preserve sock reference when scrubbing the skb."). This is similar to bpf_skb_cgroup_id() which implements the same for v2. Kubernetes ecosystem is still operating on v1 however, hence net_cls needs to be used there until this can be dropped in with the v2 helper of bpf_skb_cgroup_id().

Quoted 4 lines
Quoted 1 line

FWIW lot's of whitespace warnings from checkpatch --strict about comments having spaces before tabs here.

Expected given the way the UAPI helper comment is formatted / done in order to then render the man page. So it's formatted the same way as the other helper descriptions.

Thanks,
Daniel

author maintainer
#4 lore raw

PATCH 2/6 bpf, net: rework cookie generator as per-cpu one

In reply to Daniel Borkmann

With its use in BPF the cookie generator can be called very frequently in particular when used out of cgroup v2 hooks (e.g. connect / sendmsg) and attached to the root cgroup, for example, when used in v1/v2 mixed environments. In particular when there's a high churn on sockets in the system there can be many parallel requests to the bpf_get_socket_cookie() and bpf_get_netns_cookie() helpers which then cause contention on the shared atomic counter. As similarly done in f991bd2e1421 ("fs: introduce a per-cpu last_ino allocator"), add a small helper library that both can then use for the 64 bit counters.

---
 include/linux/cookie.h   | 41 ++++++++++++++++++++++++++++++++++++++++
 net/core/net_namespace.c |  5 +++--
 net/core/sock_diag.c     |  7 ++++---
 3 files changed, 48 insertions(+), 5 deletions(-)
 create mode 100644 include/linux/cookie.h
include/linux/cookie.h new file +41 -0
new file mode 100644
index 000000000000..2488203dc004
@@ -0,0 +1,41 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __LINUX_COOKIE_H
+#define __LINUX_COOKIE_H
+
+#include <linux/atomic.h>
+#include <linux/percpu.h>
+
+struct gen_cookie {
+	u64 __percpu	*local_last;
+	atomic64_t	 shared_last ____cacheline_aligned_in_smp;
+};
+
+#define COOKIE_LOCAL_BATCH	4096
+
+#define DEFINE_COOKIE(name)					\
+	static DEFINE_PER_CPU(u64, __##name);			\
+	static struct gen_cookie name = {			\
+		.local_last	= &__##name,			\
+		.shared_last	= ATOMIC64_INIT(0),		\
+	}
+
+static inline u64 gen_cookie_next(struct gen_cookie *gc)
+{
+	u64 *local_last = &get_cpu_var(*gc->local_last);
+	u64 val = *local_last;
+
+	if (__is_defined(CONFIG_SMP) &&
+	    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {
+		s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
+					       &gc->shared_last);
+		val = next - COOKIE_LOCAL_BATCH;
+	}
+	val++;
+	if (unlikely(!val))
+		val++;
+	*local_last = val;
+	put_cpu_var(local_last);
+	return val;
+}
+
+#endif /* __LINUX_COOKIE_H */
net/core/net_namespace.c +3 -2
index dcd61aca343e..cf29ed8950d1 100644
@@ -19,6 +19,7 @@
 #include <linux/net_namespace.h>
 #include <linux/sched/task.h>
 #include <linux/uidgid.h>
+#include <linux/cookie.h>
 
 #include <net/sock.h>
 #include <net/netlink.h>
@@ -69,7 +70,7 @@ EXPORT_SYMBOL_GPL(pernet_ops_rwsem);
 
 static unsigned int max_gen_ptrs = INITIAL_NET_GEN_PTRS;
 
-static atomic64_t cookie_gen;
+DEFINE_COOKIE(net_cookie);
 
 u64 net_gen_cookie(struct net *net)
 {
@@ -78,7 +79,7 @@ u64 net_gen_cookie(struct net *net)
 
 		if (res)
 			return res;
-		res = atomic64_inc_return(&cookie_gen);
+		res = gen_cookie_next(&net_cookie);
 		atomic64_cmpxchg(&net->net_cookie, 0, res);
 	}
 }
net/core/sock_diag.c +4 -3
index c13ffbd33d8d..4a4180e8dd35 100644
@@ -11,7 +11,7 @@
 #include <linux/tcp.h>
 #include <linux/workqueue.h>
 #include <linux/nospec.h>
-
+#include <linux/cookie.h>
 #include <linux/inet_diag.h>
 #include <linux/sock_diag.h>
 
@@ -19,7 +19,8 @@ static const struct sock_diag_handler *sock_diag_handlers[AF_MAX];
 static int (*inet_rcv_compat)(struct sk_buff *skb, struct nlmsghdr *nlh);
 static DEFINE_MUTEX(sock_diag_table_mutex);
 static struct workqueue_struct *broadcast_wq;
-static atomic64_t cookie_gen;
+
+DEFINE_COOKIE(sock_cookie);
 
 u64 sock_gen_cookie(struct sock *sk)
 {
@@ -28,7 +29,7 @@ u64 sock_gen_cookie(struct sock *sk)
 
 		if (res)
 			return res;
-		res = atomic64_inc_return(&cookie_gen);
+		res = gen_cookie_next(&sock_cookie);
 		atomic64_cmpxchg(&sk->sk_cookie, 0, res);
 	}
 }
Signature · 1 line
2.21.0
#5 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

Quoted 52 lines · Daniel Borkmann

With its use in BPF the cookie generator can be called very frequently in particular when used out of cgroup v2 hooks (e.g. connect / sendmsg) and attached to the root cgroup, for example, when used in v1/v2 mixed environments. In particular when there's a high churn on sockets in the system there can be many parallel requests to the bpf_get_socket_cookie() and bpf_get_netns_cookie() helpers which then cause contention on the shared atomic counter. As similarly done in f991bd2e1421 ("fs: introduce a per-cpu last_ino allocator"), add a small helper library that both can then use for the 64 bit counters.

---
 include/linux/cookie.h   | 41 ++++++++++++++++++++++++++++++++++++++++
 net/core/net_namespace.c |  5 +++--
 net/core/sock_diag.c     |  7 ++++---
 3 files changed, 48 insertions(+), 5 deletions(-)
 create mode 100644 include/linux/cookie.h
include/linux/cookie.h new file +28 -0
new file mode 100644
index 000000000000..2488203dc004
@@ -0,0 +1,41 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __LINUX_COOKIE_H
+#define __LINUX_COOKIE_H
+
+#include <linux/atomic.h>
+#include <linux/percpu.h>
+
+struct gen_cookie {
+	u64 __percpu	*local_last;
+	atomic64_t	 shared_last ____cacheline_aligned_in_smp;
+};
+
+#define COOKIE_LOCAL_BATCH	4096
+
+#define DEFINE_COOKIE(name)					\
+	static DEFINE_PER_CPU(u64, __##name);			\
+	static struct gen_cookie name = {			\
+		.local_last	= &__##name,			\
+		.shared_last	= ATOMIC64_INIT(0),		\
+	}
+
+static inline u64 gen_cookie_next(struct gen_cookie *gc)
+{
+	u64 *local_last = &get_cpu_var(*gc->local_last);
+	u64 val = *local_last;
+
+	if (__is_defined(CONFIG_SMP) &&
+	    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {

include/linux/cookie.h line 38

Quoted 10 lines
include/linux/cookie.h new file +10 -0
+		s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
+					       &gc->shared_last);
+		val = next - COOKIE_LOCAL_BATCH;
+	}
+	val++;
+	if (unlikely(!val))
+		val++;
+	*local_last = val;
+	put_cpu_var(local_last);
+	return val;

This is not interrupt safe.

I think sock_gen_cookie() can be called from interrupt context.

get_next_ino() is only called from process context, that is what I used get_cpu_var() and put_cpu_var()

author maintainer
#6 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

Quoted 1 line · Eric Dumazet

On 9/24/20 8:21 PM, Daniel Borkmann wrote:

[...]

Quoted 47 lines
Quoted 34 lines
include/linux/cookie.h new file +28 -0
new file mode 100644
index 000000000000..2488203dc004
@@ -0,0 +1,41 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __LINUX_COOKIE_H
+#define __LINUX_COOKIE_H
+
+#include <linux/atomic.h>
+#include <linux/percpu.h>
+
+struct gen_cookie {
+	u64 __percpu	*local_last;
+	atomic64_t	 shared_last ____cacheline_aligned_in_smp;
+};
+
+#define COOKIE_LOCAL_BATCH	4096
+
+#define DEFINE_COOKIE(name)					\
+	static DEFINE_PER_CPU(u64, __##name);			\
+	static struct gen_cookie name = {			\
+		.local_last	= &__##name,			\
+		.shared_last	= ATOMIC64_INIT(0),		\
+	}
+
+static inline u64 gen_cookie_next(struct gen_cookie *gc)
+{
+	u64 *local_last = &get_cpu_var(*gc->local_last);
+	u64 val = *local_last;
+
+	if (__is_defined(CONFIG_SMP) &&
+	    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {

include/linux/cookie.h line 38

Quoted 10 lines
include/linux/cookie.h new file +10 -0
+		s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
+					       &gc->shared_last);
+		val = next - COOKIE_LOCAL_BATCH;
+	}
+	val++;
+	if (unlikely(!val))
+		val++;
+	*local_last = val;
+	put_cpu_var(local_last);
+	return val;

This is not interrupt safe.

Quoted 4 lines

I think sock_gen_cookie() can be called from interrupt context.

get_next_ino() is only called from process context, that is what I used get_cpu_var() and put_cpu_var()

Hmm, agree, good point. Need to experiment a bit more .. initial thinking potentially something like the below could do where we fall back to atomic counter iff we encounter nesting (which should be an extremely rare case normally).

BPF progs where this can be called from are non-preemptible, so we could actually move the temp preempt_disable/enable() from get/put_cpu_var() into a wrapper func for slow path non-BPF users as well.

static inline u64 gen_cookie_next(struct gen_cookie *gc)

{
        u64 val;

        if (likely(this_cpu_inc_return(*gc->level_nesting) == 1)) {
                u64 *local_last = this_cpu_ptr(gc->local_last);

                val = *local_last;
                if (__is_defined(CONFIG_SMP) &&
                    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {
                        s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
                                                       &gc->shared_last);
                        val = next - COOKIE_LOCAL_BATCH;
                }
                val++;
                if (unlikely(!val))
                        val++;
                *local_last = val;
        } else {
                val = atomic64_add_return(COOKIE_LOCAL_BATCH,
                                          &gc->shared_last);
        }
        this_cpu_dec(*gc->level_nesting);
        return val;
}

Thanks,
Daniel

#7 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

Quoted 60 lines · Daniel Borkmann
Quoted 1 line · Eric Dumazet

On 9/24/20 8:21 PM, Daniel Borkmann wrote:

[...]

Quoted 47 lines
Quoted 34 lines
include/linux/cookie.h new file +28 -0
new file mode 100644
index 000000000000..2488203dc004
@@ -0,0 +1,41 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __LINUX_COOKIE_H
+#define __LINUX_COOKIE_H
+
+#include <linux/atomic.h>
+#include <linux/percpu.h>
+
+struct gen_cookie {
+    u64 __percpu    *local_last;
+    atomic64_t     shared_last ____cacheline_aligned_in_smp;
+};
+
+#define COOKIE_LOCAL_BATCH    4096
+
+#define DEFINE_COOKIE(name)                    \
+    static DEFINE_PER_CPU(u64, __##name);            \
+    static struct gen_cookie name = {            \
+        .local_last    = &__##name,            \
+        .shared_last    = ATOMIC64_INIT(0),        \
+    }
+
+static inline u64 gen_cookie_next(struct gen_cookie *gc)
+{
+    u64 *local_last = &get_cpu_var(*gc->local_last);
+    u64 val = *local_last;
+
+    if (__is_defined(CONFIG_SMP) &&
+        unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {

include/linux/cookie.h line 38

Quoted 10 lines
include/linux/cookie.h new file +10 -0
+        s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
+                           &gc->shared_last);
+        val = next - COOKIE_LOCAL_BATCH;
+    }
+    val++;
+    if (unlikely(!val))
+        val++;
+    *local_last = val;
+    put_cpu_var(local_last);
+    return val;

This is not interrupt safe.

Quoted 4 lines

I think sock_gen_cookie() can be called from interrupt context.

get_next_ino() is only called from process context, that is what I used get_cpu_var() and put_cpu_var()

Hmm, agree, good point. Need to experiment a bit more .. initial thinking potentially something like the below could do where we fall back to atomic counter iff we encounter nesting (which should be an extremely rare case normally).

Quoted 8 lines

BPF progs where this can be called from are non-preemptible, so we could actually move the temp preempt_disable/enable() from get/put_cpu_var() into a wrapper func for slow path non-BPF users as well.

static inline u64 gen_cookie_next(struct gen_cookie *gc) {
        u64 val;

I presume you would use a single structure to hold level_nesting and local_last in the same cache line.

struct pcpu_gen_cookie {
    int level_nesting;
    u64 local_last;
} __aligned(16);
Quoted 14 lines

        if (likely(this_cpu_inc_return(*gc->level_nesting) == 1)) { u64 *local_last = this_cpu_ptr(gc->local_last);

                val = *local_last;
                if (__is_defined(CONFIG_SMP) &&
                    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) { s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH, &gc->shared_last); val = next - COOKIE_LOCAL_BATCH;
                }
                val++;

Quoted 2 lines

                if (unlikely(!val))
                        val++;

Note that we really expect this wrapping will never happen, with 64bit value. (We had to take care of the wrapping in get_next_ino() as it was dealing with 32bit values)

Quoted 4 lines

                *local_last = val;
        } else {
                val = atomic64_add_return(COOKIE_LOCAL_BATCH, &gc->shared_last);

Or val = atomic64_dec_return(&reverse_counter)

With reverse_counter initial value set to ATOMIC64_INIT(0) ?

This will start sending 'big cookies like 0xFFFFFFFFxxxxxxxx' to make sure applications are not breaking with them, after few months of uptime.

This would also not consume COOKIE_LOCAL_BATCH units per value, but this seems minor based on the available space.

Quoted 7 lines

        }
        this_cpu_dec(*gc->level_nesting);
        return val;
}

Thanks,
Daniel

author maintainer
#8 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

Quoted 69 lines · Eric Dumazet
Quoted 60 lines · Daniel Borkmann
Quoted 1 line · Eric Dumazet

On 9/24/20 8:21 PM, Daniel Borkmann wrote:

[...]

Quoted 47 lines
Quoted 34 lines
include/linux/cookie.h new file +28 -0
new file mode 100644
index 000000000000..2488203dc004
@@ -0,0 +1,41 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __LINUX_COOKIE_H
+#define __LINUX_COOKIE_H
+
+#include <linux/atomic.h>
+#include <linux/percpu.h>
+
+struct gen_cookie {
+    u64 __percpu    *local_last;
+    atomic64_t     shared_last ____cacheline_aligned_in_smp;
+};
+
+#define COOKIE_LOCAL_BATCH    4096
+
+#define DEFINE_COOKIE(name)                    \
+    static DEFINE_PER_CPU(u64, __##name);            \
+    static struct gen_cookie name = {            \
+        .local_last    = &__##name,            \
+        .shared_last    = ATOMIC64_INIT(0),        \
+    }
+
+static inline u64 gen_cookie_next(struct gen_cookie *gc)
+{
+    u64 *local_last = &get_cpu_var(*gc->local_last);
+    u64 val = *local_last;
+
+    if (__is_defined(CONFIG_SMP) &&
+        unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {

include/linux/cookie.h line 38

Quoted 10 lines
include/linux/cookie.h new file +10 -0
+        s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
+                           &gc->shared_last);
+        val = next - COOKIE_LOCAL_BATCH;
+    }
+    val++;
+    if (unlikely(!val))
+        val++;
+    *local_last = val;
+    put_cpu_var(local_last);
+    return val;

This is not interrupt safe.

Quoted 4 lines

I think sock_gen_cookie() can be called from interrupt context.

get_next_ino() is only called from process context, that is what I used get_cpu_var() and put_cpu_var()

Hmm, agree, good point. Need to experiment a bit more .. initial thinking potentially something like the below could do where we fall back to atomic counter iff we encounter nesting (which should be an extremely rare case normally).

Quoted 7 lines

BPF progs where this can be called from are non-preemptible, so we could actually move the temp preempt_disable/enable() from get/put_cpu_var() into a wrapper func for slow path non-BPF users as well.

static inline u64 gen_cookie_next(struct gen_cookie *gc) {
        u64 val;

Quoted 7 lines

I presume you would use a single structure to hold level_nesting and local_last in the same cache line.

struct pcpu_gen_cookie {
     int level_nesting;
     u64 local_last;
} __aligned(16);

Yes.

Quoted 12 lines
Quoted 11 lines

        if (likely(this_cpu_inc_return(*gc->level_nesting) == 1)) { u64 *local_last = this_cpu_ptr(gc->local_last);

                val = *local_last;
                if (__is_defined(CONFIG_SMP) &&
                    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) { s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH, &gc->shared_last); val = next - COOKIE_LOCAL_BATCH;
                }
                val++;

Quoted 5 lines
Quoted 2 lines

                if (unlikely(!val))
                        val++;

Note that we really expect this wrapping will never happen, with 64bit value. (We had to take care of the wrapping in get_next_ino() as it was dealing with 32bit values)

Agree, all local counters will start off at 0, but we inc right after the batch and thus never run into it anyway and neither via overflow. Will remove.

Quoted 7 lines
Quoted 4 lines

                *local_last = val;
        } else {
                val = atomic64_add_return(COOKIE_LOCAL_BATCH, &gc->shared_last);

Or val = atomic64_dec_return(&reverse_counter)

Quoted 7 lines

With reverse_counter initial value set to ATOMIC64_INIT(0) ?

This will start sending 'big cookies like 0xFFFFFFFFxxxxxxxx' to make sure applications are not breaking with them, after few months of uptime.

This would also not consume COOKIE_LOCAL_BATCH units per value, but this seems minor based on the available space.

Excellent idea, I like it given it doesn't waste COOKIE_LOCAL_BATCH space. Thanks for the feedback!

Quoted 7 lines
Quoted 7 lines

        }
        this_cpu_dec(*gc->level_nesting);
        return val;
}

Thanks,
Daniel

maintainer
#9 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

In reply to Daniel Borkmann
Quoted 5 lines · On Fri, 25 Sep 2020 00:03:14 +0200 Daniel Borkmann wrote:

static inline u64 gen_cookie_next(struct gen_cookie *gc)

{
         u64 val;

         if (likely(this_cpu_inc_return(*gc->level_nesting) == 1)) {

Is this_cpu_inc() in itself atomic?

Is there a comparison of performance of various atomic ops and locking somewhere? I wonder how this scheme would compare to a using a cmpxchg.

Quoted 5 lines
                 u64 *local_last = this_cpu_ptr(gc->local_last);

                 val = *local_last;
                 if (__is_defined(CONFIG_SMP) &&
                     unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {

Can we reasonably assume we won't have more than 4k CPUs and just statically divide this space by encoding CPU id in top bits?

Quoted 15 lines
                         s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
                                                        &gc->shared_last);
                         val = next - COOKIE_LOCAL_BATCH;
                 }
                 val++;
                 if (unlikely(!val))
                         val++;
                 *local_last = val;
         } else {
                 val = atomic64_add_return(COOKIE_LOCAL_BATCH,
                                           &gc->shared_last);
         }
         this_cpu_dec(*gc->level_nesting);
         return val;
}
#10 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

On 9/25/20 5:00 PM, Jakub Kicinski wrote:

                    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {
Quoted 3 lines

Can we reasonably assume we won't have more than 4k CPUs and just statically divide this space by encoding CPU id in top bits?

This might give some food to side channel attacks, since this would give an indication of cpu that allocated the id.

Also, I hear that some distros enabled 8K cpus.

maintainer
#11 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

Quoted 2 lines · On Fri, 25 Sep 2020 17:15:17 +0200 Eric Dumazet wrote:
Quoted 1 line · Jakub Kicinski

Is this_cpu_inc() in itself atomic?

To answer my own question - it is :)

Quoted 9 lines
Quoted 4 lines
                    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0))

Can we reasonably assume we won't have more than 4k CPUs and just statically divide this space by encoding CPU id in top bits?

This might give some food to side channel attacks, since this would give an indication of cpu that allocated the id.

Also, I hear that some distros enabled 8K cpus.

Ok :(

#12 lore raw

Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

Quoted 11 lines · Jakub Kicinski
Quoted 2 lines · On Fri, 25 Sep 2020 17:15:17 +0200 Eric Dumazet wrote:
Quoted 1 line · Jakub Kicinski

Is this_cpu_inc() in itself atomic?

To answer my own question - it is :)

Quoted 5 lines
Quoted 4 lines
                    unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0))

Can we reasonably assume we won't have more than 4k CPUs and just statically divide this space by encoding CPU id in top bits?

Quoted 7 lines
Quoted 4 lines

This might give some food to side channel attacks, since this would give an indication of cpu that allocated the id.

Also, I hear that some distros enabled 8K cpus.

Ok :(

I was not really serious about the side channel attacks, just some thought about possible implications :)

Even with 8192 max cpus, splitting space into 2^(64-13) blocks would be fine I think.

author maintainer
#13 lore raw

PATCH 3/6 bpf: add redirect_neigh helper as redirect drop-in

In reply to Daniel Borkmann

Add a redirect_neigh() helper as redirect() drop-in replacement for the xmit side. Main idea for the helper is to be very similar in semantics to the latter just that the skb gets injected into the neighboring subsystem in order to let the stack do the work it knows best anyway to populate the L2 addresses of the packet and then hand over to dev_queue_xmit() as redirect() does.

This solves two bigger items: i) skbs don't need to go up to the stack on the host facing veth ingress side for traffic egressing the container to achieve the same for populating L2 which also has the huge advantage that ii) the skb->sk won't get orphaned in ip_rcv_core() when entering the IP routing layer on the host stack.

Given that skb->sk neither gets orphaned when crossing the netns as per 9c4c325252c5 ("skbuff: preserve sock reference when scrubbing the skb.") the helper can then push the skbs directly to the phys device where FQ scheduler can do its work and TCP stack gets proper back-pressure given we hold on to skb->sk as long as skb is still residing in queues.

With the helper used in BPF data path to then push the skb to the phys device, I observed a stable/consistent TCP_STREAM improvement on veth devices for traffic going container -> host -> host -> container from ~10Gbps to ~15Gbps for a single stream in my test environment.

---
 include/linux/skbuff.h         |   5 +
 include/uapi/linux/bpf.h       |  14 ++
 net/core/filter.c              | 256 +++++++++++++++++++++++++++++++--
 tools/include/uapi/linux/bpf.h |  14 ++
 4 files changed, 276 insertions(+), 13 deletions(-)
include/linux/skbuff.h +5 -0
index 46881d902124..ec3254acca77 100644
@@ -2539,6 +2539,11 @@ static inline int skb_mac_header_was_set(const struct sk_buff *skb)
 	return skb->mac_header != (typeof(skb->mac_header))~0U;
 }
 
+static inline void skb_unset_mac_header(struct sk_buff *skb)
+{
+	skb->mac_header = (typeof(skb->mac_header))~0U;
+}
+
 static inline void skb_reset_mac_header(struct sk_buff *skb)
 {
 	skb->mac_header = skb->data - skb->head;
include/uapi/linux/bpf.h +14 -0
index 48ecf246d047..95337b650d7f 100644
@@ -3595,6 +3595,19 @@ union bpf_attr {
  * 		associated socket instead of the current process.
  * 	Return
  * 		The id is returned or 0 in case the id could not be retrieved.
+ *
+ * long bpf_redirect_neigh(u32 ifindex, u64 flags)
+ * 	Description
+ * 		Redirect the packet to another net device of index *ifindex*
+ * 		and fill in L2 addresses from neighboring subsystem. This helper
+ * 		is somewhat similar to **bpf_redirect**\ (), except that it
+ * 		fills in e.g. MAC addresses based on the L3 information from
+ * 		the packet. This helper is supported for IPv4 and IPv6 protocols.
+ * 		The *flags* argument is reserved and must be 0. The helper is
+ * 		currently only supported for tc BPF program types.
+ * 	Return
+ * 		The helper returns **TC_ACT_REDIRECT** on success or
+ * 		**TC_ACT_SHOT** on error.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3747,6 +3760,7 @@ union bpf_attr {
 	FN(d_path),			\
 	FN(copy_from_user),		\
 	FN(skb_cgroup_classid),		\
+	FN(redirect_neigh),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
net/core/filter.c +243 -13

291 lines not on this page · raw message

tools/include/uapi/linux/bpf.h +14 -0
index 48ecf246d047..95337b650d7f 100644
@@ -3595,6 +3595,19 @@ union bpf_attr {
  * 		associated socket instead of the current process.
  * 	Return
  * 		The id is returned or 0 in case the id could not be retrieved.
+ *
+ * long bpf_redirect_neigh(u32 ifindex, u64 flags)
+ * 	Description
+ * 		Redirect the packet to another net device of index *ifindex*
+ * 		and fill in L2 addresses from neighboring subsystem. This helper
+ * 		is somewhat similar to **bpf_redirect**\ (), except that it
+ * 		fills in e.g. MAC addresses based on the L3 information from
+ * 		the packet. This helper is supported for IPv4 and IPv6 protocols.
+ * 		The *flags* argument is reserved and must be 0. The helper is
+ * 		currently only supported for tc BPF program types.
+ * 	Return
+ * 		The helper returns **TC_ACT_REDIRECT** on success or
+ * 		**TC_ACT_SHOT** on error.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3747,6 +3760,7 @@ union bpf_attr {
 	FN(d_path),			\
 	FN(copy_from_user),		\
 	FN(skb_cgroup_classid),		\
+	FN(redirect_neigh),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
Signature · 1 line
2.21.0
#14 lore raw

Re: [PATCH bpf-next 3/6] bpf: add redirect_neigh helper as redirect drop-in

Quoted 14 lines · Daniel Borkmann
net/core/filter.c +6 -0
index 0f913755bcba..19caa2fc21e8 100644
@@ -2160,6 +2160,205 @@ static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
 		return __bpf_redirect_no_mac(skb, dev, flags);
 }
 
+#if IS_ENABLED(CONFIG_IPV6)
+static int bpf_out_neigh_v6(struct net *net, struct sk_buff *skb)
+{
+	struct dst_entry *dst = skb_dst(skb);
+	struct net_device *dev = dst->dev;
+	const struct in6_addr *nexthop;

net/core/filter.c static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev, · line 2178

Quoted 10 lines
net/core/filter.c +10 -0
+	struct neighbour *neigh;
+
+	if (dev_xmit_recursion())
+		goto out_rec;
+	skb->dev = dev;
+	rcu_read_lock_bh();
+	nexthop = rt6_nexthop((struct rt6_info *)dst, &ipv6_hdr(skb)->daddr);
+	neigh = __ipv6_neigh_lookup_noref_stub(dev, nexthop);
+	if (unlikely(!neigh))
+		neigh = __neigh_create(ipv6_stub->nd_tbl, nexthop, dev, false);

the last 3 lines can be replaced with ip_neigh_gw6.

Quoted 21 lines
net/core/filter.c +21 -0
+	if (likely(!IS_ERR(neigh))) {
+		int ret;
+
+		sock_confirm_neigh(skb, neigh);
+		dev_xmit_recursion_inc();
+		ret = neigh_output(neigh, skb, false);
+		dev_xmit_recursion_dec();
+		rcu_read_unlock_bh();
+		return ret;
+	}
+	rcu_read_unlock_bh();
+	IP6_INC_STATS(dev_net(dst->dev),
+		      ip6_dst_idev(dst), IPSTATS_MIB_OUTNOROUTES);
+out_drop:
+	kfree_skb(skb);
+	return -EINVAL;
+out_rec:
+	net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
+	goto out_drop;
+}
+

...

net/core/filter.c static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev, · line 2249

Quoted 6 lines
net/core/filter.c +6 -0
+
+#if IS_ENABLED(CONFIG_INET)
+static int bpf_out_neigh_v4(struct net *net, struct sk_buff *skb)
+{
+	struct dst_entry *dst = skb_dst(skb);
+	struct rtable *rt = (struct rtable *)dst;

please use container_of here; I'd like to see the typecasts get removed.

Quoted 7 lines
net/core/filter.c +7 -0
+	struct net_device *dev = dst->dev;
+	u32 hh_len = LL_RESERVED_SPACE(dev);
+	struct neighbour *neigh;
+	bool is_v6gw = false;
+
+	if (dev_xmit_recursion())
+		goto out_rec;
author maintainer
#15 lore raw

Re: [PATCH bpf-next 3/6] bpf: add redirect_neigh helper as redirect drop-in

Quoted 19 lines · David Ahern
Quoted 18 lines · Daniel Borkmann
net/core/filter.c +10 -0
index 0f913755bcba..19caa2fc21e8 100644
@@ -2160,6 +2160,205 @@ static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
  		return __bpf_redirect_no_mac(skb, dev, flags);
  }
 
+#if IS_ENABLED(CONFIG_IPV6)
+static int bpf_out_neigh_v6(struct net *net, struct sk_buff *skb)
+{
+	struct dst_entry *dst = skb_dst(skb);
+	struct net_device *dev = dst->dev;
+	const struct in6_addr *nexthop;
+	struct neighbour *neigh;
+
+	if (dev_xmit_recursion())
+		goto out_rec;
Quoted 8 lines

net/core/filter.c static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev, · line 2178

Quoted 6 lines
net/core/filter.c +6 -0
+	skb->dev = dev;
+	rcu_read_lock_bh();
+	nexthop = rt6_nexthop((struct rt6_info *)dst, &ipv6_hdr(skb)->daddr);
+	neigh = __ipv6_neigh_lookup_noref_stub(dev, nexthop);
+	if (unlikely(!neigh))
+		neigh = __neigh_create(ipv6_stub->nd_tbl, nexthop, dev, false);

the last 3 lines can be replaced with ip_neigh_gw6.

Ah, nice, I wasn't aware of that one. I'll take it. :)

Quoted 24 lines
Quoted 21 lines
net/core/filter.c +21 -0
+	if (likely(!IS_ERR(neigh))) {
+		int ret;
+
+		sock_confirm_neigh(skb, neigh);
+		dev_xmit_recursion_inc();
+		ret = neigh_output(neigh, skb, false);
+		dev_xmit_recursion_dec();
+		rcu_read_unlock_bh();
+		return ret;
+	}
+	rcu_read_unlock_bh();
+	IP6_INC_STATS(dev_net(dst->dev),
+		      ip6_dst_idev(dst), IPSTATS_MIB_OUTNOROUTES);
+out_drop:
+	kfree_skb(skb);
+	return -EINVAL;
+out_rec:
+	net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
+	goto out_drop;
+}
+

...

Quoted 7 lines

net/core/filter.c static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev, · line 2249

Quoted 5 lines
net/core/filter.c +5 -0
+#if IS_ENABLED(CONFIG_INET)
+static int bpf_out_neigh_v4(struct net *net, struct sk_buff *skb)
+{
+	struct dst_entry *dst = skb_dst(skb);
+	struct rtable *rt = (struct rtable *)dst;

please use container_of here; I'd like to see the typecasts get removed.

Will do, thx!

Quoted 7 lines
Quoted 7 lines
net/core/filter.c +7 -0
+	struct net_device *dev = dst->dev;
+	u32 hh_len = LL_RESERVED_SPACE(dev);
+	struct neighbour *neigh;
+	bool is_v6gw = false;
+
+	if (dev_xmit_recursion())
+		goto out_rec;
author maintainer
#16 lore raw

PATCH 4/6 bpf, libbpf: add bpf_tail_call_static helper for bpf programs

In reply to Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

[0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035

---
 tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +32 -0
index 1106777df00b..18b75a4c82e6 100644
@@ -53,6 +53,38 @@
 	})
 #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()	__builtin_trap()
+#endif
+
+static __always_inline void
+bpf_tail_call_static(void *ctx, const void *map, const __u32 slot)
+{
+	if (!__builtin_constant_p(slot))
+		__throw_build_bug();
+
+	/*
+	 * Don't gamble, but _guarantee_ that LLVM won't optimize setting
+	 * r2 and r3 from different paths ending up at the same call insn as
+	 * otherwise we won't be able to use the jmpq/nopl retpoline-free
+	 * patching by the x86-64 JIT in the kernel.
+	 *
+	 * Note on clobber list: we need to stay in-line with BPF calling
+	 * convention, so even if we don't end up using r0, r4, r5, we need
+	 * to mark them as clobber so that LLVM doesn't end up using them
+	 * before / after the call.
+	 */
+	asm volatile("r1 = %[ctx]\n\t"
+		     "r2 = %[map]\n\t"
+		     "r3 = %[slot]\n\t"
+		     "call 12\n\t"
+		     :: [ctx]"r"(ctx), [map]"r"(map), [slot]"i"(slot)
+		     : "r0", "r1", "r2", "r3", "r4", "r5");
+}
+
 /*
  * Helper structure used by eBPF C program
  * to describe BPF map attributes to libbpf loader
Signature · 1 line
2.21.0
#17 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

Quoted 27 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

[0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035

---
 tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++
 1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 61

Quoted 9 lines
tools/lib/bpf/bpf_helpers.h +6 -0
@@ -53,6 +53,38 @@
        })
 #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()   __builtin_trap()
+#endif

this will become part of libbpf stable API, do we want/need to expose it? If we want to expose it, then we should probably provide a better description.

But also curious, how is it better than _Static_assert() (see test_cls_redirect.c), which also allows to provide a better error message?

tools/lib/bpf/bpf_helpers.h line 74

Quoted 13 lines
tools/lib/bpf/bpf_helpers.h +13 -0
+
+static __always_inline void
+bpf_tail_call_static(void *ctx, const void *map, const __u32 slot)
+{
+       if (!__builtin_constant_p(slot))
+               __throw_build_bug();
+
+       /*
+        * Don't gamble, but _guarantee_ that LLVM won't optimize setting
+        * r2 and r3 from different paths ending up at the same call insn as
+        * otherwise we won't be able to use the jmpq/nopl retpoline-free
+        * patching by the x86-64 JIT in the kernel.
+        *

So the clobbering comment below is completely clear. But this one is less clear without some sort of example situation in which bad things happen. Do you mind providing some pseudo-C example in which the compiler will optimize things in such a way that the tail call patching won't happen?

Quoted 19 lines
tools/lib/bpf/bpf_helpers.h +13 -1
+        * Note on clobber list: we need to stay in-line with BPF calling
+        * convention, so even if we don't end up using r0, r4, r5, we need
+        * to mark them as clobber so that LLVM doesn't end up using them
+        * before / after the call.
+        */
+       asm volatile("r1 = %[ctx]\n\t"
+                    "r2 = %[map]\n\t"
+                    "r3 = %[slot]\n\t"
+                    "call 12\n\t"
+                    :: [ctx]"r"(ctx), [map]"r"(map), [slot]"i"(slot)
+                    : "r0", "r1", "r2", "r3", "r4", "r5");
+}
+
 /*
  * Helper structure used by eBPF C program
  * to describe BPF map attributes to libbpf loader
--

2.21.0

author maintainer
#18 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

Quoted 42 lines · Andrii Nakryiko
Quoted 27 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

   [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035
---
  tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++
  1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 61

Quoted 9 lines
tools/lib/bpf/bpf_helpers.h +6 -0
@@ -53,6 +53,38 @@
         })
  #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()   __builtin_trap()
+#endif

this will become part of libbpf stable API, do we want/need to expose it? If we want to expose it, then we should probably provide a better description.

Quoted 3 lines

But also curious, how is it better than _Static_assert() (see test_cls_redirect.c), which also allows to provide a better error message?

Need to get back to you whether that has same semantics. We use the __throw_build_bug() also in __bpf_memzero() and friends [0] as a way to trigger a hard build bug if we hit a default switch-case [0], so we detect unsupported sizes which are not covered by the implementation yet. If _Static_assert (0, "foo") does the trick, we could also use that; will check with our code base.

[0] https://github.com/cilium/cilium/blob/master/bpf/include/bpf/builtins.h

Quoted 10 lines
Quoted 10 lines
tools/lib/bpf/bpf_helpers.h +10 -0
+static __always_inline void
+bpf_tail_call_static(void *ctx, const void *map, const __u32 slot)
+{
+       if (!__builtin_constant_p(slot))
+               __throw_build_bug();
+
+       /*
+        * Don't gamble, but _guarantee_ that LLVM won't optimize setting
+        * r2 and r3 from different paths ending up at the same call insn as
+        * otherwise we won't be able to use the jmpq/nopl retpoline-free
Quoted 8 lines

tools/lib/bpf/bpf_helpers.h line 74

Quoted 2 lines
tools/lib/bpf/bpf_helpers.h +2 -0
+        * patching by the x86-64 JIT in the kernel.
+        *

So the clobbering comment below is completely clear. But this one is less clear without some sort of example situation in which bad things happen. Do you mind providing some pseudo-C example in which the compiler will optimize things in such a way that the tail call patching won't happen?

The details are pretty much here [1] and mentioned at plumbers, so if we end up from different paths with different map or const key at the same tail-call call insn, then the record_func_key() will determine that we need to emit retpoline instead of patching which could have occured with two tail-call call insns, for example. This helper is just to guarantee that the latter will always happen.

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=d2e4c1e6c2947269346054ac8937ccfe9e0bcc6b

Quoted 19 lines
Quoted 19 lines
tools/lib/bpf/bpf_helpers.h +13 -1
+        * Note on clobber list: we need to stay in-line with BPF calling
+        * convention, so even if we don't end up using r0, r4, r5, we need
+        * to mark them as clobber so that LLVM doesn't end up using them
+        * before / after the call.
+        */
+       asm volatile("r1 = %[ctx]\n\t"
+                    "r2 = %[map]\n\t"
+                    "r3 = %[slot]\n\t"
+                    "call 12\n\t"
+                    :: [ctx]"r"(ctx), [map]"r"(map), [slot]"i"(slot)
+                    : "r0", "r1", "r2", "r3", "r4", "r5");
+}
+
  /*
   * Helper structure used by eBPF C program
   * to describe BPF map attributes to libbpf loader
--

2.21.0

author maintainer
#19 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

Quoted 47 lines · Daniel Borkmann
Quoted 42 lines · Andrii Nakryiko
Quoted 27 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

   [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035

---
  tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++

  1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 61

Quoted 9 lines
tools/lib/bpf/bpf_helpers.h +6 -0
@@ -53,6 +53,38 @@
          })
   #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()   __builtin_trap()
+#endif

this will become part of libbpf stable API, do we want/need to expose it? If we want to expose it, then we should probably provide a better description.

Quoted 3 lines

But also curious, how is it better than _Static_assert() (see test_cls_redirect.c), which also allows to provide a better error message?

Quoted 5 lines

Need to get back to you whether that has same semantics. We use the __throw_build_bug() also in __bpf_memzero() and friends [0] as a way to trigger a hard build bug if we hit a default switch-case [0], so we detect unsupported sizes which are not covered by the implementation yet. If _Static_assert (0, "foo") does the trick, we could also use that; will check with our code base.

So _Static_assert() won't work here, for example consider:

  # cat f1.c
  int main(void)
  {
	if (0)
		_Static_assert(0, "foo");
	return 0;
  }
  # clang -target bpf -Wall -O2 -c f1.c -o f1.o
  f1.c:4:3: error: expected expression
                _Static_assert(0, "foo");
                ^
  1 error generated.

In order for it to work as required form the use-case, the _Static_assert() must not trigger here given the path is unreachable and will be optimized away. I'll add a comment to the __throw_build_bug() helper. Given libbpf we should probably also prefix with bpf_. If you see a better name that would fit, pls let me know.

Quoted 1 line

  [0] https://github.com/cilium/cilium/blob/master/bpf/include/bpf/builtins.h

Thanks,
Daniel

author maintainer
#20 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

Quoted 56 lines · Daniel Borkmann
Quoted 47 lines · Daniel Borkmann
Quoted 42 lines · Andrii Nakryiko
Quoted 27 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

   [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035

---
  tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++

  1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 61

Quoted 9 lines
tools/lib/bpf/bpf_helpers.h +6 -0
@@ -53,6 +53,38 @@
          })
   #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()   __builtin_trap()
+#endif

this will become part of libbpf stable API, do we want/need to expose it? If we want to expose it, then we should probably provide a better description.

Quoted 3 lines

But also curious, how is it better than _Static_assert() (see test_cls_redirect.c), which also allows to provide a better error message?

Quoted 5 lines

Need to get back to you whether that has same semantics. We use the __throw_build_bug() also in __bpf_memzero() and friends [0] as a way to trigger a hard build bug if we hit a default switch-case [0], so we detect unsupported sizes which are not covered by the implementation yet. If _Static_assert (0, "foo") does the trick, we could also use that; will check with our code base.

So _Static_assert() won't work here, for example consider:

Quoted 12 lines
   # cat f1.c
   int main(void)
   {
     if (0)
         _Static_assert(0, "foo");
     return 0;
   }
   # clang -target bpf -Wall -O2 -c f1.c -o f1.o
   f1.c:4:3: error: expected expression
                 _Static_assert(0, "foo");
                 ^
   1 error generated.

.. aaand it looks like I need some more coffee. ;-) But result is the same after all:

  # clang -target bpf -Wall -O2 -c f1.c -o f1.o
  f1.c:4:3: error: static_assert failed "foo"
                _Static_assert(0, "foo");
                ^              ~
  1 error generated.
  # cat f1.c
  int main(void)
  {
	if (0) {
		_Static_assert(0, "foo");
	}
	return 0;
  }
Quoted 8 lines

In order for it to work as required form the use-case, the _Static_assert() must not trigger here given the path is unreachable and will be optimized away. I'll add a comment to the __throw_build_bug() helper. Given libbpf we should probably also prefix with bpf_. If you see a better name that would fit, pls let me know.

Quoted 1 line

   [0] https://github.com/cilium/cilium/blob/master/bpf/include/bpf/builtins.h

Thanks,
Daniel

#21 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

Quoted 79 lines · Daniel Borkmann
Quoted 56 lines · Daniel Borkmann
Quoted 47 lines · Daniel Borkmann
Quoted 42 lines · Andrii Nakryiko
Quoted 27 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

   [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035
---
  tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++
  1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 61

Quoted 9 lines
tools/lib/bpf/bpf_helpers.h +6 -0
@@ -53,6 +53,38 @@
         })
  #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()   __builtin_trap()
+#endif

this will become part of libbpf stable API, do we want/need to expose it? If we want to expose it, then we should probably provide a better description.

Quoted 3 lines

But also curious, how is it better than _Static_assert() (see test_cls_redirect.c), which also allows to provide a better error message?

Quoted 5 lines

Need to get back to you whether that has same semantics. We use the __throw_build_bug() also in __bpf_memzero() and friends [0] as a way to trigger a hard build bug if we hit a default switch-case [0], so we detect unsupported sizes which are not covered by the implementation yet. If _Static_assert (0, "foo") does the trick, we could also use that; will check with our code base.

So _Static_assert() won't work here, for example consider:

Quoted 12 lines
   # cat f1.c
   int main(void)
   {
     if (0)
         _Static_assert(0, "foo");
     return 0;
   }
   # clang -target bpf -Wall -O2 -c f1.c -o f1.o
   f1.c:4:3: error: expected expression
                 _Static_assert(0, "foo");
                 ^
   1 error generated.

.. aaand it looks like I need some more coffee. ;-) But result is the same after all:

   # clang -target bpf -Wall -O2 -c f1.c -o f1.o
   f1.c:4:3: error: static_assert failed "foo"
                 _Static_assert(0, "foo");
                 ^              ~
   1 error generated.
Quoted 8 lines
   # cat f1.c
   int main(void)
   {
        if (0) {
                _Static_assert(0, "foo");
        }
        return 0;
   }

You need still more :-P. For you use case it will look like this:

$ cat test-bla.c
int bar(int x) {
       _Static_assert(!__builtin_constant_p(x), "not a constant!");
       return x;
}
int foo() {
        bar(123);
        return 0;
}
$ clang -target bpf -O2 -c test-bla.c -o test-bla.o
$ echo $?
0

But in general to ensure unreachable code it's probably useful anyway to have this. How about calling it __bpf_build_error() or maybe even __bpf_unreachable()?

Quoted 10 lines
Quoted 9 lines

In order for it to work as required form the use-case, the _Static_assert() must not trigger here given the path is unreachable and will be optimized away. I'll add a comment to the __throw_build_bug() helper. Given libbpf we should probably also prefix with bpf_. If you see a better name that would fit, pls let me know.

Quoted 1 line

[0] https://github.com/cilium/cilium/blob/master/bpf/include/bpf/builtins.h

Thanks,
Daniel

author maintainer
#22 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

Quoted 96 lines · Andrii Nakryiko
Quoted 78 lines · Daniel Borkmann
Quoted 56 lines · Daniel Borkmann
Quoted 47 lines · Daniel Borkmann
Quoted 42 lines · Andrii Nakryiko
Quoted 27 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

    [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035
---
   tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++
   1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 61

Quoted 9 lines
tools/lib/bpf/bpf_helpers.h +6 -0
@@ -53,6 +53,38 @@
          })
   #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()   __builtin_trap()
+#endif

this will become part of libbpf stable API, do we want/need to expose it? If we want to expose it, then we should probably provide a better description.

Quoted 3 lines

But also curious, how is it better than _Static_assert() (see test_cls_redirect.c), which also allows to provide a better error message?

Quoted 5 lines

Need to get back to you whether that has same semantics. We use the __throw_build_bug() also in __bpf_memzero() and friends [0] as a way to trigger a hard build bug if we hit a default switch-case [0], so we detect unsupported sizes which are not covered by the implementation yet. If _Static_assert (0, "foo") does the trick, we could also use that; will check with our code base.

So _Static_assert() won't work here, for example consider:

Quoted 12 lines
    # cat f1.c
    int main(void)
    {
      if (0)
          _Static_assert(0, "foo");
      return 0;
    }
    # clang -target bpf -Wall -O2 -c f1.c -o f1.o
    f1.c:4:3: error: expected expression
                  _Static_assert(0, "foo");
                  ^
    1 error generated.

.. aaand it looks like I need some more coffee. ;-) But result is the same after all:

    # clang -target bpf -Wall -O2 -c f1.c -o f1.o
    f1.c:4:3: error: static_assert failed "foo"
                  _Static_assert(0, "foo");
                  ^              ~
    1 error generated.
Quoted 8 lines
    # cat f1.c
    int main(void)
    {
         if (0) {
                 _Static_assert(0, "foo");
         }
         return 0;
    }

You need still more :-P. For you use case it will look like this:

$ cat test-bla.c
int bar(int x) {
        _Static_assert(!__builtin_constant_p(x), "not a constant!");
        return x;
}
Quoted 7 lines
int foo() {
         bar(123);
         return 0;
}
$ clang -target bpf -O2 -c test-bla.c -o test-bla.o
$ echo $?
0

Right, but that won't work for example for the use case to detect switch cases which fall into default case as mentioned with the mem* optimizations earlier in this thread.

Quoted 3 lines

But in general to ensure unreachable code it's probably useful anyway to have this. How about calling it __bpf_build_error() or maybe even __bpf_unreachable()?

I think the __bpf_unreachable() sounds best to me, will use that.

Quoted 9 lines
Quoted 9 lines
Quoted 8 lines

In order for it to work as required form the use-case, the _Static_assert() must not trigger here given the path is unreachable and will be optimized away. I'll add a comment to the __throw_build_bug() helper. Given libbpf we should probably also prefix with bpf_. If you see a better name that would fit, pls let me know.

Quoted 1 line

[0] https://github.com/cilium/cilium/blob/master/bpf/include/bpf/builtins.h

Thanks,
Daniel

#23 lore raw

Re: [PATCH bpf-next 4/6] bpf, libbpf: add bpf_tail_call_static helper for bpf programs

In reply to Daniel Borkmann
Quoted 26 lines · Daniel Borkmann

Port of tail_call_static() helper function from Cilium's BPF code base [0] to libbpf, so others can easily consume it as well. We've been using this in production code for some time now. The main idea is that we guarantee that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the JITed BPF insns with direct jumps instead of having to fall back to using expensive retpolines. By using inline asm, we guarantee that the compiler won't merge the call from different paths with potentially different content of r2/r3.

We're also using __throw_build_bug() macro in different places as a neat trick to trigger compilation errors when compiler does not remove code at compilation time. This works for the BPF backend as it does not implement the __builtin_trap().

   [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035
---
  tools/lib/bpf/bpf_helpers.h | 32 ++++++++++++++++++++++++++++++++
  1 file changed, 32 insertions(+)
tools/lib/bpf/bpf_helpers.h +0 -0
index 1106777df00b..18b75a4c82e6 100644

tools/lib/bpf/bpf_helpers.h line 60

Quoted 8 lines
tools/lib/bpf/bpf_helpers.h +5 -0
@@ -53,6 +53,38 @@
  	})
  #endif
 
+/*
+ * Misc useful helper macros
+ */
+#ifndef __throw_build_bug
+# define __throw_build_bug()	__builtin_trap()

Just some general comments below. The patch itself is fine to me.

I guess we will never implement a 'trap' insn? The only possible use I know is for failed CORE relocation, which currently encoded as an illegal call insn.

Quoted 15 lines
tools/lib/bpf/bpf_helpers.h +15 -0
+#endif
+
+static __always_inline void
+bpf_tail_call_static(void *ctx, const void *map, const __u32 slot)
+{
+	if (!__builtin_constant_p(slot))
+		__throw_build_bug();
+
+	/*
+	 * Don't gamble, but _guarantee_ that LLVM won't optimize setting
+	 * r2 and r3 from different paths ending up at the same call insn as
+	 * otherwise we won't be able to use the jmpq/nopl retpoline-free
+	 * patching by the x86-64 JIT in the kernel.
+	 *
+	 * Note on clobber list: we need to stay in-line with BPF calling

tools/lib/bpf/bpf_helpers.h line 85

Quoted 10 lines
tools/lib/bpf/bpf_helpers.h +10 -0
+	 * convention, so even if we don't end up using r0, r4, r5, we need
+	 * to mark them as clobber so that LLVM doesn't end up using them
+	 * before / after the call.
+	 */
+	asm volatile("r1 = %[ctx]\n\t"
+		     "r2 = %[map]\n\t"
+		     "r3 = %[slot]\n\t"
+		     "call 12\n\t"
+		     :: [ctx]"r"(ctx), [map]"r"(map), [slot]"i"(slot)
+		     : "r0", "r1", "r2", "r3", "r4", "r5");

Clever scheme to enforce a constant!

This particular case is to avoid sinking common code.
We have a similar use case in https://reviews.llvm.org/D87153
which is to avoid PHI merging of two relocation globals like
phi = [reloc_global1, reloc_global2] // reachable from two paths ... phi ...

The solution is to insert bpf internal __builtin
val = __builtin_bpf_passthrough(seq_num, val)
in proper places to prevent sinking common code.

I guess in this example, we could prevent this with compiler inserting passthrough builtin's like
...
ret = bpf_tail_call(ctx, map, slot)
ret = __builtin_bpf_passthrough(seq_num, ret)
...

If in the future, such a builtin is proved useful for bpf program writers as well. we could just define
val = __builtin_bpf_passthrough(val)
and internally, it will be transformed to
val = llvm.bpf.passthrough(seq_num, val)

Quoted 6 lines
tools/lib/bpf/bpf_helpers.h +2 -0
+}
+
  /*
   * Helper structure used by eBPF C program
   * to describe BPF map attributes to libbpf loader
author maintainer
#24 lore raw

PATCH 5/6 bpf, selftests: use bpf_tail_call_static where appropriate

In reply to Daniel Borkmann

For those locations where we use an immediate tail call map index use the newly added bpf_tail_call_static() helper.

---
 tools/testing/selftests/bpf/progs/bpf_flow.c  | 12 ++++----
 tools/testing/selftests/bpf/progs/tailcall1.c | 28 +++++++++----------
 tools/testing/selftests/bpf/progs/tailcall2.c | 14 +++++-----
 tools/testing/selftests/bpf/progs/tailcall3.c |  4 +--
 .../selftests/bpf/progs/tailcall_bpf2bpf1.c   |  4 +--
 .../selftests/bpf/progs/tailcall_bpf2bpf2.c   |  6 ++--
 .../selftests/bpf/progs/tailcall_bpf2bpf3.c   |  6 ++--
 .../selftests/bpf/progs/tailcall_bpf2bpf4.c   |  6 ++--
 8 files changed, 40 insertions(+), 40 deletions(-)
tools/testing/selftests/bpf/progs/bpf_flow.c +6 -6
index de6de9221518..5a65f6b51377 100644
@@ -118,18 +118,18 @@ static __always_inline int parse_eth_proto(struct __sk_buff *skb, __be16 proto)
 
 	switch (proto) {
 	case bpf_htons(ETH_P_IP):
-		bpf_tail_call(skb, &jmp_table, IP);
+		bpf_tail_call_static(skb, &jmp_table, IP);
 		break;
 	case bpf_htons(ETH_P_IPV6):
-		bpf_tail_call(skb, &jmp_table, IPV6);
+		bpf_tail_call_static(skb, &jmp_table, IPV6);
 		break;
 	case bpf_htons(ETH_P_MPLS_MC):
 	case bpf_htons(ETH_P_MPLS_UC):
-		bpf_tail_call(skb, &jmp_table, MPLS);
+		bpf_tail_call_static(skb, &jmp_table, MPLS);
 		break;
 	case bpf_htons(ETH_P_8021Q):
 	case bpf_htons(ETH_P_8021AD):
-		bpf_tail_call(skb, &jmp_table, VLAN);
+		bpf_tail_call_static(skb, &jmp_table, VLAN);
 		break;
 	default:
 		/* Protocol not supported */
@@ -246,10 +246,10 @@ static __always_inline int parse_ipv6_proto(struct __sk_buff *skb, __u8 nexthdr)
 	switch (nexthdr) {
 	case IPPROTO_HOPOPTS:
 	case IPPROTO_DSTOPTS:
-		bpf_tail_call(skb, &jmp_table, IPV6OP);
+		bpf_tail_call_static(skb, &jmp_table, IPV6OP);
 		break;
 	case IPPROTO_FRAGMENT:
-		bpf_tail_call(skb, &jmp_table, IPV6FR);
+		bpf_tail_call_static(skb, &jmp_table, IPV6FR);
 		break;
 	default:
 		return parse_ip_proto(skb, nexthdr);
tools/testing/selftests/bpf/progs/tailcall1.c +14 -14
index 1f407e65ae52..7115bcefbe8a 100644
@@ -26,20 +26,20 @@ int entry(struct __sk_buff *skb)
 	/* Multiple locations to make sure we patch
 	 * all of them.
 	 */
-	bpf_tail_call(skb, &jmp_table, 0);
-	bpf_tail_call(skb, &jmp_table, 0);
-	bpf_tail_call(skb, &jmp_table, 0);
-	bpf_tail_call(skb, &jmp_table, 0);
-
-	bpf_tail_call(skb, &jmp_table, 1);
-	bpf_tail_call(skb, &jmp_table, 1);
-	bpf_tail_call(skb, &jmp_table, 1);
-	bpf_tail_call(skb, &jmp_table, 1);
-
-	bpf_tail_call(skb, &jmp_table, 2);
-	bpf_tail_call(skb, &jmp_table, 2);
-	bpf_tail_call(skb, &jmp_table, 2);
-	bpf_tail_call(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
+
+	bpf_tail_call_static(skb, &jmp_table, 1);
+	bpf_tail_call_static(skb, &jmp_table, 1);
+	bpf_tail_call_static(skb, &jmp_table, 1);
+	bpf_tail_call_static(skb, &jmp_table, 1);
+
+	bpf_tail_call_static(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 2);
 
 	return 3;
 }
tools/testing/selftests/bpf/progs/tailcall2.c +7 -7
index a093e739cf0e..0431e4fe7efd 100644
@@ -13,14 +13,14 @@ struct {
 SEC("classifier/0")
 int bpf_func_0(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 1);
+	bpf_tail_call_static(skb, &jmp_table, 1);
 	return 0;
 }
 
 SEC("classifier/1")
 int bpf_func_1(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 2);
 	return 1;
 }
 
@@ -33,25 +33,25 @@ int bpf_func_2(struct __sk_buff *skb)
 SEC("classifier/3")
 int bpf_func_3(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 4);
+	bpf_tail_call_static(skb, &jmp_table, 4);
 	return 3;
 }
 
 SEC("classifier/4")
 int bpf_func_4(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 3);
+	bpf_tail_call_static(skb, &jmp_table, 3);
 	return 4;
 }
 
 SEC("classifier")
 int entry(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 	/* Check multi-prog update. */
-	bpf_tail_call(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 2);
 	/* Check tail call limit. */
-	bpf_tail_call(skb, &jmp_table, 3);
+	bpf_tail_call_static(skb, &jmp_table, 3);
 	return 3;
 }
 
tools/testing/selftests/bpf/progs/tailcall3.c +2 -2
index cabda877cf0a..739dc2a51e74 100644
@@ -16,14 +16,14 @@ SEC("classifier/0")
 int bpf_func_0(struct __sk_buff *skb)
 {
 	count++;
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 	return 1;
 }
 
 SEC("classifier")
 int entry(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 	return 0;
 }
 
tools/testing/selftests/bpf/progs/tailcall_bpf2bpf1.c +2 -2
index b5d9c8e778ae..0103f3dd9f02 100644
@@ -21,7 +21,7 @@ TAIL_FUNC(1)
 static __noinline
 int subprog_tail(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 
 	return skb->len * 2;
 }
@@ -29,7 +29,7 @@ int subprog_tail(struct __sk_buff *skb)
 SEC("classifier")
 int entry(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 1);
+	bpf_tail_call_static(skb, &jmp_table, 1);
 
 	return subprog_tail(skb);
 }
tools/testing/selftests/bpf/progs/tailcall_bpf2bpf2.c +3 -3
index a004ab28ce28..7b1c04183824 100644
@@ -14,9 +14,9 @@ static __noinline
 int subprog_tail(struct __sk_buff *skb)
 {
 	if (load_byte(skb, 0))
-		bpf_tail_call(skb, &jmp_table, 1);
+		bpf_tail_call_static(skb, &jmp_table, 1);
 	else
-		bpf_tail_call(skb, &jmp_table, 0);
+		bpf_tail_call_static(skb, &jmp_table, 0);
 	return 1;
 }
 
@@ -32,7 +32,7 @@ int bpf_func_0(struct __sk_buff *skb)
 SEC("classifier")
 int entry(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 
 	return 0;
 }
tools/testing/selftests/bpf/progs/tailcall_bpf2bpf3.c +3 -3
index 96dbef2b6b7c..0d5482bea6c9 100644
@@ -16,9 +16,9 @@ int subprog_tail2(struct __sk_buff *skb)
 	volatile char arr[64] = {};
 
 	if (load_word(skb, 0) || load_half(skb, 0))
-		bpf_tail_call(skb, &jmp_table, 10);
+		bpf_tail_call_static(skb, &jmp_table, 10);
 	else
-		bpf_tail_call(skb, &jmp_table, 1);
+		bpf_tail_call_static(skb, &jmp_table, 1);
 
 	return skb->len;
 }
@@ -28,7 +28,7 @@ int subprog_tail(struct __sk_buff *skb)
 {
 	volatile char arr[64] = {};
 
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 
 	return skb->len * 2;
 }
tools/testing/selftests/bpf/progs/tailcall_bpf2bpf4.c +3 -3
index 98b40a95bc67..9a1b166b7fbe 100644
@@ -14,21 +14,21 @@ static volatile int count;
 __noinline
 int subprog_tail_2(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 2);
+	bpf_tail_call_static(skb, &jmp_table, 2);
 	return skb->len * 3;
 }
 
 __noinline
 int subprog_tail_1(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 1);
+	bpf_tail_call_static(skb, &jmp_table, 1);
 	return skb->len * 2;
 }
 
 __noinline
 int subprog_tail(struct __sk_buff *skb)
 {
-	bpf_tail_call(skb, &jmp_table, 0);
+	bpf_tail_call_static(skb, &jmp_table, 0);
 	return skb->len;
 }
 
Signature · 1 line
2.21.0
#25 lore raw

Re: [PATCH bpf-next 5/6] bpf, selftests: use bpf_tail_call_static where appropriate

Quoted 4 lines · Daniel Borkmann

For those locations where we use an immediate tail call map index use the newly added bpf_tail_call_static() helper.

Quoted 10 lines
---
 tools/testing/selftests/bpf/progs/bpf_flow.c  | 12 ++++----
 tools/testing/selftests/bpf/progs/tailcall1.c | 28 +++++++++----------
 tools/testing/selftests/bpf/progs/tailcall2.c | 14 +++++-----
 tools/testing/selftests/bpf/progs/tailcall3.c |  4 +--
 .../selftests/bpf/progs/tailcall_bpf2bpf1.c   |  4 +--
 .../selftests/bpf/progs/tailcall_bpf2bpf2.c   |  6 ++--
 .../selftests/bpf/progs/tailcall_bpf2bpf3.c   |  6 ++--
 .../selftests/bpf/progs/tailcall_bpf2bpf4.c   |  6 ++--
 8 files changed, 40 insertions(+), 40 deletions(-)

One nit, while you're at it, maybe it would be good to also address the samples/bpf/sockex3_kern.c that is also using the immediate map index?

[...]

author maintainer
#26 lore raw

Re: [PATCH bpf-next 5/6] bpf, selftests: use bpf_tail_call_static where appropriate

Quoted 16 lines · Maciej Fijalkowski
Quoted 4 lines · Daniel Borkmann

For those locations where we use an immediate tail call map index use the newly added bpf_tail_call_static() helper.

Quoted 10 lines
---
  tools/testing/selftests/bpf/progs/bpf_flow.c  | 12 ++++----
  tools/testing/selftests/bpf/progs/tailcall1.c | 28 +++++++++----------
  tools/testing/selftests/bpf/progs/tailcall2.c | 14 +++++-----
  tools/testing/selftests/bpf/progs/tailcall3.c |  4 +--
  .../selftests/bpf/progs/tailcall_bpf2bpf1.c   |  4 +--
  .../selftests/bpf/progs/tailcall_bpf2bpf2.c   |  6 ++--
  .../selftests/bpf/progs/tailcall_bpf2bpf3.c   |  6 ++--
  .../selftests/bpf/progs/tailcall_bpf2bpf4.c   |  6 ++--
  8 files changed, 40 insertions(+), 40 deletions(-)
Quoted 2 lines

One nit, while you're at it, maybe it would be good to also address the samples/bpf/sockex3_kern.c that is also using the immediate map index?

Sure, np, will add.

author maintainer
#27 lore raw

PATCH 6/6 bpf, selftests: add redirect_neigh selftest

In reply to Daniel Borkmann

Add a small test that excercises the new redirect_neigh() helper for the IPv4 and IPv6 case.

---
 .../selftests/bpf/progs/test_tc_neigh.c       | 144 +++++++++++++++
 tools/testing/selftests/bpf/test_tc_neigh.sh  | 168 ++++++++++++++++++
 2 files changed, 312 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/progs/test_tc_neigh.c
 create mode 100755 tools/testing/selftests/bpf/test_tc_neigh.sh
tools/testing/selftests/bpf/progs/test_tc_neigh.c new file +144 -0
new file mode 100644
index 000000000000..889a72c3024f
@@ -0,0 +1,144 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <stdint.h>
+#include <stdbool.h>
+
+#include <linux/bpf.h>
+#include <linux/stddef.h>
+#include <linux/pkt_cls.h>
+#include <linux/if_ether.h>
+#include <linux/in.h>
+#include <linux/ip.h>
+#include <linux/ipv6.h>
+
+#include <bpf/bpf_helpers.h>
+#include <bpf/bpf_endian.h>
+
+#ifndef barrier_data
+# define barrier_data(ptr)	asm volatile("": :"r"(ptr) :"memory")
+#endif
+
+#ifndef ctx_ptr
+# define ctx_ptr(field)		(void *)(long)(field)
+#endif
+
+#define dst_to_src_tmp		0xeeddddeeU
+#define src_to_dst_tmp		0xeeffffeeU
+
+#define ip4_src			0xac100164 /* 172.16.1.100 */
+#define ip4_dst			0xac100264 /* 172.16.2.100 */
+
+#define ip6_src			{ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, \
+				  0x00, 0x01, 0xde, 0xad, 0xbe, 0xef, 0xca, 0xfe }
+#define ip6_dst			{ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, \
+				  0x00, 0x02, 0xde, 0xad, 0xbe, 0xef, 0xca, 0xfe }
+
+#ifndef v6_equal
+# define v6_equal(a, b)		(a.s6_addr32[0] == b.s6_addr32[0] && \
+				 a.s6_addr32[1] == b.s6_addr32[1] && \
+				 a.s6_addr32[2] == b.s6_addr32[2] && \
+				 a.s6_addr32[3] == b.s6_addr32[3])
+#endif
+
+static __always_inline bool is_remote_ep_v4(struct __sk_buff *skb,
+					    __be32 addr)
+{
+	void *data_end = ctx_ptr(skb->data_end);
+	void *data = ctx_ptr(skb->data);
+	struct iphdr *ip4h;
+
+	if (data + sizeof(struct ethhdr) > data_end)
+		return false;
+
+	ip4h = (struct iphdr *)(data + sizeof(struct ethhdr));
+	if ((void *)(ip4h + 1) > data_end)
+		return false;
+
+	return ip4h->daddr == addr;
+}
+
+static __always_inline bool is_remote_ep_v6(struct __sk_buff *skb,
+					    struct in6_addr addr)
+{
+	void *data_end = ctx_ptr(skb->data_end);
+	void *data = ctx_ptr(skb->data);
+	struct ipv6hdr *ip6h;
+
+	if (data + sizeof(struct ethhdr) > data_end)
+		return false;
+
+	ip6h = (struct ipv6hdr *)(data + sizeof(struct ethhdr));
+	if ((void *)(ip6h + 1) > data_end)
+		return false;
+
+	return v6_equal(ip6h->daddr, addr);
+}
+
+SEC("chk_neigh") int tc_chk(struct __sk_buff *skb)
+{
+	void *data_end = ctx_ptr(skb->data_end);
+	void *data = ctx_ptr(skb->data);
+	__u32 *raw = data;
+
+	if (data + sizeof(struct ethhdr) > data_end)
+		return TC_ACT_SHOT;
+
+	return !raw[0] && !raw[1] && !raw[2] ? TC_ACT_SHOT : TC_ACT_OK;
+}
+
+SEC("dst_ingress") int tc_dst(struct __sk_buff *skb)
+{
+	int idx = dst_to_src_tmp;
+	__u8 zero[ETH_ALEN * 2];
+	bool redirect = false;
+
+	switch (skb->protocol) {
+	case __bpf_constant_htons(ETH_P_IP):
+		redirect = is_remote_ep_v4(skb, __bpf_constant_htonl(ip4_src));
+		break;
+	case __bpf_constant_htons(ETH_P_IPV6):
+		redirect = is_remote_ep_v6(skb, (struct in6_addr)ip6_src);
+		break;
+	}
+
+	if (!redirect)
+		return TC_ACT_OK;
+
+	barrier_data(&idx);
+	idx = bpf_ntohl(idx);
+
+	__builtin_memset(&zero, 0, sizeof(zero));
+	if (bpf_skb_store_bytes(skb, 0, &zero, sizeof(zero), 0) < 0)
+		return TC_ACT_SHOT;
+
+	return bpf_redirect_neigh(idx, 0);
+}
+
+SEC("src_ingress") int tc_src(struct __sk_buff *skb)
+{
+	int idx = src_to_dst_tmp;
+	__u8 zero[ETH_ALEN * 2];
+	bool redirect = false;
+
+	switch (skb->protocol) {
+	case __bpf_constant_htons(ETH_P_IP):
+		redirect = is_remote_ep_v4(skb, __bpf_constant_htonl(ip4_dst));
+		break;
+	case __bpf_constant_htons(ETH_P_IPV6):
+		redirect = is_remote_ep_v6(skb, (struct in6_addr)ip6_dst);
+		break;
+	}
+
+	if (!redirect)
+		return TC_ACT_OK;
+
+	barrier_data(&idx);
+	idx = bpf_ntohl(idx);
+
+	__builtin_memset(&zero, 0, sizeof(zero));
+	if (bpf_skb_store_bytes(skb, 0, &zero, sizeof(zero), 0) < 0)
+		return TC_ACT_SHOT;
+
+	return bpf_redirect_neigh(idx, 0);
+}
+
+char __license[] SEC("license") = "GPL";
tools/testing/selftests/bpf/test_tc_neigh.sh new file +168 -0

168 lines not on this page · raw message

Signature · 1 line
2.21.0