bpf: Introduce BPF dispatcher

The BPF dispatcher is a multi-way branch code generator, mainly
targeted for XDP programs. When an XDP program is executed via the
bpf_prog_run_xdp(), it is invoked via an indirect call. The indirect
call has a substantial performance impact, when retpolines are
enabled. The dispatcher transform indirect calls to direct calls, and
therefore avoids the retpoline. The dispatcher is generated using the
BPF JIT, and relies on text poking provided by bpf_arch_text_poke().

The dispatcher hijacks a trampoline function it via the __fentry__ nop
of the trampoline. One dispatcher instance currently supports up to 64
dispatch points. A user creates a dispatcher with its corresponding
trampoline with the DEFINE_BPF_DISPATCHER macro.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20191213175112.30208-3-bjorn.topel@gmail.com

arch/x86/net/bpf_jit_comp.c

@@ -10,10 +10,12 @@
 #include <linux/if_vlan.h>
 #include <linux/bpf.h>
 #include <linux/memory.h>
+#include <linux/sort.h>
 #include <asm/extable.h>
 #include <asm/set_memory.h>
 #include <asm/nospec-branch.h>
 #include <asm/text-patching.h>
+#include <asm/asm-prototypes.h>
 
 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
 {
@@ -1530,6 +1532,126 @@ int arch_prepare_bpf_trampoline(void *image, struct btf_func_model *m, u32 flags
 	return 0;
 }
 
+static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
+{
+	u8 *prog = *pprog;
+	int cnt = 0;
+	s64 offset;
+
+	offset = func - (ip + 2 + 4);
+	if (!is_simm32(offset)) {
+		pr_err("Target %p is out of range\n", func);
+		return -EINVAL;
+	}
+	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
+	*pprog = prog;
+	return 0;
+}
+
+static int emit_fallback_jump(u8 **pprog)
+{
+	u8 *prog = *pprog;
+	int err = 0;
+
+#ifdef CONFIG_RETPOLINE
+	/* Note that this assumes the the compiler uses external
+	 * thunks for indirect calls. Both clang and GCC use the same
+	 * naming convention for external thunks.
+	 */
+	err = emit_jump(&prog, __x86_indirect_thunk_rdx, prog);
+#else
+	int cnt = 0;
+
+	EMIT2(0xFF, 0xE2);	/* jmp rdx */
+#endif
+	*pprog = prog;
+	return err;
+}
+
+static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
+{
+	int pivot, err, jg_bytes = 1, cnt = 0;
+	u8 *jg_reloc, *prog = *pprog;
+	s64 jg_offset;
+
+	if (a == b) {
+		/* Leaf node of recursion, i.e. not a range of indices
+		 * anymore.
+		 */
+		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
+		if (!is_simm32(progs[a]))
+			return -1;
+		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
+			    progs[a]);
+		err = emit_cond_near_jump(&prog,	/* je func */
+					  (void *)progs[a], prog,
+					  X86_JE);
+		if (err)
+			return err;
+
+		err = emit_fallback_jump(&prog);	/* jmp thunk/indirect */
+		if (err)
+			return err;
+
+		*pprog = prog;
+		return 0;
+	}
+
+	/* Not a leaf node, so we pivot, and recursively descend into
+	 * the lower and upper ranges.
+	 */
+	pivot = (b - a) / 2;
+	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
+	if (!is_simm32(progs[a + pivot]))
+		return -1;
+	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
+
+	if (pivot > 2) {				/* jg upper_part */
+		/* Require near jump. */
+		jg_bytes = 4;
+		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
+	} else {
+		EMIT2(X86_JG, 0);
+	}
+	jg_reloc = prog;
+
+	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
+				  progs);
+	if (err)
+		return err;
+
+	jg_offset = prog - jg_reloc;
+	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
+
+	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
+				  b, progs);
+	if (err)
+		return err;
+
+	*pprog = prog;
+	return 0;
+}
+
+static int cmp_ips(const void *a, const void *b)
+{
+	const s64 *ipa = a;
+	const s64 *ipb = b;
+
+	if (*ipa > *ipb)
+		return 1;
+	if (*ipa < *ipb)
+		return -1;
+	return 0;
+}
+
+int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
+{
+	u8 *prog = image;
+
+	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
+	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
+}
+
 struct x64_jit_data {
 	struct bpf_binary_header *header;
 	int *addrs;

include/linux/bpf.h

@@ -470,12 +470,61 @@ struct bpf_trampoline {
 	void *image;
 	u64 selector;
 };
+
+#define BPF_DISPATCHER_MAX 64 /* Fits in 2048B */
+
+struct bpf_dispatcher_prog {
+	struct bpf_prog *prog;
+	refcount_t users;
+};
+
+struct bpf_dispatcher {
+	/* dispatcher mutex */
+	struct mutex mutex;
+	void *func;
+	struct bpf_dispatcher_prog progs[BPF_DISPATCHER_MAX];
+	int num_progs;
+	void *image;
+	u32 image_off;
+};
+
 #ifdef CONFIG_BPF_JIT
 struct bpf_trampoline *bpf_trampoline_lookup(u64 key);
 int bpf_trampoline_link_prog(struct bpf_prog *prog);
 int bpf_trampoline_unlink_prog(struct bpf_prog *prog);
 void bpf_trampoline_put(struct bpf_trampoline *tr);
 void *bpf_jit_alloc_exec_page(void);
+#define BPF_DISPATCHER_INIT(name) {			\
+	.mutex = __MUTEX_INITIALIZER(name.mutex),	\
+	.func = &name##func,				\
+	.progs = {},					\
+	.num_progs = 0,					\
+	.image = NULL,					\
+	.image_off = 0					\
+}
+
+#define DEFINE_BPF_DISPATCHER(name)					\
+	noinline unsigned int name##func(				\
+		const void *ctx,					\
+		const struct bpf_insn *insnsi,				\
+		unsigned int (*bpf_func)(const void *,			\
+					 const struct bpf_insn *))	\
+	{								\
+		return bpf_func(ctx, insnsi);				\
+	}								\
+	EXPORT_SYMBOL(name##func);			\
+	struct bpf_dispatcher name = BPF_DISPATCHER_INIT(name);
+#define DECLARE_BPF_DISPATCHER(name)					\
+	unsigned int name##func(					\
+		const void *ctx,					\
+		const struct bpf_insn *insnsi,				\
+		unsigned int (*bpf_func)(const void *,			\
+					 const struct bpf_insn *));	\
+	extern struct bpf_dispatcher name;
+#define BPF_DISPATCHER_FUNC(name) name##func
+#define BPF_DISPATCHER_PTR(name) (&name)
+void bpf_dispatcher_change_prog(struct bpf_dispatcher *d, struct bpf_prog *from,
+				struct bpf_prog *to);
 #else
 static inline struct bpf_trampoline *bpf_trampoline_lookup(u64 key)
 {
@@ -490,6 +539,13 @@ static inline int bpf_trampoline_unlink_prog(struct bpf_prog *prog)
 	return -ENOTSUPP;
 }
 static inline void bpf_trampoline_put(struct bpf_trampoline *tr) {}
+#define DEFINE_BPF_DISPATCHER(name)
+#define DECLARE_BPF_DISPATCHER(name)
+#define BPF_DISPATCHER_FUNC(name) bpf_dispatcher_nopfunc
+#define BPF_DISPATCHER_PTR(name) NULL
+static inline void bpf_dispatcher_change_prog(struct bpf_dispatcher *d,
+					      struct bpf_prog *from,
+					      struct bpf_prog *to) {}
 #endif
 
 struct bpf_func_info_aux {

kernel/bpf/Makefile

@@ -8,6 +8,7 @@ obj-$(CONFIG_BPF_SYSCALL) += local_storage.o queue_stack_maps.o
 obj-$(CONFIG_BPF_SYSCALL) += disasm.o
 obj-$(CONFIG_BPF_JIT) += trampoline.o
 obj-$(CONFIG_BPF_SYSCALL) += btf.o
+obj-$(CONFIG_BPF_JIT) += dispatcher.o
 ifeq ($(CONFIG_NET),y)
 obj-$(CONFIG_BPF_SYSCALL) += devmap.o
 obj-$(CONFIG_BPF_SYSCALL) += cpumap.o

kernel/bpf/dispatcher.c

+// SPDX-License-Identifier: GPL-2.0-only
+/* Copyright(c) 2019 Intel Corporation. */
+
+#include <linux/hash.h>
+#include <linux/bpf.h>
+#include <linux/filter.h>
+
+/* The BPF dispatcher is a multiway branch code generator. The
+ * dispatcher is a mechanism to avoid the performance penalty of an
+ * indirect call, which is expensive when retpolines are enabled. A
+ * dispatch client registers a BPF program into the dispatcher, and if
+ * there is available room in the dispatcher a direct call to the BPF
+ * program will be generated. All calls to the BPF programs called via
+ * the dispatcher will then be a direct call, instead of an
+ * indirect. The dispatcher hijacks a trampoline function it via the
+ * __fentry__ of the trampoline. The trampoline function has the
+ * following signature:
+ *
+ * unsigned int trampoline(const void *ctx, const struct bpf_insn *insnsi,
+ *                         unsigned int (*bpf_func)(const void *,
+ *                                                  const struct bpf_insn *));
+ */
+
+static struct bpf_dispatcher_prog *bpf_dispatcher_find_prog(
+	struct bpf_dispatcher *d, struct bpf_prog *prog)
+{
+	int i;
+
+	for (i = 0; i < BPF_DISPATCHER_MAX; i++) {
+		if (prog == d->progs[i].prog)
+			return &d->progs[i];
+	}
+	return NULL;
+}
+
+static struct bpf_dispatcher_prog *bpf_dispatcher_find_free(
+	struct bpf_dispatcher *d)
+{
+	return bpf_dispatcher_find_prog(d, NULL);
+}
+
+static bool bpf_dispatcher_add_prog(struct bpf_dispatcher *d,
+				    struct bpf_prog *prog)
+{
+	struct bpf_dispatcher_prog *entry;
+
+	if (!prog)
+		return false;
+
+	entry = bpf_dispatcher_find_prog(d, prog);
+	if (entry) {
+		refcount_inc(&entry->users);
+		return false;
+	}
+
+	entry = bpf_dispatcher_find_free(d);
+	if (!entry)
+		return false;
+
+	bpf_prog_inc(prog);
+	entry->prog = prog;
+	refcount_set(&entry->users, 1);
+	d->num_progs++;
+	return true;
+}
+
+static bool bpf_dispatcher_remove_prog(struct bpf_dispatcher *d,
+				       struct bpf_prog *prog)
+{
+	struct bpf_dispatcher_prog *entry;
+
+	if (!prog)
+		return false;
+
+	entry = bpf_dispatcher_find_prog(d, prog);
+	if (!entry)
+		return false;
+
+	if (refcount_dec_and_test(&entry->users)) {
+		entry->prog = NULL;
+		bpf_prog_put(prog);
+		d->num_progs--;
+		return true;
+	}
+	return false;
+}
+
+int __weak arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
+{
+	return -ENOTSUPP;
+}
+
+static int bpf_dispatcher_prepare(struct bpf_dispatcher *d, void *image)
+{
+	s64 ips[BPF_DISPATCHER_MAX] = {}, *ipsp = &ips[0];
+	int i;
+
+	for (i = 0; i < BPF_DISPATCHER_MAX; i++) {
+		if (d->progs[i].prog)
+			*ipsp++ = (s64)(uintptr_t)d->progs[i].prog->bpf_func;
+	}
+	return arch_prepare_bpf_dispatcher(image, &ips[0], d->num_progs);
+}
+
+static void bpf_dispatcher_update(struct bpf_dispatcher *d, int prev_num_progs)
+{
+	void *old, *new;
+	u32 noff;
+	int err;
+
+	if (!prev_num_progs) {
+		old = NULL;
+		noff = 0;
+	} else {
+		old = d->image + d->image_off;
+		noff = d->image_off ^ (PAGE_SIZE / 2);
+	}
+
+	new = d->num_progs ? d->image + noff : NULL;
+	if (new) {
+		if (bpf_dispatcher_prepare(d, new))
+			return;
+	}
+
+	err = bpf_arch_text_poke(d->func, BPF_MOD_JUMP, old, new);
+	if (err || !new)
+		return;
+
+	d->image_off = noff;
+}
+
+void bpf_dispatcher_change_prog(struct bpf_dispatcher *d, struct bpf_prog *from,
+				struct bpf_prog *to)
+{
+	bool changed = false;
+	int prev_num_progs;
+
+	if (from == to)
+		return;
+
+	mutex_lock(&d->mutex);
+	if (!d->image) {
+		d->image = bpf_jit_alloc_exec_page();
+		if (!d->image)
+			goto out;
+	}
+
+	prev_num_progs = d->num_progs;
+	changed |= bpf_dispatcher_remove_prog(d, from);
+	changed |= bpf_dispatcher_add_prog(d, to);
+
+	if (!changed)
+		goto out;
+
+	bpf_dispatcher_update(d, prev_num_progs);
+out:
+	mutex_unlock(&d->mutex);
+}