bpf: Inline calls to bpf_loop when callback is known

Calls to `bpf_loop` are replaced with direct loops to avoid
indirection. E.g. the following:

  bpf_loop(10, foo, NULL, 0);

Is replaced by equivalent of the following:

  for (int i = 0; i < 10; ++i)
    foo(i, NULL);

This transformation could be applied when:
- callback is known and does not change during program execution;
- flags passed to `bpf_loop` are always zero.

Inlining logic works as follows:

- During execution simulation function `update_loop_inline_state`
  tracks the following information for each `bpf_loop` call
  instruction:
  - is callback known and constant?
  - are flags constant and zero?
- Function `optimize_bpf_loop` increases stack depth for functions
  where `bpf_loop` calls can be inlined and invokes `inline_bpf_loop`
  to apply the inlining. The additional stack space is used to spill
  registers R6, R7 and R8. These registers are used as loop counter,
  loop maximal bound and callback context parameter;

Measurements using `benchs/run_bench_bpf_loop.sh` inside QEMU / KVM on
i7-4710HQ CPU show a drop in latency from 14 ns/op to 2 ns/op.

Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Acked-by: Song Liu <songliubraving@fb.com>
Link: https://lore.kernel.org/r/20220620235344.569325-4-eddyz87@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

2 files changed, 180 insertions, 9 deletions

diff --git a/kernel/bpf/bpf_iter.c b/kernel/bpf/bpf_iter.c
index d5d96ceca105..7e8fd49406f6 100644
--- a/kernel/bpf/bpf_iter.c
+++ b/kernel/bpf/bpf_iter.c
@@ -723,9 +723,6 @@ const struct bpf_func_proto bpf_for_each_map_elem_proto = {
 	.arg4_type	= ARG_ANYTHING,
 };
 
-/* maximum number of loops */
-#define MAX_LOOPS	BIT(23)
-
 BPF_CALL_4(bpf_loop, u32, nr_loops, void *, callback_fn, void *, callback_ctx,
 	   u64, flags)
 {
@@ -733,9 +730,13 @@ BPF_CALL_4(bpf_loop, u32, nr_loops, void *, callback_fn, void *, callback_ctx,
 	u64 ret;
 	u32 i;
 
+	/* Note: these safety checks are also verified when bpf_loop
+	 * is inlined, be careful to modify this code in sync. See
+	 * function verifier.c:inline_bpf_loop.
+	 */
 	if (flags)
 		return -EINVAL;
-	if (nr_loops > MAX_LOOPS)
+	if (nr_loops > BPF_MAX_LOOPS)
 		return -E2BIG;
 
 	for (i = 0; i < nr_loops; i++) {
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index 2859901ffbe3..bf72dc511df6 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -7124,6 +7124,41 @@ static int check_get_func_ip(struct bpf_verifier_env *env)
 	return -ENOTSUPP;
 }
 
+static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
+{
+	return &env->insn_aux_data[env->insn_idx];
+}
+
+static bool loop_flag_is_zero(struct bpf_verifier_env *env)
+{
+	struct bpf_reg_state *regs = cur_regs(env);
+	struct bpf_reg_state *reg = &regs[BPF_REG_4];
+	bool reg_is_null = register_is_null(reg);
+
+	if (reg_is_null)
+		mark_chain_precision(env, BPF_REG_4);
+
+	return reg_is_null;
+}
+
+static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
+{
+	struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
+
+	if (!state->initialized) {
+		state->initialized = 1;
+		state->fit_for_inline = loop_flag_is_zero(env);
+		state->callback_subprogno = subprogno;
+		return;
+	}
+
+	if (!state->fit_for_inline)
+		return;
+
+	state->fit_for_inline = (loop_flag_is_zero(env) &&
+				 state->callback_subprogno == subprogno);
+}
+
 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
 			     int *insn_idx_p)
 {
@@ -7276,6 +7311,7 @@ static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn
 		err = check_bpf_snprintf_call(env, regs);
 		break;
 	case BPF_FUNC_loop:
+		update_loop_inline_state(env, meta.subprogno);
 		err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
 					set_loop_callback_state);
 		break;
@@ -7682,11 +7718,6 @@ static bool check_reg_sane_offset(struct bpf_verifier_env *env,
 	return true;
 }
 
-static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
-{
-	return &env->insn_aux_data[env->insn_idx];
-}
-
 enum {
 	REASON_BOUNDS	= -1,
 	REASON_TYPE	= -2,
@@ -14315,6 +14346,142 @@ patch_call_imm:
 	return 0;
 }
 
+static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
+					int position,
+					s32 stack_base,
+					u32 callback_subprogno,
+					u32 *cnt)
+{
+	s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
+	s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
+	s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
+	int reg_loop_max = BPF_REG_6;
+	int reg_loop_cnt = BPF_REG_7;
+	int reg_loop_ctx = BPF_REG_8;
+
+	struct bpf_prog *new_prog;
+	u32 callback_start;
+	u32 call_insn_offset;
+	s32 callback_offset;
+
+	/* This represents an inlined version of bpf_iter.c:bpf_loop,
+	 * be careful to modify this code in sync.
+	 */
+	struct bpf_insn insn_buf[] = {
+		/* Return error and jump to the end of the patch if
+		 * expected number of iterations is too big.
+		 */
+		BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2),
+		BPF_MOV32_IMM(BPF_REG_0, -E2BIG),
+		BPF_JMP_IMM(BPF_JA, 0, 0, 16),
+		/* spill R6, R7, R8 to use these as loop vars */
+		BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset),
+		BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset),
+		BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset),
+		/* initialize loop vars */
+		BPF_MOV64_REG(reg_loop_max, BPF_REG_1),
+		BPF_MOV32_IMM(reg_loop_cnt, 0),
+		BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3),
+		/* loop header,
+		 * if reg_loop_cnt >= reg_loop_max skip the loop body
+		 */
+		BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5),
+		/* callback call,
+		 * correct callback offset would be set after patching
+		 */
+		BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt),
+		BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx),
+		BPF_CALL_REL(0),
+		/* increment loop counter */
+		BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1),
+		/* jump to loop header if callback returned 0 */
+		BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6),
+		/* return value of bpf_loop,
+		 * set R0 to the number of iterations
+		 */
+		BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt),
+		/* restore original values of R6, R7, R8 */
+		BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset),
+		BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset),
+		BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset),
+	};
+
+	*cnt = ARRAY_SIZE(insn_buf);
+	new_prog = bpf_patch_insn_data(env, position, insn_buf, *cnt);
+	if (!new_prog)
+		return new_prog;
+
+	/* callback start is known only after patching */
+	callback_start = env->subprog_info[callback_subprogno].start;
+	/* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
+	call_insn_offset = position + 12;
+	callback_offset = callback_start - call_insn_offset - 1;
+	env->prog->insnsi[call_insn_offset].imm = callback_offset;
+
+	return new_prog;
+}
+
+static bool is_bpf_loop_call(struct bpf_insn *insn)
+{
+	return insn->code == (BPF_JMP | BPF_CALL) &&
+		insn->src_reg == 0 &&
+		insn->imm == BPF_FUNC_loop;
+}
+
+/* For all sub-programs in the program (including main) check
+ * insn_aux_data to see if there are bpf_loop calls that require
+ * inlining. If such calls are found the calls are replaced with a
+ * sequence of instructions produced by `inline_bpf_loop` function and
+ * subprog stack_depth is increased by the size of 3 registers.
+ * This stack space is used to spill values of the R6, R7, R8.  These
+ * registers are used to store the loop bound, counter and context
+ * variables.
+ */
+static int optimize_bpf_loop(struct bpf_verifier_env *env)
+{
+	struct bpf_subprog_info *subprogs = env->subprog_info;
+	int i, cur_subprog = 0, cnt, delta = 0;
+	struct bpf_insn *insn = env->prog->insnsi;
+	int insn_cnt = env->prog->len;
+	u16 stack_depth = subprogs[cur_subprog].stack_depth;
+	u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
+	u16 stack_depth_extra = 0;
+
+	for (i = 0; i < insn_cnt; i++, insn++) {
+		struct bpf_loop_inline_state *inline_state =
+			&env->insn_aux_data[i + delta].loop_inline_state;
+
+		if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
+			struct bpf_prog *new_prog;
+
+			stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
+			new_prog = inline_bpf_loop(env,
+						   i + delta,
+						   -(stack_depth + stack_depth_extra),
+						   inline_state->callback_subprogno,
+						   &cnt);
+			if (!new_prog)
+				return -ENOMEM;
+
+			delta     += cnt - 1;
+			env->prog  = new_prog;
+			insn       = new_prog->insnsi + i + delta;
+		}
+
+		if (subprogs[cur_subprog + 1].start == i + delta + 1) {
+			subprogs[cur_subprog].stack_depth += stack_depth_extra;
+			cur_subprog++;
+			stack_depth = subprogs[cur_subprog].stack_depth;
+			stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
+			stack_depth_extra = 0;
+		}
+	}
+
+	env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
+
+	return 0;
+}
+
 static void free_states(struct bpf_verifier_env *env)
 {
 	struct bpf_verifier_state_list *sl, *sln;
@@ -15052,6 +15219,9 @@ skip_full_check:
 		ret = check_max_stack_depth(env);
 
 	/* instruction rewrites happen after this point */
+	if (ret == 0)
+		ret = optimize_bpf_loop(env);
+
 	if (is_priv) {
 		if (ret == 0)
 			opt_hard_wire_dead_code_branches(env);