Skip to contentSkip to conversation

bpf, x64: Introduce two tailcall enhancements

PATCH In review Leon Hwang 3 participants 13 messages Oct 21, 2024 – Oct 25, 2024

Subsystems (4)
lore
Share & tools
lore

Leon Hwang

author
#0 lore raw

cover bpf, x64: Introduce two tailcall enhancements

This patch set introduces two enhancements aimed at improving tailcall handling in the x64 JIT:

1. Tailcall info is propagated to a subprog only if the subprog is tail_call_reachable.
2. Tailcall info is propagated through the trampoline only when the target is a subprog and it is tail_call_reachable.

Leon Hwang (2):
bpf, x64: Propagate tailcall info only for tail_call_reachable
subprogs
bpf, verifier: Check trampoline attach target is tail_call_reachable
subprog

 arch/x86/net/bpf_jit_comp.c |  4 +++-
 include/linux/bpf.h         |  1 +
 kernel/bpf/verifier.c       | 10 +++++++++-
 3 files changed, 13 insertions(+), 2 deletions(-)
Signature · 1 line
2.44.0
author
#1 lore raw

PATCH 1/2 bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

In the x86_64 JIT, when calling a function, tailcall info is propagated if the program is tail_call_reachable, regardless of whether the function is a subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore, it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
 arch/x86/net/bpf_jit_comp.c | 4 +++-
 kernel/bpf/verifier.c       | 6 ++++++
 2 files changed, 9 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:			if (is_imm8(insn->off))
 
 			/* call */
 		case BPF_JMP | BPF_CALL: {
+			bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+			bool subprog_tail_call_reachable = dst_reg;
 			u8 *ip = image + addrs[i - 1];
 
 			func = (u8 *) __bpf_call_base + imm32;
-			if (tail_call_reachable) {
+			if (pseudo_call && subprog_tail_call_reachable) {
 				LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
 				ip += 7;
 			}
kernel/bpf/verifier.c +6 -0
index f514247ba8ba8..6e7e42c7bc7b1 100644
@@ -19990,6 +19990,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
 			insn[0].imm = (u32)addr;
 			insn[1].imm = addr >> 32;
 		}
+
+		if (bpf_pseudo_call(insn))
+			/* In the x86_64 JIT, tailcall information can only be
+			 * propagated if the subprog is tail_call_reachable.
+			 */
+			insn->dst_reg = env->subprog_info[subprog].tail_call_reachable;
 	}
 
 	err = bpf_prog_alloc_jited_linfo(prog);
Signature · 1 line
2.44.0
author
#11 lore raw

PATCH 2/2 bpf, verifier: Check trampoline target is tail_call_reachable subprog

In reply to Leon Hwang

In the x86_64 JIT, tailcall info is propagated through the trampoline when the target program is tail_call_reachable. However, this propagation is unnecessary if the target is a main prog, or a subprog that is not tail_call_reachable.

Since the verifier can determine if a subprog is tail_call_reachable, it should only propagate tailcall info when the target is subprog and the subprog is actually tail_call_reachable.

---
 include/linux/bpf.h   | 1 +
 kernel/bpf/verifier.c | 4 +++-
 2 files changed, 4 insertions(+), 1 deletion(-)
include/linux/bpf.h +1 -0
index 0c216e71cec76..32b56a759d788 100644
@@ -1220,6 +1220,7 @@ struct bpf_attach_target_info {
 	struct module *tgt_mod;
 	const char *tgt_name;
 	const struct btf_type *tgt_type;
+	bool tgt_tail_call_reachable;
 };
 
 #define BPF_DISPATCHER_MAX 48 /* Fits in 2048B */
kernel/bpf/verifier.c +3 -1
index 6e7e42c7bc7b1..9a09621a8832c 100644
@@ -21916,6 +21916,8 @@ int bpf_check_attach_target(struct bpf_verifier_log *log,
 			bpf_log(log, "Subprog %s doesn't exist\n", tname);
 			return -EINVAL;
 		}
+		tgt_info->tgt_tail_call_reachable = subprog &&
+						    aux->func[subprog]->aux->tail_call_reachable;
 		if (aux->func && aux->func[subprog]->aux->exception_cb) {
 			bpf_log(log,
 				"%s programs cannot attach to exception callback\n",
@@ -22285,7 +22287,7 @@ static int check_attach_btf_id(struct bpf_verifier_env *env)
 	if (!tr)
 		return -ENOMEM;
 
-	if (tgt_prog && tgt_prog->aux->tail_call_reachable)
+	if (tgt_prog && tgt_info.tgt_tail_call_reachable)
 		tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
 
 	prog->aux->dst_trampoline = tr;
Signature · 1 line
2.44.0
reviewer
#2 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

In reply to Leon Hwang
Quoted 20 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if the program is tail_call_reachable, regardless of whether the function is a subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore, it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
  arch/x86/net/bpf_jit_comp.c | 4 +++-
  kernel/bpf/verifier.c       | 6 ++++++
  2 files changed, 9 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +0 -0
index 06b080b61aa57..6ad6886ecfc88 100644

arch/x86/net/bpf_jit_comp.c st: if (is_imm8(insn->off)) · line 2132

Quoted 10 lines
arch/x86/net/bpf_jit_comp.c +3 -1
@@ -2124,10 +2124,12 @@ st:			if (is_imm8(insn->off))
 
  			/* call */
  		case BPF_JMP | BPF_CALL: {
+			bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+			bool subprog_tail_call_reachable = dst_reg;
  			u8 *ip = image + addrs[i - 1];
 
  			func = (u8 *) __bpf_call_base + imm32;
-			if (tail_call_reachable) {
+			if (pseudo_call && subprog_tail_call_reachable) {

Why we need subprog_tail_call_reachable? Does

	tail_call_reachable && psueudo_call

work the same way?

Quoted 20 lines
arch/x86/net/bpf_jit_comp.c +0 -0
  				LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
  				ip += 7;
  			}
kernel/bpf/verifier.c +6 -0
index f514247ba8ba8..6e7e42c7bc7b1 100644
@@ -19990,6 +19990,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
  			insn[0].imm = (u32)addr;
  			insn[1].imm = addr >> 32;
  		}
+
+		if (bpf_pseudo_call(insn))
+			/* In the x86_64 JIT, tailcall information can only be
+			 * propagated if the subprog is tail_call_reachable.
+			 */
+			insn->dst_reg = env->subprog_info[subprog].tail_call_reachable;
  	}
 
  	err = bpf_prog_alloc_jited_linfo(prog);
author
#3 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

Quoted 29 lines · Yonghong Song
Quoted 27 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if
the program is tail_call_reachable, regardless of whether the function is a
subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore,
it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
  arch/x86/net/bpf_jit_comp.c | 4 +++-
  kernel/bpf/verifier.c       | 6 ++++++

  2 files changed, 9 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +2 -0
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:            if (is_imm8(insn->off))
                 /* call */
           case BPF_JMP | BPF_CALL: {
+            bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+            bool subprog_tail_call_reachable = dst_reg;
Quoted 9 lines

arch/x86/net/bpf_jit_comp.c st: if (is_imm8(insn->off)) · line 2132

Quoted 4 lines
arch/x86/net/bpf_jit_comp.c +1 -1
               u8 *ip = image + addrs[i - 1];
                 func = (u8 *) __bpf_call_base + imm32;
-            if (tail_call_reachable) {
+            if (pseudo_call && subprog_tail_call_reachable) {

Why we need subprog_tail_call_reachable? Does
    tail_call_reachable && psueudo_call
work the same way?

'tail_call_reachable && pseudo_call' works too. However, it will propagate tailcall info to subprog even if the subprog is not tail_call_reachable.

subprog_tail_call_reachable indicates the subprog requires tailcall info from its caller.
So, 'pseudo_call && subprog_tail_call_reachable' is better.

Thanks,
Leon

reviewer
#4 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

Quoted 38 lines · Leon Hwang
Quoted 19 lines · Yonghong Song
Quoted 18 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if
the program is tail_call_reachable, regardless of whether the function is a
subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore,
it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
   arch/x86/net/bpf_jit_comp.c | 4 +++-
   kernel/bpf/verifier.c       | 6 ++++++
   2 files changed, 9 insertions(+), 1 deletion(-)
Quoted 17 lines

arch/x86/net/bpf_jit_comp.c st:            if (is_imm8(insn->off)) · line 2130

Quoted 13 lines
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:            if (is_imm8(insn->off))
                 /* call */
           case BPF_JMP | BPF_CALL: {
+            bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+            bool subprog_tail_call_reachable = dst_reg;
               u8 *ip = image + addrs[i - 1];
                 func = (u8 *) __bpf_call_base + imm32;
-            if (tail_call_reachable) {
+            if (pseudo_call && subprog_tail_call_reachable) {

Why we need subprog_tail_call_reachable? Does
    tail_call_reachable && psueudo_call
work the same way?

Quoted 7 lines

'tail_call_reachable && pseudo_call' works too. However, it will propagate tailcall info to subprog even if the subprog is not tail_call_reachable.

subprog_tail_call_reachable indicates the subprog requires tailcall info from its caller.
So, 'pseudo_call && subprog_tail_call_reachable' is better.

In verifier.c, we have
func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
that is subprog_info tail_call_reachable has been transferred to func[i] tail_call_reachable.

In x86 do_jit() func, we have
bool tail_call_reachable = bpf_prog->aux->tail_call_reachable

So looks like we do not need verifier.c change here.
Did I miss anything? Could you give a concrete example to show subprog_tail_call_reachable approach is better than tail_call_reachable?

Quoted 4 lines

Thanks,
Leon

reviewer Acked-by on v1 2/2
#12 lore raw

Re: [PATCH bpf-next 2/2] bpf, verifier: Check trampoline target is tail_call_reachable subprog

In reply to Leon Hwang
Quoted 10 lines · Leon Hwang

In the x86_64 JIT, tailcall info is propagated through the trampoline when the target program is tail_call_reachable. However, this propagation is unnecessary if the target is a main prog, or a subprog that is not tail_call_reachable.

Since the verifier can determine if a subprog is tail_call_reachable, it should only propagate tailcall info when the target is subprog and the subprog is actually tail_call_reachable.

Looks correct to me.

author
#5 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

In reply to Yonghong Song
Quoted 54 lines · Yonghong Song
Quoted 38 lines · Leon Hwang
Quoted 19 lines · Yonghong Song
Quoted 18 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if
the program is tail_call_reachable, regardless of whether the function is a
subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore,
it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
   arch/x86/net/bpf_jit_comp.c | 4 +++-
   kernel/bpf/verifier.c       | 6 ++++++

   2 files changed, 9 insertions(+), 1 deletion(-)
Quoted 17 lines

arch/x86/net/bpf_jit_comp.c st:            if (is_imm8(insn->off)) · line 2130

Quoted 13 lines
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:            if (is_imm8(insn->off))
                  /* call */
            case BPF_JMP | BPF_CALL: {
+            bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+            bool subprog_tail_call_reachable = dst_reg;
                u8 *ip = image + addrs[i - 1];
                  func = (u8 *) __bpf_call_base + imm32;
-            if (tail_call_reachable) {
+            if (pseudo_call && subprog_tail_call_reachable) {

Why we need subprog_tail_call_reachable? Does
     tail_call_reachable && psueudo_call
work the same way?

Quoted 7 lines

'tail_call_reachable && pseudo_call' works too. However, it will propagate tailcall info to subprog even if the subprog is not tail_call_reachable.

subprog_tail_call_reachable indicates the subprog requires tailcall info from its caller.
So, 'pseudo_call && subprog_tail_call_reachable' is better.

In verifier.c, we have
  func[i]->aux->tail_call_reachable = env-

Quoted 1 line

subprog_info[i].tail_call_reachable;

that is subprog_info tail_call_reachable has been transferred to func[i] tail_call_reachable.

Quoted 7 lines

In x86 do_jit() func, we have
  bool tail_call_reachable = bpf_prog->aux->tail_call_reachable

So looks like we do not need verifier.c change here.
Did I miss anything? Could you give a concrete example to show subprog_tail_call_reachable approach is better than tail_call_reachable?

Sure, here's an example:

struct {
	__uint(type, BPF_MAP_TYPE_PROG_ARRAY);
	__uint(key_size, sizeof(u32));
	__uint(value_size, sizeof(u32));
	__uint(max_entries, 1);
} jmp_table SEC(".maps");

static __noinline int
subprog_tc1(struct __sk_buff *skb)

{
	volatile int retval = TC_ACT_OK;

	bpf_tail_call_static(skb, jmp_table, 0);
	return retval;
}

static __noinline int
subprog_tc2(struct __sk_buff *skb)

{
	volatile int retval = TC_ACT_OK;

	return retval;
}

SEC("tc")
int entry_tc(struct __sk_buff *skb)

{
	u32 pid = bpf_get_smp_processor_id();
	// do something with pid
	subprog_tc2(skb);
	return subprog_tc1(skb);
}

From the verifier's perspective, both entry_tc and subprog_tc1 are tail_call_reachable.

When handling 'BPF_JMP | BPF_CALL' in the x86 do_jit() for entry_tc, three cases arise:

1. bpf_get_smp_processor_id()
2. subprog_tc1()
3. subprog_tc2()

At this point in x86 do_jit() for entry_tc, entry_tc is considered tail_call_reachable. The check 'bool pseudo_call = src_reg == BPF_PSEUDO_CALL' is used to determine whether to call a subprogram.

The question is: when should tailcall info be propagated? Should it be when entry_tc is tail_call_reachable, even if subprog_tc2 is called, or when subprog_tc1 is specifically tail_call_reachable?

I believe it is better to propagate the tailcall info when subprog_tc1 is tail_call_reachable.

Thanks,
Leon

reviewer
#6 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

Quoted 109 lines · Leon Hwang
Quoted 51 lines · Yonghong Song
Quoted 37 lines · Leon Hwang
Quoted 19 lines · Yonghong Song
Quoted 18 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if
the program is tail_call_reachable, regardless of whether the function is a
subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore,
it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
    arch/x86/net/bpf_jit_comp.c | 4 +++-
    kernel/bpf/verifier.c       | 6 ++++++
    2 files changed, 9 insertions(+), 1 deletion(-)
Quoted 17 lines

arch/x86/net/bpf_jit_comp.c st:            if (is_imm8(insn->off)) · line 2130

Quoted 13 lines
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:            if (is_imm8(insn->off))
                  /* call */
            case BPF_JMP | BPF_CALL: {
+            bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+            bool subprog_tail_call_reachable = dst_reg;
                u8 *ip = image + addrs[i - 1];
                  func = (u8 *) __bpf_call_base + imm32;
-            if (tail_call_reachable) {
+            if (pseudo_call && subprog_tail_call_reachable) {

Why we need subprog_tail_call_reachable? Does
     tail_call_reachable && psueudo_call
work the same way?

Quoted 7 lines

'tail_call_reachable && pseudo_call' works too. However, it will propagate tailcall info to subprog even if the subprog is not tail_call_reachable.

subprog_tail_call_reachable indicates the subprog requires tailcall info from its caller.
So, 'pseudo_call && subprog_tail_call_reachable' is better.

In verifier.c, we have
  func[i]->aux->tail_call_reachable = env-

Quoted 1 line

subprog_info[i].tail_call_reachable;

that is subprog_info tail_call_reachable has been transferred to func[i] tail_call_reachable.

Quoted 7 lines

In x86 do_jit() func, we have
  bool tail_call_reachable = bpf_prog->aux->tail_call_reachable

So looks like we do not need verifier.c change here.
Did I miss anything? Could you give a concrete example to show subprog_tail_call_reachable approach is better than tail_call_reachable?

Sure, here's an example:

struct {
	__uint(type, BPF_MAP_TYPE_PROG_ARRAY);
	__uint(key_size, sizeof(u32));
	__uint(value_size, sizeof(u32));
	__uint(max_entries, 1);
} jmp_table SEC(".maps");

static __noinline int
subprog_tc1(struct __sk_buff *skb)

{
	volatile int retval = TC_ACT_OK;

	bpf_tail_call_static(skb, jmp_table, 0);
	return retval;
}

static __noinline int
subprog_tc2(struct __sk_buff *skb)

{
	volatile int retval = TC_ACT_OK;

	return retval;
}

SEC("tc")
int entry_tc(struct __sk_buff *skb)

{
	u32 pid = bpf_get_smp_processor_id();
	// do something with pid
	subprog_tc2(skb);
	return subprog_tc1(skb);
}

From the verifier's perspective, both entry_tc and subprog_tc1 are
tail_call_reachable.

When handling 'BPF_JMP | BPF_CALL' in the x86 do_jit() for entry_tc, three cases arise:

1. bpf_get_smp_processor_id()
2. subprog_tc1()
3. subprog_tc2()

At this point in x86 do_jit() for entry_tc, entry_tc is considered tail_call_reachable. The check 'bool pseudo_call = src_reg == BPF_PSEUDO_CALL' is used to determine whether to call a subprogram.

Quoted 6 lines

The question is: when should tailcall info be propagated? Should it be when entry_tc is tail_call_reachable, even if subprog_tc2 is called, or when subprog_tc1 is specifically tail_call_reachable?

I believe it is better to propagate the tailcall info when subprog_tc1 is tail_call_reachable.

Okay, I see. Thanks for explanation.

You use the insn->dst_reg to record whether callee is tail call reachable or not. I think you can reuse insn->off which currently represents subprog number but it is not used for jit. We can use that to indicate callee is tail call reachable or not.

Something like below:

arch/x86/net/bpf_jit_comp.c +2 -1
index 06b080b61aa5..b3c76bf59e65 100644
@@ -2127,7 +2127,8 @@ st:                       if (is_imm8(insn->off))
                        u8 *ip = image + addrs[i - 1];
 
                        func = (u8 *) __bpf_call_base + imm32;
-                       if (tail_call_reachable) {
+                       /* insn->off == 1 means the callee is tail call reachable */
+                       if (src_reg == BPF_PSEUDO_CALL && insn->off == 1) {
                                LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
                                ip += 7;
                        }
kernel/bpf/verifier.c +2 -0
index f514247ba8ba..2ccadc1ac22e 100644
@@ -20096,6 +20096,8 @@ static int jit_subprogs(struct bpf_verifier_env *env)
                                continue;
                        subprog = insn->off;
                        insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
+                       /* Indicate whether callee is tail call reachable or not */
+                       insn->off = func[subprog]->aux->tail_call_reachable;
                }
 

WDYT?

Quoted 4 lines

Thanks,
Leon

reviewer
#7 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

Quoted 157 lines · Yonghong Song
Quoted 108 lines · Leon Hwang
Quoted 51 lines · Yonghong Song
Quoted 37 lines · Leon Hwang
Quoted 19 lines · Yonghong Song
Quoted 18 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if
the program is tail_call_reachable, regardless of whether the function is a
subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore,
it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
    arch/x86/net/bpf_jit_comp.c | 4 +++-
    kernel/bpf/verifier.c       | 6 ++++++

    2 files changed, 9 insertions(+), 1 deletion(-)
Quoted 17 lines

arch/x86/net/bpf_jit_comp.c st:            if (is_imm8(insn->off)) · line 2130

Quoted 13 lines
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:            if (is_imm8(insn->off))
                   /* call */
             case BPF_JMP | BPF_CALL: {
+            bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+            bool subprog_tail_call_reachable = dst_reg;
                 u8 *ip = image + addrs[i - 1];
                   func = (u8 *) __bpf_call_base + imm32;
-            if (tail_call_reachable) {
+            if (pseudo_call && subprog_tail_call_reachable) {

Why we need subprog_tail_call_reachable? Does
      tail_call_reachable && psueudo_call
work the same way?

Quoted 7 lines

'tail_call_reachable && pseudo_call' works too. However, it will propagate tailcall info to subprog even if the subprog is not tail_call_reachable.

subprog_tail_call_reachable indicates the subprog requires tailcall info from its caller.
So, 'pseudo_call && subprog_tail_call_reachable' is better.

In verifier.c, we have
   func[i]->aux->tail_call_reachable = env-

Quoted 1 line

subprog_info[i].tail_call_reachable;

that is subprog_info tail_call_reachable has been transferred to func[i] tail_call_reachable.

Quoted 6 lines

In x86 do_jit() func, we have
   bool tail_call_reachable = bpf_prog->aux->tail_call_reachable

So looks like we do not need verifier.c change here.
Did I miss anything? Could you give a concrete example to show subprog_tail_call_reachable approach is better than tail_call_reachable?

Sure, here's an example:

struct {
    __uint(type, BPF_MAP_TYPE_PROG_ARRAY);
    __uint(key_size, sizeof(u32));
    __uint(value_size, sizeof(u32));
    __uint(max_entries, 1);
} jmp_table SEC(".maps");

static __noinline int
subprog_tc1(struct __sk_buff *skb)
{
    volatile int retval = TC_ACT_OK;

    bpf_tail_call_static(skb, jmp_table, 0);
    return retval;
}

static __noinline int
subprog_tc2(struct __sk_buff *skb)
{
    volatile int retval = TC_ACT_OK;

    return retval;
}

SEC("tc")
int entry_tc(struct __sk_buff *skb)
{
    u32 pid = bpf_get_smp_processor_id();
    // do something with pid
    subprog_tc2(skb);
    return subprog_tc1(skb);
}

 From the verifier's perspective, both entry_tc and subprog_tc1 are tail_call_reachable.

When handling 'BPF_JMP | BPF_CALL' in the x86 do_jit() for entry_tc, three cases arise:

1. bpf_get_smp_processor_id()
2. subprog_tc1()
3. subprog_tc2()

At this point in x86 do_jit() for entry_tc, entry_tc is considered tail_call_reachable. The check 'bool pseudo_call = src_reg == BPF_PSEUDO_CALL' is used to determine whether to call a subprogram.

Quoted 6 lines

The question is: when should tailcall info be propagated? Should it be when entry_tc is tail_call_reachable, even if subprog_tc2 is called, or when subprog_tc1 is specifically tail_call_reachable?

I believe it is better to propagate the tailcall info when subprog_tc1 is tail_call_reachable.

Okay, I see. Thanks for explanation.

You use the insn->dst_reg to record whether callee is tail call reachable or not. I think you can reuse insn->off which currently represents subprog number but it is not used for jit. We can use that to indicate callee is tail call reachable or not.

Something like below:

arch/x86/net/bpf_jit_comp.c +1 -1
index 06b080b61aa5..b3c76bf59e65 100644
@@ -2127,7 +2127,8 @@ st:                       if (is_imm8(insn->off))
                         u8 *ip = image + addrs[i - 1];
 
                         func = (u8 *) __bpf_call_base + imm32;
-                       if (tail_call_reachable) {
+                       /* insn->off == 1 means the callee is tail 

call reachable */
+                       if (src_reg == BPF_PSEUDO_CALL && insn->off ==
1) {
LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
                                ip += 7;
                        }

kernel/bpf/verifier.c +1 -0
index f514247ba8ba..2ccadc1ac22e 100644
@@ -20096,6 +20096,8 @@ static int jit_subprogs(struct 
 bpf_verifier_env *env)
                                 continue;
                         subprog = insn->off;
                         insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
+                       /* Indicate whether callee is tail call 

reachable or not */
+                       insn->off =
func[subprog]->aux->tail_call_reachable;
                }

Quoted 1 line

WDYT?

Sorry, the above seems not working since verifier do jit twice for the same prog and two jit'ed results need to be the same. The above change could make jit result different between two passes.

Quoted 6 lines
Quoted 5 lines

Thanks,
Leon

reviewer Acked-by on v1 1/2
#8 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

In reply to Leon Hwang
Quoted 10 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if the program is tail_call_reachable, regardless of whether the function is a subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore, it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

LGTM with a nit below.

Quoted 28 lines
---
  arch/x86/net/bpf_jit_comp.c | 4 +++-
  kernel/bpf/verifier.c       | 6 ++++++
  2 files changed, 9 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:			if (is_imm8(insn->off))
 
  			/* call */
  		case BPF_JMP | BPF_CALL: {
+			bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+			bool subprog_tail_call_reachable = dst_reg;
  			u8 *ip = image + addrs[i - 1];
 
  			func = (u8 *) __bpf_call_base + imm32;
-			if (tail_call_reachable) {
+			if (pseudo_call && subprog_tail_call_reachable) {
  				LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
  				ip += 7;
  			}
kernel/bpf/verifier.c +0 -0
index f514247ba8ba8..6e7e42c7bc7b1 100644

kernel/bpf/verifier.c static int jit_subprogs(struct bpf_verifier_env *env) · line 19998

Quoted 9 lines
kernel/bpf/verifier.c +6 -0
@@ -19990,6 +19990,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
  			insn[0].imm = (u32)addr;
  			insn[1].imm = addr >> 32;
  		}
+
+		if (bpf_pseudo_call(insn))
+			/* In the x86_64 JIT, tailcall information can only be
+			 * propagated if the subprog is tail_call_reachable.
+			 */
+			insn->dst_reg = env->subprog_info[subprog].tail_call_reachable;

The comment can be simplied as

	/* Indicate whether callee is tail call

reachable or not */

Quoted 3 lines
kernel/bpf/verifier.c +0 -0
  	}
 
  	err = bpf_prog_alloc_jited_linfo(prog);
#9 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

In reply to Leon Hwang
Quoted 39 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if the program is tail_call_reachable, regardless of whether the function is a subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore, it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
 arch/x86/net/bpf_jit_comp.c | 4 +++-
 kernel/bpf/verifier.c       | 6 ++++++
 2 files changed, 9 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:                     if (is_imm8(insn->off))
 
                        /* call */
                case BPF_JMP | BPF_CALL: {
+                       bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+                       bool subprog_tail_call_reachable = dst_reg;
                        u8 *ip = image + addrs[i - 1];
 
                        func = (u8 *) __bpf_call_base + imm32;
-                       if (tail_call_reachable) {
+                       if (pseudo_call && subprog_tail_call_reachable) {
                                LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
                                ip += 7;
                        }
kernel/bpf/verifier.c +0 -0
index f514247ba8ba8..6e7e42c7bc7b1 100644

kernel/bpf/verifier.c static int jit_subprogs(struct bpf_verifier_env *env) · line 19998

Quoted 9 lines
kernel/bpf/verifier.c +6 -0
@@ -19990,6 +19990,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
                        insn[0].imm = (u32)addr;
                        insn[1].imm = addr >> 32;
                }
+
+               if (bpf_pseudo_call(insn))
+                       /* In the x86_64 JIT, tailcall information can only be
+                        * propagated if the subprog is tail_call_reachable.
+                        */
+                       insn->dst_reg = env->subprog_info[subprog].tail_call_reachable;

I really don't like hacking flags into dst_reg.
We already abuse insn->off which is ugly too,
but at least we clean insns later after JIT.

I'd rather live with this tail call inefficiency than abuse insns fields further.

pw-bot: cr

author
#10 lore raw

Re: [PATCH bpf-next 1/2] bpf, x64: Propagate tailcall info only for tail_call_reachable subprogs

Quoted 54 lines · Alexei Starovoitov
Quoted 39 lines · Leon Hwang

In the x86_64 JIT, when calling a function, tailcall info is propagated if the program is tail_call_reachable, regardless of whether the function is a subprog, helper, or kfunc. However, this propagation is unnecessary for not-tail_call_reachable subprogs, helpers, or kfuncs.

The verifier can determine if a subprog is tail_call_reachable. Therefore, it can be optimized to only propagate tailcall info when the callee is subprog and the subprog is actually tail_call_reachable.

---
 arch/x86/net/bpf_jit_comp.c | 4 +++-
 kernel/bpf/verifier.c       | 6 ++++++
 2 files changed, 9 insertions(+), 1 deletion(-)
arch/x86/net/bpf_jit_comp.c +3 -1
index 06b080b61aa57..6ad6886ecfc88 100644
@@ -2124,10 +2124,12 @@ st:                     if (is_imm8(insn->off))
 
                        /* call */
                case BPF_JMP | BPF_CALL: {
+                       bool pseudo_call = src_reg == BPF_PSEUDO_CALL;
+                       bool subprog_tail_call_reachable = dst_reg;
                        u8 *ip = image + addrs[i - 1];
 
                        func = (u8 *) __bpf_call_base + imm32;
-                       if (tail_call_reachable) {
+                       if (pseudo_call && subprog_tail_call_reachable) {
                                LOAD_TAIL_CALL_CNT_PTR(bpf_prog->aux->stack_depth);
                                ip += 7;
                        }
kernel/bpf/verifier.c +0 -0
index f514247ba8ba8..6e7e42c7bc7b1 100644

kernel/bpf/verifier.c static int jit_subprogs(struct bpf_verifier_env *env) · line 19998

Quoted 9 lines
kernel/bpf/verifier.c +6 -0
@@ -19990,6 +19990,12 @@ static int jit_subprogs(struct bpf_verifier_env *env)
                        insn[0].imm = (u32)addr;
                        insn[1].imm = addr >> 32;
                }
+
+               if (bpf_pseudo_call(insn))
+                       /* In the x86_64 JIT, tailcall information can only be
+                        * propagated if the subprog is tail_call_reachable.
+                        */
+                       insn->dst_reg = env->subprog_info[subprog].tail_call_reachable;

I really don't like hacking flags into dst_reg.
We already abuse insn->off which is ugly too,
but at least we clean insns later after JIT.

Quoted 3 lines

I'd rather live with this tail call inefficiency than abuse insns fields further.

OK, let us use 'pseudo_call && tail_call_reachable' in x86 JIT to avoid touching 'insn->dst_reg'.

Thanks,
Leon

Quoted 1 line

pw-bot: cr