[PATCH bpf-next v4 0/2] Pass external callchain entry to get_perf_callchain

[PATCH bpf-next v4 0/2] Pass external callchain entry to get_perf_callchain

[Tao Chen]

Background
==========
Alexei noted we should use preempt_disable to protect get_perf_callchain
in bpf stackmap.
https://lore.kernel.org/bpf/CAADnVQ+s8B7-fvR1TNO-bniSyKv57cH_ihRszmZV7pQDyV=VDQ@mail.gmail.com

A previous patch was submitted to attempt fixing this issue. And Andrii
suggested teach get_perf_callchain to let us pass that buffer directly to
avoid that unnecessary copy.
https://lore.kernel.org/bpf/20250926153952.1661146-1-chen.dylane@linux.dev

Proposed Solution
=================
Add external perf_callchain_entry parameter for get_perf_callchain to
allow us to use external buffer from BPF side. The biggest advantage is
that it can reduce unnecessary copies.

Todo
====
But I'm not sure if this modification is appropriate. After all, the
implementation of get_callchain_entry in the perf subsystem seems much more
complex than directly using an external buffer.

Comments and suggestions are always welcome.

Change list:
 - v1 -> v2
   From Jiri
   - rebase code, fix conflict
 - v1: https://lore.kernel.org/bpf/20251013174721.2681091-1-chen.dylane@linux.dev
 
 - v2 -> v3:
   From Andrii
   - entries per CPU used in a stack-like fashion
 - v2: https://lore.kernel.org/bpf/20251014100128.2721104-1-chen.dylane@linux.dev

 v3 -> v4:
   From AI
   - fix double-put issue
 - v3: https://lore.kernel.org/bpf/4c35feba-e075-42c1-82c5-5589c16f088e@linux.dev

Tao Chen (2):
  perf: Use extern perf_callchain_entry for get_perf_callchain
  bpf: Use per-cpu BPF callchain entry to save callchain

 include/linux/perf_event.h |  4 +-
 kernel/bpf/stackmap.c      | 98 ++++++++++++++++++++++++++++----------
 kernel/events/callchain.c  | 13 +++--
 kernel/events/core.c       |  2 +-
 4 files changed, 85 insertions(+), 32 deletions(-)

-- 
2.48.1

[PATCH bpf-next v4 1/2] perf: Use extern perf_callchain_entry for get_perf_callchain

[Tao Chen]

From bpf stack map, we want to use our own buffers to avoid unnecessary
copy, so let us pass it directly. BPF will use this in the next patch.

Signed-off-by: Tao Chen <chen.dylane@linux.dev>
---
 include/linux/perf_event.h |  4 ++--
 kernel/bpf/stackmap.c      |  4 ++--
 kernel/events/callchain.c  | 13 +++++++++----
 kernel/events/core.c       |  2 +-
 4 files changed, 14 insertions(+), 9 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index fd1d91017b9..b144da7d803 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -1719,8 +1719,8 @@ DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
 extern struct perf_callchain_entry *
-get_perf_callchain(struct pt_regs *regs, bool kernel, bool user,
-		   u32 max_stack, bool crosstask, bool add_mark);
+get_perf_callchain(struct pt_regs *regs, struct perf_callchain_entry *external_entry,
+		   bool kernel, bool user, u32 max_stack, bool crosstask, bool add_mark);
 extern int get_callchain_buffers(int max_stack);
 extern void put_callchain_buffers(void);
 extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
index 4d53cdd1374..94e46b7f340 100644
--- a/kernel/bpf/stackmap.c
+++ b/kernel/bpf/stackmap.c
@@ -314,7 +314,7 @@ BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 	if (max_depth > sysctl_perf_event_max_stack)
 		max_depth = sysctl_perf_event_max_stack;
 
-	trace = get_perf_callchain(regs, kernel, user, max_depth,
+	trace = get_perf_callchain(regs, NULL, kernel, user, max_depth,
 				   false, false);
 
 	if (unlikely(!trace))
@@ -451,7 +451,7 @@ static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 	else if (kernel && task)
 		trace = get_callchain_entry_for_task(task, max_depth);
 	else
-		trace = get_perf_callchain(regs, kernel, user, max_depth,
+		trace = get_perf_callchain(regs, NULL, kernel, user, max_depth,
 					   crosstask, false);
 
 	if (unlikely(!trace) || trace->nr < skip) {
diff --git a/kernel/events/callchain.c b/kernel/events/callchain.c
index 808c0d7a31f..851e8f9d026 100644
--- a/kernel/events/callchain.c
+++ b/kernel/events/callchain.c
@@ -217,8 +217,8 @@ static void fixup_uretprobe_trampoline_entries(struct perf_callchain_entry *entr
 }
 
 struct perf_callchain_entry *
-get_perf_callchain(struct pt_regs *regs, bool kernel, bool user,
-		   u32 max_stack, bool crosstask, bool add_mark)
+get_perf_callchain(struct pt_regs *regs, struct perf_callchain_entry *external_entry,
+		   bool kernel, bool user, u32 max_stack, bool crosstask, bool add_mark)
 {
 	struct perf_callchain_entry *entry;
 	struct perf_callchain_entry_ctx ctx;
@@ -228,7 +228,11 @@ get_perf_callchain(struct pt_regs *regs, bool kernel, bool user,
 	if (crosstask && user && !kernel)
 		return NULL;
 
-	entry = get_callchain_entry(&rctx);
+	if (external_entry)
+		entry = external_entry;
+	else
+		entry = get_callchain_entry(&rctx);
+
 	if (!entry)
 		return NULL;
 
@@ -260,7 +264,8 @@ get_perf_callchain(struct pt_regs *regs, bool kernel, bool user,
 	}
 
 exit_put:
-	put_callchain_entry(rctx);
+	if (!external_entry)
+		put_callchain_entry(rctx);
 
 	return entry;
 }
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 7541f6f85fc..5d8e146003a 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -8217,7 +8217,7 @@ perf_callchain(struct perf_event *event, struct pt_regs *regs)
 	if (!kernel && !user)
 		return &__empty_callchain;
 
-	callchain = get_perf_callchain(regs, kernel, user,
+	callchain = get_perf_callchain(regs, NULL, kernel, user,
 				       max_stack, crosstask, true);
 	return callchain ?: &__empty_callchain;
 }
-- 
2.48.1

[PATCH bpf-next v4 2/2] bpf: Use per-cpu BPF callchain entry to save callchain

[Tao Chen]

As Alexei noted, get_perf_callchain() return values may be reused
if a task is preempted after the BPF program enters migrate disable
mode. Drawing on the per-cpu design of bpf_bprintf_buffers,
per-cpu BPF callchain entry is used here.

Signed-off-by: Tao Chen <chen.dylane@linux.dev>
---
 kernel/bpf/stackmap.c | 98 ++++++++++++++++++++++++++++++++-----------
 1 file changed, 73 insertions(+), 25 deletions(-)

diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
index 94e46b7f340..97028d39df1 100644
--- a/kernel/bpf/stackmap.c
+++ b/kernel/bpf/stackmap.c
@@ -31,6 +31,52 @@ struct bpf_stack_map {
 	struct stack_map_bucket *buckets[] __counted_by(n_buckets);
 };
 
+struct bpf_perf_callchain_entry {
+	u64 nr;
+	u64 ip[PERF_MAX_STACK_DEPTH];
+};
+
+#define MAX_PERF_CALLCHAIN_PREEMPT 3
+static DEFINE_PER_CPU(struct bpf_perf_callchain_entry[MAX_PERF_CALLCHAIN_PREEMPT],
+		      bpf_perf_callchain_entries);
+static DEFINE_PER_CPU(int, bpf_perf_callchain_preempt_cnt);
+
+static int bpf_get_perf_callchain_or_entry(struct perf_callchain_entry **entry,
+					   struct pt_regs *regs, bool kernel,
+					   bool user, u32 max_stack, bool crosstack,
+					   bool add_mark, bool get_callchain)
+{
+	struct bpf_perf_callchain_entry *bpf_entry;
+	struct perf_callchain_entry *perf_entry;
+	int preempt_cnt;
+
+	preempt_cnt = this_cpu_inc_return(bpf_perf_callchain_preempt_cnt);
+	if (WARN_ON_ONCE(preempt_cnt > MAX_PERF_CALLCHAIN_PREEMPT)) {
+		this_cpu_dec(bpf_perf_callchain_preempt_cnt);
+		return -EBUSY;
+	}
+
+	bpf_entry = this_cpu_ptr(&bpf_perf_callchain_entries[preempt_cnt - 1]);
+	if (!get_callchain) {
+		*entry = (struct perf_callchain_entry *)bpf_entry;
+		return 0;
+	}
+
+	perf_entry = get_perf_callchain(regs, (struct perf_callchain_entry *)bpf_entry,
+					kernel, user, max_stack,
+					crosstack, add_mark);
+	*entry = perf_entry;
+
+	return 0;
+}
+
+static void bpf_put_perf_callchain(void)
+{
+	if (WARN_ON_ONCE(this_cpu_read(bpf_perf_callchain_preempt_cnt) == 0))
+		return;
+	this_cpu_dec(bpf_perf_callchain_preempt_cnt);
+}
+
 static inline bool stack_map_use_build_id(struct bpf_map *map)
 {
 	return (map->map_flags & BPF_F_STACK_BUILD_ID);
@@ -192,11 +238,11 @@ get_callchain_entry_for_task(struct task_struct *task, u32 max_depth)
 {
 #ifdef CONFIG_STACKTRACE
 	struct perf_callchain_entry *entry;
-	int rctx;
-
-	entry = get_callchain_entry(&rctx);
+	int ret;
 
-	if (!entry)
+	ret = bpf_get_perf_callchain_or_entry(&entry, NULL, false, false, 0, false, false,
+					      false);
+	if (ret)
 		return NULL;
 
 	entry->nr = stack_trace_save_tsk(task, (unsigned long *)entry->ip,
@@ -216,8 +262,6 @@ get_callchain_entry_for_task(struct task_struct *task, u32 max_depth)
 			to[i] = (u64)(from[i]);
 	}
 
-	put_callchain_entry(rctx);
-
 	return entry;
 #else /* CONFIG_STACKTRACE */
 	return NULL;
@@ -305,6 +349,7 @@ BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 	bool user = flags & BPF_F_USER_STACK;
 	struct perf_callchain_entry *trace;
 	bool kernel = !user;
+	int err;
 
 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 			       BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
@@ -314,14 +359,15 @@ BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 	if (max_depth > sysctl_perf_event_max_stack)
 		max_depth = sysctl_perf_event_max_stack;
 
-	trace = get_perf_callchain(regs, NULL, kernel, user, max_depth,
-				   false, false);
+	err = bpf_get_perf_callchain_or_entry(&trace, regs, kernel, user, max_depth,
+					      false, false, true);
+	if (err)
+		return err;
 
-	if (unlikely(!trace))
-		/* couldn't fetch the stack trace */
-		return -EFAULT;
+	err = __bpf_get_stackid(map, trace, flags);
+	bpf_put_perf_callchain();
 
-	return __bpf_get_stackid(map, trace, flags);
+	return err;
 }
 
 const struct bpf_func_proto bpf_get_stackid_proto = {
@@ -443,20 +489,23 @@ static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 	if (sysctl_perf_event_max_stack < max_depth)
 		max_depth = sysctl_perf_event_max_stack;
 
-	if (may_fault)
-		rcu_read_lock(); /* need RCU for perf's callchain below */
-
 	if (trace_in)
 		trace = trace_in;
-	else if (kernel && task)
+	else if (kernel && task) {
 		trace = get_callchain_entry_for_task(task, max_depth);
-	else
-		trace = get_perf_callchain(regs, NULL, kernel, user, max_depth,
-					   crosstask, false);
+	} else {
+		err = bpf_get_perf_callchain_or_entry(&trace, regs, kernel, user, max_depth,
+						      false, false, true);
+		if (err)
+			return err;
+	}
+
+	if (unlikely(!trace))
+		goto err_fault;
 
-	if (unlikely(!trace) || trace->nr < skip) {
-		if (may_fault)
-			rcu_read_unlock();
+	if (trace->nr < skip) {
+		if (!trace_in)
+			bpf_put_perf_callchain();
 		goto err_fault;
 	}
 
@@ -475,9 +524,8 @@ static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 		memcpy(buf, ips, copy_len);
 	}
 
-	/* trace/ips should not be dereferenced after this point */
-	if (may_fault)
-		rcu_read_unlock();
+	if (!trace_in)
+		bpf_put_perf_callchain();
 
 	if (user_build_id)
 		stack_map_get_build_id_offset(buf, trace_nr, user, may_fault);
-- 
2.48.1

Re: [PATCH bpf-next v4 2/2] bpf: Use per-cpu BPF callchain entry to save callchain

[Alexei Starovoitov]

On Tue, Oct 21, 2025 at 9:07 AM Tao Chen <chen.dylane@linux.dev> wrote:
>
> As Alexei noted, get_perf_callchain() return values may be reused
> if a task is preempted after the BPF program enters migrate disable
> mode. Drawing on the per-cpu design of bpf_bprintf_buffers,
> per-cpu BPF callchain entry is used here.
>
> Signed-off-by: Tao Chen <chen.dylane@linux.dev>
> ---
>  kernel/bpf/stackmap.c | 98 ++++++++++++++++++++++++++++++++-----------
>  1 file changed, 73 insertions(+), 25 deletions(-)
>
> diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
> index 94e46b7f340..97028d39df1 100644
> --- a/kernel/bpf/stackmap.c
> +++ b/kernel/bpf/stackmap.c
> @@ -31,6 +31,52 @@ struct bpf_stack_map {
>         struct stack_map_bucket *buckets[] __counted_by(n_buckets);
>  };
>
> +struct bpf_perf_callchain_entry {
> +       u64 nr;
> +       u64 ip[PERF_MAX_STACK_DEPTH];
> +};
> +
> +#define MAX_PERF_CALLCHAIN_PREEMPT 3
> +static DEFINE_PER_CPU(struct bpf_perf_callchain_entry[MAX_PERF_CALLCHAIN_PREEMPT],
> +                     bpf_perf_callchain_entries);
> +static DEFINE_PER_CPU(int, bpf_perf_callchain_preempt_cnt);

This is too much extra memory. Above adds 1k * 3 * num_cpus.
Let's reuse perf callchains.
Especially since they're controlled by perf_event_max_stack sysctl.
See Peter's suggestion in v3.
And for the future don't respin so quickly.

Re: [PATCH bpf-next v4 2/2] bpf: Use per-cpu BPF callchain entry to save callchain

[Tao Chen]

在 2025/10/22 01:28, Alexei Starovoitov 写道:
> On Tue, Oct 21, 2025 at 9:07 AM Tao Chen <chen.dylane@linux.dev> wrote:
>>
>> As Alexei noted, get_perf_callchain() return values may be reused
>> if a task is preempted after the BPF program enters migrate disable
>> mode. Drawing on the per-cpu design of bpf_bprintf_buffers,
>> per-cpu BPF callchain entry is used here.
>>
>> Signed-off-by: Tao Chen <chen.dylane@linux.dev>
>> ---
>>   kernel/bpf/stackmap.c | 98 ++++++++++++++++++++++++++++++++-----------
>>   1 file changed, 73 insertions(+), 25 deletions(-)
>>
>> diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
>> index 94e46b7f340..97028d39df1 100644
>> --- a/kernel/bpf/stackmap.c
>> +++ b/kernel/bpf/stackmap.c
>> @@ -31,6 +31,52 @@ struct bpf_stack_map {
>>          struct stack_map_bucket *buckets[] __counted_by(n_buckets);
>>   };
>>
>> +struct bpf_perf_callchain_entry {
>> +       u64 nr;
>> +       u64 ip[PERF_MAX_STACK_DEPTH];
>> +};
>> +
>> +#define MAX_PERF_CALLCHAIN_PREEMPT 3
>> +static DEFINE_PER_CPU(struct bpf_perf_callchain_entry[MAX_PERF_CALLCHAIN_PREEMPT],
>> +                     bpf_perf_callchain_entries);
>> +static DEFINE_PER_CPU(int, bpf_perf_callchain_preempt_cnt);
> 
> This is too much extra memory. Above adds 1k * 3 * num_cpus.
> Let's reuse perf callchains.
> Especially since they're controlled by perf_event_max_stack sysctl.
> See Peter's suggestion in v3.
> And for the future don't respin so quickly.

Ok, let's base our discussion on v3, sorry for the overly rapid version 
iterations impacted the maintainers' review process, i will respin 
slowly, thanks.

-- 
Best Regards
Tao Chen

[PATCH bpf-next v4 0/2] Pass external callchain entry to get_perf_callchain

[Tao Chen]

Background
==========
Alexei noted we should use preempt_disable to protect get_perf_callchain
in bpf stackmap.
https://lore.kernel.org/bpf/CAADnVQ+s8B7-fvR1TNO-bniSyKv57cH_ihRszmZV7pQDyV=VDQ@mail.gmail.com

A previous patch was submitted to attempt fixing this issue. And Andrii
suggested teach get_perf_callchain to let us pass that buffer directly to
avoid that unnecessary copy.
https://lore.kernel.org/bpf/20250926153952.1661146-1-chen.dylane@linux.dev

Proposed Solution
=================
Add external perf_callchain_entry parameter for get_perf_callchain to
allow us to use external buffer from BPF side. The biggest advantage is
that it can reduce unnecessary copies.

Todo
====
But I'm not sure if this modification is appropriate. After all, the
implementation of get_callchain_entry in the perf subsystem seems much more
complex than directly using an external buffer.

Comments and suggestions are always welcome.

Change list:
 - v1 -> v2
   From Jiri
   - rebase code, fix conflict
 - v1: https://lore.kernel.org/bpf/20251013174721.2681091-1-chen.dylane@linux.dev
 
 - v2 -> v3:
   From Andrii
   - entries per CPU used in a stack-like fashion
 - v2: https://lore.kernel.org/bpf/20251014100128.2721104-1-chen.dylane@linux.dev

 - v3 -> v4:
   From Peter
   - refactor get_perf_callchain and add three new APIs to use perf
     callchain easily.
   From Andrii
   - reuse the perf callchain management.

   - rename patch1 and patch2.
 - v3: https://lore.kernel.org/bpf/20251019170118.2955346-1-chen.dylane@linux.dev

Tao Chen (2):
  perf: Refactor get_perf_callchain
  bpf: Hold the perf callchain entry until used completely

 include/linux/perf_event.h | 11 +++++-
 kernel/bpf/stackmap.c      | 61 ++++++++++++++++++++++++-------
 kernel/events/callchain.c  | 75 ++++++++++++++++++++++++--------------
 kernel/events/core.c       |  2 +-
 4 files changed, 107 insertions(+), 42 deletions(-)

-- 
2.48.1