Re: [PATCH bpf-next v2] bpf: Add preempt_disable to protect get_perf_callchain
From: Tao Chen
Date: Sun Sep 28 2025 - 00:11:41 EST
在 2025/9/27 02:52, Andrii Nakryiko 写道:
On Fri, Sep 26, 2025 at 8:40 AM Tao Chen <chen.dylane@xxxxxxxxx> 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. We therefore use bpf_perf_callchain_entries percpu entries
similarly to bpf_try_get_buffers to preserve the current task's
callchain and prevent overwriting by preempting tasks. And we also
add preempt_disable to protect get_perf_callchain.
Reported-by: Alexei Starovoitov <ast@xxxxxxxxxx>
Closes: https://lore.kernel.org/bpf/CAADnVQ+s8B7-fvR1TNO-bniSyKv57cH_ihRszmZV7pQDyV=VDQ@xxxxxxxxxxxxxx
Signed-off-by: Tao Chen <chen.dylane@xxxxxxxxx>
---
kernel/bpf/stackmap.c | 76 ++++++++++++++++++++++++++++++++++---------
1 file changed, 61 insertions(+), 15 deletions(-)
Change list:
v1 -> v2:
From Alexei
- create percpu entris to preserve current task's callchain
similarly to bpf_try_get_buffers.
v1: https://lore.kernel.org/bpf/20250922075333.1452803-1-chen.dylane@xxxxxxxxx
diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
index 2e182a3ac4c..8788c219926 100644
--- a/kernel/bpf/stackmap.c
+++ b/kernel/bpf/stackmap.c
@@ -31,6 +31,55 @@ 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(struct bpf_perf_callchain_entry **entry,
+ struct pt_regs *regs, u32 init_nr, bool kernel,
+ bool user, u32 max_stack, bool crosstack,
+ bool add_mark)
+{
+ 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]);
+
+ preempt_disable();
+ perf_entry = get_perf_callchain(regs, init_nr, kernel, user, max_stack,
+ crosstack, add_mark);
+ if (unlikely(!perf_entry)) {
+ preempt_enable();
+ this_cpu_dec(bpf_perf_callchain_preempt_cnt);
+ return -EFAULT;
+ }
+ memcpy(bpf_entry, perf_entry, sizeof(u64) * (perf_entry->nr + 1));
N copies of a stack trace is not good enough, let's have N + 1 now :)
If we are going with our own buffers, we need to teach
get_perf_callchain to let us pass that buffer directly to avoid that
unnecessary copy.
Also, I know it's about 1KB, but it would be so simple and efficient
to just have this bpf_perf_callchain_entry on the stack. Kernel has a
16KB stack, right? It feels like for something like this using 1KB of
the stack to simplify and speed up stack trace capture is a good
enough reason.
It's a good idea, if that's the case, do we also not need preempt_disable for get_perf_callchain?
+ *entry = bpf_entry;
+ preempt_enable();
+
+ 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);
@@ -303,8 +352,9 @@ BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
u32 max_depth = map->value_size / stack_map_data_size(map);
u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
bool user = flags & BPF_F_USER_STACK;
- struct perf_callchain_entry *trace;
+ struct bpf_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 +364,15 @@ 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, 0, kernel, user, max_depth,
- false, false);
+ err = bpf_get_perf_callchain(&trace, regs, 0, kernel, user, max_depth,
+ false, false);
+ if (err)
+ return err;
- if (unlikely(!trace))
- /* couldn't fetch the stack trace */
- return -EFAULT;
+ err = __bpf_get_stackid(map, (struct perf_callchain_entry *)trace, flags);
+ bpf_put_perf_callchain();
- return __bpf_get_stackid(map, trace, flags);
+ return err;
}
const struct bpf_func_proto bpf_get_stackid_proto = {
@@ -443,8 +494,7 @@ 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 */
+ preempt_disable();
if (trace_in)
trace = trace_in;
@@ -455,8 +505,7 @@ static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
crosstask, false);
if (unlikely(!trace) || trace->nr < skip) {
- if (may_fault)
- rcu_read_unlock();
+ preempt_enable();
goto err_fault;
}
@@ -474,10 +523,7 @@ static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
} else {
memcpy(buf, ips, copy_len);
}
-
- /* trace/ips should not be dereferenced after this point */
- if (may_fault)
- rcu_read_unlock();
+ preempt_enable();
if (user_build_id)
stack_map_get_build_id_offset(buf, trace_nr, user, may_fault);
really it's just build_id resolution that can take a while, which is
why we are trying to avoid preemption around it. But for non-build_id
case, can we avoid extra copying?
--
2.48.1
--
Best Regards
Tao Chen