Re: [PATCH bpf-next v2 1/2] bpf: Add lookup_and_delete_elem for BPF_MAP_STACK_TRACE

[Tao Chen <chen.dylane@linux.dev>]

在 2025/9/18 09:35, Alexei Starovoitov 写道:
> On Wed, Sep 17, 2025 at 3:16 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
>>
>>
>> P.S. It seems like a good idea to switch STACKMAP to open addressing
>> instead of the current kind-of-bucket-chain-but-not-really
>> implementation. It's fixed size and pre-allocated already, so open
>> addressing seems like a great approach here, IMO.
> 
> That makes sense. It won't have backward compat issues.
> Just more reliable stack_id.
> 
> Fixed value_size is another footgun there.
> Especially for collecting user stack traces.
> We can switch the whole stackmap to bpf_mem_alloc()
> or wait for kmalloc_nolock().
> But it's probably a diminishing return.
> 
> bpf_get_stack() also isn't great with a copy into
> perf_callchain_entry, then 2nd copy into on stack/percpu buf/ringbuf,
> and 3rd copy of correct size into ringbuf (optional).
> 
> Also, I just realized we have another nasty race there.
> In the past bpf progs were run in preempt disabled context,
> but we forgot to adjust bpf_get_stack[id]() helpers when everything
> switched to migrate disable.
> 
> The return value from get_perf_callchain() may be reused
> if another task preempts and requests the stack.
> We have partially incorrect comment in __bpf_get_stack() too:
>          if (may_fault)
>                  rcu_read_lock(); /* need RCU for perf's callchain below */
> 
> rcu can be preemptable. so rcu_read_lock() makes
> trace = get_perf_callchain(...)
> accessible, but that per-cpu trace buffer can be overwritten.
> It's not an issue for CONFIG_PREEMPT_NONE=y, but that doesn't
> give much comfort.

Hi Alexei,

Can we fix it like this?

-       if (may_fault)
-               rcu_read_lock(); /* need RCU for perf's callchain below */
+       preempt_diable();

         if (trace_in)
                 trace = trace_in;
@@ -455,8 +454,7 @@ 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;
         }

@@ -475,9 +473,7 @@ 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();
+       preempt_enable();

> 
> Modern day bpf api would probably be
> - get_callchain_entry()/put() kfuncs to expose low level mechanism
> with safe acq/rel of temp buffer.
> - then another kfuncs to perf_callchain_kernel/user into that buffer.
> 
> and with bpf_mem_alloc and hash kfuncs the bpf prog can
> implement either bpf_get_stack() equivalent or much better
> bpf_get_stackid() with variable length stack traces and so on.


-- 
Best Regards
Tao Chen