Re: [PATCH bpf-next RESEND v4 1/2] bpf: refactor max_depth computation in bpf_get_stack()

[Yonghong Song <yonghong.song@linux.dev>]

On 8/25/25 1:07 PM, Lecomte, Arnaud wrote:
>
> On 25/08/2025 19:27, Yonghong Song wrote:
>>
>>
>> On 8/25/25 9:39 AM, Lecomte, Arnaud wrote:
>>>
>>> On 19/08/2025 22:15, Martin KaFai Lau wrote:
>>>> On 8/19/25 9:26 AM, Arnaud Lecomte wrote:
>>>>> A new helper function stack_map_calculate_max_depth() that
>>>>> computes the max depth for a stackmap.
>>>>>
>>>>> Changes in v2:
>>>>>   - Removed the checking 'map_size % map_elem_size' from
>>>>>     stack_map_calculate_max_depth
>>>>>   - Changed stack_map_calculate_max_depth params name to be more 
>>>>> generic
>>>>>
>>>>> Changes in v3:
>>>>>   - Changed map size param to size in max depth helper
>>>>>
>>>>> Changes in v4:
>>>>>   - Fixed indentation in max depth helper for args
>>>>>
>>>>> Link to v3: 
>>>>> https://lore.kernel.org/all/09dc40eb-a84e-472a-8a68-36a2b1835308@linux.dev/
>>>>>
>>>>> Signed-off-by: Arnaud Lecomte <contact@arnaud-lcm.com>
>>>>> Acked-by: Yonghong Song <yonghong.song@linux.dev>
>>>>> ---
>>>>>   kernel/bpf/stackmap.c | 30 ++++++++++++++++++++++++------
>>>>>   1 file changed, 24 insertions(+), 6 deletions(-)
>>>>>
>>>>> diff --git a/kernel/bpf/stackmap.c b/kernel/bpf/stackmap.c
>>>>> index 3615c06b7dfa..b9cc6c72a2a5 100644
>>>>> --- a/kernel/bpf/stackmap.c
>>>>> +++ b/kernel/bpf/stackmap.c
>>>>> @@ -42,6 +42,27 @@ static inline int stack_map_data_size(struct 
>>>>> bpf_map *map)
>>>>>           sizeof(struct bpf_stack_build_id) : sizeof(u64);
>>>>>   }
>>>>>   +/**
>>>>> + * stack_map_calculate_max_depth - Calculate maximum allowed 
>>>>> stack trace depth
>>>>> + * @size:  Size of the buffer/map value in bytes
>>>>> + * @elem_size:  Size of each stack trace element
>>>>> + * @flags:  BPF stack trace flags (BPF_F_USER_STACK, 
>>>>> BPF_F_USER_BUILD_ID, ...)
>>>>> + *
>>>>> + * Return: Maximum number of stack trace entries that can be 
>>>>> safely stored
>>>>> + */
>>>>> +static u32 stack_map_calculate_max_depth(u32 size, u32 elem_size, 
>>>>> u64 flags)
>>>>> +{
>>>>> +    u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
>>>>> +    u32 max_depth;
>>>>> +
>>>>> +    max_depth = size / elem_size;
>>>>> +    max_depth += skip;
>>>>> +    if (max_depth > sysctl_perf_event_max_stack)
>>>>> +        return sysctl_perf_event_max_stack;
>>>>
>>>> hmm... this looks a bit suspicious. Is it possible that 
>>>> sysctl_perf_event_max_stack is being changed to a larger value in 
>>>> parallel?
>>>>
>>> Hi Martin, this is a valid concern as sysctl_perf_event_max_stack 
>>> can be modified at runtime through 
>>> /proc/sys/kernel/perf_event_max_stack.
>>> What we could maybe do instead is to create a copy: u32 current_max 
>>> = READ_ONCE(sysctl_perf_event_max_stack);
>>> Any thoughts on this ?
>>
>> There is no need to have READ_ONCE. Jut do
>>     int curr_sysctl_max_stack = sysctl_perf_event_max_stack;
>>     if (max_depth > curr_sysctl_max_stack)
>>       return curr_sysctl_max_stack;
>>
>> Because of the above change, the patch is not a refactoring change 
>> any more.
>>
> Why would you not consider it as a refactoring change anymore ?

Sorry, I think I made a couple of mistakes in the above.

First, yes, we do want READ_ONCE, other potentially compiler may optimization
the above back to the original code with two references to sysctl_perf_event_max_stack.

Second, yes, it is indeed a refactoring.

>>>
>>>>> +
>>>>> +    return max_depth;
>>>>> +}
>>>>> +
>>>>>   static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
>>>>>   {
>>>>>       u64 elem_size = sizeof(struct stack_map_bucket) +
>>>>> @@ -406,7 +427,7 @@ static long __bpf_get_stack(struct pt_regs 
>>>>> *regs, struct task_struct *task,
>>>>>                   struct perf_callchain_entry *trace_in,
>>>>>                   void *buf, u32 size, u64 flags, bool may_fault)
>>>>>   {
>>>>> -    u32 trace_nr, copy_len, elem_size, num_elem, max_depth;
>>>>> +    u32 trace_nr, copy_len, elem_size, max_depth;
>>>>>       bool user_build_id = flags & BPF_F_USER_BUILD_ID;
>>>>>       bool crosstask = task && task != current;
>>>>>       u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
>>>>> @@ -438,10 +459,7 @@ static long __bpf_get_stack(struct pt_regs 
>>>>> *regs, struct task_struct *task,
>>>>>           goto clear;
>>>>>       }
>>>>>   -    num_elem = size / elem_size;
>>>>> -    max_depth = num_elem + skip;
>>>>> -    if (sysctl_perf_event_max_stack < max_depth)
>>>>> -        max_depth = sysctl_perf_event_max_stack;
>>>>> +    max_depth = stack_map_calculate_max_depth(size, elem_size, 
>>>>> flags);
>>>>>         if (may_fault)
>>>>>           rcu_read_lock(); /* need RCU for perf's callchain below */
>>>>> @@ -461,7 +479,7 @@ static long __bpf_get_stack(struct pt_regs 
>>>>> *regs, struct task_struct *task,
>>>>>       }
>>>>>         trace_nr = trace->nr - skip;
>>>>> -    trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
>>>>
>>>> I suspect it was fine because trace_nr was still bounded by num_elem.
>>>>
>>> We should bring back the num_elem bound as an additional safe net.
>>>>> +    trace_nr = min(trace_nr, max_depth - skip);
>>>>
>>>> but now the min() is also based on max_depth which could be 
>>>> sysctl_perf_event_max_stack.
>>>>
>>>> beside, if I read it correctly, in "max_depth - skip", the 
>>>> max_depth could also be less than skip. I assume trace->nr is bound 
>>>> by max_depth, so should be less of a problem but still a bit 
>>>> unintuitive to read.
>>>>
>>>>>       copy_len = trace_nr * elem_size;
>>>>>         ips = trace->ip + skip;
>>>>
>>
>>