[PATCH v3 3/5] malloc: Arena is not needed for tcache path in free()

[Guo, Wangyang]

On 11/26/2024 2:20 PM, H.J. Lu wrote:

> On Tue, Nov 26, 2024, 11:41 AM Guo, Wangyang <wangyang.guo@intel.com> wrote:
>> On 11/26/2024 11:23 AM, H.J. Lu wrote:
>>
>>
>>
>> On Tue, Nov 26, 2024, 10:55 AM Guo, Wangyang <wangyang.guo@intel.com> wrote:
>>> On 11/25/2024 7:58 PM, H.J. Lu wrote:
>>>
>>> On Mon, Nov 25, 2024 at 6:34 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>
>>> On Thu, Aug 29, 2024 at 2:31 PM Wangyang Guo <wangyang.guo@intel.com> wrote:
>>>
>>> Arena is not needed for _int_free_check() in non-DEBUG mode. This commit
>>> defers arena deference to _int_free_chunk() thus accelerate tcache path.
>>> When DEBUG enabled, arena can be obtained from p in do_check_inuse_chunk().
>>>
>>> Result of bench-malloc-thread benchmark
>>>
>>> Test Platform: Xeon-8380
>>> Ratio: New / Original time_per_iteration (Lower is Better)
>>>
>>> Threads#   | Ratio
>>> -----------|------
>>> 1 thread   | 0.994
>>> 4 threads  | 0.968
>>>
>>> The data shows it can brings 3% performance gain in multi-thread scenario.
>>>
>>> ---
>>> Changes in v2:
>>> - _int_free_check() should be put outside of USE_TCACHE.
>>> - Link to v1: https://sourceware.org/pipermail/libc-alpha/2024-August/159360.html
>>> ---
>>>   malloc/malloc.c | 10 ++++++++--
>>>   1 file changed, 8 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/malloc/malloc.c b/malloc/malloc.c
>>> index 264f35e1a3..efb5292e9f 100644
>>> --- a/malloc/malloc.c
>>> +++ b/malloc/malloc.c
>>> @@ -2143,6 +2143,9 @@ do_check_inuse_chunk (mstate av, mchunkptr p)
>>>   {
>>>     mchunkptr next;
>>>
>>> +  if (av == NULL)
>>> +    av = arena_for_chunk (p);
>>> +
>>>     do_check_chunk (av, p);
>>>
>>>     if (chunk_is_mmapped (p))
>>> @@ -3447,9 +3450,10 @@ __libc_free (void *mem)
>>>         /* Mark the chunk as belonging to the library again.  */
>>>         (void)tag_region (chunk2mem (p), memsize (p));
>>>
>>> -      ar_ptr = arena_for_chunk (p);
>>>         INTERNAL_SIZE_T size = chunksize (p);
>>> -      _int_free_check (ar_ptr, p, size);
>>> +      /* av is not needed for _int_free_check in non-DEBUG mode,
>>> +        in DEBUG mode, av will fetch from p in do_check_inuse_chunk.  */
>>> +      _int_free_check (NULL, p, size);
>>>
>>> If _int_free_check doesn't need av, can you push it further down?
>>>
>>> Something like this?
>>>
>>> The problem is _init_free is also used in other places. We can not make sure `av` always got from `arena_for_chunk (p)`.
>>
>> Have you looked at my change?
>>
>> Your patch try to remove `av` parameters in _int_free_check:
>>
>> @@ -4684,7 +4693,7 @@ _int_free (mstate av, mchunkptr p, int have_lock)
>>
>>     size = chunksize (p);
>>
>> -  _int_free_check (av, p, size);
>> +  _int_free_check (p, size);
>>
>> And fetch `av` by `arena_for_chunk (p)`.
>>
>> +static inline void +do_check_inuse_chunk_fast (mstate av, mchunkptr p) +{ + mstate av = arena_for_chunk (p); + do_check_inuse_chunk (av, p); +}
>>
>> My point is for _int_free(av, p, have_lock), if `av` is not equals to `arena_for_chunk (p)`, then we can not simply omit it from `_int_free_check`.
> Is inlining _int_free good enough for performance improvement?
>
> H.J.
The major performance gain comes from inlining. I think that is good 
enough and can drop this patch.