[PATCH] malloc: Enable THP always support on hugetlb tunable

[Adhemerval Zanella Netto]

On 08/07/25 12:10, William Hunt wrote:
> Enable support for THP always when glibc.malloc.hugetlb=1, as the tunable 
> currently only gives explicit support in malloc for the THP madvise mode 
> by aligning to a huge page size. Add a thp_mode parameter to mp_ and check 
> in madvise_thp whether the system is using madvise mode, otherwise the 
> __madvise call is useless. Set the thp_mode to be unsupported by default, 
> but if the hugetlb tunable is set this updates thp_mode. Performance of 
> xalancbmk improves by 4.9% on Neoverse V2 when THP always mode is set on the 
> system and glibc.malloc.hugetlb=1.
> 
> Passed regress, OK for commit?

LGTM, thanks.  Although, as Andrea as noted, only after 2.42 release.

Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>

> ---
>  malloc/malloc.c | 18 +++++++++++-------
>  1 file changed, 11 insertions(+), 7 deletions(-)
> 
> diff --git a/malloc/malloc.c b/malloc/malloc.c
> index 6da40ad602..e40c6c6cb0 100644
> --- a/malloc/malloc.c
> +++ b/malloc/malloc.c
> @@ -1871,6 +1871,7 @@ struct malloc_par
>    INTERNAL_SIZE_T arena_max;
>  
>    /* Transparent Large Page support.  */
> +  enum malloc_thp_mode_t thp_mode;
>    INTERNAL_SIZE_T thp_pagesize;
>    /* A value different than 0 means to align mmap allocation to hp_pagesize
>       add hp_flags on flags.  */
> @@ -1927,7 +1928,8 @@ static struct malloc_par mp_ =
>    .mmap_threshold = DEFAULT_MMAP_THRESHOLD,
>    .trim_threshold = DEFAULT_TRIM_THRESHOLD,
>  #define NARENAS_FROM_NCORES(n) ((n) * (sizeof (long) == 4 ? 2 : 8))
> -  .arena_test = NARENAS_FROM_NCORES (1)
> +  .arena_test = NARENAS_FROM_NCORES (1),
> +  .thp_mode = malloc_thp_mode_not_supported
>  #if USE_TCACHE
>    ,
>    .tcache_count = TCACHE_FILL_COUNT,
> @@ -2011,6 +2013,11 @@ static inline void
>  madvise_thp (void *p, INTERNAL_SIZE_T size)
>  {
>  #ifdef MADV_HUGEPAGE
> +  /* Only use __madvise if the system is using 'madvise' mode.
> +     Otherwise the call is wasteful. */
> +  if (mp_.thp_mode != malloc_thp_mode_madvise)
> +    return;
> +
Does __glibc_likely make any difference here?

>    /* Do not consider areas smaller than a huge page or if the tunable is
>       not active.  */
>    if (mp_.thp_pagesize == 0 || size < mp_.thp_pagesize)
> @@ -5649,12 +5656,9 @@ do_set_hugetlb (size_t value)
>  {
>    if (value == 1)
>      {
> -      enum malloc_thp_mode_t thp_mode = __malloc_thp_mode ();
> -      /*
> -	 Only enable THP madvise usage if system does support it and
> -	 has 'madvise' mode.  Otherwise the madvise() call is wasteful.
> -       */
> -      if (thp_mode == malloc_thp_mode_madvise)
> +      mp_.thp_mode = __malloc_thp_mode ();
> +      if (mp_.thp_mode == malloc_thp_mode_madvise
> +          || mp_.thp_mode == malloc_thp_mode_always)
>  	mp_.thp_pagesize = __malloc_default_thp_pagesize ();
>      }
>    else if (value >= 2)