[PATCH v6 5/6] elf: Align large load segments to PMD huge page size for THP

[Adhemerval Zanella Netto]

On 09/03/26 13:32, Wilco Dijkstra wrote:
> Hi Rui,
> 
>> Mapping segments that are at least the size of a PMD huge page to
>> huge-page-aligned addresses helps make them eligible for Transparent
>> Huge Pages (THP).
>>
>> This patch introduces a Linux-specific helper, `_dl_map_segment_align`,
>> to determine an appropriate maximum alignment for ELF load segments based
>> on the system THP policy. The optimization is enabled only when the glibc
>> tunable `glibc.elf.hugetlb=1` is set and THP is configured to be used
>> unconditionally.
>>
>> The optimization depends on Linux kernel support for file-backed THP,
>> specifically:
>>
>> * `CONFIG_READ_ONLY_THP_FOR_FS` (available since Linux kernel 5.4), and
> 
> Do you know the status of this in the kernel? It used to be experimental, and
> I think it is still off by default in most distros.
> 
>> * `CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS`.
>>
>> When enabled, the helper queries the default THP page size and uses it
>> to align sufficiently large load segments that are already properly
>> aligned in both virtual address and file offset (e.g., zero).
>>
>> For eligible segments, the alignment is bumped to the THP page size,
>> which improves THP eligibility, reduces TLB pressure, and improves
>> performance for large objects. To avoid excessive address space padding
>> on systems with very large THP sizes, the alignment is capped at 32MB.
>> The optimization is applied only to non-writable segments, matching
>> typical THP usage.
> 
> This is all reasonable, but note there will be an issue with enabling this by
> default (see below).
> 
> +#include <dl-map-segment-align.h>
> +#include <dl-tunables.h>
> +#include <hugepages.h>
> +
> +#define MAX_THP_PAGESIZE (32 * 1024 * 1024)
> 
> Should this be in generic hugepages.h? In malloc there is currently no check
> that rejects really huge pages, so something like this is useful. Having it in the
> generic header also allows targets to override it if necessary.
> 
> +ElfW (Addr)
> +_dl_map_segment_align (const struct loadcmd *c, ElfW (Addr) p_align_max)
> +{
> +  static enum thp_mode_t thp_mode = thp_mode_not_supported;
> +  static unsigned long int thp_pagesize;
> +
> +  if (TUNABLE_GET (glibc, elf, hugetlb, int32_t, NULL) == 0)
> +    return p_align_max;
> +
> +  if (__glibc_unlikely (thp_mode == thp_mode_not_supported
> +                       || thp_pagesize == 0))
> +    {
> +      thp_mode = __get_thp_mode ();
> +      thp_pagesize = DL_MAP_DEFAULT_THP_PAGESIZE ? : __get_thp_size ();
> +    }
> 
> When the tunable is enabled by default, this will always call __get_thp_mode()
> which may fail (or even abort) on some containers. I think we could add a
> similar workaround for this: if DL_MAP_DEFAULT_THP_PAGESIZE is set, use that
> and force thp_mode_always without calling __get_thp_mode(). The alignment is
> essentially for free, it's just wasting virtual space.
> 
> +  /* Aligning load segments that are large enough to the PMD size helps
> +     improve THP eligibility and reduces TLB pressure.
> +     We cap the huge page size at MAX_THP_PAGESIZE to avoid over-aligning
> +     on systems with very large normal pages (like 64K pages with 512M
> +     huge pages). */
> +  if (thp_mode == thp_mode_always && thp_pagesize <= MAX_THP_PAGESIZE
> +      && ((c->mapstart | c->mapoff) & (thp_pagesize - 1)) == 0
> +      && (c->mapend - c->mapstart) >= thp_pagesize
> +      && p_align_max < thp_pagesize && (c->prot & PROT_WRITE) == 0)
> +    return thp_pagesize;
> +
> +  return p_align_max;
> +}

There is also the potential memory usage increase of using this tunable as
default, as we saw a report for the malloc one on AArch64.  I think it would
less than the malloc, but on multi-layer environment with a lot of process
this could compound to a significant value.

I think we should *not* make this the default for now.