Re: [PATCH v3 0/7] prctl: introduce PR_SET/GET_THP_POLICY

[Mike Rapoport <rppt@kernel.org>]

(cc'ing linux-api)

On Mon, May 19, 2025 at 11:29:52PM +0100, Usama Arif wrote:
> This series allows to change the THP policy of a process, according to the
> value set in arg2, all of which will be inherited during fork+exec:
> - PR_DEFAULT_MADV_HUGEPAGE: This will set VM_HUGEPAGE and clear VM_NOHUGEPAGE
>   for the default VMA flags. It will also iterate through every VMA in the
>   process and call hugepage_madvise on it, with MADV_HUGEPAGE policy.
>   This effectively allows setting MADV_HUGEPAGE on the entire process.
>   In an environment where different types of workloads are run on the
>   same machine, this will allow workloads that benefit from always having
>   hugepages to do so, without regressing those that don't.
> - PR_DEFAULT_MADV_NOHUGEPAGE: This will set VM_NOHUGEPAGE and clear VM_HUGEPAGE
>   for the default VMA flags. It will also iterate through every VMA in the
>   process and call hugepage_madvise on it, with MADV_NOHUGEPAGE policy.
>   This effectively allows setting MADV_NOHUGEPAGE on the entire process.
>   In an environment where different types of workloads are run on the
>   same machine,this will allow workloads that benefit from having
>   hugepages on an madvise basis only to do so, without regressing those
>   that benefit from having hugepages always.
> - PR_THP_POLICY_SYSTEM: This will reset (clear) both VM_HUGEPAGE and
>   VM_NOHUGEPAGE process for the default flags.
> 
> In hyperscalers, we have a single THP policy for the entire fleet.
> We have different types of workloads (e.g. AI/compute/databases/etc)
> running on a single server.
> Some of these workloads will benefit from always getting THP at fault
> (or collapsed by khugepaged), some of them will benefit by only getting
> them at madvise.
> 
> This series is useful for 2 usecases:
> 1) global system policy = madvise, while we want some workloads to get THPs
> at fault and by khugepaged :- some processes (e.g. AI workloads) benefits
> from getting THPs at fault (and collapsed by khugepaged). Other workloads
> like databases will incur regression (either a performance regression or
> they are completely memory bound and even a very slight increase in memory
> will cause them to OOM). So what these patches will do is allow setting
> prctl(PR_DEFAULT_MADV_HUGEPAGE) on the AI workloads, (This is how
> workloads are deployed in our (Meta's/Facebook) fleet at this moment).
> 
> 2) global system policy = always, while we want some workloads to get THPs
> only on madvise basis :- Same reason as 1). What these patches
> will do is allow setting prctl(PR_DEFAULT_MADV_NOHUGEPAGE) on the database
> workloads. (We hope this is us (Meta) in the near future, if a majority of
> workloads show that they benefit from always, we flip the default host
> setting to "always" across the fleet and workloads that regress can opt-out
> and be "madvise". New services developed will then be tested with always by
> default. "always" is also the default defconfig option upstream, so I would
> imagine this is faced by others as well.)
> 
> v2->v3: (Thanks Lorenzo for all the below feedback!)
> v2: https://lore.kernel.org/all/20250515133519.2779639-1-usamaarif642@gmail.com/
> - no more flags2.
> - no more MMF2_...
> - renamed policy to PR_DEFAULT_MADV_(NO)HUGEPAGE
> - mmap_write_lock_killable acquired in PR_GET_THP_POLICY
> - mmap_write lock fixed in PR_SET_THP_POLICY
> - mmap assert check in process_default_madv_hugepage
> - check if hugepage_global_enabled is enabled in the call and account for s390
> - set mm->def_flags VM_HUGEPAGE and VM_NOHUGEPAGE according to the policy in
>   the way done by madvise(). I believe VM merge will not be broken in
>   this way.
> - process_default_madv_hugepage function that does for_each_vma and calls
>   hugepage_madvise.
> 
> v1->v2:
> - change from modifying the THP decision making for the process, to modifying
>   VMA flags only. This prevents further complicating the logic used to
>   determine THP order (Thanks David!)
> - change from using a prctl per policy change to just using PR_SET_THP_POLICY
>   and arg2 to set the policy. (Zi Yan)
> - Introduce PR_THP_POLICY_DEFAULT_NOHUGE and PR_THP_POLICY_DEFAULT_SYSTEM
> - Add selftests and documentation.
>  
> Usama Arif (7):
>   mm: khugepaged: extract vm flag setting outside of hugepage_madvise
>   prctl: introduce PR_DEFAULT_MADV_HUGEPAGE for the process
>   prctl: introduce PR_DEFAULT_MADV_NOHUGEPAGE for the process
>   prctl: introduce PR_THP_POLICY_SYSTEM for the process
>   selftests: prctl: introduce tests for PR_DEFAULT_MADV_NOHUGEPAGE
>   selftests: prctl: introduce tests for PR_THP_POLICY_DEFAULT_HUGE
>   docs: transhuge: document process level THP controls
> 
>  Documentation/admin-guide/mm/transhuge.rst    |  42 +++
>  include/linux/huge_mm.h                       |   2 +
>  include/linux/mm.h                            |   2 +-
>  include/linux/mm_types.h                      |   4 +-
>  include/uapi/linux/prctl.h                    |   6 +
>  kernel/sys.c                                  |  53 ++++
>  mm/huge_memory.c                              |  13 +
>  mm/khugepaged.c                               |  26 +-
>  tools/include/uapi/linux/prctl.h              |   6 +
>  .../trace/beauty/include/uapi/linux/prctl.h   |   6 +
>  tools/testing/selftests/prctl/Makefile        |   2 +-
>  tools/testing/selftests/prctl/thp_policy.c    | 286 ++++++++++++++++++
>  12 files changed, 436 insertions(+), 12 deletions(-)
>  create mode 100644 tools/testing/selftests/prctl/thp_policy.c
> 
> -- 
> 2.47.1
> 
> 

-- 
Sincerely yours,
Mike.