Re: [PATCH] vmalloc: support __GFP_RETRY_MAYFAIL and __GFP_NORETRY

[Mikulas Patocka <mpatocka@redhat.com>]

On Mon, 2 Mar 2026, Uladzislau Rezki (Sony) wrote:

> From: Michal Hocko <mhocko@suse.com>
> 
> __GFP_RETRY_MAYFAIL and __GFP_NORETRY haven't been supported so far
> because their semantic (i.e. to not trigger OOM killer) is not possible
> with the existing vmalloc page table allocation which is allowing for
> the OOM killer.
> 
> Example: __vmalloc(size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
> 
> <snip>
>  vmalloc_test/55 invoked oom-killer:
>  gfp_mask=0x40dc0(
>  GFP_KERNEL|__GFP_ZERO|__GFP_COMP), order=0, oom_score_adj=0
>  active_anon:0 inactive_anon:0 isolated_anon:0
>   active_file:0 inactive_file:0 isolated_file:0
>   unevictable:0 dirty:0 writeback:0
>   slab_reclaimable:700 slab_unreclaimable:33708
>   mapped:0 shmem:0 pagetables:5174
>   sec_pagetables:0 bounce:0
>   kernel_misc_reclaimable:0
>   free:850 free_pcp:319 free_cma:0
>  CPU: 4 UID: 0 PID: 639 Comm: vmalloc_test/55 ...
>  Hardware name: QEMU Standard PC (i440FX + PIIX, ...
>  Call Trace:
>   <TASK>
>   dump_stack_lvl+0x5d/0x80
>   dump_header+0x43/0x1b3
>   out_of_memory.cold+0x8/0x78
>   __alloc_pages_slowpath.constprop.0+0xef5/0x1130
>   __alloc_frozen_pages_noprof+0x312/0x330
>   alloc_pages_mpol+0x7d/0x160
>   alloc_pages_noprof+0x50/0xa0
>   __pte_alloc_kernel+0x1e/0x1f0
>   ...
> <snip>
> 
> There are usecases for these modifiers when a large allocation request
> should rather fail than trigger OOM killer which wouldn't be able to
> handle the situation anyway [1].
> 
> While we cannot change existing page table allocation code easily we can
> piggy back on scoped NOWAIT allocation for them that we already have in
> place. The rationale is that the bulk of the consumed memory is sitting
> in pages backing the vmalloc allocation. Page tables are only
> participating a tiny fraction. Moreover page tables for virtually allocated
> areas are never reclaimed so the longer the system runs to less likely
> they are. It makes sense to allow an approximation of __GFP_RETRY_MAYFAIL
> and __GFP_NORETRY even if the page table allocation part is much weaker.
> This doesn't break the failure mode while it allows for the no OOM
> semantic.
> 
> [1] https://lore.kernel.org/all/32bd9bed-a939-69c4-696d-f7f9a5fe31d8@redhat.com/T/#u
> 
> Tested-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> Signed-off-by: Michal Hocko <mhocko@suse.com>
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> ---
>  mm/vmalloc.c | 17 ++++++++++++-----
>  1 file changed, 12 insertions(+), 5 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index a06f4b3ea367..975592b0ec89 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -3798,6 +3798,8 @@ static void defer_vm_area_cleanup(struct vm_struct *area)
>   * non-blocking (no __GFP_DIRECT_RECLAIM) - memalloc_noreclaim_save()
>   * GFP_NOFS - memalloc_nofs_save()
>   * GFP_NOIO - memalloc_noio_save()
> + * __GFP_RETRY_MAYFAIL, __GFP_NORETRY - memalloc_noreclaim_save()
> + * to prevent OOMs
>   *
>   * Returns a flag cookie to pair with restore.
>   */
> @@ -3806,7 +3808,8 @@ memalloc_apply_gfp_scope(gfp_t gfp_mask)
>  {
>  	unsigned int flags = 0;
>  
> -	if (!gfpflags_allow_blocking(gfp_mask))
> +	if (!gfpflags_allow_blocking(gfp_mask) ||
> +			(gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_NORETRY)))
>  		flags = memalloc_noreclaim_save();

I wouldn't do this because:

1. it makes the __GFP_RETRY_MAYFAIL allocations unreliable.

2. The comment at memalloc_noreclaim_save says that it may deplete memory 
reserves: "This should only be used when the caller guarantees the 
allocation will allow more memory to be freed very shortly, i.e. it needs 
to allocate some memory in the process of freeing memory, and cannot 
reclaim due to potential recursion."


I think that the cleanest solution to this problem would be to get rid of 
PF_MEMALLOC_NOFS and PF_MEMALLOC_NOIO and instead introduce two per-thread 
variables "gfp_t set_flags" and "gfp_t clear_flags" and set and clear gfp 
flags according to them in the allocator: "gfp = (gfp | 
current->set_flags) & ~current->clear_flags";

Mikulas