Re: [PATCH] mm: centralize and fix max map count limit checking

[Kalesh Singh <kaleshsingh@google.com>]

On Wed, Sep 3, 2025 at 4:46 PM Pedro Falcato <pfalcato@suse.de> wrote:
>
> On Wed, Sep 03, 2025 at 04:24:35PM -0700, Kalesh Singh wrote:
> > The check against the max map count (sysctl_max_map_count) was
> > open-coded in several places. This led to inconsistent enforcement
> > and subtle bugs where the limit could be exceeded.
> >
> > For example, some paths would check map_count > sysctl_max_map_count
> > before allocating a new VMA and incrementing the count, allowing the
> > process to reach sysctl_max_map_count + 1:
> >
> >     int do_brk_flags(...)
> >     {
> >         if (mm->map_count > sysctl_max_map_count)
> >             return -ENOMEM;
> >
> >         /* We can get here with mm->map_count == sysctl_max_map_count */
> >
> >         vma = vm_area_alloc(mm);
> >         ...
> >         mm->map_count++   /* We've now exceeded the threshold. */
> >     }
>
> I think this should be fixed separately, and sent for stable.

Hi Pedro, thanks for the review. I can split this out separate.

>
> >
> > To fix this and unify the logic, introduce a new function,
> > exceeds_max_map_count(), to consolidate the check. All open-coded
> > checks are replaced with calls to this new function, ensuring the
> > limit is applied uniformly and correctly.
>
> Thanks! In general I like the idea.
>
> >
> > To improve encapsulation, sysctl_max_map_count is now static to
> > mm/mmap.c. The new helper also adds a rate-limited warning to make
> > debugging applications that exhaust their VMA limit easier.
> >
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Cc: Minchan Kim <minchan@kernel.org>
> > Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> > Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
> > ---
> >  include/linux/mm.h | 11 ++++++++++-
> >  mm/mmap.c          | 15 ++++++++++++++-
> >  mm/mremap.c        |  7 ++++---
> >  mm/nommu.c         |  2 +-
> >  mm/util.c          |  1 -
> >  mm/vma.c           |  6 +++---
> >  6 files changed, 32 insertions(+), 10 deletions(-)
> >
> > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > index 1ae97a0b8ec7..d4e64e6a9814 100644
> > --- a/include/linux/mm.h
> > +++ b/include/linux/mm.h
> > @@ -192,7 +192,16 @@ static inline void __mm_zero_struct_page(struct page *page)
> >  #define MAPCOUNT_ELF_CORE_MARGIN     (5)
> >  #define DEFAULT_MAX_MAP_COUNT        (USHRT_MAX - MAPCOUNT_ELF_CORE_MARGIN)
> >
> > -extern int sysctl_max_map_count;
> > +/**
> > + * exceeds_max_map_count - check if a VMA operation would exceed max_map_count
> > + * @mm: The memory descriptor for the process.
> > + * @new_vmas: The number of new VMAs the operation will create.
> > + *
> > + * Returns true if the operation would cause the number of VMAs to exceed
> > + * the sysctl_max_map_count limit, false otherwise. A rate-limited warning
> > + * is logged if the limit is exceeded.
> > + */
> > +extern bool exceeds_max_map_count(struct mm_struct *mm, unsigned int new_vmas);
>
> No new "extern" in func declarations please.

Ack

>
> >
> >  extern unsigned long sysctl_user_reserve_kbytes;
> >  extern unsigned long sysctl_admin_reserve_kbytes;
> > diff --git a/mm/mmap.c b/mm/mmap.c
> > index 7306253cc3b5..693a0105e6a5 100644
> > --- a/mm/mmap.c
> > +++ b/mm/mmap.c
> > @@ -374,7 +374,7 @@ unsigned long do_mmap(struct file *file, unsigned long addr,
> >               return -EOVERFLOW;
> >
> >       /* Too many mappings? */
> > -     if (mm->map_count > sysctl_max_map_count)
> > +     if (exceeds_max_map_count(mm, 0))
> >               return -ENOMEM;
>
> If the brk example is incorrect, isn't this also wrong? /me is confused

Ahh you are right, this will also go over by 1 once we return from
mmap_region(). I'll batch this with the do_brk_flags() fix.

> >
> >       /*
> > @@ -1504,6 +1504,19 @@ struct vm_area_struct *_install_special_mapping(
> >  int sysctl_legacy_va_layout;
> >  #endif
> >
> > +static int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
> > +
> > +bool exceeds_max_map_count(struct mm_struct *mm, unsigned int new_vmas)
> > +{
> > +     if (unlikely(mm->map_count + new_vmas > sysctl_max_map_count)) {
> > +             pr_warn_ratelimited("%s (%d): Map count limit %u exceeded\n",
> > +                                 current->comm, current->pid,
> > +                                 sysctl_max_map_count);
>
> I'm not entirely sold on the map count warn, even if it's rate limited. It
> sounds like something you can hit in nasty edge cases and nevertheless flood
> your dmesg (more frustrating if you can't fix the damn program).

I don't feel strongly about this, I can drop it in the next revision.

>
> In any case, if we are to make the checks more strict, we should also add
> asserts around the place. Though there's a little case in munmap() we can indeed
> go over temporarily, on purpose.

To confirm, do you mean we should WARN_ON() checks where map count is
being incremented?

>  Though there's a little case in munmap() we can indeed
>  go over temporarily, on purpose.

For the 3 way split we need 1 additional VMA after munmap completed as
one of the 3 gets unmapped. The check is done in the caller beforehand
as __split_vma() explicitly doesn't check map_count. Though if we add
asserts we'll need a variant of vma_complete() or the like that
doesn't enforce the threshold.

Thanks,
Kalesh

>
> --
> Pedro