On Tue, Aug 26, 2025 at 05:24:22PM +0200, David Hildenbrand wrote:
> On 26.08.25 16:32, Lorenzo Stoakes wrote:
> > On Tue, Aug 26, 2025 at 04:28:20PM +0200, David Hildenbrand wrote:
> > > On 26.08.25 16:21, Lorenzo Stoakes wrote:
> > > > On Tue, Aug 26, 2025 at 03:12:08PM +0200, David Hildenbrand wrote:
> > > > > On 12.08.25 17:44, Lorenzo Stoakes wrote:
> > > > > > We now need to account for flag initialisation on fork. We retain the
> > > > > > existing logic as much as we can, but dub the existing flag mask legacy.
> > > > > >
> > > > > > These flags are therefore required to fit in the first 32-bits of the flags
> > > > > > field.
> > > > > >
> > > > > > However, further flag propagation upon fork can be implemented in mm_init()
> > > > > > on a per-flag basis.
> > > > > >
> > > > > > We ensure we clear the entire bitmap prior to setting it, and use
> > > > > > __mm_flags_get_word() and __mm_flags_set_word() to manipulate these legacy
> > > > > > fields efficiently.
> > > > > >
> > > > > > Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx>
> > > > > > ---
> > > > > > include/linux/mm_types.h | 13 ++++++++++---
> > > > > > kernel/fork.c | 7 +++++--
> > > > > > 2 files changed, 15 insertions(+), 5 deletions(-)
> > > > > >
> > > > > > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > > > > > index 38b3fa927997..25577ab39094 100644
> > > > > > --- a/include/linux/mm_types.h
> > > > > > +++ b/include/linux/mm_types.h
> > > > > > @@ -1820,16 +1820,23 @@ enum {
> > > > > > #define MMF_TOPDOWN 31 /* mm searches top down by default */
> > > > > > #define MMF_TOPDOWN_MASK _BITUL(MMF_TOPDOWN)
> > > > > > -#define MMF_INIT_MASK (MMF_DUMPABLE_MASK | MMF_DUMP_FILTER_MASK |\
> > > > > > +#define MMF_INIT_LEGACY_MASK (MMF_DUMPABLE_MASK | MMF_DUMP_FILTER_MASK |\
> > > > > > MMF_DISABLE_THP_MASK | MMF_HAS_MDWE_MASK |\
> > > > > > MMF_VM_MERGE_ANY_MASK | MMF_TOPDOWN_MASK)
> > > > > > -static inline unsigned long mmf_init_flags(unsigned long flags)
> > > > > > +/* Legacy flags must fit within 32 bits. */
> > > > > > +static_assert((u64)MMF_INIT_LEGACY_MASK <= (u64)UINT_MAX);
> > > > >
> > > > > Why not use the magic number 32 you are mentioning in the comment? :)
> > > >
> > > > Meh I mean UINT_MAX works as a good 'any bit' mask and this will work on
> > > > both 32-bit and 64-bit systems.
> > > >
> > > > >
> > > > > static_assert((u32)MMF_INIT_LEGACY_MASK != MMF_INIT_LEGACY_MASK);
> > > >
> > > > On 32-bit that'd not work would it?
> > >
> > > On 32bit, BIT(32) would exceed the shift width of unsigned long -> undefined
> > > behavior.
> > >
> > > The compiler should naturally complain.
> >
> > Yeah, I don't love that sorry. Firstly it's a warning, so you may well miss it
> > (I just tried),
>
> Upstream bots usually complain at you for warnings :P
Fine, but it's not a static assert and they can be delayed.
>
> > and secondly you're making the static assert not have any
> > meaning except that you expect to trigger a compiler warning, it's a bit
> > bizarre.
>
> On 64 bit where BIT(32) *makes any sense* it triggers as expected, no?
It's not a static assert.
>
> >
> > My solution works (unless you can see a reason it shouldn't) and I don't find
> > this approach any simpler.
>
> Please explain to me like I am a 5 yo how your approach works with BIT(32)
> on 32bit when the behavior on 32bit is undefined. :P
OK right I see, in both cases BIT(32) is going to cause a warning on 32-bit.
I was wrong in thinking (u64)(1UL << 32) would get fixed up because of the
outer cast I guess.
This was the mistake here, so fine, we could do it this way.
I guess I'll have to respin the series at this point.
