Re: [PATCH 08/10] mm: update fork mm->flags initialisation to use bitmap

[David Hildenbrand <david@redhat.com>]

On 26.08.25 17:39, Lorenzo Stoakes wrote:
> 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@oracle.com>
>>>>>>> ---
>>>>>>>      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.

Let me think again: assuming someone would mess up the BIT() thing and 
convert back to 1 << NR, your variant would catch it on 32bit I guess.

So given I was primarily confused by the "u64" when talking about 32it, 
no need for a full resend of this series just to clean this up.

-- 
Cheers

David / dhildenb