Re: [PATCH 03/22] mm: introduce MMF_KERNEL flag and set it for init_mm

[Kevin Brodsky]

On 11/08/2026 10:10, David Laight wrote:
> On Wed, 5 Aug 2026 14:07:01 +0200
> "David Hildenbrand (Arm)" <david@xxxxxxxxxx> wrote:
>
>> On 8/3/26 15:59, Christophe Leroy (CS GROUP) wrote:
>>>
>>> Le 14/07/2026 à 16:03, Kevin Brodsky a écrit :  
>>>> mm code often needs to know whether some mm represents a kernel or
>>>> user address space. This is currently done by comparing the mm
>>>> pointer with &init_mm; besides not being particularly elegant, this
>>>> ignores the fact that other mm's (e.g. efi_mm) may also represent
>>>> parts of the kernel address space.
>>>>
>>>> Introduce a new mm flag MMF_KERNEL and set it for init_mm.
>>>> Subsequent patches will use this flag to replace comparisons with
>>>> &init_mm. No functional change is introduced for now.  
>>> Did you consider performance impact ? This test is usually done in quite
>>> critical memory handling functions.
>>>
>>> init_mm is known at link time. Before your patch 08/22 there is just a
>>> comparison of mm (r3) with a constant (loaded in r10):
>>>
>>> c0014048 <assert_pte_locked>:
>>> c0014048:    3d 40 c1 09     lis     r10,-16119
>>> c001404c:    39 4a 03 98     addi    r10,r10,920
>>> c0014050:    7c 03 50 00     cmpw    r3,r10
>>> c0014054:    4d 82 00 20     beqlr
>>> ...
>>>
>>> After patch 08/22 we have, it first checks that mm is not 0, then it loads the
>>> word located at mm+528 then AND it with 0x1. This load might be costly.
>>>
>>> c0014048 <assert_pte_locked>:
>>> c0014048:    2c 03 00 00     cmpwi   r3,0
>>> c001404c:    7c 85 23 78     mr      r5,r4
>>> c0014050:    41 82 00 10     beq     c0014060 <assert_pte_locked+0x18>
>>> c0014054:    81 23 02 10     lwz     r9,528(r3)
>>> c0014058:    71 29 00 01     andi.   r9,r9,1
>>> c001405c:    4c 82 00 20     bnelr  
>> Is that a real problem, though?
> The cost of the read is likely to matter most if the branch gets mispredicted.
> I'd also guess the mm isn't usually NULL - so that branch needs to statically
> predicted correctly as well.
> I'd also not assume that the mm is in the cache (unless the surrounding code
> has already accessed it), the L1 data caches are small.

You're right, I didn't think about this the right way round - in general
it may not be easy to predict that mm points to init_mm, and an
arbitrary user mm isn't guaranteed to be cached. Repeated calls to
mm_is_kernel(mm) in the same function shouldn't cost much though.

Would some (micro)benchmark be particularly relevant to measure the
overhead, if any?

- Kevin