Re: [RFC PATCH 0/6] mm: pass alloc_flags through folio, filemap, and bulk allocators

[Vlastimil Babka (SUSE)]

On 10/10/26 01:28, Gregory Price wrote:
> On Fri, Oct 09, 2026 at 10:55:05PM +0200, David Hildenbrand (Arm) wrote:
>> I do agree that two sets of flags is suboptimal, but likely more flexible. I
>> guess an alloc_flags only interface is not easily possible ...

Agree with the alloc_flags option of all the other options.

>> I do wonder whether it should be:
>> 
>> 	typedef int __bitwise alloc_flags_t;
>> 
>> instead of "unsigned int alloc_flags".
>> 
>> ... while at it

Can't hurt, but given how it's internal only it didn't seem that urgent as
for gfp.

> Does it make sense to leave alloc_flags to internal-only state
> and just implement a new flag field?
> 
> gfp_t       -> describe the page (fully public)
> req_t       -> describe the allocator behavior (mm-internal only)
> alloc_flags -> internal allocator state only

Maybe in that case the remaining alloc_flags should just become 1-bit fields
directly in struct alloc_context so we don't manage 3 separate bitwise
types. Or more bits in case of ALLOC_WMARK?

Depends on how much work this would be (and if you want to do it) and how
the result would like.
> typedef int __bitwise pa_req_t; /* Page Allocator Request Flags */
> 
> #define PA_REQ_DEFAULT        0
> #define PA_REQ_NOLOCK         BIT(0) /* from ALLOC_NOLOCK */
> #define PA_REQ_NO_CODETAG     BIT(1) /* from ALLOC_NO_CODETAG */
> #define PA_REQ_SPREAD_DIRTY   BIT(2) /* from __GFP_WRITE */
> #define PA_REQ_DEFER_INIT     BIT(3) /* from __GFP_SKIP_ZERO */
> #define PA_REQ_ZERO_TAGS      BIT(4) /* from __GFP_ZERO_TAGS */
> #define PA_REQ_NO_FALLBACK    BIT(5) /* select no fallback list */
> #define PA_REQ_NO_OOM         BIT(6) /* do not oom */
> 
> Off the bat we recover 2 ALLOC flags and 3 GFP flags. Possibly more with
> a bit more work.
> 
> Then we get:
> 
> #define PA_REQ_PRIVATE        BIT(7) /* allow private node allocation */
> #define PA_REQ_UNMAPPED       BIT(8) /* page must be unmapped */
> 
> And we might not even need a new zonelist for private nodes in this case
> as long as we require internal users to pass PA_REQ_PRIVATE and add a
> function like
> 
> folio_alloc_private(void *owner, ...) {
>    /* validate (pgdat->owner == owner) before allowing allocation */
>    __folio_alloc(..., req | PA_REQ_PRIVATE | PA_REQ_NOFALLBACK, ...);
> }
> 
> User still has the option to request NO_OOM via __GFP_THISNODE :]
> 
> ~Gregory