Re: [PATCH v2] mm/page_alloc: apply per-task GFP context in bulk allocator

[Qiqi Liu]

Thanks for the suggestion. I'll fold this into the commit message and send v3.

Best regards,
Qiqi

On Mon, Sep 14, 2026 at 03:01:14PM +0200, Vlastimil Babka (SUSE) wrote:
> On 9/14/26 10:04, Qiqi Liu wrote:
> > alloc_pages_bulk_noprof() does not call current_gfp_context(), so
> > per-task scoped allocation constraints (PF_MEMALLOC_NOIO,
> > PF_MEMALLOC_NOFS, PF_MEMALLOC_PIN) are not applied on the bulk fast
> > path.
> > 
> > The most direct consequence is PF_MEMALLOC_PIN: without clearing
> 
> I'd say ignoring PF_MEMALLOC_NOFS/PF_MEMALLOC_NOIO also can have a direct
> consequence?
> 
> Let's replace the line above with:
> 
> By ignoring PF_MEMALLOC_NOIO and PF_MEMALLOC_NOFS, the allocation can
> theoretically result in a deadlock.
> 
> Ignoring PF_MEMALLOC_PIN also has consequences: without clearing ...
> 
> > __GFP_MOVABLE, prepare_alloc_pages() selects MIGRATE_MOVABLE for
> > the PCP list, and a task with PF_MEMALLOC_PIN set receives movable
> > pages from the bulk allocator.  Once pinned, these pages can no
> > longer be migrated but remain in MOVABLE pageblocks, violating the
> > mobility contract.  This can increase fragmentation and interfere
> > with compaction or contiguous-memory allocations, eventually
> > surfacing as higher allocation latency or allocation failures under
> > memory pressure.
> > 
> > Found via review of the bulk allocation tracepoint hooks [1].
> > 
> > [1] https://lore.kernel.org/all/20260907120949.418450-1-liuqiqi@xxxxxxxxxx/
> > Link: https://sashiko.dev/#/patchset/20260907120949.418450-1-liuqiqi%40kylinos.cn
> > Fixes: 387ba26fb1cb ("mm/page_alloc: add a bulk page allocator")
> > Cc: stable@xxxxxxxxxxxxxxx
> > Signed-off-by: Qiqi Liu <liuqiqi@xxxxxxxxxx>
> 
> LGTM.
> 
> Reviewed-by: Vlastimil Babka (SUSE) <vbabka@xxxxxxxxxx>
> 
> > ---
> >  mm/page_alloc.c | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 12fac9084c48..3be7288a1aa7 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -5207,6 +5207,7 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
> >  
> >  	/* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
> >  	gfp &= gfp_allowed_mask;
> > +	gfp = current_gfp_context(gfp);
> >  	if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &gfp, &alloc_flags))
> >  		goto out;
> >