mm, swap: add support for stable large allocation in swap cache directly
Hello Dan
Currently, no caller in kernel will pass NULL mpol with NULL vmf with CONFIG_NUMA. Do do_swap_page always have vmf set, and shmem always have either vmf or mpol set (unless !CONFIG_NUMA).
I used to have a WARN_ON for this but some AI reviewers complained that the WARN_ON is redundant so I removed it. I can add that back later just to be safer, I will add a Reported-by for you. Is that OK?
Dan Carpenter<[email protected]> 在 2026年6月2日 周二 16:02 写道:
Hello Kairui Song,
Commit 128460ca95cf ("mm, swap: add support for stable large allocation in swap cache directly") from May 17, 2026 (linux-next), leads to the following Smatch static checker warning:
mm/swap_state.c:434 __swap_cache_alloc() error: we previously assumed 'mpol' could be null (see line 433)
mm/swap_state.c
403 static struct folio *__swap_cache_alloc(struct swap_cluster_info *ci,
404 swp_entry_t targ_entry, gfp_t gfp, 405 unsigned int order, struct vm_fault *vmf, 406 struct mempolicy *mpol, pgoff_t ilx) 407 {
408 int err;
409 swp_entry_t entry;
410 struct folio *folio;
411 void *shadow = NULL;
412 unsigned short memcg_id;
413 unsigned long address, nr_pages = 1UL << order;
414 struct vm_area_struct *vma = vmf ? vmf->vma : NULL;
415
416 VM_WARN_ON_ONCE(nr_pages > SWAPFILE_CLUSTER);
417 entry.val = round_down(targ_entry.val, nr_pages);
418
419 /* Check if the slot and range are available, skip allocation if not */
420 spin_lock(&ci->lock);
421 err = __swap_cache_add_check(ci, targ_entry, nr_pages, NULL, NULL);
422 spin_unlock(&ci->lock);
423 if (unlikely(err))
424 return ERR_PTR(err);
425
426 /*
427 * Limit THP gfp. The limitation is a no-op for typical
428 * GFP_HIGHUSER_MOVABLE but matters for shmem.
429 */
430 if (order)
431 gfp = thp_shmem_limit_gfp_mask(vma_thp_gfp_mask(vma), gfp);
432
433 if (mpol || !vmf) {
^^^^
imagine mpol is mpol is NULL and so is vmf.
--> 434 folio = folio_alloc_mpol(gfp, order, mpol, ilx, numa_node_id()); ^^^^ We can't pass a NULL mpol to folio_alloc_mpol() or it will crash. Unless
mpol can't be NULL when CONFIG_NUMA is disabled? I think mpol can be NULL
from do_swap_page() so this looks pretty suspect.
435 } else {
436 address = round_down(vmf->address, PAGE_SIZE << order); 437 folio = vma_alloc_folio(gfp, order, vmf->vma, address); 438 }
439 if (unlikely(!folio))
This email is a free service from the Smatch-CI project [smatch.sf.net].
regards,
dan carpenter
Quoted 8 lines · kasong(宋恺睿)
Hello Dan
Currently, no caller in kernel will pass NULL mpol with NULL vmf with
CONFIG_NUMA. Dodo_swap_pagealways have vmf set, and shmem always have either vmf or mpol set (unless !CONFIG_NUMA).I used to have a
WARN_ONfor this but some AI reviewers complained that theWARN_ONis redundant so I removed it. I can add that back later just to be safer, I will add a Reported-by for you. Is that OK?
Nah, don't worry. If it's a non-NUMA thing then that makes sense.
I was hoping we could just remove the !vmf part fo the check.
regards,
dan carpenter