Re: [PATCH v5 07/12] khugepaged: add mTHP support

From: Hugh Dickins
Date: Thu May 01 2025 - 08:59:33 EST

Next message: Alexandre Courbot: "[PATCH v2 02/21] rust: dma: expose the count and size of CoherentAllocation"
Previous message: Alexandre Courbot: "[PATCH v2 01/21] rust: devres: allow to borrow a reference to the resource's Device"
Next in thread: Andrew Morton: "Re: [PATCH v5 07/12] khugepaged: add mTHP support"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Mon, 28 Apr 2025, Nico Pache wrote:

> Introduce the ability for khugepaged to collapse to different mTHP sizes.
> While scanning PMD ranges for potential collapse candidates, keep track
> of pages in KHUGEPAGED_MIN_MTHP_ORDER chunks via a bitmap. Each bit
> represents a utilized region of order KHUGEPAGED_MIN_MTHP_ORDER ptes. If
> mTHPs are enabled we remove the restriction of max_ptes_none during the
> scan phase so we dont bailout early and miss potential mTHP candidates.
>
> After the scan is complete we will perform binary recursion on the
> bitmap to determine which mTHP size would be most efficient to collapse
> to. max_ptes_none will be scaled by the attempted collapse order to
> determine how full a THP must be to be eligible.
>
> If a mTHP collapse is attempted, but contains swapped out, or shared
> pages, we dont perform the collapse.
>
> Signed-off-by: Nico Pache <npache@xxxxxxxxxx>

There are locking errors in this patch. Let me comment inline below,
then at the end append the fix patch I'm using, to keep mm-new usable
for me. But that's more of an emergency rescue than a recommended fixup:
I don't much like your approach here, and hope it will change in v6.

> ---
> mm/khugepaged.c | 125 ++++++++++++++++++++++++++++++++++--------------
> 1 file changed, 88 insertions(+), 37 deletions(-)
>
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 6e67db86409a..3a846cd70c66 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1136,13 +1136,14 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> {
> LIST_HEAD(compound_pagelist);
> pmd_t *pmd, _pmd;
> - pte_t *pte;
> + pte_t *pte, mthp_pte;

I didn't wait to see the problem, just noticed that in the v4->v5
update, pte gets used at out_up_write, but there are gotos before
pte has been initialized. Declare pte = NULL here.

> pgtable_t pgtable;
> struct folio *folio;
> spinlock_t *pmd_ptl, *pte_ptl;
> int result = SCAN_FAIL;
> struct vm_area_struct *vma;
> struct mmu_notifier_range range;
> + unsigned long _address = address + offset * PAGE_SIZE;
>
> VM_BUG_ON(address & ~HPAGE_PMD_MASK);
>
> @@ -1158,12 +1159,13 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> *mmap_locked = false;
> }
>
> - result = alloc_charge_folio(&folio, mm, cc, HPAGE_PMD_ORDER);
> + result = alloc_charge_folio(&folio, mm, cc, order);
> if (result != SCAN_SUCCEED)
> goto out_nolock;
>
> mmap_read_lock(mm);
> - result = hugepage_vma_revalidate(mm, address, true, &vma, cc, HPAGE_PMD_ORDER);
> + *mmap_locked = true;
> + result = hugepage_vma_revalidate(mm, address, true, &vma, cc, order);
> if (result != SCAN_SUCCEED) {
> mmap_read_unlock(mm);
> goto out_nolock;
> @@ -1181,13 +1183,14 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> * released when it fails. So we jump out_nolock directly in
> * that case. Continuing to collapse causes inconsistency.
> */
> - result = __collapse_huge_page_swapin(mm, vma, address, pmd,
> - referenced, HPAGE_PMD_ORDER);
> + result = __collapse_huge_page_swapin(mm, vma, _address, pmd,
> + referenced, order);
> if (result != SCAN_SUCCEED)
> goto out_nolock;
> }
>
> mmap_read_unlock(mm);
> + *mmap_locked = false;
> /*
> * Prevent all access to pagetables with the exception of
> * gup_fast later handled by the ptep_clear_flush and the VM
> @@ -1197,7 +1200,7 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> * mmap_lock.
> */
> mmap_write_lock(mm);
> - result = hugepage_vma_revalidate(mm, address, true, &vma, cc, HPAGE_PMD_ORDER);
> + result = hugepage_vma_revalidate(mm, address, true, &vma, cc, order);
> if (result != SCAN_SUCCEED)
> goto out_up_write;
> /* check if the pmd is still valid */
> @@ -1208,11 +1211,12 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,

I spent a long time trying to work out why the include/linux/swapops.h:511
BUG is soon hit - the BUG which tells there's a migration entry left behind
after its folio has been unlocked.

In the patch at the end you'll see that I've inserted a check here, to
abort if the VMA following the revalidated "vma" is sharing the page table
(and so also affected by clearing *pmd). That turned out not to be the
problem (WARN_ONs inserted never fired in my limited testing), but it still
looks to me as if some such check is needed. Or I may be wrong, and
"revalidate" (a better place for the check) does actually check that, but
it wasn't obvious, and I haven't spent more time looking at it (but it did
appear to rule out the case of a VMA before "vma" sharing the page table).

> vma_start_write(vma);
> anon_vma_lock_write(vma->anon_vma);
>
> - mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
> - address + HPAGE_PMD_SIZE);
> + mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, _address,
> + _address + (PAGE_SIZE << order));

mmu_notifiers tend to be rather a mystery to me, so I've made no change
below, but it's not obvious whether it's correct clear the *pmd but only
notify of clearing a subset of that range: what's outside the range soon
gets replaced as it was, but is that good enough? I don't know.

> mmu_notifier_invalidate_range_start(&range);
>
> pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
> +
> /*
> * This removes any huge TLB entry from the CPU so we won't allow
> * huge and small TLB entries for the same virtual address to

The line I want to comment on does not appear in this diff context:
_pmd = pmdp_collapse_flush(vma, address, pmd);

That is appropriate for a THP occupying the whole range of the page table,
but is a surprising way to handle an "mTHP" of just some of its ptes:
I would expect you to be invalidating and replacing just those.

And that is the cause of the swapops:511 BUGs: "uninvolved" ptes are
being temporarily hidden, so not found when remove_migration_ptes()
goes looking for them.

This reliance on pmdp_collapse_flush() can be rescued, with stricter
locking (comment below); but I don't like it, and notice Jann has
picked up on it too. I hope v6 switches to handling ptes by ptes.

> @@ -1226,18 +1230,16 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> mmu_notifier_invalidate_range_end(&range);
> tlb_remove_table_sync_one();
>
> - pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);
> + pte = pte_offset_map_lock(mm, &_pmd, _address, &pte_ptl);
> if (pte) {
> - result = __collapse_huge_page_isolate(vma, address, pte, cc,
> - &compound_pagelist, HPAGE_PMD_ORDER);
> + result = __collapse_huge_page_isolate(vma, _address, pte, cc,
> + &compound_pagelist, order);
> spin_unlock(pte_ptl);
> } else {
> result = SCAN_PMD_NULL;
> }
>
> if (unlikely(result != SCAN_SUCCEED)) {
> - if (pte)
> - pte_unmap(pte);
> spin_lock(pmd_ptl);
> BUG_ON(!pmd_none(*pmd));
> /*
> @@ -1258,9 +1260,8 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> anon_vma_unlock_write(vma->anon_vma);

Phew, it's just visible there in the context. The anon_vma lock is what
keeps out racing lookups; so, that anon_vma_unlock_write() (and its
"All pages are isolated and locked" comment) is appropriate in the
HPAGE_PMD_SIZEd THP case, but has to be left until later for mTHP ptes.

But the anon_vma lock may well span a much larger range than the pte
lock, and the pmd lock certainly spans a much larger range than the
pte lock; so we really prefer to release anon_vma lock and pmd lock
as soon as is safe, and use pte lock in preference where possible.

>
> result = __collapse_huge_page_copy(pte, folio, pmd, _pmd,
> - vma, address, pte_ptl,
> - &compound_pagelist, HPAGE_PMD_ORDER);
> - pte_unmap(pte);
> + vma, _address, pte_ptl,
> + &compound_pagelist, order);
> if (unlikely(result != SCAN_SUCCEED))
> goto out_up_write;
>
> @@ -1270,25 +1271,42 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
> * write.
> */
> __folio_mark_uptodate(folio);
> - pgtable = pmd_pgtable(_pmd);
> -
> - _pmd = folio_mk_pmd(folio, vma->vm_page_prot);
> - _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
> -
> - spin_lock(pmd_ptl);
> - BUG_ON(!pmd_none(*pmd));
> - folio_add_new_anon_rmap(folio, vma, address, RMAP_EXCLUSIVE);
> - folio_add_lru_vma(folio, vma);
> - pgtable_trans_huge_deposit(mm, pmd, pgtable);
> - set_pmd_at(mm, address, pmd, _pmd);
> - update_mmu_cache_pmd(vma, address, pmd);
> - deferred_split_folio(folio, false);
> - spin_unlock(pmd_ptl);
> + if (order == HPAGE_PMD_ORDER) {
> + pgtable = pmd_pgtable(_pmd);
> + _pmd = folio_mk_pmd(folio, vma->vm_page_prot);
> + _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
> +
> + spin_lock(pmd_ptl);
> + BUG_ON(!pmd_none(*pmd));
> + folio_add_new_anon_rmap(folio, vma, _address, RMAP_EXCLUSIVE);
> + folio_add_lru_vma(folio, vma);
> + pgtable_trans_huge_deposit(mm, pmd, pgtable);
> + set_pmd_at(mm, address, pmd, _pmd);
> + update_mmu_cache_pmd(vma, address, pmd);
> + deferred_split_folio(folio, false);
> + spin_unlock(pmd_ptl);
> + } else { /* mTHP collapse */
> + mthp_pte = mk_pte(&folio->page, vma->vm_page_prot);
> + mthp_pte = maybe_mkwrite(pte_mkdirty(mthp_pte), vma);
> +
> + spin_lock(pmd_ptl);

I haven't changed that, but it is odd: yes, pmd_ptl will be required
when doing the pmd_populate(), but it serves no purpose here when
fiddling around with ptes in a disconnected page table.

> + folio_ref_add(folio, (1 << order) - 1);
> + folio_add_new_anon_rmap(folio, vma, _address, RMAP_EXCLUSIVE);
> + folio_add_lru_vma(folio, vma);
> + set_ptes(vma->vm_mm, _address, pte, mthp_pte, (1 << order));
> + update_mmu_cache_range(NULL, vma, _address, pte, (1 << order));
> +
> + smp_wmb(); /* make pte visible before pmd */
> + pmd_populate(mm, pmd, pmd_pgtable(_pmd));
> + spin_unlock(pmd_ptl);
> + }
>
> folio = NULL;
>
> result = SCAN_SUCCEED;

Somewhere around here it becomes safe for mTHP to anon_vma_unlock_write().

> out_up_write:
> + if (pte)
> + pte_unmap(pte);
> mmap_write_unlock(mm);
> out_nolock:
> *mmap_locked = false;
> @@ -1364,31 +1382,58 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> {
> pmd_t *pmd;
> pte_t *pte, *_pte;
> + int i;
> int result = SCAN_FAIL, referenced = 0;
> int none_or_zero = 0, shared = 0;
> struct page *page = NULL;
> struct folio *folio = NULL;
> unsigned long _address;
> + unsigned long enabled_orders;
> spinlock_t *ptl;
> int node = NUMA_NO_NODE, unmapped = 0;
> + bool is_pmd_only;
> bool writable = false;
> -
> + int chunk_none_count = 0;
> + int scaled_none = khugepaged_max_ptes_none >> (HPAGE_PMD_ORDER - KHUGEPAGED_MIN_MTHP_ORDER);
> + unsigned long tva_flags = cc->is_khugepaged ? TVA_ENFORCE_SYSFS : 0;
> VM_BUG_ON(address & ~HPAGE_PMD_MASK);
>
> result = find_pmd_or_thp_or_none(mm, address, &pmd);
> if (result != SCAN_SUCCEED)
> goto out;
>
> + bitmap_zero(cc->mthp_bitmap, MAX_MTHP_BITMAP_SIZE);
> + bitmap_zero(cc->mthp_bitmap_temp, MAX_MTHP_BITMAP_SIZE);
> memset(cc->node_load, 0, sizeof(cc->node_load));
> nodes_clear(cc->alloc_nmask);
> +
> + enabled_orders = thp_vma_allowable_orders(vma, vma->vm_flags,
> + tva_flags, THP_ORDERS_ALL_ANON);
> +
> + is_pmd_only = (enabled_orders == (1 << HPAGE_PMD_ORDER));
> +
> pte = pte_offset_map_lock(mm, pmd, address, &ptl);
> if (!pte) {
> result = SCAN_PMD_NULL;
> goto out;
> }
>
> - for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;
> - _pte++, _address += PAGE_SIZE) {
> + for (i = 0; i < HPAGE_PMD_NR; i++) {
> + /*
> + * we are reading in KHUGEPAGED_MIN_MTHP_NR page chunks. if
> + * there are pages in this chunk keep track of it in the bitmap
> + * for mTHP collapsing.
> + */
> + if (i % KHUGEPAGED_MIN_MTHP_NR == 0) {
> + if (chunk_none_count <= scaled_none)
> + bitmap_set(cc->mthp_bitmap,
> + i / KHUGEPAGED_MIN_MTHP_NR, 1);
> +
> + chunk_none_count = 0;
> + }
> +
> + _pte = pte + i;
> + _address = address + i * PAGE_SIZE;
> pte_t pteval = ptep_get(_pte);
> if (is_swap_pte(pteval)) {
> ++unmapped;
> @@ -1411,10 +1456,11 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> }
> }
> if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
> + ++chunk_none_count;
> ++none_or_zero;
> if (!userfaultfd_armed(vma) &&
> - (!cc->is_khugepaged ||
> - none_or_zero <= khugepaged_max_ptes_none)) {
> + (!cc->is_khugepaged || !is_pmd_only ||
> + none_or_zero <= khugepaged_max_ptes_none)) {
> continue;
> } else {
> result = SCAN_EXCEED_NONE_PTE;
> @@ -1510,6 +1556,7 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> address)))
> referenced++;
> }
> +
> if (!writable) {
> result = SCAN_PAGE_RO;
> } else if (cc->is_khugepaged &&
> @@ -1522,8 +1569,12 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> out_unmap:
> pte_unmap_unlock(pte, ptl);
> if (result == SCAN_SUCCEED) {
> - result = collapse_huge_page(mm, address, referenced,
> - unmapped, cc, mmap_locked, HPAGE_PMD_ORDER, 0);
> + result = khugepaged_scan_bitmap(mm, address, referenced, unmapped, cc,
> + mmap_locked, enabled_orders);
> + if (result > 0)
> + result = SCAN_SUCCEED;
> + else
> + result = SCAN_FAIL;
> }
> out:
> trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced,
> --
> 2.48.1

Fixes to 07/12 "khugepaged: add mTHP support".
But I see now that the first hunk is actually not to this 07/12, but to
05/12 "khugepaged: generalize __collapse_huge_page_* for mTHP support":
the mTHP check added in __collapse_huge_page_swapin() forgets to unmap
and unlock before returning, causing RCU imbalance warnings and lockups.
I won't separate it out here, let me leave that to you.

And I had other fixes to v4, which you've fixed differently in v5,
I haven't looked up which patch: where khugepaged_collapse_single_pmd()
does mmap_read_(un)lock() around collapse_pte_mapped_thp(). I dislike
your special use of result SCAN_ANY_PROCESS there, because mmap_locked
is precisely the tool for that job, so just lock and unlock without
setting *mmap_locked true (but I'd agree that mmap_locked is confusing,
and offhand wouldn't want to assert exactly what it means - does it
mean that mmap lock was *never* dropped, so "vma" is safe without
revalidation? depends on where it's used perhaps).

Hugh

---
mm/khugepaged.c | 32 +++++++++++++++++++++++++++-----
1 file changed, 27 insertions(+), 5 deletions(-)

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index c1c637dbcb81..2c814c239d65 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1054,6 +1054,8 @@ static int __collapse_huge_page_swapin(struct mm_struct *mm,

/* Dont swapin for mTHP collapse */
if (order != HPAGE_PMD_ORDER) {
+ pte_unmap(pte);
+ mmap_read_unlock(mm);
result = SCAN_EXCEED_SWAP_PTE;
goto out;
}
@@ -1136,7 +1138,7 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
{
LIST_HEAD(compound_pagelist);
pmd_t *pmd, _pmd;
- pte_t *pte, mthp_pte;
+ pte_t *pte = NULL, mthp_pte;
pgtable_t pgtable;
struct folio *folio;
spinlock_t *pmd_ptl, *pte_ptl;
@@ -1208,6 +1210,21 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
if (result != SCAN_SUCCEED)
goto out_up_write;

+ if (vma->vm_end < address + HPAGE_PMD_SIZE) {
+ struct vm_area_struct *next_vma = find_vma(mm, vma->vm_end);
+ /*
+ * We must not clear *pmd if it is used by the following VMA.
+ * Well, perhaps we could if it, and all following VMAs using
+ * this same page table, share the same anon_vma, and so are
+ * locked out together: but keep it simple for now (and this
+ * code might better belong in hugepage_vma_revalidate()).
+ */
+ if (next_vma && next_vma->vm_start < address + HPAGE_PMD_SIZE) {
+ result = SCAN_ADDRESS_RANGE;
+ goto out_up_write;
+ }
+ }
+
vma_start_write(vma);
anon_vma_lock_write(vma->anon_vma);

@@ -1255,15 +1272,17 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,

/*
* All pages are isolated and locked so anon_vma rmap
- * can't run anymore.
- */
- anon_vma_unlock_write(vma->anon_vma);
+ * can't run anymore - IF the entire extent has been isolated.
+ * anon_vma lock may cover a large area: better unlock a.s.a.p.
+ */
+ if (order == HPAGE_PMD_ORDER)
+ anon_vma_unlock_write(vma->anon_vma);

result = __collapse_huge_page_copy(pte, folio, pmd, _pmd,
vma, _address, pte_ptl,
&compound_pagelist, order);
if (unlikely(result != SCAN_SUCCEED))
- goto out_up_write;
+ goto out_unlock_anon_vma;

/*
* The smp_wmb() inside __folio_mark_uptodate() ensures the
@@ -1304,6 +1323,9 @@ static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
folio = NULL;

result = SCAN_SUCCEED;
+out_unlock_anon_vma:
+ if (order != HPAGE_PMD_ORDER)
+ anon_vma_unlock_write(vma->anon_vma);
out_up_write:
if (pte)
pte_unmap(pte);
--
2.43.0

Next message: Alexandre Courbot: "[PATCH v2 02/21] rust: dma: expose the count and size of CoherentAllocation"
Previous message: Alexandre Courbot: "[PATCH v2 01/21] rust: devres: allow to borrow a reference to the resource's Device"
Next in thread: Andrew Morton: "Re: [PATCH v5 07/12] khugepaged: add mTHP support"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]