include/linux/mm.h | 15 ++++++++++ mm/huge_memory.c | 72 +++++++++++++++++++++++++++++++++++----------- 2 files changed, 70 insertions(+), 17 deletions(-)
Based on Linus' master.
While working on improving vm_normal_page() and friends, I stumbled
over this issues: refcounted "normal" pages must not be marked
using pmd_special() / pud_special().
Fortunately, so far there doesn't seem to be serious damage.
This is only compile-tested so far. Still looking for an easy way to test
PMD/PUD mappings with DAX. Any tests I can easily run?
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Nico Pache <npache@redhat.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Dan Williams <dan.j.williams@intel.com>
David Hildenbrand (2):
mm/huge_memory: don't mark refcounted pages special in
vmf_insert_folio_pmd()
mm/huge_memory: don't mark refcounted pages special in
vmf_insert_folio_pud()
include/linux/mm.h | 15 ++++++++++
mm/huge_memory.c | 72 +++++++++++++++++++++++++++++++++++-----------
2 files changed, 70 insertions(+), 17 deletions(-)
base-commit: a9dfb7db96f7bc1f30feae673aab7fdbfbc94e9c
--
2.49.0
David Hildenbrand wrote:
> Based on Linus' master.
>
> While working on improving vm_normal_page() and friends, I stumbled
> over this issues: refcounted "normal" pages must not be marked
> using pmd_special() / pud_special().
>
> Fortunately, so far there doesn't seem to be serious damage.
>
> This is only compile-tested so far. Still looking for an easy way to test
> PMD/PUD mappings with DAX. Any tests I can easily run?
The way I test this I would not classify as "easy", it is a bit of a pain
to setup, but it is passing here:
[root@host ndctl]# meson test -C build --suite ndctl:dax
ninja: Entering directory `/root/git/ndctl/build'
[168/168] Linking target cxl/cxl
1/13 ndctl:dax / daxdev-errors.sh OK 14.30s
2/13 ndctl:dax / multi-dax.sh OK 2.89s
3/13 ndctl:dax / sub-section.sh OK 8.40s
4/13 ndctl:dax / dax-dev OK 0.06s
5/13 ndctl:dax / dax-ext4.sh OK 20.53s
6/13 ndctl:dax / dax-xfs.sh OK 20.34s
7/13 ndctl:dax / device-dax OK 11.67s
8/13 ndctl:dax / revoke-devmem OK 0.25s
9/13 ndctl:dax / device-dax-fio.sh OK 34.02s
10/13 ndctl:dax / daxctl-devices.sh OK 3.44s
11/13 ndctl:dax / daxctl-create.sh SKIP 0.32s exit status 77
12/13 ndctl:dax / dm.sh OK 1.33s
13/13 ndctl:dax / mmap.sh OK 85.12s
...ignore the SKIP, that seems to be caused by an acpi-einj regression.
However, how about not duplicating the internals of insert_pfn_p[mu]d()
with something like the below. Either way you can add:
Tested-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
base-commit: a9dfb7db96f7bc1f30feae673aab7fdbfbc94e9c
-- 8< --
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index d3e66136e41a..cce4456aa62b 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1372,9 +1372,9 @@ vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
return __do_huge_pmd_anonymous_page(vmf);
}
-static int insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
- pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
- pgtable_t pgtable)
+static int insert_pmd(struct vm_area_struct *vma, unsigned long addr,
+ pmd_t *pmd, pfn_t pfn, struct folio *folio,
+ pgprot_t prot, bool write, pgtable_t pgtable)
{
struct mm_struct *mm = vma->vm_mm;
pmd_t entry;
@@ -1397,9 +1397,7 @@ static int insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
}
entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
- if (pfn_t_devmap(pfn))
- entry = pmd_mkdevmap(entry);
- else
+ if (!folio)
entry = pmd_mkspecial(entry);
if (write) {
entry = pmd_mkyoung(pmd_mkdirty(entry));
@@ -1458,8 +1456,8 @@ vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
pfnmap_setup_cachemode_pfn(pfn_t_to_pfn(pfn), &pgprot);
ptl = pmd_lock(vma->vm_mm, vmf->pmd);
- error = insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write,
- pgtable);
+ error = insert_pmd(vma, addr, vmf->pmd, pfn, NULL, pgprot, write,
+ pgtable);
spin_unlock(ptl);
if (error && pgtable)
pte_free(vma->vm_mm, pgtable);
@@ -1496,9 +1494,8 @@ vm_fault_t vmf_insert_folio_pmd(struct vm_fault *vmf, struct folio *folio,
folio_add_file_rmap_pmd(folio, &folio->page, vma);
add_mm_counter(mm, mm_counter_file(folio), HPAGE_PMD_NR);
}
- error = insert_pfn_pmd(vma, addr, vmf->pmd,
- pfn_to_pfn_t(folio_pfn(folio)), vma->vm_page_prot,
- write, pgtable);
+ error = insert_pmd(vma, addr, vmf->pmd, pfn_to_pfn_t(folio_pfn(folio)),
+ folio, vma->vm_page_prot, write, pgtable);
spin_unlock(ptl);
if (error && pgtable)
pte_free(mm, pgtable);
@@ -1515,8 +1512,8 @@ static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
return pud;
}
-static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
- pud_t *pud, pfn_t pfn, bool write)
+static void insert_pud(struct vm_area_struct *vma, unsigned long addr,
+ pud_t *pud, pfn_t pfn, struct folio *folio, bool write)
{
struct mm_struct *mm = vma->vm_mm;
pgprot_t prot = vma->vm_page_prot;
@@ -1535,9 +1532,7 @@ static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
}
entry = pud_mkhuge(pfn_t_pud(pfn, prot));
- if (pfn_t_devmap(pfn))
- entry = pud_mkdevmap(entry);
- else
+ if (!folio)
entry = pud_mkspecial(entry);
if (write) {
entry = pud_mkyoung(pud_mkdirty(entry));
@@ -1581,7 +1576,7 @@ vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
pfnmap_setup_cachemode_pfn(pfn_t_to_pfn(pfn), &pgprot);
ptl = pud_lock(vma->vm_mm, vmf->pud);
- insert_pfn_pud(vma, addr, vmf->pud, pfn, write);
+ insert_pud(vma, addr, vmf->pud, pfn, NULL, write);
spin_unlock(ptl);
return VM_FAULT_NOPAGE;
@@ -1616,7 +1611,7 @@ vm_fault_t vmf_insert_folio_pud(struct vm_fault *vmf, struct folio *folio,
/*
* If there is already an entry present we assume the folio is
* already mapped, hence no need to take another reference. We
- * still call insert_pfn_pud() though in case the mapping needs
+ * still call insert_pud() though in case the mapping needs
* upgrading to writeable.
*/
if (pud_none(*vmf->pud)) {
@@ -1624,8 +1619,8 @@ vm_fault_t vmf_insert_folio_pud(struct vm_fault *vmf, struct folio *folio,
folio_add_file_rmap_pud(folio, &folio->page, vma);
add_mm_counter(mm, mm_counter_file(folio), HPAGE_PUD_NR);
}
- insert_pfn_pud(vma, addr, vmf->pud, pfn_to_pfn_t(folio_pfn(folio)),
- write);
+ insert_pud(vma, addr, vmf->pud, pfn_to_pfn_t(folio_pfn(folio)), folio,
+ write);
spin_unlock(ptl);
return VM_FAULT_NOPAGE;
Hi Dan, On 06.06.25 01:47, Dan Williams wrote: > David Hildenbrand wrote: >> Based on Linus' master. >> >> While working on improving vm_normal_page() and friends, I stumbled >> over this issues: refcounted "normal" pages must not be marked >> using pmd_special() / pud_special(). >> >> Fortunately, so far there doesn't seem to be serious damage. >> >> This is only compile-tested so far. Still looking for an easy way to test >> PMD/PUD mappings with DAX. Any tests I can easily run? > > The way I test this I would not classify as "easy", it is a bit of a pain > to setup, but it is passing here: I guess most of the instructions are in https://github.com/pmem/ndctl ? I would assume that we need to set aside some special dax area using early boot params (memmap=). Might come in handy in the future. > > [root@host ndctl]# meson test -C build --suite ndctl:dax > ninja: Entering directory `/root/git/ndctl/build' > [168/168] Linking target cxl/cxl > 1/13 ndctl:dax / daxdev-errors.sh OK 14.30s > 2/13 ndctl:dax / multi-dax.sh OK 2.89s > 3/13 ndctl:dax / sub-section.sh OK 8.40s > 4/13 ndctl:dax / dax-dev OK 0.06s > 5/13 ndctl:dax / dax-ext4.sh OK 20.53s > 6/13 ndctl:dax / dax-xfs.sh OK 20.34s > 7/13 ndctl:dax / device-dax OK 11.67s > 8/13 ndctl:dax / revoke-devmem OK 0.25s > 9/13 ndctl:dax / device-dax-fio.sh OK 34.02s > 10/13 ndctl:dax / daxctl-devices.sh OK 3.44s > 11/13 ndctl:dax / daxctl-create.sh SKIP 0.32s exit status 77 > 12/13 ndctl:dax / dm.sh OK 1.33s > 13/13 ndctl:dax / mmap.sh OK 85.12s > > ...ignore the SKIP, that seems to be caused by an acpi-einj regression. Thanks for running these tests! > > However, how about not duplicating the internals of insert_pfn_p[mu]d() > with something like the below. Either way you can add: I considered that, but I prefer the current end result where we cleanup the pmd_none() handling and not mess with folios and pfns at the same time. ... just like we do for insert_pfn() vs. insert_page(), I don't think these code paths should be unified. (we should do more sanity checks like validate_page_before_insert() etc., but that's something for another day :) ) > > Tested-by: Dan Williams <dan.j.williams@intel.com> > Reviewed-by: Dan Williams <dan.j.williams@intel.com> Thanks! Let me resend with the fixup squashed. -- Cheers, David / dhildenb
On 03.06.25 23:16, David Hildenbrand wrote: > Based on Linus' master. > > While working on improving vm_normal_page() and friends, I stumbled > over this issues: refcounted "normal" pages must not be marked > using pmd_special() / pud_special(). > > Fortunately, so far there doesn't seem to be serious damage. > > This is only compile-tested so far. Still looking for an easy way to test > PMD/PUD mappings with DAX. Any tests I can easily run? ... aaaand I should have waited for the cross compiles. s390x and loongarch are not happy about pud_mkhuge() and pfn_pud(). Need to fence folio_mk_pud() by CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD. -- Cheers, David / dhildenb
© 2016 - 2025 Red Hat, Inc.