[PATCH v6 0/1] kasan: Avoid sleepable page allocation from atomic context

[PATCH v6 0/1] kasan: Avoid sleepable page allocation from atomic context

[Alexander Gordeev]

Hi All,

Chages since v5:
- full error message included into commit description

Chages since v4:
- unused pages leak is avoided

Chages since v3:
- pfn_to_virt() changed to page_to_virt() due to compile error

Chages since v2:
- page allocation moved out of the atomic context

Chages since v1:
- Fixes: and -stable tags added to the patch description

Thanks!

Alexander Gordeev (1):
  kasan: Avoid sleepable page allocation from atomic context

 mm/kasan/shadow.c | 77 ++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 63 insertions(+), 14 deletions(-)

-- 
2.45.2

[PATCH v6 1/1] kasan: Avoid sleepable page allocation from atomic context

[Alexander Gordeev]

apply_to_pte_range() enters the lazy MMU mode and then invokes
kasan_populate_vmalloc_pte() callback on each page table walk
iteration. However, the callback can go into sleep when trying
to allocate a single page, e.g. if an architecutre disables
preemption on lazy MMU mode enter.

On s390 if make arch_enter_lazy_mmu_mode() -> preempt_enable()
and arch_leave_lazy_mmu_mode() -> preempt_disable(), such crash
occurs:

[    0.663336] BUG: sleeping function called from invalid context at ./include/linux/sched/mm.h:321
[    0.663348] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 2, name: kthreadd
[    0.663358] preempt_count: 1, expected: 0
[    0.663366] RCU nest depth: 0, expected: 0
[    0.663375] no locks held by kthreadd/2.
[    0.663383] Preemption disabled at:
[    0.663386] [<0002f3284cbb4eda>] apply_to_pte_range+0xfa/0x4a0
[    0.663405] CPU: 0 UID: 0 PID: 2 Comm: kthreadd Not tainted 6.15.0-rc5-gcc-kasan-00043-gd76bb1ebb558-dirty #162 PREEMPT
[    0.663408] Hardware name: IBM 3931 A01 701 (KVM/Linux)
[    0.663409] Call Trace:
[    0.663410]  [<0002f3284c385f58>] dump_stack_lvl+0xe8/0x140
[    0.663413]  [<0002f3284c507b9e>] __might_resched+0x66e/0x700
[    0.663415]  [<0002f3284cc4f6c0>] __alloc_frozen_pages_noprof+0x370/0x4b0
[    0.663419]  [<0002f3284ccc73c0>] alloc_pages_mpol+0x1a0/0x4a0
[    0.663421]  [<0002f3284ccc8518>] alloc_frozen_pages_noprof+0x88/0xc0
[    0.663424]  [<0002f3284ccc8572>] alloc_pages_noprof+0x22/0x120
[    0.663427]  [<0002f3284cc341ac>] get_free_pages_noprof+0x2c/0xc0
[    0.663429]  [<0002f3284cceba70>] kasan_populate_vmalloc_pte+0x50/0x120
[    0.663433]  [<0002f3284cbb4ef8>] apply_to_pte_range+0x118/0x4a0
[    0.663435]  [<0002f3284cbc7c14>] apply_to_pmd_range+0x194/0x3e0
[    0.663437]  [<0002f3284cbc99be>] __apply_to_page_range+0x2fe/0x7a0
[    0.663440]  [<0002f3284cbc9e88>] apply_to_page_range+0x28/0x40
[    0.663442]  [<0002f3284ccebf12>] kasan_populate_vmalloc+0x82/0xa0
[    0.663445]  [<0002f3284cc1578c>] alloc_vmap_area+0x34c/0xc10
[    0.663448]  [<0002f3284cc1c2a6>] __get_vm_area_node+0x186/0x2a0
[    0.663451]  [<0002f3284cc1e696>] __vmalloc_node_range_noprof+0x116/0x310
[    0.663454]  [<0002f3284cc1d950>] __vmalloc_node_noprof+0xd0/0x110
[    0.663457]  [<0002f3284c454b88>] alloc_thread_stack_node+0xf8/0x330
[    0.663460]  [<0002f3284c458d56>] dup_task_struct+0x66/0x4d0
[    0.663463]  [<0002f3284c45be90>] copy_process+0x280/0x4b90
[    0.663465]  [<0002f3284c460940>] kernel_clone+0xd0/0x4b0
[    0.663467]  [<0002f3284c46115e>] kernel_thread+0xbe/0xe0
[    0.663469]  [<0002f3284c4e440e>] kthreadd+0x50e/0x7f0
[    0.663472]  [<0002f3284c38c04a>] __ret_from_fork+0x8a/0xf0
[    0.663475]  [<0002f3284ed57ff2>] ret_from_fork+0xa/0x38

Instead of allocating single pages per-PTE, bulk-allocate the
shadow memory prior to applying kasan_populate_vmalloc_pte()
callback on a page range.

Suggested-by: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: stable@vger.kernel.org
Fixes: 3c5c3cfb9ef4 ("kasan: support backing vmalloc space with real shadow memory")

Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
---
 mm/kasan/shadow.c | 77 ++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 63 insertions(+), 14 deletions(-)

diff --git a/mm/kasan/shadow.c b/mm/kasan/shadow.c
index 88d1c9dcb507..660cc2148575 100644
--- a/mm/kasan/shadow.c
+++ b/mm/kasan/shadow.c
@@ -292,30 +292,81 @@ void __init __weak kasan_populate_early_vm_area_shadow(void *start,
 {
 }
 
+struct vmalloc_populate_data {
+	unsigned long start;
+	struct page **pages;
+};
+
 static int kasan_populate_vmalloc_pte(pte_t *ptep, unsigned long addr,
-				      void *unused)
+				      void *_data)
 {
-	unsigned long page;
+	struct vmalloc_populate_data *data = _data;
+	struct page *page;
 	pte_t pte;
+	int index;
 
 	if (likely(!pte_none(ptep_get(ptep))))
 		return 0;
 
-	page = __get_free_page(GFP_KERNEL);
-	if (!page)
-		return -ENOMEM;
-
-	__memset((void *)page, KASAN_VMALLOC_INVALID, PAGE_SIZE);
-	pte = pfn_pte(PFN_DOWN(__pa(page)), PAGE_KERNEL);
+	index = PFN_DOWN(addr - data->start);
+	page = data->pages[index];
+	__memset(page_to_virt(page), KASAN_VMALLOC_INVALID, PAGE_SIZE);
+	pte = pfn_pte(page_to_pfn(page), PAGE_KERNEL);
 
 	spin_lock(&init_mm.page_table_lock);
 	if (likely(pte_none(ptep_get(ptep)))) {
 		set_pte_at(&init_mm, addr, ptep, pte);
-		page = 0;
+		data->pages[index] = NULL;
 	}
 	spin_unlock(&init_mm.page_table_lock);
-	if (page)
-		free_page(page);
+
+	return 0;
+}
+
+static inline void free_pages_bulk(struct page **pages, int nr_pages)
+{
+	int i;
+
+	for (i = 0; i < nr_pages; i++) {
+		if (pages[i]) {
+			__free_pages(pages[i], 0);
+			pages[i] = NULL;
+		}
+	}
+}
+
+static int __kasan_populate_vmalloc(unsigned long start, unsigned long end)
+{
+	unsigned long nr_populated, nr_pages, nr_total = PFN_UP(end - start);
+	struct vmalloc_populate_data data;
+	int ret;
+
+	data.pages = (struct page **)__get_free_page(GFP_KERNEL | __GFP_ZERO);
+	if (!data.pages)
+		return -ENOMEM;
+
+	while (nr_total) {
+		nr_pages = min(nr_total, PAGE_SIZE / sizeof(data.pages[0]));
+		nr_populated = alloc_pages_bulk(GFP_KERNEL, nr_pages, data.pages);
+		if (nr_populated != nr_pages) {
+			free_pages_bulk(data.pages, nr_populated);
+			free_page((unsigned long)data.pages);
+			return -ENOMEM;
+		}
+
+		data.start = start;
+		ret = apply_to_page_range(&init_mm, start, nr_pages * PAGE_SIZE,
+					  kasan_populate_vmalloc_pte, &data);
+		free_pages_bulk(data.pages, nr_pages);
+		if (ret)
+			return ret;
+
+		start += nr_pages * PAGE_SIZE;
+		nr_total -= nr_pages;
+	}
+
+	free_page((unsigned long)data.pages);
+
 	return 0;
 }
 
@@ -348,9 +399,7 @@ int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
 	shadow_start = PAGE_ALIGN_DOWN(shadow_start);
 	shadow_end = PAGE_ALIGN(shadow_end);
 
-	ret = apply_to_page_range(&init_mm, shadow_start,
-				  shadow_end - shadow_start,
-				  kasan_populate_vmalloc_pte, NULL);
+	ret = __kasan_populate_vmalloc(shadow_start, shadow_end);
 	if (ret)
 		return ret;
 
-- 
2.45.2

Re: [PATCH v6 1/1] kasan: Avoid sleepable page allocation from atomic context

[Harry Yoo]

On Thu, May 08, 2025 at 04:15:46PM +0200, Alexander Gordeev wrote:
> apply_to_pte_range() enters the lazy MMU mode and then invokes
> kasan_populate_vmalloc_pte() callback on each page table walk
> iteration. However, the callback can go into sleep when trying
> to allocate a single page, e.g. if an architecutre disables
> preemption on lazy MMU mode enter.
> 
> On s390 if make arch_enter_lazy_mmu_mode() -> preempt_enable()
> and arch_leave_lazy_mmu_mode() -> preempt_disable(), such crash
> occurs:
> 
> [    0.663336] BUG: sleeping function called from invalid context at ./include/linux/sched/mm.h:321
> [    0.663348] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 2, name: kthreadd
> [    0.663358] preempt_count: 1, expected: 0
> [    0.663366] RCU nest depth: 0, expected: 0
> [    0.663375] no locks held by kthreadd/2.
> [    0.663383] Preemption disabled at:
> [    0.663386] [<0002f3284cbb4eda>] apply_to_pte_range+0xfa/0x4a0
> [    0.663405] CPU: 0 UID: 0 PID: 2 Comm: kthreadd Not tainted 6.15.0-rc5-gcc-kasan-00043-gd76bb1ebb558-dirty #162 PREEMPT
> [    0.663408] Hardware name: IBM 3931 A01 701 (KVM/Linux)
> [    0.663409] Call Trace:
> [    0.663410]  [<0002f3284c385f58>] dump_stack_lvl+0xe8/0x140
> [    0.663413]  [<0002f3284c507b9e>] __might_resched+0x66e/0x700
> [    0.663415]  [<0002f3284cc4f6c0>] __alloc_frozen_pages_noprof+0x370/0x4b0
> [    0.663419]  [<0002f3284ccc73c0>] alloc_pages_mpol+0x1a0/0x4a0
> [    0.663421]  [<0002f3284ccc8518>] alloc_frozen_pages_noprof+0x88/0xc0
> [    0.663424]  [<0002f3284ccc8572>] alloc_pages_noprof+0x22/0x120
> [    0.663427]  [<0002f3284cc341ac>] get_free_pages_noprof+0x2c/0xc0
> [    0.663429]  [<0002f3284cceba70>] kasan_populate_vmalloc_pte+0x50/0x120
> [    0.663433]  [<0002f3284cbb4ef8>] apply_to_pte_range+0x118/0x4a0
> [    0.663435]  [<0002f3284cbc7c14>] apply_to_pmd_range+0x194/0x3e0
> [    0.663437]  [<0002f3284cbc99be>] __apply_to_page_range+0x2fe/0x7a0
> [    0.663440]  [<0002f3284cbc9e88>] apply_to_page_range+0x28/0x40
> [    0.663442]  [<0002f3284ccebf12>] kasan_populate_vmalloc+0x82/0xa0
> [    0.663445]  [<0002f3284cc1578c>] alloc_vmap_area+0x34c/0xc10
> [    0.663448]  [<0002f3284cc1c2a6>] __get_vm_area_node+0x186/0x2a0
> [    0.663451]  [<0002f3284cc1e696>] __vmalloc_node_range_noprof+0x116/0x310
> [    0.663454]  [<0002f3284cc1d950>] __vmalloc_node_noprof+0xd0/0x110
> [    0.663457]  [<0002f3284c454b88>] alloc_thread_stack_node+0xf8/0x330
> [    0.663460]  [<0002f3284c458d56>] dup_task_struct+0x66/0x4d0
> [    0.663463]  [<0002f3284c45be90>] copy_process+0x280/0x4b90
> [    0.663465]  [<0002f3284c460940>] kernel_clone+0xd0/0x4b0
> [    0.663467]  [<0002f3284c46115e>] kernel_thread+0xbe/0xe0
> [    0.663469]  [<0002f3284c4e440e>] kthreadd+0x50e/0x7f0
> [    0.663472]  [<0002f3284c38c04a>] __ret_from_fork+0x8a/0xf0
> [    0.663475]  [<0002f3284ed57ff2>] ret_from_fork+0xa/0x38
> 
> Instead of allocating single pages per-PTE, bulk-allocate the
> shadow memory prior to applying kasan_populate_vmalloc_pte()
> callback on a page range.
> 
> Suggested-by: Andrey Ryabinin <ryabinin.a.a@gmail.com>
> Cc: stable@vger.kernel.org
> Fixes: 3c5c3cfb9ef4 ("kasan: support backing vmalloc space with real shadow memory")
> 
> Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
> ---

FWIW, this patch looks good to me.
Reviewed-by: Harry Yoo <harry.yoo@oracle.com>

With a minor suggestion below.

>  mm/kasan/shadow.c | 77 ++++++++++++++++++++++++++++++++++++++---------
>  1 file changed, 63 insertions(+), 14 deletions(-)
> 
> diff --git a/mm/kasan/shadow.c b/mm/kasan/shadow.c
> index 88d1c9dcb507..660cc2148575 100644
> --- a/mm/kasan/shadow.c
> +++ b/mm/kasan/shadow.c
> @@ -292,30 +292,81 @@ void __init __weak kasan_populate_early_vm_area_shadow(void *start,

... snip ...

> +static int __kasan_populate_vmalloc(unsigned long start, unsigned long end)
> +{
> +	unsigned long nr_populated, nr_pages, nr_total = PFN_UP(end - start);
> +	struct vmalloc_populate_data data;
> +	int ret;
> +
> +	data.pages = (struct page **)__get_free_page(GFP_KERNEL | __GFP_ZERO);
> +	if (!data.pages)
> +		return -ENOMEM;
> +
> +	while (nr_total) {
> +		nr_pages = min(nr_total, PAGE_SIZE / sizeof(data.pages[0]));
> +		nr_populated = alloc_pages_bulk(GFP_KERNEL, nr_pages, data.pages);
> +		if (nr_populated != nr_pages) {
> +			free_pages_bulk(data.pages, nr_populated);
> +			free_page((unsigned long)data.pages);
> +			return -ENOMEM;
> +		}
> +
> +		data.start = start;
> +		ret = apply_to_page_range(&init_mm, start, nr_pages * PAGE_SIZE,
> +					  kasan_populate_vmalloc_pte, &data);
> +		free_pages_bulk(data.pages, nr_pages);

A minor suggestion:

I think this free_pages_bulk() can be moved outside the loop
(but with PAGE_SIZE / sizeof(data.pages[0]) instead of nr_pages),
because alloc_pages_bulk() simply skips allocating pages for any
non-NULL entries.

If some pages in the array were not used, it doesn't have to be freed;
on the next iteration of the loop alloc_pages_bulk() can skip
allocating pages for the non-NULL entries.

> +		if (ret)
> +			return ret;
> +
> +		start += nr_pages * PAGE_SIZE;
> +		nr_total -= nr_pages;
> +	}
> +
> +	free_page((unsigned long)data.pages);
> +
>  	return 0;
>  }

-- 
Cheers,
Harry / Hyeonggon

Re: [PATCH v6 1/1] kasan: Avoid sleepable page allocation from atomic context

[Alexander Gordeev]

On Fri, May 09, 2025 at 07:05:56PM +0900, Harry Yoo wrote:
> > +	while (nr_total) {
> > +		nr_pages = min(nr_total, PAGE_SIZE / sizeof(data.pages[0]));
> > +		nr_populated = alloc_pages_bulk(GFP_KERNEL, nr_pages, data.pages);
> > +		if (nr_populated != nr_pages) {
> > +			free_pages_bulk(data.pages, nr_populated);
> > +			free_page((unsigned long)data.pages);
> > +			return -ENOMEM;
> > +		}
> > +
> > +		data.start = start;
> > +		ret = apply_to_page_range(&init_mm, start, nr_pages * PAGE_SIZE,
> > +					  kasan_populate_vmalloc_pte, &data);
> > +		free_pages_bulk(data.pages, nr_pages);
> 
> A minor suggestion:
> 
> I think this free_pages_bulk() can be moved outside the loop
> (but with PAGE_SIZE / sizeof(data.pages[0]) instead of nr_pages),

Because we know the number of populated pages I think we could
use it instead of maximal (PAGE_SIZE / sizeof(data.pages[0])).

> because alloc_pages_bulk() simply skips allocating pages for any
> non-NULL entries.
> 
> If some pages in the array were not used, it doesn't have to be freed;
> on the next iteration of the loop alloc_pages_bulk() can skip
> allocating pages for the non-NULL entries.

Thanks for the suggestion! I will send an updated version.