kasan: Avoid sleepable page allocation from atomic context

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 <[email protected]>

Cc [email protected]

Fixes 3c5c3cfb9ef4 ("kasan: support backing vmalloc space with real shadow memory")

Signed-off-by Alexander Gordeev <[email protected]>

---
 mm/kasan/shadow.c | 76 ++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 62 insertions(+), 14 deletions(-)

mm/kasan/shadow.c +46 -11

index 88d1c9dcb507..2bf00bf7e545 100644
@@ -292,33 +292,83 @@ 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_pages, nr_populated = 0, nr_total = PFN_UP(end - start);
+	struct vmalloc_populate_data data;
+	int ret = 0;
+
+	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) {
+			ret = -ENOMEM;
+			break;
+		}
+
+		data.start = start;

mm/kasan/shadow.c +6 -3

+	free_page((unsigned long)data.pages);
+
+	return ret;
+}
+
 int kasan_populate_vmalloc(unsigned long addr, unsigned long size)
 {
 	unsigned long shadow_start, shadow_end;

@@ -348,9 +398,7 @@ 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;
 

Signature · 1 line

2.45.2

Thanks for the update, but I don't think nr_populated is sufficient here. If nr_populated in the last iteration is smaller than its value in any previous iteration, it could lead to a memory leak.

That's why I suggested (PAGE_SIZE / sizeof(data.pages[0])). ...but on second thought maybe touching the whole array is not efficient either.

If this ends up making things complicated probably we should just merge v6 instead (v6 looks good)? micro-optimizing vmalloc shadow memory population doesn't seem worth it if it comes at the cost of complexity :)

Signature · 2 lines

Cheers,
Harry / Hyeonggon

Quoted 7 lines · Harry Yoo

Thanks for the update, but I don't think nr_populated is sufficient here. If nr_populated in the last iteration is smaller than its value in any previous iteration, it could lead to a memory leak.

That's why I suggested (PAGE_SIZE / sizeof(data.pages[0])). ...but on second thought maybe touching the whole array is not efficient either.

Yes, I did not like it and wanted to limit the number of pages, but did not realize that using nr_populated still could produce leaks. In addition I could simply do:

	max_populted = max(max_populted, nr_populated);
	...
	free_pages_bulk(data.pages, max_populated);

Quoted 3 lines

If this ends up making things complicated probably we should just merge v6 instead (v6 looks good)? micro-optimizing vmalloc shadow memory population doesn't seem worth it if it comes at the cost of complexity :)

v6 is okay, except that in v7 I use break instead of return:

	ret = apply_to_page_range(...);
	if (ret)
		break;

and as result can call the final:

	free_page((unsigned long)data.pages);