Re: [PATCH v2 3/5] mm/vmalloc: undo partial mappings inside the mapping functions

[Hao Ge <hao.ge@linux.dev>]

Hi Uladzislau and Suren

On 2026/10/20 23:08, Uladzislau Rezki wrote:
> On Tue, Oct 20, 2026 at 09:27:14PM +0800, Hao Ge wrote:
>>
>> On 10/20/26 18:03, Hao Ge wrote:
>>> Hi Suren and Ulad
>>>
>>> Sashiko reported several issues, and they all look real.
>>> I have some ideas for fixing the first few;
>>> of course, please correct me if I get anything wrong.
>>>
>>> I'd like to discuss the last issue with you all.
>>>
>>>
>>> On 2026/10/20 13:45, sashiko-bot@kernel.org wrote:
>>>> Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
>>>> - [High] mm/vmalloc: infinite loop in vunmap_pte_range() on unaligned error rollback
>>>> - [High] mm/vmalloc: __vmap_pages_range_noflush() calls cond_resched() in atomic context
>>>> - [Medium] mm/vmalloc: vmap_page_range() calls flush_cache_vmap() before error rollback
>>>> --
>>>>
>>>> commit 043026437ba63b051df5d14470fe69ce8fa8825c
>>>> Author: Hao Ge <hao.ge@linux.dev>
>>>>
>>>> mm/vmalloc: undo partial mappings inside the mapping functions
>>>>
>>>> This commit adds rollback logic using __vunmap_range_noflush() and
>>>> kmsan_vunmap_range_noflush() into vmap_page_range(),
>>>> __vmap_pages_range_noflush(), and vmap_pages_range_noflush() when page
>>>> table installation or KMSAN metadata setup fails. It also adds range and
>>>> page_shift validation in __vmap_pages_range_noflush().
>>>>
>>>>> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
>>>>> index 44e0eb32708f1..75aa5acf55660 100644
>>>>> --- a/mm/vmalloc.c
>>>>> +++ b/mm/vmalloc.c
>>>>> @@ -363,6 +363,9 @@ int vmap_page_range(unsigned long addr, unsigned long end,
>>>>>   	if (!err)
>>>>>   		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
>>>>>   					       ioremap_max_page_shift);
>>>>> +	if (err)
>>>>> +		__vunmap_range_noflush(addr, end);
>>>>> +
>>>>>   	return err;
>>>>>   }
>>>> [Severity: High]
>>>> What happens here when vmap_range_noflush() returns -EINVAL for a range
>>>> where end - addr is not PAGE_SIZE-aligned?
>>>>
>>>> In vmap_pte_range(), commit a739e6b557af09 added a check to bail out on
>>>> unaligned ranges specifically to avoid an infinite loop:
>>>>
>>>> mm/vmalloc.c:vmap_pte_range() {
>>>> 	...
>>>> 	if (WARN_ON_ONCE(!PAGE_ALIGNED(end - addr)))
>>>> 		return -EINVAL;
>>>> }
>>>>
>>>> This can be reached on 64KB page kernels via pci_remap_iospace() ->
>>>> vmap_page_range() with 16KB/32KB/48KB PCI I/O ranges (for example via
>>>> loongson_pci_probe() or acpi_pci_root_remap_iospace()).
>>>>
>>>> When vmap_pte_range() returns -EINVAL, vmap_page_range() now calls
>>>> __vunmap_range_noflush(addr, end) with the unaligned range. Whenever a PMD
>>>> in [addr, end) is already populated with a PTE table (such as the
>>>> PCI_IOBASE PMD populated during early boot by reserve_pio_range(), or an
>>>> earlier PMD mapped in the same call), __vunmap_range_noflush() descends
>>>> into vunmap_pte_range(), which has no alignment check:
>>>>
>>>> mm/vmalloc.c:vunmap_pte_range() {
>>>> 	...
>>>> 	do {
>>>> 		...
>>>> 			ptent = ptep_get_and_clear(&init_mm, addr, pte);
>>>> 		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
>>>> 	} while (pte += (size >> PAGE_SHIFT), addr += size, addr != end);
>>>> }
>>>>
>>>> Because end - addr is not a multiple of PAGE_SIZE, won't addr step past end
>>>> and loop infinitely, clearing unrelated kernel PTEs and corrupting memory
>>>> past the PTE table?
>>>>
>>> That sounds reasonable. I will add a separate patch in the next version to fix
>>> this problem.
>>>
>>>> [Severity: Medium]
>>>> Also in vmap_page_range(), flush_cache_vmap(addr, end) is called
>>>> unconditionally before checking err and before rolling back with
>>>> __vunmap_range_noflush(addr, end):
>>>>
>>>> mm/vmalloc.c:vmap_page_range() {
>>>> 	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
>>>> 				 ioremap_max_page_shift);
>>>> 	flush_cache_vmap(addr, end);
>>>> 	if (!err)
>>>> 		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
>>>> 					       ioremap_max_page_shift);
>>>> 	if (err)
>>>> 		__vunmap_range_noflush(addr, end);
>>>>
>>>> 	return err;
>>>> }
>>>>
>>>> On UML, flush_cache_vmap is defined as flush_tlb_kernel_range:
>>>>
>>>> arch/um/include/asm/cacheflush.h:
>>>> #define flush_cache_vmap flush_tlb_kernel_range
>>>>
>>>> When vmap_range_noflush() fails partway through or
>>>> kmsan_ioremap_page_range() fails, flush_cache_vmap() runs while the
>>>> partially installed PTEs are still pte_present(), causing
>>>> update_pte_range() in arch/um/kernel/tlb.c to map them into the host OS via
>>>> ops->mmap() and clear _PAGE_NEEDSYNC:
>>>>
>>>> arch/um/kernel/tlb.c:update_pte_range() {
>>>> 	...
>>>> 		if (pte_present(*pte)) {
>>>> 			...
>>>> 			ret = ops->mmap(ops->mm_idp, addr, PAGE_SIZE,
>>>> 					prot, fd, offset);
>>>> 		} else
>>>> 			ret = ops->unmap(ops->mm_idp, addr, PAGE_SIZE);
>>>>
>>>> 		*pte = pte_mkuptodate(*pte);
>>>> }
>>>>
>>>> When __vunmap_range_noflush(addr, end) then clears the software PTEs
>>>> without a subsequent TLB flush or unmap notification, won't those host OS
>>>> mmap mappings be left behind? Should flush_cache_vmap(addr, end) only be
>>>> called after both vmap_range_noflush() and kmsan_ioremap_page_range()
>>>> succeed?
>>>>
>>> That sounds reasonable. I will add this logic in the next version.
>>>
>>>>> @@ -683,27 +686,35 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
>>>>>   		pgprot_t prot, struct page **pages, unsigned int page_shift)
>>>>>   {
>>>>>   	unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
>>>>> +	unsigned long start = addr;
>>>>> +	int err = 0;
>>>>> +
>>>>> +	if (WARN_ON_ONCE(addr >= end))
>>>>> +		return -EINVAL;
>>>>>   	if (WARN_ON_ONCE(page_shift < PAGE_SHIFT))
>>>>>   		return -EINVAL;
>>>>>   	if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
>>>>> -			page_shift == PAGE_SHIFT)
>>>>> -		return vmap_small_pages_range_noflush(addr, end, prot, pages);
>>>>> -
>>>>> -	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
>>>>> -		int err;
>>>>> -
>>>>> -		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
>>>>> +			page_shift == PAGE_SHIFT) {
>>>>> +		err = vmap_small_pages_range_noflush(addr, end, prot, pages);
>>>>> +	} else {
>>>>> +		for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
>>>>> +			err = vmap_range_noflush(addr, addr + (1UL << page_shift),
>>>>>   					page_to_phys(pages[i]), prot,
>>>>>   					page_shift);
>>>>> -		if (err)
>>>>> -			return err;
>>>>> +			if (err)
>>>>> +				break;
>>>>> -		addr += 1UL << page_shift;
>>>>> +			addr += 1UL << page_shift;
>>>>> +		}
>>>>>   	}
>>>>> -	return 0;
>>>>> +	/* Undo the PTEs installed before the failure. */
>>>>> +	if (err)
>>>>> +		__vunmap_range_noflush(start, end);
>>>>> +
>>>>> +	return err;
>>>>>   }
>>>> [Severity: High]
>>>> Similarly here, __vmap_pages_range_noflush() checks addr >= end, but does
>>>> not check PAGE_ALIGNED(addr) or PAGE_ALIGNED(end).
>>>>
>>>> If a caller passes a sub-page or unaligned range (for example, 0 < end -
>>>> addr < PAGE_SIZE), nr = (end - addr) >> PAGE_SHIFT evaluates to 0, so the
>>>> huge-page path still maps nothing and returns 0, while the small-page path
>>>> loops infinitely in vmap_pages_pte_range(). And if the huge-page path fails
>>>> on an unaligned range, __vunmap_range_noflush(start, end) will hit the same
>>>> infinite loop in vunmap_pte_range().
>>>>
>>>> Should both vmap_page_range() and __vmap_pages_range_noflush() validate
>>>> PAGE_ALIGNED(addr) && PAGE_ALIGNED(end) upfront before attempting to map or
>>>> unmap?
>>>>
>>> I think We need to check both PAGE_ALIGNED(addr) and PAGE_ALIGNED(end).
>>> As an example: if addr is PAGE-aligned, end is not, and end - addr < PAGE_SIZE,
>>> then nr becomes zero. On the huge page path, we simply return 0 with no work performed.
>>>
>>> At the same time, vmap_pages_pte_range() needs the same guard as commit
>>> a739e6b557af ("mm: vmalloc: WARN_ON if mapping size is not PAGE_SIZE aligned"),
>>> to protect against unaligned ranges.
>>>
>>> So will do next version.
>>>
>>>>>   int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
>>>>> @@ -715,7 +726,16 @@ int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
>>>>>   	if (ret)
>>>>>   		return ret;
>>>>> -	return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
>>>>> +
>>>>> +	ret = __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
>>>>> +	/*
>>>>> +	 * The page tables undo themselves on failure. Tear down the
>>>>> +	 * metadata that was fully set up before the mapping failed.
>>>>> +	 */
>>>>> +	if (ret)
>>>>> +		kmsan_vunmap_range_noflush(addr, end);
>>>>> +
>>>>> +	return ret;
>>>>>   }
>>>> [Severity: High]
>>>> Can calling __vunmap_range_noflush() and kmsan_vunmap_range_noflush() here
>>>> trigger a sleeping-in-atomic-context bug when __vmalloc() or kvmalloc() is
>>>> called with GFP_ATOMIC or GFP_NOWAIT?
>>>>
>>>> __vmalloc_area_node() supports non-blocking allocations by wrapping
>>>> __vmap_pages_range() in memalloc_apply_gfp_scope(gfp_mask):
>>>>
>>>> mm/vmalloc.c:__vmalloc_area_node() {
>>>> 	...
>>>> 	flags = memalloc_apply_gfp_scope(gfp_mask);
>>>> 	do {
>>>> 		ret = __vmap_pages_range(addr, addr + size, prot, area->pages,
>>>> 				page_shift, nested_gfp);
>>>> 	} while ((gfp_mask & __GFP_NOFAIL) && (ret < 0));
>>>> 	memalloc_restore_gfp_scope(flags);
>>>> }
>>>>
>>>> When a GFP_ATOMIC or GFP_NOWAIT allocation runs in atomic context (such as
>>>> under a spinlock, preempt_disable(), or rcu_read_lock()) and page table
>>>> allocation fails in vmap_small_pages_range_noflush() after at least one PMD
>>>> in [start, end) is populated (or after kmsan_vmap_pages_range_noflush()
>>>> populated the KMSAN shadow/origin ranges), __vmap_pages_range_noflush() and
>>>> kmsan_vunmap_range_noflush() synchronously call __vunmap_range_noflush().
>>>>
>>>> __vunmap_range_noflush() walks the populated PMDs via vunmap_pmd_range(),
>>>> which unconditionally calls cond_resched():
>>>>
>>>> mm/vmalloc.c:vunmap_pmd_range() {
>>>> 	...
>>>> 	do {
>>>> 		next = pmd_addr_end(addr, end);
>>>> 		if (pmd_none_or_clear_bad(pmd))
>>>> 			continue;
>>>> 		vunmap_pte_range(pmd, addr, next, mask);
>>>>
>>>> 		cond_resched();
>>>> 	} while (pmd++, addr = next, addr != end);
>>>> }
>>>>
>>>> Commit 9c47753167a6 ("mm/vmalloc: defer freeing partly initialized
>>>> vm_struct") deferred error-path cleanup in __vmalloc_area_node() to a
>>>> workqueue specifically because unmapping on failure can sleep. Doesn't
>>>> calling __vunmap_range_noflush() synchronously inside
>>>> __vmap_pages_range_noflush() and vmap_pages_range_noflush() reintroduce
>>>> cond_resched() in atomic context?
>>>>
>>> This issue also appears real. One possible approach is to plumb the caller's gfp_mask into vunmap_pmd_range()
>>> to conditionally skip cond_resched().
>>> Suren, Ulad, do you have any preferred or cleaner way to handle this?
>>
>>
>> Alternatively, we can follow the approach from commit 9c47753167a6
>> ("mm/vmalloc: defer freeing partly initialized vm_struct"):
>>
>> defer the cleanup to a workqueue when |gfp_mask| has GFP_ATOMIC or
>> GFP_NOWAIT，and perform cleanup synchronously
>>
>> if we are already in process context. Anyway, I will look into this further
>> tomorrow.
>>
> I will check the AI concern.
> 

Thank you very much for the quick reply and for offering to check
the AI concern. Much appreciated!

Here is a summary of what I have found over the last few days.
I went through the history first. The cond_resched() in vunmap_pmd_range()
was added by commit e47110e90584 ("mm/vunmap: add cond_resched() in vunmap_pmd_range"),
modeled after zap_pte_range(). It fixed a real soft lockup: on a non-preemptible
kernel, unmapping a large persistent-memory namespace (hence a large I/O map region)
kept a CPU busy for 23 seconds inside the walk:

	watchdog: BUG: soft lockup - CPU#49 stuck for 23s! [ndctl:50777]

Then I looked at how vfree() handles atomic contexts: it defers the
whole free to a workqueue from interrupt context (vfree_atomic()),
and 9c47753167a6 does the same for a partly initialized vm_struct.

Unfortunately I do not think deferral works for this rollback case.
A deferred rollback returns with the PTEs still installed, and if
that range is mapped again before the workqueue runs, the mapping
hits the leftovers: a warning and -EBUSY on the small-page path, a
BUG() on the huge path.

So I kept it simple: plumb the caller's gfp_mask down to the unmap
walk and call cond_resched() conditionally, like
commit 660654f90e7f ("mm/vmalloc.c: add a schedule point to vmalloc()")
did:
	if (gfp_mask & __GFP_WAIT)
		cond_resched();

Of course, this may not be the perfect solution either. I would
gladly hear your thoughts on it.

Thanks
Best Regards
Hao

> --
> Uladzislau Rezki