[PATCH 1/1] device-dax: Correct pgoff align in dax_set_mapping()

[PATCH 1/1] device-dax: Correct pgoff align in dax_set_mapping()

[Kun(llfl)]

pgoff should be aligned using ALIGN_DOWN() instead of ALIGN(). Otherwise,
vmf->address not aligned to fault_size will be aligned to the next
alignment, that can result in memory failure getting the wrong address.

Fixes: b9b5777f09be ("device-dax: use ALIGN() for determining pgoff")
Signed-off-by: Kun(llfl) <llfl@linux.alibaba.com>
Tested-by: JianXiong Zhao <zhaojianxiong.zjx@alibaba-inc.com>
---
 drivers/dax/device.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/dax/device.c b/drivers/dax/device.c
index 9c1a729cd77e..6d74e62bbee0 100644
--- a/drivers/dax/device.c
+++ b/drivers/dax/device.c
@@ -86,7 +86,7 @@ static void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
 		nr_pages = 1;
 
 	pgoff = linear_page_index(vmf->vma,
-			ALIGN(vmf->address, fault_size));
+			ALIGN_DOWN(vmf->address, fault_size));
 
 	for (i = 0; i < nr_pages; i++) {
 		struct page *page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
-- 
2.43.0

Re: [PATCH 1/1] device-dax: Correct pgoff align in dax_set_mapping()

[Andrew Morton]

(cc's added)

On Fri, 27 Sep 2024 15:45:09 +0800 "Kun(llfl)" <llfl@linux.alibaba.com> wrote:

> pgoff should be aligned using ALIGN_DOWN() instead of ALIGN(). Otherwise,
> vmf->address not aligned to fault_size will be aligned to the next
> alignment, that can result in memory failure getting the wrong address.
> 
> Fixes: b9b5777f09be ("device-dax: use ALIGN() for determining pgoff")

That's quite an old change.  Can you suggest why it took this long to
be discovered?  What is your userspace doing to trigger this?

> Signed-off-by: Kun(llfl) <llfl@linux.alibaba.com>
> Tested-by: JianXiong Zhao <zhaojianxiong.zjx@alibaba-inc.com>
> ---
>  drivers/dax/device.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/dax/device.c b/drivers/dax/device.c
> index 9c1a729cd77e..6d74e62bbee0 100644
> --- a/drivers/dax/device.c
> +++ b/drivers/dax/device.c
> @@ -86,7 +86,7 @@ static void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
>  		nr_pages = 1;
>  
>  	pgoff = linear_page_index(vmf->vma,
> -			ALIGN(vmf->address, fault_size));
> +			ALIGN_DOWN(vmf->address, fault_size));
>  
>  	for (i = 0; i < nr_pages; i++) {
>  		struct page *page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
> -- 
> 2.43.0

Re: [PATCH 1/1] device-dax: Correct pgoff align in dax_set_mapping()

[Kun(llfl)]

That's a subtle situation that only can be observed in 
page_mapped_in_vma() after the page is page fault handled by 
dev_dax_huge_fault. Generally, there is little chance to perform 
page_mapped_in_vma in dev-dax's page unless in specific error injection 
to the dax device to trigger an MCE - memory-failure. In that case, 
page_mapped_in_vma() will be triggered to determine which task is 
accessing the failure address and kill that task in the end.


We used self-developed dax device (which is 2M aligned mapping) , to 
perform error injection to random address. It turned out that error 
injected to non-2M-aligned address was causing endless MCE until panic. 
Because page_mapped_in_vma() kept resulting wrong address and the task 
accessing the failure address was never killed properly:


[ 3783.719419] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3784.049006] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3784.049190] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3784.448042] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3784.448186] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3784.792026] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3784.792179] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3785.162502] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3785.162633] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3785.461116] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3785.461247] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3785.764730] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3785.764859] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3786.042128] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3786.042259] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3786.464293] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3786.464423] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3786.818090] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3786.818217] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered
[ 3787.085297] mce: Uncorrected hardware memory error in user-access at 
200c9742380
[ 3787.085424] Memory failure: 0x200c9742: recovery action for dax page: 
Recovered

It took us several weeks to pinpoint this problem,  but we eventually 
used bpftrace to trace the page fault and mce address and successfully 
identified the issue.

On 9/28/24 1:46 AM, Andrew Morton wrote:
> (cc's added)
>
> On Fri, 27 Sep 2024 15:45:09 +0800 "Kun(llfl)" <llfl@linux.alibaba.com> wrote:
>
>> pgoff should be aligned using ALIGN_DOWN() instead of ALIGN(). Otherwise,
>> vmf->address not aligned to fault_size will be aligned to the next
>> alignment, that can result in memory failure getting the wrong address.
>>
>> Fixes: b9b5777f09be ("device-dax: use ALIGN() for determining pgoff")
> That's quite an old change.  Can you suggest why it took this long to
> be discovered?  What is your userspace doing to trigger this?
>
>> Signed-off-by: Kun(llfl) <llfl@linux.alibaba.com>
>> Tested-by: JianXiong Zhao <zhaojianxiong.zjx@alibaba-inc.com>
>> ---
>>   drivers/dax/device.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/dax/device.c b/drivers/dax/device.c
>> index 9c1a729cd77e..6d74e62bbee0 100644
>> --- a/drivers/dax/device.c
>> +++ b/drivers/dax/device.c
>> @@ -86,7 +86,7 @@ static void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
>>   		nr_pages = 1;
>>   
>>   	pgoff = linear_page_index(vmf->vma,
>> -			ALIGN(vmf->address, fault_size));
>> +			ALIGN_DOWN(vmf->address, fault_size));
>>   
>>   	for (i = 0; i < nr_pages; i++) {
>>   		struct page *page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
>> -- 
>> 2.43.0

-- 
Best,
KUN

Re: [PATCH 1/1] device-dax: Correct pgoff align in dax_set_mapping()

[Joao Martins]

On 27/09/2024 18:46, Andrew Morton wrote:
> (cc's added)
> 
> On Fri, 27 Sep 2024 15:45:09 +0800 "Kun(llfl)" <llfl@linux.alibaba.com> wrote:
> 
>> pgoff should be aligned using ALIGN_DOWN() instead of ALIGN(). Otherwise,
>> vmf->address not aligned to fault_size will be aligned to the next
>> alignment, that can result in memory failure getting the wrong address.
>>
>> Fixes: b9b5777f09be ("device-dax: use ALIGN() for determining pgoff")
> 
> That's quite an old change.  Can you suggest why it took this long to
> be discovered?  What is your userspace doing to trigger this?
> 
Likely we never reproduce in production because we always pin device-dax regions
in the region align they provide (Qemu does similarly with prealloc in
hugetlb/file backed memory). I think this bug requires that we touch *unpinned*
device-dax regions unaligned to the device-dax selected alignment (page size
i.e. 4K/2M/1G)

>> Signed-off-by: Kun(llfl) <llfl@linux.alibaba.com>
>> Tested-by: JianXiong Zhao <zhaojianxiong.zjx@alibaba-inc.com>

Thanks a lot for the catch and sorry for the pain that it may have caused you:

	Reviewed-by: Joao Martins <joao.m.martins@oracle.com>

>> ---
>>  drivers/dax/device.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/dax/device.c b/drivers/dax/device.c
>> index 9c1a729cd77e..6d74e62bbee0 100644
>> --- a/drivers/dax/device.c
>> +++ b/drivers/dax/device.c
>> @@ -86,7 +86,7 @@ static void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
>>  		nr_pages = 1;
>>  
>>  	pgoff = linear_page_index(vmf->vma,
>> -			ALIGN(vmf->address, fault_size));
>> +			ALIGN_DOWN(vmf->address, fault_size));
>>  
>>  	for (i = 0; i < nr_pages; i++) {
>>  		struct page *page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
>> -- 
>> 2.43.0