[PATCH] book3s64/radix : Align section vmemmap start address to PAGE_SIZE

[Donet Tom]

On 3/6/25 9:41 AM, Aneesh Kumar K.V wrote:
> Donet Tom <donettom at linux.ibm.com> writes:
>
>> On 3/3/25 18:32, Aneesh Kumar K.V wrote:
>>> Donet Tom <donettom at linux.ibm.com> writes:
>>>
>>>> A vmemmap altmap is a device-provided region used to provide
>>>> backing storage for struct pages. For each namespace, the altmap
>>>> should belong to that same namespace. If the namespaces are
>>>> created unaligned, there is a chance that the section vmemmap
>>>> start address could also be unaligned. If the section vmemmap
>>>> start address is unaligned, the altmap page allocated from the
>>>> current namespace might be used by the previous namespace also.
>>>> During the free operation, since the altmap is shared between two
>>>> namespaces, the previous namespace may detect that the page does
>>>> not belong to its altmap and incorrectly assume that the page is a
>>>> normal page. It then attempts to free the normal page, which leads
>>>> to a kernel crash.
>>>>
>>>> In this patch, we are aligning the section vmemmap start address
>>>> to PAGE_SIZE. After alignment, the start address will not be
>>>> part of the current namespace, and a normal page will be allocated
>>>> for the vmemmap mapping of the current section. For the remaining
>>>> sections, altmaps will be allocated. During the free operation,
>>>> the normal page will be correctly freed.
>>>>
>>>> Without this patch
>>>> ==================
>>>> NS1 start               NS2 start
>>>>    _________________________________________________________
>>>> |         NS1               |            NS2              |
>>>>    ---------------------------------------------------------
>>>> | Altmap| Altmap | .....|Altmap| Altmap | ...........
>>>> |  NS1  |  NS1   |      | NS2  |  NS2   |
>>>>
>>>                               ^^^ this should be allocated in ram?
>>>
>> Yes, it should be allocated from RAM. However, in the current
>> implementation, an altmap page gets allocated. This is because the
>> NS2 vmemmap section's start address is unaligned. There is an
>> altmap_cross_boundary() check. Here, from the vmemmap section
>> start, we identify the namespace start and check if the namespace start
>> is within the boundary. Since it is within the boundary, it returns false,
>> causing an altmap page to be allocated. During the PTE update, the
>> vmemmap start address is aligned down to PAGE_SIZE, and the PTE is
>> updated. As a result, the altmap page is shared between the current
>> and previous namespaces.
>>
>> If we had aligned the vmemmap start address, the
>> altmap_cross_boundary() function would return true because the
>> vmemmap section's start address belongs to the previous
>> namespace. Therefore normal page gets allocated. During the
>> PTE set operation, since the address is already aligned, the
>> PTE will updated.
>>
> So the nvdimm driver should ensure that alignment right? I assume other things
> will also require that to be properly aligned.?

  #cat /proc/iomem
00000000-63ffffffff : System RAM
40340000000-403401fffff : namespace1.0
40340200000-403a0ffffff : dax1.0
403a1000000-403a11fffff : namespace1.1
403a1200000-40401ffffff : dax1.1
40402000000-404021fffff : namespace1.2
40402200000-40462ffffff : dax1.2
40463000000-404631fffff : namespace1.3
40463200000-404c3ffffff : dax1.3
  #

I have created 4 namespaces with a size of 1552M. As you can see, the 
start of
namespace1.0 is 1G aligned, while namespace1.1, namespace1.2, and 
namespace1.3
are not 1G aligned. If I had created the namespace with a size of 1536M 
(1.5G), then
all the namespaces would have started 1G aligned.

I believe that based on the size we are requesting, the namespaces 
alignments are
being created. They do not always need to be 1G aligned.

Now, if we calculate the vmemmap start for namespace1.1..

Phy start - 0x403a1000000
pfn start - 0x403a1000000 / PAGE_SIZE = 0x403a100

vmemmap start = 0xc00c000000000000 + (0x403a100 * 0x40)
                           =0xC00C000100E84000

This address is not page aligned. This will trigger this issue.

>
> -aneesh