On 3/6/25 9:41 AM, Aneesh Kumar K.V wrote:
Donet Tom <donet...@linux.ibm.com> writes:
On 3/3/25 18:32, Aneesh Kumar K.V wrote:
Donet Tom <donet...@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
