2012/2/7, Catalin Marinas [off-list ref]:

On Mon, Jan 30, 2012 at 07:57:12AM +0000, bill4carson at gmail.com wrote:

quoted

+/* 2M and 16M hugepage linux ptes are stored in an array
+ *
+ * 2M hugepage
+ * ===========
+ * one linux pte caters to two HW ptes,
+ * so the maximum huge linux pte needed is 4096M/2M = 2048 entry
pointers.
+ * Two 4K page is used to store these entry pointers(2048 * 4 = 8192
bytes)
+ * in a two-dimension array, huge_2m_pte[2][1024].

Actually we only need to cover TASK_SIZE so for a 2:2 split you only
need half of the above.

I haven't polish here yet.
Yes, to be precise, this array depends on TASK_SIZE configuration.
The maximum array size is 1536 when TASK_SIZE is 0xc0000000
then kmalloc instead of __get_free_page can be used to allocate
1536 * sizeof(pte_t *). In this way VA[31:21] could be used as index
to get rid of row/column stuff.

I will keep this on V2 todo list :)

quoted

+ *
+ * How to find the hugepage linux pte corresponding to a specific address
?
+ * VA[31] is used as row index;
+ * VA[30:21] is used as column  index;

I haven't fully reviewed the code but can we not drop this row/column
set up and just use a VA[31:21] as the index?

Please see above reply.

quoted

+ *
+ * 16M hugepage
+ * ============
+ * one linux pte caters for one HW pte,

Actually that's a bit misleading as we need 16 consecutive pmd entries
for a supersection. So one Linux pmd caters for 16 HW pmds.

You are right, I will correct this in V2.

quoted

+ * so maxium huge linux pte needed is 4096M/16M = 256 entry pointers,
+ * 256 * 4 = 1024 bytes spaces is allocated to store these linux pte;
+ * this is a simple one-dimension array huge_16m_pte[256].
+ *
+ * VA[31:24] is used to index this array;

Maybe we should call them Linux pmd rather than pte in the comments? It
is less confusing (I know that the generic hugetlb code calls them
ptes).

OK, I will correct this in V2.

--
Catalin