Grygorii,

On Thursday 05 December 2013 01:48 PM, Strashko, Grygorii wrote:

Hi Tejun,

quoted

On Thu, Dec 05, 2013 at 06:35:00PM +0200, Grygorii Strashko wrote:

quoted

quoted

quoted

+#define memblock_virt_alloc_align(x, align) \
+  memblock_virt_alloc_try_nid(x, align, BOOTMEM_LOW_LIMIT, \
+                               BOOTMEM_ALLOC_ACCESSIBLE, MAX_NUMNODES)

Also, do we really need this align variant separate when the caller
can simply specify 0 for the default?

Unfortunately Yes.
We need it to keep compatibility with bootmem/nobootmem
which don't handle 0 as default align value.

Hmm... why wouldn't just interpreting 0 to SMP_CACHE_BYTES in the
memblock_virt*() function work?

Problem is not with memblock_virt*(). The issue will happen in case if
memblock or nobootmem are disabled in below code (memblock_virt*() is disabled).

+/* Fall back to all the existing bootmem APIs */
+#define memblock_virt_alloc(x) \
+       __alloc_bootmem(x, SMP_CACHE_BYTES, BOOTMEM_LOW_LIMIT)

which will be transformed to 
+/* Fall back to all the existing bootmem APIs */
+#define memblock_virt_alloc(x, align) \
+       __alloc_bootmem(x, align, BOOTMEM_LOW_LIMIT)

and used as

memblock_virt_alloc(size, 0);

so, by default bootmem code will use 0 as default alignment and not SMP_CACHE_BYTES
and that is wrong.

Looks like you didn't understood the suggestion completely.
The fall back inline will look like below .....

static inline memblock_virt_alloc(x, align)
{
	if (align == 0)
		align = SMP_CACHE_BYTES
	__alloc_bootmem(x, align, BOOTMEM_LOW_LIMIT);
}

regards,
Santosh