On 2016/8/5 5:24, Andrew Morton wrote:

quoted

it causes double align requirement for __get_vm_area_node() if parameter
size is power of 2 and VM_IOREMAP is set in parameter flags

it is fixed by handling the specail case manually due to lack of
get_count_order() for long parameter

...

--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c

@@ -1357,11 +1357,16 @@ static struct vm_struct *__get_vm_area_node(unsigned long size,
 {
 	struct vmap_area *va;
 	struct vm_struct *area;
+	int ioremap_size_order;
 
 	BUG_ON(in_interrupt());
-	if (flags & VM_IOREMAP)
-		align = 1ul << clamp_t(int, fls_long(size),
-				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
+	if (flags & VM_IOREMAP) {
+		ioremap_size_order = fls_long(size);
+		if (is_power_of_2(size) && size != 1)
+			ioremap_size_order--;
+		align = 1ul << clamp_t(int, ioremap_size_order, PAGE_SHIFT,
+				IOREMAP_MAX_ORDER);
+	}
 
 	size = PAGE_ALIGN(size);
 	if (unlikely(!size))

I'm having trouble with this, and a more complete description would
have helped!

As far as I can tell, the current code will decide the following:

size=0x10000: alignment=0x10000
size=0x0f000: alignment=0x8000

no, the current code doesn't achieve the above results as shown below
size=0x10000 -> fls_long(0x10000)=17 -> alignment=0x20000
size=0x0f000 -> fls_long(0x0f000)=16 -> alignment=0x10000
it is wrong for power of 2 value such as size=0x10000

And your patch will change it so that

size=0x10000: alignment=0x8000
size=0x0f000: alignment=0x8000

Correct?

no, my patch will results in the following calculations
size=0x10000: alignment=0x10000
size=0x0f000: alignment=0x10000

If so, I'm struggling to see the sense in this.  Shouldn't we be
changing things so that

size=0x10000: alignment=0x10000
size=0x0f000: alignment=0x10000

?

okay, it is the aim of my patch as explained above
i provide another solution as shown below
i appreciate it since it is more canonical
please help to review and apply it kindly