On Mon, Sep 15, 2014 at 12:33:25PM +0100, Russell King - ARM Linux wrote:

quoted

On Mon, Sep 15, 2014 at 07:07:20PM +0800, Wang, Yalin wrote:

quoted

@@ -636,6 +646,11 @@ static int keep_initrd;
 void free_initrd_mem(unsigned long start, unsigned long end)
 {
 	if (!keep_initrd) {
+		if (start == initrd_start)
+			start = round_down(start, PAGE_SIZE);
+		if (end == initrd_end)
+			end = round_up(end, PAGE_SIZE);
+
 		poison_init_mem((void *)start, PAGE_ALIGN(end) - start);
 		free_reserved_area((void *)start, (void *)end, -1, "initrd");
 	}

is the only bit of code you likely need to achieve your goal.

Thinking about this, I think that you are quite right to align these.
The memory around the initrd is defined to be system memory, and we
already free the pages around it, so it *is* wrong not to free the
partial initrd pages.

Actually, I think we have a problem, at least on arm64 (raised by Peter
Maydell). There is no guarantee that the page around start/end of initrd
is free, it may contain the dtb for example. This is even more obvious
when we have a 64KB page kernel (the boot loader doesn't know the page
size that the kernel is going to use).

The bug was there before as we had poison_init_mem() already (not it
disappeared since free_reserved_area does the poisoning).

So as a quick fix I think we need the rounding the other way (and in the
general case we probably lose a page at the end of initrd):

diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
index 494297c698ca..39fd080683e7 100644
--- a/arch/arm64/mm/init.c
+++ b/arch/arm64/mm/init.c

@@ -335,9 +335,9 @@ void free_initrd_mem(unsigned long start, unsigned long end)
 {
 	if (!keep_initrd) {
 		if (start == initrd_start)
-			start = round_down(start, PAGE_SIZE);
+			start = round_up(start, PAGE_SIZE);
 		if (end == initrd_end)
-			end = round_up(end, PAGE_SIZE);
+			end = round_down(end, PAGE_SIZE);
 
 		free_reserved_area((void *)start, (void *)end, 0, "initrd");
 	}

A better fix would be to check what else is around the start/end of
initrd.