Re: [patch 1/1] [PATCH] include storage keys in hibernation image.

On Tuesday, June 14, 2011, Martin Schwidefsky wrote:

Hi Rafael,

Hi,

first of all: thanks for the review.

On Sun, 12 Jun 2011 14:41:34 +0200
"Rafael J. Wysocki" [off-list ref] wrote:

quoted

quoted

diff -urpN linux-2.6/arch/s390/kernel/swsusp_asm64.S linux-2.6-patched/arch/s390/kernel/swsusp_asm64.S

--- linux-2.6/arch/s390/kernel/swsusp_asm64.S	2011-05-19 06:06:34.000000000 +0200
+++ linux-2.6-patched/arch/s390/kernel/swsusp_asm64.S	2011-06-06 11:14:43.000000000 +0200

@@ -138,11 +138,14 @@ swsusp_arch_resume:
 0:
 	lg	%r2,8(%r1)
 	lg	%r4,0(%r1)
+	iske	%r0,%r4
 	lghi	%r3,PAGE_SIZE
 	lghi	%r5,PAGE_SIZE
 1:
 	mvcle	%r2,%r4,0
 	jo	1b
+	lg	%r2,8(%r1)
+	sske	%r0,%r2
 	lg	%r1,16(%r1)
 	ltgr	%r1,%r1
 	jnz	0b

I cannot comment on the assembly.

The swsusp_arch_resume code copies the safe pages to their final target
page. Writing to a page sets the dirty bit in the storage key. Which
means that the storage key needs to be restored >after< the copy has
been done. To avoid having to deal with additional pages where the
storage key are kept while the copy is still pending I simply set the
storage key of the safe page and transfer it from the safe page to the
final target page after the copy has been done. That is what the iske
and the sske instruction are doing.

I see.

quoted

quoted

diff -urpN linux-2.6/kernel/power/snapshot.c linux-2.6-patched/kernel/power/snapshot.c

--- linux-2.6/kernel/power/snapshot.c	2011-06-06 11:14:39.000000000 +0200
+++ linux-2.6-patched/kernel/power/snapshot.c	2011-06-06 11:14:43.000000000 +0200

@@ -1022,6 +1022,137 @@ static inline void copy_data_page(unsign
 }
 #endif /* CONFIG_HIGHMEM */
 
+#ifdef CONFIG_S390
+/*
+ * For s390 there is one additional byte associated with each page,
+ * the storage key. The key consists of the access-control bits
+ * (alias the key), the fetch-protection bit, the referenced bit
+ * and the change bit (alias dirty bit). Linux uses only the
+ * referenced bit and the change bit for pages in the page cache.
+ * Any modification of a page will set the change bit, any access
+ * sets the referenced bit. Overindication of the referenced bits
+ * after a hibernation cycle does not cause any harm but the
+ * overindication of the change bits would cause trouble.
+ * Therefore it is necessary to include the storage keys in the
+ * hibernation image. The storage keys are saved to the most
+ * significant byte of each page frame number in the list of pfns
+ * in the hibernation image.
+ */

Let me say that is not exactly straightforward. :-)

One thing I don't really understand is where those storage keys are normally
stored so that they aren't present in the image without this additional
mechanism.  Could you explain that a bit more, please?

The storage keys are outside of the directly addressable memory, the only
way to get to them is with special instructions:
iske - insert storage key extended, ivsk - insert virtual storage key,
sske - set storage key extended, and rrbe - reset reference bit extended.
The storage key of a page has 7 bits, 4 bit for access-control, the
fetch-protection-bit, the reference-bit and the change-bit. In Linux only
the reference and change bits are used.

OK, so it looks like we would only need to store 2 bits per page in additional
allocations.

These bits are per physical page, one of the major differences of the s390
machines compared to other architectures. We had a number of interesting
problems because of pte dirty & referenced vs page dirty & referenced.
For e.g. x86 a page is implicitly clean if it has no mapper. On s390 is
depends on the contents of the storage key. Which is why SetPageUptodate
clears the storage key.

In regard to hibernation the important aspect is that each and every write
to a page sets the dirty bit even if it is done by I/O. The restore of the
hibernation image therefore makes all pages dirty. To get the correct
original combination of page content and storage key content the storage
key needs to be set after the last access to the page content.

What, in your opinion, is the right time for restoring the original storage
keys after the image has been loaded?  It seems that should happen after
passing control to the hibernated kernel.

quoted

quoted

+
+/*
+ * Key storage is allocated as a linked list of pages.
+ * The size of the keys array is (PAGE_SIZE - sizeof(long))
+ */
+struct page_key_data {
+	struct page_key_data *next;
+	unsigned char data[];
+};

This seems to be similar to the data structure for the saving of ACPI NVS
memory, so perhaps it's possible to combine the two.

Almost the same, I thought about ways to merge them. I dropped the idea in
order to keep the patch as small as possible.

I think, however, that we really should try to merge them.  The only
difference seems to be how the additionally allocated pages will be populated
and what's going to happen to their contents during restore.

ACPI will simply copy the NVS memory to those pages, while S390 will save
the relevant storage key bits in there.

quoted

quoted

+#define PAGE_KEY_DATA_SIZE	(PAGE_SIZE - sizeof(struct page_key_data *))
+
+static struct page_key_data *page_key_data;
+static struct page_key_data *page_key_rp, *page_key_wp;
+static unsigned long page_key_rx, page_key_wx;
+
+/*
+ * For each page in the hibernation image one additional byte is
+ * stored in the most significant byte of the page frame number.

I'm not sure it's really worth the complication.  If you simply store those
keys in separete pages, you'd need one additional page per PAGE_SIZE
page frames, so given PAGE_SIZE = 4096 that's one page per 16 MB of memory,
which is not a whole lot (64 extra pages per GB if I'm not mistaken).

It seems that doing it will simplify things quite a git.

One of the versions of the patch actually did that and it turned out to
be much larger then the proposed solution. That older version had another
step in snapshot_read_next / snapshot_write_next to copy the storage keys
to pages allocated for that purpose. I had some trouble with the way how
the code deals with the orig_bm and the copy_bm though. Not trivial code..

Well, I'm sure we can handle that. :-)

I thought of the following approach:

* Allocate memory for saving the relevant bits of the storage keys before
  the memory preallocation kicks in.

* Save the storage key bits to that memory before creating the image (be
  careful about the storage keys of the pages they are being stored into).

* After getting control from the boot kernel, restore the storage key bits
  saved before creating the image.

* Free additional memory used for storing them.

This way we wouldn't have to worry about any interactions with
snapshot_read_next / snapshot_write_next etc.

Thanks,
Rafael