[Crash-utility] Re: kmem -p mapping and index

[David Anderson]

Bob Montgomery wrote:

>Dave,
>
>We're working on a problem where two physical pages think they're mapped
>to the same virtual page (one has been orphaned)
>
>I wanted to see if any other cases occurred, so I wanted to use kmem -p
>to dump all the page structs.  But I got the lines like this:
>
>      PAGE       PHYSICAL      MAPPING       INDEX CNT FLAGS
>ffff81012fbb7720 11c6fc000      -------       -----   3 200000000000004
>...
>ffff81012fc8d3b8 120411000      -------       -----   2 200000000000064
>
>which led to the twisty passages of anonymous struct members, and gdb
>nastiness, and datatype_info etc.  
>
>But the real key was that 
>
>MEMBER_OFFSET_INIT(page_mapping, "page", "mapping");
>
>was failing because on our kernel, mapping is in an anonymous struct in
>a union in the page struct.
>
>When I figured out that the following would still work:
>crash> gdb p &((struct page*)0x0).mapping
>$4 = (struct address_space **) 0x18
>
>I set about to writing a patch to find it the hard way by sending that
>command to gdb and then parsing the result, blah, blah, blah.  Imagine
>my surprise when I looked for an example somewhere in crash and happened
>upon anon_member_offset, doing exactly the same thing!  That and a few
>discovered macros and all I had to do was change the above to:
>
>   MEMBER_OFFSET_INIT(page_mapping, "page", "mapping");
>   if (INVALID_MEMBER(page_mapping))
>                ANON_MEMBER_OFFSET_INIT(page_mapping, "page","mapping");
>
>So now, I can see that these two page structs have the same mapping and
>offset, (and that this was the only case of it in my entire dump).  I'm
>so happy.  
>
>      PAGE       PHYSICAL      MAPPING       INDEX CNT FLAGS
>ffff81012fbb7720 11c6fc000 ffff81012a0993f1 11453246214  3 200000000000004
>...
>ffff81012fc8d3b8 120411000 ffff81012a0993f1 11453246214  2 200000000000064
>
>Patch is attached.  Any problem with doing it this way?
>
Absolutely not -- and in fact, later kernels with CONFIG_SLUB support
require the same thing to be done for the page index offset as well.
I just added that interface in 4.0-4.8 for the CONFIG_SLUB support.

Thanks for catching this -- queued for the next release.

Dave

>
>
>Thanks,
>Bob Montgomery
>
>
>
>------------------------------------------------------------------------
>
>--- memory.c.orig	2007-11-07 13:34:34.000000000 -0700
>+++ memory.c	2007-11-07 13:37:39.000000000 -0700
>@@ -277,6 +277,8 @@ vm_init(void)
> 		MEMBER_OFFSET_INIT(page_count, "page", "_count");
> 	MEMBER_OFFSET_INIT(page_flags, "page", "flags");
>         MEMBER_OFFSET_INIT(page_mapping, "page", "mapping");
>+	if (INVALID_MEMBER(page_mapping))
>+		ANON_MEMBER_OFFSET_INIT(page_mapping, "page", "mapping");
>         MEMBER_OFFSET_INIT(page_index, "page", "index");
>         MEMBER_OFFSET_INIT(page_buffers, "page", "buffers");
> 	MEMBER_OFFSET_INIT(page_lru, "page", "lru");
>