----- "Daisuke Nishimura" <nishimura at mxp.nes.nec.co.jp> wrote:
> In !SPARSEMEM_EX case, the symbol mem_section points a array of struct mem_section,
> doesn't point a array of pointer to mem_section[], so I think the
> check:
>
> if ((mem_sec[SECTION_NR_TO_ROOT(nr)] == 0) ||
> !IS_KVADDR(mem_sec[SECTION_NR_TO_ROOT(nr)]))
> return 0;
>
> in nr_to_section() is not a valid check in this case.
>
> To make the matters worse, if CONFIG_CGROUP_MEM_RES_CTLR is enabled, the end of
> struct mem_section is padded with 0.
>
> So, reading mem_section by "rd" will look like:
>
> crash> rd mem_section 128
> c08b71e0: c9002003 c9001200 f7000000 00000000 . ..............
> c08b71f0: c9002003 c9001260 f6980000 00000000 . ..`...........
> c08b7200: c9002003 c90012c0 f6300000 00000000 . ........0.....
> c08b7210: c9002003 c9001320 f5c80000 00000000 . .. ...........
> ...
>
> This means nr_to_section() will return 0 when "nr" is 3,7,11,... because it meets
> the condition(mem_sec[SECTION_NR_TO_ROOT(nr)] == mem_sec[nr] == 0).
>
> As a result, mem_map for section 3,7,11,... cannot be handled properly.
> Actually, "kmem -n" doesn't show a mem_map for them.
>
> crash> kmem -n
> ...
> NR SECTION CODED_MEM_MAP MEM_MAP PFN
> 0 c08b71e0 c9002000 c9002000 0
> 1 c08b71f0 c9002000 c9402000 131072
> 2 c08b7200 c9002000 c9802000 262144
> 4 c08b7220 c9002000 ca002000 524288
> 5 c08b7230 c9002000 ca402000 655360
> 6 c08b7240 c9002000 ca802000 786432
> 8 c08b7260 c8c02000 cac02000 1048576
> 9 c08b7270 c8c02000 cb002000 1179648
> 10 c08b7280 c8c02000 cb402000 1310720
> 12 c08b72a0 c8c02000 cbc02000 1572864
> ...
>
> This patch is a fix for this problem. nr_to_section() will check "addr" by
> IS_KVADDR() later anyway, so this patch just removes the problematic check.
I cannot recall exactly why that sanity check was put in place,
but it's clear that it only applies for CONFIG_SPARSEMEM_EXTREME
kernels. Thanks for catching that.
However, at least theoretically, the s390/s390x architectures
could be configured with CONFIG_SPARSEMEM_EXTREME, and in that
case "0" is considered a legitimate kernel virtual address.
But it would never be a legitimate mem_sec[] address.
So -- out of sheer paranoia -- instead of removing the check
completely, I will move it further down inside the subsequent
"if (IS_SPARSEMEM_EX())" section:
--- memory.c 5 Apr 2010 19:04:29 -0000 1.217
+++ memory.c 6 Apr 2010 14:23:21 -0000
@@ -13058,14 +13058,13 @@
}
}
- if ((mem_sec[SECTION_NR_TO_ROOT(nr)] == 0) ||
- !IS_KVADDR(mem_sec[SECTION_NR_TO_ROOT(nr)]))
- return 0;
-
- if (IS_SPARSEMEM_EX())
+ if (IS_SPARSEMEM_EX()) {
+ if ((mem_sec[SECTION_NR_TO_ROOT(nr)] == 0) ||
+ !IS_KVADDR(mem_sec[SECTION_NR_TO_ROOT(nr)]))
+ return 0;
addr = mem_sec[SECTION_NR_TO_ROOT(nr)] +
(nr & SECTION_ROOT_MASK()) * SIZE(mem_section);
- else
+ } else
addr = symbol_value("mem_section") +
(SECTIONS_PER_ROOT() * SECTION_NR_TO_ROOT(nr) +
(nr & SECTION_ROOT_MASK())) * SIZE(mem_section);
Thanks,
Dave
> Signed-off-by: Daisuke Nishimura <nishimura at mxp.nes.nec.co.jp>
> ---
> diff -uprN crash-5.0.2.orig/memory.c crash-5.0.2/memory.c
> --- crash-5.0.2.orig/memory.c 2010-04-05 16:08:52.000000000 +0900
> +++ crash-5.0.2/memory.c 2010-04-06 09:56:43.000000000 +0900
> @@ -13030,10 +13030,6 @@ nr_to_section(ulong nr)
> }
> }
>
> - if ((mem_sec[SECTION_NR_TO_ROOT(nr)] == 0) ||
> - !IS_KVADDR(mem_sec[SECTION_NR_TO_ROOT(nr)]))
> - return 0;
> -
> if (IS_SPARSEMEM_EX())
> addr = mem_sec[SECTION_NR_TO_ROOT(nr)] +
> (nr & SECTION_ROOT_MASK()) * SIZE(mem_section);
