EUNBONG SONG <[email protected]> reported a bug on MIPS64 platforms
caused by free_initmem() as below:
[ 132.134719] CPU: 0 PID: 1 Comm: swapper/0 Tainted: G W 3.9.0+ #29
[ 132.141678] Stack : 0000000000000004 000000000000003f ffffffff80fa0000 ffffffff802924a8
0000000000000000 ffffffff80fa0000 00000000000000ff ffffffff80293760
0000000000000000 0000000000000000 ffffffff81080000 ffffffff81080000
ffffffff80e2baf0 ffffffff80f93977 a80000004146cbb8 0000000000000020
0000000000000003 0000000000000020 a800000041473da8 ffffffff810f0000
a800000041473a10 ffffffff806ef910 a800000041473828 ffffffff80290920
0000000000000000 ffffffff80293b90 000000000000000a ffffffff80e2baf0
0000000000000000 a800000041473750 000000004146cef8 ffffffff805e7794
0000000000000000 0000000000000000 0000000000000000 0000000000000000
0000000000000000 ffffffff80272498 0000000000000000 0000000000000000
...
[ 132.207201] Call Trace:
[ 132.209655] [<ffffffff80272498>] show_stack+0x68/0x80
[ 132.225943] [<ffffffff802bd4ac>] notifier_call_chain+0x5c/0xa8
[ 132.231776] [<ffffffff802bdb84>] __atomic_notifier_call_chain+0x3c/0x58
[ 132.238391] [<ffffffff802bdbe8>] notify_die+0x38/0x48
[ 132.243442] [<ffffffff802716cc>] die+0x4c/0x148
[ 132.247974] [<ffffffff8027f998>] do_page_fault+0x4b8/0x500
[ 132.253461] [<ffffffff8026c764>] resume_userspace_check+0x0/0x10
[ 132.259469] [<ffffffff80324a54>] free_reserved_area+0x8c/0x178
[ 132.265304] [<ffffffff806e0dc8>] kernel_init+0x20/0x100
[ 132.270529] [<ffffffff8026c7e0>] ret_from_kernel_thread+0x10/0x18
The root cause is that virt_to_page()/virt_to_phys() can't be used to
handle virtual address from compatible segments on MIPS64 because
virt_to_phys() is defined as:
static inline unsigned long virt_to_phys(volatile const void *address)
{
return (unsigned long)address - PAGE_OFFSET + PHYS_OFFSET;
}
x86_64 platforms have a similar situation, but __pa() has been enhanced
to handle virtual address space used for kernel code and data segments.
static inline unsigned long __phys_addr_nodebug(unsigned long x)
{
unsigned long y = x - __START_KERNEL_map;
/* use the carry flag to determine if x was < __START_KERNEL_map */
x = y + ((x > y) ? phys_base : (__START_KERNEL_map - PAGE_OFFSET));
return x;
}
So we have two possible solutions here. The quick solution is to revert
to the original implementation by using __pa_symbal(). The long term
solution is to enhance virt_to_phys() to correctly handle virtual
address from compatible segments.
This patch adopts the quick solution to fix the bug for v3.10, and we
need guidance from MIPS64 experts on whether we should go with the long
term solution.
Signed-off-by: Jiang Liu <[email protected]>
Signed-off-by: EUNBONG SONG <[email protected]>
Cc: David Daney <[email protected]>
Cc: Ralf Baechle <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Arnd Bergmann <[email protected]>
Cc: [email protected]
Cc: [email protected]
---
arch/mips/mm/init.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/arch/mips/mm/init.c b/arch/mips/mm/init.c
index 3d0346d..3648768 100644
--- a/arch/mips/mm/init.c
+++ b/arch/mips/mm/init.c
@@ -446,7 +446,9 @@ void free_initrd_mem(unsigned long start, unsigned long end)
void __init_refok free_initmem(void)
{
prom_free_prom_memory();
- free_initmem_default(POISON_FREE_INITMEM);
+ free_init_pages("unused kernel memory",
+ __pa_symbol(&__init_begin),
+ __pa_symbol(&__init_end));
}
#ifndef CONFIG_MIPS_PGD_C0_CONTEXT
--
1.7.9.5
On 05/09/2013 09:10 AM, Jiang Liu wrote:
> EUNBONG SONG <[email protected]> reported a bug on MIPS64 platforms
> caused by free_initmem() as below:
> [ 132.134719] CPU: 0 PID: 1 Comm: swapper/0 Tainted: G W 3.9.0+ #29
> [ 132.141678] Stack : 0000000000000004 000000000000003f ffffffff80fa0000 ffffffff802924a8
> 0000000000000000 ffffffff80fa0000 00000000000000ff ffffffff80293760
> 0000000000000000 0000000000000000 ffffffff81080000 ffffffff81080000
> ffffffff80e2baf0 ffffffff80f93977 a80000004146cbb8 0000000000000020
> 0000000000000003 0000000000000020 a800000041473da8 ffffffff810f0000
> a800000041473a10 ffffffff806ef910 a800000041473828 ffffffff80290920
> 0000000000000000 ffffffff80293b90 000000000000000a ffffffff80e2baf0
> 0000000000000000 a800000041473750 000000004146cef8 ffffffff805e7794
> 0000000000000000 0000000000000000 0000000000000000 0000000000000000
> 0000000000000000 ffffffff80272498 0000000000000000 0000000000000000
> ...
> [ 132.207201] Call Trace:
> [ 132.209655] [<ffffffff80272498>] show_stack+0x68/0x80
> [ 132.225943] [<ffffffff802bd4ac>] notifier_call_chain+0x5c/0xa8
> [ 132.231776] [<ffffffff802bdb84>] __atomic_notifier_call_chain+0x3c/0x58
> [ 132.238391] [<ffffffff802bdbe8>] notify_die+0x38/0x48
> [ 132.243442] [<ffffffff802716cc>] die+0x4c/0x148
> [ 132.247974] [<ffffffff8027f998>] do_page_fault+0x4b8/0x500
> [ 132.253461] [<ffffffff8026c764>] resume_userspace_check+0x0/0x10
> [ 132.259469] [<ffffffff80324a54>] free_reserved_area+0x8c/0x178
> [ 132.265304] [<ffffffff806e0dc8>] kernel_init+0x20/0x100
> [ 132.270529] [<ffffffff8026c7e0>] ret_from_kernel_thread+0x10/0x18
>
> The root cause is that virt_to_page()/virt_to_phys() can't be used to
> handle virtual address from compatible segments on MIPS64 because
> virt_to_phys() is defined as:
> static inline unsigned long virt_to_phys(volatile const void *address)
> {
> return (unsigned long)address - PAGE_OFFSET + PHYS_OFFSET;
> }
>
> x86_64 platforms have a similar situation, but __pa() has been enhanced
> to handle virtual address space used for kernel code and data segments.
> static inline unsigned long __phys_addr_nodebug(unsigned long x)
> {
> unsigned long y = x - __START_KERNEL_map;
>
> /* use the carry flag to determine if x was < __START_KERNEL_map */
> x = y + ((x > y) ? phys_base : (__START_KERNEL_map - PAGE_OFFSET));
>
> return x;
> }
>
> So we have two possible solutions here. The quick solution is to revert
> to the original implementation by using __pa_symbal(). The long term
> solution is to enhance virt_to_phys() to correctly handle virtual
> address from compatible segments.
We have a patch that does exactly this for MIPS:
http://www.linux-mips.org/archives/linux-mips/2013-05/msg00048.html
or if you prefer patchwork:
http://patchwork.linux-mips.org/patch/5220/
>
> This patch adopts the quick solution to fix the bug for v3.10, and we
> need guidance from MIPS64 experts on whether we should go with the long
> term solution.
>
I (as the author of the Long Term Solution) prefer that we apply the fix
for virt_to_phys(), which is also needed for the reasons stated in the
patch changelog.
So I would have to say NACK to this patch, use the virt_to_phys() fix
instead.
David Daney
> Signed-off-by: Jiang Liu <[email protected]>
> Signed-off-by: EUNBONG SONG <[email protected]>
> Cc: David Daney <[email protected]>
> Cc: Ralf Baechle <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: Arnd Bergmann <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> ---
> arch/mips/mm/init.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/arch/mips/mm/init.c b/arch/mips/mm/init.c
> index 3d0346d..3648768 100644
> --- a/arch/mips/mm/init.c
> +++ b/arch/mips/mm/init.c
> @@ -446,7 +446,9 @@ void free_initrd_mem(unsigned long start, unsigned long end)
> void __init_refok free_initmem(void)
> {
> prom_free_prom_memory();
> - free_initmem_default(POISON_FREE_INITMEM);
> + free_init_pages("unused kernel memory",
> + __pa_symbol(&__init_begin),
> + __pa_symbol(&__init_end));
> }
>
> #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
>