Re: [PATCH v2 RESEND 2/2] x86/mm/KASLR: Do not adapt the size of the direct mapping section for SGI UV system

From: Ingo Molnar
Date: Thu Sep 28 2017 - 03:56:16 EST



* Baoquan He <bhe@xxxxxxxxxx> wrote:

> On SGI UV system, kernel often hangs when KASLR is enabled. Disabling
> KASLR makes kernel work well.
>
> The back trace is:
>
> kernel BUG at arch/x86/mm/init_64.c:311!
> invalid opcode: 0000 [#1] SMP
> [...]
> RIP: 0010:__init_extra_mapping+0x188/0x196
> [...]
> Call Trace:
> init_extra_mapping_uc+0x13/0x15
> map_high+0x67/0x75
> map_mmioh_high_uv3+0x20a/0x219
> uv_system_init_hub+0x12d9/0x1496
> uv_system_init+0x27/0x29
> native_smp_prepare_cpus+0x28d/0x2d8
> kernel_init_freeable+0xdd/0x253
> ? rest_init+0x80/0x80
> kernel_init+0xe/0x110
> ret_from_fork+0x2c/0x40
>
> This is because the SGI UV system need map its MMIOH region to the direct
> mapping section, and the mapping happens in rest_init() which is much
> later than the calling of kernel_randomize_memory() to do mm KASLR. So
> mm KASLR can't count in the size of the MMIOH region when caculate the
> needed size of address space for the direct mapping section.
>
> When KASLR is disabled, there are 64TB address space for both system RAM
> and the MMIOH regions to share. When KASLR is enabled, the current code
> of mm KASLR only reserves the actual size of system RAM plus extra 10TB
> for the direct mapping. Thus later the MMIOH mapping could go beyond
> the upper bound of the direct mapping to step into VMALLOC or VMEMMAP area.
> Then BUG_ON() in __init_extra_mapping() will be triggered.
>
> E.g on the SGI UV3 machine where this bug is reported , there are two MMIOH
> regions:
>
> [ 1.519001] UV: Map MMIOH0_HI 0xffc00000000 - 0x100000000000
> [ 1.523001] UV: Map MMIOH1_HI 0x100000000000 - 0x200000000000
>
> They are [16TB-16G, 16TB) and [16TB, 32TB). On this machine, 512G RAM are
> spread out to 1TB regions. Then above two SGI MMIOH regions also will be
> mapped into the direct mapping section.
>
> To fix it, we need check if it's SGI UV system by calling is_early_uv_system()
> in kernel_randomize_memory(). If yes, do not adapt the size of the direct
> mapping section, just keep it as 64TB.
>
> Signed-off-by: Baoquan He <bhe@xxxxxxxxxx>
> Reviewed-by: Thomas Garnier <thgarnie@xxxxxxxxxx>
> Acked-by: Mike Travis <travis@xxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> Cc: x86@xxxxxxxxxx
> Cc: Thomas Garnier <thgarnie@xxxxxxxxxx>
> Cc: Kees Cook <keescook@xxxxxxxxxxxx>
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> ---
> arch/x86/mm/kaslr.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/mm/kaslr.c b/arch/x86/mm/kaslr.c
> index af599167fe3c..4d68c08df82d 100644
> --- a/arch/x86/mm/kaslr.c
> +++ b/arch/x86/mm/kaslr.c
> @@ -27,6 +27,7 @@
> #include <asm/pgtable.h>
> #include <asm/setup.h>
> #include <asm/kaslr.h>
> +#include <asm/uv/uv.h>
>
> #include "mm_internal.h"
>
> @@ -123,7 +124,7 @@ void __init kernel_randomize_memory(void)
> CONFIG_RANDOMIZE_MEMORY_PHYSICAL_PADDING;
>
> /* Adapt phyiscal memory region size based on available memory */
> - if (memory_tb < kaslr_regions[0].size_tb)
> + if (memory_tb < kaslr_regions[0].size_tb && !is_early_uv_system())
> kaslr_regions[0].size_tb = memory_tb;

This is really an ugly hack. Is kaslr_regions[] incorrect? If so then it should be
corrected instead of uglifying the code that uses it...

Thanks,

Ingo