[External] Re: kasan-for-arm32

[李文杰]

Dear Linus Walleij,

Thanks for reply,

1. I know why my kernel-5.4's vmalloc region not shadowed already,
because the CONFIG_ENABLE_VMALLOC_SAVING=y was enabled.

2. But I could not find it in mainline，so I guess it was added by Qualcomm.

3. Once it was enabled, the VMALLOC_START and VMALLOC_END will be
invalid.The code is some like below:

553 #ifdef CONFIG_ENABLE_VMALLOC_SAVING
554         print_vmalloc_lowmem_info();
555 #else
556         pr_notice(
557                    "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
558                    "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n",
559                         MLM(VMALLOC_START, VMALLOC_END),
560                         MLM(PAGE_OFFSET, (unsigned long)high_memory));
561 #endif

4. Instead, the vmalloc region follows lowmem region one by one, so my
kernel virtual memory layout as below:
[    0.000000] Virtual kernel memory layout:
[    0.000000]     vector  : 0xffff0000 - 0xffff1000   (   4 kB)
[    0.000000]     fixmap  : 0xffc80000 - 0xfff00000   (2560 kB)
[    0.000000]     vmalloc : 0xf8500000 - 0xff800000   ( 115 MB)  //highmem
[    0.000000]     lowmem  : 0xe1f00000 - 0xf8500000   ( 358 MB)
[    0.000000]     vmalloc : 0xe0e00000 - 0xe1f00000   (  17 MB)
[    0.000000]     lowmem  : 0xe0100000 - 0xe0e00000   (  13 MB)
[    0.000000]     vmalloc : 0xe0000000 - 0xe0100000   (   1 MB)
[    0.000000]     lowmem  : 0xcff17000 - 0xe0000000   ( 256 MB)
[    0.000000]     vmalloc : 0xcab00000 - 0xcff17000   (  84 MB)
[    0.000000]     lowmem  : 0xc6300000 - 0xcab00000   (  72 MB)
[    0.000000]     vmalloc : 0xc5fff000 - 0xc6300000   (   3 MB)
[    0.000000]     lowmem  : 0xc5f10000 - 0xc5fff000   (   0 MB)
[    0.000000]     vmalloc : 0xc5a00000 - 0xc5f10000   (   5 MB)
[    0.000000]     lowmem  : 0xc0000000 - 0xc5a00000   (  90 MB)
[    0.000000]     pkmap   : 0xbfe00000 - 0xc0000000   (   2 MB)
[    0.000000]     modules : 0xbf000000 - 0xbfe00000   (  14 MB)
[    0.000000]       .text : 0x(ptrval) - 0x(ptrval)   (27616 kB)
[    0.000000]       .init : 0x(ptrval) - 0x(ptrval)   (1024 kB)
[    0.000000]       .data : 0x(ptrval) - 0x(ptrval)   (2221 kB)
[    0.000000]        .bss : 0x(ptrval) - 0x(ptrval)   (8100 kB)

5. Now my solution is that, add all vmalloc regions above except
highmem to the mapping. Then my machine bringup successfully.
arch/arm/mm/kasan_init.c -> kasan_init() -> create_mapping().

6. Next I will continue to try and enjoy your kasan-for-arm32 patch in
my machine.

Thanks again dear Linus Walleij and Lecopzer.chen,

Linus Walleij <linus.walleij at linaro.org> 于2022年11月10日周四 17:51写道：
>
> On Wed, Nov 9, 2022 at 12:14 PM 李文杰 <liwenjie.liwenjie at bytedance.com> wrote:
>
> [Contex: porting kasan to kernel v5.4]
>
> > 1. I know why my machine crash after porting the patch.
> > 2. Because the kernel access the vmalloc region, which not shadowed.
> > 3. Now I found at least two places where used vmalloc region:
> >    1). arch/arm/mm/fault-armv.c ->check_writebuffer_bugs() -> vmap();
> >    2). lib/genalloc.c -> chunk = vzalloc_node(nbytes, nid);
> > 4. could I add the vmalloc to shadow memory region?
>
> I think you need to backport at least the following patches from upstream:
>
> 823f606ab6b4 ARM: 9242/1: kasan: Only map modules if CONFIG_KASAN_VMALLOC=n
> 565cbaad83d8 ARM: 9202/1: kasan: support CONFIG_KASAN_VMALLOC
>
> Both from Lecopzer Chen.
>
> Yours,
> Linus Walleij