ARM64: Regression with commit e3067861ba66 ("arm64: add basic VMAP_STACK support")

[Ard Biesheuvel]

On 17 October 2017 at 10:36, Leo Yan <leo.yan at linaro.org> wrote:
> On Tue, Oct 17, 2017 at 10:32:21AM +0100, Ard Biesheuvel wrote:
>
> [...]
>
>> > AFAICT, erratum 836870 results in livelock rather than memory
>> > corruption, so I think we can ignore that.
>> >
>> > I'm a little worried by erratum 843419. The VMAP_STACK patches changed
>> > {adr,ldr}_this_cpu (and some users thereof), and it's possible we're
>> > managing to tickle that issue.
>> >
>> > If you still have an affected kernel, could you dump the output of:
>> >
>> > $ aarch64-linux-gnu-objdump -d vmlinux | grep -A 3 'ff[8c]:\s\+[a-f0-9]\+\s\+adrp'
>> >
>> > ... that would show us if there are any affected sequences.
>> >
>> > From a quick scan of my own vmlinux build from commit e3067861ba66, I
>> > didn't see any, but it's possible this depends on the config used.
>> >
>>
>> The linker should take care of that: it scans the entire executable,
>> and inserts a veneer if an adrp happens to end up at a vulnerable
>> offset in the page.
>
> Is this dependent on any GCC version?
>
> I am using GCC 6.2.1, so I get many affected sequences with Mark's command:
>
> leoy at leoy-linaro:~/Work/reference/opensource/linux$ aarch64-linux-gnu-objdump -d vmlinux | grep -A 3 'ff[8c]:\s\+[a-f0-9]\+\s\+adrp'
> ffff0000080a1ffc:       90007340        adrp    x0, ffff000008f09000 <page_wait_table+0x1280>
> ffff0000080a2000:       a900fedf        stp     xzr, xzr, [x22,#8]
> ffff0000080a2004:       91374000        add     x0, x0, #0xdd0
> ffff0000080a2008:       f9000ec0        str     x0, [x22,#24]
> --
> ffff0000080b6ff8:       b0007ce0        adrp    x0, ffff000009053000 <chunk_hash_heads+0x680>
> ffff0000080b6ffc:       52901801        mov     w1, #0x80c0                     // #32960
> ffff0000080b7000:       72a02801        movk    w1, #0x140, lsl #16
> ffff0000080b7004:       9102e273        add     x19, x19, #0xb8
> --
> ffff0000080f1ff8:       d00070a1        adrp    x1, ffff000008f07000 <bit_wait_table+0xd80>
> ffff0000080f1ffc:       f9406402        ldr     x2, [x0,#200]
> ffff0000080f2000:       f9405c03        ldr     x3, [x0,#184]
> ffff0000080f2004:       14002915        b       ffff0000080fc458 <e843419 at 00ce_00000e94_3c4>

This is a branch to a veneer

> --
> ffff0000080feff8:       90000002        adrp    x2, ffff0000080fe000 <prio_changed_rt+0x88>
> ffff0000080feffc:       9136e042        add     x2, x2, #0xdb8
> ffff0000080ff000:       f9000462        str     x2, [x3,#8]
> ffff0000080ff004:       f9448e62        ldr     x2, [x19,#2328]
> --
> ffff00000810affc:       f0006fe1        adrp    x1, ffff000008f09000 <page_wait_table+0x1280>
> ffff00000810b000:       f94017a2        ldr     x2, [x29,#40]
> ffff00000810b004:       aa1303e0        mov     x0, x19
> ffff00000810b008:       1400013d        b       ffff00000810b4fc <e843419 at 00e7_000010c5_338>

And this as well

> --
> ffff00000811dff8:       d0005c80        adrp    x0, ffff000008caf000 <kallsyms_token_index+0xaf00>
> ffff00000811dffc:       912be021        add     x1, x1, #0xaf8
> ffff00000811e000:       912ae000        add     x0, x0, #0xab8
> ffff00000811e004:       97ffe6ab        bl      ffff000008117ab0 <printk>
> --
> ffff000008137ff8:       f0004443        adrp    x3, ffff0000089c2000 <clock_monotonic+0x50>
> ffff000008137ffc:       910ac042        add     x2, x2, #0x2b0
> ffff000008138000:       9113c063        add     x3, x3, #0x4f0
> ffff000008138004:       71000c9f        cmp     w4, #0x3
> --
> ffff00000815eff8:       b0005aa1        adrp    x1, ffff000008cb3000 <kallsyms_token_index+0xef00>
> ffff00000815effc:       91362021        add     x1, x1, #0xd88
> ffff00000815f000:       97ffffc8        bl      ffff00000815ef20 <audit_log_format>
> ffff00000815f004:       b94023a1        ldr     w1, [x29,#32]
>

... so it seems the linker is doing its job, and updating the affected
sequences, and leaving the other ones alone.