Message-Id: <E821A395-77FE-476C-8806-001395274AC4@linaro.org>
Date:	Sun, 24 Jan 2016 18:20:47 +0100
From:	Ard Biesheuvel <ard.biesheuvel@...aro.org>
To:	Guenter Roeck <linux@...ck-us.net>
Cc:	Heiko Carstens <heiko.carstens@...ibm.com>,
	Michael Ellerman <mpe@...erman.id.au>,
	Kees Cook <keescook@...omium.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: Problems with commit 'kallsyms: add support for relative offsets in kallsyms address table' (in mmotm)


> On 24 jan. 2016, at 18:05, Guenter Roeck <linux@...ck-us.net> wrote:
> 
>> On 01/24/2016 12:21 AM, Ard Biesheuvel wrote:
>>> On 24 January 2016 at 08:06, Guenter Roeck <linux@...ck-us.net> wrote:
>>>> On 01/23/2016 10:10 PM, Ard Biesheuvel wrote:
>>>> 
>>>> 
>>>> 
>>>>>> On 24 jan. 2016, at 03:35, Guenter Roeck <linux@...ck-us.net> wrote:
>>>>>> 
>>>>>> On 01/23/2016 06:06 PM, Guenter Roeck wrote:
>>>>>> Hi,
>>>>>> 
>>>>>> I see runtime problems with the current mmotm branch. All qemu mips
>>>>>> targets
>>>>>> (32 and 64 bit, big and little endian) are stuck in boot after this
>>>>>> commit.
>>>>>> 
>>>>>> Bisect points to commit d13682e4d9d2 ("kallsyms: add support for
>>>>>> relative offsets
>>>>>> in kallsyms address table". Disabling CONFIG_KALLSYMS_BASE_RELATIVE
>>>>>> fixes the problem,
>>>>>> ie I can boot the image with qemu.
>>>>>> 
>>>>>> Bisect log is attached.
>>>>>> 
>>>>>> Playing with the problem, I found the following:
>>>>>> 
>>>>>> 1) The problem is only seen with a toolchain using binutils 2.22, but
>>>>>> not
>>>>>>     with a toolchain using binutils 2.25. The compiler configuration may
>>>>>> be
>>>>>>     different for both toolchains.
>>>>>> 2) Message "kallsyms failure: absolute symbol value 0xffffffff807afd14
>>>>>> out of range
>>>>>>     in relative mode" (twice) when using the toolchain with binutils
>>>>>> 2.22.
>>>>>>     This does not cause the build to fail, though.
>>>>>> 3) kallsyms_sym_address() parameter variable type is "int". In the
>>>>>> calling code,
>>>>>>     the variable type used is "unsigned long". That has no impact on the
>>>>>> problem,
>>>>>>     though.
>>>>> 
>>>>> 
>>>>> An additional data point: When using the older toolchain, many symbols in
>>>>> System.map
>>>>> are marked "A".
>>>>>     ffffffff80100000 A _text
>>>>> With the more recent toolchain, the same symbols are marked "T".
>>>>>     ffffffff80100000 T _text
>>>> 
>>>> Thanks for the analysis. It is surprising that the build does not fail
>>>> when this occurs, and the subsequent hangs themselves are probably caused by
>>>> missing kallsyms data.
>>> Yes, I wondered why the build doesn't fail. Seems odd.
>>> 
>>>> scripts/kallsyms.c ignores all A symbols except _text, which is actually a
>>>> relative symbol by nature so we can simply assume it is relative (i.e.,
>>>> override it as T)
>>>> 
>>>> Re x86_64 !SMP, any build time errors there as well? Likewise for sparc32?
>>> 
>>> Yes, same kind of errors for both. For x86_64/nosmp I also get the error
>>> message
>>> when using the Ubuntu native toolchain, so it doesn't seem to be (directly)
>>> related to binutils 2.22 vs. 2.25 for that architecture.
>>> 
>>> Runtime behavior is a bit different for the different architectures.
>>> x86_64 dies silently without any console output, mips just hangs,
>>> and sparc32 gets a panic with NULL pointer access.
>>> Of course, with missing kallsyms data all bets are off.
>>> 
>>>> 
>>>> Thanks again, and sorry for the trouble,
>>> 
>>> 
>>> No worries. Hope you'll get this sorted out.
>> 
>> OK, there's an additional issue in my latest version: the
>> kallsyms_relative_base value itself is not relocated.
>> 
>> If you have more time to burn on this, could you try the following on
>> top? (If not, that is also fine, I will look into it myself on Monday)
>> 
>> diff --git a/scripts/kallsyms.c b/scripts/kallsyms.c
>> index 5ab13394dfd9..0f43f0751d47 100644
>> --- a/scripts/kallsyms.c
>> +++ b/scripts/kallsyms.c
>> @@ -137,8 +137,10 @@ static int read_symbol(FILE *in, struct sym_entry *s)
>>                 sym++;
>> 
>>         /* Ignore most absolute/undefined (?) symbols. */
>> -       if (strcmp(sym, "_text") == 0)
>> +       if (strcmp(sym, "_text") == 0) {
>>                 _text = s->addr;
>> +               stype = 'T';
>> +       }
>>         else if (check_symbol_range(sym, s->addr, text_ranges,
>>                                     ARRAY_SIZE(text_ranges)) == 0)
>>                 /* nothing to do */;
>> @@ -406,7 +408,7 @@ static void write_src(void)
>> 
>>         if (base_relative) {
>>                 output_label("kallsyms_relative_base");
>> -               printf("\tPTR\t%#llx\n", relative_base);
>> +               printf("\tPTR\t_text - %#llx\n", _text - relative_base);
>>                 printf("\n");
>>         }
> 
> Does not help.
> 

For x86? Or none of them?

> Here is part of the problem. This is from a log message added to make_percpus_absolute().
> 
> Marking symbol 'B__bss_start' as absolute
> Marking symbol '?__init_end' as absolute
> Marking symbol 'D__nosave_begin' as absolute
> Marking symbol 'D__nosave_end' as absolute
> Marking symbol 'D__per_cpu_end' as absolute
> Marking symbol 'D__per_cpu_load' as absolute
> Marking symbol 'D__per_cpu_start' as absolute
> Marking symbol '?__smp_locks' as absolute
> Marking symbol '?__smp_locks_end' as absolute
> Marking symbol 'Bempty_zero_page' as absolute
> 
> This is with x86_64/nosmp. At least some of those symbols don't really reflect
> 'percpu' values. Maybe the distinction between percpu and non-percpu variables
> gets lost if SMP is not configured.
> 

Yes, sounds plausible, and that probably means some latent issue gets uncovered here rather than created. I suppose few people are testing x86_64+!SMP+CONFIG_RELOCATABLE thoroughly.

> On top of that, older versions of binutils mark additional symbols as absolute,
> even with x86_64.
> 
> ffffffff81a00000 A __end_rodata_hpage_align
> ffffffff81b19000 A __vvar_page
> ffffffff81d3d000 A _end
> 

Yes, but _text is the *only* symbol that is natively A that does not get filtered out (save for some ia64 specific ones) so these should not matter. Only _text and the percpu ones that get marked A explicitly should end up in the final table.

> Hope this helps,

A great deal, thanks a lot
Ard.

Powered by blists - more mailing lists