Message-ID: <bb34147d-4e67-456a-b0d6-965699cda596@gmx.de>
Date: Fri, 22 Dec 2023 09:23:14 +0100
From: Helge Deller <deller@....de>
To: Masahiro Yamada <masahiroy@...nel.org>, deller@...nel.org
Cc: linux-kernel@...r.kernel.org, Arnd Bergmann <arnd@...db.de>,
 linux-modules@...r.kernel.org, linux-arch@...r.kernel.org,
 Luis Chamberlain <mcgrof@...nel.org>
Subject: Re: [PATCH 0/4] Section alignment issues?

On 12/21/23 16:42, Masahiro Yamada wrote:
> On Thu, Dec 21, 2023 at 10:40 PM Masahiro Yamada <masahiroy@...nel.org> wrote:
>>
>> On Thu, Nov 23, 2023 at 7:18 AM <deller@...nel.org> wrote:
>>>
>>> From: Helge Deller <deller@....de>
>>>
>>> While working on the 64-bit parisc kernel, I noticed that the __ksymtab[]
>>> table was not correctly 64-bit aligned in many modules.
>>> The following patches do fix some of those issues in the generic code.
>>>
>>> But further investigation shows that multiple sections in the kernel and in
>>> modules are possibly not correctly aligned, and thus may lead to performance
>>> degregations at runtime (small on x86, huge on parisc, sparc and others which
>>> need exception handlers). Sometimes wrong alignments may also be simply hidden
>>> by the linker or kernel module loader which pulls in the sections by luck with
>>> a correct alignment (e.g. because the previous section was aligned already).
>>>
>>> An objdump on a x86 module shows e.g.:
>>>
>>> ./kernel/net/netfilter/nf_log_syslog.ko:     file format elf64-x86-64
>>> Sections:
>>> Idx Name          Size      VMA               LMA               File off  Algn
>>>    0 .text         00001fdf  0000000000000000  0000000000000000  00000040  2**4
>>>                    CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
>>>    1 .init.text    000000f6  0000000000000000  0000000000000000  00002020  2**4
>>>                    CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
>>>    2 .exit.text    0000005c  0000000000000000  0000000000000000  00002120  2**4
>>>                    CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
>>>    3 .rodata.str1.8 000000dc  0000000000000000  0000000000000000  00002180  2**3
>>>                    CONTENTS, ALLOC, LOAD, READONLY, DATA
>>>    4 .rodata.str1.1 0000030a  0000000000000000  0000000000000000  0000225c  2**0
>>>                    CONTENTS, ALLOC, LOAD, READONLY, DATA
>>>    5 .rodata       000000b0  0000000000000000  0000000000000000  00002580  2**5
>>>                    CONTENTS, ALLOC, LOAD, READONLY, DATA
>>>    6 .modinfo      0000019e  0000000000000000  0000000000000000  00002630  2**0
>>>                    CONTENTS, ALLOC, LOAD, READONLY, DATA
>>>    7 .return_sites 00000034  0000000000000000  0000000000000000  000027ce  2**0
>>>                    CONTENTS, ALLOC, LOAD, RELOC, READONLY, DATA
>>>    8 .call_sites   0000029c  0000000000000000  0000000000000000  00002802  2**0
>>>                    CONTENTS, ALLOC, LOAD, RELOC, READONLY, DATA
>>>
>>> In this example I believe the ".return_sites" and ".call_sites" should have
>>> an alignment of at least 32-bit (4 bytes).
>>>
>>> On other architectures or modules other sections like ".altinstructions" or
>>> "__ex_table" may show up wrongly instead.
>>>
>>> In general I think it would be beneficial to search for wrong alignments at
>>> link time, and maybe at runtime.
>>>
>>> The patch at the end of this cover letter
>>> - adds compile time checks to the "modpost" tool, and
>>> - adds a runtime check to the kernel module loader at runtime.
>>> And it will possibly show false positives too (!!!)
>>> I do understand that some of those sections are not performce critical
>>> and thus any alignment is OK.
>>>
>>> The modpost patch will emit at compile time such warnings (on x86-64 kernel build):
>>>
>>> WARNING: modpost: vmlinux: section .initcall7.init (type 1, flags 2) has alignment of 1, expected at least 4.
>>> Maybe you need to add ALIGN() to the modules.lds file (or fix modpost) ?
>>> WARNING: modpost: vmlinux: section .altinstructions (type 1, flags 2) has alignment of 1, expected at least 2.
>>> WARNING: modpost: vmlinux: section .initcall6.init (type 1, flags 2) has alignment of 1, expected at least 4.
>>> WARNING: modpost: vmlinux: section .initcallearly.init (type 1, flags 2) has alignment of 1, expected at least 4.
>>> WARNING: modpost: vmlinux: section .rodata.cst2 (type 1, flags 18) has alignment of 2, expected at least 64.
>>> WARNING: modpost: vmlinux: section .static_call_tramp_key (type 1, flags 2) has alignment of 1, expected at least 8.
>>> WARNING: modpost: vmlinux: section .con_initcall.init (type 1, flags 2) has alignment of 1, expected at least 8.
>>> WARNING: modpost: vmlinux: section __bug_table (type 1, flags 3) has alignment of 1, expected at least 4.
>>> ...
>>
>>
>>
>>
>> modpost acts on vmlinux.o instead of vmlinux.
>>
>>
>> vmlinux.o is a relocatable ELF, which is not a real layout
>> because no linker script has been considered yet at this
>> point.
>>
>>
>> vmlinux is an executable ELF, produced by a linker,
>> with the linker script taken into consideration
>> to determine the final section/symbol layout.
>>
>>
>> So, checking this in modpost is meaningless.
>>
>>
>>
>> I did not check the module checking code, but
>> modules are also relocatable ELF.
>
>
>
> Sorry, I replied too early.
> (Actually I replied without reading your modpost code).
>
> Now, I understand what your checker is doing.
>
>
> I did not test how many false positives are produced,
> but it catches several suspicious mis-alignments.

Yes.

> However, I am not convinced with this warning.
>
>
> +               warn("%s: section %s (type %d, flags %lu) has
> alignment of %d, expected at least %d.\n"
> +                    "Maybe you need to add ALIGN() to the modules.lds
> file (or fix modpost) ?\n",
> +                    modname, sec, sechdr->sh_type, sechdr->sh_flags,
> is_shalign, should_shalign);
> +       }
>
>
> Adding ALGIN() hides the real problem.

Right.
It took me some time to understand the effects here too.
See below...

> I think the real problem is that not enough alignment was requested
> in the code.
>
> For example, the right fix for ".initcall7.init" should be this:
>
> diff --git a/include/linux/init.h b/include/linux/init.h
> index 3fa3f6241350..650311e4b215 100644
> --- a/include/linux/init.h
> +++ b/include/linux/init.h
> @@ -264,6 +264,7 @@ extern struct module __this_module;
>   #define ____define_initcall(fn, __stub, __name, __sec)         \
>          __define_initcall_stub(__stub, fn)                      \
>          asm(".section   \"" __sec "\", \"a\"            \n"     \
> +           ".balign 4                                  \n"     \
>              __stringify(__name) ":                      \n"     \
>              ".long      " __stringify(__stub) " - .     \n"     \
>              ".previous                                  \n");   \
>
> Then, "this section requires at least 4 byte alignment"
> is recorded in the sh_addralign field.

Yes, this is the important part.

> Then, the rest is the linker's job.
>
> We should not tweak the linker script.

That's right, but let's phrase it slightly different...
There is *no need* to tweak the linker script, *if* the alignment
gets correctly assigned by the inline assembly (like your
initcall patch above).
But on some platforms (e.g. on parisc) I noticed that this .balign
was missing for some other sections, in which case the other (not preferred)
possible option is to tweak the linker script.

So I think we agree that fixing the inline assembly is the right
way to go?

Either way, a link-time check like the proposed modpost patch
may catch section issue for upcoming/newly added sections too.

Helge

Powered by blists - more mailing lists