Re: [PATCH] Add non-zero module sections to sysfs

Re: [PATCH] Add non-zero module sections to sysfs

[Rusty Russell]

Sebastian Wankerl <sisewank@cip.cs.fau.de> writes:
> On 04/04/13 03:00, Rusty Russell wrote:
>> Sebastian Wankerl <sisewank@cip.cs.fau.de> writes:
>>> Add non-zero module sections to sysfs on architectures unequal to PARISC.
>>> KGDB needs all module sections for proper module debugging. Therefore, commit 
>>> 35dead4235e2b67da7275b4122fed37099c2f462 is revoked except for PARISC
>>> architecture.
>> #ifdef CONFIG_PARISC in the middle of kernel/module.c is super-ugly, and
>> wrong.
>
> I don't see why this is wrong. It used to load all sections to sysfs
> until the patch mentioned. Actually, it is the PARISC build chain which
> is broken.

Exactly.  Don't workaround it here, revert it and put the
duplicate-section-name fixup in parisc where it belongs.

Assuming parisc still produces these dup sections: that patch is 4 years
old now.

Untested:

diff --git a/arch/parisc/kernel/module.c b/arch/parisc/kernel/module.c
index 2a625fb..28d32a2 100644
--- a/arch/parisc/kernel/module.c
+++ b/arch/parisc/kernel/module.c
@@ -341,6 +341,11 @@ int module_frob_arch_sections(CONST Elf_Ehdr *hdr,
 			    ".PARISC.unwind", 14) == 0)
 			me->arch.unwind_section = i;
 
+		/* we produce multiple, empty .text sections, and kallsyms
+		 * gets upset.  make non-alloc so it doesn't see them. */
+		if (sechdrs[i].sh_size == 0)
+			sechdrs[i].sh_flags &= ~SHF_ALLOC;
+
 		if (sechdrs[i].sh_type != SHT_RELA)
 			continue;
 
>> My preference would be to fix kgdb.  If the section is empty, what need
>> does it have to examine it?
>
> GDB needs to know all sections of the binary and its addresses.

Why?  Does something refer to this empty section?  Why has noone noticed
this since 2009?

> It is generally useful to be able to check up all sections of the binary
> regardless if they are empty or not so one can see the binary's
> structure.

A zero-length section doesn't change the binary's structure.  You don't
see non-SHF_ALLOC sections either.

Cheers,
Rusty.

Re: [PATCH] Add non-zero module sections to sysfs

[Sebastian Wankerl]

On 04/05/13 06:00, Rusty Russell wrote:
> Exactly.  Don't workaround it here, revert it and put the
> duplicate-section-name fixup in parisc where it belongs.
>
> Assuming parisc still produces these dup sections: that patch is 4 years
> old now.
>
> Untested:
>
> diff --git a/arch/parisc/kernel/module.c b/arch/parisc/kernel/module.c
> index 2a625fb..28d32a2 100644
> --- a/arch/parisc/kernel/module.c
> +++ b/arch/parisc/kernel/module.c
> @@ -341,6 +341,11 @@ int module_frob_arch_sections(CONST Elf_Ehdr *hdr,
>  			    ".PARISC.unwind", 14) == 0)
>  			me->arch.unwind_section = i;
>  
> +		/* we produce multiple, empty .text sections, and kallsyms
> +		 * gets upset.  make non-alloc so it doesn't see them. */
> +		if (sechdrs[i].sh_size == 0)
> +			sechdrs[i].sh_flags &= ~SHF_ALLOC;
> +
>  		if (sechdrs[i].sh_type != SHT_RELA)
>  			continue;
We just worked your suggested patch in.

> Why?  Does something refer to this empty section?  Why has noone noticed
> this since 2009?

GDB wants to know all section with attribute ALLOC, regardless whether
they are empty or not. Thus, it is useful if all of them appear in sysfs.
> A zero-length section doesn't change the binary's structure.  You don't
> see non-SHF_ALLOC sections either.

Yes, but they do occupy an index in the section headers of the binary.
GDB needs to know all of them in the right order.

Re: [PATCH] Add non-zero module sections to sysfs

[James Bottomley]

On Fri, 2013-04-05 at 14:30 +1030, Rusty Russell wrote:
> Sebastian Wankerl <sisewank@cip.cs.fau.de> writes:
> > On 04/04/13 03:00, Rusty Russell wrote:
> >> Sebastian Wankerl <sisewank@cip.cs.fau.de> writes:
> >>> Add non-zero module sections to sysfs on architectures unequal to PARISC.
> >>> KGDB needs all module sections for proper module debugging. Therefore, commit 
> >>> 35dead4235e2b67da7275b4122fed37099c2f462 is revoked except for PARISC
> >>> architecture.

Thanks for actually cc'ing us.

> >> #ifdef CONFIG_PARISC in the middle of kernel/module.c is super-ugly, and
> >> wrong.
> >
> > I don't see why this is wrong. It used to load all sections to sysfs
> > until the patch mentioned. Actually, it is the PARISC build chain which
> > is broken.
> 
> Exactly.  Don't workaround it here, revert it and put the
> duplicate-section-name fixup in parisc where it belongs.
> 
> Assuming parisc still produces these dup sections: that patch is 4 years
> old now.

Just so you know: this isn't a parisc specific problem.  Gcc produces
duplicate section names under various circumstances, but the one that
bites us is -ffunction-sections.  Note that there are proposals to use
-ffunction-sections on all architectures (so we can garbage collect
unused functions) in which case you'll induce the bug identified in
35dead4235e2b67da7275b4122fed37099c2f462 on every architecture

The problem is our assumption that section names be unique.  This
assumption is wrong.  The ELF spec says (version 1.1 page 1-15): "An
object file may have more than one section with the same name."  We need
to fix the kernel not to rely on a bogus assumption ... but we had no
idea how to do that in a way that preserved the backwards compatibility
of sections subdirectory.

I admit that 35dead4235e2b67da7275b4122fed37099c2f462 is a hack, but now
the problem has got attention, can we fix it properly?

James

Re: [PATCH] Add non-zero module sections to sysfs

[Rusty Russell]

James Bottomley <James.Bottomley@HansenPartnership.com> writes:
> On Fri, 2013-04-05 at 14:30 +1030, Rusty Russell wrote:
>> Sebastian Wankerl <sisewank@cip.cs.fau.de> writes:
>> > On 04/04/13 03:00, Rusty Russell wrote:
>> >> Sebastian Wankerl <sisewank@cip.cs.fau.de> writes:
>> >>> Add non-zero module sections to sysfs on architectures unequal to PARISC.
>> >>> KGDB needs all module sections for proper module debugging. Therefore, commit 
>> >>> 35dead4235e2b67da7275b4122fed37099c2f462 is revoked except for PARISC
>> >>> architecture.
>
> Thanks for actually cc'ing us.
>
>> >> #ifdef CONFIG_PARISC in the middle of kernel/module.c is super-ugly, and
>> >> wrong.
>> >
>> > I don't see why this is wrong. It used to load all sections to sysfs
>> > until the patch mentioned. Actually, it is the PARISC build chain which
>> > is broken.
>> 
>> Exactly.  Don't workaround it here, revert it and put the
>> duplicate-section-name fixup in parisc where it belongs.
>> 
>> Assuming parisc still produces these dup sections: that patch is 4 years
>> old now.
>
> Just so you know: this isn't a parisc specific problem.  Gcc produces
> duplicate section names under various circumstances, but the one that
> bites us is -ffunction-sections.

*This* is a PA-RISC specific issue.  -ffunction-sections is a different
problem, which this hack wouldn't help.

> Note that there are proposals to use
> -ffunction-sections on all architectures (so we can garbage collect
> unused functions) in which case you'll induce the bug identified in
> 35dead4235e2b67da7275b4122fed37099c2f462 on every architecture

Good point, though I note that we seem to have stalled on
-ffunction-sections.  (And I vaguely recall an issue with
-ffunction-sections and using ld -o which would fold duplicate named
sections back together reducing elimination opportunities).

> The problem is our assumption that section names be unique.  This
> assumption is wrong.  The ELF spec says (version 1.1 page 1-15): "An
> object file may have more than one section with the same name."  We need
> to fix the kernel not to rely on a bogus assumption ... but we had no
> idea how to do that in a way that preserved the backwards compatibility
> of sections subdirectory.
>
> I admit that 35dead4235e2b67da7275b4122fed37099c2f462 is a hack, but now
> the problem has got attention, can we fix it properly?

Yep.  The original patch didn't go through me, or we would have had this
discussion back then...

The use of section names in sysfs goes back to one Mr. Corbet.  Why did
he do it that way?  Because gdb's add-symbol-file makes the same
assumption.  So if we fixed the sysfs somehow, it still wouldn't be
useful, since there's no way to tell gdb :(

The real answer don't use -ffunction-sections on modules: probably not
as important as the rest of the kernel.  And the new shiny is
-flto anyway.

And that leaves us with a PA-RISC specific issue, for which we should
move the fix to PA-RISC.

Thoughts?
Rusty.

Re: [PATCH] Add non-zero module sections to sysfs

[James Bottomley]

On Sat, 2013-04-06 at 15:22 +1030, Rusty Russell wrote:
> > The problem is our assumption that section names be unique.  This
> > assumption is wrong.  The ELF spec says (version 1.1 page 1-15): "An
> > object file may have more than one section with the same name."  We need
> > to fix the kernel not to rely on a bogus assumption ... but we had no
> > idea how to do that in a way that preserved the backwards compatibility
> > of sections subdirectory.
> >
> > I admit that 35dead4235e2b67da7275b4122fed37099c2f462 is a hack, but now
> > the problem has got attention, can we fix it properly?
> 
> Yep.  The original patch didn't go through me, or we would have had this
> discussion back then...
>
> The use of section names in sysfs goes back to one Mr. Corbet.  Why did
> he do it that way?  Because gdb's add-symbol-file makes the same
> assumption.  So if we fixed the sysfs somehow, it still wouldn't be
> useful, since there's no way to tell gdb :(
> 
> The real answer don't use -ffunction-sections on modules: probably not
> as important as the rest of the kernel.  And the new shiny is
> -flto anyway.
> 
> And that leaves us with a PA-RISC specific issue, for which we should
> move the fix to PA-RISC.
> 
> Thoughts?

Well, we don't have much of a choice.  Our ELF stub jump on 32 bits is a
PCREL17.  That means once a module size is over 128k there's a chance we
might not be able to link it because the jump is too big for the
instruction.  IPV6 is one such big module today, but I'm sure there are
others.  The only way I know to fix this is to allow the linker to
insert stubs between functions, so we only fail at linking if a single
function is >128k big.  The way to do this is -ffunction-sections,
unless there's something else we could do (all we really need is a way
to ensure we can insert ELF stubs every 128k).

We're not the only architecture that has these problems: frv, metag and
score seem to as well.

James

Re: [PATCH] Add non-zero module sections to sysfs

[John David Anglin]

On 6-Apr-13, at 6:52 AM, James Bottomley wrote:

> On Sat, 2013-04-06 at 15:22 +1030, Rusty Russell wrote:
>>> The problem is our assumption that section names be unique.  This
>>> assumption is wrong.  The ELF spec says (version 1.1 page 1-15): "An
>>> object file may have more than one section with the same name."   
>>> We need
>>> to fix the kernel not to rely on a bogus assumption ... but we had  
>>> no
>>> idea how to do that in a way that preserved the backwards  
>>> compatibility
>>> of sections subdirectory.
>>>
>>> I admit that 35dead4235e2b67da7275b4122fed37099c2f462 is a hack,  
>>> but now
>>> the problem has got attention, can we fix it properly?
>>
>> Yep.  The original patch didn't go through me, or we would have had  
>> this
>> discussion back then...
>>
>> The use of section names in sysfs goes back to one Mr. Corbet.  Why  
>> did
>> he do it that way?  Because gdb's add-symbol-file makes the same
>> assumption.  So if we fixed the sysfs somehow, it still wouldn't be
>> useful, since there's no way to tell gdb :(
>>
>> The real answer don't use -ffunction-sections on modules: probably  
>> not
>> as important as the rest of the kernel.  And the new shiny is
>> -flto anyway.
>>
>> And that leaves us with a PA-RISC specific issue, for which we should
>> move the fix to PA-RISC.
>>
>> Thoughts?
>
> Well, we don't have much of a choice.  Our ELF stub jump on 32 bits  
> is a
> PCREL17.  That means once a module size is over 128k there's a  
> chance we
> might not be able to link it because the jump is too big for the
> instruction.  IPV6 is one such big module today, but I'm sure there  
> are
> others.  The only way I know to fix this is to allow the linker to
> insert stubs between functions, so we only fail at linking if a single
> function is >128k big.  The way to do this is -ffunction-sections,
> unless there's something else we could do (all we really need is a way
> to ensure we can insert ELF stubs every 128k).

There is now a config work around for this.  See:
http://www.spinics.net/lists/linux-parisc/msg04521.html

Dave
--
John David Anglin	dave.anglin@bell.net

Re: [PATCH] Add non-zero module sections to sysfs

[James Bottomley]

John David Anglin <dave.anglin@bell.net> wrote:

>On 6-Apr-13, at 6:52 AM, James Bottomley wrote:
>
>> On Sat, 2013-04-06 at 15:22 +1030, Rusty Russell wrote:
>>>> The problem is our assumption that section names be unique.  This
>>>> assumption is wrong.  The ELF spec says (version 1.1 page 1-15):
>"An
>>>> object file may have more than one section with the same name."   
>>>> We need
>>>> to fix the kernel not to rely on a bogus assumption ... but we had 
>
>>>> no
>>>> idea how to do that in a way that preserved the backwards  
>>>> compatibility
>>>> of sections subdirectory.
>>>>
>>>> I admit that 35dead4235e2b67da7275b4122fed37099c2f462 is a hack,  
>>>> but now
>>>> the problem has got attention, can we fix it properly?
>>>
>>> Yep.  The original patch didn't go through me, or we would have had 
>
>>> this
>>> discussion back then...
>>>
>>> The use of section names in sysfs goes back to one Mr. Corbet.  Why 
>
>>> did
>>> he do it that way?  Because gdb's add-symbol-file makes the same
>>> assumption.  So if we fixed the sysfs somehow, it still wouldn't be
>>> useful, since there's no way to tell gdb :(
>>>
>>> The real answer don't use -ffunction-sections on modules: probably  
>>> not
>>> as important as the rest of the kernel.  And the new shiny is
>>> -flto anyway.
>>>
>>> And that leaves us with a PA-RISC specific issue, for which we
>should
>>> move the fix to PA-RISC.
>>>
>>> Thoughts?
>>
>> Well, we don't have much of a choice.  Our ELF stub jump on 32 bits  
>> is a
>> PCREL17.  That means once a module size is over 128k there's a  
>> chance we
>> might not be able to link it because the jump is too big for the
>> instruction.  IPV6 is one such big module today, but I'm sure there  
>> are
>> others.  The only way I know to fix this is to allow the linker to
>> insert stubs between functions, so we only fail at linking if a
>single
>> function is >128k big.  The way to do this is -ffunction-sections,
>> unless there's something else we could do (all we really need is a
>way
>> to ensure we can insert ELF stubs every 128k).
>
>There is now a config work around for this.  See:
>http://www.spinics.net/lists/linux-parisc/msg04521.html

The longcalls config option only works on pa2 doesn't it? Although we could just deprecate pa1.

James
-- 
Sent from my Android phone with K-9 Mail. Please excuse my brevity and top posting.

Re: [PATCH] Add non-zero module sections to sysfs

[John David Anglin]

On 6-Apr-13, at 9:22 PM, James Bottomley wrote:

>
>
> John David Anglin <dave.anglin@bell.net> wrote:
>
>> On 6-Apr-13, at 6:52 AM, James Bottomley wrote:
>>
>>> On Sat, 2013-04-06 at 15:22 +1030, Rusty Russell wrote:
>>>>> The problem is our assumption that section names be unique.  This
>>>>> assumption is wrong.  The ELF spec says (version 1.1 page 1-15):
>> "An
>>>>> object file may have more than one section with the same name."
>>>>> We need
>>>>> to fix the kernel not to rely on a bogus assumption ... but we had
>>
>>>>> no
>>>>> idea how to do that in a way that preserved the backwards
>>>>> compatibility
>>>>> of sections subdirectory.
>>>>>
>>>>> I admit that 35dead4235e2b67da7275b4122fed37099c2f462 is a hack,
>>>>> but now
>>>>> the problem has got attention, can we fix it properly?
>>>>
>>>> Yep.  The original patch didn't go through me, or we would have had
>>
>>>> this
>>>> discussion back then...
>>>>
>>>> The use of section names in sysfs goes back to one Mr. Corbet.  Why
>>
>>>> did
>>>> he do it that way?  Because gdb's add-symbol-file makes the same
>>>> assumption.  So if we fixed the sysfs somehow, it still wouldn't be
>>>> useful, since there's no way to tell gdb :(
>>>>
>>>> The real answer don't use -ffunction-sections on modules: probably
>>>> not
>>>> as important as the rest of the kernel.  And the new shiny is
>>>> -flto anyway.
>>>>
>>>> And that leaves us with a PA-RISC specific issue, for which we
>> should
>>>> move the fix to PA-RISC.
>>>>
>>>> Thoughts?
>>>
>>> Well, we don't have much of a choice.  Our ELF stub jump on 32 bits
>>> is a
>>> PCREL17.  That means once a module size is over 128k there's a
>>> chance we
>>> might not be able to link it because the jump is too big for the
>>> instruction.  IPV6 is one such big module today, but I'm sure there
>>> are
>>> others.  The only way I know to fix this is to allow the linker to
>>> insert stubs between functions, so we only fail at linking if a
>> single
>>> function is >128k big.  The way to do this is -ffunction-sections,
>>> unless there's something else we could do (all we really need is a
>> way
>>> to ensure we can insert ELF stubs every 128k).
>>
>> There is now a config work around for this.  See:
>> http://www.spinics.net/lists/linux-parisc/msg04521.html
>
> The longcalls config option only works on pa2 doesn't it? Although  
> we could just deprecate pa1.


No, it works on pa1.1 but the calls are more efficient on pa2.  On  
linux with a flat space, they should be
about the same in terms of instruction count.  On HP-UX, an additional  
space register load is needed.
For calls within the same space, the be instruction is pretty  
efficient but we never implemented linker
support for it in binutils.  If I recall correctly, it works if the  
call is local to a module but not for global calls.
It does work in HP-UX.

Dave
--
John David Anglin	dave.anglin@bell.net

Re: [PATCH] Add non-zero module sections to sysfs

[Philip Kranz]

Hello.

On Fri, Apr 05, 2013 at 12:07:15PM +0200, James Bottomley wrote:
> Just so you know: this isn't a parisc specific problem.  Gcc produces
> duplicate section names under various circumstances, but the one that
> bites us is -ffunction-sections.  Note that there are proposals to use
> -ffunction-sections on all architectures (so we can garbage collect
> unused functions) in which case you'll induce the bug identified in
> 35dead4235e2b67da7275b4122fed37099c2f462 on every architecture

I am not able to produce an object file with duplicate section names
using gcc on x86. Even with -ffunction-sections, every section gets a
unique name. Is this architecture-specific behaviour of gcc?

Greetings,
Philip

Re: [PATCH] Add non-zero module sections to sysfs

[Rusty Russell]

Philip Kranz <philip.kranz@googlemail.com> writes:
> Hello.
>
> On Fri, Apr 05, 2013 at 12:07:15PM +0200, James Bottomley wrote:
>> Just so you know: this isn't a parisc specific problem.  Gcc produces
>> duplicate section names under various circumstances, but the one that
>> bites us is -ffunction-sections.  Note that there are proposals to use
>> -ffunction-sections on all architectures (so we can garbage collect
>> unused functions) in which case you'll induce the bug identified in
>> 35dead4235e2b67da7275b4122fed37099c2f462 on every architecture
>
> I am not able to produce an object file with duplicate section names
> using gcc on x86. Even with -ffunction-sections, every section gets a
> unique name. Is this architecture-specific behaviour of gcc?

Good point.  ld -r will collapse them into the same section (since gcc
produces them they have to have the same section attributes).

You can do it with --unique, but no arch uses that.  PARISC has a
platform-specific toolchain hack which does that for .text sections.
(Thanks to Alan Modra for that clue...)

Thanks,
Rusty.

Re: [PATCH] Add non-zero module sections to sysfs

[Philip Kranz]

On Mon, Apr 08, 2013 at 01:44:45PM +0930, Rusty Russell wrote:
> Philip Kranz <philip.kranz@googlemail.com> writes:
> > I am not able to produce an object file with duplicate section names
> > using gcc on x86. Even with -ffunction-sections, every section gets a
> > unique name. Is this architecture-specific behaviour of gcc?
> 
> Good point.  ld -r will collapse them into the same section (since gcc
> produces them they have to have the same section attributes).
> 
> You can do it with --unique, but no arch uses that.  PARISC has a
> platform-specific toolchain hack which does that for .text sections.
> (Thanks to Alan Modra for that clue...)

So that problem is indeed platform-specific. If it is safe to assume
that kernel modules don't have duplicate section names (except on
PARISC), it would make sense to simply move the check for empty sections
to arch/parisc as you suggested.

James, what do you think about that?

Greetings,
Philip

Re: [PATCH] Add non-zero module sections to sysfs

[Philip Kranz]

On Mon, Apr 08, 2013 at 01:44:45PM +0930, Rusty Russell wrote:
> Philip Kranz <philip.kranz@googlemail.com> writes:
> > Hello.
> >
> > On Fri, Apr 05, 2013 at 12:07:15PM +0200, James Bottomley wrote:
> >> Just so you know: this isn't a parisc specific problem.  Gcc produces
> >> duplicate section names under various circumstances, but the one that
> >> bites us is -ffunction-sections.  Note that there are proposals to use
> >> -ffunction-sections on all architectures (so we can garbage collect
> >> unused functions) in which case you'll induce the bug identified in
> >> 35dead4235e2b67da7275b4122fed37099c2f462 on every architecture
> >
> > I am not able to produce an object file with duplicate section names
> > using gcc on x86. Even with -ffunction-sections, every section gets a
> > unique name. Is this architecture-specific behaviour of gcc?
> 
> Good point.  ld -r will collapse them into the same section (since gcc
> produces them they have to have the same section attributes).
> 
> You can do it with --unique, but no arch uses that.  PARISC has a
> platform-specific toolchain hack which does that for .text sections.
> (Thanks to Alan Modra for that clue...)

Just for clarification, as we are currently preparing a patch set that
depends on this: Would the patch below be an acceptable solution for
this?

Thanks,
Philip



diff --git a/arch/parisc/kernel/module.c b/arch/parisc/kernel/module.c
index 2a625fb..1fd4411 100644
--- a/arch/parisc/kernel/module.c
+++ b/arch/parisc/kernel/module.c
@@ -341,6 +341,11 @@ int module_frob_arch_sections(CONST Elf_Ehdr *hdr,
                ".PARISC.unwind", 14) == 0)
            me->arch.unwind_section = i;
 
+       /* we produce multiple, empty .text sections, and kallsyms
+       * gets upset.  make non-alloc so it doesn't see them. */
+       if (sechdrs[i].sh_size == 0)
+           sechdrs[i].sh_flags &= ~SHF_ALLOC;
+
        if (sechdrs[i].sh_type != SHT_RELA)
            continue;
 
diff --git a/kernel/module.c b/kernel/module.c
index 3c2c72d..42e0d5a 100644
--- a/kernel/module.c
+++ b/kernel/module.c
@@ -1316,7 +1316,7 @@ resolve_symbol_wait(struct module *mod,
 #ifdef CONFIG_KALLSYMS
 static inline bool sect_empty(const Elf_Shdr *sect)
 {
-   return !(sect->sh_flags & SHF_ALLOC) || sect->sh_size == 0;
+   return !(sect->sh_flags & SHF_ALLOC);
 }
 
 struct module_sect_attr