Re: [Patch 4/4] module: trim exception table in module_free()

Re: [Patch 4/4] module: trim exception table in module_free()

[Rusty Russell]

On Tue, 26 May 2009 06:05:39 pm Amerigo Wang wrote:
>  void module_free(struct module *mod, void *module_region)
>  {
>  	vfree(module_region);
> -	/* FIXME: If module_region == mod->init_region, trim exception
> -	   table entries. */
> +	if (module_region == mod->module_init)
> +		mod->num_exentries = 0;
>  }

Hi Amerigo,

   This looks wrong.  The extable covers both init and core exception entries.   
We want to remove the ones in the module_init section.  The good news is that 
it's sorted, so they're either at the start or the end (except sparc 32).

  The bad news is that this is really a generic problem, and deserves a 
generic fix.  That's easy for archs which use the generic sort_extable(), but 
alpha, ia64 and sparc(32) will need to define their own trim_extable().

Thanks!
Rusty.

Re: [Patch 4/4] module: trim exception table in module_free()

[Amerigo Wang]

Rusty Russell wrote:
> On Tue, 26 May 2009 06:05:39 pm Amerigo Wang wrote:
>   
>>  void module_free(struct module *mod, void *module_region)
>>  {
>>  	vfree(module_region);
>> -	/* FIXME: If module_region == mod->init_region, trim exception
>> -	   table entries. */
>> +	if (module_region == mod->module_init)
>> +		mod->num_exentries = 0;
>>  }
>>     
>
> Hi Amerigo,
>
>    This looks wrong.  The extable covers both init and core exception entries.   
> We want to remove the ones in the module_init section.  The good news is that 
> it's sorted, so they're either at the start or the end (except sparc 32).
>   

Hi, Rusty.

Yes? The extable of a module is in '__ex_table' section, and during the 
section transfer, one
section will be either in module_init or module_core, so its entries are 
only in one of them,
not both, right?

Thank you.

Re: [Patch 4/4] module: trim exception table in module_free()

[Rusty Russell]

On Wed, 27 May 2009 12:41:00 pm Amerigo Wang wrote:
> Rusty Russell wrote:
> > On Tue, 26 May 2009 06:05:39 pm Amerigo Wang wrote:
> >>  void module_free(struct module *mod, void *module_region)
> >>  {
> >>  	vfree(module_region);
> >> -	/* FIXME: If module_region == mod->init_region, trim exception
> >> -	   table entries. */
> >> +	if (module_region == mod->module_init)
> >> +		mod->num_exentries = 0;
> >>  }
> >
> > Hi Amerigo,
> >
> >    This looks wrong.  The extable covers both init and core exception
> > entries. We want to remove the ones in the module_init section.  The good
> > news is that it's sorted, so they're either at the start or the end
> > (except sparc 32).
>
> Hi, Rusty.
>
> Yes? The extable of a module is in '__ex_table' section, and during the
> section transfer, one
> section will be either in module_init or module_core, so its entries are
> only in one of them,
> not both, right?

32-bit example:

#include <linux/module.h>
#include <linux/uaccess.h>
#include <linux/moduleparam.h>

static unsigned long uaddr;
module_param(uaddr, ulong, 0600);

void extable_not_init(u64 val)
{
	__put_user(val, (u64 *)uaddr);
}

static int __init init(void)
{
	__put_user(0, (u64 *)uaddr);
	return 0;
}
module_init(init);

__ex_table ends up with two entries:

Contents of section __ex_table:
 0000 0c000000 00000000 0e000000 00000000  ................
 0010 10000000 0a000000 12000000 0a000000  ................

The first is for the __put_user in .text (extable_not_init()) and the second is 
for the one in .init.text (init()).

Depending on how the module gets allocated, the one referring to .init.text 
may be first or last.

(You can see here why we haven't fixed this: exceptions in __init in modules 
are rare, perhaps non-existent).

Hope that clarifies?
Rusty.

Re: [Patch 4/4] module: trim exception table in module_free()

[Amerigo Wang]

Rusty Russell wrote:
> __ex_table ends up with two entries:
>
> Contents of section __ex_table:
>  0000 0c000000 00000000 0e000000 00000000  ................
>  0010 10000000 0a000000 12000000 0a000000  ................
>
> The first is for the __put_user in .text (extable_not_init()) and the second is 
> for the one in .init.text (init()).
>
> Depending on how the module gets allocated, the one referring to .init.text 
> may be first or last.
>   

Hmm, I understand now. The problem is that we don't know which entries
are for __init, and which are not...

> (You can see here why we haven't fixed this: exceptions in __init in modules 
> are rare, perhaps non-existent).
>   
Agreed.
Is it possible to put extable for __init in a separate section?

Thanks for your explanation!

Re: [Patch 4/4] module: trim exception table in module_free()

[Amerigo Wang]

Rusty Russell wrote:
> __ex_table ends up with two entries:
>
> Contents of section __ex_table:
>  0000 0c000000 00000000 0e000000 00000000  ................
>  0010 10000000 0a000000 12000000 0a000000  ................
>
> The first is for the __put_user in .text (extable_not_init()) and the second is 
> for the one in .init.text (init()).
>
> Depending on how the module gets allocated, the one referring to .init.text 
> may be first or last.
>   

Hmm, how about the following? :-)

struct exception_table_entry *p = mod->extable;

for (;p <= mod->extable+mod->num_exentries; p++ )
         if (with_in_module_init(p->insn, mod))
                   trim_it(p);

Re: [Patch 4/4] module: trim exception table in module_free()

[Rusty Russell]

On Wed, 27 May 2009 05:16:11 pm Amerigo Wang wrote:
> Rusty Russell wrote:
> > __ex_table ends up with two entries:
> >
> > Contents of section __ex_table:
> >  0000 0c000000 00000000 0e000000 00000000  ................
> >  0010 10000000 0a000000 12000000 0a000000  ................
> >
> > The first is for the __put_user in .text (extable_not_init()) and the
> > second is for the one in .init.text (init()).
> >
> > Depending on how the module gets allocated, the one referring to
> > .init.text may be first or last.
>
> Hmm, how about the following? :-)
>
> struct exception_table_entry *p = mod->extable;
>
> for (;p <= mod->extable+mod->num_exentries; p++ )
>          if (with_in_module_init(p->insn, mod))
>                    trim_it(p);

More like this:

void trim_init_extable(struct module *m)
{
	/* Since entries are sorted, init entries are at the start... */
	while (m->num_exentries && within_module_init(m->extable[0].insn)) {
		m->extable++;
		m->num_exentries--;
	}

	/* ... or the end. */
	while (m->num_exentries && within_module_init(m->extable[m->num_exentries-1].insn))
		m->num_exentries--;
}

Cheers,
Rusty.

Re: [Patch 4/4] module: trim exception table in module_free()

[Amerigo Wang]

Rusty Russell wrote:
> On Wed, 27 May 2009 05:16:11 pm Amerigo Wang wrote:
>   
>> Rusty Russell wrote:
>>     
>>> __ex_table ends up with two entries:
>>>
>>> Contents of section __ex_table:
>>>  0000 0c000000 00000000 0e000000 00000000  ................
>>>  0010 10000000 0a000000 12000000 0a000000  ................
>>>
>>> The first is for the __put_user in .text (extable_not_init()) and the
>>> second is for the one in .init.text (init()).
>>>
>>> Depending on how the module gets allocated, the one referring to
>>> .init.text may be first or last.
>>>       
>> Hmm, how about the following? :-)
>>
>> struct exception_table_entry *p = mod->extable;
>>
>> for (;p <= mod->extable+mod->num_exentries; p++ )
>>          if (with_in_module_init(p->insn, mod))
>>                    trim_it(p);
>>     
>
> More like this:
>
> void trim_init_extable(struct module *m)
> {
> 	/* Since entries are sorted, init entries are at the start... */
> 	while (m->num_exentries && within_module_init(m->extable[0].insn)) {
> 		m->extable++;
> 		m->num_exentries--;
> 	}
>
> 	/* ... or the end. */
> 	while (m->num_exentries && within_module_init(m->extable[m->num_exentries-1].insn))
> 		m->num_exentries--;
> }
>   

Great! Thank you!
But does this works on all arch? Or only except sparc32?

Re: [Patch 4/4] module: trim exception table in module_free()

[Rusty Russell]

On Sun, 31 May 2009 02:45:44 pm Amerigo Wang wrote:
> Rusty Russell wrote:
> > On Wed, 27 May 2009 05:16:11 pm Amerigo Wang wrote:
> >> Rusty Russell wrote:
> >>> __ex_table ends up with two entries:
> >>>
> >>> Contents of section __ex_table:
> >>>  0000 0c000000 00000000 0e000000 00000000  ................
> >>>  0010 10000000 0a000000 12000000 0a000000  ................
> >>>
> >>> The first is for the __put_user in .text (extable_not_init()) and the
> >>> second is for the one in .init.text (init()).
> >>>
> >>> Depending on how the module gets allocated, the one referring to
> >>> .init.text may be first or last.
> >>
> >> Hmm, how about the following? :-)
> >>
> >> struct exception_table_entry *p = mod->extable;
> >>
> >> for (;p <= mod->extable+mod->num_exentries; p++ )
> >>          if (with_in_module_init(p->insn, mod))
> >>                    trim_it(p);
> >
> > More like this:
> >
> > void trim_init_extable(struct module *m)
> > {
> > 	/* Since entries are sorted, init entries are at the start... */
> > 	while (m->num_exentries && within_module_init(m->extable[0].insn)) {
> > 		m->extable++;
> > 		m->num_exentries--;
> > 	}
> >
> > 	/* ... or the end. */
> > 	while (m->num_exentries &&
> > within_module_init(m->extable[m->num_exentries-1].insn))
> > m->num_exentries--;
> > }
>
> Great! Thank you!
> But does this works on all arch? Or only except sparc32?

Variations will work on all archs.  sparc32 doesn't sort its exception table 
AFAICT, so it will need something more sophisticated (ISTR they have exception 
areas, so perhaps setting length to 0 would 'delete' them).

Hope that helps,
Rusty.