Skip to contentSkip to conversation

module: fix __module_ref_addr()

PATCH In review Mathieu Desnoyers 4 participants 11 messages Mar 27, 2010 – Apr 20, 2010 Module support

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 module: fix __module_ref_addr()

__module_ref_addr() should use per_cpu_ptr() to obfuscate the pointer (RELOC_HIDE is needed for per cpu pointers).

This non-standard per-cpu pointer use has been introduced by commit 720eba31f47aeade8ec130ca7f4353223c49170f

It causes a NULL pointer exception on some configurations when CONFIG_TRACING is enabled on 2.6.33. This patch fixes the problem (acknowledged by Randy who reported the bug).

It did not appear to hurt previously because most of the accesses were done through local_inc, which probably obfuscated the access enough that no compiler optimizations were done. But with local_read() done when CONFIG_TRACING is active, this becomes a problem. Non-CONFIG_TRACING is probably affected as well (module.c contains local_set and local_read that use __module_ref_addr()), but I guess nobody noticed because we've been lucky enough that the compiler did not generate the inappropriate optimization pattern there.

This patch should be queued for the 2.6.29.x through 2.6.33.x stable branches. (tested on 2.6.33.1 x86_64)

The __module_ref_addr() problem disappears in 2.6.34-rc kernels because these percpu accesses were re-factored.

It makes me wonder about other non-standard uses of per_cpu_offset: there is one in module.c and two in lockdep.c, which are still in 2.6.34-rc. This should probably be fixed by the code authors in separate patches.

lockdep.c: commit 8e18257d29238311e82085152741f0c3aa18b74d module.c: commit 6b588c18f8dacfa6d7957c33c5ff832096e752d3

CC Eric Dumazet <[email protected]>
CC Rusty Russell <[email protected]>
CC Peter Zijlstra <[email protected]>
CC Tejun Heo <[email protected]>
CC Ingo Molnar <[email protected]>
CC Andrew Morton <[email protected]>
CC Linus Torvalds <[email protected]>
CC Greg Kroah-Hartman <[email protected]>
CC Steven Rostedt <[email protected]>
---
 include/linux/module.h |    2 +-
 kernel/module.c        |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
include/linux/module.h +1 -1
Index: linux-2.6-lttng/include/linux/module.h
===================================================================
@@ -467,7 +467,7 @@ void symbol_put_addr(void *addr);
 static inline local_t *__module_ref_addr(struct module *mod, int cpu)
 {
 #ifdef CONFIG_SMP
-	return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+	return (local_t *) per_cpu_ptr(mod->refptr, cpu);
 #else
 	return &mod->ref;
 #endif
Signature · 4 lines
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
#1 lore raw

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

-- Steve

Quoted 61 lines · Mathieu Desnoyers

__module_ref_addr() should use per_cpu_ptr() to obfuscate the pointer (RELOC_HIDE is needed for per cpu pointers).

This non-standard per-cpu pointer use has been introduced by commit 720eba31f47aeade8ec130ca7f4353223c49170f

It causes a NULL pointer exception on some configurations when CONFIG_TRACING is enabled on 2.6.33. This patch fixes the problem (acknowledged by Randy who reported the bug).

It did not appear to hurt previously because most of the accesses were done through local_inc, which probably obfuscated the access enough that no compiler optimizations were done. But with local_read() done when CONFIG_TRACING is active, this becomes a problem. Non-CONFIG_TRACING is probably affected as well (module.c contains local_set and local_read that use __module_ref_addr()), but I guess nobody noticed because we've been lucky enough that the compiler did not generate the inappropriate optimization pattern there.

This patch should be queued for the 2.6.29.x through 2.6.33.x stable branches. (tested on 2.6.33.1 x86_64)

The __module_ref_addr() problem disappears in 2.6.34-rc kernels because these percpu accesses were re-factored.

It makes me wonder about other non-standard uses of per_cpu_offset: there is one in module.c and two in lockdep.c, which are still in 2.6.34-rc. This should probably be fixed by the code authors in separate patches.

lockdep.c: commit 8e18257d29238311e82085152741f0c3aa18b74d module.c: commit 6b588c18f8dacfa6d7957c33c5ff832096e752d3

CC Eric Dumazet <[email protected]>
CC Rusty Russell <[email protected]>
CC Peter Zijlstra <[email protected]>
CC Tejun Heo <[email protected]>
CC Ingo Molnar <[email protected]>
CC Andrew Morton <[email protected]>
CC Linus Torvalds <[email protected]>
CC Greg Kroah-Hartman <[email protected]>
CC Steven Rostedt <[email protected]>
---
 include/linux/module.h |    2 +-
 kernel/module.c        |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

Index: linux-2.6-lttng/include/linux/module.h
===================================================================

include/linux/module.h +1 -1
@@ -467,7 +467,7 @@ void symbol_put_addr(void *addr);
 static inline local_t *__module_ref_addr(struct module *mod, int cpu)
 {
 #ifdef CONFIG_SMP
-	return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+	return (local_t *) per_cpu_ptr(mod->refptr, cpu);
 #else
 	return &mod->ref;
 #endif
#2 lore raw
Quoted 2 lines · * Steven Rostedt ([email protected]) wrote:

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Thanks,

Mathieu

Quoted 68 lines

-- Steve

Quoted 61 lines · Mathieu Desnoyers

__module_ref_addr() should use per_cpu_ptr() to obfuscate the pointer (RELOC_HIDE is needed for per cpu pointers).

This non-standard per-cpu pointer use has been introduced by commit 720eba31f47aeade8ec130ca7f4353223c49170f

It causes a NULL pointer exception on some configurations when CONFIG_TRACING is enabled on 2.6.33. This patch fixes the problem (acknowledged by Randy who reported the bug).

It did not appear to hurt previously because most of the accesses were done through local_inc, which probably obfuscated the access enough that no compiler optimizations were done. But with local_read() done when CONFIG_TRACING is active, this becomes a problem. Non-CONFIG_TRACING is probably affected as well (module.c contains local_set and local_read that use __module_ref_addr()), but I guess nobody noticed because we've been lucky enough that the compiler did not generate the inappropriate optimization pattern there.

This patch should be queued for the 2.6.29.x through 2.6.33.x stable branches. (tested on 2.6.33.1 x86_64)

The __module_ref_addr() problem disappears in 2.6.34-rc kernels because these percpu accesses were re-factored.

It makes me wonder about other non-standard uses of per_cpu_offset: there is one in module.c and two in lockdep.c, which are still in 2.6.34-rc. This should probably be fixed by the code authors in separate patches.

lockdep.c: commit 8e18257d29238311e82085152741f0c3aa18b74d module.c: commit 6b588c18f8dacfa6d7957c33c5ff832096e752d3

CC Eric Dumazet <[email protected]>
CC Rusty Russell <[email protected]>
CC Peter Zijlstra <[email protected]>
CC Tejun Heo <[email protected]>
CC Ingo Molnar <[email protected]>
CC Andrew Morton <[email protected]>
CC Linus Torvalds <[email protected]>
CC Greg Kroah-Hartman <[email protected]>
CC Steven Rostedt <[email protected]>
---
 include/linux/module.h |    2 +-
 kernel/module.c        |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

Index: linux-2.6-lttng/include/linux/module.h
===================================================================

include/linux/module.h +1 -1
@@ -467,7 +467,7 @@ void symbol_put_addr(void *addr);
 static inline local_t *__module_ref_addr(struct module *mod, int cpu)
 {
 #ifdef CONFIG_SMP
-	return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+	return (local_t *) per_cpu_ptr(mod->refptr, cpu);
 #else
 	return &mod->ref;
 #endif
Signature · 4 lines
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
#3 lore raw
Quoted 6 lines · Mathieu Desnoyers
Quoted 2 lines · * Steven Rostedt ([email protected]) wrote:

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

thanks,

greg k-h

#4 lore raw
Quoted 11 lines · Greg KH
Quoted 6 lines · Mathieu Desnoyers
Quoted 2 lines · * Steven Rostedt ([email protected]) wrote:

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

Looks like it was this commit:

commit e1783a240f491fb233f04edc042e16b18a7a79ba

Author: Christoph Lameter <[email protected]>
Date:   Tue Jan 5 15:34:50 2010 +0900
module: Use this_cpu_xx to dynamically allocate counters

Mathieu's fix was:

- return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+ return (local_t *) per_cpu_ptr(mod->refptr, cpu);

As Mathieu states in his change log, the bug is that the mod->refptr is outside the assembly obfuscation of the per_cpu_offset(). This allows the compiler to optimize and cause a NULL pointer dereference with the manipulation of per cpu data.

Christoph Lameter's change fixes this bug as a side effect:

-static inline local_t *__module_ref_addr(struct module *mod, int cpu) -{
-#ifdef CONFIG_SMP
- return (local_t *) (mod->refptr + per_cpu_offset(cpu));
-#else
- return &mod->ref;
-#endif
-}
-
/* Sometimes we know we already have a refcount, and it's easier not
to handle the error case (which only happens with rmmod --wait). */
static inline void __module_get(struct module *module)

 {
        if (module) {
-               unsigned int cpu = get_cpu();
-               local_inc(__module_ref_addr(module, cpu));
+               preempt_disable();
+               __this_cpu_inc(module->refptr->count);

trace_module_get(module, _THIS_IP_,
- local_read(__module_ref_addr(module, cpu)));
- put_cpu();
+ __this_cpu_read(module->refptr->count));

+               preempt_enable();
        }

}

By removing the buggy code all together.

-- Steve

#5 lore raw
Quoted 58 lines · * Steven Rostedt ([email protected]) wrote:
Quoted 11 lines · Greg KH
Quoted 6 lines · Mathieu Desnoyers
Quoted 2 lines · * Steven Rostedt ([email protected]) wrote:

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

Looks like it was this commit:

commit e1783a240f491fb233f04edc042e16b18a7a79ba

Author: Christoph Lameter <[email protected]>
Date:   Tue Jan 5 15:34:50 2010 +0900
module: Use this_cpu_xx to dynamically allocate counters

Mathieu's fix was:

include/linux/module.h +1 -1
-   return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+   return (local_t *) per_cpu_ptr(mod->refptr, cpu);

As Mathieu states in his change log, the bug is that the mod->refptr is outside the assembly obfuscation of the per_cpu_offset(). This allows the compiler to optimize and cause a NULL pointer dereference with the manipulation of per cpu data.

Christoph Lameter's change fixes this bug as a side effect:

(quoted diff) +4 -13
-static inline local_t *__module_ref_addr(struct module *mod, int cpu)
-{
-#ifdef CONFIG_SMP
-       return (local_t *) (mod->refptr + per_cpu_offset(cpu));
-#else
-       return &mod->ref;
-#endif
-}
-
 /* Sometimes we know we already have a refcount, and it's easier not
    to handle the error case (which only happens with rmmod --wait). */
 static inline void __module_get(struct module *module)
 {
        if (module) {
-               unsigned int cpu = get_cpu();
-               local_inc(__module_ref_addr(module, cpu));
+               preempt_disable();
+               __this_cpu_inc(module->refptr->count);
                trace_module_get(module, _THIS_IP_,
-                                local_read(__module_ref_addr(module, cpu)));
-               put_cpu();
+                                __this_cpu_read(module->refptr->count));
+               preempt_enable();
        }
 }
Quoted 1 line

By removing the buggy code all together.

Exactly. Steven has beaten me on the start line on this one. ;)

Thanks,

Mathieu

Quoted 4 lines

-- Steve

Signature · 4 lines
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
#6 lore raw
Quoted 59 lines · Mathieu Desnoyers
Quoted 58 lines · * Steven Rostedt ([email protected]) wrote:
Quoted 11 lines · Greg KH
Quoted 6 lines · Mathieu Desnoyers
Quoted 2 lines · * Steven Rostedt ([email protected]) wrote:

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

Looks like it was this commit:

commit e1783a240f491fb233f04edc042e16b18a7a79ba

Author: Christoph Lameter <[email protected]>
Date:   Tue Jan 5 15:34:50 2010 +0900
module: Use this_cpu_xx to dynamically allocate counters

Mathieu's fix was:

include/linux/module.h +1 -1
-   return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+   return (local_t *) per_cpu_ptr(mod->refptr, cpu);

As Mathieu states in his change log, the bug is that the mod->refptr is outside the assembly obfuscation of the per_cpu_offset(). This allows the compiler to optimize and cause a NULL pointer dereference with the manipulation of per cpu data.

Christoph Lameter's change fixes this bug as a side effect:

(quoted diff) +4 -13
-static inline local_t *__module_ref_addr(struct module *mod, int cpu)
-{
-#ifdef CONFIG_SMP
-       return (local_t *) (mod->refptr + per_cpu_offset(cpu));
-#else
-       return &mod->ref;
-#endif
-}
-
 /* Sometimes we know we already have a refcount, and it's easier not
    to handle the error case (which only happens with rmmod --wait). */
 static inline void __module_get(struct module *module)
 {
        if (module) {
-               unsigned int cpu = get_cpu();
-               local_inc(__module_ref_addr(module, cpu));
+               preempt_disable();
+               __this_cpu_inc(module->refptr->count);
                trace_module_get(module, _THIS_IP_,
-                                local_read(__module_ref_addr(module, cpu)));
-               put_cpu();
+                                __this_cpu_read(module->refptr->count));
+               preempt_enable();
        }
 }
Quoted 3 lines
Quoted 1 line

By removing the buggy code all together.

Exactly. Steven has beaten me on the start line on this one. ;)

Ok, I'm totally confused now :(

What patch should I apply to a stable release, and which stable release?

thanks,

greg k-h

author Tested-by Randy Dunlap
#7 lore raw
Quoted 64 lines · * Greg KH ([email protected]) wrote:
Quoted 59 lines · Mathieu Desnoyers
Quoted 58 lines · * Steven Rostedt ([email protected]) wrote:
Quoted 11 lines · Greg KH
Quoted 6 lines · Mathieu Desnoyers
Quoted 2 lines · * Steven Rostedt ([email protected]) wrote:

This patch does not apply to 2.6.34-rc, and the code in upstream looks to have been fixed. Should this go to stable?

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

Looks like it was this commit:

commit e1783a240f491fb233f04edc042e16b18a7a79ba

Author: Christoph Lameter <[email protected]>
Date:   Tue Jan 5 15:34:50 2010 +0900
module: Use this_cpu_xx to dynamically allocate counters

Mathieu's fix was:

(quoted diff) +1 -1
-   return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+   return (local_t *) per_cpu_ptr(mod->refptr, cpu);

As Mathieu states in his change log, the bug is that the mod->refptr is outside the assembly obfuscation of the per_cpu_offset(). This allows the compiler to optimize and cause a NULL pointer dereference with the manipulation of per cpu data.

Christoph Lameter's change fixes this bug as a side effect:

(quoted diff) +4 -13
-static inline local_t *__module_ref_addr(struct module *mod, int cpu)
-{
-#ifdef CONFIG_SMP
-       return (local_t *) (mod->refptr + per_cpu_offset(cpu));
-#else
-       return &mod->ref;
-#endif
-}
-
 /* Sometimes we know we already have a refcount, and it's easier not
    to handle the error case (which only happens with rmmod --wait). */
 static inline void __module_get(struct module *module)
 {
        if (module) {
-               unsigned int cpu = get_cpu();
-               local_inc(__module_ref_addr(module, cpu));
+               preempt_disable();
+               __this_cpu_inc(module->refptr->count);
                trace_module_get(module, _THIS_IP_,
-                                local_read(__module_ref_addr(module, cpu)));
-               put_cpu();
+                                __this_cpu_read(module->refptr->count));
+               preempt_enable();
        }
 }
Quoted 3 lines
Quoted 1 line

By removing the buggy code all together.

Exactly. Steven has beaten me on the start line on this one. ;)

Quoted 7 lines

Ok, I'm totally confused now :(

What patch should I apply to a stable release, and which stable release?

thanks,

greg k-h

Unless you take all the per cpu refactoring from 2.6.34-rc, the following patch should be taken into -stable. Please see the changelog for the list of stable releases that need it.

module: fix __module_ref_addr()

__module_ref_addr() should use per_cpu_ptr() to obfuscate the pointer (RELOC_HIDE is needed for per cpu pointers).

This non-standard per-cpu pointer use has been introduced by commit 720eba31f47aeade8ec130ca7f4353223c49170f

It causes a NULL pointer exception on some configurations when CONFIG_TRACING is enabled on 2.6.33. This patch fixes the problem (acknowledged by Randy who reported the bug).

It did not appear to hurt previously because most of the accesses were done through local_inc, which probably obfuscated the access enough that no compiler optimizations were done. But with local_read() done when CONFIG_TRACING is active, this becomes a problem. Non-CONFIG_TRACING is probably affected as well (module.c contains local_set and local_read that use __module_ref_addr()), but I guess nobody noticed because we've been lucky enough that the compiler did not generate the inappropriate optimization pattern there.

This patch should be queued for the 2.6.29.x through 2.6.33.x stable branches. (tested on 2.6.33.1 x86_64)

The __module_ref_addr() problem disappears in 2.6.34-rc kernels because these percpu accesses were re-factored.

CC Eric Dumazet <[email protected]>
CC Rusty Russell <[email protected]>
CC Peter Zijlstra <[email protected]>
CC Tejun Heo <[email protected]>
CC Ingo Molnar <[email protected]>
CC Andrew Morton <[email protected]>
CC Linus Torvalds <[email protected]>
CC Greg Kroah-Hartman <[email protected]>
CC Steven Rostedt <[email protected]>
---
 include/linux/module.h |    2 +-
 kernel/module.c        |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
include/linux/module.h +1 -1
Index: linux-2.6-lttng/include/linux/module.h
===================================================================
@@ -467,7 +467,7 @@ void symbol_put_addr(void *addr);
 static inline local_t *__module_ref_addr(struct module *mod, int cpu)
 {
 #ifdef CONFIG_SMP
-	return (local_t *) (mod->refptr + per_cpu_offset(cpu));
+	return (local_t *) per_cpu_ptr(mod->refptr, cpu);
 #else
 	return &mod->ref;
 #endif
Signature · 4 lines
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
#8 lore raw
In reply to Greg KH

Hello, Greg.

Quoted 5 lines · Greg KH
Quoted 2 lines

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

For 2.6.34, the following two patches should remove the problem. They're not in mainline yet but should appear in linux-next soon and after a few days, I'll push them to Linus.

http://thread.gmane.org/gmane.linux.kernel/958794/focus=959493

Thanks.

Signature · 1 line
tejun
#9 lore raw
Quoted 6 lines · Tejun Heo

Hello, Greg.

Quoted 3 lines · Greg KH
Quoted 2 lines

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Quoted 8 lines
Quoted 2 lines

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

For 2.6.34, the following two patches should remove the problem. They're not in mainline yet but should appear in linux-next soon and after a few days, I'll push them to Linus.

http://thread.gmane.org/gmane.linux.kernel/958794/focus=959493

Oh, Steven is right. The ones which are removed by the above patch are different ones. That said, it might be a good idea to apply fixes for them to -stable too. Mathieu, are you interested in submitting similar -stable fixes for kernel/module.c::percpu_modcopy() and lockdep.c::static_obj()?

Thanks.

Signature · 1 line
tejun
#10 lore raw
Quoted 14 lines · * Tejun Heo ([email protected]) wrote:
Quoted 13 lines · Tejun Heo

Hello, Greg.

Quoted 5 lines · Greg KH
Quoted 2 lines

Yes. 2.6.34-rc does not have this issue anymore, but the patch is needed in -stable.

Why is this not in .34-rc2? Can you find the specific patch in Linus's tree that solves this and let [email protected] know about it?

For 2.6.34, the following two patches should remove the problem. They're not in mainline yet but should appear in linux-next soon and after a few days, I'll push them to Linus.

Quoted 7 lines
Quoted 1 line

http://thread.gmane.org/gmane.linux.kernel/958794/focus=959493

Oh, Steven is right. The ones which are removed by the above patch are different ones. That said, it might be a good idea to apply fixes for them to -stable too. Mathieu, are you interested in submitting similar -stable fixes for kernel/module.c::percpu_modcopy() and lockdep.c::static_obj()?

Sure, will do. I'm doing a bit of testing and I'll post these patches shortly.

Thanks,

Mathieu

Quoted 5 lines

Thanks.

Signature · 1 line
tejun
Signature · 4 lines
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com