liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Mathieu Desnoyers]

Hi Paul, Will, Peter,

I noticed in this discussion https://lkml.org/lkml/2021/4/16/118 that LTO
is able to break rcu_dereference. This seems to be taken care of by
arch/arm64/include/asm/rwonce.h on arm64 in the Linux kernel tree.

In the liburcu user-space library, we have this comment near rcu_dereference() in
include/urcu/static/pointer.h:

 * The compiler memory barrier in CMM_LOAD_SHARED() ensures that value-speculative
 * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
 * data read before the pointer read by speculating the value of the pointer.
 * Correct ordering is ensured because the pointer is read as a volatile access.
 * This acts as a global side-effect operation, which forbids reordering of
 * dependent memory operations. Note that such concern about dependency-breaking
 * optimizations will eventually be taken care of by the "memory_order_consume"
 * addition to forthcoming C++ standard.

(note: CMM_LOAD_SHARED() is the equivalent of READ_ONCE(), but was introduced in
liburcu as a public API before READ_ONCE() existed in the Linux kernel)

Peter tells me the "memory_order_consume" is not something which can be used today.
Any information on its status at C/C++ standard levels and implementation-wise ?

Pragmatically speaking, what should we change in liburcu to ensure we don't generate
broken code when LTO is enabled ? I suspect there are a few options here:

1) Fail to build if LTO is enabled,
2) Generate slower code for rcu_dereference, either on all architectures or only
   on weakly-ordered architectures,
3) Generate different code depending on whether LTO is enabled or not. AFAIU this would only
   work if every compile unit is aware that it will end up being optimized with LTO. Not sure
   how this could be done in the context of user-space.
4) [ Insert better idea here. ]

Thoughts ?

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

Re: liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Peter Zijlstra]

On Fri, Apr 16, 2021 at 10:52:16AM -0400, Mathieu Desnoyers wrote:
> Hi Paul, Will, Peter,
> 
> I noticed in this discussion https://lkml.org/lkml/2021/4/16/118 that LTO
> is able to break rcu_dereference. This seems to be taken care of by
> arch/arm64/include/asm/rwonce.h on arm64 in the Linux kernel tree.
> 
> In the liburcu user-space library, we have this comment near rcu_dereference() in
> include/urcu/static/pointer.h:
> 
>  * The compiler memory barrier in CMM_LOAD_SHARED() ensures that value-speculative
>  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
>  * data read before the pointer read by speculating the value of the pointer.
>  * Correct ordering is ensured because the pointer is read as a volatile access.
>  * This acts as a global side-effect operation, which forbids reordering of
>  * dependent memory operations. Note that such concern about dependency-breaking
>  * optimizations will eventually be taken care of by the "memory_order_consume"
>  * addition to forthcoming C++ standard.
> 
> (note: CMM_LOAD_SHARED() is the equivalent of READ_ONCE(), but was introduced in
> liburcu as a public API before READ_ONCE() existed in the Linux kernel)
> 
> Peter tells me the "memory_order_consume" is not something which can be used today.
> Any information on its status at C/C++ standard levels and implementation-wise ?
> 
> Pragmatically speaking, what should we change in liburcu to ensure we don't generate
> broken code when LTO is enabled ? I suspect there are a few options here:
> 
> 1) Fail to build if LTO is enabled,
> 2) Generate slower code for rcu_dereference, either on all architectures or only
>    on weakly-ordered architectures,
> 3) Generate different code depending on whether LTO is enabled or not. AFAIU this would only
>    work if every compile unit is aware that it will end up being optimized with LTO. Not sure
>    how this could be done in the context of user-space.
> 4) [ Insert better idea here. ]
> 
> Thoughts ?

Using memory_order_acquire is safe; and is basically what Will did for
ARM64.

The problematic tranformations are possible even without LTO, although
less likely due to less visibility, but everybody agrees they're
possible and allowed.

OTOH we do not have a positive sighting of it actually happening (I
think), we're all just being cautious and not willing to debug the
resulting wreckage if it does indeed happen.

Re: liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Paul E. McKenney]

On Fri, Apr 16, 2021 at 05:17:11PM +0200, Peter Zijlstra wrote:
> On Fri, Apr 16, 2021 at 10:52:16AM -0400, Mathieu Desnoyers wrote:
> > Hi Paul, Will, Peter,
> > 
> > I noticed in this discussion https://lkml.org/lkml/2021/4/16/118 that LTO
> > is able to break rcu_dereference. This seems to be taken care of by
> > arch/arm64/include/asm/rwonce.h on arm64 in the Linux kernel tree.
> > 
> > In the liburcu user-space library, we have this comment near rcu_dereference() in
> > include/urcu/static/pointer.h:
> > 
> >  * The compiler memory barrier in CMM_LOAD_SHARED() ensures that value-speculative
> >  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
> >  * data read before the pointer read by speculating the value of the pointer.
> >  * Correct ordering is ensured because the pointer is read as a volatile access.
> >  * This acts as a global side-effect operation, which forbids reordering of
> >  * dependent memory operations. Note that such concern about dependency-breaking
> >  * optimizations will eventually be taken care of by the "memory_order_consume"
> >  * addition to forthcoming C++ standard.
> > 
> > (note: CMM_LOAD_SHARED() is the equivalent of READ_ONCE(), but was introduced in
> > liburcu as a public API before READ_ONCE() existed in the Linux kernel)
> > 
> > Peter tells me the "memory_order_consume" is not something which can be used today.
> > Any information on its status at C/C++ standard levels and implementation-wise ?

Actually, you really can use memory_order_consume.  All current
implementations will compile it as if it was memory_order_acquire.
This will work correctly, but may be slower than you would like on ARM,
PowerPC, and so on.

On things like x86, the penalty is forgone optimizations, so less
of a problem there.

> > Pragmatically speaking, what should we change in liburcu to ensure we don't generate
> > broken code when LTO is enabled ? I suspect there are a few options here:
> > 
> > 1) Fail to build if LTO is enabled,
> > 2) Generate slower code for rcu_dereference, either on all architectures or only
> >    on weakly-ordered architectures,
> > 3) Generate different code depending on whether LTO is enabled or not. AFAIU this would only
> >    work if every compile unit is aware that it will end up being optimized with LTO. Not sure
> >    how this could be done in the context of user-space.
> > 4) [ Insert better idea here. ]

Use memory_order_consume if LTO is enabled.  That will work now, and
might generate good code in some hoped-for future.

> > Thoughts ?
> 
> Using memory_order_acquire is safe; and is basically what Will did for
> ARM64.
> 
> The problematic tranformations are possible even without LTO, although
> less likely due to less visibility, but everybody agrees they're
> possible and allowed.
> 
> OTOH we do not have a positive sighting of it actually happening (I
> think), we're all just being cautious and not willing to debug the
> resulting wreckage if it does indeed happen.

And yes, you can also use memory_order_acquire.

							Thanx, Paul

Re: liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Mathieu Desnoyers]

----- On Apr 16, 2021, at 12:01 PM, paulmck paulmck@kernel.org wrote:

> On Fri, Apr 16, 2021 at 05:17:11PM +0200, Peter Zijlstra wrote:
>> On Fri, Apr 16, 2021 at 10:52:16AM -0400, Mathieu Desnoyers wrote:
>> > Hi Paul, Will, Peter,
>> > 
>> > I noticed in this discussion https://lkml.org/lkml/2021/4/16/118 that LTO
>> > is able to break rcu_dereference. This seems to be taken care of by
>> > arch/arm64/include/asm/rwonce.h on arm64 in the Linux kernel tree.
>> > 
>> > In the liburcu user-space library, we have this comment near rcu_dereference()
>> > in
>> > include/urcu/static/pointer.h:
>> > 
>> >  * The compiler memory barrier in CMM_LOAD_SHARED() ensures that
>> >  value-speculative
>> >  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
>> >  * data read before the pointer read by speculating the value of the pointer.
>> >  * Correct ordering is ensured because the pointer is read as a volatile access.
>> >  * This acts as a global side-effect operation, which forbids reordering of
>> >  * dependent memory operations. Note that such concern about dependency-breaking
>> >  * optimizations will eventually be taken care of by the "memory_order_consume"
>> >  * addition to forthcoming C++ standard.
>> > 
>> > (note: CMM_LOAD_SHARED() is the equivalent of READ_ONCE(), but was introduced in
>> > liburcu as a public API before READ_ONCE() existed in the Linux kernel)
>> > 
>> > Peter tells me the "memory_order_consume" is not something which can be used
>> > today.
>> > Any information on its status at C/C++ standard levels and implementation-wise ?
> 
> Actually, you really can use memory_order_consume.  All current
> implementations will compile it as if it was memory_order_acquire.
> This will work correctly, but may be slower than you would like on ARM,
> PowerPC, and so on.
> 
> On things like x86, the penalty is forgone optimizations, so less
> of a problem there.

OK

> 
>> > Pragmatically speaking, what should we change in liburcu to ensure we don't
>> > generate
>> > broken code when LTO is enabled ? I suspect there are a few options here:
>> > 
>> > 1) Fail to build if LTO is enabled,
>> > 2) Generate slower code for rcu_dereference, either on all architectures or only
>> >    on weakly-ordered architectures,
>> > 3) Generate different code depending on whether LTO is enabled or not. AFAIU
>> > this would only
>> >    work if every compile unit is aware that it will end up being optimized with
>> >    LTO. Not sure
>> >    how this could be done in the context of user-space.
>> > 4) [ Insert better idea here. ]
> 
> Use memory_order_consume if LTO is enabled.  That will work now, and
> might generate good code in some hoped-for future.

In the context of a user-space library, how does one check whether LTO is enabled with
preprocessor directives ? A quick test with gcc seems to show that both with and without
-flto cannot be distinguished from a preprocessor POV, e.g. the output of both

gcc --std=c11 -O2 -dM -E - < /dev/null
and
gcc --std=c11 -O2 -flto -dM -E - < /dev/null

is exactly the same. Am I missing something here ?

If we accept to use memory_order_consume all the time in both C and C++ code starting from
C11 and C++11, the following code snippet could do the trick:

#define CMM_ACCESS_ONCE(x) (*(__volatile__  __typeof__(x) *)&(x))
#define CMM_LOAD_SHARED(p) CMM_ACCESS_ONCE(p)

#if defined (__cplusplus)
# if __cplusplus >= 201103L
#  include <atomic>
#  define rcu_dereference(x)    ((std::atomic<__typeof__(x)>)(x)).load(std::memory_order_consume)
# else
#  define rcu_dereference(x)    CMM_LOAD_SHARED(x)
# endif
#else
# if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L)
#  include <stdatomic.h>
#  define rcu_dereference(x)    atomic_load_explicit(&(x), memory_order_consume)
# else
#  define rcu_dereference(x)    CMM_LOAD_SHARED(x)
# endif
#endif

This uses the volatile approach prior to C11/C++11, and moves to memory_order_consume
afterwards. This will bring a performance penalty on weakly-ordered architectures even
when -flto is not specified though.

Then the burden is pushed on the compiler people to eventually implement an efficient
memory_order_consume.

Is that acceptable ?

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

Re: liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Paul E. McKenney]

On Fri, Apr 16, 2021 at 02:40:08PM -0400, Mathieu Desnoyers wrote:
> ----- On Apr 16, 2021, at 12:01 PM, paulmck paulmck@kernel.org wrote:
> 
> > On Fri, Apr 16, 2021 at 05:17:11PM +0200, Peter Zijlstra wrote:
> >> On Fri, Apr 16, 2021 at 10:52:16AM -0400, Mathieu Desnoyers wrote:
> >> > Hi Paul, Will, Peter,
> >> > 
> >> > I noticed in this discussion https://lkml.org/lkml/2021/4/16/118 that LTO
> >> > is able to break rcu_dereference. This seems to be taken care of by
> >> > arch/arm64/include/asm/rwonce.h on arm64 in the Linux kernel tree.
> >> > 
> >> > In the liburcu user-space library, we have this comment near rcu_dereference()
> >> > in
> >> > include/urcu/static/pointer.h:
> >> > 
> >> >  * The compiler memory barrier in CMM_LOAD_SHARED() ensures that
> >> >  value-speculative
> >> >  * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
> >> >  * data read before the pointer read by speculating the value of the pointer.
> >> >  * Correct ordering is ensured because the pointer is read as a volatile access.
> >> >  * This acts as a global side-effect operation, which forbids reordering of
> >> >  * dependent memory operations. Note that such concern about dependency-breaking
> >> >  * optimizations will eventually be taken care of by the "memory_order_consume"
> >> >  * addition to forthcoming C++ standard.
> >> > 
> >> > (note: CMM_LOAD_SHARED() is the equivalent of READ_ONCE(), but was introduced in
> >> > liburcu as a public API before READ_ONCE() existed in the Linux kernel)
> >> > 
> >> > Peter tells me the "memory_order_consume" is not something which can be used
> >> > today.
> >> > Any information on its status at C/C++ standard levels and implementation-wise ?
> > 
> > Actually, you really can use memory_order_consume.  All current
> > implementations will compile it as if it was memory_order_acquire.
> > This will work correctly, but may be slower than you would like on ARM,
> > PowerPC, and so on.
> > 
> > On things like x86, the penalty is forgone optimizations, so less
> > of a problem there.
> 
> OK
> 
> > 
> >> > Pragmatically speaking, what should we change in liburcu to ensure we don't
> >> > generate
> >> > broken code when LTO is enabled ? I suspect there are a few options here:
> >> > 
> >> > 1) Fail to build if LTO is enabled,
> >> > 2) Generate slower code for rcu_dereference, either on all architectures or only
> >> >    on weakly-ordered architectures,
> >> > 3) Generate different code depending on whether LTO is enabled or not. AFAIU
> >> > this would only
> >> >    work if every compile unit is aware that it will end up being optimized with
> >> >    LTO. Not sure
> >> >    how this could be done in the context of user-space.
> >> > 4) [ Insert better idea here. ]
> > 
> > Use memory_order_consume if LTO is enabled.  That will work now, and
> > might generate good code in some hoped-for future.
> 
> In the context of a user-space library, how does one check whether LTO is enabled with
> preprocessor directives ? A quick test with gcc seems to show that both with and without
> -flto cannot be distinguished from a preprocessor POV, e.g. the output of both
> 
> gcc --std=c11 -O2 -dM -E - < /dev/null
> and
> gcc --std=c11 -O2 -flto -dM -E - < /dev/null
> 
> is exactly the same. Am I missing something here ?

No idea.  ;-)

> If we accept to use memory_order_consume all the time in both C and C++ code starting from
> C11 and C++11, the following code snippet could do the trick:
> 
> #define CMM_ACCESS_ONCE(x) (*(__volatile__  __typeof__(x) *)&(x))
> #define CMM_LOAD_SHARED(p) CMM_ACCESS_ONCE(p)
> 
> #if defined (__cplusplus)
> # if __cplusplus >= 201103L
> #  include <atomic>
> #  define rcu_dereference(x)    ((std::atomic<__typeof__(x)>)(x)).load(std::memory_order_consume)
> # else
> #  define rcu_dereference(x)    CMM_LOAD_SHARED(x)
> # endif
> #else
> # if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L)
> #  include <stdatomic.h>
> #  define rcu_dereference(x)    atomic_load_explicit(&(x), memory_order_consume)
> # else
> #  define rcu_dereference(x)    CMM_LOAD_SHARED(x)
> # endif
> #endif
> 
> This uses the volatile approach prior to C11/C++11, and moves to memory_order_consume
> afterwards. This will bring a performance penalty on weakly-ordered architectures even
> when -flto is not specified though.
> 
> Then the burden is pushed on the compiler people to eventually implement an efficient
> memory_order_consume.
> 
> Is that acceptable ?

That makes sense to me!

If it can be done reasonably, I suggest also having some way for the
person building userspace RCU to say "I know what I am doing, so do
it with volatile rather than memory_order_consume."

							Thanx, Paul

Re: liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Mathieu Desnoyers]

----- On Apr 16, 2021, at 3:02 PM, paulmck paulmck@kernel.org wrote:
[...]
> 
> If it can be done reasonably, I suggest also having some way for the
> person building userspace RCU to say "I know what I am doing, so do
> it with volatile rather than memory_order_consume."

Like so ?

#define CMM_ACCESS_ONCE(x) (*(__volatile__  __typeof__(x) *)&(x))
#define CMM_LOAD_SHARED(p) CMM_ACCESS_ONCE(p)

/*
 * By defining URCU_DEREFERENCE_USE_VOLATILE, the user requires use of
 * volatile access to implement rcu_dereference rather than
 * memory_order_consume load from the C11/C++11 standards.
 *
 * This may improve performance on weakly-ordered architectures where
 * the compiler implements memory_order_consume as a
 * memory_order_acquire, which is stricter than required by the
 * standard.
 *
 * Note that using volatile accesses for rcu_dereference may cause
 * LTO to generate incorrectly ordered code starting from C11/C++11.
 */

#ifdef URCU_DEREFERENCE_USE_VOLATILE
# define rcu_dereference(x)     CMM_LOAD_SHARED(x)
#else
# if defined (__cplusplus)
#  if __cplusplus >= 201103L
#   include <atomic>
#   define rcu_dereference(x)   ((std::atomic<__typeof__(x)>)(x)).load(std::memory_order_consume)
#  else
#   define rcu_dereference(x)   CMM_LOAD_SHARED(x)
#  endif
# else
#  if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L)
#   include <stdatomic.h>
#   define rcu_dereference(x)   atomic_load_explicit(&(x), memory_order_consume)
#  else
#   define rcu_dereference(x)   CMM_LOAD_SHARED(x)
#  endif
# endif
#endif

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

Re: liburcu: LTO breaking rcu_dereference on arm64 and possibly other architectures ?

[Paul E. McKenney]

On Fri, Apr 16, 2021 at 03:30:53PM -0400, Mathieu Desnoyers wrote:
> ----- On Apr 16, 2021, at 3:02 PM, paulmck paulmck@kernel.org wrote:
> [...]
> > 
> > If it can be done reasonably, I suggest also having some way for the
> > person building userspace RCU to say "I know what I am doing, so do
> > it with volatile rather than memory_order_consume."
> 
> Like so ?
> 
> #define CMM_ACCESS_ONCE(x) (*(__volatile__  __typeof__(x) *)&(x))
> #define CMM_LOAD_SHARED(p) CMM_ACCESS_ONCE(p)
> 
> /*
>  * By defining URCU_DEREFERENCE_USE_VOLATILE, the user requires use of
>  * volatile access to implement rcu_dereference rather than
>  * memory_order_consume load from the C11/C++11 standards.
>  *
>  * This may improve performance on weakly-ordered architectures where
>  * the compiler implements memory_order_consume as a
>  * memory_order_acquire, which is stricter than required by the
>  * standard.
>  *
>  * Note that using volatile accesses for rcu_dereference may cause
>  * LTO to generate incorrectly ordered code starting from C11/C++11.
>  */
> 
> #ifdef URCU_DEREFERENCE_USE_VOLATILE
> # define rcu_dereference(x)     CMM_LOAD_SHARED(x)
> #else
> # if defined (__cplusplus)
> #  if __cplusplus >= 201103L
> #   include <atomic>
> #   define rcu_dereference(x)   ((std::atomic<__typeof__(x)>)(x)).load(std::memory_order_consume)
> #  else
> #   define rcu_dereference(x)   CMM_LOAD_SHARED(x)
> #  endif
> # else
> #  if (defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L)
> #   include <stdatomic.h>
> #   define rcu_dereference(x)   atomic_load_explicit(&(x), memory_order_consume)
> #  else
> #   define rcu_dereference(x)   CMM_LOAD_SHARED(x)
> #  endif
> # endif
> #endif

Looks good to me!

							Thanx, Paul