Re: different kind of memory reordering clarification

Re: different kind of memory reordering clarification

[Yubin Ruan]

On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
[...]
> > > 
> > > Can you please provide me with some examples or references for different kinds
> > > of memory reordering in a SMP system? You know, there are different kinds of
> > > reordering:
> > > 
> > >    - Loads reordered after loads
> > >    - Loads reordered after stores
> > >    - Stores reordered after stores
> > >    - Stores reordered after loads
> > >    - Atomic reordered with loads
> > >    - Atomic reordered with stores
> > >    - Dependent loads reordered (DEC alpha)
> > 
> > I remember there is open-std.org webpage containing comparision of C++'s
> > memory model to those primitives used in the Linux kernel. But I just can't
> > find that page.
> 
> Here you go!
> 
> http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> 
> There will be an update in a month or so, but the above is pretty
> close.  Also, the Linux-kernel memory model was presented at
> ASPLOS and accepted into the Linux kernel itself:
> 
> https://paulmck.livejournal.com/49667.html

Many thanks. But I am currently confused about the relationship between
terminologies used in the Linux kernel and those used in some programming
languages (e.g., C++), i.e., the relationships between

    memory_order_release
    memory_order_relaxed
    memory_order_acquire
    memory_order_seq_cst
    ...

and those used in the kernel:

    READ_ONCE() / WRITE_ONCE()
    rmb() / wmb() / mb() / smp_mb()
    ...

Any materials for that?

Yubin

Re: different kind of memory reordering clarification

[Yubin Ruan]

On Wed, Apr 11, 2018 at 10:46:28AM +0800, Yubin Ruan wrote:
> On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> > On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
> [...]
> > > > 
> > > > Can you please provide me with some examples or references for different kinds
> > > > of memory reordering in a SMP system? You know, there are different kinds of
> > > > reordering:
> > > > 
> > > >    - Loads reordered after loads
> > > >    - Loads reordered after stores
> > > >    - Stores reordered after stores
> > > >    - Stores reordered after loads
> > > >    - Atomic reordered with loads
> > > >    - Atomic reordered with stores
> > > >    - Dependent loads reordered (DEC alpha)
> > > 
> > > I remember there is open-std.org webpage containing comparision of C++'s
> > > memory model to those primitives used in the Linux kernel. But I just can't
> > > find that page.
> > 
> > Here you go!
> > 
> > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> > 
> > There will be an update in a month or so, but the above is pretty
> > close.  Also, the Linux-kernel memory model was presented at
> > ASPLOS and accepted into the Linux kernel itself:
> > 
> > https://paulmck.livejournal.com/49667.html
> 
> Many thanks. But I am currently confused about the relationship between
> terminologies used in the Linux kernel and those used in some programming
> languages (e.g., C++), i.e., the relationships between
> 
>     memory_order_release
>     memory_order_relaxed
>     memory_order_acquire
>     memory_order_seq_cst
>     ...
> 
> and those used in the kernel:
> 
>     READ_ONCE() / WRITE_ONCE()
>     rmb() / wmb() / mb() / smp_mb()
>     ...
> 
> Any materials for that?

Hmm, to be more exact, what I want is something like this:

    “These primitives can be expressed directly in terms of the upcoming
    C++0x standard. For the smp_mb() primitive this correspondence is not
    exact; our memory barriers are somewhat stronger than the standard’s
    atomic_thread_fence(memory_order_seq_cst). The LOAD_SHARED() primitive
    maps to x.load(memory_order_relaxed) and STORE_SHARED() to
    x.store(memory_order_relaxed). The barrier() primitive maps to
    atomic_signal_fence(memory_order_seq_cst). In addition, rcu_dereference()
    maps to x.load(memory_order_consume) and rcu_assign_pointer() maps to
    x.store(v, memory_order_release).”

This is extracted from the paper "User-Level Implementations of Read-Copy
Update" by M. Desnoyers and P. McKenney,  A. S. Stern, M. R. Dagenais and J.
Walpole[1]. And the LOAD_SHARED() and STORE_SHARED() above are READ_ONCE() and
WRITE_ONCE(), respectively. (BTW, LOAD_SHARED/STORE_SHARED seem to be better
names than READ_ONCE/WRITE_ONCE, which are a bit confusing. How come people
adopted those name?)

(I find this after digging into a whole bunch of emails...hmm..email is a good
thing)

[1]: https://www.efficios.com/pub/rcu/urcu-main-accepted.pdf

Yubin

Re: different kind of memory reordering clarification

[Paul E. McKenney]

On Wed, Apr 11, 2018 at 10:46:28AM +0800, Yubin Ruan wrote:
> On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> > On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
> [...]
> > > > 
> > > > Can you please provide me with some examples or references for different kinds
> > > > of memory reordering in a SMP system? You know, there are different kinds of
> > > > reordering:
> > > > 
> > > >    - Loads reordered after loads
> > > >    - Loads reordered after stores
> > > >    - Stores reordered after stores
> > > >    - Stores reordered after loads
> > > >    - Atomic reordered with loads
> > > >    - Atomic reordered with stores
> > > >    - Dependent loads reordered (DEC alpha)
> > > 
> > > I remember there is open-std.org webpage containing comparision of C++'s
> > > memory model to those primitives used in the Linux kernel. But I just can't
> > > find that page.
> > 
> > Here you go!
> > 
> > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> > 
> > There will be an update in a month or so, but the above is pretty
> > close.  Also, the Linux-kernel memory model was presented at
> > ASPLOS and accepted into the Linux kernel itself:
> > 
> > https://paulmck.livejournal.com/49667.html
> 
> Many thanks. But I am currently confused about the relationship between
> terminologies used in the Linux kernel and those used in some programming
> languages (e.g., C++), i.e., the relationships between
> 
>     memory_order_release
>     memory_order_relaxed
>     memory_order_acquire
>     memory_order_seq_cst
>     ...
> 
> and those used in the kernel:
> 
>     READ_ONCE() / WRITE_ONCE()
>     rmb() / wmb() / mb() / smp_mb()
>     ...
> 
> Any materials for that?

The URL I posted above:

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html

							Thanx, Paul

Re: different kind of memory reordering clarification

[Paul E. McKenney]

On Wed, Apr 11, 2018 at 11:00:58AM +0800, Yubin Ruan wrote:
> On Wed, Apr 11, 2018 at 10:46:28AM +0800, Yubin Ruan wrote:
> > On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> > > On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > > > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
> > [...]
> > > > > 
> > > > > Can you please provide me with some examples or references for different kinds
> > > > > of memory reordering in a SMP system? You know, there are different kinds of
> > > > > reordering:
> > > > > 
> > > > >    - Loads reordered after loads
> > > > >    - Loads reordered after stores
> > > > >    - Stores reordered after stores
> > > > >    - Stores reordered after loads
> > > > >    - Atomic reordered with loads
> > > > >    - Atomic reordered with stores
> > > > >    - Dependent loads reordered (DEC alpha)
> > > > 
> > > > I remember there is open-std.org webpage containing comparision of C++'s
> > > > memory model to those primitives used in the Linux kernel. But I just can't
> > > > find that page.
> > > 
> > > Here you go!
> > > 
> > > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> > > 
> > > There will be an update in a month or so, but the above is pretty
> > > close.  Also, the Linux-kernel memory model was presented at
> > > ASPLOS and accepted into the Linux kernel itself:
> > > 
> > > https://paulmck.livejournal.com/49667.html
> > 
> > Many thanks. But I am currently confused about the relationship between
> > terminologies used in the Linux kernel and those used in some programming
> > languages (e.g., C++), i.e., the relationships between
> > 
> >     memory_order_release
> >     memory_order_relaxed
> >     memory_order_acquire
> >     memory_order_seq_cst
> >     ...
> > 
> > and those used in the kernel:
> > 
> >     READ_ONCE() / WRITE_ONCE()
> >     rmb() / wmb() / mb() / smp_mb()
> >     ...
> > 
> > Any materials for that?
> 
> Hmm, to be more exact, what I want is something like this:
> 
>     “These primitives can be expressed directly in terms of the upcoming
>     C++0x standard. For the smp_mb() primitive this correspondence is not
>     exact; our memory barriers are somewhat stronger than the standard’s
>     atomic_thread_fence(memory_order_seq_cst). The LOAD_SHARED() primitive
>     maps to x.load(memory_order_relaxed) and STORE_SHARED() to
>     x.store(memory_order_relaxed). The barrier() primitive maps to
>     atomic_signal_fence(memory_order_seq_cst). In addition, rcu_dereference()
>     maps to x.load(memory_order_consume) and rcu_assign_pointer() maps to
>     x.store(v, memory_order_release).”

Those are still valid.  Again, the other paper from my earlier email
has more mappings.

> This is extracted from the paper "User-Level Implementations of Read-Copy
> Update" by M. Desnoyers and P. McKenney,  A. S. Stern, M. R. Dagenais and J.
> Walpole[1]. And the LOAD_SHARED() and STORE_SHARED() above are READ_ONCE() and
> WRITE_ONCE(), respectively. (BTW, LOAD_SHARED/STORE_SHARED seem to be better
> names than READ_ONCE/WRITE_ONCE, which are a bit confusing. How come people
> adopted those name?)

They came from ACCESS_ONCE().  They are _ONCE() because they prevent the
compiler from fusing and splitting accesses, which it can do with normal
loads and stores.  This Linux Weekly News article has some of this history:

	http://lwn.net/Articles/508991/

> (I find this after digging into a whole bunch of emails...hmm..email is a good
> thing)

;-) ;-) ;-)

							Thanx, Paul

> [1]: https://www.efficios.com/pub/rcu/urcu-main-accepted.pdf
> 
> Yubin
> 

Re: different kind of memory reordering clarification

[Yubin Ruan]

On Tue, Apr 10, 2018 at 08:09:04PM -0700, Paul E. McKenney wrote:
> On Wed, Apr 11, 2018 at 11:00:58AM +0800, Yubin Ruan wrote:
> > On Wed, Apr 11, 2018 at 10:46:28AM +0800, Yubin Ruan wrote:
> > > On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> > > > On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > > > > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
> > > [...]
> > > > > > 
> > > > > > Can you please provide me with some examples or references for different kinds
> > > > > > of memory reordering in a SMP system? You know, there are different kinds of
> > > > > > reordering:
> > > > > > 
> > > > > >    - Loads reordered after loads
> > > > > >    - Loads reordered after stores
> > > > > >    - Stores reordered after stores
> > > > > >    - Stores reordered after loads
> > > > > >    - Atomic reordered with loads
> > > > > >    - Atomic reordered with stores
> > > > > >    - Dependent loads reordered (DEC alpha)
> > > > > 
> > > > > I remember there is open-std.org webpage containing comparision of C++'s
> > > > > memory model to those primitives used in the Linux kernel. But I just can't
> > > > > find that page.
> > > > 
> > > > Here you go!
> > > > 
> > > > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> > > > 
> > > > There will be an update in a month or so, but the above is pretty
> > > > close.  Also, the Linux-kernel memory model was presented at
> > > > ASPLOS and accepted into the Linux kernel itself:
> > > > 
> > > > https://paulmck.livejournal.com/49667.html
> > > 
> > > Many thanks. But I am currently confused about the relationship between
> > > terminologies used in the Linux kernel and those used in some programming
> > > languages (e.g., C++), i.e., the relationships between
> > > 
> > >     memory_order_release
> > >     memory_order_relaxed
> > >     memory_order_acquire
> > >     memory_order_seq_cst
> > >     ...
> > > 
> > > and those used in the kernel:
> > > 
> > >     READ_ONCE() / WRITE_ONCE()
> > >     rmb() / wmb() / mb() / smp_mb()
> > >     ...
> > > 
> > > Any materials for that?
> > 
> > Hmm, to be more exact, what I want is something like this:
> > 
> >     “These primitives can be expressed directly in terms of the upcoming
> >     C++0x standard. For the smp_mb() primitive this correspondence is not
> >     exact; our memory barriers are somewhat stronger than the standard’s
> >     atomic_thread_fence(memory_order_seq_cst). The LOAD_SHARED() primitive
> >     maps to x.load(memory_order_relaxed) and STORE_SHARED() to
> >     x.store(memory_order_relaxed). The barrier() primitive maps to
> >     atomic_signal_fence(memory_order_seq_cst). In addition, rcu_dereference()
> >     maps to x.load(memory_order_consume) and rcu_assign_pointer() maps to
> >     x.store(v, memory_order_release).”
> 
> Those are still valid.  Again, the other paper from my earlier email
> has more mappings.

Sorry I missed that! (trying to read too much all at once)

I read about atomic_ops here

    https://www.kernel.org/doc/html/v4.14/core-api/atomic_ops.html

and find that many "atomic" operations such as
atomic_set/atomic_read/atomic_write does not require volatile semantic, nor
does it require alignment constraints that force the CPU to do load/store "at
once". In this situation, both the compiler and the processor are all allowed
to tear apart a read/write atomic operation. How can it be "atomic" in this
case?

Yubin

Re: different kind of memory reordering clarification

[Yubin Ruan]

On Wed, Apr 11, 2018 at 11:43:24AM +0800, Yubin Ruan wrote:
> On Tue, Apr 10, 2018 at 08:09:04PM -0700, Paul E. McKenney wrote:
> > On Wed, Apr 11, 2018 at 11:00:58AM +0800, Yubin Ruan wrote:
> > > On Wed, Apr 11, 2018 at 10:46:28AM +0800, Yubin Ruan wrote:
> > > > On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> > > > > On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > > > > > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
> > > > [...]
> > > > > > > 
> > > > > > > Can you please provide me with some examples or references for different kinds
> > > > > > > of memory reordering in a SMP system? You know, there are different kinds of
> > > > > > > reordering:
> > > > > > > 
> > > > > > >    - Loads reordered after loads
> > > > > > >    - Loads reordered after stores
> > > > > > >    - Stores reordered after stores
> > > > > > >    - Stores reordered after loads
> > > > > > >    - Atomic reordered with loads
> > > > > > >    - Atomic reordered with stores
> > > > > > >    - Dependent loads reordered (DEC alpha)
> > > > > > 
> > > > > > I remember there is open-std.org webpage containing comparision of C++'s
> > > > > > memory model to those primitives used in the Linux kernel. But I just can't
> > > > > > find that page.
> > > > > 
> > > > > Here you go!
> > > > > 
> > > > > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> > > > > 
> > > > > There will be an update in a month or so, but the above is pretty
> > > > > close.  Also, the Linux-kernel memory model was presented at
> > > > > ASPLOS and accepted into the Linux kernel itself:
> > > > > 
> > > > > https://paulmck.livejournal.com/49667.html
> > > > 
> > > > Many thanks. But I am currently confused about the relationship between
> > > > terminologies used in the Linux kernel and those used in some programming
> > > > languages (e.g., C++), i.e., the relationships between
> > > > 
> > > >     memory_order_release
> > > >     memory_order_relaxed
> > > >     memory_order_acquire
> > > >     memory_order_seq_cst
> > > >     ...
> > > > 
> > > > and those used in the kernel:
> > > > 
> > > >     READ_ONCE() / WRITE_ONCE()
> > > >     rmb() / wmb() / mb() / smp_mb()
> > > >     ...
> > > > 
> > > > Any materials for that?
> > > 
> > > Hmm, to be more exact, what I want is something like this:
> > > 
> > >     “These primitives can be expressed directly in terms of the upcoming
> > >     C++0x standard. For the smp_mb() primitive this correspondence is not
> > >     exact; our memory barriers are somewhat stronger than the standard’s
> > >     atomic_thread_fence(memory_order_seq_cst). The LOAD_SHARED() primitive
> > >     maps to x.load(memory_order_relaxed) and STORE_SHARED() to
> > >     x.store(memory_order_relaxed). The barrier() primitive maps to
> > >     atomic_signal_fence(memory_order_seq_cst). In addition, rcu_dereference()
> > >     maps to x.load(memory_order_consume) and rcu_assign_pointer() maps to
> > >     x.store(v, memory_order_release).”
> > 
> > Those are still valid.  Again, the other paper from my earlier email
> > has more mappings.
> 
> Sorry I missed that! (trying to read too much all at once)
> 
> I read about atomic_ops here
> 
>     https://www.kernel.org/doc/html/v4.14/core-api/atomic_ops.html
> 
> and find that many "atomic" operations such as
> atomic_set/atomic_read/atomic_write does not require volatile semantic, nor
> does it require alignment constraints that force the CPU to do load/store "at
> once". In this situation, both the compiler and the processor are all allowed
> to tear apart a read/write atomic operation. How can it be "atomic" in this
> case?

Hmm... Reading more source code confirms that there are READ_ONCE/WRITE_ONCE
used in atomic_set/atomic_read/atomic_write. But these does not prevent the
processor from tearing apart the reads/write. Can the definition

    typedef struct { int counter; } atomic_t;

guarantee necessary alignment constraints required by the processor to perform
atomic operations?

Yubin

Re: different kind of memory reordering clarification

[Paul E. McKenney]

On Wed, Apr 11, 2018 at 11:59:08AM +0800, Yubin Ruan wrote:
> On Wed, Apr 11, 2018 at 11:43:24AM +0800, Yubin Ruan wrote:
> > On Tue, Apr 10, 2018 at 08:09:04PM -0700, Paul E. McKenney wrote:
> > > On Wed, Apr 11, 2018 at 11:00:58AM +0800, Yubin Ruan wrote:
> > > > On Wed, Apr 11, 2018 at 10:46:28AM +0800, Yubin Ruan wrote:
> > > > > On Tue, Apr 10, 2018 at 10:04:09AM -0700, Paul E. McKenney wrote:
> > > > > > On Tue, Apr 10, 2018 at 11:20:24PM +0800, Yubin Ruan wrote:
> > > > > > > On Tue, Apr 10, 2018 at 08:14:08PM +0800, Yubin Ruan wrote:
> > > > > [...]
> > > > > > > > 
> > > > > > > > Can you please provide me with some examples or references for different kinds
> > > > > > > > of memory reordering in a SMP system? You know, there are different kinds of
> > > > > > > > reordering:
> > > > > > > > 
> > > > > > > >    - Loads reordered after loads
> > > > > > > >    - Loads reordered after stores
> > > > > > > >    - Stores reordered after stores
> > > > > > > >    - Stores reordered after loads
> > > > > > > >    - Atomic reordered with loads
> > > > > > > >    - Atomic reordered with stores
> > > > > > > >    - Dependent loads reordered (DEC alpha)
> > > > > > > 
> > > > > > > I remember there is open-std.org webpage containing comparision of C++'s
> > > > > > > memory model to those primitives used in the Linux kernel. But I just can't
> > > > > > > find that page.
> > > > > > 
> > > > > > Here you go!
> > > > > > 
> > > > > > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0124r4.html
> > > > > > 
> > > > > > There will be an update in a month or so, but the above is pretty
> > > > > > close.  Also, the Linux-kernel memory model was presented at
> > > > > > ASPLOS and accepted into the Linux kernel itself:
> > > > > > 
> > > > > > https://paulmck.livejournal.com/49667.html
> > > > > 
> > > > > Many thanks. But I am currently confused about the relationship between
> > > > > terminologies used in the Linux kernel and those used in some programming
> > > > > languages (e.g., C++), i.e., the relationships between
> > > > > 
> > > > >     memory_order_release
> > > > >     memory_order_relaxed
> > > > >     memory_order_acquire
> > > > >     memory_order_seq_cst
> > > > >     ...
> > > > > 
> > > > > and those used in the kernel:
> > > > > 
> > > > >     READ_ONCE() / WRITE_ONCE()
> > > > >     rmb() / wmb() / mb() / smp_mb()
> > > > >     ...
> > > > > 
> > > > > Any materials for that?
> > > > 
> > > > Hmm, to be more exact, what I want is something like this:
> > > > 
> > > >     “These primitives can be expressed directly in terms of the upcoming
> > > >     C++0x standard. For the smp_mb() primitive this correspondence is not
> > > >     exact; our memory barriers are somewhat stronger than the standard’s
> > > >     atomic_thread_fence(memory_order_seq_cst). The LOAD_SHARED() primitive
> > > >     maps to x.load(memory_order_relaxed) and STORE_SHARED() to
> > > >     x.store(memory_order_relaxed). The barrier() primitive maps to
> > > >     atomic_signal_fence(memory_order_seq_cst). In addition, rcu_dereference()
> > > >     maps to x.load(memory_order_consume) and rcu_assign_pointer() maps to
> > > >     x.store(v, memory_order_release).”
> > > 
> > > Those are still valid.  Again, the other paper from my earlier email
> > > has more mappings.
> > 
> > Sorry I missed that! (trying to read too much all at once)
> > 
> > I read about atomic_ops here
> > 
> >     https://www.kernel.org/doc/html/v4.14/core-api/atomic_ops.html
> > 
> > and find that many "atomic" operations such as
> > atomic_set/atomic_read/atomic_write does not require volatile semantic, nor
> > does it require alignment constraints that force the CPU to do load/store "at
> > once". In this situation, both the compiler and the processor are all allowed
> > to tear apart a read/write atomic operation. How can it be "atomic" in this
> > case?
> 
> Hmm... Reading more source code confirms that there are READ_ONCE/WRITE_ONCE
> used in atomic_set/atomic_read/atomic_write. But these does not prevent the
> processor from tearing apart the reads/write.

But it does prevent the compiler from doing so if the variable is
small enough to be accessed with a single load/store instruction.
And if the variable is aligned, the processor won't split a single
load/store instruction.  And if the processor doesn't have (say) a
load-byte instruction, the current C-language standard requires it to use
an atomic read-modify-write sequence, which also avoids the tearing apart.

>                                               Can the definition
> 
>     typedef struct { int counter; } atomic_t;
> 
> guarantee necessary alignment constraints required by the processor to perform
> atomic operations?

Yes, the compiler aligns machine types (including int).  Unless you use
"packed" or some such, but if you do that, you get what you deserve.  ;-)

							Thanx, Paul