|Authors:||Hans-J. Boehm, Olivier Giroux, Viktor Vafeiades, with input from from Will Deacon, Doug Lea, Daniel Lustig, Paul McKenney and others|
Although the current C++ memory model, adopted essentially in C++11, has served our user community reasonably well in practice, a number of problems have come to light. The first one of these is particularly new and troubling:
memory_order_seq_cst, especially for fences, is too weak. This was caused by historical assumptions that have since been disproven.
memory_order_relaxedoperations to interfere with synchronizes-with relationships.
memory_order_consumeis widely acknowledged to be unusable, and implementations generally treat it as
memory_order_acquire. The current draft "temporarily discourages" it. See http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0371r1.html . There are proposals to repair it (cf. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0190r3.pdf and http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0462r1.pdf ), but nothing that is ready to go.
Here we concentrate on, and outline proposals for, the first three, and merely keep the last two in mind. These three are listed in decreasing order of severity, or at least our perception of severity. It is critical to address the first, where a fix for the third one would be "nice to have".
Although we previously believed otherwise, it has recently been shown that the standard implementations of memory_order_acquire and memory_order_release on Power are insufficient. Very briefly, these are compiled using "lightweight" fences, which are insufficient to enforce required properties for memory_order_sq_cst accesses to the same location.
To illustrate, we borrow the Z6.U example, from S 2.1 of http://plv.mpi-sws.org/scfix/paper.pdf
(pseudo-C++ syntax, memory orders indicated by subscript, e.g. y =rel
Thread 1 Thread 2 Thread 3
x =sc 1
y =rel 1
b = fetch_add(y)sc//1
c = yrlx//3
y =sc 3
a = xsc//0
The comments here indicate the observed assigned value.
The indicated outcome here is disallowed by the current standard: All
memory_order_seq_cst (sc) accesses must occur in a single total
order, which is constrained to have
a = xsc //0
x =sc 1 (since it doesn't observe the store),
which must be before
b = fetch_add(y)sc //1
(since it happens before it), which must be before
y =sc 3
(since the fetch_add does not observe the store, which is
forced to be last in modification order by the load in Thread 2). But this is
disallowed since the standard requires the happens before ordering to be
consistent with the sequential consistency ordering,
y =sc 3, the
last element of the sc order, happens before
a = xsc //0, the first one.
On the other hand, this outcome is allowed by the Power implementation. Power
normally uses the "leading fence" convention for sequentially consistent
atomics. ( See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
) This means that there is only an lwsync fence between Thread 1's instructions.
This does not have the "cumulativity"/transitivity properties that would be
required to make the store to
x visible to Thread 3 under
This issue was missed in earlier analyses. This example is not a problem for the "trailing fence" mapping that could also have been used on Power. But Lahav et al contains examples that fail for that mapping as well, for similar reasons.
This example relies crucially on the fact that a
memory_order_release operation synchronizes with a
memory_order_seq_cst operation on the same location. Code that
consistently accesses each location only with
or only with weaker ordering, is not affected, and works correctly.
Whether or not such code occurs in practice depends on coding style. One reasonable coding style is to initially use only seq_cst operations, and then selectively weaken those that are performance critical; it does result in such cases. Even in such cases, it seems clear that the current compilation strategy does not result in frequent failures; this problem was discovered through careful theoretical analysis, not bug reports. It is unclear whether there is any real code that can fail as a result of the current mapping; it would require careful analysis of the use cases to determine whether the weaker ordering provided by the hardware is in fact sufficient for these use cases.
For ARM, the situation is theoretically similar, but appears to be much less severe in practice. On ARMv8, the usual compilation mode for loads and stores is currently to compile acquire/release operations as seq_cst operations, so there is currently no issue. On ARMv7, some compilation schemes for acquire/release have the same issues as for Power, but the most common scheme seems to be to use "dmb ish", which does not share this problem.
Nvidia GPUs have a memory model similar to Power, and share the same issues, probably with larger cost differences. For very large numbers of cores, it is natural to share store buffers between threads, which may make stores visible to different threads in inconsistent orders. This lack of "multi-copy atomicity" is also the core distinguishing property of the Power and ARM memory models. We suspect that other GPUs are also affected, but cannot say that definitively.
We are not aware of issues on other existing CPU architectures. Since it appears more attractive to drop multi-copy atomicity with higher core counts, we expect the same issue to recur with some future processor designs.
Repairing this on Power without changing the specification would prevent us from
generating the lighter weight "lwsync" fence instruction for a
memory_order_release operation (unless we either know it will never
synchronize with a
memory_order_seq_cst operation, or we make
memory_order_seq_cst operations even more expensive), which would
have a significant performance impact on acquire/release synchronization. It
would also defeat a significant part, though not all, of the motivation for for
memory_order_release to start with.
The cost on GPUs is likely to be higher.
Among people we informally surveyed, this was not a popular option. Many people felt that we would be penalizing a subset of the available machine architectures for an issue with little practical impact. The language would no longer be able to express pure acquire-release synchronization, which many people feel is essential.
We could regain acquire-release synchronization by adding a new "weak" atomic
type that does not support
memory_order_seq_cst, and requires
memory_order arguments. Again there was general concern
that this significantly increases the library API footprint for an issue without
much practical impact.
This is the approach taken by Lahav et al, and the one we pursue here.
The proposal in Lahav et al is mathematically elegant. Currently the standard requires that the sequential consistency total order S is consistent with the happens before relation. Essentially, if any two sc operations are ordered by happens before, then they must be ordered the same way by S. In our example, this requires x =sc 1 to be ordered before b = fetch_add(y)sc //1 in spite of the fact that the hardware mapping does not sufficiently enforce it. The core fix (S1fix in the paper) is to relax the restriction that a happens before ordering implies ordering in S to only the sequenced before case, or the case in which the happens before ordering between a and b is produced by a chain
a is sequenced before x happens before y is sequenced before b
The downside of this is that "happens before" now has a rather strange meaning, since sequentially consistent operations can appear to execute in an order that's not consistent with it.
In the Z6.U example,
x =sc 1
must no longer precede
b = fetch_add(y)sc //1 in the
sequential consistency order S, in spite of
the fact that the former "happens before" the latter. And in the questionable
execution that we now wish to allow, they indeed have the opposite order in S.
We propose to make this somewhat less confusing by suitably renaming the relations in the standard as follows:
Currently the initialization rules, etc. use "strongly happens before" in guaranteeing ordering. The current intent is to also use that to specify library ordering, such as for mutexes. We propose to modify that definition to require "sequenced before" ordering at both ends. This new improved "strongly happens before" would be used in the same contexts as now, and would remain strong enough to ensure that if a happens before b, and they both participate in the sc ordering S, then a also precedes b in S. "Strongly happens before" would continue to exclude any ordering via memory_order_consume, since such ordering is much more restrictive, and must be explicitly accommodated by the caller.
Thus we would propose to change 4.7.1p11 [intro.races, a.k.a. 1.10] roughly as follows (this is an early attempt at wording, which is still under discussion):
An evaluation A
stronglyhappens before an evaluation B if either
- A is sequenced before B, or
- A synchronizes with B, or
- A simply happens before X and X simply happens before B.
[ Note: In the absence of consume operations, the happens before and
stronglyhappens before relations are identical. Stronglyhappens before essentially excludes consume operations. — end note ]
We would then adjust 32.4p3 [atomics.order] correspondingly:
There shall be a single total order S on all
memory_order_seq_cstoperations, consistent with the " happens before" order, and modification orders for all affected locations, such that each
memory_order_seq_cstoperation B that loads a value from an atomic object M observes one of the following values: …
and add a second note at the end of p3:
Note for editor: If the following section is also adopted, we just need the note added to the text below. The normative change avove is also included in the text below.]
memory_order_seq_cst fence semantics
do not guarantee that a program with only
memory_order_relaxed accesses and
between each pair actually exhibits sequential consistency. This was, at one
point, intentional. The goal was to ensure that architectures like Itanium that
allow stores to become visible to different processors in different orders, and
do not provide fences to rectify this, could be supported. But it subsequently
became clear that Itanium, as a result of failing to provide strong ordering for
accesses to a single location) would need to use stronger primitives for
memory_order_relaxed anyway. All known SC fence implementations provide the
stronger semantics, and we should acknowledge that.
We propose to strengthen the
memory_order_seq_cst fence semantics
as suggested in Lahav et al. (Note edit conflict with last section, and the
resolution suggested there.)
Replace Section 32.4 [atomics.order] paragraphs 3-7 (the definition of SC
The note from the previous section would go after this.
The definition of "coherence-ordered before" is essentially standard terminology, but was not previously part of our standard. The third bullet corresponds to what's often called "reads before": A reads a write earlier than B.
This new wording takes a significantly different approach with respect to the sequential consistency ordering S: Instead of defining visibility in terms of S and the other orders in the standard, this essentially defines constraints on S in terms of visibility in a particular execution, as expressed by the coherence order. If these constraints are not satisfiable by any total order S, then the candidate execution which gave rise to the coherence order is not valid.
This discussion largely follows Section 4.3 of Vafeiades et al, Common Compiler Optimisations are Invalid in the C11 Memory Model and what we can do about it, POPL 2015.
However, it became clear in Toronto that the proposal there, reflected in Option 3 below, is controversial, even among the authors of this proposal. We thus present the issues and three possible approaches.
We can illustrate the problem with the example from the above paper:
Thread 1 Thread 2 Thread 3
x =rlx 2
y =na 1
x =rel 1
x =rlx 3
if (xacq == 3)
Without Thread 1, this program is data-race-free.
Thread 3 accesses
y only if it sees a value of 3 for
which must mean that it saw the second assignment to
x by Thread 2,
which is in the release sequence of the release store. Hence the second assignment
synchronizes with the conditional in Thread 3.
Surprisingly, this no longer holds if we add Thread 1. If the Thread 1
assignment occurs (in
x's modification order) between the
two assignments in Thread 2, then the release sequence is broken by this
intervening assignment. There is no longer a synchronizes with relationship,
and thuis there is a data race.
This is highly counterintuitive, since Thread 1 should have no impact on memory ordering in this case. The current definition of release sequence was added fairly late in the C++11 cycle to compensate for clearly artificial weaknesses in the memory model without it, but before we had a clear understanding of actual hardware constraints, or a full understanding of programming requirements. On current actual hardware Thread 1 does not invalidate memory ordering guarantees.
The existing definition also greatly complicates reasoning about C++ programs. Consider the above example with non-atomic operations replaced by relaxed operations. That program allows Thread 3 to print zero, since again the release sequence can be broken by Thread 1. However the program without Thread 1 does not allow this execution, in spite of the fact that no Thread actually observes the write by Thread 1. In all reasonable senses of the word, the program without Thread 1 is a prefix of the whole program. But the execution of the whole program is not an extension of the program prefix. This is problematic for both formal and informal reasoning about programs.
Unfortunately, there are no painless solutions. The options include:
This seriously complicates reasoning about C++ programs in order to support a feature for which it is very hard to construct good use cases. It further seems that the current definition of release sequence presents some (relatively minor) poorly motivated obstacles for hardware designers.
Release sequences were originally introduced to prevent relaxed read-modify-write operations, from breaking synchronizes-with relationships. If a thread initializes a data structure, and then sets an initialized bit in a word to signal that it has done so, other threads using a compare_exchange to set other bits should not interfere with that signaling mechanism. Same-thread stores were then added to release sequences because it seemed at the time that this was a natural thing to do, and didn't introduce implementation issues on the platforms we knew about.
However, it is becoming increasingly common to design hardware features (e.g. ARMv8 acquire loads and release stores) to match the C++ memory model specification. For such new designs, release sequences do impose additional constraints, as was pointed out by the hardware architects participating in the Toronto discussion. It remains unclear that there are practical use cases that justify these constraints.
One possible solution, and the only one that appears to generate consensus among the paper authors, is to just remove same-thread-stores from release sequences. It is unclear whether this breaks actual code, but on paper this is a breaking change. Thus the actual proposal would be to deprecate the use of same-thread-store in release sequences, with the intent to remove that provision entirely around 2025.
Wording for this option will be provided in a later version of this paper.
We could, without breaking compatibility for existing code, change the release sequence definition so that unobserved assignments in other threads no longer break release sequences. This was the primary approach suggested by earlier academic work. It was controversial at the Toronto meeting, with architects expressing concern about adding to existing poorly motivated restrictions on hardware implementations.
This can be implemented with the following wording changes:
Update Section 4.7.1p5 [intro.races]:
sequenceheaded by a release operation A on an atomicobject M is a maximal sub-sequenceof side effects in the modification order of M. , where the first operation isA, and every subsequentoperation
- is performed by the same thread that performed A, or
- is an atomic read-modify-write operation that gets its value from
the immediately previous operation in the sequence.
In Section 4.7.1p8 [intro.races], rename "release sequence" to "release set" to maintain consistency with tha above. Similarly, replace the two occurrences of "release sequence" in 32.9 [atomics.fences] by "release set". Aside from the index, where the same replacement should be made, these are the only other occurrence of "release sequence".
In Toronto, we discussed an update D0668R1 of P0668R0 that added wording for the sequentially consistent fence changes, and that added the release sequence proposal. The first two proposals received strong support for the core idea; it was understood that the precise wording needed more bake time.
The vote on the release sequence proposal was delayed after hardware architects pointed out that it potentially imposed significant hardware constraints. It made sense to reexamine the significance of the underlying problem to make sure that the change was justified, particularly since the entire argument for allowing same thread stores to extend a release sequence now seems suspect.
This version incorporates the changes from the draft document we discussed, fixes some serious editing mistakes in D0668R1, adds further discussion for the release sequence proposal, and adjusts the desired straw poll list.