9 “A” Standard Extension for Atomic Instructions, Version 2.1
The standard atomic-instruction extension, named “A”, contains instructions that atomically read-modify-write memory to support synchronization between multiple RISC-V harts running in the same memory space. The two forms of atomic instruction provided are load-reserved/store-conditional instructions and atomic fetch-and-op memory instructions. Both types of atomic instruction support various memory consistency orderings including unordered, acquire, release, and sequentially consistent semantics. These instructions allow RISC-V to support the RCsc memory consistency model [Gharachorloo90memoryconsistency].
9.1 Specifying Ordering of Atomic Instructions
The base RISC-V ISA has a relaxed memory model, with the FENCE instruction used to impose additional ordering constraints. The address space is divided by the execution environment into memory and I/O domains, and the FENCE instruction provides options to order accesses to one or both of these two address domains.
To provide more efficient support for release consistency [Gharachorloo90memoryconsistency], each atomic instruction has two bits, aq and rl, used to specify additional memory ordering constraints as viewed by other RISC-V harts. The bits order accesses to one of the two address domains, memory or I/O, depending on which address domain the atomic instruction is accessing. No ordering constraint is implied to accesses to the other domain, and a FENCE instruction should be used to order across both domains.
If both bits are clear, no additional ordering constraints are imposed on the atomic memory operation. If only the aq bit is set, the atomic memory operation is treated as an acquire access, i.e., no following memory operations on this RISC-V hart can be observed to take place before the acquire memory operation. If only the rl bit is set, the atomic memory operation is treated as a release access, i.e., the release memory operation cannot be observed to take place before any earlier memory operations on this RISC-V hart. If both the aq and rl bits are set, the atomic memory operation is sequentially consistent and cannot be observed to happen before any earlier memory operations or after any later memory operations in the same RISC-V hart and to the same address domain.
9.2 Load-Reserved/Store-Conditional Instructions
Complex atomic memory operations on a single memory word or doubleword are performed with the load-reserved (LR) and store-conditional (SC) instructions. LR.W loads a word from the address in rs1, places the sign-extended value in rd, and registers a reservation set—a set of bytes that subsumes the bytes in the addressed word. SC.W conditionally writes a word in rs2 to the address in rs1: the SC.W succeeds only if the reservation is still valid and the reservation set contains the bytes being written. If the SC.W succeeds, the instruction writes the word in rs2 to memory, and it writes zero to rd. If the SC.W fails, the instruction does not write to memory, and it writes a nonzero value to rd. Regardless of success or failure, executing an SC.W instruction invalidates any reservation held by this hart. LR.D and SC.D act analogously on doublewords and are only available on RV64. For RV64, LR.W and SC.W sign-extend the value placed in rd.
Both compare-and-swap (CAS) and LR/SC can be used to build lock-free data structures. After extensive discussion, we opted for LR/SC for several reasons: 1) CAS suffers from the ABA problem, which LR/SC avoids because it monitors all writes to the address rather than only checking for changes in the data value; 2) CAS would also require a new integer instruction format to support three source operands (address, compare value, swap value) as well as a different memory system message format, which would complicate microarchitectures; 3) Furthermore, to avoid the ABA problem, other systems provide a double-wide CAS (DW-CAS) to allow a counter to be tested and incremented along with a data word. This requires reading five registers and writing two in one instruction, and also a new larger memory system message type, further complicating implementations; 4) LR/SC provides a more efficient implementation of many primitives as it only requires one load as opposed to two with CAS (one load before the CAS instruction to obtain a value for speculative computation, then a second load as part of the CAS instruction to check if value is unchanged before updating).
The main disadvantage of LR/SC over CAS is livelock, which we avoid, under certain circumstances, with an architected guarantee of eventual forward progress as described below. Another concern is whether the influence of the current x86 architecture, with its DW-CAS, will complicate porting of synchronization libraries and other software that assumes DW-CAS is the basic machine primitive. A possible mitigating factor is the recent addition of transactional memory instructions to x86, which might cause a move away from DW-CAS.
More generally, a multi-word atomic primitive is desirable, but there is still considerable debate about what form this should take, and guaranteeing forward progress adds complexity to a system.
The failure code with value 1 encodes an unspecified failure. Other failure codes are reserved at this time. Portable software should only assume the failure code will be non-zero.
We reserve a failure code of 1 to mean “unspecified” so that simple implementations may return this value using the existing mux required for the SLT/SLTU instructions. More specific failure codes might be defined in future versions or extensions to the ISA.
For LR and SC, the A extension requires that the address held in rs1 be naturally aligned to the size of the operand (i.e., eight-byte aligned for 64-bit words and four-byte aligned for 32-bit words). If the address is not naturally aligned, an address-misaligned exception or an access-fault exception will be generated. The access-fault exception can be generated for a memory access that would otherwise be able to complete except for the misalignment, if the misaligned access should not be emulated.
Emulating misaligned LR/SC sequences is impractical in most systems.
Misaligned LR/SC sequences also raise the possibility of accessing multiple reservation sets at once, which present definitions do not provide for.
An implementation can register an arbitrarily large reservation set on each LR, provided the reservation set includes all bytes of the addressed data word or doubleword. An SC can only pair with the most recent LR in program order. An SC may succeed only if no store from another hart to the reservation set can be observed to have occurred between the LR and the SC, and if there is no other SC between the LR and itself in program order. An SC may succeed only if no write from a device other than a hart to the bytes accessed by the LR instruction can be observed to have occurred between the LR and SC. Note this LR might have had a different effective address and data size, but reserved the SC’s address as part of the reservation set.
Following this model, in systems with memory translation, an SC is allowed to succeed if the earlier LR reserved the same location using an alias with a different virtual address, but is also allowed to fail if the virtual address is different.
To accommodate legacy devices and buses, writes from devices other than RISC-V harts are only required to invalidate reservations when they overlap the bytes accessed by the LR. These writes are not required to invalidate the reservation when they access other bytes in the reservation set.
The SC must fail if the address is not within the reservation set of the most recent LR in program order. The SC must fail if a store to the reservation set from another hart can be observed to occur between the LR and SC. The SC must fail if a write from some other device to the bytes accessed by the LR can be observed to occur between the LR and SC. (If such a device writes the reservation set but does not write the bytes accessed by the LR, the SC may or may not fail.) An SC must fail if there is another SC (to any address) between the LR and the SC in program order. The precise statement of the atomicity requirements for successful LR/SC sequences is defined by the Atomicity Axiom in Section [sec:rvwmo].
The platform should provide a means to determine the size and shape of the reservation set.
A platform specification may constrain the size and shape of the reservation set.
A store-conditional instruction to a scratch word of memory should be used to forcibly invalidate any existing load reservation:
during a preemptive context switch, and
if necessary when changing virtual to physical address mappings, such as when migrating pages that might contain an active reservation.
The invalidation of a hart’s reservation when it executes an LR or SC imply that a hart can only hold one reservation at a time, and that an SC can only pair with the most recent LR, and LR with the next following SC, in program order. This is a restriction to the Atomicity Axiom in Section [sec:rvwmo] that ensures software runs correctly on expected common implementations that operate in this manner.
An SC instruction can never be observed by another RISC-V hart before the LR instruction that established the reservation. The LR/SC sequence can be given acquire semantics by setting the aq bit on the LR instruction. The LR/SC sequence can be given release semantics by setting the rl bit on the SC instruction. Setting the aq bit on the LR instruction, and setting both the aq and the rl bit on the SC instruction makes the LR/SC sequence sequentially consistent, meaning that it cannot be reordered with earlier or later memory operations from the same hart.
If neither bit is set on both LR and SC, the LR/SC sequence can be observed to occur before or after surrounding memory operations from the same RISC-V hart. This can be appropriate when the LR/SC sequence is used to implement a parallel reduction operation.
Software should not set the rl bit on an LR instruction unless the aq bit is also set, nor should software set the aq bit on an SC instruction unless the rl bit is also set. LR.rl and SC.aq instructions are not guaranteed to provide any stronger ordering than those with both bits clear, but may result in lower performance.
LR/SC can be used to construct lock-free data structures. An example using LR/SC to implement a compare-and-swap function is shown in Figure 1.1. If inlined, compare-and-swap functionality need only take four instructions.
9.3 Eventual Success of Store-Conditional Instructions
The standard A extension defines constrained LR/SC loops, which have the following properties:
The loop comprises only an LR/SC sequence and code to retry the sequence in the case of failure, and must comprise at most 16 instructions placed sequentially in memory.
An LR/SC sequence begins with an LR instruction and ends with an SC instruction. The dynamic code executed between the LR and SC instructions can only contain instructions from the base “I” instruction set, excluding loads, stores, backward jumps, taken backward branches, JALR, FENCE, and SYSTEM instructions. If the “C” extension is supported, then compressed forms of the aforementioned “I” instructions are also permitted.
The code to retry a failing LR/SC sequence can contain backwards jumps and/or branches to repeat the LR/SC sequence, but otherwise has the same constraint as the code between the LR and SC.
The LR and SC addresses must lie within a memory region with the LR/SC eventuality property. The execution environment is responsible for communicating which regions have this property.
The SC must be to the same effective address and of the same data size as the latest LR executed by the same hart.
LR/SC sequences that do not lie within constrained LR/SC loops are unconstrained. Unconstrained LR/SC sequences might succeed on some attempts on some implementations, but might never succeed on other implementations.
We restricted the length of LR/SC loops to fit within 64 contiguous instruction bytes in the base ISA to avoid undue restrictions on instruction cache and TLB size and associativity. Similarly, we disallowed other loads and stores within the loops to avoid restrictions on data-cache associativity in simple implementations that track the reservation within a private cache. The restrictions on branches and jumps limit the time that can be spent in the sequence. Floating-point operations and integer multiply/divide were disallowed to simplify the operating system’s emulation of these instructions on implementations lacking appropriate hardware support.
Software is not forbidden from using unconstrained LR/SC sequences, but portable software must detect the case that the sequence repeatedly fails, then fall back to an alternate code sequence that does not rely on an unconstrained LR/SC sequence. Implementations are permitted to unconditionally fail any unconstrained LR/SC sequence.
If a hart H enters a constrained LR/SC loop, the execution environment must guarantee that one of the following events eventually occurs:
H or some other hart executes a successful SC to the reservation set of the LR instruction in H’s constrained LR/SC loops.
Some other hart executes an unconditional store or AMO instruction to the reservation set of the LR instruction in H’s constrained LR/SC loop, or some other device in the system writes to that reservation set.
H executes a branch or jump that exits the constrained LR/SC loop.
H traps.
Note that these definitions permit an implementation to fail an SC instruction occasionally for any reason, provided the aforementioned guarantee is not violated.
As a consequence of the eventuality guarantee, if some harts in an execution environment are executing constrained LR/SC loops, and no other harts or devices in the execution environment execute an unconditional store or AMO to that reservation set, then at least one hart will eventually exit its constrained LR/SC loop. By contrast, if other harts or devices continue to write to that reservation set, it is not guaranteed that any hart will exit its LR/SC loop.
Loads and load-reserved instructions do not by themselves impede the progress of other harts’ LR/SC sequences. We note this constraint implies, among other things, that loads and load-reserved instructions executed by other harts (possibly within the same core) cannot impede LR/SC progress indefinitely. For example, cache evictions caused by another hart sharing the cache cannot impede LR/SC progress indefinitely. Typically, this implies reservations are tracked independently of evictions from any shared cache. Similarly, cache misses caused by speculative execution within a hart cannot impede LR/SC progress indefinitely.
These definitions admit the possibility that SC instructions may spuriously fail for implementation reasons, provided progress is eventually made.
One advantage of CAS is that it guarantees that some hart eventually makes progress, whereas an LR/SC atomic sequence could livelock indefinitely on some systems. To avoid this concern, we added an architectural guarantee of livelock freedom for certain LR/SC sequences.
Earlier versions of this specification imposed a stronger starvation-freedom guarantee. However, the weaker livelock-freedom guarantee is sufficient to implement the C11 and C++11 languages, and is substantially easier to provide in some microarchitectural styles.
9.4 Atomic Memory Operations
The atomic memory operation (AMO) instructions perform read-modify-write operations for multiprocessor synchronization and are encoded with an R-type instruction format. These AMO instructions atomically load a data value from the address in rs1, place the value into register rd, apply a binary operator to the loaded value and the original value in rs2, then store the result back to the original address in rs1. AMOs can either operate on 64-bit (RV64 only) or 32-bit words in memory. For RV64, 32-bit AMOs always sign-extend the value placed in rd, and ignore the upper 32 bits of the original value of rs2.
For AMOs, the A extension requires that the address held in rs1 be naturally aligned to the size of the operand (i.e., eight-byte aligned for 64-bit words and four-byte aligned for 32-bit words). If the address is not naturally aligned, an address-misaligned exception or an access-fault exception will be generated. The access-fault exception can be generated for a memory access that would otherwise be able to complete except for the misalignment, if the misaligned access should not be emulated. The “Zam” extension, described in Chapter [sec:zam], relaxes this requirement and specifies the semantics of misaligned AMOs.
The operations supported are swap, integer add, bitwise AND, bitwise
OR, bitwise XOR, and signed and unsigned integer maximum and minimum.
Without ordering constraints, these AMOs can be used to implement
parallel reduction operations, where typically the return value would
be discarded by writing to x0.
We provided fetch-and-op style atomic primitives as they scale to
highly parallel systems better than LR/SC or CAS.
A simple microarchitecture can implement AMOs using the LR/SC primitives,
provided the implementation can guarantee the AMO eventually completes.
More complex implementations might also implement AMOs at memory
controllers, and can optimize away fetching the original value when
the destination is x0.
The set of AMOs was chosen to support the C11/C++11 atomic memory operations efficiently, and also to support parallel reductions in memory. Another use of AMOs is to provide atomic updates to memory-mapped device registers (e.g., setting, clearing, or toggling bits) in the I/O space.
To help implement multiprocessor synchronization, the AMOs optionally provide release consistency semantics. If the aq bit is set, then no later memory operations in this RISC-V hart can be observed to take place before the AMO. Conversely, if the rl bit is set, then other RISC-V harts will not observe the AMO before memory accesses preceding the AMO in this RISC-V hart. Setting both the aq and the rl bit on an AMO makes the sequence sequentially consistent, meaning that it cannot be reordered with earlier or later memory operations from the same hart.
The AMOs were designed to implement the C11 and C++11 memory models efficiently. Although the FENCE R, RW instruction suffices to implement the acquire operation and FENCE RW, W suffices to implement release, both imply additional unnecessary ordering as compared to AMOs with the corresponding aq or rl bit set.
An example code sequence for a critical section guarded by a test-and-test-and-set spinlock is shown in Figure 1.2. Note the first AMO is marked aq to order the lock acquisition before the critical section, and the second AMO is marked rl to order the critical section before the lock relinquishment.
a0 contains the address of the lock.We recommend the use of the AMO Swap idiom shown above for both lock acquire and release to simplify the implementation of speculative lock elision [Rajwar:2001:SLE].
The instructions in the “A” extension can also be used to provide sequentially consistent loads and stores. A sequentially consistent load can be implemented as an LR with both aq and rl set. A sequentially consistent store can be implemented as an AMOSWAP that writes the old value to x0 and has both aq and rl set.
After much debate, the language community and architecture community appear to have finally settled on release consistency as the standard memory consistency model and so the RISC-V atomic support is built around this model.