1 =========================
2 Atomic operations in QEMU
3 =========================
5 CPUs perform independent memory operations effectively in random order.
6 but this can be a problem for CPU-CPU interaction (including interactions
7 between QEMU and the guest). Multi-threaded programs use various tools
8 to instruct the compiler and the CPU to restrict the order to something
9 that is consistent with the expectations of the programmer.
11 The most basic tool is locking. Mutexes, condition variables and
12 semaphores are used in QEMU, and should be the default approach to
13 synchronization. Anything else is considerably harder, but it's
14 also justified more often than one would like;
15 the most performance-critical parts of QEMU in particular require
16 a very low level approach to concurrency, involving memory barriers
17 and atomic operations. The semantics of concurrent memory accesses are governed
18 by the C11 memory model.
20 QEMU provides a header, ``qemu/atomic.h``, which wraps C11 atomics to
21 provide better portability and a less verbose syntax. ``qemu/atomic.h``
22 provides macros that fall in three camps:
24 - compiler barriers: ``barrier()``;
26 - weak atomic access and manual memory barriers: ``atomic_read()``,
27 ``atomic_set()``, ``smp_rmb()``, ``smp_wmb()``, ``smp_mb()``, ``smp_mb_acquire()``,
28 ``smp_mb_release()``, ``smp_read_barrier_depends()``;
30 - sequentially consistent atomic access: everything else.
32 In general, use of ``qemu/atomic.h`` should be wrapped with more easily
33 used data structures (e.g. the lock-free singly-linked list operations
34 ``QSLIST_INSERT_HEAD_ATOMIC`` and ``QSLIST_MOVE_ATOMIC``) or synchronization
35 primitives (such as RCU, ``QemuEvent`` or ``QemuLockCnt``). Bare use of
36 atomic operations and memory barriers should be limited to inter-thread
37 checking of flags and documented thoroughly.
41 Compiler memory barrier
42 =======================
44 ``barrier()`` prevents the compiler from moving the memory accesses on
45 either side of it to the other side. The compiler barrier has no direct
46 effect on the CPU, which may then reorder things however it wishes.
48 ``barrier()`` is mostly used within ``qemu/atomic.h`` itself. On some
49 architectures, CPU guarantees are strong enough that blocking compiler
50 optimizations already ensures the correct order of execution. In this
51 case, ``qemu/atomic.h`` will reduce stronger memory barriers to simple
54 Still, ``barrier()`` can be useful when writing code that can be interrupted
58 Sequentially consistent atomic access
59 =====================================
61 Most of the operations in the ``qemu/atomic.h`` header ensure *sequential
62 consistency*, where "the result of any execution is the same as if the
63 operations of all the processors were executed in some sequential order,
64 and the operations of each individual processor appear in this sequence
65 in the order specified by its program".
67 ``qemu/atomic.h`` provides the following set of atomic read-modify-write
72 void atomic_add(ptr, val)
73 void atomic_sub(ptr, val)
74 void atomic_and(ptr, val)
75 void atomic_or(ptr, val)
77 typeof(*ptr) atomic_fetch_inc(ptr)
78 typeof(*ptr) atomic_fetch_dec(ptr)
79 typeof(*ptr) atomic_fetch_add(ptr, val)
80 typeof(*ptr) atomic_fetch_sub(ptr, val)
81 typeof(*ptr) atomic_fetch_and(ptr, val)
82 typeof(*ptr) atomic_fetch_or(ptr, val)
83 typeof(*ptr) atomic_fetch_xor(ptr, val)
84 typeof(*ptr) atomic_fetch_inc_nonzero(ptr)
85 typeof(*ptr) atomic_xchg(ptr, val)
86 typeof(*ptr) atomic_cmpxchg(ptr, old, new)
88 all of which return the old value of ``*ptr``. These operations are
89 polymorphic; they operate on any type that is as wide as a pointer or
92 Similar operations return the new value of ``*ptr``::
94 typeof(*ptr) atomic_inc_fetch(ptr)
95 typeof(*ptr) atomic_dec_fetch(ptr)
96 typeof(*ptr) atomic_add_fetch(ptr, val)
97 typeof(*ptr) atomic_sub_fetch(ptr, val)
98 typeof(*ptr) atomic_and_fetch(ptr, val)
99 typeof(*ptr) atomic_or_fetch(ptr, val)
100 typeof(*ptr) atomic_xor_fetch(ptr, val)
102 ``qemu/atomic.h`` also provides loads and stores that cannot be reordered
105 typeof(*ptr) atomic_mb_read(ptr)
106 void atomic_mb_set(ptr, val)
108 However these do not provide sequential consistency and, in particular,
109 they do not participate in the total ordering enforced by
110 sequentially-consistent operations. For this reason they are deprecated.
111 They should instead be replaced with any of the following (ordered from
114 - accesses inside a mutex or spinlock
116 - lightweight synchronization primitives such as ``QemuEvent``
118 - RCU operations (``atomic_rcu_read``, ``atomic_rcu_set``) when publishing
119 or accessing a new version of a data structure
121 - other atomic accesses: ``atomic_read`` and ``atomic_load_acquire`` for
122 loads, ``atomic_set`` and ``atomic_store_release`` for stores, ``smp_mb``
123 to forbid reordering subsequent loads before a store.
126 Weak atomic access and manual memory barriers
127 =============================================
129 Compared to sequentially consistent atomic access, programming with
130 weaker consistency models can be considerably more complicated.
131 The only guarantees that you can rely upon in this case are:
133 - atomic accesses will not cause data races (and hence undefined behavior);
134 ordinary accesses instead cause data races if they are concurrent with
135 other accesses of which at least one is a write. In order to ensure this,
136 the compiler will not optimize accesses out of existence, create unsolicited
137 accesses, or perform other similar optimzations.
139 - acquire operations will appear to happen, with respect to the other
140 components of the system, before all the LOAD or STORE operations
141 specified afterwards.
143 - release operations will appear to happen, with respect to the other
144 components of the system, after all the LOAD or STORE operations
147 - release operations will *synchronize with* acquire operations;
148 see :ref:`acqrel` for a detailed explanation.
150 When using this model, variables are accessed with:
152 - ``atomic_read()`` and ``atomic_set()``; these prevent the compiler from
153 optimizing accesses out of existence and creating unsolicited
154 accesses, but do not otherwise impose any ordering on loads and
155 stores: both the compiler and the processor are free to reorder
158 - ``atomic_load_acquire()``, which guarantees the LOAD to appear to
159 happen, with respect to the other components of the system,
160 before all the LOAD or STORE operations specified afterwards.
161 Operations coming before ``atomic_load_acquire()`` can still be
164 - ``atomic_store_release()``, which guarantees the STORE to appear to
165 happen, with respect to the other components of the system,
166 after all the LOAD or STORE operations specified before.
167 Operations coming after ``atomic_store_release()`` can still be
170 Restrictions to the ordering of accesses can also be specified
171 using the memory barrier macros: ``smp_rmb()``, ``smp_wmb()``, ``smp_mb()``,
172 ``smp_mb_acquire()``, ``smp_mb_release()``, ``smp_read_barrier_depends()``.
174 Memory barriers control the order of references to shared memory.
175 They come in six kinds:
177 - ``smp_rmb()`` guarantees that all the LOAD operations specified before
178 the barrier will appear to happen before all the LOAD operations
179 specified after the barrier with respect to the other components of
182 In other words, ``smp_rmb()`` puts a partial ordering on loads, but is not
183 required to have any effect on stores.
185 - ``smp_wmb()`` guarantees that all the STORE operations specified before
186 the barrier will appear to happen before all the STORE operations
187 specified after the barrier with respect to the other components of
190 In other words, ``smp_wmb()`` puts a partial ordering on stores, but is not
191 required to have any effect on loads.
193 - ``smp_mb_acquire()`` guarantees that all the LOAD operations specified before
194 the barrier will appear to happen before all the LOAD or STORE operations
195 specified after the barrier with respect to the other components of
198 - ``smp_mb_release()`` guarantees that all the STORE operations specified *after*
199 the barrier will appear to happen after all the LOAD or STORE operations
200 specified *before* the barrier with respect to the other components of
203 - ``smp_mb()`` guarantees that all the LOAD and STORE operations specified
204 before the barrier will appear to happen before all the LOAD and
205 STORE operations specified after the barrier with respect to the other
206 components of the system.
208 ``smp_mb()`` puts a partial ordering on both loads and stores. It is
209 stronger than both a read and a write memory barrier; it implies both
210 ``smp_mb_acquire()`` and ``smp_mb_release()``, but it also prevents STOREs
211 coming before the barrier from overtaking LOADs coming after the
212 barrier and vice versa.
214 - ``smp_read_barrier_depends()`` is a weaker kind of read barrier. On
215 most processors, whenever two loads are performed such that the
216 second depends on the result of the first (e.g., the first load
217 retrieves the address to which the second load will be directed),
218 the processor will guarantee that the first LOAD will appear to happen
219 before the second with respect to the other components of the system.
220 However, this is not always true---for example, it was not true on
221 Alpha processors. Whenever this kind of access happens to shared
222 memory (that is not protected by a lock), a read barrier is needed,
223 and ``smp_read_barrier_depends()`` can be used instead of ``smp_rmb()``.
225 Note that the first load really has to have a _data_ dependency and not
226 a control dependency. If the address for the second load is dependent
227 on the first load, but the dependency is through a conditional rather
228 than actually loading the address itself, then it's a _control_
229 dependency and a full read barrier or better is required.
232 Memory barriers and ``atomic_load_acquire``/``atomic_store_release`` are
233 mostly used when a data structure has one thread that is always a writer
234 and one thread that is always a reader:
236 +----------------------------------+----------------------------------+
237 | thread 1 | thread 2 |
238 +==================================+==================================+
241 | atomic_store_release(&a, x); | y = atomic_load_acquire(&b); |
242 | atomic_store_release(&b, y); | x = atomic_load_acquire(&a); |
243 +----------------------------------+----------------------------------+
245 In this case, correctness is easy to check for using the "pairing"
246 trick that is explained below.
248 Sometimes, a thread is accessing many variables that are otherwise
249 unrelated to each other (for example because, apart from the current
250 thread, exactly one other thread will read or write each of these
251 variables). In this case, it is possible to "hoist" the barriers
252 outside a loop. For example:
254 +------------------------------------------+----------------------------------+
256 +==========================================+==================================+
260 | for (i = 0; i < 10; i++) | for (i = 0; i < 10; i++) |
261 | n += atomic_load_acquire(&a[i]); | n += atomic_read(&a[i]); |
262 | | smp_mb_acquire(); |
263 +------------------------------------------+----------------------------------+
266 | | smp_mb_release(); |
267 | for (i = 0; i < 10; i++) | for (i = 0; i < 10; i++) |
268 | atomic_store_release(&a[i], false); | atomic_set(&a[i], false); |
269 +------------------------------------------+----------------------------------+
271 Splitting a loop can also be useful to reduce the number of barriers:
273 +------------------------------------------+----------------------------------+
275 +==========================================+==================================+
278 | n = 0; | smp_mb_release(); |
279 | for (i = 0; i < 10; i++) { | for (i = 0; i < 10; i++) |
280 | atomic_store_release(&a[i], false); | atomic_set(&a[i], false); |
281 | smp_mb(); | smb_mb(); |
282 | n += atomic_read(&b[i]); | n = 0; |
283 | } | for (i = 0; i < 10; i++) |
284 | | n += atomic_read(&b[i]); |
285 +------------------------------------------+----------------------------------+
287 In this case, a ``smp_mb_release()`` is also replaced with a (possibly cheaper, and clearer
288 as well) ``smp_wmb()``:
290 +------------------------------------------+----------------------------------+
292 +==========================================+==================================+
295 | | smp_mb_release(); |
296 | for (i = 0; i < 10; i++) { | for (i = 0; i < 10; i++) |
297 | atomic_store_release(&a[i], false); | atomic_set(&a[i], false); |
298 | atomic_store_release(&b[i], false); | smb_wmb(); |
299 | } | for (i = 0; i < 10; i++) |
300 | | atomic_set(&b[i], false); |
301 +------------------------------------------+----------------------------------+
306 Acquire/release pairing and the *synchronizes-with* relation
307 ------------------------------------------------------------
309 Atomic operations other than ``atomic_set()`` and ``atomic_read()`` have
310 either *acquire* or *release* semantics [#rmw]_. This has two effects:
312 .. [#rmw] Read-modify-write operations can have both---acquire applies to the
313 read part, and release to the write.
315 - within a thread, they are ordered either before subsequent operations
316 (for acquire) or after previous operations (for release).
318 - if a release operation in one thread *synchronizes with* an acquire operation
319 in another thread, the ordering constraints propagates from the first to the
320 second thread. That is, everything before the release operation in the
321 first thread is guaranteed to *happen before* everything after the
322 acquire operation in the second thread.
324 The concept of acquire and release semantics is not exclusive to atomic
325 operations; almost all higher-level synchronization primitives also have
326 acquire or release semantics. For example:
328 - ``pthread_mutex_lock`` has acquire semantics, ``pthread_mutex_unlock`` has
329 release semantics and synchronizes with a ``pthread_mutex_lock`` for the
332 - ``pthread_cond_signal`` and ``pthread_cond_broadcast`` have release semantics;
333 ``pthread_cond_wait`` has both release semantics (synchronizing with
334 ``pthread_mutex_lock``) and acquire semantics (synchronizing with
335 ``pthread_mutex_unlock`` and signaling of the condition variable).
337 - ``pthread_create`` has release semantics and synchronizes with the start
338 of the new thread; ``pthread_join`` has acquire semantics and synchronizes
339 with the exiting of the thread.
341 - ``qemu_event_set`` has release semantics, ``qemu_event_wait`` has
344 For example, in the following example there are no atomic accesses, but still
345 thread 2 is relying on the *synchronizes-with* relation between ``pthread_exit``
346 (release) and ``pthread_join`` (acquire):
348 +----------------------+-------------------------------+
349 | thread 1 | thread 2 |
350 +======================+===============================+
354 | pthread_exit(a); | pthread_join(thread1, &a); |
356 +----------------------+-------------------------------+
358 Synchronization between threads basically descends from this pairing of
359 a release operation and an acquire operation. Therefore, atomic operations
360 other than ``atomic_set()`` and ``atomic_read()`` will almost always be
361 paired with another operation of the opposite kind: an acquire operation
362 will pair with a release operation and vice versa. This rule of thumb is
363 extremely useful; in the case of QEMU, however, note that the other
364 operation may actually be in a driver that runs in the guest!
366 ``smp_read_barrier_depends()``, ``smp_rmb()``, ``smp_mb_acquire()``,
367 ``atomic_load_acquire()`` and ``atomic_rcu_read()`` all count
368 as acquire operations. ``smp_wmb()``, ``smp_mb_release()``,
369 ``atomic_store_release()`` and ``atomic_rcu_set()`` all count as release
370 operations. ``smp_mb()`` counts as both acquire and release, therefore
371 it can pair with any other atomic operation. Here is an example:
373 +----------------------+------------------------------+
374 | thread 1 | thread 2 |
375 +======================+==============================+
378 | atomic_set(&a, 1); | |
380 | atomic_set(&b, 2); | x = atomic_read(&b); |
382 | | y = atomic_read(&a); |
383 +----------------------+------------------------------+
385 Note that a load-store pair only counts if the two operations access the
386 same variable: that is, a store-release on a variable ``x`` *synchronizes
387 with* a load-acquire on a variable ``x``, while a release barrier
388 synchronizes with any acquire operation. The following example shows
389 correct synchronization:
391 +--------------------------------+--------------------------------+
392 | thread 1 | thread 2 |
393 +================================+================================+
396 | atomic_set(&a, 1); | |
397 | atomic_store_release(&b, 2); | x = atomic_load_acquire(&b); |
398 | | y = atomic_read(&a); |
399 +--------------------------------+--------------------------------+
401 Acquire and release semantics of higher-level primitives can also be
402 relied upon for the purpose of establishing the *synchronizes with*
405 Note that the "writing" thread is accessing the variables in the
406 opposite order as the "reading" thread. This is expected: stores
407 before a release operation will normally match the loads after
408 the acquire operation, and vice versa. In fact, this happened already
409 in the ``pthread_exit``/``pthread_join`` example above.
411 Finally, this more complex example has more than two accesses and data
412 dependency barriers. It also does not use atomic accesses whenever there
413 cannot be a data race:
415 +----------------------+------------------------------+
416 | thread 1 | thread 2 |
417 +======================+==============================+
424 | atomic_set(&a, x); | x = atomic_read(&a); |
425 | | smp_read_barrier_depends(); |
427 | | smp_read_barrier_depends(); |
429 +----------------------+------------------------------+
431 Comparison with Linux kernel primitives
432 =======================================
434 Here is a list of differences between Linux kernel atomic operations
435 and memory barriers, and the equivalents in QEMU:
437 - atomic operations in Linux are always on a 32-bit int type and
438 use a boxed ``atomic_t`` type; atomic operations in QEMU are polymorphic
439 and use normal C types.
441 - Originally, ``atomic_read`` and ``atomic_set`` in Linux gave no guarantee
442 at all. Linux 4.1 updated them to implement volatile
443 semantics via ``ACCESS_ONCE`` (or the more recent ``READ``/``WRITE_ONCE``).
445 QEMU's ``atomic_read`` and ``atomic_set`` implement C11 atomic relaxed
446 semantics if the compiler supports it, and volatile semantics otherwise.
447 Both semantics prevent the compiler from doing certain transformations;
448 the difference is that atomic accesses are guaranteed to be atomic,
449 while volatile accesses aren't. Thus, in the volatile case we just cross
450 our fingers hoping that the compiler will generate atomic accesses,
451 since we assume the variables passed are machine-word sized and
454 No barriers are implied by ``atomic_read`` and ``atomic_set`` in either Linux
457 - atomic read-modify-write operations in Linux are of three kinds:
459 ===================== =========================================
460 ``atomic_OP`` returns void
461 ``atomic_OP_return`` returns new value of the variable
462 ``atomic_fetch_OP`` returns the old value of the variable
463 ``atomic_cmpxchg`` returns the old value of the variable
464 ===================== =========================================
466 In QEMU, the second kind is named ``atomic_OP_fetch``.
468 - different atomic read-modify-write operations in Linux imply
469 a different set of memory barriers; in QEMU, all of them enforce
470 sequential consistency.
472 - in QEMU, ``atomic_read()`` and ``atomic_set()`` do not participate in
473 the total ordering enforced by sequentially-consistent operations.
474 This is because QEMU uses the C11 memory model. The following example
475 is correct in Linux but not in QEMU:
477 +----------------------------------+--------------------------------+
478 | Linux (correct) | QEMU (incorrect) |
479 +==================================+================================+
482 | a = atomic_fetch_add(&x, 2); | a = atomic_fetch_add(&x, 2); |
483 | b = READ_ONCE(&y); | b = atomic_read(&y); |
484 +----------------------------------+--------------------------------+
486 because the read of ``y`` can be moved (by either the processor or the
487 compiler) before the write of ``x``.
489 Fixing this requires an ``smp_mb()`` memory barrier between the write
490 of ``x`` and the read of ``y``. In the common case where only one thread
491 writes ``x``, it is also possible to write it like this:
493 +--------------------------------+
495 +================================+
498 | a = atomic_read(&x); |
499 | atomic_set(&x, a + 2); |
501 | b = atomic_read(&y); |
502 +--------------------------------+
507 - ``Documentation/memory-barriers.txt`` from the Linux kernel