xref: /openbmc/qemu/docs/devel/atomics.rst (revision ffe98631)
1.. _atomics-ref:
2
3=========================
4Atomic operations in QEMU
5=========================
6
7CPUs perform independent memory operations effectively in random order.
8but this can be a problem for CPU-CPU interaction (including interactions
9between QEMU and the guest).  Multi-threaded programs use various tools
10to instruct the compiler and the CPU to restrict the order to something
11that is consistent with the expectations of the programmer.
12
13The most basic tool is locking.  Mutexes, condition variables and
14semaphores are used in QEMU, and should be the default approach to
15synchronization.  Anything else is considerably harder, but it's
16also justified more often than one would like;
17the most performance-critical parts of QEMU in particular require
18a very low level approach to concurrency, involving memory barriers
19and atomic operations.  The semantics of concurrent memory accesses are governed
20by the C11 memory model.
21
22QEMU provides a header, ``qemu/atomic.h``, which wraps C11 atomics to
23provide better portability and a less verbose syntax.  ``qemu/atomic.h``
24provides macros that fall in three camps:
25
26- compiler barriers: ``barrier()``;
27
28- weak atomic access and manual memory barriers: ``qatomic_read()``,
29  ``qatomic_set()``, ``smp_rmb()``, ``smp_wmb()``, ``smp_mb()``,
30  ``smp_mb_acquire()``, ``smp_mb_release()``, ``smp_read_barrier_depends()``;
31
32- sequentially consistent atomic access: everything else.
33
34In general, use of ``qemu/atomic.h`` should be wrapped with more easily
35used data structures (e.g. the lock-free singly-linked list operations
36``QSLIST_INSERT_HEAD_ATOMIC`` and ``QSLIST_MOVE_ATOMIC``) or synchronization
37primitives (such as RCU, ``QemuEvent`` or ``QemuLockCnt``).  Bare use of
38atomic operations and memory barriers should be limited to inter-thread
39checking of flags and documented thoroughly.
40
41
42
43Compiler memory barrier
44=======================
45
46``barrier()`` prevents the compiler from moving the memory accesses on
47either side of it to the other side.  The compiler barrier has no direct
48effect on the CPU, which may then reorder things however it wishes.
49
50``barrier()`` is mostly used within ``qemu/atomic.h`` itself.  On some
51architectures, CPU guarantees are strong enough that blocking compiler
52optimizations already ensures the correct order of execution.  In this
53case, ``qemu/atomic.h`` will reduce stronger memory barriers to simple
54compiler barriers.
55
56Still, ``barrier()`` can be useful when writing code that can be interrupted
57by signal handlers.
58
59
60Sequentially consistent atomic access
61=====================================
62
63Most of the operations in the ``qemu/atomic.h`` header ensure *sequential
64consistency*, where "the result of any execution is the same as if the
65operations of all the processors were executed in some sequential order,
66and the operations of each individual processor appear in this sequence
67in the order specified by its program".
68
69``qemu/atomic.h`` provides the following set of atomic read-modify-write
70operations::
71
72    void qatomic_inc(ptr)
73    void qatomic_dec(ptr)
74    void qatomic_add(ptr, val)
75    void qatomic_sub(ptr, val)
76    void qatomic_and(ptr, val)
77    void qatomic_or(ptr, val)
78
79    typeof(*ptr) qatomic_fetch_inc(ptr)
80    typeof(*ptr) qatomic_fetch_dec(ptr)
81    typeof(*ptr) qatomic_fetch_add(ptr, val)
82    typeof(*ptr) qatomic_fetch_sub(ptr, val)
83    typeof(*ptr) qatomic_fetch_and(ptr, val)
84    typeof(*ptr) qatomic_fetch_or(ptr, val)
85    typeof(*ptr) qatomic_fetch_xor(ptr, val)
86    typeof(*ptr) qatomic_fetch_inc_nonzero(ptr)
87    typeof(*ptr) qatomic_xchg(ptr, val)
88    typeof(*ptr) qatomic_cmpxchg(ptr, old, new)
89
90all of which return the old value of ``*ptr``.  These operations are
91polymorphic; they operate on any type that is as wide as a pointer or
92smaller.
93
94Similar operations return the new value of ``*ptr``::
95
96    typeof(*ptr) qatomic_inc_fetch(ptr)
97    typeof(*ptr) qatomic_dec_fetch(ptr)
98    typeof(*ptr) qatomic_add_fetch(ptr, val)
99    typeof(*ptr) qatomic_sub_fetch(ptr, val)
100    typeof(*ptr) qatomic_and_fetch(ptr, val)
101    typeof(*ptr) qatomic_or_fetch(ptr, val)
102    typeof(*ptr) qatomic_xor_fetch(ptr, val)
103
104``qemu/atomic.h`` also provides loads and stores that cannot be reordered
105with each other::
106
107    typeof(*ptr) qatomic_mb_read(ptr)
108    void         qatomic_mb_set(ptr, val)
109
110However these do not provide sequential consistency and, in particular,
111they do not participate in the total ordering enforced by
112sequentially-consistent operations.  For this reason they are deprecated.
113They should instead be replaced with any of the following (ordered from
114easiest to hardest):
115
116- accesses inside a mutex or spinlock
117
118- lightweight synchronization primitives such as ``QemuEvent``
119
120- RCU operations (``qatomic_rcu_read``, ``qatomic_rcu_set``) when publishing
121  or accessing a new version of a data structure
122
123- other atomic accesses: ``qatomic_read`` and ``qatomic_load_acquire`` for
124  loads, ``qatomic_set`` and ``qatomic_store_release`` for stores, ``smp_mb``
125  to forbid reordering subsequent loads before a store.
126
127
128Weak atomic access and manual memory barriers
129=============================================
130
131Compared to sequentially consistent atomic access, programming with
132weaker consistency models can be considerably more complicated.
133The only guarantees that you can rely upon in this case are:
134
135- atomic accesses will not cause data races (and hence undefined behavior);
136  ordinary accesses instead cause data races if they are concurrent with
137  other accesses of which at least one is a write.  In order to ensure this,
138  the compiler will not optimize accesses out of existence, create unsolicited
139  accesses, or perform other similar optimzations.
140
141- acquire operations will appear to happen, with respect to the other
142  components of the system, before all the LOAD or STORE operations
143  specified afterwards.
144
145- release operations will appear to happen, with respect to the other
146  components of the system, after all the LOAD or STORE operations
147  specified before.
148
149- release operations will *synchronize with* acquire operations;
150  see :ref:`acqrel` for a detailed explanation.
151
152When using this model, variables are accessed with:
153
154- ``qatomic_read()`` and ``qatomic_set()``; these prevent the compiler from
155  optimizing accesses out of existence and creating unsolicited
156  accesses, but do not otherwise impose any ordering on loads and
157  stores: both the compiler and the processor are free to reorder
158  them.
159
160- ``qatomic_load_acquire()``, which guarantees the LOAD to appear to
161  happen, with respect to the other components of the system,
162  before all the LOAD or STORE operations specified afterwards.
163  Operations coming before ``qatomic_load_acquire()`` can still be
164  reordered after it.
165
166- ``qatomic_store_release()``, which guarantees the STORE to appear to
167  happen, with respect to the other components of the system,
168  after all the LOAD or STORE operations specified before.
169  Operations coming after ``qatomic_store_release()`` can still be
170  reordered before it.
171
172Restrictions to the ordering of accesses can also be specified
173using the memory barrier macros: ``smp_rmb()``, ``smp_wmb()``, ``smp_mb()``,
174``smp_mb_acquire()``, ``smp_mb_release()``, ``smp_read_barrier_depends()``.
175
176Memory barriers control the order of references to shared memory.
177They come in six kinds:
178
179- ``smp_rmb()`` guarantees that all the LOAD operations specified before
180  the barrier will appear to happen before all the LOAD operations
181  specified after the barrier with respect to the other components of
182  the system.
183
184  In other words, ``smp_rmb()`` puts a partial ordering on loads, but is not
185  required to have any effect on stores.
186
187- ``smp_wmb()`` guarantees that all the STORE operations specified before
188  the barrier will appear to happen before all the STORE operations
189  specified after the barrier with respect to the other components of
190  the system.
191
192  In other words, ``smp_wmb()`` puts a partial ordering on stores, but is not
193  required to have any effect on loads.
194
195- ``smp_mb_acquire()`` guarantees that all the LOAD operations specified before
196  the barrier will appear to happen before all the LOAD or STORE operations
197  specified after the barrier with respect to the other components of
198  the system.
199
200- ``smp_mb_release()`` guarantees that all the STORE operations specified *after*
201  the barrier will appear to happen after all the LOAD or STORE operations
202  specified *before* the barrier with respect to the other components of
203  the system.
204
205- ``smp_mb()`` guarantees that all the LOAD and STORE operations specified
206  before the barrier will appear to happen before all the LOAD and
207  STORE operations specified after the barrier with respect to the other
208  components of the system.
209
210  ``smp_mb()`` puts a partial ordering on both loads and stores.  It is
211  stronger than both a read and a write memory barrier; it implies both
212  ``smp_mb_acquire()`` and ``smp_mb_release()``, but it also prevents STOREs
213  coming before the barrier from overtaking LOADs coming after the
214  barrier and vice versa.
215
216- ``smp_read_barrier_depends()`` is a weaker kind of read barrier.  On
217  most processors, whenever two loads are performed such that the
218  second depends on the result of the first (e.g., the first load
219  retrieves the address to which the second load will be directed),
220  the processor will guarantee that the first LOAD will appear to happen
221  before the second with respect to the other components of the system.
222  However, this is not always true---for example, it was not true on
223  Alpha processors.  Whenever this kind of access happens to shared
224  memory (that is not protected by a lock), a read barrier is needed,
225  and ``smp_read_barrier_depends()`` can be used instead of ``smp_rmb()``.
226
227  Note that the first load really has to have a _data_ dependency and not
228  a control dependency.  If the address for the second load is dependent
229  on the first load, but the dependency is through a conditional rather
230  than actually loading the address itself, then it's a _control_
231  dependency and a full read barrier or better is required.
232
233
234Memory barriers and ``qatomic_load_acquire``/``qatomic_store_release`` are
235mostly used when a data structure has one thread that is always a writer
236and one thread that is always a reader:
237
238    +----------------------------------+----------------------------------+
239    | thread 1                         | thread 2                         |
240    +==================================+==================================+
241    | ::                               | ::                               |
242    |                                  |                                  |
243    |   qatomic_store_release(&a, x);  |   y = qatomic_load_acquire(&b);  |
244    |   qatomic_store_release(&b, y);  |   x = qatomic_load_acquire(&a);  |
245    +----------------------------------+----------------------------------+
246
247In this case, correctness is easy to check for using the "pairing"
248trick that is explained below.
249
250Sometimes, a thread is accessing many variables that are otherwise
251unrelated to each other (for example because, apart from the current
252thread, exactly one other thread will read or write each of these
253variables).  In this case, it is possible to "hoist" the barriers
254outside a loop.  For example:
255
256    +------------------------------------------+----------------------------------+
257    | before                                   | after                            |
258    +==========================================+==================================+
259    | ::                                       | ::                               |
260    |                                          |                                  |
261    |   n = 0;                                 |   n = 0;                         |
262    |   for (i = 0; i < 10; i++)               |   for (i = 0; i < 10; i++)       |
263    |     n += qatomic_load_acquire(&a[i]);    |     n += qatomic_read(&a[i]);    |
264    |                                          |   smp_mb_acquire();              |
265    +------------------------------------------+----------------------------------+
266    | ::                                       | ::                               |
267    |                                          |                                  |
268    |                                          |   smp_mb_release();              |
269    |   for (i = 0; i < 10; i++)               |   for (i = 0; i < 10; i++)       |
270    |     qatomic_store_release(&a[i], false); |     qatomic_set(&a[i], false);   |
271    +------------------------------------------+----------------------------------+
272
273Splitting a loop can also be useful to reduce the number of barriers:
274
275    +------------------------------------------+----------------------------------+
276    | before                                   | after                            |
277    +==========================================+==================================+
278    | ::                                       | ::                               |
279    |                                          |                                  |
280    |   n = 0;                                 |     smp_mb_release();            |
281    |   for (i = 0; i < 10; i++) {             |     for (i = 0; i < 10; i++)     |
282    |     qatomic_store_release(&a[i], false); |       qatomic_set(&a[i], false); |
283    |     smp_mb();                            |     smb_mb();                    |
284    |     n += qatomic_read(&b[i]);            |     n = 0;                       |
285    |   }                                      |     for (i = 0; i < 10; i++)     |
286    |                                          |       n += qatomic_read(&b[i]);  |
287    +------------------------------------------+----------------------------------+
288
289In this case, a ``smp_mb_release()`` is also replaced with a (possibly cheaper, and clearer
290as well) ``smp_wmb()``:
291
292    +------------------------------------------+----------------------------------+
293    | before                                   | after                            |
294    +==========================================+==================================+
295    | ::                                       | ::                               |
296    |                                          |                                  |
297    |                                          |     smp_mb_release();            |
298    |   for (i = 0; i < 10; i++) {             |     for (i = 0; i < 10; i++)     |
299    |     qatomic_store_release(&a[i], false); |       qatomic_set(&a[i], false); |
300    |     qatomic_store_release(&b[i], false); |     smb_wmb();                   |
301    |   }                                      |     for (i = 0; i < 10; i++)     |
302    |                                          |       qatomic_set(&b[i], false); |
303    +------------------------------------------+----------------------------------+
304
305
306.. _acqrel:
307
308Acquire/release pairing and the *synchronizes-with* relation
309------------------------------------------------------------
310
311Atomic operations other than ``qatomic_set()`` and ``qatomic_read()`` have
312either *acquire* or *release* semantics [#rmw]_.  This has two effects:
313
314.. [#rmw] Read-modify-write operations can have both---acquire applies to the
315          read part, and release to the write.
316
317- within a thread, they are ordered either before subsequent operations
318  (for acquire) or after previous operations (for release).
319
320- if a release operation in one thread *synchronizes with* an acquire operation
321  in another thread, the ordering constraints propagates from the first to the
322  second thread.  That is, everything before the release operation in the
323  first thread is guaranteed to *happen before* everything after the
324  acquire operation in the second thread.
325
326The concept of acquire and release semantics is not exclusive to atomic
327operations; almost all higher-level synchronization primitives also have
328acquire or release semantics.  For example:
329
330- ``pthread_mutex_lock`` has acquire semantics, ``pthread_mutex_unlock`` has
331  release semantics and synchronizes with a ``pthread_mutex_lock`` for the
332  same mutex.
333
334- ``pthread_cond_signal`` and ``pthread_cond_broadcast`` have release semantics;
335  ``pthread_cond_wait`` has both release semantics (synchronizing with
336  ``pthread_mutex_lock``) and acquire semantics (synchronizing with
337  ``pthread_mutex_unlock`` and signaling of the condition variable).
338
339- ``pthread_create`` has release semantics and synchronizes with the start
340  of the new thread; ``pthread_join`` has acquire semantics and synchronizes
341  with the exiting of the thread.
342
343- ``qemu_event_set`` has release semantics, ``qemu_event_wait`` has
344  acquire semantics.
345
346For example, in the following example there are no atomic accesses, but still
347thread 2 is relying on the *synchronizes-with* relation between ``pthread_exit``
348(release) and ``pthread_join`` (acquire):
349
350      +----------------------+-------------------------------+
351      | thread 1             | thread 2                      |
352      +======================+===============================+
353      | ::                   | ::                            |
354      |                      |                               |
355      |   *a = 1;            |                               |
356      |   pthread_exit(a);   |   pthread_join(thread1, &a);  |
357      |                      |   x = *a;                     |
358      +----------------------+-------------------------------+
359
360Synchronization between threads basically descends from this pairing of
361a release operation and an acquire operation.  Therefore, atomic operations
362other than ``qatomic_set()`` and ``qatomic_read()`` will almost always be
363paired with another operation of the opposite kind: an acquire operation
364will pair with a release operation and vice versa.  This rule of thumb is
365extremely useful; in the case of QEMU, however, note that the other
366operation may actually be in a driver that runs in the guest!
367
368``smp_read_barrier_depends()``, ``smp_rmb()``, ``smp_mb_acquire()``,
369``qatomic_load_acquire()`` and ``qatomic_rcu_read()`` all count
370as acquire operations.  ``smp_wmb()``, ``smp_mb_release()``,
371``qatomic_store_release()`` and ``qatomic_rcu_set()`` all count as release
372operations.  ``smp_mb()`` counts as both acquire and release, therefore
373it can pair with any other atomic operation.  Here is an example:
374
375      +----------------------+------------------------------+
376      | thread 1             | thread 2                     |
377      +======================+==============================+
378      | ::                   | ::                           |
379      |                      |                              |
380      |   qatomic_set(&a, 1);|                              |
381      |   smp_wmb();         |                              |
382      |   qatomic_set(&b, 2);|   x = qatomic_read(&b);      |
383      |                      |   smp_rmb();                 |
384      |                      |   y = qatomic_read(&a);      |
385      +----------------------+------------------------------+
386
387Note that a load-store pair only counts if the two operations access the
388same variable: that is, a store-release on a variable ``x`` *synchronizes
389with* a load-acquire on a variable ``x``, while a release barrier
390synchronizes with any acquire operation.  The following example shows
391correct synchronization:
392
393      +--------------------------------+--------------------------------+
394      | thread 1                       | thread 2                       |
395      +================================+================================+
396      | ::                             | ::                             |
397      |                                |                                |
398      |   qatomic_set(&a, 1);          |                                |
399      |   qatomic_store_release(&b, 2);|   x = qatomic_load_acquire(&b);|
400      |                                |   y = qatomic_read(&a);        |
401      +--------------------------------+--------------------------------+
402
403Acquire and release semantics of higher-level primitives can also be
404relied upon for the purpose of establishing the *synchronizes with*
405relation.
406
407Note that the "writing" thread is accessing the variables in the
408opposite order as the "reading" thread.  This is expected: stores
409before a release operation will normally match the loads after
410the acquire operation, and vice versa.  In fact, this happened already
411in the ``pthread_exit``/``pthread_join`` example above.
412
413Finally, this more complex example has more than two accesses and data
414dependency barriers.  It also does not use atomic accesses whenever there
415cannot be a data race:
416
417      +----------------------+------------------------------+
418      | thread 1             | thread 2                     |
419      +======================+==============================+
420      | ::                   | ::                           |
421      |                      |                              |
422      |   b[2] = 1;          |                              |
423      |   smp_wmb();         |                              |
424      |   x->i = 2;          |                              |
425      |   smp_wmb();         |                              |
426      |   qatomic_set(&a, x);|  x = qatomic_read(&a);       |
427      |                      |  smp_read_barrier_depends(); |
428      |                      |  y = x->i;                   |
429      |                      |  smp_read_barrier_depends(); |
430      |                      |  z = b[y];                   |
431      +----------------------+------------------------------+
432
433Comparison with Linux kernel primitives
434=======================================
435
436Here is a list of differences between Linux kernel atomic operations
437and memory barriers, and the equivalents in QEMU:
438
439- atomic operations in Linux are always on a 32-bit int type and
440  use a boxed ``atomic_t`` type; atomic operations in QEMU are polymorphic
441  and use normal C types.
442
443- Originally, ``atomic_read`` and ``atomic_set`` in Linux gave no guarantee
444  at all. Linux 4.1 updated them to implement volatile
445  semantics via ``ACCESS_ONCE`` (or the more recent ``READ``/``WRITE_ONCE``).
446
447  QEMU's ``qatomic_read`` and ``qatomic_set`` implement C11 atomic relaxed
448  semantics if the compiler supports it, and volatile semantics otherwise.
449  Both semantics prevent the compiler from doing certain transformations;
450  the difference is that atomic accesses are guaranteed to be atomic,
451  while volatile accesses aren't. Thus, in the volatile case we just cross
452  our fingers hoping that the compiler will generate atomic accesses,
453  since we assume the variables passed are machine-word sized and
454  properly aligned.
455
456  No barriers are implied by ``qatomic_read`` and ``qatomic_set`` in either
457  Linux or QEMU.
458
459- atomic read-modify-write operations in Linux are of three kinds:
460
461         ===================== =========================================
462         ``atomic_OP``         returns void
463         ``atomic_OP_return``  returns new value of the variable
464         ``atomic_fetch_OP``   returns the old value of the variable
465         ``atomic_cmpxchg``    returns the old value of the variable
466         ===================== =========================================
467
468  In QEMU, the second kind is named ``atomic_OP_fetch``.
469
470- different atomic read-modify-write operations in Linux imply
471  a different set of memory barriers; in QEMU, all of them enforce
472  sequential consistency.
473
474- in QEMU, ``qatomic_read()`` and ``qatomic_set()`` do not participate in
475  the total ordering enforced by sequentially-consistent operations.
476  This is because QEMU uses the C11 memory model.  The following example
477  is correct in Linux but not in QEMU:
478
479      +----------------------------------+--------------------------------+
480      | Linux (correct)                  | QEMU (incorrect)               |
481      +==================================+================================+
482      | ::                               | ::                             |
483      |                                  |                                |
484      |   a = atomic_fetch_add(&x, 2);   |   a = qatomic_fetch_add(&x, 2);|
485      |   b = READ_ONCE(&y);             |   b = qatomic_read(&y);        |
486      +----------------------------------+--------------------------------+
487
488  because the read of ``y`` can be moved (by either the processor or the
489  compiler) before the write of ``x``.
490
491  Fixing this requires an ``smp_mb()`` memory barrier between the write
492  of ``x`` and the read of ``y``.  In the common case where only one thread
493  writes ``x``, it is also possible to write it like this:
494
495      +--------------------------------+
496      | QEMU (correct)                 |
497      +================================+
498      | ::                             |
499      |                                |
500      |   a = qatomic_read(&x);        |
501      |   qatomic_set(&x, a + 2);      |
502      |   smp_mb();                    |
503      |   b = qatomic_read(&y);        |
504      +--------------------------------+
505
506Sources
507=======
508
509- ``Documentation/memory-barriers.txt`` from the Linux kernel
510