1 CPUs perform independent memory operations effectively in random order.
2 but this can be a problem for CPU-CPU interaction (including interactions
3 between QEMU and the guest). Multi-threaded programs use various tools
4 to instruct the compiler and the CPU to restrict the order to something
5 that is consistent with the expectations of the programmer.
7 The most basic tool is locking. Mutexes, condition variables and
8 semaphores are used in QEMU, and should be the default approach to
9 synchronization. Anything else is considerably harder, but it's
10 also justified more often than one would like. The two tools that
11 are provided by qemu/atomic.h are memory barriers and atomic operations.
13 Macros defined by qemu/atomic.h fall in three camps:
15 - compiler barriers: barrier();
17 - weak atomic access and manual memory barriers: atomic_read(),
18 atomic_set(), smp_rmb(), smp_wmb(), smp_mb(), smp_mb_acquire(),
19 smp_mb_release(), smp_read_barrier_depends();
21 - sequentially consistent atomic access: everything else.
24 COMPILER MEMORY BARRIER
25 =======================
27 barrier() prevents the compiler from moving the memory accesses either
28 side of it to the other side. The compiler barrier has no direct effect
29 on the CPU, which may then reorder things however it wishes.
31 barrier() is mostly used within qemu/atomic.h itself. On some
32 architectures, CPU guarantees are strong enough that blocking compiler
33 optimizations already ensures the correct order of execution. In this
34 case, qemu/atomic.h will reduce stronger memory barriers to simple
37 Still, barrier() can be useful when writing code that can be interrupted
41 SEQUENTIALLY CONSISTENT ATOMIC ACCESS
42 =====================================
44 Most of the operations in the qemu/atomic.h header ensure *sequential
45 consistency*, where "the result of any execution is the same as if the
46 operations of all the processors were executed in some sequential order,
47 and the operations of each individual processor appear in this sequence
48 in the order specified by its program".
50 qemu/atomic.h provides the following set of atomic read-modify-write
55 void atomic_add(ptr, val)
56 void atomic_sub(ptr, val)
57 void atomic_and(ptr, val)
58 void atomic_or(ptr, val)
60 typeof(*ptr) atomic_fetch_inc(ptr)
61 typeof(*ptr) atomic_fetch_dec(ptr)
62 typeof(*ptr) atomic_fetch_add(ptr, val)
63 typeof(*ptr) atomic_fetch_sub(ptr, val)
64 typeof(*ptr) atomic_fetch_and(ptr, val)
65 typeof(*ptr) atomic_fetch_or(ptr, val)
66 typeof(*ptr) atomic_fetch_xor(ptr, val)
67 typeof(*ptr) atomic_fetch_inc_nonzero(ptr)
68 typeof(*ptr) atomic_xchg(ptr, val)
69 typeof(*ptr) atomic_cmpxchg(ptr, old, new)
71 all of which return the old value of *ptr. These operations are
72 polymorphic; they operate on any type that is as wide as a pointer.
74 Similar operations return the new value of *ptr:
76 typeof(*ptr) atomic_inc_fetch(ptr)
77 typeof(*ptr) atomic_dec_fetch(ptr)
78 typeof(*ptr) atomic_add_fetch(ptr, val)
79 typeof(*ptr) atomic_sub_fetch(ptr, val)
80 typeof(*ptr) atomic_and_fetch(ptr, val)
81 typeof(*ptr) atomic_or_fetch(ptr, val)
82 typeof(*ptr) atomic_xor_fetch(ptr, val)
84 Sequentially consistent loads and stores can be done using:
86 atomic_fetch_add(ptr, 0) for loads
87 atomic_xchg(ptr, val) for stores
89 However, they are quite expensive on some platforms, notably POWER and
90 ARM. Therefore, qemu/atomic.h provides two primitives with slightly
93 typeof(*ptr) atomic_mb_read(ptr)
94 void atomic_mb_set(ptr, val)
96 The semantics of these primitives map to Java volatile variables,
97 and are strongly related to memory barriers as used in the Linux
100 As long as you use atomic_mb_read and atomic_mb_set, accesses cannot
101 be reordered with each other, and it is also not possible to reorder
102 "normal" accesses around them.
104 However, and this is the important difference between
105 atomic_mb_read/atomic_mb_set and sequential consistency, it is important
106 for both threads to access the same volatile variable. It is not the
107 case that everything visible to thread A when it writes volatile field f
108 becomes visible to thread B after it reads volatile field g. The store
109 and load have to "match" (i.e., be performed on the same volatile
110 field) to achieve the right semantics.
113 These operations operate on any type that is as wide as an int or smaller.
116 WEAK ATOMIC ACCESS AND MANUAL MEMORY BARRIERS
117 =============================================
119 Compared to sequentially consistent atomic access, programming with
120 weaker consistency models can be considerably more complicated.
121 In general, if the algorithm you are writing includes both writes
122 and reads on the same side, it is generally simpler to use sequentially
123 consistent primitives.
125 When using this model, variables are accessed with:
127 - atomic_read() and atomic_set(); these prevent the compiler from
128 optimizing accesses out of existence and creating unsolicited
129 accesses, but do not otherwise impose any ordering on loads and
130 stores: both the compiler and the processor are free to reorder
133 - atomic_load_acquire(), which guarantees the LOAD to appear to
134 happen, with respect to the other components of the system,
135 before all the LOAD or STORE operations specified afterwards.
136 Operations coming before atomic_load_acquire() can still be
139 - atomic_store_release(), which guarantees the STORE to appear to
140 happen, with respect to the other components of the system,
141 after all the LOAD or STORE operations specified afterwards.
142 Operations coming after atomic_store_release() can still be
145 Restrictions to the ordering of accesses can also be specified
146 using the memory barrier macros: smp_rmb(), smp_wmb(), smp_mb(),
147 smp_mb_acquire(), smp_mb_release(), smp_read_barrier_depends().
149 Memory barriers control the order of references to shared memory.
150 They come in six kinds:
152 - smp_rmb() guarantees that all the LOAD operations specified before
153 the barrier will appear to happen before all the LOAD operations
154 specified after the barrier with respect to the other components of
157 In other words, smp_rmb() puts a partial ordering on loads, but is not
158 required to have any effect on stores.
160 - smp_wmb() guarantees that all the STORE operations specified before
161 the barrier will appear to happen before all the STORE operations
162 specified after the barrier with respect to the other components of
165 In other words, smp_wmb() puts a partial ordering on stores, but is not
166 required to have any effect on loads.
168 - smp_mb_acquire() guarantees that all the LOAD operations specified before
169 the barrier will appear to happen before all the LOAD or STORE operations
170 specified after the barrier with respect to the other components of
173 - smp_mb_release() guarantees that all the STORE operations specified *after*
174 the barrier will appear to happen after all the LOAD or STORE operations
175 specified *before* the barrier with respect to the other components of
178 - smp_mb() guarantees that all the LOAD and STORE operations specified
179 before the barrier will appear to happen before all the LOAD and
180 STORE operations specified after the barrier with respect to the other
181 components of the system.
183 smp_mb() puts a partial ordering on both loads and stores. It is
184 stronger than both a read and a write memory barrier; it implies both
185 smp_mb_acquire() and smp_mb_release(), but it also prevents STOREs
186 coming before the barrier from overtaking LOADs coming after the
187 barrier and vice versa.
189 - smp_read_barrier_depends() is a weaker kind of read barrier. On
190 most processors, whenever two loads are performed such that the
191 second depends on the result of the first (e.g., the first load
192 retrieves the address to which the second load will be directed),
193 the processor will guarantee that the first LOAD will appear to happen
194 before the second with respect to the other components of the system.
195 However, this is not always true---for example, it was not true on
196 Alpha processors. Whenever this kind of access happens to shared
197 memory (that is not protected by a lock), a read barrier is needed,
198 and smp_read_barrier_depends() can be used instead of smp_rmb().
200 Note that the first load really has to have a _data_ dependency and not
201 a control dependency. If the address for the second load is dependent
202 on the first load, but the dependency is through a conditional rather
203 than actually loading the address itself, then it's a _control_
204 dependency and a full read barrier or better is required.
207 This is the set of barriers that is required *between* two atomic_read()
208 and atomic_set() operations to achieve sequential consistency:
211 |-----------------------------------------------|
212 1st operation | (after last) | atomic_read | atomic_set |
213 ---------------+----------------+-------------+----------------|
214 (before first) | | none | smp_mb_release |
215 ---------------+----------------+-------------+----------------|
216 atomic_read | smp_mb_acquire | smp_rmb | ** |
217 ---------------+----------------+-------------+----------------|
218 atomic_set | none | smp_mb()*** | smp_wmb() |
219 ---------------+----------------+-------------+----------------|
221 * Or smp_read_barrier_depends().
223 ** This requires a load-store barrier. This is achieved by
224 either smp_mb_acquire() or smp_mb_release().
226 *** This requires a store-load barrier. On most machines, the only
227 way to achieve this is a full barrier.
230 You can see that the two possible definitions of atomic_mb_read()
231 and atomic_mb_set() are the following:
233 1) atomic_mb_read(p) = atomic_read(p); smp_mb_acquire()
234 atomic_mb_set(p, v) = smp_mb_release(); atomic_set(p, v); smp_mb()
236 2) atomic_mb_read(p) = smp_mb() atomic_read(p); smp_mb_acquire()
237 atomic_mb_set(p, v) = smp_mb_release(); atomic_set(p, v);
239 Usually the former is used, because smp_mb() is expensive and a program
240 normally has more reads than writes. Therefore it makes more sense to
241 make atomic_mb_set() the more expensive operation.
243 There are two common cases in which atomic_mb_read and atomic_mb_set
244 generate too many memory barriers, and thus it can be useful to manually
245 place barriers, or use atomic_load_acquire/atomic_store_release instead:
247 - when a data structure has one thread that is always a writer
248 and one thread that is always a reader, manual placement of
249 memory barriers makes the write side faster. Furthermore,
250 correctness is easy to check for in this case using the "pairing"
251 trick that is explained below:
254 ------------------------- ------------------------
256 atomic_mb_set(&a, x) atomic_store_release(&a, x)
257 atomic_mb_set(&b, y) atomic_store_release(&b, y)
261 ------------------------- ------------------------
262 y = atomic_mb_read(&b) y = atomic_load_acquire(&b)
263 x = atomic_mb_read(&a) x = atomic_load_acquire(&a)
266 Note that the barrier between the stores in thread 1, and between
267 the loads in thread 2, has been optimized here to a write or a
268 read memory barrier respectively. On some architectures, notably
269 ARMv7, smp_mb_acquire and smp_mb_release are just as expensive as
270 smp_mb, but smp_rmb and/or smp_wmb are more efficient.
272 - sometimes, a thread is accessing many variables that are otherwise
273 unrelated to each other (for example because, apart from the current
274 thread, exactly one other thread will read or write each of these
275 variables). In this case, it is possible to "hoist" the implicit
276 barriers provided by atomic_mb_read() and atomic_mb_set() outside
277 a loop. For example, the above definition atomic_mb_read() gives
278 the following transformation:
281 for (i = 0; i < 10; i++) => for (i = 0; i < 10; i++)
282 n += atomic_mb_read(&a[i]); n += atomic_read(&a[i]);
285 Similarly, atomic_mb_set() can be transformed as follows:
288 for (i = 0; i < 10; i++) => for (i = 0; i < 10; i++)
289 atomic_mb_set(&a[i], false); atomic_set(&a[i], false);
293 The other thread can still use atomic_mb_read()/atomic_mb_set().
295 The two tricks can be combined. In this case, splitting a loop in
296 two lets you hoist the barriers out of the loops _and_ eliminate the
300 for (i = 0; i < 10; i++) { => for (i = 0; i < 10; i++)
301 atomic_mb_set(&a[i], false); atomic_set(&a[i], false);
302 atomic_mb_set(&b[i], false); smb_wmb();
303 } for (i = 0; i < 10; i++)
304 atomic_set(&a[i], false);
308 Memory barrier pairing
309 ----------------------
311 A useful rule of thumb is that memory barriers should always, or almost
312 always, be paired with another barrier. In the case of QEMU, however,
313 note that the other barrier may actually be in a driver that runs in
316 For the purposes of pairing, smp_read_barrier_depends() and smp_rmb()
317 both count as read barriers. A read barrier shall pair with a write
318 barrier or a full barrier; a write barrier shall pair with a read
319 barrier or a full barrier. A full barrier can pair with anything.
323 =============== ===============
330 Note that the "writing" thread is accessing the variables in the
331 opposite order as the "reading" thread. This is expected: stores
332 before the write barrier will normally match the loads after the
333 read barrier, and vice versa. The same is true for more than 2
334 access and for data dependency barriers:
337 =============== ===============
343 smp_read_barrier_depends();
345 smp_read_barrier_depends();
348 smp_wmb() also pairs with atomic_mb_read() and smp_mb_acquire().
349 and smp_rmb() also pairs with atomic_mb_set() and smp_mb_release().
352 COMPARISON WITH LINUX KERNEL MEMORY BARRIERS
353 ============================================
355 Here is a list of differences between Linux kernel atomic operations
356 and memory barriers, and the equivalents in QEMU:
358 - atomic operations in Linux are always on a 32-bit int type and
359 use a boxed atomic_t type; atomic operations in QEMU are polymorphic
360 and use normal C types.
362 - Originally, atomic_read and atomic_set in Linux gave no guarantee
363 at all. Linux 4.1 updated them to implement volatile
364 semantics via ACCESS_ONCE (or the more recent READ/WRITE_ONCE).
366 QEMU's atomic_read/set implement, if the compiler supports it, C11
367 atomic relaxed semantics, and volatile semantics otherwise.
368 Both semantics prevent the compiler from doing certain transformations;
369 the difference is that atomic accesses are guaranteed to be atomic,
370 while volatile accesses aren't. Thus, in the volatile case we just cross
371 our fingers hoping that the compiler will generate atomic accesses,
372 since we assume the variables passed are machine-word sized and
374 No barriers are implied by atomic_read/set in either Linux or QEMU.
376 - atomic read-modify-write operations in Linux are of three kinds:
378 atomic_OP returns void
379 atomic_OP_return returns new value of the variable
380 atomic_fetch_OP returns the old value of the variable
381 atomic_cmpxchg returns the old value of the variable
383 In QEMU, the second kind does not exist. Currently Linux has
384 atomic_fetch_or only. QEMU provides and, or, inc, dec, add, sub.
386 - different atomic read-modify-write operations in Linux imply
387 a different set of memory barriers; in QEMU, all of them enforce
388 sequential consistency, which means they imply full memory barriers
389 before and after the operation.
391 - Linux does not have an equivalent of atomic_mb_set(). In particular,
392 note that smp_store_mb() is a little weaker than atomic_mb_set().
393 atomic_mb_read() compiles to the same instructions as Linux's
394 smp_load_acquire(), but this should be treated as an implementation
400 * Documentation/memory-barriers.txt from the Linux kernel
402 * "The JSR-133 Cookbook for Compiler Writers", available at
403 http://g.oswego.edu/dl/jmm/cookbook.html