blob: 556f951f86260211879295309bff88302c66dd98 [file] [log] [blame]
David Howells108b42b2006-03-31 16:00:29 +01001 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
4
5By: David Howells <dhowells@redhat.com>
David Howells90fddab2010-03-24 09:43:00 +00006 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
David Howells108b42b2006-03-31 16:00:29 +01007
8Contents:
9
10 (*) Abstract memory access model.
11
12 - Device operations.
13 - Guarantees.
14
15 (*) What are memory barriers?
16
17 - Varieties of memory barrier.
18 - What may not be assumed about memory barriers?
19 - Data dependency barriers.
20 - Control dependencies.
21 - SMP barrier pairing.
22 - Examples of memory barrier sequences.
David Howells670bd952006-06-10 09:54:12 -070023 - Read memory barriers vs load speculation.
Paul E. McKenney241e6662011-02-10 16:54:50 -080024 - Transitivity
David Howells108b42b2006-03-31 16:00:29 +010025
26 (*) Explicit kernel barriers.
27
28 - Compiler barrier.
Jarek Poplawski81fc6322007-05-23 13:58:20 -070029 - CPU memory barriers.
David Howells108b42b2006-03-31 16:00:29 +010030 - MMIO write barrier.
31
32 (*) Implicit kernel memory barriers.
33
34 - Locking functions.
35 - Interrupt disabling functions.
David Howells50fa6102009-04-28 15:01:38 +010036 - Sleep and wake-up functions.
David Howells108b42b2006-03-31 16:00:29 +010037 - Miscellaneous functions.
38
39 (*) Inter-CPU locking barrier effects.
40
41 - Locks vs memory accesses.
42 - Locks vs I/O accesses.
43
44 (*) Where are memory barriers needed?
45
46 - Interprocessor interaction.
47 - Atomic operations.
48 - Accessing devices.
49 - Interrupts.
50
51 (*) Kernel I/O barrier effects.
52
53 (*) Assumed minimum execution ordering model.
54
55 (*) The effects of the cpu cache.
56
57 - Cache coherency.
58 - Cache coherency vs DMA.
59 - Cache coherency vs MMIO.
60
61 (*) The things CPUs get up to.
62
63 - And then there's the Alpha.
64
David Howells90fddab2010-03-24 09:43:00 +000065 (*) Example uses.
66
67 - Circular buffers.
68
David Howells108b42b2006-03-31 16:00:29 +010069 (*) References.
70
71
72============================
73ABSTRACT MEMORY ACCESS MODEL
74============================
75
76Consider the following abstract model of the system:
77
78 : :
79 : :
80 : :
81 +-------+ : +--------+ : +-------+
82 | | : | | : | |
83 | | : | | : | |
84 | CPU 1 |<----->| Memory |<----->| CPU 2 |
85 | | : | | : | |
86 | | : | | : | |
87 +-------+ : +--------+ : +-------+
88 ^ : ^ : ^
89 | : | : |
90 | : | : |
91 | : v : |
92 | : +--------+ : |
93 | : | | : |
94 | : | | : |
95 +---------->| Device |<----------+
96 : | | :
97 : | | :
98 : +--------+ :
99 : :
100
101Each CPU executes a program that generates memory access operations. In the
102abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
103perform the memory operations in any order it likes, provided program causality
104appears to be maintained. Similarly, the compiler may also arrange the
105instructions it emits in any order it likes, provided it doesn't affect the
106apparent operation of the program.
107
108So in the above diagram, the effects of the memory operations performed by a
109CPU are perceived by the rest of the system as the operations cross the
110interface between the CPU and rest of the system (the dotted lines).
111
112
113For example, consider the following sequence of events:
114
115 CPU 1 CPU 2
116 =============== ===============
117 { A == 1; B == 2 }
118 A = 3; x = A;
119 B = 4; y = B;
120
121The set of accesses as seen by the memory system in the middle can be arranged
122in 24 different combinations:
123
124 STORE A=3, STORE B=4, x=LOAD A->3, y=LOAD B->4
125 STORE A=3, STORE B=4, y=LOAD B->4, x=LOAD A->3
126 STORE A=3, x=LOAD A->3, STORE B=4, y=LOAD B->4
127 STORE A=3, x=LOAD A->3, y=LOAD B->2, STORE B=4
128 STORE A=3, y=LOAD B->2, STORE B=4, x=LOAD A->3
129 STORE A=3, y=LOAD B->2, x=LOAD A->3, STORE B=4
130 STORE B=4, STORE A=3, x=LOAD A->3, y=LOAD B->4
131 STORE B=4, ...
132 ...
133
134and can thus result in four different combinations of values:
135
136 x == 1, y == 2
137 x == 1, y == 4
138 x == 3, y == 2
139 x == 3, y == 4
140
141
142Furthermore, the stores committed by a CPU to the memory system may not be
143perceived by the loads made by another CPU in the same order as the stores were
144committed.
145
146
147As a further example, consider this sequence of events:
148
149 CPU 1 CPU 2
150 =============== ===============
151 { A == 1, B == 2, C = 3, P == &A, Q == &C }
152 B = 4; Q = P;
153 P = &B D = *Q;
154
155There is an obvious data dependency here, as the value loaded into D depends on
156the address retrieved from P by CPU 2. At the end of the sequence, any of the
157following results are possible:
158
159 (Q == &A) and (D == 1)
160 (Q == &B) and (D == 2)
161 (Q == &B) and (D == 4)
162
163Note that CPU 2 will never try and load C into D because the CPU will load P
164into Q before issuing the load of *Q.
165
166
167DEVICE OPERATIONS
168-----------------
169
170Some devices present their control interfaces as collections of memory
171locations, but the order in which the control registers are accessed is very
172important. For instance, imagine an ethernet card with a set of internal
173registers that are accessed through an address port register (A) and a data
174port register (D). To read internal register 5, the following code might then
175be used:
176
177 *A = 5;
178 x = *D;
179
180but this might show up as either of the following two sequences:
181
182 STORE *A = 5, x = LOAD *D
183 x = LOAD *D, STORE *A = 5
184
185the second of which will almost certainly result in a malfunction, since it set
186the address _after_ attempting to read the register.
187
188
189GUARANTEES
190----------
191
192There are some minimal guarantees that may be expected of a CPU:
193
194 (*) On any given CPU, dependent memory accesses will be issued in order, with
195 respect to itself. This means that for:
196
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800197 ACCESS_ONCE(Q) = P; smp_read_barrier_depends(); D = ACCESS_ONCE(*Q);
David Howells108b42b2006-03-31 16:00:29 +0100198
199 the CPU will issue the following memory operations:
200
201 Q = LOAD P, D = LOAD *Q
202
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800203 and always in that order. On most systems, smp_read_barrier_depends()
204 does nothing, but it is required for DEC Alpha. The ACCESS_ONCE()
205 is required to prevent compiler mischief. Please note that you
206 should normally use something like rcu_dereference() instead of
207 open-coding smp_read_barrier_depends().
David Howells108b42b2006-03-31 16:00:29 +0100208
209 (*) Overlapping loads and stores within a particular CPU will appear to be
210 ordered within that CPU. This means that for:
211
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800212 a = ACCESS_ONCE(*X); ACCESS_ONCE(*X) = b;
David Howells108b42b2006-03-31 16:00:29 +0100213
214 the CPU will only issue the following sequence of memory operations:
215
216 a = LOAD *X, STORE *X = b
217
218 And for:
219
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800220 ACCESS_ONCE(*X) = c; d = ACCESS_ONCE(*X);
David Howells108b42b2006-03-31 16:00:29 +0100221
222 the CPU will only issue:
223
224 STORE *X = c, d = LOAD *X
225
Matt LaPlantefa00e7e2006-11-30 04:55:36 +0100226 (Loads and stores overlap if they are targeted at overlapping pieces of
David Howells108b42b2006-03-31 16:00:29 +0100227 memory).
228
229And there are a number of things that _must_ or _must_not_ be assumed:
230
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800231 (*) It _must_not_ be assumed that the compiler will do what you want with
232 memory references that are not protected by ACCESS_ONCE(). Without
233 ACCESS_ONCE(), the compiler is within its rights to do all sorts
Paul E. McKenney692118d2013-12-11 13:59:07 -0800234 of "creative" transformations, which are covered in the Compiler
235 Barrier section.
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800236
David Howells108b42b2006-03-31 16:00:29 +0100237 (*) It _must_not_ be assumed that independent loads and stores will be issued
238 in the order given. This means that for:
239
240 X = *A; Y = *B; *D = Z;
241
242 we may get any of the following sequences:
243
244 X = LOAD *A, Y = LOAD *B, STORE *D = Z
245 X = LOAD *A, STORE *D = Z, Y = LOAD *B
246 Y = LOAD *B, X = LOAD *A, STORE *D = Z
247 Y = LOAD *B, STORE *D = Z, X = LOAD *A
248 STORE *D = Z, X = LOAD *A, Y = LOAD *B
249 STORE *D = Z, Y = LOAD *B, X = LOAD *A
250
251 (*) It _must_ be assumed that overlapping memory accesses may be merged or
252 discarded. This means that for:
253
254 X = *A; Y = *(A + 4);
255
256 we may get any one of the following sequences:
257
258 X = LOAD *A; Y = LOAD *(A + 4);
259 Y = LOAD *(A + 4); X = LOAD *A;
260 {X, Y} = LOAD {*A, *(A + 4) };
261
262 And for:
263
Paul E. McKenneyf191eec2012-10-03 10:28:30 -0700264 *A = X; *(A + 4) = Y;
David Howells108b42b2006-03-31 16:00:29 +0100265
Paul E. McKenneyf191eec2012-10-03 10:28:30 -0700266 we may get any of:
David Howells108b42b2006-03-31 16:00:29 +0100267
Paul E. McKenneyf191eec2012-10-03 10:28:30 -0700268 STORE *A = X; STORE *(A + 4) = Y;
269 STORE *(A + 4) = Y; STORE *A = X;
270 STORE {*A, *(A + 4) } = {X, Y};
David Howells108b42b2006-03-31 16:00:29 +0100271
272
273=========================
274WHAT ARE MEMORY BARRIERS?
275=========================
276
277As can be seen above, independent memory operations are effectively performed
278in random order, but this can be a problem for CPU-CPU interaction and for I/O.
279What is required is some way of intervening to instruct the compiler and the
280CPU to restrict the order.
281
282Memory barriers are such interventions. They impose a perceived partial
David Howells2b948952006-06-25 05:48:49 -0700283ordering over the memory operations on either side of the barrier.
284
285Such enforcement is important because the CPUs and other devices in a system
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700286can use a variety of tricks to improve performance, including reordering,
David Howells2b948952006-06-25 05:48:49 -0700287deferral and combination of memory operations; speculative loads; speculative
288branch prediction and various types of caching. Memory barriers are used to
289override or suppress these tricks, allowing the code to sanely control the
290interaction of multiple CPUs and/or devices.
David Howells108b42b2006-03-31 16:00:29 +0100291
292
293VARIETIES OF MEMORY BARRIER
294---------------------------
295
296Memory barriers come in four basic varieties:
297
298 (1) Write (or store) memory barriers.
299
300 A write memory barrier gives a guarantee that all the STORE operations
301 specified before the barrier will appear to happen before all the STORE
302 operations specified after the barrier with respect to the other
303 components of the system.
304
305 A write barrier is a partial ordering on stores only; it is not required
306 to have any effect on loads.
307
David Howells6bc39272006-06-25 05:49:22 -0700308 A CPU can be viewed as committing a sequence of store operations to the
David Howells108b42b2006-03-31 16:00:29 +0100309 memory system as time progresses. All stores before a write barrier will
310 occur in the sequence _before_ all the stores after the write barrier.
311
312 [!] Note that write barriers should normally be paired with read or data
313 dependency barriers; see the "SMP barrier pairing" subsection.
314
315
316 (2) Data dependency barriers.
317
318 A data dependency barrier is a weaker form of read barrier. In the case
319 where two loads are performed such that the second depends on the result
320 of the first (eg: the first load retrieves the address to which the second
321 load will be directed), a data dependency barrier would be required to
322 make sure that the target of the second load is updated before the address
323 obtained by the first load is accessed.
324
325 A data dependency barrier is a partial ordering on interdependent loads
326 only; it is not required to have any effect on stores, independent loads
327 or overlapping loads.
328
329 As mentioned in (1), the other CPUs in the system can be viewed as
330 committing sequences of stores to the memory system that the CPU being
331 considered can then perceive. A data dependency barrier issued by the CPU
332 under consideration guarantees that for any load preceding it, if that
333 load touches one of a sequence of stores from another CPU, then by the
334 time the barrier completes, the effects of all the stores prior to that
335 touched by the load will be perceptible to any loads issued after the data
336 dependency barrier.
337
338 See the "Examples of memory barrier sequences" subsection for diagrams
339 showing the ordering constraints.
340
341 [!] Note that the first load really has to have a _data_ dependency and
342 not a control dependency. If the address for the second load is dependent
343 on the first load, but the dependency is through a conditional rather than
344 actually loading the address itself, then it's a _control_ dependency and
345 a full read barrier or better is required. See the "Control dependencies"
346 subsection for more information.
347
348 [!] Note that data dependency barriers should normally be paired with
349 write barriers; see the "SMP barrier pairing" subsection.
350
351
352 (3) Read (or load) memory barriers.
353
354 A read barrier is a data dependency barrier plus a guarantee that all the
355 LOAD operations specified before the barrier will appear to happen before
356 all the LOAD operations specified after the barrier with respect to the
357 other components of the system.
358
359 A read barrier is a partial ordering on loads only; it is not required to
360 have any effect on stores.
361
362 Read memory barriers imply data dependency barriers, and so can substitute
363 for them.
364
365 [!] Note that read barriers should normally be paired with write barriers;
366 see the "SMP barrier pairing" subsection.
367
368
369 (4) General memory barriers.
370
David Howells670bd952006-06-10 09:54:12 -0700371 A general memory barrier gives a guarantee that all the LOAD and STORE
372 operations specified before the barrier will appear to happen before all
373 the LOAD and STORE operations specified after the barrier with respect to
374 the other components of the system.
375
376 A general memory barrier is a partial ordering over both loads and stores.
David Howells108b42b2006-03-31 16:00:29 +0100377
378 General memory barriers imply both read and write memory barriers, and so
379 can substitute for either.
380
381
382And a couple of implicit varieties:
383
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100384 (5) ACQUIRE operations.
David Howells108b42b2006-03-31 16:00:29 +0100385
386 This acts as a one-way permeable barrier. It guarantees that all memory
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100387 operations after the ACQUIRE operation will appear to happen after the
388 ACQUIRE operation with respect to the other components of the system.
389 ACQUIRE operations include LOCK operations and smp_load_acquire()
390 operations.
David Howells108b42b2006-03-31 16:00:29 +0100391
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100392 Memory operations that occur before an ACQUIRE operation may appear to
393 happen after it completes.
David Howells108b42b2006-03-31 16:00:29 +0100394
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100395 An ACQUIRE operation should almost always be paired with a RELEASE
396 operation.
David Howells108b42b2006-03-31 16:00:29 +0100397
398
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100399 (6) RELEASE operations.
David Howells108b42b2006-03-31 16:00:29 +0100400
401 This also acts as a one-way permeable barrier. It guarantees that all
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100402 memory operations before the RELEASE operation will appear to happen
403 before the RELEASE operation with respect to the other components of the
404 system. RELEASE operations include UNLOCK operations and
405 smp_store_release() operations.
David Howells108b42b2006-03-31 16:00:29 +0100406
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100407 Memory operations that occur after a RELEASE operation may appear to
David Howells108b42b2006-03-31 16:00:29 +0100408 happen before it completes.
409
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100410 The use of ACQUIRE and RELEASE operations generally precludes the need
411 for other sorts of memory barrier (but note the exceptions mentioned in
412 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
413 pair is -not- guaranteed to act as a full memory barrier. However, after
414 an ACQUIRE on a given variable, all memory accesses preceding any prior
415 RELEASE on that same variable are guaranteed to be visible. In other
416 words, within a given variable's critical section, all accesses of all
417 previous critical sections for that variable are guaranteed to have
418 completed.
Paul E. McKenney17eb88e2013-12-11 13:59:09 -0800419
Peter Zijlstra2e4f5382013-11-06 14:57:36 +0100420 This means that ACQUIRE acts as a minimal "acquire" operation and
421 RELEASE acts as a minimal "release" operation.
David Howells108b42b2006-03-31 16:00:29 +0100422
423
424Memory barriers are only required where there's a possibility of interaction
425between two CPUs or between a CPU and a device. If it can be guaranteed that
426there won't be any such interaction in any particular piece of code, then
427memory barriers are unnecessary in that piece of code.
428
429
430Note that these are the _minimum_ guarantees. Different architectures may give
431more substantial guarantees, but they may _not_ be relied upon outside of arch
432specific code.
433
434
435WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
436----------------------------------------------
437
438There are certain things that the Linux kernel memory barriers do not guarantee:
439
440 (*) There is no guarantee that any of the memory accesses specified before a
441 memory barrier will be _complete_ by the completion of a memory barrier
442 instruction; the barrier can be considered to draw a line in that CPU's
443 access queue that accesses of the appropriate type may not cross.
444
445 (*) There is no guarantee that issuing a memory barrier on one CPU will have
446 any direct effect on another CPU or any other hardware in the system. The
447 indirect effect will be the order in which the second CPU sees the effects
448 of the first CPU's accesses occur, but see the next point:
449
David Howells6bc39272006-06-25 05:49:22 -0700450 (*) There is no guarantee that a CPU will see the correct order of effects
David Howells108b42b2006-03-31 16:00:29 +0100451 from a second CPU's accesses, even _if_ the second CPU uses a memory
452 barrier, unless the first CPU _also_ uses a matching memory barrier (see
453 the subsection on "SMP Barrier Pairing").
454
455 (*) There is no guarantee that some intervening piece of off-the-CPU
456 hardware[*] will not reorder the memory accesses. CPU cache coherency
457 mechanisms should propagate the indirect effects of a memory barrier
458 between CPUs, but might not do so in order.
459
460 [*] For information on bus mastering DMA and coherency please read:
461
Randy Dunlap4b5ff462008-03-10 17:16:32 -0700462 Documentation/PCI/pci.txt
Paul Bolle395cf962011-08-15 02:02:26 +0200463 Documentation/DMA-API-HOWTO.txt
David Howells108b42b2006-03-31 16:00:29 +0100464 Documentation/DMA-API.txt
465
466
467DATA DEPENDENCY BARRIERS
468------------------------
469
470The usage requirements of data dependency barriers are a little subtle, and
471it's not always obvious that they're needed. To illustrate, consider the
472following sequence of events:
473
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800474 CPU 1 CPU 2
475 =============== ===============
David Howells108b42b2006-03-31 16:00:29 +0100476 { A == 1, B == 2, C = 3, P == &A, Q == &C }
477 B = 4;
478 <write barrier>
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800479 ACCESS_ONCE(P) = &B
480 Q = ACCESS_ONCE(P);
481 D = *Q;
David Howells108b42b2006-03-31 16:00:29 +0100482
483There's a clear data dependency here, and it would seem that by the end of the
484sequence, Q must be either &A or &B, and that:
485
486 (Q == &A) implies (D == 1)
487 (Q == &B) implies (D == 4)
488
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700489But! CPU 2's perception of P may be updated _before_ its perception of B, thus
David Howells108b42b2006-03-31 16:00:29 +0100490leading to the following situation:
491
492 (Q == &B) and (D == 2) ????
493
494Whilst this may seem like a failure of coherency or causality maintenance, it
495isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
496Alpha).
497
David Howells2b948952006-06-25 05:48:49 -0700498To deal with this, a data dependency barrier or better must be inserted
499between the address load and the data load:
David Howells108b42b2006-03-31 16:00:29 +0100500
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800501 CPU 1 CPU 2
502 =============== ===============
David Howells108b42b2006-03-31 16:00:29 +0100503 { A == 1, B == 2, C = 3, P == &A, Q == &C }
504 B = 4;
505 <write barrier>
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800506 ACCESS_ONCE(P) = &B
507 Q = ACCESS_ONCE(P);
508 <data dependency barrier>
509 D = *Q;
David Howells108b42b2006-03-31 16:00:29 +0100510
511This enforces the occurrence of one of the two implications, and prevents the
512third possibility from arising.
513
514[!] Note that this extremely counterintuitive situation arises most easily on
515machines with split caches, so that, for example, one cache bank processes
516even-numbered cache lines and the other bank processes odd-numbered cache
517lines. The pointer P might be stored in an odd-numbered cache line, and the
518variable B might be stored in an even-numbered cache line. Then, if the
519even-numbered bank of the reading CPU's cache is extremely busy while the
520odd-numbered bank is idle, one can see the new value of the pointer P (&B),
David Howells6bc39272006-06-25 05:49:22 -0700521but the old value of the variable B (2).
David Howells108b42b2006-03-31 16:00:29 +0100522
523
Ingo Molnare0edc782013-11-22 11:24:53 +0100524Another example of where data dependency barriers might be required is where a
David Howells108b42b2006-03-31 16:00:29 +0100525number is read from memory and then used to calculate the index for an array
526access:
527
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800528 CPU 1 CPU 2
529 =============== ===============
David Howells108b42b2006-03-31 16:00:29 +0100530 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
531 M[1] = 4;
532 <write barrier>
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800533 ACCESS_ONCE(P) = 1
534 Q = ACCESS_ONCE(P);
535 <data dependency barrier>
536 D = M[Q];
David Howells108b42b2006-03-31 16:00:29 +0100537
538
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800539The data dependency barrier is very important to the RCU system,
540for example. See rcu_assign_pointer() and rcu_dereference() in
541include/linux/rcupdate.h. This permits the current target of an RCU'd
542pointer to be replaced with a new modified target, without the replacement
543target appearing to be incompletely initialised.
David Howells108b42b2006-03-31 16:00:29 +0100544
545See also the subsection on "Cache Coherency" for a more thorough example.
546
547
548CONTROL DEPENDENCIES
549--------------------
550
551A control dependency requires a full read memory barrier, not simply a data
552dependency barrier to make it work correctly. Consider the following bit of
553code:
554
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800555 q = ACCESS_ONCE(a);
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800556 if (q) {
557 <data dependency barrier> /* BUG: No data dependency!!! */
558 p = ACCESS_ONCE(b);
Paul E. McKenney45c8a362013-07-02 15:24:09 -0700559 }
David Howells108b42b2006-03-31 16:00:29 +0100560
561This will not have the desired effect because there is no actual data
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800562dependency, but rather a control dependency that the CPU may short-circuit
563by attempting to predict the outcome in advance, so that other CPUs see
564the load from b as having happened before the load from a. In such a
565case what's actually required is:
David Howells108b42b2006-03-31 16:00:29 +0100566
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800567 q = ACCESS_ONCE(a);
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800568 if (q) {
Paul E. McKenney45c8a362013-07-02 15:24:09 -0700569 <read barrier>
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800570 p = ACCESS_ONCE(b);
Paul E. McKenney45c8a362013-07-02 15:24:09 -0700571 }
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800572
573However, stores are not speculated. This means that ordering -is- provided
574in the following example:
575
576 q = ACCESS_ONCE(a);
577 if (ACCESS_ONCE(q)) {
578 ACCESS_ONCE(b) = p;
579 }
580
581Please note that ACCESS_ONCE() is not optional! Without the ACCESS_ONCE(),
582the compiler is within its rights to transform this example:
583
584 q = a;
585 if (q) {
586 b = p; /* BUG: Compiler can reorder!!! */
587 do_something();
588 } else {
589 b = p; /* BUG: Compiler can reorder!!! */
590 do_something_else();
591 }
592
593into this, which of course defeats the ordering:
594
595 b = p;
596 q = a;
597 if (q)
598 do_something();
599 else
600 do_something_else();
601
602Worse yet, if the compiler is able to prove (say) that the value of
603variable 'a' is always non-zero, it would be well within its rights
604to optimize the original example by eliminating the "if" statement
605as follows:
606
607 q = a;
608 b = p; /* BUG: Compiler can reorder!!! */
609 do_something();
610
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800611The solution is again ACCESS_ONCE() and barrier(), which preserves the
612ordering between the load from variable 'a' and the store to variable 'b':
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800613
614 q = ACCESS_ONCE(a);
615 if (q) {
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800616 barrier();
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800617 ACCESS_ONCE(b) = p;
618 do_something();
619 } else {
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800620 barrier();
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800621 ACCESS_ONCE(b) = p;
622 do_something_else();
623 }
624
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800625The initial ACCESS_ONCE() is required to prevent the compiler from
626proving the value of 'a', and the pair of barrier() invocations are
627required to prevent the compiler from pulling the two identical stores
628to 'b' out from the legs of the "if" statement.
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800629
630It is important to note that control dependencies absolutely require a
631a conditional. For example, the following "optimized" version of
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800632the above example breaks ordering, which is why the barrier() invocations
633are absolutely required if you have identical stores in both legs of
634the "if" statement:
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800635
636 q = ACCESS_ONCE(a);
637 ACCESS_ONCE(b) = p; /* BUG: No ordering vs. load from a!!! */
638 if (q) {
639 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
640 do_something();
641 } else {
642 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
643 do_something_else();
644 }
645
646It is of course legal for the prior load to be part of the conditional,
647for example, as follows:
648
649 if (ACCESS_ONCE(a) > 0) {
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800650 barrier();
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800651 ACCESS_ONCE(b) = q / 2;
652 do_something();
653 } else {
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800654 barrier();
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800655 ACCESS_ONCE(b) = q / 3;
656 do_something_else();
657 }
658
659This will again ensure that the load from variable 'a' is ordered before the
660stores to variable 'b'.
661
662In addition, you need to be careful what you do with the local variable 'q',
663otherwise the compiler might be able to guess the value and again remove
664the needed conditional. For example:
665
666 q = ACCESS_ONCE(a);
667 if (q % MAX) {
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800668 barrier();
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800669 ACCESS_ONCE(b) = p;
670 do_something();
671 } else {
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800672 barrier();
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800673 ACCESS_ONCE(b) = p;
674 do_something_else();
675 }
676
677If MAX is defined to be 1, then the compiler knows that (q % MAX) is
678equal to zero, in which case the compiler is within its rights to
679transform the above code into the following:
680
681 q = ACCESS_ONCE(a);
682 ACCESS_ONCE(b) = p;
683 do_something_else();
684
685This transformation loses the ordering between the load from variable 'a'
686and the store to variable 'b'. If you are relying on this ordering, you
687should do something like the following:
688
689 q = ACCESS_ONCE(a);
690 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
691 if (q % MAX) {
692 ACCESS_ONCE(b) = p;
693 do_something();
694 } else {
695 ACCESS_ONCE(b) = p;
696 do_something_else();
697 }
698
699Finally, control dependencies do -not- provide transitivity. This is
700demonstrated by two related examples:
701
702 CPU 0 CPU 1
703 ===================== =====================
704 r1 = ACCESS_ONCE(x); r2 = ACCESS_ONCE(y);
705 if (r1 >= 0) if (r2 >= 0)
706 ACCESS_ONCE(y) = 1; ACCESS_ONCE(x) = 1;
707
708 assert(!(r1 == 1 && r2 == 1));
709
710The above two-CPU example will never trigger the assert(). However,
711if control dependencies guaranteed transitivity (which they do not),
712then adding the following two CPUs would guarantee a related assertion:
713
714 CPU 2 CPU 3
715 ===================== =====================
716 ACCESS_ONCE(x) = 2; ACCESS_ONCE(y) = 2;
717
718 assert(!(r1 == 2 && r2 == 2 && x == 1 && y == 1)); /* FAILS!!! */
719
720But because control dependencies do -not- provide transitivity, the
721above assertion can fail after the combined four-CPU example completes.
722If you need the four-CPU example to provide ordering, you will need
723smp_mb() between the loads and stores in the CPU 0 and CPU 1 code fragments.
724
725In summary:
726
727 (*) Control dependencies can order prior loads against later stores.
728 However, they do -not- guarantee any other sort of ordering:
729 Not prior loads against later loads, nor prior stores against
730 later anything. If you need these other forms of ordering,
731 use smb_rmb(), smp_wmb(), or, in the case of prior stores and
732 later loads, smp_mb().
733
Paul E. McKenney9b2b3bf2014-02-12 20:19:47 -0800734 (*) If both legs of the "if" statement begin with identical stores
735 to the same variable, a barrier() statement is required at the
736 beginning of each leg of the "if" statement.
737
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800738 (*) Control dependencies require at least one run-time conditional
Paul E. McKenney586dd562014-02-11 12:28:06 -0800739 between the prior load and the subsequent store, and this
740 conditional must involve the prior load. If the compiler
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800741 is able to optimize the conditional away, it will have also
742 optimized away the ordering. Careful use of ACCESS_ONCE() can
743 help to preserve the needed conditional.
744
745 (*) Control dependencies require that the compiler avoid reordering the
746 dependency into nonexistence. Careful use of ACCESS_ONCE() or
Paul E. McKenney692118d2013-12-11 13:59:07 -0800747 barrier() can help to preserve your control dependency. Please
748 see the Compiler Barrier section for more information.
Peter Zijlstra18c03c62013-12-11 13:59:06 -0800749
750 (*) Control dependencies do -not- provide transitivity. If you
751 need transitivity, use smp_mb().
David Howells108b42b2006-03-31 16:00:29 +0100752
753
754SMP BARRIER PAIRING
755-------------------
756
757When dealing with CPU-CPU interactions, certain types of memory barrier should
758always be paired. A lack of appropriate pairing is almost certainly an error.
759
760A write barrier should always be paired with a data dependency barrier or read
761barrier, though a general barrier would also be viable. Similarly a read
762barrier or a data dependency barrier should always be paired with at least an
763write barrier, though, again, a general barrier is viable:
764
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800765 CPU 1 CPU 2
766 =============== ===============
767 ACCESS_ONCE(a) = 1;
David Howells108b42b2006-03-31 16:00:29 +0100768 <write barrier>
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800769 ACCESS_ONCE(b) = 2; x = ACCESS_ONCE(b);
770 <read barrier>
771 y = ACCESS_ONCE(a);
David Howells108b42b2006-03-31 16:00:29 +0100772
773Or:
774
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800775 CPU 1 CPU 2
776 =============== ===============================
David Howells108b42b2006-03-31 16:00:29 +0100777 a = 1;
778 <write barrier>
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800779 ACCESS_ONCE(b) = &a; x = ACCESS_ONCE(b);
780 <data dependency barrier>
781 y = *x;
David Howells108b42b2006-03-31 16:00:29 +0100782
783Basically, the read barrier always has to be there, even though it can be of
784the "weaker" type.
785
David Howells670bd952006-06-10 09:54:12 -0700786[!] Note that the stores before the write barrier would normally be expected to
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700787match the loads after the read barrier or the data dependency barrier, and vice
David Howells670bd952006-06-10 09:54:12 -0700788versa:
789
Paul E. McKenney2ecf8102013-12-11 13:59:04 -0800790 CPU 1 CPU 2
791 =================== ===================
792 ACCESS_ONCE(a) = 1; }---- --->{ v = ACCESS_ONCE(c);
793 ACCESS_ONCE(b) = 2; } \ / { w = ACCESS_ONCE(d);
794 <write barrier> \ <read barrier>
795 ACCESS_ONCE(c) = 3; } / \ { x = ACCESS_ONCE(a);
796 ACCESS_ONCE(d) = 4; }---- --->{ y = ACCESS_ONCE(b);
David Howells670bd952006-06-10 09:54:12 -0700797
David Howells108b42b2006-03-31 16:00:29 +0100798
799EXAMPLES OF MEMORY BARRIER SEQUENCES
800------------------------------------
801
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700802Firstly, write barriers act as partial orderings on store operations.
David Howells108b42b2006-03-31 16:00:29 +0100803Consider the following sequence of events:
804
805 CPU 1
806 =======================
807 STORE A = 1
808 STORE B = 2
809 STORE C = 3
810 <write barrier>
811 STORE D = 4
812 STORE E = 5
813
814This sequence of events is committed to the memory coherence system in an order
815that the rest of the system might perceive as the unordered set of { STORE A,
Adrian Bunk80f72282006-06-30 18:27:16 +0200816STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
David Howells108b42b2006-03-31 16:00:29 +0100817}:
818
819 +-------+ : :
820 | | +------+
821 | |------>| C=3 | } /\
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700822 | | : +------+ }----- \ -----> Events perceptible to
823 | | : | A=1 | } \/ the rest of the system
David Howells108b42b2006-03-31 16:00:29 +0100824 | | : +------+ }
825 | CPU 1 | : | B=2 | }
826 | | +------+ }
827 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
828 | | +------+ } requires all stores prior to the
829 | | : | E=5 | } barrier to be committed before
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700830 | | : +------+ } further stores may take place
David Howells108b42b2006-03-31 16:00:29 +0100831 | |------>| D=4 | }
832 | | +------+
833 +-------+ : :
834 |
David Howells670bd952006-06-10 09:54:12 -0700835 | Sequence in which stores are committed to the
836 | memory system by CPU 1
David Howells108b42b2006-03-31 16:00:29 +0100837 V
838
839
Jarek Poplawski81fc6322007-05-23 13:58:20 -0700840Secondly, data dependency barriers act as partial orderings on data-dependent
David Howells108b42b2006-03-31 16:00:29 +0100841loads. Consider the following sequence of events:
842
843 CPU 1 CPU 2
844 ======================= =======================
David Howellsc14038c2006-04-10 22:54:24 -0700845 { B = 7; X = 9; Y = 8; C = &Y }
David Howells108b42b2006-03-31 16:00:29 +0100846 STORE A = 1
847 STORE B = 2
848 <write barrier>
849 STORE C = &B LOAD X
850 STORE D = 4 LOAD C (gets &B)
851 LOAD *C (reads B)
852
853Without intervention, CPU 2 may perceive the events on CPU 1 in some
854effectively random order, despite the write barrier issued by CPU 1:
855
856 +-------+ : : : :
857 | | +------+ +-------+ | Sequence of update
858 | |------>| B=2 |----- --->| Y->8 | | of perception on
859 | | : +------+ \ +-------+ | CPU 2
860 | CPU 1 | : | A=1 | \ --->| C->&Y | V
861 | | +------+ | +-------+
862 | | wwwwwwwwwwwwwwww | : :
863 | | +------+ | : :
864 | | : | C=&B |--- | : : +-------+
865 | | : +------+ \ | +-------+ | |
866 | |------>| D=4 | ----------->| C->&B |------>| |
867 | | +------+ | +-------+ | |
868 +-------+ : : | : : | |
869 | : : | |
870 | : : | CPU 2 |
871 | +-------+ | |
872 Apparently incorrect ---> | | B->7 |------>| |
873 perception of B (!) | +-------+ | |
874 | : : | |
875 | +-------+ | |
876 The load of X holds ---> \ | X->9 |------>| |
877 up the maintenance \ +-------+ | |
878 of coherence of B ----->| B->2 | +-------+
879 +-------+
880 : :
881
882
883In the above example, CPU 2 perceives that B is 7, despite the load of *C
Paolo Ornati670e9f32006-10-03 22:57:56 +0200884(which would be B) coming after the LOAD of C.
David Howells108b42b2006-03-31 16:00:29 +0100885
886If, however, a data dependency barrier were to be placed between the load of C
David Howellsc14038c2006-04-10 22:54:24 -0700887and the load of *C (ie: B) on CPU 2:
888
889 CPU 1 CPU 2
890 ======================= =======================
891 { B = 7; X = 9; Y = 8; C = &Y }
892 STORE A = 1
893 STORE B = 2
894 <write barrier>
895 STORE C = &B LOAD X
896 STORE D = 4 LOAD C (gets &B)
897 <data dependency barrier>
898 LOAD *C (reads B)
899
900then the following will occur:
David Howells108b42b2006-03-31 16:00:29 +0100901
902 +-------+ : : : :
903 | | +------+ +-------+
904 | |------>| B=2 |----- --->| Y->8 |
905 | | : +------+ \ +-------+
906 | CPU 1 | : | A=1 | \ --->| C->&Y |
907 | | +------+ | +-------+
908 | | wwwwwwwwwwwwwwww | : :
909 | | +------+ | : :
910 | | : | C=&B |--- | : : +-------+
911 | | : +------+ \ | +-------+ | |
912 | |------>| D=4 | ----------->| C->&B |------>| |
913 | | +------+ | +-------+ | |
914 +-------+ : : | : : | |
915 | : : | |
916 | : : | CPU 2 |
917 | +-------+ | |
David Howells670bd952006-06-10 09:54:12 -0700918 | | X->9 |------>| |
919 | +-------+ | |
920 Makes sure all effects ---> \ ddddddddddddddddd | |
921 prior to the store of C \ +-------+ | |
922 are perceptible to ----->| B->2 |------>| |
923 subsequent loads +-------+ | |
David Howells108b42b2006-03-31 16:00:29 +0100924 : : +-------+
925
926
927And thirdly, a read barrier acts as a partial order on loads. Consider the
928following sequence of events:
929
930 CPU 1 CPU 2
931 ======================= =======================
David Howells670bd952006-06-10 09:54:12 -0700932 { A = 0, B = 9 }
David Howells108b42b2006-03-31 16:00:29 +0100933 STORE A=1
David Howells108b42b2006-03-31 16:00:29 +0100934 <write barrier>
David Howells670bd952006-06-10 09:54:12 -0700935 STORE B=2
David Howells108b42b2006-03-31 16:00:29 +0100936 LOAD B
David Howells670bd952006-06-10 09:54:12 -0700937 LOAD A
David Howells108b42b2006-03-31 16:00:29 +0100938
939Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
940some effectively random order, despite the write barrier issued by CPU 1:
941
David Howells670bd952006-06-10 09:54:12 -0700942 +-------+ : : : :
943 | | +------+ +-------+
944 | |------>| A=1 |------ --->| A->0 |
945 | | +------+ \ +-------+
946 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
947 | | +------+ | +-------+
948 | |------>| B=2 |--- | : :
949 | | +------+ \ | : : +-------+
950 +-------+ : : \ | +-------+ | |
951 ---------->| B->2 |------>| |
952 | +-------+ | CPU 2 |
953 | | A->0 |------>| |
954 | +-------+ | |
955 | : : +-------+
956 \ : :
957 \ +-------+
958 ---->| A->1 |
959 +-------+
960 : :
David Howells108b42b2006-03-31 16:00:29 +0100961
962
David Howells6bc39272006-06-25 05:49:22 -0700963If, however, a read barrier were to be placed between the load of B and the
David Howells670bd952006-06-10 09:54:12 -0700964load of A on CPU 2:
David Howells108b42b2006-03-31 16:00:29 +0100965
David Howells670bd952006-06-10 09:54:12 -0700966 CPU 1 CPU 2
967 ======================= =======================
968 { A = 0, B = 9 }
969 STORE A=1
970 <write barrier>
971 STORE B=2
972 LOAD B
973 <read barrier>
974 LOAD A
975
976then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
9772:
978
979 +-------+ : : : :
980 | | +------+ +-------+
981 | |------>| A=1 |------ --->| A->0 |
982 | | +------+ \ +-------+
983 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
984 | | +------+ | +-------+
985 | |------>| B=2 |--- | : :
986 | | +------+ \ | : : +-------+
987 +-------+ : : \ | +-------+ | |
988 ---------->| B->2 |------>| |
989 | +-------+ | CPU 2 |
990 | : : | |
991 | : : | |
992 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
993 barrier causes all effects \ +-------+ | |
994 prior to the storage of B ---->| A->1 |------>| |
995 to be perceptible to CPU 2 +-------+ | |
996 : : +-------+
997
998
999To illustrate this more completely, consider what could happen if the code
1000contained a load of A either side of the read barrier:
1001
1002 CPU 1 CPU 2
1003 ======================= =======================
1004 { A = 0, B = 9 }
1005 STORE A=1
1006 <write barrier>
1007 STORE B=2
1008 LOAD B
1009 LOAD A [first load of A]
1010 <read barrier>
1011 LOAD A [second load of A]
1012
1013Even though the two loads of A both occur after the load of B, they may both
1014come up with different values:
1015
1016 +-------+ : : : :
1017 | | +------+ +-------+
1018 | |------>| A=1 |------ --->| A->0 |
1019 | | +------+ \ +-------+
1020 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1021 | | +------+ | +-------+
1022 | |------>| B=2 |--- | : :
1023 | | +------+ \ | : : +-------+
1024 +-------+ : : \ | +-------+ | |
1025 ---------->| B->2 |------>| |
1026 | +-------+ | CPU 2 |
1027 | : : | |
1028 | : : | |
1029 | +-------+ | |
1030 | | A->0 |------>| 1st |
1031 | +-------+ | |
1032 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1033 barrier causes all effects \ +-------+ | |
1034 prior to the storage of B ---->| A->1 |------>| 2nd |
1035 to be perceptible to CPU 2 +-------+ | |
1036 : : +-------+
1037
1038
1039But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1040before the read barrier completes anyway:
1041
1042 +-------+ : : : :
1043 | | +------+ +-------+
1044 | |------>| A=1 |------ --->| A->0 |
1045 | | +------+ \ +-------+
1046 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1047 | | +------+ | +-------+
1048 | |------>| B=2 |--- | : :
1049 | | +------+ \ | : : +-------+
1050 +-------+ : : \ | +-------+ | |
1051 ---------->| B->2 |------>| |
1052 | +-------+ | CPU 2 |
1053 | : : | |
1054 \ : : | |
1055 \ +-------+ | |
1056 ---->| A->1 |------>| 1st |
1057 +-------+ | |
1058 rrrrrrrrrrrrrrrrr | |
1059 +-------+ | |
1060 | A->1 |------>| 2nd |
1061 +-------+ | |
1062 : : +-------+
1063
1064
1065The guarantee is that the second load will always come up with A == 1 if the
1066load of B came up with B == 2. No such guarantee exists for the first load of
1067A; that may come up with either A == 0 or A == 1.
1068
1069
1070READ MEMORY BARRIERS VS LOAD SPECULATION
1071----------------------------------------
1072
1073Many CPUs speculate with loads: that is they see that they will need to load an
1074item from memory, and they find a time where they're not using the bus for any
1075other loads, and so do the load in advance - even though they haven't actually
1076got to that point in the instruction execution flow yet. This permits the
1077actual load instruction to potentially complete immediately because the CPU
1078already has the value to hand.
1079
1080It may turn out that the CPU didn't actually need the value - perhaps because a
1081branch circumvented the load - in which case it can discard the value or just
1082cache it for later use.
1083
1084Consider:
1085
Ingo Molnare0edc782013-11-22 11:24:53 +01001086 CPU 1 CPU 2
David Howells670bd952006-06-10 09:54:12 -07001087 ======================= =======================
Ingo Molnare0edc782013-11-22 11:24:53 +01001088 LOAD B
1089 DIVIDE } Divide instructions generally
1090 DIVIDE } take a long time to perform
1091 LOAD A
David Howells670bd952006-06-10 09:54:12 -07001092
1093Which might appear as this:
1094
1095 : : +-------+
1096 +-------+ | |
1097 --->| B->2 |------>| |
1098 +-------+ | CPU 2 |
1099 : :DIVIDE | |
1100 +-------+ | |
1101 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1102 division speculates on the +-------+ ~ | |
1103 LOAD of A : : ~ | |
1104 : :DIVIDE | |
1105 : : ~ | |
1106 Once the divisions are complete --> : : ~-->| |
1107 the CPU can then perform the : : | |
1108 LOAD with immediate effect : : +-------+
1109
1110
1111Placing a read barrier or a data dependency barrier just before the second
1112load:
1113
Ingo Molnare0edc782013-11-22 11:24:53 +01001114 CPU 1 CPU 2
David Howells670bd952006-06-10 09:54:12 -07001115 ======================= =======================
Ingo Molnare0edc782013-11-22 11:24:53 +01001116 LOAD B
1117 DIVIDE
1118 DIVIDE
David Howells670bd952006-06-10 09:54:12 -07001119 <read barrier>
Ingo Molnare0edc782013-11-22 11:24:53 +01001120 LOAD A
David Howells670bd952006-06-10 09:54:12 -07001121
1122will force any value speculatively obtained to be reconsidered to an extent
1123dependent on the type of barrier used. If there was no change made to the
1124speculated memory location, then the speculated value will just be used:
1125
1126 : : +-------+
1127 +-------+ | |
1128 --->| B->2 |------>| |
1129 +-------+ | CPU 2 |
1130 : :DIVIDE | |
1131 +-------+ | |
1132 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1133 division speculates on the +-------+ ~ | |
1134 LOAD of A : : ~ | |
1135 : :DIVIDE | |
1136 : : ~ | |
1137 : : ~ | |
1138 rrrrrrrrrrrrrrrr~ | |
1139 : : ~ | |
1140 : : ~-->| |
1141 : : | |
1142 : : +-------+
1143
1144
1145but if there was an update or an invalidation from another CPU pending, then
1146the speculation will be cancelled and the value reloaded:
1147
1148 : : +-------+
1149 +-------+ | |
1150 --->| B->2 |------>| |
1151 +-------+ | CPU 2 |
1152 : :DIVIDE | |
1153 +-------+ | |
1154 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1155 division speculates on the +-------+ ~ | |
1156 LOAD of A : : ~ | |
1157 : :DIVIDE | |
1158 : : ~ | |
1159 : : ~ | |
1160 rrrrrrrrrrrrrrrrr | |
1161 +-------+ | |
1162 The speculation is discarded ---> --->| A->1 |------>| |
1163 and an updated value is +-------+ | |
1164 retrieved : : +-------+
David Howells108b42b2006-03-31 16:00:29 +01001165
1166
Paul E. McKenney241e6662011-02-10 16:54:50 -08001167TRANSITIVITY
1168------------
1169
1170Transitivity is a deeply intuitive notion about ordering that is not
1171always provided by real computer systems. The following example
1172demonstrates transitivity (also called "cumulativity"):
1173
1174 CPU 1 CPU 2 CPU 3
1175 ======================= ======================= =======================
1176 { X = 0, Y = 0 }
1177 STORE X=1 LOAD X STORE Y=1
1178 <general barrier> <general barrier>
1179 LOAD Y LOAD X
1180
1181Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1182This indicates that CPU 2's load from X in some sense follows CPU 1's
1183store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1184store to Y. The question is then "Can CPU 3's load from X return 0?"
1185
1186Because CPU 2's load from X in some sense came after CPU 1's store, it
1187is natural to expect that CPU 3's load from X must therefore return 1.
1188This expectation is an example of transitivity: if a load executing on
1189CPU A follows a load from the same variable executing on CPU B, then
1190CPU A's load must either return the same value that CPU B's load did,
1191or must return some later value.
1192
1193In the Linux kernel, use of general memory barriers guarantees
1194transitivity. Therefore, in the above example, if CPU 2's load from X
1195returns 1 and its load from Y returns 0, then CPU 3's load from X must
1196also return 1.
1197
1198However, transitivity is -not- guaranteed for read or write barriers.
1199For example, suppose that CPU 2's general barrier in the above example
1200is changed to a read barrier as shown below:
1201
1202 CPU 1 CPU 2 CPU 3
1203 ======================= ======================= =======================
1204 { X = 0, Y = 0 }
1205 STORE X=1 LOAD X STORE Y=1
1206 <read barrier> <general barrier>
1207 LOAD Y LOAD X
1208
1209This substitution destroys transitivity: in this example, it is perfectly
1210legal for CPU 2's load from X to return 1, its load from Y to return 0,
1211and CPU 3's load from X to return 0.
1212
1213The key point is that although CPU 2's read barrier orders its pair
1214of loads, it does not guarantee to order CPU 1's store. Therefore, if
1215this example runs on a system where CPUs 1 and 2 share a store buffer
1216or a level of cache, CPU 2 might have early access to CPU 1's writes.
1217General barriers are therefore required to ensure that all CPUs agree
1218on the combined order of CPU 1's and CPU 2's accesses.
1219
1220To reiterate, if your code requires transitivity, use general barriers
1221throughout.
1222
1223
David Howells108b42b2006-03-31 16:00:29 +01001224========================
1225EXPLICIT KERNEL BARRIERS
1226========================
1227
1228The Linux kernel has a variety of different barriers that act at different
1229levels:
1230
1231 (*) Compiler barrier.
1232
1233 (*) CPU memory barriers.
1234
1235 (*) MMIO write barrier.
1236
1237
1238COMPILER BARRIER
1239----------------
1240
1241The Linux kernel has an explicit compiler barrier function that prevents the
1242compiler from moving the memory accesses either side of it to the other side:
1243
1244 barrier();
1245
Peter Zijlstra18c03c62013-12-11 13:59:06 -08001246This is a general barrier -- there are no read-read or write-write variants
Paul E. McKenney692118d2013-12-11 13:59:07 -08001247of barrier(). However, ACCESS_ONCE() can be thought of as a weak form
Peter Zijlstra18c03c62013-12-11 13:59:06 -08001248for barrier() that affects only the specific accesses flagged by the
1249ACCESS_ONCE().
David Howells108b42b2006-03-31 16:00:29 +01001250
Paul E. McKenney692118d2013-12-11 13:59:07 -08001251The barrier() function has the following effects:
1252
1253 (*) Prevents the compiler from reordering accesses following the
1254 barrier() to precede any accesses preceding the barrier().
1255 One example use for this property is to ease communication between
1256 interrupt-handler code and the code that was interrupted.
1257
1258 (*) Within a loop, forces the compiler to load the variables used
1259 in that loop's conditional on each pass through that loop.
1260
1261The ACCESS_ONCE() function can prevent any number of optimizations that,
1262while perfectly safe in single-threaded code, can be fatal in concurrent
1263code. Here are some examples of these sorts of optimizations:
1264
Paul E. McKenney449f7412014-01-02 15:03:50 -08001265 (*) The compiler is within its rights to reorder loads and stores
1266 to the same variable, and in some cases, the CPU is within its
1267 rights to reorder loads to the same variable. This means that
1268 the following code:
1269
1270 a[0] = x;
1271 a[1] = x;
1272
1273 Might result in an older value of x stored in a[1] than in a[0].
1274 Prevent both the compiler and the CPU from doing this as follows:
1275
1276 a[0] = ACCESS_ONCE(x);
1277 a[1] = ACCESS_ONCE(x);
1278
1279 In short, ACCESS_ONCE() provides cache coherence for accesses from
1280 multiple CPUs to a single variable.
1281
Paul E. McKenney692118d2013-12-11 13:59:07 -08001282 (*) The compiler is within its rights to merge successive loads from
1283 the same variable. Such merging can cause the compiler to "optimize"
1284 the following code:
1285
1286 while (tmp = a)
1287 do_something_with(tmp);
1288
1289 into the following code, which, although in some sense legitimate
1290 for single-threaded code, is almost certainly not what the developer
1291 intended:
1292
1293 if (tmp = a)
1294 for (;;)
1295 do_something_with(tmp);
1296
1297 Use ACCESS_ONCE() to prevent the compiler from doing this to you:
1298
1299 while (tmp = ACCESS_ONCE(a))
1300 do_something_with(tmp);
1301
1302 (*) The compiler is within its rights to reload a variable, for example,
1303 in cases where high register pressure prevents the compiler from
1304 keeping all data of interest in registers. The compiler might
1305 therefore optimize the variable 'tmp' out of our previous example:
1306
1307 while (tmp = a)
1308 do_something_with(tmp);
1309
1310 This could result in the following code, which is perfectly safe in
1311 single-threaded code, but can be fatal in concurrent code:
1312
1313 while (a)
1314 do_something_with(a);
1315
1316 For example, the optimized version of this code could result in
1317 passing a zero to do_something_with() in the case where the variable
1318 a was modified by some other CPU between the "while" statement and
1319 the call to do_something_with().
1320
1321 Again, use ACCESS_ONCE() to prevent the compiler from doing this:
1322
1323 while (tmp = ACCESS_ONCE(a))
1324 do_something_with(tmp);
1325
1326 Note that if the compiler runs short of registers, it might save
1327 tmp onto the stack. The overhead of this saving and later restoring
1328 is why compilers reload variables. Doing so is perfectly safe for
1329 single-threaded code, so you need to tell the compiler about cases
1330 where it is not safe.
1331
1332 (*) The compiler is within its rights to omit a load entirely if it knows
1333 what the value will be. For example, if the compiler can prove that
1334 the value of variable 'a' is always zero, it can optimize this code:
1335
1336 while (tmp = a)
1337 do_something_with(tmp);
1338
1339 Into this:
1340
1341 do { } while (0);
1342
1343 This transformation is a win for single-threaded code because it gets
1344 rid of a load and a branch. The problem is that the compiler will
1345 carry out its proof assuming that the current CPU is the only one
1346 updating variable 'a'. If variable 'a' is shared, then the compiler's
1347 proof will be erroneous. Use ACCESS_ONCE() to tell the compiler
1348 that it doesn't know as much as it thinks it does:
1349
1350 while (tmp = ACCESS_ONCE(a))
1351 do_something_with(tmp);
1352
1353 But please note that the compiler is also closely watching what you
1354 do with the value after the ACCESS_ONCE(). For example, suppose you
1355 do the following and MAX is a preprocessor macro with the value 1:
1356
1357 while ((tmp = ACCESS_ONCE(a)) % MAX)
1358 do_something_with(tmp);
1359
1360 Then the compiler knows that the result of the "%" operator applied
1361 to MAX will always be zero, again allowing the compiler to optimize
1362 the code into near-nonexistence. (It will still load from the
1363 variable 'a'.)
1364
1365 (*) Similarly, the compiler is within its rights to omit a store entirely
1366 if it knows that the variable already has the value being stored.
1367 Again, the compiler assumes that the current CPU is the only one
1368 storing into the variable, which can cause the compiler to do the
1369 wrong thing for shared variables. For example, suppose you have
1370 the following:
1371
1372 a = 0;
1373 /* Code that does not store to variable a. */
1374 a = 0;
1375
1376 The compiler sees that the value of variable 'a' is already zero, so
1377 it might well omit the second store. This would come as a fatal
1378 surprise if some other CPU might have stored to variable 'a' in the
1379 meantime.
1380
1381 Use ACCESS_ONCE() to prevent the compiler from making this sort of
1382 wrong guess:
1383
1384 ACCESS_ONCE(a) = 0;
1385 /* Code that does not store to variable a. */
1386 ACCESS_ONCE(a) = 0;
1387
1388 (*) The compiler is within its rights to reorder memory accesses unless
1389 you tell it not to. For example, consider the following interaction
1390 between process-level code and an interrupt handler:
1391
1392 void process_level(void)
1393 {
1394 msg = get_message();
1395 flag = true;
1396 }
1397
1398 void interrupt_handler(void)
1399 {
1400 if (flag)
1401 process_message(msg);
1402 }
1403
Masanari Iidadf5cbb22014-03-21 10:04:30 +09001404 There is nothing to prevent the compiler from transforming
Paul E. McKenney692118d2013-12-11 13:59:07 -08001405 process_level() to the following, in fact, this might well be a
1406 win for single-threaded code:
1407
1408 void process_level(void)
1409 {
1410 flag = true;
1411 msg = get_message();
1412 }
1413
1414 If the interrupt occurs between these two statement, then
1415 interrupt_handler() might be passed a garbled msg. Use ACCESS_ONCE()
1416 to prevent this as follows:
1417
1418 void process_level(void)
1419 {
1420 ACCESS_ONCE(msg) = get_message();
1421 ACCESS_ONCE(flag) = true;
1422 }
1423
1424 void interrupt_handler(void)
1425 {
1426 if (ACCESS_ONCE(flag))
1427 process_message(ACCESS_ONCE(msg));
1428 }
1429
1430 Note that the ACCESS_ONCE() wrappers in interrupt_handler()
1431 are needed if this interrupt handler can itself be interrupted
1432 by something that also accesses 'flag' and 'msg', for example,
1433 a nested interrupt or an NMI. Otherwise, ACCESS_ONCE() is not
1434 needed in interrupt_handler() other than for documentation purposes.
1435 (Note also that nested interrupts do not typically occur in modern
1436 Linux kernels, in fact, if an interrupt handler returns with
1437 interrupts enabled, you will get a WARN_ONCE() splat.)
1438
1439 You should assume that the compiler can move ACCESS_ONCE() past
1440 code not containing ACCESS_ONCE(), barrier(), or similar primitives.
1441
1442 This effect could also be achieved using barrier(), but ACCESS_ONCE()
1443 is more selective: With ACCESS_ONCE(), the compiler need only forget
1444 the contents of the indicated memory locations, while with barrier()
1445 the compiler must discard the value of all memory locations that
1446 it has currented cached in any machine registers. Of course,
1447 the compiler must also respect the order in which the ACCESS_ONCE()s
1448 occur, though the CPU of course need not do so.
1449
1450 (*) The compiler is within its rights to invent stores to a variable,
1451 as in the following example:
1452
1453 if (a)
1454 b = a;
1455 else
1456 b = 42;
1457
1458 The compiler might save a branch by optimizing this as follows:
1459
1460 b = 42;
1461 if (a)
1462 b = a;
1463
1464 In single-threaded code, this is not only safe, but also saves
1465 a branch. Unfortunately, in concurrent code, this optimization
1466 could cause some other CPU to see a spurious value of 42 -- even
1467 if variable 'a' was never zero -- when loading variable 'b'.
1468 Use ACCESS_ONCE() to prevent this as follows:
1469
1470 if (a)
1471 ACCESS_ONCE(b) = a;
1472 else
1473 ACCESS_ONCE(b) = 42;
1474
1475 The compiler can also invent loads. These are usually less
1476 damaging, but they can result in cache-line bouncing and thus in
1477 poor performance and scalability. Use ACCESS_ONCE() to prevent
1478 invented loads.
1479
1480 (*) For aligned memory locations whose size allows them to be accessed
1481 with a single memory-reference instruction, prevents "load tearing"
1482 and "store tearing," in which a single large access is replaced by
1483 multiple smaller accesses. For example, given an architecture having
1484 16-bit store instructions with 7-bit immediate fields, the compiler
1485 might be tempted to use two 16-bit store-immediate instructions to
1486 implement the following 32-bit store:
1487
1488 p = 0x00010002;
1489
1490 Please note that GCC really does use this sort of optimization,
1491 which is not surprising given that it would likely take more
1492 than two instructions to build the constant and then store it.
1493 This optimization can therefore be a win in single-threaded code.
1494 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1495 this optimization in a volatile store. In the absence of such bugs,
1496 use of ACCESS_ONCE() prevents store tearing in the following example:
1497
1498 ACCESS_ONCE(p) = 0x00010002;
1499
1500 Use of packed structures can also result in load and store tearing,
1501 as in this example:
1502
1503 struct __attribute__((__packed__)) foo {
1504 short a;
1505 int b;
1506 short c;
1507 };
1508 struct foo foo1, foo2;
1509 ...
1510
1511 foo2.a = foo1.a;
1512 foo2.b = foo1.b;
1513 foo2.c = foo1.c;
1514
1515 Because there are no ACCESS_ONCE() wrappers and no volatile markings,
1516 the compiler would be well within its rights to implement these three
1517 assignment statements as a pair of 32-bit loads followed by a pair
1518 of 32-bit stores. This would result in load tearing on 'foo1.b'
1519 and store tearing on 'foo2.b'. ACCESS_ONCE() again prevents tearing
1520 in this example:
1521
1522 foo2.a = foo1.a;
1523 ACCESS_ONCE(foo2.b) = ACCESS_ONCE(foo1.b);
1524 foo2.c = foo1.c;
1525
1526All that aside, it is never necessary to use ACCESS_ONCE() on a variable
1527that has been marked volatile. For example, because 'jiffies' is marked
1528volatile, it is never necessary to say ACCESS_ONCE(jiffies). The reason
1529for this is that ACCESS_ONCE() is implemented as a volatile cast, which
1530has no effect when its argument is already marked volatile.
1531
1532Please note that these compiler barriers have no direct effect on the CPU,
1533which may then reorder things however it wishes.
David Howells108b42b2006-03-31 16:00:29 +01001534
1535
1536CPU MEMORY BARRIERS
1537-------------------
1538
1539The Linux kernel has eight basic CPU memory barriers:
1540
1541 TYPE MANDATORY SMP CONDITIONAL
1542 =============== ======================= ===========================
1543 GENERAL mb() smp_mb()
1544 WRITE wmb() smp_wmb()
1545 READ rmb() smp_rmb()
1546 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1547
1548
Nick Piggin73f10282008-05-14 06:35:11 +02001549All memory barriers except the data dependency barriers imply a compiler
1550barrier. Data dependencies do not impose any additional compiler ordering.
1551
1552Aside: In the case of data dependencies, the compiler would be expected to
1553issue the loads in the correct order (eg. `a[b]` would have to load the value
1554of b before loading a[b]), however there is no guarantee in the C specification
1555that the compiler may not speculate the value of b (eg. is equal to 1) and load
1556a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1557problem of a compiler reloading b after having loaded a[b], thus having a newer
1558copy of b than a[b]. A consensus has not yet been reached about these problems,
1559however the ACCESS_ONCE macro is a good place to start looking.
David Howells108b42b2006-03-31 16:00:29 +01001560
1561SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
Jarek Poplawski81fc6322007-05-23 13:58:20 -07001562systems because it is assumed that a CPU will appear to be self-consistent,
David Howells108b42b2006-03-31 16:00:29 +01001563and will order overlapping accesses correctly with respect to itself.
1564
1565[!] Note that SMP memory barriers _must_ be used to control the ordering of
1566references to shared memory on SMP systems, though the use of locking instead
1567is sufficient.
1568
1569Mandatory barriers should not be used to control SMP effects, since mandatory
1570barriers unnecessarily impose overhead on UP systems. They may, however, be
1571used to control MMIO effects on accesses through relaxed memory I/O windows.
1572These are required even on non-SMP systems as they affect the order in which
1573memory operations appear to a device by prohibiting both the compiler and the
1574CPU from reordering them.
1575
1576
1577There are some more advanced barrier functions:
1578
1579 (*) set_mb(var, value)
David Howells108b42b2006-03-31 16:00:29 +01001580
Oleg Nesterov75b2bd52006-11-08 17:44:38 -08001581 This assigns the value to the variable and then inserts a full memory
Steven Rostedtf92213b2006-07-14 16:05:01 -04001582 barrier after it, depending on the function. It isn't guaranteed to
David Howells108b42b2006-03-31 16:00:29 +01001583 insert anything more than a compiler barrier in a UP compilation.
1584
1585
1586 (*) smp_mb__before_atomic_dec();
1587 (*) smp_mb__after_atomic_dec();
1588 (*) smp_mb__before_atomic_inc();
1589 (*) smp_mb__after_atomic_inc();
1590
1591 These are for use with atomic add, subtract, increment and decrement
David Howellsdbc87002006-04-10 22:54:23 -07001592 functions that don't return a value, especially when used for reference
1593 counting. These functions do not imply memory barriers.
David Howells108b42b2006-03-31 16:00:29 +01001594
1595 As an example, consider a piece of code that marks an object as being dead
1596 and then decrements the object's reference count:
1597
1598 obj->dead = 1;
1599 smp_mb__before_atomic_dec();
1600 atomic_dec(&obj->ref_count);
1601
1602 This makes sure that the death mark on the object is perceived to be set
1603 *before* the reference counter is decremented.
1604
1605 See Documentation/atomic_ops.txt for more information. See the "Atomic
1606 operations" subsection for information on where to use these.
1607
1608
1609 (*) smp_mb__before_clear_bit(void);
1610 (*) smp_mb__after_clear_bit(void);
1611
1612 These are for use similar to the atomic inc/dec barriers. These are
1613 typically used for bitwise unlocking operations, so care must be taken as
1614 there are no implicit memory barriers here either.
1615
1616 Consider implementing an unlock operation of some nature by clearing a
1617 locking bit. The clear_bit() would then need to be barriered like this:
1618
1619 smp_mb__before_clear_bit();
1620 clear_bit( ... );
1621
1622 This prevents memory operations before the clear leaking to after it. See
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001623 the subsection on "Locking Functions" with reference to RELEASE operation
David Howells108b42b2006-03-31 16:00:29 +01001624 implications.
1625
1626 See Documentation/atomic_ops.txt for more information. See the "Atomic
1627 operations" subsection for information on where to use these.
1628
1629
1630MMIO WRITE BARRIER
1631------------------
1632
1633The Linux kernel also has a special barrier for use with memory-mapped I/O
1634writes:
1635
1636 mmiowb();
1637
1638This is a variation on the mandatory write barrier that causes writes to weakly
1639ordered I/O regions to be partially ordered. Its effects may go beyond the
1640CPU->Hardware interface and actually affect the hardware at some level.
1641
1642See the subsection "Locks vs I/O accesses" for more information.
1643
1644
1645===============================
1646IMPLICIT KERNEL MEMORY BARRIERS
1647===============================
1648
1649Some of the other functions in the linux kernel imply memory barriers, amongst
David Howells670bd952006-06-10 09:54:12 -07001650which are locking and scheduling functions.
David Howells108b42b2006-03-31 16:00:29 +01001651
1652This specification is a _minimum_ guarantee; any particular architecture may
1653provide more substantial guarantees, but these may not be relied upon outside
1654of arch specific code.
1655
1656
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001657ACQUIRING FUNCTIONS
1658-------------------
David Howells108b42b2006-03-31 16:00:29 +01001659
1660The Linux kernel has a number of locking constructs:
1661
1662 (*) spin locks
1663 (*) R/W spin locks
1664 (*) mutexes
1665 (*) semaphores
1666 (*) R/W semaphores
1667 (*) RCU
1668
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001669In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
David Howells108b42b2006-03-31 16:00:29 +01001670for each construct. These operations all imply certain barriers:
1671
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001672 (1) ACQUIRE operation implication:
David Howells108b42b2006-03-31 16:00:29 +01001673
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001674 Memory operations issued after the ACQUIRE will be completed after the
1675 ACQUIRE operation has completed.
David Howells108b42b2006-03-31 16:00:29 +01001676
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001677 Memory operations issued before the ACQUIRE may be completed after
1678 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1679 combined with a following ACQUIRE, orders prior loads against
1680 subsequent loads and stores and also orders prior stores against
1681 subsequent stores. Note that this is weaker than smp_mb()! The
1682 smp_mb__before_spinlock() primitive is free on many architectures.
David Howells108b42b2006-03-31 16:00:29 +01001683
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001684 (2) RELEASE operation implication:
David Howells108b42b2006-03-31 16:00:29 +01001685
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001686 Memory operations issued before the RELEASE will be completed before the
1687 RELEASE operation has completed.
David Howells108b42b2006-03-31 16:00:29 +01001688
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001689 Memory operations issued after the RELEASE may be completed before the
1690 RELEASE operation has completed.
David Howells108b42b2006-03-31 16:00:29 +01001691
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001692 (3) ACQUIRE vs ACQUIRE implication:
David Howells108b42b2006-03-31 16:00:29 +01001693
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001694 All ACQUIRE operations issued before another ACQUIRE operation will be
1695 completed before that ACQUIRE operation.
David Howells108b42b2006-03-31 16:00:29 +01001696
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001697 (4) ACQUIRE vs RELEASE implication:
David Howells108b42b2006-03-31 16:00:29 +01001698
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001699 All ACQUIRE operations issued before a RELEASE operation will be
1700 completed before the RELEASE operation.
David Howells108b42b2006-03-31 16:00:29 +01001701
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001702 (5) Failed conditional ACQUIRE implication:
David Howells108b42b2006-03-31 16:00:29 +01001703
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001704 Certain locking variants of the ACQUIRE operation may fail, either due to
1705 being unable to get the lock immediately, or due to receiving an unblocked
David Howells108b42b2006-03-31 16:00:29 +01001706 signal whilst asleep waiting for the lock to become available. Failed
1707 locks do not imply any sort of barrier.
1708
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001709[!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1710one-way barriers is that the effects of instructions outside of a critical
1711section may seep into the inside of the critical section.
David Howells108b42b2006-03-31 16:00:29 +01001712
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001713An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1714because it is possible for an access preceding the ACQUIRE to happen after the
1715ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
1716the two accesses can themselves then cross:
David Howells670bd952006-06-10 09:54:12 -07001717
1718 *A = a;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001719 ACQUIRE M
1720 RELEASE M
David Howells670bd952006-06-10 09:54:12 -07001721 *B = b;
1722
1723may occur as:
1724
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001725 ACQUIRE M, STORE *B, STORE *A, RELEASE M
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001726
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001727When the ACQUIRE and RELEASE are a lock acquisition and release,
1728respectively, this same reordering can occur if the lock's ACQUIRE and
1729RELEASE are to the same lock variable, but only from the perspective of
1730another CPU not holding that lock. In short, a ACQUIRE followed by an
1731RELEASE may -not- be assumed to be a full memory barrier.
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001732
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001733Similarly, the reverse case of a RELEASE followed by an ACQUIRE does not
1734imply a full memory barrier. If it is necessary for a RELEASE-ACQUIRE
1735pair to produce a full barrier, the ACQUIRE can be followed by an
1736smp_mb__after_unlock_lock() invocation. This will produce a full barrier
1737if either (a) the RELEASE and the ACQUIRE are executed by the same
1738CPU or task, or (b) the RELEASE and ACQUIRE act on the same variable.
1739The smp_mb__after_unlock_lock() primitive is free on many architectures.
1740Without smp_mb__after_unlock_lock(), the CPU's execution of the critical
1741sections corresponding to the RELEASE and the ACQUIRE can cross, so that:
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001742
1743 *A = a;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001744 RELEASE M
1745 ACQUIRE N
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001746 *B = b;
1747
1748could occur as:
1749
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001750 ACQUIRE N, STORE *B, STORE *A, RELEASE M
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001751
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001752It might appear that this reordering could introduce a deadlock.
1753However, this cannot happen because if such a deadlock threatened,
1754the RELEASE would simply complete, thereby avoiding the deadlock.
1755
1756 Why does this work?
1757
1758 One key point is that we are only talking about the CPU doing
1759 the reordering, not the compiler. If the compiler (or, for
1760 that matter, the developer) switched the operations, deadlock
1761 -could- occur.
1762
1763 But suppose the CPU reordered the operations. In this case,
1764 the unlock precedes the lock in the assembly code. The CPU
1765 simply elected to try executing the later lock operation first.
1766 If there is a deadlock, this lock operation will simply spin (or
1767 try to sleep, but more on that later). The CPU will eventually
1768 execute the unlock operation (which preceded the lock operation
1769 in the assembly code), which will unravel the potential deadlock,
1770 allowing the lock operation to succeed.
1771
1772 But what if the lock is a sleeplock? In that case, the code will
1773 try to enter the scheduler, where it will eventually encounter
1774 a memory barrier, which will force the earlier unlock operation
1775 to complete, again unraveling the deadlock. There might be
1776 a sleep-unlock race, but the locking primitive needs to resolve
1777 such races properly in any case.
1778
1779With smp_mb__after_unlock_lock(), the two critical sections cannot overlap.
1780For example, with the following code, the store to *A will always be
1781seen by other CPUs before the store to *B:
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001782
1783 *A = a;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001784 RELEASE M
1785 ACQUIRE N
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001786 smp_mb__after_unlock_lock();
1787 *B = b;
1788
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001789The operations will always occur in one of the following orders:
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001790
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001791 STORE *A, RELEASE, ACQUIRE, smp_mb__after_unlock_lock(), STORE *B
1792 STORE *A, ACQUIRE, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1793 ACQUIRE, STORE *A, RELEASE, smp_mb__after_unlock_lock(), STORE *B
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08001794
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001795If the RELEASE and ACQUIRE were instead both operating on the same lock
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08001796variable, only the first of these alternatives can occur. In addition,
1797the more strongly ordered systems may rule out some of the above orders.
1798But in any case, as noted earlier, the smp_mb__after_unlock_lock()
1799ensures that the store to *A will always be seen as happening before
1800the store to *B.
David Howells670bd952006-06-10 09:54:12 -07001801
David Howells108b42b2006-03-31 16:00:29 +01001802Locks and semaphores may not provide any guarantee of ordering on UP compiled
1803systems, and so cannot be counted on in such a situation to actually achieve
1804anything at all - especially with respect to I/O accesses - unless combined
1805with interrupt disabling operations.
1806
1807See also the section on "Inter-CPU locking barrier effects".
1808
1809
1810As an example, consider the following:
1811
1812 *A = a;
1813 *B = b;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001814 ACQUIRE
David Howells108b42b2006-03-31 16:00:29 +01001815 *C = c;
1816 *D = d;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001817 RELEASE
David Howells108b42b2006-03-31 16:00:29 +01001818 *E = e;
1819 *F = f;
1820
1821The following sequence of events is acceptable:
1822
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001823 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
David Howells108b42b2006-03-31 16:00:29 +01001824
1825 [+] Note that {*F,*A} indicates a combined access.
1826
1827But none of the following are:
1828
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001829 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
1830 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
1831 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
1832 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
David Howells108b42b2006-03-31 16:00:29 +01001833
1834
1835
1836INTERRUPT DISABLING FUNCTIONS
1837-----------------------------
1838
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001839Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
1840(RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
David Howells108b42b2006-03-31 16:00:29 +01001841barriers are required in such a situation, they must be provided from some
1842other means.
1843
1844
David Howells50fa6102009-04-28 15:01:38 +01001845SLEEP AND WAKE-UP FUNCTIONS
1846---------------------------
1847
1848Sleeping and waking on an event flagged in global data can be viewed as an
1849interaction between two pieces of data: the task state of the task waiting for
1850the event and the global data used to indicate the event. To make sure that
1851these appear to happen in the right order, the primitives to begin the process
1852of going to sleep, and the primitives to initiate a wake up imply certain
1853barriers.
1854
1855Firstly, the sleeper normally follows something like this sequence of events:
1856
1857 for (;;) {
1858 set_current_state(TASK_UNINTERRUPTIBLE);
1859 if (event_indicated)
1860 break;
1861 schedule();
1862 }
1863
1864A general memory barrier is interpolated automatically by set_current_state()
1865after it has altered the task state:
1866
1867 CPU 1
1868 ===============================
1869 set_current_state();
1870 set_mb();
1871 STORE current->state
1872 <general barrier>
1873 LOAD event_indicated
1874
1875set_current_state() may be wrapped by:
1876
1877 prepare_to_wait();
1878 prepare_to_wait_exclusive();
1879
1880which therefore also imply a general memory barrier after setting the state.
1881The whole sequence above is available in various canned forms, all of which
1882interpolate the memory barrier in the right place:
1883
1884 wait_event();
1885 wait_event_interruptible();
1886 wait_event_interruptible_exclusive();
1887 wait_event_interruptible_timeout();
1888 wait_event_killable();
1889 wait_event_timeout();
1890 wait_on_bit();
1891 wait_on_bit_lock();
1892
1893
1894Secondly, code that performs a wake up normally follows something like this:
1895
1896 event_indicated = 1;
1897 wake_up(&event_wait_queue);
1898
1899or:
1900
1901 event_indicated = 1;
1902 wake_up_process(event_daemon);
1903
1904A write memory barrier is implied by wake_up() and co. if and only if they wake
1905something up. The barrier occurs before the task state is cleared, and so sits
1906between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1907
1908 CPU 1 CPU 2
1909 =============================== ===============================
1910 set_current_state(); STORE event_indicated
1911 set_mb(); wake_up();
1912 STORE current->state <write barrier>
1913 <general barrier> STORE current->state
1914 LOAD event_indicated
1915
1916The available waker functions include:
1917
1918 complete();
1919 wake_up();
1920 wake_up_all();
1921 wake_up_bit();
1922 wake_up_interruptible();
1923 wake_up_interruptible_all();
1924 wake_up_interruptible_nr();
1925 wake_up_interruptible_poll();
1926 wake_up_interruptible_sync();
1927 wake_up_interruptible_sync_poll();
1928 wake_up_locked();
1929 wake_up_locked_poll();
1930 wake_up_nr();
1931 wake_up_poll();
1932 wake_up_process();
1933
1934
1935[!] Note that the memory barriers implied by the sleeper and the waker do _not_
1936order multiple stores before the wake-up with respect to loads of those stored
1937values after the sleeper has called set_current_state(). For instance, if the
1938sleeper does:
1939
1940 set_current_state(TASK_INTERRUPTIBLE);
1941 if (event_indicated)
1942 break;
1943 __set_current_state(TASK_RUNNING);
1944 do_something(my_data);
1945
1946and the waker does:
1947
1948 my_data = value;
1949 event_indicated = 1;
1950 wake_up(&event_wait_queue);
1951
1952there's no guarantee that the change to event_indicated will be perceived by
1953the sleeper as coming after the change to my_data. In such a circumstance, the
1954code on both sides must interpolate its own memory barriers between the
1955separate data accesses. Thus the above sleeper ought to do:
1956
1957 set_current_state(TASK_INTERRUPTIBLE);
1958 if (event_indicated) {
1959 smp_rmb();
1960 do_something(my_data);
1961 }
1962
1963and the waker should do:
1964
1965 my_data = value;
1966 smp_wmb();
1967 event_indicated = 1;
1968 wake_up(&event_wait_queue);
1969
1970
David Howells108b42b2006-03-31 16:00:29 +01001971MISCELLANEOUS FUNCTIONS
1972-----------------------
1973
1974Other functions that imply barriers:
1975
1976 (*) schedule() and similar imply full memory barriers.
1977
David Howells108b42b2006-03-31 16:00:29 +01001978
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001979===================================
1980INTER-CPU ACQUIRING BARRIER EFFECTS
1981===================================
David Howells108b42b2006-03-31 16:00:29 +01001982
1983On SMP systems locking primitives give a more substantial form of barrier: one
1984that does affect memory access ordering on other CPUs, within the context of
1985conflict on any particular lock.
1986
1987
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001988ACQUIRES VS MEMORY ACCESSES
1989---------------------------
David Howells108b42b2006-03-31 16:00:29 +01001990
Aneesh Kumar79afecf2006-05-15 09:44:36 -07001991Consider the following: the system has a pair of spinlocks (M) and (Q), and
David Howells108b42b2006-03-31 16:00:29 +01001992three CPUs; then should the following sequence of events occur:
1993
1994 CPU 1 CPU 2
1995 =============================== ===============================
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08001996 ACCESS_ONCE(*A) = a; ACCESS_ONCE(*E) = e;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01001997 ACQUIRE M ACQUIRE Q
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08001998 ACCESS_ONCE(*B) = b; ACCESS_ONCE(*F) = f;
1999 ACCESS_ONCE(*C) = c; ACCESS_ONCE(*G) = g;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002000 RELEASE M RELEASE Q
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002001 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*H) = h;
David Howells108b42b2006-03-31 16:00:29 +01002002
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002003Then there is no guarantee as to what order CPU 3 will see the accesses to *A
David Howells108b42b2006-03-31 16:00:29 +01002004through *H occur in, other than the constraints imposed by the separate locks
2005on the separate CPUs. It might, for example, see:
2006
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002007 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
David Howells108b42b2006-03-31 16:00:29 +01002008
2009But it won't see any of:
2010
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002011 *B, *C or *D preceding ACQUIRE M
2012 *A, *B or *C following RELEASE M
2013 *F, *G or *H preceding ACQUIRE Q
2014 *E, *F or *G following RELEASE Q
David Howells108b42b2006-03-31 16:00:29 +01002015
2016
2017However, if the following occurs:
2018
2019 CPU 1 CPU 2
2020 =============================== ===============================
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002021 ACCESS_ONCE(*A) = a;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002022 ACQUIRE M [1]
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002023 ACCESS_ONCE(*B) = b;
2024 ACCESS_ONCE(*C) = c;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002025 RELEASE M [1]
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002026 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*E) = e;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002027 ACQUIRE M [2]
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08002028 smp_mb__after_unlock_lock();
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002029 ACCESS_ONCE(*F) = f;
2030 ACCESS_ONCE(*G) = g;
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002031 RELEASE M [2]
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002032 ACCESS_ONCE(*H) = h;
David Howells108b42b2006-03-31 16:00:29 +01002033
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002034CPU 3 might see:
David Howells108b42b2006-03-31 16:00:29 +01002035
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002036 *E, ACQUIRE M [1], *C, *B, *A, RELEASE M [1],
2037 ACQUIRE M [2], *H, *F, *G, RELEASE M [2], *D
David Howells108b42b2006-03-31 16:00:29 +01002038
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002039But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
David Howells108b42b2006-03-31 16:00:29 +01002040
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002041 *B, *C, *D, *F, *G or *H preceding ACQUIRE M [1]
2042 *A, *B or *C following RELEASE M [1]
2043 *F, *G or *H preceding ACQUIRE M [2]
2044 *A, *B, *C, *E, *F or *G following RELEASE M [2]
David Howells108b42b2006-03-31 16:00:29 +01002045
Paul E. McKenney17eb88e2013-12-11 13:59:09 -08002046Note that the smp_mb__after_unlock_lock() is critically important
2047here: Without it CPU 3 might see some of the above orderings.
2048Without smp_mb__after_unlock_lock(), the accesses are not guaranteed
2049to be seen in order unless CPU 3 holds lock M.
2050
David Howells108b42b2006-03-31 16:00:29 +01002051
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002052ACQUIRES VS I/O ACCESSES
2053------------------------
David Howells108b42b2006-03-31 16:00:29 +01002054
2055Under certain circumstances (especially involving NUMA), I/O accesses within
2056two spinlocked sections on two different CPUs may be seen as interleaved by the
2057PCI bridge, because the PCI bridge does not necessarily participate in the
2058cache-coherence protocol, and is therefore incapable of issuing the required
2059read memory barriers.
2060
2061For example:
2062
2063 CPU 1 CPU 2
2064 =============================== ===============================
2065 spin_lock(Q)
2066 writel(0, ADDR)
2067 writel(1, DATA);
2068 spin_unlock(Q);
2069 spin_lock(Q);
2070 writel(4, ADDR);
2071 writel(5, DATA);
2072 spin_unlock(Q);
2073
2074may be seen by the PCI bridge as follows:
2075
2076 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2077
2078which would probably cause the hardware to malfunction.
2079
2080
2081What is necessary here is to intervene with an mmiowb() before dropping the
2082spinlock, for example:
2083
2084 CPU 1 CPU 2
2085 =============================== ===============================
2086 spin_lock(Q)
2087 writel(0, ADDR)
2088 writel(1, DATA);
2089 mmiowb();
2090 spin_unlock(Q);
2091 spin_lock(Q);
2092 writel(4, ADDR);
2093 writel(5, DATA);
2094 mmiowb();
2095 spin_unlock(Q);
2096
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002097this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2098before either of the stores issued on CPU 2.
David Howells108b42b2006-03-31 16:00:29 +01002099
2100
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002101Furthermore, following a store by a load from the same device obviates the need
2102for the mmiowb(), because the load forces the store to complete before the load
David Howells108b42b2006-03-31 16:00:29 +01002103is performed:
2104
2105 CPU 1 CPU 2
2106 =============================== ===============================
2107 spin_lock(Q)
2108 writel(0, ADDR)
2109 a = readl(DATA);
2110 spin_unlock(Q);
2111 spin_lock(Q);
2112 writel(4, ADDR);
2113 b = readl(DATA);
2114 spin_unlock(Q);
2115
2116
2117See Documentation/DocBook/deviceiobook.tmpl for more information.
2118
2119
2120=================================
2121WHERE ARE MEMORY BARRIERS NEEDED?
2122=================================
2123
2124Under normal operation, memory operation reordering is generally not going to
2125be a problem as a single-threaded linear piece of code will still appear to
David Howells50fa6102009-04-28 15:01:38 +01002126work correctly, even if it's in an SMP kernel. There are, however, four
David Howells108b42b2006-03-31 16:00:29 +01002127circumstances in which reordering definitely _could_ be a problem:
2128
2129 (*) Interprocessor interaction.
2130
2131 (*) Atomic operations.
2132
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002133 (*) Accessing devices.
David Howells108b42b2006-03-31 16:00:29 +01002134
2135 (*) Interrupts.
2136
2137
2138INTERPROCESSOR INTERACTION
2139--------------------------
2140
2141When there's a system with more than one processor, more than one CPU in the
2142system may be working on the same data set at the same time. This can cause
2143synchronisation problems, and the usual way of dealing with them is to use
2144locks. Locks, however, are quite expensive, and so it may be preferable to
2145operate without the use of a lock if at all possible. In such a case
2146operations that affect both CPUs may have to be carefully ordered to prevent
2147a malfunction.
2148
2149Consider, for example, the R/W semaphore slow path. Here a waiting process is
2150queued on the semaphore, by virtue of it having a piece of its stack linked to
2151the semaphore's list of waiting processes:
2152
2153 struct rw_semaphore {
2154 ...
2155 spinlock_t lock;
2156 struct list_head waiters;
2157 };
2158
2159 struct rwsem_waiter {
2160 struct list_head list;
2161 struct task_struct *task;
2162 };
2163
2164To wake up a particular waiter, the up_read() or up_write() functions have to:
2165
2166 (1) read the next pointer from this waiter's record to know as to where the
2167 next waiter record is;
2168
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002169 (2) read the pointer to the waiter's task structure;
David Howells108b42b2006-03-31 16:00:29 +01002170
2171 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2172
2173 (4) call wake_up_process() on the task; and
2174
2175 (5) release the reference held on the waiter's task struct.
2176
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002177In other words, it has to perform this sequence of events:
David Howells108b42b2006-03-31 16:00:29 +01002178
2179 LOAD waiter->list.next;
2180 LOAD waiter->task;
2181 STORE waiter->task;
2182 CALL wakeup
2183 RELEASE task
2184
2185and if any of these steps occur out of order, then the whole thing may
2186malfunction.
2187
2188Once it has queued itself and dropped the semaphore lock, the waiter does not
2189get the lock again; it instead just waits for its task pointer to be cleared
2190before proceeding. Since the record is on the waiter's stack, this means that
2191if the task pointer is cleared _before_ the next pointer in the list is read,
2192another CPU might start processing the waiter and might clobber the waiter's
2193stack before the up*() function has a chance to read the next pointer.
2194
2195Consider then what might happen to the above sequence of events:
2196
2197 CPU 1 CPU 2
2198 =============================== ===============================
2199 down_xxx()
2200 Queue waiter
2201 Sleep
2202 up_yyy()
2203 LOAD waiter->task;
2204 STORE waiter->task;
2205 Woken up by other event
2206 <preempt>
2207 Resume processing
2208 down_xxx() returns
2209 call foo()
2210 foo() clobbers *waiter
2211 </preempt>
2212 LOAD waiter->list.next;
2213 --- OOPS ---
2214
2215This could be dealt with using the semaphore lock, but then the down_xxx()
2216function has to needlessly get the spinlock again after being woken up.
2217
2218The way to deal with this is to insert a general SMP memory barrier:
2219
2220 LOAD waiter->list.next;
2221 LOAD waiter->task;
2222 smp_mb();
2223 STORE waiter->task;
2224 CALL wakeup
2225 RELEASE task
2226
2227In this case, the barrier makes a guarantee that all memory accesses before the
2228barrier will appear to happen before all the memory accesses after the barrier
2229with respect to the other CPUs on the system. It does _not_ guarantee that all
2230the memory accesses before the barrier will be complete by the time the barrier
2231instruction itself is complete.
2232
2233On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2234compiler barrier, thus making sure the compiler emits the instructions in the
David Howells6bc39272006-06-25 05:49:22 -07002235right order without actually intervening in the CPU. Since there's only one
2236CPU, that CPU's dependency ordering logic will take care of everything else.
David Howells108b42b2006-03-31 16:00:29 +01002237
2238
2239ATOMIC OPERATIONS
2240-----------------
2241
David Howellsdbc87002006-04-10 22:54:23 -07002242Whilst they are technically interprocessor interaction considerations, atomic
2243operations are noted specially as some of them imply full memory barriers and
2244some don't, but they're very heavily relied on as a group throughout the
2245kernel.
2246
2247Any atomic operation that modifies some state in memory and returns information
2248about the state (old or new) implies an SMP-conditional general memory barrier
Nick Piggin26333572007-10-18 03:06:39 -07002249(smp_mb()) on each side of the actual operation (with the exception of
2250explicit lock operations, described later). These include:
David Howells108b42b2006-03-31 16:00:29 +01002251
2252 xchg();
2253 cmpxchg();
Paul E. McKenneyfb2b5812013-12-11 13:59:05 -08002254 atomic_xchg(); atomic_long_xchg();
2255 atomic_cmpxchg(); atomic_long_cmpxchg();
2256 atomic_inc_return(); atomic_long_inc_return();
2257 atomic_dec_return(); atomic_long_dec_return();
2258 atomic_add_return(); atomic_long_add_return();
2259 atomic_sub_return(); atomic_long_sub_return();
2260 atomic_inc_and_test(); atomic_long_inc_and_test();
2261 atomic_dec_and_test(); atomic_long_dec_and_test();
2262 atomic_sub_and_test(); atomic_long_sub_and_test();
2263 atomic_add_negative(); atomic_long_add_negative();
David Howellsdbc87002006-04-10 22:54:23 -07002264 test_and_set_bit();
2265 test_and_clear_bit();
2266 test_and_change_bit();
David Howells108b42b2006-03-31 16:00:29 +01002267
Paul E. McKenneyfb2b5812013-12-11 13:59:05 -08002268 /* when succeeds (returns 1) */
2269 atomic_add_unless(); atomic_long_add_unless();
2270
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002271These are used for such things as implementing ACQUIRE-class and RELEASE-class
David Howellsdbc87002006-04-10 22:54:23 -07002272operations and adjusting reference counters towards object destruction, and as
2273such the implicit memory barrier effects are necessary.
David Howells108b42b2006-03-31 16:00:29 +01002274
David Howells108b42b2006-03-31 16:00:29 +01002275
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002276The following operations are potential problems as they do _not_ imply memory
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002277barriers, but might be used for implementing such things as RELEASE-class
David Howellsdbc87002006-04-10 22:54:23 -07002278operations:
2279
2280 atomic_set();
David Howells108b42b2006-03-31 16:00:29 +01002281 set_bit();
2282 clear_bit();
2283 change_bit();
David Howellsdbc87002006-04-10 22:54:23 -07002284
2285With these the appropriate explicit memory barrier should be used if necessary
2286(smp_mb__before_clear_bit() for instance).
David Howells108b42b2006-03-31 16:00:29 +01002287
2288
David Howellsdbc87002006-04-10 22:54:23 -07002289The following also do _not_ imply memory barriers, and so may require explicit
2290memory barriers under some circumstances (smp_mb__before_atomic_dec() for
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002291instance):
David Howells108b42b2006-03-31 16:00:29 +01002292
2293 atomic_add();
2294 atomic_sub();
2295 atomic_inc();
2296 atomic_dec();
2297
2298If they're used for statistics generation, then they probably don't need memory
2299barriers, unless there's a coupling between statistical data.
2300
2301If they're used for reference counting on an object to control its lifetime,
2302they probably don't need memory barriers because either the reference count
2303will be adjusted inside a locked section, or the caller will already hold
2304sufficient references to make the lock, and thus a memory barrier unnecessary.
2305
2306If they're used for constructing a lock of some description, then they probably
2307do need memory barriers as a lock primitive generally has to do things in a
2308specific order.
2309
David Howells108b42b2006-03-31 16:00:29 +01002310Basically, each usage case has to be carefully considered as to whether memory
David Howellsdbc87002006-04-10 22:54:23 -07002311barriers are needed or not.
2312
Nick Piggin26333572007-10-18 03:06:39 -07002313The following operations are special locking primitives:
2314
2315 test_and_set_bit_lock();
2316 clear_bit_unlock();
2317 __clear_bit_unlock();
2318
Peter Zijlstra2e4f5382013-11-06 14:57:36 +01002319These implement ACQUIRE-class and RELEASE-class operations. These should be used in
Nick Piggin26333572007-10-18 03:06:39 -07002320preference to other operations when implementing locking primitives, because
2321their implementations can be optimised on many architectures.
2322
David Howellsdbc87002006-04-10 22:54:23 -07002323[!] Note that special memory barrier primitives are available for these
2324situations because on some CPUs the atomic instructions used imply full memory
2325barriers, and so barrier instructions are superfluous in conjunction with them,
2326and in such cases the special barrier primitives will be no-ops.
David Howells108b42b2006-03-31 16:00:29 +01002327
2328See Documentation/atomic_ops.txt for more information.
2329
2330
2331ACCESSING DEVICES
2332-----------------
2333
2334Many devices can be memory mapped, and so appear to the CPU as if they're just
2335a set of memory locations. To control such a device, the driver usually has to
2336make the right memory accesses in exactly the right order.
2337
2338However, having a clever CPU or a clever compiler creates a potential problem
2339in that the carefully sequenced accesses in the driver code won't reach the
2340device in the requisite order if the CPU or the compiler thinks it is more
2341efficient to reorder, combine or merge accesses - something that would cause
2342the device to malfunction.
2343
2344Inside of the Linux kernel, I/O should be done through the appropriate accessor
2345routines - such as inb() or writel() - which know how to make such accesses
2346appropriately sequential. Whilst this, for the most part, renders the explicit
2347use of memory barriers unnecessary, there are a couple of situations where they
2348might be needed:
2349
2350 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2351 so for _all_ general drivers locks should be used and mmiowb() must be
2352 issued prior to unlocking the critical section.
2353
2354 (2) If the accessor functions are used to refer to an I/O memory window with
2355 relaxed memory access properties, then _mandatory_ memory barriers are
2356 required to enforce ordering.
2357
2358See Documentation/DocBook/deviceiobook.tmpl for more information.
2359
2360
2361INTERRUPTS
2362----------
2363
2364A driver may be interrupted by its own interrupt service routine, and thus the
2365two parts of the driver may interfere with each other's attempts to control or
2366access the device.
2367
2368This may be alleviated - at least in part - by disabling local interrupts (a
2369form of locking), such that the critical operations are all contained within
2370the interrupt-disabled section in the driver. Whilst the driver's interrupt
2371routine is executing, the driver's core may not run on the same CPU, and its
2372interrupt is not permitted to happen again until the current interrupt has been
2373handled, thus the interrupt handler does not need to lock against that.
2374
2375However, consider a driver that was talking to an ethernet card that sports an
2376address register and a data register. If that driver's core talks to the card
2377under interrupt-disablement and then the driver's interrupt handler is invoked:
2378
2379 LOCAL IRQ DISABLE
2380 writew(ADDR, 3);
2381 writew(DATA, y);
2382 LOCAL IRQ ENABLE
2383 <interrupt>
2384 writew(ADDR, 4);
2385 q = readw(DATA);
2386 </interrupt>
2387
2388The store to the data register might happen after the second store to the
2389address register if ordering rules are sufficiently relaxed:
2390
2391 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2392
2393
2394If ordering rules are relaxed, it must be assumed that accesses done inside an
2395interrupt disabled section may leak outside of it and may interleave with
2396accesses performed in an interrupt - and vice versa - unless implicit or
2397explicit barriers are used.
2398
2399Normally this won't be a problem because the I/O accesses done inside such
2400sections will include synchronous load operations on strictly ordered I/O
2401registers that form implicit I/O barriers. If this isn't sufficient then an
2402mmiowb() may need to be used explicitly.
2403
2404
2405A similar situation may occur between an interrupt routine and two routines
2406running on separate CPUs that communicate with each other. If such a case is
2407likely, then interrupt-disabling locks should be used to guarantee ordering.
2408
2409
2410==========================
2411KERNEL I/O BARRIER EFFECTS
2412==========================
2413
2414When accessing I/O memory, drivers should use the appropriate accessor
2415functions:
2416
2417 (*) inX(), outX():
2418
2419 These are intended to talk to I/O space rather than memory space, but
2420 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
2421 indeed have special I/O space access cycles and instructions, but many
2422 CPUs don't have such a concept.
2423
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002424 The PCI bus, amongst others, defines an I/O space concept which - on such
2425 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
David Howells6bc39272006-06-25 05:49:22 -07002426 space. However, it may also be mapped as a virtual I/O space in the CPU's
2427 memory map, particularly on those CPUs that don't support alternate I/O
2428 spaces.
David Howells108b42b2006-03-31 16:00:29 +01002429
2430 Accesses to this space may be fully synchronous (as on i386), but
2431 intermediary bridges (such as the PCI host bridge) may not fully honour
2432 that.
2433
2434 They are guaranteed to be fully ordered with respect to each other.
2435
2436 They are not guaranteed to be fully ordered with respect to other types of
2437 memory and I/O operation.
2438
2439 (*) readX(), writeX():
2440
2441 Whether these are guaranteed to be fully ordered and uncombined with
2442 respect to each other on the issuing CPU depends on the characteristics
2443 defined for the memory window through which they're accessing. On later
2444 i386 architecture machines, for example, this is controlled by way of the
2445 MTRR registers.
2446
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002447 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
David Howells108b42b2006-03-31 16:00:29 +01002448 provided they're not accessing a prefetchable device.
2449
2450 However, intermediary hardware (such as a PCI bridge) may indulge in
2451 deferral if it so wishes; to flush a store, a load from the same location
2452 is preferred[*], but a load from the same device or from configuration
2453 space should suffice for PCI.
2454
2455 [*] NOTE! attempting to load from the same location as was written to may
Ingo Molnare0edc782013-11-22 11:24:53 +01002456 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2457 example.
David Howells108b42b2006-03-31 16:00:29 +01002458
2459 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2460 force stores to be ordered.
2461
2462 Please refer to the PCI specification for more information on interactions
2463 between PCI transactions.
2464
2465 (*) readX_relaxed()
2466
2467 These are similar to readX(), but are not guaranteed to be ordered in any
2468 way. Be aware that there is no I/O read barrier available.
2469
2470 (*) ioreadX(), iowriteX()
2471
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002472 These will perform appropriately for the type of access they're actually
David Howells108b42b2006-03-31 16:00:29 +01002473 doing, be it inX()/outX() or readX()/writeX().
2474
2475
2476========================================
2477ASSUMED MINIMUM EXECUTION ORDERING MODEL
2478========================================
2479
2480It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2481maintain the appearance of program causality with respect to itself. Some CPUs
2482(such as i386 or x86_64) are more constrained than others (such as powerpc or
2483frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2484of arch-specific code.
2485
2486This means that it must be considered that the CPU will execute its instruction
2487stream in any order it feels like - or even in parallel - provided that if an
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002488instruction in the stream depends on an earlier instruction, then that
David Howells108b42b2006-03-31 16:00:29 +01002489earlier instruction must be sufficiently complete[*] before the later
2490instruction may proceed; in other words: provided that the appearance of
2491causality is maintained.
2492
2493 [*] Some instructions have more than one effect - such as changing the
2494 condition codes, changing registers or changing memory - and different
2495 instructions may depend on different effects.
2496
2497A CPU may also discard any instruction sequence that winds up having no
2498ultimate effect. For example, if two adjacent instructions both load an
2499immediate value into the same register, the first may be discarded.
2500
2501
2502Similarly, it has to be assumed that compiler might reorder the instruction
2503stream in any way it sees fit, again provided the appearance of causality is
2504maintained.
2505
2506
2507============================
2508THE EFFECTS OF THE CPU CACHE
2509============================
2510
2511The way cached memory operations are perceived across the system is affected to
2512a certain extent by the caches that lie between CPUs and memory, and by the
2513memory coherence system that maintains the consistency of state in the system.
2514
2515As far as the way a CPU interacts with another part of the system through the
2516caches goes, the memory system has to include the CPU's caches, and memory
2517barriers for the most part act at the interface between the CPU and its cache
2518(memory barriers logically act on the dotted line in the following diagram):
2519
2520 <--- CPU ---> : <----------- Memory ----------->
2521 :
2522 +--------+ +--------+ : +--------+ +-----------+
2523 | | | | : | | | | +--------+
Ingo Molnare0edc782013-11-22 11:24:53 +01002524 | CPU | | Memory | : | CPU | | | | |
2525 | Core |--->| Access |----->| Cache |<-->| | | |
David Howells108b42b2006-03-31 16:00:29 +01002526 | | | Queue | : | | | |--->| Memory |
Ingo Molnare0edc782013-11-22 11:24:53 +01002527 | | | | : | | | | | |
2528 +--------+ +--------+ : +--------+ | | | |
David Howells108b42b2006-03-31 16:00:29 +01002529 : | Cache | +--------+
2530 : | Coherency |
2531 : | Mechanism | +--------+
2532 +--------+ +--------+ : +--------+ | | | |
2533 | | | | : | | | | | |
2534 | CPU | | Memory | : | CPU | | |--->| Device |
Ingo Molnare0edc782013-11-22 11:24:53 +01002535 | Core |--->| Access |----->| Cache |<-->| | | |
2536 | | | Queue | : | | | | | |
David Howells108b42b2006-03-31 16:00:29 +01002537 | | | | : | | | | +--------+
2538 +--------+ +--------+ : +--------+ +-----------+
2539 :
2540 :
2541
2542Although any particular load or store may not actually appear outside of the
2543CPU that issued it since it may have been satisfied within the CPU's own cache,
2544it will still appear as if the full memory access had taken place as far as the
2545other CPUs are concerned since the cache coherency mechanisms will migrate the
2546cacheline over to the accessing CPU and propagate the effects upon conflict.
2547
2548The CPU core may execute instructions in any order it deems fit, provided the
2549expected program causality appears to be maintained. Some of the instructions
2550generate load and store operations which then go into the queue of memory
2551accesses to be performed. The core may place these in the queue in any order
2552it wishes, and continue execution until it is forced to wait for an instruction
2553to complete.
2554
2555What memory barriers are concerned with is controlling the order in which
2556accesses cross from the CPU side of things to the memory side of things, and
2557the order in which the effects are perceived to happen by the other observers
2558in the system.
2559
2560[!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2561their own loads and stores as if they had happened in program order.
2562
2563[!] MMIO or other device accesses may bypass the cache system. This depends on
2564the properties of the memory window through which devices are accessed and/or
2565the use of any special device communication instructions the CPU may have.
2566
2567
2568CACHE COHERENCY
2569---------------
2570
2571Life isn't quite as simple as it may appear above, however: for while the
2572caches are expected to be coherent, there's no guarantee that that coherency
2573will be ordered. This means that whilst changes made on one CPU will
2574eventually become visible on all CPUs, there's no guarantee that they will
2575become apparent in the same order on those other CPUs.
2576
2577
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002578Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2579has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
David Howells108b42b2006-03-31 16:00:29 +01002580
2581 :
2582 : +--------+
2583 : +---------+ | |
2584 +--------+ : +--->| Cache A |<------->| |
2585 | | : | +---------+ | |
2586 | CPU 1 |<---+ | |
2587 | | : | +---------+ | |
2588 +--------+ : +--->| Cache B |<------->| |
2589 : +---------+ | |
2590 : | Memory |
2591 : +---------+ | System |
2592 +--------+ : +--->| Cache C |<------->| |
2593 | | : | +---------+ | |
2594 | CPU 2 |<---+ | |
2595 | | : | +---------+ | |
2596 +--------+ : +--->| Cache D |<------->| |
2597 : +---------+ | |
2598 : +--------+
2599 :
2600
2601Imagine the system has the following properties:
2602
2603 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2604 resident in memory;
2605
2606 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2607 resident in memory;
2608
2609 (*) whilst the CPU core is interrogating one cache, the other cache may be
2610 making use of the bus to access the rest of the system - perhaps to
2611 displace a dirty cacheline or to do a speculative load;
2612
2613 (*) each cache has a queue of operations that need to be applied to that cache
2614 to maintain coherency with the rest of the system;
2615
2616 (*) the coherency queue is not flushed by normal loads to lines already
2617 present in the cache, even though the contents of the queue may
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002618 potentially affect those loads.
David Howells108b42b2006-03-31 16:00:29 +01002619
2620Imagine, then, that two writes are made on the first CPU, with a write barrier
2621between them to guarantee that they will appear to reach that CPU's caches in
2622the requisite order:
2623
2624 CPU 1 CPU 2 COMMENT
2625 =============== =============== =======================================
2626 u == 0, v == 1 and p == &u, q == &u
2627 v = 2;
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002628 smp_wmb(); Make sure change to v is visible before
David Howells108b42b2006-03-31 16:00:29 +01002629 change to p
2630 <A:modify v=2> v is now in cache A exclusively
2631 p = &v;
2632 <B:modify p=&v> p is now in cache B exclusively
2633
2634The write memory barrier forces the other CPUs in the system to perceive that
2635the local CPU's caches have apparently been updated in the correct order. But
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002636now imagine that the second CPU wants to read those values:
David Howells108b42b2006-03-31 16:00:29 +01002637
2638 CPU 1 CPU 2 COMMENT
2639 =============== =============== =======================================
2640 ...
2641 q = p;
2642 x = *q;
2643
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002644The above pair of reads may then fail to happen in the expected order, as the
David Howells108b42b2006-03-31 16:00:29 +01002645cacheline holding p may get updated in one of the second CPU's caches whilst
2646the update to the cacheline holding v is delayed in the other of the second
2647CPU's caches by some other cache event:
2648
2649 CPU 1 CPU 2 COMMENT
2650 =============== =============== =======================================
2651 u == 0, v == 1 and p == &u, q == &u
2652 v = 2;
2653 smp_wmb();
2654 <A:modify v=2> <C:busy>
2655 <C:queue v=2>
Aneesh Kumar79afecf2006-05-15 09:44:36 -07002656 p = &v; q = p;
David Howells108b42b2006-03-31 16:00:29 +01002657 <D:request p>
2658 <B:modify p=&v> <D:commit p=&v>
Ingo Molnare0edc782013-11-22 11:24:53 +01002659 <D:read p>
David Howells108b42b2006-03-31 16:00:29 +01002660 x = *q;
2661 <C:read *q> Reads from v before v updated in cache
2662 <C:unbusy>
2663 <C:commit v=2>
2664
2665Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2666no guarantee that, without intervention, the order of update will be the same
2667as that committed on CPU 1.
2668
2669
2670To intervene, we need to interpolate a data dependency barrier or a read
2671barrier between the loads. This will force the cache to commit its coherency
2672queue before processing any further requests:
2673
2674 CPU 1 CPU 2 COMMENT
2675 =============== =============== =======================================
2676 u == 0, v == 1 and p == &u, q == &u
2677 v = 2;
2678 smp_wmb();
2679 <A:modify v=2> <C:busy>
2680 <C:queue v=2>
Paolo 'Blaisorblade' Giarrusso3fda9822006-10-19 23:28:19 -07002681 p = &v; q = p;
David Howells108b42b2006-03-31 16:00:29 +01002682 <D:request p>
2683 <B:modify p=&v> <D:commit p=&v>
Ingo Molnare0edc782013-11-22 11:24:53 +01002684 <D:read p>
David Howells108b42b2006-03-31 16:00:29 +01002685 smp_read_barrier_depends()
2686 <C:unbusy>
2687 <C:commit v=2>
2688 x = *q;
2689 <C:read *q> Reads from v after v updated in cache
2690
2691
2692This sort of problem can be encountered on DEC Alpha processors as they have a
2693split cache that improves performance by making better use of the data bus.
2694Whilst most CPUs do imply a data dependency barrier on the read when a memory
2695access depends on a read, not all do, so it may not be relied on.
2696
2697Other CPUs may also have split caches, but must coordinate between the various
Matt LaPlante3f6dee92006-10-03 22:45:33 +02002698cachelets for normal memory accesses. The semantics of the Alpha removes the
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002699need for coordination in the absence of memory barriers.
David Howells108b42b2006-03-31 16:00:29 +01002700
2701
2702CACHE COHERENCY VS DMA
2703----------------------
2704
2705Not all systems maintain cache coherency with respect to devices doing DMA. In
2706such cases, a device attempting DMA may obtain stale data from RAM because
2707dirty cache lines may be resident in the caches of various CPUs, and may not
2708have been written back to RAM yet. To deal with this, the appropriate part of
2709the kernel must flush the overlapping bits of cache on each CPU (and maybe
2710invalidate them as well).
2711
2712In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2713cache lines being written back to RAM from a CPU's cache after the device has
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002714installed its own data, or cache lines present in the CPU's cache may simply
2715obscure the fact that RAM has been updated, until at such time as the cacheline
2716is discarded from the CPU's cache and reloaded. To deal with this, the
2717appropriate part of the kernel must invalidate the overlapping bits of the
David Howells108b42b2006-03-31 16:00:29 +01002718cache on each CPU.
2719
2720See Documentation/cachetlb.txt for more information on cache management.
2721
2722
2723CACHE COHERENCY VS MMIO
2724-----------------------
2725
2726Memory mapped I/O usually takes place through memory locations that are part of
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002727a window in the CPU's memory space that has different properties assigned than
David Howells108b42b2006-03-31 16:00:29 +01002728the usual RAM directed window.
2729
2730Amongst these properties is usually the fact that such accesses bypass the
2731caching entirely and go directly to the device buses. This means MMIO accesses
2732may, in effect, overtake accesses to cached memory that were emitted earlier.
2733A memory barrier isn't sufficient in such a case, but rather the cache must be
2734flushed between the cached memory write and the MMIO access if the two are in
2735any way dependent.
2736
2737
2738=========================
2739THE THINGS CPUS GET UP TO
2740=========================
2741
2742A programmer might take it for granted that the CPU will perform memory
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002743operations in exactly the order specified, so that if the CPU is, for example,
David Howells108b42b2006-03-31 16:00:29 +01002744given the following piece of code to execute:
2745
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002746 a = ACCESS_ONCE(*A);
2747 ACCESS_ONCE(*B) = b;
2748 c = ACCESS_ONCE(*C);
2749 d = ACCESS_ONCE(*D);
2750 ACCESS_ONCE(*E) = e;
David Howells108b42b2006-03-31 16:00:29 +01002751
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002752they would then expect that the CPU will complete the memory operation for each
David Howells108b42b2006-03-31 16:00:29 +01002753instruction before moving on to the next one, leading to a definite sequence of
2754operations as seen by external observers in the system:
2755
2756 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2757
2758
2759Reality is, of course, much messier. With many CPUs and compilers, the above
2760assumption doesn't hold because:
2761
2762 (*) loads are more likely to need to be completed immediately to permit
2763 execution progress, whereas stores can often be deferred without a
2764 problem;
2765
2766 (*) loads may be done speculatively, and the result discarded should it prove
2767 to have been unnecessary;
2768
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002769 (*) loads may be done speculatively, leading to the result having been fetched
2770 at the wrong time in the expected sequence of events;
David Howells108b42b2006-03-31 16:00:29 +01002771
2772 (*) the order of the memory accesses may be rearranged to promote better use
2773 of the CPU buses and caches;
2774
2775 (*) loads and stores may be combined to improve performance when talking to
2776 memory or I/O hardware that can do batched accesses of adjacent locations,
2777 thus cutting down on transaction setup costs (memory and PCI devices may
2778 both be able to do this); and
2779
2780 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2781 mechanisms may alleviate this - once the store has actually hit the cache
2782 - there's no guarantee that the coherency management will be propagated in
2783 order to other CPUs.
2784
2785So what another CPU, say, might actually observe from the above piece of code
2786is:
2787
2788 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2789
2790 (Where "LOAD {*C,*D}" is a combined load)
2791
2792
2793However, it is guaranteed that a CPU will be self-consistent: it will see its
2794_own_ accesses appear to be correctly ordered, without the need for a memory
2795barrier. For instance with the following code:
2796
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002797 U = ACCESS_ONCE(*A);
2798 ACCESS_ONCE(*A) = V;
2799 ACCESS_ONCE(*A) = W;
2800 X = ACCESS_ONCE(*A);
2801 ACCESS_ONCE(*A) = Y;
2802 Z = ACCESS_ONCE(*A);
David Howells108b42b2006-03-31 16:00:29 +01002803
2804and assuming no intervention by an external influence, it can be assumed that
2805the final result will appear to be:
2806
2807 U == the original value of *A
2808 X == W
2809 Z == Y
2810 *A == Y
2811
2812The code above may cause the CPU to generate the full sequence of memory
2813accesses:
2814
2815 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2816
2817in that order, but, without intervention, the sequence may have almost any
2818combination of elements combined or discarded, provided the program's view of
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002819the world remains consistent. Note that ACCESS_ONCE() is -not- optional
2820in the above example, as there are architectures where a given CPU might
Paul E. McKenney8dd853d2014-02-23 08:34:24 -08002821reorder successive loads to the same location. On such architectures,
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002822ACCESS_ONCE() does whatever is necessary to prevent this, for example, on
2823Itanium the volatile casts used by ACCESS_ONCE() cause GCC to emit the
2824special ld.acq and st.rel instructions that prevent such reordering.
David Howells108b42b2006-03-31 16:00:29 +01002825
2826The compiler may also combine, discard or defer elements of the sequence before
2827the CPU even sees them.
2828
2829For instance:
2830
2831 *A = V;
2832 *A = W;
2833
2834may be reduced to:
2835
2836 *A = W;
2837
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002838since, without either a write barrier or an ACCESS_ONCE(), it can be
2839assumed that the effect of the storage of V to *A is lost. Similarly:
David Howells108b42b2006-03-31 16:00:29 +01002840
2841 *A = Y;
2842 Z = *A;
2843
Paul E. McKenney2ecf8102013-12-11 13:59:04 -08002844may, without a memory barrier or an ACCESS_ONCE(), be reduced to:
David Howells108b42b2006-03-31 16:00:29 +01002845
2846 *A = Y;
2847 Z = Y;
2848
2849and the LOAD operation never appear outside of the CPU.
2850
2851
2852AND THEN THERE'S THE ALPHA
2853--------------------------
2854
2855The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2856some versions of the Alpha CPU have a split data cache, permitting them to have
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002857two semantically-related cache lines updated at separate times. This is where
David Howells108b42b2006-03-31 16:00:29 +01002858the data dependency barrier really becomes necessary as this synchronises both
2859caches with the memory coherence system, thus making it seem like pointer
2860changes vs new data occur in the right order.
2861
Jarek Poplawski81fc6322007-05-23 13:58:20 -07002862The Alpha defines the Linux kernel's memory barrier model.
David Howells108b42b2006-03-31 16:00:29 +01002863
2864See the subsection on "Cache Coherency" above.
2865
2866
David Howells90fddab2010-03-24 09:43:00 +00002867============
2868EXAMPLE USES
2869============
2870
2871CIRCULAR BUFFERS
2872----------------
2873
2874Memory barriers can be used to implement circular buffering without the need
2875of a lock to serialise the producer with the consumer. See:
2876
2877 Documentation/circular-buffers.txt
2878
2879for details.
2880
2881
David Howells108b42b2006-03-31 16:00:29 +01002882==========
2883REFERENCES
2884==========
2885
2886Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2887Digital Press)
2888 Chapter 5.2: Physical Address Space Characteristics
2889 Chapter 5.4: Caches and Write Buffers
2890 Chapter 5.5: Data Sharing
2891 Chapter 5.6: Read/Write Ordering
2892
2893AMD64 Architecture Programmer's Manual Volume 2: System Programming
2894 Chapter 7.1: Memory-Access Ordering
2895 Chapter 7.4: Buffering and Combining Memory Writes
2896
2897IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2898System Programming Guide
2899 Chapter 7.1: Locked Atomic Operations
2900 Chapter 7.2: Memory Ordering
2901 Chapter 7.4: Serializing Instructions
2902
2903The SPARC Architecture Manual, Version 9
2904 Chapter 8: Memory Models
2905 Appendix D: Formal Specification of the Memory Models
2906 Appendix J: Programming with the Memory Models
2907
2908UltraSPARC Programmer Reference Manual
2909 Chapter 5: Memory Accesses and Cacheability
2910 Chapter 15: Sparc-V9 Memory Models
2911
2912UltraSPARC III Cu User's Manual
2913 Chapter 9: Memory Models
2914
2915UltraSPARC IIIi Processor User's Manual
2916 Chapter 8: Memory Models
2917
2918UltraSPARC Architecture 2005
2919 Chapter 9: Memory
2920 Appendix D: Formal Specifications of the Memory Models
2921
2922UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2923 Chapter 8: Memory Models
2924 Appendix F: Caches and Cache Coherency
2925
2926Solaris Internals, Core Kernel Architecture, p63-68:
2927 Chapter 3.3: Hardware Considerations for Locks and
2928 Synchronization
2929
2930Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2931for Kernel Programmers:
2932 Chapter 13: Other Memory Models
2933
2934Intel Itanium Architecture Software Developer's Manual: Volume 1:
2935 Section 2.6: Speculation
2936 Section 4.4: Memory Access