]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - Documentation/memory-barriers.txt
Merge branch 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[mirror_ubuntu-zesty-kernel.git] / Documentation / memory-barriers.txt
1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
4
5 By: David Howells <dhowells@redhat.com>
6 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
7
8 Contents:
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.
23 - Read memory barriers vs load speculation.
24 - Transitivity
25
26 (*) Explicit kernel barriers.
27
28 - Compiler barrier.
29 - CPU memory barriers.
30 - MMIO write barrier.
31
32 (*) Implicit kernel memory barriers.
33
34 - Locking functions.
35 - Interrupt disabling functions.
36 - Sleep and wake-up functions.
37 - 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
65 (*) Example uses.
66
67 - Circular buffers.
68
69 (*) References.
70
71
72 ============================
73 ABSTRACT MEMORY ACCESS MODEL
74 ============================
75
76 Consider 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
101 Each CPU executes a program that generates memory access operations. In the
102 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
103 perform the memory operations in any order it likes, provided program causality
104 appears to be maintained. Similarly, the compiler may also arrange the
105 instructions it emits in any order it likes, provided it doesn't affect the
106 apparent operation of the program.
107
108 So in the above diagram, the effects of the memory operations performed by a
109 CPU are perceived by the rest of the system as the operations cross the
110 interface between the CPU and rest of the system (the dotted lines).
111
112
113 For example, consider the following sequence of events:
114
115 CPU 1 CPU 2
116 =============== ===============
117 { A == 1; B == 2 }
118 A = 3; x = B;
119 B = 4; y = A;
120
121 The set of accesses as seen by the memory system in the middle can be arranged
122 in 24 different combinations:
123
124 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4
125 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3
126 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4
127 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4
128 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3
129 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4
130 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4
131 STORE B=4, ...
132 ...
133
134 and can thus result in four different combinations of values:
135
136 x == 2, y == 1
137 x == 2, y == 3
138 x == 4, y == 1
139 x == 4, y == 3
140
141
142 Furthermore, the stores committed by a CPU to the memory system may not be
143 perceived by the loads made by another CPU in the same order as the stores were
144 committed.
145
146
147 As 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
155 There is an obvious data dependency here, as the value loaded into D depends on
156 the address retrieved from P by CPU 2. At the end of the sequence, any of the
157 following results are possible:
158
159 (Q == &A) and (D == 1)
160 (Q == &B) and (D == 2)
161 (Q == &B) and (D == 4)
162
163 Note that CPU 2 will never try and load C into D because the CPU will load P
164 into Q before issuing the load of *Q.
165
166
167 DEVICE OPERATIONS
168 -----------------
169
170 Some devices present their control interfaces as collections of memory
171 locations, but the order in which the control registers are accessed is very
172 important. For instance, imagine an ethernet card with a set of internal
173 registers that are accessed through an address port register (A) and a data
174 port register (D). To read internal register 5, the following code might then
175 be used:
176
177 *A = 5;
178 x = *D;
179
180 but 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
185 the second of which will almost certainly result in a malfunction, since it set
186 the address _after_ attempting to read the register.
187
188
189 GUARANTEES
190 ----------
191
192 There 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
197 ACCESS_ONCE(Q) = P; smp_read_barrier_depends(); D = ACCESS_ONCE(*Q);
198
199 the CPU will issue the following memory operations:
200
201 Q = LOAD P, D = LOAD *Q
202
203 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().
208
209 (*) Overlapping loads and stores within a particular CPU will appear to be
210 ordered within that CPU. This means that for:
211
212 a = ACCESS_ONCE(*X); ACCESS_ONCE(*X) = b;
213
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
220 ACCESS_ONCE(*X) = c; d = ACCESS_ONCE(*X);
221
222 the CPU will only issue:
223
224 STORE *X = c, d = LOAD *X
225
226 (Loads and stores overlap if they are targeted at overlapping pieces of
227 memory).
228
229 And there are a number of things that _must_ or _must_not_ be assumed:
230
231 (*) 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
234 of "creative" transformations, which are covered in the Compiler
235 Barrier section.
236
237 (*) 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
264 *A = X; *(A + 4) = Y;
265
266 we may get any of:
267
268 STORE *A = X; STORE *(A + 4) = Y;
269 STORE *(A + 4) = Y; STORE *A = X;
270 STORE {*A, *(A + 4) } = {X, Y};
271
272 And there are anti-guarantees:
273
274 (*) These guarantees do not apply to bitfields, because compilers often
275 generate code to modify these using non-atomic read-modify-write
276 sequences. Do not attempt to use bitfields to synchronize parallel
277 algorithms.
278
279 (*) Even in cases where bitfields are protected by locks, all fields
280 in a given bitfield must be protected by one lock. If two fields
281 in a given bitfield are protected by different locks, the compiler's
282 non-atomic read-modify-write sequences can cause an update to one
283 field to corrupt the value of an adjacent field.
284
285 (*) These guarantees apply only to properly aligned and sized scalar
286 variables. "Properly sized" currently means variables that are
287 the same size as "char", "short", "int" and "long". "Properly
288 aligned" means the natural alignment, thus no constraints for
289 "char", two-byte alignment for "short", four-byte alignment for
290 "int", and either four-byte or eight-byte alignment for "long",
291 on 32-bit and 64-bit systems, respectively. Note that these
292 guarantees were introduced into the C11 standard, so beware when
293 using older pre-C11 compilers (for example, gcc 4.6). The portion
294 of the standard containing this guarantee is Section 3.14, which
295 defines "memory location" as follows:
296
297 memory location
298 either an object of scalar type, or a maximal sequence
299 of adjacent bit-fields all having nonzero width
300
301 NOTE 1: Two threads of execution can update and access
302 separate memory locations without interfering with
303 each other.
304
305 NOTE 2: A bit-field and an adjacent non-bit-field member
306 are in separate memory locations. The same applies
307 to two bit-fields, if one is declared inside a nested
308 structure declaration and the other is not, or if the two
309 are separated by a zero-length bit-field declaration,
310 or if they are separated by a non-bit-field member
311 declaration. It is not safe to concurrently update two
312 bit-fields in the same structure if all members declared
313 between them are also bit-fields, no matter what the
314 sizes of those intervening bit-fields happen to be.
315
316
317 =========================
318 WHAT ARE MEMORY BARRIERS?
319 =========================
320
321 As can be seen above, independent memory operations are effectively performed
322 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
323 What is required is some way of intervening to instruct the compiler and the
324 CPU to restrict the order.
325
326 Memory barriers are such interventions. They impose a perceived partial
327 ordering over the memory operations on either side of the barrier.
328
329 Such enforcement is important because the CPUs and other devices in a system
330 can use a variety of tricks to improve performance, including reordering,
331 deferral and combination of memory operations; speculative loads; speculative
332 branch prediction and various types of caching. Memory barriers are used to
333 override or suppress these tricks, allowing the code to sanely control the
334 interaction of multiple CPUs and/or devices.
335
336
337 VARIETIES OF MEMORY BARRIER
338 ---------------------------
339
340 Memory barriers come in four basic varieties:
341
342 (1) Write (or store) memory barriers.
343
344 A write memory barrier gives a guarantee that all the STORE operations
345 specified before the barrier will appear to happen before all the STORE
346 operations specified after the barrier with respect to the other
347 components of the system.
348
349 A write barrier is a partial ordering on stores only; it is not required
350 to have any effect on loads.
351
352 A CPU can be viewed as committing a sequence of store operations to the
353 memory system as time progresses. All stores before a write barrier will
354 occur in the sequence _before_ all the stores after the write barrier.
355
356 [!] Note that write barriers should normally be paired with read or data
357 dependency barriers; see the "SMP barrier pairing" subsection.
358
359
360 (2) Data dependency barriers.
361
362 A data dependency barrier is a weaker form of read barrier. In the case
363 where two loads are performed such that the second depends on the result
364 of the first (eg: the first load retrieves the address to which the second
365 load will be directed), a data dependency barrier would be required to
366 make sure that the target of the second load is updated before the address
367 obtained by the first load is accessed.
368
369 A data dependency barrier is a partial ordering on interdependent loads
370 only; it is not required to have any effect on stores, independent loads
371 or overlapping loads.
372
373 As mentioned in (1), the other CPUs in the system can be viewed as
374 committing sequences of stores to the memory system that the CPU being
375 considered can then perceive. A data dependency barrier issued by the CPU
376 under consideration guarantees that for any load preceding it, if that
377 load touches one of a sequence of stores from another CPU, then by the
378 time the barrier completes, the effects of all the stores prior to that
379 touched by the load will be perceptible to any loads issued after the data
380 dependency barrier.
381
382 See the "Examples of memory barrier sequences" subsection for diagrams
383 showing the ordering constraints.
384
385 [!] Note that the first load really has to have a _data_ dependency and
386 not a control dependency. If the address for the second load is dependent
387 on the first load, but the dependency is through a conditional rather than
388 actually loading the address itself, then it's a _control_ dependency and
389 a full read barrier or better is required. See the "Control dependencies"
390 subsection for more information.
391
392 [!] Note that data dependency barriers should normally be paired with
393 write barriers; see the "SMP barrier pairing" subsection.
394
395
396 (3) Read (or load) memory barriers.
397
398 A read barrier is a data dependency barrier plus a guarantee that all the
399 LOAD operations specified before the barrier will appear to happen before
400 all the LOAD operations specified after the barrier with respect to the
401 other components of the system.
402
403 A read barrier is a partial ordering on loads only; it is not required to
404 have any effect on stores.
405
406 Read memory barriers imply data dependency barriers, and so can substitute
407 for them.
408
409 [!] Note that read barriers should normally be paired with write barriers;
410 see the "SMP barrier pairing" subsection.
411
412
413 (4) General memory barriers.
414
415 A general memory barrier gives a guarantee that all the LOAD and STORE
416 operations specified before the barrier will appear to happen before all
417 the LOAD and STORE operations specified after the barrier with respect to
418 the other components of the system.
419
420 A general memory barrier is a partial ordering over both loads and stores.
421
422 General memory barriers imply both read and write memory barriers, and so
423 can substitute for either.
424
425
426 And a couple of implicit varieties:
427
428 (5) ACQUIRE operations.
429
430 This acts as a one-way permeable barrier. It guarantees that all memory
431 operations after the ACQUIRE operation will appear to happen after the
432 ACQUIRE operation with respect to the other components of the system.
433 ACQUIRE operations include LOCK operations and smp_load_acquire()
434 operations.
435
436 Memory operations that occur before an ACQUIRE operation may appear to
437 happen after it completes.
438
439 An ACQUIRE operation should almost always be paired with a RELEASE
440 operation.
441
442
443 (6) RELEASE operations.
444
445 This also acts as a one-way permeable barrier. It guarantees that all
446 memory operations before the RELEASE operation will appear to happen
447 before the RELEASE operation with respect to the other components of the
448 system. RELEASE operations include UNLOCK operations and
449 smp_store_release() operations.
450
451 Memory operations that occur after a RELEASE operation may appear to
452 happen before it completes.
453
454 The use of ACQUIRE and RELEASE operations generally precludes the need
455 for other sorts of memory barrier (but note the exceptions mentioned in
456 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
457 pair is -not- guaranteed to act as a full memory barrier. However, after
458 an ACQUIRE on a given variable, all memory accesses preceding any prior
459 RELEASE on that same variable are guaranteed to be visible. In other
460 words, within a given variable's critical section, all accesses of all
461 previous critical sections for that variable are guaranteed to have
462 completed.
463
464 This means that ACQUIRE acts as a minimal "acquire" operation and
465 RELEASE acts as a minimal "release" operation.
466
467
468 Memory barriers are only required where there's a possibility of interaction
469 between two CPUs or between a CPU and a device. If it can be guaranteed that
470 there won't be any such interaction in any particular piece of code, then
471 memory barriers are unnecessary in that piece of code.
472
473
474 Note that these are the _minimum_ guarantees. Different architectures may give
475 more substantial guarantees, but they may _not_ be relied upon outside of arch
476 specific code.
477
478
479 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
480 ----------------------------------------------
481
482 There are certain things that the Linux kernel memory barriers do not guarantee:
483
484 (*) There is no guarantee that any of the memory accesses specified before a
485 memory barrier will be _complete_ by the completion of a memory barrier
486 instruction; the barrier can be considered to draw a line in that CPU's
487 access queue that accesses of the appropriate type may not cross.
488
489 (*) There is no guarantee that issuing a memory barrier on one CPU will have
490 any direct effect on another CPU or any other hardware in the system. The
491 indirect effect will be the order in which the second CPU sees the effects
492 of the first CPU's accesses occur, but see the next point:
493
494 (*) There is no guarantee that a CPU will see the correct order of effects
495 from a second CPU's accesses, even _if_ the second CPU uses a memory
496 barrier, unless the first CPU _also_ uses a matching memory barrier (see
497 the subsection on "SMP Barrier Pairing").
498
499 (*) There is no guarantee that some intervening piece of off-the-CPU
500 hardware[*] will not reorder the memory accesses. CPU cache coherency
501 mechanisms should propagate the indirect effects of a memory barrier
502 between CPUs, but might not do so in order.
503
504 [*] For information on bus mastering DMA and coherency please read:
505
506 Documentation/PCI/pci.txt
507 Documentation/DMA-API-HOWTO.txt
508 Documentation/DMA-API.txt
509
510
511 DATA DEPENDENCY BARRIERS
512 ------------------------
513
514 The usage requirements of data dependency barriers are a little subtle, and
515 it's not always obvious that they're needed. To illustrate, consider the
516 following sequence of events:
517
518 CPU 1 CPU 2
519 =============== ===============
520 { A == 1, B == 2, C = 3, P == &A, Q == &C }
521 B = 4;
522 <write barrier>
523 ACCESS_ONCE(P) = &B
524 Q = ACCESS_ONCE(P);
525 D = *Q;
526
527 There's a clear data dependency here, and it would seem that by the end of the
528 sequence, Q must be either &A or &B, and that:
529
530 (Q == &A) implies (D == 1)
531 (Q == &B) implies (D == 4)
532
533 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
534 leading to the following situation:
535
536 (Q == &B) and (D == 2) ????
537
538 Whilst this may seem like a failure of coherency or causality maintenance, it
539 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
540 Alpha).
541
542 To deal with this, a data dependency barrier or better must be inserted
543 between the address load and the data load:
544
545 CPU 1 CPU 2
546 =============== ===============
547 { A == 1, B == 2, C = 3, P == &A, Q == &C }
548 B = 4;
549 <write barrier>
550 ACCESS_ONCE(P) = &B
551 Q = ACCESS_ONCE(P);
552 <data dependency barrier>
553 D = *Q;
554
555 This enforces the occurrence of one of the two implications, and prevents the
556 third possibility from arising.
557
558 [!] Note that this extremely counterintuitive situation arises most easily on
559 machines with split caches, so that, for example, one cache bank processes
560 even-numbered cache lines and the other bank processes odd-numbered cache
561 lines. The pointer P might be stored in an odd-numbered cache line, and the
562 variable B might be stored in an even-numbered cache line. Then, if the
563 even-numbered bank of the reading CPU's cache is extremely busy while the
564 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
565 but the old value of the variable B (2).
566
567
568 Another example of where data dependency barriers might be required is where a
569 number is read from memory and then used to calculate the index for an array
570 access:
571
572 CPU 1 CPU 2
573 =============== ===============
574 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
575 M[1] = 4;
576 <write barrier>
577 ACCESS_ONCE(P) = 1
578 Q = ACCESS_ONCE(P);
579 <data dependency barrier>
580 D = M[Q];
581
582
583 The data dependency barrier is very important to the RCU system,
584 for example. See rcu_assign_pointer() and rcu_dereference() in
585 include/linux/rcupdate.h. This permits the current target of an RCU'd
586 pointer to be replaced with a new modified target, without the replacement
587 target appearing to be incompletely initialised.
588
589 See also the subsection on "Cache Coherency" for a more thorough example.
590
591
592 CONTROL DEPENDENCIES
593 --------------------
594
595 A load-load control dependency requires a full read memory barrier, not
596 simply a data dependency barrier to make it work correctly. Consider the
597 following bit of code:
598
599 q = ACCESS_ONCE(a);
600 if (q) {
601 <data dependency barrier> /* BUG: No data dependency!!! */
602 p = ACCESS_ONCE(b);
603 }
604
605 This will not have the desired effect because there is no actual data
606 dependency, but rather a control dependency that the CPU may short-circuit
607 by attempting to predict the outcome in advance, so that other CPUs see
608 the load from b as having happened before the load from a. In such a
609 case what's actually required is:
610
611 q = ACCESS_ONCE(a);
612 if (q) {
613 <read barrier>
614 p = ACCESS_ONCE(b);
615 }
616
617 However, stores are not speculated. This means that ordering -is- provided
618 for load-store control dependencies, as in the following example:
619
620 q = READ_ONCE_CTRL(a);
621 if (q) {
622 ACCESS_ONCE(b) = p;
623 }
624
625 Control dependencies pair normally with other types of barriers. That
626 said, please note that READ_ONCE_CTRL() is not optional! Without the
627 READ_ONCE_CTRL(), the compiler might combine the load from 'a' with
628 other loads from 'a', and the store to 'b' with other stores to 'b',
629 with possible highly counterintuitive effects on ordering.
630
631 Worse yet, if the compiler is able to prove (say) that the value of
632 variable 'a' is always non-zero, it would be well within its rights
633 to optimize the original example by eliminating the "if" statement
634 as follows:
635
636 q = a;
637 b = p; /* BUG: Compiler and CPU can both reorder!!! */
638
639 Finally, the READ_ONCE_CTRL() includes an smp_read_barrier_depends()
640 that DEC Alpha needs in order to respect control depedencies.
641
642 So don't leave out the READ_ONCE_CTRL().
643
644 It is tempting to try to enforce ordering on identical stores on both
645 branches of the "if" statement as follows:
646
647 q = READ_ONCE_CTRL(a);
648 if (q) {
649 barrier();
650 ACCESS_ONCE(b) = p;
651 do_something();
652 } else {
653 barrier();
654 ACCESS_ONCE(b) = p;
655 do_something_else();
656 }
657
658 Unfortunately, current compilers will transform this as follows at high
659 optimization levels:
660
661 q = READ_ONCE_CTRL(a);
662 barrier();
663 ACCESS_ONCE(b) = p; /* BUG: No ordering vs. load from a!!! */
664 if (q) {
665 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
666 do_something();
667 } else {
668 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
669 do_something_else();
670 }
671
672 Now there is no conditional between the load from 'a' and the store to
673 'b', which means that the CPU is within its rights to reorder them:
674 The conditional is absolutely required, and must be present in the
675 assembly code even after all compiler optimizations have been applied.
676 Therefore, if you need ordering in this example, you need explicit
677 memory barriers, for example, smp_store_release():
678
679 q = ACCESS_ONCE(a);
680 if (q) {
681 smp_store_release(&b, p);
682 do_something();
683 } else {
684 smp_store_release(&b, p);
685 do_something_else();
686 }
687
688 In contrast, without explicit memory barriers, two-legged-if control
689 ordering is guaranteed only when the stores differ, for example:
690
691 q = READ_ONCE_CTRL(a);
692 if (q) {
693 ACCESS_ONCE(b) = p;
694 do_something();
695 } else {
696 ACCESS_ONCE(b) = r;
697 do_something_else();
698 }
699
700 The initial READ_ONCE_CTRL() is still required to prevent the compiler
701 from proving the value of 'a'.
702
703 In addition, you need to be careful what you do with the local variable 'q',
704 otherwise the compiler might be able to guess the value and again remove
705 the needed conditional. For example:
706
707 q = READ_ONCE_CTRL(a);
708 if (q % MAX) {
709 ACCESS_ONCE(b) = p;
710 do_something();
711 } else {
712 ACCESS_ONCE(b) = r;
713 do_something_else();
714 }
715
716 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
717 equal to zero, in which case the compiler is within its rights to
718 transform the above code into the following:
719
720 q = READ_ONCE_CTRL(a);
721 ACCESS_ONCE(b) = p;
722 do_something_else();
723
724 Given this transformation, the CPU is not required to respect the ordering
725 between the load from variable 'a' and the store to variable 'b'. It is
726 tempting to add a barrier(), but this does not help. The conditional
727 is gone, and the barrier won't bring it back. Therefore, if you are
728 relying on this ordering, you should make sure that MAX is greater than
729 one, perhaps as follows:
730
731 q = READ_ONCE_CTRL(a);
732 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
733 if (q % MAX) {
734 ACCESS_ONCE(b) = p;
735 do_something();
736 } else {
737 ACCESS_ONCE(b) = r;
738 do_something_else();
739 }
740
741 Please note once again that the stores to 'b' differ. If they were
742 identical, as noted earlier, the compiler could pull this store outside
743 of the 'if' statement.
744
745 You must also be careful not to rely too much on boolean short-circuit
746 evaluation. Consider this example:
747
748 q = READ_ONCE_CTRL(a);
749 if (a || 1 > 0)
750 ACCESS_ONCE(b) = 1;
751
752 Because the first condition cannot fault and the second condition is
753 always true, the compiler can transform this example as following,
754 defeating control dependency:
755
756 q = READ_ONCE_CTRL(a);
757 ACCESS_ONCE(b) = 1;
758
759 This example underscores the need to ensure that the compiler cannot
760 out-guess your code. More generally, although ACCESS_ONCE() does force
761 the compiler to actually emit code for a given load, it does not force
762 the compiler to use the results.
763
764 Finally, control dependencies do -not- provide transitivity. This is
765 demonstrated by two related examples, with the initial values of
766 x and y both being zero:
767
768 CPU 0 CPU 1
769 ======================= =======================
770 r1 = READ_ONCE_CTRL(x); r2 = READ_ONCE_CTRL(y);
771 if (r1 > 0) if (r2 > 0)
772 ACCESS_ONCE(y) = 1; ACCESS_ONCE(x) = 1;
773
774 assert(!(r1 == 1 && r2 == 1));
775
776 The above two-CPU example will never trigger the assert(). However,
777 if control dependencies guaranteed transitivity (which they do not),
778 then adding the following CPU would guarantee a related assertion:
779
780 CPU 2
781 =====================
782 ACCESS_ONCE(x) = 2;
783
784 assert(!(r1 == 2 && r2 == 1 && x == 2)); /* FAILS!!! */
785
786 But because control dependencies do -not- provide transitivity, the above
787 assertion can fail after the combined three-CPU example completes. If you
788 need the three-CPU example to provide ordering, you will need smp_mb()
789 between the loads and stores in the CPU 0 and CPU 1 code fragments,
790 that is, just before or just after the "if" statements. Furthermore,
791 the original two-CPU example is very fragile and should be avoided.
792
793 These two examples are the LB and WWC litmus tests from this paper:
794 http://www.cl.cam.ac.uk/users/pes20/ppc-supplemental/test6.pdf and this
795 site: https://www.cl.cam.ac.uk/~pes20/ppcmem/index.html.
796
797 In summary:
798
799 (*) Control dependencies must be headed by READ_ONCE_CTRL().
800 Or, as a much less preferable alternative, interpose
801 be headed by READ_ONCE() or an ACCESS_ONCE() read and must
802 have smp_read_barrier_depends() between this read and the
803 control-dependent write.
804
805 (*) Control dependencies can order prior loads against later stores.
806 However, they do -not- guarantee any other sort of ordering:
807 Not prior loads against later loads, nor prior stores against
808 later anything. If you need these other forms of ordering,
809 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
810 later loads, smp_mb().
811
812 (*) If both legs of the "if" statement begin with identical stores
813 to the same variable, a barrier() statement is required at the
814 beginning of each leg of the "if" statement.
815
816 (*) Control dependencies require at least one run-time conditional
817 between the prior load and the subsequent store, and this
818 conditional must involve the prior load. If the compiler
819 is able to optimize the conditional away, it will have also
820 optimized away the ordering. Careful use of ACCESS_ONCE() can
821 help to preserve the needed conditional.
822
823 (*) Control dependencies require that the compiler avoid reordering the
824 dependency into nonexistence. Careful use of ACCESS_ONCE() or
825 barrier() can help to preserve your control dependency. Please
826 see the Compiler Barrier section for more information.
827
828 (*) Control dependencies pair normally with other types of barriers.
829
830 (*) Control dependencies do -not- provide transitivity. If you
831 need transitivity, use smp_mb().
832
833
834 SMP BARRIER PAIRING
835 -------------------
836
837 When dealing with CPU-CPU interactions, certain types of memory barrier should
838 always be paired. A lack of appropriate pairing is almost certainly an error.
839
840 General barriers pair with each other, though they also pair with most
841 other types of barriers, albeit without transitivity. An acquire barrier
842 pairs with a release barrier, but both may also pair with other barriers,
843 including of course general barriers. A write barrier pairs with a data
844 dependency barrier, a control dependency, an acquire barrier, a release
845 barrier, a read barrier, or a general barrier. Similarly a read barrier,
846 control dependency, or a data dependency barrier pairs with a write
847 barrier, an acquire barrier, a release barrier, or a general barrier:
848
849 CPU 1 CPU 2
850 =============== ===============
851 ACCESS_ONCE(a) = 1;
852 <write barrier>
853 ACCESS_ONCE(b) = 2; x = ACCESS_ONCE(b);
854 <read barrier>
855 y = ACCESS_ONCE(a);
856
857 Or:
858
859 CPU 1 CPU 2
860 =============== ===============================
861 a = 1;
862 <write barrier>
863 ACCESS_ONCE(b) = &a; x = ACCESS_ONCE(b);
864 <data dependency barrier>
865 y = *x;
866
867 Or even:
868
869 CPU 1 CPU 2
870 =============== ===============================
871 r1 = ACCESS_ONCE(y);
872 <general barrier>
873 ACCESS_ONCE(y) = 1; if (r2 = ACCESS_ONCE(x)) {
874 <implicit control dependency>
875 ACCESS_ONCE(y) = 1;
876 }
877
878 assert(r1 == 0 || r2 == 0);
879
880 Basically, the read barrier always has to be there, even though it can be of
881 the "weaker" type.
882
883 [!] Note that the stores before the write barrier would normally be expected to
884 match the loads after the read barrier or the data dependency barrier, and vice
885 versa:
886
887 CPU 1 CPU 2
888 =================== ===================
889 ACCESS_ONCE(a) = 1; }---- --->{ v = ACCESS_ONCE(c);
890 ACCESS_ONCE(b) = 2; } \ / { w = ACCESS_ONCE(d);
891 <write barrier> \ <read barrier>
892 ACCESS_ONCE(c) = 3; } / \ { x = ACCESS_ONCE(a);
893 ACCESS_ONCE(d) = 4; }---- --->{ y = ACCESS_ONCE(b);
894
895
896 EXAMPLES OF MEMORY BARRIER SEQUENCES
897 ------------------------------------
898
899 Firstly, write barriers act as partial orderings on store operations.
900 Consider the following sequence of events:
901
902 CPU 1
903 =======================
904 STORE A = 1
905 STORE B = 2
906 STORE C = 3
907 <write barrier>
908 STORE D = 4
909 STORE E = 5
910
911 This sequence of events is committed to the memory coherence system in an order
912 that the rest of the system might perceive as the unordered set of { STORE A,
913 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
914 }:
915
916 +-------+ : :
917 | | +------+
918 | |------>| C=3 | } /\
919 | | : +------+ }----- \ -----> Events perceptible to
920 | | : | A=1 | } \/ the rest of the system
921 | | : +------+ }
922 | CPU 1 | : | B=2 | }
923 | | +------+ }
924 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
925 | | +------+ } requires all stores prior to the
926 | | : | E=5 | } barrier to be committed before
927 | | : +------+ } further stores may take place
928 | |------>| D=4 | }
929 | | +------+
930 +-------+ : :
931 |
932 | Sequence in which stores are committed to the
933 | memory system by CPU 1
934 V
935
936
937 Secondly, data dependency barriers act as partial orderings on data-dependent
938 loads. Consider the following sequence of events:
939
940 CPU 1 CPU 2
941 ======================= =======================
942 { B = 7; X = 9; Y = 8; C = &Y }
943 STORE A = 1
944 STORE B = 2
945 <write barrier>
946 STORE C = &B LOAD X
947 STORE D = 4 LOAD C (gets &B)
948 LOAD *C (reads B)
949
950 Without intervention, CPU 2 may perceive the events on CPU 1 in some
951 effectively random order, despite the write barrier issued by CPU 1:
952
953 +-------+ : : : :
954 | | +------+ +-------+ | Sequence of update
955 | |------>| B=2 |----- --->| Y->8 | | of perception on
956 | | : +------+ \ +-------+ | CPU 2
957 | CPU 1 | : | A=1 | \ --->| C->&Y | V
958 | | +------+ | +-------+
959 | | wwwwwwwwwwwwwwww | : :
960 | | +------+ | : :
961 | | : | C=&B |--- | : : +-------+
962 | | : +------+ \ | +-------+ | |
963 | |------>| D=4 | ----------->| C->&B |------>| |
964 | | +------+ | +-------+ | |
965 +-------+ : : | : : | |
966 | : : | |
967 | : : | CPU 2 |
968 | +-------+ | |
969 Apparently incorrect ---> | | B->7 |------>| |
970 perception of B (!) | +-------+ | |
971 | : : | |
972 | +-------+ | |
973 The load of X holds ---> \ | X->9 |------>| |
974 up the maintenance \ +-------+ | |
975 of coherence of B ----->| B->2 | +-------+
976 +-------+
977 : :
978
979
980 In the above example, CPU 2 perceives that B is 7, despite the load of *C
981 (which would be B) coming after the LOAD of C.
982
983 If, however, a data dependency barrier were to be placed between the load of C
984 and the load of *C (ie: B) on CPU 2:
985
986 CPU 1 CPU 2
987 ======================= =======================
988 { B = 7; X = 9; Y = 8; C = &Y }
989 STORE A = 1
990 STORE B = 2
991 <write barrier>
992 STORE C = &B LOAD X
993 STORE D = 4 LOAD C (gets &B)
994 <data dependency barrier>
995 LOAD *C (reads B)
996
997 then the following will occur:
998
999 +-------+ : : : :
1000 | | +------+ +-------+
1001 | |------>| B=2 |----- --->| Y->8 |
1002 | | : +------+ \ +-------+
1003 | CPU 1 | : | A=1 | \ --->| C->&Y |
1004 | | +------+ | +-------+
1005 | | wwwwwwwwwwwwwwww | : :
1006 | | +------+ | : :
1007 | | : | C=&B |--- | : : +-------+
1008 | | : +------+ \ | +-------+ | |
1009 | |------>| D=4 | ----------->| C->&B |------>| |
1010 | | +------+ | +-------+ | |
1011 +-------+ : : | : : | |
1012 | : : | |
1013 | : : | CPU 2 |
1014 | +-------+ | |
1015 | | X->9 |------>| |
1016 | +-------+ | |
1017 Makes sure all effects ---> \ ddddddddddddddddd | |
1018 prior to the store of C \ +-------+ | |
1019 are perceptible to ----->| B->2 |------>| |
1020 subsequent loads +-------+ | |
1021 : : +-------+
1022
1023
1024 And thirdly, a read barrier acts as a partial order on loads. Consider the
1025 following sequence of events:
1026
1027 CPU 1 CPU 2
1028 ======================= =======================
1029 { A = 0, B = 9 }
1030 STORE A=1
1031 <write barrier>
1032 STORE B=2
1033 LOAD B
1034 LOAD A
1035
1036 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1037 some effectively random order, despite the write barrier issued by CPU 1:
1038
1039 +-------+ : : : :
1040 | | +------+ +-------+
1041 | |------>| A=1 |------ --->| A->0 |
1042 | | +------+ \ +-------+
1043 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1044 | | +------+ | +-------+
1045 | |------>| B=2 |--- | : :
1046 | | +------+ \ | : : +-------+
1047 +-------+ : : \ | +-------+ | |
1048 ---------->| B->2 |------>| |
1049 | +-------+ | CPU 2 |
1050 | | A->0 |------>| |
1051 | +-------+ | |
1052 | : : +-------+
1053 \ : :
1054 \ +-------+
1055 ---->| A->1 |
1056 +-------+
1057 : :
1058
1059
1060 If, however, a read barrier were to be placed between the load of B and the
1061 load of A on CPU 2:
1062
1063 CPU 1 CPU 2
1064 ======================= =======================
1065 { A = 0, B = 9 }
1066 STORE A=1
1067 <write barrier>
1068 STORE B=2
1069 LOAD B
1070 <read barrier>
1071 LOAD A
1072
1073 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1074 2:
1075
1076 +-------+ : : : :
1077 | | +------+ +-------+
1078 | |------>| A=1 |------ --->| A->0 |
1079 | | +------+ \ +-------+
1080 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1081 | | +------+ | +-------+
1082 | |------>| B=2 |--- | : :
1083 | | +------+ \ | : : +-------+
1084 +-------+ : : \ | +-------+ | |
1085 ---------->| B->2 |------>| |
1086 | +-------+ | CPU 2 |
1087 | : : | |
1088 | : : | |
1089 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1090 barrier causes all effects \ +-------+ | |
1091 prior to the storage of B ---->| A->1 |------>| |
1092 to be perceptible to CPU 2 +-------+ | |
1093 : : +-------+
1094
1095
1096 To illustrate this more completely, consider what could happen if the code
1097 contained a load of A either side of the read barrier:
1098
1099 CPU 1 CPU 2
1100 ======================= =======================
1101 { A = 0, B = 9 }
1102 STORE A=1
1103 <write barrier>
1104 STORE B=2
1105 LOAD B
1106 LOAD A [first load of A]
1107 <read barrier>
1108 LOAD A [second load of A]
1109
1110 Even though the two loads of A both occur after the load of B, they may both
1111 come up with different values:
1112
1113 +-------+ : : : :
1114 | | +------+ +-------+
1115 | |------>| A=1 |------ --->| A->0 |
1116 | | +------+ \ +-------+
1117 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1118 | | +------+ | +-------+
1119 | |------>| B=2 |--- | : :
1120 | | +------+ \ | : : +-------+
1121 +-------+ : : \ | +-------+ | |
1122 ---------->| B->2 |------>| |
1123 | +-------+ | CPU 2 |
1124 | : : | |
1125 | : : | |
1126 | +-------+ | |
1127 | | A->0 |------>| 1st |
1128 | +-------+ | |
1129 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1130 barrier causes all effects \ +-------+ | |
1131 prior to the storage of B ---->| A->1 |------>| 2nd |
1132 to be perceptible to CPU 2 +-------+ | |
1133 : : +-------+
1134
1135
1136 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1137 before the read barrier completes anyway:
1138
1139 +-------+ : : : :
1140 | | +------+ +-------+
1141 | |------>| A=1 |------ --->| A->0 |
1142 | | +------+ \ +-------+
1143 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1144 | | +------+ | +-------+
1145 | |------>| B=2 |--- | : :
1146 | | +------+ \ | : : +-------+
1147 +-------+ : : \ | +-------+ | |
1148 ---------->| B->2 |------>| |
1149 | +-------+ | CPU 2 |
1150 | : : | |
1151 \ : : | |
1152 \ +-------+ | |
1153 ---->| A->1 |------>| 1st |
1154 +-------+ | |
1155 rrrrrrrrrrrrrrrrr | |
1156 +-------+ | |
1157 | A->1 |------>| 2nd |
1158 +-------+ | |
1159 : : +-------+
1160
1161
1162 The guarantee is that the second load will always come up with A == 1 if the
1163 load of B came up with B == 2. No such guarantee exists for the first load of
1164 A; that may come up with either A == 0 or A == 1.
1165
1166
1167 READ MEMORY BARRIERS VS LOAD SPECULATION
1168 ----------------------------------------
1169
1170 Many CPUs speculate with loads: that is they see that they will need to load an
1171 item from memory, and they find a time where they're not using the bus for any
1172 other loads, and so do the load in advance - even though they haven't actually
1173 got to that point in the instruction execution flow yet. This permits the
1174 actual load instruction to potentially complete immediately because the CPU
1175 already has the value to hand.
1176
1177 It may turn out that the CPU didn't actually need the value - perhaps because a
1178 branch circumvented the load - in which case it can discard the value or just
1179 cache it for later use.
1180
1181 Consider:
1182
1183 CPU 1 CPU 2
1184 ======================= =======================
1185 LOAD B
1186 DIVIDE } Divide instructions generally
1187 DIVIDE } take a long time to perform
1188 LOAD A
1189
1190 Which might appear as this:
1191
1192 : : +-------+
1193 +-------+ | |
1194 --->| B->2 |------>| |
1195 +-------+ | CPU 2 |
1196 : :DIVIDE | |
1197 +-------+ | |
1198 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1199 division speculates on the +-------+ ~ | |
1200 LOAD of A : : ~ | |
1201 : :DIVIDE | |
1202 : : ~ | |
1203 Once the divisions are complete --> : : ~-->| |
1204 the CPU can then perform the : : | |
1205 LOAD with immediate effect : : +-------+
1206
1207
1208 Placing a read barrier or a data dependency barrier just before the second
1209 load:
1210
1211 CPU 1 CPU 2
1212 ======================= =======================
1213 LOAD B
1214 DIVIDE
1215 DIVIDE
1216 <read barrier>
1217 LOAD A
1218
1219 will force any value speculatively obtained to be reconsidered to an extent
1220 dependent on the type of barrier used. If there was no change made to the
1221 speculated memory location, then the speculated value will just be used:
1222
1223 : : +-------+
1224 +-------+ | |
1225 --->| B->2 |------>| |
1226 +-------+ | CPU 2 |
1227 : :DIVIDE | |
1228 +-------+ | |
1229 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1230 division speculates on the +-------+ ~ | |
1231 LOAD of A : : ~ | |
1232 : :DIVIDE | |
1233 : : ~ | |
1234 : : ~ | |
1235 rrrrrrrrrrrrrrrr~ | |
1236 : : ~ | |
1237 : : ~-->| |
1238 : : | |
1239 : : +-------+
1240
1241
1242 but if there was an update or an invalidation from another CPU pending, then
1243 the speculation will be cancelled and the value reloaded:
1244
1245 : : +-------+
1246 +-------+ | |
1247 --->| B->2 |------>| |
1248 +-------+ | CPU 2 |
1249 : :DIVIDE | |
1250 +-------+ | |
1251 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1252 division speculates on the +-------+ ~ | |
1253 LOAD of A : : ~ | |
1254 : :DIVIDE | |
1255 : : ~ | |
1256 : : ~ | |
1257 rrrrrrrrrrrrrrrrr | |
1258 +-------+ | |
1259 The speculation is discarded ---> --->| A->1 |------>| |
1260 and an updated value is +-------+ | |
1261 retrieved : : +-------+
1262
1263
1264 TRANSITIVITY
1265 ------------
1266
1267 Transitivity is a deeply intuitive notion about ordering that is not
1268 always provided by real computer systems. The following example
1269 demonstrates transitivity (also called "cumulativity"):
1270
1271 CPU 1 CPU 2 CPU 3
1272 ======================= ======================= =======================
1273 { X = 0, Y = 0 }
1274 STORE X=1 LOAD X STORE Y=1
1275 <general barrier> <general barrier>
1276 LOAD Y LOAD X
1277
1278 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1279 This indicates that CPU 2's load from X in some sense follows CPU 1's
1280 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1281 store to Y. The question is then "Can CPU 3's load from X return 0?"
1282
1283 Because CPU 2's load from X in some sense came after CPU 1's store, it
1284 is natural to expect that CPU 3's load from X must therefore return 1.
1285 This expectation is an example of transitivity: if a load executing on
1286 CPU A follows a load from the same variable executing on CPU B, then
1287 CPU A's load must either return the same value that CPU B's load did,
1288 or must return some later value.
1289
1290 In the Linux kernel, use of general memory barriers guarantees
1291 transitivity. Therefore, in the above example, if CPU 2's load from X
1292 returns 1 and its load from Y returns 0, then CPU 3's load from X must
1293 also return 1.
1294
1295 However, transitivity is -not- guaranteed for read or write barriers.
1296 For example, suppose that CPU 2's general barrier in the above example
1297 is changed to a read barrier as shown below:
1298
1299 CPU 1 CPU 2 CPU 3
1300 ======================= ======================= =======================
1301 { X = 0, Y = 0 }
1302 STORE X=1 LOAD X STORE Y=1
1303 <read barrier> <general barrier>
1304 LOAD Y LOAD X
1305
1306 This substitution destroys transitivity: in this example, it is perfectly
1307 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1308 and CPU 3's load from X to return 0.
1309
1310 The key point is that although CPU 2's read barrier orders its pair
1311 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1312 this example runs on a system where CPUs 1 and 2 share a store buffer
1313 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1314 General barriers are therefore required to ensure that all CPUs agree
1315 on the combined order of CPU 1's and CPU 2's accesses.
1316
1317 To reiterate, if your code requires transitivity, use general barriers
1318 throughout.
1319
1320
1321 ========================
1322 EXPLICIT KERNEL BARRIERS
1323 ========================
1324
1325 The Linux kernel has a variety of different barriers that act at different
1326 levels:
1327
1328 (*) Compiler barrier.
1329
1330 (*) CPU memory barriers.
1331
1332 (*) MMIO write barrier.
1333
1334
1335 COMPILER BARRIER
1336 ----------------
1337
1338 The Linux kernel has an explicit compiler barrier function that prevents the
1339 compiler from moving the memory accesses either side of it to the other side:
1340
1341 barrier();
1342
1343 This is a general barrier -- there are no read-read or write-write variants
1344 of barrier(). However, ACCESS_ONCE() can be thought of as a weak form
1345 for barrier() that affects only the specific accesses flagged by the
1346 ACCESS_ONCE().
1347
1348 The barrier() function has the following effects:
1349
1350 (*) Prevents the compiler from reordering accesses following the
1351 barrier() to precede any accesses preceding the barrier().
1352 One example use for this property is to ease communication between
1353 interrupt-handler code and the code that was interrupted.
1354
1355 (*) Within a loop, forces the compiler to load the variables used
1356 in that loop's conditional on each pass through that loop.
1357
1358 The ACCESS_ONCE() function can prevent any number of optimizations that,
1359 while perfectly safe in single-threaded code, can be fatal in concurrent
1360 code. Here are some examples of these sorts of optimizations:
1361
1362 (*) The compiler is within its rights to reorder loads and stores
1363 to the same variable, and in some cases, the CPU is within its
1364 rights to reorder loads to the same variable. This means that
1365 the following code:
1366
1367 a[0] = x;
1368 a[1] = x;
1369
1370 Might result in an older value of x stored in a[1] than in a[0].
1371 Prevent both the compiler and the CPU from doing this as follows:
1372
1373 a[0] = ACCESS_ONCE(x);
1374 a[1] = ACCESS_ONCE(x);
1375
1376 In short, ACCESS_ONCE() provides cache coherence for accesses from
1377 multiple CPUs to a single variable.
1378
1379 (*) The compiler is within its rights to merge successive loads from
1380 the same variable. Such merging can cause the compiler to "optimize"
1381 the following code:
1382
1383 while (tmp = a)
1384 do_something_with(tmp);
1385
1386 into the following code, which, although in some sense legitimate
1387 for single-threaded code, is almost certainly not what the developer
1388 intended:
1389
1390 if (tmp = a)
1391 for (;;)
1392 do_something_with(tmp);
1393
1394 Use ACCESS_ONCE() to prevent the compiler from doing this to you:
1395
1396 while (tmp = ACCESS_ONCE(a))
1397 do_something_with(tmp);
1398
1399 (*) The compiler is within its rights to reload a variable, for example,
1400 in cases where high register pressure prevents the compiler from
1401 keeping all data of interest in registers. The compiler might
1402 therefore optimize the variable 'tmp' out of our previous example:
1403
1404 while (tmp = a)
1405 do_something_with(tmp);
1406
1407 This could result in the following code, which is perfectly safe in
1408 single-threaded code, but can be fatal in concurrent code:
1409
1410 while (a)
1411 do_something_with(a);
1412
1413 For example, the optimized version of this code could result in
1414 passing a zero to do_something_with() in the case where the variable
1415 a was modified by some other CPU between the "while" statement and
1416 the call to do_something_with().
1417
1418 Again, use ACCESS_ONCE() to prevent the compiler from doing this:
1419
1420 while (tmp = ACCESS_ONCE(a))
1421 do_something_with(tmp);
1422
1423 Note that if the compiler runs short of registers, it might save
1424 tmp onto the stack. The overhead of this saving and later restoring
1425 is why compilers reload variables. Doing so is perfectly safe for
1426 single-threaded code, so you need to tell the compiler about cases
1427 where it is not safe.
1428
1429 (*) The compiler is within its rights to omit a load entirely if it knows
1430 what the value will be. For example, if the compiler can prove that
1431 the value of variable 'a' is always zero, it can optimize this code:
1432
1433 while (tmp = a)
1434 do_something_with(tmp);
1435
1436 Into this:
1437
1438 do { } while (0);
1439
1440 This transformation is a win for single-threaded code because it gets
1441 rid of a load and a branch. The problem is that the compiler will
1442 carry out its proof assuming that the current CPU is the only one
1443 updating variable 'a'. If variable 'a' is shared, then the compiler's
1444 proof will be erroneous. Use ACCESS_ONCE() to tell the compiler
1445 that it doesn't know as much as it thinks it does:
1446
1447 while (tmp = ACCESS_ONCE(a))
1448 do_something_with(tmp);
1449
1450 But please note that the compiler is also closely watching what you
1451 do with the value after the ACCESS_ONCE(). For example, suppose you
1452 do the following and MAX is a preprocessor macro with the value 1:
1453
1454 while ((tmp = ACCESS_ONCE(a)) % MAX)
1455 do_something_with(tmp);
1456
1457 Then the compiler knows that the result of the "%" operator applied
1458 to MAX will always be zero, again allowing the compiler to optimize
1459 the code into near-nonexistence. (It will still load from the
1460 variable 'a'.)
1461
1462 (*) Similarly, the compiler is within its rights to omit a store entirely
1463 if it knows that the variable already has the value being stored.
1464 Again, the compiler assumes that the current CPU is the only one
1465 storing into the variable, which can cause the compiler to do the
1466 wrong thing for shared variables. For example, suppose you have
1467 the following:
1468
1469 a = 0;
1470 /* Code that does not store to variable a. */
1471 a = 0;
1472
1473 The compiler sees that the value of variable 'a' is already zero, so
1474 it might well omit the second store. This would come as a fatal
1475 surprise if some other CPU might have stored to variable 'a' in the
1476 meantime.
1477
1478 Use ACCESS_ONCE() to prevent the compiler from making this sort of
1479 wrong guess:
1480
1481 ACCESS_ONCE(a) = 0;
1482 /* Code that does not store to variable a. */
1483 ACCESS_ONCE(a) = 0;
1484
1485 (*) The compiler is within its rights to reorder memory accesses unless
1486 you tell it not to. For example, consider the following interaction
1487 between process-level code and an interrupt handler:
1488
1489 void process_level(void)
1490 {
1491 msg = get_message();
1492 flag = true;
1493 }
1494
1495 void interrupt_handler(void)
1496 {
1497 if (flag)
1498 process_message(msg);
1499 }
1500
1501 There is nothing to prevent the compiler from transforming
1502 process_level() to the following, in fact, this might well be a
1503 win for single-threaded code:
1504
1505 void process_level(void)
1506 {
1507 flag = true;
1508 msg = get_message();
1509 }
1510
1511 If the interrupt occurs between these two statement, then
1512 interrupt_handler() might be passed a garbled msg. Use ACCESS_ONCE()
1513 to prevent this as follows:
1514
1515 void process_level(void)
1516 {
1517 ACCESS_ONCE(msg) = get_message();
1518 ACCESS_ONCE(flag) = true;
1519 }
1520
1521 void interrupt_handler(void)
1522 {
1523 if (ACCESS_ONCE(flag))
1524 process_message(ACCESS_ONCE(msg));
1525 }
1526
1527 Note that the ACCESS_ONCE() wrappers in interrupt_handler()
1528 are needed if this interrupt handler can itself be interrupted
1529 by something that also accesses 'flag' and 'msg', for example,
1530 a nested interrupt or an NMI. Otherwise, ACCESS_ONCE() is not
1531 needed in interrupt_handler() other than for documentation purposes.
1532 (Note also that nested interrupts do not typically occur in modern
1533 Linux kernels, in fact, if an interrupt handler returns with
1534 interrupts enabled, you will get a WARN_ONCE() splat.)
1535
1536 You should assume that the compiler can move ACCESS_ONCE() past
1537 code not containing ACCESS_ONCE(), barrier(), or similar primitives.
1538
1539 This effect could also be achieved using barrier(), but ACCESS_ONCE()
1540 is more selective: With ACCESS_ONCE(), the compiler need only forget
1541 the contents of the indicated memory locations, while with barrier()
1542 the compiler must discard the value of all memory locations that
1543 it has currented cached in any machine registers. Of course,
1544 the compiler must also respect the order in which the ACCESS_ONCE()s
1545 occur, though the CPU of course need not do so.
1546
1547 (*) The compiler is within its rights to invent stores to a variable,
1548 as in the following example:
1549
1550 if (a)
1551 b = a;
1552 else
1553 b = 42;
1554
1555 The compiler might save a branch by optimizing this as follows:
1556
1557 b = 42;
1558 if (a)
1559 b = a;
1560
1561 In single-threaded code, this is not only safe, but also saves
1562 a branch. Unfortunately, in concurrent code, this optimization
1563 could cause some other CPU to see a spurious value of 42 -- even
1564 if variable 'a' was never zero -- when loading variable 'b'.
1565 Use ACCESS_ONCE() to prevent this as follows:
1566
1567 if (a)
1568 ACCESS_ONCE(b) = a;
1569 else
1570 ACCESS_ONCE(b) = 42;
1571
1572 The compiler can also invent loads. These are usually less
1573 damaging, but they can result in cache-line bouncing and thus in
1574 poor performance and scalability. Use ACCESS_ONCE() to prevent
1575 invented loads.
1576
1577 (*) For aligned memory locations whose size allows them to be accessed
1578 with a single memory-reference instruction, prevents "load tearing"
1579 and "store tearing," in which a single large access is replaced by
1580 multiple smaller accesses. For example, given an architecture having
1581 16-bit store instructions with 7-bit immediate fields, the compiler
1582 might be tempted to use two 16-bit store-immediate instructions to
1583 implement the following 32-bit store:
1584
1585 p = 0x00010002;
1586
1587 Please note that GCC really does use this sort of optimization,
1588 which is not surprising given that it would likely take more
1589 than two instructions to build the constant and then store it.
1590 This optimization can therefore be a win in single-threaded code.
1591 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1592 this optimization in a volatile store. In the absence of such bugs,
1593 use of ACCESS_ONCE() prevents store tearing in the following example:
1594
1595 ACCESS_ONCE(p) = 0x00010002;
1596
1597 Use of packed structures can also result in load and store tearing,
1598 as in this example:
1599
1600 struct __attribute__((__packed__)) foo {
1601 short a;
1602 int b;
1603 short c;
1604 };
1605 struct foo foo1, foo2;
1606 ...
1607
1608 foo2.a = foo1.a;
1609 foo2.b = foo1.b;
1610 foo2.c = foo1.c;
1611
1612 Because there are no ACCESS_ONCE() wrappers and no volatile markings,
1613 the compiler would be well within its rights to implement these three
1614 assignment statements as a pair of 32-bit loads followed by a pair
1615 of 32-bit stores. This would result in load tearing on 'foo1.b'
1616 and store tearing on 'foo2.b'. ACCESS_ONCE() again prevents tearing
1617 in this example:
1618
1619 foo2.a = foo1.a;
1620 ACCESS_ONCE(foo2.b) = ACCESS_ONCE(foo1.b);
1621 foo2.c = foo1.c;
1622
1623 All that aside, it is never necessary to use ACCESS_ONCE() on a variable
1624 that has been marked volatile. For example, because 'jiffies' is marked
1625 volatile, it is never necessary to say ACCESS_ONCE(jiffies). The reason
1626 for this is that ACCESS_ONCE() is implemented as a volatile cast, which
1627 has no effect when its argument is already marked volatile.
1628
1629 Please note that these compiler barriers have no direct effect on the CPU,
1630 which may then reorder things however it wishes.
1631
1632
1633 CPU MEMORY BARRIERS
1634 -------------------
1635
1636 The Linux kernel has eight basic CPU memory barriers:
1637
1638 TYPE MANDATORY SMP CONDITIONAL
1639 =============== ======================= ===========================
1640 GENERAL mb() smp_mb()
1641 WRITE wmb() smp_wmb()
1642 READ rmb() smp_rmb()
1643 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1644
1645
1646 All memory barriers except the data dependency barriers imply a compiler
1647 barrier. Data dependencies do not impose any additional compiler ordering.
1648
1649 Aside: In the case of data dependencies, the compiler would be expected to
1650 issue the loads in the correct order (eg. `a[b]` would have to load the value
1651 of b before loading a[b]), however there is no guarantee in the C specification
1652 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1653 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1654 problem of a compiler reloading b after having loaded a[b], thus having a newer
1655 copy of b than a[b]. A consensus has not yet been reached about these problems,
1656 however the ACCESS_ONCE macro is a good place to start looking.
1657
1658 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1659 systems because it is assumed that a CPU will appear to be self-consistent,
1660 and will order overlapping accesses correctly with respect to itself.
1661
1662 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1663 references to shared memory on SMP systems, though the use of locking instead
1664 is sufficient.
1665
1666 Mandatory barriers should not be used to control SMP effects, since mandatory
1667 barriers unnecessarily impose overhead on UP systems. They may, however, be
1668 used to control MMIO effects on accesses through relaxed memory I/O windows.
1669 These are required even on non-SMP systems as they affect the order in which
1670 memory operations appear to a device by prohibiting both the compiler and the
1671 CPU from reordering them.
1672
1673
1674 There are some more advanced barrier functions:
1675
1676 (*) smp_store_mb(var, value)
1677
1678 This assigns the value to the variable and then inserts a full memory
1679 barrier after it, depending on the function. It isn't guaranteed to
1680 insert anything more than a compiler barrier in a UP compilation.
1681
1682
1683 (*) smp_mb__before_atomic();
1684 (*) smp_mb__after_atomic();
1685
1686 These are for use with atomic (such as add, subtract, increment and
1687 decrement) functions that don't return a value, especially when used for
1688 reference counting. These functions do not imply memory barriers.
1689
1690 These are also used for atomic bitop functions that do not return a
1691 value (such as set_bit and clear_bit).
1692
1693 As an example, consider a piece of code that marks an object as being dead
1694 and then decrements the object's reference count:
1695
1696 obj->dead = 1;
1697 smp_mb__before_atomic();
1698 atomic_dec(&obj->ref_count);
1699
1700 This makes sure that the death mark on the object is perceived to be set
1701 *before* the reference counter is decremented.
1702
1703 See Documentation/atomic_ops.txt for more information. See the "Atomic
1704 operations" subsection for information on where to use these.
1705
1706
1707 (*) dma_wmb();
1708 (*) dma_rmb();
1709
1710 These are for use with consistent memory to guarantee the ordering
1711 of writes or reads of shared memory accessible to both the CPU and a
1712 DMA capable device.
1713
1714 For example, consider a device driver that shares memory with a device
1715 and uses a descriptor status value to indicate if the descriptor belongs
1716 to the device or the CPU, and a doorbell to notify it when new
1717 descriptors are available:
1718
1719 if (desc->status != DEVICE_OWN) {
1720 /* do not read data until we own descriptor */
1721 dma_rmb();
1722
1723 /* read/modify data */
1724 read_data = desc->data;
1725 desc->data = write_data;
1726
1727 /* flush modifications before status update */
1728 dma_wmb();
1729
1730 /* assign ownership */
1731 desc->status = DEVICE_OWN;
1732
1733 /* force memory to sync before notifying device via MMIO */
1734 wmb();
1735
1736 /* notify device of new descriptors */
1737 writel(DESC_NOTIFY, doorbell);
1738 }
1739
1740 The dma_rmb() allows us guarantee the device has released ownership
1741 before we read the data from the descriptor, and the dma_wmb() allows
1742 us to guarantee the data is written to the descriptor before the device
1743 can see it now has ownership. The wmb() is needed to guarantee that the
1744 cache coherent memory writes have completed before attempting a write to
1745 the cache incoherent MMIO region.
1746
1747 See Documentation/DMA-API.txt for more information on consistent memory.
1748
1749 MMIO WRITE BARRIER
1750 ------------------
1751
1752 The Linux kernel also has a special barrier for use with memory-mapped I/O
1753 writes:
1754
1755 mmiowb();
1756
1757 This is a variation on the mandatory write barrier that causes writes to weakly
1758 ordered I/O regions to be partially ordered. Its effects may go beyond the
1759 CPU->Hardware interface and actually affect the hardware at some level.
1760
1761 See the subsection "Locks vs I/O accesses" for more information.
1762
1763
1764 ===============================
1765 IMPLICIT KERNEL MEMORY BARRIERS
1766 ===============================
1767
1768 Some of the other functions in the linux kernel imply memory barriers, amongst
1769 which are locking and scheduling functions.
1770
1771 This specification is a _minimum_ guarantee; any particular architecture may
1772 provide more substantial guarantees, but these may not be relied upon outside
1773 of arch specific code.
1774
1775
1776 ACQUIRING FUNCTIONS
1777 -------------------
1778
1779 The Linux kernel has a number of locking constructs:
1780
1781 (*) spin locks
1782 (*) R/W spin locks
1783 (*) mutexes
1784 (*) semaphores
1785 (*) R/W semaphores
1786 (*) RCU
1787
1788 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1789 for each construct. These operations all imply certain barriers:
1790
1791 (1) ACQUIRE operation implication:
1792
1793 Memory operations issued after the ACQUIRE will be completed after the
1794 ACQUIRE operation has completed.
1795
1796 Memory operations issued before the ACQUIRE may be completed after
1797 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1798 combined with a following ACQUIRE, orders prior stores against
1799 subsequent loads and stores. Note that this is weaker than smp_mb()!
1800 The smp_mb__before_spinlock() primitive is free on many architectures.
1801
1802 (2) RELEASE operation implication:
1803
1804 Memory operations issued before the RELEASE will be completed before the
1805 RELEASE operation has completed.
1806
1807 Memory operations issued after the RELEASE may be completed before the
1808 RELEASE operation has completed.
1809
1810 (3) ACQUIRE vs ACQUIRE implication:
1811
1812 All ACQUIRE operations issued before another ACQUIRE operation will be
1813 completed before that ACQUIRE operation.
1814
1815 (4) ACQUIRE vs RELEASE implication:
1816
1817 All ACQUIRE operations issued before a RELEASE operation will be
1818 completed before the RELEASE operation.
1819
1820 (5) Failed conditional ACQUIRE implication:
1821
1822 Certain locking variants of the ACQUIRE operation may fail, either due to
1823 being unable to get the lock immediately, or due to receiving an unblocked
1824 signal whilst asleep waiting for the lock to become available. Failed
1825 locks do not imply any sort of barrier.
1826
1827 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1828 one-way barriers is that the effects of instructions outside of a critical
1829 section may seep into the inside of the critical section.
1830
1831 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1832 because it is possible for an access preceding the ACQUIRE to happen after the
1833 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
1834 the two accesses can themselves then cross:
1835
1836 *A = a;
1837 ACQUIRE M
1838 RELEASE M
1839 *B = b;
1840
1841 may occur as:
1842
1843 ACQUIRE M, STORE *B, STORE *A, RELEASE M
1844
1845 When the ACQUIRE and RELEASE are a lock acquisition and release,
1846 respectively, this same reordering can occur if the lock's ACQUIRE and
1847 RELEASE are to the same lock variable, but only from the perspective of
1848 another CPU not holding that lock. In short, a ACQUIRE followed by an
1849 RELEASE may -not- be assumed to be a full memory barrier.
1850
1851 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does not
1852 imply a full memory barrier. If it is necessary for a RELEASE-ACQUIRE
1853 pair to produce a full barrier, the ACQUIRE can be followed by an
1854 smp_mb__after_unlock_lock() invocation. This will produce a full barrier
1855 if either (a) the RELEASE and the ACQUIRE are executed by the same
1856 CPU or task, or (b) the RELEASE and ACQUIRE act on the same variable.
1857 The smp_mb__after_unlock_lock() primitive is free on many architectures.
1858 Without smp_mb__after_unlock_lock(), the CPU's execution of the critical
1859 sections corresponding to the RELEASE and the ACQUIRE can cross, so that:
1860
1861 *A = a;
1862 RELEASE M
1863 ACQUIRE N
1864 *B = b;
1865
1866 could occur as:
1867
1868 ACQUIRE N, STORE *B, STORE *A, RELEASE M
1869
1870 It might appear that this reordering could introduce a deadlock.
1871 However, this cannot happen because if such a deadlock threatened,
1872 the RELEASE would simply complete, thereby avoiding the deadlock.
1873
1874 Why does this work?
1875
1876 One key point is that we are only talking about the CPU doing
1877 the reordering, not the compiler. If the compiler (or, for
1878 that matter, the developer) switched the operations, deadlock
1879 -could- occur.
1880
1881 But suppose the CPU reordered the operations. In this case,
1882 the unlock precedes the lock in the assembly code. The CPU
1883 simply elected to try executing the later lock operation first.
1884 If there is a deadlock, this lock operation will simply spin (or
1885 try to sleep, but more on that later). The CPU will eventually
1886 execute the unlock operation (which preceded the lock operation
1887 in the assembly code), which will unravel the potential deadlock,
1888 allowing the lock operation to succeed.
1889
1890 But what if the lock is a sleeplock? In that case, the code will
1891 try to enter the scheduler, where it will eventually encounter
1892 a memory barrier, which will force the earlier unlock operation
1893 to complete, again unraveling the deadlock. There might be
1894 a sleep-unlock race, but the locking primitive needs to resolve
1895 such races properly in any case.
1896
1897 With smp_mb__after_unlock_lock(), the two critical sections cannot overlap.
1898 For example, with the following code, the store to *A will always be
1899 seen by other CPUs before the store to *B:
1900
1901 *A = a;
1902 RELEASE M
1903 ACQUIRE N
1904 smp_mb__after_unlock_lock();
1905 *B = b;
1906
1907 The operations will always occur in one of the following orders:
1908
1909 STORE *A, RELEASE, ACQUIRE, smp_mb__after_unlock_lock(), STORE *B
1910 STORE *A, ACQUIRE, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1911 ACQUIRE, STORE *A, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1912
1913 If the RELEASE and ACQUIRE were instead both operating on the same lock
1914 variable, only the first of these alternatives can occur. In addition,
1915 the more strongly ordered systems may rule out some of the above orders.
1916 But in any case, as noted earlier, the smp_mb__after_unlock_lock()
1917 ensures that the store to *A will always be seen as happening before
1918 the store to *B.
1919
1920 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1921 systems, and so cannot be counted on in such a situation to actually achieve
1922 anything at all - especially with respect to I/O accesses - unless combined
1923 with interrupt disabling operations.
1924
1925 See also the section on "Inter-CPU locking barrier effects".
1926
1927
1928 As an example, consider the following:
1929
1930 *A = a;
1931 *B = b;
1932 ACQUIRE
1933 *C = c;
1934 *D = d;
1935 RELEASE
1936 *E = e;
1937 *F = f;
1938
1939 The following sequence of events is acceptable:
1940
1941 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
1942
1943 [+] Note that {*F,*A} indicates a combined access.
1944
1945 But none of the following are:
1946
1947 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
1948 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
1949 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
1950 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
1951
1952
1953
1954 INTERRUPT DISABLING FUNCTIONS
1955 -----------------------------
1956
1957 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
1958 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
1959 barriers are required in such a situation, they must be provided from some
1960 other means.
1961
1962
1963 SLEEP AND WAKE-UP FUNCTIONS
1964 ---------------------------
1965
1966 Sleeping and waking on an event flagged in global data can be viewed as an
1967 interaction between two pieces of data: the task state of the task waiting for
1968 the event and the global data used to indicate the event. To make sure that
1969 these appear to happen in the right order, the primitives to begin the process
1970 of going to sleep, and the primitives to initiate a wake up imply certain
1971 barriers.
1972
1973 Firstly, the sleeper normally follows something like this sequence of events:
1974
1975 for (;;) {
1976 set_current_state(TASK_UNINTERRUPTIBLE);
1977 if (event_indicated)
1978 break;
1979 schedule();
1980 }
1981
1982 A general memory barrier is interpolated automatically by set_current_state()
1983 after it has altered the task state:
1984
1985 CPU 1
1986 ===============================
1987 set_current_state();
1988 smp_store_mb();
1989 STORE current->state
1990 <general barrier>
1991 LOAD event_indicated
1992
1993 set_current_state() may be wrapped by:
1994
1995 prepare_to_wait();
1996 prepare_to_wait_exclusive();
1997
1998 which therefore also imply a general memory barrier after setting the state.
1999 The whole sequence above is available in various canned forms, all of which
2000 interpolate the memory barrier in the right place:
2001
2002 wait_event();
2003 wait_event_interruptible();
2004 wait_event_interruptible_exclusive();
2005 wait_event_interruptible_timeout();
2006 wait_event_killable();
2007 wait_event_timeout();
2008 wait_on_bit();
2009 wait_on_bit_lock();
2010
2011
2012 Secondly, code that performs a wake up normally follows something like this:
2013
2014 event_indicated = 1;
2015 wake_up(&event_wait_queue);
2016
2017 or:
2018
2019 event_indicated = 1;
2020 wake_up_process(event_daemon);
2021
2022 A write memory barrier is implied by wake_up() and co. if and only if they wake
2023 something up. The barrier occurs before the task state is cleared, and so sits
2024 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
2025
2026 CPU 1 CPU 2
2027 =============================== ===============================
2028 set_current_state(); STORE event_indicated
2029 smp_store_mb(); wake_up();
2030 STORE current->state <write barrier>
2031 <general barrier> STORE current->state
2032 LOAD event_indicated
2033
2034 To repeat, this write memory barrier is present if and only if something
2035 is actually awakened. To see this, consider the following sequence of
2036 events, where X and Y are both initially zero:
2037
2038 CPU 1 CPU 2
2039 =============================== ===============================
2040 X = 1; STORE event_indicated
2041 smp_mb(); wake_up();
2042 Y = 1; wait_event(wq, Y == 1);
2043 wake_up(); load from Y sees 1, no memory barrier
2044 load from X might see 0
2045
2046 In contrast, if a wakeup does occur, CPU 2's load from X would be guaranteed
2047 to see 1.
2048
2049 The available waker functions include:
2050
2051 complete();
2052 wake_up();
2053 wake_up_all();
2054 wake_up_bit();
2055 wake_up_interruptible();
2056 wake_up_interruptible_all();
2057 wake_up_interruptible_nr();
2058 wake_up_interruptible_poll();
2059 wake_up_interruptible_sync();
2060 wake_up_interruptible_sync_poll();
2061 wake_up_locked();
2062 wake_up_locked_poll();
2063 wake_up_nr();
2064 wake_up_poll();
2065 wake_up_process();
2066
2067
2068 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2069 order multiple stores before the wake-up with respect to loads of those stored
2070 values after the sleeper has called set_current_state(). For instance, if the
2071 sleeper does:
2072
2073 set_current_state(TASK_INTERRUPTIBLE);
2074 if (event_indicated)
2075 break;
2076 __set_current_state(TASK_RUNNING);
2077 do_something(my_data);
2078
2079 and the waker does:
2080
2081 my_data = value;
2082 event_indicated = 1;
2083 wake_up(&event_wait_queue);
2084
2085 there's no guarantee that the change to event_indicated will be perceived by
2086 the sleeper as coming after the change to my_data. In such a circumstance, the
2087 code on both sides must interpolate its own memory barriers between the
2088 separate data accesses. Thus the above sleeper ought to do:
2089
2090 set_current_state(TASK_INTERRUPTIBLE);
2091 if (event_indicated) {
2092 smp_rmb();
2093 do_something(my_data);
2094 }
2095
2096 and the waker should do:
2097
2098 my_data = value;
2099 smp_wmb();
2100 event_indicated = 1;
2101 wake_up(&event_wait_queue);
2102
2103
2104 MISCELLANEOUS FUNCTIONS
2105 -----------------------
2106
2107 Other functions that imply barriers:
2108
2109 (*) schedule() and similar imply full memory barriers.
2110
2111
2112 ===================================
2113 INTER-CPU ACQUIRING BARRIER EFFECTS
2114 ===================================
2115
2116 On SMP systems locking primitives give a more substantial form of barrier: one
2117 that does affect memory access ordering on other CPUs, within the context of
2118 conflict on any particular lock.
2119
2120
2121 ACQUIRES VS MEMORY ACCESSES
2122 ---------------------------
2123
2124 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2125 three CPUs; then should the following sequence of events occur:
2126
2127 CPU 1 CPU 2
2128 =============================== ===============================
2129 ACCESS_ONCE(*A) = a; ACCESS_ONCE(*E) = e;
2130 ACQUIRE M ACQUIRE Q
2131 ACCESS_ONCE(*B) = b; ACCESS_ONCE(*F) = f;
2132 ACCESS_ONCE(*C) = c; ACCESS_ONCE(*G) = g;
2133 RELEASE M RELEASE Q
2134 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*H) = h;
2135
2136 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2137 through *H occur in, other than the constraints imposed by the separate locks
2138 on the separate CPUs. It might, for example, see:
2139
2140 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2141
2142 But it won't see any of:
2143
2144 *B, *C or *D preceding ACQUIRE M
2145 *A, *B or *C following RELEASE M
2146 *F, *G or *H preceding ACQUIRE Q
2147 *E, *F or *G following RELEASE Q
2148
2149
2150 However, if the following occurs:
2151
2152 CPU 1 CPU 2
2153 =============================== ===============================
2154 ACCESS_ONCE(*A) = a;
2155 ACQUIRE M [1]
2156 ACCESS_ONCE(*B) = b;
2157 ACCESS_ONCE(*C) = c;
2158 RELEASE M [1]
2159 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*E) = e;
2160 ACQUIRE M [2]
2161 smp_mb__after_unlock_lock();
2162 ACCESS_ONCE(*F) = f;
2163 ACCESS_ONCE(*G) = g;
2164 RELEASE M [2]
2165 ACCESS_ONCE(*H) = h;
2166
2167 CPU 3 might see:
2168
2169 *E, ACQUIRE M [1], *C, *B, *A, RELEASE M [1],
2170 ACQUIRE M [2], *H, *F, *G, RELEASE M [2], *D
2171
2172 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
2173
2174 *B, *C, *D, *F, *G or *H preceding ACQUIRE M [1]
2175 *A, *B or *C following RELEASE M [1]
2176 *F, *G or *H preceding ACQUIRE M [2]
2177 *A, *B, *C, *E, *F or *G following RELEASE M [2]
2178
2179 Note that the smp_mb__after_unlock_lock() is critically important
2180 here: Without it CPU 3 might see some of the above orderings.
2181 Without smp_mb__after_unlock_lock(), the accesses are not guaranteed
2182 to be seen in order unless CPU 3 holds lock M.
2183
2184
2185 ACQUIRES VS I/O ACCESSES
2186 ------------------------
2187
2188 Under certain circumstances (especially involving NUMA), I/O accesses within
2189 two spinlocked sections on two different CPUs may be seen as interleaved by the
2190 PCI bridge, because the PCI bridge does not necessarily participate in the
2191 cache-coherence protocol, and is therefore incapable of issuing the required
2192 read memory barriers.
2193
2194 For example:
2195
2196 CPU 1 CPU 2
2197 =============================== ===============================
2198 spin_lock(Q)
2199 writel(0, ADDR)
2200 writel(1, DATA);
2201 spin_unlock(Q);
2202 spin_lock(Q);
2203 writel(4, ADDR);
2204 writel(5, DATA);
2205 spin_unlock(Q);
2206
2207 may be seen by the PCI bridge as follows:
2208
2209 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2210
2211 which would probably cause the hardware to malfunction.
2212
2213
2214 What is necessary here is to intervene with an mmiowb() before dropping the
2215 spinlock, for example:
2216
2217 CPU 1 CPU 2
2218 =============================== ===============================
2219 spin_lock(Q)
2220 writel(0, ADDR)
2221 writel(1, DATA);
2222 mmiowb();
2223 spin_unlock(Q);
2224 spin_lock(Q);
2225 writel(4, ADDR);
2226 writel(5, DATA);
2227 mmiowb();
2228 spin_unlock(Q);
2229
2230 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2231 before either of the stores issued on CPU 2.
2232
2233
2234 Furthermore, following a store by a load from the same device obviates the need
2235 for the mmiowb(), because the load forces the store to complete before the load
2236 is performed:
2237
2238 CPU 1 CPU 2
2239 =============================== ===============================
2240 spin_lock(Q)
2241 writel(0, ADDR)
2242 a = readl(DATA);
2243 spin_unlock(Q);
2244 spin_lock(Q);
2245 writel(4, ADDR);
2246 b = readl(DATA);
2247 spin_unlock(Q);
2248
2249
2250 See Documentation/DocBook/deviceiobook.tmpl for more information.
2251
2252
2253 =================================
2254 WHERE ARE MEMORY BARRIERS NEEDED?
2255 =================================
2256
2257 Under normal operation, memory operation reordering is generally not going to
2258 be a problem as a single-threaded linear piece of code will still appear to
2259 work correctly, even if it's in an SMP kernel. There are, however, four
2260 circumstances in which reordering definitely _could_ be a problem:
2261
2262 (*) Interprocessor interaction.
2263
2264 (*) Atomic operations.
2265
2266 (*) Accessing devices.
2267
2268 (*) Interrupts.
2269
2270
2271 INTERPROCESSOR INTERACTION
2272 --------------------------
2273
2274 When there's a system with more than one processor, more than one CPU in the
2275 system may be working on the same data set at the same time. This can cause
2276 synchronisation problems, and the usual way of dealing with them is to use
2277 locks. Locks, however, are quite expensive, and so it may be preferable to
2278 operate without the use of a lock if at all possible. In such a case
2279 operations that affect both CPUs may have to be carefully ordered to prevent
2280 a malfunction.
2281
2282 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2283 queued on the semaphore, by virtue of it having a piece of its stack linked to
2284 the semaphore's list of waiting processes:
2285
2286 struct rw_semaphore {
2287 ...
2288 spinlock_t lock;
2289 struct list_head waiters;
2290 };
2291
2292 struct rwsem_waiter {
2293 struct list_head list;
2294 struct task_struct *task;
2295 };
2296
2297 To wake up a particular waiter, the up_read() or up_write() functions have to:
2298
2299 (1) read the next pointer from this waiter's record to know as to where the
2300 next waiter record is;
2301
2302 (2) read the pointer to the waiter's task structure;
2303
2304 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2305
2306 (4) call wake_up_process() on the task; and
2307
2308 (5) release the reference held on the waiter's task struct.
2309
2310 In other words, it has to perform this sequence of events:
2311
2312 LOAD waiter->list.next;
2313 LOAD waiter->task;
2314 STORE waiter->task;
2315 CALL wakeup
2316 RELEASE task
2317
2318 and if any of these steps occur out of order, then the whole thing may
2319 malfunction.
2320
2321 Once it has queued itself and dropped the semaphore lock, the waiter does not
2322 get the lock again; it instead just waits for its task pointer to be cleared
2323 before proceeding. Since the record is on the waiter's stack, this means that
2324 if the task pointer is cleared _before_ the next pointer in the list is read,
2325 another CPU might start processing the waiter and might clobber the waiter's
2326 stack before the up*() function has a chance to read the next pointer.
2327
2328 Consider then what might happen to the above sequence of events:
2329
2330 CPU 1 CPU 2
2331 =============================== ===============================
2332 down_xxx()
2333 Queue waiter
2334 Sleep
2335 up_yyy()
2336 LOAD waiter->task;
2337 STORE waiter->task;
2338 Woken up by other event
2339 <preempt>
2340 Resume processing
2341 down_xxx() returns
2342 call foo()
2343 foo() clobbers *waiter
2344 </preempt>
2345 LOAD waiter->list.next;
2346 --- OOPS ---
2347
2348 This could be dealt with using the semaphore lock, but then the down_xxx()
2349 function has to needlessly get the spinlock again after being woken up.
2350
2351 The way to deal with this is to insert a general SMP memory barrier:
2352
2353 LOAD waiter->list.next;
2354 LOAD waiter->task;
2355 smp_mb();
2356 STORE waiter->task;
2357 CALL wakeup
2358 RELEASE task
2359
2360 In this case, the barrier makes a guarantee that all memory accesses before the
2361 barrier will appear to happen before all the memory accesses after the barrier
2362 with respect to the other CPUs on the system. It does _not_ guarantee that all
2363 the memory accesses before the barrier will be complete by the time the barrier
2364 instruction itself is complete.
2365
2366 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2367 compiler barrier, thus making sure the compiler emits the instructions in the
2368 right order without actually intervening in the CPU. Since there's only one
2369 CPU, that CPU's dependency ordering logic will take care of everything else.
2370
2371
2372 ATOMIC OPERATIONS
2373 -----------------
2374
2375 Whilst they are technically interprocessor interaction considerations, atomic
2376 operations are noted specially as some of them imply full memory barriers and
2377 some don't, but they're very heavily relied on as a group throughout the
2378 kernel.
2379
2380 Any atomic operation that modifies some state in memory and returns information
2381 about the state (old or new) implies an SMP-conditional general memory barrier
2382 (smp_mb()) on each side of the actual operation (with the exception of
2383 explicit lock operations, described later). These include:
2384
2385 xchg();
2386 cmpxchg();
2387 atomic_xchg(); atomic_long_xchg();
2388 atomic_cmpxchg(); atomic_long_cmpxchg();
2389 atomic_inc_return(); atomic_long_inc_return();
2390 atomic_dec_return(); atomic_long_dec_return();
2391 atomic_add_return(); atomic_long_add_return();
2392 atomic_sub_return(); atomic_long_sub_return();
2393 atomic_inc_and_test(); atomic_long_inc_and_test();
2394 atomic_dec_and_test(); atomic_long_dec_and_test();
2395 atomic_sub_and_test(); atomic_long_sub_and_test();
2396 atomic_add_negative(); atomic_long_add_negative();
2397 test_and_set_bit();
2398 test_and_clear_bit();
2399 test_and_change_bit();
2400
2401 /* when succeeds (returns 1) */
2402 atomic_add_unless(); atomic_long_add_unless();
2403
2404 These are used for such things as implementing ACQUIRE-class and RELEASE-class
2405 operations and adjusting reference counters towards object destruction, and as
2406 such the implicit memory barrier effects are necessary.
2407
2408
2409 The following operations are potential problems as they do _not_ imply memory
2410 barriers, but might be used for implementing such things as RELEASE-class
2411 operations:
2412
2413 atomic_set();
2414 set_bit();
2415 clear_bit();
2416 change_bit();
2417
2418 With these the appropriate explicit memory barrier should be used if necessary
2419 (smp_mb__before_atomic() for instance).
2420
2421
2422 The following also do _not_ imply memory barriers, and so may require explicit
2423 memory barriers under some circumstances (smp_mb__before_atomic() for
2424 instance):
2425
2426 atomic_add();
2427 atomic_sub();
2428 atomic_inc();
2429 atomic_dec();
2430
2431 If they're used for statistics generation, then they probably don't need memory
2432 barriers, unless there's a coupling between statistical data.
2433
2434 If they're used for reference counting on an object to control its lifetime,
2435 they probably don't need memory barriers because either the reference count
2436 will be adjusted inside a locked section, or the caller will already hold
2437 sufficient references to make the lock, and thus a memory barrier unnecessary.
2438
2439 If they're used for constructing a lock of some description, then they probably
2440 do need memory barriers as a lock primitive generally has to do things in a
2441 specific order.
2442
2443 Basically, each usage case has to be carefully considered as to whether memory
2444 barriers are needed or not.
2445
2446 The following operations are special locking primitives:
2447
2448 test_and_set_bit_lock();
2449 clear_bit_unlock();
2450 __clear_bit_unlock();
2451
2452 These implement ACQUIRE-class and RELEASE-class operations. These should be used in
2453 preference to other operations when implementing locking primitives, because
2454 their implementations can be optimised on many architectures.
2455
2456 [!] Note that special memory barrier primitives are available for these
2457 situations because on some CPUs the atomic instructions used imply full memory
2458 barriers, and so barrier instructions are superfluous in conjunction with them,
2459 and in such cases the special barrier primitives will be no-ops.
2460
2461 See Documentation/atomic_ops.txt for more information.
2462
2463
2464 ACCESSING DEVICES
2465 -----------------
2466
2467 Many devices can be memory mapped, and so appear to the CPU as if they're just
2468 a set of memory locations. To control such a device, the driver usually has to
2469 make the right memory accesses in exactly the right order.
2470
2471 However, having a clever CPU or a clever compiler creates a potential problem
2472 in that the carefully sequenced accesses in the driver code won't reach the
2473 device in the requisite order if the CPU or the compiler thinks it is more
2474 efficient to reorder, combine or merge accesses - something that would cause
2475 the device to malfunction.
2476
2477 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2478 routines - such as inb() or writel() - which know how to make such accesses
2479 appropriately sequential. Whilst this, for the most part, renders the explicit
2480 use of memory barriers unnecessary, there are a couple of situations where they
2481 might be needed:
2482
2483 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2484 so for _all_ general drivers locks should be used and mmiowb() must be
2485 issued prior to unlocking the critical section.
2486
2487 (2) If the accessor functions are used to refer to an I/O memory window with
2488 relaxed memory access properties, then _mandatory_ memory barriers are
2489 required to enforce ordering.
2490
2491 See Documentation/DocBook/deviceiobook.tmpl for more information.
2492
2493
2494 INTERRUPTS
2495 ----------
2496
2497 A driver may be interrupted by its own interrupt service routine, and thus the
2498 two parts of the driver may interfere with each other's attempts to control or
2499 access the device.
2500
2501 This may be alleviated - at least in part - by disabling local interrupts (a
2502 form of locking), such that the critical operations are all contained within
2503 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2504 routine is executing, the driver's core may not run on the same CPU, and its
2505 interrupt is not permitted to happen again until the current interrupt has been
2506 handled, thus the interrupt handler does not need to lock against that.
2507
2508 However, consider a driver that was talking to an ethernet card that sports an
2509 address register and a data register. If that driver's core talks to the card
2510 under interrupt-disablement and then the driver's interrupt handler is invoked:
2511
2512 LOCAL IRQ DISABLE
2513 writew(ADDR, 3);
2514 writew(DATA, y);
2515 LOCAL IRQ ENABLE
2516 <interrupt>
2517 writew(ADDR, 4);
2518 q = readw(DATA);
2519 </interrupt>
2520
2521 The store to the data register might happen after the second store to the
2522 address register if ordering rules are sufficiently relaxed:
2523
2524 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2525
2526
2527 If ordering rules are relaxed, it must be assumed that accesses done inside an
2528 interrupt disabled section may leak outside of it and may interleave with
2529 accesses performed in an interrupt - and vice versa - unless implicit or
2530 explicit barriers are used.
2531
2532 Normally this won't be a problem because the I/O accesses done inside such
2533 sections will include synchronous load operations on strictly ordered I/O
2534 registers that form implicit I/O barriers. If this isn't sufficient then an
2535 mmiowb() may need to be used explicitly.
2536
2537
2538 A similar situation may occur between an interrupt routine and two routines
2539 running on separate CPUs that communicate with each other. If such a case is
2540 likely, then interrupt-disabling locks should be used to guarantee ordering.
2541
2542
2543 ==========================
2544 KERNEL I/O BARRIER EFFECTS
2545 ==========================
2546
2547 When accessing I/O memory, drivers should use the appropriate accessor
2548 functions:
2549
2550 (*) inX(), outX():
2551
2552 These are intended to talk to I/O space rather than memory space, but
2553 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
2554 indeed have special I/O space access cycles and instructions, but many
2555 CPUs don't have such a concept.
2556
2557 The PCI bus, amongst others, defines an I/O space concept which - on such
2558 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2559 space. However, it may also be mapped as a virtual I/O space in the CPU's
2560 memory map, particularly on those CPUs that don't support alternate I/O
2561 spaces.
2562
2563 Accesses to this space may be fully synchronous (as on i386), but
2564 intermediary bridges (such as the PCI host bridge) may not fully honour
2565 that.
2566
2567 They are guaranteed to be fully ordered with respect to each other.
2568
2569 They are not guaranteed to be fully ordered with respect to other types of
2570 memory and I/O operation.
2571
2572 (*) readX(), writeX():
2573
2574 Whether these are guaranteed to be fully ordered and uncombined with
2575 respect to each other on the issuing CPU depends on the characteristics
2576 defined for the memory window through which they're accessing. On later
2577 i386 architecture machines, for example, this is controlled by way of the
2578 MTRR registers.
2579
2580 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2581 provided they're not accessing a prefetchable device.
2582
2583 However, intermediary hardware (such as a PCI bridge) may indulge in
2584 deferral if it so wishes; to flush a store, a load from the same location
2585 is preferred[*], but a load from the same device or from configuration
2586 space should suffice for PCI.
2587
2588 [*] NOTE! attempting to load from the same location as was written to may
2589 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2590 example.
2591
2592 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2593 force stores to be ordered.
2594
2595 Please refer to the PCI specification for more information on interactions
2596 between PCI transactions.
2597
2598 (*) readX_relaxed(), writeX_relaxed()
2599
2600 These are similar to readX() and writeX(), but provide weaker memory
2601 ordering guarantees. Specifically, they do not guarantee ordering with
2602 respect to normal memory accesses (e.g. DMA buffers) nor do they guarantee
2603 ordering with respect to LOCK or UNLOCK operations. If the latter is
2604 required, an mmiowb() barrier can be used. Note that relaxed accesses to
2605 the same peripheral are guaranteed to be ordered with respect to each
2606 other.
2607
2608 (*) ioreadX(), iowriteX()
2609
2610 These will perform appropriately for the type of access they're actually
2611 doing, be it inX()/outX() or readX()/writeX().
2612
2613
2614 ========================================
2615 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2616 ========================================
2617
2618 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2619 maintain the appearance of program causality with respect to itself. Some CPUs
2620 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2621 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2622 of arch-specific code.
2623
2624 This means that it must be considered that the CPU will execute its instruction
2625 stream in any order it feels like - or even in parallel - provided that if an
2626 instruction in the stream depends on an earlier instruction, then that
2627 earlier instruction must be sufficiently complete[*] before the later
2628 instruction may proceed; in other words: provided that the appearance of
2629 causality is maintained.
2630
2631 [*] Some instructions have more than one effect - such as changing the
2632 condition codes, changing registers or changing memory - and different
2633 instructions may depend on different effects.
2634
2635 A CPU may also discard any instruction sequence that winds up having no
2636 ultimate effect. For example, if two adjacent instructions both load an
2637 immediate value into the same register, the first may be discarded.
2638
2639
2640 Similarly, it has to be assumed that compiler might reorder the instruction
2641 stream in any way it sees fit, again provided the appearance of causality is
2642 maintained.
2643
2644
2645 ============================
2646 THE EFFECTS OF THE CPU CACHE
2647 ============================
2648
2649 The way cached memory operations are perceived across the system is affected to
2650 a certain extent by the caches that lie between CPUs and memory, and by the
2651 memory coherence system that maintains the consistency of state in the system.
2652
2653 As far as the way a CPU interacts with another part of the system through the
2654 caches goes, the memory system has to include the CPU's caches, and memory
2655 barriers for the most part act at the interface between the CPU and its cache
2656 (memory barriers logically act on the dotted line in the following diagram):
2657
2658 <--- CPU ---> : <----------- Memory ----------->
2659 :
2660 +--------+ +--------+ : +--------+ +-----------+
2661 | | | | : | | | | +--------+
2662 | CPU | | Memory | : | CPU | | | | |
2663 | Core |--->| Access |----->| Cache |<-->| | | |
2664 | | | Queue | : | | | |--->| Memory |
2665 | | | | : | | | | | |
2666 +--------+ +--------+ : +--------+ | | | |
2667 : | Cache | +--------+
2668 : | Coherency |
2669 : | Mechanism | +--------+
2670 +--------+ +--------+ : +--------+ | | | |
2671 | | | | : | | | | | |
2672 | CPU | | Memory | : | CPU | | |--->| Device |
2673 | Core |--->| Access |----->| Cache |<-->| | | |
2674 | | | Queue | : | | | | | |
2675 | | | | : | | | | +--------+
2676 +--------+ +--------+ : +--------+ +-----------+
2677 :
2678 :
2679
2680 Although any particular load or store may not actually appear outside of the
2681 CPU that issued it since it may have been satisfied within the CPU's own cache,
2682 it will still appear as if the full memory access had taken place as far as the
2683 other CPUs are concerned since the cache coherency mechanisms will migrate the
2684 cacheline over to the accessing CPU and propagate the effects upon conflict.
2685
2686 The CPU core may execute instructions in any order it deems fit, provided the
2687 expected program causality appears to be maintained. Some of the instructions
2688 generate load and store operations which then go into the queue of memory
2689 accesses to be performed. The core may place these in the queue in any order
2690 it wishes, and continue execution until it is forced to wait for an instruction
2691 to complete.
2692
2693 What memory barriers are concerned with is controlling the order in which
2694 accesses cross from the CPU side of things to the memory side of things, and
2695 the order in which the effects are perceived to happen by the other observers
2696 in the system.
2697
2698 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2699 their own loads and stores as if they had happened in program order.
2700
2701 [!] MMIO or other device accesses may bypass the cache system. This depends on
2702 the properties of the memory window through which devices are accessed and/or
2703 the use of any special device communication instructions the CPU may have.
2704
2705
2706 CACHE COHERENCY
2707 ---------------
2708
2709 Life isn't quite as simple as it may appear above, however: for while the
2710 caches are expected to be coherent, there's no guarantee that that coherency
2711 will be ordered. This means that whilst changes made on one CPU will
2712 eventually become visible on all CPUs, there's no guarantee that they will
2713 become apparent in the same order on those other CPUs.
2714
2715
2716 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2717 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2718
2719 :
2720 : +--------+
2721 : +---------+ | |
2722 +--------+ : +--->| Cache A |<------->| |
2723 | | : | +---------+ | |
2724 | CPU 1 |<---+ | |
2725 | | : | +---------+ | |
2726 +--------+ : +--->| Cache B |<------->| |
2727 : +---------+ | |
2728 : | Memory |
2729 : +---------+ | System |
2730 +--------+ : +--->| Cache C |<------->| |
2731 | | : | +---------+ | |
2732 | CPU 2 |<---+ | |
2733 | | : | +---------+ | |
2734 +--------+ : +--->| Cache D |<------->| |
2735 : +---------+ | |
2736 : +--------+
2737 :
2738
2739 Imagine the system has the following properties:
2740
2741 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2742 resident in memory;
2743
2744 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2745 resident in memory;
2746
2747 (*) whilst the CPU core is interrogating one cache, the other cache may be
2748 making use of the bus to access the rest of the system - perhaps to
2749 displace a dirty cacheline or to do a speculative load;
2750
2751 (*) each cache has a queue of operations that need to be applied to that cache
2752 to maintain coherency with the rest of the system;
2753
2754 (*) the coherency queue is not flushed by normal loads to lines already
2755 present in the cache, even though the contents of the queue may
2756 potentially affect those loads.
2757
2758 Imagine, then, that two writes are made on the first CPU, with a write barrier
2759 between them to guarantee that they will appear to reach that CPU's caches in
2760 the requisite order:
2761
2762 CPU 1 CPU 2 COMMENT
2763 =============== =============== =======================================
2764 u == 0, v == 1 and p == &u, q == &u
2765 v = 2;
2766 smp_wmb(); Make sure change to v is visible before
2767 change to p
2768 <A:modify v=2> v is now in cache A exclusively
2769 p = &v;
2770 <B:modify p=&v> p is now in cache B exclusively
2771
2772 The write memory barrier forces the other CPUs in the system to perceive that
2773 the local CPU's caches have apparently been updated in the correct order. But
2774 now imagine that the second CPU wants to read those values:
2775
2776 CPU 1 CPU 2 COMMENT
2777 =============== =============== =======================================
2778 ...
2779 q = p;
2780 x = *q;
2781
2782 The above pair of reads may then fail to happen in the expected order, as the
2783 cacheline holding p may get updated in one of the second CPU's caches whilst
2784 the update to the cacheline holding v is delayed in the other of the second
2785 CPU's caches by some other cache event:
2786
2787 CPU 1 CPU 2 COMMENT
2788 =============== =============== =======================================
2789 u == 0, v == 1 and p == &u, q == &u
2790 v = 2;
2791 smp_wmb();
2792 <A:modify v=2> <C:busy>
2793 <C:queue v=2>
2794 p = &v; q = p;
2795 <D:request p>
2796 <B:modify p=&v> <D:commit p=&v>
2797 <D:read p>
2798 x = *q;
2799 <C:read *q> Reads from v before v updated in cache
2800 <C:unbusy>
2801 <C:commit v=2>
2802
2803 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2804 no guarantee that, without intervention, the order of update will be the same
2805 as that committed on CPU 1.
2806
2807
2808 To intervene, we need to interpolate a data dependency barrier or a read
2809 barrier between the loads. This will force the cache to commit its coherency
2810 queue before processing any further requests:
2811
2812 CPU 1 CPU 2 COMMENT
2813 =============== =============== =======================================
2814 u == 0, v == 1 and p == &u, q == &u
2815 v = 2;
2816 smp_wmb();
2817 <A:modify v=2> <C:busy>
2818 <C:queue v=2>
2819 p = &v; q = p;
2820 <D:request p>
2821 <B:modify p=&v> <D:commit p=&v>
2822 <D:read p>
2823 smp_read_barrier_depends()
2824 <C:unbusy>
2825 <C:commit v=2>
2826 x = *q;
2827 <C:read *q> Reads from v after v updated in cache
2828
2829
2830 This sort of problem can be encountered on DEC Alpha processors as they have a
2831 split cache that improves performance by making better use of the data bus.
2832 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2833 access depends on a read, not all do, so it may not be relied on.
2834
2835 Other CPUs may also have split caches, but must coordinate between the various
2836 cachelets for normal memory accesses. The semantics of the Alpha removes the
2837 need for coordination in the absence of memory barriers.
2838
2839
2840 CACHE COHERENCY VS DMA
2841 ----------------------
2842
2843 Not all systems maintain cache coherency with respect to devices doing DMA. In
2844 such cases, a device attempting DMA may obtain stale data from RAM because
2845 dirty cache lines may be resident in the caches of various CPUs, and may not
2846 have been written back to RAM yet. To deal with this, the appropriate part of
2847 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2848 invalidate them as well).
2849
2850 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2851 cache lines being written back to RAM from a CPU's cache after the device has
2852 installed its own data, or cache lines present in the CPU's cache may simply
2853 obscure the fact that RAM has been updated, until at such time as the cacheline
2854 is discarded from the CPU's cache and reloaded. To deal with this, the
2855 appropriate part of the kernel must invalidate the overlapping bits of the
2856 cache on each CPU.
2857
2858 See Documentation/cachetlb.txt for more information on cache management.
2859
2860
2861 CACHE COHERENCY VS MMIO
2862 -----------------------
2863
2864 Memory mapped I/O usually takes place through memory locations that are part of
2865 a window in the CPU's memory space that has different properties assigned than
2866 the usual RAM directed window.
2867
2868 Amongst these properties is usually the fact that such accesses bypass the
2869 caching entirely and go directly to the device buses. This means MMIO accesses
2870 may, in effect, overtake accesses to cached memory that were emitted earlier.
2871 A memory barrier isn't sufficient in such a case, but rather the cache must be
2872 flushed between the cached memory write and the MMIO access if the two are in
2873 any way dependent.
2874
2875
2876 =========================
2877 THE THINGS CPUS GET UP TO
2878 =========================
2879
2880 A programmer might take it for granted that the CPU will perform memory
2881 operations in exactly the order specified, so that if the CPU is, for example,
2882 given the following piece of code to execute:
2883
2884 a = ACCESS_ONCE(*A);
2885 ACCESS_ONCE(*B) = b;
2886 c = ACCESS_ONCE(*C);
2887 d = ACCESS_ONCE(*D);
2888 ACCESS_ONCE(*E) = e;
2889
2890 they would then expect that the CPU will complete the memory operation for each
2891 instruction before moving on to the next one, leading to a definite sequence of
2892 operations as seen by external observers in the system:
2893
2894 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2895
2896
2897 Reality is, of course, much messier. With many CPUs and compilers, the above
2898 assumption doesn't hold because:
2899
2900 (*) loads are more likely to need to be completed immediately to permit
2901 execution progress, whereas stores can often be deferred without a
2902 problem;
2903
2904 (*) loads may be done speculatively, and the result discarded should it prove
2905 to have been unnecessary;
2906
2907 (*) loads may be done speculatively, leading to the result having been fetched
2908 at the wrong time in the expected sequence of events;
2909
2910 (*) the order of the memory accesses may be rearranged to promote better use
2911 of the CPU buses and caches;
2912
2913 (*) loads and stores may be combined to improve performance when talking to
2914 memory or I/O hardware that can do batched accesses of adjacent locations,
2915 thus cutting down on transaction setup costs (memory and PCI devices may
2916 both be able to do this); and
2917
2918 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2919 mechanisms may alleviate this - once the store has actually hit the cache
2920 - there's no guarantee that the coherency management will be propagated in
2921 order to other CPUs.
2922
2923 So what another CPU, say, might actually observe from the above piece of code
2924 is:
2925
2926 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2927
2928 (Where "LOAD {*C,*D}" is a combined load)
2929
2930
2931 However, it is guaranteed that a CPU will be self-consistent: it will see its
2932 _own_ accesses appear to be correctly ordered, without the need for a memory
2933 barrier. For instance with the following code:
2934
2935 U = ACCESS_ONCE(*A);
2936 ACCESS_ONCE(*A) = V;
2937 ACCESS_ONCE(*A) = W;
2938 X = ACCESS_ONCE(*A);
2939 ACCESS_ONCE(*A) = Y;
2940 Z = ACCESS_ONCE(*A);
2941
2942 and assuming no intervention by an external influence, it can be assumed that
2943 the final result will appear to be:
2944
2945 U == the original value of *A
2946 X == W
2947 Z == Y
2948 *A == Y
2949
2950 The code above may cause the CPU to generate the full sequence of memory
2951 accesses:
2952
2953 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2954
2955 in that order, but, without intervention, the sequence may have almost any
2956 combination of elements combined or discarded, provided the program's view of
2957 the world remains consistent. Note that ACCESS_ONCE() is -not- optional
2958 in the above example, as there are architectures where a given CPU might
2959 reorder successive loads to the same location. On such architectures,
2960 ACCESS_ONCE() does whatever is necessary to prevent this, for example, on
2961 Itanium the volatile casts used by ACCESS_ONCE() cause GCC to emit the
2962 special ld.acq and st.rel instructions that prevent such reordering.
2963
2964 The compiler may also combine, discard or defer elements of the sequence before
2965 the CPU even sees them.
2966
2967 For instance:
2968
2969 *A = V;
2970 *A = W;
2971
2972 may be reduced to:
2973
2974 *A = W;
2975
2976 since, without either a write barrier or an ACCESS_ONCE(), it can be
2977 assumed that the effect of the storage of V to *A is lost. Similarly:
2978
2979 *A = Y;
2980 Z = *A;
2981
2982 may, without a memory barrier or an ACCESS_ONCE(), be reduced to:
2983
2984 *A = Y;
2985 Z = Y;
2986
2987 and the LOAD operation never appear outside of the CPU.
2988
2989
2990 AND THEN THERE'S THE ALPHA
2991 --------------------------
2992
2993 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2994 some versions of the Alpha CPU have a split data cache, permitting them to have
2995 two semantically-related cache lines updated at separate times. This is where
2996 the data dependency barrier really becomes necessary as this synchronises both
2997 caches with the memory coherence system, thus making it seem like pointer
2998 changes vs new data occur in the right order.
2999
3000 The Alpha defines the Linux kernel's memory barrier model.
3001
3002 See the subsection on "Cache Coherency" above.
3003
3004
3005 ============
3006 EXAMPLE USES
3007 ============
3008
3009 CIRCULAR BUFFERS
3010 ----------------
3011
3012 Memory barriers can be used to implement circular buffering without the need
3013 of a lock to serialise the producer with the consumer. See:
3014
3015 Documentation/circular-buffers.txt
3016
3017 for details.
3018
3019
3020 ==========
3021 REFERENCES
3022 ==========
3023
3024 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
3025 Digital Press)
3026 Chapter 5.2: Physical Address Space Characteristics
3027 Chapter 5.4: Caches and Write Buffers
3028 Chapter 5.5: Data Sharing
3029 Chapter 5.6: Read/Write Ordering
3030
3031 AMD64 Architecture Programmer's Manual Volume 2: System Programming
3032 Chapter 7.1: Memory-Access Ordering
3033 Chapter 7.4: Buffering and Combining Memory Writes
3034
3035 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
3036 System Programming Guide
3037 Chapter 7.1: Locked Atomic Operations
3038 Chapter 7.2: Memory Ordering
3039 Chapter 7.4: Serializing Instructions
3040
3041 The SPARC Architecture Manual, Version 9
3042 Chapter 8: Memory Models
3043 Appendix D: Formal Specification of the Memory Models
3044 Appendix J: Programming with the Memory Models
3045
3046 UltraSPARC Programmer Reference Manual
3047 Chapter 5: Memory Accesses and Cacheability
3048 Chapter 15: Sparc-V9 Memory Models
3049
3050 UltraSPARC III Cu User's Manual
3051 Chapter 9: Memory Models
3052
3053 UltraSPARC IIIi Processor User's Manual
3054 Chapter 8: Memory Models
3055
3056 UltraSPARC Architecture 2005
3057 Chapter 9: Memory
3058 Appendix D: Formal Specifications of the Memory Models
3059
3060 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
3061 Chapter 8: Memory Models
3062 Appendix F: Caches and Cache Coherency
3063
3064 Solaris Internals, Core Kernel Architecture, p63-68:
3065 Chapter 3.3: Hardware Considerations for Locks and
3066 Synchronization
3067
3068 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3069 for Kernel Programmers:
3070 Chapter 13: Other Memory Models
3071
3072 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3073 Section 2.6: Speculation
3074 Section 4.4: Memory Access