| blob | ec3b5865c1bec7e5b051b33be1f4805fff222b4e |
1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
5 By: David Howells <dhowells@redhat.com>
6 Paul E. McKenney <paulmck@linux.ibm.com>
7 Will Deacon <will.deacon@arm.com>
8 Peter Zijlstra <peterz@infradead.org>
10 ==========
11 DISCLAIMER
12 ==========
14 This document is not a specification; it is intentionally (for the sake of
15 brevity) and unintentionally (due to being human) incomplete. This document is
16 meant as a guide to using the various memory barriers provided by Linux, but
17 in case of any doubt (and there are many) please ask. Some doubts may be
18 resolved by referring to the formal memory consistency model and related
19 documentation at tools/memory-model/. Nevertheless, even this memory
20 model should be viewed as the collective opinion of its maintainers rather
21 than as an infallible oracle.
23 To repeat, this document is not a specification of what Linux expects from
24 hardware.
26 The purpose of this document is twofold:
28 (1) to specify the minimum functionality that one can rely on for any
29 particular barrier, and
31 (2) to provide a guide as to how to use the barriers that are available.
33 Note that an architecture can provide more than the minimum requirement
34 for any particular barrier, but if the architecture provides less than
35 that, that architecture is incorrect.
37 Note also that it is possible that a barrier may be a no-op for an
38 architecture because the way that arch works renders an explicit barrier
39 unnecessary in that case.
42 ========
43 CONTENTS
44 ========
46 (*) Abstract memory access model.
48 - Device operations.
49 - Guarantees.
51 (*) What are memory barriers?
53 - Varieties of memory barrier.
54 - What may not be assumed about memory barriers?
55 - Data dependency barriers (historical).
56 - Control dependencies.
57 - SMP barrier pairing.
58 - Examples of memory barrier sequences.
59 - Read memory barriers vs load speculation.
60 - Multicopy atomicity.
62 (*) Explicit kernel barriers.
64 - Compiler barrier.
65 - CPU memory barriers.
67 (*) Implicit kernel memory barriers.
69 - Lock acquisition functions.
70 - Interrupt disabling functions.
71 - Sleep and wake-up functions.
72 - Miscellaneous functions.
74 (*) Inter-CPU acquiring barrier effects.
76 - Acquires vs memory accesses.
78 (*) Where are memory barriers needed?
80 - Interprocessor interaction.
81 - Atomic operations.
82 - Accessing devices.
83 - Interrupts.
85 (*) Kernel I/O barrier effects.
87 (*) Assumed minimum execution ordering model.
89 (*) The effects of the cpu cache.
91 - Cache coherency.
92 - Cache coherency vs DMA.
93 - Cache coherency vs MMIO.
95 (*) The things CPUs get up to.
97 - And then there's the Alpha.
98 - Virtual Machine Guests.
100 (*) Example uses.
102 - Circular buffers.
104 (*) References.
107 ============================
108 ABSTRACT MEMORY ACCESS MODEL
109 ============================
111 Consider the following abstract model of the system:
113 : :
114 : :
115 : :
116 +-------+ : +--------+ : +-------+
117 | | : | | : | |
118 | | : | | : | |
119 | CPU 1 |<----->| Memory |<----->| CPU 2 |
120 | | : | | : | |
121 | | : | | : | |
122 +-------+ : +--------+ : +-------+
123 ^ : ^ : ^
124 | : | : |
125 | : | : |
126 | : v : |
127 | : +--------+ : |
128 | : | | : |
129 | : | | : |
130 +---------->| Device |<----------+
131 : | | :
132 : | | :
133 : +--------+ :
134 : :
136 Each CPU executes a program that generates memory access operations. In the
137 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
138 perform the memory operations in any order it likes, provided program causality
139 appears to be maintained. Similarly, the compiler may also arrange the
140 instructions it emits in any order it likes, provided it doesn't affect the
141 apparent operation of the program.
143 So in the above diagram, the effects of the memory operations performed by a
144 CPU are perceived by the rest of the system as the operations cross the
145 interface between the CPU and rest of the system (the dotted lines).
148 For example, consider the following sequence of events:
150 CPU 1 CPU 2
151 =============== ===============
152 { A == 1; B == 2 }
153 A = 3; x = B;
154 B = 4; y = A;
156 The set of accesses as seen by the memory system in the middle can be arranged
157 in 24 different combinations:
159 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4
160 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3
161 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4
162 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4
163 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3
164 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4
165 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4
166 STORE B=4, ...
167 ...
169 and can thus result in four different combinations of values:
171 x == 2, y == 1
172 x == 2, y == 3
173 x == 4, y == 1
174 x == 4, y == 3
177 Furthermore, the stores committed by a CPU to the memory system may not be
178 perceived by the loads made by another CPU in the same order as the stores were
179 committed.
182 As a further example, consider this sequence of events:
184 CPU 1 CPU 2
185 =============== ===============
186 { A == 1, B == 2, C == 3, P == &A, Q == &C }
187 B = 4; Q = P;
188 P = &B D = *Q;
190 There is an obvious data dependency here, as the value loaded into D depends on
191 the address retrieved from P by CPU 2. At the end of the sequence, any of the
192 following results are possible:
194 (Q == &A) and (D == 1)
195 (Q == &B) and (D == 2)
196 (Q == &B) and (D == 4)
198 Note that CPU 2 will never try and load C into D because the CPU will load P
199 into Q before issuing the load of *Q.
202 DEVICE OPERATIONS
203 -----------------
205 Some devices present their control interfaces as collections of memory
206 locations, but the order in which the control registers are accessed is very
207 important. For instance, imagine an ethernet card with a set of internal
208 registers that are accessed through an address port register (A) and a data
209 port register (D). To read internal register 5, the following code might then
210 be used:
212 *A = 5;
213 x = *D;
215 but this might show up as either of the following two sequences:
217 STORE *A = 5, x = LOAD *D
218 x = LOAD *D, STORE *A = 5
220 the second of which will almost certainly result in a malfunction, since it set
221 the address _after_ attempting to read the register.
224 GUARANTEES
225 ----------
227 There are some minimal guarantees that may be expected of a CPU:
229 (*) On any given CPU, dependent memory accesses will be issued in order, with
230 respect to itself. This means that for:
232 Q = READ_ONCE(P); D = READ_ONCE(*Q);
234 the CPU will issue the following memory operations:
236 Q = LOAD P, D = LOAD *Q
238 and always in that order. However, on DEC Alpha, READ_ONCE() also
239 emits a memory-barrier instruction, so that a DEC Alpha CPU will
240 instead issue the following memory operations:
242 Q = LOAD P, MEMORY_BARRIER, D = LOAD *Q, MEMORY_BARRIER
244 Whether on DEC Alpha or not, the READ_ONCE() also prevents compiler
245 mischief.
247 (*) Overlapping loads and stores within a particular CPU will appear to be
248 ordered within that CPU. This means that for:
250 a = READ_ONCE(*X); WRITE_ONCE(*X, b);
252 the CPU will only issue the following sequence of memory operations:
254 a = LOAD *X, STORE *X = b
256 And for:
258 WRITE_ONCE(*X, c); d = READ_ONCE(*X);
260 the CPU will only issue:
262 STORE *X = c, d = LOAD *X
264 (Loads and stores overlap if they are targeted at overlapping pieces of
265 memory).
267 And there are a number of things that _must_ or _must_not_ be assumed:
269 (*) It _must_not_ be assumed that the compiler will do what you want
270 with memory references that are not protected by READ_ONCE() and
271 WRITE_ONCE(). Without them, the compiler is within its rights to
272 do all sorts of "creative" transformations, which are covered in
273 the COMPILER BARRIER section.
275 (*) It _must_not_ be assumed that independent loads and stores will be issued
276 in the order given. This means that for:
278 X = *A; Y = *B; *D = Z;
280 we may get any of the following sequences:
282 X = LOAD *A, Y = LOAD *B, STORE *D = Z
283 X = LOAD *A, STORE *D = Z, Y = LOAD *B
284 Y = LOAD *B, X = LOAD *A, STORE *D = Z
285 Y = LOAD *B, STORE *D = Z, X = LOAD *A
286 STORE *D = Z, X = LOAD *A, Y = LOAD *B
287 STORE *D = Z, Y = LOAD *B, X = LOAD *A
289 (*) It _must_ be assumed that overlapping memory accesses may be merged or
290 discarded. This means that for:
292 X = *A; Y = *(A + 4);
294 we may get any one of the following sequences:
296 X = LOAD *A; Y = LOAD *(A + 4);
297 Y = LOAD *(A + 4); X = LOAD *A;
298 {X, Y} = LOAD {*A, *(A + 4) };
300 And for:
302 *A = X; *(A + 4) = Y;
304 we may get any of:
306 STORE *A = X; STORE *(A + 4) = Y;
307 STORE *(A + 4) = Y; STORE *A = X;
308 STORE {*A, *(A + 4) } = {X, Y};
310 And there are anti-guarantees:
312 (*) These guarantees do not apply to bitfields, because compilers often
313 generate code to modify these using non-atomic read-modify-write
314 sequences. Do not attempt to use bitfields to synchronize parallel
315 algorithms.
317 (*) Even in cases where bitfields are protected by locks, all fields
318 in a given bitfield must be protected by one lock. If two fields
319 in a given bitfield are protected by different locks, the compiler's
320 non-atomic read-modify-write sequences can cause an update to one
321 field to corrupt the value of an adjacent field.
323 (*) These guarantees apply only to properly aligned and sized scalar
324 variables. "Properly sized" currently means variables that are
325 the same size as "char", "short", "int" and "long". "Properly
326 aligned" means the natural alignment, thus no constraints for
327 "char", two-byte alignment for "short", four-byte alignment for
328 "int", and either four-byte or eight-byte alignment for "long",
329 on 32-bit and 64-bit systems, respectively. Note that these
330 guarantees were introduced into the C11 standard, so beware when
331 using older pre-C11 compilers (for example, gcc 4.6). The portion
332 of the standard containing this guarantee is Section 3.14, which
333 defines "memory location" as follows:
335 memory location
336 either an object of scalar type, or a maximal sequence
337 of adjacent bit-fields all having nonzero width
339 NOTE 1: Two threads of execution can update and access
340 separate memory locations without interfering with
341 each other.
343 NOTE 2: A bit-field and an adjacent non-bit-field member
344 are in separate memory locations. The same applies
345 to two bit-fields, if one is declared inside a nested
346 structure declaration and the other is not, or if the two
347 are separated by a zero-length bit-field declaration,
348 or if they are separated by a non-bit-field member
349 declaration. It is not safe to concurrently update two
350 bit-fields in the same structure if all members declared
351 between them are also bit-fields, no matter what the
352 sizes of those intervening bit-fields happen to be.
355 =========================
356 WHAT ARE MEMORY BARRIERS?
357 =========================
359 As can be seen above, independent memory operations are effectively performed
360 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
361 What is required is some way of intervening to instruct the compiler and the
362 CPU to restrict the order.
364 Memory barriers are such interventions. They impose a perceived partial
365 ordering over the memory operations on either side of the barrier.
367 Such enforcement is important because the CPUs and other devices in a system
368 can use a variety of tricks to improve performance, including reordering,
369 deferral and combination of memory operations; speculative loads; speculative
370 branch prediction and various types of caching. Memory barriers are used to
371 override or suppress these tricks, allowing the code to sanely control the
372 interaction of multiple CPUs and/or devices.
375 VARIETIES OF MEMORY BARRIER
376 ---------------------------
378 Memory barriers come in four basic varieties:
380 (1) Write (or store) memory barriers.
382 A write memory barrier gives a guarantee that all the STORE operations
383 specified before the barrier will appear to happen before all the STORE
384 operations specified after the barrier with respect to the other
385 components of the system.
387 A write barrier is a partial ordering on stores only; it is not required
388 to have any effect on loads.
390 A CPU can be viewed as committing a sequence of store operations to the
391 memory system as time progresses. All stores _before_ a write barrier
392 will occur _before_ all the stores after the write barrier.
394 [!] Note that write barriers should normally be paired with read or data
395 dependency barriers; see the "SMP barrier pairing" subsection.
398 (2) Data dependency barriers.
400 A data dependency barrier is a weaker form of read barrier. In the case
401 where two loads are performed such that the second depends on the result
402 of the first (eg: the first load retrieves the address to which the second
403 load will be directed), a data dependency barrier would be required to
404 make sure that the target of the second load is updated after the address
405 obtained by the first load is accessed.
407 A data dependency barrier is a partial ordering on interdependent loads
408 only; it is not required to have any effect on stores, independent loads
409 or overlapping loads.
411 As mentioned in (1), the other CPUs in the system can be viewed as
412 committing sequences of stores to the memory system that the CPU being
413 considered can then perceive. A data dependency barrier issued by the CPU
414 under consideration guarantees that for any load preceding it, if that
415 load touches one of a sequence of stores from another CPU, then by the
416 time the barrier completes, the effects of all the stores prior to that
417 touched by the load will be perceptible to any loads issued after the data
418 dependency barrier.
420 See the "Examples of memory barrier sequences" subsection for diagrams
421 showing the ordering constraints.
423 [!] Note that the first load really has to have a _data_ dependency and
424 not a control dependency. If the address for the second load is dependent
425 on the first load, but the dependency is through a conditional rather than
426 actually loading the address itself, then it's a _control_ dependency and
427 a full read barrier or better is required. See the "Control dependencies"
428 subsection for more information.
430 [!] Note that data dependency barriers should normally be paired with
431 write barriers; see the "SMP barrier pairing" subsection.
434 (3) Read (or load) memory barriers.
436 A read barrier is a data dependency barrier plus a guarantee that all the
437 LOAD operations specified before the barrier will appear to happen before
438 all the LOAD operations specified after the barrier with respect to the
439 other components of the system.
441 A read barrier is a partial ordering on loads only; it is not required to
442 have any effect on stores.
444 Read memory barriers imply data dependency barriers, and so can substitute
445 for them.
447 [!] Note that read barriers should normally be paired with write barriers;
448 see the "SMP barrier pairing" subsection.
451 (4) General memory barriers.
453 A general memory barrier gives a guarantee that all the LOAD and STORE
454 operations specified before the barrier will appear to happen before all
455 the LOAD and STORE operations specified after the barrier with respect to
456 the other components of the system.
458 A general memory barrier is a partial ordering over both loads and stores.
460 General memory barriers imply both read and write memory barriers, and so
461 can substitute for either.
464 And a couple of implicit varieties:
466 (5) ACQUIRE operations.
468 This acts as a one-way permeable barrier. It guarantees that all memory
469 operations after the ACQUIRE operation will appear to happen after the
470 ACQUIRE operation with respect to the other components of the system.
471 ACQUIRE operations include LOCK operations and both smp_load_acquire()
472 and smp_cond_load_acquire() operations.
474 Memory operations that occur before an ACQUIRE operation may appear to
475 happen after it completes.
477 An ACQUIRE operation should almost always be paired with a RELEASE
478 operation.
481 (6) RELEASE operations.
483 This also acts as a one-way permeable barrier. It guarantees that all
484 memory operations before the RELEASE operation will appear to happen
485 before the RELEASE operation with respect to the other components of the
486 system. RELEASE operations include UNLOCK operations and
487 smp_store_release() operations.
489 Memory operations that occur after a RELEASE operation may appear to
490 happen before it completes.
492 The use of ACQUIRE and RELEASE operations generally precludes the need
493 for other sorts of memory barrier. In addition, a RELEASE+ACQUIRE pair is
494 -not- guaranteed to act as a full memory barrier. However, after an
495 ACQUIRE on a given variable, all memory accesses preceding any prior
496 RELEASE on that same variable are guaranteed to be visible. In other
497 words, within a given variable's critical section, all accesses of all
498 previous critical sections for that variable are guaranteed to have
499 completed.
501 This means that ACQUIRE acts as a minimal "acquire" operation and
502 RELEASE acts as a minimal "release" operation.
504 A subset of the atomic operations described in atomic_t.txt have ACQUIRE and
505 RELEASE variants in addition to fully-ordered and relaxed (no barrier
506 semantics) definitions. For compound atomics performing both a load and a
507 store, ACQUIRE semantics apply only to the load and RELEASE semantics apply
508 only to the store portion of the operation.
510 Memory barriers are only required where there's a possibility of interaction
511 between two CPUs or between a CPU and a device. If it can be guaranteed that
512 there won't be any such interaction in any particular piece of code, then
513 memory barriers are unnecessary in that piece of code.
516 Note that these are the _minimum_ guarantees. Different architectures may give
517 more substantial guarantees, but they may _not_ be relied upon outside of arch
518 specific code.
521 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
522 ----------------------------------------------
524 There are certain things that the Linux kernel memory barriers do not guarantee:
526 (*) There is no guarantee that any of the memory accesses specified before a
527 memory barrier will be _complete_ by the completion of a memory barrier
528 instruction; the barrier can be considered to draw a line in that CPU's
529 access queue that accesses of the appropriate type may not cross.
531 (*) There is no guarantee that issuing a memory barrier on one CPU will have
532 any direct effect on another CPU or any other hardware in the system. The
533 indirect effect will be the order in which the second CPU sees the effects
534 of the first CPU's accesses occur, but see the next point:
536 (*) There is no guarantee that a CPU will see the correct order of effects
537 from a second CPU's accesses, even _if_ the second CPU uses a memory
538 barrier, unless the first CPU _also_ uses a matching memory barrier (see
539 the subsection on "SMP Barrier Pairing").
541 (*) There is no guarantee that some intervening piece of off-the-CPU
542 hardware[*] will not reorder the memory accesses. CPU cache coherency
543 mechanisms should propagate the indirect effects of a memory barrier
544 between CPUs, but might not do so in order.
546 [*] For information on bus mastering DMA and coherency please read:
548 Documentation/driver-api/pci/pci.rst
549 Documentation/DMA-API-HOWTO.txt
550 Documentation/DMA-API.txt
553 DATA DEPENDENCY BARRIERS (HISTORICAL)
554 -------------------------------------
556 As of v4.15 of the Linux kernel, an smp_read_barrier_depends() was
557 added to READ_ONCE(), which means that about the only people who
558 need to pay attention to this section are those working on DEC Alpha
559 architecture-specific code and those working on READ_ONCE() itself.
560 For those who need it, and for those who are interested in the history,
561 here is the story of data-dependency barriers.
563 The usage requirements of data dependency barriers are a little subtle, and
564 it's not always obvious that they're needed. To illustrate, consider the
565 following sequence of events:
567 CPU 1 CPU 2
568 =============== ===============
569 { A == 1, B == 2, C == 3, P == &A, Q == &C }
570 B = 4;
571 <write barrier>
572 WRITE_ONCE(P, &B)
573 Q = READ_ONCE(P);
574 D = *Q;
576 There's a clear data dependency here, and it would seem that by the end of the
577 sequence, Q must be either &A or &B, and that:
579 (Q == &A) implies (D == 1)
580 (Q == &B) implies (D == 4)
582 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
583 leading to the following situation:
585 (Q == &B) and (D == 2) ????
587 While this may seem like a failure of coherency or causality maintenance, it
588 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
589 Alpha).
591 To deal with this, a data dependency barrier or better must be inserted
592 between the address load and the data load:
594 CPU 1 CPU 2
595 =============== ===============
596 { A == 1, B == 2, C == 3, P == &A, Q == &C }
597 B = 4;
598 <write barrier>
599 WRITE_ONCE(P, &B);
600 Q = READ_ONCE(P);
601 <data dependency barrier>
602 D = *Q;
604 This enforces the occurrence of one of the two implications, and prevents the
605 third possibility from arising.
608 [!] Note that this extremely counterintuitive situation arises most easily on
609 machines with split caches, so that, for example, one cache bank processes
610 even-numbered cache lines and the other bank processes odd-numbered cache
611 lines. The pointer P might be stored in an odd-numbered cache line, and the
612 variable B might be stored in an even-numbered cache line. Then, if the
613 even-numbered bank of the reading CPU's cache is extremely busy while the
614 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
615 but the old value of the variable B (2).
618 A data-dependency barrier is not required to order dependent writes
619 because the CPUs that the Linux kernel supports don't do writes
620 until they are certain (1) that the write will actually happen, (2)
621 of the location of the write, and (3) of the value to be written.
622 But please carefully read the "CONTROL DEPENDENCIES" section and the
623 Documentation/RCU/rcu_dereference.txt file: The compiler can and does
624 break dependencies in a great many highly creative ways.
626 CPU 1 CPU 2
627 =============== ===============
628 { A == 1, B == 2, C = 3, P == &A, Q == &C }
629 B = 4;
630 <write barrier>
631 WRITE_ONCE(P, &B);
632 Q = READ_ONCE(P);
633 WRITE_ONCE(*Q, 5);
635 Therefore, no data-dependency barrier is required to order the read into
636 Q with the store into *Q. In other words, this outcome is prohibited,
637 even without a data-dependency barrier:
639 (Q == &B) && (B == 4)
641 Please note that this pattern should be rare. After all, the whole point
642 of dependency ordering is to -prevent- writes to the data structure, along
643 with the expensive cache misses associated with those writes. This pattern
644 can be used to record rare error conditions and the like, and the CPUs'
645 naturally occurring ordering prevents such records from being lost.
648 Note well that the ordering provided by a data dependency is local to
649 the CPU containing it. See the section on "Multicopy atomicity" for
650 more information.
653 The data dependency barrier is very important to the RCU system,
654 for example. See rcu_assign_pointer() and rcu_dereference() in
655 include/linux/rcupdate.h. This permits the current target of an RCU'd
656 pointer to be replaced with a new modified target, without the replacement
657 target appearing to be incompletely initialised.
659 See also the subsection on "Cache Coherency" for a more thorough example.
662 CONTROL DEPENDENCIES
663 --------------------
665 Control dependencies can be a bit tricky because current compilers do
666 not understand them. The purpose of this section is to help you prevent
667 the compiler's ignorance from breaking your code.
669 A load-load control dependency requires a full read memory barrier, not
670 simply a data dependency barrier to make it work correctly. Consider the
671 following bit of code:
673 q = READ_ONCE(a);
674 if (q) {
675 <data dependency barrier> /* BUG: No data dependency!!! */
676 p = READ_ONCE(b);
677 }
679 This will not have the desired effect because there is no actual data
680 dependency, but rather a control dependency that the CPU may short-circuit
681 by attempting to predict the outcome in advance, so that other CPUs see
682 the load from b as having happened before the load from a. In such a
683 case what's actually required is:
685 q = READ_ONCE(a);
686 if (q) {
687 <read barrier>
688 p = READ_ONCE(b);
689 }
691 However, stores are not speculated. This means that ordering -is- provided
692 for load-store control dependencies, as in the following example:
694 q = READ_ONCE(a);
695 if (q) {
696 WRITE_ONCE(b, 1);
697 }
699 Control dependencies pair normally with other types of barriers.
700 That said, please note that neither READ_ONCE() nor WRITE_ONCE()
701 are optional! Without the READ_ONCE(), the compiler might combine the
702 load from 'a' with other loads from 'a'. Without the WRITE_ONCE(),
703 the compiler might combine the store to 'b' with other stores to 'b'.
704 Either can result in highly counterintuitive effects on ordering.
706 Worse yet, if the compiler is able to prove (say) that the value of
707 variable 'a' is always non-zero, it would be well within its rights
708 to optimize the original example by eliminating the "if" statement
709 as follows:
711 q = a;
712 b = 1; /* BUG: Compiler and CPU can both reorder!!! */
714 So don't leave out the READ_ONCE().
716 It is tempting to try to enforce ordering on identical stores on both
717 branches of the "if" statement as follows:
719 q = READ_ONCE(a);
720 if (q) {
721 barrier();
722 WRITE_ONCE(b, 1);
723 do_something();
724 } else {
725 barrier();
726 WRITE_ONCE(b, 1);
727 do_something_else();
728 }
730 Unfortunately, current compilers will transform this as follows at high
731 optimization levels:
733 q = READ_ONCE(a);
734 barrier();
735 WRITE_ONCE(b, 1); /* BUG: No ordering vs. load from a!!! */
736 if (q) {
737 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
738 do_something();
739 } else {
740 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
741 do_something_else();
742 }
744 Now there is no conditional between the load from 'a' and the store to
745 'b', which means that the CPU is within its rights to reorder them:
746 The conditional is absolutely required, and must be present in the
747 assembly code even after all compiler optimizations have been applied.
748 Therefore, if you need ordering in this example, you need explicit
749 memory barriers, for example, smp_store_release():
751 q = READ_ONCE(a);
752 if (q) {
753 smp_store_release(&b, 1);
754 do_something();
755 } else {
756 smp_store_release(&b, 1);
757 do_something_else();
758 }
760 In contrast, without explicit memory barriers, two-legged-if control
761 ordering is guaranteed only when the stores differ, for example:
763 q = READ_ONCE(a);
764 if (q) {
765 WRITE_ONCE(b, 1);
766 do_something();
767 } else {
768 WRITE_ONCE(b, 2);
769 do_something_else();
770 }
772 The initial READ_ONCE() is still required to prevent the compiler from
773 proving the value of 'a'.
775 In addition, you need to be careful what you do with the local variable 'q',
776 otherwise the compiler might be able to guess the value and again remove
777 the needed conditional. For example:
779 q = READ_ONCE(a);
780 if (q % MAX) {
781 WRITE_ONCE(b, 1);
782 do_something();
783 } else {
784 WRITE_ONCE(b, 2);
785 do_something_else();
786 }
788 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
789 equal to zero, in which case the compiler is within its rights to
790 transform the above code into the following:
792 q = READ_ONCE(a);
793 WRITE_ONCE(b, 2);
794 do_something_else();
796 Given this transformation, the CPU is not required to respect the ordering
797 between the load from variable 'a' and the store to variable 'b'. It is
798 tempting to add a barrier(), but this does not help. The conditional
799 is gone, and the barrier won't bring it back. Therefore, if you are
800 relying on this ordering, you should make sure that MAX is greater than
801 one, perhaps as follows:
803 q = READ_ONCE(a);
804 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
805 if (q % MAX) {
806 WRITE_ONCE(b, 1);
807 do_something();
808 } else {
809 WRITE_ONCE(b, 2);
810 do_something_else();
811 }
813 Please note once again that the stores to 'b' differ. If they were
814 identical, as noted earlier, the compiler could pull this store outside
815 of the 'if' statement.
817 You must also be careful not to rely too much on boolean short-circuit
818 evaluation. Consider this example:
820 q = READ_ONCE(a);
821 if (q || 1 > 0)
822 WRITE_ONCE(b, 1);
824 Because the first condition cannot fault and the second condition is
825 always true, the compiler can transform this example as following,
826 defeating control dependency:
828 q = READ_ONCE(a);
829 WRITE_ONCE(b, 1);
831 This example underscores the need to ensure that the compiler cannot
832 out-guess your code. More generally, although READ_ONCE() does force
833 the compiler to actually emit code for a given load, it does not force
834 the compiler to use the results.
836 In addition, control dependencies apply only to the then-clause and
837 else-clause of the if-statement in question. In particular, it does
838 not necessarily apply to code following the if-statement:
840 q = READ_ONCE(a);
841 if (q) {
842 WRITE_ONCE(b, 1);
843 } else {
844 WRITE_ONCE(b, 2);
845 }
846 WRITE_ONCE(c, 1); /* BUG: No ordering against the read from 'a'. */
848 It is tempting to argue that there in fact is ordering because the
849 compiler cannot reorder volatile accesses and also cannot reorder
850 the writes to 'b' with the condition. Unfortunately for this line
851 of reasoning, the compiler might compile the two writes to 'b' as
852 conditional-move instructions, as in this fanciful pseudo-assembly
853 language:
855 ld r1,a
856 cmp r1,$0
857 cmov,ne r4,$1
858 cmov,eq r4,$2
859 st r4,b
860 st $1,c
862 A weakly ordered CPU would have no dependency of any sort between the load
863 from 'a' and the store to 'c'. The control dependencies would extend
864 only to the pair of cmov instructions and the store depending on them.
865 In short, control dependencies apply only to the stores in the then-clause
866 and else-clause of the if-statement in question (including functions
867 invoked by those two clauses), not to code following that if-statement.
870 Note well that the ordering provided by a control dependency is local
871 to the CPU containing it. See the section on "Multicopy atomicity"
872 for more information.
875 In summary:
877 (*) Control dependencies can order prior loads against later stores.
878 However, they do -not- guarantee any other sort of ordering:
879 Not prior loads against later loads, nor prior stores against
880 later anything. If you need these other forms of ordering,
881 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
882 later loads, smp_mb().
884 (*) If both legs of the "if" statement begin with identical stores to
885 the same variable, then those stores must be ordered, either by
886 preceding both of them with smp_mb() or by using smp_store_release()
887 to carry out the stores. Please note that it is -not- sufficient
888 to use barrier() at beginning of each leg of the "if" statement
889 because, as shown by the example above, optimizing compilers can
890 destroy the control dependency while respecting the letter of the
891 barrier() law.
893 (*) Control dependencies require at least one run-time conditional
894 between the prior load and the subsequent store, and this
895 conditional must involve the prior load. If the compiler is able
896 to optimize the conditional away, it will have also optimized
897 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE()
898 can help to preserve the needed conditional.
900 (*) Control dependencies require that the compiler avoid reordering the
901 dependency into nonexistence. Careful use of READ_ONCE() or
902 atomic{,64}_read() can help to preserve your control dependency.
903 Please see the COMPILER BARRIER section for more information.
905 (*) Control dependencies apply only to the then-clause and else-clause
906 of the if-statement containing the control dependency, including
907 any functions that these two clauses call. Control dependencies
908 do -not- apply to code following the if-statement containing the
909 control dependency.
911 (*) Control dependencies pair normally with other types of barriers.
913 (*) Control dependencies do -not- provide multicopy atomicity. If you
914 need all the CPUs to see a given store at the same time, use smp_mb().
916 (*) Compilers do not understand control dependencies. It is therefore
917 your job to ensure that they do not break your code.
920 SMP BARRIER PAIRING
921 -------------------
923 When dealing with CPU-CPU interactions, certain types of memory barrier should
924 always be paired. A lack of appropriate pairing is almost certainly an error.
926 General barriers pair with each other, though they also pair with most
927 other types of barriers, albeit without multicopy atomicity. An acquire
928 barrier pairs with a release barrier, but both may also pair with other
929 barriers, including of course general barriers. A write barrier pairs
930 with a data dependency barrier, a control dependency, an acquire barrier,
931 a release barrier, a read barrier, or a general barrier. Similarly a
932 read barrier, control dependency, or a data dependency barrier pairs
933 with a write barrier, an acquire barrier, a release barrier, or a
934 general barrier:
936 CPU 1 CPU 2
937 =============== ===============
938 WRITE_ONCE(a, 1);
939 <write barrier>
940 WRITE_ONCE(b, 2); x = READ_ONCE(b);
941 <read barrier>
942 y = READ_ONCE(a);
944 Or:
946 CPU 1 CPU 2
947 =============== ===============================
948 a = 1;
949 <write barrier>
950 WRITE_ONCE(b, &a); x = READ_ONCE(b);
951 <data dependency barrier>
952 y = *x;
954 Or even:
956 CPU 1 CPU 2
957 =============== ===============================
958 r1 = READ_ONCE(y);
959 <general barrier>
960 WRITE_ONCE(x, 1); if (r2 = READ_ONCE(x)) {
961 <implicit control dependency>
962 WRITE_ONCE(y, 1);
963 }
965 assert(r1 == 0 || r2 == 0);
967 Basically, the read barrier always has to be there, even though it can be of
968 the "weaker" type.
970 [!] Note that the stores before the write barrier would normally be expected to
971 match the loads after the read barrier or the data dependency barrier, and vice
972 versa:
974 CPU 1 CPU 2
975 =================== ===================
976 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c);
977 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d);
978 <write barrier> \ <read barrier>
979 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a);
980 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b);
983 EXAMPLES OF MEMORY BARRIER SEQUENCES
984 ------------------------------------
986 Firstly, write barriers act as partial orderings on store operations.
987 Consider the following sequence of events:
989 CPU 1
990 =======================
991 STORE A = 1
992 STORE B = 2
993 STORE C = 3
994 <write barrier>
995 STORE D = 4
996 STORE E = 5
998 This sequence of events is committed to the memory coherence system in an order
999 that the rest of the system might perceive as the unordered set of { STORE A,
1000 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
1001 }:
1003 +-------+ : :
1004 | | +------+
1005 | |------>| C=3 | } /\
1006 | | : +------+ }----- \ -----> Events perceptible to
1007 | | : | A=1 | } \/ the rest of the system
1008 | | : +------+ }
1009 | CPU 1 | : | B=2 | }
1010 | | +------+ }
1011 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
1012 | | +------+ } requires all stores prior to the
1013 | | : | E=5 | } barrier to be committed before
1014 | | : +------+ } further stores may take place
1015 | |------>| D=4 | }
1016 | | +------+
1017 +-------+ : :
1018 |
1019 | Sequence in which stores are committed to the
1020 | memory system by CPU 1
1021 V
1024 Secondly, data dependency barriers act as partial orderings on data-dependent
1025 loads. Consider the following sequence of events:
1027 CPU 1 CPU 2
1028 ======================= =======================
1029 { B = 7; X = 9; Y = 8; C = &Y }
1030 STORE A = 1
1031 STORE B = 2
1032 <write barrier>
1033 STORE C = &B LOAD X
1034 STORE D = 4 LOAD C (gets &B)
1035 LOAD *C (reads B)
1037 Without intervention, CPU 2 may perceive the events on CPU 1 in some
1038 effectively random order, despite the write barrier issued by CPU 1:
1040 +-------+ : : : :
1041 | | +------+ +-------+ | Sequence of update
1042 | |------>| B=2 |----- --->| Y->8 | | of perception on
1043 | | : +------+ \ +-------+ | CPU 2
1044 | CPU 1 | : | A=1 | \ --->| C->&Y | V
1045 | | +------+ | +-------+
1046 | | wwwwwwwwwwwwwwww | : :
1047 | | +------+ | : :
1048 | | : | C=&B |--- | : : +-------+
1049 | | : +------+ \ | +-------+ | |
1050 | |------>| D=4 | ----------->| C->&B |------>| |
1051 | | +------+ | +-------+ | |
1052 +-------+ : : | : : | |
1053 | : : | |
1054 | : : | CPU 2 |
1055 | +-------+ | |
1056 Apparently incorrect ---> | | B->7 |------>| |
1057 perception of B (!) | +-------+ | |
1058 | : : | |
1059 | +-------+ | |
1060 The load of X holds ---> \ | X->9 |------>| |
1061 up the maintenance \ +-------+ | |
1062 of coherence of B ----->| B->2 | +-------+
1063 +-------+
1064 : :
1067 In the above example, CPU 2 perceives that B is 7, despite the load of *C
1068 (which would be B) coming after the LOAD of C.
1070 If, however, a data dependency barrier were to be placed between the load of C
1071 and the load of *C (ie: B) on CPU 2:
1073 CPU 1 CPU 2
1074 ======================= =======================
1075 { B = 7; X = 9; Y = 8; C = &Y }
1076 STORE A = 1
1077 STORE B = 2
1078 <write barrier>
1079 STORE C = &B LOAD X
1080 STORE D = 4 LOAD C (gets &B)
1081 <data dependency barrier>
1082 LOAD *C (reads B)
1084 then the following will occur:
1086 +-------+ : : : :
1087 | | +------+ +-------+
1088 | |------>| B=2 |----- --->| Y->8 |
1089 | | : +------+ \ +-------+
1090 | CPU 1 | : | A=1 | \ --->| C->&Y |
1091 | | +------+ | +-------+
1092 | | wwwwwwwwwwwwwwww | : :
1093 | | +------+ | : :
1094 | | : | C=&B |--- | : : +-------+
1095 | | : +------+ \ | +-------+ | |
1096 | |------>| D=4 | ----------->| C->&B |------>| |
1097 | | +------+ | +-------+ | |
1098 +-------+ : : | : : | |
1099 | : : | |
1100 | : : | CPU 2 |
1101 | +-------+ | |
1102 | | X->9 |------>| |
1103 | +-------+ | |
1104 Makes sure all effects ---> \ ddddddddddddddddd | |
1105 prior to the store of C \ +-------+ | |
1106 are perceptible to ----->| B->2 |------>| |
1107 subsequent loads +-------+ | |
1108 : : +-------+
1111 And thirdly, a read barrier acts as a partial order on loads. Consider the
1112 following sequence of events:
1114 CPU 1 CPU 2
1115 ======================= =======================
1116 { A = 0, B = 9 }
1117 STORE A=1
1118 <write barrier>
1119 STORE B=2
1120 LOAD B
1121 LOAD A
1123 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1124 some effectively random order, despite the write barrier issued by CPU 1:
1126 +-------+ : : : :
1127 | | +------+ +-------+
1128 | |------>| A=1 |------ --->| A->0 |
1129 | | +------+ \ +-------+
1130 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1131 | | +------+ | +-------+
1132 | |------>| B=2 |--- | : :
1133 | | +------+ \ | : : +-------+
1134 +-------+ : : \ | +-------+ | |
1135 ---------->| B->2 |------>| |
1136 | +-------+ | CPU 2 |
1137 | | A->0 |------>| |
1138 | +-------+ | |
1139 | : : +-------+
1140 \ : :
1141 \ +-------+
1142 ---->| A->1 |
1143 +-------+
1144 : :
1147 If, however, a read barrier were to be placed between the load of B and the
1148 load of A on CPU 2:
1150 CPU 1 CPU 2
1151 ======================= =======================
1152 { A = 0, B = 9 }
1153 STORE A=1
1154 <write barrier>
1155 STORE B=2
1156 LOAD B
1157 <read barrier>
1158 LOAD A
1160 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1161 2:
1163 +-------+ : : : :
1164 | | +------+ +-------+
1165 | |------>| A=1 |------ --->| A->0 |
1166 | | +------+ \ +-------+
1167 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1168 | | +------+ | +-------+
1169 | |------>| B=2 |--- | : :
1170 | | +------+ \ | : : +-------+
1171 +-------+ : : \ | +-------+ | |
1172 ---------->| B->2 |------>| |
1173 | +-------+ | CPU 2 |
1174 | : : | |
1175 | : : | |
1176 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1177 barrier causes all effects \ +-------+ | |
1178 prior to the storage of B ---->| A->1 |------>| |
1179 to be perceptible to CPU 2 +-------+ | |
1180 : : +-------+
1183 To illustrate this more completely, consider what could happen if the code
1184 contained a load of A either side of the read barrier:
1186 CPU 1 CPU 2
1187 ======================= =======================
1188 { A = 0, B = 9 }
1189 STORE A=1
1190 <write barrier>
1191 STORE B=2
1192 LOAD B
1193 LOAD A [first load of A]
1194 <read barrier>
1195 LOAD A [second load of A]
1197 Even though the two loads of A both occur after the load of B, they may both
1198 come up with different values:
1200 +-------+ : : : :
1201 | | +------+ +-------+
1202 | |------>| A=1 |------ --->| A->0 |
1203 | | +------+ \ +-------+
1204 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1205 | | +------+ | +-------+
1206 | |------>| B=2 |--- | : :
1207 | | +------+ \ | : : +-------+
1208 +-------+ : : \ | +-------+ | |
1209 ---------->| B->2 |------>| |
1210 | +-------+ | CPU 2 |
1211 | : : | |
1212 | : : | |
1213 | +-------+ | |
1214 | | A->0 |------>| 1st |
1215 | +-------+ | |
1216 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1217 barrier causes all effects \ +-------+ | |
1218 prior to the storage of B ---->| A->1 |------>| 2nd |
1219 to be perceptible to CPU 2 +-------+ | |
1220 : : +-------+
1223 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1224 before the read barrier completes anyway:
1226 +-------+ : : : :
1227 | | +------+ +-------+
1228 | |------>| A=1 |------ --->| A->0 |
1229 | | +------+ \ +-------+
1230 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1231 | | +------+ | +-------+
1232 | |------>| B=2 |--- | : :
1233 | | +------+ \ | : : +-------+
1234 +-------+ : : \ | +-------+ | |
1235 ---------->| B->2 |------>| |
1236 | +-------+ | CPU 2 |
1237 | : : | |
1238 \ : : | |
1239 \ +-------+ | |
1240 ---->| A->1 |------>| 1st |
1241 +-------+ | |
1242 rrrrrrrrrrrrrrrrr | |
1243 +-------+ | |
1244 | A->1 |------>| 2nd |
1245 +-------+ | |
1246 : : +-------+
1249 The guarantee is that the second load will always come up with A == 1 if the
1250 load of B came up with B == 2. No such guarantee exists for the first load of
1251 A; that may come up with either A == 0 or A == 1.
1254 READ MEMORY BARRIERS VS LOAD SPECULATION
1255 ----------------------------------------
1257 Many CPUs speculate with loads: that is they see that they will need to load an
1258 item from memory, and they find a time where they're not using the bus for any
1259 other loads, and so do the load in advance - even though they haven't actually
1260 got to that point in the instruction execution flow yet. This permits the
1261 actual load instruction to potentially complete immediately because the CPU
1262 already has the value to hand.
1264 It may turn out that the CPU didn't actually need the value - perhaps because a
1265 branch circumvented the load - in which case it can discard the value or just
1266 cache it for later use.
1268 Consider:
1270 CPU 1 CPU 2
1271 ======================= =======================
1272 LOAD B
1273 DIVIDE } Divide instructions generally
1274 DIVIDE } take a long time to perform
1275 LOAD A
1277 Which might appear as this:
1279 : : +-------+
1280 +-------+ | |
1281 --->| B->2 |------>| |
1282 +-------+ | CPU 2 |
1283 : :DIVIDE | |
1284 +-------+ | |
1285 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1286 division speculates on the +-------+ ~ | |
1287 LOAD of A : : ~ | |
1288 : :DIVIDE | |
1289 : : ~ | |
1290 Once the divisions are complete --> : : ~-->| |
1291 the CPU can then perform the : : | |
1292 LOAD with immediate effect : : +-------+
1295 Placing a read barrier or a data dependency barrier just before the second
1296 load:
1298 CPU 1 CPU 2
1299 ======================= =======================
1300 LOAD B
1301 DIVIDE
1302 DIVIDE
1303 <read barrier>
1304 LOAD A
1306 will force any value speculatively obtained to be reconsidered to an extent
1307 dependent on the type of barrier used. If there was no change made to the
1308 speculated memory location, then the speculated value will just be used:
1310 : : +-------+
1311 +-------+ | |
1312 --->| B->2 |------>| |
1313 +-------+ | CPU 2 |
1314 : :DIVIDE | |
1315 +-------+ | |
1316 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1317 division speculates on the +-------+ ~ | |
1318 LOAD of A : : ~ | |
1319 : :DIVIDE | |
1320 : : ~ | |
1321 : : ~ | |
1322 rrrrrrrrrrrrrrrr~ | |
1323 : : ~ | |
1324 : : ~-->| |
1325 : : | |
1326 : : +-------+
1329 but if there was an update or an invalidation from another CPU pending, then
1330 the speculation will be cancelled and the value reloaded:
1332 : : +-------+
1333 +-------+ | |
1334 --->| B->2 |------>| |
1335 +-------+ | CPU 2 |
1336 : :DIVIDE | |
1337 +-------+ | |
1338 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1339 division speculates on the +-------+ ~ | |
1340 LOAD of A : : ~ | |
1341 : :DIVIDE | |
1342 : : ~ | |
1343 : : ~ | |
1344 rrrrrrrrrrrrrrrrr | |
1345 +-------+ | |
1346 The speculation is discarded ---> --->| A->1 |------>| |
1347 and an updated value is +-------+ | |
1348 retrieved : : +-------+
1351 MULTICOPY ATOMICITY
1352 --------------------
1354 Multicopy atomicity is a deeply intuitive notion about ordering that is
1355 not always provided by real computer systems, namely that a given store
1356 becomes visible at the same time to all CPUs, or, alternatively, that all
1357 CPUs agree on the order in which all stores become visible. However,
1358 support of full multicopy atomicity would rule out valuable hardware
1359 optimizations, so a weaker form called ``other multicopy atomicity''
1360 instead guarantees only that a given store becomes visible at the same
1361 time to all -other- CPUs. The remainder of this document discusses this
1362 weaker form, but for brevity will call it simply ``multicopy atomicity''.
1364 The following example demonstrates multicopy atomicity:
1366 CPU 1 CPU 2 CPU 3
1367 ======================= ======================= =======================
1368 { X = 0, Y = 0 }
1369 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1)
1370 <general barrier> <read barrier>
1371 STORE Y=r1 LOAD X
1373 Suppose that CPU 2's load from X returns 1, which it then stores to Y,
1374 and CPU 3's load from Y returns 1. This indicates that CPU 1's store
1375 to X precedes CPU 2's load from X and that CPU 2's store to Y precedes
1376 CPU 3's load from Y. In addition, the memory barriers guarantee that
1377 CPU 2 executes its load before its store, and CPU 3 loads from Y before
1378 it loads from X. The question is then "Can CPU 3's load from X return 0?"
1380 Because CPU 3's load from X in some sense comes after CPU 2's load, it
1381 is natural to expect that CPU 3's load from X must therefore return 1.
1382 This expectation follows from multicopy atomicity: if a load executing
1383 on CPU B follows a load from the same variable executing on CPU A (and
1384 CPU A did not originally store the value which it read), then on
1385 multicopy-atomic systems, CPU B's load must return either the same value
1386 that CPU A's load did or some later value. However, the Linux kernel
1387 does not require systems to be multicopy atomic.
1389 The use of a general memory barrier in the example above compensates
1390 for any lack of multicopy atomicity. In the example, if CPU 2's load
1391 from X returns 1 and CPU 3's load from Y returns 1, then CPU 3's load
1392 from X must indeed also return 1.
1394 However, dependencies, read barriers, and write barriers are not always
1395 able to compensate for non-multicopy atomicity. For example, suppose
1396 that CPU 2's general barrier is removed from the above example, leaving
1397 only the data dependency shown below:
1399 CPU 1 CPU 2 CPU 3
1400 ======================= ======================= =======================
1401 { X = 0, Y = 0 }
1402 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1)
1403 <data dependency> <read barrier>
1404 STORE Y=r1 LOAD X (reads 0)
1406 This substitution allows non-multicopy atomicity to run rampant: in
1407 this example, it is perfectly legal for CPU 2's load from X to return 1,
1408 CPU 3's load from Y to return 1, and its load from X to return 0.
1410 The key point is that although CPU 2's data dependency orders its load
1411 and store, it does not guarantee to order CPU 1's store. Thus, if this
1412 example runs on a non-multicopy-atomic system where CPUs 1 and 2 share a
1413 store buffer or a level of cache, CPU 2 might have early access to CPU 1's
1414 writes. General barriers are therefore required to ensure that all CPUs
1415 agree on the combined order of multiple accesses.
1417 General barriers can compensate not only for non-multicopy atomicity,
1418 but can also generate additional ordering that can ensure that -all-
1419 CPUs will perceive the same order of -all- operations. In contrast, a
1420 chain of release-acquire pairs do not provide this additional ordering,
1421 which means that only those CPUs on the chain are guaranteed to agree
1422 on the combined order of the accesses. For example, switching to C code
1423 in deference to the ghost of Herman Hollerith:
1425 int u, v, x, y, z;
1427 void cpu0(void)
1428 {
1429 r0 = smp_load_acquire(&x);
1430 WRITE_ONCE(u, 1);
1431 smp_store_release(&y, 1);
1432 }
1434 void cpu1(void)
1435 {
1436 r1 = smp_load_acquire(&y);
1437 r4 = READ_ONCE(v);
1438 r5 = READ_ONCE(u);
1439 smp_store_release(&z, 1);
1440 }
1442 void cpu2(void)
1443 {
1444 r2 = smp_load_acquire(&z);
1445 smp_store_release(&x, 1);
1446 }
1448 void cpu3(void)
1449 {
1450 WRITE_ONCE(v, 1);
1451 smp_mb();
1452 r3 = READ_ONCE(u);
1453 }
1455 Because cpu0(), cpu1(), and cpu2() participate in a chain of
1456 smp_store_release()/smp_load_acquire() pairs, the following outcome
1457 is prohibited:
1459 r0 == 1 && r1 == 1 && r2 == 1
1461 Furthermore, because of the release-acquire relationship between cpu0()
1462 and cpu1(), cpu1() must see cpu0()'s writes, so that the following
1463 outcome is prohibited:
1465 r1 == 1 && r5 == 0
1467 However, the ordering provided by a release-acquire chain is local
1468 to the CPUs participating in that chain and does not apply to cpu3(),
1469 at least aside from stores. Therefore, the following outcome is possible:
1471 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0
1473 As an aside, the following outcome is also possible:
1475 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1
1477 Although cpu0(), cpu1(), and cpu2() will see their respective reads and
1478 writes in order, CPUs not involved in the release-acquire chain might
1479 well disagree on the order. This disagreement stems from the fact that
1480 the weak memory-barrier instructions used to implement smp_load_acquire()
1481 and smp_store_release() are not required to order prior stores against
1482 subsequent loads in all cases. This means that cpu3() can see cpu0()'s
1483 store to u as happening -after- cpu1()'s load from v, even though
1484 both cpu0() and cpu1() agree that these two operations occurred in the
1485 intended order.
1487 However, please keep in mind that smp_load_acquire() is not magic.
1488 In particular, it simply reads from its argument with ordering. It does
1489 -not- ensure that any particular value will be read. Therefore, the
1490 following outcome is possible:
1492 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0
1494 Note that this outcome can happen even on a mythical sequentially
1495 consistent system where nothing is ever reordered.
1497 To reiterate, if your code requires full ordering of all operations,
1498 use general barriers throughout.
1501 ========================
1502 EXPLICIT KERNEL BARRIERS
1503 ========================
1505 The Linux kernel has a variety of different barriers that act at different
1506 levels:
1508 (*) Compiler barrier.
1510 (*) CPU memory barriers.
1513 COMPILER BARRIER
1514 ----------------
1516 The Linux kernel has an explicit compiler barrier function that prevents the
1517 compiler from moving the memory accesses either side of it to the other side:
1519 barrier();
1521 This is a general barrier -- there are no read-read or write-write
1522 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be
1523 thought of as weak forms of barrier() that affect only the specific
1524 accesses flagged by the READ_ONCE() or WRITE_ONCE().
1526 The barrier() function has the following effects:
1528 (*) Prevents the compiler from reordering accesses following the
1529 barrier() to precede any accesses preceding the barrier().
1530 One example use for this property is to ease communication between
1531 interrupt-handler code and the code that was interrupted.
1533 (*) Within a loop, forces the compiler to load the variables used
1534 in that loop's conditional on each pass through that loop.
1536 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of
1537 optimizations that, while perfectly safe in single-threaded code, can
1538 be fatal in concurrent code. Here are some examples of these sorts
1539 of optimizations:
1541 (*) The compiler is within its rights to reorder loads and stores
1542 to the same variable, and in some cases, the CPU is within its
1543 rights to reorder loads to the same variable. This means that
1544 the following code:
1546 a[0] = x;
1547 a[1] = x;
1549 Might result in an older value of x stored in a[1] than in a[0].
1550 Prevent both the compiler and the CPU from doing this as follows:
1552 a[0] = READ_ONCE(x);
1553 a[1] = READ_ONCE(x);
1555 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for
1556 accesses from multiple CPUs to a single variable.
1558 (*) The compiler is within its rights to merge successive loads from
1559 the same variable. Such merging can cause the compiler to "optimize"
1560 the following code:
1562 while (tmp = a)
1563 do_something_with(tmp);
1565 into the following code, which, although in some sense legitimate
1566 for single-threaded code, is almost certainly not what the developer
1567 intended:
1569 if (tmp = a)
1570 for (;;)
1571 do_something_with(tmp);
1573 Use READ_ONCE() to prevent the compiler from doing this to you:
1575 while (tmp = READ_ONCE(a))
1576 do_something_with(tmp);
1578 (*) The compiler is within its rights to reload a variable, for example,
1579 in cases where high register pressure prevents the compiler from
1580 keeping all data of interest in registers. The compiler might
1581 therefore optimize the variable 'tmp' out of our previous example:
1583 while (tmp = a)
1584 do_something_with(tmp);
1586 This could result in the following code, which is perfectly safe in
1587 single-threaded code, but can be fatal in concurrent code:
1589 while (a)
1590 do_something_with(a);
1592 For example, the optimized version of this code could result in
1593 passing a zero to do_something_with() in the case where the variable
1594 a was modified by some other CPU between the "while" statement and
1595 the call to do_something_with().
1597 Again, use READ_ONCE() to prevent the compiler from doing this:
1599 while (tmp = READ_ONCE(a))
1600 do_something_with(tmp);
1602 Note that if the compiler runs short of registers, it might save
1603 tmp onto the stack. The overhead of this saving and later restoring
1604 is why compilers reload variables. Doing so is perfectly safe for
1605 single-threaded code, so you need to tell the compiler about cases
1606 where it is not safe.
1608 (*) The compiler is within its rights to omit a load entirely if it knows
1609 what the value will be. For example, if the compiler can prove that
1610 the value of variable 'a' is always zero, it can optimize this code:
1612 while (tmp = a)
1613 do_something_with(tmp);
1615 Into this:
1617 do { } while (0);
1619 This transformation is a win for single-threaded code because it
1620 gets rid of a load and a branch. The problem is that the compiler
1621 will carry out its proof assuming that the current CPU is the only
1622 one updating variable 'a'. If variable 'a' is shared, then the
1623 compiler's proof will be erroneous. Use READ_ONCE() to tell the
1624 compiler that it doesn't know as much as it thinks it does:
1626 while (tmp = READ_ONCE(a))
1627 do_something_with(tmp);
1629 But please note that the compiler is also closely watching what you
1630 do with the value after the READ_ONCE(). For example, suppose you
1631 do the following and MAX is a preprocessor macro with the value 1:
1633 while ((tmp = READ_ONCE(a)) % MAX)
1634 do_something_with(tmp);
1636 Then the compiler knows that the result of the "%" operator applied
1637 to MAX will always be zero, again allowing the compiler to optimize
1638 the code into near-nonexistence. (It will still load from the
1639 variable 'a'.)
1641 (*) Similarly, the compiler is within its rights to omit a store entirely
1642 if it knows that the variable already has the value being stored.
1643 Again, the compiler assumes that the current CPU is the only one
1644 storing into the variable, which can cause the compiler to do the
1645 wrong thing for shared variables. For example, suppose you have
1646 the following:
1648 a = 0;
1649 ... Code that does not store to variable a ...
1650 a = 0;
1652 The compiler sees that the value of variable 'a' is already zero, so
1653 it might well omit the second store. This would come as a fatal
1654 surprise if some other CPU might have stored to variable 'a' in the
1655 meantime.
1657 Use WRITE_ONCE() to prevent the compiler from making this sort of
1658 wrong guess:
1660 WRITE_ONCE(a, 0);
1661 ... Code that does not store to variable a ...
1662 WRITE_ONCE(a, 0);
1664 (*) The compiler is within its rights to reorder memory accesses unless
1665 you tell it not to. For example, consider the following interaction
1666 between process-level code and an interrupt handler:
1668 void process_level(void)
1669 {
1670 msg = get_message();
1671 flag = true;
1672 }
1674 void interrupt_handler(void)
1675 {
1676 if (flag)
1677 process_message(msg);
1678 }
1680 There is nothing to prevent the compiler from transforming
1681 process_level() to the following, in fact, this might well be a
1682 win for single-threaded code:
1684 void process_level(void)
1685 {
1686 flag = true;
1687 msg = get_message();
1688 }
1690 If the interrupt occurs between these two statement, then
1691 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE()
1692 to prevent this as follows:
1694 void process_level(void)
1695 {
1696 WRITE_ONCE(msg, get_message());
1697 WRITE_ONCE(flag, true);
1698 }
1700 void interrupt_handler(void)
1701 {
1702 if (READ_ONCE(flag))
1703 process_message(READ_ONCE(msg));
1704 }
1706 Note that the READ_ONCE() and WRITE_ONCE() wrappers in
1707 interrupt_handler() are needed if this interrupt handler can itself
1708 be interrupted by something that also accesses 'flag' and 'msg',
1709 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE()
1710 and WRITE_ONCE() are not needed in interrupt_handler() other than
1711 for documentation purposes. (Note also that nested interrupts
1712 do not typically occur in modern Linux kernels, in fact, if an
1713 interrupt handler returns with interrupts enabled, you will get a
1714 WARN_ONCE() splat.)
1716 You should assume that the compiler can move READ_ONCE() and
1717 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(),
1718 barrier(), or similar primitives.
1720 This effect could also be achieved using barrier(), but READ_ONCE()
1721 and WRITE_ONCE() are more selective: With READ_ONCE() and
1722 WRITE_ONCE(), the compiler need only forget the contents of the
1723 indicated memory locations, while with barrier() the compiler must
1724 discard the value of all memory locations that it has currented
1725 cached in any machine registers. Of course, the compiler must also
1726 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
1727 though the CPU of course need not do so.
1729 (*) The compiler is within its rights to invent stores to a variable,
1730 as in the following example:
1732 if (a)
1733 b = a;
1734 else
1735 b = 42;
1737 The compiler might save a branch by optimizing this as follows:
1739 b = 42;
1740 if (a)
1741 b = a;
1743 In single-threaded code, this is not only safe, but also saves
1744 a branch. Unfortunately, in concurrent code, this optimization
1745 could cause some other CPU to see a spurious value of 42 -- even
1746 if variable 'a' was never zero -- when loading variable 'b'.
1747 Use WRITE_ONCE() to prevent this as follows:
1749 if (a)
1750 WRITE_ONCE(b, a);
1751 else
1752 WRITE_ONCE(b, 42);
1754 The compiler can also invent loads. These are usually less
1755 damaging, but they can result in cache-line bouncing and thus in
1756 poor performance and scalability. Use READ_ONCE() to prevent
1757 invented loads.
1759 (*) For aligned memory locations whose size allows them to be accessed
1760 with a single memory-reference instruction, prevents "load tearing"
1761 and "store tearing," in which a single large access is replaced by
1762 multiple smaller accesses. For example, given an architecture having
1763 16-bit store instructions with 7-bit immediate fields, the compiler
1764 might be tempted to use two 16-bit store-immediate instructions to
1765 implement the following 32-bit store:
1767 p = 0x00010002;
1769 Please note that GCC really does use this sort of optimization,
1770 which is not surprising given that it would likely take more
1771 than two instructions to build the constant and then store it.
1772 This optimization can therefore be a win in single-threaded code.
1773 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1774 this optimization in a volatile store. In the absence of such bugs,
1775 use of WRITE_ONCE() prevents store tearing in the following example:
1777 WRITE_ONCE(p, 0x00010002);
1779 Use of packed structures can also result in load and store tearing,
1780 as in this example:
1782 struct __attribute__((__packed__)) foo {
1783 short a;
1784 int b;
1785 short c;
1786 };
1787 struct foo foo1, foo2;
1788 ...
1790 foo2.a = foo1.a;
1791 foo2.b = foo1.b;
1792 foo2.c = foo1.c;
1794 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no
1795 volatile markings, the compiler would be well within its rights to
1796 implement these three assignment statements as a pair of 32-bit
1797 loads followed by a pair of 32-bit stores. This would result in
1798 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE()
1799 and WRITE_ONCE() again prevent tearing in this example:
1801 foo2.a = foo1.a;
1802 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b));
1803 foo2.c = foo1.c;
1805 All that aside, it is never necessary to use READ_ONCE() and
1806 WRITE_ONCE() on a variable that has been marked volatile. For example,
1807 because 'jiffies' is marked volatile, it is never necessary to
1808 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and
1809 WRITE_ONCE() are implemented as volatile casts, which has no effect when
1810 its argument is already marked volatile.
1812 Please note that these compiler barriers have no direct effect on the CPU,
1813 which may then reorder things however it wishes.
1816 CPU MEMORY BARRIERS
1817 -------------------
1819 The Linux kernel has eight basic CPU memory barriers:
1821 TYPE MANDATORY SMP CONDITIONAL
1822 =============== ======================= ===========================
1823 GENERAL mb() smp_mb()
1824 WRITE wmb() smp_wmb()
1825 READ rmb() smp_rmb()
1826 DATA DEPENDENCY READ_ONCE()
1829 All memory barriers except the data dependency barriers imply a compiler
1830 barrier. Data dependencies do not impose any additional compiler ordering.
1832 Aside: In the case of data dependencies, the compiler would be expected
1833 to issue the loads in the correct order (eg. `a[b]` would have to load
1834 the value of b before loading a[b]), however there is no guarantee in
1835 the C specification that the compiler may not speculate the value of b
1836 (eg. is equal to 1) and load a before b (eg. tmp = a[1]; if (b != 1)
1837 tmp = a[b]; ). There is also the problem of a compiler reloading b after
1838 having loaded a[b], thus having a newer copy of b than a[b]. A consensus
1839 has not yet been reached about these problems, however the READ_ONCE()
1840 macro is a good place to start looking.
1842 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1843 systems because it is assumed that a CPU will appear to be self-consistent,
1844 and will order overlapping accesses correctly with respect to itself.
1845 However, see the subsection on "Virtual Machine Guests" below.
1847 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1848 references to shared memory on SMP systems, though the use of locking instead
1849 is sufficient.
1851 Mandatory barriers should not be used to control SMP effects, since mandatory
1852 barriers impose unnecessary overhead on both SMP and UP systems. They may,
1853 however, be used to control MMIO effects on accesses through relaxed memory I/O
1854 windows. These barriers are required even on non-SMP systems as they affect
1855 the order in which memory operations appear to a device by prohibiting both the
1856 compiler and the CPU from reordering them.
1859 There are some more advanced barrier functions:
1861 (*) smp_store_mb(var, value)
1863 This assigns the value to the variable and then inserts a full memory
1864 barrier after it. It isn't guaranteed to insert anything more than a
1865 compiler barrier in a UP compilation.
1868 (*) smp_mb__before_atomic();
1869 (*) smp_mb__after_atomic();
1871 These are for use with atomic (such as add, subtract, increment and
1872 decrement) functions that don't return a value, especially when used for
1873 reference counting. These functions do not imply memory barriers.
1875 These are also used for atomic bitop functions that do not return a
1876 value (such as set_bit and clear_bit).
1878 As an example, consider a piece of code that marks an object as being dead
1879 and then decrements the object's reference count:
1881 obj->dead = 1;
1882 smp_mb__before_atomic();
1883 atomic_dec(&obj->ref_count);
1885 This makes sure that the death mark on the object is perceived to be set
1886 *before* the reference counter is decremented.
1888 See Documentation/atomic_{t,bitops}.txt for more information.
1891 (*) dma_wmb();
1892 (*) dma_rmb();
1894 These are for use with consistent memory to guarantee the ordering
1895 of writes or reads of shared memory accessible to both the CPU and a
1896 DMA capable device.
1898 For example, consider a device driver that shares memory with a device
1899 and uses a descriptor status value to indicate if the descriptor belongs
1900 to the device or the CPU, and a doorbell to notify it when new
1901 descriptors are available:
1903 if (desc->status != DEVICE_OWN) {
1904 /* do not read data until we own descriptor */
1905 dma_rmb();
1907 /* read/modify data */
1908 read_data = desc->data;
1909 desc->data = write_data;
1911 /* flush modifications before status update */
1912 dma_wmb();
1914 /* assign ownership */
1915 desc->status = DEVICE_OWN;
1917 /* notify device of new descriptors */
1918 writel(DESC_NOTIFY, doorbell);
1919 }
1921 The dma_rmb() allows us guarantee the device has released ownership
1922 before we read the data from the descriptor, and the dma_wmb() allows
1923 us to guarantee the data is written to the descriptor before the device
1924 can see it now has ownership. Note that, when using writel(), a prior
1925 wmb() is not needed to guarantee that the cache coherent memory writes
1926 have completed before writing to the MMIO region. The cheaper
1927 writel_relaxed() does not provide this guarantee and must not be used
1928 here.
1930 See the subsection "Kernel I/O barrier effects" for more information on
1931 relaxed I/O accessors and the Documentation/DMA-API.txt file for more
1932 information on consistent memory.
1935 ===============================
1936 IMPLICIT KERNEL MEMORY BARRIERS
1937 ===============================
1939 Some of the other functions in the linux kernel imply memory barriers, amongst
1940 which are locking and scheduling functions.
1942 This specification is a _minimum_ guarantee; any particular architecture may
1943 provide more substantial guarantees, but these may not be relied upon outside
1944 of arch specific code.
1947 LOCK ACQUISITION FUNCTIONS
1948 --------------------------
1950 The Linux kernel has a number of locking constructs:
1952 (*) spin locks
1953 (*) R/W spin locks
1954 (*) mutexes
1955 (*) semaphores
1956 (*) R/W semaphores
1958 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1959 for each construct. These operations all imply certain barriers:
1961 (1) ACQUIRE operation implication:
1963 Memory operations issued after the ACQUIRE will be completed after the
1964 ACQUIRE operation has completed.
1966 Memory operations issued before the ACQUIRE may be completed after
1967 the ACQUIRE operation has completed.
1969 (2) RELEASE operation implication:
1971 Memory operations issued before the RELEASE will be completed before the
1972 RELEASE operation has completed.
1974 Memory operations issued after the RELEASE may be completed before the
1975 RELEASE operation has completed.
1977 (3) ACQUIRE vs ACQUIRE implication:
1979 All ACQUIRE operations issued before another ACQUIRE operation will be
1980 completed before that ACQUIRE operation.
1982 (4) ACQUIRE vs RELEASE implication:
1984 All ACQUIRE operations issued before a RELEASE operation will be
1985 completed before the RELEASE operation.
1987 (5) Failed conditional ACQUIRE implication:
1989 Certain locking variants of the ACQUIRE operation may fail, either due to
1990 being unable to get the lock immediately, or due to receiving an unblocked
1991 signal while asleep waiting for the lock to become available. Failed
1992 locks do not imply any sort of barrier.
1994 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1995 one-way barriers is that the effects of instructions outside of a critical
1996 section may seep into the inside of the critical section.
1998 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1999 because it is possible for an access preceding the ACQUIRE to happen after the
2000 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
2001 the two accesses can themselves then cross:
2003 *A = a;
2004 ACQUIRE M
2005 RELEASE M
2006 *B = b;
2008 may occur as:
2010 ACQUIRE M, STORE *B, STORE *A, RELEASE M
2012 When the ACQUIRE and RELEASE are a lock acquisition and release,
2013 respectively, this same reordering can occur if the lock's ACQUIRE and
2014 RELEASE are to the same lock variable, but only from the perspective of
2015 another CPU not holding that lock. In short, a ACQUIRE followed by an
2016 RELEASE may -not- be assumed to be a full memory barrier.
2018 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does
2019 not imply a full memory barrier. Therefore, the CPU's execution of the
2020 critical sections corresponding to the RELEASE and the ACQUIRE can cross,
2021 so that:
2023 *A = a;
2024 RELEASE M
2025 ACQUIRE N
2026 *B = b;
2028 could occur as:
2030 ACQUIRE N, STORE *B, STORE *A, RELEASE M
2032 It might appear that this reordering could introduce a deadlock.
2033 However, this cannot happen because if such a deadlock threatened,
2034 the RELEASE would simply complete, thereby avoiding the deadlock.
2036 Why does this work?
2038 One key point is that we are only talking about the CPU doing
2039 the reordering, not the compiler. If the compiler (or, for
2040 that matter, the developer) switched the operations, deadlock
2041 -could- occur.
2043 But suppose the CPU reordered the operations. In this case,
2044 the unlock precedes the lock in the assembly code. The CPU
2045 simply elected to try executing the later lock operation first.
2046 If there is a deadlock, this lock operation will simply spin (or
2047 try to sleep, but more on that later). The CPU will eventually
2048 execute the unlock operation (which preceded the lock operation
2049 in the assembly code), which will unravel the potential deadlock,
2050 allowing the lock operation to succeed.
2052 But what if the lock is a sleeplock? In that case, the code will
2053 try to enter the scheduler, where it will eventually encounter
2054 a memory barrier, which will force the earlier unlock operation
2055 to complete, again unraveling the deadlock. There might be
2056 a sleep-unlock race, but the locking primitive needs to resolve
2057 such races properly in any case.
2059 Locks and semaphores may not provide any guarantee of ordering on UP compiled
2060 systems, and so cannot be counted on in such a situation to actually achieve
2061 anything at all - especially with respect to I/O accesses - unless combined
2062 with interrupt disabling operations.
2064 See also the section on "Inter-CPU acquiring barrier effects".
2067 As an example, consider the following:
2069 *A = a;
2070 *B = b;
2071 ACQUIRE
2072 *C = c;
2073 *D = d;
2074 RELEASE
2075 *E = e;
2076 *F = f;
2078 The following sequence of events is acceptable:
2080 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
2082 [+] Note that {*F,*A} indicates a combined access.
2084 But none of the following are:
2086 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
2087 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
2088 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
2089 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
2093 INTERRUPT DISABLING FUNCTIONS
2094 -----------------------------
2096 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
2097 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
2098 barriers are required in such a situation, they must be provided from some
2099 other means.
2102 SLEEP AND WAKE-UP FUNCTIONS
2103 ---------------------------
2105 Sleeping and waking on an event flagged in global data can be viewed as an
2106 interaction between two pieces of data: the task state of the task waiting for
2107 the event and the global data used to indicate the event. To make sure that
2108 these appear to happen in the right order, the primitives to begin the process
2109 of going to sleep, and the primitives to initiate a wake up imply certain
2110 barriers.
2112 Firstly, the sleeper normally follows something like this sequence of events:
2114 for (;;) {
2115 set_current_state(TASK_UNINTERRUPTIBLE);
2116 if (event_indicated)
2117 break;
2118 schedule();
2119 }
2121 A general memory barrier is interpolated automatically by set_current_state()
2122 after it has altered the task state:
2124 CPU 1
2125 ===============================
2126 set_current_state();
2127 smp_store_mb();
2128 STORE current->state
2129 <general barrier>
2130 LOAD event_indicated
2132 set_current_state() may be wrapped by:
2134 prepare_to_wait();
2135 prepare_to_wait_exclusive();
2137 which therefore also imply a general memory barrier after setting the state.
2138 The whole sequence above is available in various canned forms, all of which
2139 interpolate the memory barrier in the right place:
2141 wait_event();
2142 wait_event_interruptible();
2143 wait_event_interruptible_exclusive();
2144 wait_event_interruptible_timeout();
2145 wait_event_killable();
2146 wait_event_timeout();
2147 wait_on_bit();
2148 wait_on_bit_lock();
2151 Secondly, code that performs a wake up normally follows something like this:
2153 event_indicated = 1;
2154 wake_up(&event_wait_queue);
2156 or:
2158 event_indicated = 1;
2159 wake_up_process(event_daemon);
2161 A general memory barrier is executed by wake_up() if it wakes something up.
2162 If it doesn't wake anything up then a memory barrier may or may not be
2163 executed; you must not rely on it. The barrier occurs before the task state
2164 is accessed, in particular, it sits between the STORE to indicate the event
2165 and the STORE to set TASK_RUNNING:
2167 CPU 1 (Sleeper) CPU 2 (Waker)
2168 =============================== ===============================
2169 set_current_state(); STORE event_indicated
2170 smp_store_mb(); wake_up();
2171 STORE current->state ...
2172 <general barrier> <general barrier>
2173 LOAD event_indicated if ((LOAD task->state) & TASK_NORMAL)
2174 STORE task->state
2176 where "task" is the thread being woken up and it equals CPU 1's "current".
2178 To repeat, a general memory barrier is guaranteed to be executed by wake_up()
2179 if something is actually awakened, but otherwise there is no such guarantee.
2180 To see this, consider the following sequence of events, where X and Y are both
2181 initially zero:
2183 CPU 1 CPU 2
2184 =============================== ===============================
2185 X = 1; Y = 1;
2186 smp_mb(); wake_up();
2187 LOAD Y LOAD X
2189 If a wakeup does occur, one (at least) of the two loads must see 1. If, on
2190 the other hand, a wakeup does not occur, both loads might see 0.
2192 wake_up_process() always executes a general memory barrier. The barrier again
2193 occurs before the task state is accessed. In particular, if the wake_up() in
2194 the previous snippet were replaced by a call to wake_up_process() then one of
2195 the two loads would be guaranteed to see 1.
2197 The available waker functions include:
2199 complete();
2200 wake_up();
2201 wake_up_all();
2202 wake_up_bit();
2203 wake_up_interruptible();
2204 wake_up_interruptible_all();
2205 wake_up_interruptible_nr();
2206 wake_up_interruptible_poll();
2207 wake_up_interruptible_sync();
2208 wake_up_interruptible_sync_poll();
2209 wake_up_locked();
2210 wake_up_locked_poll();
2211 wake_up_nr();
2212 wake_up_poll();
2213 wake_up_process();
2215 In terms of memory ordering, these functions all provide the same guarantees of
2216 a wake_up() (or stronger).
2218 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2219 order multiple stores before the wake-up with respect to loads of those stored
2220 values after the sleeper has called set_current_state(). For instance, if the
2221 sleeper does:
2223 set_current_state(TASK_INTERRUPTIBLE);
2224 if (event_indicated)
2225 break;
2226 __set_current_state(TASK_RUNNING);
2227 do_something(my_data);
2229 and the waker does:
2231 my_data = value;
2232 event_indicated = 1;
2233 wake_up(&event_wait_queue);
2235 there's no guarantee that the change to event_indicated will be perceived by
2236 the sleeper as coming after the change to my_data. In such a circumstance, the
2237 code on both sides must interpolate its own memory barriers between the
2238 separate data accesses. Thus the above sleeper ought to do:
2240 set_current_state(TASK_INTERRUPTIBLE);
2241 if (event_indicated) {
2242 smp_rmb();
2243 do_something(my_data);
2244 }
2246 and the waker should do:
2248 my_data = value;
2249 smp_wmb();
2250 event_indicated = 1;
2251 wake_up(&event_wait_queue);
2254 MISCELLANEOUS FUNCTIONS
2255 -----------------------
2257 Other functions that imply barriers:
2259 (*) schedule() and similar imply full memory barriers.
2262 ===================================
2263 INTER-CPU ACQUIRING BARRIER EFFECTS
2264 ===================================
2266 On SMP systems locking primitives give a more substantial form of barrier: one
2267 that does affect memory access ordering on other CPUs, within the context of
2268 conflict on any particular lock.
2271 ACQUIRES VS MEMORY ACCESSES
2272 ---------------------------
2274 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2275 three CPUs; then should the following sequence of events occur:
2277 CPU 1 CPU 2
2278 =============================== ===============================
2279 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e);
2280 ACQUIRE M ACQUIRE Q
2281 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f);
2282 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g);
2283 RELEASE M RELEASE Q
2284 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h);
2286 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2287 through *H occur in, other than the constraints imposed by the separate locks
2288 on the separate CPUs. It might, for example, see:
2290 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2292 But it won't see any of:
2294 *B, *C or *D preceding ACQUIRE M
2295 *A, *B or *C following RELEASE M
2296 *F, *G or *H preceding ACQUIRE Q
2297 *E, *F or *G following RELEASE Q
2300 =================================
2301 WHERE ARE MEMORY BARRIERS NEEDED?
2302 =================================
2304 Under normal operation, memory operation reordering is generally not going to
2305 be a problem as a single-threaded linear piece of code will still appear to
2306 work correctly, even if it's in an SMP kernel. There are, however, four
2307 circumstances in which reordering definitely _could_ be a problem:
2309 (*) Interprocessor interaction.
2311 (*) Atomic operations.
2313 (*) Accessing devices.
2315 (*) Interrupts.
2318 INTERPROCESSOR INTERACTION
2319 --------------------------
2321 When there's a system with more than one processor, more than one CPU in the
2322 system may be working on the same data set at the same time. This can cause
2323 synchronisation problems, and the usual way of dealing with them is to use
2324 locks. Locks, however, are quite expensive, and so it may be preferable to
2325 operate without the use of a lock if at all possible. In such a case
2326 operations that affect both CPUs may have to be carefully ordered to prevent
2327 a malfunction.
2329 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2330 queued on the semaphore, by virtue of it having a piece of its stack linked to
2331 the semaphore's list of waiting processes:
2333 struct rw_semaphore {
2334 ...
2335 spinlock_t lock;
2336 struct list_head waiters;
2337 };
2339 struct rwsem_waiter {
2340 struct list_head list;
2341 struct task_struct *task;
2342 };
2344 To wake up a particular waiter, the up_read() or up_write() functions have to:
2346 (1) read the next pointer from this waiter's record to know as to where the
2347 next waiter record is;
2349 (2) read the pointer to the waiter's task structure;
2351 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2353 (4) call wake_up_process() on the task; and
2355 (5) release the reference held on the waiter's task struct.
2357 In other words, it has to perform this sequence of events:
2359 LOAD waiter->list.next;
2360 LOAD waiter->task;
2361 STORE waiter->task;
2362 CALL wakeup
2363 RELEASE task
2365 and if any of these steps occur out of order, then the whole thing may
2366 malfunction.
2368 Once it has queued itself and dropped the semaphore lock, the waiter does not
2369 get the lock again; it instead just waits for its task pointer to be cleared
2370 before proceeding. Since the record is on the waiter's stack, this means that
2371 if the task pointer is cleared _before_ the next pointer in the list is read,
2372 another CPU might start processing the waiter and might clobber the waiter's
2373 stack before the up*() function has a chance to read the next pointer.
2375 Consider then what might happen to the above sequence of events:
2377 CPU 1 CPU 2
2378 =============================== ===============================
2379 down_xxx()
2380 Queue waiter
2381 Sleep
2382 up_yyy()
2383 LOAD waiter->task;
2384 STORE waiter->task;
2385 Woken up by other event
2386 <preempt>
2387 Resume processing
2388 down_xxx() returns
2389 call foo()
2390 foo() clobbers *waiter
2391 </preempt>
2392 LOAD waiter->list.next;
2393 --- OOPS ---
2395 This could be dealt with using the semaphore lock, but then the down_xxx()
2396 function has to needlessly get the spinlock again after being woken up.
2398 The way to deal with this is to insert a general SMP memory barrier:
2400 LOAD waiter->list.next;
2401 LOAD waiter->task;
2402 smp_mb();
2403 STORE waiter->task;
2404 CALL wakeup
2405 RELEASE task
2407 In this case, the barrier makes a guarantee that all memory accesses before the
2408 barrier will appear to happen before all the memory accesses after the barrier
2409 with respect to the other CPUs on the system. It does _not_ guarantee that all
2410 the memory accesses before the barrier will be complete by the time the barrier
2411 instruction itself is complete.
2413 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2414 compiler barrier, thus making sure the compiler emits the instructions in the
2415 right order without actually intervening in the CPU. Since there's only one
2416 CPU, that CPU's dependency ordering logic will take care of everything else.
2419 ATOMIC OPERATIONS
2420 -----------------
2422 While they are technically interprocessor interaction considerations, atomic
2423 operations are noted specially as some of them imply full memory barriers and
2424 some don't, but they're very heavily relied on as a group throughout the
2425 kernel.
2427 See Documentation/atomic_t.txt for more information.
2430 ACCESSING DEVICES
2431 -----------------
2433 Many devices can be memory mapped, and so appear to the CPU as if they're just
2434 a set of memory locations. To control such a device, the driver usually has to
2435 make the right memory accesses in exactly the right order.
2437 However, having a clever CPU or a clever compiler creates a potential problem
2438 in that the carefully sequenced accesses in the driver code won't reach the
2439 device in the requisite order if the CPU or the compiler thinks it is more
2440 efficient to reorder, combine or merge accesses - something that would cause
2441 the device to malfunction.
2443 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2444 routines - such as inb() or writel() - which know how to make such accesses
2445 appropriately sequential. While this, for the most part, renders the explicit
2446 use of memory barriers unnecessary, if the accessor functions are used to refer
2447 to an I/O memory window with relaxed memory access properties, then _mandatory_
2448 memory barriers are required to enforce ordering.
2450 See Documentation/driver-api/device-io.rst for more information.
2453 INTERRUPTS
2454 ----------
2456 A driver may be interrupted by its own interrupt service routine, and thus the
2457 two parts of the driver may interfere with each other's attempts to control or
2458 access the device.
2460 This may be alleviated - at least in part - by disabling local interrupts (a
2461 form of locking), such that the critical operations are all contained within
2462 the interrupt-disabled section in the driver. While the driver's interrupt
2463 routine is executing, the driver's core may not run on the same CPU, and its
2464 interrupt is not permitted to happen again until the current interrupt has been
2465 handled, thus the interrupt handler does not need to lock against that.
2467 However, consider a driver that was talking to an ethernet card that sports an
2468 address register and a data register. If that driver's core talks to the card
2469 under interrupt-disablement and then the driver's interrupt handler is invoked:
2471 LOCAL IRQ DISABLE
2472 writew(ADDR, 3);
2473 writew(DATA, y);
2474 LOCAL IRQ ENABLE
2475 <interrupt>
2476 writew(ADDR, 4);
2477 q = readw(DATA);
2478 </interrupt>
2480 The store to the data register might happen after the second store to the
2481 address register if ordering rules are sufficiently relaxed:
2483 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2486 If ordering rules are relaxed, it must be assumed that accesses done inside an
2487 interrupt disabled section may leak outside of it and may interleave with
2488 accesses performed in an interrupt - and vice versa - unless implicit or
2489 explicit barriers are used.
2491 Normally this won't be a problem because the I/O accesses done inside such
2492 sections will include synchronous load operations on strictly ordered I/O
2493 registers that form implicit I/O barriers.
2496 A similar situation may occur between an interrupt routine and two routines
2497 running on separate CPUs that communicate with each other. If such a case is
2498 likely, then interrupt-disabling locks should be used to guarantee ordering.
2501 ==========================
2502 KERNEL I/O BARRIER EFFECTS
2503 ==========================
2505 Interfacing with peripherals via I/O accesses is deeply architecture and device
2506 specific. Therefore, drivers which are inherently non-portable may rely on
2507 specific behaviours of their target systems in order to achieve synchronization
2508 in the most lightweight manner possible. For drivers intending to be portable
2509 between multiple architectures and bus implementations, the kernel offers a
2510 series of accessor functions that provide various degrees of ordering
2511 guarantees:
2513 (*) readX(), writeX():
2515 The readX() and writeX() MMIO accessors take a pointer to the
2516 peripheral being accessed as an __iomem * parameter. For pointers
2517 mapped with the default I/O attributes (e.g. those returned by
2518 ioremap()), the ordering guarantees are as follows:
2520 1. All readX() and writeX() accesses to the same peripheral are ordered
2521 with respect to each other. This ensures that MMIO register accesses
2522 by the same CPU thread to a particular device will arrive in program
2523 order.
2525 2. A writeX() issued by a CPU thread holding a spinlock is ordered
2526 before a writeX() to the same peripheral from another CPU thread
2527 issued after a later acquisition of the same spinlock. This ensures
2528 that MMIO register writes to a particular device issued while holding
2529 a spinlock will arrive in an order consistent with acquisitions of
2530 the lock.
2532 3. A writeX() by a CPU thread to the peripheral will first wait for the
2533 completion of all prior writes to memory either issued by, or
2534 propagated to, the same thread. This ensures that writes by the CPU
2535 to an outbound DMA buffer allocated by dma_alloc_coherent() will be
2536 visible to a DMA engine when the CPU writes to its MMIO control
2537 register to trigger the transfer.
2539 4. A readX() by a CPU thread from the peripheral will complete before
2540 any subsequent reads from memory by the same thread can begin. This
2541 ensures that reads by the CPU from an incoming DMA buffer allocated
2542 by dma_alloc_coherent() will not see stale data after reading from
2543 the DMA engine's MMIO status register to establish that the DMA
2544 transfer has completed.
2546 5. A readX() by a CPU thread from the peripheral will complete before
2547 any subsequent delay() loop can begin execution on the same thread.
2548 This ensures that two MMIO register writes by the CPU to a peripheral
2549 will arrive at least 1us apart if the first write is immediately read
2550 back with readX() and udelay(1) is called prior to the second
2551 writeX():
2553 writel(42, DEVICE_REGISTER_0); // Arrives at the device...
2554 readl(DEVICE_REGISTER_0);
2555 udelay(1);
2556 writel(42, DEVICE_REGISTER_1); // ...at least 1us before this.
2558 The ordering properties of __iomem pointers obtained with non-default
2559 attributes (e.g. those returned by ioremap_wc()) are specific to the
2560 underlying architecture and therefore the guarantees listed above cannot
2561 generally be relied upon for accesses to these types of mappings.
2563 (*) readX_relaxed(), writeX_relaxed():
2565 These are similar to readX() and writeX(), but provide weaker memory
2566 ordering guarantees. Specifically, they do not guarantee ordering with
2567 respect to locking, normal memory accesses or delay() loops (i.e.
2568 bullets 2-5 above) but they are still guaranteed to be ordered with
2569 respect to other accesses from the same CPU thread to the same
2570 peripheral when operating on __iomem pointers mapped with the default
2571 I/O attributes.
2573 (*) readsX(), writesX():
2575 The readsX() and writesX() MMIO accessors are designed for accessing
2576 register-based, memory-mapped FIFOs residing on peripherals that are not
2577 capable of performing DMA. Consequently, they provide only the ordering
2578 guarantees of readX_relaxed() and writeX_relaxed(), as documented above.
2580 (*) inX(), outX():
2582 The inX() and outX() accessors are intended to access legacy port-mapped
2583 I/O peripherals, which may require special instructions on some
2584 architectures (notably x86). The port number of the peripheral being
2585 accessed is passed as an argument.
2587 Since many CPU architectures ultimately access these peripherals via an
2588 internal virtual memory mapping, the portable ordering guarantees
2589 provided by inX() and outX() are the same as those provided by readX()
2590 and writeX() respectively when accessing a mapping with the default I/O
2591 attributes.
2593 Device drivers may expect outX() to emit a non-posted write transaction
2594 that waits for a completion response from the I/O peripheral before
2595 returning. This is not guaranteed by all architectures and is therefore
2596 not part of the portable ordering semantics.
2598 (*) insX(), outsX():
2600 As above, the insX() and outsX() accessors provide the same ordering
2601 guarantees as readsX() and writesX() respectively when accessing a
2602 mapping with the default I/O attributes.
2604 (*) ioreadX(), iowriteX():
2606 These will perform appropriately for the type of access they're actually
2607 doing, be it inX()/outX() or readX()/writeX().
2609 With the exception of the string accessors (insX(), outsX(), readsX() and
2610 writesX()), all of the above assume that the underlying peripheral is
2611 little-endian and will therefore perform byte-swapping operations on big-endian
2612 architectures.
2615 ========================================
2616 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2617 ========================================
2619 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2620 maintain the appearance of program causality with respect to itself. Some CPUs
2621 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2622 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2623 of arch-specific code.
2625 This means that it must be considered that the CPU will execute its instruction
2626 stream in any order it feels like - or even in parallel - provided that if an
2627 instruction in the stream depends on an earlier instruction, then that
2628 earlier instruction must be sufficiently complete[*] before the later
2629 instruction may proceed; in other words: provided that the appearance of
2630 causality is maintained.
2632 [*] Some instructions have more than one effect - such as changing the
2633 condition codes, changing registers or changing memory - and different
2634 instructions may depend on different effects.
2636 A CPU may also discard any instruction sequence that winds up having no
2637 ultimate effect. For example, if two adjacent instructions both load an
2638 immediate value into the same register, the first may be discarded.
2641 Similarly, it has to be assumed that compiler might reorder the instruction
2642 stream in any way it sees fit, again provided the appearance of causality is
2643 maintained.
2646 ============================
2647 THE EFFECTS OF THE CPU CACHE
2648 ============================
2650 The way cached memory operations are perceived across the system is affected to
2651 a certain extent by the caches that lie between CPUs and memory, and by the
2652 memory coherence system that maintains the consistency of state in the system.
2654 As far as the way a CPU interacts with another part of the system through the
2655 caches goes, the memory system has to include the CPU's caches, and memory
2656 barriers for the most part act at the interface between the CPU and its cache
2657 (memory barriers logically act on the dotted line in the following diagram):
2659 <--- CPU ---> : <----------- Memory ----------->
2660 :
2661 +--------+ +--------+ : +--------+ +-----------+
2662 | | | | : | | | | +--------+
2663 | CPU | | Memory | : | CPU | | | | |
2664 | Core |--->| Access |----->| Cache |<-->| | | |
2665 | | | Queue | : | | | |--->| Memory |
2666 | | | | : | | | | | |
2667 +--------+ +--------+ : +--------+ | | | |
2668 : | Cache | +--------+
2669 : | Coherency |
2670 : | Mechanism | +--------+
2671 +--------+ +--------+ : +--------+ | | | |
2672 | | | | : | | | | | |
2673 | CPU | | Memory | : | CPU | | |--->| Device |
2674 | Core |--->| Access |----->| Cache |<-->| | | |
2675 | | | Queue | : | | | | | |
2676 | | | | : | | | | +--------+
2677 +--------+ +--------+ : +--------+ +-----------+
2678 :
2679 :
2681 Although any particular load or store may not actually appear outside of the
2682 CPU that issued it since it may have been satisfied within the CPU's own cache,
2683 it will still appear as if the full memory access had taken place as far as the
2684 other CPUs are concerned since the cache coherency mechanisms will migrate the
2685 cacheline over to the accessing CPU and propagate the effects upon conflict.
2687 The CPU core may execute instructions in any order it deems fit, provided the
2688 expected program causality appears to be maintained. Some of the instructions
2689 generate load and store operations which then go into the queue of memory
2690 accesses to be performed. The core may place these in the queue in any order
2691 it wishes, and continue execution until it is forced to wait for an instruction
2692 to complete.
2694 What memory barriers are concerned with is controlling the order in which
2695 accesses cross from the CPU side of things to the memory side of things, and
2696 the order in which the effects are perceived to happen by the other observers
2697 in the system.
2699 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2700 their own loads and stores as if they had happened in program order.
2702 [!] MMIO or other device accesses may bypass the cache system. This depends on
2703 the properties of the memory window through which devices are accessed and/or
2704 the use of any special device communication instructions the CPU may have.
2707 CACHE COHERENCY
2708 ---------------
2710 Life isn't quite as simple as it may appear above, however: for while the
2711 caches are expected to be coherent, there's no guarantee that that coherency
2712 will be ordered. This means that while changes made on one CPU will
2713 eventually become visible on all CPUs, there's no guarantee that they will
2714 become apparent in the same order on those other CPUs.
2717 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2718 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2720 :
2721 : +--------+
2722 : +---------+ | |
2723 +--------+ : +--->| Cache A |<------->| |
2724 | | : | +---------+ | |
2725 | CPU 1 |<---+ | |
2726 | | : | +---------+ | |
2727 +--------+ : +--->| Cache B |<------->| |
2728 : +---------+ | |
2729 : | Memory |
2730 : +---------+ | System |
2731 +--------+ : +--->| Cache C |<------->| |
2732 | | : | +---------+ | |
2733 | CPU 2 |<---+ | |
2734 | | : | +---------+ | |
2735 +--------+ : +--->| Cache D |<------->| |
2736 : +---------+ | |
2737 : +--------+
2738 :
2740 Imagine the system has the following properties:
2742 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2743 resident in memory;
2745 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2746 resident in memory;
2748 (*) while the CPU core is interrogating one cache, the other cache may be
2749 making use of the bus to access the rest of the system - perhaps to
2750 displace a dirty cacheline or to do a speculative load;
2752 (*) each cache has a queue of operations that need to be applied to that cache
2753 to maintain coherency with the rest of the system;
2755 (*) the coherency queue is not flushed by normal loads to lines already
2756 present in the cache, even though the contents of the queue may
2757 potentially affect those loads.
2759 Imagine, then, that two writes are made on the first CPU, with a write barrier
2760 between them to guarantee that they will appear to reach that CPU's caches in
2761 the requisite order:
2763 CPU 1 CPU 2 COMMENT
2764 =============== =============== =======================================
2765 u == 0, v == 1 and p == &u, q == &u
2766 v = 2;
2767 smp_wmb(); Make sure change to v is visible before
2768 change to p
2769 <A:modify v=2> v is now in cache A exclusively
2770 p = &v;
2771 <B:modify p=&v> p is now in cache B exclusively
2773 The write memory barrier forces the other CPUs in the system to perceive that
2774 the local CPU's caches have apparently been updated in the correct order. But
2775 now imagine that the second CPU wants to read those values:
2777 CPU 1 CPU 2 COMMENT
2778 =============== =============== =======================================
2779 ...
2780 q = p;
2781 x = *q;
2783 The above pair of reads may then fail to happen in the expected order, as the
2784 cacheline holding p may get updated in one of the second CPU's caches while
2785 the update to the cacheline holding v is delayed in the other of the second
2786 CPU's caches by some other cache event:
2788 CPU 1 CPU 2 COMMENT
2789 =============== =============== =======================================
2790 u == 0, v == 1 and p == &u, q == &u
2791 v = 2;
2792 smp_wmb();
2793 <A:modify v=2> <C:busy>
2794 <C:queue v=2>
2795 p = &v; q = p;
2796 <D:request p>
2797 <B:modify p=&v> <D:commit p=&v>
2798 <D:read p>
2799 x = *q;
2800 <C:read *q> Reads from v before v updated in cache
2801 <C:unbusy>
2802 <C:commit v=2>
2804 Basically, while both cachelines will be updated on CPU 2 eventually, there's
2805 no guarantee that, without intervention, the order of update will be the same
2806 as that committed on CPU 1.
2809 To intervene, we need to interpolate a data dependency barrier or a read
2810 barrier between the loads (which as of v4.15 is supplied unconditionally
2811 by the READ_ONCE() macro). This will force the cache to commit its
2812 coherency queue before processing any further requests:
2814 CPU 1 CPU 2 COMMENT
2815 =============== =============== =======================================
2816 u == 0, v == 1 and p == &u, q == &u
2817 v = 2;
2818 smp_wmb();
2819 <A:modify v=2> <C:busy>
2820 <C:queue v=2>
2821 p = &v; q = p;
2822 <D:request p>
2823 <B:modify p=&v> <D:commit p=&v>
2824 <D:read p>
2825 smp_read_barrier_depends()
2826 <C:unbusy>
2827 <C:commit v=2>
2828 x = *q;
2829 <C:read *q> Reads from v after v updated in cache
2832 This sort of problem can be encountered on DEC Alpha processors as they have a
2833 split cache that improves performance by making better use of the data bus.
2834 While most CPUs do imply a data dependency barrier on the read when a memory
2835 access depends on a read, not all do, so it may not be relied on.
2837 Other CPUs may also have split caches, but must coordinate between the various
2838 cachelets for normal memory accesses. The semantics of the Alpha removes the
2839 need for hardware coordination in the absence of memory barriers, which
2840 permitted Alpha to sport higher CPU clock rates back in the day. However,
2841 please note that (again, as of v4.15) smp_read_barrier_depends() should not
2842 be used except in Alpha arch-specific code and within the READ_ONCE() macro.
2845 CACHE COHERENCY VS DMA
2846 ----------------------
2848 Not all systems maintain cache coherency with respect to devices doing DMA. In
2849 such cases, a device attempting DMA may obtain stale data from RAM because
2850 dirty cache lines may be resident in the caches of various CPUs, and may not
2851 have been written back to RAM yet. To deal with this, the appropriate part of
2852 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2853 invalidate them as well).
2855 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2856 cache lines being written back to RAM from a CPU's cache after the device has
2857 installed its own data, or cache lines present in the CPU's cache may simply
2858 obscure the fact that RAM has been updated, until at such time as the cacheline
2859 is discarded from the CPU's cache and reloaded. To deal with this, the
2860 appropriate part of the kernel must invalidate the overlapping bits of the
2861 cache on each CPU.
2863 See Documentation/core-api/cachetlb.rst for more information on cache management.
2866 CACHE COHERENCY VS MMIO
2867 -----------------------
2869 Memory mapped I/O usually takes place through memory locations that are part of
2870 a window in the CPU's memory space that has different properties assigned than
2871 the usual RAM directed window.
2873 Amongst these properties is usually the fact that such accesses bypass the
2874 caching entirely and go directly to the device buses. This means MMIO accesses
2875 may, in effect, overtake accesses to cached memory that were emitted earlier.
2876 A memory barrier isn't sufficient in such a case, but rather the cache must be
2877 flushed between the cached memory write and the MMIO access if the two are in
2878 any way dependent.
2881 =========================
2882 THE THINGS CPUS GET UP TO
2883 =========================
2885 A programmer might take it for granted that the CPU will perform memory
2886 operations in exactly the order specified, so that if the CPU is, for example,
2887 given the following piece of code to execute:
2889 a = READ_ONCE(*A);
2890 WRITE_ONCE(*B, b);
2891 c = READ_ONCE(*C);
2892 d = READ_ONCE(*D);
2893 WRITE_ONCE(*E, e);
2895 they would then expect that the CPU will complete the memory operation for each
2896 instruction before moving on to the next one, leading to a definite sequence of
2897 operations as seen by external observers in the system:
2899 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2902 Reality is, of course, much messier. With many CPUs and compilers, the above
2903 assumption doesn't hold because:
2905 (*) loads are more likely to need to be completed immediately to permit
2906 execution progress, whereas stores can often be deferred without a
2907 problem;
2909 (*) loads may be done speculatively, and the result discarded should it prove
2910 to have been unnecessary;
2912 (*) loads may be done speculatively, leading to the result having been fetched
2913 at the wrong time in the expected sequence of events;
2915 (*) the order of the memory accesses may be rearranged to promote better use
2916 of the CPU buses and caches;
2918 (*) loads and stores may be combined to improve performance when talking to
2919 memory or I/O hardware that can do batched accesses of adjacent locations,
2920 thus cutting down on transaction setup costs (memory and PCI devices may
2921 both be able to do this); and
2923 (*) the CPU's data cache may affect the ordering, and while cache-coherency
2924 mechanisms may alleviate this - once the store has actually hit the cache
2925 - there's no guarantee that the coherency management will be propagated in
2926 order to other CPUs.
2928 So what another CPU, say, might actually observe from the above piece of code
2929 is:
2931 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2933 (Where "LOAD {*C,*D}" is a combined load)
2936 However, it is guaranteed that a CPU will be self-consistent: it will see its
2937 _own_ accesses appear to be correctly ordered, without the need for a memory
2938 barrier. For instance with the following code:
2940 U = READ_ONCE(*A);
2941 WRITE_ONCE(*A, V);
2942 WRITE_ONCE(*A, W);
2943 X = READ_ONCE(*A);
2944 WRITE_ONCE(*A, Y);
2945 Z = READ_ONCE(*A);
2947 and assuming no intervention by an external influence, it can be assumed that
2948 the final result will appear to be:
2950 U == the original value of *A
2951 X == W
2952 Z == Y
2953 *A == Y
2955 The code above may cause the CPU to generate the full sequence of memory
2956 accesses:
2958 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2960 in that order, but, without intervention, the sequence may have almost any
2961 combination of elements combined or discarded, provided the program's view
2962 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE()
2963 are -not- optional in the above example, as there are architectures
2964 where a given CPU might reorder successive loads to the same location.
2965 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is
2966 necessary to prevent this, for example, on Itanium the volatile casts
2967 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq
2968 and st.rel instructions (respectively) that prevent such reordering.
2970 The compiler may also combine, discard or defer elements of the sequence before
2971 the CPU even sees them.
2973 For instance:
2975 *A = V;
2976 *A = W;
2978 may be reduced to:
2980 *A = W;
2982 since, without either a write barrier or an WRITE_ONCE(), it can be
2983 assumed that the effect of the storage of V to *A is lost. Similarly:
2985 *A = Y;
2986 Z = *A;
2988 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be
2989 reduced to:
2991 *A = Y;
2992 Z = Y;
2994 and the LOAD operation never appear outside of the CPU.
2997 AND THEN THERE'S THE ALPHA
2998 --------------------------
3000 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
3001 some versions of the Alpha CPU have a split data cache, permitting them to have
3002 two semantically-related cache lines updated at separate times. This is where
3003 the data dependency barrier really becomes necessary as this synchronises both
3004 caches with the memory coherence system, thus making it seem like pointer
3005 changes vs new data occur in the right order.
3007 The Alpha defines the Linux kernel's memory model, although as of v4.15
3008 the Linux kernel's addition of smp_read_barrier_depends() to READ_ONCE()
3009 greatly reduced Alpha's impact on the memory model.
3011 See the subsection on "Cache Coherency" above.
3014 VIRTUAL MACHINE GUESTS
3015 ----------------------
3017 Guests running within virtual machines might be affected by SMP effects even if
3018 the guest itself is compiled without SMP support. This is an artifact of
3019 interfacing with an SMP host while running an UP kernel. Using mandatory
3020 barriers for this use-case would be possible but is often suboptimal.
3022 To handle this case optimally, low-level virt_mb() etc macros are available.
3023 These have the same effect as smp_mb() etc when SMP is enabled, but generate
3024 identical code for SMP and non-SMP systems. For example, virtual machine guests
3025 should use virt_mb() rather than smp_mb() when synchronizing against a
3026 (possibly SMP) host.
3028 These are equivalent to smp_mb() etc counterparts in all other respects,
3029 in particular, they do not control MMIO effects: to control
3030 MMIO effects, use mandatory barriers.
3033 ============
3034 EXAMPLE USES
3035 ============
3037 CIRCULAR BUFFERS
3038 ----------------
3040 Memory barriers can be used to implement circular buffering without the need
3041 of a lock to serialise the producer with the consumer. See:
3043 Documentation/core-api/circular-buffers.rst
3045 for details.
3048 ==========
3049 REFERENCES
3050 ==========
3052 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
3053 Digital Press)
3054 Chapter 5.2: Physical Address Space Characteristics
3055 Chapter 5.4: Caches and Write Buffers
3056 Chapter 5.5: Data Sharing
3057 Chapter 5.6: Read/Write Ordering
3059 AMD64 Architecture Programmer's Manual Volume 2: System Programming
3060 Chapter 7.1: Memory-Access Ordering
3061 Chapter 7.4: Buffering and Combining Memory Writes
3063 ARM Architecture Reference Manual (ARMv8, for ARMv8-A architecture profile)
3064 Chapter B2: The AArch64 Application Level Memory Model
3066 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
3067 System Programming Guide
3068 Chapter 7.1: Locked Atomic Operations
3069 Chapter 7.2: Memory Ordering
3070 Chapter 7.4: Serializing Instructions
3072 The SPARC Architecture Manual, Version 9
3073 Chapter 8: Memory Models
3074 Appendix D: Formal Specification of the Memory Models
3075 Appendix J: Programming with the Memory Models
3077 Storage in the PowerPC (Stone and Fitzgerald)
3079 UltraSPARC Programmer Reference Manual
3080 Chapter 5: Memory Accesses and Cacheability
3081 Chapter 15: Sparc-V9 Memory Models
3083 UltraSPARC III Cu User's Manual
3084 Chapter 9: Memory Models
3086 UltraSPARC IIIi Processor User's Manual
3087 Chapter 8: Memory Models
3089 UltraSPARC Architecture 2005
3090 Chapter 9: Memory
3091 Appendix D: Formal Specifications of the Memory Models
3093 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
3094 Chapter 8: Memory Models
3095 Appendix F: Caches and Cache Coherency
3097 Solaris Internals, Core Kernel Architecture, p63-68:
3098 Chapter 3.3: Hardware Considerations for Locks and
3099 Synchronization
3101 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3102 for Kernel Programmers:
3103 Chapter 13: Other Memory Models
3105 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3106 Section 2.6: Speculation
3107 Section 4.4: Memory Access