| blob | c4ddfcd5ee3265788f37061e6b739b99b707df16 |
1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
5 By: David Howells <dhowells@redhat.com>
6 Paul E. McKenney <paulmck@linux.vnet.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.
19 To repeat, this document is not a specification of what Linux expects from
20 hardware.
22 The purpose of this document is twofold:
24 (1) to specify the minimum functionality that one can rely on for any
25 particular barrier, and
27 (2) to provide a guide as to how to use the barriers that are available.
29 Note that an architecture can provide more than the minimum requirement
30 for any particular barrier, but if the architecture provides less than
31 that, that architecture is incorrect.
33 Note also that it is possible that a barrier may be a no-op for an
34 architecture because the way that arch works renders an explicit barrier
35 unnecessary in that case.
38 ========
39 CONTENTS
40 ========
42 (*) Abstract memory access model.
44 - Device operations.
45 - Guarantees.
47 (*) What are memory barriers?
49 - Varieties of memory barrier.
50 - What may not be assumed about memory barriers?
51 - Data dependency barriers.
52 - Control dependencies.
53 - SMP barrier pairing.
54 - Examples of memory barrier sequences.
55 - Read memory barriers vs load speculation.
56 - Transitivity
58 (*) Explicit kernel barriers.
60 - Compiler barrier.
61 - CPU memory barriers.
62 - MMIO write barrier.
64 (*) Implicit kernel memory barriers.
66 - Lock acquisition functions.
67 - Interrupt disabling functions.
68 - Sleep and wake-up functions.
69 - Miscellaneous functions.
71 (*) Inter-CPU acquiring barrier effects.
73 - Acquires vs memory accesses.
74 - Acquires vs I/O accesses.
76 (*) Where are memory barriers needed?
78 - Interprocessor interaction.
79 - Atomic operations.
80 - Accessing devices.
81 - Interrupts.
83 (*) Kernel I/O barrier effects.
85 (*) Assumed minimum execution ordering model.
87 (*) The effects of the cpu cache.
89 - Cache coherency.
90 - Cache coherency vs DMA.
91 - Cache coherency vs MMIO.
93 (*) The things CPUs get up to.
95 - And then there's the Alpha.
96 - Virtual Machine Guests.
98 (*) Example uses.
100 - Circular buffers.
102 (*) References.
105 ============================
106 ABSTRACT MEMORY ACCESS MODEL
107 ============================
109 Consider the following abstract model of the system:
111 : :
112 : :
113 : :
114 +-------+ : +--------+ : +-------+
115 | | : | | : | |
116 | | : | | : | |
117 | CPU 1 |<----->| Memory |<----->| CPU 2 |
118 | | : | | : | |
119 | | : | | : | |
120 +-------+ : +--------+ : +-------+
121 ^ : ^ : ^
122 | : | : |
123 | : | : |
124 | : v : |
125 | : +--------+ : |
126 | : | | : |
127 | : | | : |
128 +---------->| Device |<----------+
129 : | | :
130 : | | :
131 : +--------+ :
132 : :
134 Each CPU executes a program that generates memory access operations. In the
135 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
136 perform the memory operations in any order it likes, provided program causality
137 appears to be maintained. Similarly, the compiler may also arrange the
138 instructions it emits in any order it likes, provided it doesn't affect the
139 apparent operation of the program.
141 So in the above diagram, the effects of the memory operations performed by a
142 CPU are perceived by the rest of the system as the operations cross the
143 interface between the CPU and rest of the system (the dotted lines).
146 For example, consider the following sequence of events:
148 CPU 1 CPU 2
149 =============== ===============
150 { A == 1; B == 2 }
151 A = 3; x = B;
152 B = 4; y = A;
154 The set of accesses as seen by the memory system in the middle can be arranged
155 in 24 different combinations:
157 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4
158 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3
159 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4
160 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4
161 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3
162 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4
163 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4
164 STORE B=4, ...
165 ...
167 and can thus result in four different combinations of values:
169 x == 2, y == 1
170 x == 2, y == 3
171 x == 4, y == 1
172 x == 4, y == 3
175 Furthermore, the stores committed by a CPU to the memory system may not be
176 perceived by the loads made by another CPU in the same order as the stores were
177 committed.
180 As a further example, consider this sequence of events:
182 CPU 1 CPU 2
183 =============== ===============
184 { A == 1, B == 2, C == 3, P == &A, Q == &C }
185 B = 4; Q = P;
186 P = &B D = *Q;
188 There is an obvious data dependency here, as the value loaded into D depends on
189 the address retrieved from P by CPU 2. At the end of the sequence, any of the
190 following results are possible:
192 (Q == &A) and (D == 1)
193 (Q == &B) and (D == 2)
194 (Q == &B) and (D == 4)
196 Note that CPU 2 will never try and load C into D because the CPU will load P
197 into Q before issuing the load of *Q.
200 DEVICE OPERATIONS
201 -----------------
203 Some devices present their control interfaces as collections of memory
204 locations, but the order in which the control registers are accessed is very
205 important. For instance, imagine an ethernet card with a set of internal
206 registers that are accessed through an address port register (A) and a data
207 port register (D). To read internal register 5, the following code might then
208 be used:
210 *A = 5;
211 x = *D;
213 but this might show up as either of the following two sequences:
215 STORE *A = 5, x = LOAD *D
216 x = LOAD *D, STORE *A = 5
218 the second of which will almost certainly result in a malfunction, since it set
219 the address _after_ attempting to read the register.
222 GUARANTEES
223 ----------
225 There are some minimal guarantees that may be expected of a CPU:
227 (*) On any given CPU, dependent memory accesses will be issued in order, with
228 respect to itself. This means that for:
230 Q = READ_ONCE(P); smp_read_barrier_depends(); D = READ_ONCE(*Q);
232 the CPU will issue the following memory operations:
234 Q = LOAD P, D = LOAD *Q
236 and always in that order. On most systems, smp_read_barrier_depends()
237 does nothing, but it is required for DEC Alpha. The READ_ONCE()
238 is required to prevent compiler mischief. Please note that you
239 should normally use something like rcu_dereference() instead of
240 open-coding smp_read_barrier_depends().
242 (*) Overlapping loads and stores within a particular CPU will appear to be
243 ordered within that CPU. This means that for:
245 a = READ_ONCE(*X); WRITE_ONCE(*X, b);
247 the CPU will only issue the following sequence of memory operations:
249 a = LOAD *X, STORE *X = b
251 And for:
253 WRITE_ONCE(*X, c); d = READ_ONCE(*X);
255 the CPU will only issue:
257 STORE *X = c, d = LOAD *X
259 (Loads and stores overlap if they are targeted at overlapping pieces of
260 memory).
262 And there are a number of things that _must_ or _must_not_ be assumed:
264 (*) It _must_not_ be assumed that the compiler will do what you want
265 with memory references that are not protected by READ_ONCE() and
266 WRITE_ONCE(). Without them, the compiler is within its rights to
267 do all sorts of "creative" transformations, which are covered in
268 the COMPILER BARRIER section.
270 (*) It _must_not_ be assumed that independent loads and stores will be issued
271 in the order given. This means that for:
273 X = *A; Y = *B; *D = Z;
275 we may get any of the following sequences:
277 X = LOAD *A, Y = LOAD *B, STORE *D = Z
278 X = LOAD *A, STORE *D = Z, Y = LOAD *B
279 Y = LOAD *B, X = LOAD *A, STORE *D = Z
280 Y = LOAD *B, STORE *D = Z, X = LOAD *A
281 STORE *D = Z, X = LOAD *A, Y = LOAD *B
282 STORE *D = Z, Y = LOAD *B, X = LOAD *A
284 (*) It _must_ be assumed that overlapping memory accesses may be merged or
285 discarded. This means that for:
287 X = *A; Y = *(A + 4);
289 we may get any one of the following sequences:
291 X = LOAD *A; Y = LOAD *(A + 4);
292 Y = LOAD *(A + 4); X = LOAD *A;
293 {X, Y} = LOAD {*A, *(A + 4) };
295 And for:
297 *A = X; *(A + 4) = Y;
299 we may get any of:
301 STORE *A = X; STORE *(A + 4) = Y;
302 STORE *(A + 4) = Y; STORE *A = X;
303 STORE {*A, *(A + 4) } = {X, Y};
305 And there are anti-guarantees:
307 (*) These guarantees do not apply to bitfields, because compilers often
308 generate code to modify these using non-atomic read-modify-write
309 sequences. Do not attempt to use bitfields to synchronize parallel
310 algorithms.
312 (*) Even in cases where bitfields are protected by locks, all fields
313 in a given bitfield must be protected by one lock. If two fields
314 in a given bitfield are protected by different locks, the compiler's
315 non-atomic read-modify-write sequences can cause an update to one
316 field to corrupt the value of an adjacent field.
318 (*) These guarantees apply only to properly aligned and sized scalar
319 variables. "Properly sized" currently means variables that are
320 the same size as "char", "short", "int" and "long". "Properly
321 aligned" means the natural alignment, thus no constraints for
322 "char", two-byte alignment for "short", four-byte alignment for
323 "int", and either four-byte or eight-byte alignment for "long",
324 on 32-bit and 64-bit systems, respectively. Note that these
325 guarantees were introduced into the C11 standard, so beware when
326 using older pre-C11 compilers (for example, gcc 4.6). The portion
327 of the standard containing this guarantee is Section 3.14, which
328 defines "memory location" as follows:
330 memory location
331 either an object of scalar type, or a maximal sequence
332 of adjacent bit-fields all having nonzero width
334 NOTE 1: Two threads of execution can update and access
335 separate memory locations without interfering with
336 each other.
338 NOTE 2: A bit-field and an adjacent non-bit-field member
339 are in separate memory locations. The same applies
340 to two bit-fields, if one is declared inside a nested
341 structure declaration and the other is not, or if the two
342 are separated by a zero-length bit-field declaration,
343 or if they are separated by a non-bit-field member
344 declaration. It is not safe to concurrently update two
345 bit-fields in the same structure if all members declared
346 between them are also bit-fields, no matter what the
347 sizes of those intervening bit-fields happen to be.
350 =========================
351 WHAT ARE MEMORY BARRIERS?
352 =========================
354 As can be seen above, independent memory operations are effectively performed
355 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
356 What is required is some way of intervening to instruct the compiler and the
357 CPU to restrict the order.
359 Memory barriers are such interventions. They impose a perceived partial
360 ordering over the memory operations on either side of the barrier.
362 Such enforcement is important because the CPUs and other devices in a system
363 can use a variety of tricks to improve performance, including reordering,
364 deferral and combination of memory operations; speculative loads; speculative
365 branch prediction and various types of caching. Memory barriers are used to
366 override or suppress these tricks, allowing the code to sanely control the
367 interaction of multiple CPUs and/or devices.
370 VARIETIES OF MEMORY BARRIER
371 ---------------------------
373 Memory barriers come in four basic varieties:
375 (1) Write (or store) memory barriers.
377 A write memory barrier gives a guarantee that all the STORE operations
378 specified before the barrier will appear to happen before all the STORE
379 operations specified after the barrier with respect to the other
380 components of the system.
382 A write barrier is a partial ordering on stores only; it is not required
383 to have any effect on loads.
385 A CPU can be viewed as committing a sequence of store operations to the
386 memory system as time progresses. All stores before a write barrier will
387 occur in the sequence _before_ all the stores after the write barrier.
389 [!] Note that write barriers should normally be paired with read or data
390 dependency barriers; see the "SMP barrier pairing" subsection.
393 (2) Data dependency barriers.
395 A data dependency barrier is a weaker form of read barrier. In the case
396 where two loads are performed such that the second depends on the result
397 of the first (eg: the first load retrieves the address to which the second
398 load will be directed), a data dependency barrier would be required to
399 make sure that the target of the second load is updated before the address
400 obtained by the first load is accessed.
402 A data dependency barrier is a partial ordering on interdependent loads
403 only; it is not required to have any effect on stores, independent loads
404 or overlapping loads.
406 As mentioned in (1), the other CPUs in the system can be viewed as
407 committing sequences of stores to the memory system that the CPU being
408 considered can then perceive. A data dependency barrier issued by the CPU
409 under consideration guarantees that for any load preceding it, if that
410 load touches one of a sequence of stores from another CPU, then by the
411 time the barrier completes, the effects of all the stores prior to that
412 touched by the load will be perceptible to any loads issued after the data
413 dependency barrier.
415 See the "Examples of memory barrier sequences" subsection for diagrams
416 showing the ordering constraints.
418 [!] Note that the first load really has to have a _data_ dependency and
419 not a control dependency. If the address for the second load is dependent
420 on the first load, but the dependency is through a conditional rather than
421 actually loading the address itself, then it's a _control_ dependency and
422 a full read barrier or better is required. See the "Control dependencies"
423 subsection for more information.
425 [!] Note that data dependency barriers should normally be paired with
426 write barriers; see the "SMP barrier pairing" subsection.
429 (3) Read (or load) memory barriers.
431 A read barrier is a data dependency barrier plus a guarantee that all the
432 LOAD operations specified before the barrier will appear to happen before
433 all the LOAD operations specified after the barrier with respect to the
434 other components of the system.
436 A read barrier is a partial ordering on loads only; it is not required to
437 have any effect on stores.
439 Read memory barriers imply data dependency barriers, and so can substitute
440 for them.
442 [!] Note that read barriers should normally be paired with write barriers;
443 see the "SMP barrier pairing" subsection.
446 (4) General memory barriers.
448 A general memory barrier gives a guarantee that all the LOAD and STORE
449 operations specified before the barrier will appear to happen before all
450 the LOAD and STORE operations specified after the barrier with respect to
451 the other components of the system.
453 A general memory barrier is a partial ordering over both loads and stores.
455 General memory barriers imply both read and write memory barriers, and so
456 can substitute for either.
459 And a couple of implicit varieties:
461 (5) ACQUIRE operations.
463 This acts as a one-way permeable barrier. It guarantees that all memory
464 operations after the ACQUIRE operation will appear to happen after the
465 ACQUIRE operation with respect to the other components of the system.
466 ACQUIRE operations include LOCK operations and both smp_load_acquire()
467 and smp_cond_acquire() operations. The later builds the necessary ACQUIRE
468 semantics from relying on a control dependency and smp_rmb().
470 Memory operations that occur before an ACQUIRE operation may appear to
471 happen after it completes.
473 An ACQUIRE operation should almost always be paired with a RELEASE
474 operation.
477 (6) RELEASE operations.
479 This also acts as a one-way permeable barrier. It guarantees that all
480 memory operations before the RELEASE operation will appear to happen
481 before the RELEASE operation with respect to the other components of the
482 system. RELEASE operations include UNLOCK operations and
483 smp_store_release() operations.
485 Memory operations that occur after a RELEASE operation may appear to
486 happen before it completes.
488 The use of ACQUIRE and RELEASE operations generally precludes the need
489 for other sorts of memory barrier (but note the exceptions mentioned in
490 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
491 pair is -not- guaranteed to act as a full memory barrier. However, after
492 an ACQUIRE on a given variable, all memory accesses preceding any prior
493 RELEASE on that same variable are guaranteed to be visible. In other
494 words, within a given variable's critical section, all accesses of all
495 previous critical sections for that variable are guaranteed to have
496 completed.
498 This means that ACQUIRE acts as a minimal "acquire" operation and
499 RELEASE acts as a minimal "release" operation.
501 A subset of the atomic operations described in core-api/atomic_ops.rst have
502 ACQUIRE and RELEASE variants in addition to fully-ordered and relaxed (no
503 barrier semantics) definitions. For compound atomics performing both a load
504 and a store, ACQUIRE semantics apply only to the load and RELEASE semantics
505 apply only to the store portion of the operation.
507 Memory barriers are only required where there's a possibility of interaction
508 between two CPUs or between a CPU and a device. If it can be guaranteed that
509 there won't be any such interaction in any particular piece of code, then
510 memory barriers are unnecessary in that piece of code.
513 Note that these are the _minimum_ guarantees. Different architectures may give
514 more substantial guarantees, but they may _not_ be relied upon outside of arch
515 specific code.
518 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
519 ----------------------------------------------
521 There are certain things that the Linux kernel memory barriers do not guarantee:
523 (*) There is no guarantee that any of the memory accesses specified before a
524 memory barrier will be _complete_ by the completion of a memory barrier
525 instruction; the barrier can be considered to draw a line in that CPU's
526 access queue that accesses of the appropriate type may not cross.
528 (*) There is no guarantee that issuing a memory barrier on one CPU will have
529 any direct effect on another CPU or any other hardware in the system. The
530 indirect effect will be the order in which the second CPU sees the effects
531 of the first CPU's accesses occur, but see the next point:
533 (*) There is no guarantee that a CPU will see the correct order of effects
534 from a second CPU's accesses, even _if_ the second CPU uses a memory
535 barrier, unless the first CPU _also_ uses a matching memory barrier (see
536 the subsection on "SMP Barrier Pairing").
538 (*) There is no guarantee that some intervening piece of off-the-CPU
539 hardware[*] will not reorder the memory accesses. CPU cache coherency
540 mechanisms should propagate the indirect effects of a memory barrier
541 between CPUs, but might not do so in order.
543 [*] For information on bus mastering DMA and coherency please read:
545 Documentation/PCI/pci.txt
546 Documentation/DMA-API-HOWTO.txt
547 Documentation/DMA-API.txt
550 DATA DEPENDENCY BARRIERS
551 ------------------------
553 The usage requirements of data dependency barriers are a little subtle, and
554 it's not always obvious that they're needed. To illustrate, consider the
555 following sequence of events:
557 CPU 1 CPU 2
558 =============== ===============
559 { A == 1, B == 2, C == 3, P == &A, Q == &C }
560 B = 4;
561 <write barrier>
562 WRITE_ONCE(P, &B)
563 Q = READ_ONCE(P);
564 D = *Q;
566 There's a clear data dependency here, and it would seem that by the end of the
567 sequence, Q must be either &A or &B, and that:
569 (Q == &A) implies (D == 1)
570 (Q == &B) implies (D == 4)
572 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
573 leading to the following situation:
575 (Q == &B) and (D == 2) ????
577 Whilst this may seem like a failure of coherency or causality maintenance, it
578 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
579 Alpha).
581 To deal with this, a data dependency barrier or better must be inserted
582 between the address load and the data load:
584 CPU 1 CPU 2
585 =============== ===============
586 { A == 1, B == 2, C == 3, P == &A, Q == &C }
587 B = 4;
588 <write barrier>
589 WRITE_ONCE(P, &B);
590 Q = READ_ONCE(P);
591 <data dependency barrier>
592 D = *Q;
594 This enforces the occurrence of one of the two implications, and prevents the
595 third possibility from arising.
597 A data-dependency barrier must also order against dependent writes:
599 CPU 1 CPU 2
600 =============== ===============
601 { A == 1, B == 2, C = 3, P == &A, Q == &C }
602 B = 4;
603 <write barrier>
604 WRITE_ONCE(P, &B);
605 Q = READ_ONCE(P);
606 <data dependency barrier>
607 *Q = 5;
609 The data-dependency barrier must order the read into Q with the store
610 into *Q. This prohibits this outcome:
612 (Q == &B) && (B == 4)
614 Please note that this pattern should be rare. After all, the whole point
615 of dependency ordering is to -prevent- writes to the data structure, along
616 with the expensive cache misses associated with those writes. This pattern
617 can be used to record rare error conditions and the like, and the ordering
618 prevents such records from being lost.
621 [!] Note that this extremely counterintuitive situation arises most easily on
622 machines with split caches, so that, for example, one cache bank processes
623 even-numbered cache lines and the other bank processes odd-numbered cache
624 lines. The pointer P might be stored in an odd-numbered cache line, and the
625 variable B might be stored in an even-numbered cache line. Then, if the
626 even-numbered bank of the reading CPU's cache is extremely busy while the
627 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
628 but the old value of the variable B (2).
631 The data dependency barrier is very important to the RCU system,
632 for example. See rcu_assign_pointer() and rcu_dereference() in
633 include/linux/rcupdate.h. This permits the current target of an RCU'd
634 pointer to be replaced with a new modified target, without the replacement
635 target appearing to be incompletely initialised.
637 See also the subsection on "Cache Coherency" for a more thorough example.
640 CONTROL DEPENDENCIES
641 --------------------
643 Control dependencies can be a bit tricky because current compilers do
644 not understand them. The purpose of this section is to help you prevent
645 the compiler's ignorance from breaking your code.
647 A load-load control dependency requires a full read memory barrier, not
648 simply a data dependency barrier to make it work correctly. Consider the
649 following bit of code:
651 q = READ_ONCE(a);
652 if (q) {
653 <data dependency barrier> /* BUG: No data dependency!!! */
654 p = READ_ONCE(b);
655 }
657 This will not have the desired effect because there is no actual data
658 dependency, but rather a control dependency that the CPU may short-circuit
659 by attempting to predict the outcome in advance, so that other CPUs see
660 the load from b as having happened before the load from a. In such a
661 case what's actually required is:
663 q = READ_ONCE(a);
664 if (q) {
665 <read barrier>
666 p = READ_ONCE(b);
667 }
669 However, stores are not speculated. This means that ordering -is- provided
670 for load-store control dependencies, as in the following example:
672 q = READ_ONCE(a);
673 if (q) {
674 WRITE_ONCE(b, 1);
675 }
677 Control dependencies pair normally with other types of barriers.
678 That said, please note that neither READ_ONCE() nor WRITE_ONCE()
679 are optional! Without the READ_ONCE(), the compiler might combine the
680 load from 'a' with other loads from 'a'. Without the WRITE_ONCE(),
681 the compiler might combine the store to 'b' with other stores to 'b'.
682 Either can result in highly counterintuitive effects on ordering.
684 Worse yet, if the compiler is able to prove (say) that the value of
685 variable 'a' is always non-zero, it would be well within its rights
686 to optimize the original example by eliminating the "if" statement
687 as follows:
689 q = a;
690 b = 1; /* BUG: Compiler and CPU can both reorder!!! */
692 So don't leave out the READ_ONCE().
694 It is tempting to try to enforce ordering on identical stores on both
695 branches of the "if" statement as follows:
697 q = READ_ONCE(a);
698 if (q) {
699 barrier();
700 WRITE_ONCE(b, 1);
701 do_something();
702 } else {
703 barrier();
704 WRITE_ONCE(b, 1);
705 do_something_else();
706 }
708 Unfortunately, current compilers will transform this as follows at high
709 optimization levels:
711 q = READ_ONCE(a);
712 barrier();
713 WRITE_ONCE(b, 1); /* BUG: No ordering vs. load from a!!! */
714 if (q) {
715 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
716 do_something();
717 } else {
718 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
719 do_something_else();
720 }
722 Now there is no conditional between the load from 'a' and the store to
723 'b', which means that the CPU is within its rights to reorder them:
724 The conditional is absolutely required, and must be present in the
725 assembly code even after all compiler optimizations have been applied.
726 Therefore, if you need ordering in this example, you need explicit
727 memory barriers, for example, smp_store_release():
729 q = READ_ONCE(a);
730 if (q) {
731 smp_store_release(&b, 1);
732 do_something();
733 } else {
734 smp_store_release(&b, 1);
735 do_something_else();
736 }
738 In contrast, without explicit memory barriers, two-legged-if control
739 ordering is guaranteed only when the stores differ, for example:
741 q = READ_ONCE(a);
742 if (q) {
743 WRITE_ONCE(b, 1);
744 do_something();
745 } else {
746 WRITE_ONCE(b, 2);
747 do_something_else();
748 }
750 The initial READ_ONCE() is still required to prevent the compiler from
751 proving the value of 'a'.
753 In addition, you need to be careful what you do with the local variable 'q',
754 otherwise the compiler might be able to guess the value and again remove
755 the needed conditional. For example:
757 q = READ_ONCE(a);
758 if (q % MAX) {
759 WRITE_ONCE(b, 1);
760 do_something();
761 } else {
762 WRITE_ONCE(b, 2);
763 do_something_else();
764 }
766 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
767 equal to zero, in which case the compiler is within its rights to
768 transform the above code into the following:
770 q = READ_ONCE(a);
771 WRITE_ONCE(b, 2);
772 do_something_else();
774 Given this transformation, the CPU is not required to respect the ordering
775 between the load from variable 'a' and the store to variable 'b'. It is
776 tempting to add a barrier(), but this does not help. The conditional
777 is gone, and the barrier won't bring it back. Therefore, if you are
778 relying on this ordering, you should make sure that MAX is greater than
779 one, perhaps as follows:
781 q = READ_ONCE(a);
782 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
783 if (q % MAX) {
784 WRITE_ONCE(b, 1);
785 do_something();
786 } else {
787 WRITE_ONCE(b, 2);
788 do_something_else();
789 }
791 Please note once again that the stores to 'b' differ. If they were
792 identical, as noted earlier, the compiler could pull this store outside
793 of the 'if' statement.
795 You must also be careful not to rely too much on boolean short-circuit
796 evaluation. Consider this example:
798 q = READ_ONCE(a);
799 if (q || 1 > 0)
800 WRITE_ONCE(b, 1);
802 Because the first condition cannot fault and the second condition is
803 always true, the compiler can transform this example as following,
804 defeating control dependency:
806 q = READ_ONCE(a);
807 WRITE_ONCE(b, 1);
809 This example underscores the need to ensure that the compiler cannot
810 out-guess your code. More generally, although READ_ONCE() does force
811 the compiler to actually emit code for a given load, it does not force
812 the compiler to use the results.
814 In addition, control dependencies apply only to the then-clause and
815 else-clause of the if-statement in question. In particular, it does
816 not necessarily apply to code following the if-statement:
818 q = READ_ONCE(a);
819 if (q) {
820 WRITE_ONCE(b, 1);
821 } else {
822 WRITE_ONCE(b, 2);
823 }
824 WRITE_ONCE(c, 1); /* BUG: No ordering against the read from 'a'. */
826 It is tempting to argue that there in fact is ordering because the
827 compiler cannot reorder volatile accesses and also cannot reorder
828 the writes to 'b' with the condition. Unfortunately for this line
829 of reasoning, the compiler might compile the two writes to 'b' as
830 conditional-move instructions, as in this fanciful pseudo-assembly
831 language:
833 ld r1,a
834 cmp r1,$0
835 cmov,ne r4,$1
836 cmov,eq r4,$2
837 st r4,b
838 st $1,c
840 A weakly ordered CPU would have no dependency of any sort between the load
841 from 'a' and the store to 'c'. The control dependencies would extend
842 only to the pair of cmov instructions and the store depending on them.
843 In short, control dependencies apply only to the stores in the then-clause
844 and else-clause of the if-statement in question (including functions
845 invoked by those two clauses), not to code following that if-statement.
847 Finally, control dependencies do -not- provide transitivity. This is
848 demonstrated by two related examples, with the initial values of
849 'x' and 'y' both being zero:
851 CPU 0 CPU 1
852 ======================= =======================
853 r1 = READ_ONCE(x); r2 = READ_ONCE(y);
854 if (r1 > 0) if (r2 > 0)
855 WRITE_ONCE(y, 1); WRITE_ONCE(x, 1);
857 assert(!(r1 == 1 && r2 == 1));
859 The above two-CPU example will never trigger the assert(). However,
860 if control dependencies guaranteed transitivity (which they do not),
861 then adding the following CPU would guarantee a related assertion:
863 CPU 2
864 =====================
865 WRITE_ONCE(x, 2);
867 assert(!(r1 == 2 && r2 == 1 && x == 2)); /* FAILS!!! */
869 But because control dependencies do -not- provide transitivity, the above
870 assertion can fail after the combined three-CPU example completes. If you
871 need the three-CPU example to provide ordering, you will need smp_mb()
872 between the loads and stores in the CPU 0 and CPU 1 code fragments,
873 that is, just before or just after the "if" statements. Furthermore,
874 the original two-CPU example is very fragile and should be avoided.
876 These two examples are the LB and WWC litmus tests from this paper:
877 http://www.cl.cam.ac.uk/users/pes20/ppc-supplemental/test6.pdf and this
878 site: https://www.cl.cam.ac.uk/~pes20/ppcmem/index.html.
880 In summary:
882 (*) Control dependencies can order prior loads against later stores.
883 However, they do -not- guarantee any other sort of ordering:
884 Not prior loads against later loads, nor prior stores against
885 later anything. If you need these other forms of ordering,
886 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
887 later loads, smp_mb().
889 (*) If both legs of the "if" statement begin with identical stores to
890 the same variable, then those stores must be ordered, either by
891 preceding both of them with smp_mb() or by using smp_store_release()
892 to carry out the stores. Please note that it is -not- sufficient
893 to use barrier() at beginning of each leg of the "if" statement
894 because, as shown by the example above, optimizing compilers can
895 destroy the control dependency while respecting the letter of the
896 barrier() law.
898 (*) Control dependencies require at least one run-time conditional
899 between the prior load and the subsequent store, and this
900 conditional must involve the prior load. If the compiler is able
901 to optimize the conditional away, it will have also optimized
902 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE()
903 can help to preserve the needed conditional.
905 (*) Control dependencies require that the compiler avoid reordering the
906 dependency into nonexistence. Careful use of READ_ONCE() or
907 atomic{,64}_read() can help to preserve your control dependency.
908 Please see the COMPILER BARRIER section for more information.
910 (*) Control dependencies apply only to the then-clause and else-clause
911 of the if-statement containing the control dependency, including
912 any functions that these two clauses call. Control dependencies
913 do -not- apply to code following the if-statement containing the
914 control dependency.
916 (*) Control dependencies pair normally with other types of barriers.
918 (*) Control dependencies do -not- provide transitivity. If you
919 need transitivity, use smp_mb().
921 (*) Compilers do not understand control dependencies. It is therefore
922 your job to ensure that they do not break your code.
925 SMP BARRIER PAIRING
926 -------------------
928 When dealing with CPU-CPU interactions, certain types of memory barrier should
929 always be paired. A lack of appropriate pairing is almost certainly an error.
931 General barriers pair with each other, though they also pair with most
932 other types of barriers, albeit without transitivity. An acquire barrier
933 pairs with a release barrier, but both may also pair with other barriers,
934 including of course general barriers. A write barrier pairs with a data
935 dependency barrier, a control dependency, an acquire barrier, a release
936 barrier, a read barrier, or a general barrier. Similarly a read barrier,
937 control dependency, or a data dependency barrier pairs with a write
938 barrier, an acquire barrier, a release barrier, or a general barrier:
940 CPU 1 CPU 2
941 =============== ===============
942 WRITE_ONCE(a, 1);
943 <write barrier>
944 WRITE_ONCE(b, 2); x = READ_ONCE(b);
945 <read barrier>
946 y = READ_ONCE(a);
948 Or:
950 CPU 1 CPU 2
951 =============== ===============================
952 a = 1;
953 <write barrier>
954 WRITE_ONCE(b, &a); x = READ_ONCE(b);
955 <data dependency barrier>
956 y = *x;
958 Or even:
960 CPU 1 CPU 2
961 =============== ===============================
962 r1 = READ_ONCE(y);
963 <general barrier>
964 WRITE_ONCE(y, 1); if (r2 = READ_ONCE(x)) {
965 <implicit control dependency>
966 WRITE_ONCE(y, 1);
967 }
969 assert(r1 == 0 || r2 == 0);
971 Basically, the read barrier always has to be there, even though it can be of
972 the "weaker" type.
974 [!] Note that the stores before the write barrier would normally be expected to
975 match the loads after the read barrier or the data dependency barrier, and vice
976 versa:
978 CPU 1 CPU 2
979 =================== ===================
980 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c);
981 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d);
982 <write barrier> \ <read barrier>
983 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a);
984 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b);
987 EXAMPLES OF MEMORY BARRIER SEQUENCES
988 ------------------------------------
990 Firstly, write barriers act as partial orderings on store operations.
991 Consider the following sequence of events:
993 CPU 1
994 =======================
995 STORE A = 1
996 STORE B = 2
997 STORE C = 3
998 <write barrier>
999 STORE D = 4
1000 STORE E = 5
1002 This sequence of events is committed to the memory coherence system in an order
1003 that the rest of the system might perceive as the unordered set of { STORE A,
1004 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
1005 }:
1007 +-------+ : :
1008 | | +------+
1009 | |------>| C=3 | } /\
1010 | | : +------+ }----- \ -----> Events perceptible to
1011 | | : | A=1 | } \/ the rest of the system
1012 | | : +------+ }
1013 | CPU 1 | : | B=2 | }
1014 | | +------+ }
1015 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
1016 | | +------+ } requires all stores prior to the
1017 | | : | E=5 | } barrier to be committed before
1018 | | : +------+ } further stores may take place
1019 | |------>| D=4 | }
1020 | | +------+
1021 +-------+ : :
1022 |
1023 | Sequence in which stores are committed to the
1024 | memory system by CPU 1
1025 V
1028 Secondly, data dependency barriers act as partial orderings on data-dependent
1029 loads. Consider the following sequence of events:
1031 CPU 1 CPU 2
1032 ======================= =======================
1033 { B = 7; X = 9; Y = 8; C = &Y }
1034 STORE A = 1
1035 STORE B = 2
1036 <write barrier>
1037 STORE C = &B LOAD X
1038 STORE D = 4 LOAD C (gets &B)
1039 LOAD *C (reads B)
1041 Without intervention, CPU 2 may perceive the events on CPU 1 in some
1042 effectively random order, despite the write barrier issued by CPU 1:
1044 +-------+ : : : :
1045 | | +------+ +-------+ | Sequence of update
1046 | |------>| B=2 |----- --->| Y->8 | | of perception on
1047 | | : +------+ \ +-------+ | CPU 2
1048 | CPU 1 | : | A=1 | \ --->| C->&Y | V
1049 | | +------+ | +-------+
1050 | | wwwwwwwwwwwwwwww | : :
1051 | | +------+ | : :
1052 | | : | C=&B |--- | : : +-------+
1053 | | : +------+ \ | +-------+ | |
1054 | |------>| D=4 | ----------->| C->&B |------>| |
1055 | | +------+ | +-------+ | |
1056 +-------+ : : | : : | |
1057 | : : | |
1058 | : : | CPU 2 |
1059 | +-------+ | |
1060 Apparently incorrect ---> | | B->7 |------>| |
1061 perception of B (!) | +-------+ | |
1062 | : : | |
1063 | +-------+ | |
1064 The load of X holds ---> \ | X->9 |------>| |
1065 up the maintenance \ +-------+ | |
1066 of coherence of B ----->| B->2 | +-------+
1067 +-------+
1068 : :
1071 In the above example, CPU 2 perceives that B is 7, despite the load of *C
1072 (which would be B) coming after the LOAD of C.
1074 If, however, a data dependency barrier were to be placed between the load of C
1075 and the load of *C (ie: B) on CPU 2:
1077 CPU 1 CPU 2
1078 ======================= =======================
1079 { B = 7; X = 9; Y = 8; C = &Y }
1080 STORE A = 1
1081 STORE B = 2
1082 <write barrier>
1083 STORE C = &B LOAD X
1084 STORE D = 4 LOAD C (gets &B)
1085 <data dependency barrier>
1086 LOAD *C (reads B)
1088 then the following will occur:
1090 +-------+ : : : :
1091 | | +------+ +-------+
1092 | |------>| B=2 |----- --->| Y->8 |
1093 | | : +------+ \ +-------+
1094 | CPU 1 | : | A=1 | \ --->| C->&Y |
1095 | | +------+ | +-------+
1096 | | wwwwwwwwwwwwwwww | : :
1097 | | +------+ | : :
1098 | | : | C=&B |--- | : : +-------+
1099 | | : +------+ \ | +-------+ | |
1100 | |------>| D=4 | ----------->| C->&B |------>| |
1101 | | +------+ | +-------+ | |
1102 +-------+ : : | : : | |
1103 | : : | |
1104 | : : | CPU 2 |
1105 | +-------+ | |
1106 | | X->9 |------>| |
1107 | +-------+ | |
1108 Makes sure all effects ---> \ ddddddddddddddddd | |
1109 prior to the store of C \ +-------+ | |
1110 are perceptible to ----->| B->2 |------>| |
1111 subsequent loads +-------+ | |
1112 : : +-------+
1115 And thirdly, a read barrier acts as a partial order on loads. Consider the
1116 following sequence of events:
1118 CPU 1 CPU 2
1119 ======================= =======================
1120 { A = 0, B = 9 }
1121 STORE A=1
1122 <write barrier>
1123 STORE B=2
1124 LOAD B
1125 LOAD A
1127 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1128 some effectively random order, despite the write barrier issued by CPU 1:
1130 +-------+ : : : :
1131 | | +------+ +-------+
1132 | |------>| A=1 |------ --->| A->0 |
1133 | | +------+ \ +-------+
1134 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1135 | | +------+ | +-------+
1136 | |------>| B=2 |--- | : :
1137 | | +------+ \ | : : +-------+
1138 +-------+ : : \ | +-------+ | |
1139 ---------->| B->2 |------>| |
1140 | +-------+ | CPU 2 |
1141 | | A->0 |------>| |
1142 | +-------+ | |
1143 | : : +-------+
1144 \ : :
1145 \ +-------+
1146 ---->| A->1 |
1147 +-------+
1148 : :
1151 If, however, a read barrier were to be placed between the load of B and the
1152 load of A on CPU 2:
1154 CPU 1 CPU 2
1155 ======================= =======================
1156 { A = 0, B = 9 }
1157 STORE A=1
1158 <write barrier>
1159 STORE B=2
1160 LOAD B
1161 <read barrier>
1162 LOAD A
1164 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1165 2:
1167 +-------+ : : : :
1168 | | +------+ +-------+
1169 | |------>| A=1 |------ --->| A->0 |
1170 | | +------+ \ +-------+
1171 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1172 | | +------+ | +-------+
1173 | |------>| B=2 |--- | : :
1174 | | +------+ \ | : : +-------+
1175 +-------+ : : \ | +-------+ | |
1176 ---------->| B->2 |------>| |
1177 | +-------+ | CPU 2 |
1178 | : : | |
1179 | : : | |
1180 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1181 barrier causes all effects \ +-------+ | |
1182 prior to the storage of B ---->| A->1 |------>| |
1183 to be perceptible to CPU 2 +-------+ | |
1184 : : +-------+
1187 To illustrate this more completely, consider what could happen if the code
1188 contained a load of A either side of the read barrier:
1190 CPU 1 CPU 2
1191 ======================= =======================
1192 { A = 0, B = 9 }
1193 STORE A=1
1194 <write barrier>
1195 STORE B=2
1196 LOAD B
1197 LOAD A [first load of A]
1198 <read barrier>
1199 LOAD A [second load of A]
1201 Even though the two loads of A both occur after the load of B, they may both
1202 come up with different values:
1204 +-------+ : : : :
1205 | | +------+ +-------+
1206 | |------>| A=1 |------ --->| A->0 |
1207 | | +------+ \ +-------+
1208 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1209 | | +------+ | +-------+
1210 | |------>| B=2 |--- | : :
1211 | | +------+ \ | : : +-------+
1212 +-------+ : : \ | +-------+ | |
1213 ---------->| B->2 |------>| |
1214 | +-------+ | CPU 2 |
1215 | : : | |
1216 | : : | |
1217 | +-------+ | |
1218 | | A->0 |------>| 1st |
1219 | +-------+ | |
1220 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1221 barrier causes all effects \ +-------+ | |
1222 prior to the storage of B ---->| A->1 |------>| 2nd |
1223 to be perceptible to CPU 2 +-------+ | |
1224 : : +-------+
1227 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1228 before the read barrier completes anyway:
1230 +-------+ : : : :
1231 | | +------+ +-------+
1232 | |------>| A=1 |------ --->| A->0 |
1233 | | +------+ \ +-------+
1234 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1235 | | +------+ | +-------+
1236 | |------>| B=2 |--- | : :
1237 | | +------+ \ | : : +-------+
1238 +-------+ : : \ | +-------+ | |
1239 ---------->| B->2 |------>| |
1240 | +-------+ | CPU 2 |
1241 | : : | |
1242 \ : : | |
1243 \ +-------+ | |
1244 ---->| A->1 |------>| 1st |
1245 +-------+ | |
1246 rrrrrrrrrrrrrrrrr | |
1247 +-------+ | |
1248 | A->1 |------>| 2nd |
1249 +-------+ | |
1250 : : +-------+
1253 The guarantee is that the second load will always come up with A == 1 if the
1254 load of B came up with B == 2. No such guarantee exists for the first load of
1255 A; that may come up with either A == 0 or A == 1.
1258 READ MEMORY BARRIERS VS LOAD SPECULATION
1259 ----------------------------------------
1261 Many CPUs speculate with loads: that is they see that they will need to load an
1262 item from memory, and they find a time where they're not using the bus for any
1263 other loads, and so do the load in advance - even though they haven't actually
1264 got to that point in the instruction execution flow yet. This permits the
1265 actual load instruction to potentially complete immediately because the CPU
1266 already has the value to hand.
1268 It may turn out that the CPU didn't actually need the value - perhaps because a
1269 branch circumvented the load - in which case it can discard the value or just
1270 cache it for later use.
1272 Consider:
1274 CPU 1 CPU 2
1275 ======================= =======================
1276 LOAD B
1277 DIVIDE } Divide instructions generally
1278 DIVIDE } take a long time to perform
1279 LOAD A
1281 Which might appear as this:
1283 : : +-------+
1284 +-------+ | |
1285 --->| B->2 |------>| |
1286 +-------+ | CPU 2 |
1287 : :DIVIDE | |
1288 +-------+ | |
1289 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1290 division speculates on the +-------+ ~ | |
1291 LOAD of A : : ~ | |
1292 : :DIVIDE | |
1293 : : ~ | |
1294 Once the divisions are complete --> : : ~-->| |
1295 the CPU can then perform the : : | |
1296 LOAD with immediate effect : : +-------+
1299 Placing a read barrier or a data dependency barrier just before the second
1300 load:
1302 CPU 1 CPU 2
1303 ======================= =======================
1304 LOAD B
1305 DIVIDE
1306 DIVIDE
1307 <read barrier>
1308 LOAD A
1310 will force any value speculatively obtained to be reconsidered to an extent
1311 dependent on the type of barrier used. If there was no change made to the
1312 speculated memory location, then the speculated value will just be used:
1314 : : +-------+
1315 +-------+ | |
1316 --->| B->2 |------>| |
1317 +-------+ | CPU 2 |
1318 : :DIVIDE | |
1319 +-------+ | |
1320 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1321 division speculates on the +-------+ ~ | |
1322 LOAD of A : : ~ | |
1323 : :DIVIDE | |
1324 : : ~ | |
1325 : : ~ | |
1326 rrrrrrrrrrrrrrrr~ | |
1327 : : ~ | |
1328 : : ~-->| |
1329 : : | |
1330 : : +-------+
1333 but if there was an update or an invalidation from another CPU pending, then
1334 the speculation will be cancelled and the value reloaded:
1336 : : +-------+
1337 +-------+ | |
1338 --->| B->2 |------>| |
1339 +-------+ | CPU 2 |
1340 : :DIVIDE | |
1341 +-------+ | |
1342 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1343 division speculates on the +-------+ ~ | |
1344 LOAD of A : : ~ | |
1345 : :DIVIDE | |
1346 : : ~ | |
1347 : : ~ | |
1348 rrrrrrrrrrrrrrrrr | |
1349 +-------+ | |
1350 The speculation is discarded ---> --->| A->1 |------>| |
1351 and an updated value is +-------+ | |
1352 retrieved : : +-------+
1355 TRANSITIVITY
1356 ------------
1358 Transitivity is a deeply intuitive notion about ordering that is not
1359 always provided by real computer systems. The following example
1360 demonstrates transitivity:
1362 CPU 1 CPU 2 CPU 3
1363 ======================= ======================= =======================
1364 { X = 0, Y = 0 }
1365 STORE X=1 LOAD X STORE Y=1
1366 <general barrier> <general barrier>
1367 LOAD Y LOAD X
1369 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1370 This indicates that CPU 2's load from X in some sense follows CPU 1's
1371 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1372 store to Y. The question is then "Can CPU 3's load from X return 0?"
1374 Because CPU 2's load from X in some sense came after CPU 1's store, it
1375 is natural to expect that CPU 3's load from X must therefore return 1.
1376 This expectation is an example of transitivity: if a load executing on
1377 CPU A follows a load from the same variable executing on CPU B, then
1378 CPU A's load must either return the same value that CPU B's load did,
1379 or must return some later value.
1381 In the Linux kernel, use of general memory barriers guarantees
1382 transitivity. Therefore, in the above example, if CPU 2's load from X
1383 returns 1 and its load from Y returns 0, then CPU 3's load from X must
1384 also return 1.
1386 However, transitivity is -not- guaranteed for read or write barriers.
1387 For example, suppose that CPU 2's general barrier in the above example
1388 is changed to a read barrier as shown below:
1390 CPU 1 CPU 2 CPU 3
1391 ======================= ======================= =======================
1392 { X = 0, Y = 0 }
1393 STORE X=1 LOAD X STORE Y=1
1394 <read barrier> <general barrier>
1395 LOAD Y LOAD X
1397 This substitution destroys transitivity: in this example, it is perfectly
1398 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1399 and CPU 3's load from X to return 0.
1401 The key point is that although CPU 2's read barrier orders its pair
1402 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1403 this example runs on a system where CPUs 1 and 2 share a store buffer
1404 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1405 General barriers are therefore required to ensure that all CPUs agree
1406 on the combined order of CPU 1's and CPU 2's accesses.
1408 General barriers provide "global transitivity", so that all CPUs will
1409 agree on the order of operations. In contrast, a chain of release-acquire
1410 pairs provides only "local transitivity", so that only those CPUs on
1411 the chain are guaranteed to agree on the combined order of the accesses.
1412 For example, switching to C code in deference to Herman Hollerith:
1414 int u, v, x, y, z;
1416 void cpu0(void)
1417 {
1418 r0 = smp_load_acquire(&x);
1419 WRITE_ONCE(u, 1);
1420 smp_store_release(&y, 1);
1421 }
1423 void cpu1(void)
1424 {
1425 r1 = smp_load_acquire(&y);
1426 r4 = READ_ONCE(v);
1427 r5 = READ_ONCE(u);
1428 smp_store_release(&z, 1);
1429 }
1431 void cpu2(void)
1432 {
1433 r2 = smp_load_acquire(&z);
1434 smp_store_release(&x, 1);
1435 }
1437 void cpu3(void)
1438 {
1439 WRITE_ONCE(v, 1);
1440 smp_mb();
1441 r3 = READ_ONCE(u);
1442 }
1444 Because cpu0(), cpu1(), and cpu2() participate in a local transitive
1445 chain of smp_store_release()/smp_load_acquire() pairs, the following
1446 outcome is prohibited:
1448 r0 == 1 && r1 == 1 && r2 == 1
1450 Furthermore, because of the release-acquire relationship between cpu0()
1451 and cpu1(), cpu1() must see cpu0()'s writes, so that the following
1452 outcome is prohibited:
1454 r1 == 1 && r5 == 0
1456 However, the transitivity of release-acquire is local to the participating
1457 CPUs and does not apply to cpu3(). Therefore, the following outcome
1458 is possible:
1460 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0
1462 As an aside, the following outcome is also possible:
1464 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1
1466 Although cpu0(), cpu1(), and cpu2() will see their respective reads and
1467 writes in order, CPUs not involved in the release-acquire chain might
1468 well disagree on the order. This disagreement stems from the fact that
1469 the weak memory-barrier instructions used to implement smp_load_acquire()
1470 and smp_store_release() are not required to order prior stores against
1471 subsequent loads in all cases. This means that cpu3() can see cpu0()'s
1472 store to u as happening -after- cpu1()'s load from v, even though
1473 both cpu0() and cpu1() agree that these two operations occurred in the
1474 intended order.
1476 However, please keep in mind that smp_load_acquire() is not magic.
1477 In particular, it simply reads from its argument with ordering. It does
1478 -not- ensure that any particular value will be read. Therefore, the
1479 following outcome is possible:
1481 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0
1483 Note that this outcome can happen even on a mythical sequentially
1484 consistent system where nothing is ever reordered.
1486 To reiterate, if your code requires global transitivity, use general
1487 barriers throughout.
1490 ========================
1491 EXPLICIT KERNEL BARRIERS
1492 ========================
1494 The Linux kernel has a variety of different barriers that act at different
1495 levels:
1497 (*) Compiler barrier.
1499 (*) CPU memory barriers.
1501 (*) MMIO write barrier.
1504 COMPILER BARRIER
1505 ----------------
1507 The Linux kernel has an explicit compiler barrier function that prevents the
1508 compiler from moving the memory accesses either side of it to the other side:
1510 barrier();
1512 This is a general barrier -- there are no read-read or write-write
1513 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be
1514 thought of as weak forms of barrier() that affect only the specific
1515 accesses flagged by the READ_ONCE() or WRITE_ONCE().
1517 The barrier() function has the following effects:
1519 (*) Prevents the compiler from reordering accesses following the
1520 barrier() to precede any accesses preceding the barrier().
1521 One example use for this property is to ease communication between
1522 interrupt-handler code and the code that was interrupted.
1524 (*) Within a loop, forces the compiler to load the variables used
1525 in that loop's conditional on each pass through that loop.
1527 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of
1528 optimizations that, while perfectly safe in single-threaded code, can
1529 be fatal in concurrent code. Here are some examples of these sorts
1530 of optimizations:
1532 (*) The compiler is within its rights to reorder loads and stores
1533 to the same variable, and in some cases, the CPU is within its
1534 rights to reorder loads to the same variable. This means that
1535 the following code:
1537 a[0] = x;
1538 a[1] = x;
1540 Might result in an older value of x stored in a[1] than in a[0].
1541 Prevent both the compiler and the CPU from doing this as follows:
1543 a[0] = READ_ONCE(x);
1544 a[1] = READ_ONCE(x);
1546 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for
1547 accesses from multiple CPUs to a single variable.
1549 (*) The compiler is within its rights to merge successive loads from
1550 the same variable. Such merging can cause the compiler to "optimize"
1551 the following code:
1553 while (tmp = a)
1554 do_something_with(tmp);
1556 into the following code, which, although in some sense legitimate
1557 for single-threaded code, is almost certainly not what the developer
1558 intended:
1560 if (tmp = a)
1561 for (;;)
1562 do_something_with(tmp);
1564 Use READ_ONCE() to prevent the compiler from doing this to you:
1566 while (tmp = READ_ONCE(a))
1567 do_something_with(tmp);
1569 (*) The compiler is within its rights to reload a variable, for example,
1570 in cases where high register pressure prevents the compiler from
1571 keeping all data of interest in registers. The compiler might
1572 therefore optimize the variable 'tmp' out of our previous example:
1574 while (tmp = a)
1575 do_something_with(tmp);
1577 This could result in the following code, which is perfectly safe in
1578 single-threaded code, but can be fatal in concurrent code:
1580 while (a)
1581 do_something_with(a);
1583 For example, the optimized version of this code could result in
1584 passing a zero to do_something_with() in the case where the variable
1585 a was modified by some other CPU between the "while" statement and
1586 the call to do_something_with().
1588 Again, use READ_ONCE() to prevent the compiler from doing this:
1590 while (tmp = READ_ONCE(a))
1591 do_something_with(tmp);
1593 Note that if the compiler runs short of registers, it might save
1594 tmp onto the stack. The overhead of this saving and later restoring
1595 is why compilers reload variables. Doing so is perfectly safe for
1596 single-threaded code, so you need to tell the compiler about cases
1597 where it is not safe.
1599 (*) The compiler is within its rights to omit a load entirely if it knows
1600 what the value will be. For example, if the compiler can prove that
1601 the value of variable 'a' is always zero, it can optimize this code:
1603 while (tmp = a)
1604 do_something_with(tmp);
1606 Into this:
1608 do { } while (0);
1610 This transformation is a win for single-threaded code because it
1611 gets rid of a load and a branch. The problem is that the compiler
1612 will carry out its proof assuming that the current CPU is the only
1613 one updating variable 'a'. If variable 'a' is shared, then the
1614 compiler's proof will be erroneous. Use READ_ONCE() to tell the
1615 compiler that it doesn't know as much as it thinks it does:
1617 while (tmp = READ_ONCE(a))
1618 do_something_with(tmp);
1620 But please note that the compiler is also closely watching what you
1621 do with the value after the READ_ONCE(). For example, suppose you
1622 do the following and MAX is a preprocessor macro with the value 1:
1624 while ((tmp = READ_ONCE(a)) % MAX)
1625 do_something_with(tmp);
1627 Then the compiler knows that the result of the "%" operator applied
1628 to MAX will always be zero, again allowing the compiler to optimize
1629 the code into near-nonexistence. (It will still load from the
1630 variable 'a'.)
1632 (*) Similarly, the compiler is within its rights to omit a store entirely
1633 if it knows that the variable already has the value being stored.
1634 Again, the compiler assumes that the current CPU is the only one
1635 storing into the variable, which can cause the compiler to do the
1636 wrong thing for shared variables. For example, suppose you have
1637 the following:
1639 a = 0;
1640 ... Code that does not store to variable a ...
1641 a = 0;
1643 The compiler sees that the value of variable 'a' is already zero, so
1644 it might well omit the second store. This would come as a fatal
1645 surprise if some other CPU might have stored to variable 'a' in the
1646 meantime.
1648 Use WRITE_ONCE() to prevent the compiler from making this sort of
1649 wrong guess:
1651 WRITE_ONCE(a, 0);
1652 ... Code that does not store to variable a ...
1653 WRITE_ONCE(a, 0);
1655 (*) The compiler is within its rights to reorder memory accesses unless
1656 you tell it not to. For example, consider the following interaction
1657 between process-level code and an interrupt handler:
1659 void process_level(void)
1660 {
1661 msg = get_message();
1662 flag = true;
1663 }
1665 void interrupt_handler(void)
1666 {
1667 if (flag)
1668 process_message(msg);
1669 }
1671 There is nothing to prevent the compiler from transforming
1672 process_level() to the following, in fact, this might well be a
1673 win for single-threaded code:
1675 void process_level(void)
1676 {
1677 flag = true;
1678 msg = get_message();
1679 }
1681 If the interrupt occurs between these two statement, then
1682 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE()
1683 to prevent this as follows:
1685 void process_level(void)
1686 {
1687 WRITE_ONCE(msg, get_message());
1688 WRITE_ONCE(flag, true);
1689 }
1691 void interrupt_handler(void)
1692 {
1693 if (READ_ONCE(flag))
1694 process_message(READ_ONCE(msg));
1695 }
1697 Note that the READ_ONCE() and WRITE_ONCE() wrappers in
1698 interrupt_handler() are needed if this interrupt handler can itself
1699 be interrupted by something that also accesses 'flag' and 'msg',
1700 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE()
1701 and WRITE_ONCE() are not needed in interrupt_handler() other than
1702 for documentation purposes. (Note also that nested interrupts
1703 do not typically occur in modern Linux kernels, in fact, if an
1704 interrupt handler returns with interrupts enabled, you will get a
1705 WARN_ONCE() splat.)
1707 You should assume that the compiler can move READ_ONCE() and
1708 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(),
1709 barrier(), or similar primitives.
1711 This effect could also be achieved using barrier(), but READ_ONCE()
1712 and WRITE_ONCE() are more selective: With READ_ONCE() and
1713 WRITE_ONCE(), the compiler need only forget the contents of the
1714 indicated memory locations, while with barrier() the compiler must
1715 discard the value of all memory locations that it has currented
1716 cached in any machine registers. Of course, the compiler must also
1717 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
1718 though the CPU of course need not do so.
1720 (*) The compiler is within its rights to invent stores to a variable,
1721 as in the following example:
1723 if (a)
1724 b = a;
1725 else
1726 b = 42;
1728 The compiler might save a branch by optimizing this as follows:
1730 b = 42;
1731 if (a)
1732 b = a;
1734 In single-threaded code, this is not only safe, but also saves
1735 a branch. Unfortunately, in concurrent code, this optimization
1736 could cause some other CPU to see a spurious value of 42 -- even
1737 if variable 'a' was never zero -- when loading variable 'b'.
1738 Use WRITE_ONCE() to prevent this as follows:
1740 if (a)
1741 WRITE_ONCE(b, a);
1742 else
1743 WRITE_ONCE(b, 42);
1745 The compiler can also invent loads. These are usually less
1746 damaging, but they can result in cache-line bouncing and thus in
1747 poor performance and scalability. Use READ_ONCE() to prevent
1748 invented loads.
1750 (*) For aligned memory locations whose size allows them to be accessed
1751 with a single memory-reference instruction, prevents "load tearing"
1752 and "store tearing," in which a single large access is replaced by
1753 multiple smaller accesses. For example, given an architecture having
1754 16-bit store instructions with 7-bit immediate fields, the compiler
1755 might be tempted to use two 16-bit store-immediate instructions to
1756 implement the following 32-bit store:
1758 p = 0x00010002;
1760 Please note that GCC really does use this sort of optimization,
1761 which is not surprising given that it would likely take more
1762 than two instructions to build the constant and then store it.
1763 This optimization can therefore be a win in single-threaded code.
1764 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1765 this optimization in a volatile store. In the absence of such bugs,
1766 use of WRITE_ONCE() prevents store tearing in the following example:
1768 WRITE_ONCE(p, 0x00010002);
1770 Use of packed structures can also result in load and store tearing,
1771 as in this example:
1773 struct __attribute__((__packed__)) foo {
1774 short a;
1775 int b;
1776 short c;
1777 };
1778 struct foo foo1, foo2;
1779 ...
1781 foo2.a = foo1.a;
1782 foo2.b = foo1.b;
1783 foo2.c = foo1.c;
1785 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no
1786 volatile markings, the compiler would be well within its rights to
1787 implement these three assignment statements as a pair of 32-bit
1788 loads followed by a pair of 32-bit stores. This would result in
1789 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE()
1790 and WRITE_ONCE() again prevent tearing in this example:
1792 foo2.a = foo1.a;
1793 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b));
1794 foo2.c = foo1.c;
1796 All that aside, it is never necessary to use READ_ONCE() and
1797 WRITE_ONCE() on a variable that has been marked volatile. For example,
1798 because 'jiffies' is marked volatile, it is never necessary to
1799 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and
1800 WRITE_ONCE() are implemented as volatile casts, which has no effect when
1801 its argument is already marked volatile.
1803 Please note that these compiler barriers have no direct effect on the CPU,
1804 which may then reorder things however it wishes.
1807 CPU MEMORY BARRIERS
1808 -------------------
1810 The Linux kernel has eight basic CPU memory barriers:
1812 TYPE MANDATORY SMP CONDITIONAL
1813 =============== ======================= ===========================
1814 GENERAL mb() smp_mb()
1815 WRITE wmb() smp_wmb()
1816 READ rmb() smp_rmb()
1817 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1820 All memory barriers except the data dependency barriers imply a compiler
1821 barrier. Data dependencies do not impose any additional compiler ordering.
1823 Aside: In the case of data dependencies, the compiler would be expected
1824 to issue the loads in the correct order (eg. `a[b]` would have to load
1825 the value of b before loading a[b]), however there is no guarantee in
1826 the C specification that the compiler may not speculate the value of b
1827 (eg. is equal to 1) and load a before b (eg. tmp = a[1]; if (b != 1)
1828 tmp = a[b]; ). There is also the problem of a compiler reloading b after
1829 having loaded a[b], thus having a newer copy of b than a[b]. A consensus
1830 has not yet been reached about these problems, however the READ_ONCE()
1831 macro is a good place to start looking.
1833 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1834 systems because it is assumed that a CPU will appear to be self-consistent,
1835 and will order overlapping accesses correctly with respect to itself.
1836 However, see the subsection on "Virtual Machine Guests" below.
1838 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1839 references to shared memory on SMP systems, though the use of locking instead
1840 is sufficient.
1842 Mandatory barriers should not be used to control SMP effects, since mandatory
1843 barriers impose unnecessary overhead on both SMP and UP systems. They may,
1844 however, be used to control MMIO effects on accesses through relaxed memory I/O
1845 windows. These barriers are required even on non-SMP systems as they affect
1846 the order in which memory operations appear to a device by prohibiting both the
1847 compiler and the CPU from reordering them.
1850 There are some more advanced barrier functions:
1852 (*) smp_store_mb(var, value)
1854 This assigns the value to the variable and then inserts a full memory
1855 barrier after it. It isn't guaranteed to insert anything more than a
1856 compiler barrier in a UP compilation.
1859 (*) smp_mb__before_atomic();
1860 (*) smp_mb__after_atomic();
1862 These are for use with atomic (such as add, subtract, increment and
1863 decrement) functions that don't return a value, especially when used for
1864 reference counting. These functions do not imply memory barriers.
1866 These are also used for atomic bitop functions that do not return a
1867 value (such as set_bit and clear_bit).
1869 As an example, consider a piece of code that marks an object as being dead
1870 and then decrements the object's reference count:
1872 obj->dead = 1;
1873 smp_mb__before_atomic();
1874 atomic_dec(&obj->ref_count);
1876 This makes sure that the death mark on the object is perceived to be set
1877 *before* the reference counter is decremented.
1879 See Documentation/core-api/atomic_ops.rst for more information. See the
1880 "Atomic operations" subsection for information on where to use these.
1883 (*) lockless_dereference();
1885 This can be thought of as a pointer-fetch wrapper around the
1886 smp_read_barrier_depends() data-dependency barrier.
1888 This is also similar to rcu_dereference(), but in cases where
1889 object lifetime is handled by some mechanism other than RCU, for
1890 example, when the objects removed only when the system goes down.
1891 In addition, lockless_dereference() is used in some data structures
1892 that can be used both with and without RCU.
1895 (*) dma_wmb();
1896 (*) dma_rmb();
1898 These are for use with consistent memory to guarantee the ordering
1899 of writes or reads of shared memory accessible to both the CPU and a
1900 DMA capable device.
1902 For example, consider a device driver that shares memory with a device
1903 and uses a descriptor status value to indicate if the descriptor belongs
1904 to the device or the CPU, and a doorbell to notify it when new
1905 descriptors are available:
1907 if (desc->status != DEVICE_OWN) {
1908 /* do not read data until we own descriptor */
1909 dma_rmb();
1911 /* read/modify data */
1912 read_data = desc->data;
1913 desc->data = write_data;
1915 /* flush modifications before status update */
1916 dma_wmb();
1918 /* assign ownership */
1919 desc->status = DEVICE_OWN;
1921 /* force memory to sync before notifying device via MMIO */
1922 wmb();
1924 /* notify device of new descriptors */
1925 writel(DESC_NOTIFY, doorbell);
1926 }
1928 The dma_rmb() allows us guarantee the device has released ownership
1929 before we read the data from the descriptor, and the dma_wmb() allows
1930 us to guarantee the data is written to the descriptor before the device
1931 can see it now has ownership. The wmb() is needed to guarantee that the
1932 cache coherent memory writes have completed before attempting a write to
1933 the cache incoherent MMIO region.
1935 See Documentation/DMA-API.txt for more information on consistent memory.
1938 MMIO WRITE BARRIER
1939 ------------------
1941 The Linux kernel also has a special barrier for use with memory-mapped I/O
1942 writes:
1944 mmiowb();
1946 This is a variation on the mandatory write barrier that causes writes to weakly
1947 ordered I/O regions to be partially ordered. Its effects may go beyond the
1948 CPU->Hardware interface and actually affect the hardware at some level.
1950 See the subsection "Acquires vs I/O accesses" for more information.
1953 ===============================
1954 IMPLICIT KERNEL MEMORY BARRIERS
1955 ===============================
1957 Some of the other functions in the linux kernel imply memory barriers, amongst
1958 which are locking and scheduling functions.
1960 This specification is a _minimum_ guarantee; any particular architecture may
1961 provide more substantial guarantees, but these may not be relied upon outside
1962 of arch specific code.
1965 LOCK ACQUISITION FUNCTIONS
1966 --------------------------
1968 The Linux kernel has a number of locking constructs:
1970 (*) spin locks
1971 (*) R/W spin locks
1972 (*) mutexes
1973 (*) semaphores
1974 (*) R/W semaphores
1976 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1977 for each construct. These operations all imply certain barriers:
1979 (1) ACQUIRE operation implication:
1981 Memory operations issued after the ACQUIRE will be completed after the
1982 ACQUIRE operation has completed.
1984 Memory operations issued before the ACQUIRE may be completed after
1985 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1986 combined with a following ACQUIRE, orders prior stores against
1987 subsequent loads and stores. Note that this is weaker than smp_mb()!
1988 The smp_mb__before_spinlock() primitive is free on many architectures.
1990 (2) RELEASE operation implication:
1992 Memory operations issued before the RELEASE will be completed before the
1993 RELEASE operation has completed.
1995 Memory operations issued after the RELEASE may be completed before the
1996 RELEASE operation has completed.
1998 (3) ACQUIRE vs ACQUIRE implication:
2000 All ACQUIRE operations issued before another ACQUIRE operation will be
2001 completed before that ACQUIRE operation.
2003 (4) ACQUIRE vs RELEASE implication:
2005 All ACQUIRE operations issued before a RELEASE operation will be
2006 completed before the RELEASE operation.
2008 (5) Failed conditional ACQUIRE implication:
2010 Certain locking variants of the ACQUIRE operation may fail, either due to
2011 being unable to get the lock immediately, or due to receiving an unblocked
2012 signal whilst asleep waiting for the lock to become available. Failed
2013 locks do not imply any sort of barrier.
2015 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
2016 one-way barriers is that the effects of instructions outside of a critical
2017 section may seep into the inside of the critical section.
2019 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
2020 because it is possible for an access preceding the ACQUIRE to happen after the
2021 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
2022 the two accesses can themselves then cross:
2024 *A = a;
2025 ACQUIRE M
2026 RELEASE M
2027 *B = b;
2029 may occur as:
2031 ACQUIRE M, STORE *B, STORE *A, RELEASE M
2033 When the ACQUIRE and RELEASE are a lock acquisition and release,
2034 respectively, this same reordering can occur if the lock's ACQUIRE and
2035 RELEASE are to the same lock variable, but only from the perspective of
2036 another CPU not holding that lock. In short, a ACQUIRE followed by an
2037 RELEASE may -not- be assumed to be a full memory barrier.
2039 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does
2040 not imply a full memory barrier. Therefore, the CPU's execution of the
2041 critical sections corresponding to the RELEASE and the ACQUIRE can cross,
2042 so that:
2044 *A = a;
2045 RELEASE M
2046 ACQUIRE N
2047 *B = b;
2049 could occur as:
2051 ACQUIRE N, STORE *B, STORE *A, RELEASE M
2053 It might appear that this reordering could introduce a deadlock.
2054 However, this cannot happen because if such a deadlock threatened,
2055 the RELEASE would simply complete, thereby avoiding the deadlock.
2057 Why does this work?
2059 One key point is that we are only talking about the CPU doing
2060 the reordering, not the compiler. If the compiler (or, for
2061 that matter, the developer) switched the operations, deadlock
2062 -could- occur.
2064 But suppose the CPU reordered the operations. In this case,
2065 the unlock precedes the lock in the assembly code. The CPU
2066 simply elected to try executing the later lock operation first.
2067 If there is a deadlock, this lock operation will simply spin (or
2068 try to sleep, but more on that later). The CPU will eventually
2069 execute the unlock operation (which preceded the lock operation
2070 in the assembly code), which will unravel the potential deadlock,
2071 allowing the lock operation to succeed.
2073 But what if the lock is a sleeplock? In that case, the code will
2074 try to enter the scheduler, where it will eventually encounter
2075 a memory barrier, which will force the earlier unlock operation
2076 to complete, again unraveling the deadlock. There might be
2077 a sleep-unlock race, but the locking primitive needs to resolve
2078 such races properly in any case.
2080 Locks and semaphores may not provide any guarantee of ordering on UP compiled
2081 systems, and so cannot be counted on in such a situation to actually achieve
2082 anything at all - especially with respect to I/O accesses - unless combined
2083 with interrupt disabling operations.
2085 See also the section on "Inter-CPU acquiring barrier effects".
2088 As an example, consider the following:
2090 *A = a;
2091 *B = b;
2092 ACQUIRE
2093 *C = c;
2094 *D = d;
2095 RELEASE
2096 *E = e;
2097 *F = f;
2099 The following sequence of events is acceptable:
2101 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
2103 [+] Note that {*F,*A} indicates a combined access.
2105 But none of the following are:
2107 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
2108 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
2109 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
2110 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
2114 INTERRUPT DISABLING FUNCTIONS
2115 -----------------------------
2117 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
2118 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
2119 barriers are required in such a situation, they must be provided from some
2120 other means.
2123 SLEEP AND WAKE-UP FUNCTIONS
2124 ---------------------------
2126 Sleeping and waking on an event flagged in global data can be viewed as an
2127 interaction between two pieces of data: the task state of the task waiting for
2128 the event and the global data used to indicate the event. To make sure that
2129 these appear to happen in the right order, the primitives to begin the process
2130 of going to sleep, and the primitives to initiate a wake up imply certain
2131 barriers.
2133 Firstly, the sleeper normally follows something like this sequence of events:
2135 for (;;) {
2136 set_current_state(TASK_UNINTERRUPTIBLE);
2137 if (event_indicated)
2138 break;
2139 schedule();
2140 }
2142 A general memory barrier is interpolated automatically by set_current_state()
2143 after it has altered the task state:
2145 CPU 1
2146 ===============================
2147 set_current_state();
2148 smp_store_mb();
2149 STORE current->state
2150 <general barrier>
2151 LOAD event_indicated
2153 set_current_state() may be wrapped by:
2155 prepare_to_wait();
2156 prepare_to_wait_exclusive();
2158 which therefore also imply a general memory barrier after setting the state.
2159 The whole sequence above is available in various canned forms, all of which
2160 interpolate the memory barrier in the right place:
2162 wait_event();
2163 wait_event_interruptible();
2164 wait_event_interruptible_exclusive();
2165 wait_event_interruptible_timeout();
2166 wait_event_killable();
2167 wait_event_timeout();
2168 wait_on_bit();
2169 wait_on_bit_lock();
2172 Secondly, code that performs a wake up normally follows something like this:
2174 event_indicated = 1;
2175 wake_up(&event_wait_queue);
2177 or:
2179 event_indicated = 1;
2180 wake_up_process(event_daemon);
2182 A write memory barrier is implied by wake_up() and co. if and only if they
2183 wake something up. The barrier occurs before the task state is cleared, and so
2184 sits between the STORE to indicate the event and the STORE to set TASK_RUNNING:
2186 CPU 1 CPU 2
2187 =============================== ===============================
2188 set_current_state(); STORE event_indicated
2189 smp_store_mb(); wake_up();
2190 STORE current->state <write barrier>
2191 <general barrier> STORE current->state
2192 LOAD event_indicated
2194 To repeat, this write memory barrier is present if and only if something
2195 is actually awakened. To see this, consider the following sequence of
2196 events, where X and Y are both initially zero:
2198 CPU 1 CPU 2
2199 =============================== ===============================
2200 X = 1; STORE event_indicated
2201 smp_mb(); wake_up();
2202 Y = 1; wait_event(wq, Y == 1);
2203 wake_up(); load from Y sees 1, no memory barrier
2204 load from X might see 0
2206 In contrast, if a wakeup does occur, CPU 2's load from X would be guaranteed
2207 to see 1.
2209 The available waker functions include:
2211 complete();
2212 wake_up();
2213 wake_up_all();
2214 wake_up_bit();
2215 wake_up_interruptible();
2216 wake_up_interruptible_all();
2217 wake_up_interruptible_nr();
2218 wake_up_interruptible_poll();
2219 wake_up_interruptible_sync();
2220 wake_up_interruptible_sync_poll();
2221 wake_up_locked();
2222 wake_up_locked_poll();
2223 wake_up_nr();
2224 wake_up_poll();
2225 wake_up_process();
2228 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2229 order multiple stores before the wake-up with respect to loads of those stored
2230 values after the sleeper has called set_current_state(). For instance, if the
2231 sleeper does:
2233 set_current_state(TASK_INTERRUPTIBLE);
2234 if (event_indicated)
2235 break;
2236 __set_current_state(TASK_RUNNING);
2237 do_something(my_data);
2239 and the waker does:
2241 my_data = value;
2242 event_indicated = 1;
2243 wake_up(&event_wait_queue);
2245 there's no guarantee that the change to event_indicated will be perceived by
2246 the sleeper as coming after the change to my_data. In such a circumstance, the
2247 code on both sides must interpolate its own memory barriers between the
2248 separate data accesses. Thus the above sleeper ought to do:
2250 set_current_state(TASK_INTERRUPTIBLE);
2251 if (event_indicated) {
2252 smp_rmb();
2253 do_something(my_data);
2254 }
2256 and the waker should do:
2258 my_data = value;
2259 smp_wmb();
2260 event_indicated = 1;
2261 wake_up(&event_wait_queue);
2264 MISCELLANEOUS FUNCTIONS
2265 -----------------------
2267 Other functions that imply barriers:
2269 (*) schedule() and similar imply full memory barriers.
2272 ===================================
2273 INTER-CPU ACQUIRING BARRIER EFFECTS
2274 ===================================
2276 On SMP systems locking primitives give a more substantial form of barrier: one
2277 that does affect memory access ordering on other CPUs, within the context of
2278 conflict on any particular lock.
2281 ACQUIRES VS MEMORY ACCESSES
2282 ---------------------------
2284 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2285 three CPUs; then should the following sequence of events occur:
2287 CPU 1 CPU 2
2288 =============================== ===============================
2289 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e);
2290 ACQUIRE M ACQUIRE Q
2291 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f);
2292 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g);
2293 RELEASE M RELEASE Q
2294 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h);
2296 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2297 through *H occur in, other than the constraints imposed by the separate locks
2298 on the separate CPUs. It might, for example, see:
2300 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2302 But it won't see any of:
2304 *B, *C or *D preceding ACQUIRE M
2305 *A, *B or *C following RELEASE M
2306 *F, *G or *H preceding ACQUIRE Q
2307 *E, *F or *G following RELEASE Q
2311 ACQUIRES VS I/O ACCESSES
2312 ------------------------
2314 Under certain circumstances (especially involving NUMA), I/O accesses within
2315 two spinlocked sections on two different CPUs may be seen as interleaved by the
2316 PCI bridge, because the PCI bridge does not necessarily participate in the
2317 cache-coherence protocol, and is therefore incapable of issuing the required
2318 read memory barriers.
2320 For example:
2322 CPU 1 CPU 2
2323 =============================== ===============================
2324 spin_lock(Q)
2325 writel(0, ADDR)
2326 writel(1, DATA);
2327 spin_unlock(Q);
2328 spin_lock(Q);
2329 writel(4, ADDR);
2330 writel(5, DATA);
2331 spin_unlock(Q);
2333 may be seen by the PCI bridge as follows:
2335 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2337 which would probably cause the hardware to malfunction.
2340 What is necessary here is to intervene with an mmiowb() before dropping the
2341 spinlock, for example:
2343 CPU 1 CPU 2
2344 =============================== ===============================
2345 spin_lock(Q)
2346 writel(0, ADDR)
2347 writel(1, DATA);
2348 mmiowb();
2349 spin_unlock(Q);
2350 spin_lock(Q);
2351 writel(4, ADDR);
2352 writel(5, DATA);
2353 mmiowb();
2354 spin_unlock(Q);
2356 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2357 before either of the stores issued on CPU 2.
2360 Furthermore, following a store by a load from the same device obviates the need
2361 for the mmiowb(), because the load forces the store to complete before the load
2362 is performed:
2364 CPU 1 CPU 2
2365 =============================== ===============================
2366 spin_lock(Q)
2367 writel(0, ADDR)
2368 a = readl(DATA);
2369 spin_unlock(Q);
2370 spin_lock(Q);
2371 writel(4, ADDR);
2372 b = readl(DATA);
2373 spin_unlock(Q);
2376 See Documentation/driver-api/device-io.rst for more information.
2379 =================================
2380 WHERE ARE MEMORY BARRIERS NEEDED?
2381 =================================
2383 Under normal operation, memory operation reordering is generally not going to
2384 be a problem as a single-threaded linear piece of code will still appear to
2385 work correctly, even if it's in an SMP kernel. There are, however, four
2386 circumstances in which reordering definitely _could_ be a problem:
2388 (*) Interprocessor interaction.
2390 (*) Atomic operations.
2392 (*) Accessing devices.
2394 (*) Interrupts.
2397 INTERPROCESSOR INTERACTION
2398 --------------------------
2400 When there's a system with more than one processor, more than one CPU in the
2401 system may be working on the same data set at the same time. This can cause
2402 synchronisation problems, and the usual way of dealing with them is to use
2403 locks. Locks, however, are quite expensive, and so it may be preferable to
2404 operate without the use of a lock if at all possible. In such a case
2405 operations that affect both CPUs may have to be carefully ordered to prevent
2406 a malfunction.
2408 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2409 queued on the semaphore, by virtue of it having a piece of its stack linked to
2410 the semaphore's list of waiting processes:
2412 struct rw_semaphore {
2413 ...
2414 spinlock_t lock;
2415 struct list_head waiters;
2416 };
2418 struct rwsem_waiter {
2419 struct list_head list;
2420 struct task_struct *task;
2421 };
2423 To wake up a particular waiter, the up_read() or up_write() functions have to:
2425 (1) read the next pointer from this waiter's record to know as to where the
2426 next waiter record is;
2428 (2) read the pointer to the waiter's task structure;
2430 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2432 (4) call wake_up_process() on the task; and
2434 (5) release the reference held on the waiter's task struct.
2436 In other words, it has to perform this sequence of events:
2438 LOAD waiter->list.next;
2439 LOAD waiter->task;
2440 STORE waiter->task;
2441 CALL wakeup
2442 RELEASE task
2444 and if any of these steps occur out of order, then the whole thing may
2445 malfunction.
2447 Once it has queued itself and dropped the semaphore lock, the waiter does not
2448 get the lock again; it instead just waits for its task pointer to be cleared
2449 before proceeding. Since the record is on the waiter's stack, this means that
2450 if the task pointer is cleared _before_ the next pointer in the list is read,
2451 another CPU might start processing the waiter and might clobber the waiter's
2452 stack before the up*() function has a chance to read the next pointer.
2454 Consider then what might happen to the above sequence of events:
2456 CPU 1 CPU 2
2457 =============================== ===============================
2458 down_xxx()
2459 Queue waiter
2460 Sleep
2461 up_yyy()
2462 LOAD waiter->task;
2463 STORE waiter->task;
2464 Woken up by other event
2465 <preempt>
2466 Resume processing
2467 down_xxx() returns
2468 call foo()
2469 foo() clobbers *waiter
2470 </preempt>
2471 LOAD waiter->list.next;
2472 --- OOPS ---
2474 This could be dealt with using the semaphore lock, but then the down_xxx()
2475 function has to needlessly get the spinlock again after being woken up.
2477 The way to deal with this is to insert a general SMP memory barrier:
2479 LOAD waiter->list.next;
2480 LOAD waiter->task;
2481 smp_mb();
2482 STORE waiter->task;
2483 CALL wakeup
2484 RELEASE task
2486 In this case, the barrier makes a guarantee that all memory accesses before the
2487 barrier will appear to happen before all the memory accesses after the barrier
2488 with respect to the other CPUs on the system. It does _not_ guarantee that all
2489 the memory accesses before the barrier will be complete by the time the barrier
2490 instruction itself is complete.
2492 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2493 compiler barrier, thus making sure the compiler emits the instructions in the
2494 right order without actually intervening in the CPU. Since there's only one
2495 CPU, that CPU's dependency ordering logic will take care of everything else.
2498 ATOMIC OPERATIONS
2499 -----------------
2501 Whilst they are technically interprocessor interaction considerations, atomic
2502 operations are noted specially as some of them imply full memory barriers and
2503 some don't, but they're very heavily relied on as a group throughout the
2504 kernel.
2506 Any atomic operation that modifies some state in memory and returns information
2507 about the state (old or new) implies an SMP-conditional general memory barrier
2508 (smp_mb()) on each side of the actual operation (with the exception of
2509 explicit lock operations, described later). These include:
2511 xchg();
2512 atomic_xchg(); atomic_long_xchg();
2513 atomic_inc_return(); atomic_long_inc_return();
2514 atomic_dec_return(); atomic_long_dec_return();
2515 atomic_add_return(); atomic_long_add_return();
2516 atomic_sub_return(); atomic_long_sub_return();
2517 atomic_inc_and_test(); atomic_long_inc_and_test();
2518 atomic_dec_and_test(); atomic_long_dec_and_test();
2519 atomic_sub_and_test(); atomic_long_sub_and_test();
2520 atomic_add_negative(); atomic_long_add_negative();
2521 test_and_set_bit();
2522 test_and_clear_bit();
2523 test_and_change_bit();
2525 /* when succeeds */
2526 cmpxchg();
2527 atomic_cmpxchg(); atomic_long_cmpxchg();
2528 atomic_add_unless(); atomic_long_add_unless();
2530 These are used for such things as implementing ACQUIRE-class and RELEASE-class
2531 operations and adjusting reference counters towards object destruction, and as
2532 such the implicit memory barrier effects are necessary.
2535 The following operations are potential problems as they do _not_ imply memory
2536 barriers, but might be used for implementing such things as RELEASE-class
2537 operations:
2539 atomic_set();
2540 set_bit();
2541 clear_bit();
2542 change_bit();
2544 With these the appropriate explicit memory barrier should be used if necessary
2545 (smp_mb__before_atomic() for instance).
2548 The following also do _not_ imply memory barriers, and so may require explicit
2549 memory barriers under some circumstances (smp_mb__before_atomic() for
2550 instance):
2552 atomic_add();
2553 atomic_sub();
2554 atomic_inc();
2555 atomic_dec();
2557 If they're used for statistics generation, then they probably don't need memory
2558 barriers, unless there's a coupling between statistical data.
2560 If they're used for reference counting on an object to control its lifetime,
2561 they probably don't need memory barriers because either the reference count
2562 will be adjusted inside a locked section, or the caller will already hold
2563 sufficient references to make the lock, and thus a memory barrier unnecessary.
2565 If they're used for constructing a lock of some description, then they probably
2566 do need memory barriers as a lock primitive generally has to do things in a
2567 specific order.
2569 Basically, each usage case has to be carefully considered as to whether memory
2570 barriers are needed or not.
2572 The following operations are special locking primitives:
2574 test_and_set_bit_lock();
2575 clear_bit_unlock();
2576 __clear_bit_unlock();
2578 These implement ACQUIRE-class and RELEASE-class operations. These should be
2579 used in preference to other operations when implementing locking primitives,
2580 because their implementations can be optimised on many architectures.
2582 [!] Note that special memory barrier primitives are available for these
2583 situations because on some CPUs the atomic instructions used imply full memory
2584 barriers, and so barrier instructions are superfluous in conjunction with them,
2585 and in such cases the special barrier primitives will be no-ops.
2587 See Documentation/core-api/atomic_ops.rst for more information.
2590 ACCESSING DEVICES
2591 -----------------
2593 Many devices can be memory mapped, and so appear to the CPU as if they're just
2594 a set of memory locations. To control such a device, the driver usually has to
2595 make the right memory accesses in exactly the right order.
2597 However, having a clever CPU or a clever compiler creates a potential problem
2598 in that the carefully sequenced accesses in the driver code won't reach the
2599 device in the requisite order if the CPU or the compiler thinks it is more
2600 efficient to reorder, combine or merge accesses - something that would cause
2601 the device to malfunction.
2603 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2604 routines - such as inb() or writel() - which know how to make such accesses
2605 appropriately sequential. Whilst this, for the most part, renders the explicit
2606 use of memory barriers unnecessary, there are a couple of situations where they
2607 might be needed:
2609 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2610 so for _all_ general drivers locks should be used and mmiowb() must be
2611 issued prior to unlocking the critical section.
2613 (2) If the accessor functions are used to refer to an I/O memory window with
2614 relaxed memory access properties, then _mandatory_ memory barriers are
2615 required to enforce ordering.
2617 See Documentation/driver-api/device-io.rst for more information.
2620 INTERRUPTS
2621 ----------
2623 A driver may be interrupted by its own interrupt service routine, and thus the
2624 two parts of the driver may interfere with each other's attempts to control or
2625 access the device.
2627 This may be alleviated - at least in part - by disabling local interrupts (a
2628 form of locking), such that the critical operations are all contained within
2629 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2630 routine is executing, the driver's core may not run on the same CPU, and its
2631 interrupt is not permitted to happen again until the current interrupt has been
2632 handled, thus the interrupt handler does not need to lock against that.
2634 However, consider a driver that was talking to an ethernet card that sports an
2635 address register and a data register. If that driver's core talks to the card
2636 under interrupt-disablement and then the driver's interrupt handler is invoked:
2638 LOCAL IRQ DISABLE
2639 writew(ADDR, 3);
2640 writew(DATA, y);
2641 LOCAL IRQ ENABLE
2642 <interrupt>
2643 writew(ADDR, 4);
2644 q = readw(DATA);
2645 </interrupt>
2647 The store to the data register might happen after the second store to the
2648 address register if ordering rules are sufficiently relaxed:
2650 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2653 If ordering rules are relaxed, it must be assumed that accesses done inside an
2654 interrupt disabled section may leak outside of it and may interleave with
2655 accesses performed in an interrupt - and vice versa - unless implicit or
2656 explicit barriers are used.
2658 Normally this won't be a problem because the I/O accesses done inside such
2659 sections will include synchronous load operations on strictly ordered I/O
2660 registers that form implicit I/O barriers. If this isn't sufficient then an
2661 mmiowb() may need to be used explicitly.
2664 A similar situation may occur between an interrupt routine and two routines
2665 running on separate CPUs that communicate with each other. If such a case is
2666 likely, then interrupt-disabling locks should be used to guarantee ordering.
2669 ==========================
2670 KERNEL I/O BARRIER EFFECTS
2671 ==========================
2673 When accessing I/O memory, drivers should use the appropriate accessor
2674 functions:
2676 (*) inX(), outX():
2678 These are intended to talk to I/O space rather than memory space, but
2679 that's primarily a CPU-specific concept. The i386 and x86_64 processors
2680 do indeed have special I/O space access cycles and instructions, but many
2681 CPUs don't have such a concept.
2683 The PCI bus, amongst others, defines an I/O space concept which - on such
2684 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2685 space. However, it may also be mapped as a virtual I/O space in the CPU's
2686 memory map, particularly on those CPUs that don't support alternate I/O
2687 spaces.
2689 Accesses to this space may be fully synchronous (as on i386), but
2690 intermediary bridges (such as the PCI host bridge) may not fully honour
2691 that.
2693 They are guaranteed to be fully ordered with respect to each other.
2695 They are not guaranteed to be fully ordered with respect to other types of
2696 memory and I/O operation.
2698 (*) readX(), writeX():
2700 Whether these are guaranteed to be fully ordered and uncombined with
2701 respect to each other on the issuing CPU depends on the characteristics
2702 defined for the memory window through which they're accessing. On later
2703 i386 architecture machines, for example, this is controlled by way of the
2704 MTRR registers.
2706 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2707 provided they're not accessing a prefetchable device.
2709 However, intermediary hardware (such as a PCI bridge) may indulge in
2710 deferral if it so wishes; to flush a store, a load from the same location
2711 is preferred[*], but a load from the same device or from configuration
2712 space should suffice for PCI.
2714 [*] NOTE! attempting to load from the same location as was written to may
2715 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2716 example.
2718 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2719 force stores to be ordered.
2721 Please refer to the PCI specification for more information on interactions
2722 between PCI transactions.
2724 (*) readX_relaxed(), writeX_relaxed()
2726 These are similar to readX() and writeX(), but provide weaker memory
2727 ordering guarantees. Specifically, they do not guarantee ordering with
2728 respect to normal memory accesses (e.g. DMA buffers) nor do they guarantee
2729 ordering with respect to LOCK or UNLOCK operations. If the latter is
2730 required, an mmiowb() barrier can be used. Note that relaxed accesses to
2731 the same peripheral are guaranteed to be ordered with respect to each
2732 other.
2734 (*) ioreadX(), iowriteX()
2736 These will perform appropriately for the type of access they're actually
2737 doing, be it inX()/outX() or readX()/writeX().
2740 ========================================
2741 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2742 ========================================
2744 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2745 maintain the appearance of program causality with respect to itself. Some CPUs
2746 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2747 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2748 of arch-specific code.
2750 This means that it must be considered that the CPU will execute its instruction
2751 stream in any order it feels like - or even in parallel - provided that if an
2752 instruction in the stream depends on an earlier instruction, then that
2753 earlier instruction must be sufficiently complete[*] before the later
2754 instruction may proceed; in other words: provided that the appearance of
2755 causality is maintained.
2757 [*] Some instructions have more than one effect - such as changing the
2758 condition codes, changing registers or changing memory - and different
2759 instructions may depend on different effects.
2761 A CPU may also discard any instruction sequence that winds up having no
2762 ultimate effect. For example, if two adjacent instructions both load an
2763 immediate value into the same register, the first may be discarded.
2766 Similarly, it has to be assumed that compiler might reorder the instruction
2767 stream in any way it sees fit, again provided the appearance of causality is
2768 maintained.
2771 ============================
2772 THE EFFECTS OF THE CPU CACHE
2773 ============================
2775 The way cached memory operations are perceived across the system is affected to
2776 a certain extent by the caches that lie between CPUs and memory, and by the
2777 memory coherence system that maintains the consistency of state in the system.
2779 As far as the way a CPU interacts with another part of the system through the
2780 caches goes, the memory system has to include the CPU's caches, and memory
2781 barriers for the most part act at the interface between the CPU and its cache
2782 (memory barriers logically act on the dotted line in the following diagram):
2784 <--- CPU ---> : <----------- Memory ----------->
2785 :
2786 +--------+ +--------+ : +--------+ +-----------+
2787 | | | | : | | | | +--------+
2788 | CPU | | Memory | : | CPU | | | | |
2789 | Core |--->| Access |----->| Cache |<-->| | | |
2790 | | | Queue | : | | | |--->| Memory |
2791 | | | | : | | | | | |
2792 +--------+ +--------+ : +--------+ | | | |
2793 : | Cache | +--------+
2794 : | Coherency |
2795 : | Mechanism | +--------+
2796 +--------+ +--------+ : +--------+ | | | |
2797 | | | | : | | | | | |
2798 | CPU | | Memory | : | CPU | | |--->| Device |
2799 | Core |--->| Access |----->| Cache |<-->| | | |
2800 | | | Queue | : | | | | | |
2801 | | | | : | | | | +--------+
2802 +--------+ +--------+ : +--------+ +-----------+
2803 :
2804 :
2806 Although any particular load or store may not actually appear outside of the
2807 CPU that issued it since it may have been satisfied within the CPU's own cache,
2808 it will still appear as if the full memory access had taken place as far as the
2809 other CPUs are concerned since the cache coherency mechanisms will migrate the
2810 cacheline over to the accessing CPU and propagate the effects upon conflict.
2812 The CPU core may execute instructions in any order it deems fit, provided the
2813 expected program causality appears to be maintained. Some of the instructions
2814 generate load and store operations which then go into the queue of memory
2815 accesses to be performed. The core may place these in the queue in any order
2816 it wishes, and continue execution until it is forced to wait for an instruction
2817 to complete.
2819 What memory barriers are concerned with is controlling the order in which
2820 accesses cross from the CPU side of things to the memory side of things, and
2821 the order in which the effects are perceived to happen by the other observers
2822 in the system.
2824 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2825 their own loads and stores as if they had happened in program order.
2827 [!] MMIO or other device accesses may bypass the cache system. This depends on
2828 the properties of the memory window through which devices are accessed and/or
2829 the use of any special device communication instructions the CPU may have.
2832 CACHE COHERENCY
2833 ---------------
2835 Life isn't quite as simple as it may appear above, however: for while the
2836 caches are expected to be coherent, there's no guarantee that that coherency
2837 will be ordered. This means that whilst changes made on one CPU will
2838 eventually become visible on all CPUs, there's no guarantee that they will
2839 become apparent in the same order on those other CPUs.
2842 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2843 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2845 :
2846 : +--------+
2847 : +---------+ | |
2848 +--------+ : +--->| Cache A |<------->| |
2849 | | : | +---------+ | |
2850 | CPU 1 |<---+ | |
2851 | | : | +---------+ | |
2852 +--------+ : +--->| Cache B |<------->| |
2853 : +---------+ | |
2854 : | Memory |
2855 : +---------+ | System |
2856 +--------+ : +--->| Cache C |<------->| |
2857 | | : | +---------+ | |
2858 | CPU 2 |<---+ | |
2859 | | : | +---------+ | |
2860 +--------+ : +--->| Cache D |<------->| |
2861 : +---------+ | |
2862 : +--------+
2863 :
2865 Imagine the system has the following properties:
2867 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2868 resident in memory;
2870 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2871 resident in memory;
2873 (*) whilst the CPU core is interrogating one cache, the other cache may be
2874 making use of the bus to access the rest of the system - perhaps to
2875 displace a dirty cacheline or to do a speculative load;
2877 (*) each cache has a queue of operations that need to be applied to that cache
2878 to maintain coherency with the rest of the system;
2880 (*) the coherency queue is not flushed by normal loads to lines already
2881 present in the cache, even though the contents of the queue may
2882 potentially affect those loads.
2884 Imagine, then, that two writes are made on the first CPU, with a write barrier
2885 between them to guarantee that they will appear to reach that CPU's caches in
2886 the requisite order:
2888 CPU 1 CPU 2 COMMENT
2889 =============== =============== =======================================
2890 u == 0, v == 1 and p == &u, q == &u
2891 v = 2;
2892 smp_wmb(); Make sure change to v is visible before
2893 change to p
2894 <A:modify v=2> v is now in cache A exclusively
2895 p = &v;
2896 <B:modify p=&v> p is now in cache B exclusively
2898 The write memory barrier forces the other CPUs in the system to perceive that
2899 the local CPU's caches have apparently been updated in the correct order. But
2900 now imagine that the second CPU wants to read those values:
2902 CPU 1 CPU 2 COMMENT
2903 =============== =============== =======================================
2904 ...
2905 q = p;
2906 x = *q;
2908 The above pair of reads may then fail to happen in the expected order, as the
2909 cacheline holding p may get updated in one of the second CPU's caches whilst
2910 the update to the cacheline holding v is delayed in the other of the second
2911 CPU's caches by some other cache event:
2913 CPU 1 CPU 2 COMMENT
2914 =============== =============== =======================================
2915 u == 0, v == 1 and p == &u, q == &u
2916 v = 2;
2917 smp_wmb();
2918 <A:modify v=2> <C:busy>
2919 <C:queue v=2>
2920 p = &v; q = p;
2921 <D:request p>
2922 <B:modify p=&v> <D:commit p=&v>
2923 <D:read p>
2924 x = *q;
2925 <C:read *q> Reads from v before v updated in cache
2926 <C:unbusy>
2927 <C:commit v=2>
2929 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2930 no guarantee that, without intervention, the order of update will be the same
2931 as that committed on CPU 1.
2934 To intervene, we need to interpolate a data dependency barrier or a read
2935 barrier between the loads. This will force the cache to commit its coherency
2936 queue before processing any further requests:
2938 CPU 1 CPU 2 COMMENT
2939 =============== =============== =======================================
2940 u == 0, v == 1 and p == &u, q == &u
2941 v = 2;
2942 smp_wmb();
2943 <A:modify v=2> <C:busy>
2944 <C:queue v=2>
2945 p = &v; q = p;
2946 <D:request p>
2947 <B:modify p=&v> <D:commit p=&v>
2948 <D:read p>
2949 smp_read_barrier_depends()
2950 <C:unbusy>
2951 <C:commit v=2>
2952 x = *q;
2953 <C:read *q> Reads from v after v updated in cache
2956 This sort of problem can be encountered on DEC Alpha processors as they have a
2957 split cache that improves performance by making better use of the data bus.
2958 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2959 access depends on a read, not all do, so it may not be relied on.
2961 Other CPUs may also have split caches, but must coordinate between the various
2962 cachelets for normal memory accesses. The semantics of the Alpha removes the
2963 need for coordination in the absence of memory barriers.
2966 CACHE COHERENCY VS DMA
2967 ----------------------
2969 Not all systems maintain cache coherency with respect to devices doing DMA. In
2970 such cases, a device attempting DMA may obtain stale data from RAM because
2971 dirty cache lines may be resident in the caches of various CPUs, and may not
2972 have been written back to RAM yet. To deal with this, the appropriate part of
2973 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2974 invalidate them as well).
2976 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2977 cache lines being written back to RAM from a CPU's cache after the device has
2978 installed its own data, or cache lines present in the CPU's cache may simply
2979 obscure the fact that RAM has been updated, until at such time as the cacheline
2980 is discarded from the CPU's cache and reloaded. To deal with this, the
2981 appropriate part of the kernel must invalidate the overlapping bits of the
2982 cache on each CPU.
2984 See Documentation/cachetlb.txt for more information on cache management.
2987 CACHE COHERENCY VS MMIO
2988 -----------------------
2990 Memory mapped I/O usually takes place through memory locations that are part of
2991 a window in the CPU's memory space that has different properties assigned than
2992 the usual RAM directed window.
2994 Amongst these properties is usually the fact that such accesses bypass the
2995 caching entirely and go directly to the device buses. This means MMIO accesses
2996 may, in effect, overtake accesses to cached memory that were emitted earlier.
2997 A memory barrier isn't sufficient in such a case, but rather the cache must be
2998 flushed between the cached memory write and the MMIO access if the two are in
2999 any way dependent.
3002 =========================
3003 THE THINGS CPUS GET UP TO
3004 =========================
3006 A programmer might take it for granted that the CPU will perform memory
3007 operations in exactly the order specified, so that if the CPU is, for example,
3008 given the following piece of code to execute:
3010 a = READ_ONCE(*A);
3011 WRITE_ONCE(*B, b);
3012 c = READ_ONCE(*C);
3013 d = READ_ONCE(*D);
3014 WRITE_ONCE(*E, e);
3016 they would then expect that the CPU will complete the memory operation for each
3017 instruction before moving on to the next one, leading to a definite sequence of
3018 operations as seen by external observers in the system:
3020 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
3023 Reality is, of course, much messier. With many CPUs and compilers, the above
3024 assumption doesn't hold because:
3026 (*) loads are more likely to need to be completed immediately to permit
3027 execution progress, whereas stores can often be deferred without a
3028 problem;
3030 (*) loads may be done speculatively, and the result discarded should it prove
3031 to have been unnecessary;
3033 (*) loads may be done speculatively, leading to the result having been fetched
3034 at the wrong time in the expected sequence of events;
3036 (*) the order of the memory accesses may be rearranged to promote better use
3037 of the CPU buses and caches;
3039 (*) loads and stores may be combined to improve performance when talking to
3040 memory or I/O hardware that can do batched accesses of adjacent locations,
3041 thus cutting down on transaction setup costs (memory and PCI devices may
3042 both be able to do this); and
3044 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
3045 mechanisms may alleviate this - once the store has actually hit the cache
3046 - there's no guarantee that the coherency management will be propagated in
3047 order to other CPUs.
3049 So what another CPU, say, might actually observe from the above piece of code
3050 is:
3052 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
3054 (Where "LOAD {*C,*D}" is a combined load)
3057 However, it is guaranteed that a CPU will be self-consistent: it will see its
3058 _own_ accesses appear to be correctly ordered, without the need for a memory
3059 barrier. For instance with the following code:
3061 U = READ_ONCE(*A);
3062 WRITE_ONCE(*A, V);
3063 WRITE_ONCE(*A, W);
3064 X = READ_ONCE(*A);
3065 WRITE_ONCE(*A, Y);
3066 Z = READ_ONCE(*A);
3068 and assuming no intervention by an external influence, it can be assumed that
3069 the final result will appear to be:
3071 U == the original value of *A
3072 X == W
3073 Z == Y
3074 *A == Y
3076 The code above may cause the CPU to generate the full sequence of memory
3077 accesses:
3079 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
3081 in that order, but, without intervention, the sequence may have almost any
3082 combination of elements combined or discarded, provided the program's view
3083 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE()
3084 are -not- optional in the above example, as there are architectures
3085 where a given CPU might reorder successive loads to the same location.
3086 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is
3087 necessary to prevent this, for example, on Itanium the volatile casts
3088 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq
3089 and st.rel instructions (respectively) that prevent such reordering.
3091 The compiler may also combine, discard or defer elements of the sequence before
3092 the CPU even sees them.
3094 For instance:
3096 *A = V;
3097 *A = W;
3099 may be reduced to:
3101 *A = W;
3103 since, without either a write barrier or an WRITE_ONCE(), it can be
3104 assumed that the effect of the storage of V to *A is lost. Similarly:
3106 *A = Y;
3107 Z = *A;
3109 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be
3110 reduced to:
3112 *A = Y;
3113 Z = Y;
3115 and the LOAD operation never appear outside of the CPU.
3118 AND THEN THERE'S THE ALPHA
3119 --------------------------
3121 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
3122 some versions of the Alpha CPU have a split data cache, permitting them to have
3123 two semantically-related cache lines updated at separate times. This is where
3124 the data dependency barrier really becomes necessary as this synchronises both
3125 caches with the memory coherence system, thus making it seem like pointer
3126 changes vs new data occur in the right order.
3128 The Alpha defines the Linux kernel's memory barrier model.
3130 See the subsection on "Cache Coherency" above.
3133 VIRTUAL MACHINE GUESTS
3134 ----------------------
3136 Guests running within virtual machines might be affected by SMP effects even if
3137 the guest itself is compiled without SMP support. This is an artifact of
3138 interfacing with an SMP host while running an UP kernel. Using mandatory
3139 barriers for this use-case would be possible but is often suboptimal.
3141 To handle this case optimally, low-level virt_mb() etc macros are available.
3142 These have the same effect as smp_mb() etc when SMP is enabled, but generate
3143 identical code for SMP and non-SMP systems. For example, virtual machine guests
3144 should use virt_mb() rather than smp_mb() when synchronizing against a
3145 (possibly SMP) host.
3147 These are equivalent to smp_mb() etc counterparts in all other respects,
3148 in particular, they do not control MMIO effects: to control
3149 MMIO effects, use mandatory barriers.
3152 ============
3153 EXAMPLE USES
3154 ============
3156 CIRCULAR BUFFERS
3157 ----------------
3159 Memory barriers can be used to implement circular buffering without the need
3160 of a lock to serialise the producer with the consumer. See:
3162 Documentation/circular-buffers.txt
3164 for details.
3167 ==========
3168 REFERENCES
3169 ==========
3171 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
3172 Digital Press)
3173 Chapter 5.2: Physical Address Space Characteristics
3174 Chapter 5.4: Caches and Write Buffers
3175 Chapter 5.5: Data Sharing
3176 Chapter 5.6: Read/Write Ordering
3178 AMD64 Architecture Programmer's Manual Volume 2: System Programming
3179 Chapter 7.1: Memory-Access Ordering
3180 Chapter 7.4: Buffering and Combining Memory Writes
3182 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
3183 System Programming Guide
3184 Chapter 7.1: Locked Atomic Operations
3185 Chapter 7.2: Memory Ordering
3186 Chapter 7.4: Serializing Instructions
3188 The SPARC Architecture Manual, Version 9
3189 Chapter 8: Memory Models
3190 Appendix D: Formal Specification of the Memory Models
3191 Appendix J: Programming with the Memory Models
3193 UltraSPARC Programmer Reference Manual
3194 Chapter 5: Memory Accesses and Cacheability
3195 Chapter 15: Sparc-V9 Memory Models
3197 UltraSPARC III Cu User's Manual
3198 Chapter 9: Memory Models
3200 UltraSPARC IIIi Processor User's Manual
3201 Chapter 8: Memory Models
3203 UltraSPARC Architecture 2005
3204 Chapter 9: Memory
3205 Appendix D: Formal Specifications of the Memory Models
3207 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
3208 Chapter 8: Memory Models
3209 Appendix F: Caches and Cache Coherency
3211 Solaris Internals, Core Kernel Architecture, p63-68:
3212 Chapter 3.3: Hardware Considerations for Locks and
3213 Synchronization
3215 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3216 for Kernel Programmers:
3217 Chapter 13: Other Memory Models
3219 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3220 Section 2.6: Speculation
3221 Section 4.4: Memory Access