1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
5 By: David Howells <dhowells@redhat.com>
6 Paul E. McKenney <paulmck@linux.ibm.com>
7 Will Deacon <will.deacon@arm.com>
8 Peter Zijlstra <peterz@infradead.org>
14 This document is not a specification; it is intentionally (for the sake of
15 brevity) and unintentionally (due to being human) incomplete. This document is
16 meant as a guide to using the various memory barriers provided by Linux, but
17 in case of any doubt (and there are many) please ask. Some doubts may be
18 resolved by referring to the formal memory consistency model and related
19 documentation at tools/memory-model/. Nevertheless, even this memory
20 model should be viewed as the collective opinion of its maintainers rather
21 than as an infallible oracle.
23 To repeat, this document is not a specification of what Linux expects from
26 The purpose of this document is twofold:
28 (1) to specify the minimum functionality that one can rely on for any
29 particular barrier, and
31 (2) to provide a guide as to how to use the barriers that are available.
33 Note that an architecture can provide more than the minimum requirement
34 for any particular barrier, but if the architecture provides less than
35 that, that architecture is incorrect.
37 Note also that it is possible that a barrier may be a no-op for an
38 architecture because the way that arch works renders an explicit barrier
39 unnecessary in that case.
46 (*) Abstract memory access model.
51 (*) What are memory barriers?
53 - Varieties of memory barrier.
54 - What may not be assumed about memory barriers?
55 - Address-dependency barriers (historical).
56 - Control dependencies.
57 - SMP barrier pairing.
58 - Examples of memory barrier sequences.
59 - Read memory barriers vs load speculation.
60 - Multicopy atomicity.
62 (*) Explicit kernel barriers.
65 - CPU memory barriers.
67 (*) Implicit kernel memory barriers.
69 - Lock acquisition functions.
70 - Interrupt disabling functions.
71 - Sleep and wake-up functions.
72 - Miscellaneous functions.
74 (*) Inter-CPU acquiring barrier effects.
76 - Acquires vs memory accesses.
78 (*) Where are memory barriers needed?
80 - Interprocessor interaction.
85 (*) Kernel I/O barrier effects.
87 (*) Assumed minimum execution ordering model.
89 (*) The effects of the cpu cache.
92 - Cache coherency vs DMA.
93 - Cache coherency vs MMIO.
95 (*) The things CPUs get up to.
97 - And then there's the Alpha.
98 - Virtual Machine Guests.
107 ============================
108 ABSTRACT MEMORY ACCESS MODEL
109 ============================
111 Consider the following abstract model of the system:
116 +-------+ : +--------+ : +-------+
119 | CPU 1 |<----->| Memory |<----->| CPU 2 |
122 +-------+ : +--------+ : +-------+
130 +---------->| Device |<----------+
136 Each CPU executes a program that generates memory access operations. In the
137 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
138 perform the memory operations in any order it likes, provided program causality
139 appears to be maintained. Similarly, the compiler may also arrange the
140 instructions it emits in any order it likes, provided it doesn't affect the
141 apparent operation of the program.
143 So in the above diagram, the effects of the memory operations performed by a
144 CPU are perceived by the rest of the system as the operations cross the
145 interface between the CPU and rest of the system (the dotted lines).
148 For example, consider the following sequence of events:
151 =============== ===============
156 The set of accesses as seen by the memory system in the middle can be arranged
157 in 24 different combinations:
159 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4
160 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3
161 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4
162 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4
163 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3
164 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4
165 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4
169 and can thus result in four different combinations of values:
177 Furthermore, the stores committed by a CPU to the memory system may not be
178 perceived by the loads made by another CPU in the same order as the stores were
182 As a further example, consider this sequence of events:
185 =============== ===============
186 { A == 1, B == 2, C == 3, P == &A, Q == &C }
190 There is an obvious address dependency here, as the value loaded into D depends
191 on the address retrieved from P by CPU 2. At the end of the sequence, any of
192 the following results are possible:
194 (Q == &A) and (D == 1)
195 (Q == &B) and (D == 2)
196 (Q == &B) and (D == 4)
198 Note that CPU 2 will never try and load C into D because the CPU will load P
199 into Q before issuing the load of *Q.
205 Some devices present their control interfaces as collections of memory
206 locations, but the order in which the control registers are accessed is very
207 important. For instance, imagine an ethernet card with a set of internal
208 registers that are accessed through an address port register (A) and a data
209 port register (D). To read internal register 5, the following code might then
215 but this might show up as either of the following two sequences:
217 STORE *A = 5, x = LOAD *D
218 x = LOAD *D, STORE *A = 5
220 the second of which will almost certainly result in a malfunction, since it set
221 the address _after_ attempting to read the register.
227 There are some minimal guarantees that may be expected of a CPU:
229 (*) On any given CPU, dependent memory accesses will be issued in order, with
230 respect to itself. This means that for:
232 Q = READ_ONCE(P); D = READ_ONCE(*Q);
234 the CPU will issue the following memory operations:
236 Q = LOAD P, D = LOAD *Q
238 and always in that order. However, on DEC Alpha, READ_ONCE() also
239 emits a memory-barrier instruction, so that a DEC Alpha CPU will
240 instead issue the following memory operations:
242 Q = LOAD P, MEMORY_BARRIER, D = LOAD *Q, MEMORY_BARRIER
244 Whether on DEC Alpha or not, the READ_ONCE() also prevents compiler
247 (*) Overlapping loads and stores within a particular CPU will appear to be
248 ordered within that CPU. This means that for:
250 a = READ_ONCE(*X); WRITE_ONCE(*X, b);
252 the CPU will only issue the following sequence of memory operations:
254 a = LOAD *X, STORE *X = b
258 WRITE_ONCE(*X, c); d = READ_ONCE(*X);
260 the CPU will only issue:
262 STORE *X = c, d = LOAD *X
264 (Loads and stores overlap if they are targeted at overlapping pieces of
267 And there are a number of things that _must_ or _must_not_ be assumed:
269 (*) It _must_not_ be assumed that the compiler will do what you want
270 with memory references that are not protected by READ_ONCE() and
271 WRITE_ONCE(). Without them, the compiler is within its rights to
272 do all sorts of "creative" transformations, which are covered in
273 the COMPILER BARRIER section.
275 (*) It _must_not_ be assumed that independent loads and stores will be issued
276 in the order given. This means that for:
278 X = *A; Y = *B; *D = Z;
280 we may get any of the following sequences:
282 X = LOAD *A, Y = LOAD *B, STORE *D = Z
283 X = LOAD *A, STORE *D = Z, Y = LOAD *B
284 Y = LOAD *B, X = LOAD *A, STORE *D = Z
285 Y = LOAD *B, STORE *D = Z, X = LOAD *A
286 STORE *D = Z, X = LOAD *A, Y = LOAD *B
287 STORE *D = Z, Y = LOAD *B, X = LOAD *A
289 (*) It _must_ be assumed that overlapping memory accesses may be merged or
290 discarded. This means that for:
292 X = *A; Y = *(A + 4);
294 we may get any one of the following sequences:
296 X = LOAD *A; Y = LOAD *(A + 4);
297 Y = LOAD *(A + 4); X = LOAD *A;
298 {X, Y} = LOAD {*A, *(A + 4) };
302 *A = X; *(A + 4) = Y;
306 STORE *A = X; STORE *(A + 4) = Y;
307 STORE *(A + 4) = Y; STORE *A = X;
308 STORE {*A, *(A + 4) } = {X, Y};
310 And there are anti-guarantees:
312 (*) These guarantees do not apply to bitfields, because compilers often
313 generate code to modify these using non-atomic read-modify-write
314 sequences. Do not attempt to use bitfields to synchronize parallel
317 (*) Even in cases where bitfields are protected by locks, all fields
318 in a given bitfield must be protected by one lock. If two fields
319 in a given bitfield are protected by different locks, the compiler's
320 non-atomic read-modify-write sequences can cause an update to one
321 field to corrupt the value of an adjacent field.
323 (*) These guarantees apply only to properly aligned and sized scalar
324 variables. "Properly sized" currently means variables that are
325 the same size as "char", "short", "int" and "long". "Properly
326 aligned" means the natural alignment, thus no constraints for
327 "char", two-byte alignment for "short", four-byte alignment for
328 "int", and either four-byte or eight-byte alignment for "long",
329 on 32-bit and 64-bit systems, respectively. Note that these
330 guarantees were introduced into the C11 standard, so beware when
331 using older pre-C11 compilers (for example, gcc 4.6). The portion
332 of the standard containing this guarantee is Section 3.14, which
333 defines "memory location" as follows:
336 either an object of scalar type, or a maximal sequence
337 of adjacent bit-fields all having nonzero width
339 NOTE 1: Two threads of execution can update and access
340 separate memory locations without interfering with
343 NOTE 2: A bit-field and an adjacent non-bit-field member
344 are in separate memory locations. The same applies
345 to two bit-fields, if one is declared inside a nested
346 structure declaration and the other is not, or if the two
347 are separated by a zero-length bit-field declaration,
348 or if they are separated by a non-bit-field member
349 declaration. It is not safe to concurrently update two
350 bit-fields in the same structure if all members declared
351 between them are also bit-fields, no matter what the
352 sizes of those intervening bit-fields happen to be.
355 =========================
356 WHAT ARE MEMORY BARRIERS?
357 =========================
359 As can be seen above, independent memory operations are effectively performed
360 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
361 What is required is some way of intervening to instruct the compiler and the
362 CPU to restrict the order.
364 Memory barriers are such interventions. They impose a perceived partial
365 ordering over the memory operations on either side of the barrier.
367 Such enforcement is important because the CPUs and other devices in a system
368 can use a variety of tricks to improve performance, including reordering,
369 deferral and combination of memory operations; speculative loads; speculative
370 branch prediction and various types of caching. Memory barriers are used to
371 override or suppress these tricks, allowing the code to sanely control the
372 interaction of multiple CPUs and/or devices.
375 VARIETIES OF MEMORY BARRIER
376 ---------------------------
378 Memory barriers come in four basic varieties:
380 (1) Write (or store) memory barriers.
382 A write memory barrier gives a guarantee that all the STORE operations
383 specified before the barrier will appear to happen before all the STORE
384 operations specified after the barrier with respect to the other
385 components of the system.
387 A write barrier is a partial ordering on stores only; it is not required
388 to have any effect on loads.
390 A CPU can be viewed as committing a sequence of store operations to the
391 memory system as time progresses. All stores _before_ a write barrier
392 will occur _before_ all the stores after the write barrier.
394 [!] Note that write barriers should normally be paired with read or
395 address-dependency barriers; see the "SMP barrier pairing" subsection.
398 (2) Address-dependency barriers (historical).
400 An address-dependency barrier is a weaker form of read barrier. In the
401 case where two loads are performed such that the second depends on the
402 result of the first (eg: the first load retrieves the address to which
403 the second load will be directed), an address-dependency barrier would
404 be required to make sure that the target of the second load is updated
405 after the address obtained by the first load is accessed.
407 An address-dependency barrier is a partial ordering on interdependent
408 loads only; it is not required to have any effect on stores, independent
409 loads or overlapping loads.
411 As mentioned in (1), the other CPUs in the system can be viewed as
412 committing sequences of stores to the memory system that the CPU being
413 considered can then perceive. An address-dependency barrier issued by
414 the CPU under consideration guarantees that for any load preceding it,
415 if that load touches one of a sequence of stores from another CPU, then
416 by the time the barrier completes, the effects of all the stores prior to
417 that touched by the load will be perceptible to any loads issued after
418 the address-dependency barrier.
420 See the "Examples of memory barrier sequences" subsection for diagrams
421 showing the ordering constraints.
423 [!] Note that the first load really has to have an _address_ dependency and
424 not a control dependency. If the address for the second load is dependent
425 on the first load, but the dependency is through a conditional rather than
426 actually loading the address itself, then it's a _control_ dependency and
427 a full read barrier or better is required. See the "Control dependencies"
428 subsection for more information.
430 [!] Note that address-dependency barriers should normally be paired with
431 write barriers; see the "SMP barrier pairing" subsection.
433 [!] Kernel release v5.9 removed kernel APIs for explicit address-
434 dependency barriers. Nowadays, APIs for marking loads from shared
435 variables such as READ_ONCE() and rcu_dereference() provide implicit
436 address-dependency barriers.
438 (3) Read (or load) memory barriers.
440 A read barrier is an address-dependency barrier plus a guarantee that all
441 the LOAD operations specified before the barrier will appear to happen
442 before all the LOAD operations specified after the barrier with respect to
443 the other components of the system.
445 A read barrier is a partial ordering on loads only; it is not required to
446 have any effect on stores.
448 Read memory barriers imply address-dependency barriers, and so can
451 [!] Note that read barriers should normally be paired with write barriers;
452 see the "SMP barrier pairing" subsection.
455 (4) General memory barriers.
457 A general memory barrier gives a guarantee that all the LOAD and STORE
458 operations specified before the barrier will appear to happen before all
459 the LOAD and STORE operations specified after the barrier with respect to
460 the other components of the system.
462 A general memory barrier is a partial ordering over both loads and stores.
464 General memory barriers imply both read and write memory barriers, and so
465 can substitute for either.
468 And a couple of implicit varieties:
470 (5) ACQUIRE operations.
472 This acts as a one-way permeable barrier. It guarantees that all memory
473 operations after the ACQUIRE operation will appear to happen after the
474 ACQUIRE operation with respect to the other components of the system.
475 ACQUIRE operations include LOCK operations and both smp_load_acquire()
476 and smp_cond_load_acquire() operations.
478 Memory operations that occur before an ACQUIRE operation may appear to
479 happen after it completes.
481 An ACQUIRE operation should almost always be paired with a RELEASE
485 (6) RELEASE operations.
487 This also acts as a one-way permeable barrier. It guarantees that all
488 memory operations before the RELEASE operation will appear to happen
489 before the RELEASE operation with respect to the other components of the
490 system. RELEASE operations include UNLOCK operations and
491 smp_store_release() operations.
493 Memory operations that occur after a RELEASE operation may appear to
494 happen before it completes.
496 The use of ACQUIRE and RELEASE operations generally precludes the need
497 for other sorts of memory barrier. In addition, a RELEASE+ACQUIRE pair is
498 -not- guaranteed to act as a full memory barrier. However, after an
499 ACQUIRE on a given variable, all memory accesses preceding any prior
500 RELEASE on that same variable are guaranteed to be visible. In other
501 words, within a given variable's critical section, all accesses of all
502 previous critical sections for that variable are guaranteed to have
505 This means that ACQUIRE acts as a minimal "acquire" operation and
506 RELEASE acts as a minimal "release" operation.
508 A subset of the atomic operations described in atomic_t.txt have ACQUIRE and
509 RELEASE variants in addition to fully-ordered and relaxed (no barrier
510 semantics) definitions. For compound atomics performing both a load and a
511 store, ACQUIRE semantics apply only to the load and RELEASE semantics apply
512 only to the store portion of the operation.
514 Memory barriers are only required where there's a possibility of interaction
515 between two CPUs or between a CPU and a device. If it can be guaranteed that
516 there won't be any such interaction in any particular piece of code, then
517 memory barriers are unnecessary in that piece of code.
520 Note that these are the _minimum_ guarantees. Different architectures may give
521 more substantial guarantees, but they may _not_ be relied upon outside of arch
525 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
526 ----------------------------------------------
528 There are certain things that the Linux kernel memory barriers do not guarantee:
530 (*) There is no guarantee that any of the memory accesses specified before a
531 memory barrier will be _complete_ by the completion of a memory barrier
532 instruction; the barrier can be considered to draw a line in that CPU's
533 access queue that accesses of the appropriate type may not cross.
535 (*) There is no guarantee that issuing a memory barrier on one CPU will have
536 any direct effect on another CPU or any other hardware in the system. The
537 indirect effect will be the order in which the second CPU sees the effects
538 of the first CPU's accesses occur, but see the next point:
540 (*) There is no guarantee that a CPU will see the correct order of effects
541 from a second CPU's accesses, even _if_ the second CPU uses a memory
542 barrier, unless the first CPU _also_ uses a matching memory barrier (see
543 the subsection on "SMP Barrier Pairing").
545 (*) There is no guarantee that some intervening piece of off-the-CPU
546 hardware[*] will not reorder the memory accesses. CPU cache coherency
547 mechanisms should propagate the indirect effects of a memory barrier
548 between CPUs, but might not do so in order.
550 [*] For information on bus mastering DMA and coherency please read:
552 Documentation/driver-api/pci/pci.rst
553 Documentation/core-api/dma-api-howto.rst
554 Documentation/core-api/dma-api.rst
557 ADDRESS-DEPENDENCY BARRIERS (HISTORICAL)
558 ----------------------------------------
560 As of v4.15 of the Linux kernel, an smp_mb() was added to READ_ONCE() for
561 DEC Alpha, which means that about the only people who need to pay attention
562 to this section are those working on DEC Alpha architecture-specific code
563 and those working on READ_ONCE() itself. For those who need it, and for
564 those who are interested in the history, here is the story of
565 address-dependency barriers.
567 [!] While address dependencies are observed in both load-to-load and
568 load-to-store relations, address-dependency barriers are not necessary
569 for load-to-store situations.
571 The requirement of address-dependency barriers is a little subtle, and
572 it's not always obvious that they're needed. To illustrate, consider the
573 following sequence of events:
576 =============== ===============
577 { A == 1, B == 2, C == 3, P == &A, Q == &C }
581 Q = READ_ONCE_OLD(P);
584 [!] READ_ONCE_OLD() corresponds to READ_ONCE() of pre-4.15 kernel, which
585 doesn't imply an address-dependency barrier.
587 There's a clear address dependency here, and it would seem that by the end of
588 the sequence, Q must be either &A or &B, and that:
590 (Q == &A) implies (D == 1)
591 (Q == &B) implies (D == 4)
593 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
594 leading to the following situation:
596 (Q == &B) and (D == 2) ????
598 While this may seem like a failure of coherency or causality maintenance, it
599 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
602 To deal with this, READ_ONCE() provides an implicit address-dependency barrier
603 since kernel release v4.15:
606 =============== ===============
607 { A == 1, B == 2, C == 3, P == &A, Q == &C }
612 <implicit address-dependency barrier>
615 This enforces the occurrence of one of the two implications, and prevents the
616 third possibility from arising.
619 [!] Note that this extremely counterintuitive situation arises most easily on
620 machines with split caches, so that, for example, one cache bank processes
621 even-numbered cache lines and the other bank processes odd-numbered cache
622 lines. The pointer P might be stored in an odd-numbered cache line, and the
623 variable B might be stored in an even-numbered cache line. Then, if the
624 even-numbered bank of the reading CPU's cache is extremely busy while the
625 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
626 but the old value of the variable B (2).
629 An address-dependency barrier is not required to order dependent writes
630 because the CPUs that the Linux kernel supports don't do writes until they
631 are certain (1) that the write will actually happen, (2) of the location of
632 the write, and (3) of the value to be written.
633 But please carefully read the "CONTROL DEPENDENCIES" section and the
634 Documentation/RCU/rcu_dereference.rst file: The compiler can and does break
635 dependencies in a great many highly creative ways.
638 =============== ===============
639 { A == 1, B == 2, C = 3, P == &A, Q == &C }
643 Q = READ_ONCE_OLD(P);
646 Therefore, no address-dependency barrier is required to order the read into
647 Q with the store into *Q. In other words, this outcome is prohibited,
648 even without an implicit address-dependency barrier of modern READ_ONCE():
650 (Q == &B) && (B == 4)
652 Please note that this pattern should be rare. After all, the whole point
653 of dependency ordering is to -prevent- writes to the data structure, along
654 with the expensive cache misses associated with those writes. This pattern
655 can be used to record rare error conditions and the like, and the CPUs'
656 naturally occurring ordering prevents such records from being lost.
659 Note well that the ordering provided by an address dependency is local to
660 the CPU containing it. See the section on "Multicopy atomicity" for
664 The address-dependency barrier is very important to the RCU system,
665 for example. See rcu_assign_pointer() and rcu_dereference() in
666 include/linux/rcupdate.h. This permits the current target of an RCU'd
667 pointer to be replaced with a new modified target, without the replacement
668 target appearing to be incompletely initialised.
670 See also the subsection on "Cache Coherency" for a more thorough example.
676 Control dependencies can be a bit tricky because current compilers do
677 not understand them. The purpose of this section is to help you prevent
678 the compiler's ignorance from breaking your code.
680 A load-load control dependency requires a full read memory barrier, not
681 simply an (implicit) address-dependency barrier to make it work correctly.
682 Consider the following bit of code:
685 <implicit address-dependency barrier>
687 /* BUG: No address dependency!!! */
691 This will not have the desired effect because there is no actual address
692 dependency, but rather a control dependency that the CPU may short-circuit
693 by attempting to predict the outcome in advance, so that other CPUs see
694 the load from b as having happened before the load from a. In such a case
695 what's actually required is:
703 However, stores are not speculated. This means that ordering -is- provided
704 for load-store control dependencies, as in the following example:
711 Control dependencies pair normally with other types of barriers.
712 That said, please note that neither READ_ONCE() nor WRITE_ONCE()
713 are optional! Without the READ_ONCE(), the compiler might combine the
714 load from 'a' with other loads from 'a'. Without the WRITE_ONCE(),
715 the compiler might combine the store to 'b' with other stores to 'b'.
716 Either can result in highly counterintuitive effects on ordering.
718 Worse yet, if the compiler is able to prove (say) that the value of
719 variable 'a' is always non-zero, it would be well within its rights
720 to optimize the original example by eliminating the "if" statement
724 b = 1; /* BUG: Compiler and CPU can both reorder!!! */
726 So don't leave out the READ_ONCE().
728 It is tempting to try to enforce ordering on identical stores on both
729 branches of the "if" statement as follows:
742 Unfortunately, current compilers will transform this as follows at high
747 WRITE_ONCE(b, 1); /* BUG: No ordering vs. load from a!!! */
749 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
752 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
756 Now there is no conditional between the load from 'a' and the store to
757 'b', which means that the CPU is within its rights to reorder them:
758 The conditional is absolutely required, and must be present in the
759 assembly code even after all compiler optimizations have been applied.
760 Therefore, if you need ordering in this example, you need explicit
761 memory barriers, for example, smp_store_release():
765 smp_store_release(&b, 1);
768 smp_store_release(&b, 1);
772 In contrast, without explicit memory barriers, two-legged-if control
773 ordering is guaranteed only when the stores differ, for example:
784 The initial READ_ONCE() is still required to prevent the compiler from
785 proving the value of 'a'.
787 In addition, you need to be careful what you do with the local variable 'q',
788 otherwise the compiler might be able to guess the value and again remove
789 the needed conditional. For example:
800 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
801 equal to zero, in which case the compiler is within its rights to
802 transform the above code into the following:
808 Given this transformation, the CPU is not required to respect the ordering
809 between the load from variable 'a' and the store to variable 'b'. It is
810 tempting to add a barrier(), but this does not help. The conditional
811 is gone, and the barrier won't bring it back. Therefore, if you are
812 relying on this ordering, you should make sure that MAX is greater than
813 one, perhaps as follows:
816 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
825 Please note once again that the stores to 'b' differ. If they were
826 identical, as noted earlier, the compiler could pull this store outside
827 of the 'if' statement.
829 You must also be careful not to rely too much on boolean short-circuit
830 evaluation. Consider this example:
836 Because the first condition cannot fault and the second condition is
837 always true, the compiler can transform this example as following,
838 defeating control dependency:
843 This example underscores the need to ensure that the compiler cannot
844 out-guess your code. More generally, although READ_ONCE() does force
845 the compiler to actually emit code for a given load, it does not force
846 the compiler to use the results.
848 In addition, control dependencies apply only to the then-clause and
849 else-clause of the if-statement in question. In particular, it does
850 not necessarily apply to code following the if-statement:
858 WRITE_ONCE(c, 1); /* BUG: No ordering against the read from 'a'. */
860 It is tempting to argue that there in fact is ordering because the
861 compiler cannot reorder volatile accesses and also cannot reorder
862 the writes to 'b' with the condition. Unfortunately for this line
863 of reasoning, the compiler might compile the two writes to 'b' as
864 conditional-move instructions, as in this fanciful pseudo-assembly
874 A weakly ordered CPU would have no dependency of any sort between the load
875 from 'a' and the store to 'c'. The control dependencies would extend
876 only to the pair of cmov instructions and the store depending on them.
877 In short, control dependencies apply only to the stores in the then-clause
878 and else-clause of the if-statement in question (including functions
879 invoked by those two clauses), not to code following that if-statement.
882 Note well that the ordering provided by a control dependency is local
883 to the CPU containing it. See the section on "Multicopy atomicity"
884 for more information.
889 (*) Control dependencies can order prior loads against later stores.
890 However, they do -not- guarantee any other sort of ordering:
891 Not prior loads against later loads, nor prior stores against
892 later anything. If you need these other forms of ordering,
893 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
894 later loads, smp_mb().
896 (*) If both legs of the "if" statement begin with identical stores to
897 the same variable, then those stores must be ordered, either by
898 preceding both of them with smp_mb() or by using smp_store_release()
899 to carry out the stores. Please note that it is -not- sufficient
900 to use barrier() at beginning of each leg of the "if" statement
901 because, as shown by the example above, optimizing compilers can
902 destroy the control dependency while respecting the letter of the
905 (*) Control dependencies require at least one run-time conditional
906 between the prior load and the subsequent store, and this
907 conditional must involve the prior load. If the compiler is able
908 to optimize the conditional away, it will have also optimized
909 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE()
910 can help to preserve the needed conditional.
912 (*) Control dependencies require that the compiler avoid reordering the
913 dependency into nonexistence. Careful use of READ_ONCE() or
914 atomic{,64}_read() can help to preserve your control dependency.
915 Please see the COMPILER BARRIER section for more information.
917 (*) Control dependencies apply only to the then-clause and else-clause
918 of the if-statement containing the control dependency, including
919 any functions that these two clauses call. Control dependencies
920 do -not- apply to code following the if-statement containing the
923 (*) Control dependencies pair normally with other types of barriers.
925 (*) Control dependencies do -not- provide multicopy atomicity. If you
926 need all the CPUs to see a given store at the same time, use smp_mb().
928 (*) Compilers do not understand control dependencies. It is therefore
929 your job to ensure that they do not break your code.
935 When dealing with CPU-CPU interactions, certain types of memory barrier should
936 always be paired. A lack of appropriate pairing is almost certainly an error.
938 General barriers pair with each other, though they also pair with most
939 other types of barriers, albeit without multicopy atomicity. An acquire
940 barrier pairs with a release barrier, but both may also pair with other
941 barriers, including of course general barriers. A write barrier pairs
942 with an address-dependency barrier, a control dependency, an acquire barrier,
943 a release barrier, a read barrier, or a general barrier. Similarly a
944 read barrier, control dependency, or an address-dependency barrier pairs
945 with a write barrier, an acquire barrier, a release barrier, or a
949 =============== ===============
952 WRITE_ONCE(b, 2); x = READ_ONCE(b);
959 =============== ===============================
962 WRITE_ONCE(b, &a); x = READ_ONCE(b);
963 <implicit address-dependency barrier>
969 =============== ===============================
972 WRITE_ONCE(x, 1); if (r2 = READ_ONCE(x)) {
973 <implicit control dependency>
977 assert(r1 == 0 || r2 == 0);
979 Basically, the read barrier always has to be there, even though it can be of
982 [!] Note that the stores before the write barrier would normally be expected to
983 match the loads after the read barrier or the address-dependency barrier, and
987 =================== ===================
988 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c);
989 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d);
990 <write barrier> \ <read barrier>
991 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a);
992 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b);
995 EXAMPLES OF MEMORY BARRIER SEQUENCES
996 ------------------------------------
998 Firstly, write barriers act as partial orderings on store operations.
999 Consider the following sequence of events:
1002 =======================
1010 This sequence of events is committed to the memory coherence system in an order
1011 that the rest of the system might perceive as the unordered set of { STORE A,
1012 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
1017 | |------>| C=3 | } /\
1018 | | : +------+ }----- \ -----> Events perceptible to
1019 | | : | A=1 | } \/ the rest of the system
1021 | CPU 1 | : | B=2 | }
1023 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
1024 | | +------+ } requires all stores prior to the
1025 | | : | E=5 | } barrier to be committed before
1026 | | : +------+ } further stores may take place
1031 | Sequence in which stores are committed to the
1032 | memory system by CPU 1
1036 Secondly, address-dependency barriers act as partial orderings on address-
1037 dependent loads. Consider the following sequence of events:
1040 ======================= =======================
1041 { B = 7; X = 9; Y = 8; C = &Y }
1046 STORE D = 4 LOAD C (gets &B)
1049 Without intervention, CPU 2 may perceive the events on CPU 1 in some
1050 effectively random order, despite the write barrier issued by CPU 1:
1053 | | +------+ +-------+ | Sequence of update
1054 | |------>| B=2 |----- --->| Y->8 | | of perception on
1055 | | : +------+ \ +-------+ | CPU 2
1056 | CPU 1 | : | A=1 | \ --->| C->&Y | V
1057 | | +------+ | +-------+
1058 | | wwwwwwwwwwwwwwww | : :
1060 | | : | C=&B |--- | : : +-------+
1061 | | : +------+ \ | +-------+ | |
1062 | |------>| D=4 | ----------->| C->&B |------>| |
1063 | | +------+ | +-------+ | |
1064 +-------+ : : | : : | |
1068 Apparently incorrect ---> | | B->7 |------>| |
1069 perception of B (!) | +-------+ | |
1072 The load of X holds ---> \ | X->9 |------>| |
1073 up the maintenance \ +-------+ | |
1074 of coherence of B ----->| B->2 | +-------+
1079 In the above example, CPU 2 perceives that B is 7, despite the load of *C
1080 (which would be B) coming after the LOAD of C.
1082 If, however, an address-dependency barrier were to be placed between the load
1083 of C and the load of *C (ie: B) on CPU 2:
1086 ======================= =======================
1087 { B = 7; X = 9; Y = 8; C = &Y }
1092 STORE D = 4 LOAD C (gets &B)
1093 <address-dependency barrier>
1096 then the following will occur:
1099 | | +------+ +-------+
1100 | |------>| B=2 |----- --->| Y->8 |
1101 | | : +------+ \ +-------+
1102 | CPU 1 | : | A=1 | \ --->| C->&Y |
1103 | | +------+ | +-------+
1104 | | wwwwwwwwwwwwwwww | : :
1106 | | : | C=&B |--- | : : +-------+
1107 | | : +------+ \ | +-------+ | |
1108 | |------>| D=4 | ----------->| C->&B |------>| |
1109 | | +------+ | +-------+ | |
1110 +-------+ : : | : : | |
1114 | | X->9 |------>| |
1116 Makes sure all effects ---> \ aaaaaaaaaaaaaaaaa | |
1117 prior to the store of C \ +-------+ | |
1118 are perceptible to ----->| B->2 |------>| |
1119 subsequent loads +-------+ | |
1123 And thirdly, a read barrier acts as a partial order on loads. Consider the
1124 following sequence of events:
1127 ======================= =======================
1135 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1136 some effectively random order, despite the write barrier issued by CPU 1:
1139 | | +------+ +-------+
1140 | |------>| A=1 |------ --->| A->0 |
1141 | | +------+ \ +-------+
1142 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1143 | | +------+ | +-------+
1144 | |------>| B=2 |--- | : :
1145 | | +------+ \ | : : +-------+
1146 +-------+ : : \ | +-------+ | |
1147 ---------->| B->2 |------>| |
1148 | +-------+ | CPU 2 |
1149 | | A->0 |------>| |
1159 If, however, a read barrier were to be placed between the load of B and the
1163 ======================= =======================
1172 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1176 | | +------+ +-------+
1177 | |------>| A=1 |------ --->| A->0 |
1178 | | +------+ \ +-------+
1179 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1180 | | +------+ | +-------+
1181 | |------>| B=2 |--- | : :
1182 | | +------+ \ | : : +-------+
1183 +-------+ : : \ | +-------+ | |
1184 ---------->| B->2 |------>| |
1185 | +-------+ | CPU 2 |
1188 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1189 barrier causes all effects \ +-------+ | |
1190 prior to the storage of B ---->| A->1 |------>| |
1191 to be perceptible to CPU 2 +-------+ | |
1195 To illustrate this more completely, consider what could happen if the code
1196 contained a load of A either side of the read barrier:
1199 ======================= =======================
1205 LOAD A [first load of A]
1207 LOAD A [second load of A]
1209 Even though the two loads of A both occur after the load of B, they may both
1210 come up with different values:
1213 | | +------+ +-------+
1214 | |------>| A=1 |------ --->| A->0 |
1215 | | +------+ \ +-------+
1216 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1217 | | +------+ | +-------+
1218 | |------>| B=2 |--- | : :
1219 | | +------+ \ | : : +-------+
1220 +-------+ : : \ | +-------+ | |
1221 ---------->| B->2 |------>| |
1222 | +-------+ | CPU 2 |
1226 | | A->0 |------>| 1st |
1228 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1229 barrier causes all effects \ +-------+ | |
1230 prior to the storage of B ---->| A->1 |------>| 2nd |
1231 to be perceptible to CPU 2 +-------+ | |
1235 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1236 before the read barrier completes anyway:
1239 | | +------+ +-------+
1240 | |------>| A=1 |------ --->| A->0 |
1241 | | +------+ \ +-------+
1242 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1243 | | +------+ | +-------+
1244 | |------>| B=2 |--- | : :
1245 | | +------+ \ | : : +-------+
1246 +-------+ : : \ | +-------+ | |
1247 ---------->| B->2 |------>| |
1248 | +-------+ | CPU 2 |
1252 ---->| A->1 |------>| 1st |
1254 rrrrrrrrrrrrrrrrr | |
1256 | A->1 |------>| 2nd |
1261 The guarantee is that the second load will always come up with A == 1 if the
1262 load of B came up with B == 2. No such guarantee exists for the first load of
1263 A; that may come up with either A == 0 or A == 1.
1266 READ MEMORY BARRIERS VS LOAD SPECULATION
1267 ----------------------------------------
1269 Many CPUs speculate with loads: that is they see that they will need to load an
1270 item from memory, and they find a time where they're not using the bus for any
1271 other loads, and so do the load in advance - even though they haven't actually
1272 got to that point in the instruction execution flow yet. This permits the
1273 actual load instruction to potentially complete immediately because the CPU
1274 already has the value to hand.
1276 It may turn out that the CPU didn't actually need the value - perhaps because a
1277 branch circumvented the load - in which case it can discard the value or just
1278 cache it for later use.
1283 ======================= =======================
1285 DIVIDE } Divide instructions generally
1286 DIVIDE } take a long time to perform
1289 Which might appear as this:
1293 --->| B->2 |------>| |
1297 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1298 division speculates on the +-------+ ~ | |
1302 Once the divisions are complete --> : : ~-->| |
1303 the CPU can then perform the : : | |
1304 LOAD with immediate effect : : +-------+
1307 Placing a read barrier or an address-dependency barrier just before the second
1311 ======================= =======================
1318 will force any value speculatively obtained to be reconsidered to an extent
1319 dependent on the type of barrier used. If there was no change made to the
1320 speculated memory location, then the speculated value will just be used:
1324 --->| B->2 |------>| |
1328 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1329 division speculates on the +-------+ ~ | |
1334 rrrrrrrrrrrrrrrr~ | |
1341 but if there was an update or an invalidation from another CPU pending, then
1342 the speculation will be cancelled and the value reloaded:
1346 --->| B->2 |------>| |
1350 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1351 division speculates on the +-------+ ~ | |
1356 rrrrrrrrrrrrrrrrr | |
1358 The speculation is discarded ---> --->| A->1 |------>| |
1359 and an updated value is +-------+ | |
1360 retrieved : : +-------+
1364 --------------------
1366 Multicopy atomicity is a deeply intuitive notion about ordering that is
1367 not always provided by real computer systems, namely that a given store
1368 becomes visible at the same time to all CPUs, or, alternatively, that all
1369 CPUs agree on the order in which all stores become visible. However,
1370 support of full multicopy atomicity would rule out valuable hardware
1371 optimizations, so a weaker form called ``other multicopy atomicity''
1372 instead guarantees only that a given store becomes visible at the same
1373 time to all -other- CPUs. The remainder of this document discusses this
1374 weaker form, but for brevity will call it simply ``multicopy atomicity''.
1376 The following example demonstrates multicopy atomicity:
1379 ======================= ======================= =======================
1381 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1)
1382 <general barrier> <read barrier>
1385 Suppose that CPU 2's load from X returns 1, which it then stores to Y,
1386 and CPU 3's load from Y returns 1. This indicates that CPU 1's store
1387 to X precedes CPU 2's load from X and that CPU 2's store to Y precedes
1388 CPU 3's load from Y. In addition, the memory barriers guarantee that
1389 CPU 2 executes its load before its store, and CPU 3 loads from Y before
1390 it loads from X. The question is then "Can CPU 3's load from X return 0?"
1392 Because CPU 3's load from X in some sense comes after CPU 2's load, it
1393 is natural to expect that CPU 3's load from X must therefore return 1.
1394 This expectation follows from multicopy atomicity: if a load executing
1395 on CPU B follows a load from the same variable executing on CPU A (and
1396 CPU A did not originally store the value which it read), then on
1397 multicopy-atomic systems, CPU B's load must return either the same value
1398 that CPU A's load did or some later value. However, the Linux kernel
1399 does not require systems to be multicopy atomic.
1401 The use of a general memory barrier in the example above compensates
1402 for any lack of multicopy atomicity. In the example, if CPU 2's load
1403 from X returns 1 and CPU 3's load from Y returns 1, then CPU 3's load
1404 from X must indeed also return 1.
1406 However, dependencies, read barriers, and write barriers are not always
1407 able to compensate for non-multicopy atomicity. For example, suppose
1408 that CPU 2's general barrier is removed from the above example, leaving
1409 only the data dependency shown below:
1412 ======================= ======================= =======================
1414 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1)
1415 <data dependency> <read barrier>
1416 STORE Y=r1 LOAD X (reads 0)
1418 This substitution allows non-multicopy atomicity to run rampant: in
1419 this example, it is perfectly legal for CPU 2's load from X to return 1,
1420 CPU 3's load from Y to return 1, and its load from X to return 0.
1422 The key point is that although CPU 2's data dependency orders its load
1423 and store, it does not guarantee to order CPU 1's store. Thus, if this
1424 example runs on a non-multicopy-atomic system where CPUs 1 and 2 share a
1425 store buffer or a level of cache, CPU 2 might have early access to CPU 1's
1426 writes. General barriers are therefore required to ensure that all CPUs
1427 agree on the combined order of multiple accesses.
1429 General barriers can compensate not only for non-multicopy atomicity,
1430 but can also generate additional ordering that can ensure that -all-
1431 CPUs will perceive the same order of -all- operations. In contrast, a
1432 chain of release-acquire pairs do not provide this additional ordering,
1433 which means that only those CPUs on the chain are guaranteed to agree
1434 on the combined order of the accesses. For example, switching to C code
1435 in deference to the ghost of Herman Hollerith:
1441 r0 = smp_load_acquire(&x);
1443 smp_store_release(&y, 1);
1448 r1 = smp_load_acquire(&y);
1451 smp_store_release(&z, 1);
1456 r2 = smp_load_acquire(&z);
1457 smp_store_release(&x, 1);
1467 Because cpu0(), cpu1(), and cpu2() participate in a chain of
1468 smp_store_release()/smp_load_acquire() pairs, the following outcome
1471 r0 == 1 && r1 == 1 && r2 == 1
1473 Furthermore, because of the release-acquire relationship between cpu0()
1474 and cpu1(), cpu1() must see cpu0()'s writes, so that the following
1475 outcome is prohibited:
1479 However, the ordering provided by a release-acquire chain is local
1480 to the CPUs participating in that chain and does not apply to cpu3(),
1481 at least aside from stores. Therefore, the following outcome is possible:
1483 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0
1485 As an aside, the following outcome is also possible:
1487 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1
1489 Although cpu0(), cpu1(), and cpu2() will see their respective reads and
1490 writes in order, CPUs not involved in the release-acquire chain might
1491 well disagree on the order. This disagreement stems from the fact that
1492 the weak memory-barrier instructions used to implement smp_load_acquire()
1493 and smp_store_release() are not required to order prior stores against
1494 subsequent loads in all cases. This means that cpu3() can see cpu0()'s
1495 store to u as happening -after- cpu1()'s load from v, even though
1496 both cpu0() and cpu1() agree that these two operations occurred in the
1499 However, please keep in mind that smp_load_acquire() is not magic.
1500 In particular, it simply reads from its argument with ordering. It does
1501 -not- ensure that any particular value will be read. Therefore, the
1502 following outcome is possible:
1504 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0
1506 Note that this outcome can happen even on a mythical sequentially
1507 consistent system where nothing is ever reordered.
1509 To reiterate, if your code requires full ordering of all operations,
1510 use general barriers throughout.
1513 ========================
1514 EXPLICIT KERNEL BARRIERS
1515 ========================
1517 The Linux kernel has a variety of different barriers that act at different
1520 (*) Compiler barrier.
1522 (*) CPU memory barriers.
1528 The Linux kernel has an explicit compiler barrier function that prevents the
1529 compiler from moving the memory accesses either side of it to the other side:
1533 This is a general barrier -- there are no read-read or write-write
1534 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be
1535 thought of as weak forms of barrier() that affect only the specific
1536 accesses flagged by the READ_ONCE() or WRITE_ONCE().
1538 The barrier() function has the following effects:
1540 (*) Prevents the compiler from reordering accesses following the
1541 barrier() to precede any accesses preceding the barrier().
1542 One example use for this property is to ease communication between
1543 interrupt-handler code and the code that was interrupted.
1545 (*) Within a loop, forces the compiler to load the variables used
1546 in that loop's conditional on each pass through that loop.
1548 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of
1549 optimizations that, while perfectly safe in single-threaded code, can
1550 be fatal in concurrent code. Here are some examples of these sorts
1553 (*) The compiler is within its rights to reorder loads and stores
1554 to the same variable, and in some cases, the CPU is within its
1555 rights to reorder loads to the same variable. This means that
1561 Might result in an older value of x stored in a[1] than in a[0].
1562 Prevent both the compiler and the CPU from doing this as follows:
1564 a[0] = READ_ONCE(x);
1565 a[1] = READ_ONCE(x);
1567 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for
1568 accesses from multiple CPUs to a single variable.
1570 (*) The compiler is within its rights to merge successive loads from
1571 the same variable. Such merging can cause the compiler to "optimize"
1575 do_something_with(tmp);
1577 into the following code, which, although in some sense legitimate
1578 for single-threaded code, is almost certainly not what the developer
1583 do_something_with(tmp);
1585 Use READ_ONCE() to prevent the compiler from doing this to you:
1587 while (tmp = READ_ONCE(a))
1588 do_something_with(tmp);
1590 (*) The compiler is within its rights to reload a variable, for example,
1591 in cases where high register pressure prevents the compiler from
1592 keeping all data of interest in registers. The compiler might
1593 therefore optimize the variable 'tmp' out of our previous example:
1596 do_something_with(tmp);
1598 This could result in the following code, which is perfectly safe in
1599 single-threaded code, but can be fatal in concurrent code:
1602 do_something_with(a);
1604 For example, the optimized version of this code could result in
1605 passing a zero to do_something_with() in the case where the variable
1606 a was modified by some other CPU between the "while" statement and
1607 the call to do_something_with().
1609 Again, use READ_ONCE() to prevent the compiler from doing this:
1611 while (tmp = READ_ONCE(a))
1612 do_something_with(tmp);
1614 Note that if the compiler runs short of registers, it might save
1615 tmp onto the stack. The overhead of this saving and later restoring
1616 is why compilers reload variables. Doing so is perfectly safe for
1617 single-threaded code, so you need to tell the compiler about cases
1618 where it is not safe.
1620 (*) The compiler is within its rights to omit a load entirely if it knows
1621 what the value will be. For example, if the compiler can prove that
1622 the value of variable 'a' is always zero, it can optimize this code:
1625 do_something_with(tmp);
1631 This transformation is a win for single-threaded code because it
1632 gets rid of a load and a branch. The problem is that the compiler
1633 will carry out its proof assuming that the current CPU is the only
1634 one updating variable 'a'. If variable 'a' is shared, then the
1635 compiler's proof will be erroneous. Use READ_ONCE() to tell the
1636 compiler that it doesn't know as much as it thinks it does:
1638 while (tmp = READ_ONCE(a))
1639 do_something_with(tmp);
1641 But please note that the compiler is also closely watching what you
1642 do with the value after the READ_ONCE(). For example, suppose you
1643 do the following and MAX is a preprocessor macro with the value 1:
1645 while ((tmp = READ_ONCE(a)) % MAX)
1646 do_something_with(tmp);
1648 Then the compiler knows that the result of the "%" operator applied
1649 to MAX will always be zero, again allowing the compiler to optimize
1650 the code into near-nonexistence. (It will still load from the
1653 (*) Similarly, the compiler is within its rights to omit a store entirely
1654 if it knows that the variable already has the value being stored.
1655 Again, the compiler assumes that the current CPU is the only one
1656 storing into the variable, which can cause the compiler to do the
1657 wrong thing for shared variables. For example, suppose you have
1661 ... Code that does not store to variable a ...
1664 The compiler sees that the value of variable 'a' is already zero, so
1665 it might well omit the second store. This would come as a fatal
1666 surprise if some other CPU might have stored to variable 'a' in the
1669 Use WRITE_ONCE() to prevent the compiler from making this sort of
1673 ... Code that does not store to variable a ...
1676 (*) The compiler is within its rights to reorder memory accesses unless
1677 you tell it not to. For example, consider the following interaction
1678 between process-level code and an interrupt handler:
1680 void process_level(void)
1682 msg = get_message();
1686 void interrupt_handler(void)
1689 process_message(msg);
1692 There is nothing to prevent the compiler from transforming
1693 process_level() to the following, in fact, this might well be a
1694 win for single-threaded code:
1696 void process_level(void)
1699 msg = get_message();
1702 If the interrupt occurs between these two statement, then
1703 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE()
1704 to prevent this as follows:
1706 void process_level(void)
1708 WRITE_ONCE(msg, get_message());
1709 WRITE_ONCE(flag, true);
1712 void interrupt_handler(void)
1714 if (READ_ONCE(flag))
1715 process_message(READ_ONCE(msg));
1718 Note that the READ_ONCE() and WRITE_ONCE() wrappers in
1719 interrupt_handler() are needed if this interrupt handler can itself
1720 be interrupted by something that also accesses 'flag' and 'msg',
1721 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE()
1722 and WRITE_ONCE() are not needed in interrupt_handler() other than
1723 for documentation purposes. (Note also that nested interrupts
1724 do not typically occur in modern Linux kernels, in fact, if an
1725 interrupt handler returns with interrupts enabled, you will get a
1728 You should assume that the compiler can move READ_ONCE() and
1729 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(),
1730 barrier(), or similar primitives.
1732 This effect could also be achieved using barrier(), but READ_ONCE()
1733 and WRITE_ONCE() are more selective: With READ_ONCE() and
1734 WRITE_ONCE(), the compiler need only forget the contents of the
1735 indicated memory locations, while with barrier() the compiler must
1736 discard the value of all memory locations that it has currently
1737 cached in any machine registers. Of course, the compiler must also
1738 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
1739 though the CPU of course need not do so.
1741 (*) The compiler is within its rights to invent stores to a variable,
1742 as in the following example:
1749 The compiler might save a branch by optimizing this as follows:
1755 In single-threaded code, this is not only safe, but also saves
1756 a branch. Unfortunately, in concurrent code, this optimization
1757 could cause some other CPU to see a spurious value of 42 -- even
1758 if variable 'a' was never zero -- when loading variable 'b'.
1759 Use WRITE_ONCE() to prevent this as follows:
1766 The compiler can also invent loads. These are usually less
1767 damaging, but they can result in cache-line bouncing and thus in
1768 poor performance and scalability. Use READ_ONCE() to prevent
1771 (*) For aligned memory locations whose size allows them to be accessed
1772 with a single memory-reference instruction, prevents "load tearing"
1773 and "store tearing," in which a single large access is replaced by
1774 multiple smaller accesses. For example, given an architecture having
1775 16-bit store instructions with 7-bit immediate fields, the compiler
1776 might be tempted to use two 16-bit store-immediate instructions to
1777 implement the following 32-bit store:
1781 Please note that GCC really does use this sort of optimization,
1782 which is not surprising given that it would likely take more
1783 than two instructions to build the constant and then store it.
1784 This optimization can therefore be a win in single-threaded code.
1785 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1786 this optimization in a volatile store. In the absence of such bugs,
1787 use of WRITE_ONCE() prevents store tearing in the following example:
1789 WRITE_ONCE(p, 0x00010002);
1791 Use of packed structures can also result in load and store tearing,
1794 struct __attribute__((__packed__)) foo {
1799 struct foo foo1, foo2;
1806 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no
1807 volatile markings, the compiler would be well within its rights to
1808 implement these three assignment statements as a pair of 32-bit
1809 loads followed by a pair of 32-bit stores. This would result in
1810 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE()
1811 and WRITE_ONCE() again prevent tearing in this example:
1814 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b));
1817 All that aside, it is never necessary to use READ_ONCE() and
1818 WRITE_ONCE() on a variable that has been marked volatile. For example,
1819 because 'jiffies' is marked volatile, it is never necessary to
1820 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and
1821 WRITE_ONCE() are implemented as volatile casts, which has no effect when
1822 its argument is already marked volatile.
1824 Please note that these compiler barriers have no direct effect on the CPU,
1825 which may then reorder things however it wishes.
1831 The Linux kernel has seven basic CPU memory barriers:
1833 TYPE MANDATORY SMP CONDITIONAL
1834 ======================= =============== ===============
1835 GENERAL mb() smp_mb()
1836 WRITE wmb() smp_wmb()
1837 READ rmb() smp_rmb()
1838 ADDRESS DEPENDENCY READ_ONCE()
1841 All memory barriers except the address-dependency barriers imply a compiler
1842 barrier. Address dependencies do not impose any additional compiler ordering.
1844 Aside: In the case of address dependencies, the compiler would be expected
1845 to issue the loads in the correct order (eg. `a[b]` would have to load
1846 the value of b before loading a[b]), however there is no guarantee in
1847 the C specification that the compiler may not speculate the value of b
1848 (eg. is equal to 1) and load a[b] before b (eg. tmp = a[1]; if (b != 1)
1849 tmp = a[b]; ). There is also the problem of a compiler reloading b after
1850 having loaded a[b], thus having a newer copy of b than a[b]. A consensus
1851 has not yet been reached about these problems, however the READ_ONCE()
1852 macro is a good place to start looking.
1854 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1855 systems because it is assumed that a CPU will appear to be self-consistent,
1856 and will order overlapping accesses correctly with respect to itself.
1857 However, see the subsection on "Virtual Machine Guests" below.
1859 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1860 references to shared memory on SMP systems, though the use of locking instead
1863 Mandatory barriers should not be used to control SMP effects, since mandatory
1864 barriers impose unnecessary overhead on both SMP and UP systems. They may,
1865 however, be used to control MMIO effects on accesses through relaxed memory I/O
1866 windows. These barriers are required even on non-SMP systems as they affect
1867 the order in which memory operations appear to a device by prohibiting both the
1868 compiler and the CPU from reordering them.
1871 There are some more advanced barrier functions:
1873 (*) smp_store_mb(var, value)
1875 This assigns the value to the variable and then inserts a full memory
1876 barrier after it. It isn't guaranteed to insert anything more than a
1877 compiler barrier in a UP compilation.
1880 (*) smp_mb__before_atomic();
1881 (*) smp_mb__after_atomic();
1883 These are for use with atomic RMW functions that do not imply memory
1884 barriers, but where the code needs a memory barrier. Examples for atomic
1885 RMW functions that do not imply a memory barrier are e.g. add,
1886 subtract, (failed) conditional operations, _relaxed functions,
1887 but not atomic_read or atomic_set. A common example where a memory
1888 barrier may be required is when atomic ops are used for reference
1891 These are also used for atomic RMW bitop functions that do not imply a
1892 memory barrier (such as set_bit and clear_bit).
1894 As an example, consider a piece of code that marks an object as being dead
1895 and then decrements the object's reference count:
1898 smp_mb__before_atomic();
1899 atomic_dec(&obj->ref_count);
1901 This makes sure that the death mark on the object is perceived to be set
1902 *before* the reference counter is decremented.
1904 See Documentation/atomic_{t,bitops}.txt for more information.
1911 These are for use with consistent memory to guarantee the ordering
1912 of writes or reads of shared memory accessible to both the CPU and a
1913 DMA capable device. See Documentation/core-api/dma-api.rst file for more
1914 information about consistent memory.
1916 For example, consider a device driver that shares memory with a device
1917 and uses a descriptor status value to indicate if the descriptor belongs
1918 to the device or the CPU, and a doorbell to notify it when new
1919 descriptors are available:
1921 if (desc->status != DEVICE_OWN) {
1922 /* do not read data until we own descriptor */
1925 /* read/modify data */
1926 read_data = desc->data;
1927 desc->data = write_data;
1929 /* flush modifications before status update */
1932 /* assign ownership */
1933 desc->status = DEVICE_OWN;
1935 /* Make descriptor status visible to the device followed by
1936 * notify device of new descriptor
1938 writel(DESC_NOTIFY, doorbell);
1941 The dma_rmb() allows us to guarantee that the device has released ownership
1942 before we read the data from the descriptor, and the dma_wmb() allows
1943 us to guarantee the data is written to the descriptor before the device
1944 can see it now has ownership. The dma_mb() implies both a dma_rmb() and
1947 Note that the dma_*() barriers do not provide any ordering guarantees for
1948 accesses to MMIO regions. See the later "KERNEL I/O BARRIER EFFECTS"
1949 subsection for more information about I/O accessors and MMIO ordering.
1953 This is for use with persistent memory to ensure that stores for which
1954 modifications are written to persistent storage reached a platform
1957 For example, after a non-temporal write to pmem region, we use pmem_wmb()
1958 to ensure that stores have reached a platform durability domain. This ensures
1959 that stores have updated persistent storage before any data access or
1960 data transfer caused by subsequent instructions is initiated. This is
1961 in addition to the ordering done by wmb().
1963 For load from persistent memory, existing read memory barriers are sufficient
1964 to ensure read ordering.
1968 For memory accesses with write-combining attributes (e.g. those returned
1969 by ioremap_wc()), the CPU may wait for prior accesses to be merged with
1970 subsequent ones. io_stop_wc() can be used to prevent the merging of
1971 write-combining memory accesses before this macro with those after it when
1972 such wait has performance implications.
1974 ===============================
1975 IMPLICIT KERNEL MEMORY BARRIERS
1976 ===============================
1978 Some of the other functions in the linux kernel imply memory barriers, amongst
1979 which are locking and scheduling functions.
1981 This specification is a _minimum_ guarantee; any particular architecture may
1982 provide more substantial guarantees, but these may not be relied upon outside
1983 of arch specific code.
1986 LOCK ACQUISITION FUNCTIONS
1987 --------------------------
1989 The Linux kernel has a number of locking constructs:
1997 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1998 for each construct. These operations all imply certain barriers:
2000 (1) ACQUIRE operation implication:
2002 Memory operations issued after the ACQUIRE will be completed after the
2003 ACQUIRE operation has completed.
2005 Memory operations issued before the ACQUIRE may be completed after
2006 the ACQUIRE operation has completed.
2008 (2) RELEASE operation implication:
2010 Memory operations issued before the RELEASE will be completed before the
2011 RELEASE operation has completed.
2013 Memory operations issued after the RELEASE may be completed before the
2014 RELEASE operation has completed.
2016 (3) ACQUIRE vs ACQUIRE implication:
2018 All ACQUIRE operations issued before another ACQUIRE operation will be
2019 completed before that ACQUIRE operation.
2021 (4) ACQUIRE vs RELEASE implication:
2023 All ACQUIRE operations issued before a RELEASE operation will be
2024 completed before the RELEASE operation.
2026 (5) Failed conditional ACQUIRE implication:
2028 Certain locking variants of the ACQUIRE operation may fail, either due to
2029 being unable to get the lock immediately, or due to receiving an unblocked
2030 signal while asleep waiting for the lock to become available. Failed
2031 locks do not imply any sort of barrier.
2033 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
2034 one-way barriers is that the effects of instructions outside of a critical
2035 section may seep into the inside of the critical section.
2037 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
2038 because it is possible for an access preceding the ACQUIRE to happen after the
2039 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
2040 the two accesses can themselves then cross:
2049 ACQUIRE M, STORE *B, STORE *A, RELEASE M
2051 When the ACQUIRE and RELEASE are a lock acquisition and release,
2052 respectively, this same reordering can occur if the lock's ACQUIRE and
2053 RELEASE are to the same lock variable, but only from the perspective of
2054 another CPU not holding that lock. In short, a ACQUIRE followed by an
2055 RELEASE may -not- be assumed to be a full memory barrier.
2057 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does
2058 not imply a full memory barrier. Therefore, the CPU's execution of the
2059 critical sections corresponding to the RELEASE and the ACQUIRE can cross,
2069 ACQUIRE N, STORE *B, STORE *A, RELEASE M
2071 It might appear that this reordering could introduce a deadlock.
2072 However, this cannot happen because if such a deadlock threatened,
2073 the RELEASE would simply complete, thereby avoiding the deadlock.
2077 One key point is that we are only talking about the CPU doing
2078 the reordering, not the compiler. If the compiler (or, for
2079 that matter, the developer) switched the operations, deadlock
2082 But suppose the CPU reordered the operations. In this case,
2083 the unlock precedes the lock in the assembly code. The CPU
2084 simply elected to try executing the later lock operation first.
2085 If there is a deadlock, this lock operation will simply spin (or
2086 try to sleep, but more on that later). The CPU will eventually
2087 execute the unlock operation (which preceded the lock operation
2088 in the assembly code), which will unravel the potential deadlock,
2089 allowing the lock operation to succeed.
2091 But what if the lock is a sleeplock? In that case, the code will
2092 try to enter the scheduler, where it will eventually encounter
2093 a memory barrier, which will force the earlier unlock operation
2094 to complete, again unraveling the deadlock. There might be
2095 a sleep-unlock race, but the locking primitive needs to resolve
2096 such races properly in any case.
2098 Locks and semaphores may not provide any guarantee of ordering on UP compiled
2099 systems, and so cannot be counted on in such a situation to actually achieve
2100 anything at all - especially with respect to I/O accesses - unless combined
2101 with interrupt disabling operations.
2103 See also the section on "Inter-CPU acquiring barrier effects".
2106 As an example, consider the following:
2117 The following sequence of events is acceptable:
2119 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
2121 [+] Note that {*F,*A} indicates a combined access.
2123 But none of the following are:
2125 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
2126 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
2127 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
2128 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
2132 INTERRUPT DISABLING FUNCTIONS
2133 -----------------------------
2135 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
2136 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
2137 barriers are required in such a situation, they must be provided from some
2141 SLEEP AND WAKE-UP FUNCTIONS
2142 ---------------------------
2144 Sleeping and waking on an event flagged in global data can be viewed as an
2145 interaction between two pieces of data: the task state of the task waiting for
2146 the event and the global data used to indicate the event. To make sure that
2147 these appear to happen in the right order, the primitives to begin the process
2148 of going to sleep, and the primitives to initiate a wake up imply certain
2151 Firstly, the sleeper normally follows something like this sequence of events:
2154 set_current_state(TASK_UNINTERRUPTIBLE);
2155 if (event_indicated)
2160 A general memory barrier is interpolated automatically by set_current_state()
2161 after it has altered the task state:
2164 ===============================
2165 set_current_state();
2167 STORE current->state
2169 LOAD event_indicated
2171 set_current_state() may be wrapped by:
2174 prepare_to_wait_exclusive();
2176 which therefore also imply a general memory barrier after setting the state.
2177 The whole sequence above is available in various canned forms, all of which
2178 interpolate the memory barrier in the right place:
2181 wait_event_interruptible();
2182 wait_event_interruptible_exclusive();
2183 wait_event_interruptible_timeout();
2184 wait_event_killable();
2185 wait_event_timeout();
2190 Secondly, code that performs a wake up normally follows something like this:
2192 event_indicated = 1;
2193 wake_up(&event_wait_queue);
2197 event_indicated = 1;
2198 wake_up_process(event_daemon);
2200 A general memory barrier is executed by wake_up() if it wakes something up.
2201 If it doesn't wake anything up then a memory barrier may or may not be
2202 executed; you must not rely on it. The barrier occurs before the task state
2203 is accessed, in particular, it sits between the STORE to indicate the event
2204 and the STORE to set TASK_RUNNING:
2206 CPU 1 (Sleeper) CPU 2 (Waker)
2207 =============================== ===============================
2208 set_current_state(); STORE event_indicated
2209 smp_store_mb(); wake_up();
2210 STORE current->state ...
2211 <general barrier> <general barrier>
2212 LOAD event_indicated if ((LOAD task->state) & TASK_NORMAL)
2215 where "task" is the thread being woken up and it equals CPU 1's "current".
2217 To repeat, a general memory barrier is guaranteed to be executed by wake_up()
2218 if something is actually awakened, but otherwise there is no such guarantee.
2219 To see this, consider the following sequence of events, where X and Y are both
2223 =============================== ===============================
2225 smp_mb(); wake_up();
2228 If a wakeup does occur, one (at least) of the two loads must see 1. If, on
2229 the other hand, a wakeup does not occur, both loads might see 0.
2231 wake_up_process() always executes a general memory barrier. The barrier again
2232 occurs before the task state is accessed. In particular, if the wake_up() in
2233 the previous snippet were replaced by a call to wake_up_process() then one of
2234 the two loads would be guaranteed to see 1.
2236 The available waker functions include:
2242 wake_up_interruptible();
2243 wake_up_interruptible_all();
2244 wake_up_interruptible_nr();
2245 wake_up_interruptible_poll();
2246 wake_up_interruptible_sync();
2247 wake_up_interruptible_sync_poll();
2249 wake_up_locked_poll();
2254 In terms of memory ordering, these functions all provide the same guarantees of
2255 a wake_up() (or stronger).
2257 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2258 order multiple stores before the wake-up with respect to loads of those stored
2259 values after the sleeper has called set_current_state(). For instance, if the
2262 set_current_state(TASK_INTERRUPTIBLE);
2263 if (event_indicated)
2265 __set_current_state(TASK_RUNNING);
2266 do_something(my_data);
2271 event_indicated = 1;
2272 wake_up(&event_wait_queue);
2274 there's no guarantee that the change to event_indicated will be perceived by
2275 the sleeper as coming after the change to my_data. In such a circumstance, the
2276 code on both sides must interpolate its own memory barriers between the
2277 separate data accesses. Thus the above sleeper ought to do:
2279 set_current_state(TASK_INTERRUPTIBLE);
2280 if (event_indicated) {
2282 do_something(my_data);
2285 and the waker should do:
2289 event_indicated = 1;
2290 wake_up(&event_wait_queue);
2293 MISCELLANEOUS FUNCTIONS
2294 -----------------------
2296 Other functions that imply barriers:
2298 (*) schedule() and similar imply full memory barriers.
2301 ===================================
2302 INTER-CPU ACQUIRING BARRIER EFFECTS
2303 ===================================
2305 On SMP systems locking primitives give a more substantial form of barrier: one
2306 that does affect memory access ordering on other CPUs, within the context of
2307 conflict on any particular lock.
2310 ACQUIRES VS MEMORY ACCESSES
2311 ---------------------------
2313 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2314 three CPUs; then should the following sequence of events occur:
2317 =============================== ===============================
2318 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e);
2320 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f);
2321 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g);
2323 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h);
2325 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2326 through *H occur in, other than the constraints imposed by the separate locks
2327 on the separate CPUs. It might, for example, see:
2329 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2331 But it won't see any of:
2333 *B, *C or *D preceding ACQUIRE M
2334 *A, *B or *C following RELEASE M
2335 *F, *G or *H preceding ACQUIRE Q
2336 *E, *F or *G following RELEASE Q
2339 =================================
2340 WHERE ARE MEMORY BARRIERS NEEDED?
2341 =================================
2343 Under normal operation, memory operation reordering is generally not going to
2344 be a problem as a single-threaded linear piece of code will still appear to
2345 work correctly, even if it's in an SMP kernel. There are, however, four
2346 circumstances in which reordering definitely _could_ be a problem:
2348 (*) Interprocessor interaction.
2350 (*) Atomic operations.
2352 (*) Accessing devices.
2357 INTERPROCESSOR INTERACTION
2358 --------------------------
2360 When there's a system with more than one processor, more than one CPU in the
2361 system may be working on the same data set at the same time. This can cause
2362 synchronisation problems, and the usual way of dealing with them is to use
2363 locks. Locks, however, are quite expensive, and so it may be preferable to
2364 operate without the use of a lock if at all possible. In such a case
2365 operations that affect both CPUs may have to be carefully ordered to prevent
2368 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2369 queued on the semaphore, by virtue of it having a piece of its stack linked to
2370 the semaphore's list of waiting processes:
2372 struct rw_semaphore {
2375 struct list_head waiters;
2378 struct rwsem_waiter {
2379 struct list_head list;
2380 struct task_struct *task;
2383 To wake up a particular waiter, the up_read() or up_write() functions have to:
2385 (1) read the next pointer from this waiter's record to know as to where the
2386 next waiter record is;
2388 (2) read the pointer to the waiter's task structure;
2390 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2392 (4) call wake_up_process() on the task; and
2394 (5) release the reference held on the waiter's task struct.
2396 In other words, it has to perform this sequence of events:
2398 LOAD waiter->list.next;
2404 and if any of these steps occur out of order, then the whole thing may
2407 Once it has queued itself and dropped the semaphore lock, the waiter does not
2408 get the lock again; it instead just waits for its task pointer to be cleared
2409 before proceeding. Since the record is on the waiter's stack, this means that
2410 if the task pointer is cleared _before_ the next pointer in the list is read,
2411 another CPU might start processing the waiter and might clobber the waiter's
2412 stack before the up*() function has a chance to read the next pointer.
2414 Consider then what might happen to the above sequence of events:
2417 =============================== ===============================
2424 Woken up by other event
2429 foo() clobbers *waiter
2431 LOAD waiter->list.next;
2434 This could be dealt with using the semaphore lock, but then the down_xxx()
2435 function has to needlessly get the spinlock again after being woken up.
2437 The way to deal with this is to insert a general SMP memory barrier:
2439 LOAD waiter->list.next;
2446 In this case, the barrier makes a guarantee that all memory accesses before the
2447 barrier will appear to happen before all the memory accesses after the barrier
2448 with respect to the other CPUs on the system. It does _not_ guarantee that all
2449 the memory accesses before the barrier will be complete by the time the barrier
2450 instruction itself is complete.
2452 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2453 compiler barrier, thus making sure the compiler emits the instructions in the
2454 right order without actually intervening in the CPU. Since there's only one
2455 CPU, that CPU's dependency ordering logic will take care of everything else.
2461 While they are technically interprocessor interaction considerations, atomic
2462 operations are noted specially as some of them imply full memory barriers and
2463 some don't, but they're very heavily relied on as a group throughout the
2466 See Documentation/atomic_t.txt for more information.
2472 Many devices can be memory mapped, and so appear to the CPU as if they're just
2473 a set of memory locations. To control such a device, the driver usually has to
2474 make the right memory accesses in exactly the right order.
2476 However, having a clever CPU or a clever compiler creates a potential problem
2477 in that the carefully sequenced accesses in the driver code won't reach the
2478 device in the requisite order if the CPU or the compiler thinks it is more
2479 efficient to reorder, combine or merge accesses - something that would cause
2480 the device to malfunction.
2482 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2483 routines - such as inb() or writel() - which know how to make such accesses
2484 appropriately sequential. While this, for the most part, renders the explicit
2485 use of memory barriers unnecessary, if the accessor functions are used to refer
2486 to an I/O memory window with relaxed memory access properties, then _mandatory_
2487 memory barriers are required to enforce ordering.
2489 See Documentation/driver-api/device-io.rst for more information.
2495 A driver may be interrupted by its own interrupt service routine, and thus the
2496 two parts of the driver may interfere with each other's attempts to control or
2499 This may be alleviated - at least in part - by disabling local interrupts (a
2500 form of locking), such that the critical operations are all contained within
2501 the interrupt-disabled section in the driver. While the driver's interrupt
2502 routine is executing, the driver's core may not run on the same CPU, and its
2503 interrupt is not permitted to happen again until the current interrupt has been
2504 handled, thus the interrupt handler does not need to lock against that.
2506 However, consider a driver that was talking to an ethernet card that sports an
2507 address register and a data register. If that driver's core talks to the card
2508 under interrupt-disablement and then the driver's interrupt handler is invoked:
2519 The store to the data register might happen after the second store to the
2520 address register if ordering rules are sufficiently relaxed:
2522 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2525 If ordering rules are relaxed, it must be assumed that accesses done inside an
2526 interrupt disabled section may leak outside of it and may interleave with
2527 accesses performed in an interrupt - and vice versa - unless implicit or
2528 explicit barriers are used.
2530 Normally this won't be a problem because the I/O accesses done inside such
2531 sections will include synchronous load operations on strictly ordered I/O
2532 registers that form implicit I/O barriers.
2535 A similar situation may occur between an interrupt routine and two routines
2536 running on separate CPUs that communicate with each other. If such a case is
2537 likely, then interrupt-disabling locks should be used to guarantee ordering.
2540 ==========================
2541 KERNEL I/O BARRIER EFFECTS
2542 ==========================
2544 Interfacing with peripherals via I/O accesses is deeply architecture and device
2545 specific. Therefore, drivers which are inherently non-portable may rely on
2546 specific behaviours of their target systems in order to achieve synchronization
2547 in the most lightweight manner possible. For drivers intending to be portable
2548 between multiple architectures and bus implementations, the kernel offers a
2549 series of accessor functions that provide various degrees of ordering
2552 (*) readX(), writeX():
2554 The readX() and writeX() MMIO accessors take a pointer to the
2555 peripheral being accessed as an __iomem * parameter. For pointers
2556 mapped with the default I/O attributes (e.g. those returned by
2557 ioremap()), the ordering guarantees are as follows:
2559 1. All readX() and writeX() accesses to the same peripheral are ordered
2560 with respect to each other. This ensures that MMIO register accesses
2561 by the same CPU thread to a particular device will arrive in program
2564 2. A writeX() issued by a CPU thread holding a spinlock is ordered
2565 before a writeX() to the same peripheral from another CPU thread
2566 issued after a later acquisition of the same spinlock. This ensures
2567 that MMIO register writes to a particular device issued while holding
2568 a spinlock will arrive in an order consistent with acquisitions of
2571 3. A writeX() by a CPU thread to the peripheral will first wait for the
2572 completion of all prior writes to memory either issued by, or
2573 propagated to, the same thread. This ensures that writes by the CPU
2574 to an outbound DMA buffer allocated by dma_alloc_coherent() will be
2575 visible to a DMA engine when the CPU writes to its MMIO control
2576 register to trigger the transfer.
2578 4. A readX() by a CPU thread from the peripheral will complete before
2579 any subsequent reads from memory by the same thread can begin. This
2580 ensures that reads by the CPU from an incoming DMA buffer allocated
2581 by dma_alloc_coherent() will not see stale data after reading from
2582 the DMA engine's MMIO status register to establish that the DMA
2583 transfer has completed.
2585 5. A readX() by a CPU thread from the peripheral will complete before
2586 any subsequent delay() loop can begin execution on the same thread.
2587 This ensures that two MMIO register writes by the CPU to a peripheral
2588 will arrive at least 1us apart if the first write is immediately read
2589 back with readX() and udelay(1) is called prior to the second
2592 writel(42, DEVICE_REGISTER_0); // Arrives at the device...
2593 readl(DEVICE_REGISTER_0);
2595 writel(42, DEVICE_REGISTER_1); // ...at least 1us before this.
2597 The ordering properties of __iomem pointers obtained with non-default
2598 attributes (e.g. those returned by ioremap_wc()) are specific to the
2599 underlying architecture and therefore the guarantees listed above cannot
2600 generally be relied upon for accesses to these types of mappings.
2602 (*) readX_relaxed(), writeX_relaxed():
2604 These are similar to readX() and writeX(), but provide weaker memory
2605 ordering guarantees. Specifically, they do not guarantee ordering with
2606 respect to locking, normal memory accesses or delay() loops (i.e.
2607 bullets 2-5 above) but they are still guaranteed to be ordered with
2608 respect to other accesses from the same CPU thread to the same
2609 peripheral when operating on __iomem pointers mapped with the default
2612 (*) readsX(), writesX():
2614 The readsX() and writesX() MMIO accessors are designed for accessing
2615 register-based, memory-mapped FIFOs residing on peripherals that are not
2616 capable of performing DMA. Consequently, they provide only the ordering
2617 guarantees of readX_relaxed() and writeX_relaxed(), as documented above.
2621 The inX() and outX() accessors are intended to access legacy port-mapped
2622 I/O peripherals, which may require special instructions on some
2623 architectures (notably x86). The port number of the peripheral being
2624 accessed is passed as an argument.
2626 Since many CPU architectures ultimately access these peripherals via an
2627 internal virtual memory mapping, the portable ordering guarantees
2628 provided by inX() and outX() are the same as those provided by readX()
2629 and writeX() respectively when accessing a mapping with the default I/O
2632 Device drivers may expect outX() to emit a non-posted write transaction
2633 that waits for a completion response from the I/O peripheral before
2634 returning. This is not guaranteed by all architectures and is therefore
2635 not part of the portable ordering semantics.
2637 (*) insX(), outsX():
2639 As above, the insX() and outsX() accessors provide the same ordering
2640 guarantees as readsX() and writesX() respectively when accessing a
2641 mapping with the default I/O attributes.
2643 (*) ioreadX(), iowriteX():
2645 These will perform appropriately for the type of access they're actually
2646 doing, be it inX()/outX() or readX()/writeX().
2648 With the exception of the string accessors (insX(), outsX(), readsX() and
2649 writesX()), all of the above assume that the underlying peripheral is
2650 little-endian and will therefore perform byte-swapping operations on big-endian
2654 ========================================
2655 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2656 ========================================
2658 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2659 maintain the appearance of program causality with respect to itself. Some CPUs
2660 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2661 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2662 of arch-specific code.
2664 This means that it must be considered that the CPU will execute its instruction
2665 stream in any order it feels like - or even in parallel - provided that if an
2666 instruction in the stream depends on an earlier instruction, then that
2667 earlier instruction must be sufficiently complete[*] before the later
2668 instruction may proceed; in other words: provided that the appearance of
2669 causality is maintained.
2671 [*] Some instructions have more than one effect - such as changing the
2672 condition codes, changing registers or changing memory - and different
2673 instructions may depend on different effects.
2675 A CPU may also discard any instruction sequence that winds up having no
2676 ultimate effect. For example, if two adjacent instructions both load an
2677 immediate value into the same register, the first may be discarded.
2680 Similarly, it has to be assumed that compiler might reorder the instruction
2681 stream in any way it sees fit, again provided the appearance of causality is
2685 ============================
2686 THE EFFECTS OF THE CPU CACHE
2687 ============================
2689 The way cached memory operations are perceived across the system is affected to
2690 a certain extent by the caches that lie between CPUs and memory, and by the
2691 memory coherence system that maintains the consistency of state in the system.
2693 As far as the way a CPU interacts with another part of the system through the
2694 caches goes, the memory system has to include the CPU's caches, and memory
2695 barriers for the most part act at the interface between the CPU and its cache
2696 (memory barriers logically act on the dotted line in the following diagram):
2698 <--- CPU ---> : <----------- Memory ----------->
2700 +--------+ +--------+ : +--------+ +-----------+
2701 | | | | : | | | | +--------+
2702 | CPU | | Memory | : | CPU | | | | |
2703 | Core |--->| Access |----->| Cache |<-->| | | |
2704 | | | Queue | : | | | |--->| Memory |
2705 | | | | : | | | | | |
2706 +--------+ +--------+ : +--------+ | | | |
2707 : | Cache | +--------+
2709 : | Mechanism | +--------+
2710 +--------+ +--------+ : +--------+ | | | |
2711 | | | | : | | | | | |
2712 | CPU | | Memory | : | CPU | | |--->| Device |
2713 | Core |--->| Access |----->| Cache |<-->| | | |
2714 | | | Queue | : | | | | | |
2715 | | | | : | | | | +--------+
2716 +--------+ +--------+ : +--------+ +-----------+
2720 Although any particular load or store may not actually appear outside of the
2721 CPU that issued it since it may have been satisfied within the CPU's own cache,
2722 it will still appear as if the full memory access had taken place as far as the
2723 other CPUs are concerned since the cache coherency mechanisms will migrate the
2724 cacheline over to the accessing CPU and propagate the effects upon conflict.
2726 The CPU core may execute instructions in any order it deems fit, provided the
2727 expected program causality appears to be maintained. Some of the instructions
2728 generate load and store operations which then go into the queue of memory
2729 accesses to be performed. The core may place these in the queue in any order
2730 it wishes, and continue execution until it is forced to wait for an instruction
2733 What memory barriers are concerned with is controlling the order in which
2734 accesses cross from the CPU side of things to the memory side of things, and
2735 the order in which the effects are perceived to happen by the other observers
2738 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2739 their own loads and stores as if they had happened in program order.
2741 [!] MMIO or other device accesses may bypass the cache system. This depends on
2742 the properties of the memory window through which devices are accessed and/or
2743 the use of any special device communication instructions the CPU may have.
2746 CACHE COHERENCY VS DMA
2747 ----------------------
2749 Not all systems maintain cache coherency with respect to devices doing DMA. In
2750 such cases, a device attempting DMA may obtain stale data from RAM because
2751 dirty cache lines may be resident in the caches of various CPUs, and may not
2752 have been written back to RAM yet. To deal with this, the appropriate part of
2753 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2754 invalidate them as well).
2756 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2757 cache lines being written back to RAM from a CPU's cache after the device has
2758 installed its own data, or cache lines present in the CPU's cache may simply
2759 obscure the fact that RAM has been updated, until at such time as the cacheline
2760 is discarded from the CPU's cache and reloaded. To deal with this, the
2761 appropriate part of the kernel must invalidate the overlapping bits of the
2764 See Documentation/core-api/cachetlb.rst for more information on cache
2768 CACHE COHERENCY VS MMIO
2769 -----------------------
2771 Memory mapped I/O usually takes place through memory locations that are part of
2772 a window in the CPU's memory space that has different properties assigned than
2773 the usual RAM directed window.
2775 Amongst these properties is usually the fact that such accesses bypass the
2776 caching entirely and go directly to the device buses. This means MMIO accesses
2777 may, in effect, overtake accesses to cached memory that were emitted earlier.
2778 A memory barrier isn't sufficient in such a case, but rather the cache must be
2779 flushed between the cached memory write and the MMIO access if the two are in
2783 =========================
2784 THE THINGS CPUS GET UP TO
2785 =========================
2787 A programmer might take it for granted that the CPU will perform memory
2788 operations in exactly the order specified, so that if the CPU is, for example,
2789 given the following piece of code to execute:
2797 they would then expect that the CPU will complete the memory operation for each
2798 instruction before moving on to the next one, leading to a definite sequence of
2799 operations as seen by external observers in the system:
2801 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2804 Reality is, of course, much messier. With many CPUs and compilers, the above
2805 assumption doesn't hold because:
2807 (*) loads are more likely to need to be completed immediately to permit
2808 execution progress, whereas stores can often be deferred without a
2811 (*) loads may be done speculatively, and the result discarded should it prove
2812 to have been unnecessary;
2814 (*) loads may be done speculatively, leading to the result having been fetched
2815 at the wrong time in the expected sequence of events;
2817 (*) the order of the memory accesses may be rearranged to promote better use
2818 of the CPU buses and caches;
2820 (*) loads and stores may be combined to improve performance when talking to
2821 memory or I/O hardware that can do batched accesses of adjacent locations,
2822 thus cutting down on transaction setup costs (memory and PCI devices may
2823 both be able to do this); and
2825 (*) the CPU's data cache may affect the ordering, and while cache-coherency
2826 mechanisms may alleviate this - once the store has actually hit the cache
2827 - there's no guarantee that the coherency management will be propagated in
2828 order to other CPUs.
2830 So what another CPU, say, might actually observe from the above piece of code
2833 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2835 (Where "LOAD {*C,*D}" is a combined load)
2838 However, it is guaranteed that a CPU will be self-consistent: it will see its
2839 _own_ accesses appear to be correctly ordered, without the need for a memory
2840 barrier. For instance with the following code:
2849 and assuming no intervention by an external influence, it can be assumed that
2850 the final result will appear to be:
2852 U == the original value of *A
2857 The code above may cause the CPU to generate the full sequence of memory
2860 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2862 in that order, but, without intervention, the sequence may have almost any
2863 combination of elements combined or discarded, provided the program's view
2864 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE()
2865 are -not- optional in the above example, as there are architectures
2866 where a given CPU might reorder successive loads to the same location.
2867 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is
2868 necessary to prevent this, for example, on Itanium the volatile casts
2869 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq
2870 and st.rel instructions (respectively) that prevent such reordering.
2872 The compiler may also combine, discard or defer elements of the sequence before
2873 the CPU even sees them.
2884 since, without either a write barrier or an WRITE_ONCE(), it can be
2885 assumed that the effect of the storage of V to *A is lost. Similarly:
2890 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be
2896 and the LOAD operation never appear outside of the CPU.
2899 AND THEN THERE'S THE ALPHA
2900 --------------------------
2902 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2903 some versions of the Alpha CPU have a split data cache, permitting them to have
2904 two semantically-related cache lines updated at separate times. This is where
2905 the address-dependency barrier really becomes necessary as this synchronises
2906 both caches with the memory coherence system, thus making it seem like pointer
2907 changes vs new data occur in the right order.
2909 The Alpha defines the Linux kernel's memory model, although as of v4.15
2910 the Linux kernel's addition of smp_mb() to READ_ONCE() on Alpha greatly
2911 reduced its impact on the memory model.
2914 VIRTUAL MACHINE GUESTS
2915 ----------------------
2917 Guests running within virtual machines might be affected by SMP effects even if
2918 the guest itself is compiled without SMP support. This is an artifact of
2919 interfacing with an SMP host while running an UP kernel. Using mandatory
2920 barriers for this use-case would be possible but is often suboptimal.
2922 To handle this case optimally, low-level virt_mb() etc macros are available.
2923 These have the same effect as smp_mb() etc when SMP is enabled, but generate
2924 identical code for SMP and non-SMP systems. For example, virtual machine guests
2925 should use virt_mb() rather than smp_mb() when synchronizing against a
2926 (possibly SMP) host.
2928 These are equivalent to smp_mb() etc counterparts in all other respects,
2929 in particular, they do not control MMIO effects: to control
2930 MMIO effects, use mandatory barriers.
2940 Memory barriers can be used to implement circular buffering without the need
2941 of a lock to serialise the producer with the consumer. See:
2943 Documentation/core-api/circular-buffers.rst
2952 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2954 Chapter 5.2: Physical Address Space Characteristics
2955 Chapter 5.4: Caches and Write Buffers
2956 Chapter 5.5: Data Sharing
2957 Chapter 5.6: Read/Write Ordering
2959 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2960 Chapter 7.1: Memory-Access Ordering
2961 Chapter 7.4: Buffering and Combining Memory Writes
2963 ARM Architecture Reference Manual (ARMv8, for ARMv8-A architecture profile)
2964 Chapter B2: The AArch64 Application Level Memory Model
2966 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2967 System Programming Guide
2968 Chapter 7.1: Locked Atomic Operations
2969 Chapter 7.2: Memory Ordering
2970 Chapter 7.4: Serializing Instructions
2972 The SPARC Architecture Manual, Version 9
2973 Chapter 8: Memory Models
2974 Appendix D: Formal Specification of the Memory Models
2975 Appendix J: Programming with the Memory Models
2977 Storage in the PowerPC (Stone and Fitzgerald)
2979 UltraSPARC Programmer Reference Manual
2980 Chapter 5: Memory Accesses and Cacheability
2981 Chapter 15: Sparc-V9 Memory Models
2983 UltraSPARC III Cu User's Manual
2984 Chapter 9: Memory Models
2986 UltraSPARC IIIi Processor User's Manual
2987 Chapter 8: Memory Models
2989 UltraSPARC Architecture 2005
2991 Appendix D: Formal Specifications of the Memory Models
2993 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2994 Chapter 8: Memory Models
2995 Appendix F: Caches and Cache Coherency
2997 Solaris Internals, Core Kernel Architecture, p63-68:
2998 Chapter 3.3: Hardware Considerations for Locks and
3001 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3002 for Kernel Programmers:
3003 Chapter 13: Other Memory Models
3005 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3006 Section 2.6: Speculation
3007 Section 4.4: Memory Access
projects / linux-block.git / blob