1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
5 By: David Howells <dhowells@redhat.com>
6 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
7 Will Deacon <will.deacon@arm.com>
8 Peter Zijlstra <peterz@infradead.org>
14 This document is not a specification; it is intentionally (for the sake of
15 brevity) and unintentionally (due to being human) incomplete. This document is
16 meant as a guide to using the various memory barriers provided by Linux, but
17 in case of any doubt (and there are many) please ask.
19 To repeat, this document is not a specification of what Linux expects from
22 The purpose of this document is twofold:
24 (1) to specify the minimum functionality that one can rely on for any
25 particular barrier, and
27 (2) to provide a guide as to how to use the barriers that are available.
29 Note that an architecture can provide more than the minimum requirement
30 for any particular barrier, but if the architecture provides less than
31 that, that architecture is incorrect.
33 Note also that it is possible that a barrier may be a no-op for an
34 architecture because the way that arch works renders an explicit barrier
35 unnecessary in that case.
42 (*) Abstract memory access model.
47 (*) What are memory barriers?
49 - Varieties of memory barrier.
50 - What may not be assumed about memory barriers?
51 - Data dependency barriers.
52 - Control dependencies.
53 - SMP barrier pairing.
54 - Examples of memory barrier sequences.
55 - Read memory barriers vs load speculation.
56 - Multicopy atomicity.
58 (*) Explicit kernel barriers.
61 - CPU memory barriers.
64 (*) Implicit kernel memory barriers.
66 - Lock acquisition functions.
67 - Interrupt disabling functions.
68 - Sleep and wake-up functions.
69 - Miscellaneous functions.
71 (*) Inter-CPU acquiring barrier effects.
73 - Acquires vs memory accesses.
74 - Acquires vs I/O accesses.
76 (*) Where are memory barriers needed?
78 - Interprocessor interaction.
83 (*) Kernel I/O barrier effects.
85 (*) Assumed minimum execution ordering model.
87 (*) The effects of the cpu cache.
90 - Cache coherency vs DMA.
91 - Cache coherency vs MMIO.
93 (*) The things CPUs get up to.
95 - And then there's the Alpha.
96 - Virtual Machine Guests.
105 ============================
106 ABSTRACT MEMORY ACCESS MODEL
107 ============================
109 Consider the following abstract model of the system:
114 +-------+ : +--------+ : +-------+
117 | CPU 1 |<----->| Memory |<----->| CPU 2 |
120 +-------+ : +--------+ : +-------+
128 +---------->| Device |<----------+
134 Each CPU executes a program that generates memory access operations. In the
135 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
136 perform the memory operations in any order it likes, provided program causality
137 appears to be maintained. Similarly, the compiler may also arrange the
138 instructions it emits in any order it likes, provided it doesn't affect the
139 apparent operation of the program.
141 So in the above diagram, the effects of the memory operations performed by a
142 CPU are perceived by the rest of the system as the operations cross the
143 interface between the CPU and rest of the system (the dotted lines).
146 For example, consider the following sequence of events:
149 =============== ===============
154 The set of accesses as seen by the memory system in the middle can be arranged
155 in 24 different combinations:
157 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4
158 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3
159 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4
160 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4
161 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3
162 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4
163 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4
167 and can thus result in four different combinations of values:
175 Furthermore, the stores committed by a CPU to the memory system may not be
176 perceived by the loads made by another CPU in the same order as the stores were
180 As a further example, consider this sequence of events:
183 =============== ===============
184 { A == 1, B == 2, C == 3, P == &A, Q == &C }
188 There is an obvious data dependency here, as the value loaded into D depends on
189 the address retrieved from P by CPU 2. At the end of the sequence, any of the
190 following results are possible:
192 (Q == &A) and (D == 1)
193 (Q == &B) and (D == 2)
194 (Q == &B) and (D == 4)
196 Note that CPU 2 will never try and load C into D because the CPU will load P
197 into Q before issuing the load of *Q.
203 Some devices present their control interfaces as collections of memory
204 locations, but the order in which the control registers are accessed is very
205 important. For instance, imagine an ethernet card with a set of internal
206 registers that are accessed through an address port register (A) and a data
207 port register (D). To read internal register 5, the following code might then
213 but this might show up as either of the following two sequences:
215 STORE *A = 5, x = LOAD *D
216 x = LOAD *D, STORE *A = 5
218 the second of which will almost certainly result in a malfunction, since it set
219 the address _after_ attempting to read the register.
225 There are some minimal guarantees that may be expected of a CPU:
227 (*) On any given CPU, dependent memory accesses will be issued in order, with
228 respect to itself. This means that for:
230 Q = READ_ONCE(P); D = READ_ONCE(*Q);
232 the CPU will issue the following memory operations:
234 Q = LOAD P, D = LOAD *Q
236 and always in that order. However, on DEC Alpha, READ_ONCE() also
237 emits a memory-barrier instruction, so that a DEC Alpha CPU will
238 instead issue the following memory operations:
240 Q = LOAD P, MEMORY_BARRIER, D = LOAD *Q, MEMORY_BARRIER
242 Whether on DEC Alpha or not, the READ_ONCE() also prevents compiler
245 (*) Overlapping loads and stores within a particular CPU will appear to be
246 ordered within that CPU. This means that for:
248 a = READ_ONCE(*X); WRITE_ONCE(*X, b);
250 the CPU will only issue the following sequence of memory operations:
252 a = LOAD *X, STORE *X = b
256 WRITE_ONCE(*X, c); d = READ_ONCE(*X);
258 the CPU will only issue:
260 STORE *X = c, d = LOAD *X
262 (Loads and stores overlap if they are targeted at overlapping pieces of
265 And there are a number of things that _must_ or _must_not_ be assumed:
267 (*) It _must_not_ be assumed that the compiler will do what you want
268 with memory references that are not protected by READ_ONCE() and
269 WRITE_ONCE(). Without them, the compiler is within its rights to
270 do all sorts of "creative" transformations, which are covered in
271 the COMPILER BARRIER section.
273 (*) It _must_not_ be assumed that independent loads and stores will be issued
274 in the order given. This means that for:
276 X = *A; Y = *B; *D = Z;
278 we may get any of the following sequences:
280 X = LOAD *A, Y = LOAD *B, STORE *D = Z
281 X = LOAD *A, STORE *D = Z, Y = LOAD *B
282 Y = LOAD *B, X = LOAD *A, STORE *D = Z
283 Y = LOAD *B, STORE *D = Z, X = LOAD *A
284 STORE *D = Z, X = LOAD *A, Y = LOAD *B
285 STORE *D = Z, Y = LOAD *B, X = LOAD *A
287 (*) It _must_ be assumed that overlapping memory accesses may be merged or
288 discarded. This means that for:
290 X = *A; Y = *(A + 4);
292 we may get any one of the following sequences:
294 X = LOAD *A; Y = LOAD *(A + 4);
295 Y = LOAD *(A + 4); X = LOAD *A;
296 {X, Y} = LOAD {*A, *(A + 4) };
300 *A = X; *(A + 4) = Y;
304 STORE *A = X; STORE *(A + 4) = Y;
305 STORE *(A + 4) = Y; STORE *A = X;
306 STORE {*A, *(A + 4) } = {X, Y};
308 And there are anti-guarantees:
310 (*) These guarantees do not apply to bitfields, because compilers often
311 generate code to modify these using non-atomic read-modify-write
312 sequences. Do not attempt to use bitfields to synchronize parallel
315 (*) Even in cases where bitfields are protected by locks, all fields
316 in a given bitfield must be protected by one lock. If two fields
317 in a given bitfield are protected by different locks, the compiler's
318 non-atomic read-modify-write sequences can cause an update to one
319 field to corrupt the value of an adjacent field.
321 (*) These guarantees apply only to properly aligned and sized scalar
322 variables. "Properly sized" currently means variables that are
323 the same size as "char", "short", "int" and "long". "Properly
324 aligned" means the natural alignment, thus no constraints for
325 "char", two-byte alignment for "short", four-byte alignment for
326 "int", and either four-byte or eight-byte alignment for "long",
327 on 32-bit and 64-bit systems, respectively. Note that these
328 guarantees were introduced into the C11 standard, so beware when
329 using older pre-C11 compilers (for example, gcc 4.6). The portion
330 of the standard containing this guarantee is Section 3.14, which
331 defines "memory location" as follows:
334 either an object of scalar type, or a maximal sequence
335 of adjacent bit-fields all having nonzero width
337 NOTE 1: Two threads of execution can update and access
338 separate memory locations without interfering with
341 NOTE 2: A bit-field and an adjacent non-bit-field member
342 are in separate memory locations. The same applies
343 to two bit-fields, if one is declared inside a nested
344 structure declaration and the other is not, or if the two
345 are separated by a zero-length bit-field declaration,
346 or if they are separated by a non-bit-field member
347 declaration. It is not safe to concurrently update two
348 bit-fields in the same structure if all members declared
349 between them are also bit-fields, no matter what the
350 sizes of those intervening bit-fields happen to be.
353 =========================
354 WHAT ARE MEMORY BARRIERS?
355 =========================
357 As can be seen above, independent memory operations are effectively performed
358 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
359 What is required is some way of intervening to instruct the compiler and the
360 CPU to restrict the order.
362 Memory barriers are such interventions. They impose a perceived partial
363 ordering over the memory operations on either side of the barrier.
365 Such enforcement is important because the CPUs and other devices in a system
366 can use a variety of tricks to improve performance, including reordering,
367 deferral and combination of memory operations; speculative loads; speculative
368 branch prediction and various types of caching. Memory barriers are used to
369 override or suppress these tricks, allowing the code to sanely control the
370 interaction of multiple CPUs and/or devices.
373 VARIETIES OF MEMORY BARRIER
374 ---------------------------
376 Memory barriers come in four basic varieties:
378 (1) Write (or store) memory barriers.
380 A write memory barrier gives a guarantee that all the STORE operations
381 specified before the barrier will appear to happen before all the STORE
382 operations specified after the barrier with respect to the other
383 components of the system.
385 A write barrier is a partial ordering on stores only; it is not required
386 to have any effect on loads.
388 A CPU can be viewed as committing a sequence of store operations to the
389 memory system as time progresses. All stores _before_ a write barrier
390 will occur _before_ all the stores after the write barrier.
392 [!] Note that write barriers should normally be paired with read or data
393 dependency barriers; see the "SMP barrier pairing" subsection.
396 (2) Data dependency barriers.
398 A data dependency barrier is a weaker form of read barrier. In the case
399 where two loads are performed such that the second depends on the result
400 of the first (eg: the first load retrieves the address to which the second
401 load will be directed), a data dependency barrier would be required to
402 make sure that the target of the second load is updated before the address
403 obtained by the first load is accessed.
405 A data dependency barrier is a partial ordering on interdependent loads
406 only; it is not required to have any effect on stores, independent loads
407 or overlapping loads.
409 As mentioned in (1), the other CPUs in the system can be viewed as
410 committing sequences of stores to the memory system that the CPU being
411 considered can then perceive. A data dependency barrier issued by the CPU
412 under consideration guarantees that for any load preceding it, if that
413 load touches one of a sequence of stores from another CPU, then by the
414 time the barrier completes, the effects of all the stores prior to that
415 touched by the load will be perceptible to any loads issued after the data
418 See the "Examples of memory barrier sequences" subsection for diagrams
419 showing the ordering constraints.
421 [!] Note that the first load really has to have a _data_ dependency and
422 not a control dependency. If the address for the second load is dependent
423 on the first load, but the dependency is through a conditional rather than
424 actually loading the address itself, then it's a _control_ dependency and
425 a full read barrier or better is required. See the "Control dependencies"
426 subsection for more information.
428 [!] Note that data dependency barriers should normally be paired with
429 write barriers; see the "SMP barrier pairing" subsection.
432 (3) Read (or load) memory barriers.
434 A read barrier is a data dependency barrier plus a guarantee that all the
435 LOAD operations specified before the barrier will appear to happen before
436 all the LOAD operations specified after the barrier with respect to the
437 other components of the system.
439 A read barrier is a partial ordering on loads only; it is not required to
440 have any effect on stores.
442 Read memory barriers imply data dependency barriers, and so can substitute
445 [!] Note that read barriers should normally be paired with write barriers;
446 see the "SMP barrier pairing" subsection.
449 (4) General memory barriers.
451 A general memory barrier gives a guarantee that all the LOAD and STORE
452 operations specified before the barrier will appear to happen before all
453 the LOAD and STORE operations specified after the barrier with respect to
454 the other components of the system.
456 A general memory barrier is a partial ordering over both loads and stores.
458 General memory barriers imply both read and write memory barriers, and so
459 can substitute for either.
462 And a couple of implicit varieties:
464 (5) ACQUIRE operations.
466 This acts as a one-way permeable barrier. It guarantees that all memory
467 operations after the ACQUIRE operation will appear to happen after the
468 ACQUIRE operation with respect to the other components of the system.
469 ACQUIRE operations include LOCK operations and both smp_load_acquire()
470 and smp_cond_acquire() operations. The later builds the necessary ACQUIRE
471 semantics from relying on a control dependency and smp_rmb().
473 Memory operations that occur before an ACQUIRE operation may appear to
474 happen after it completes.
476 An ACQUIRE operation should almost always be paired with a RELEASE
480 (6) RELEASE operations.
482 This also acts as a one-way permeable barrier. It guarantees that all
483 memory operations before the RELEASE operation will appear to happen
484 before the RELEASE operation with respect to the other components of the
485 system. RELEASE operations include UNLOCK operations and
486 smp_store_release() operations.
488 Memory operations that occur after a RELEASE operation may appear to
489 happen before it completes.
491 The use of ACQUIRE and RELEASE operations generally precludes the need
492 for other sorts of memory barrier (but note the exceptions mentioned in
493 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
494 pair is -not- guaranteed to act as a full memory barrier. However, after
495 an ACQUIRE on a given variable, all memory accesses preceding any prior
496 RELEASE on that same variable are guaranteed to be visible. In other
497 words, within a given variable's critical section, all accesses of all
498 previous critical sections for that variable are guaranteed to have
501 This means that ACQUIRE acts as a minimal "acquire" operation and
502 RELEASE acts as a minimal "release" operation.
504 A subset of the atomic operations described in atomic_t.txt have ACQUIRE and
505 RELEASE variants in addition to fully-ordered and relaxed (no barrier
506 semantics) definitions. For compound atomics performing both a load and a
507 store, ACQUIRE semantics apply only to the load and RELEASE semantics apply
508 only to the store portion of the operation.
510 Memory barriers are only required where there's a possibility of interaction
511 between two CPUs or between a CPU and a device. If it can be guaranteed that
512 there won't be any such interaction in any particular piece of code, then
513 memory barriers are unnecessary in that piece of code.
516 Note that these are the _minimum_ guarantees. Different architectures may give
517 more substantial guarantees, but they may _not_ be relied upon outside of arch
521 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
522 ----------------------------------------------
524 There are certain things that the Linux kernel memory barriers do not guarantee:
526 (*) There is no guarantee that any of the memory accesses specified before a
527 memory barrier will be _complete_ by the completion of a memory barrier
528 instruction; the barrier can be considered to draw a line in that CPU's
529 access queue that accesses of the appropriate type may not cross.
531 (*) There is no guarantee that issuing a memory barrier on one CPU will have
532 any direct effect on another CPU or any other hardware in the system. The
533 indirect effect will be the order in which the second CPU sees the effects
534 of the first CPU's accesses occur, but see the next point:
536 (*) There is no guarantee that a CPU will see the correct order of effects
537 from a second CPU's accesses, even _if_ the second CPU uses a memory
538 barrier, unless the first CPU _also_ uses a matching memory barrier (see
539 the subsection on "SMP Barrier Pairing").
541 (*) There is no guarantee that some intervening piece of off-the-CPU
542 hardware[*] will not reorder the memory accesses. CPU cache coherency
543 mechanisms should propagate the indirect effects of a memory barrier
544 between CPUs, but might not do so in order.
546 [*] For information on bus mastering DMA and coherency please read:
548 Documentation/PCI/pci.txt
549 Documentation/DMA-API-HOWTO.txt
550 Documentation/DMA-API.txt
553 DATA DEPENDENCY BARRIERS
554 ------------------------
556 The usage requirements of data dependency barriers are a little subtle, and
557 it's not always obvious that they're needed. To illustrate, consider the
558 following sequence of events:
561 =============== ===============
562 { A == 1, B == 2, C == 3, P == &A, Q == &C }
569 There's a clear data dependency here, and it would seem that by the end of the
570 sequence, Q must be either &A or &B, and that:
572 (Q == &A) implies (D == 1)
573 (Q == &B) implies (D == 4)
575 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
576 leading to the following situation:
578 (Q == &B) and (D == 2) ????
580 Whilst this may seem like a failure of coherency or causality maintenance, it
581 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
584 To deal with this, a data dependency barrier or better must be inserted
585 between the address load and the data load:
588 =============== ===============
589 { A == 1, B == 2, C == 3, P == &A, Q == &C }
594 <data dependency barrier>
597 This enforces the occurrence of one of the two implications, and prevents the
598 third possibility from arising.
601 [!] Note that this extremely counterintuitive situation arises most easily on
602 machines with split caches, so that, for example, one cache bank processes
603 even-numbered cache lines and the other bank processes odd-numbered cache
604 lines. The pointer P might be stored in an odd-numbered cache line, and the
605 variable B might be stored in an even-numbered cache line. Then, if the
606 even-numbered bank of the reading CPU's cache is extremely busy while the
607 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
608 but the old value of the variable B (2).
611 A data-dependency barrier is not required to order dependent writes
612 because the CPUs that the Linux kernel supports don't do writes
613 until they are certain (1) that the write will actually happen, (2)
614 of the location of the write, and (3) of the value to be written.
615 But please carefully read the "CONTROL DEPENDENCIES" section and the
616 Documentation/RCU/rcu_dereference.txt file: The compiler can and does
617 break dependencies in a great many highly creative ways.
620 =============== ===============
621 { A == 1, B == 2, C = 3, P == &A, Q == &C }
628 Therefore, no data-dependency barrier is required to order the read into
629 Q with the store into *Q. In other words, this outcome is prohibited,
630 even without a data-dependency barrier:
632 (Q == &B) && (B == 4)
634 Please note that this pattern should be rare. After all, the whole point
635 of dependency ordering is to -prevent- writes to the data structure, along
636 with the expensive cache misses associated with those writes. This pattern
637 can be used to record rare error conditions and the like, and the CPUs'
638 naturally occurring ordering prevents such records from being lost.
641 Note well that the ordering provided by a data dependency is local to
642 the CPU containing it. See the section on "Multicopy atomicity" for
646 The data dependency barrier is very important to the RCU system,
647 for example. See rcu_assign_pointer() and rcu_dereference() in
648 include/linux/rcupdate.h. This permits the current target of an RCU'd
649 pointer to be replaced with a new modified target, without the replacement
650 target appearing to be incompletely initialised.
652 See also the subsection on "Cache Coherency" for a more thorough example.
658 Control dependencies can be a bit tricky because current compilers do
659 not understand them. The purpose of this section is to help you prevent
660 the compiler's ignorance from breaking your code.
662 A load-load control dependency requires a full read memory barrier, not
663 simply a data dependency barrier to make it work correctly. Consider the
664 following bit of code:
668 <data dependency barrier> /* BUG: No data dependency!!! */
672 This will not have the desired effect because there is no actual data
673 dependency, but rather a control dependency that the CPU may short-circuit
674 by attempting to predict the outcome in advance, so that other CPUs see
675 the load from b as having happened before the load from a. In such a
676 case what's actually required is:
684 However, stores are not speculated. This means that ordering -is- provided
685 for load-store control dependencies, as in the following example:
692 Control dependencies pair normally with other types of barriers.
693 That said, please note that neither READ_ONCE() nor WRITE_ONCE()
694 are optional! Without the READ_ONCE(), the compiler might combine the
695 load from 'a' with other loads from 'a'. Without the WRITE_ONCE(),
696 the compiler might combine the store to 'b' with other stores to 'b'.
697 Either can result in highly counterintuitive effects on ordering.
699 Worse yet, if the compiler is able to prove (say) that the value of
700 variable 'a' is always non-zero, it would be well within its rights
701 to optimize the original example by eliminating the "if" statement
705 b = 1; /* BUG: Compiler and CPU can both reorder!!! */
707 So don't leave out the READ_ONCE().
709 It is tempting to try to enforce ordering on identical stores on both
710 branches of the "if" statement as follows:
723 Unfortunately, current compilers will transform this as follows at high
728 WRITE_ONCE(b, 1); /* BUG: No ordering vs. load from a!!! */
730 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
733 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */
737 Now there is no conditional between the load from 'a' and the store to
738 'b', which means that the CPU is within its rights to reorder them:
739 The conditional is absolutely required, and must be present in the
740 assembly code even after all compiler optimizations have been applied.
741 Therefore, if you need ordering in this example, you need explicit
742 memory barriers, for example, smp_store_release():
746 smp_store_release(&b, 1);
749 smp_store_release(&b, 1);
753 In contrast, without explicit memory barriers, two-legged-if control
754 ordering is guaranteed only when the stores differ, for example:
765 The initial READ_ONCE() is still required to prevent the compiler from
766 proving the value of 'a'.
768 In addition, you need to be careful what you do with the local variable 'q',
769 otherwise the compiler might be able to guess the value and again remove
770 the needed conditional. For example:
781 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
782 equal to zero, in which case the compiler is within its rights to
783 transform the above code into the following:
789 Given this transformation, the CPU is not required to respect the ordering
790 between the load from variable 'a' and the store to variable 'b'. It is
791 tempting to add a barrier(), but this does not help. The conditional
792 is gone, and the barrier won't bring it back. Therefore, if you are
793 relying on this ordering, you should make sure that MAX is greater than
794 one, perhaps as follows:
797 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
806 Please note once again that the stores to 'b' differ. If they were
807 identical, as noted earlier, the compiler could pull this store outside
808 of the 'if' statement.
810 You must also be careful not to rely too much on boolean short-circuit
811 evaluation. Consider this example:
817 Because the first condition cannot fault and the second condition is
818 always true, the compiler can transform this example as following,
819 defeating control dependency:
824 This example underscores the need to ensure that the compiler cannot
825 out-guess your code. More generally, although READ_ONCE() does force
826 the compiler to actually emit code for a given load, it does not force
827 the compiler to use the results.
829 In addition, control dependencies apply only to the then-clause and
830 else-clause of the if-statement in question. In particular, it does
831 not necessarily apply to code following the if-statement:
839 WRITE_ONCE(c, 1); /* BUG: No ordering against the read from 'a'. */
841 It is tempting to argue that there in fact is ordering because the
842 compiler cannot reorder volatile accesses and also cannot reorder
843 the writes to 'b' with the condition. Unfortunately for this line
844 of reasoning, the compiler might compile the two writes to 'b' as
845 conditional-move instructions, as in this fanciful pseudo-assembly
855 A weakly ordered CPU would have no dependency of any sort between the load
856 from 'a' and the store to 'c'. The control dependencies would extend
857 only to the pair of cmov instructions and the store depending on them.
858 In short, control dependencies apply only to the stores in the then-clause
859 and else-clause of the if-statement in question (including functions
860 invoked by those two clauses), not to code following that if-statement.
863 Note well that the ordering provided by a control dependency is local
864 to the CPU containing it. See the section on "Multicopy atomicity"
865 for more information.
870 (*) Control dependencies can order prior loads against later stores.
871 However, they do -not- guarantee any other sort of ordering:
872 Not prior loads against later loads, nor prior stores against
873 later anything. If you need these other forms of ordering,
874 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
875 later loads, smp_mb().
877 (*) If both legs of the "if" statement begin with identical stores to
878 the same variable, then those stores must be ordered, either by
879 preceding both of them with smp_mb() or by using smp_store_release()
880 to carry out the stores. Please note that it is -not- sufficient
881 to use barrier() at beginning of each leg of the "if" statement
882 because, as shown by the example above, optimizing compilers can
883 destroy the control dependency while respecting the letter of the
886 (*) Control dependencies require at least one run-time conditional
887 between the prior load and the subsequent store, and this
888 conditional must involve the prior load. If the compiler is able
889 to optimize the conditional away, it will have also optimized
890 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE()
891 can help to preserve the needed conditional.
893 (*) Control dependencies require that the compiler avoid reordering the
894 dependency into nonexistence. Careful use of READ_ONCE() or
895 atomic{,64}_read() can help to preserve your control dependency.
896 Please see the COMPILER BARRIER section for more information.
898 (*) Control dependencies apply only to the then-clause and else-clause
899 of the if-statement containing the control dependency, including
900 any functions that these two clauses call. Control dependencies
901 do -not- apply to code following the if-statement containing the
904 (*) Control dependencies pair normally with other types of barriers.
906 (*) Control dependencies do -not- provide multicopy atomicity. If you
907 need all the CPUs to see a given store at the same time, use smp_mb().
909 (*) Compilers do not understand control dependencies. It is therefore
910 your job to ensure that they do not break your code.
916 When dealing with CPU-CPU interactions, certain types of memory barrier should
917 always be paired. A lack of appropriate pairing is almost certainly an error.
919 General barriers pair with each other, though they also pair with most
920 other types of barriers, albeit without multicopy atomicity. An acquire
921 barrier pairs with a release barrier, but both may also pair with other
922 barriers, including of course general barriers. A write barrier pairs
923 with a data dependency barrier, a control dependency, an acquire barrier,
924 a release barrier, a read barrier, or a general barrier. Similarly a
925 read barrier, control dependency, or a data dependency barrier pairs
926 with a write barrier, an acquire barrier, a release barrier, or a
930 =============== ===============
933 WRITE_ONCE(b, 2); x = READ_ONCE(b);
940 =============== ===============================
943 WRITE_ONCE(b, &a); x = READ_ONCE(b);
944 <data dependency barrier>
950 =============== ===============================
953 WRITE_ONCE(x, 1); if (r2 = READ_ONCE(x)) {
954 <implicit control dependency>
958 assert(r1 == 0 || r2 == 0);
960 Basically, the read barrier always has to be there, even though it can be of
963 [!] Note that the stores before the write barrier would normally be expected to
964 match the loads after the read barrier or the data dependency barrier, and vice
968 =================== ===================
969 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c);
970 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d);
971 <write barrier> \ <read barrier>
972 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a);
973 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b);
976 EXAMPLES OF MEMORY BARRIER SEQUENCES
977 ------------------------------------
979 Firstly, write barriers act as partial orderings on store operations.
980 Consider the following sequence of events:
983 =======================
991 This sequence of events is committed to the memory coherence system in an order
992 that the rest of the system might perceive as the unordered set of { STORE A,
993 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
998 | |------>| C=3 | } /\
999 | | : +------+ }----- \ -----> Events perceptible to
1000 | | : | A=1 | } \/ the rest of the system
1002 | CPU 1 | : | B=2 | }
1004 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
1005 | | +------+ } requires all stores prior to the
1006 | | : | E=5 | } barrier to be committed before
1007 | | : +------+ } further stores may take place
1012 | Sequence in which stores are committed to the
1013 | memory system by CPU 1
1017 Secondly, data dependency barriers act as partial orderings on data-dependent
1018 loads. Consider the following sequence of events:
1021 ======================= =======================
1022 { B = 7; X = 9; Y = 8; C = &Y }
1027 STORE D = 4 LOAD C (gets &B)
1030 Without intervention, CPU 2 may perceive the events on CPU 1 in some
1031 effectively random order, despite the write barrier issued by CPU 1:
1034 | | +------+ +-------+ | Sequence of update
1035 | |------>| B=2 |----- --->| Y->8 | | of perception on
1036 | | : +------+ \ +-------+ | CPU 2
1037 | CPU 1 | : | A=1 | \ --->| C->&Y | V
1038 | | +------+ | +-------+
1039 | | wwwwwwwwwwwwwwww | : :
1041 | | : | C=&B |--- | : : +-------+
1042 | | : +------+ \ | +-------+ | |
1043 | |------>| D=4 | ----------->| C->&B |------>| |
1044 | | +------+ | +-------+ | |
1045 +-------+ : : | : : | |
1049 Apparently incorrect ---> | | B->7 |------>| |
1050 perception of B (!) | +-------+ | |
1053 The load of X holds ---> \ | X->9 |------>| |
1054 up the maintenance \ +-------+ | |
1055 of coherence of B ----->| B->2 | +-------+
1060 In the above example, CPU 2 perceives that B is 7, despite the load of *C
1061 (which would be B) coming after the LOAD of C.
1063 If, however, a data dependency barrier were to be placed between the load of C
1064 and the load of *C (ie: B) on CPU 2:
1067 ======================= =======================
1068 { B = 7; X = 9; Y = 8; C = &Y }
1073 STORE D = 4 LOAD C (gets &B)
1074 <data dependency barrier>
1077 then the following will occur:
1080 | | +------+ +-------+
1081 | |------>| B=2 |----- --->| Y->8 |
1082 | | : +------+ \ +-------+
1083 | CPU 1 | : | A=1 | \ --->| C->&Y |
1084 | | +------+ | +-------+
1085 | | wwwwwwwwwwwwwwww | : :
1087 | | : | C=&B |--- | : : +-------+
1088 | | : +------+ \ | +-------+ | |
1089 | |------>| D=4 | ----------->| C->&B |------>| |
1090 | | +------+ | +-------+ | |
1091 +-------+ : : | : : | |
1095 | | X->9 |------>| |
1097 Makes sure all effects ---> \ ddddddddddddddddd | |
1098 prior to the store of C \ +-------+ | |
1099 are perceptible to ----->| B->2 |------>| |
1100 subsequent loads +-------+ | |
1104 And thirdly, a read barrier acts as a partial order on loads. Consider the
1105 following sequence of events:
1108 ======================= =======================
1116 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1117 some effectively random order, despite the write barrier issued by CPU 1:
1120 | | +------+ +-------+
1121 | |------>| A=1 |------ --->| A->0 |
1122 | | +------+ \ +-------+
1123 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1124 | | +------+ | +-------+
1125 | |------>| B=2 |--- | : :
1126 | | +------+ \ | : : +-------+
1127 +-------+ : : \ | +-------+ | |
1128 ---------->| B->2 |------>| |
1129 | +-------+ | CPU 2 |
1130 | | A->0 |------>| |
1140 If, however, a read barrier were to be placed between the load of B and the
1144 ======================= =======================
1153 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1157 | | +------+ +-------+
1158 | |------>| A=1 |------ --->| A->0 |
1159 | | +------+ \ +-------+
1160 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1161 | | +------+ | +-------+
1162 | |------>| B=2 |--- | : :
1163 | | +------+ \ | : : +-------+
1164 +-------+ : : \ | +-------+ | |
1165 ---------->| B->2 |------>| |
1166 | +-------+ | CPU 2 |
1169 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1170 barrier causes all effects \ +-------+ | |
1171 prior to the storage of B ---->| A->1 |------>| |
1172 to be perceptible to CPU 2 +-------+ | |
1176 To illustrate this more completely, consider what could happen if the code
1177 contained a load of A either side of the read barrier:
1180 ======================= =======================
1186 LOAD A [first load of A]
1188 LOAD A [second load of A]
1190 Even though the two loads of A both occur after the load of B, they may both
1191 come up with different values:
1194 | | +------+ +-------+
1195 | |------>| A=1 |------ --->| A->0 |
1196 | | +------+ \ +-------+
1197 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1198 | | +------+ | +-------+
1199 | |------>| B=2 |--- | : :
1200 | | +------+ \ | : : +-------+
1201 +-------+ : : \ | +-------+ | |
1202 ---------->| B->2 |------>| |
1203 | +-------+ | CPU 2 |
1207 | | A->0 |------>| 1st |
1209 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1210 barrier causes all effects \ +-------+ | |
1211 prior to the storage of B ---->| A->1 |------>| 2nd |
1212 to be perceptible to CPU 2 +-------+ | |
1216 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1217 before the read barrier completes anyway:
1220 | | +------+ +-------+
1221 | |------>| A=1 |------ --->| A->0 |
1222 | | +------+ \ +-------+
1223 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1224 | | +------+ | +-------+
1225 | |------>| B=2 |--- | : :
1226 | | +------+ \ | : : +-------+
1227 +-------+ : : \ | +-------+ | |
1228 ---------->| B->2 |------>| |
1229 | +-------+ | CPU 2 |
1233 ---->| A->1 |------>| 1st |
1235 rrrrrrrrrrrrrrrrr | |
1237 | A->1 |------>| 2nd |
1242 The guarantee is that the second load will always come up with A == 1 if the
1243 load of B came up with B == 2. No such guarantee exists for the first load of
1244 A; that may come up with either A == 0 or A == 1.
1247 READ MEMORY BARRIERS VS LOAD SPECULATION
1248 ----------------------------------------
1250 Many CPUs speculate with loads: that is they see that they will need to load an
1251 item from memory, and they find a time where they're not using the bus for any
1252 other loads, and so do the load in advance - even though they haven't actually
1253 got to that point in the instruction execution flow yet. This permits the
1254 actual load instruction to potentially complete immediately because the CPU
1255 already has the value to hand.
1257 It may turn out that the CPU didn't actually need the value - perhaps because a
1258 branch circumvented the load - in which case it can discard the value or just
1259 cache it for later use.
1264 ======================= =======================
1266 DIVIDE } Divide instructions generally
1267 DIVIDE } take a long time to perform
1270 Which might appear as this:
1274 --->| B->2 |------>| |
1278 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1279 division speculates on the +-------+ ~ | |
1283 Once the divisions are complete --> : : ~-->| |
1284 the CPU can then perform the : : | |
1285 LOAD with immediate effect : : +-------+
1288 Placing a read barrier or a data dependency barrier just before the second
1292 ======================= =======================
1299 will force any value speculatively obtained to be reconsidered to an extent
1300 dependent on the type of barrier used. If there was no change made to the
1301 speculated memory location, then the speculated value will just be used:
1305 --->| B->2 |------>| |
1309 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1310 division speculates on the +-------+ ~ | |
1315 rrrrrrrrrrrrrrrr~ | |
1322 but if there was an update or an invalidation from another CPU pending, then
1323 the speculation will be cancelled and the value reloaded:
1327 --->| B->2 |------>| |
1331 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1332 division speculates on the +-------+ ~ | |
1337 rrrrrrrrrrrrrrrrr | |
1339 The speculation is discarded ---> --->| A->1 |------>| |
1340 and an updated value is +-------+ | |
1341 retrieved : : +-------+
1345 --------------------
1347 Multicopy atomicity is a deeply intuitive notion about ordering that is
1348 not always provided by real computer systems, namely that a given store
1349 becomes visible at the same time to all CPUs, or, alternatively, that all
1350 CPUs agree on the order in which all stores become visible. However,
1351 support of full multicopy atomicity would rule out valuable hardware
1352 optimizations, so a weaker form called ``other multicopy atomicity''
1353 instead guarantees only that a given store becomes visible at the same
1354 time to all -other- CPUs. The remainder of this document discusses this
1355 weaker form, but for brevity will call it simply ``multicopy atomicity''.
1357 The following example demonstrates multicopy atomicity:
1360 ======================= ======================= =======================
1362 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1)
1363 <general barrier> <read barrier>
1366 Suppose that CPU 2's load from X returns 1, which it then stores to Y,
1367 and CPU 3's load from Y returns 1. This indicates that CPU 1's store
1368 to X precedes CPU 2's load from X and that CPU 2's store to Y precedes
1369 CPU 3's load from Y. In addition, the memory barriers guarantee that
1370 CPU 2 executes its load before its store, and CPU 3 loads from Y before
1371 it loads from X. The question is then "Can CPU 3's load from X return 0?"
1373 Because CPU 3's load from X in some sense comes after CPU 2's load, it
1374 is natural to expect that CPU 3's load from X must therefore return 1.
1375 This expectation follows from multicopy atomicity: if a load executing
1376 on CPU B follows a load from the same variable executing on CPU A (and
1377 CPU A did not originally store the value which it read), then on
1378 multicopy-atomic systems, CPU B's load must return either the same value
1379 that CPU A's load did or some later value. However, the Linux kernel
1380 does not require systems to be multicopy atomic.
1382 The use of a general memory barrier in the example above compensates
1383 for any lack of multicopy atomicity. In the example, if CPU 2's load
1384 from X returns 1 and CPU 3's load from Y returns 1, then CPU 3's load
1385 from X must indeed also return 1.
1387 However, dependencies, read barriers, and write barriers are not always
1388 able to compensate for non-multicopy atomicity. For example, suppose
1389 that CPU 2's general barrier is removed from the above example, leaving
1390 only the data dependency shown below:
1393 ======================= ======================= =======================
1395 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1)
1396 <data dependency> <read barrier>
1397 STORE Y=r1 LOAD X (reads 0)
1399 This substitution allows non-multicopy atomicity to run rampant: in
1400 this example, it is perfectly legal for CPU 2's load from X to return 1,
1401 CPU 3's load from Y to return 1, and its load from X to return 0.
1403 The key point is that although CPU 2's data dependency orders its load
1404 and store, it does not guarantee to order CPU 1's store. Thus, if this
1405 example runs on a non-multicopy-atomic system where CPUs 1 and 2 share a
1406 store buffer or a level of cache, CPU 2 might have early access to CPU 1's
1407 writes. General barriers are therefore required to ensure that all CPUs
1408 agree on the combined order of multiple accesses.
1410 General barriers can compensate not only for non-multicopy atomicity,
1411 but can also generate additional ordering that can ensure that -all-
1412 CPUs will perceive the same order of -all- operations. In contrast, a
1413 chain of release-acquire pairs do not provide this additional ordering,
1414 which means that only those CPUs on the chain are guaranteed to agree
1415 on the combined order of the accesses. For example, switching to C code
1416 in deference to the ghost of Herman Hollerith:
1422 r0 = smp_load_acquire(&x);
1424 smp_store_release(&y, 1);
1429 r1 = smp_load_acquire(&y);
1432 smp_store_release(&z, 1);
1437 r2 = smp_load_acquire(&z);
1438 smp_store_release(&x, 1);
1448 Because cpu0(), cpu1(), and cpu2() participate in a chain of
1449 smp_store_release()/smp_load_acquire() pairs, the following outcome
1452 r0 == 1 && r1 == 1 && r2 == 1
1454 Furthermore, because of the release-acquire relationship between cpu0()
1455 and cpu1(), cpu1() must see cpu0()'s writes, so that the following
1456 outcome is prohibited:
1460 However, the ordering provided by a release-acquire chain is local
1461 to the CPUs participating in that chain and does not apply to cpu3(),
1462 at least aside from stores. Therefore, the following outcome is possible:
1464 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0
1466 As an aside, the following outcome is also possible:
1468 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1
1470 Although cpu0(), cpu1(), and cpu2() will see their respective reads and
1471 writes in order, CPUs not involved in the release-acquire chain might
1472 well disagree on the order. This disagreement stems from the fact that
1473 the weak memory-barrier instructions used to implement smp_load_acquire()
1474 and smp_store_release() are not required to order prior stores against
1475 subsequent loads in all cases. This means that cpu3() can see cpu0()'s
1476 store to u as happening -after- cpu1()'s load from v, even though
1477 both cpu0() and cpu1() agree that these two operations occurred in the
1480 However, please keep in mind that smp_load_acquire() is not magic.
1481 In particular, it simply reads from its argument with ordering. It does
1482 -not- ensure that any particular value will be read. Therefore, the
1483 following outcome is possible:
1485 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0
1487 Note that this outcome can happen even on a mythical sequentially
1488 consistent system where nothing is ever reordered.
1490 To reiterate, if your code requires full ordering of all operations,
1491 use general barriers throughout.
1494 ========================
1495 EXPLICIT KERNEL BARRIERS
1496 ========================
1498 The Linux kernel has a variety of different barriers that act at different
1501 (*) Compiler barrier.
1503 (*) CPU memory barriers.
1505 (*) MMIO write barrier.
1511 The Linux kernel has an explicit compiler barrier function that prevents the
1512 compiler from moving the memory accesses either side of it to the other side:
1516 This is a general barrier -- there are no read-read or write-write
1517 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be
1518 thought of as weak forms of barrier() that affect only the specific
1519 accesses flagged by the READ_ONCE() or WRITE_ONCE().
1521 The barrier() function has the following effects:
1523 (*) Prevents the compiler from reordering accesses following the
1524 barrier() to precede any accesses preceding the barrier().
1525 One example use for this property is to ease communication between
1526 interrupt-handler code and the code that was interrupted.
1528 (*) Within a loop, forces the compiler to load the variables used
1529 in that loop's conditional on each pass through that loop.
1531 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of
1532 optimizations that, while perfectly safe in single-threaded code, can
1533 be fatal in concurrent code. Here are some examples of these sorts
1536 (*) The compiler is within its rights to reorder loads and stores
1537 to the same variable, and in some cases, the CPU is within its
1538 rights to reorder loads to the same variable. This means that
1544 Might result in an older value of x stored in a[1] than in a[0].
1545 Prevent both the compiler and the CPU from doing this as follows:
1547 a[0] = READ_ONCE(x);
1548 a[1] = READ_ONCE(x);
1550 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for
1551 accesses from multiple CPUs to a single variable.
1553 (*) The compiler is within its rights to merge successive loads from
1554 the same variable. Such merging can cause the compiler to "optimize"
1558 do_something_with(tmp);
1560 into the following code, which, although in some sense legitimate
1561 for single-threaded code, is almost certainly not what the developer
1566 do_something_with(tmp);
1568 Use READ_ONCE() to prevent the compiler from doing this to you:
1570 while (tmp = READ_ONCE(a))
1571 do_something_with(tmp);
1573 (*) The compiler is within its rights to reload a variable, for example,
1574 in cases where high register pressure prevents the compiler from
1575 keeping all data of interest in registers. The compiler might
1576 therefore optimize the variable 'tmp' out of our previous example:
1579 do_something_with(tmp);
1581 This could result in the following code, which is perfectly safe in
1582 single-threaded code, but can be fatal in concurrent code:
1585 do_something_with(a);
1587 For example, the optimized version of this code could result in
1588 passing a zero to do_something_with() in the case where the variable
1589 a was modified by some other CPU between the "while" statement and
1590 the call to do_something_with().
1592 Again, use READ_ONCE() to prevent the compiler from doing this:
1594 while (tmp = READ_ONCE(a))
1595 do_something_with(tmp);
1597 Note that if the compiler runs short of registers, it might save
1598 tmp onto the stack. The overhead of this saving and later restoring
1599 is why compilers reload variables. Doing so is perfectly safe for
1600 single-threaded code, so you need to tell the compiler about cases
1601 where it is not safe.
1603 (*) The compiler is within its rights to omit a load entirely if it knows
1604 what the value will be. For example, if the compiler can prove that
1605 the value of variable 'a' is always zero, it can optimize this code:
1608 do_something_with(tmp);
1614 This transformation is a win for single-threaded code because it
1615 gets rid of a load and a branch. The problem is that the compiler
1616 will carry out its proof assuming that the current CPU is the only
1617 one updating variable 'a'. If variable 'a' is shared, then the
1618 compiler's proof will be erroneous. Use READ_ONCE() to tell the
1619 compiler that it doesn't know as much as it thinks it does:
1621 while (tmp = READ_ONCE(a))
1622 do_something_with(tmp);
1624 But please note that the compiler is also closely watching what you
1625 do with the value after the READ_ONCE(). For example, suppose you
1626 do the following and MAX is a preprocessor macro with the value 1:
1628 while ((tmp = READ_ONCE(a)) % MAX)
1629 do_something_with(tmp);
1631 Then the compiler knows that the result of the "%" operator applied
1632 to MAX will always be zero, again allowing the compiler to optimize
1633 the code into near-nonexistence. (It will still load from the
1636 (*) Similarly, the compiler is within its rights to omit a store entirely
1637 if it knows that the variable already has the value being stored.
1638 Again, the compiler assumes that the current CPU is the only one
1639 storing into the variable, which can cause the compiler to do the
1640 wrong thing for shared variables. For example, suppose you have
1644 ... Code that does not store to variable a ...
1647 The compiler sees that the value of variable 'a' is already zero, so
1648 it might well omit the second store. This would come as a fatal
1649 surprise if some other CPU might have stored to variable 'a' in the
1652 Use WRITE_ONCE() to prevent the compiler from making this sort of
1656 ... Code that does not store to variable a ...
1659 (*) The compiler is within its rights to reorder memory accesses unless
1660 you tell it not to. For example, consider the following interaction
1661 between process-level code and an interrupt handler:
1663 void process_level(void)
1665 msg = get_message();
1669 void interrupt_handler(void)
1672 process_message(msg);
1675 There is nothing to prevent the compiler from transforming
1676 process_level() to the following, in fact, this might well be a
1677 win for single-threaded code:
1679 void process_level(void)
1682 msg = get_message();
1685 If the interrupt occurs between these two statement, then
1686 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE()
1687 to prevent this as follows:
1689 void process_level(void)
1691 WRITE_ONCE(msg, get_message());
1692 WRITE_ONCE(flag, true);
1695 void interrupt_handler(void)
1697 if (READ_ONCE(flag))
1698 process_message(READ_ONCE(msg));
1701 Note that the READ_ONCE() and WRITE_ONCE() wrappers in
1702 interrupt_handler() are needed if this interrupt handler can itself
1703 be interrupted by something that also accesses 'flag' and 'msg',
1704 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE()
1705 and WRITE_ONCE() are not needed in interrupt_handler() other than
1706 for documentation purposes. (Note also that nested interrupts
1707 do not typically occur in modern Linux kernels, in fact, if an
1708 interrupt handler returns with interrupts enabled, you will get a
1711 You should assume that the compiler can move READ_ONCE() and
1712 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(),
1713 barrier(), or similar primitives.
1715 This effect could also be achieved using barrier(), but READ_ONCE()
1716 and WRITE_ONCE() are more selective: With READ_ONCE() and
1717 WRITE_ONCE(), the compiler need only forget the contents of the
1718 indicated memory locations, while with barrier() the compiler must
1719 discard the value of all memory locations that it has currented
1720 cached in any machine registers. Of course, the compiler must also
1721 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
1722 though the CPU of course need not do so.
1724 (*) The compiler is within its rights to invent stores to a variable,
1725 as in the following example:
1732 The compiler might save a branch by optimizing this as follows:
1738 In single-threaded code, this is not only safe, but also saves
1739 a branch. Unfortunately, in concurrent code, this optimization
1740 could cause some other CPU to see a spurious value of 42 -- even
1741 if variable 'a' was never zero -- when loading variable 'b'.
1742 Use WRITE_ONCE() to prevent this as follows:
1749 The compiler can also invent loads. These are usually less
1750 damaging, but they can result in cache-line bouncing and thus in
1751 poor performance and scalability. Use READ_ONCE() to prevent
1754 (*) For aligned memory locations whose size allows them to be accessed
1755 with a single memory-reference instruction, prevents "load tearing"
1756 and "store tearing," in which a single large access is replaced by
1757 multiple smaller accesses. For example, given an architecture having
1758 16-bit store instructions with 7-bit immediate fields, the compiler
1759 might be tempted to use two 16-bit store-immediate instructions to
1760 implement the following 32-bit store:
1764 Please note that GCC really does use this sort of optimization,
1765 which is not surprising given that it would likely take more
1766 than two instructions to build the constant and then store it.
1767 This optimization can therefore be a win in single-threaded code.
1768 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1769 this optimization in a volatile store. In the absence of such bugs,
1770 use of WRITE_ONCE() prevents store tearing in the following example:
1772 WRITE_ONCE(p, 0x00010002);
1774 Use of packed structures can also result in load and store tearing,
1777 struct __attribute__((__packed__)) foo {
1782 struct foo foo1, foo2;
1789 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no
1790 volatile markings, the compiler would be well within its rights to
1791 implement these three assignment statements as a pair of 32-bit
1792 loads followed by a pair of 32-bit stores. This would result in
1793 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE()
1794 and WRITE_ONCE() again prevent tearing in this example:
1797 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b));
1800 All that aside, it is never necessary to use READ_ONCE() and
1801 WRITE_ONCE() on a variable that has been marked volatile. For example,
1802 because 'jiffies' is marked volatile, it is never necessary to
1803 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and
1804 WRITE_ONCE() are implemented as volatile casts, which has no effect when
1805 its argument is already marked volatile.
1807 Please note that these compiler barriers have no direct effect on the CPU,
1808 which may then reorder things however it wishes.
1814 The Linux kernel has eight basic CPU memory barriers:
1816 TYPE MANDATORY SMP CONDITIONAL
1817 =============== ======================= ===========================
1818 GENERAL mb() smp_mb()
1819 WRITE wmb() smp_wmb()
1820 READ rmb() smp_rmb()
1821 DATA DEPENDENCY READ_ONCE()
1824 All memory barriers except the data dependency barriers imply a compiler
1825 barrier. Data dependencies do not impose any additional compiler ordering.
1827 Aside: In the case of data dependencies, the compiler would be expected
1828 to issue the loads in the correct order (eg. `a[b]` would have to load
1829 the value of b before loading a[b]), however there is no guarantee in
1830 the C specification that the compiler may not speculate the value of b
1831 (eg. is equal to 1) and load a before b (eg. tmp = a[1]; if (b != 1)
1832 tmp = a[b]; ). There is also the problem of a compiler reloading b after
1833 having loaded a[b], thus having a newer copy of b than a[b]. A consensus
1834 has not yet been reached about these problems, however the READ_ONCE()
1835 macro is a good place to start looking.
1837 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1838 systems because it is assumed that a CPU will appear to be self-consistent,
1839 and will order overlapping accesses correctly with respect to itself.
1840 However, see the subsection on "Virtual Machine Guests" below.
1842 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1843 references to shared memory on SMP systems, though the use of locking instead
1846 Mandatory barriers should not be used to control SMP effects, since mandatory
1847 barriers impose unnecessary overhead on both SMP and UP systems. They may,
1848 however, be used to control MMIO effects on accesses through relaxed memory I/O
1849 windows. These barriers are required even on non-SMP systems as they affect
1850 the order in which memory operations appear to a device by prohibiting both the
1851 compiler and the CPU from reordering them.
1854 There are some more advanced barrier functions:
1856 (*) smp_store_mb(var, value)
1858 This assigns the value to the variable and then inserts a full memory
1859 barrier after it. It isn't guaranteed to insert anything more than a
1860 compiler barrier in a UP compilation.
1863 (*) smp_mb__before_atomic();
1864 (*) smp_mb__after_atomic();
1866 These are for use with atomic (such as add, subtract, increment and
1867 decrement) functions that don't return a value, especially when used for
1868 reference counting. These functions do not imply memory barriers.
1870 These are also used for atomic bitop functions that do not return a
1871 value (such as set_bit and clear_bit).
1873 As an example, consider a piece of code that marks an object as being dead
1874 and then decrements the object's reference count:
1877 smp_mb__before_atomic();
1878 atomic_dec(&obj->ref_count);
1880 This makes sure that the death mark on the object is perceived to be set
1881 *before* the reference counter is decremented.
1883 See Documentation/atomic_{t,bitops}.txt for more information.
1889 These are for use with consistent memory to guarantee the ordering
1890 of writes or reads of shared memory accessible to both the CPU and a
1893 For example, consider a device driver that shares memory with a device
1894 and uses a descriptor status value to indicate if the descriptor belongs
1895 to the device or the CPU, and a doorbell to notify it when new
1896 descriptors are available:
1898 if (desc->status != DEVICE_OWN) {
1899 /* do not read data until we own descriptor */
1902 /* read/modify data */
1903 read_data = desc->data;
1904 desc->data = write_data;
1906 /* flush modifications before status update */
1909 /* assign ownership */
1910 desc->status = DEVICE_OWN;
1912 /* force memory to sync before notifying device via MMIO */
1915 /* notify device of new descriptors */
1916 writel(DESC_NOTIFY, doorbell);
1919 The dma_rmb() allows us guarantee the device has released ownership
1920 before we read the data from the descriptor, and the dma_wmb() allows
1921 us to guarantee the data is written to the descriptor before the device
1922 can see it now has ownership. The wmb() is needed to guarantee that the
1923 cache coherent memory writes have completed before attempting a write to
1924 the cache incoherent MMIO region.
1926 See Documentation/DMA-API.txt for more information on consistent memory.
1932 The Linux kernel also has a special barrier for use with memory-mapped I/O
1937 This is a variation on the mandatory write barrier that causes writes to weakly
1938 ordered I/O regions to be partially ordered. Its effects may go beyond the
1939 CPU->Hardware interface and actually affect the hardware at some level.
1941 See the subsection "Acquires vs I/O accesses" for more information.
1944 ===============================
1945 IMPLICIT KERNEL MEMORY BARRIERS
1946 ===============================
1948 Some of the other functions in the linux kernel imply memory barriers, amongst
1949 which are locking and scheduling functions.
1951 This specification is a _minimum_ guarantee; any particular architecture may
1952 provide more substantial guarantees, but these may not be relied upon outside
1953 of arch specific code.
1956 LOCK ACQUISITION FUNCTIONS
1957 --------------------------
1959 The Linux kernel has a number of locking constructs:
1967 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1968 for each construct. These operations all imply certain barriers:
1970 (1) ACQUIRE operation implication:
1972 Memory operations issued after the ACQUIRE will be completed after the
1973 ACQUIRE operation has completed.
1975 Memory operations issued before the ACQUIRE may be completed after
1976 the ACQUIRE operation has completed.
1978 (2) RELEASE operation implication:
1980 Memory operations issued before the RELEASE will be completed before the
1981 RELEASE operation has completed.
1983 Memory operations issued after the RELEASE may be completed before the
1984 RELEASE operation has completed.
1986 (3) ACQUIRE vs ACQUIRE implication:
1988 All ACQUIRE operations issued before another ACQUIRE operation will be
1989 completed before that ACQUIRE operation.
1991 (4) ACQUIRE vs RELEASE implication:
1993 All ACQUIRE operations issued before a RELEASE operation will be
1994 completed before the RELEASE operation.
1996 (5) Failed conditional ACQUIRE implication:
1998 Certain locking variants of the ACQUIRE operation may fail, either due to
1999 being unable to get the lock immediately, or due to receiving an unblocked
2000 signal whilst asleep waiting for the lock to become available. Failed
2001 locks do not imply any sort of barrier.
2003 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
2004 one-way barriers is that the effects of instructions outside of a critical
2005 section may seep into the inside of the critical section.
2007 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
2008 because it is possible for an access preceding the ACQUIRE to happen after the
2009 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
2010 the two accesses can themselves then cross:
2019 ACQUIRE M, STORE *B, STORE *A, RELEASE M
2021 When the ACQUIRE and RELEASE are a lock acquisition and release,
2022 respectively, this same reordering can occur if the lock's ACQUIRE and
2023 RELEASE are to the same lock variable, but only from the perspective of
2024 another CPU not holding that lock. In short, a ACQUIRE followed by an
2025 RELEASE may -not- be assumed to be a full memory barrier.
2027 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does
2028 not imply a full memory barrier. Therefore, the CPU's execution of the
2029 critical sections corresponding to the RELEASE and the ACQUIRE can cross,
2039 ACQUIRE N, STORE *B, STORE *A, RELEASE M
2041 It might appear that this reordering could introduce a deadlock.
2042 However, this cannot happen because if such a deadlock threatened,
2043 the RELEASE would simply complete, thereby avoiding the deadlock.
2047 One key point is that we are only talking about the CPU doing
2048 the reordering, not the compiler. If the compiler (or, for
2049 that matter, the developer) switched the operations, deadlock
2052 But suppose the CPU reordered the operations. In this case,
2053 the unlock precedes the lock in the assembly code. The CPU
2054 simply elected to try executing the later lock operation first.
2055 If there is a deadlock, this lock operation will simply spin (or
2056 try to sleep, but more on that later). The CPU will eventually
2057 execute the unlock operation (which preceded the lock operation
2058 in the assembly code), which will unravel the potential deadlock,
2059 allowing the lock operation to succeed.
2061 But what if the lock is a sleeplock? In that case, the code will
2062 try to enter the scheduler, where it will eventually encounter
2063 a memory barrier, which will force the earlier unlock operation
2064 to complete, again unraveling the deadlock. There might be
2065 a sleep-unlock race, but the locking primitive needs to resolve
2066 such races properly in any case.
2068 Locks and semaphores may not provide any guarantee of ordering on UP compiled
2069 systems, and so cannot be counted on in such a situation to actually achieve
2070 anything at all - especially with respect to I/O accesses - unless combined
2071 with interrupt disabling operations.
2073 See also the section on "Inter-CPU acquiring barrier effects".
2076 As an example, consider the following:
2087 The following sequence of events is acceptable:
2089 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
2091 [+] Note that {*F,*A} indicates a combined access.
2093 But none of the following are:
2095 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
2096 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
2097 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
2098 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
2102 INTERRUPT DISABLING FUNCTIONS
2103 -----------------------------
2105 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
2106 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
2107 barriers are required in such a situation, they must be provided from some
2111 SLEEP AND WAKE-UP FUNCTIONS
2112 ---------------------------
2114 Sleeping and waking on an event flagged in global data can be viewed as an
2115 interaction between two pieces of data: the task state of the task waiting for
2116 the event and the global data used to indicate the event. To make sure that
2117 these appear to happen in the right order, the primitives to begin the process
2118 of going to sleep, and the primitives to initiate a wake up imply certain
2121 Firstly, the sleeper normally follows something like this sequence of events:
2124 set_current_state(TASK_UNINTERRUPTIBLE);
2125 if (event_indicated)
2130 A general memory barrier is interpolated automatically by set_current_state()
2131 after it has altered the task state:
2134 ===============================
2135 set_current_state();
2137 STORE current->state
2139 LOAD event_indicated
2141 set_current_state() may be wrapped by:
2144 prepare_to_wait_exclusive();
2146 which therefore also imply a general memory barrier after setting the state.
2147 The whole sequence above is available in various canned forms, all of which
2148 interpolate the memory barrier in the right place:
2151 wait_event_interruptible();
2152 wait_event_interruptible_exclusive();
2153 wait_event_interruptible_timeout();
2154 wait_event_killable();
2155 wait_event_timeout();
2160 Secondly, code that performs a wake up normally follows something like this:
2162 event_indicated = 1;
2163 wake_up(&event_wait_queue);
2167 event_indicated = 1;
2168 wake_up_process(event_daemon);
2170 A write memory barrier is implied by wake_up() and co. if and only if they
2171 wake something up. The barrier occurs before the task state is cleared, and so
2172 sits between the STORE to indicate the event and the STORE to set TASK_RUNNING:
2175 =============================== ===============================
2176 set_current_state(); STORE event_indicated
2177 smp_store_mb(); wake_up();
2178 STORE current->state <write barrier>
2179 <general barrier> STORE current->state
2180 LOAD event_indicated
2182 To repeat, this write memory barrier is present if and only if something
2183 is actually awakened. To see this, consider the following sequence of
2184 events, where X and Y are both initially zero:
2187 =============================== ===============================
2188 X = 1; STORE event_indicated
2189 smp_mb(); wake_up();
2190 Y = 1; wait_event(wq, Y == 1);
2191 wake_up(); load from Y sees 1, no memory barrier
2192 load from X might see 0
2194 In contrast, if a wakeup does occur, CPU 2's load from X would be guaranteed
2197 The available waker functions include:
2203 wake_up_interruptible();
2204 wake_up_interruptible_all();
2205 wake_up_interruptible_nr();
2206 wake_up_interruptible_poll();
2207 wake_up_interruptible_sync();
2208 wake_up_interruptible_sync_poll();
2210 wake_up_locked_poll();
2216 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2217 order multiple stores before the wake-up with respect to loads of those stored
2218 values after the sleeper has called set_current_state(). For instance, if the
2221 set_current_state(TASK_INTERRUPTIBLE);
2222 if (event_indicated)
2224 __set_current_state(TASK_RUNNING);
2225 do_something(my_data);
2230 event_indicated = 1;
2231 wake_up(&event_wait_queue);
2233 there's no guarantee that the change to event_indicated will be perceived by
2234 the sleeper as coming after the change to my_data. In such a circumstance, the
2235 code on both sides must interpolate its own memory barriers between the
2236 separate data accesses. Thus the above sleeper ought to do:
2238 set_current_state(TASK_INTERRUPTIBLE);
2239 if (event_indicated) {
2241 do_something(my_data);
2244 and the waker should do:
2248 event_indicated = 1;
2249 wake_up(&event_wait_queue);
2252 MISCELLANEOUS FUNCTIONS
2253 -----------------------
2255 Other functions that imply barriers:
2257 (*) schedule() and similar imply full memory barriers.
2260 ===================================
2261 INTER-CPU ACQUIRING BARRIER EFFECTS
2262 ===================================
2264 On SMP systems locking primitives give a more substantial form of barrier: one
2265 that does affect memory access ordering on other CPUs, within the context of
2266 conflict on any particular lock.
2269 ACQUIRES VS MEMORY ACCESSES
2270 ---------------------------
2272 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2273 three CPUs; then should the following sequence of events occur:
2276 =============================== ===============================
2277 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e);
2279 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f);
2280 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g);
2282 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h);
2284 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2285 through *H occur in, other than the constraints imposed by the separate locks
2286 on the separate CPUs. It might, for example, see:
2288 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2290 But it won't see any of:
2292 *B, *C or *D preceding ACQUIRE M
2293 *A, *B or *C following RELEASE M
2294 *F, *G or *H preceding ACQUIRE Q
2295 *E, *F or *G following RELEASE Q
2299 ACQUIRES VS I/O ACCESSES
2300 ------------------------
2302 Under certain circumstances (especially involving NUMA), I/O accesses within
2303 two spinlocked sections on two different CPUs may be seen as interleaved by the
2304 PCI bridge, because the PCI bridge does not necessarily participate in the
2305 cache-coherence protocol, and is therefore incapable of issuing the required
2306 read memory barriers.
2311 =============================== ===============================
2321 may be seen by the PCI bridge as follows:
2323 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2325 which would probably cause the hardware to malfunction.
2328 What is necessary here is to intervene with an mmiowb() before dropping the
2329 spinlock, for example:
2332 =============================== ===============================
2344 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2345 before either of the stores issued on CPU 2.
2348 Furthermore, following a store by a load from the same device obviates the need
2349 for the mmiowb(), because the load forces the store to complete before the load
2353 =============================== ===============================
2364 See Documentation/driver-api/device-io.rst for more information.
2367 =================================
2368 WHERE ARE MEMORY BARRIERS NEEDED?
2369 =================================
2371 Under normal operation, memory operation reordering is generally not going to
2372 be a problem as a single-threaded linear piece of code will still appear to
2373 work correctly, even if it's in an SMP kernel. There are, however, four
2374 circumstances in which reordering definitely _could_ be a problem:
2376 (*) Interprocessor interaction.
2378 (*) Atomic operations.
2380 (*) Accessing devices.
2385 INTERPROCESSOR INTERACTION
2386 --------------------------
2388 When there's a system with more than one processor, more than one CPU in the
2389 system may be working on the same data set at the same time. This can cause
2390 synchronisation problems, and the usual way of dealing with them is to use
2391 locks. Locks, however, are quite expensive, and so it may be preferable to
2392 operate without the use of a lock if at all possible. In such a case
2393 operations that affect both CPUs may have to be carefully ordered to prevent
2396 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2397 queued on the semaphore, by virtue of it having a piece of its stack linked to
2398 the semaphore's list of waiting processes:
2400 struct rw_semaphore {
2403 struct list_head waiters;
2406 struct rwsem_waiter {
2407 struct list_head list;
2408 struct task_struct *task;
2411 To wake up a particular waiter, the up_read() or up_write() functions have to:
2413 (1) read the next pointer from this waiter's record to know as to where the
2414 next waiter record is;
2416 (2) read the pointer to the waiter's task structure;
2418 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2420 (4) call wake_up_process() on the task; and
2422 (5) release the reference held on the waiter's task struct.
2424 In other words, it has to perform this sequence of events:
2426 LOAD waiter->list.next;
2432 and if any of these steps occur out of order, then the whole thing may
2435 Once it has queued itself and dropped the semaphore lock, the waiter does not
2436 get the lock again; it instead just waits for its task pointer to be cleared
2437 before proceeding. Since the record is on the waiter's stack, this means that
2438 if the task pointer is cleared _before_ the next pointer in the list is read,
2439 another CPU might start processing the waiter and might clobber the waiter's
2440 stack before the up*() function has a chance to read the next pointer.
2442 Consider then what might happen to the above sequence of events:
2445 =============================== ===============================
2452 Woken up by other event
2457 foo() clobbers *waiter
2459 LOAD waiter->list.next;
2462 This could be dealt with using the semaphore lock, but then the down_xxx()
2463 function has to needlessly get the spinlock again after being woken up.
2465 The way to deal with this is to insert a general SMP memory barrier:
2467 LOAD waiter->list.next;
2474 In this case, the barrier makes a guarantee that all memory accesses before the
2475 barrier will appear to happen before all the memory accesses after the barrier
2476 with respect to the other CPUs on the system. It does _not_ guarantee that all
2477 the memory accesses before the barrier will be complete by the time the barrier
2478 instruction itself is complete.
2480 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2481 compiler barrier, thus making sure the compiler emits the instructions in the
2482 right order without actually intervening in the CPU. Since there's only one
2483 CPU, that CPU's dependency ordering logic will take care of everything else.
2489 Whilst they are technically interprocessor interaction considerations, atomic
2490 operations are noted specially as some of them imply full memory barriers and
2491 some don't, but they're very heavily relied on as a group throughout the
2494 See Documentation/atomic_t.txt for more information.
2500 Many devices can be memory mapped, and so appear to the CPU as if they're just
2501 a set of memory locations. To control such a device, the driver usually has to
2502 make the right memory accesses in exactly the right order.
2504 However, having a clever CPU or a clever compiler creates a potential problem
2505 in that the carefully sequenced accesses in the driver code won't reach the
2506 device in the requisite order if the CPU or the compiler thinks it is more
2507 efficient to reorder, combine or merge accesses - something that would cause
2508 the device to malfunction.
2510 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2511 routines - such as inb() or writel() - which know how to make such accesses
2512 appropriately sequential. Whilst this, for the most part, renders the explicit
2513 use of memory barriers unnecessary, there are a couple of situations where they
2516 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2517 so for _all_ general drivers locks should be used and mmiowb() must be
2518 issued prior to unlocking the critical section.
2520 (2) If the accessor functions are used to refer to an I/O memory window with
2521 relaxed memory access properties, then _mandatory_ memory barriers are
2522 required to enforce ordering.
2524 See Documentation/driver-api/device-io.rst for more information.
2530 A driver may be interrupted by its own interrupt service routine, and thus the
2531 two parts of the driver may interfere with each other's attempts to control or
2534 This may be alleviated - at least in part - by disabling local interrupts (a
2535 form of locking), such that the critical operations are all contained within
2536 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2537 routine is executing, the driver's core may not run on the same CPU, and its
2538 interrupt is not permitted to happen again until the current interrupt has been
2539 handled, thus the interrupt handler does not need to lock against that.
2541 However, consider a driver that was talking to an ethernet card that sports an
2542 address register and a data register. If that driver's core talks to the card
2543 under interrupt-disablement and then the driver's interrupt handler is invoked:
2554 The store to the data register might happen after the second store to the
2555 address register if ordering rules are sufficiently relaxed:
2557 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2560 If ordering rules are relaxed, it must be assumed that accesses done inside an
2561 interrupt disabled section may leak outside of it and may interleave with
2562 accesses performed in an interrupt - and vice versa - unless implicit or
2563 explicit barriers are used.
2565 Normally this won't be a problem because the I/O accesses done inside such
2566 sections will include synchronous load operations on strictly ordered I/O
2567 registers that form implicit I/O barriers. If this isn't sufficient then an
2568 mmiowb() may need to be used explicitly.
2571 A similar situation may occur between an interrupt routine and two routines
2572 running on separate CPUs that communicate with each other. If such a case is
2573 likely, then interrupt-disabling locks should be used to guarantee ordering.
2576 ==========================
2577 KERNEL I/O BARRIER EFFECTS
2578 ==========================
2580 When accessing I/O memory, drivers should use the appropriate accessor
2585 These are intended to talk to I/O space rather than memory space, but
2586 that's primarily a CPU-specific concept. The i386 and x86_64 processors
2587 do indeed have special I/O space access cycles and instructions, but many
2588 CPUs don't have such a concept.
2590 The PCI bus, amongst others, defines an I/O space concept which - on such
2591 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2592 space. However, it may also be mapped as a virtual I/O space in the CPU's
2593 memory map, particularly on those CPUs that don't support alternate I/O
2596 Accesses to this space may be fully synchronous (as on i386), but
2597 intermediary bridges (such as the PCI host bridge) may not fully honour
2600 They are guaranteed to be fully ordered with respect to each other.
2602 They are not guaranteed to be fully ordered with respect to other types of
2603 memory and I/O operation.
2605 (*) readX(), writeX():
2607 Whether these are guaranteed to be fully ordered and uncombined with
2608 respect to each other on the issuing CPU depends on the characteristics
2609 defined for the memory window through which they're accessing. On later
2610 i386 architecture machines, for example, this is controlled by way of the
2613 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2614 provided they're not accessing a prefetchable device.
2616 However, intermediary hardware (such as a PCI bridge) may indulge in
2617 deferral if it so wishes; to flush a store, a load from the same location
2618 is preferred[*], but a load from the same device or from configuration
2619 space should suffice for PCI.
2621 [*] NOTE! attempting to load from the same location as was written to may
2622 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2625 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2626 force stores to be ordered.
2628 Please refer to the PCI specification for more information on interactions
2629 between PCI transactions.
2631 (*) readX_relaxed(), writeX_relaxed()
2633 These are similar to readX() and writeX(), but provide weaker memory
2634 ordering guarantees. Specifically, they do not guarantee ordering with
2635 respect to normal memory accesses (e.g. DMA buffers) nor do they guarantee
2636 ordering with respect to LOCK or UNLOCK operations. If the latter is
2637 required, an mmiowb() barrier can be used. Note that relaxed accesses to
2638 the same peripheral are guaranteed to be ordered with respect to each
2641 (*) ioreadX(), iowriteX()
2643 These will perform appropriately for the type of access they're actually
2644 doing, be it inX()/outX() or readX()/writeX().
2647 ========================================
2648 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2649 ========================================
2651 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2652 maintain the appearance of program causality with respect to itself. Some CPUs
2653 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2654 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2655 of arch-specific code.
2657 This means that it must be considered that the CPU will execute its instruction
2658 stream in any order it feels like - or even in parallel - provided that if an
2659 instruction in the stream depends on an earlier instruction, then that
2660 earlier instruction must be sufficiently complete[*] before the later
2661 instruction may proceed; in other words: provided that the appearance of
2662 causality is maintained.
2664 [*] Some instructions have more than one effect - such as changing the
2665 condition codes, changing registers or changing memory - and different
2666 instructions may depend on different effects.
2668 A CPU may also discard any instruction sequence that winds up having no
2669 ultimate effect. For example, if two adjacent instructions both load an
2670 immediate value into the same register, the first may be discarded.
2673 Similarly, it has to be assumed that compiler might reorder the instruction
2674 stream in any way it sees fit, again provided the appearance of causality is
2678 ============================
2679 THE EFFECTS OF THE CPU CACHE
2680 ============================
2682 The way cached memory operations are perceived across the system is affected to
2683 a certain extent by the caches that lie between CPUs and memory, and by the
2684 memory coherence system that maintains the consistency of state in the system.
2686 As far as the way a CPU interacts with another part of the system through the
2687 caches goes, the memory system has to include the CPU's caches, and memory
2688 barriers for the most part act at the interface between the CPU and its cache
2689 (memory barriers logically act on the dotted line in the following diagram):
2691 <--- CPU ---> : <----------- Memory ----------->
2693 +--------+ +--------+ : +--------+ +-----------+
2694 | | | | : | | | | +--------+
2695 | CPU | | Memory | : | CPU | | | | |
2696 | Core |--->| Access |----->| Cache |<-->| | | |
2697 | | | Queue | : | | | |--->| Memory |
2698 | | | | : | | | | | |
2699 +--------+ +--------+ : +--------+ | | | |
2700 : | Cache | +--------+
2702 : | Mechanism | +--------+
2703 +--------+ +--------+ : +--------+ | | | |
2704 | | | | : | | | | | |
2705 | CPU | | Memory | : | CPU | | |--->| Device |
2706 | Core |--->| Access |----->| Cache |<-->| | | |
2707 | | | Queue | : | | | | | |
2708 | | | | : | | | | +--------+
2709 +--------+ +--------+ : +--------+ +-----------+
2713 Although any particular load or store may not actually appear outside of the
2714 CPU that issued it since it may have been satisfied within the CPU's own cache,
2715 it will still appear as if the full memory access had taken place as far as the
2716 other CPUs are concerned since the cache coherency mechanisms will migrate the
2717 cacheline over to the accessing CPU and propagate the effects upon conflict.
2719 The CPU core may execute instructions in any order it deems fit, provided the
2720 expected program causality appears to be maintained. Some of the instructions
2721 generate load and store operations which then go into the queue of memory
2722 accesses to be performed. The core may place these in the queue in any order
2723 it wishes, and continue execution until it is forced to wait for an instruction
2726 What memory barriers are concerned with is controlling the order in which
2727 accesses cross from the CPU side of things to the memory side of things, and
2728 the order in which the effects are perceived to happen by the other observers
2731 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2732 their own loads and stores as if they had happened in program order.
2734 [!] MMIO or other device accesses may bypass the cache system. This depends on
2735 the properties of the memory window through which devices are accessed and/or
2736 the use of any special device communication instructions the CPU may have.
2742 Life isn't quite as simple as it may appear above, however: for while the
2743 caches are expected to be coherent, there's no guarantee that that coherency
2744 will be ordered. This means that whilst changes made on one CPU will
2745 eventually become visible on all CPUs, there's no guarantee that they will
2746 become apparent in the same order on those other CPUs.
2749 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2750 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2755 +--------+ : +--->| Cache A |<------->| |
2756 | | : | +---------+ | |
2758 | | : | +---------+ | |
2759 +--------+ : +--->| Cache B |<------->| |
2762 : +---------+ | System |
2763 +--------+ : +--->| Cache C |<------->| |
2764 | | : | +---------+ | |
2766 | | : | +---------+ | |
2767 +--------+ : +--->| Cache D |<------->| |
2772 Imagine the system has the following properties:
2774 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2777 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2780 (*) whilst the CPU core is interrogating one cache, the other cache may be
2781 making use of the bus to access the rest of the system - perhaps to
2782 displace a dirty cacheline or to do a speculative load;
2784 (*) each cache has a queue of operations that need to be applied to that cache
2785 to maintain coherency with the rest of the system;
2787 (*) the coherency queue is not flushed by normal loads to lines already
2788 present in the cache, even though the contents of the queue may
2789 potentially affect those loads.
2791 Imagine, then, that two writes are made on the first CPU, with a write barrier
2792 between them to guarantee that they will appear to reach that CPU's caches in
2793 the requisite order:
2796 =============== =============== =======================================
2797 u == 0, v == 1 and p == &u, q == &u
2799 smp_wmb(); Make sure change to v is visible before
2801 <A:modify v=2> v is now in cache A exclusively
2803 <B:modify p=&v> p is now in cache B exclusively
2805 The write memory barrier forces the other CPUs in the system to perceive that
2806 the local CPU's caches have apparently been updated in the correct order. But
2807 now imagine that the second CPU wants to read those values:
2810 =============== =============== =======================================
2815 The above pair of reads may then fail to happen in the expected order, as the
2816 cacheline holding p may get updated in one of the second CPU's caches whilst
2817 the update to the cacheline holding v is delayed in the other of the second
2818 CPU's caches by some other cache event:
2821 =============== =============== =======================================
2822 u == 0, v == 1 and p == &u, q == &u
2825 <A:modify v=2> <C:busy>
2829 <B:modify p=&v> <D:commit p=&v>
2832 <C:read *q> Reads from v before v updated in cache
2836 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2837 no guarantee that, without intervention, the order of update will be the same
2838 as that committed on CPU 1.
2841 To intervene, we need to interpolate a data dependency barrier or a read
2842 barrier between the loads. This will force the cache to commit its coherency
2843 queue before processing any further requests:
2846 =============== =============== =======================================
2847 u == 0, v == 1 and p == &u, q == &u
2850 <A:modify v=2> <C:busy>
2854 <B:modify p=&v> <D:commit p=&v>
2856 smp_read_barrier_depends()
2860 <C:read *q> Reads from v after v updated in cache
2863 This sort of problem can be encountered on DEC Alpha processors as they have a
2864 split cache that improves performance by making better use of the data bus.
2865 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2866 access depends on a read, not all do, so it may not be relied on.
2868 Other CPUs may also have split caches, but must coordinate between the various
2869 cachelets for normal memory accesses. The semantics of the Alpha removes the
2870 need for hardware coordination in the absence of memory barriers, which
2871 permitted Alpha to sport higher CPU clock rates back in the day. However,
2872 please note that smp_read_barrier_depends() should not be used except in
2873 Alpha arch-specific code and within the READ_ONCE() macro.
2876 CACHE COHERENCY VS DMA
2877 ----------------------
2879 Not all systems maintain cache coherency with respect to devices doing DMA. In
2880 such cases, a device attempting DMA may obtain stale data from RAM because
2881 dirty cache lines may be resident in the caches of various CPUs, and may not
2882 have been written back to RAM yet. To deal with this, the appropriate part of
2883 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2884 invalidate them as well).
2886 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2887 cache lines being written back to RAM from a CPU's cache after the device has
2888 installed its own data, or cache lines present in the CPU's cache may simply
2889 obscure the fact that RAM has been updated, until at such time as the cacheline
2890 is discarded from the CPU's cache and reloaded. To deal with this, the
2891 appropriate part of the kernel must invalidate the overlapping bits of the
2894 See Documentation/cachetlb.txt for more information on cache management.
2897 CACHE COHERENCY VS MMIO
2898 -----------------------
2900 Memory mapped I/O usually takes place through memory locations that are part of
2901 a window in the CPU's memory space that has different properties assigned than
2902 the usual RAM directed window.
2904 Amongst these properties is usually the fact that such accesses bypass the
2905 caching entirely and go directly to the device buses. This means MMIO accesses
2906 may, in effect, overtake accesses to cached memory that were emitted earlier.
2907 A memory barrier isn't sufficient in such a case, but rather the cache must be
2908 flushed between the cached memory write and the MMIO access if the two are in
2912 =========================
2913 THE THINGS CPUS GET UP TO
2914 =========================
2916 A programmer might take it for granted that the CPU will perform memory
2917 operations in exactly the order specified, so that if the CPU is, for example,
2918 given the following piece of code to execute:
2926 they would then expect that the CPU will complete the memory operation for each
2927 instruction before moving on to the next one, leading to a definite sequence of
2928 operations as seen by external observers in the system:
2930 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2933 Reality is, of course, much messier. With many CPUs and compilers, the above
2934 assumption doesn't hold because:
2936 (*) loads are more likely to need to be completed immediately to permit
2937 execution progress, whereas stores can often be deferred without a
2940 (*) loads may be done speculatively, and the result discarded should it prove
2941 to have been unnecessary;
2943 (*) loads may be done speculatively, leading to the result having been fetched
2944 at the wrong time in the expected sequence of events;
2946 (*) the order of the memory accesses may be rearranged to promote better use
2947 of the CPU buses and caches;
2949 (*) loads and stores may be combined to improve performance when talking to
2950 memory or I/O hardware that can do batched accesses of adjacent locations,
2951 thus cutting down on transaction setup costs (memory and PCI devices may
2952 both be able to do this); and
2954 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2955 mechanisms may alleviate this - once the store has actually hit the cache
2956 - there's no guarantee that the coherency management will be propagated in
2957 order to other CPUs.
2959 So what another CPU, say, might actually observe from the above piece of code
2962 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2964 (Where "LOAD {*C,*D}" is a combined load)
2967 However, it is guaranteed that a CPU will be self-consistent: it will see its
2968 _own_ accesses appear to be correctly ordered, without the need for a memory
2969 barrier. For instance with the following code:
2978 and assuming no intervention by an external influence, it can be assumed that
2979 the final result will appear to be:
2981 U == the original value of *A
2986 The code above may cause the CPU to generate the full sequence of memory
2989 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2991 in that order, but, without intervention, the sequence may have almost any
2992 combination of elements combined or discarded, provided the program's view
2993 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE()
2994 are -not- optional in the above example, as there are architectures
2995 where a given CPU might reorder successive loads to the same location.
2996 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is
2997 necessary to prevent this, for example, on Itanium the volatile casts
2998 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq
2999 and st.rel instructions (respectively) that prevent such reordering.
3001 The compiler may also combine, discard or defer elements of the sequence before
3002 the CPU even sees them.
3013 since, without either a write barrier or an WRITE_ONCE(), it can be
3014 assumed that the effect of the storage of V to *A is lost. Similarly:
3019 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be
3025 and the LOAD operation never appear outside of the CPU.
3028 AND THEN THERE'S THE ALPHA
3029 --------------------------
3031 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
3032 some versions of the Alpha CPU have a split data cache, permitting them to have
3033 two semantically-related cache lines updated at separate times. This is where
3034 the data dependency barrier really becomes necessary as this synchronises both
3035 caches with the memory coherence system, thus making it seem like pointer
3036 changes vs new data occur in the right order.
3038 The Alpha defines the Linux kernel's memory barrier model.
3040 See the subsection on "Cache Coherency" above.
3043 VIRTUAL MACHINE GUESTS
3044 ----------------------
3046 Guests running within virtual machines might be affected by SMP effects even if
3047 the guest itself is compiled without SMP support. This is an artifact of
3048 interfacing with an SMP host while running an UP kernel. Using mandatory
3049 barriers for this use-case would be possible but is often suboptimal.
3051 To handle this case optimally, low-level virt_mb() etc macros are available.
3052 These have the same effect as smp_mb() etc when SMP is enabled, but generate
3053 identical code for SMP and non-SMP systems. For example, virtual machine guests
3054 should use virt_mb() rather than smp_mb() when synchronizing against a
3055 (possibly SMP) host.
3057 These are equivalent to smp_mb() etc counterparts in all other respects,
3058 in particular, they do not control MMIO effects: to control
3059 MMIO effects, use mandatory barriers.
3069 Memory barriers can be used to implement circular buffering without the need
3070 of a lock to serialise the producer with the consumer. See:
3072 Documentation/circular-buffers.txt
3081 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
3083 Chapter 5.2: Physical Address Space Characteristics
3084 Chapter 5.4: Caches and Write Buffers
3085 Chapter 5.5: Data Sharing
3086 Chapter 5.6: Read/Write Ordering
3088 AMD64 Architecture Programmer's Manual Volume 2: System Programming
3089 Chapter 7.1: Memory-Access Ordering
3090 Chapter 7.4: Buffering and Combining Memory Writes
3092 ARM Architecture Reference Manual (ARMv8, for ARMv8-A architecture profile)
3093 Chapter B2: The AArch64 Application Level Memory Model
3095 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
3096 System Programming Guide
3097 Chapter 7.1: Locked Atomic Operations
3098 Chapter 7.2: Memory Ordering
3099 Chapter 7.4: Serializing Instructions
3101 The SPARC Architecture Manual, Version 9
3102 Chapter 8: Memory Models
3103 Appendix D: Formal Specification of the Memory Models
3104 Appendix J: Programming with the Memory Models
3106 Storage in the PowerPC (Stone and Fitzgerald)
3108 UltraSPARC Programmer Reference Manual
3109 Chapter 5: Memory Accesses and Cacheability
3110 Chapter 15: Sparc-V9 Memory Models
3112 UltraSPARC III Cu User's Manual
3113 Chapter 9: Memory Models
3115 UltraSPARC IIIi Processor User's Manual
3116 Chapter 8: Memory Models
3118 UltraSPARC Architecture 2005
3120 Appendix D: Formal Specifications of the Memory Models
3122 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
3123 Chapter 8: Memory Models
3124 Appendix F: Caches and Cache Coherency
3126 Solaris Internals, Core Kernel Architecture, p63-68:
3127 Chapter 3.3: Hardware Considerations for Locks and
3130 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3131 for Kernel Programmers:
3132 Chapter 13: Other Memory Models
3134 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3135 Section 2.6: Speculation
3136 Section 4.4: Memory Access
git.samba.org / sfrench / cifs-2.6.git / blob