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 architecure 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.
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); smp_read_barrier_depends(); D = READ_ONCE(*Q);
232 the CPU will issue the following memory operations:
234 Q = LOAD P, D = LOAD *Q
236 and always in that order. On most systems, smp_read_barrier_depends()
237 does nothing, but it is required for DEC Alpha. The READ_ONCE()
238 is required to prevent compiler mischief. Please note that you
239 should normally use something like rcu_dereference() instead of
240 open-coding smp_read_barrier_depends().
242 (*) Overlapping loads and stores within a particular CPU will appear to be
243 ordered within that CPU. This means that for:
245 a = READ_ONCE(*X); WRITE_ONCE(*X, b);
247 the CPU will only issue the following sequence of memory operations:
249 a = LOAD *X, STORE *X = b
253 WRITE_ONCE(*X, c); d = READ_ONCE(*X);
255 the CPU will only issue:
257 STORE *X = c, d = LOAD *X
259 (Loads and stores overlap if they are targeted at overlapping pieces of
262 And there are a number of things that _must_ or _must_not_ be assumed:
264 (*) It _must_not_ be assumed that the compiler will do what you want
265 with memory references that are not protected by READ_ONCE() and
266 WRITE_ONCE(). Without them, the compiler is within its rights to
267 do all sorts of "creative" transformations, which are covered in
268 the COMPILER BARRIER section.
270 (*) It _must_not_ be assumed that independent loads and stores will be issued
271 in the order given. This means that for:
273 X = *A; Y = *B; *D = Z;
275 we may get any of the following sequences:
277 X = LOAD *A, Y = LOAD *B, STORE *D = Z
278 X = LOAD *A, STORE *D = Z, Y = LOAD *B
279 Y = LOAD *B, X = LOAD *A, STORE *D = Z
280 Y = LOAD *B, STORE *D = Z, X = LOAD *A
281 STORE *D = Z, X = LOAD *A, Y = LOAD *B
282 STORE *D = Z, Y = LOAD *B, X = LOAD *A
284 (*) It _must_ be assumed that overlapping memory accesses may be merged or
285 discarded. This means that for:
287 X = *A; Y = *(A + 4);
289 we may get any one of the following sequences:
291 X = LOAD *A; Y = LOAD *(A + 4);
292 Y = LOAD *(A + 4); X = LOAD *A;
293 {X, Y} = LOAD {*A, *(A + 4) };
297 *A = X; *(A + 4) = Y;
301 STORE *A = X; STORE *(A + 4) = Y;
302 STORE *(A + 4) = Y; STORE *A = X;
303 STORE {*A, *(A + 4) } = {X, Y};
305 And there are anti-guarantees:
307 (*) These guarantees do not apply to bitfields, because compilers often
308 generate code to modify these using non-atomic read-modify-write
309 sequences. Do not attempt to use bitfields to synchronize parallel
312 (*) Even in cases where bitfields are protected by locks, all fields
313 in a given bitfield must be protected by one lock. If two fields
314 in a given bitfield are protected by different locks, the compiler's
315 non-atomic read-modify-write sequences can cause an update to one
316 field to corrupt the value of an adjacent field.
318 (*) These guarantees apply only to properly aligned and sized scalar
319 variables. "Properly sized" currently means variables that are
320 the same size as "char", "short", "int" and "long". "Properly
321 aligned" means the natural alignment, thus no constraints for
322 "char", two-byte alignment for "short", four-byte alignment for
323 "int", and either four-byte or eight-byte alignment for "long",
324 on 32-bit and 64-bit systems, respectively. Note that these
325 guarantees were introduced into the C11 standard, so beware when
326 using older pre-C11 compilers (for example, gcc 4.6). The portion
327 of the standard containing this guarantee is Section 3.14, which
328 defines "memory location" as follows:
331 either an object of scalar type, or a maximal sequence
332 of adjacent bit-fields all having nonzero width
334 NOTE 1: Two threads of execution can update and access
335 separate memory locations without interfering with
338 NOTE 2: A bit-field and an adjacent non-bit-field member
339 are in separate memory locations. The same applies
340 to two bit-fields, if one is declared inside a nested
341 structure declaration and the other is not, or if the two
342 are separated by a zero-length bit-field declaration,
343 or if they are separated by a non-bit-field member
344 declaration. It is not safe to concurrently update two
345 bit-fields in the same structure if all members declared
346 between them are also bit-fields, no matter what the
347 sizes of those intervening bit-fields happen to be.
350 =========================
351 WHAT ARE MEMORY BARRIERS?
352 =========================
354 As can be seen above, independent memory operations are effectively performed
355 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
356 What is required is some way of intervening to instruct the compiler and the
357 CPU to restrict the order.
359 Memory barriers are such interventions. They impose a perceived partial
360 ordering over the memory operations on either side of the barrier.
362 Such enforcement is important because the CPUs and other devices in a system
363 can use a variety of tricks to improve performance, including reordering,
364 deferral and combination of memory operations; speculative loads; speculative
365 branch prediction and various types of caching. Memory barriers are used to
366 override or suppress these tricks, allowing the code to sanely control the
367 interaction of multiple CPUs and/or devices.
370 VARIETIES OF MEMORY BARRIER
371 ---------------------------
373 Memory barriers come in four basic varieties:
375 (1) Write (or store) memory barriers.
377 A write memory barrier gives a guarantee that all the STORE operations
378 specified before the barrier will appear to happen before all the STORE
379 operations specified after the barrier with respect to the other
380 components of the system.
382 A write barrier is a partial ordering on stores only; it is not required
383 to have any effect on loads.
385 A CPU can be viewed as committing a sequence of store operations to the
386 memory system as time progresses. All stores before a write barrier will
387 occur in the sequence _before_ all the stores after the write barrier.
389 [!] Note that write barriers should normally be paired with read or data
390 dependency barriers; see the "SMP barrier pairing" subsection.
393 (2) Data dependency barriers.
395 A data dependency barrier is a weaker form of read barrier. In the case
396 where two loads are performed such that the second depends on the result
397 of the first (eg: the first load retrieves the address to which the second
398 load will be directed), a data dependency barrier would be required to
399 make sure that the target of the second load is updated before the address
400 obtained by the first load is accessed.
402 A data dependency barrier is a partial ordering on interdependent loads
403 only; it is not required to have any effect on stores, independent loads
404 or overlapping loads.
406 As mentioned in (1), the other CPUs in the system can be viewed as
407 committing sequences of stores to the memory system that the CPU being
408 considered can then perceive. A data dependency barrier issued by the CPU
409 under consideration guarantees that for any load preceding it, if that
410 load touches one of a sequence of stores from another CPU, then by the
411 time the barrier completes, the effects of all the stores prior to that
412 touched by the load will be perceptible to any loads issued after the data
415 See the "Examples of memory barrier sequences" subsection for diagrams
416 showing the ordering constraints.
418 [!] Note that the first load really has to have a _data_ dependency and
419 not a control dependency. If the address for the second load is dependent
420 on the first load, but the dependency is through a conditional rather than
421 actually loading the address itself, then it's a _control_ dependency and
422 a full read barrier or better is required. See the "Control dependencies"
423 subsection for more information.
425 [!] Note that data dependency barriers should normally be paired with
426 write barriers; see the "SMP barrier pairing" subsection.
429 (3) Read (or load) memory barriers.
431 A read barrier is a data dependency barrier plus a guarantee that all the
432 LOAD operations specified before the barrier will appear to happen before
433 all the LOAD operations specified after the barrier with respect to the
434 other components of the system.
436 A read barrier is a partial ordering on loads only; it is not required to
437 have any effect on stores.
439 Read memory barriers imply data dependency barriers, and so can substitute
442 [!] Note that read barriers should normally be paired with write barriers;
443 see the "SMP barrier pairing" subsection.
446 (4) General memory barriers.
448 A general memory barrier gives a guarantee that all the LOAD and STORE
449 operations specified before the barrier will appear to happen before all
450 the LOAD and STORE operations specified after the barrier with respect to
451 the other components of the system.
453 A general memory barrier is a partial ordering over both loads and stores.
455 General memory barriers imply both read and write memory barriers, and so
456 can substitute for either.
459 And a couple of implicit varieties:
461 (5) ACQUIRE operations.
463 This acts as a one-way permeable barrier. It guarantees that all memory
464 operations after the ACQUIRE operation will appear to happen after the
465 ACQUIRE operation with respect to the other components of the system.
466 ACQUIRE operations include LOCK operations and both smp_load_acquire()
467 and smp_cond_acquire() operations. The later builds the necessary ACQUIRE
468 semantics from relying on a control dependency and smp_rmb().
470 Memory operations that occur before an ACQUIRE operation may appear to
471 happen after it completes.
473 An ACQUIRE operation should almost always be paired with a RELEASE
477 (6) RELEASE operations.
479 This also acts as a one-way permeable barrier. It guarantees that all
480 memory operations before the RELEASE operation will appear to happen
481 before the RELEASE operation with respect to the other components of the
482 system. RELEASE operations include UNLOCK operations and
483 smp_store_release() operations.
485 Memory operations that occur after a RELEASE operation may appear to
486 happen before it completes.
488 The use of ACQUIRE and RELEASE operations generally precludes the need
489 for other sorts of memory barrier (but note the exceptions mentioned in
490 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
491 pair is -not- guaranteed to act as a full memory barrier. However, after
492 an ACQUIRE on a given variable, all memory accesses preceding any prior
493 RELEASE on that same variable are guaranteed to be visible. In other
494 words, within a given variable's critical section, all accesses of all
495 previous critical sections for that variable are guaranteed to have
498 This means that ACQUIRE acts as a minimal "acquire" operation and
499 RELEASE acts as a minimal "release" operation.
501 A subset of the atomic operations described in atomic_ops.txt have ACQUIRE
502 and RELEASE variants in addition to fully-ordered and relaxed (no barrier
503 semantics) definitions. For compound atomics performing both a load and a
504 store, ACQUIRE semantics apply only to the load and RELEASE semantics apply
505 only to the store portion of the operation.
507 Memory barriers are only required where there's a possibility of interaction
508 between two CPUs or between a CPU and a device. If it can be guaranteed that
509 there won't be any such interaction in any particular piece of code, then
510 memory barriers are unnecessary in that piece of code.
513 Note that these are the _minimum_ guarantees. Different architectures may give
514 more substantial guarantees, but they may _not_ be relied upon outside of arch
518 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
519 ----------------------------------------------
521 There are certain things that the Linux kernel memory barriers do not guarantee:
523 (*) There is no guarantee that any of the memory accesses specified before a
524 memory barrier will be _complete_ by the completion of a memory barrier
525 instruction; the barrier can be considered to draw a line in that CPU's
526 access queue that accesses of the appropriate type may not cross.
528 (*) There is no guarantee that issuing a memory barrier on one CPU will have
529 any direct effect on another CPU or any other hardware in the system. The
530 indirect effect will be the order in which the second CPU sees the effects
531 of the first CPU's accesses occur, but see the next point:
533 (*) There is no guarantee that a CPU will see the correct order of effects
534 from a second CPU's accesses, even _if_ the second CPU uses a memory
535 barrier, unless the first CPU _also_ uses a matching memory barrier (see
536 the subsection on "SMP Barrier Pairing").
538 (*) There is no guarantee that some intervening piece of off-the-CPU
539 hardware[*] will not reorder the memory accesses. CPU cache coherency
540 mechanisms should propagate the indirect effects of a memory barrier
541 between CPUs, but might not do so in order.
543 [*] For information on bus mastering DMA and coherency please read:
545 Documentation/PCI/pci.txt
546 Documentation/DMA-API-HOWTO.txt
547 Documentation/DMA-API.txt
550 DATA DEPENDENCY BARRIERS
551 ------------------------
553 The usage requirements of data dependency barriers are a little subtle, and
554 it's not always obvious that they're needed. To illustrate, consider the
555 following sequence of events:
558 =============== ===============
559 { A == 1, B == 2, C == 3, P == &A, Q == &C }
566 There's a clear data dependency here, and it would seem that by the end of the
567 sequence, Q must be either &A or &B, and that:
569 (Q == &A) implies (D == 1)
570 (Q == &B) implies (D == 4)
572 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
573 leading to the following situation:
575 (Q == &B) and (D == 2) ????
577 Whilst this may seem like a failure of coherency or causality maintenance, it
578 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
581 To deal with this, a data dependency barrier or better must be inserted
582 between the address load and the data load:
585 =============== ===============
586 { A == 1, B == 2, C == 3, P == &A, Q == &C }
591 <data dependency barrier>
594 This enforces the occurrence of one of the two implications, and prevents the
595 third possibility from arising.
597 A data-dependency barrier must also order against dependent writes:
600 =============== ===============
601 { A == 1, B == 2, C = 3, P == &A, Q == &C }
606 <data dependency barrier>
609 The data-dependency barrier must order the read into Q with the store
610 into *Q. This prohibits this outcome:
614 Please note that this pattern should be rare. After all, the whole point
615 of dependency ordering is to -prevent- writes to the data structure, along
616 with the expensive cache misses associated with those writes. This pattern
617 can be used to record rare error conditions and the like, and the ordering
618 prevents such records from being lost.
621 [!] Note that this extremely counterintuitive situation arises most easily on
622 machines with split caches, so that, for example, one cache bank processes
623 even-numbered cache lines and the other bank processes odd-numbered cache
624 lines. The pointer P might be stored in an odd-numbered cache line, and the
625 variable B might be stored in an even-numbered cache line. Then, if the
626 even-numbered bank of the reading CPU's cache is extremely busy while the
627 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
628 but the old value of the variable B (2).
631 The data dependency barrier is very important to the RCU system,
632 for example. See rcu_assign_pointer() and rcu_dereference() in
633 include/linux/rcupdate.h. This permits the current target of an RCU'd
634 pointer to be replaced with a new modified target, without the replacement
635 target appearing to be incompletely initialised.
637 See also the subsection on "Cache Coherency" for a more thorough example.
643 A load-load control dependency requires a full read memory barrier, not
644 simply a data dependency barrier to make it work correctly. Consider the
645 following bit of code:
649 <data dependency barrier> /* BUG: No data dependency!!! */
653 This will not have the desired effect because there is no actual data
654 dependency, but rather a control dependency that the CPU may short-circuit
655 by attempting to predict the outcome in advance, so that other CPUs see
656 the load from b as having happened before the load from a. In such a
657 case what's actually required is:
665 However, stores are not speculated. This means that ordering -is- provided
666 for load-store control dependencies, as in the following example:
673 Control dependencies pair normally with other types of barriers. That
674 said, please note that READ_ONCE() is not optional! Without the
675 READ_ONCE(), the compiler might combine the load from 'a' with other
676 loads from 'a', and the store to 'b' with other stores to 'b', with
677 possible highly counterintuitive effects on ordering.
679 Worse yet, if the compiler is able to prove (say) that the value of
680 variable 'a' is always non-zero, it would be well within its rights
681 to optimize the original example by eliminating the "if" statement
685 b = p; /* BUG: Compiler and CPU can both reorder!!! */
687 So don't leave out the READ_ONCE().
689 It is tempting to try to enforce ordering on identical stores on both
690 branches of the "if" statement as follows:
703 Unfortunately, current compilers will transform this as follows at high
708 WRITE_ONCE(b, p); /* BUG: No ordering vs. load from a!!! */
710 /* WRITE_ONCE(b, p); -- moved up, BUG!!! */
713 /* WRITE_ONCE(b, p); -- moved up, BUG!!! */
717 Now there is no conditional between the load from 'a' and the store to
718 'b', which means that the CPU is within its rights to reorder them:
719 The conditional is absolutely required, and must be present in the
720 assembly code even after all compiler optimizations have been applied.
721 Therefore, if you need ordering in this example, you need explicit
722 memory barriers, for example, smp_store_release():
726 smp_store_release(&b, p);
729 smp_store_release(&b, p);
733 In contrast, without explicit memory barriers, two-legged-if control
734 ordering is guaranteed only when the stores differ, for example:
745 The initial READ_ONCE() is still required to prevent the compiler from
746 proving the value of 'a'.
748 In addition, you need to be careful what you do with the local variable 'q',
749 otherwise the compiler might be able to guess the value and again remove
750 the needed conditional. For example:
761 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
762 equal to zero, in which case the compiler is within its rights to
763 transform the above code into the following:
769 Given this transformation, the CPU is not required to respect the ordering
770 between the load from variable 'a' and the store to variable 'b'. It is
771 tempting to add a barrier(), but this does not help. The conditional
772 is gone, and the barrier won't bring it back. Therefore, if you are
773 relying on this ordering, you should make sure that MAX is greater than
774 one, perhaps as follows:
777 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
786 Please note once again that the stores to 'b' differ. If they were
787 identical, as noted earlier, the compiler could pull this store outside
788 of the 'if' statement.
790 You must also be careful not to rely too much on boolean short-circuit
791 evaluation. Consider this example:
797 Because the first condition cannot fault and the second condition is
798 always true, the compiler can transform this example as following,
799 defeating control dependency:
804 This example underscores the need to ensure that the compiler cannot
805 out-guess your code. More generally, although READ_ONCE() does force
806 the compiler to actually emit code for a given load, it does not force
807 the compiler to use the results.
809 Finally, control dependencies do -not- provide transitivity. This is
810 demonstrated by two related examples, with the initial values of
811 x and y both being zero:
814 ======================= =======================
815 r1 = READ_ONCE(x); r2 = READ_ONCE(y);
816 if (r1 > 0) if (r2 > 0)
817 WRITE_ONCE(y, 1); WRITE_ONCE(x, 1);
819 assert(!(r1 == 1 && r2 == 1));
821 The above two-CPU example will never trigger the assert(). However,
822 if control dependencies guaranteed transitivity (which they do not),
823 then adding the following CPU would guarantee a related assertion:
826 =====================
829 assert(!(r1 == 2 && r2 == 1 && x == 2)); /* FAILS!!! */
831 But because control dependencies do -not- provide transitivity, the above
832 assertion can fail after the combined three-CPU example completes. If you
833 need the three-CPU example to provide ordering, you will need smp_mb()
834 between the loads and stores in the CPU 0 and CPU 1 code fragments,
835 that is, just before or just after the "if" statements. Furthermore,
836 the original two-CPU example is very fragile and should be avoided.
838 These two examples are the LB and WWC litmus tests from this paper:
839 http://www.cl.cam.ac.uk/users/pes20/ppc-supplemental/test6.pdf and this
840 site: https://www.cl.cam.ac.uk/~pes20/ppcmem/index.html.
844 (*) Control dependencies can order prior loads against later stores.
845 However, they do -not- guarantee any other sort of ordering:
846 Not prior loads against later loads, nor prior stores against
847 later anything. If you need these other forms of ordering,
848 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
849 later loads, smp_mb().
851 (*) If both legs of the "if" statement begin with identical stores to
852 the same variable, then those stores must be ordered, either by
853 preceding both of them with smp_mb() or by using smp_store_release()
854 to carry out the stores. Please note that it is -not- sufficient
855 to use barrier() at beginning of each leg of the "if" statement
856 because, as shown by the example above, optimizing compilers can
857 destroy the control dependency while respecting the letter of the
860 (*) Control dependencies require at least one run-time conditional
861 between the prior load and the subsequent store, and this
862 conditional must involve the prior load. If the compiler is able
863 to optimize the conditional away, it will have also optimized
864 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE()
865 can help to preserve the needed conditional.
867 (*) Control dependencies require that the compiler avoid reordering the
868 dependency into nonexistence. Careful use of READ_ONCE() or
869 atomic{,64}_read() can help to preserve your control dependency.
870 Please see the COMPILER BARRIER section for more information.
872 (*) Control dependencies pair normally with other types of barriers.
874 (*) Control dependencies do -not- provide transitivity. If you
875 need transitivity, use smp_mb().
881 When dealing with CPU-CPU interactions, certain types of memory barrier should
882 always be paired. A lack of appropriate pairing is almost certainly an error.
884 General barriers pair with each other, though they also pair with most
885 other types of barriers, albeit without transitivity. An acquire barrier
886 pairs with a release barrier, but both may also pair with other barriers,
887 including of course general barriers. A write barrier pairs with a data
888 dependency barrier, a control dependency, an acquire barrier, a release
889 barrier, a read barrier, or a general barrier. Similarly a read barrier,
890 control dependency, or a data dependency barrier pairs with a write
891 barrier, an acquire barrier, a release barrier, or a general barrier:
894 =============== ===============
897 WRITE_ONCE(b, 2); x = READ_ONCE(b);
904 =============== ===============================
907 WRITE_ONCE(b, &a); x = READ_ONCE(b);
908 <data dependency barrier>
914 =============== ===============================
917 WRITE_ONCE(y, 1); if (r2 = READ_ONCE(x)) {
918 <implicit control dependency>
922 assert(r1 == 0 || r2 == 0);
924 Basically, the read barrier always has to be there, even though it can be of
927 [!] Note that the stores before the write barrier would normally be expected to
928 match the loads after the read barrier or the data dependency barrier, and vice
932 =================== ===================
933 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c);
934 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d);
935 <write barrier> \ <read barrier>
936 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a);
937 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b);
940 EXAMPLES OF MEMORY BARRIER SEQUENCES
941 ------------------------------------
943 Firstly, write barriers act as partial orderings on store operations.
944 Consider the following sequence of events:
947 =======================
955 This sequence of events is committed to the memory coherence system in an order
956 that the rest of the system might perceive as the unordered set of { STORE A,
957 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
962 | |------>| C=3 | } /\
963 | | : +------+ }----- \ -----> Events perceptible to
964 | | : | A=1 | } \/ the rest of the system
966 | CPU 1 | : | B=2 | }
968 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
969 | | +------+ } requires all stores prior to the
970 | | : | E=5 | } barrier to be committed before
971 | | : +------+ } further stores may take place
976 | Sequence in which stores are committed to the
977 | memory system by CPU 1
981 Secondly, data dependency barriers act as partial orderings on data-dependent
982 loads. Consider the following sequence of events:
985 ======================= =======================
986 { B = 7; X = 9; Y = 8; C = &Y }
991 STORE D = 4 LOAD C (gets &B)
994 Without intervention, CPU 2 may perceive the events on CPU 1 in some
995 effectively random order, despite the write barrier issued by CPU 1:
998 | | +------+ +-------+ | Sequence of update
999 | |------>| B=2 |----- --->| Y->8 | | of perception on
1000 | | : +------+ \ +-------+ | CPU 2
1001 | CPU 1 | : | A=1 | \ --->| C->&Y | V
1002 | | +------+ | +-------+
1003 | | wwwwwwwwwwwwwwww | : :
1005 | | : | C=&B |--- | : : +-------+
1006 | | : +------+ \ | +-------+ | |
1007 | |------>| D=4 | ----------->| C->&B |------>| |
1008 | | +------+ | +-------+ | |
1009 +-------+ : : | : : | |
1013 Apparently incorrect ---> | | B->7 |------>| |
1014 perception of B (!) | +-------+ | |
1017 The load of X holds ---> \ | X->9 |------>| |
1018 up the maintenance \ +-------+ | |
1019 of coherence of B ----->| B->2 | +-------+
1024 In the above example, CPU 2 perceives that B is 7, despite the load of *C
1025 (which would be B) coming after the LOAD of C.
1027 If, however, a data dependency barrier were to be placed between the load of C
1028 and the load of *C (ie: B) on CPU 2:
1031 ======================= =======================
1032 { B = 7; X = 9; Y = 8; C = &Y }
1037 STORE D = 4 LOAD C (gets &B)
1038 <data dependency barrier>
1041 then the following will occur:
1044 | | +------+ +-------+
1045 | |------>| B=2 |----- --->| Y->8 |
1046 | | : +------+ \ +-------+
1047 | CPU 1 | : | A=1 | \ --->| C->&Y |
1048 | | +------+ | +-------+
1049 | | wwwwwwwwwwwwwwww | : :
1051 | | : | C=&B |--- | : : +-------+
1052 | | : +------+ \ | +-------+ | |
1053 | |------>| D=4 | ----------->| C->&B |------>| |
1054 | | +------+ | +-------+ | |
1055 +-------+ : : | : : | |
1059 | | X->9 |------>| |
1061 Makes sure all effects ---> \ ddddddddddddddddd | |
1062 prior to the store of C \ +-------+ | |
1063 are perceptible to ----->| B->2 |------>| |
1064 subsequent loads +-------+ | |
1068 And thirdly, a read barrier acts as a partial order on loads. Consider the
1069 following sequence of events:
1072 ======================= =======================
1080 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1081 some effectively random order, despite the write barrier issued by CPU 1:
1084 | | +------+ +-------+
1085 | |------>| A=1 |------ --->| A->0 |
1086 | | +------+ \ +-------+
1087 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1088 | | +------+ | +-------+
1089 | |------>| B=2 |--- | : :
1090 | | +------+ \ | : : +-------+
1091 +-------+ : : \ | +-------+ | |
1092 ---------->| B->2 |------>| |
1093 | +-------+ | CPU 2 |
1094 | | A->0 |------>| |
1104 If, however, a read barrier were to be placed between the load of B and the
1108 ======================= =======================
1117 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1121 | | +------+ +-------+
1122 | |------>| A=1 |------ --->| A->0 |
1123 | | +------+ \ +-------+
1124 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1125 | | +------+ | +-------+
1126 | |------>| B=2 |--- | : :
1127 | | +------+ \ | : : +-------+
1128 +-------+ : : \ | +-------+ | |
1129 ---------->| B->2 |------>| |
1130 | +-------+ | CPU 2 |
1133 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1134 barrier causes all effects \ +-------+ | |
1135 prior to the storage of B ---->| A->1 |------>| |
1136 to be perceptible to CPU 2 +-------+ | |
1140 To illustrate this more completely, consider what could happen if the code
1141 contained a load of A either side of the read barrier:
1144 ======================= =======================
1150 LOAD A [first load of A]
1152 LOAD A [second load of A]
1154 Even though the two loads of A both occur after the load of B, they may both
1155 come up with different values:
1158 | | +------+ +-------+
1159 | |------>| A=1 |------ --->| A->0 |
1160 | | +------+ \ +-------+
1161 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1162 | | +------+ | +-------+
1163 | |------>| B=2 |--- | : :
1164 | | +------+ \ | : : +-------+
1165 +-------+ : : \ | +-------+ | |
1166 ---------->| B->2 |------>| |
1167 | +-------+ | CPU 2 |
1171 | | A->0 |------>| 1st |
1173 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1174 barrier causes all effects \ +-------+ | |
1175 prior to the storage of B ---->| A->1 |------>| 2nd |
1176 to be perceptible to CPU 2 +-------+ | |
1180 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1181 before the read barrier completes anyway:
1184 | | +------+ +-------+
1185 | |------>| A=1 |------ --->| A->0 |
1186 | | +------+ \ +-------+
1187 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1188 | | +------+ | +-------+
1189 | |------>| B=2 |--- | : :
1190 | | +------+ \ | : : +-------+
1191 +-------+ : : \ | +-------+ | |
1192 ---------->| B->2 |------>| |
1193 | +-------+ | CPU 2 |
1197 ---->| A->1 |------>| 1st |
1199 rrrrrrrrrrrrrrrrr | |
1201 | A->1 |------>| 2nd |
1206 The guarantee is that the second load will always come up with A == 1 if the
1207 load of B came up with B == 2. No such guarantee exists for the first load of
1208 A; that may come up with either A == 0 or A == 1.
1211 READ MEMORY BARRIERS VS LOAD SPECULATION
1212 ----------------------------------------
1214 Many CPUs speculate with loads: that is they see that they will need to load an
1215 item from memory, and they find a time where they're not using the bus for any
1216 other loads, and so do the load in advance - even though they haven't actually
1217 got to that point in the instruction execution flow yet. This permits the
1218 actual load instruction to potentially complete immediately because the CPU
1219 already has the value to hand.
1221 It may turn out that the CPU didn't actually need the value - perhaps because a
1222 branch circumvented the load - in which case it can discard the value or just
1223 cache it for later use.
1228 ======================= =======================
1230 DIVIDE } Divide instructions generally
1231 DIVIDE } take a long time to perform
1234 Which might appear as this:
1238 --->| B->2 |------>| |
1242 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1243 division speculates on the +-------+ ~ | |
1247 Once the divisions are complete --> : : ~-->| |
1248 the CPU can then perform the : : | |
1249 LOAD with immediate effect : : +-------+
1252 Placing a read barrier or a data dependency barrier just before the second
1256 ======================= =======================
1263 will force any value speculatively obtained to be reconsidered to an extent
1264 dependent on the type of barrier used. If there was no change made to the
1265 speculated memory location, then the speculated value will just be used:
1269 --->| B->2 |------>| |
1273 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1274 division speculates on the +-------+ ~ | |
1279 rrrrrrrrrrrrrrrr~ | |
1286 but if there was an update or an invalidation from another CPU pending, then
1287 the speculation will be cancelled and the value reloaded:
1291 --->| B->2 |------>| |
1295 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1296 division speculates on the +-------+ ~ | |
1301 rrrrrrrrrrrrrrrrr | |
1303 The speculation is discarded ---> --->| A->1 |------>| |
1304 and an updated value is +-------+ | |
1305 retrieved : : +-------+
1311 Transitivity is a deeply intuitive notion about ordering that is not
1312 always provided by real computer systems. The following example
1313 demonstrates transitivity:
1316 ======================= ======================= =======================
1318 STORE X=1 LOAD X STORE Y=1
1319 <general barrier> <general barrier>
1322 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1323 This indicates that CPU 2's load from X in some sense follows CPU 1's
1324 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1325 store to Y. The question is then "Can CPU 3's load from X return 0?"
1327 Because CPU 2's load from X in some sense came after CPU 1's store, it
1328 is natural to expect that CPU 3's load from X must therefore return 1.
1329 This expectation is an example of transitivity: if a load executing on
1330 CPU A follows a load from the same variable executing on CPU B, then
1331 CPU A's load must either return the same value that CPU B's load did,
1332 or must return some later value.
1334 In the Linux kernel, use of general memory barriers guarantees
1335 transitivity. Therefore, in the above example, if CPU 2's load from X
1336 returns 1 and its load from Y returns 0, then CPU 3's load from X must
1339 However, transitivity is -not- guaranteed for read or write barriers.
1340 For example, suppose that CPU 2's general barrier in the above example
1341 is changed to a read barrier as shown below:
1344 ======================= ======================= =======================
1346 STORE X=1 LOAD X STORE Y=1
1347 <read barrier> <general barrier>
1350 This substitution destroys transitivity: in this example, it is perfectly
1351 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1352 and CPU 3's load from X to return 0.
1354 The key point is that although CPU 2's read barrier orders its pair
1355 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1356 this example runs on a system where CPUs 1 and 2 share a store buffer
1357 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1358 General barriers are therefore required to ensure that all CPUs agree
1359 on the combined order of CPU 1's and CPU 2's accesses.
1361 General barriers provide "global transitivity", so that all CPUs will
1362 agree on the order of operations. In contrast, a chain of release-acquire
1363 pairs provides only "local transitivity", so that only those CPUs on
1364 the chain are guaranteed to agree on the combined order of the accesses.
1365 For example, switching to C code in deference to Herman Hollerith:
1371 r0 = smp_load_acquire(&x);
1373 smp_store_release(&y, 1);
1378 r1 = smp_load_acquire(&y);
1381 smp_store_release(&z, 1);
1386 r2 = smp_load_acquire(&z);
1387 smp_store_release(&x, 1);
1397 Because cpu0(), cpu1(), and cpu2() participate in a local transitive
1398 chain of smp_store_release()/smp_load_acquire() pairs, the following
1399 outcome is prohibited:
1401 r0 == 1 && r1 == 1 && r2 == 1
1403 Furthermore, because of the release-acquire relationship between cpu0()
1404 and cpu1(), cpu1() must see cpu0()'s writes, so that the following
1405 outcome is prohibited:
1409 However, the transitivity of release-acquire is local to the participating
1410 CPUs and does not apply to cpu3(). Therefore, the following outcome
1413 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0
1415 As an aside, the following outcome is also possible:
1417 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1
1419 Although cpu0(), cpu1(), and cpu2() will see their respective reads and
1420 writes in order, CPUs not involved in the release-acquire chain might
1421 well disagree on the order. This disagreement stems from the fact that
1422 the weak memory-barrier instructions used to implement smp_load_acquire()
1423 and smp_store_release() are not required to order prior stores against
1424 subsequent loads in all cases. This means that cpu3() can see cpu0()'s
1425 store to u as happening -after- cpu1()'s load from v, even though
1426 both cpu0() and cpu1() agree that these two operations occurred in the
1429 However, please keep in mind that smp_load_acquire() is not magic.
1430 In particular, it simply reads from its argument with ordering. It does
1431 -not- ensure that any particular value will be read. Therefore, the
1432 following outcome is possible:
1434 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0
1436 Note that this outcome can happen even on a mythical sequentially
1437 consistent system where nothing is ever reordered.
1439 To reiterate, if your code requires global transitivity, use general
1440 barriers throughout.
1443 ========================
1444 EXPLICIT KERNEL BARRIERS
1445 ========================
1447 The Linux kernel has a variety of different barriers that act at different
1450 (*) Compiler barrier.
1452 (*) CPU memory barriers.
1454 (*) MMIO write barrier.
1460 The Linux kernel has an explicit compiler barrier function that prevents the
1461 compiler from moving the memory accesses either side of it to the other side:
1465 This is a general barrier -- there are no read-read or write-write
1466 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be
1467 thought of as weak forms of barrier() that affect only the specific
1468 accesses flagged by the READ_ONCE() or WRITE_ONCE().
1470 The barrier() function has the following effects:
1472 (*) Prevents the compiler from reordering accesses following the
1473 barrier() to precede any accesses preceding the barrier().
1474 One example use for this property is to ease communication between
1475 interrupt-handler code and the code that was interrupted.
1477 (*) Within a loop, forces the compiler to load the variables used
1478 in that loop's conditional on each pass through that loop.
1480 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of
1481 optimizations that, while perfectly safe in single-threaded code, can
1482 be fatal in concurrent code. Here are some examples of these sorts
1485 (*) The compiler is within its rights to reorder loads and stores
1486 to the same variable, and in some cases, the CPU is within its
1487 rights to reorder loads to the same variable. This means that
1493 Might result in an older value of x stored in a[1] than in a[0].
1494 Prevent both the compiler and the CPU from doing this as follows:
1496 a[0] = READ_ONCE(x);
1497 a[1] = READ_ONCE(x);
1499 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for
1500 accesses from multiple CPUs to a single variable.
1502 (*) The compiler is within its rights to merge successive loads from
1503 the same variable. Such merging can cause the compiler to "optimize"
1507 do_something_with(tmp);
1509 into the following code, which, although in some sense legitimate
1510 for single-threaded code, is almost certainly not what the developer
1515 do_something_with(tmp);
1517 Use READ_ONCE() to prevent the compiler from doing this to you:
1519 while (tmp = READ_ONCE(a))
1520 do_something_with(tmp);
1522 (*) The compiler is within its rights to reload a variable, for example,
1523 in cases where high register pressure prevents the compiler from
1524 keeping all data of interest in registers. The compiler might
1525 therefore optimize the variable 'tmp' out of our previous example:
1528 do_something_with(tmp);
1530 This could result in the following code, which is perfectly safe in
1531 single-threaded code, but can be fatal in concurrent code:
1534 do_something_with(a);
1536 For example, the optimized version of this code could result in
1537 passing a zero to do_something_with() in the case where the variable
1538 a was modified by some other CPU between the "while" statement and
1539 the call to do_something_with().
1541 Again, use READ_ONCE() to prevent the compiler from doing this:
1543 while (tmp = READ_ONCE(a))
1544 do_something_with(tmp);
1546 Note that if the compiler runs short of registers, it might save
1547 tmp onto the stack. The overhead of this saving and later restoring
1548 is why compilers reload variables. Doing so is perfectly safe for
1549 single-threaded code, so you need to tell the compiler about cases
1550 where it is not safe.
1552 (*) The compiler is within its rights to omit a load entirely if it knows
1553 what the value will be. For example, if the compiler can prove that
1554 the value of variable 'a' is always zero, it can optimize this code:
1557 do_something_with(tmp);
1563 This transformation is a win for single-threaded code because it
1564 gets rid of a load and a branch. The problem is that the compiler
1565 will carry out its proof assuming that the current CPU is the only
1566 one updating variable 'a'. If variable 'a' is shared, then the
1567 compiler's proof will be erroneous. Use READ_ONCE() to tell the
1568 compiler that it doesn't know as much as it thinks it does:
1570 while (tmp = READ_ONCE(a))
1571 do_something_with(tmp);
1573 But please note that the compiler is also closely watching what you
1574 do with the value after the READ_ONCE(). For example, suppose you
1575 do the following and MAX is a preprocessor macro with the value 1:
1577 while ((tmp = READ_ONCE(a)) % MAX)
1578 do_something_with(tmp);
1580 Then the compiler knows that the result of the "%" operator applied
1581 to MAX will always be zero, again allowing the compiler to optimize
1582 the code into near-nonexistence. (It will still load from the
1585 (*) Similarly, the compiler is within its rights to omit a store entirely
1586 if it knows that the variable already has the value being stored.
1587 Again, the compiler assumes that the current CPU is the only one
1588 storing into the variable, which can cause the compiler to do the
1589 wrong thing for shared variables. For example, suppose you have
1593 ... Code that does not store to variable a ...
1596 The compiler sees that the value of variable 'a' is already zero, so
1597 it might well omit the second store. This would come as a fatal
1598 surprise if some other CPU might have stored to variable 'a' in the
1601 Use WRITE_ONCE() to prevent the compiler from making this sort of
1605 ... Code that does not store to variable a ...
1608 (*) The compiler is within its rights to reorder memory accesses unless
1609 you tell it not to. For example, consider the following interaction
1610 between process-level code and an interrupt handler:
1612 void process_level(void)
1614 msg = get_message();
1618 void interrupt_handler(void)
1621 process_message(msg);
1624 There is nothing to prevent the compiler from transforming
1625 process_level() to the following, in fact, this might well be a
1626 win for single-threaded code:
1628 void process_level(void)
1631 msg = get_message();
1634 If the interrupt occurs between these two statement, then
1635 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE()
1636 to prevent this as follows:
1638 void process_level(void)
1640 WRITE_ONCE(msg, get_message());
1641 WRITE_ONCE(flag, true);
1644 void interrupt_handler(void)
1646 if (READ_ONCE(flag))
1647 process_message(READ_ONCE(msg));
1650 Note that the READ_ONCE() and WRITE_ONCE() wrappers in
1651 interrupt_handler() are needed if this interrupt handler can itself
1652 be interrupted by something that also accesses 'flag' and 'msg',
1653 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE()
1654 and WRITE_ONCE() are not needed in interrupt_handler() other than
1655 for documentation purposes. (Note also that nested interrupts
1656 do not typically occur in modern Linux kernels, in fact, if an
1657 interrupt handler returns with interrupts enabled, you will get a
1660 You should assume that the compiler can move READ_ONCE() and
1661 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(),
1662 barrier(), or similar primitives.
1664 This effect could also be achieved using barrier(), but READ_ONCE()
1665 and WRITE_ONCE() are more selective: With READ_ONCE() and
1666 WRITE_ONCE(), the compiler need only forget the contents of the
1667 indicated memory locations, while with barrier() the compiler must
1668 discard the value of all memory locations that it has currented
1669 cached in any machine registers. Of course, the compiler must also
1670 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
1671 though the CPU of course need not do so.
1673 (*) The compiler is within its rights to invent stores to a variable,
1674 as in the following example:
1681 The compiler might save a branch by optimizing this as follows:
1687 In single-threaded code, this is not only safe, but also saves
1688 a branch. Unfortunately, in concurrent code, this optimization
1689 could cause some other CPU to see a spurious value of 42 -- even
1690 if variable 'a' was never zero -- when loading variable 'b'.
1691 Use WRITE_ONCE() to prevent this as follows:
1698 The compiler can also invent loads. These are usually less
1699 damaging, but they can result in cache-line bouncing and thus in
1700 poor performance and scalability. Use READ_ONCE() to prevent
1703 (*) For aligned memory locations whose size allows them to be accessed
1704 with a single memory-reference instruction, prevents "load tearing"
1705 and "store tearing," in which a single large access is replaced by
1706 multiple smaller accesses. For example, given an architecture having
1707 16-bit store instructions with 7-bit immediate fields, the compiler
1708 might be tempted to use two 16-bit store-immediate instructions to
1709 implement the following 32-bit store:
1713 Please note that GCC really does use this sort of optimization,
1714 which is not surprising given that it would likely take more
1715 than two instructions to build the constant and then store it.
1716 This optimization can therefore be a win in single-threaded code.
1717 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1718 this optimization in a volatile store. In the absence of such bugs,
1719 use of WRITE_ONCE() prevents store tearing in the following example:
1721 WRITE_ONCE(p, 0x00010002);
1723 Use of packed structures can also result in load and store tearing,
1726 struct __attribute__((__packed__)) foo {
1731 struct foo foo1, foo2;
1738 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no
1739 volatile markings, the compiler would be well within its rights to
1740 implement these three assignment statements as a pair of 32-bit
1741 loads followed by a pair of 32-bit stores. This would result in
1742 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE()
1743 and WRITE_ONCE() again prevent tearing in this example:
1746 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b));
1749 All that aside, it is never necessary to use READ_ONCE() and
1750 WRITE_ONCE() on a variable that has been marked volatile. For example,
1751 because 'jiffies' is marked volatile, it is never necessary to
1752 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and
1753 WRITE_ONCE() are implemented as volatile casts, which has no effect when
1754 its argument is already marked volatile.
1756 Please note that these compiler barriers have no direct effect on the CPU,
1757 which may then reorder things however it wishes.
1763 The Linux kernel has eight basic CPU memory barriers:
1765 TYPE MANDATORY SMP CONDITIONAL
1766 =============== ======================= ===========================
1767 GENERAL mb() smp_mb()
1768 WRITE wmb() smp_wmb()
1769 READ rmb() smp_rmb()
1770 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1773 All memory barriers except the data dependency barriers imply a compiler
1774 barrier. Data dependencies do not impose any additional compiler ordering.
1776 Aside: In the case of data dependencies, the compiler would be expected
1777 to issue the loads in the correct order (eg. `a[b]` would have to load
1778 the value of b before loading a[b]), however there is no guarantee in
1779 the C specification that the compiler may not speculate the value of b
1780 (eg. is equal to 1) and load a before b (eg. tmp = a[1]; if (b != 1)
1781 tmp = a[b]; ). There is also the problem of a compiler reloading b after
1782 having loaded a[b], thus having a newer copy of b than a[b]. A consensus
1783 has not yet been reached about these problems, however the READ_ONCE()
1784 macro is a good place to start looking.
1786 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1787 systems because it is assumed that a CPU will appear to be self-consistent,
1788 and will order overlapping accesses correctly with respect to itself.
1789 However, see the subsection on "Virtual Machine Guests" below.
1791 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1792 references to shared memory on SMP systems, though the use of locking instead
1795 Mandatory barriers should not be used to control SMP effects, since mandatory
1796 barriers impose unnecessary overhead on both SMP and UP systems. They may,
1797 however, be used to control MMIO effects on accesses through relaxed memory I/O
1798 windows. These barriers are required even on non-SMP systems as they affect
1799 the order in which memory operations appear to a device by prohibiting both the
1800 compiler and the CPU from reordering them.
1803 There are some more advanced barrier functions:
1805 (*) smp_store_mb(var, value)
1807 This assigns the value to the variable and then inserts a full memory
1808 barrier after it. It isn't guaranteed to insert anything more than a
1809 compiler barrier in a UP compilation.
1812 (*) smp_mb__before_atomic();
1813 (*) smp_mb__after_atomic();
1815 These are for use with atomic (such as add, subtract, increment and
1816 decrement) functions that don't return a value, especially when used for
1817 reference counting. These functions do not imply memory barriers.
1819 These are also used for atomic bitop functions that do not return a
1820 value (such as set_bit and clear_bit).
1822 As an example, consider a piece of code that marks an object as being dead
1823 and then decrements the object's reference count:
1826 smp_mb__before_atomic();
1827 atomic_dec(&obj->ref_count);
1829 This makes sure that the death mark on the object is perceived to be set
1830 *before* the reference counter is decremented.
1832 See Documentation/atomic_ops.txt for more information. See the "Atomic
1833 operations" subsection for information on where to use these.
1836 (*) lockless_dereference();
1838 This can be thought of as a pointer-fetch wrapper around the
1839 smp_read_barrier_depends() data-dependency barrier.
1841 This is also similar to rcu_dereference(), but in cases where
1842 object lifetime is handled by some mechanism other than RCU, for
1843 example, when the objects removed only when the system goes down.
1844 In addition, lockless_dereference() is used in some data structures
1845 that can be used both with and without RCU.
1851 These are for use with consistent memory to guarantee the ordering
1852 of writes or reads of shared memory accessible to both the CPU and a
1855 For example, consider a device driver that shares memory with a device
1856 and uses a descriptor status value to indicate if the descriptor belongs
1857 to the device or the CPU, and a doorbell to notify it when new
1858 descriptors are available:
1860 if (desc->status != DEVICE_OWN) {
1861 /* do not read data until we own descriptor */
1864 /* read/modify data */
1865 read_data = desc->data;
1866 desc->data = write_data;
1868 /* flush modifications before status update */
1871 /* assign ownership */
1872 desc->status = DEVICE_OWN;
1874 /* force memory to sync before notifying device via MMIO */
1877 /* notify device of new descriptors */
1878 writel(DESC_NOTIFY, doorbell);
1881 The dma_rmb() allows us guarantee the device has released ownership
1882 before we read the data from the descriptor, and the dma_wmb() allows
1883 us to guarantee the data is written to the descriptor before the device
1884 can see it now has ownership. The wmb() is needed to guarantee that the
1885 cache coherent memory writes have completed before attempting a write to
1886 the cache incoherent MMIO region.
1888 See Documentation/DMA-API.txt for more information on consistent memory.
1893 The Linux kernel also has a special barrier for use with memory-mapped I/O
1898 This is a variation on the mandatory write barrier that causes writes to weakly
1899 ordered I/O regions to be partially ordered. Its effects may go beyond the
1900 CPU->Hardware interface and actually affect the hardware at some level.
1902 See the subsection "Acquires vs I/O accesses" for more information.
1905 ===============================
1906 IMPLICIT KERNEL MEMORY BARRIERS
1907 ===============================
1909 Some of the other functions in the linux kernel imply memory barriers, amongst
1910 which are locking and scheduling functions.
1912 This specification is a _minimum_ guarantee; any particular architecture may
1913 provide more substantial guarantees, but these may not be relied upon outside
1914 of arch specific code.
1917 LOCK ACQUISITION FUNCTIONS
1918 --------------------------
1920 The Linux kernel has a number of locking constructs:
1928 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1929 for each construct. These operations all imply certain barriers:
1931 (1) ACQUIRE operation implication:
1933 Memory operations issued after the ACQUIRE will be completed after the
1934 ACQUIRE operation has completed.
1936 Memory operations issued before the ACQUIRE may be completed after
1937 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1938 combined with a following ACQUIRE, orders prior stores against
1939 subsequent loads and stores. Note that this is weaker than smp_mb()!
1940 The smp_mb__before_spinlock() primitive is free on many architectures.
1942 (2) RELEASE operation implication:
1944 Memory operations issued before the RELEASE will be completed before the
1945 RELEASE operation has completed.
1947 Memory operations issued after the RELEASE may be completed before the
1948 RELEASE operation has completed.
1950 (3) ACQUIRE vs ACQUIRE implication:
1952 All ACQUIRE operations issued before another ACQUIRE operation will be
1953 completed before that ACQUIRE operation.
1955 (4) ACQUIRE vs RELEASE implication:
1957 All ACQUIRE operations issued before a RELEASE operation will be
1958 completed before the RELEASE operation.
1960 (5) Failed conditional ACQUIRE implication:
1962 Certain locking variants of the ACQUIRE operation may fail, either due to
1963 being unable to get the lock immediately, or due to receiving an unblocked
1964 signal whilst asleep waiting for the lock to become available. Failed
1965 locks do not imply any sort of barrier.
1967 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1968 one-way barriers is that the effects of instructions outside of a critical
1969 section may seep into the inside of the critical section.
1971 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1972 because it is possible for an access preceding the ACQUIRE to happen after the
1973 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
1974 the two accesses can themselves then cross:
1983 ACQUIRE M, STORE *B, STORE *A, RELEASE M
1985 When the ACQUIRE and RELEASE are a lock acquisition and release,
1986 respectively, this same reordering can occur if the lock's ACQUIRE and
1987 RELEASE are to the same lock variable, but only from the perspective of
1988 another CPU not holding that lock. In short, a ACQUIRE followed by an
1989 RELEASE may -not- be assumed to be a full memory barrier.
1991 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does
1992 not imply a full memory barrier. Therefore, the CPU's execution of the
1993 critical sections corresponding to the RELEASE and the ACQUIRE can cross,
2003 ACQUIRE N, STORE *B, STORE *A, RELEASE M
2005 It might appear that this reordering could introduce a deadlock.
2006 However, this cannot happen because if such a deadlock threatened,
2007 the RELEASE would simply complete, thereby avoiding the deadlock.
2011 One key point is that we are only talking about the CPU doing
2012 the reordering, not the compiler. If the compiler (or, for
2013 that matter, the developer) switched the operations, deadlock
2016 But suppose the CPU reordered the operations. In this case,
2017 the unlock precedes the lock in the assembly code. The CPU
2018 simply elected to try executing the later lock operation first.
2019 If there is a deadlock, this lock operation will simply spin (or
2020 try to sleep, but more on that later). The CPU will eventually
2021 execute the unlock operation (which preceded the lock operation
2022 in the assembly code), which will unravel the potential deadlock,
2023 allowing the lock operation to succeed.
2025 But what if the lock is a sleeplock? In that case, the code will
2026 try to enter the scheduler, where it will eventually encounter
2027 a memory barrier, which will force the earlier unlock operation
2028 to complete, again unraveling the deadlock. There might be
2029 a sleep-unlock race, but the locking primitive needs to resolve
2030 such races properly in any case.
2032 Locks and semaphores may not provide any guarantee of ordering on UP compiled
2033 systems, and so cannot be counted on in such a situation to actually achieve
2034 anything at all - especially with respect to I/O accesses - unless combined
2035 with interrupt disabling operations.
2037 See also the section on "Inter-CPU locking barrier effects".
2040 As an example, consider the following:
2051 The following sequence of events is acceptable:
2053 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
2055 [+] Note that {*F,*A} indicates a combined access.
2057 But none of the following are:
2059 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
2060 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
2061 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
2062 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
2066 INTERRUPT DISABLING FUNCTIONS
2067 -----------------------------
2069 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
2070 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
2071 barriers are required in such a situation, they must be provided from some
2075 SLEEP AND WAKE-UP FUNCTIONS
2076 ---------------------------
2078 Sleeping and waking on an event flagged in global data can be viewed as an
2079 interaction between two pieces of data: the task state of the task waiting for
2080 the event and the global data used to indicate the event. To make sure that
2081 these appear to happen in the right order, the primitives to begin the process
2082 of going to sleep, and the primitives to initiate a wake up imply certain
2085 Firstly, the sleeper normally follows something like this sequence of events:
2088 set_current_state(TASK_UNINTERRUPTIBLE);
2089 if (event_indicated)
2094 A general memory barrier is interpolated automatically by set_current_state()
2095 after it has altered the task state:
2098 ===============================
2099 set_current_state();
2101 STORE current->state
2103 LOAD event_indicated
2105 set_current_state() may be wrapped by:
2108 prepare_to_wait_exclusive();
2110 which therefore also imply a general memory barrier after setting the state.
2111 The whole sequence above is available in various canned forms, all of which
2112 interpolate the memory barrier in the right place:
2115 wait_event_interruptible();
2116 wait_event_interruptible_exclusive();
2117 wait_event_interruptible_timeout();
2118 wait_event_killable();
2119 wait_event_timeout();
2124 Secondly, code that performs a wake up normally follows something like this:
2126 event_indicated = 1;
2127 wake_up(&event_wait_queue);
2131 event_indicated = 1;
2132 wake_up_process(event_daemon);
2134 A write memory barrier is implied by wake_up() and co. if and only if they
2135 wake something up. The barrier occurs before the task state is cleared, and so
2136 sits between the STORE to indicate the event and the STORE to set TASK_RUNNING:
2139 =============================== ===============================
2140 set_current_state(); STORE event_indicated
2141 smp_store_mb(); wake_up();
2142 STORE current->state <write barrier>
2143 <general barrier> STORE current->state
2144 LOAD event_indicated
2146 To repeat, this write memory barrier is present if and only if something
2147 is actually awakened. To see this, consider the following sequence of
2148 events, where X and Y are both initially zero:
2151 =============================== ===============================
2152 X = 1; STORE event_indicated
2153 smp_mb(); wake_up();
2154 Y = 1; wait_event(wq, Y == 1);
2155 wake_up(); load from Y sees 1, no memory barrier
2156 load from X might see 0
2158 In contrast, if a wakeup does occur, CPU 2's load from X would be guaranteed
2161 The available waker functions include:
2167 wake_up_interruptible();
2168 wake_up_interruptible_all();
2169 wake_up_interruptible_nr();
2170 wake_up_interruptible_poll();
2171 wake_up_interruptible_sync();
2172 wake_up_interruptible_sync_poll();
2174 wake_up_locked_poll();
2180 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2181 order multiple stores before the wake-up with respect to loads of those stored
2182 values after the sleeper has called set_current_state(). For instance, if the
2185 set_current_state(TASK_INTERRUPTIBLE);
2186 if (event_indicated)
2188 __set_current_state(TASK_RUNNING);
2189 do_something(my_data);
2194 event_indicated = 1;
2195 wake_up(&event_wait_queue);
2197 there's no guarantee that the change to event_indicated will be perceived by
2198 the sleeper as coming after the change to my_data. In such a circumstance, the
2199 code on both sides must interpolate its own memory barriers between the
2200 separate data accesses. Thus the above sleeper ought to do:
2202 set_current_state(TASK_INTERRUPTIBLE);
2203 if (event_indicated) {
2205 do_something(my_data);
2208 and the waker should do:
2212 event_indicated = 1;
2213 wake_up(&event_wait_queue);
2216 MISCELLANEOUS FUNCTIONS
2217 -----------------------
2219 Other functions that imply barriers:
2221 (*) schedule() and similar imply full memory barriers.
2224 ===================================
2225 INTER-CPU ACQUIRING BARRIER EFFECTS
2226 ===================================
2228 On SMP systems locking primitives give a more substantial form of barrier: one
2229 that does affect memory access ordering on other CPUs, within the context of
2230 conflict on any particular lock.
2233 ACQUIRES VS MEMORY ACCESSES
2234 ---------------------------
2236 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2237 three CPUs; then should the following sequence of events occur:
2240 =============================== ===============================
2241 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e);
2243 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f);
2244 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g);
2246 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h);
2248 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2249 through *H occur in, other than the constraints imposed by the separate locks
2250 on the separate CPUs. It might, for example, see:
2252 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2254 But it won't see any of:
2256 *B, *C or *D preceding ACQUIRE M
2257 *A, *B or *C following RELEASE M
2258 *F, *G or *H preceding ACQUIRE Q
2259 *E, *F or *G following RELEASE Q
2263 ACQUIRES VS I/O ACCESSES
2264 ------------------------
2266 Under certain circumstances (especially involving NUMA), I/O accesses within
2267 two spinlocked sections on two different CPUs may be seen as interleaved by the
2268 PCI bridge, because the PCI bridge does not necessarily participate in the
2269 cache-coherence protocol, and is therefore incapable of issuing the required
2270 read memory barriers.
2275 =============================== ===============================
2285 may be seen by the PCI bridge as follows:
2287 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2289 which would probably cause the hardware to malfunction.
2292 What is necessary here is to intervene with an mmiowb() before dropping the
2293 spinlock, for example:
2296 =============================== ===============================
2308 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2309 before either of the stores issued on CPU 2.
2312 Furthermore, following a store by a load from the same device obviates the need
2313 for the mmiowb(), because the load forces the store to complete before the load
2317 =============================== ===============================
2328 See Documentation/DocBook/deviceiobook.tmpl for more information.
2331 =================================
2332 WHERE ARE MEMORY BARRIERS NEEDED?
2333 =================================
2335 Under normal operation, memory operation reordering is generally not going to
2336 be a problem as a single-threaded linear piece of code will still appear to
2337 work correctly, even if it's in an SMP kernel. There are, however, four
2338 circumstances in which reordering definitely _could_ be a problem:
2340 (*) Interprocessor interaction.
2342 (*) Atomic operations.
2344 (*) Accessing devices.
2349 INTERPROCESSOR INTERACTION
2350 --------------------------
2352 When there's a system with more than one processor, more than one CPU in the
2353 system may be working on the same data set at the same time. This can cause
2354 synchronisation problems, and the usual way of dealing with them is to use
2355 locks. Locks, however, are quite expensive, and so it may be preferable to
2356 operate without the use of a lock if at all possible. In such a case
2357 operations that affect both CPUs may have to be carefully ordered to prevent
2360 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2361 queued on the semaphore, by virtue of it having a piece of its stack linked to
2362 the semaphore's list of waiting processes:
2364 struct rw_semaphore {
2367 struct list_head waiters;
2370 struct rwsem_waiter {
2371 struct list_head list;
2372 struct task_struct *task;
2375 To wake up a particular waiter, the up_read() or up_write() functions have to:
2377 (1) read the next pointer from this waiter's record to know as to where the
2378 next waiter record is;
2380 (2) read the pointer to the waiter's task structure;
2382 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2384 (4) call wake_up_process() on the task; and
2386 (5) release the reference held on the waiter's task struct.
2388 In other words, it has to perform this sequence of events:
2390 LOAD waiter->list.next;
2396 and if any of these steps occur out of order, then the whole thing may
2399 Once it has queued itself and dropped the semaphore lock, the waiter does not
2400 get the lock again; it instead just waits for its task pointer to be cleared
2401 before proceeding. Since the record is on the waiter's stack, this means that
2402 if the task pointer is cleared _before_ the next pointer in the list is read,
2403 another CPU might start processing the waiter and might clobber the waiter's
2404 stack before the up*() function has a chance to read the next pointer.
2406 Consider then what might happen to the above sequence of events:
2409 =============================== ===============================
2416 Woken up by other event
2421 foo() clobbers *waiter
2423 LOAD waiter->list.next;
2426 This could be dealt with using the semaphore lock, but then the down_xxx()
2427 function has to needlessly get the spinlock again after being woken up.
2429 The way to deal with this is to insert a general SMP memory barrier:
2431 LOAD waiter->list.next;
2438 In this case, the barrier makes a guarantee that all memory accesses before the
2439 barrier will appear to happen before all the memory accesses after the barrier
2440 with respect to the other CPUs on the system. It does _not_ guarantee that all
2441 the memory accesses before the barrier will be complete by the time the barrier
2442 instruction itself is complete.
2444 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2445 compiler barrier, thus making sure the compiler emits the instructions in the
2446 right order without actually intervening in the CPU. Since there's only one
2447 CPU, that CPU's dependency ordering logic will take care of everything else.
2453 Whilst they are technically interprocessor interaction considerations, atomic
2454 operations are noted specially as some of them imply full memory barriers and
2455 some don't, but they're very heavily relied on as a group throughout the
2458 Any atomic operation that modifies some state in memory and returns information
2459 about the state (old or new) implies an SMP-conditional general memory barrier
2460 (smp_mb()) on each side of the actual operation (with the exception of
2461 explicit lock operations, described later). These include:
2464 atomic_xchg(); atomic_long_xchg();
2465 atomic_inc_return(); atomic_long_inc_return();
2466 atomic_dec_return(); atomic_long_dec_return();
2467 atomic_add_return(); atomic_long_add_return();
2468 atomic_sub_return(); atomic_long_sub_return();
2469 atomic_inc_and_test(); atomic_long_inc_and_test();
2470 atomic_dec_and_test(); atomic_long_dec_and_test();
2471 atomic_sub_and_test(); atomic_long_sub_and_test();
2472 atomic_add_negative(); atomic_long_add_negative();
2474 test_and_clear_bit();
2475 test_and_change_bit();
2479 atomic_cmpxchg(); atomic_long_cmpxchg();
2480 atomic_add_unless(); atomic_long_add_unless();
2482 These are used for such things as implementing ACQUIRE-class and RELEASE-class
2483 operations and adjusting reference counters towards object destruction, and as
2484 such the implicit memory barrier effects are necessary.
2487 The following operations are potential problems as they do _not_ imply memory
2488 barriers, but might be used for implementing such things as RELEASE-class
2496 With these the appropriate explicit memory barrier should be used if necessary
2497 (smp_mb__before_atomic() for instance).
2500 The following also do _not_ imply memory barriers, and so may require explicit
2501 memory barriers under some circumstances (smp_mb__before_atomic() for
2509 If they're used for statistics generation, then they probably don't need memory
2510 barriers, unless there's a coupling between statistical data.
2512 If they're used for reference counting on an object to control its lifetime,
2513 they probably don't need memory barriers because either the reference count
2514 will be adjusted inside a locked section, or the caller will already hold
2515 sufficient references to make the lock, and thus a memory barrier unnecessary.
2517 If they're used for constructing a lock of some description, then they probably
2518 do need memory barriers as a lock primitive generally has to do things in a
2521 Basically, each usage case has to be carefully considered as to whether memory
2522 barriers are needed or not.
2524 The following operations are special locking primitives:
2526 test_and_set_bit_lock();
2528 __clear_bit_unlock();
2530 These implement ACQUIRE-class and RELEASE-class operations. These should be
2531 used in preference to other operations when implementing locking primitives,
2532 because their implementations can be optimised on many architectures.
2534 [!] Note that special memory barrier primitives are available for these
2535 situations because on some CPUs the atomic instructions used imply full memory
2536 barriers, and so barrier instructions are superfluous in conjunction with them,
2537 and in such cases the special barrier primitives will be no-ops.
2539 See Documentation/atomic_ops.txt for more information.
2545 Many devices can be memory mapped, and so appear to the CPU as if they're just
2546 a set of memory locations. To control such a device, the driver usually has to
2547 make the right memory accesses in exactly the right order.
2549 However, having a clever CPU or a clever compiler creates a potential problem
2550 in that the carefully sequenced accesses in the driver code won't reach the
2551 device in the requisite order if the CPU or the compiler thinks it is more
2552 efficient to reorder, combine or merge accesses - something that would cause
2553 the device to malfunction.
2555 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2556 routines - such as inb() or writel() - which know how to make such accesses
2557 appropriately sequential. Whilst this, for the most part, renders the explicit
2558 use of memory barriers unnecessary, there are a couple of situations where they
2561 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2562 so for _all_ general drivers locks should be used and mmiowb() must be
2563 issued prior to unlocking the critical section.
2565 (2) If the accessor functions are used to refer to an I/O memory window with
2566 relaxed memory access properties, then _mandatory_ memory barriers are
2567 required to enforce ordering.
2569 See Documentation/DocBook/deviceiobook.tmpl for more information.
2575 A driver may be interrupted by its own interrupt service routine, and thus the
2576 two parts of the driver may interfere with each other's attempts to control or
2579 This may be alleviated - at least in part - by disabling local interrupts (a
2580 form of locking), such that the critical operations are all contained within
2581 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2582 routine is executing, the driver's core may not run on the same CPU, and its
2583 interrupt is not permitted to happen again until the current interrupt has been
2584 handled, thus the interrupt handler does not need to lock against that.
2586 However, consider a driver that was talking to an ethernet card that sports an
2587 address register and a data register. If that driver's core talks to the card
2588 under interrupt-disablement and then the driver's interrupt handler is invoked:
2599 The store to the data register might happen after the second store to the
2600 address register if ordering rules are sufficiently relaxed:
2602 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2605 If ordering rules are relaxed, it must be assumed that accesses done inside an
2606 interrupt disabled section may leak outside of it and may interleave with
2607 accesses performed in an interrupt - and vice versa - unless implicit or
2608 explicit barriers are used.
2610 Normally this won't be a problem because the I/O accesses done inside such
2611 sections will include synchronous load operations on strictly ordered I/O
2612 registers that form implicit I/O barriers. If this isn't sufficient then an
2613 mmiowb() may need to be used explicitly.
2616 A similar situation may occur between an interrupt routine and two routines
2617 running on separate CPUs that communicate with each other. If such a case is
2618 likely, then interrupt-disabling locks should be used to guarantee ordering.
2621 ==========================
2622 KERNEL I/O BARRIER EFFECTS
2623 ==========================
2625 When accessing I/O memory, drivers should use the appropriate accessor
2630 These are intended to talk to I/O space rather than memory space, but
2631 that's primarily a CPU-specific concept. The i386 and x86_64 processors
2632 do indeed have special I/O space access cycles and instructions, but many
2633 CPUs don't have such a concept.
2635 The PCI bus, amongst others, defines an I/O space concept which - on such
2636 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2637 space. However, it may also be mapped as a virtual I/O space in the CPU's
2638 memory map, particularly on those CPUs that don't support alternate I/O
2641 Accesses to this space may be fully synchronous (as on i386), but
2642 intermediary bridges (such as the PCI host bridge) may not fully honour
2645 They are guaranteed to be fully ordered with respect to each other.
2647 They are not guaranteed to be fully ordered with respect to other types of
2648 memory and I/O operation.
2650 (*) readX(), writeX():
2652 Whether these are guaranteed to be fully ordered and uncombined with
2653 respect to each other on the issuing CPU depends on the characteristics
2654 defined for the memory window through which they're accessing. On later
2655 i386 architecture machines, for example, this is controlled by way of the
2658 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2659 provided they're not accessing a prefetchable device.
2661 However, intermediary hardware (such as a PCI bridge) may indulge in
2662 deferral if it so wishes; to flush a store, a load from the same location
2663 is preferred[*], but a load from the same device or from configuration
2664 space should suffice for PCI.
2666 [*] NOTE! attempting to load from the same location as was written to may
2667 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2670 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2671 force stores to be ordered.
2673 Please refer to the PCI specification for more information on interactions
2674 between PCI transactions.
2676 (*) readX_relaxed(), writeX_relaxed()
2678 These are similar to readX() and writeX(), but provide weaker memory
2679 ordering guarantees. Specifically, they do not guarantee ordering with
2680 respect to normal memory accesses (e.g. DMA buffers) nor do they guarantee
2681 ordering with respect to LOCK or UNLOCK operations. If the latter is
2682 required, an mmiowb() barrier can be used. Note that relaxed accesses to
2683 the same peripheral are guaranteed to be ordered with respect to each
2686 (*) ioreadX(), iowriteX()
2688 These will perform appropriately for the type of access they're actually
2689 doing, be it inX()/outX() or readX()/writeX().
2692 ========================================
2693 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2694 ========================================
2696 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2697 maintain the appearance of program causality with respect to itself. Some CPUs
2698 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2699 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2700 of arch-specific code.
2702 This means that it must be considered that the CPU will execute its instruction
2703 stream in any order it feels like - or even in parallel - provided that if an
2704 instruction in the stream depends on an earlier instruction, then that
2705 earlier instruction must be sufficiently complete[*] before the later
2706 instruction may proceed; in other words: provided that the appearance of
2707 causality is maintained.
2709 [*] Some instructions have more than one effect - such as changing the
2710 condition codes, changing registers or changing memory - and different
2711 instructions may depend on different effects.
2713 A CPU may also discard any instruction sequence that winds up having no
2714 ultimate effect. For example, if two adjacent instructions both load an
2715 immediate value into the same register, the first may be discarded.
2718 Similarly, it has to be assumed that compiler might reorder the instruction
2719 stream in any way it sees fit, again provided the appearance of causality is
2723 ============================
2724 THE EFFECTS OF THE CPU CACHE
2725 ============================
2727 The way cached memory operations are perceived across the system is affected to
2728 a certain extent by the caches that lie between CPUs and memory, and by the
2729 memory coherence system that maintains the consistency of state in the system.
2731 As far as the way a CPU interacts with another part of the system through the
2732 caches goes, the memory system has to include the CPU's caches, and memory
2733 barriers for the most part act at the interface between the CPU and its cache
2734 (memory barriers logically act on the dotted line in the following diagram):
2736 <--- CPU ---> : <----------- Memory ----------->
2738 +--------+ +--------+ : +--------+ +-----------+
2739 | | | | : | | | | +--------+
2740 | CPU | | Memory | : | CPU | | | | |
2741 | Core |--->| Access |----->| Cache |<-->| | | |
2742 | | | Queue | : | | | |--->| Memory |
2743 | | | | : | | | | | |
2744 +--------+ +--------+ : +--------+ | | | |
2745 : | Cache | +--------+
2747 : | Mechanism | +--------+
2748 +--------+ +--------+ : +--------+ | | | |
2749 | | | | : | | | | | |
2750 | CPU | | Memory | : | CPU | | |--->| Device |
2751 | Core |--->| Access |----->| Cache |<-->| | | |
2752 | | | Queue | : | | | | | |
2753 | | | | : | | | | +--------+
2754 +--------+ +--------+ : +--------+ +-----------+
2758 Although any particular load or store may not actually appear outside of the
2759 CPU that issued it since it may have been satisfied within the CPU's own cache,
2760 it will still appear as if the full memory access had taken place as far as the
2761 other CPUs are concerned since the cache coherency mechanisms will migrate the
2762 cacheline over to the accessing CPU and propagate the effects upon conflict.
2764 The CPU core may execute instructions in any order it deems fit, provided the
2765 expected program causality appears to be maintained. Some of the instructions
2766 generate load and store operations which then go into the queue of memory
2767 accesses to be performed. The core may place these in the queue in any order
2768 it wishes, and continue execution until it is forced to wait for an instruction
2771 What memory barriers are concerned with is controlling the order in which
2772 accesses cross from the CPU side of things to the memory side of things, and
2773 the order in which the effects are perceived to happen by the other observers
2776 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2777 their own loads and stores as if they had happened in program order.
2779 [!] MMIO or other device accesses may bypass the cache system. This depends on
2780 the properties of the memory window through which devices are accessed and/or
2781 the use of any special device communication instructions the CPU may have.
2787 Life isn't quite as simple as it may appear above, however: for while the
2788 caches are expected to be coherent, there's no guarantee that that coherency
2789 will be ordered. This means that whilst changes made on one CPU will
2790 eventually become visible on all CPUs, there's no guarantee that they will
2791 become apparent in the same order on those other CPUs.
2794 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2795 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2800 +--------+ : +--->| Cache A |<------->| |
2801 | | : | +---------+ | |
2803 | | : | +---------+ | |
2804 +--------+ : +--->| Cache B |<------->| |
2807 : +---------+ | System |
2808 +--------+ : +--->| Cache C |<------->| |
2809 | | : | +---------+ | |
2811 | | : | +---------+ | |
2812 +--------+ : +--->| Cache D |<------->| |
2817 Imagine the system has the following properties:
2819 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2822 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2825 (*) whilst the CPU core is interrogating one cache, the other cache may be
2826 making use of the bus to access the rest of the system - perhaps to
2827 displace a dirty cacheline or to do a speculative load;
2829 (*) each cache has a queue of operations that need to be applied to that cache
2830 to maintain coherency with the rest of the system;
2832 (*) the coherency queue is not flushed by normal loads to lines already
2833 present in the cache, even though the contents of the queue may
2834 potentially affect those loads.
2836 Imagine, then, that two writes are made on the first CPU, with a write barrier
2837 between them to guarantee that they will appear to reach that CPU's caches in
2838 the requisite order:
2841 =============== =============== =======================================
2842 u == 0, v == 1 and p == &u, q == &u
2844 smp_wmb(); Make sure change to v is visible before
2846 <A:modify v=2> v is now in cache A exclusively
2848 <B:modify p=&v> p is now in cache B exclusively
2850 The write memory barrier forces the other CPUs in the system to perceive that
2851 the local CPU's caches have apparently been updated in the correct order. But
2852 now imagine that the second CPU wants to read those values:
2855 =============== =============== =======================================
2860 The above pair of reads may then fail to happen in the expected order, as the
2861 cacheline holding p may get updated in one of the second CPU's caches whilst
2862 the update to the cacheline holding v is delayed in the other of the second
2863 CPU's caches by some other cache event:
2866 =============== =============== =======================================
2867 u == 0, v == 1 and p == &u, q == &u
2870 <A:modify v=2> <C:busy>
2874 <B:modify p=&v> <D:commit p=&v>
2877 <C:read *q> Reads from v before v updated in cache
2881 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2882 no guarantee that, without intervention, the order of update will be the same
2883 as that committed on CPU 1.
2886 To intervene, we need to interpolate a data dependency barrier or a read
2887 barrier between the loads. This will force the cache to commit its coherency
2888 queue before processing any further requests:
2891 =============== =============== =======================================
2892 u == 0, v == 1 and p == &u, q == &u
2895 <A:modify v=2> <C:busy>
2899 <B:modify p=&v> <D:commit p=&v>
2901 smp_read_barrier_depends()
2905 <C:read *q> Reads from v after v updated in cache
2908 This sort of problem can be encountered on DEC Alpha processors as they have a
2909 split cache that improves performance by making better use of the data bus.
2910 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2911 access depends on a read, not all do, so it may not be relied on.
2913 Other CPUs may also have split caches, but must coordinate between the various
2914 cachelets for normal memory accesses. The semantics of the Alpha removes the
2915 need for coordination in the absence of memory barriers.
2918 CACHE COHERENCY VS DMA
2919 ----------------------
2921 Not all systems maintain cache coherency with respect to devices doing DMA. In
2922 such cases, a device attempting DMA may obtain stale data from RAM because
2923 dirty cache lines may be resident in the caches of various CPUs, and may not
2924 have been written back to RAM yet. To deal with this, the appropriate part of
2925 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2926 invalidate them as well).
2928 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2929 cache lines being written back to RAM from a CPU's cache after the device has
2930 installed its own data, or cache lines present in the CPU's cache may simply
2931 obscure the fact that RAM has been updated, until at such time as the cacheline
2932 is discarded from the CPU's cache and reloaded. To deal with this, the
2933 appropriate part of the kernel must invalidate the overlapping bits of the
2936 See Documentation/cachetlb.txt for more information on cache management.
2939 CACHE COHERENCY VS MMIO
2940 -----------------------
2942 Memory mapped I/O usually takes place through memory locations that are part of
2943 a window in the CPU's memory space that has different properties assigned than
2944 the usual RAM directed window.
2946 Amongst these properties is usually the fact that such accesses bypass the
2947 caching entirely and go directly to the device buses. This means MMIO accesses
2948 may, in effect, overtake accesses to cached memory that were emitted earlier.
2949 A memory barrier isn't sufficient in such a case, but rather the cache must be
2950 flushed between the cached memory write and the MMIO access if the two are in
2954 =========================
2955 THE THINGS CPUS GET UP TO
2956 =========================
2958 A programmer might take it for granted that the CPU will perform memory
2959 operations in exactly the order specified, so that if the CPU is, for example,
2960 given the following piece of code to execute:
2968 they would then expect that the CPU will complete the memory operation for each
2969 instruction before moving on to the next one, leading to a definite sequence of
2970 operations as seen by external observers in the system:
2972 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2975 Reality is, of course, much messier. With many CPUs and compilers, the above
2976 assumption doesn't hold because:
2978 (*) loads are more likely to need to be completed immediately to permit
2979 execution progress, whereas stores can often be deferred without a
2982 (*) loads may be done speculatively, and the result discarded should it prove
2983 to have been unnecessary;
2985 (*) loads may be done speculatively, leading to the result having been fetched
2986 at the wrong time in the expected sequence of events;
2988 (*) the order of the memory accesses may be rearranged to promote better use
2989 of the CPU buses and caches;
2991 (*) loads and stores may be combined to improve performance when talking to
2992 memory or I/O hardware that can do batched accesses of adjacent locations,
2993 thus cutting down on transaction setup costs (memory and PCI devices may
2994 both be able to do this); and
2996 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2997 mechanisms may alleviate this - once the store has actually hit the cache
2998 - there's no guarantee that the coherency management will be propagated in
2999 order to other CPUs.
3001 So what another CPU, say, might actually observe from the above piece of code
3004 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
3006 (Where "LOAD {*C,*D}" is a combined load)
3009 However, it is guaranteed that a CPU will be self-consistent: it will see its
3010 _own_ accesses appear to be correctly ordered, without the need for a memory
3011 barrier. For instance with the following code:
3020 and assuming no intervention by an external influence, it can be assumed that
3021 the final result will appear to be:
3023 U == the original value of *A
3028 The code above may cause the CPU to generate the full sequence of memory
3031 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
3033 in that order, but, without intervention, the sequence may have almost any
3034 combination of elements combined or discarded, provided the program's view
3035 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE()
3036 are -not- optional in the above example, as there are architectures
3037 where a given CPU might reorder successive loads to the same location.
3038 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is
3039 necessary to prevent this, for example, on Itanium the volatile casts
3040 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq
3041 and st.rel instructions (respectively) that prevent such reordering.
3043 The compiler may also combine, discard or defer elements of the sequence before
3044 the CPU even sees them.
3055 since, without either a write barrier or an WRITE_ONCE(), it can be
3056 assumed that the effect of the storage of V to *A is lost. Similarly:
3061 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be
3067 and the LOAD operation never appear outside of the CPU.
3070 AND THEN THERE'S THE ALPHA
3071 --------------------------
3073 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
3074 some versions of the Alpha CPU have a split data cache, permitting them to have
3075 two semantically-related cache lines updated at separate times. This is where
3076 the data dependency barrier really becomes necessary as this synchronises both
3077 caches with the memory coherence system, thus making it seem like pointer
3078 changes vs new data occur in the right order.
3080 The Alpha defines the Linux kernel's memory barrier model.
3082 See the subsection on "Cache Coherency" above.
3085 VIRTUAL MACHINE GUESTS
3086 ----------------------
3088 Guests running within virtual machines might be affected by SMP effects even if
3089 the guest itself is compiled without SMP support. This is an artifact of
3090 interfacing with an SMP host while running an UP kernel. Using mandatory
3091 barriers for this use-case would be possible but is often suboptimal.
3093 To handle this case optimally, low-level virt_mb() etc macros are available.
3094 These have the same effect as smp_mb() etc when SMP is enabled, but generate
3095 identical code for SMP and non-SMP systems. For example, virtual machine guests
3096 should use virt_mb() rather than smp_mb() when synchronizing against a
3097 (possibly SMP) host.
3099 These are equivalent to smp_mb() etc counterparts in all other respects,
3100 in particular, they do not control MMIO effects: to control
3101 MMIO effects, use mandatory barriers.
3111 Memory barriers can be used to implement circular buffering without the need
3112 of a lock to serialise the producer with the consumer. See:
3114 Documentation/circular-buffers.txt
3123 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
3125 Chapter 5.2: Physical Address Space Characteristics
3126 Chapter 5.4: Caches and Write Buffers
3127 Chapter 5.5: Data Sharing
3128 Chapter 5.6: Read/Write Ordering
3130 AMD64 Architecture Programmer's Manual Volume 2: System Programming
3131 Chapter 7.1: Memory-Access Ordering
3132 Chapter 7.4: Buffering and Combining Memory Writes
3134 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
3135 System Programming Guide
3136 Chapter 7.1: Locked Atomic Operations
3137 Chapter 7.2: Memory Ordering
3138 Chapter 7.4: Serializing Instructions
3140 The SPARC Architecture Manual, Version 9
3141 Chapter 8: Memory Models
3142 Appendix D: Formal Specification of the Memory Models
3143 Appendix J: Programming with the Memory Models
3145 UltraSPARC Programmer Reference Manual
3146 Chapter 5: Memory Accesses and Cacheability
3147 Chapter 15: Sparc-V9 Memory Models
3149 UltraSPARC III Cu User's Manual
3150 Chapter 9: Memory Models
3152 UltraSPARC IIIi Processor User's Manual
3153 Chapter 8: Memory Models
3155 UltraSPARC Architecture 2005
3157 Appendix D: Formal Specifications of the Memory Models
3159 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
3160 Chapter 8: Memory Models
3161 Appendix F: Caches and Cache Coherency
3163 Solaris Internals, Core Kernel Architecture, p63-68:
3164 Chapter 3.3: Hardware Considerations for Locks and
3167 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3168 for Kernel Programmers:
3169 Chapter 13: Other Memory Models
3171 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3172 Section 2.6: Speculation
3173 Section 4.4: Memory Access
projects / cascardo / linux.git / blob