1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
5 By: David Howells <dhowells@redhat.com>
6 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
10 (*) Abstract memory access model.
15 (*) What are memory barriers?
17 - Varieties of memory barrier.
18 - What may not be assumed about memory barriers?
19 - Data dependency barriers.
20 - Control dependencies.
21 - SMP barrier pairing.
22 - Examples of memory barrier sequences.
23 - Read memory barriers vs load speculation.
26 (*) Explicit kernel barriers.
29 - CPU memory barriers.
32 (*) Implicit kernel memory barriers.
35 - Interrupt disabling functions.
36 - Sleep and wake-up functions.
37 - Miscellaneous functions.
39 (*) Inter-CPU locking barrier effects.
41 - Locks vs memory accesses.
42 - Locks vs I/O accesses.
44 (*) Where are memory barriers needed?
46 - Interprocessor interaction.
51 (*) Kernel I/O barrier effects.
53 (*) Assumed minimum execution ordering model.
55 (*) The effects of the cpu cache.
58 - Cache coherency vs DMA.
59 - Cache coherency vs MMIO.
61 (*) The things CPUs get up to.
63 - And then there's the Alpha.
72 ============================
73 ABSTRACT MEMORY ACCESS MODEL
74 ============================
76 Consider the following abstract model of the system:
81 +-------+ : +--------+ : +-------+
84 | CPU 1 |<----->| Memory |<----->| CPU 2 |
87 +-------+ : +--------+ : +-------+
95 +---------->| Device |<----------+
101 Each CPU executes a program that generates memory access operations. In the
102 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
103 perform the memory operations in any order it likes, provided program causality
104 appears to be maintained. Similarly, the compiler may also arrange the
105 instructions it emits in any order it likes, provided it doesn't affect the
106 apparent operation of the program.
108 So in the above diagram, the effects of the memory operations performed by a
109 CPU are perceived by the rest of the system as the operations cross the
110 interface between the CPU and rest of the system (the dotted lines).
113 For example, consider the following sequence of events:
116 =============== ===============
121 The set of accesses as seen by the memory system in the middle can be arranged
122 in 24 different combinations:
124 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4
125 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3
126 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4
127 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4
128 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3
129 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4
130 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4
134 and can thus result in four different combinations of values:
142 Furthermore, the stores committed by a CPU to the memory system may not be
143 perceived by the loads made by another CPU in the same order as the stores were
147 As a further example, consider this sequence of events:
150 =============== ===============
151 { A == 1, B == 2, C = 3, P == &A, Q == &C }
155 There is an obvious data dependency here, as the value loaded into D depends on
156 the address retrieved from P by CPU 2. At the end of the sequence, any of the
157 following results are possible:
159 (Q == &A) and (D == 1)
160 (Q == &B) and (D == 2)
161 (Q == &B) and (D == 4)
163 Note that CPU 2 will never try and load C into D because the CPU will load P
164 into Q before issuing the load of *Q.
170 Some devices present their control interfaces as collections of memory
171 locations, but the order in which the control registers are accessed is very
172 important. For instance, imagine an ethernet card with a set of internal
173 registers that are accessed through an address port register (A) and a data
174 port register (D). To read internal register 5, the following code might then
180 but this might show up as either of the following two sequences:
182 STORE *A = 5, x = LOAD *D
183 x = LOAD *D, STORE *A = 5
185 the second of which will almost certainly result in a malfunction, since it set
186 the address _after_ attempting to read the register.
192 There are some minimal guarantees that may be expected of a CPU:
194 (*) On any given CPU, dependent memory accesses will be issued in order, with
195 respect to itself. This means that for:
197 Q = READ_ONCE(P); smp_read_barrier_depends(); D = READ_ONCE(*Q);
199 the CPU will issue the following memory operations:
201 Q = LOAD P, D = LOAD *Q
203 and always in that order. On most systems, smp_read_barrier_depends()
204 does nothing, but it is required for DEC Alpha. The READ_ONCE()
205 is required to prevent compiler mischief. Please note that you
206 should normally use something like rcu_dereference() instead of
207 open-coding smp_read_barrier_depends().
209 (*) Overlapping loads and stores within a particular CPU will appear to be
210 ordered within that CPU. This means that for:
212 a = READ_ONCE(*X); WRITE_ONCE(*X, b);
214 the CPU will only issue the following sequence of memory operations:
216 a = LOAD *X, STORE *X = b
220 WRITE_ONCE(*X, c); d = READ_ONCE(*X);
222 the CPU will only issue:
224 STORE *X = c, d = LOAD *X
226 (Loads and stores overlap if they are targeted at overlapping pieces of
229 And there are a number of things that _must_ or _must_not_ be assumed:
231 (*) It _must_not_ be assumed that the compiler will do what you want
232 with memory references that are not protected by READ_ONCE() and
233 WRITE_ONCE(). Without them, the compiler is within its rights to
234 do all sorts of "creative" transformations, which are covered in
235 the COMPILER BARRIER section.
237 (*) It _must_not_ be assumed that independent loads and stores will be issued
238 in the order given. This means that for:
240 X = *A; Y = *B; *D = Z;
242 we may get any of the following sequences:
244 X = LOAD *A, Y = LOAD *B, STORE *D = Z
245 X = LOAD *A, STORE *D = Z, Y = LOAD *B
246 Y = LOAD *B, X = LOAD *A, STORE *D = Z
247 Y = LOAD *B, STORE *D = Z, X = LOAD *A
248 STORE *D = Z, X = LOAD *A, Y = LOAD *B
249 STORE *D = Z, Y = LOAD *B, X = LOAD *A
251 (*) It _must_ be assumed that overlapping memory accesses may be merged or
252 discarded. This means that for:
254 X = *A; Y = *(A + 4);
256 we may get any one of the following sequences:
258 X = LOAD *A; Y = LOAD *(A + 4);
259 Y = LOAD *(A + 4); X = LOAD *A;
260 {X, Y} = LOAD {*A, *(A + 4) };
264 *A = X; *(A + 4) = Y;
268 STORE *A = X; STORE *(A + 4) = Y;
269 STORE *(A + 4) = Y; STORE *A = X;
270 STORE {*A, *(A + 4) } = {X, Y};
272 And there are anti-guarantees:
274 (*) These guarantees do not apply to bitfields, because compilers often
275 generate code to modify these using non-atomic read-modify-write
276 sequences. Do not attempt to use bitfields to synchronize parallel
279 (*) Even in cases where bitfields are protected by locks, all fields
280 in a given bitfield must be protected by one lock. If two fields
281 in a given bitfield are protected by different locks, the compiler's
282 non-atomic read-modify-write sequences can cause an update to one
283 field to corrupt the value of an adjacent field.
285 (*) These guarantees apply only to properly aligned and sized scalar
286 variables. "Properly sized" currently means variables that are
287 the same size as "char", "short", "int" and "long". "Properly
288 aligned" means the natural alignment, thus no constraints for
289 "char", two-byte alignment for "short", four-byte alignment for
290 "int", and either four-byte or eight-byte alignment for "long",
291 on 32-bit and 64-bit systems, respectively. Note that these
292 guarantees were introduced into the C11 standard, so beware when
293 using older pre-C11 compilers (for example, gcc 4.6). The portion
294 of the standard containing this guarantee is Section 3.14, which
295 defines "memory location" as follows:
298 either an object of scalar type, or a maximal sequence
299 of adjacent bit-fields all having nonzero width
301 NOTE 1: Two threads of execution can update and access
302 separate memory locations without interfering with
305 NOTE 2: A bit-field and an adjacent non-bit-field member
306 are in separate memory locations. The same applies
307 to two bit-fields, if one is declared inside a nested
308 structure declaration and the other is not, or if the two
309 are separated by a zero-length bit-field declaration,
310 or if they are separated by a non-bit-field member
311 declaration. It is not safe to concurrently update two
312 bit-fields in the same structure if all members declared
313 between them are also bit-fields, no matter what the
314 sizes of those intervening bit-fields happen to be.
317 =========================
318 WHAT ARE MEMORY BARRIERS?
319 =========================
321 As can be seen above, independent memory operations are effectively performed
322 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
323 What is required is some way of intervening to instruct the compiler and the
324 CPU to restrict the order.
326 Memory barriers are such interventions. They impose a perceived partial
327 ordering over the memory operations on either side of the barrier.
329 Such enforcement is important because the CPUs and other devices in a system
330 can use a variety of tricks to improve performance, including reordering,
331 deferral and combination of memory operations; speculative loads; speculative
332 branch prediction and various types of caching. Memory barriers are used to
333 override or suppress these tricks, allowing the code to sanely control the
334 interaction of multiple CPUs and/or devices.
337 VARIETIES OF MEMORY BARRIER
338 ---------------------------
340 Memory barriers come in four basic varieties:
342 (1) Write (or store) memory barriers.
344 A write memory barrier gives a guarantee that all the STORE operations
345 specified before the barrier will appear to happen before all the STORE
346 operations specified after the barrier with respect to the other
347 components of the system.
349 A write barrier is a partial ordering on stores only; it is not required
350 to have any effect on loads.
352 A CPU can be viewed as committing a sequence of store operations to the
353 memory system as time progresses. All stores before a write barrier will
354 occur in the sequence _before_ all the stores after the write barrier.
356 [!] Note that write barriers should normally be paired with read or data
357 dependency barriers; see the "SMP barrier pairing" subsection.
360 (2) Data dependency barriers.
362 A data dependency barrier is a weaker form of read barrier. In the case
363 where two loads are performed such that the second depends on the result
364 of the first (eg: the first load retrieves the address to which the second
365 load will be directed), a data dependency barrier would be required to
366 make sure that the target of the second load is updated before the address
367 obtained by the first load is accessed.
369 A data dependency barrier is a partial ordering on interdependent loads
370 only; it is not required to have any effect on stores, independent loads
371 or overlapping loads.
373 As mentioned in (1), the other CPUs in the system can be viewed as
374 committing sequences of stores to the memory system that the CPU being
375 considered can then perceive. A data dependency barrier issued by the CPU
376 under consideration guarantees that for any load preceding it, if that
377 load touches one of a sequence of stores from another CPU, then by the
378 time the barrier completes, the effects of all the stores prior to that
379 touched by the load will be perceptible to any loads issued after the data
382 See the "Examples of memory barrier sequences" subsection for diagrams
383 showing the ordering constraints.
385 [!] Note that the first load really has to have a _data_ dependency and
386 not a control dependency. If the address for the second load is dependent
387 on the first load, but the dependency is through a conditional rather than
388 actually loading the address itself, then it's a _control_ dependency and
389 a full read barrier or better is required. See the "Control dependencies"
390 subsection for more information.
392 [!] Note that data dependency barriers should normally be paired with
393 write barriers; see the "SMP barrier pairing" subsection.
396 (3) Read (or load) memory barriers.
398 A read barrier is a data dependency barrier plus a guarantee that all the
399 LOAD operations specified before the barrier will appear to happen before
400 all the LOAD operations specified after the barrier with respect to the
401 other components of the system.
403 A read barrier is a partial ordering on loads only; it is not required to
404 have any effect on stores.
406 Read memory barriers imply data dependency barriers, and so can substitute
409 [!] Note that read barriers should normally be paired with write barriers;
410 see the "SMP barrier pairing" subsection.
413 (4) General memory barriers.
415 A general memory barrier gives a guarantee that all the LOAD and STORE
416 operations specified before the barrier will appear to happen before all
417 the LOAD and STORE operations specified after the barrier with respect to
418 the other components of the system.
420 A general memory barrier is a partial ordering over both loads and stores.
422 General memory barriers imply both read and write memory barriers, and so
423 can substitute for either.
426 And a couple of implicit varieties:
428 (5) ACQUIRE operations.
430 This acts as a one-way permeable barrier. It guarantees that all memory
431 operations after the ACQUIRE operation will appear to happen after the
432 ACQUIRE operation with respect to the other components of the system.
433 ACQUIRE operations include LOCK operations and smp_load_acquire()
436 Memory operations that occur before an ACQUIRE operation may appear to
437 happen after it completes.
439 An ACQUIRE operation should almost always be paired with a RELEASE
443 (6) RELEASE operations.
445 This also acts as a one-way permeable barrier. It guarantees that all
446 memory operations before the RELEASE operation will appear to happen
447 before the RELEASE operation with respect to the other components of the
448 system. RELEASE operations include UNLOCK operations and
449 smp_store_release() operations.
451 Memory operations that occur after a RELEASE operation may appear to
452 happen before it completes.
454 The use of ACQUIRE and RELEASE operations generally precludes the need
455 for other sorts of memory barrier (but note the exceptions mentioned in
456 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
457 pair is -not- guaranteed to act as a full memory barrier. However, after
458 an ACQUIRE on a given variable, all memory accesses preceding any prior
459 RELEASE on that same variable are guaranteed to be visible. In other
460 words, within a given variable's critical section, all accesses of all
461 previous critical sections for that variable are guaranteed to have
464 This means that ACQUIRE acts as a minimal "acquire" operation and
465 RELEASE acts as a minimal "release" operation.
468 Memory barriers are only required where there's a possibility of interaction
469 between two CPUs or between a CPU and a device. If it can be guaranteed that
470 there won't be any such interaction in any particular piece of code, then
471 memory barriers are unnecessary in that piece of code.
474 Note that these are the _minimum_ guarantees. Different architectures may give
475 more substantial guarantees, but they may _not_ be relied upon outside of arch
479 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
480 ----------------------------------------------
482 There are certain things that the Linux kernel memory barriers do not guarantee:
484 (*) There is no guarantee that any of the memory accesses specified before a
485 memory barrier will be _complete_ by the completion of a memory barrier
486 instruction; the barrier can be considered to draw a line in that CPU's
487 access queue that accesses of the appropriate type may not cross.
489 (*) There is no guarantee that issuing a memory barrier on one CPU will have
490 any direct effect on another CPU or any other hardware in the system. The
491 indirect effect will be the order in which the second CPU sees the effects
492 of the first CPU's accesses occur, but see the next point:
494 (*) There is no guarantee that a CPU will see the correct order of effects
495 from a second CPU's accesses, even _if_ the second CPU uses a memory
496 barrier, unless the first CPU _also_ uses a matching memory barrier (see
497 the subsection on "SMP Barrier Pairing").
499 (*) There is no guarantee that some intervening piece of off-the-CPU
500 hardware[*] will not reorder the memory accesses. CPU cache coherency
501 mechanisms should propagate the indirect effects of a memory barrier
502 between CPUs, but might not do so in order.
504 [*] For information on bus mastering DMA and coherency please read:
506 Documentation/PCI/pci.txt
507 Documentation/DMA-API-HOWTO.txt
508 Documentation/DMA-API.txt
511 DATA DEPENDENCY BARRIERS
512 ------------------------
514 The usage requirements of data dependency barriers are a little subtle, and
515 it's not always obvious that they're needed. To illustrate, consider the
516 following sequence of events:
519 =============== ===============
520 { A == 1, B == 2, C = 3, P == &A, Q == &C }
527 There's a clear data dependency here, and it would seem that by the end of the
528 sequence, Q must be either &A or &B, and that:
530 (Q == &A) implies (D == 1)
531 (Q == &B) implies (D == 4)
533 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
534 leading to the following situation:
536 (Q == &B) and (D == 2) ????
538 Whilst this may seem like a failure of coherency or causality maintenance, it
539 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
542 To deal with this, a data dependency barrier or better must be inserted
543 between the address load and the data load:
546 =============== ===============
547 { A == 1, B == 2, C = 3, P == &A, Q == &C }
552 <data dependency barrier>
555 This enforces the occurrence of one of the two implications, and prevents the
556 third possibility from arising.
558 [!] Note that this extremely counterintuitive situation arises most easily on
559 machines with split caches, so that, for example, one cache bank processes
560 even-numbered cache lines and the other bank processes odd-numbered cache
561 lines. The pointer P might be stored in an odd-numbered cache line, and the
562 variable B might be stored in an even-numbered cache line. Then, if the
563 even-numbered bank of the reading CPU's cache is extremely busy while the
564 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
565 but the old value of the variable B (2).
568 The data dependency barrier is very important to the RCU system,
569 for example. See rcu_assign_pointer() and rcu_dereference() in
570 include/linux/rcupdate.h. This permits the current target of an RCU'd
571 pointer to be replaced with a new modified target, without the replacement
572 target appearing to be incompletely initialised.
574 See also the subsection on "Cache Coherency" for a more thorough example.
580 A load-load control dependency requires a full read memory barrier, not
581 simply a data dependency barrier to make it work correctly. Consider the
582 following bit of code:
586 <data dependency barrier> /* BUG: No data dependency!!! */
590 This will not have the desired effect because there is no actual data
591 dependency, but rather a control dependency that the CPU may short-circuit
592 by attempting to predict the outcome in advance, so that other CPUs see
593 the load from b as having happened before the load from a. In such a
594 case what's actually required is:
602 However, stores are not speculated. This means that ordering -is- provided
603 for load-store control dependencies, as in the following example:
610 Control dependencies pair normally with other types of barriers. That
611 said, please note that READ_ONCE() is not optional! Without the
612 READ_ONCE(), the compiler might combine the load from 'a' with other
613 loads from 'a', and the store to 'b' with other stores to 'b', with
614 possible highly counterintuitive effects on ordering.
616 Worse yet, if the compiler is able to prove (say) that the value of
617 variable 'a' is always non-zero, it would be well within its rights
618 to optimize the original example by eliminating the "if" statement
622 b = p; /* BUG: Compiler and CPU can both reorder!!! */
624 So don't leave out the READ_ONCE().
626 It is tempting to try to enforce ordering on identical stores on both
627 branches of the "if" statement as follows:
640 Unfortunately, current compilers will transform this as follows at high
645 WRITE_ONCE(b, p); /* BUG: No ordering vs. load from a!!! */
647 /* WRITE_ONCE(b, p); -- moved up, BUG!!! */
650 /* WRITE_ONCE(b, p); -- moved up, BUG!!! */
654 Now there is no conditional between the load from 'a' and the store to
655 'b', which means that the CPU is within its rights to reorder them:
656 The conditional is absolutely required, and must be present in the
657 assembly code even after all compiler optimizations have been applied.
658 Therefore, if you need ordering in this example, you need explicit
659 memory barriers, for example, smp_store_release():
663 smp_store_release(&b, p);
666 smp_store_release(&b, p);
670 In contrast, without explicit memory barriers, two-legged-if control
671 ordering is guaranteed only when the stores differ, for example:
682 The initial READ_ONCE() is still required to prevent the compiler from
683 proving the value of 'a'.
685 In addition, you need to be careful what you do with the local variable 'q',
686 otherwise the compiler might be able to guess the value and again remove
687 the needed conditional. For example:
698 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
699 equal to zero, in which case the compiler is within its rights to
700 transform the above code into the following:
706 Given this transformation, the CPU is not required to respect the ordering
707 between the load from variable 'a' and the store to variable 'b'. It is
708 tempting to add a barrier(), but this does not help. The conditional
709 is gone, and the barrier won't bring it back. Therefore, if you are
710 relying on this ordering, you should make sure that MAX is greater than
711 one, perhaps as follows:
714 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
723 Please note once again that the stores to 'b' differ. If they were
724 identical, as noted earlier, the compiler could pull this store outside
725 of the 'if' statement.
727 You must also be careful not to rely too much on boolean short-circuit
728 evaluation. Consider this example:
734 Because the first condition cannot fault and the second condition is
735 always true, the compiler can transform this example as following,
736 defeating control dependency:
741 This example underscores the need to ensure that the compiler cannot
742 out-guess your code. More generally, although READ_ONCE() does force
743 the compiler to actually emit code for a given load, it does not force
744 the compiler to use the results.
746 Finally, control dependencies do -not- provide transitivity. This is
747 demonstrated by two related examples, with the initial values of
748 x and y both being zero:
751 ======================= =======================
752 r1 = READ_ONCE(x); r2 = READ_ONCE(y);
753 if (r1 > 0) if (r2 > 0)
754 WRITE_ONCE(y, 1); WRITE_ONCE(x, 1);
756 assert(!(r1 == 1 && r2 == 1));
758 The above two-CPU example will never trigger the assert(). However,
759 if control dependencies guaranteed transitivity (which they do not),
760 then adding the following CPU would guarantee a related assertion:
763 =====================
766 assert(!(r1 == 2 && r2 == 1 && x == 2)); /* FAILS!!! */
768 But because control dependencies do -not- provide transitivity, the above
769 assertion can fail after the combined three-CPU example completes. If you
770 need the three-CPU example to provide ordering, you will need smp_mb()
771 between the loads and stores in the CPU 0 and CPU 1 code fragments,
772 that is, just before or just after the "if" statements. Furthermore,
773 the original two-CPU example is very fragile and should be avoided.
775 These two examples are the LB and WWC litmus tests from this paper:
776 http://www.cl.cam.ac.uk/users/pes20/ppc-supplemental/test6.pdf and this
777 site: https://www.cl.cam.ac.uk/~pes20/ppcmem/index.html.
781 (*) Control dependencies can order prior loads against later stores.
782 However, they do -not- guarantee any other sort of ordering:
783 Not prior loads against later loads, nor prior stores against
784 later anything. If you need these other forms of ordering,
785 use smp_rmb(), smp_wmb(), or, in the case of prior stores and
786 later loads, smp_mb().
788 (*) If both legs of the "if" statement begin with identical stores to
789 the same variable, then those stores must be ordered, either by
790 preceding both of them with smp_mb() or by using smp_store_release()
791 to carry out the stores. Please note that it is -not- sufficient
792 to use barrier() at beginning of each leg of the "if" statement,
793 as optimizing compilers do not necessarily respect barrier()
796 (*) Control dependencies require at least one run-time conditional
797 between the prior load and the subsequent store, and this
798 conditional must involve the prior load. If the compiler is able
799 to optimize the conditional away, it will have also optimized
800 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE()
801 can help to preserve the needed conditional.
803 (*) Control dependencies require that the compiler avoid reordering the
804 dependency into nonexistence. Careful use of READ_ONCE() or
805 atomic{,64}_read() can help to preserve your control dependency.
806 Please see the COMPILER BARRIER section for more information.
808 (*) Control dependencies pair normally with other types of barriers.
810 (*) Control dependencies do -not- provide transitivity. If you
811 need transitivity, use smp_mb().
817 When dealing with CPU-CPU interactions, certain types of memory barrier should
818 always be paired. A lack of appropriate pairing is almost certainly an error.
820 General barriers pair with each other, though they also pair with most
821 other types of barriers, albeit without transitivity. An acquire barrier
822 pairs with a release barrier, but both may also pair with other barriers,
823 including of course general barriers. A write barrier pairs with a data
824 dependency barrier, a control dependency, an acquire barrier, a release
825 barrier, a read barrier, or a general barrier. Similarly a read barrier,
826 control dependency, or a data dependency barrier pairs with a write
827 barrier, an acquire barrier, a release barrier, or a general barrier:
830 =============== ===============
833 WRITE_ONCE(b, 2); x = READ_ONCE(b);
840 =============== ===============================
843 WRITE_ONCE(b, &a); x = READ_ONCE(b);
844 <data dependency barrier>
850 =============== ===============================
853 WRITE_ONCE(y, 1); if (r2 = READ_ONCE(x)) {
854 <implicit control dependency>
858 assert(r1 == 0 || r2 == 0);
860 Basically, the read barrier always has to be there, even though it can be of
863 [!] Note that the stores before the write barrier would normally be expected to
864 match the loads after the read barrier or the data dependency barrier, and vice
868 =================== ===================
869 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c);
870 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d);
871 <write barrier> \ <read barrier>
872 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a);
873 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b);
876 EXAMPLES OF MEMORY BARRIER SEQUENCES
877 ------------------------------------
879 Firstly, write barriers act as partial orderings on store operations.
880 Consider the following sequence of events:
883 =======================
891 This sequence of events is committed to the memory coherence system in an order
892 that the rest of the system might perceive as the unordered set of { STORE A,
893 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
898 | |------>| C=3 | } /\
899 | | : +------+ }----- \ -----> Events perceptible to
900 | | : | A=1 | } \/ the rest of the system
902 | CPU 1 | : | B=2 | }
904 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
905 | | +------+ } requires all stores prior to the
906 | | : | E=5 | } barrier to be committed before
907 | | : +------+ } further stores may take place
912 | Sequence in which stores are committed to the
913 | memory system by CPU 1
917 Secondly, data dependency barriers act as partial orderings on data-dependent
918 loads. Consider the following sequence of events:
921 ======================= =======================
922 { B = 7; X = 9; Y = 8; C = &Y }
927 STORE D = 4 LOAD C (gets &B)
930 Without intervention, CPU 2 may perceive the events on CPU 1 in some
931 effectively random order, despite the write barrier issued by CPU 1:
934 | | +------+ +-------+ | Sequence of update
935 | |------>| B=2 |----- --->| Y->8 | | of perception on
936 | | : +------+ \ +-------+ | CPU 2
937 | CPU 1 | : | A=1 | \ --->| C->&Y | V
938 | | +------+ | +-------+
939 | | wwwwwwwwwwwwwwww | : :
941 | | : | C=&B |--- | : : +-------+
942 | | : +------+ \ | +-------+ | |
943 | |------>| D=4 | ----------->| C->&B |------>| |
944 | | +------+ | +-------+ | |
945 +-------+ : : | : : | |
949 Apparently incorrect ---> | | B->7 |------>| |
950 perception of B (!) | +-------+ | |
953 The load of X holds ---> \ | X->9 |------>| |
954 up the maintenance \ +-------+ | |
955 of coherence of B ----->| B->2 | +-------+
960 In the above example, CPU 2 perceives that B is 7, despite the load of *C
961 (which would be B) coming after the LOAD of C.
963 If, however, a data dependency barrier were to be placed between the load of C
964 and the load of *C (ie: B) on CPU 2:
967 ======================= =======================
968 { B = 7; X = 9; Y = 8; C = &Y }
973 STORE D = 4 LOAD C (gets &B)
974 <data dependency barrier>
977 then the following will occur:
980 | | +------+ +-------+
981 | |------>| B=2 |----- --->| Y->8 |
982 | | : +------+ \ +-------+
983 | CPU 1 | : | A=1 | \ --->| C->&Y |
984 | | +------+ | +-------+
985 | | wwwwwwwwwwwwwwww | : :
987 | | : | C=&B |--- | : : +-------+
988 | | : +------+ \ | +-------+ | |
989 | |------>| D=4 | ----------->| C->&B |------>| |
990 | | +------+ | +-------+ | |
991 +-------+ : : | : : | |
997 Makes sure all effects ---> \ ddddddddddddddddd | |
998 prior to the store of C \ +-------+ | |
999 are perceptible to ----->| B->2 |------>| |
1000 subsequent loads +-------+ | |
1004 And thirdly, a read barrier acts as a partial order on loads. Consider the
1005 following sequence of events:
1008 ======================= =======================
1016 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
1017 some effectively random order, despite the write barrier issued by CPU 1:
1020 | | +------+ +-------+
1021 | |------>| A=1 |------ --->| A->0 |
1022 | | +------+ \ +-------+
1023 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1024 | | +------+ | +-------+
1025 | |------>| B=2 |--- | : :
1026 | | +------+ \ | : : +-------+
1027 +-------+ : : \ | +-------+ | |
1028 ---------->| B->2 |------>| |
1029 | +-------+ | CPU 2 |
1030 | | A->0 |------>| |
1040 If, however, a read barrier were to be placed between the load of B and the
1044 ======================= =======================
1053 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
1057 | | +------+ +-------+
1058 | |------>| A=1 |------ --->| A->0 |
1059 | | +------+ \ +-------+
1060 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1061 | | +------+ | +-------+
1062 | |------>| B=2 |--- | : :
1063 | | +------+ \ | : : +-------+
1064 +-------+ : : \ | +-------+ | |
1065 ---------->| B->2 |------>| |
1066 | +-------+ | CPU 2 |
1069 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1070 barrier causes all effects \ +-------+ | |
1071 prior to the storage of B ---->| A->1 |------>| |
1072 to be perceptible to CPU 2 +-------+ | |
1076 To illustrate this more completely, consider what could happen if the code
1077 contained a load of A either side of the read barrier:
1080 ======================= =======================
1086 LOAD A [first load of A]
1088 LOAD A [second load of A]
1090 Even though the two loads of A both occur after the load of B, they may both
1091 come up with different values:
1094 | | +------+ +-------+
1095 | |------>| A=1 |------ --->| A->0 |
1096 | | +------+ \ +-------+
1097 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1098 | | +------+ | +-------+
1099 | |------>| B=2 |--- | : :
1100 | | +------+ \ | : : +-------+
1101 +-------+ : : \ | +-------+ | |
1102 ---------->| B->2 |------>| |
1103 | +-------+ | CPU 2 |
1107 | | A->0 |------>| 1st |
1109 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1110 barrier causes all effects \ +-------+ | |
1111 prior to the storage of B ---->| A->1 |------>| 2nd |
1112 to be perceptible to CPU 2 +-------+ | |
1116 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1117 before the read barrier completes anyway:
1120 | | +------+ +-------+
1121 | |------>| A=1 |------ --->| A->0 |
1122 | | +------+ \ +-------+
1123 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1124 | | +------+ | +-------+
1125 | |------>| B=2 |--- | : :
1126 | | +------+ \ | : : +-------+
1127 +-------+ : : \ | +-------+ | |
1128 ---------->| B->2 |------>| |
1129 | +-------+ | CPU 2 |
1133 ---->| A->1 |------>| 1st |
1135 rrrrrrrrrrrrrrrrr | |
1137 | A->1 |------>| 2nd |
1142 The guarantee is that the second load will always come up with A == 1 if the
1143 load of B came up with B == 2. No such guarantee exists for the first load of
1144 A; that may come up with either A == 0 or A == 1.
1147 READ MEMORY BARRIERS VS LOAD SPECULATION
1148 ----------------------------------------
1150 Many CPUs speculate with loads: that is they see that they will need to load an
1151 item from memory, and they find a time where they're not using the bus for any
1152 other loads, and so do the load in advance - even though they haven't actually
1153 got to that point in the instruction execution flow yet. This permits the
1154 actual load instruction to potentially complete immediately because the CPU
1155 already has the value to hand.
1157 It may turn out that the CPU didn't actually need the value - perhaps because a
1158 branch circumvented the load - in which case it can discard the value or just
1159 cache it for later use.
1164 ======================= =======================
1166 DIVIDE } Divide instructions generally
1167 DIVIDE } take a long time to perform
1170 Which might appear as this:
1174 --->| B->2 |------>| |
1178 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1179 division speculates on the +-------+ ~ | |
1183 Once the divisions are complete --> : : ~-->| |
1184 the CPU can then perform the : : | |
1185 LOAD with immediate effect : : +-------+
1188 Placing a read barrier or a data dependency barrier just before the second
1192 ======================= =======================
1199 will force any value speculatively obtained to be reconsidered to an extent
1200 dependent on the type of barrier used. If there was no change made to the
1201 speculated memory location, then the speculated value will just be used:
1205 --->| B->2 |------>| |
1209 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1210 division speculates on the +-------+ ~ | |
1215 rrrrrrrrrrrrrrrr~ | |
1222 but if there was an update or an invalidation from another CPU pending, then
1223 the speculation will be cancelled and the value reloaded:
1227 --->| B->2 |------>| |
1231 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1232 division speculates on the +-------+ ~ | |
1237 rrrrrrrrrrrrrrrrr | |
1239 The speculation is discarded ---> --->| A->1 |------>| |
1240 and an updated value is +-------+ | |
1241 retrieved : : +-------+
1247 Transitivity is a deeply intuitive notion about ordering that is not
1248 always provided by real computer systems. The following example
1249 demonstrates transitivity (also called "cumulativity"):
1252 ======================= ======================= =======================
1254 STORE X=1 LOAD X STORE Y=1
1255 <general barrier> <general barrier>
1258 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1259 This indicates that CPU 2's load from X in some sense follows CPU 1's
1260 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1261 store to Y. The question is then "Can CPU 3's load from X return 0?"
1263 Because CPU 2's load from X in some sense came after CPU 1's store, it
1264 is natural to expect that CPU 3's load from X must therefore return 1.
1265 This expectation is an example of transitivity: if a load executing on
1266 CPU A follows a load from the same variable executing on CPU B, then
1267 CPU A's load must either return the same value that CPU B's load did,
1268 or must return some later value.
1270 In the Linux kernel, use of general memory barriers guarantees
1271 transitivity. Therefore, in the above example, if CPU 2's load from X
1272 returns 1 and its load from Y returns 0, then CPU 3's load from X must
1275 However, transitivity is -not- guaranteed for read or write barriers.
1276 For example, suppose that CPU 2's general barrier in the above example
1277 is changed to a read barrier as shown below:
1280 ======================= ======================= =======================
1282 STORE X=1 LOAD X STORE Y=1
1283 <read barrier> <general barrier>
1286 This substitution destroys transitivity: in this example, it is perfectly
1287 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1288 and CPU 3's load from X to return 0.
1290 The key point is that although CPU 2's read barrier orders its pair
1291 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1292 this example runs on a system where CPUs 1 and 2 share a store buffer
1293 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1294 General barriers are therefore required to ensure that all CPUs agree
1295 on the combined order of CPU 1's and CPU 2's accesses.
1297 To reiterate, if your code requires transitivity, use general barriers
1301 ========================
1302 EXPLICIT KERNEL BARRIERS
1303 ========================
1305 The Linux kernel has a variety of different barriers that act at different
1308 (*) Compiler barrier.
1310 (*) CPU memory barriers.
1312 (*) MMIO write barrier.
1318 The Linux kernel has an explicit compiler barrier function that prevents the
1319 compiler from moving the memory accesses either side of it to the other side:
1323 This is a general barrier -- there are no read-read or write-write
1324 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be
1325 thought of as weak forms of barrier() that affect only the specific
1326 accesses flagged by the READ_ONCE() or WRITE_ONCE().
1328 The barrier() function has the following effects:
1330 (*) Prevents the compiler from reordering accesses following the
1331 barrier() to precede any accesses preceding the barrier().
1332 One example use for this property is to ease communication between
1333 interrupt-handler code and the code that was interrupted.
1335 (*) Within a loop, forces the compiler to load the variables used
1336 in that loop's conditional on each pass through that loop.
1338 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of
1339 optimizations that, while perfectly safe in single-threaded code, can
1340 be fatal in concurrent code. Here are some examples of these sorts
1343 (*) The compiler is within its rights to reorder loads and stores
1344 to the same variable, and in some cases, the CPU is within its
1345 rights to reorder loads to the same variable. This means that
1351 Might result in an older value of x stored in a[1] than in a[0].
1352 Prevent both the compiler and the CPU from doing this as follows:
1354 a[0] = READ_ONCE(x);
1355 a[1] = READ_ONCE(x);
1357 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for
1358 accesses from multiple CPUs to a single variable.
1360 (*) The compiler is within its rights to merge successive loads from
1361 the same variable. Such merging can cause the compiler to "optimize"
1365 do_something_with(tmp);
1367 into the following code, which, although in some sense legitimate
1368 for single-threaded code, is almost certainly not what the developer
1373 do_something_with(tmp);
1375 Use READ_ONCE() to prevent the compiler from doing this to you:
1377 while (tmp = READ_ONCE(a))
1378 do_something_with(tmp);
1380 (*) The compiler is within its rights to reload a variable, for example,
1381 in cases where high register pressure prevents the compiler from
1382 keeping all data of interest in registers. The compiler might
1383 therefore optimize the variable 'tmp' out of our previous example:
1386 do_something_with(tmp);
1388 This could result in the following code, which is perfectly safe in
1389 single-threaded code, but can be fatal in concurrent code:
1392 do_something_with(a);
1394 For example, the optimized version of this code could result in
1395 passing a zero to do_something_with() in the case where the variable
1396 a was modified by some other CPU between the "while" statement and
1397 the call to do_something_with().
1399 Again, use READ_ONCE() to prevent the compiler from doing this:
1401 while (tmp = READ_ONCE(a))
1402 do_something_with(tmp);
1404 Note that if the compiler runs short of registers, it might save
1405 tmp onto the stack. The overhead of this saving and later restoring
1406 is why compilers reload variables. Doing so is perfectly safe for
1407 single-threaded code, so you need to tell the compiler about cases
1408 where it is not safe.
1410 (*) The compiler is within its rights to omit a load entirely if it knows
1411 what the value will be. For example, if the compiler can prove that
1412 the value of variable 'a' is always zero, it can optimize this code:
1415 do_something_with(tmp);
1421 This transformation is a win for single-threaded code because it
1422 gets rid of a load and a branch. The problem is that the compiler
1423 will carry out its proof assuming that the current CPU is the only
1424 one updating variable 'a'. If variable 'a' is shared, then the
1425 compiler's proof will be erroneous. Use READ_ONCE() to tell the
1426 compiler that it doesn't know as much as it thinks it does:
1428 while (tmp = READ_ONCE(a))
1429 do_something_with(tmp);
1431 But please note that the compiler is also closely watching what you
1432 do with the value after the READ_ONCE(). For example, suppose you
1433 do the following and MAX is a preprocessor macro with the value 1:
1435 while ((tmp = READ_ONCE(a)) % MAX)
1436 do_something_with(tmp);
1438 Then the compiler knows that the result of the "%" operator applied
1439 to MAX will always be zero, again allowing the compiler to optimize
1440 the code into near-nonexistence. (It will still load from the
1443 (*) Similarly, the compiler is within its rights to omit a store entirely
1444 if it knows that the variable already has the value being stored.
1445 Again, the compiler assumes that the current CPU is the only one
1446 storing into the variable, which can cause the compiler to do the
1447 wrong thing for shared variables. For example, suppose you have
1451 /* Code that does not store to variable a. */
1454 The compiler sees that the value of variable 'a' is already zero, so
1455 it might well omit the second store. This would come as a fatal
1456 surprise if some other CPU might have stored to variable 'a' in the
1459 Use WRITE_ONCE() to prevent the compiler from making this sort of
1463 /* Code that does not store to variable a. */
1466 (*) The compiler is within its rights to reorder memory accesses unless
1467 you tell it not to. For example, consider the following interaction
1468 between process-level code and an interrupt handler:
1470 void process_level(void)
1472 msg = get_message();
1476 void interrupt_handler(void)
1479 process_message(msg);
1482 There is nothing to prevent the compiler from transforming
1483 process_level() to the following, in fact, this might well be a
1484 win for single-threaded code:
1486 void process_level(void)
1489 msg = get_message();
1492 If the interrupt occurs between these two statement, then
1493 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE()
1494 to prevent this as follows:
1496 void process_level(void)
1498 WRITE_ONCE(msg, get_message());
1499 WRITE_ONCE(flag, true);
1502 void interrupt_handler(void)
1504 if (READ_ONCE(flag))
1505 process_message(READ_ONCE(msg));
1508 Note that the READ_ONCE() and WRITE_ONCE() wrappers in
1509 interrupt_handler() are needed if this interrupt handler can itself
1510 be interrupted by something that also accesses 'flag' and 'msg',
1511 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE()
1512 and WRITE_ONCE() are not needed in interrupt_handler() other than
1513 for documentation purposes. (Note also that nested interrupts
1514 do not typically occur in modern Linux kernels, in fact, if an
1515 interrupt handler returns with interrupts enabled, you will get a
1518 You should assume that the compiler can move READ_ONCE() and
1519 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(),
1520 barrier(), or similar primitives.
1522 This effect could also be achieved using barrier(), but READ_ONCE()
1523 and WRITE_ONCE() are more selective: With READ_ONCE() and
1524 WRITE_ONCE(), the compiler need only forget the contents of the
1525 indicated memory locations, while with barrier() the compiler must
1526 discard the value of all memory locations that it has currented
1527 cached in any machine registers. Of course, the compiler must also
1528 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
1529 though the CPU of course need not do so.
1531 (*) The compiler is within its rights to invent stores to a variable,
1532 as in the following example:
1539 The compiler might save a branch by optimizing this as follows:
1545 In single-threaded code, this is not only safe, but also saves
1546 a branch. Unfortunately, in concurrent code, this optimization
1547 could cause some other CPU to see a spurious value of 42 -- even
1548 if variable 'a' was never zero -- when loading variable 'b'.
1549 Use WRITE_ONCE() to prevent this as follows:
1556 The compiler can also invent loads. These are usually less
1557 damaging, but they can result in cache-line bouncing and thus in
1558 poor performance and scalability. Use READ_ONCE() to prevent
1561 (*) For aligned memory locations whose size allows them to be accessed
1562 with a single memory-reference instruction, prevents "load tearing"
1563 and "store tearing," in which a single large access is replaced by
1564 multiple smaller accesses. For example, given an architecture having
1565 16-bit store instructions with 7-bit immediate fields, the compiler
1566 might be tempted to use two 16-bit store-immediate instructions to
1567 implement the following 32-bit store:
1571 Please note that GCC really does use this sort of optimization,
1572 which is not surprising given that it would likely take more
1573 than two instructions to build the constant and then store it.
1574 This optimization can therefore be a win in single-threaded code.
1575 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1576 this optimization in a volatile store. In the absence of such bugs,
1577 use of WRITE_ONCE() prevents store tearing in the following example:
1579 WRITE_ONCE(p, 0x00010002);
1581 Use of packed structures can also result in load and store tearing,
1584 struct __attribute__((__packed__)) foo {
1589 struct foo foo1, foo2;
1596 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no
1597 volatile markings, the compiler would be well within its rights to
1598 implement these three assignment statements as a pair of 32-bit
1599 loads followed by a pair of 32-bit stores. This would result in
1600 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE()
1601 and WRITE_ONCE() again prevent tearing in this example:
1604 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b));
1607 All that aside, it is never necessary to use READ_ONCE() and
1608 WRITE_ONCE() on a variable that has been marked volatile. For example,
1609 because 'jiffies' is marked volatile, it is never necessary to
1610 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and
1611 WRITE_ONCE() are implemented as volatile casts, which has no effect when
1612 its argument is already marked volatile.
1614 Please note that these compiler barriers have no direct effect on the CPU,
1615 which may then reorder things however it wishes.
1621 The Linux kernel has eight basic CPU memory barriers:
1623 TYPE MANDATORY SMP CONDITIONAL
1624 =============== ======================= ===========================
1625 GENERAL mb() smp_mb()
1626 WRITE wmb() smp_wmb()
1627 READ rmb() smp_rmb()
1628 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1631 All memory barriers except the data dependency barriers imply a compiler
1632 barrier. Data dependencies do not impose any additional compiler ordering.
1634 Aside: In the case of data dependencies, the compiler would be expected
1635 to issue the loads in the correct order (eg. `a[b]` would have to load
1636 the value of b before loading a[b]), however there is no guarantee in
1637 the C specification that the compiler may not speculate the value of b
1638 (eg. is equal to 1) and load a before b (eg. tmp = a[1]; if (b != 1)
1639 tmp = a[b]; ). There is also the problem of a compiler reloading b after
1640 having loaded a[b], thus having a newer copy of b than a[b]. A consensus
1641 has not yet been reached about these problems, however the READ_ONCE()
1642 macro is a good place to start looking.
1644 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1645 systems because it is assumed that a CPU will appear to be self-consistent,
1646 and will order overlapping accesses correctly with respect to itself.
1647 However, see the subsection on "Virtual Machine Guests" below.
1649 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1650 references to shared memory on SMP systems, though the use of locking instead
1653 Mandatory barriers should not be used to control SMP effects, since mandatory
1654 barriers impose unnecessary overhead on both SMP and UP systems. They may,
1655 however, be used to control MMIO effects on accesses through relaxed memory I/O
1656 windows. These barriers are required even on non-SMP systems as they affect
1657 the order in which memory operations appear to a device by prohibiting both the
1658 compiler and the CPU from reordering them.
1661 There are some more advanced barrier functions:
1663 (*) smp_store_mb(var, value)
1665 This assigns the value to the variable and then inserts a full memory
1666 barrier after it. It isn't guaranteed to insert anything more than a
1667 compiler barrier in a UP compilation.
1670 (*) smp_mb__before_atomic();
1671 (*) smp_mb__after_atomic();
1673 These are for use with atomic (such as add, subtract, increment and
1674 decrement) functions that don't return a value, especially when used for
1675 reference counting. These functions do not imply memory barriers.
1677 These are also used for atomic bitop functions that do not return a
1678 value (such as set_bit and clear_bit).
1680 As an example, consider a piece of code that marks an object as being dead
1681 and then decrements the object's reference count:
1684 smp_mb__before_atomic();
1685 atomic_dec(&obj->ref_count);
1687 This makes sure that the death mark on the object is perceived to be set
1688 *before* the reference counter is decremented.
1690 See Documentation/atomic_ops.txt for more information. See the "Atomic
1691 operations" subsection for information on where to use these.
1694 (*) lockless_dereference();
1695 This can be thought of as a pointer-fetch wrapper around the
1696 smp_read_barrier_depends() data-dependency barrier.
1698 This is also similar to rcu_dereference(), but in cases where
1699 object lifetime is handled by some mechanism other than RCU, for
1700 example, when the objects removed only when the system goes down.
1701 In addition, lockless_dereference() is used in some data structures
1702 that can be used both with and without RCU.
1708 These are for use with consistent memory to guarantee the ordering
1709 of writes or reads of shared memory accessible to both the CPU and a
1712 For example, consider a device driver that shares memory with a device
1713 and uses a descriptor status value to indicate if the descriptor belongs
1714 to the device or the CPU, and a doorbell to notify it when new
1715 descriptors are available:
1717 if (desc->status != DEVICE_OWN) {
1718 /* do not read data until we own descriptor */
1721 /* read/modify data */
1722 read_data = desc->data;
1723 desc->data = write_data;
1725 /* flush modifications before status update */
1728 /* assign ownership */
1729 desc->status = DEVICE_OWN;
1731 /* force memory to sync before notifying device via MMIO */
1734 /* notify device of new descriptors */
1735 writel(DESC_NOTIFY, doorbell);
1738 The dma_rmb() allows us guarantee the device has released ownership
1739 before we read the data from the descriptor, and the dma_wmb() allows
1740 us to guarantee the data is written to the descriptor before the device
1741 can see it now has ownership. The wmb() is needed to guarantee that the
1742 cache coherent memory writes have completed before attempting a write to
1743 the cache incoherent MMIO region.
1745 See Documentation/DMA-API.txt for more information on consistent memory.
1750 The Linux kernel also has a special barrier for use with memory-mapped I/O
1755 This is a variation on the mandatory write barrier that causes writes to weakly
1756 ordered I/O regions to be partially ordered. Its effects may go beyond the
1757 CPU->Hardware interface and actually affect the hardware at some level.
1759 See the subsection "Locks vs I/O accesses" for more information.
1762 ===============================
1763 IMPLICIT KERNEL MEMORY BARRIERS
1764 ===============================
1766 Some of the other functions in the linux kernel imply memory barriers, amongst
1767 which are locking and scheduling functions.
1769 This specification is a _minimum_ guarantee; any particular architecture may
1770 provide more substantial guarantees, but these may not be relied upon outside
1771 of arch specific code.
1777 The Linux kernel has a number of locking constructs:
1785 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1786 for each construct. These operations all imply certain barriers:
1788 (1) ACQUIRE operation implication:
1790 Memory operations issued after the ACQUIRE will be completed after the
1791 ACQUIRE operation has completed.
1793 Memory operations issued before the ACQUIRE may be completed after
1794 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1795 combined with a following ACQUIRE, orders prior stores against
1796 subsequent loads and stores. Note that this is weaker than smp_mb()!
1797 The smp_mb__before_spinlock() primitive is free on many architectures.
1799 (2) RELEASE operation implication:
1801 Memory operations issued before the RELEASE will be completed before the
1802 RELEASE operation has completed.
1804 Memory operations issued after the RELEASE may be completed before the
1805 RELEASE operation has completed.
1807 (3) ACQUIRE vs ACQUIRE implication:
1809 All ACQUIRE operations issued before another ACQUIRE operation will be
1810 completed before that ACQUIRE operation.
1812 (4) ACQUIRE vs RELEASE implication:
1814 All ACQUIRE operations issued before a RELEASE operation will be
1815 completed before the RELEASE operation.
1817 (5) Failed conditional ACQUIRE implication:
1819 Certain locking variants of the ACQUIRE operation may fail, either due to
1820 being unable to get the lock immediately, or due to receiving an unblocked
1821 signal whilst asleep waiting for the lock to become available. Failed
1822 locks do not imply any sort of barrier.
1824 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1825 one-way barriers is that the effects of instructions outside of a critical
1826 section may seep into the inside of the critical section.
1828 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1829 because it is possible for an access preceding the ACQUIRE to happen after the
1830 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
1831 the two accesses can themselves then cross:
1840 ACQUIRE M, STORE *B, STORE *A, RELEASE M
1842 When the ACQUIRE and RELEASE are a lock acquisition and release,
1843 respectively, this same reordering can occur if the lock's ACQUIRE and
1844 RELEASE are to the same lock variable, but only from the perspective of
1845 another CPU not holding that lock. In short, a ACQUIRE followed by an
1846 RELEASE may -not- be assumed to be a full memory barrier.
1848 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does
1849 not imply a full memory barrier. Therefore, the CPU's execution of the
1850 critical sections corresponding to the RELEASE and the ACQUIRE can cross,
1860 ACQUIRE N, STORE *B, STORE *A, RELEASE M
1862 It might appear that this reordering could introduce a deadlock.
1863 However, this cannot happen because if such a deadlock threatened,
1864 the RELEASE would simply complete, thereby avoiding the deadlock.
1868 One key point is that we are only talking about the CPU doing
1869 the reordering, not the compiler. If the compiler (or, for
1870 that matter, the developer) switched the operations, deadlock
1873 But suppose the CPU reordered the operations. In this case,
1874 the unlock precedes the lock in the assembly code. The CPU
1875 simply elected to try executing the later lock operation first.
1876 If there is a deadlock, this lock operation will simply spin (or
1877 try to sleep, but more on that later). The CPU will eventually
1878 execute the unlock operation (which preceded the lock operation
1879 in the assembly code), which will unravel the potential deadlock,
1880 allowing the lock operation to succeed.
1882 But what if the lock is a sleeplock? In that case, the code will
1883 try to enter the scheduler, where it will eventually encounter
1884 a memory barrier, which will force the earlier unlock operation
1885 to complete, again unraveling the deadlock. There might be
1886 a sleep-unlock race, but the locking primitive needs to resolve
1887 such races properly in any case.
1889 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1890 systems, and so cannot be counted on in such a situation to actually achieve
1891 anything at all - especially with respect to I/O accesses - unless combined
1892 with interrupt disabling operations.
1894 See also the section on "Inter-CPU locking barrier effects".
1897 As an example, consider the following:
1908 The following sequence of events is acceptable:
1910 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
1912 [+] Note that {*F,*A} indicates a combined access.
1914 But none of the following are:
1916 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
1917 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
1918 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
1919 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
1923 INTERRUPT DISABLING FUNCTIONS
1924 -----------------------------
1926 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
1927 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
1928 barriers are required in such a situation, they must be provided from some
1932 SLEEP AND WAKE-UP FUNCTIONS
1933 ---------------------------
1935 Sleeping and waking on an event flagged in global data can be viewed as an
1936 interaction between two pieces of data: the task state of the task waiting for
1937 the event and the global data used to indicate the event. To make sure that
1938 these appear to happen in the right order, the primitives to begin the process
1939 of going to sleep, and the primitives to initiate a wake up imply certain
1942 Firstly, the sleeper normally follows something like this sequence of events:
1945 set_current_state(TASK_UNINTERRUPTIBLE);
1946 if (event_indicated)
1951 A general memory barrier is interpolated automatically by set_current_state()
1952 after it has altered the task state:
1955 ===============================
1956 set_current_state();
1958 STORE current->state
1960 LOAD event_indicated
1962 set_current_state() may be wrapped by:
1965 prepare_to_wait_exclusive();
1967 which therefore also imply a general memory barrier after setting the state.
1968 The whole sequence above is available in various canned forms, all of which
1969 interpolate the memory barrier in the right place:
1972 wait_event_interruptible();
1973 wait_event_interruptible_exclusive();
1974 wait_event_interruptible_timeout();
1975 wait_event_killable();
1976 wait_event_timeout();
1981 Secondly, code that performs a wake up normally follows something like this:
1983 event_indicated = 1;
1984 wake_up(&event_wait_queue);
1988 event_indicated = 1;
1989 wake_up_process(event_daemon);
1991 A write memory barrier is implied by wake_up() and co. if and only if they wake
1992 something up. The barrier occurs before the task state is cleared, and so sits
1993 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1996 =============================== ===============================
1997 set_current_state(); STORE event_indicated
1998 smp_store_mb(); wake_up();
1999 STORE current->state <write barrier>
2000 <general barrier> STORE current->state
2001 LOAD event_indicated
2003 To repeat, this write memory barrier is present if and only if something
2004 is actually awakened. To see this, consider the following sequence of
2005 events, where X and Y are both initially zero:
2008 =============================== ===============================
2009 X = 1; STORE event_indicated
2010 smp_mb(); wake_up();
2011 Y = 1; wait_event(wq, Y == 1);
2012 wake_up(); load from Y sees 1, no memory barrier
2013 load from X might see 0
2015 In contrast, if a wakeup does occur, CPU 2's load from X would be guaranteed
2018 The available waker functions include:
2024 wake_up_interruptible();
2025 wake_up_interruptible_all();
2026 wake_up_interruptible_nr();
2027 wake_up_interruptible_poll();
2028 wake_up_interruptible_sync();
2029 wake_up_interruptible_sync_poll();
2031 wake_up_locked_poll();
2037 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
2038 order multiple stores before the wake-up with respect to loads of those stored
2039 values after the sleeper has called set_current_state(). For instance, if the
2042 set_current_state(TASK_INTERRUPTIBLE);
2043 if (event_indicated)
2045 __set_current_state(TASK_RUNNING);
2046 do_something(my_data);
2051 event_indicated = 1;
2052 wake_up(&event_wait_queue);
2054 there's no guarantee that the change to event_indicated will be perceived by
2055 the sleeper as coming after the change to my_data. In such a circumstance, the
2056 code on both sides must interpolate its own memory barriers between the
2057 separate data accesses. Thus the above sleeper ought to do:
2059 set_current_state(TASK_INTERRUPTIBLE);
2060 if (event_indicated) {
2062 do_something(my_data);
2065 and the waker should do:
2069 event_indicated = 1;
2070 wake_up(&event_wait_queue);
2073 MISCELLANEOUS FUNCTIONS
2074 -----------------------
2076 Other functions that imply barriers:
2078 (*) schedule() and similar imply full memory barriers.
2081 ===================================
2082 INTER-CPU ACQUIRING BARRIER EFFECTS
2083 ===================================
2085 On SMP systems locking primitives give a more substantial form of barrier: one
2086 that does affect memory access ordering on other CPUs, within the context of
2087 conflict on any particular lock.
2090 ACQUIRES VS MEMORY ACCESSES
2091 ---------------------------
2093 Consider the following: the system has a pair of spinlocks (M) and (Q), and
2094 three CPUs; then should the following sequence of events occur:
2097 =============================== ===============================
2098 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e);
2100 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f);
2101 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g);
2103 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h);
2105 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2106 through *H occur in, other than the constraints imposed by the separate locks
2107 on the separate CPUs. It might, for example, see:
2109 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2111 But it won't see any of:
2113 *B, *C or *D preceding ACQUIRE M
2114 *A, *B or *C following RELEASE M
2115 *F, *G or *H preceding ACQUIRE Q
2116 *E, *F or *G following RELEASE Q
2120 ACQUIRES VS I/O ACCESSES
2121 ------------------------
2123 Under certain circumstances (especially involving NUMA), I/O accesses within
2124 two spinlocked sections on two different CPUs may be seen as interleaved by the
2125 PCI bridge, because the PCI bridge does not necessarily participate in the
2126 cache-coherence protocol, and is therefore incapable of issuing the required
2127 read memory barriers.
2132 =============================== ===============================
2142 may be seen by the PCI bridge as follows:
2144 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2146 which would probably cause the hardware to malfunction.
2149 What is necessary here is to intervene with an mmiowb() before dropping the
2150 spinlock, for example:
2153 =============================== ===============================
2165 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2166 before either of the stores issued on CPU 2.
2169 Furthermore, following a store by a load from the same device obviates the need
2170 for the mmiowb(), because the load forces the store to complete before the load
2174 =============================== ===============================
2185 See Documentation/DocBook/deviceiobook.tmpl for more information.
2188 =================================
2189 WHERE ARE MEMORY BARRIERS NEEDED?
2190 =================================
2192 Under normal operation, memory operation reordering is generally not going to
2193 be a problem as a single-threaded linear piece of code will still appear to
2194 work correctly, even if it's in an SMP kernel. There are, however, four
2195 circumstances in which reordering definitely _could_ be a problem:
2197 (*) Interprocessor interaction.
2199 (*) Atomic operations.
2201 (*) Accessing devices.
2206 INTERPROCESSOR INTERACTION
2207 --------------------------
2209 When there's a system with more than one processor, more than one CPU in the
2210 system may be working on the same data set at the same time. This can cause
2211 synchronisation problems, and the usual way of dealing with them is to use
2212 locks. Locks, however, are quite expensive, and so it may be preferable to
2213 operate without the use of a lock if at all possible. In such a case
2214 operations that affect both CPUs may have to be carefully ordered to prevent
2217 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2218 queued on the semaphore, by virtue of it having a piece of its stack linked to
2219 the semaphore's list of waiting processes:
2221 struct rw_semaphore {
2224 struct list_head waiters;
2227 struct rwsem_waiter {
2228 struct list_head list;
2229 struct task_struct *task;
2232 To wake up a particular waiter, the up_read() or up_write() functions have to:
2234 (1) read the next pointer from this waiter's record to know as to where the
2235 next waiter record is;
2237 (2) read the pointer to the waiter's task structure;
2239 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2241 (4) call wake_up_process() on the task; and
2243 (5) release the reference held on the waiter's task struct.
2245 In other words, it has to perform this sequence of events:
2247 LOAD waiter->list.next;
2253 and if any of these steps occur out of order, then the whole thing may
2256 Once it has queued itself and dropped the semaphore lock, the waiter does not
2257 get the lock again; it instead just waits for its task pointer to be cleared
2258 before proceeding. Since the record is on the waiter's stack, this means that
2259 if the task pointer is cleared _before_ the next pointer in the list is read,
2260 another CPU might start processing the waiter and might clobber the waiter's
2261 stack before the up*() function has a chance to read the next pointer.
2263 Consider then what might happen to the above sequence of events:
2266 =============================== ===============================
2273 Woken up by other event
2278 foo() clobbers *waiter
2280 LOAD waiter->list.next;
2283 This could be dealt with using the semaphore lock, but then the down_xxx()
2284 function has to needlessly get the spinlock again after being woken up.
2286 The way to deal with this is to insert a general SMP memory barrier:
2288 LOAD waiter->list.next;
2295 In this case, the barrier makes a guarantee that all memory accesses before the
2296 barrier will appear to happen before all the memory accesses after the barrier
2297 with respect to the other CPUs on the system. It does _not_ guarantee that all
2298 the memory accesses before the barrier will be complete by the time the barrier
2299 instruction itself is complete.
2301 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2302 compiler barrier, thus making sure the compiler emits the instructions in the
2303 right order without actually intervening in the CPU. Since there's only one
2304 CPU, that CPU's dependency ordering logic will take care of everything else.
2310 Whilst they are technically interprocessor interaction considerations, atomic
2311 operations are noted specially as some of them imply full memory barriers and
2312 some don't, but they're very heavily relied on as a group throughout the
2315 Any atomic operation that modifies some state in memory and returns information
2316 about the state (old or new) implies an SMP-conditional general memory barrier
2317 (smp_mb()) on each side of the actual operation (with the exception of
2318 explicit lock operations, described later). These include:
2321 atomic_xchg(); atomic_long_xchg();
2322 atomic_inc_return(); atomic_long_inc_return();
2323 atomic_dec_return(); atomic_long_dec_return();
2324 atomic_add_return(); atomic_long_add_return();
2325 atomic_sub_return(); atomic_long_sub_return();
2326 atomic_inc_and_test(); atomic_long_inc_and_test();
2327 atomic_dec_and_test(); atomic_long_dec_and_test();
2328 atomic_sub_and_test(); atomic_long_sub_and_test();
2329 atomic_add_negative(); atomic_long_add_negative();
2331 test_and_clear_bit();
2332 test_and_change_bit();
2336 atomic_cmpxchg(); atomic_long_cmpxchg();
2337 atomic_add_unless(); atomic_long_add_unless();
2339 These are used for such things as implementing ACQUIRE-class and RELEASE-class
2340 operations and adjusting reference counters towards object destruction, and as
2341 such the implicit memory barrier effects are necessary.
2344 The following operations are potential problems as they do _not_ imply memory
2345 barriers, but might be used for implementing such things as RELEASE-class
2353 With these the appropriate explicit memory barrier should be used if necessary
2354 (smp_mb__before_atomic() for instance).
2357 The following also do _not_ imply memory barriers, and so may require explicit
2358 memory barriers under some circumstances (smp_mb__before_atomic() for
2366 If they're used for statistics generation, then they probably don't need memory
2367 barriers, unless there's a coupling between statistical data.
2369 If they're used for reference counting on an object to control its lifetime,
2370 they probably don't need memory barriers because either the reference count
2371 will be adjusted inside a locked section, or the caller will already hold
2372 sufficient references to make the lock, and thus a memory barrier unnecessary.
2374 If they're used for constructing a lock of some description, then they probably
2375 do need memory barriers as a lock primitive generally has to do things in a
2378 Basically, each usage case has to be carefully considered as to whether memory
2379 barriers are needed or not.
2381 The following operations are special locking primitives:
2383 test_and_set_bit_lock();
2385 __clear_bit_unlock();
2387 These implement ACQUIRE-class and RELEASE-class operations. These should be used in
2388 preference to other operations when implementing locking primitives, because
2389 their implementations can be optimised on many architectures.
2391 [!] Note that special memory barrier primitives are available for these
2392 situations because on some CPUs the atomic instructions used imply full memory
2393 barriers, and so barrier instructions are superfluous in conjunction with them,
2394 and in such cases the special barrier primitives will be no-ops.
2396 See Documentation/atomic_ops.txt for more information.
2402 Many devices can be memory mapped, and so appear to the CPU as if they're just
2403 a set of memory locations. To control such a device, the driver usually has to
2404 make the right memory accesses in exactly the right order.
2406 However, having a clever CPU or a clever compiler creates a potential problem
2407 in that the carefully sequenced accesses in the driver code won't reach the
2408 device in the requisite order if the CPU or the compiler thinks it is more
2409 efficient to reorder, combine or merge accesses - something that would cause
2410 the device to malfunction.
2412 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2413 routines - such as inb() or writel() - which know how to make such accesses
2414 appropriately sequential. Whilst this, for the most part, renders the explicit
2415 use of memory barriers unnecessary, there are a couple of situations where they
2418 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2419 so for _all_ general drivers locks should be used and mmiowb() must be
2420 issued prior to unlocking the critical section.
2422 (2) If the accessor functions are used to refer to an I/O memory window with
2423 relaxed memory access properties, then _mandatory_ memory barriers are
2424 required to enforce ordering.
2426 See Documentation/DocBook/deviceiobook.tmpl for more information.
2432 A driver may be interrupted by its own interrupt service routine, and thus the
2433 two parts of the driver may interfere with each other's attempts to control or
2436 This may be alleviated - at least in part - by disabling local interrupts (a
2437 form of locking), such that the critical operations are all contained within
2438 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2439 routine is executing, the driver's core may not run on the same CPU, and its
2440 interrupt is not permitted to happen again until the current interrupt has been
2441 handled, thus the interrupt handler does not need to lock against that.
2443 However, consider a driver that was talking to an ethernet card that sports an
2444 address register and a data register. If that driver's core talks to the card
2445 under interrupt-disablement and then the driver's interrupt handler is invoked:
2456 The store to the data register might happen after the second store to the
2457 address register if ordering rules are sufficiently relaxed:
2459 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2462 If ordering rules are relaxed, it must be assumed that accesses done inside an
2463 interrupt disabled section may leak outside of it and may interleave with
2464 accesses performed in an interrupt - and vice versa - unless implicit or
2465 explicit barriers are used.
2467 Normally this won't be a problem because the I/O accesses done inside such
2468 sections will include synchronous load operations on strictly ordered I/O
2469 registers that form implicit I/O barriers. If this isn't sufficient then an
2470 mmiowb() may need to be used explicitly.
2473 A similar situation may occur between an interrupt routine and two routines
2474 running on separate CPUs that communicate with each other. If such a case is
2475 likely, then interrupt-disabling locks should be used to guarantee ordering.
2478 ==========================
2479 KERNEL I/O BARRIER EFFECTS
2480 ==========================
2482 When accessing I/O memory, drivers should use the appropriate accessor
2487 These are intended to talk to I/O space rather than memory space, but
2488 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
2489 indeed have special I/O space access cycles and instructions, but many
2490 CPUs don't have such a concept.
2492 The PCI bus, amongst others, defines an I/O space concept which - on such
2493 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2494 space. However, it may also be mapped as a virtual I/O space in the CPU's
2495 memory map, particularly on those CPUs that don't support alternate I/O
2498 Accesses to this space may be fully synchronous (as on i386), but
2499 intermediary bridges (such as the PCI host bridge) may not fully honour
2502 They are guaranteed to be fully ordered with respect to each other.
2504 They are not guaranteed to be fully ordered with respect to other types of
2505 memory and I/O operation.
2507 (*) readX(), writeX():
2509 Whether these are guaranteed to be fully ordered and uncombined with
2510 respect to each other on the issuing CPU depends on the characteristics
2511 defined for the memory window through which they're accessing. On later
2512 i386 architecture machines, for example, this is controlled by way of the
2515 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2516 provided they're not accessing a prefetchable device.
2518 However, intermediary hardware (such as a PCI bridge) may indulge in
2519 deferral if it so wishes; to flush a store, a load from the same location
2520 is preferred[*], but a load from the same device or from configuration
2521 space should suffice for PCI.
2523 [*] NOTE! attempting to load from the same location as was written to may
2524 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2527 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2528 force stores to be ordered.
2530 Please refer to the PCI specification for more information on interactions
2531 between PCI transactions.
2533 (*) readX_relaxed(), writeX_relaxed()
2535 These are similar to readX() and writeX(), but provide weaker memory
2536 ordering guarantees. Specifically, they do not guarantee ordering with
2537 respect to normal memory accesses (e.g. DMA buffers) nor do they guarantee
2538 ordering with respect to LOCK or UNLOCK operations. If the latter is
2539 required, an mmiowb() barrier can be used. Note that relaxed accesses to
2540 the same peripheral are guaranteed to be ordered with respect to each
2543 (*) ioreadX(), iowriteX()
2545 These will perform appropriately for the type of access they're actually
2546 doing, be it inX()/outX() or readX()/writeX().
2549 ========================================
2550 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2551 ========================================
2553 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2554 maintain the appearance of program causality with respect to itself. Some CPUs
2555 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2556 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2557 of arch-specific code.
2559 This means that it must be considered that the CPU will execute its instruction
2560 stream in any order it feels like - or even in parallel - provided that if an
2561 instruction in the stream depends on an earlier instruction, then that
2562 earlier instruction must be sufficiently complete[*] before the later
2563 instruction may proceed; in other words: provided that the appearance of
2564 causality is maintained.
2566 [*] Some instructions have more than one effect - such as changing the
2567 condition codes, changing registers or changing memory - and different
2568 instructions may depend on different effects.
2570 A CPU may also discard any instruction sequence that winds up having no
2571 ultimate effect. For example, if two adjacent instructions both load an
2572 immediate value into the same register, the first may be discarded.
2575 Similarly, it has to be assumed that compiler might reorder the instruction
2576 stream in any way it sees fit, again provided the appearance of causality is
2580 ============================
2581 THE EFFECTS OF THE CPU CACHE
2582 ============================
2584 The way cached memory operations are perceived across the system is affected to
2585 a certain extent by the caches that lie between CPUs and memory, and by the
2586 memory coherence system that maintains the consistency of state in the system.
2588 As far as the way a CPU interacts with another part of the system through the
2589 caches goes, the memory system has to include the CPU's caches, and memory
2590 barriers for the most part act at the interface between the CPU and its cache
2591 (memory barriers logically act on the dotted line in the following diagram):
2593 <--- CPU ---> : <----------- Memory ----------->
2595 +--------+ +--------+ : +--------+ +-----------+
2596 | | | | : | | | | +--------+
2597 | CPU | | Memory | : | CPU | | | | |
2598 | Core |--->| Access |----->| Cache |<-->| | | |
2599 | | | Queue | : | | | |--->| Memory |
2600 | | | | : | | | | | |
2601 +--------+ +--------+ : +--------+ | | | |
2602 : | Cache | +--------+
2604 : | Mechanism | +--------+
2605 +--------+ +--------+ : +--------+ | | | |
2606 | | | | : | | | | | |
2607 | CPU | | Memory | : | CPU | | |--->| Device |
2608 | Core |--->| Access |----->| Cache |<-->| | | |
2609 | | | Queue | : | | | | | |
2610 | | | | : | | | | +--------+
2611 +--------+ +--------+ : +--------+ +-----------+
2615 Although any particular load or store may not actually appear outside of the
2616 CPU that issued it since it may have been satisfied within the CPU's own cache,
2617 it will still appear as if the full memory access had taken place as far as the
2618 other CPUs are concerned since the cache coherency mechanisms will migrate the
2619 cacheline over to the accessing CPU and propagate the effects upon conflict.
2621 The CPU core may execute instructions in any order it deems fit, provided the
2622 expected program causality appears to be maintained. Some of the instructions
2623 generate load and store operations which then go into the queue of memory
2624 accesses to be performed. The core may place these in the queue in any order
2625 it wishes, and continue execution until it is forced to wait for an instruction
2628 What memory barriers are concerned with is controlling the order in which
2629 accesses cross from the CPU side of things to the memory side of things, and
2630 the order in which the effects are perceived to happen by the other observers
2633 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2634 their own loads and stores as if they had happened in program order.
2636 [!] MMIO or other device accesses may bypass the cache system. This depends on
2637 the properties of the memory window through which devices are accessed and/or
2638 the use of any special device communication instructions the CPU may have.
2644 Life isn't quite as simple as it may appear above, however: for while the
2645 caches are expected to be coherent, there's no guarantee that that coherency
2646 will be ordered. This means that whilst changes made on one CPU will
2647 eventually become visible on all CPUs, there's no guarantee that they will
2648 become apparent in the same order on those other CPUs.
2651 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2652 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2657 +--------+ : +--->| Cache A |<------->| |
2658 | | : | +---------+ | |
2660 | | : | +---------+ | |
2661 +--------+ : +--->| Cache B |<------->| |
2664 : +---------+ | System |
2665 +--------+ : +--->| Cache C |<------->| |
2666 | | : | +---------+ | |
2668 | | : | +---------+ | |
2669 +--------+ : +--->| Cache D |<------->| |
2674 Imagine the system has the following properties:
2676 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2679 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2682 (*) whilst the CPU core is interrogating one cache, the other cache may be
2683 making use of the bus to access the rest of the system - perhaps to
2684 displace a dirty cacheline or to do a speculative load;
2686 (*) each cache has a queue of operations that need to be applied to that cache
2687 to maintain coherency with the rest of the system;
2689 (*) the coherency queue is not flushed by normal loads to lines already
2690 present in the cache, even though the contents of the queue may
2691 potentially affect those loads.
2693 Imagine, then, that two writes are made on the first CPU, with a write barrier
2694 between them to guarantee that they will appear to reach that CPU's caches in
2695 the requisite order:
2698 =============== =============== =======================================
2699 u == 0, v == 1 and p == &u, q == &u
2701 smp_wmb(); Make sure change to v is visible before
2703 <A:modify v=2> v is now in cache A exclusively
2705 <B:modify p=&v> p is now in cache B exclusively
2707 The write memory barrier forces the other CPUs in the system to perceive that
2708 the local CPU's caches have apparently been updated in the correct order. But
2709 now imagine that the second CPU wants to read those values:
2712 =============== =============== =======================================
2717 The above pair of reads may then fail to happen in the expected order, as the
2718 cacheline holding p may get updated in one of the second CPU's caches whilst
2719 the update to the cacheline holding v is delayed in the other of the second
2720 CPU's caches by some other cache event:
2723 =============== =============== =======================================
2724 u == 0, v == 1 and p == &u, q == &u
2727 <A:modify v=2> <C:busy>
2731 <B:modify p=&v> <D:commit p=&v>
2734 <C:read *q> Reads from v before v updated in cache
2738 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2739 no guarantee that, without intervention, the order of update will be the same
2740 as that committed on CPU 1.
2743 To intervene, we need to interpolate a data dependency barrier or a read
2744 barrier between the loads. This will force the cache to commit its coherency
2745 queue before processing any further requests:
2748 =============== =============== =======================================
2749 u == 0, v == 1 and p == &u, q == &u
2752 <A:modify v=2> <C:busy>
2756 <B:modify p=&v> <D:commit p=&v>
2758 smp_read_barrier_depends()
2762 <C:read *q> Reads from v after v updated in cache
2765 This sort of problem can be encountered on DEC Alpha processors as they have a
2766 split cache that improves performance by making better use of the data bus.
2767 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2768 access depends on a read, not all do, so it may not be relied on.
2770 Other CPUs may also have split caches, but must coordinate between the various
2771 cachelets for normal memory accesses. The semantics of the Alpha removes the
2772 need for coordination in the absence of memory barriers.
2775 CACHE COHERENCY VS DMA
2776 ----------------------
2778 Not all systems maintain cache coherency with respect to devices doing DMA. In
2779 such cases, a device attempting DMA may obtain stale data from RAM because
2780 dirty cache lines may be resident in the caches of various CPUs, and may not
2781 have been written back to RAM yet. To deal with this, the appropriate part of
2782 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2783 invalidate them as well).
2785 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2786 cache lines being written back to RAM from a CPU's cache after the device has
2787 installed its own data, or cache lines present in the CPU's cache may simply
2788 obscure the fact that RAM has been updated, until at such time as the cacheline
2789 is discarded from the CPU's cache and reloaded. To deal with this, the
2790 appropriate part of the kernel must invalidate the overlapping bits of the
2793 See Documentation/cachetlb.txt for more information on cache management.
2796 CACHE COHERENCY VS MMIO
2797 -----------------------
2799 Memory mapped I/O usually takes place through memory locations that are part of
2800 a window in the CPU's memory space that has different properties assigned than
2801 the usual RAM directed window.
2803 Amongst these properties is usually the fact that such accesses bypass the
2804 caching entirely and go directly to the device buses. This means MMIO accesses
2805 may, in effect, overtake accesses to cached memory that were emitted earlier.
2806 A memory barrier isn't sufficient in such a case, but rather the cache must be
2807 flushed between the cached memory write and the MMIO access if the two are in
2811 =========================
2812 THE THINGS CPUS GET UP TO
2813 =========================
2815 A programmer might take it for granted that the CPU will perform memory
2816 operations in exactly the order specified, so that if the CPU is, for example,
2817 given the following piece of code to execute:
2825 they would then expect that the CPU will complete the memory operation for each
2826 instruction before moving on to the next one, leading to a definite sequence of
2827 operations as seen by external observers in the system:
2829 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2832 Reality is, of course, much messier. With many CPUs and compilers, the above
2833 assumption doesn't hold because:
2835 (*) loads are more likely to need to be completed immediately to permit
2836 execution progress, whereas stores can often be deferred without a
2839 (*) loads may be done speculatively, and the result discarded should it prove
2840 to have been unnecessary;
2842 (*) loads may be done speculatively, leading to the result having been fetched
2843 at the wrong time in the expected sequence of events;
2845 (*) the order of the memory accesses may be rearranged to promote better use
2846 of the CPU buses and caches;
2848 (*) loads and stores may be combined to improve performance when talking to
2849 memory or I/O hardware that can do batched accesses of adjacent locations,
2850 thus cutting down on transaction setup costs (memory and PCI devices may
2851 both be able to do this); and
2853 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2854 mechanisms may alleviate this - once the store has actually hit the cache
2855 - there's no guarantee that the coherency management will be propagated in
2856 order to other CPUs.
2858 So what another CPU, say, might actually observe from the above piece of code
2861 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2863 (Where "LOAD {*C,*D}" is a combined load)
2866 However, it is guaranteed that a CPU will be self-consistent: it will see its
2867 _own_ accesses appear to be correctly ordered, without the need for a memory
2868 barrier. For instance with the following code:
2877 and assuming no intervention by an external influence, it can be assumed that
2878 the final result will appear to be:
2880 U == the original value of *A
2885 The code above may cause the CPU to generate the full sequence of memory
2888 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2890 in that order, but, without intervention, the sequence may have almost any
2891 combination of elements combined or discarded, provided the program's view
2892 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE()
2893 are -not- optional in the above example, as there are architectures
2894 where a given CPU might reorder successive loads to the same location.
2895 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is
2896 necessary to prevent this, for example, on Itanium the volatile casts
2897 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq
2898 and st.rel instructions (respectively) that prevent such reordering.
2900 The compiler may also combine, discard or defer elements of the sequence before
2901 the CPU even sees them.
2912 since, without either a write barrier or an WRITE_ONCE(), it can be
2913 assumed that the effect of the storage of V to *A is lost. Similarly:
2918 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be
2924 and the LOAD operation never appear outside of the CPU.
2927 AND THEN THERE'S THE ALPHA
2928 --------------------------
2930 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2931 some versions of the Alpha CPU have a split data cache, permitting them to have
2932 two semantically-related cache lines updated at separate times. This is where
2933 the data dependency barrier really becomes necessary as this synchronises both
2934 caches with the memory coherence system, thus making it seem like pointer
2935 changes vs new data occur in the right order.
2937 The Alpha defines the Linux kernel's memory barrier model.
2939 See the subsection on "Cache Coherency" above.
2941 VIRTUAL MACHINE GUESTS
2944 Guests running within virtual machines might be affected by SMP effects even if
2945 the guest itself is compiled without SMP support. This is an artifact of
2946 interfacing with an SMP host while running an UP kernel. Using mandatory
2947 barriers for this use-case would be possible but is often suboptimal.
2949 To handle this case optimally, low-level virt_mb() etc macros are available.
2950 These have the same effect as smp_mb() etc when SMP is enabled, but generate
2951 identical code for SMP and non-SMP systems. For example, virtual machine guests
2952 should use virt_mb() rather than smp_mb() when synchronizing against a
2953 (possibly SMP) host.
2955 These are equivalent to smp_mb() etc counterparts in all other respects,
2956 in particular, they do not control MMIO effects: to control
2957 MMIO effects, use mandatory barriers.
2966 Memory barriers can be used to implement circular buffering without the need
2967 of a lock to serialise the producer with the consumer. See:
2969 Documentation/circular-buffers.txt
2978 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2980 Chapter 5.2: Physical Address Space Characteristics
2981 Chapter 5.4: Caches and Write Buffers
2982 Chapter 5.5: Data Sharing
2983 Chapter 5.6: Read/Write Ordering
2985 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2986 Chapter 7.1: Memory-Access Ordering
2987 Chapter 7.4: Buffering and Combining Memory Writes
2989 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2990 System Programming Guide
2991 Chapter 7.1: Locked Atomic Operations
2992 Chapter 7.2: Memory Ordering
2993 Chapter 7.4: Serializing Instructions
2995 The SPARC Architecture Manual, Version 9
2996 Chapter 8: Memory Models
2997 Appendix D: Formal Specification of the Memory Models
2998 Appendix J: Programming with the Memory Models
3000 UltraSPARC Programmer Reference Manual
3001 Chapter 5: Memory Accesses and Cacheability
3002 Chapter 15: Sparc-V9 Memory Models
3004 UltraSPARC III Cu User's Manual
3005 Chapter 9: Memory Models
3007 UltraSPARC IIIi Processor User's Manual
3008 Chapter 8: Memory Models
3010 UltraSPARC Architecture 2005
3012 Appendix D: Formal Specifications of the Memory Models
3014 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
3015 Chapter 8: Memory Models
3016 Appendix F: Caches and Cache Coherency
3018 Solaris Internals, Core Kernel Architecture, p63-68:
3019 Chapter 3.3: Hardware Considerations for Locks and
3022 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
3023 for Kernel Programmers:
3024 Chapter 13: Other Memory Models
3026 Intel Itanium Architecture Software Developer's Manual: Volume 1:
3027 Section 2.6: Speculation
3028 Section 4.4: Memory Access