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, x=LOAD A->3, y=LOAD B->4
125 STORE A=3, STORE B=4, y=LOAD B->4, x=LOAD A->3
126 STORE A=3, x=LOAD A->3, STORE B=4, y=LOAD B->4
127 STORE A=3, x=LOAD A->3, y=LOAD B->2, STORE B=4
128 STORE A=3, y=LOAD B->2, STORE B=4, x=LOAD A->3
129 STORE A=3, y=LOAD B->2, x=LOAD A->3, STORE B=4
130 STORE B=4, STORE A=3, x=LOAD A->3, y=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 ACCESS_ONCE(Q) = P; smp_read_barrier_depends(); D = ACCESS_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 ACCESS_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 = ACCESS_ONCE(*X); ACCESS_ONCE(*X) = b;
214 the CPU will only issue the following sequence of memory operations:
216 a = LOAD *X, STORE *X = b
220 ACCESS_ONCE(*X) = c; d = ACCESS_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 with
232 memory references that are not protected by ACCESS_ONCE(). Without
233 ACCESS_ONCE(), the compiler is within its rights to do all sorts
234 of "creative" transformations, which are covered in the Compiler
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};
273 =========================
274 WHAT ARE MEMORY BARRIERS?
275 =========================
277 As can be seen above, independent memory operations are effectively performed
278 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
279 What is required is some way of intervening to instruct the compiler and the
280 CPU to restrict the order.
282 Memory barriers are such interventions. They impose a perceived partial
283 ordering over the memory operations on either side of the barrier.
285 Such enforcement is important because the CPUs and other devices in a system
286 can use a variety of tricks to improve performance, including reordering,
287 deferral and combination of memory operations; speculative loads; speculative
288 branch prediction and various types of caching. Memory barriers are used to
289 override or suppress these tricks, allowing the code to sanely control the
290 interaction of multiple CPUs and/or devices.
293 VARIETIES OF MEMORY BARRIER
294 ---------------------------
296 Memory barriers come in four basic varieties:
298 (1) Write (or store) memory barriers.
300 A write memory barrier gives a guarantee that all the STORE operations
301 specified before the barrier will appear to happen before all the STORE
302 operations specified after the barrier with respect to the other
303 components of the system.
305 A write barrier is a partial ordering on stores only; it is not required
306 to have any effect on loads.
308 A CPU can be viewed as committing a sequence of store operations to the
309 memory system as time progresses. All stores before a write barrier will
310 occur in the sequence _before_ all the stores after the write barrier.
312 [!] Note that write barriers should normally be paired with read or data
313 dependency barriers; see the "SMP barrier pairing" subsection.
316 (2) Data dependency barriers.
318 A data dependency barrier is a weaker form of read barrier. In the case
319 where two loads are performed such that the second depends on the result
320 of the first (eg: the first load retrieves the address to which the second
321 load will be directed), a data dependency barrier would be required to
322 make sure that the target of the second load is updated before the address
323 obtained by the first load is accessed.
325 A data dependency barrier is a partial ordering on interdependent loads
326 only; it is not required to have any effect on stores, independent loads
327 or overlapping loads.
329 As mentioned in (1), the other CPUs in the system can be viewed as
330 committing sequences of stores to the memory system that the CPU being
331 considered can then perceive. A data dependency barrier issued by the CPU
332 under consideration guarantees that for any load preceding it, if that
333 load touches one of a sequence of stores from another CPU, then by the
334 time the barrier completes, the effects of all the stores prior to that
335 touched by the load will be perceptible to any loads issued after the data
338 See the "Examples of memory barrier sequences" subsection for diagrams
339 showing the ordering constraints.
341 [!] Note that the first load really has to have a _data_ dependency and
342 not a control dependency. If the address for the second load is dependent
343 on the first load, but the dependency is through a conditional rather than
344 actually loading the address itself, then it's a _control_ dependency and
345 a full read barrier or better is required. See the "Control dependencies"
346 subsection for more information.
348 [!] Note that data dependency barriers should normally be paired with
349 write barriers; see the "SMP barrier pairing" subsection.
352 (3) Read (or load) memory barriers.
354 A read barrier is a data dependency barrier plus a guarantee that all the
355 LOAD operations specified before the barrier will appear to happen before
356 all the LOAD operations specified after the barrier with respect to the
357 other components of the system.
359 A read barrier is a partial ordering on loads only; it is not required to
360 have any effect on stores.
362 Read memory barriers imply data dependency barriers, and so can substitute
365 [!] Note that read barriers should normally be paired with write barriers;
366 see the "SMP barrier pairing" subsection.
369 (4) General memory barriers.
371 A general memory barrier gives a guarantee that all the LOAD and STORE
372 operations specified before the barrier will appear to happen before all
373 the LOAD and STORE operations specified after the barrier with respect to
374 the other components of the system.
376 A general memory barrier is a partial ordering over both loads and stores.
378 General memory barriers imply both read and write memory barriers, and so
379 can substitute for either.
382 And a couple of implicit varieties:
384 (5) ACQUIRE operations.
386 This acts as a one-way permeable barrier. It guarantees that all memory
387 operations after the ACQUIRE operation will appear to happen after the
388 ACQUIRE operation with respect to the other components of the system.
389 ACQUIRE operations include LOCK operations and smp_load_acquire()
392 Memory operations that occur before an ACQUIRE operation may appear to
393 happen after it completes.
395 An ACQUIRE operation should almost always be paired with a RELEASE
399 (6) RELEASE operations.
401 This also acts as a one-way permeable barrier. It guarantees that all
402 memory operations before the RELEASE operation will appear to happen
403 before the RELEASE operation with respect to the other components of the
404 system. RELEASE operations include UNLOCK operations and
405 smp_store_release() operations.
407 Memory operations that occur after a RELEASE operation may appear to
408 happen before it completes.
410 The use of ACQUIRE and RELEASE operations generally precludes the need
411 for other sorts of memory barrier (but note the exceptions mentioned in
412 the subsection "MMIO write barrier"). In addition, a RELEASE+ACQUIRE
413 pair is -not- guaranteed to act as a full memory barrier. However, after
414 an ACQUIRE on a given variable, all memory accesses preceding any prior
415 RELEASE on that same variable are guaranteed to be visible. In other
416 words, within a given variable's critical section, all accesses of all
417 previous critical sections for that variable are guaranteed to have
420 This means that ACQUIRE acts as a minimal "acquire" operation and
421 RELEASE acts as a minimal "release" operation.
424 Memory barriers are only required where there's a possibility of interaction
425 between two CPUs or between a CPU and a device. If it can be guaranteed that
426 there won't be any such interaction in any particular piece of code, then
427 memory barriers are unnecessary in that piece of code.
430 Note that these are the _minimum_ guarantees. Different architectures may give
431 more substantial guarantees, but they may _not_ be relied upon outside of arch
435 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
436 ----------------------------------------------
438 There are certain things that the Linux kernel memory barriers do not guarantee:
440 (*) There is no guarantee that any of the memory accesses specified before a
441 memory barrier will be _complete_ by the completion of a memory barrier
442 instruction; the barrier can be considered to draw a line in that CPU's
443 access queue that accesses of the appropriate type may not cross.
445 (*) There is no guarantee that issuing a memory barrier on one CPU will have
446 any direct effect on another CPU or any other hardware in the system. The
447 indirect effect will be the order in which the second CPU sees the effects
448 of the first CPU's accesses occur, but see the next point:
450 (*) There is no guarantee that a CPU will see the correct order of effects
451 from a second CPU's accesses, even _if_ the second CPU uses a memory
452 barrier, unless the first CPU _also_ uses a matching memory barrier (see
453 the subsection on "SMP Barrier Pairing").
455 (*) There is no guarantee that some intervening piece of off-the-CPU
456 hardware[*] will not reorder the memory accesses. CPU cache coherency
457 mechanisms should propagate the indirect effects of a memory barrier
458 between CPUs, but might not do so in order.
460 [*] For information on bus mastering DMA and coherency please read:
462 Documentation/PCI/pci.txt
463 Documentation/DMA-API-HOWTO.txt
464 Documentation/DMA-API.txt
467 DATA DEPENDENCY BARRIERS
468 ------------------------
470 The usage requirements of data dependency barriers are a little subtle, and
471 it's not always obvious that they're needed. To illustrate, consider the
472 following sequence of events:
475 =============== ===============
476 { A == 1, B == 2, C = 3, P == &A, Q == &C }
483 There's a clear data dependency here, and it would seem that by the end of the
484 sequence, Q must be either &A or &B, and that:
486 (Q == &A) implies (D == 1)
487 (Q == &B) implies (D == 4)
489 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
490 leading to the following situation:
492 (Q == &B) and (D == 2) ????
494 Whilst this may seem like a failure of coherency or causality maintenance, it
495 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
498 To deal with this, a data dependency barrier or better must be inserted
499 between the address load and the data load:
502 =============== ===============
503 { A == 1, B == 2, C = 3, P == &A, Q == &C }
508 <data dependency barrier>
511 This enforces the occurrence of one of the two implications, and prevents the
512 third possibility from arising.
514 [!] Note that this extremely counterintuitive situation arises most easily on
515 machines with split caches, so that, for example, one cache bank processes
516 even-numbered cache lines and the other bank processes odd-numbered cache
517 lines. The pointer P might be stored in an odd-numbered cache line, and the
518 variable B might be stored in an even-numbered cache line. Then, if the
519 even-numbered bank of the reading CPU's cache is extremely busy while the
520 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
521 but the old value of the variable B (2).
524 Another example of where data dependency barriers might be required is where a
525 number is read from memory and then used to calculate the index for an array
529 =============== ===============
530 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
535 <data dependency barrier>
539 The data dependency barrier is very important to the RCU system,
540 for example. See rcu_assign_pointer() and rcu_dereference() in
541 include/linux/rcupdate.h. This permits the current target of an RCU'd
542 pointer to be replaced with a new modified target, without the replacement
543 target appearing to be incompletely initialised.
545 See also the subsection on "Cache Coherency" for a more thorough example.
551 A control dependency requires a full read memory barrier, not simply a data
552 dependency barrier to make it work correctly. Consider the following bit of
557 <data dependency barrier> /* BUG: No data dependency!!! */
561 This will not have the desired effect because there is no actual data
562 dependency, but rather a control dependency that the CPU may short-circuit
563 by attempting to predict the outcome in advance, so that other CPUs see
564 the load from b as having happened before the load from a. In such a
565 case what's actually required is:
573 However, stores are not speculated. This means that ordering -is- provided
574 in the following example:
581 Please note that ACCESS_ONCE() is not optional! Without the
582 ACCESS_ONCE(), might combine the load from 'a' with other loads from
583 'a', and the store to 'b' with other stores to 'b', with possible highly
584 counterintuitive effects on ordering.
586 Worse yet, if the compiler is able to prove (say) that the value of
587 variable 'a' is always non-zero, it would be well within its rights
588 to optimize the original example by eliminating the "if" statement
592 b = p; /* BUG: Compiler and CPU can both reorder!!! */
594 So don't leave out the ACCESS_ONCE().
596 It is tempting to try to enforce ordering on identical stores on both
597 branches of the "if" statement as follows:
610 Unfortunately, current compilers will transform this as follows at high
615 ACCESS_ONCE(b) = p; /* BUG: No ordering vs. load from a!!! */
617 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
620 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
624 Now there is no conditional between the load from 'a' and the store to
625 'b', which means that the CPU is within its rights to reorder them:
626 The conditional is absolutely required, and must be present in the
627 assembly code even after all compiler optimizations have been applied.
628 Therefore, if you need ordering in this example, you need explicit
629 memory barriers, for example, smp_store_release():
633 smp_store_release(&b, p);
636 smp_store_release(&b, p);
640 In contrast, without explicit memory barriers, two-legged-if control
641 ordering is guaranteed only when the stores differ, for example:
652 The initial ACCESS_ONCE() is still required to prevent the compiler from
653 proving the value of 'a'.
655 In addition, you need to be careful what you do with the local variable 'q',
656 otherwise the compiler might be able to guess the value and again remove
657 the needed conditional. For example:
668 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
669 equal to zero, in which case the compiler is within its rights to
670 transform the above code into the following:
676 Given this transformation, the CPU is not required to respect the ordering
677 between the load from variable 'a' and the store to variable 'b'. It is
678 tempting to add a barrier(), but this does not help. The conditional
679 is gone, and the barrier won't bring it back. Therefore, if you are
680 relying on this ordering, you should make sure that MAX is greater than
681 one, perhaps as follows:
684 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
693 Please note once again that the stores to 'b' differ. If they were
694 identical, as noted earlier, the compiler could pull this store outside
695 of the 'if' statement.
697 Finally, control dependencies do -not- provide transitivity. This is
698 demonstrated by two related examples, with the initial values of
699 x and y both being zero:
702 ===================== =====================
703 r1 = ACCESS_ONCE(x); r2 = ACCESS_ONCE(y);
704 if (r1 > 0) if (r2 > 0)
705 ACCESS_ONCE(y) = 1; ACCESS_ONCE(x) = 1;
707 assert(!(r1 == 1 && r2 == 1));
709 The above two-CPU example will never trigger the assert(). However,
710 if control dependencies guaranteed transitivity (which they do not),
711 then adding the following CPU would guarantee a related assertion:
714 =====================
717 assert(!(r1 == 2 && r2 == 1 && x == 2)); /* FAILS!!! */
719 But because control dependencies do -not- provide transitivity, the above
720 assertion can fail after the combined three-CPU example completes. If you
721 need the three-CPU example to provide ordering, you will need smp_mb()
722 between the loads and stores in the CPU 0 and CPU 1 code fragments,
723 that is, just before or just after the "if" statements.
725 These two examples are the LB and WWC litmus tests from this paper:
726 http://www.cl.cam.ac.uk/users/pes20/ppc-supplemental/test6.pdf and this
727 site: https://www.cl.cam.ac.uk/~pes20/ppcmem/index.html.
731 (*) Control dependencies can order prior loads against later stores.
732 However, they do -not- guarantee any other sort of ordering:
733 Not prior loads against later loads, nor prior stores against
734 later anything. If you need these other forms of ordering,
735 use smb_rmb(), smp_wmb(), or, in the case of prior stores and
736 later loads, smp_mb().
738 (*) If both legs of the "if" statement begin with identical stores
739 to the same variable, a barrier() statement is required at the
740 beginning of each leg of the "if" statement.
742 (*) Control dependencies require at least one run-time conditional
743 between the prior load and the subsequent store, and this
744 conditional must involve the prior load. If the compiler
745 is able to optimize the conditional away, it will have also
746 optimized away the ordering. Careful use of ACCESS_ONCE() can
747 help to preserve the needed conditional.
749 (*) Control dependencies require that the compiler avoid reordering the
750 dependency into nonexistence. Careful use of ACCESS_ONCE() or
751 barrier() can help to preserve your control dependency. Please
752 see the Compiler Barrier section for more information.
754 (*) Control dependencies do -not- provide transitivity. If you
755 need transitivity, use smp_mb().
761 When dealing with CPU-CPU interactions, certain types of memory barrier should
762 always be paired. A lack of appropriate pairing is almost certainly an error.
764 General barriers pair with each other, though they also pair with
765 most other types of barriers, albeit without transitivity. An acquire
766 barrier pairs with a release barrier, but both may also pair with other
767 barriers, including of course general barriers. A write barrier pairs
768 with a data dependency barrier, an acquire barrier, a release barrier,
769 a read barrier, or a general barrier. Similarly a read barrier or a
770 data dependency barrier pairs with a write barrier, an acquire barrier,
771 a release barrier, or a general barrier:
774 =============== ===============
777 ACCESS_ONCE(b) = 2; x = ACCESS_ONCE(b);
784 =============== ===============================
787 ACCESS_ONCE(b) = &a; x = ACCESS_ONCE(b);
788 <data dependency barrier>
791 Basically, the read barrier always has to be there, even though it can be of
794 [!] Note that the stores before the write barrier would normally be expected to
795 match the loads after the read barrier or the data dependency barrier, and vice
799 =================== ===================
800 ACCESS_ONCE(a) = 1; }---- --->{ v = ACCESS_ONCE(c);
801 ACCESS_ONCE(b) = 2; } \ / { w = ACCESS_ONCE(d);
802 <write barrier> \ <read barrier>
803 ACCESS_ONCE(c) = 3; } / \ { x = ACCESS_ONCE(a);
804 ACCESS_ONCE(d) = 4; }---- --->{ y = ACCESS_ONCE(b);
807 EXAMPLES OF MEMORY BARRIER SEQUENCES
808 ------------------------------------
810 Firstly, write barriers act as partial orderings on store operations.
811 Consider the following sequence of events:
814 =======================
822 This sequence of events is committed to the memory coherence system in an order
823 that the rest of the system might perceive as the unordered set of { STORE A,
824 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
829 | |------>| C=3 | } /\
830 | | : +------+ }----- \ -----> Events perceptible to
831 | | : | A=1 | } \/ the rest of the system
833 | CPU 1 | : | B=2 | }
835 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
836 | | +------+ } requires all stores prior to the
837 | | : | E=5 | } barrier to be committed before
838 | | : +------+ } further stores may take place
843 | Sequence in which stores are committed to the
844 | memory system by CPU 1
848 Secondly, data dependency barriers act as partial orderings on data-dependent
849 loads. Consider the following sequence of events:
852 ======================= =======================
853 { B = 7; X = 9; Y = 8; C = &Y }
858 STORE D = 4 LOAD C (gets &B)
861 Without intervention, CPU 2 may perceive the events on CPU 1 in some
862 effectively random order, despite the write barrier issued by CPU 1:
865 | | +------+ +-------+ | Sequence of update
866 | |------>| B=2 |----- --->| Y->8 | | of perception on
867 | | : +------+ \ +-------+ | CPU 2
868 | CPU 1 | : | A=1 | \ --->| C->&Y | V
869 | | +------+ | +-------+
870 | | wwwwwwwwwwwwwwww | : :
872 | | : | C=&B |--- | : : +-------+
873 | | : +------+ \ | +-------+ | |
874 | |------>| D=4 | ----------->| C->&B |------>| |
875 | | +------+ | +-------+ | |
876 +-------+ : : | : : | |
880 Apparently incorrect ---> | | B->7 |------>| |
881 perception of B (!) | +-------+ | |
884 The load of X holds ---> \ | X->9 |------>| |
885 up the maintenance \ +-------+ | |
886 of coherence of B ----->| B->2 | +-------+
891 In the above example, CPU 2 perceives that B is 7, despite the load of *C
892 (which would be B) coming after the LOAD of C.
894 If, however, a data dependency barrier were to be placed between the load of C
895 and the load of *C (ie: B) on CPU 2:
898 ======================= =======================
899 { B = 7; X = 9; Y = 8; C = &Y }
904 STORE D = 4 LOAD C (gets &B)
905 <data dependency barrier>
908 then the following will occur:
911 | | +------+ +-------+
912 | |------>| B=2 |----- --->| Y->8 |
913 | | : +------+ \ +-------+
914 | CPU 1 | : | A=1 | \ --->| C->&Y |
915 | | +------+ | +-------+
916 | | wwwwwwwwwwwwwwww | : :
918 | | : | C=&B |--- | : : +-------+
919 | | : +------+ \ | +-------+ | |
920 | |------>| D=4 | ----------->| C->&B |------>| |
921 | | +------+ | +-------+ | |
922 +-------+ : : | : : | |
928 Makes sure all effects ---> \ ddddddddddddddddd | |
929 prior to the store of C \ +-------+ | |
930 are perceptible to ----->| B->2 |------>| |
931 subsequent loads +-------+ | |
935 And thirdly, a read barrier acts as a partial order on loads. Consider the
936 following sequence of events:
939 ======================= =======================
947 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
948 some effectively random order, despite the write barrier issued by CPU 1:
951 | | +------+ +-------+
952 | |------>| A=1 |------ --->| A->0 |
953 | | +------+ \ +-------+
954 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
955 | | +------+ | +-------+
956 | |------>| B=2 |--- | : :
957 | | +------+ \ | : : +-------+
958 +-------+ : : \ | +-------+ | |
959 ---------->| B->2 |------>| |
960 | +-------+ | CPU 2 |
971 If, however, a read barrier were to be placed between the load of B and the
975 ======================= =======================
984 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
988 | | +------+ +-------+
989 | |------>| A=1 |------ --->| A->0 |
990 | | +------+ \ +-------+
991 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
992 | | +------+ | +-------+
993 | |------>| B=2 |--- | : :
994 | | +------+ \ | : : +-------+
995 +-------+ : : \ | +-------+ | |
996 ---------->| B->2 |------>| |
997 | +-------+ | CPU 2 |
1000 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1001 barrier causes all effects \ +-------+ | |
1002 prior to the storage of B ---->| A->1 |------>| |
1003 to be perceptible to CPU 2 +-------+ | |
1007 To illustrate this more completely, consider what could happen if the code
1008 contained a load of A either side of the read barrier:
1011 ======================= =======================
1017 LOAD A [first load of A]
1019 LOAD A [second load of A]
1021 Even though the two loads of A both occur after the load of B, they may both
1022 come up with different values:
1025 | | +------+ +-------+
1026 | |------>| A=1 |------ --->| A->0 |
1027 | | +------+ \ +-------+
1028 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1029 | | +------+ | +-------+
1030 | |------>| B=2 |--- | : :
1031 | | +------+ \ | : : +-------+
1032 +-------+ : : \ | +-------+ | |
1033 ---------->| B->2 |------>| |
1034 | +-------+ | CPU 2 |
1038 | | A->0 |------>| 1st |
1040 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1041 barrier causes all effects \ +-------+ | |
1042 prior to the storage of B ---->| A->1 |------>| 2nd |
1043 to be perceptible to CPU 2 +-------+ | |
1047 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1048 before the read barrier completes anyway:
1051 | | +------+ +-------+
1052 | |------>| A=1 |------ --->| A->0 |
1053 | | +------+ \ +-------+
1054 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1055 | | +------+ | +-------+
1056 | |------>| B=2 |--- | : :
1057 | | +------+ \ | : : +-------+
1058 +-------+ : : \ | +-------+ | |
1059 ---------->| B->2 |------>| |
1060 | +-------+ | CPU 2 |
1064 ---->| A->1 |------>| 1st |
1066 rrrrrrrrrrrrrrrrr | |
1068 | A->1 |------>| 2nd |
1073 The guarantee is that the second load will always come up with A == 1 if the
1074 load of B came up with B == 2. No such guarantee exists for the first load of
1075 A; that may come up with either A == 0 or A == 1.
1078 READ MEMORY BARRIERS VS LOAD SPECULATION
1079 ----------------------------------------
1081 Many CPUs speculate with loads: that is they see that they will need to load an
1082 item from memory, and they find a time where they're not using the bus for any
1083 other loads, and so do the load in advance - even though they haven't actually
1084 got to that point in the instruction execution flow yet. This permits the
1085 actual load instruction to potentially complete immediately because the CPU
1086 already has the value to hand.
1088 It may turn out that the CPU didn't actually need the value - perhaps because a
1089 branch circumvented the load - in which case it can discard the value or just
1090 cache it for later use.
1095 ======================= =======================
1097 DIVIDE } Divide instructions generally
1098 DIVIDE } take a long time to perform
1101 Which might appear as this:
1105 --->| B->2 |------>| |
1109 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1110 division speculates on the +-------+ ~ | |
1114 Once the divisions are complete --> : : ~-->| |
1115 the CPU can then perform the : : | |
1116 LOAD with immediate effect : : +-------+
1119 Placing a read barrier or a data dependency barrier just before the second
1123 ======================= =======================
1130 will force any value speculatively obtained to be reconsidered to an extent
1131 dependent on the type of barrier used. If there was no change made to the
1132 speculated memory location, then the speculated value will just be used:
1136 --->| B->2 |------>| |
1140 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1141 division speculates on the +-------+ ~ | |
1146 rrrrrrrrrrrrrrrr~ | |
1153 but if there was an update or an invalidation from another CPU pending, then
1154 the speculation will be cancelled and the value reloaded:
1158 --->| B->2 |------>| |
1162 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1163 division speculates on the +-------+ ~ | |
1168 rrrrrrrrrrrrrrrrr | |
1170 The speculation is discarded ---> --->| A->1 |------>| |
1171 and an updated value is +-------+ | |
1172 retrieved : : +-------+
1178 Transitivity is a deeply intuitive notion about ordering that is not
1179 always provided by real computer systems. The following example
1180 demonstrates transitivity (also called "cumulativity"):
1183 ======================= ======================= =======================
1185 STORE X=1 LOAD X STORE Y=1
1186 <general barrier> <general barrier>
1189 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1190 This indicates that CPU 2's load from X in some sense follows CPU 1's
1191 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1192 store to Y. The question is then "Can CPU 3's load from X return 0?"
1194 Because CPU 2's load from X in some sense came after CPU 1's store, it
1195 is natural to expect that CPU 3's load from X must therefore return 1.
1196 This expectation is an example of transitivity: if a load executing on
1197 CPU A follows a load from the same variable executing on CPU B, then
1198 CPU A's load must either return the same value that CPU B's load did,
1199 or must return some later value.
1201 In the Linux kernel, use of general memory barriers guarantees
1202 transitivity. Therefore, in the above example, if CPU 2's load from X
1203 returns 1 and its load from Y returns 0, then CPU 3's load from X must
1206 However, transitivity is -not- guaranteed for read or write barriers.
1207 For example, suppose that CPU 2's general barrier in the above example
1208 is changed to a read barrier as shown below:
1211 ======================= ======================= =======================
1213 STORE X=1 LOAD X STORE Y=1
1214 <read barrier> <general barrier>
1217 This substitution destroys transitivity: in this example, it is perfectly
1218 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1219 and CPU 3's load from X to return 0.
1221 The key point is that although CPU 2's read barrier orders its pair
1222 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1223 this example runs on a system where CPUs 1 and 2 share a store buffer
1224 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1225 General barriers are therefore required to ensure that all CPUs agree
1226 on the combined order of CPU 1's and CPU 2's accesses.
1228 To reiterate, if your code requires transitivity, use general barriers
1232 ========================
1233 EXPLICIT KERNEL BARRIERS
1234 ========================
1236 The Linux kernel has a variety of different barriers that act at different
1239 (*) Compiler barrier.
1241 (*) CPU memory barriers.
1243 (*) MMIO write barrier.
1249 The Linux kernel has an explicit compiler barrier function that prevents the
1250 compiler from moving the memory accesses either side of it to the other side:
1254 This is a general barrier -- there are no read-read or write-write variants
1255 of barrier(). However, ACCESS_ONCE() can be thought of as a weak form
1256 for barrier() that affects only the specific accesses flagged by the
1259 The barrier() function has the following effects:
1261 (*) Prevents the compiler from reordering accesses following the
1262 barrier() to precede any accesses preceding the barrier().
1263 One example use for this property is to ease communication between
1264 interrupt-handler code and the code that was interrupted.
1266 (*) Within a loop, forces the compiler to load the variables used
1267 in that loop's conditional on each pass through that loop.
1269 The ACCESS_ONCE() function can prevent any number of optimizations that,
1270 while perfectly safe in single-threaded code, can be fatal in concurrent
1271 code. Here are some examples of these sorts of optimizations:
1273 (*) The compiler is within its rights to reorder loads and stores
1274 to the same variable, and in some cases, the CPU is within its
1275 rights to reorder loads to the same variable. This means that
1281 Might result in an older value of x stored in a[1] than in a[0].
1282 Prevent both the compiler and the CPU from doing this as follows:
1284 a[0] = ACCESS_ONCE(x);
1285 a[1] = ACCESS_ONCE(x);
1287 In short, ACCESS_ONCE() provides cache coherence for accesses from
1288 multiple CPUs to a single variable.
1290 (*) The compiler is within its rights to merge successive loads from
1291 the same variable. Such merging can cause the compiler to "optimize"
1295 do_something_with(tmp);
1297 into the following code, which, although in some sense legitimate
1298 for single-threaded code, is almost certainly not what the developer
1303 do_something_with(tmp);
1305 Use ACCESS_ONCE() to prevent the compiler from doing this to you:
1307 while (tmp = ACCESS_ONCE(a))
1308 do_something_with(tmp);
1310 (*) The compiler is within its rights to reload a variable, for example,
1311 in cases where high register pressure prevents the compiler from
1312 keeping all data of interest in registers. The compiler might
1313 therefore optimize the variable 'tmp' out of our previous example:
1316 do_something_with(tmp);
1318 This could result in the following code, which is perfectly safe in
1319 single-threaded code, but can be fatal in concurrent code:
1322 do_something_with(a);
1324 For example, the optimized version of this code could result in
1325 passing a zero to do_something_with() in the case where the variable
1326 a was modified by some other CPU between the "while" statement and
1327 the call to do_something_with().
1329 Again, use ACCESS_ONCE() to prevent the compiler from doing this:
1331 while (tmp = ACCESS_ONCE(a))
1332 do_something_with(tmp);
1334 Note that if the compiler runs short of registers, it might save
1335 tmp onto the stack. The overhead of this saving and later restoring
1336 is why compilers reload variables. Doing so is perfectly safe for
1337 single-threaded code, so you need to tell the compiler about cases
1338 where it is not safe.
1340 (*) The compiler is within its rights to omit a load entirely if it knows
1341 what the value will be. For example, if the compiler can prove that
1342 the value of variable 'a' is always zero, it can optimize this code:
1345 do_something_with(tmp);
1351 This transformation is a win for single-threaded code because it gets
1352 rid of a load and a branch. The problem is that the compiler will
1353 carry out its proof assuming that the current CPU is the only one
1354 updating variable 'a'. If variable 'a' is shared, then the compiler's
1355 proof will be erroneous. Use ACCESS_ONCE() to tell the compiler
1356 that it doesn't know as much as it thinks it does:
1358 while (tmp = ACCESS_ONCE(a))
1359 do_something_with(tmp);
1361 But please note that the compiler is also closely watching what you
1362 do with the value after the ACCESS_ONCE(). For example, suppose you
1363 do the following and MAX is a preprocessor macro with the value 1:
1365 while ((tmp = ACCESS_ONCE(a)) % MAX)
1366 do_something_with(tmp);
1368 Then the compiler knows that the result of the "%" operator applied
1369 to MAX will always be zero, again allowing the compiler to optimize
1370 the code into near-nonexistence. (It will still load from the
1373 (*) Similarly, the compiler is within its rights to omit a store entirely
1374 if it knows that the variable already has the value being stored.
1375 Again, the compiler assumes that the current CPU is the only one
1376 storing into the variable, which can cause the compiler to do the
1377 wrong thing for shared variables. For example, suppose you have
1381 /* Code that does not store to variable a. */
1384 The compiler sees that the value of variable 'a' is already zero, so
1385 it might well omit the second store. This would come as a fatal
1386 surprise if some other CPU might have stored to variable 'a' in the
1389 Use ACCESS_ONCE() to prevent the compiler from making this sort of
1393 /* Code that does not store to variable a. */
1396 (*) The compiler is within its rights to reorder memory accesses unless
1397 you tell it not to. For example, consider the following interaction
1398 between process-level code and an interrupt handler:
1400 void process_level(void)
1402 msg = get_message();
1406 void interrupt_handler(void)
1409 process_message(msg);
1412 There is nothing to prevent the compiler from transforming
1413 process_level() to the following, in fact, this might well be a
1414 win for single-threaded code:
1416 void process_level(void)
1419 msg = get_message();
1422 If the interrupt occurs between these two statement, then
1423 interrupt_handler() might be passed a garbled msg. Use ACCESS_ONCE()
1424 to prevent this as follows:
1426 void process_level(void)
1428 ACCESS_ONCE(msg) = get_message();
1429 ACCESS_ONCE(flag) = true;
1432 void interrupt_handler(void)
1434 if (ACCESS_ONCE(flag))
1435 process_message(ACCESS_ONCE(msg));
1438 Note that the ACCESS_ONCE() wrappers in interrupt_handler()
1439 are needed if this interrupt handler can itself be interrupted
1440 by something that also accesses 'flag' and 'msg', for example,
1441 a nested interrupt or an NMI. Otherwise, ACCESS_ONCE() is not
1442 needed in interrupt_handler() other than for documentation purposes.
1443 (Note also that nested interrupts do not typically occur in modern
1444 Linux kernels, in fact, if an interrupt handler returns with
1445 interrupts enabled, you will get a WARN_ONCE() splat.)
1447 You should assume that the compiler can move ACCESS_ONCE() past
1448 code not containing ACCESS_ONCE(), barrier(), or similar primitives.
1450 This effect could also be achieved using barrier(), but ACCESS_ONCE()
1451 is more selective: With ACCESS_ONCE(), the compiler need only forget
1452 the contents of the indicated memory locations, while with barrier()
1453 the compiler must discard the value of all memory locations that
1454 it has currented cached in any machine registers. Of course,
1455 the compiler must also respect the order in which the ACCESS_ONCE()s
1456 occur, though the CPU of course need not do so.
1458 (*) The compiler is within its rights to invent stores to a variable,
1459 as in the following example:
1466 The compiler might save a branch by optimizing this as follows:
1472 In single-threaded code, this is not only safe, but also saves
1473 a branch. Unfortunately, in concurrent code, this optimization
1474 could cause some other CPU to see a spurious value of 42 -- even
1475 if variable 'a' was never zero -- when loading variable 'b'.
1476 Use ACCESS_ONCE() to prevent this as follows:
1481 ACCESS_ONCE(b) = 42;
1483 The compiler can also invent loads. These are usually less
1484 damaging, but they can result in cache-line bouncing and thus in
1485 poor performance and scalability. Use ACCESS_ONCE() to prevent
1488 (*) For aligned memory locations whose size allows them to be accessed
1489 with a single memory-reference instruction, prevents "load tearing"
1490 and "store tearing," in which a single large access is replaced by
1491 multiple smaller accesses. For example, given an architecture having
1492 16-bit store instructions with 7-bit immediate fields, the compiler
1493 might be tempted to use two 16-bit store-immediate instructions to
1494 implement the following 32-bit store:
1498 Please note that GCC really does use this sort of optimization,
1499 which is not surprising given that it would likely take more
1500 than two instructions to build the constant and then store it.
1501 This optimization can therefore be a win in single-threaded code.
1502 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1503 this optimization in a volatile store. In the absence of such bugs,
1504 use of ACCESS_ONCE() prevents store tearing in the following example:
1506 ACCESS_ONCE(p) = 0x00010002;
1508 Use of packed structures can also result in load and store tearing,
1511 struct __attribute__((__packed__)) foo {
1516 struct foo foo1, foo2;
1523 Because there are no ACCESS_ONCE() wrappers and no volatile markings,
1524 the compiler would be well within its rights to implement these three
1525 assignment statements as a pair of 32-bit loads followed by a pair
1526 of 32-bit stores. This would result in load tearing on 'foo1.b'
1527 and store tearing on 'foo2.b'. ACCESS_ONCE() again prevents tearing
1531 ACCESS_ONCE(foo2.b) = ACCESS_ONCE(foo1.b);
1534 All that aside, it is never necessary to use ACCESS_ONCE() on a variable
1535 that has been marked volatile. For example, because 'jiffies' is marked
1536 volatile, it is never necessary to say ACCESS_ONCE(jiffies). The reason
1537 for this is that ACCESS_ONCE() is implemented as a volatile cast, which
1538 has no effect when its argument is already marked volatile.
1540 Please note that these compiler barriers have no direct effect on the CPU,
1541 which may then reorder things however it wishes.
1547 The Linux kernel has eight basic CPU memory barriers:
1549 TYPE MANDATORY SMP CONDITIONAL
1550 =============== ======================= ===========================
1551 GENERAL mb() smp_mb()
1552 WRITE wmb() smp_wmb()
1553 READ rmb() smp_rmb()
1554 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1557 All memory barriers except the data dependency barriers imply a compiler
1558 barrier. Data dependencies do not impose any additional compiler ordering.
1560 Aside: In the case of data dependencies, the compiler would be expected to
1561 issue the loads in the correct order (eg. `a[b]` would have to load the value
1562 of b before loading a[b]), however there is no guarantee in the C specification
1563 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1564 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1565 problem of a compiler reloading b after having loaded a[b], thus having a newer
1566 copy of b than a[b]. A consensus has not yet been reached about these problems,
1567 however the ACCESS_ONCE macro is a good place to start looking.
1569 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1570 systems because it is assumed that a CPU will appear to be self-consistent,
1571 and will order overlapping accesses correctly with respect to itself.
1573 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1574 references to shared memory on SMP systems, though the use of locking instead
1577 Mandatory barriers should not be used to control SMP effects, since mandatory
1578 barriers unnecessarily impose overhead on UP systems. They may, however, be
1579 used to control MMIO effects on accesses through relaxed memory I/O windows.
1580 These are required even on non-SMP systems as they affect the order in which
1581 memory operations appear to a device by prohibiting both the compiler and the
1582 CPU from reordering them.
1585 There are some more advanced barrier functions:
1587 (*) set_mb(var, value)
1589 This assigns the value to the variable and then inserts a full memory
1590 barrier after it, depending on the function. It isn't guaranteed to
1591 insert anything more than a compiler barrier in a UP compilation.
1594 (*) smp_mb__before_atomic();
1595 (*) smp_mb__after_atomic();
1597 These are for use with atomic (such as add, subtract, increment and
1598 decrement) functions that don't return a value, especially when used for
1599 reference counting. These functions do not imply memory barriers.
1601 These are also used for atomic bitop functions that do not return a
1602 value (such as set_bit and clear_bit).
1604 As an example, consider a piece of code that marks an object as being dead
1605 and then decrements the object's reference count:
1608 smp_mb__before_atomic();
1609 atomic_dec(&obj->ref_count);
1611 This makes sure that the death mark on the object is perceived to be set
1612 *before* the reference counter is decremented.
1614 See Documentation/atomic_ops.txt for more information. See the "Atomic
1615 operations" subsection for information on where to use these.
1621 The Linux kernel also has a special barrier for use with memory-mapped I/O
1626 This is a variation on the mandatory write barrier that causes writes to weakly
1627 ordered I/O regions to be partially ordered. Its effects may go beyond the
1628 CPU->Hardware interface and actually affect the hardware at some level.
1630 See the subsection "Locks vs I/O accesses" for more information.
1633 ===============================
1634 IMPLICIT KERNEL MEMORY BARRIERS
1635 ===============================
1637 Some of the other functions in the linux kernel imply memory barriers, amongst
1638 which are locking and scheduling functions.
1640 This specification is a _minimum_ guarantee; any particular architecture may
1641 provide more substantial guarantees, but these may not be relied upon outside
1642 of arch specific code.
1648 The Linux kernel has a number of locking constructs:
1657 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1658 for each construct. These operations all imply certain barriers:
1660 (1) ACQUIRE operation implication:
1662 Memory operations issued after the ACQUIRE will be completed after the
1663 ACQUIRE operation has completed.
1665 Memory operations issued before the ACQUIRE may be completed after
1666 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1667 combined with a following ACQUIRE, orders prior loads against
1668 subsequent loads and stores and also orders prior stores against
1669 subsequent stores. Note that this is weaker than smp_mb()! The
1670 smp_mb__before_spinlock() primitive is free on many architectures.
1672 (2) RELEASE operation implication:
1674 Memory operations issued before the RELEASE will be completed before the
1675 RELEASE operation has completed.
1677 Memory operations issued after the RELEASE may be completed before the
1678 RELEASE operation has completed.
1680 (3) ACQUIRE vs ACQUIRE implication:
1682 All ACQUIRE operations issued before another ACQUIRE operation will be
1683 completed before that ACQUIRE operation.
1685 (4) ACQUIRE vs RELEASE implication:
1687 All ACQUIRE operations issued before a RELEASE operation will be
1688 completed before the RELEASE operation.
1690 (5) Failed conditional ACQUIRE implication:
1692 Certain locking variants of the ACQUIRE operation may fail, either due to
1693 being unable to get the lock immediately, or due to receiving an unblocked
1694 signal whilst asleep waiting for the lock to become available. Failed
1695 locks do not imply any sort of barrier.
1697 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1698 one-way barriers is that the effects of instructions outside of a critical
1699 section may seep into the inside of the critical section.
1701 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1702 because it is possible for an access preceding the ACQUIRE to happen after the
1703 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
1704 the two accesses can themselves then cross:
1713 ACQUIRE M, STORE *B, STORE *A, RELEASE M
1715 When the ACQUIRE and RELEASE are a lock acquisition and release,
1716 respectively, this same reordering can occur if the lock's ACQUIRE and
1717 RELEASE are to the same lock variable, but only from the perspective of
1718 another CPU not holding that lock. In short, a ACQUIRE followed by an
1719 RELEASE may -not- be assumed to be a full memory barrier.
1721 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does not
1722 imply a full memory barrier. If it is necessary for a RELEASE-ACQUIRE
1723 pair to produce a full barrier, the ACQUIRE can be followed by an
1724 smp_mb__after_unlock_lock() invocation. This will produce a full barrier
1725 if either (a) the RELEASE and the ACQUIRE are executed by the same
1726 CPU or task, or (b) the RELEASE and ACQUIRE act on the same variable.
1727 The smp_mb__after_unlock_lock() primitive is free on many architectures.
1728 Without smp_mb__after_unlock_lock(), the CPU's execution of the critical
1729 sections corresponding to the RELEASE and the ACQUIRE can cross, so that:
1738 ACQUIRE N, STORE *B, STORE *A, RELEASE M
1740 It might appear that this reordering could introduce a deadlock.
1741 However, this cannot happen because if such a deadlock threatened,
1742 the RELEASE would simply complete, thereby avoiding the deadlock.
1746 One key point is that we are only talking about the CPU doing
1747 the reordering, not the compiler. If the compiler (or, for
1748 that matter, the developer) switched the operations, deadlock
1751 But suppose the CPU reordered the operations. In this case,
1752 the unlock precedes the lock in the assembly code. The CPU
1753 simply elected to try executing the later lock operation first.
1754 If there is a deadlock, this lock operation will simply spin (or
1755 try to sleep, but more on that later). The CPU will eventually
1756 execute the unlock operation (which preceded the lock operation
1757 in the assembly code), which will unravel the potential deadlock,
1758 allowing the lock operation to succeed.
1760 But what if the lock is a sleeplock? In that case, the code will
1761 try to enter the scheduler, where it will eventually encounter
1762 a memory barrier, which will force the earlier unlock operation
1763 to complete, again unraveling the deadlock. There might be
1764 a sleep-unlock race, but the locking primitive needs to resolve
1765 such races properly in any case.
1767 With smp_mb__after_unlock_lock(), the two critical sections cannot overlap.
1768 For example, with the following code, the store to *A will always be
1769 seen by other CPUs before the store to *B:
1774 smp_mb__after_unlock_lock();
1777 The operations will always occur in one of the following orders:
1779 STORE *A, RELEASE, ACQUIRE, smp_mb__after_unlock_lock(), STORE *B
1780 STORE *A, ACQUIRE, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1781 ACQUIRE, STORE *A, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1783 If the RELEASE and ACQUIRE were instead both operating on the same lock
1784 variable, only the first of these alternatives can occur. In addition,
1785 the more strongly ordered systems may rule out some of the above orders.
1786 But in any case, as noted earlier, the smp_mb__after_unlock_lock()
1787 ensures that the store to *A will always be seen as happening before
1790 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1791 systems, and so cannot be counted on in such a situation to actually achieve
1792 anything at all - especially with respect to I/O accesses - unless combined
1793 with interrupt disabling operations.
1795 See also the section on "Inter-CPU locking barrier effects".
1798 As an example, consider the following:
1809 The following sequence of events is acceptable:
1811 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
1813 [+] Note that {*F,*A} indicates a combined access.
1815 But none of the following are:
1817 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
1818 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
1819 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
1820 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
1824 INTERRUPT DISABLING FUNCTIONS
1825 -----------------------------
1827 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
1828 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
1829 barriers are required in such a situation, they must be provided from some
1833 SLEEP AND WAKE-UP FUNCTIONS
1834 ---------------------------
1836 Sleeping and waking on an event flagged in global data can be viewed as an
1837 interaction between two pieces of data: the task state of the task waiting for
1838 the event and the global data used to indicate the event. To make sure that
1839 these appear to happen in the right order, the primitives to begin the process
1840 of going to sleep, and the primitives to initiate a wake up imply certain
1843 Firstly, the sleeper normally follows something like this sequence of events:
1846 set_current_state(TASK_UNINTERRUPTIBLE);
1847 if (event_indicated)
1852 A general memory barrier is interpolated automatically by set_current_state()
1853 after it has altered the task state:
1856 ===============================
1857 set_current_state();
1859 STORE current->state
1861 LOAD event_indicated
1863 set_current_state() may be wrapped by:
1866 prepare_to_wait_exclusive();
1868 which therefore also imply a general memory barrier after setting the state.
1869 The whole sequence above is available in various canned forms, all of which
1870 interpolate the memory barrier in the right place:
1873 wait_event_interruptible();
1874 wait_event_interruptible_exclusive();
1875 wait_event_interruptible_timeout();
1876 wait_event_killable();
1877 wait_event_timeout();
1882 Secondly, code that performs a wake up normally follows something like this:
1884 event_indicated = 1;
1885 wake_up(&event_wait_queue);
1889 event_indicated = 1;
1890 wake_up_process(event_daemon);
1892 A write memory barrier is implied by wake_up() and co. if and only if they wake
1893 something up. The barrier occurs before the task state is cleared, and so sits
1894 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1897 =============================== ===============================
1898 set_current_state(); STORE event_indicated
1899 set_mb(); wake_up();
1900 STORE current->state <write barrier>
1901 <general barrier> STORE current->state
1902 LOAD event_indicated
1904 To repeat, this write memory barrier is present if and only if something
1905 is actually awakened. To see this, consider the following sequence of
1906 events, where X and Y are both initially zero:
1909 =============================== ===============================
1910 X = 1; STORE event_indicated
1911 smp_mb(); wake_up();
1912 Y = 1; wait_event(wq, Y == 1);
1913 wake_up(); load from Y sees 1, no memory barrier
1914 load from X might see 0
1916 In contrast, if a wakeup does occur, CPU 2's load from X would be guaranteed
1919 The available waker functions include:
1925 wake_up_interruptible();
1926 wake_up_interruptible_all();
1927 wake_up_interruptible_nr();
1928 wake_up_interruptible_poll();
1929 wake_up_interruptible_sync();
1930 wake_up_interruptible_sync_poll();
1932 wake_up_locked_poll();
1938 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1939 order multiple stores before the wake-up with respect to loads of those stored
1940 values after the sleeper has called set_current_state(). For instance, if the
1943 set_current_state(TASK_INTERRUPTIBLE);
1944 if (event_indicated)
1946 __set_current_state(TASK_RUNNING);
1947 do_something(my_data);
1952 event_indicated = 1;
1953 wake_up(&event_wait_queue);
1955 there's no guarantee that the change to event_indicated will be perceived by
1956 the sleeper as coming after the change to my_data. In such a circumstance, the
1957 code on both sides must interpolate its own memory barriers between the
1958 separate data accesses. Thus the above sleeper ought to do:
1960 set_current_state(TASK_INTERRUPTIBLE);
1961 if (event_indicated) {
1963 do_something(my_data);
1966 and the waker should do:
1970 event_indicated = 1;
1971 wake_up(&event_wait_queue);
1974 MISCELLANEOUS FUNCTIONS
1975 -----------------------
1977 Other functions that imply barriers:
1979 (*) schedule() and similar imply full memory barriers.
1982 ===================================
1983 INTER-CPU ACQUIRING BARRIER EFFECTS
1984 ===================================
1986 On SMP systems locking primitives give a more substantial form of barrier: one
1987 that does affect memory access ordering on other CPUs, within the context of
1988 conflict on any particular lock.
1991 ACQUIRES VS MEMORY ACCESSES
1992 ---------------------------
1994 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1995 three CPUs; then should the following sequence of events occur:
1998 =============================== ===============================
1999 ACCESS_ONCE(*A) = a; ACCESS_ONCE(*E) = e;
2001 ACCESS_ONCE(*B) = b; ACCESS_ONCE(*F) = f;
2002 ACCESS_ONCE(*C) = c; ACCESS_ONCE(*G) = g;
2004 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*H) = h;
2006 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
2007 through *H occur in, other than the constraints imposed by the separate locks
2008 on the separate CPUs. It might, for example, see:
2010 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
2012 But it won't see any of:
2014 *B, *C or *D preceding ACQUIRE M
2015 *A, *B or *C following RELEASE M
2016 *F, *G or *H preceding ACQUIRE Q
2017 *E, *F or *G following RELEASE Q
2020 However, if the following occurs:
2023 =============================== ===============================
2024 ACCESS_ONCE(*A) = a;
2026 ACCESS_ONCE(*B) = b;
2027 ACCESS_ONCE(*C) = c;
2029 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*E) = e;
2031 smp_mb__after_unlock_lock();
2032 ACCESS_ONCE(*F) = f;
2033 ACCESS_ONCE(*G) = g;
2035 ACCESS_ONCE(*H) = h;
2039 *E, ACQUIRE M [1], *C, *B, *A, RELEASE M [1],
2040 ACQUIRE M [2], *H, *F, *G, RELEASE M [2], *D
2042 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
2044 *B, *C, *D, *F, *G or *H preceding ACQUIRE M [1]
2045 *A, *B or *C following RELEASE M [1]
2046 *F, *G or *H preceding ACQUIRE M [2]
2047 *A, *B, *C, *E, *F or *G following RELEASE M [2]
2049 Note that the smp_mb__after_unlock_lock() is critically important
2050 here: Without it CPU 3 might see some of the above orderings.
2051 Without smp_mb__after_unlock_lock(), the accesses are not guaranteed
2052 to be seen in order unless CPU 3 holds lock M.
2055 ACQUIRES VS I/O ACCESSES
2056 ------------------------
2058 Under certain circumstances (especially involving NUMA), I/O accesses within
2059 two spinlocked sections on two different CPUs may be seen as interleaved by the
2060 PCI bridge, because the PCI bridge does not necessarily participate in the
2061 cache-coherence protocol, and is therefore incapable of issuing the required
2062 read memory barriers.
2067 =============================== ===============================
2077 may be seen by the PCI bridge as follows:
2079 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2081 which would probably cause the hardware to malfunction.
2084 What is necessary here is to intervene with an mmiowb() before dropping the
2085 spinlock, for example:
2088 =============================== ===============================
2100 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2101 before either of the stores issued on CPU 2.
2104 Furthermore, following a store by a load from the same device obviates the need
2105 for the mmiowb(), because the load forces the store to complete before the load
2109 =============================== ===============================
2120 See Documentation/DocBook/deviceiobook.tmpl for more information.
2123 =================================
2124 WHERE ARE MEMORY BARRIERS NEEDED?
2125 =================================
2127 Under normal operation, memory operation reordering is generally not going to
2128 be a problem as a single-threaded linear piece of code will still appear to
2129 work correctly, even if it's in an SMP kernel. There are, however, four
2130 circumstances in which reordering definitely _could_ be a problem:
2132 (*) Interprocessor interaction.
2134 (*) Atomic operations.
2136 (*) Accessing devices.
2141 INTERPROCESSOR INTERACTION
2142 --------------------------
2144 When there's a system with more than one processor, more than one CPU in the
2145 system may be working on the same data set at the same time. This can cause
2146 synchronisation problems, and the usual way of dealing with them is to use
2147 locks. Locks, however, are quite expensive, and so it may be preferable to
2148 operate without the use of a lock if at all possible. In such a case
2149 operations that affect both CPUs may have to be carefully ordered to prevent
2152 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2153 queued on the semaphore, by virtue of it having a piece of its stack linked to
2154 the semaphore's list of waiting processes:
2156 struct rw_semaphore {
2159 struct list_head waiters;
2162 struct rwsem_waiter {
2163 struct list_head list;
2164 struct task_struct *task;
2167 To wake up a particular waiter, the up_read() or up_write() functions have to:
2169 (1) read the next pointer from this waiter's record to know as to where the
2170 next waiter record is;
2172 (2) read the pointer to the waiter's task structure;
2174 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2176 (4) call wake_up_process() on the task; and
2178 (5) release the reference held on the waiter's task struct.
2180 In other words, it has to perform this sequence of events:
2182 LOAD waiter->list.next;
2188 and if any of these steps occur out of order, then the whole thing may
2191 Once it has queued itself and dropped the semaphore lock, the waiter does not
2192 get the lock again; it instead just waits for its task pointer to be cleared
2193 before proceeding. Since the record is on the waiter's stack, this means that
2194 if the task pointer is cleared _before_ the next pointer in the list is read,
2195 another CPU might start processing the waiter and might clobber the waiter's
2196 stack before the up*() function has a chance to read the next pointer.
2198 Consider then what might happen to the above sequence of events:
2201 =============================== ===============================
2208 Woken up by other event
2213 foo() clobbers *waiter
2215 LOAD waiter->list.next;
2218 This could be dealt with using the semaphore lock, but then the down_xxx()
2219 function has to needlessly get the spinlock again after being woken up.
2221 The way to deal with this is to insert a general SMP memory barrier:
2223 LOAD waiter->list.next;
2230 In this case, the barrier makes a guarantee that all memory accesses before the
2231 barrier will appear to happen before all the memory accesses after the barrier
2232 with respect to the other CPUs on the system. It does _not_ guarantee that all
2233 the memory accesses before the barrier will be complete by the time the barrier
2234 instruction itself is complete.
2236 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2237 compiler barrier, thus making sure the compiler emits the instructions in the
2238 right order without actually intervening in the CPU. Since there's only one
2239 CPU, that CPU's dependency ordering logic will take care of everything else.
2245 Whilst they are technically interprocessor interaction considerations, atomic
2246 operations are noted specially as some of them imply full memory barriers and
2247 some don't, but they're very heavily relied on as a group throughout the
2250 Any atomic operation that modifies some state in memory and returns information
2251 about the state (old or new) implies an SMP-conditional general memory barrier
2252 (smp_mb()) on each side of the actual operation (with the exception of
2253 explicit lock operations, described later). These include:
2257 atomic_xchg(); atomic_long_xchg();
2258 atomic_cmpxchg(); atomic_long_cmpxchg();
2259 atomic_inc_return(); atomic_long_inc_return();
2260 atomic_dec_return(); atomic_long_dec_return();
2261 atomic_add_return(); atomic_long_add_return();
2262 atomic_sub_return(); atomic_long_sub_return();
2263 atomic_inc_and_test(); atomic_long_inc_and_test();
2264 atomic_dec_and_test(); atomic_long_dec_and_test();
2265 atomic_sub_and_test(); atomic_long_sub_and_test();
2266 atomic_add_negative(); atomic_long_add_negative();
2268 test_and_clear_bit();
2269 test_and_change_bit();
2271 /* when succeeds (returns 1) */
2272 atomic_add_unless(); atomic_long_add_unless();
2274 These are used for such things as implementing ACQUIRE-class and RELEASE-class
2275 operations and adjusting reference counters towards object destruction, and as
2276 such the implicit memory barrier effects are necessary.
2279 The following operations are potential problems as they do _not_ imply memory
2280 barriers, but might be used for implementing such things as RELEASE-class
2288 With these the appropriate explicit memory barrier should be used if necessary
2289 (smp_mb__before_atomic() for instance).
2292 The following also do _not_ imply memory barriers, and so may require explicit
2293 memory barriers under some circumstances (smp_mb__before_atomic() for
2301 If they're used for statistics generation, then they probably don't need memory
2302 barriers, unless there's a coupling between statistical data.
2304 If they're used for reference counting on an object to control its lifetime,
2305 they probably don't need memory barriers because either the reference count
2306 will be adjusted inside a locked section, or the caller will already hold
2307 sufficient references to make the lock, and thus a memory barrier unnecessary.
2309 If they're used for constructing a lock of some description, then they probably
2310 do need memory barriers as a lock primitive generally has to do things in a
2313 Basically, each usage case has to be carefully considered as to whether memory
2314 barriers are needed or not.
2316 The following operations are special locking primitives:
2318 test_and_set_bit_lock();
2320 __clear_bit_unlock();
2322 These implement ACQUIRE-class and RELEASE-class operations. These should be used in
2323 preference to other operations when implementing locking primitives, because
2324 their implementations can be optimised on many architectures.
2326 [!] Note that special memory barrier primitives are available for these
2327 situations because on some CPUs the atomic instructions used imply full memory
2328 barriers, and so barrier instructions are superfluous in conjunction with them,
2329 and in such cases the special barrier primitives will be no-ops.
2331 See Documentation/atomic_ops.txt for more information.
2337 Many devices can be memory mapped, and so appear to the CPU as if they're just
2338 a set of memory locations. To control such a device, the driver usually has to
2339 make the right memory accesses in exactly the right order.
2341 However, having a clever CPU or a clever compiler creates a potential problem
2342 in that the carefully sequenced accesses in the driver code won't reach the
2343 device in the requisite order if the CPU or the compiler thinks it is more
2344 efficient to reorder, combine or merge accesses - something that would cause
2345 the device to malfunction.
2347 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2348 routines - such as inb() or writel() - which know how to make such accesses
2349 appropriately sequential. Whilst this, for the most part, renders the explicit
2350 use of memory barriers unnecessary, there are a couple of situations where they
2353 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2354 so for _all_ general drivers locks should be used and mmiowb() must be
2355 issued prior to unlocking the critical section.
2357 (2) If the accessor functions are used to refer to an I/O memory window with
2358 relaxed memory access properties, then _mandatory_ memory barriers are
2359 required to enforce ordering.
2361 See Documentation/DocBook/deviceiobook.tmpl for more information.
2367 A driver may be interrupted by its own interrupt service routine, and thus the
2368 two parts of the driver may interfere with each other's attempts to control or
2371 This may be alleviated - at least in part - by disabling local interrupts (a
2372 form of locking), such that the critical operations are all contained within
2373 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2374 routine is executing, the driver's core may not run on the same CPU, and its
2375 interrupt is not permitted to happen again until the current interrupt has been
2376 handled, thus the interrupt handler does not need to lock against that.
2378 However, consider a driver that was talking to an ethernet card that sports an
2379 address register and a data register. If that driver's core talks to the card
2380 under interrupt-disablement and then the driver's interrupt handler is invoked:
2391 The store to the data register might happen after the second store to the
2392 address register if ordering rules are sufficiently relaxed:
2394 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2397 If ordering rules are relaxed, it must be assumed that accesses done inside an
2398 interrupt disabled section may leak outside of it and may interleave with
2399 accesses performed in an interrupt - and vice versa - unless implicit or
2400 explicit barriers are used.
2402 Normally this won't be a problem because the I/O accesses done inside such
2403 sections will include synchronous load operations on strictly ordered I/O
2404 registers that form implicit I/O barriers. If this isn't sufficient then an
2405 mmiowb() may need to be used explicitly.
2408 A similar situation may occur between an interrupt routine and two routines
2409 running on separate CPUs that communicate with each other. If such a case is
2410 likely, then interrupt-disabling locks should be used to guarantee ordering.
2413 ==========================
2414 KERNEL I/O BARRIER EFFECTS
2415 ==========================
2417 When accessing I/O memory, drivers should use the appropriate accessor
2422 These are intended to talk to I/O space rather than memory space, but
2423 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
2424 indeed have special I/O space access cycles and instructions, but many
2425 CPUs don't have such a concept.
2427 The PCI bus, amongst others, defines an I/O space concept which - on such
2428 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2429 space. However, it may also be mapped as a virtual I/O space in the CPU's
2430 memory map, particularly on those CPUs that don't support alternate I/O
2433 Accesses to this space may be fully synchronous (as on i386), but
2434 intermediary bridges (such as the PCI host bridge) may not fully honour
2437 They are guaranteed to be fully ordered with respect to each other.
2439 They are not guaranteed to be fully ordered with respect to other types of
2440 memory and I/O operation.
2442 (*) readX(), writeX():
2444 Whether these are guaranteed to be fully ordered and uncombined with
2445 respect to each other on the issuing CPU depends on the characteristics
2446 defined for the memory window through which they're accessing. On later
2447 i386 architecture machines, for example, this is controlled by way of the
2450 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2451 provided they're not accessing a prefetchable device.
2453 However, intermediary hardware (such as a PCI bridge) may indulge in
2454 deferral if it so wishes; to flush a store, a load from the same location
2455 is preferred[*], but a load from the same device or from configuration
2456 space should suffice for PCI.
2458 [*] NOTE! attempting to load from the same location as was written to may
2459 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2462 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2463 force stores to be ordered.
2465 Please refer to the PCI specification for more information on interactions
2466 between PCI transactions.
2470 These are similar to readX(), but are not guaranteed to be ordered in any
2471 way. Be aware that there is no I/O read barrier available.
2473 (*) ioreadX(), iowriteX()
2475 These will perform appropriately for the type of access they're actually
2476 doing, be it inX()/outX() or readX()/writeX().
2479 ========================================
2480 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2481 ========================================
2483 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2484 maintain the appearance of program causality with respect to itself. Some CPUs
2485 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2486 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2487 of arch-specific code.
2489 This means that it must be considered that the CPU will execute its instruction
2490 stream in any order it feels like - or even in parallel - provided that if an
2491 instruction in the stream depends on an earlier instruction, then that
2492 earlier instruction must be sufficiently complete[*] before the later
2493 instruction may proceed; in other words: provided that the appearance of
2494 causality is maintained.
2496 [*] Some instructions have more than one effect - such as changing the
2497 condition codes, changing registers or changing memory - and different
2498 instructions may depend on different effects.
2500 A CPU may also discard any instruction sequence that winds up having no
2501 ultimate effect. For example, if two adjacent instructions both load an
2502 immediate value into the same register, the first may be discarded.
2505 Similarly, it has to be assumed that compiler might reorder the instruction
2506 stream in any way it sees fit, again provided the appearance of causality is
2510 ============================
2511 THE EFFECTS OF THE CPU CACHE
2512 ============================
2514 The way cached memory operations are perceived across the system is affected to
2515 a certain extent by the caches that lie between CPUs and memory, and by the
2516 memory coherence system that maintains the consistency of state in the system.
2518 As far as the way a CPU interacts with another part of the system through the
2519 caches goes, the memory system has to include the CPU's caches, and memory
2520 barriers for the most part act at the interface between the CPU and its cache
2521 (memory barriers logically act on the dotted line in the following diagram):
2523 <--- CPU ---> : <----------- Memory ----------->
2525 +--------+ +--------+ : +--------+ +-----------+
2526 | | | | : | | | | +--------+
2527 | CPU | | Memory | : | CPU | | | | |
2528 | Core |--->| Access |----->| Cache |<-->| | | |
2529 | | | Queue | : | | | |--->| Memory |
2530 | | | | : | | | | | |
2531 +--------+ +--------+ : +--------+ | | | |
2532 : | Cache | +--------+
2534 : | Mechanism | +--------+
2535 +--------+ +--------+ : +--------+ | | | |
2536 | | | | : | | | | | |
2537 | CPU | | Memory | : | CPU | | |--->| Device |
2538 | Core |--->| Access |----->| Cache |<-->| | | |
2539 | | | Queue | : | | | | | |
2540 | | | | : | | | | +--------+
2541 +--------+ +--------+ : +--------+ +-----------+
2545 Although any particular load or store may not actually appear outside of the
2546 CPU that issued it since it may have been satisfied within the CPU's own cache,
2547 it will still appear as if the full memory access had taken place as far as the
2548 other CPUs are concerned since the cache coherency mechanisms will migrate the
2549 cacheline over to the accessing CPU and propagate the effects upon conflict.
2551 The CPU core may execute instructions in any order it deems fit, provided the
2552 expected program causality appears to be maintained. Some of the instructions
2553 generate load and store operations which then go into the queue of memory
2554 accesses to be performed. The core may place these in the queue in any order
2555 it wishes, and continue execution until it is forced to wait for an instruction
2558 What memory barriers are concerned with is controlling the order in which
2559 accesses cross from the CPU side of things to the memory side of things, and
2560 the order in which the effects are perceived to happen by the other observers
2563 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2564 their own loads and stores as if they had happened in program order.
2566 [!] MMIO or other device accesses may bypass the cache system. This depends on
2567 the properties of the memory window through which devices are accessed and/or
2568 the use of any special device communication instructions the CPU may have.
2574 Life isn't quite as simple as it may appear above, however: for while the
2575 caches are expected to be coherent, there's no guarantee that that coherency
2576 will be ordered. This means that whilst changes made on one CPU will
2577 eventually become visible on all CPUs, there's no guarantee that they will
2578 become apparent in the same order on those other CPUs.
2581 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2582 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2587 +--------+ : +--->| Cache A |<------->| |
2588 | | : | +---------+ | |
2590 | | : | +---------+ | |
2591 +--------+ : +--->| Cache B |<------->| |
2594 : +---------+ | System |
2595 +--------+ : +--->| Cache C |<------->| |
2596 | | : | +---------+ | |
2598 | | : | +---------+ | |
2599 +--------+ : +--->| Cache D |<------->| |
2604 Imagine the system has the following properties:
2606 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2609 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2612 (*) whilst the CPU core is interrogating one cache, the other cache may be
2613 making use of the bus to access the rest of the system - perhaps to
2614 displace a dirty cacheline or to do a speculative load;
2616 (*) each cache has a queue of operations that need to be applied to that cache
2617 to maintain coherency with the rest of the system;
2619 (*) the coherency queue is not flushed by normal loads to lines already
2620 present in the cache, even though the contents of the queue may
2621 potentially affect those loads.
2623 Imagine, then, that two writes are made on the first CPU, with a write barrier
2624 between them to guarantee that they will appear to reach that CPU's caches in
2625 the requisite order:
2628 =============== =============== =======================================
2629 u == 0, v == 1 and p == &u, q == &u
2631 smp_wmb(); Make sure change to v is visible before
2633 <A:modify v=2> v is now in cache A exclusively
2635 <B:modify p=&v> p is now in cache B exclusively
2637 The write memory barrier forces the other CPUs in the system to perceive that
2638 the local CPU's caches have apparently been updated in the correct order. But
2639 now imagine that the second CPU wants to read those values:
2642 =============== =============== =======================================
2647 The above pair of reads may then fail to happen in the expected order, as the
2648 cacheline holding p may get updated in one of the second CPU's caches whilst
2649 the update to the cacheline holding v is delayed in the other of the second
2650 CPU's caches by some other cache event:
2653 =============== =============== =======================================
2654 u == 0, v == 1 and p == &u, q == &u
2657 <A:modify v=2> <C:busy>
2661 <B:modify p=&v> <D:commit p=&v>
2664 <C:read *q> Reads from v before v updated in cache
2668 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2669 no guarantee that, without intervention, the order of update will be the same
2670 as that committed on CPU 1.
2673 To intervene, we need to interpolate a data dependency barrier or a read
2674 barrier between the loads. This will force the cache to commit its coherency
2675 queue before processing any further requests:
2678 =============== =============== =======================================
2679 u == 0, v == 1 and p == &u, q == &u
2682 <A:modify v=2> <C:busy>
2686 <B:modify p=&v> <D:commit p=&v>
2688 smp_read_barrier_depends()
2692 <C:read *q> Reads from v after v updated in cache
2695 This sort of problem can be encountered on DEC Alpha processors as they have a
2696 split cache that improves performance by making better use of the data bus.
2697 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2698 access depends on a read, not all do, so it may not be relied on.
2700 Other CPUs may also have split caches, but must coordinate between the various
2701 cachelets for normal memory accesses. The semantics of the Alpha removes the
2702 need for coordination in the absence of memory barriers.
2705 CACHE COHERENCY VS DMA
2706 ----------------------
2708 Not all systems maintain cache coherency with respect to devices doing DMA. In
2709 such cases, a device attempting DMA may obtain stale data from RAM because
2710 dirty cache lines may be resident in the caches of various CPUs, and may not
2711 have been written back to RAM yet. To deal with this, the appropriate part of
2712 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2713 invalidate them as well).
2715 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2716 cache lines being written back to RAM from a CPU's cache after the device has
2717 installed its own data, or cache lines present in the CPU's cache may simply
2718 obscure the fact that RAM has been updated, until at such time as the cacheline
2719 is discarded from the CPU's cache and reloaded. To deal with this, the
2720 appropriate part of the kernel must invalidate the overlapping bits of the
2723 See Documentation/cachetlb.txt for more information on cache management.
2726 CACHE COHERENCY VS MMIO
2727 -----------------------
2729 Memory mapped I/O usually takes place through memory locations that are part of
2730 a window in the CPU's memory space that has different properties assigned than
2731 the usual RAM directed window.
2733 Amongst these properties is usually the fact that such accesses bypass the
2734 caching entirely and go directly to the device buses. This means MMIO accesses
2735 may, in effect, overtake accesses to cached memory that were emitted earlier.
2736 A memory barrier isn't sufficient in such a case, but rather the cache must be
2737 flushed between the cached memory write and the MMIO access if the two are in
2741 =========================
2742 THE THINGS CPUS GET UP TO
2743 =========================
2745 A programmer might take it for granted that the CPU will perform memory
2746 operations in exactly the order specified, so that if the CPU is, for example,
2747 given the following piece of code to execute:
2749 a = ACCESS_ONCE(*A);
2750 ACCESS_ONCE(*B) = b;
2751 c = ACCESS_ONCE(*C);
2752 d = ACCESS_ONCE(*D);
2753 ACCESS_ONCE(*E) = e;
2755 they would then expect that the CPU will complete the memory operation for each
2756 instruction before moving on to the next one, leading to a definite sequence of
2757 operations as seen by external observers in the system:
2759 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2762 Reality is, of course, much messier. With many CPUs and compilers, the above
2763 assumption doesn't hold because:
2765 (*) loads are more likely to need to be completed immediately to permit
2766 execution progress, whereas stores can often be deferred without a
2769 (*) loads may be done speculatively, and the result discarded should it prove
2770 to have been unnecessary;
2772 (*) loads may be done speculatively, leading to the result having been fetched
2773 at the wrong time in the expected sequence of events;
2775 (*) the order of the memory accesses may be rearranged to promote better use
2776 of the CPU buses and caches;
2778 (*) loads and stores may be combined to improve performance when talking to
2779 memory or I/O hardware that can do batched accesses of adjacent locations,
2780 thus cutting down on transaction setup costs (memory and PCI devices may
2781 both be able to do this); and
2783 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2784 mechanisms may alleviate this - once the store has actually hit the cache
2785 - there's no guarantee that the coherency management will be propagated in
2786 order to other CPUs.
2788 So what another CPU, say, might actually observe from the above piece of code
2791 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2793 (Where "LOAD {*C,*D}" is a combined load)
2796 However, it is guaranteed that a CPU will be self-consistent: it will see its
2797 _own_ accesses appear to be correctly ordered, without the need for a memory
2798 barrier. For instance with the following code:
2800 U = ACCESS_ONCE(*A);
2801 ACCESS_ONCE(*A) = V;
2802 ACCESS_ONCE(*A) = W;
2803 X = ACCESS_ONCE(*A);
2804 ACCESS_ONCE(*A) = Y;
2805 Z = ACCESS_ONCE(*A);
2807 and assuming no intervention by an external influence, it can be assumed that
2808 the final result will appear to be:
2810 U == the original value of *A
2815 The code above may cause the CPU to generate the full sequence of memory
2818 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2820 in that order, but, without intervention, the sequence may have almost any
2821 combination of elements combined or discarded, provided the program's view of
2822 the world remains consistent. Note that ACCESS_ONCE() is -not- optional
2823 in the above example, as there are architectures where a given CPU might
2824 reorder successive loads to the same location. On such architectures,
2825 ACCESS_ONCE() does whatever is necessary to prevent this, for example, on
2826 Itanium the volatile casts used by ACCESS_ONCE() cause GCC to emit the
2827 special ld.acq and st.rel instructions that prevent such reordering.
2829 The compiler may also combine, discard or defer elements of the sequence before
2830 the CPU even sees them.
2841 since, without either a write barrier or an ACCESS_ONCE(), it can be
2842 assumed that the effect of the storage of V to *A is lost. Similarly:
2847 may, without a memory barrier or an ACCESS_ONCE(), be reduced to:
2852 and the LOAD operation never appear outside of the CPU.
2855 AND THEN THERE'S THE ALPHA
2856 --------------------------
2858 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2859 some versions of the Alpha CPU have a split data cache, permitting them to have
2860 two semantically-related cache lines updated at separate times. This is where
2861 the data dependency barrier really becomes necessary as this synchronises both
2862 caches with the memory coherence system, thus making it seem like pointer
2863 changes vs new data occur in the right order.
2865 The Alpha defines the Linux kernel's memory barrier model.
2867 See the subsection on "Cache Coherency" above.
2877 Memory barriers can be used to implement circular buffering without the need
2878 of a lock to serialise the producer with the consumer. See:
2880 Documentation/circular-buffers.txt
2889 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2891 Chapter 5.2: Physical Address Space Characteristics
2892 Chapter 5.4: Caches and Write Buffers
2893 Chapter 5.5: Data Sharing
2894 Chapter 5.6: Read/Write Ordering
2896 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2897 Chapter 7.1: Memory-Access Ordering
2898 Chapter 7.4: Buffering and Combining Memory Writes
2900 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2901 System Programming Guide
2902 Chapter 7.1: Locked Atomic Operations
2903 Chapter 7.2: Memory Ordering
2904 Chapter 7.4: Serializing Instructions
2906 The SPARC Architecture Manual, Version 9
2907 Chapter 8: Memory Models
2908 Appendix D: Formal Specification of the Memory Models
2909 Appendix J: Programming with the Memory Models
2911 UltraSPARC Programmer Reference Manual
2912 Chapter 5: Memory Accesses and Cacheability
2913 Chapter 15: Sparc-V9 Memory Models
2915 UltraSPARC III Cu User's Manual
2916 Chapter 9: Memory Models
2918 UltraSPARC IIIi Processor User's Manual
2919 Chapter 8: Memory Models
2921 UltraSPARC Architecture 2005
2923 Appendix D: Formal Specifications of the Memory Models
2925 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2926 Chapter 8: Memory Models
2927 Appendix F: Caches and Cache Coherency
2929 Solaris Internals, Core Kernel Architecture, p63-68:
2930 Chapter 3.3: Hardware Considerations for Locks and
2933 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2934 for Kernel Programmers:
2935 Chapter 13: Other Memory Models
2937 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2938 Section 2.6: Speculation
2939 Section 4.4: Memory Access
projects / cascardo / linux.git / blob