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:
577 if (ACCESS_ONCE(q)) {
581 Please note that ACCESS_ONCE() is not optional! Without the ACCESS_ONCE(),
582 the compiler is within its rights to transform this example:
586 b = p; /* BUG: Compiler can reorder!!! */
589 b = p; /* BUG: Compiler can reorder!!! */
593 into this, which of course defeats the ordering:
602 Worse yet, if the compiler is able to prove (say) that the value of
603 variable 'a' is always non-zero, it would be well within its rights
604 to optimize the original example by eliminating the "if" statement
608 b = p; /* BUG: Compiler can reorder!!! */
611 The solution is again ACCESS_ONCE() and barrier(), which preserves the
612 ordering between the load from variable 'a' and the store to variable 'b':
625 The initial ACCESS_ONCE() is required to prevent the compiler from
626 proving the value of 'a', and the pair of barrier() invocations are
627 required to prevent the compiler from pulling the two identical stores
628 to 'b' out from the legs of the "if" statement.
630 It is important to note that control dependencies absolutely require a
631 a conditional. For example, the following "optimized" version of
632 the above example breaks ordering, which is why the barrier() invocations
633 are absolutely required if you have identical stores in both legs of
637 ACCESS_ONCE(b) = p; /* BUG: No ordering vs. load from a!!! */
639 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
642 /* ACCESS_ONCE(b) = p; -- moved up, BUG!!! */
646 It is of course legal for the prior load to be part of the conditional,
647 for example, as follows:
649 if (ACCESS_ONCE(a) > 0) {
651 ACCESS_ONCE(b) = q / 2;
655 ACCESS_ONCE(b) = q / 3;
659 This will again ensure that the load from variable 'a' is ordered before the
660 stores to variable 'b'.
662 In addition, you need to be careful what you do with the local variable 'q',
663 otherwise the compiler might be able to guess the value and again remove
664 the needed conditional. For example:
677 If MAX is defined to be 1, then the compiler knows that (q % MAX) is
678 equal to zero, in which case the compiler is within its rights to
679 transform the above code into the following:
685 This transformation loses the ordering between the load from variable 'a'
686 and the store to variable 'b'. If you are relying on this ordering, you
687 should do something like the following:
690 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */
699 Finally, control dependencies do -not- provide transitivity. This is
700 demonstrated by two related examples:
703 ===================== =====================
704 r1 = ACCESS_ONCE(x); r2 = ACCESS_ONCE(y);
705 if (r1 >= 0) if (r2 >= 0)
706 ACCESS_ONCE(y) = 1; ACCESS_ONCE(x) = 1;
708 assert(!(r1 == 1 && r2 == 1));
710 The above two-CPU example will never trigger the assert(). However,
711 if control dependencies guaranteed transitivity (which they do not),
712 then adding the following two CPUs would guarantee a related assertion:
715 ===================== =====================
716 ACCESS_ONCE(x) = 2; ACCESS_ONCE(y) = 2;
718 assert(!(r1 == 2 && r2 == 2 && x == 1 && y == 1)); /* FAILS!!! */
720 But because control dependencies do -not- provide transitivity, the
721 above assertion can fail after the combined four-CPU example completes.
722 If you need the four-CPU example to provide ordering, you will need
723 smp_mb() between the loads and stores in the CPU 0 and CPU 1 code fragments.
727 (*) Control dependencies can order prior loads against later stores.
728 However, they do -not- guarantee any other sort of ordering:
729 Not prior loads against later loads, nor prior stores against
730 later anything. If you need these other forms of ordering,
731 use smb_rmb(), smp_wmb(), or, in the case of prior stores and
732 later loads, smp_mb().
734 (*) If both legs of the "if" statement begin with identical stores
735 to the same variable, a barrier() statement is required at the
736 beginning of each leg of the "if" statement.
738 (*) Control dependencies require at least one run-time conditional
739 between the prior load and the subsequent store, and this
740 conditional must involve the prior load. If the compiler
741 is able to optimize the conditional away, it will have also
742 optimized away the ordering. Careful use of ACCESS_ONCE() can
743 help to preserve the needed conditional.
745 (*) Control dependencies require that the compiler avoid reordering the
746 dependency into nonexistence. Careful use of ACCESS_ONCE() or
747 barrier() can help to preserve your control dependency. Please
748 see the Compiler Barrier section for more information.
750 (*) Control dependencies do -not- provide transitivity. If you
751 need transitivity, use smp_mb().
757 When dealing with CPU-CPU interactions, certain types of memory barrier should
758 always be paired. A lack of appropriate pairing is almost certainly an error.
760 A write barrier should always be paired with a data dependency barrier or read
761 barrier, though a general barrier would also be viable. Similarly a read
762 barrier or a data dependency barrier should always be paired with at least an
763 write barrier, though, again, a general barrier is viable:
766 =============== ===============
769 ACCESS_ONCE(b) = 2; x = ACCESS_ONCE(b);
776 =============== ===============================
779 ACCESS_ONCE(b) = &a; x = ACCESS_ONCE(b);
780 <data dependency barrier>
783 Basically, the read barrier always has to be there, even though it can be of
786 [!] Note that the stores before the write barrier would normally be expected to
787 match the loads after the read barrier or the data dependency barrier, and vice
791 =================== ===================
792 ACCESS_ONCE(a) = 1; }---- --->{ v = ACCESS_ONCE(c);
793 ACCESS_ONCE(b) = 2; } \ / { w = ACCESS_ONCE(d);
794 <write barrier> \ <read barrier>
795 ACCESS_ONCE(c) = 3; } / \ { x = ACCESS_ONCE(a);
796 ACCESS_ONCE(d) = 4; }---- --->{ y = ACCESS_ONCE(b);
799 EXAMPLES OF MEMORY BARRIER SEQUENCES
800 ------------------------------------
802 Firstly, write barriers act as partial orderings on store operations.
803 Consider the following sequence of events:
806 =======================
814 This sequence of events is committed to the memory coherence system in an order
815 that the rest of the system might perceive as the unordered set of { STORE A,
816 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
821 | |------>| C=3 | } /\
822 | | : +------+ }----- \ -----> Events perceptible to
823 | | : | A=1 | } \/ the rest of the system
825 | CPU 1 | : | B=2 | }
827 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
828 | | +------+ } requires all stores prior to the
829 | | : | E=5 | } barrier to be committed before
830 | | : +------+ } further stores may take place
835 | Sequence in which stores are committed to the
836 | memory system by CPU 1
840 Secondly, data dependency barriers act as partial orderings on data-dependent
841 loads. Consider the following sequence of events:
844 ======================= =======================
845 { B = 7; X = 9; Y = 8; C = &Y }
850 STORE D = 4 LOAD C (gets &B)
853 Without intervention, CPU 2 may perceive the events on CPU 1 in some
854 effectively random order, despite the write barrier issued by CPU 1:
857 | | +------+ +-------+ | Sequence of update
858 | |------>| B=2 |----- --->| Y->8 | | of perception on
859 | | : +------+ \ +-------+ | CPU 2
860 | CPU 1 | : | A=1 | \ --->| C->&Y | V
861 | | +------+ | +-------+
862 | | wwwwwwwwwwwwwwww | : :
864 | | : | C=&B |--- | : : +-------+
865 | | : +------+ \ | +-------+ | |
866 | |------>| D=4 | ----------->| C->&B |------>| |
867 | | +------+ | +-------+ | |
868 +-------+ : : | : : | |
872 Apparently incorrect ---> | | B->7 |------>| |
873 perception of B (!) | +-------+ | |
876 The load of X holds ---> \ | X->9 |------>| |
877 up the maintenance \ +-------+ | |
878 of coherence of B ----->| B->2 | +-------+
883 In the above example, CPU 2 perceives that B is 7, despite the load of *C
884 (which would be B) coming after the LOAD of C.
886 If, however, a data dependency barrier were to be placed between the load of C
887 and the load of *C (ie: B) on CPU 2:
890 ======================= =======================
891 { B = 7; X = 9; Y = 8; C = &Y }
896 STORE D = 4 LOAD C (gets &B)
897 <data dependency barrier>
900 then the following will occur:
903 | | +------+ +-------+
904 | |------>| B=2 |----- --->| Y->8 |
905 | | : +------+ \ +-------+
906 | CPU 1 | : | A=1 | \ --->| C->&Y |
907 | | +------+ | +-------+
908 | | wwwwwwwwwwwwwwww | : :
910 | | : | C=&B |--- | : : +-------+
911 | | : +------+ \ | +-------+ | |
912 | |------>| D=4 | ----------->| C->&B |------>| |
913 | | +------+ | +-------+ | |
914 +-------+ : : | : : | |
920 Makes sure all effects ---> \ ddddddddddddddddd | |
921 prior to the store of C \ +-------+ | |
922 are perceptible to ----->| B->2 |------>| |
923 subsequent loads +-------+ | |
927 And thirdly, a read barrier acts as a partial order on loads. Consider the
928 following sequence of events:
931 ======================= =======================
939 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
940 some effectively random order, despite the write barrier issued by CPU 1:
943 | | +------+ +-------+
944 | |------>| A=1 |------ --->| A->0 |
945 | | +------+ \ +-------+
946 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
947 | | +------+ | +-------+
948 | |------>| B=2 |--- | : :
949 | | +------+ \ | : : +-------+
950 +-------+ : : \ | +-------+ | |
951 ---------->| B->2 |------>| |
952 | +-------+ | CPU 2 |
963 If, however, a read barrier were to be placed between the load of B and the
967 ======================= =======================
976 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
980 | | +------+ +-------+
981 | |------>| A=1 |------ --->| A->0 |
982 | | +------+ \ +-------+
983 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
984 | | +------+ | +-------+
985 | |------>| B=2 |--- | : :
986 | | +------+ \ | : : +-------+
987 +-------+ : : \ | +-------+ | |
988 ---------->| B->2 |------>| |
989 | +-------+ | CPU 2 |
992 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
993 barrier causes all effects \ +-------+ | |
994 prior to the storage of B ---->| A->1 |------>| |
995 to be perceptible to CPU 2 +-------+ | |
999 To illustrate this more completely, consider what could happen if the code
1000 contained a load of A either side of the read barrier:
1003 ======================= =======================
1009 LOAD A [first load of A]
1011 LOAD A [second load of A]
1013 Even though the two loads of A both occur after the load of B, they may both
1014 come up with different values:
1017 | | +------+ +-------+
1018 | |------>| A=1 |------ --->| A->0 |
1019 | | +------+ \ +-------+
1020 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1021 | | +------+ | +-------+
1022 | |------>| B=2 |--- | : :
1023 | | +------+ \ | : : +-------+
1024 +-------+ : : \ | +-------+ | |
1025 ---------->| B->2 |------>| |
1026 | +-------+ | CPU 2 |
1030 | | A->0 |------>| 1st |
1032 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
1033 barrier causes all effects \ +-------+ | |
1034 prior to the storage of B ---->| A->1 |------>| 2nd |
1035 to be perceptible to CPU 2 +-------+ | |
1039 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
1040 before the read barrier completes anyway:
1043 | | +------+ +-------+
1044 | |------>| A=1 |------ --->| A->0 |
1045 | | +------+ \ +-------+
1046 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
1047 | | +------+ | +-------+
1048 | |------>| B=2 |--- | : :
1049 | | +------+ \ | : : +-------+
1050 +-------+ : : \ | +-------+ | |
1051 ---------->| B->2 |------>| |
1052 | +-------+ | CPU 2 |
1056 ---->| A->1 |------>| 1st |
1058 rrrrrrrrrrrrrrrrr | |
1060 | A->1 |------>| 2nd |
1065 The guarantee is that the second load will always come up with A == 1 if the
1066 load of B came up with B == 2. No such guarantee exists for the first load of
1067 A; that may come up with either A == 0 or A == 1.
1070 READ MEMORY BARRIERS VS LOAD SPECULATION
1071 ----------------------------------------
1073 Many CPUs speculate with loads: that is they see that they will need to load an
1074 item from memory, and they find a time where they're not using the bus for any
1075 other loads, and so do the load in advance - even though they haven't actually
1076 got to that point in the instruction execution flow yet. This permits the
1077 actual load instruction to potentially complete immediately because the CPU
1078 already has the value to hand.
1080 It may turn out that the CPU didn't actually need the value - perhaps because a
1081 branch circumvented the load - in which case it can discard the value or just
1082 cache it for later use.
1087 ======================= =======================
1089 DIVIDE } Divide instructions generally
1090 DIVIDE } take a long time to perform
1093 Which might appear as this:
1097 --->| B->2 |------>| |
1101 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1102 division speculates on the +-------+ ~ | |
1106 Once the divisions are complete --> : : ~-->| |
1107 the CPU can then perform the : : | |
1108 LOAD with immediate effect : : +-------+
1111 Placing a read barrier or a data dependency barrier just before the second
1115 ======================= =======================
1122 will force any value speculatively obtained to be reconsidered to an extent
1123 dependent on the type of barrier used. If there was no change made to the
1124 speculated memory location, then the speculated value will just be used:
1128 --->| B->2 |------>| |
1132 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1133 division speculates on the +-------+ ~ | |
1138 rrrrrrrrrrrrrrrr~ | |
1145 but if there was an update or an invalidation from another CPU pending, then
1146 the speculation will be cancelled and the value reloaded:
1150 --->| B->2 |------>| |
1154 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
1155 division speculates on the +-------+ ~ | |
1160 rrrrrrrrrrrrrrrrr | |
1162 The speculation is discarded ---> --->| A->1 |------>| |
1163 and an updated value is +-------+ | |
1164 retrieved : : +-------+
1170 Transitivity is a deeply intuitive notion about ordering that is not
1171 always provided by real computer systems. The following example
1172 demonstrates transitivity (also called "cumulativity"):
1175 ======================= ======================= =======================
1177 STORE X=1 LOAD X STORE Y=1
1178 <general barrier> <general barrier>
1181 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
1182 This indicates that CPU 2's load from X in some sense follows CPU 1's
1183 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
1184 store to Y. The question is then "Can CPU 3's load from X return 0?"
1186 Because CPU 2's load from X in some sense came after CPU 1's store, it
1187 is natural to expect that CPU 3's load from X must therefore return 1.
1188 This expectation is an example of transitivity: if a load executing on
1189 CPU A follows a load from the same variable executing on CPU B, then
1190 CPU A's load must either return the same value that CPU B's load did,
1191 or must return some later value.
1193 In the Linux kernel, use of general memory barriers guarantees
1194 transitivity. Therefore, in the above example, if CPU 2's load from X
1195 returns 1 and its load from Y returns 0, then CPU 3's load from X must
1198 However, transitivity is -not- guaranteed for read or write barriers.
1199 For example, suppose that CPU 2's general barrier in the above example
1200 is changed to a read barrier as shown below:
1203 ======================= ======================= =======================
1205 STORE X=1 LOAD X STORE Y=1
1206 <read barrier> <general barrier>
1209 This substitution destroys transitivity: in this example, it is perfectly
1210 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1211 and CPU 3's load from X to return 0.
1213 The key point is that although CPU 2's read barrier orders its pair
1214 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1215 this example runs on a system where CPUs 1 and 2 share a store buffer
1216 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1217 General barriers are therefore required to ensure that all CPUs agree
1218 on the combined order of CPU 1's and CPU 2's accesses.
1220 To reiterate, if your code requires transitivity, use general barriers
1224 ========================
1225 EXPLICIT KERNEL BARRIERS
1226 ========================
1228 The Linux kernel has a variety of different barriers that act at different
1231 (*) Compiler barrier.
1233 (*) CPU memory barriers.
1235 (*) MMIO write barrier.
1241 The Linux kernel has an explicit compiler barrier function that prevents the
1242 compiler from moving the memory accesses either side of it to the other side:
1246 This is a general barrier -- there are no read-read or write-write variants
1247 of barrier(). However, ACCESS_ONCE() can be thought of as a weak form
1248 for barrier() that affects only the specific accesses flagged by the
1251 The barrier() function has the following effects:
1253 (*) Prevents the compiler from reordering accesses following the
1254 barrier() to precede any accesses preceding the barrier().
1255 One example use for this property is to ease communication between
1256 interrupt-handler code and the code that was interrupted.
1258 (*) Within a loop, forces the compiler to load the variables used
1259 in that loop's conditional on each pass through that loop.
1261 The ACCESS_ONCE() function can prevent any number of optimizations that,
1262 while perfectly safe in single-threaded code, can be fatal in concurrent
1263 code. Here are some examples of these sorts of optimizations:
1265 (*) The compiler is within its rights to reorder loads and stores
1266 to the same variable, and in some cases, the CPU is within its
1267 rights to reorder loads to the same variable. This means that
1273 Might result in an older value of x stored in a[1] than in a[0].
1274 Prevent both the compiler and the CPU from doing this as follows:
1276 a[0] = ACCESS_ONCE(x);
1277 a[1] = ACCESS_ONCE(x);
1279 In short, ACCESS_ONCE() provides cache coherence for accesses from
1280 multiple CPUs to a single variable.
1282 (*) The compiler is within its rights to merge successive loads from
1283 the same variable. Such merging can cause the compiler to "optimize"
1287 do_something_with(tmp);
1289 into the following code, which, although in some sense legitimate
1290 for single-threaded code, is almost certainly not what the developer
1295 do_something_with(tmp);
1297 Use ACCESS_ONCE() to prevent the compiler from doing this to you:
1299 while (tmp = ACCESS_ONCE(a))
1300 do_something_with(tmp);
1302 (*) The compiler is within its rights to reload a variable, for example,
1303 in cases where high register pressure prevents the compiler from
1304 keeping all data of interest in registers. The compiler might
1305 therefore optimize the variable 'tmp' out of our previous example:
1308 do_something_with(tmp);
1310 This could result in the following code, which is perfectly safe in
1311 single-threaded code, but can be fatal in concurrent code:
1314 do_something_with(a);
1316 For example, the optimized version of this code could result in
1317 passing a zero to do_something_with() in the case where the variable
1318 a was modified by some other CPU between the "while" statement and
1319 the call to do_something_with().
1321 Again, use ACCESS_ONCE() to prevent the compiler from doing this:
1323 while (tmp = ACCESS_ONCE(a))
1324 do_something_with(tmp);
1326 Note that if the compiler runs short of registers, it might save
1327 tmp onto the stack. The overhead of this saving and later restoring
1328 is why compilers reload variables. Doing so is perfectly safe for
1329 single-threaded code, so you need to tell the compiler about cases
1330 where it is not safe.
1332 (*) The compiler is within its rights to omit a load entirely if it knows
1333 what the value will be. For example, if the compiler can prove that
1334 the value of variable 'a' is always zero, it can optimize this code:
1337 do_something_with(tmp);
1343 This transformation is a win for single-threaded code because it gets
1344 rid of a load and a branch. The problem is that the compiler will
1345 carry out its proof assuming that the current CPU is the only one
1346 updating variable 'a'. If variable 'a' is shared, then the compiler's
1347 proof will be erroneous. Use ACCESS_ONCE() to tell the compiler
1348 that it doesn't know as much as it thinks it does:
1350 while (tmp = ACCESS_ONCE(a))
1351 do_something_with(tmp);
1353 But please note that the compiler is also closely watching what you
1354 do with the value after the ACCESS_ONCE(). For example, suppose you
1355 do the following and MAX is a preprocessor macro with the value 1:
1357 while ((tmp = ACCESS_ONCE(a)) % MAX)
1358 do_something_with(tmp);
1360 Then the compiler knows that the result of the "%" operator applied
1361 to MAX will always be zero, again allowing the compiler to optimize
1362 the code into near-nonexistence. (It will still load from the
1365 (*) Similarly, the compiler is within its rights to omit a store entirely
1366 if it knows that the variable already has the value being stored.
1367 Again, the compiler assumes that the current CPU is the only one
1368 storing into the variable, which can cause the compiler to do the
1369 wrong thing for shared variables. For example, suppose you have
1373 /* Code that does not store to variable a. */
1376 The compiler sees that the value of variable 'a' is already zero, so
1377 it might well omit the second store. This would come as a fatal
1378 surprise if some other CPU might have stored to variable 'a' in the
1381 Use ACCESS_ONCE() to prevent the compiler from making this sort of
1385 /* Code that does not store to variable a. */
1388 (*) The compiler is within its rights to reorder memory accesses unless
1389 you tell it not to. For example, consider the following interaction
1390 between process-level code and an interrupt handler:
1392 void process_level(void)
1394 msg = get_message();
1398 void interrupt_handler(void)
1401 process_message(msg);
1404 There is nothing to prevent the compiler from transforming
1405 process_level() to the following, in fact, this might well be a
1406 win for single-threaded code:
1408 void process_level(void)
1411 msg = get_message();
1414 If the interrupt occurs between these two statement, then
1415 interrupt_handler() might be passed a garbled msg. Use ACCESS_ONCE()
1416 to prevent this as follows:
1418 void process_level(void)
1420 ACCESS_ONCE(msg) = get_message();
1421 ACCESS_ONCE(flag) = true;
1424 void interrupt_handler(void)
1426 if (ACCESS_ONCE(flag))
1427 process_message(ACCESS_ONCE(msg));
1430 Note that the ACCESS_ONCE() wrappers in interrupt_handler()
1431 are needed if this interrupt handler can itself be interrupted
1432 by something that also accesses 'flag' and 'msg', for example,
1433 a nested interrupt or an NMI. Otherwise, ACCESS_ONCE() is not
1434 needed in interrupt_handler() other than for documentation purposes.
1435 (Note also that nested interrupts do not typically occur in modern
1436 Linux kernels, in fact, if an interrupt handler returns with
1437 interrupts enabled, you will get a WARN_ONCE() splat.)
1439 You should assume that the compiler can move ACCESS_ONCE() past
1440 code not containing ACCESS_ONCE(), barrier(), or similar primitives.
1442 This effect could also be achieved using barrier(), but ACCESS_ONCE()
1443 is more selective: With ACCESS_ONCE(), the compiler need only forget
1444 the contents of the indicated memory locations, while with barrier()
1445 the compiler must discard the value of all memory locations that
1446 it has currented cached in any machine registers. Of course,
1447 the compiler must also respect the order in which the ACCESS_ONCE()s
1448 occur, though the CPU of course need not do so.
1450 (*) The compiler is within its rights to invent stores to a variable,
1451 as in the following example:
1458 The compiler might save a branch by optimizing this as follows:
1464 In single-threaded code, this is not only safe, but also saves
1465 a branch. Unfortunately, in concurrent code, this optimization
1466 could cause some other CPU to see a spurious value of 42 -- even
1467 if variable 'a' was never zero -- when loading variable 'b'.
1468 Use ACCESS_ONCE() to prevent this as follows:
1473 ACCESS_ONCE(b) = 42;
1475 The compiler can also invent loads. These are usually less
1476 damaging, but they can result in cache-line bouncing and thus in
1477 poor performance and scalability. Use ACCESS_ONCE() to prevent
1480 (*) For aligned memory locations whose size allows them to be accessed
1481 with a single memory-reference instruction, prevents "load tearing"
1482 and "store tearing," in which a single large access is replaced by
1483 multiple smaller accesses. For example, given an architecture having
1484 16-bit store instructions with 7-bit immediate fields, the compiler
1485 might be tempted to use two 16-bit store-immediate instructions to
1486 implement the following 32-bit store:
1490 Please note that GCC really does use this sort of optimization,
1491 which is not surprising given that it would likely take more
1492 than two instructions to build the constant and then store it.
1493 This optimization can therefore be a win in single-threaded code.
1494 In fact, a recent bug (since fixed) caused GCC to incorrectly use
1495 this optimization in a volatile store. In the absence of such bugs,
1496 use of ACCESS_ONCE() prevents store tearing in the following example:
1498 ACCESS_ONCE(p) = 0x00010002;
1500 Use of packed structures can also result in load and store tearing,
1503 struct __attribute__((__packed__)) foo {
1508 struct foo foo1, foo2;
1515 Because there are no ACCESS_ONCE() wrappers and no volatile markings,
1516 the compiler would be well within its rights to implement these three
1517 assignment statements as a pair of 32-bit loads followed by a pair
1518 of 32-bit stores. This would result in load tearing on 'foo1.b'
1519 and store tearing on 'foo2.b'. ACCESS_ONCE() again prevents tearing
1523 ACCESS_ONCE(foo2.b) = ACCESS_ONCE(foo1.b);
1526 All that aside, it is never necessary to use ACCESS_ONCE() on a variable
1527 that has been marked volatile. For example, because 'jiffies' is marked
1528 volatile, it is never necessary to say ACCESS_ONCE(jiffies). The reason
1529 for this is that ACCESS_ONCE() is implemented as a volatile cast, which
1530 has no effect when its argument is already marked volatile.
1532 Please note that these compiler barriers have no direct effect on the CPU,
1533 which may then reorder things however it wishes.
1539 The Linux kernel has eight basic CPU memory barriers:
1541 TYPE MANDATORY SMP CONDITIONAL
1542 =============== ======================= ===========================
1543 GENERAL mb() smp_mb()
1544 WRITE wmb() smp_wmb()
1545 READ rmb() smp_rmb()
1546 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1549 All memory barriers except the data dependency barriers imply a compiler
1550 barrier. Data dependencies do not impose any additional compiler ordering.
1552 Aside: In the case of data dependencies, the compiler would be expected to
1553 issue the loads in the correct order (eg. `a[b]` would have to load the value
1554 of b before loading a[b]), however there is no guarantee in the C specification
1555 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1556 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1557 problem of a compiler reloading b after having loaded a[b], thus having a newer
1558 copy of b than a[b]. A consensus has not yet been reached about these problems,
1559 however the ACCESS_ONCE macro is a good place to start looking.
1561 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1562 systems because it is assumed that a CPU will appear to be self-consistent,
1563 and will order overlapping accesses correctly with respect to itself.
1565 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1566 references to shared memory on SMP systems, though the use of locking instead
1569 Mandatory barriers should not be used to control SMP effects, since mandatory
1570 barriers unnecessarily impose overhead on UP systems. They may, however, be
1571 used to control MMIO effects on accesses through relaxed memory I/O windows.
1572 These are required even on non-SMP systems as they affect the order in which
1573 memory operations appear to a device by prohibiting both the compiler and the
1574 CPU from reordering them.
1577 There are some more advanced barrier functions:
1579 (*) set_mb(var, value)
1581 This assigns the value to the variable and then inserts a full memory
1582 barrier after it, depending on the function. It isn't guaranteed to
1583 insert anything more than a compiler barrier in a UP compilation.
1586 (*) smp_mb__before_atomic();
1587 (*) smp_mb__after_atomic();
1589 These are for use with atomic (such as add, subtract, increment and
1590 decrement) functions that don't return a value, especially when used for
1591 reference counting. These functions do not imply memory barriers.
1593 These are also used for atomic bitop functions that do not return a
1594 value (such as set_bit and clear_bit).
1596 As an example, consider a piece of code that marks an object as being dead
1597 and then decrements the object's reference count:
1600 smp_mb__before_atomic();
1601 atomic_dec(&obj->ref_count);
1603 This makes sure that the death mark on the object is perceived to be set
1604 *before* the reference counter is decremented.
1606 See Documentation/atomic_ops.txt for more information. See the "Atomic
1607 operations" subsection for information on where to use these.
1613 The Linux kernel also has a special barrier for use with memory-mapped I/O
1618 This is a variation on the mandatory write barrier that causes writes to weakly
1619 ordered I/O regions to be partially ordered. Its effects may go beyond the
1620 CPU->Hardware interface and actually affect the hardware at some level.
1622 See the subsection "Locks vs I/O accesses" for more information.
1625 ===============================
1626 IMPLICIT KERNEL MEMORY BARRIERS
1627 ===============================
1629 Some of the other functions in the linux kernel imply memory barriers, amongst
1630 which are locking and scheduling functions.
1632 This specification is a _minimum_ guarantee; any particular architecture may
1633 provide more substantial guarantees, but these may not be relied upon outside
1634 of arch specific code.
1640 The Linux kernel has a number of locking constructs:
1649 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations
1650 for each construct. These operations all imply certain barriers:
1652 (1) ACQUIRE operation implication:
1654 Memory operations issued after the ACQUIRE will be completed after the
1655 ACQUIRE operation has completed.
1657 Memory operations issued before the ACQUIRE may be completed after
1658 the ACQUIRE operation has completed. An smp_mb__before_spinlock(),
1659 combined with a following ACQUIRE, orders prior loads against
1660 subsequent loads and stores and also orders prior stores against
1661 subsequent stores. Note that this is weaker than smp_mb()! The
1662 smp_mb__before_spinlock() primitive is free on many architectures.
1664 (2) RELEASE operation implication:
1666 Memory operations issued before the RELEASE will be completed before the
1667 RELEASE operation has completed.
1669 Memory operations issued after the RELEASE may be completed before the
1670 RELEASE operation has completed.
1672 (3) ACQUIRE vs ACQUIRE implication:
1674 All ACQUIRE operations issued before another ACQUIRE operation will be
1675 completed before that ACQUIRE operation.
1677 (4) ACQUIRE vs RELEASE implication:
1679 All ACQUIRE operations issued before a RELEASE operation will be
1680 completed before the RELEASE operation.
1682 (5) Failed conditional ACQUIRE implication:
1684 Certain locking variants of the ACQUIRE operation may fail, either due to
1685 being unable to get the lock immediately, or due to receiving an unblocked
1686 signal whilst asleep waiting for the lock to become available. Failed
1687 locks do not imply any sort of barrier.
1689 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only
1690 one-way barriers is that the effects of instructions outside of a critical
1691 section may seep into the inside of the critical section.
1693 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier
1694 because it is possible for an access preceding the ACQUIRE to happen after the
1695 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and
1696 the two accesses can themselves then cross:
1705 ACQUIRE M, STORE *B, STORE *A, RELEASE M
1707 When the ACQUIRE and RELEASE are a lock acquisition and release,
1708 respectively, this same reordering can occur if the lock's ACQUIRE and
1709 RELEASE are to the same lock variable, but only from the perspective of
1710 another CPU not holding that lock. In short, a ACQUIRE followed by an
1711 RELEASE may -not- be assumed to be a full memory barrier.
1713 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does not
1714 imply a full memory barrier. If it is necessary for a RELEASE-ACQUIRE
1715 pair to produce a full barrier, the ACQUIRE can be followed by an
1716 smp_mb__after_unlock_lock() invocation. This will produce a full barrier
1717 if either (a) the RELEASE and the ACQUIRE are executed by the same
1718 CPU or task, or (b) the RELEASE and ACQUIRE act on the same variable.
1719 The smp_mb__after_unlock_lock() primitive is free on many architectures.
1720 Without smp_mb__after_unlock_lock(), the CPU's execution of the critical
1721 sections corresponding to the RELEASE and the ACQUIRE can cross, so that:
1730 ACQUIRE N, STORE *B, STORE *A, RELEASE M
1732 It might appear that this reordering could introduce a deadlock.
1733 However, this cannot happen because if such a deadlock threatened,
1734 the RELEASE would simply complete, thereby avoiding the deadlock.
1738 One key point is that we are only talking about the CPU doing
1739 the reordering, not the compiler. If the compiler (or, for
1740 that matter, the developer) switched the operations, deadlock
1743 But suppose the CPU reordered the operations. In this case,
1744 the unlock precedes the lock in the assembly code. The CPU
1745 simply elected to try executing the later lock operation first.
1746 If there is a deadlock, this lock operation will simply spin (or
1747 try to sleep, but more on that later). The CPU will eventually
1748 execute the unlock operation (which preceded the lock operation
1749 in the assembly code), which will unravel the potential deadlock,
1750 allowing the lock operation to succeed.
1752 But what if the lock is a sleeplock? In that case, the code will
1753 try to enter the scheduler, where it will eventually encounter
1754 a memory barrier, which will force the earlier unlock operation
1755 to complete, again unraveling the deadlock. There might be
1756 a sleep-unlock race, but the locking primitive needs to resolve
1757 such races properly in any case.
1759 With smp_mb__after_unlock_lock(), the two critical sections cannot overlap.
1760 For example, with the following code, the store to *A will always be
1761 seen by other CPUs before the store to *B:
1766 smp_mb__after_unlock_lock();
1769 The operations will always occur in one of the following orders:
1771 STORE *A, RELEASE, ACQUIRE, smp_mb__after_unlock_lock(), STORE *B
1772 STORE *A, ACQUIRE, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1773 ACQUIRE, STORE *A, RELEASE, smp_mb__after_unlock_lock(), STORE *B
1775 If the RELEASE and ACQUIRE were instead both operating on the same lock
1776 variable, only the first of these alternatives can occur. In addition,
1777 the more strongly ordered systems may rule out some of the above orders.
1778 But in any case, as noted earlier, the smp_mb__after_unlock_lock()
1779 ensures that the store to *A will always be seen as happening before
1782 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1783 systems, and so cannot be counted on in such a situation to actually achieve
1784 anything at all - especially with respect to I/O accesses - unless combined
1785 with interrupt disabling operations.
1787 See also the section on "Inter-CPU locking barrier effects".
1790 As an example, consider the following:
1801 The following sequence of events is acceptable:
1803 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE
1805 [+] Note that {*F,*A} indicates a combined access.
1807 But none of the following are:
1809 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E
1810 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F
1811 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F
1812 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E
1816 INTERRUPT DISABLING FUNCTIONS
1817 -----------------------------
1819 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts
1820 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O
1821 barriers are required in such a situation, they must be provided from some
1825 SLEEP AND WAKE-UP FUNCTIONS
1826 ---------------------------
1828 Sleeping and waking on an event flagged in global data can be viewed as an
1829 interaction between two pieces of data: the task state of the task waiting for
1830 the event and the global data used to indicate the event. To make sure that
1831 these appear to happen in the right order, the primitives to begin the process
1832 of going to sleep, and the primitives to initiate a wake up imply certain
1835 Firstly, the sleeper normally follows something like this sequence of events:
1838 set_current_state(TASK_UNINTERRUPTIBLE);
1839 if (event_indicated)
1844 A general memory barrier is interpolated automatically by set_current_state()
1845 after it has altered the task state:
1848 ===============================
1849 set_current_state();
1851 STORE current->state
1853 LOAD event_indicated
1855 set_current_state() may be wrapped by:
1858 prepare_to_wait_exclusive();
1860 which therefore also imply a general memory barrier after setting the state.
1861 The whole sequence above is available in various canned forms, all of which
1862 interpolate the memory barrier in the right place:
1865 wait_event_interruptible();
1866 wait_event_interruptible_exclusive();
1867 wait_event_interruptible_timeout();
1868 wait_event_killable();
1869 wait_event_timeout();
1874 Secondly, code that performs a wake up normally follows something like this:
1876 event_indicated = 1;
1877 wake_up(&event_wait_queue);
1881 event_indicated = 1;
1882 wake_up_process(event_daemon);
1884 A write memory barrier is implied by wake_up() and co. if and only if they wake
1885 something up. The barrier occurs before the task state is cleared, and so sits
1886 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1889 =============================== ===============================
1890 set_current_state(); STORE event_indicated
1891 set_mb(); wake_up();
1892 STORE current->state <write barrier>
1893 <general barrier> STORE current->state
1894 LOAD event_indicated
1896 The available waker functions include:
1902 wake_up_interruptible();
1903 wake_up_interruptible_all();
1904 wake_up_interruptible_nr();
1905 wake_up_interruptible_poll();
1906 wake_up_interruptible_sync();
1907 wake_up_interruptible_sync_poll();
1909 wake_up_locked_poll();
1915 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1916 order multiple stores before the wake-up with respect to loads of those stored
1917 values after the sleeper has called set_current_state(). For instance, if the
1920 set_current_state(TASK_INTERRUPTIBLE);
1921 if (event_indicated)
1923 __set_current_state(TASK_RUNNING);
1924 do_something(my_data);
1929 event_indicated = 1;
1930 wake_up(&event_wait_queue);
1932 there's no guarantee that the change to event_indicated will be perceived by
1933 the sleeper as coming after the change to my_data. In such a circumstance, the
1934 code on both sides must interpolate its own memory barriers between the
1935 separate data accesses. Thus the above sleeper ought to do:
1937 set_current_state(TASK_INTERRUPTIBLE);
1938 if (event_indicated) {
1940 do_something(my_data);
1943 and the waker should do:
1947 event_indicated = 1;
1948 wake_up(&event_wait_queue);
1951 MISCELLANEOUS FUNCTIONS
1952 -----------------------
1954 Other functions that imply barriers:
1956 (*) schedule() and similar imply full memory barriers.
1959 ===================================
1960 INTER-CPU ACQUIRING BARRIER EFFECTS
1961 ===================================
1963 On SMP systems locking primitives give a more substantial form of barrier: one
1964 that does affect memory access ordering on other CPUs, within the context of
1965 conflict on any particular lock.
1968 ACQUIRES VS MEMORY ACCESSES
1969 ---------------------------
1971 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1972 three CPUs; then should the following sequence of events occur:
1975 =============================== ===============================
1976 ACCESS_ONCE(*A) = a; ACCESS_ONCE(*E) = e;
1978 ACCESS_ONCE(*B) = b; ACCESS_ONCE(*F) = f;
1979 ACCESS_ONCE(*C) = c; ACCESS_ONCE(*G) = g;
1981 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*H) = h;
1983 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
1984 through *H occur in, other than the constraints imposed by the separate locks
1985 on the separate CPUs. It might, for example, see:
1987 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M
1989 But it won't see any of:
1991 *B, *C or *D preceding ACQUIRE M
1992 *A, *B or *C following RELEASE M
1993 *F, *G or *H preceding ACQUIRE Q
1994 *E, *F or *G following RELEASE Q
1997 However, if the following occurs:
2000 =============================== ===============================
2001 ACCESS_ONCE(*A) = a;
2003 ACCESS_ONCE(*B) = b;
2004 ACCESS_ONCE(*C) = c;
2006 ACCESS_ONCE(*D) = d; ACCESS_ONCE(*E) = e;
2008 smp_mb__after_unlock_lock();
2009 ACCESS_ONCE(*F) = f;
2010 ACCESS_ONCE(*G) = g;
2012 ACCESS_ONCE(*H) = h;
2016 *E, ACQUIRE M [1], *C, *B, *A, RELEASE M [1],
2017 ACQUIRE M [2], *H, *F, *G, RELEASE M [2], *D
2019 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
2021 *B, *C, *D, *F, *G or *H preceding ACQUIRE M [1]
2022 *A, *B or *C following RELEASE M [1]
2023 *F, *G or *H preceding ACQUIRE M [2]
2024 *A, *B, *C, *E, *F or *G following RELEASE M [2]
2026 Note that the smp_mb__after_unlock_lock() is critically important
2027 here: Without it CPU 3 might see some of the above orderings.
2028 Without smp_mb__after_unlock_lock(), the accesses are not guaranteed
2029 to be seen in order unless CPU 3 holds lock M.
2032 ACQUIRES VS I/O ACCESSES
2033 ------------------------
2035 Under certain circumstances (especially involving NUMA), I/O accesses within
2036 two spinlocked sections on two different CPUs may be seen as interleaved by the
2037 PCI bridge, because the PCI bridge does not necessarily participate in the
2038 cache-coherence protocol, and is therefore incapable of issuing the required
2039 read memory barriers.
2044 =============================== ===============================
2054 may be seen by the PCI bridge as follows:
2056 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
2058 which would probably cause the hardware to malfunction.
2061 What is necessary here is to intervene with an mmiowb() before dropping the
2062 spinlock, for example:
2065 =============================== ===============================
2077 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
2078 before either of the stores issued on CPU 2.
2081 Furthermore, following a store by a load from the same device obviates the need
2082 for the mmiowb(), because the load forces the store to complete before the load
2086 =============================== ===============================
2097 See Documentation/DocBook/deviceiobook.tmpl for more information.
2100 =================================
2101 WHERE ARE MEMORY BARRIERS NEEDED?
2102 =================================
2104 Under normal operation, memory operation reordering is generally not going to
2105 be a problem as a single-threaded linear piece of code will still appear to
2106 work correctly, even if it's in an SMP kernel. There are, however, four
2107 circumstances in which reordering definitely _could_ be a problem:
2109 (*) Interprocessor interaction.
2111 (*) Atomic operations.
2113 (*) Accessing devices.
2118 INTERPROCESSOR INTERACTION
2119 --------------------------
2121 When there's a system with more than one processor, more than one CPU in the
2122 system may be working on the same data set at the same time. This can cause
2123 synchronisation problems, and the usual way of dealing with them is to use
2124 locks. Locks, however, are quite expensive, and so it may be preferable to
2125 operate without the use of a lock if at all possible. In such a case
2126 operations that affect both CPUs may have to be carefully ordered to prevent
2129 Consider, for example, the R/W semaphore slow path. Here a waiting process is
2130 queued on the semaphore, by virtue of it having a piece of its stack linked to
2131 the semaphore's list of waiting processes:
2133 struct rw_semaphore {
2136 struct list_head waiters;
2139 struct rwsem_waiter {
2140 struct list_head list;
2141 struct task_struct *task;
2144 To wake up a particular waiter, the up_read() or up_write() functions have to:
2146 (1) read the next pointer from this waiter's record to know as to where the
2147 next waiter record is;
2149 (2) read the pointer to the waiter's task structure;
2151 (3) clear the task pointer to tell the waiter it has been given the semaphore;
2153 (4) call wake_up_process() on the task; and
2155 (5) release the reference held on the waiter's task struct.
2157 In other words, it has to perform this sequence of events:
2159 LOAD waiter->list.next;
2165 and if any of these steps occur out of order, then the whole thing may
2168 Once it has queued itself and dropped the semaphore lock, the waiter does not
2169 get the lock again; it instead just waits for its task pointer to be cleared
2170 before proceeding. Since the record is on the waiter's stack, this means that
2171 if the task pointer is cleared _before_ the next pointer in the list is read,
2172 another CPU might start processing the waiter and might clobber the waiter's
2173 stack before the up*() function has a chance to read the next pointer.
2175 Consider then what might happen to the above sequence of events:
2178 =============================== ===============================
2185 Woken up by other event
2190 foo() clobbers *waiter
2192 LOAD waiter->list.next;
2195 This could be dealt with using the semaphore lock, but then the down_xxx()
2196 function has to needlessly get the spinlock again after being woken up.
2198 The way to deal with this is to insert a general SMP memory barrier:
2200 LOAD waiter->list.next;
2207 In this case, the barrier makes a guarantee that all memory accesses before the
2208 barrier will appear to happen before all the memory accesses after the barrier
2209 with respect to the other CPUs on the system. It does _not_ guarantee that all
2210 the memory accesses before the barrier will be complete by the time the barrier
2211 instruction itself is complete.
2213 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
2214 compiler barrier, thus making sure the compiler emits the instructions in the
2215 right order without actually intervening in the CPU. Since there's only one
2216 CPU, that CPU's dependency ordering logic will take care of everything else.
2222 Whilst they are technically interprocessor interaction considerations, atomic
2223 operations are noted specially as some of them imply full memory barriers and
2224 some don't, but they're very heavily relied on as a group throughout the
2227 Any atomic operation that modifies some state in memory and returns information
2228 about the state (old or new) implies an SMP-conditional general memory barrier
2229 (smp_mb()) on each side of the actual operation (with the exception of
2230 explicit lock operations, described later). These include:
2234 atomic_xchg(); atomic_long_xchg();
2235 atomic_cmpxchg(); atomic_long_cmpxchg();
2236 atomic_inc_return(); atomic_long_inc_return();
2237 atomic_dec_return(); atomic_long_dec_return();
2238 atomic_add_return(); atomic_long_add_return();
2239 atomic_sub_return(); atomic_long_sub_return();
2240 atomic_inc_and_test(); atomic_long_inc_and_test();
2241 atomic_dec_and_test(); atomic_long_dec_and_test();
2242 atomic_sub_and_test(); atomic_long_sub_and_test();
2243 atomic_add_negative(); atomic_long_add_negative();
2245 test_and_clear_bit();
2246 test_and_change_bit();
2248 /* when succeeds (returns 1) */
2249 atomic_add_unless(); atomic_long_add_unless();
2251 These are used for such things as implementing ACQUIRE-class and RELEASE-class
2252 operations and adjusting reference counters towards object destruction, and as
2253 such the implicit memory barrier effects are necessary.
2256 The following operations are potential problems as they do _not_ imply memory
2257 barriers, but might be used for implementing such things as RELEASE-class
2265 With these the appropriate explicit memory barrier should be used if necessary
2266 (smp_mb__before_atomic() for instance).
2269 The following also do _not_ imply memory barriers, and so may require explicit
2270 memory barriers under some circumstances (smp_mb__before_atomic() for
2278 If they're used for statistics generation, then they probably don't need memory
2279 barriers, unless there's a coupling between statistical data.
2281 If they're used for reference counting on an object to control its lifetime,
2282 they probably don't need memory barriers because either the reference count
2283 will be adjusted inside a locked section, or the caller will already hold
2284 sufficient references to make the lock, and thus a memory barrier unnecessary.
2286 If they're used for constructing a lock of some description, then they probably
2287 do need memory barriers as a lock primitive generally has to do things in a
2290 Basically, each usage case has to be carefully considered as to whether memory
2291 barriers are needed or not.
2293 The following operations are special locking primitives:
2295 test_and_set_bit_lock();
2297 __clear_bit_unlock();
2299 These implement ACQUIRE-class and RELEASE-class operations. These should be used in
2300 preference to other operations when implementing locking primitives, because
2301 their implementations can be optimised on many architectures.
2303 [!] Note that special memory barrier primitives are available for these
2304 situations because on some CPUs the atomic instructions used imply full memory
2305 barriers, and so barrier instructions are superfluous in conjunction with them,
2306 and in such cases the special barrier primitives will be no-ops.
2308 See Documentation/atomic_ops.txt for more information.
2314 Many devices can be memory mapped, and so appear to the CPU as if they're just
2315 a set of memory locations. To control such a device, the driver usually has to
2316 make the right memory accesses in exactly the right order.
2318 However, having a clever CPU or a clever compiler creates a potential problem
2319 in that the carefully sequenced accesses in the driver code won't reach the
2320 device in the requisite order if the CPU or the compiler thinks it is more
2321 efficient to reorder, combine or merge accesses - something that would cause
2322 the device to malfunction.
2324 Inside of the Linux kernel, I/O should be done through the appropriate accessor
2325 routines - such as inb() or writel() - which know how to make such accesses
2326 appropriately sequential. Whilst this, for the most part, renders the explicit
2327 use of memory barriers unnecessary, there are a couple of situations where they
2330 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
2331 so for _all_ general drivers locks should be used and mmiowb() must be
2332 issued prior to unlocking the critical section.
2334 (2) If the accessor functions are used to refer to an I/O memory window with
2335 relaxed memory access properties, then _mandatory_ memory barriers are
2336 required to enforce ordering.
2338 See Documentation/DocBook/deviceiobook.tmpl for more information.
2344 A driver may be interrupted by its own interrupt service routine, and thus the
2345 two parts of the driver may interfere with each other's attempts to control or
2348 This may be alleviated - at least in part - by disabling local interrupts (a
2349 form of locking), such that the critical operations are all contained within
2350 the interrupt-disabled section in the driver. Whilst the driver's interrupt
2351 routine is executing, the driver's core may not run on the same CPU, and its
2352 interrupt is not permitted to happen again until the current interrupt has been
2353 handled, thus the interrupt handler does not need to lock against that.
2355 However, consider a driver that was talking to an ethernet card that sports an
2356 address register and a data register. If that driver's core talks to the card
2357 under interrupt-disablement and then the driver's interrupt handler is invoked:
2368 The store to the data register might happen after the second store to the
2369 address register if ordering rules are sufficiently relaxed:
2371 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
2374 If ordering rules are relaxed, it must be assumed that accesses done inside an
2375 interrupt disabled section may leak outside of it and may interleave with
2376 accesses performed in an interrupt - and vice versa - unless implicit or
2377 explicit barriers are used.
2379 Normally this won't be a problem because the I/O accesses done inside such
2380 sections will include synchronous load operations on strictly ordered I/O
2381 registers that form implicit I/O barriers. If this isn't sufficient then an
2382 mmiowb() may need to be used explicitly.
2385 A similar situation may occur between an interrupt routine and two routines
2386 running on separate CPUs that communicate with each other. If such a case is
2387 likely, then interrupt-disabling locks should be used to guarantee ordering.
2390 ==========================
2391 KERNEL I/O BARRIER EFFECTS
2392 ==========================
2394 When accessing I/O memory, drivers should use the appropriate accessor
2399 These are intended to talk to I/O space rather than memory space, but
2400 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
2401 indeed have special I/O space access cycles and instructions, but many
2402 CPUs don't have such a concept.
2404 The PCI bus, amongst others, defines an I/O space concept which - on such
2405 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
2406 space. However, it may also be mapped as a virtual I/O space in the CPU's
2407 memory map, particularly on those CPUs that don't support alternate I/O
2410 Accesses to this space may be fully synchronous (as on i386), but
2411 intermediary bridges (such as the PCI host bridge) may not fully honour
2414 They are guaranteed to be fully ordered with respect to each other.
2416 They are not guaranteed to be fully ordered with respect to other types of
2417 memory and I/O operation.
2419 (*) readX(), writeX():
2421 Whether these are guaranteed to be fully ordered and uncombined with
2422 respect to each other on the issuing CPU depends on the characteristics
2423 defined for the memory window through which they're accessing. On later
2424 i386 architecture machines, for example, this is controlled by way of the
2427 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
2428 provided they're not accessing a prefetchable device.
2430 However, intermediary hardware (such as a PCI bridge) may indulge in
2431 deferral if it so wishes; to flush a store, a load from the same location
2432 is preferred[*], but a load from the same device or from configuration
2433 space should suffice for PCI.
2435 [*] NOTE! attempting to load from the same location as was written to may
2436 cause a malfunction - consider the 16550 Rx/Tx serial registers for
2439 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
2440 force stores to be ordered.
2442 Please refer to the PCI specification for more information on interactions
2443 between PCI transactions.
2447 These are similar to readX(), but are not guaranteed to be ordered in any
2448 way. Be aware that there is no I/O read barrier available.
2450 (*) ioreadX(), iowriteX()
2452 These will perform appropriately for the type of access they're actually
2453 doing, be it inX()/outX() or readX()/writeX().
2456 ========================================
2457 ASSUMED MINIMUM EXECUTION ORDERING MODEL
2458 ========================================
2460 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
2461 maintain the appearance of program causality with respect to itself. Some CPUs
2462 (such as i386 or x86_64) are more constrained than others (such as powerpc or
2463 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
2464 of arch-specific code.
2466 This means that it must be considered that the CPU will execute its instruction
2467 stream in any order it feels like - or even in parallel - provided that if an
2468 instruction in the stream depends on an earlier instruction, then that
2469 earlier instruction must be sufficiently complete[*] before the later
2470 instruction may proceed; in other words: provided that the appearance of
2471 causality is maintained.
2473 [*] Some instructions have more than one effect - such as changing the
2474 condition codes, changing registers or changing memory - and different
2475 instructions may depend on different effects.
2477 A CPU may also discard any instruction sequence that winds up having no
2478 ultimate effect. For example, if two adjacent instructions both load an
2479 immediate value into the same register, the first may be discarded.
2482 Similarly, it has to be assumed that compiler might reorder the instruction
2483 stream in any way it sees fit, again provided the appearance of causality is
2487 ============================
2488 THE EFFECTS OF THE CPU CACHE
2489 ============================
2491 The way cached memory operations are perceived across the system is affected to
2492 a certain extent by the caches that lie between CPUs and memory, and by the
2493 memory coherence system that maintains the consistency of state in the system.
2495 As far as the way a CPU interacts with another part of the system through the
2496 caches goes, the memory system has to include the CPU's caches, and memory
2497 barriers for the most part act at the interface between the CPU and its cache
2498 (memory barriers logically act on the dotted line in the following diagram):
2500 <--- CPU ---> : <----------- Memory ----------->
2502 +--------+ +--------+ : +--------+ +-----------+
2503 | | | | : | | | | +--------+
2504 | CPU | | Memory | : | CPU | | | | |
2505 | Core |--->| Access |----->| Cache |<-->| | | |
2506 | | | Queue | : | | | |--->| Memory |
2507 | | | | : | | | | | |
2508 +--------+ +--------+ : +--------+ | | | |
2509 : | Cache | +--------+
2511 : | Mechanism | +--------+
2512 +--------+ +--------+ : +--------+ | | | |
2513 | | | | : | | | | | |
2514 | CPU | | Memory | : | CPU | | |--->| Device |
2515 | Core |--->| Access |----->| Cache |<-->| | | |
2516 | | | Queue | : | | | | | |
2517 | | | | : | | | | +--------+
2518 +--------+ +--------+ : +--------+ +-----------+
2522 Although any particular load or store may not actually appear outside of the
2523 CPU that issued it since it may have been satisfied within the CPU's own cache,
2524 it will still appear as if the full memory access had taken place as far as the
2525 other CPUs are concerned since the cache coherency mechanisms will migrate the
2526 cacheline over to the accessing CPU and propagate the effects upon conflict.
2528 The CPU core may execute instructions in any order it deems fit, provided the
2529 expected program causality appears to be maintained. Some of the instructions
2530 generate load and store operations which then go into the queue of memory
2531 accesses to be performed. The core may place these in the queue in any order
2532 it wishes, and continue execution until it is forced to wait for an instruction
2535 What memory barriers are concerned with is controlling the order in which
2536 accesses cross from the CPU side of things to the memory side of things, and
2537 the order in which the effects are perceived to happen by the other observers
2540 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
2541 their own loads and stores as if they had happened in program order.
2543 [!] MMIO or other device accesses may bypass the cache system. This depends on
2544 the properties of the memory window through which devices are accessed and/or
2545 the use of any special device communication instructions the CPU may have.
2551 Life isn't quite as simple as it may appear above, however: for while the
2552 caches are expected to be coherent, there's no guarantee that that coherency
2553 will be ordered. This means that whilst changes made on one CPU will
2554 eventually become visible on all CPUs, there's no guarantee that they will
2555 become apparent in the same order on those other CPUs.
2558 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2559 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2564 +--------+ : +--->| Cache A |<------->| |
2565 | | : | +---------+ | |
2567 | | : | +---------+ | |
2568 +--------+ : +--->| Cache B |<------->| |
2571 : +---------+ | System |
2572 +--------+ : +--->| Cache C |<------->| |
2573 | | : | +---------+ | |
2575 | | : | +---------+ | |
2576 +--------+ : +--->| Cache D |<------->| |
2581 Imagine the system has the following properties:
2583 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2586 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2589 (*) whilst the CPU core is interrogating one cache, the other cache may be
2590 making use of the bus to access the rest of the system - perhaps to
2591 displace a dirty cacheline or to do a speculative load;
2593 (*) each cache has a queue of operations that need to be applied to that cache
2594 to maintain coherency with the rest of the system;
2596 (*) the coherency queue is not flushed by normal loads to lines already
2597 present in the cache, even though the contents of the queue may
2598 potentially affect those loads.
2600 Imagine, then, that two writes are made on the first CPU, with a write barrier
2601 between them to guarantee that they will appear to reach that CPU's caches in
2602 the requisite order:
2605 =============== =============== =======================================
2606 u == 0, v == 1 and p == &u, q == &u
2608 smp_wmb(); Make sure change to v is visible before
2610 <A:modify v=2> v is now in cache A exclusively
2612 <B:modify p=&v> p is now in cache B exclusively
2614 The write memory barrier forces the other CPUs in the system to perceive that
2615 the local CPU's caches have apparently been updated in the correct order. But
2616 now imagine that the second CPU wants to read those values:
2619 =============== =============== =======================================
2624 The above pair of reads may then fail to happen in the expected order, as the
2625 cacheline holding p may get updated in one of the second CPU's caches whilst
2626 the update to the cacheline holding v is delayed in the other of the second
2627 CPU's caches by some other cache event:
2630 =============== =============== =======================================
2631 u == 0, v == 1 and p == &u, q == &u
2634 <A:modify v=2> <C:busy>
2638 <B:modify p=&v> <D:commit p=&v>
2641 <C:read *q> Reads from v before v updated in cache
2645 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2646 no guarantee that, without intervention, the order of update will be the same
2647 as that committed on CPU 1.
2650 To intervene, we need to interpolate a data dependency barrier or a read
2651 barrier between the loads. This will force the cache to commit its coherency
2652 queue before processing any further requests:
2655 =============== =============== =======================================
2656 u == 0, v == 1 and p == &u, q == &u
2659 <A:modify v=2> <C:busy>
2663 <B:modify p=&v> <D:commit p=&v>
2665 smp_read_barrier_depends()
2669 <C:read *q> Reads from v after v updated in cache
2672 This sort of problem can be encountered on DEC Alpha processors as they have a
2673 split cache that improves performance by making better use of the data bus.
2674 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2675 access depends on a read, not all do, so it may not be relied on.
2677 Other CPUs may also have split caches, but must coordinate between the various
2678 cachelets for normal memory accesses. The semantics of the Alpha removes the
2679 need for coordination in the absence of memory barriers.
2682 CACHE COHERENCY VS DMA
2683 ----------------------
2685 Not all systems maintain cache coherency with respect to devices doing DMA. In
2686 such cases, a device attempting DMA may obtain stale data from RAM because
2687 dirty cache lines may be resident in the caches of various CPUs, and may not
2688 have been written back to RAM yet. To deal with this, the appropriate part of
2689 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2690 invalidate them as well).
2692 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2693 cache lines being written back to RAM from a CPU's cache after the device has
2694 installed its own data, or cache lines present in the CPU's cache may simply
2695 obscure the fact that RAM has been updated, until at such time as the cacheline
2696 is discarded from the CPU's cache and reloaded. To deal with this, the
2697 appropriate part of the kernel must invalidate the overlapping bits of the
2700 See Documentation/cachetlb.txt for more information on cache management.
2703 CACHE COHERENCY VS MMIO
2704 -----------------------
2706 Memory mapped I/O usually takes place through memory locations that are part of
2707 a window in the CPU's memory space that has different properties assigned than
2708 the usual RAM directed window.
2710 Amongst these properties is usually the fact that such accesses bypass the
2711 caching entirely and go directly to the device buses. This means MMIO accesses
2712 may, in effect, overtake accesses to cached memory that were emitted earlier.
2713 A memory barrier isn't sufficient in such a case, but rather the cache must be
2714 flushed between the cached memory write and the MMIO access if the two are in
2718 =========================
2719 THE THINGS CPUS GET UP TO
2720 =========================
2722 A programmer might take it for granted that the CPU will perform memory
2723 operations in exactly the order specified, so that if the CPU is, for example,
2724 given the following piece of code to execute:
2726 a = ACCESS_ONCE(*A);
2727 ACCESS_ONCE(*B) = b;
2728 c = ACCESS_ONCE(*C);
2729 d = ACCESS_ONCE(*D);
2730 ACCESS_ONCE(*E) = e;
2732 they would then expect that the CPU will complete the memory operation for each
2733 instruction before moving on to the next one, leading to a definite sequence of
2734 operations as seen by external observers in the system:
2736 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2739 Reality is, of course, much messier. With many CPUs and compilers, the above
2740 assumption doesn't hold because:
2742 (*) loads are more likely to need to be completed immediately to permit
2743 execution progress, whereas stores can often be deferred without a
2746 (*) loads may be done speculatively, and the result discarded should it prove
2747 to have been unnecessary;
2749 (*) loads may be done speculatively, leading to the result having been fetched
2750 at the wrong time in the expected sequence of events;
2752 (*) the order of the memory accesses may be rearranged to promote better use
2753 of the CPU buses and caches;
2755 (*) loads and stores may be combined to improve performance when talking to
2756 memory or I/O hardware that can do batched accesses of adjacent locations,
2757 thus cutting down on transaction setup costs (memory and PCI devices may
2758 both be able to do this); and
2760 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2761 mechanisms may alleviate this - once the store has actually hit the cache
2762 - there's no guarantee that the coherency management will be propagated in
2763 order to other CPUs.
2765 So what another CPU, say, might actually observe from the above piece of code
2768 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2770 (Where "LOAD {*C,*D}" is a combined load)
2773 However, it is guaranteed that a CPU will be self-consistent: it will see its
2774 _own_ accesses appear to be correctly ordered, without the need for a memory
2775 barrier. For instance with the following code:
2777 U = ACCESS_ONCE(*A);
2778 ACCESS_ONCE(*A) = V;
2779 ACCESS_ONCE(*A) = W;
2780 X = ACCESS_ONCE(*A);
2781 ACCESS_ONCE(*A) = Y;
2782 Z = ACCESS_ONCE(*A);
2784 and assuming no intervention by an external influence, it can be assumed that
2785 the final result will appear to be:
2787 U == the original value of *A
2792 The code above may cause the CPU to generate the full sequence of memory
2795 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2797 in that order, but, without intervention, the sequence may have almost any
2798 combination of elements combined or discarded, provided the program's view of
2799 the world remains consistent. Note that ACCESS_ONCE() is -not- optional
2800 in the above example, as there are architectures where a given CPU might
2801 reorder successive loads to the same location. On such architectures,
2802 ACCESS_ONCE() does whatever is necessary to prevent this, for example, on
2803 Itanium the volatile casts used by ACCESS_ONCE() cause GCC to emit the
2804 special ld.acq and st.rel instructions that prevent such reordering.
2806 The compiler may also combine, discard or defer elements of the sequence before
2807 the CPU even sees them.
2818 since, without either a write barrier or an ACCESS_ONCE(), it can be
2819 assumed that the effect of the storage of V to *A is lost. Similarly:
2824 may, without a memory barrier or an ACCESS_ONCE(), be reduced to:
2829 and the LOAD operation never appear outside of the CPU.
2832 AND THEN THERE'S THE ALPHA
2833 --------------------------
2835 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2836 some versions of the Alpha CPU have a split data cache, permitting them to have
2837 two semantically-related cache lines updated at separate times. This is where
2838 the data dependency barrier really becomes necessary as this synchronises both
2839 caches with the memory coherence system, thus making it seem like pointer
2840 changes vs new data occur in the right order.
2842 The Alpha defines the Linux kernel's memory barrier model.
2844 See the subsection on "Cache Coherency" above.
2854 Memory barriers can be used to implement circular buffering without the need
2855 of a lock to serialise the producer with the consumer. See:
2857 Documentation/circular-buffers.txt
2866 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2868 Chapter 5.2: Physical Address Space Characteristics
2869 Chapter 5.4: Caches and Write Buffers
2870 Chapter 5.5: Data Sharing
2871 Chapter 5.6: Read/Write Ordering
2873 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2874 Chapter 7.1: Memory-Access Ordering
2875 Chapter 7.4: Buffering and Combining Memory Writes
2877 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2878 System Programming Guide
2879 Chapter 7.1: Locked Atomic Operations
2880 Chapter 7.2: Memory Ordering
2881 Chapter 7.4: Serializing Instructions
2883 The SPARC Architecture Manual, Version 9
2884 Chapter 8: Memory Models
2885 Appendix D: Formal Specification of the Memory Models
2886 Appendix J: Programming with the Memory Models
2888 UltraSPARC Programmer Reference Manual
2889 Chapter 5: Memory Accesses and Cacheability
2890 Chapter 15: Sparc-V9 Memory Models
2892 UltraSPARC III Cu User's Manual
2893 Chapter 9: Memory Models
2895 UltraSPARC IIIi Processor User's Manual
2896 Chapter 8: Memory Models
2898 UltraSPARC Architecture 2005
2900 Appendix D: Formal Specifications of the Memory Models
2902 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2903 Chapter 8: Memory Models
2904 Appendix F: Caches and Cache Coherency
2906 Solaris Internals, Core Kernel Architecture, p63-68:
2907 Chapter 3.3: Hardware Considerations for Locks and
2910 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2911 for Kernel Programmers:
2912 Chapter 13: Other Memory Models
2914 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2915 Section 2.6: Speculation
2916 Section 4.4: Memory Access
projects / firefly-linux-kernel-4.4.55.git / blob