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:
199 the CPU will issue the following memory operations:
201 Q = LOAD P, D = LOAD *Q
203 and always in that order.
205 (*) Overlapping loads and stores within a particular CPU will appear to be
206 ordered within that CPU. This means that for:
210 the CPU will only issue the following sequence of memory operations:
212 a = LOAD *X, STORE *X = b
218 the CPU will only issue:
220 STORE *X = c, d = LOAD *X
222 (Loads and stores overlap if they are targeted at overlapping pieces of
225 And there are a number of things that _must_ or _must_not_ be assumed:
227 (*) It _must_not_ be assumed that independent loads and stores will be issued
228 in the order given. This means that for:
230 X = *A; Y = *B; *D = Z;
232 we may get any of the following sequences:
234 X = LOAD *A, Y = LOAD *B, STORE *D = Z
235 X = LOAD *A, STORE *D = Z, Y = LOAD *B
236 Y = LOAD *B, X = LOAD *A, STORE *D = Z
237 Y = LOAD *B, STORE *D = Z, X = LOAD *A
238 STORE *D = Z, X = LOAD *A, Y = LOAD *B
239 STORE *D = Z, Y = LOAD *B, X = LOAD *A
241 (*) It _must_ be assumed that overlapping memory accesses may be merged or
242 discarded. This means that for:
244 X = *A; Y = *(A + 4);
246 we may get any one of the following sequences:
248 X = LOAD *A; Y = LOAD *(A + 4);
249 Y = LOAD *(A + 4); X = LOAD *A;
250 {X, Y} = LOAD {*A, *(A + 4) };
254 *A = X; *(A + 4) = Y;
258 STORE *A = X; STORE *(A + 4) = Y;
259 STORE *(A + 4) = Y; STORE *A = X;
260 STORE {*A, *(A + 4) } = {X, Y};
263 =========================
264 WHAT ARE MEMORY BARRIERS?
265 =========================
267 As can be seen above, independent memory operations are effectively performed
268 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
269 What is required is some way of intervening to instruct the compiler and the
270 CPU to restrict the order.
272 Memory barriers are such interventions. They impose a perceived partial
273 ordering over the memory operations on either side of the barrier.
275 Such enforcement is important because the CPUs and other devices in a system
276 can use a variety of tricks to improve performance, including reordering,
277 deferral and combination of memory operations; speculative loads; speculative
278 branch prediction and various types of caching. Memory barriers are used to
279 override or suppress these tricks, allowing the code to sanely control the
280 interaction of multiple CPUs and/or devices.
283 VARIETIES OF MEMORY BARRIER
284 ---------------------------
286 Memory barriers come in four basic varieties:
288 (1) Write (or store) memory barriers.
290 A write memory barrier gives a guarantee that all the STORE operations
291 specified before the barrier will appear to happen before all the STORE
292 operations specified after the barrier with respect to the other
293 components of the system.
295 A write barrier is a partial ordering on stores only; it is not required
296 to have any effect on loads.
298 A CPU can be viewed as committing a sequence of store operations to the
299 memory system as time progresses. All stores before a write barrier will
300 occur in the sequence _before_ all the stores after the write barrier.
302 [!] Note that write barriers should normally be paired with read or data
303 dependency barriers; see the "SMP barrier pairing" subsection.
306 (2) Data dependency barriers.
308 A data dependency barrier is a weaker form of read barrier. In the case
309 where two loads are performed such that the second depends on the result
310 of the first (eg: the first load retrieves the address to which the second
311 load will be directed), a data dependency barrier would be required to
312 make sure that the target of the second load is updated before the address
313 obtained by the first load is accessed.
315 A data dependency barrier is a partial ordering on interdependent loads
316 only; it is not required to have any effect on stores, independent loads
317 or overlapping loads.
319 As mentioned in (1), the other CPUs in the system can be viewed as
320 committing sequences of stores to the memory system that the CPU being
321 considered can then perceive. A data dependency barrier issued by the CPU
322 under consideration guarantees that for any load preceding it, if that
323 load touches one of a sequence of stores from another CPU, then by the
324 time the barrier completes, the effects of all the stores prior to that
325 touched by the load will be perceptible to any loads issued after the data
328 See the "Examples of memory barrier sequences" subsection for diagrams
329 showing the ordering constraints.
331 [!] Note that the first load really has to have a _data_ dependency and
332 not a control dependency. If the address for the second load is dependent
333 on the first load, but the dependency is through a conditional rather than
334 actually loading the address itself, then it's a _control_ dependency and
335 a full read barrier or better is required. See the "Control dependencies"
336 subsection for more information.
338 [!] Note that data dependency barriers should normally be paired with
339 write barriers; see the "SMP barrier pairing" subsection.
342 (3) Read (or load) memory barriers.
344 A read barrier is a data dependency barrier plus a guarantee that all the
345 LOAD operations specified before the barrier will appear to happen before
346 all the LOAD operations specified after the barrier with respect to the
347 other components of the system.
349 A read barrier is a partial ordering on loads only; it is not required to
350 have any effect on stores.
352 Read memory barriers imply data dependency barriers, and so can substitute
355 [!] Note that read barriers should normally be paired with write barriers;
356 see the "SMP barrier pairing" subsection.
359 (4) General memory barriers.
361 A general memory barrier gives a guarantee that all the LOAD and STORE
362 operations specified before the barrier will appear to happen before all
363 the LOAD and STORE operations specified after the barrier with respect to
364 the other components of the system.
366 A general memory barrier is a partial ordering over both loads and stores.
368 General memory barriers imply both read and write memory barriers, and so
369 can substitute for either.
372 And a couple of implicit varieties:
376 This acts as a one-way permeable barrier. It guarantees that all memory
377 operations after the LOCK operation will appear to happen after the LOCK
378 operation with respect to the other components of the system.
380 Memory operations that occur before a LOCK operation may appear to happen
383 A LOCK operation should almost always be paired with an UNLOCK operation.
386 (6) UNLOCK operations.
388 This also acts as a one-way permeable barrier. It guarantees that all
389 memory operations before the UNLOCK operation will appear to happen before
390 the UNLOCK operation with respect to the other components of the system.
392 Memory operations that occur after an UNLOCK operation may appear to
393 happen before it completes.
395 LOCK and UNLOCK operations are guaranteed to appear with respect to each
396 other strictly in the order specified.
398 The use of LOCK and UNLOCK operations generally precludes the need for
399 other sorts of memory barrier (but note the exceptions mentioned in the
400 subsection "MMIO write barrier").
403 Memory barriers are only required where there's a possibility of interaction
404 between two CPUs or between a CPU and a device. If it can be guaranteed that
405 there won't be any such interaction in any particular piece of code, then
406 memory barriers are unnecessary in that piece of code.
409 Note that these are the _minimum_ guarantees. Different architectures may give
410 more substantial guarantees, but they may _not_ be relied upon outside of arch
414 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
415 ----------------------------------------------
417 There are certain things that the Linux kernel memory barriers do not guarantee:
419 (*) There is no guarantee that any of the memory accesses specified before a
420 memory barrier will be _complete_ by the completion of a memory barrier
421 instruction; the barrier can be considered to draw a line in that CPU's
422 access queue that accesses of the appropriate type may not cross.
424 (*) There is no guarantee that issuing a memory barrier on one CPU will have
425 any direct effect on another CPU or any other hardware in the system. The
426 indirect effect will be the order in which the second CPU sees the effects
427 of the first CPU's accesses occur, but see the next point:
429 (*) There is no guarantee that a CPU will see the correct order of effects
430 from a second CPU's accesses, even _if_ the second CPU uses a memory
431 barrier, unless the first CPU _also_ uses a matching memory barrier (see
432 the subsection on "SMP Barrier Pairing").
434 (*) There is no guarantee that some intervening piece of off-the-CPU
435 hardware[*] will not reorder the memory accesses. CPU cache coherency
436 mechanisms should propagate the indirect effects of a memory barrier
437 between CPUs, but might not do so in order.
439 [*] For information on bus mastering DMA and coherency please read:
441 Documentation/PCI/pci.txt
442 Documentation/DMA-API-HOWTO.txt
443 Documentation/DMA-API.txt
446 DATA DEPENDENCY BARRIERS
447 ------------------------
449 The usage requirements of data dependency barriers are a little subtle, and
450 it's not always obvious that they're needed. To illustrate, consider the
451 following sequence of events:
454 =============== ===============
455 { A == 1, B == 2, C = 3, P == &A, Q == &C }
462 There's a clear data dependency here, and it would seem that by the end of the
463 sequence, Q must be either &A or &B, and that:
465 (Q == &A) implies (D == 1)
466 (Q == &B) implies (D == 4)
468 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
469 leading to the following situation:
471 (Q == &B) and (D == 2) ????
473 Whilst this may seem like a failure of coherency or causality maintenance, it
474 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
477 To deal with this, a data dependency barrier or better must be inserted
478 between the address load and the data load:
481 =============== ===============
482 { A == 1, B == 2, C = 3, P == &A, Q == &C }
487 <data dependency barrier>
490 This enforces the occurrence of one of the two implications, and prevents the
491 third possibility from arising.
493 [!] Note that this extremely counterintuitive situation arises most easily on
494 machines with split caches, so that, for example, one cache bank processes
495 even-numbered cache lines and the other bank processes odd-numbered cache
496 lines. The pointer P might be stored in an odd-numbered cache line, and the
497 variable B might be stored in an even-numbered cache line. Then, if the
498 even-numbered bank of the reading CPU's cache is extremely busy while the
499 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
500 but the old value of the variable B (2).
503 Another example of where data dependency barriers might by required is where a
504 number is read from memory and then used to calculate the index for an array
508 =============== ===============
509 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
514 <data dependency barrier>
518 The data dependency barrier is very important to the RCU system, for example.
519 See rcu_dereference() in include/linux/rcupdate.h. This permits the current
520 target of an RCU'd pointer to be replaced with a new modified target, without
521 the replacement target appearing to be incompletely initialised.
523 See also the subsection on "Cache Coherency" for a more thorough example.
529 A control dependency requires a full read memory barrier, not simply a data
530 dependency barrier to make it work correctly. Consider the following bit of
535 <data dependency barrier>
540 This will not have the desired effect because there is no actual data
541 dependency, but rather a control dependency that the CPU may short-circuit by
542 attempting to predict the outcome in advance. In such a case what's actually
556 When dealing with CPU-CPU interactions, certain types of memory barrier should
557 always be paired. A lack of appropriate pairing is almost certainly an error.
559 A write barrier should always be paired with a data dependency barrier or read
560 barrier, though a general barrier would also be viable. Similarly a read
561 barrier or a data dependency barrier should always be paired with at least an
562 write barrier, though, again, a general barrier is viable:
565 =============== ===============
575 =============== ===============================
579 <data dependency barrier>
582 Basically, the read barrier always has to be there, even though it can be of
585 [!] Note that the stores before the write barrier would normally be expected to
586 match the loads after the read barrier or the data dependency barrier, and vice
590 =============== ===============
591 a = 1; }---- --->{ v = c
593 <write barrier> \ <read barrier>
594 c = 3; } / \ { x = a;
595 d = 4; }---- --->{ y = b;
598 EXAMPLES OF MEMORY BARRIER SEQUENCES
599 ------------------------------------
601 Firstly, write barriers act as partial orderings on store operations.
602 Consider the following sequence of events:
605 =======================
613 This sequence of events is committed to the memory coherence system in an order
614 that the rest of the system might perceive as the unordered set of { STORE A,
615 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
620 | |------>| C=3 | } /\
621 | | : +------+ }----- \ -----> Events perceptible to
622 | | : | A=1 | } \/ the rest of the system
624 | CPU 1 | : | B=2 | }
626 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
627 | | +------+ } requires all stores prior to the
628 | | : | E=5 | } barrier to be committed before
629 | | : +------+ } further stores may take place
634 | Sequence in which stores are committed to the
635 | memory system by CPU 1
639 Secondly, data dependency barriers act as partial orderings on data-dependent
640 loads. Consider the following sequence of events:
643 ======================= =======================
644 { B = 7; X = 9; Y = 8; C = &Y }
649 STORE D = 4 LOAD C (gets &B)
652 Without intervention, CPU 2 may perceive the events on CPU 1 in some
653 effectively random order, despite the write barrier issued by CPU 1:
656 | | +------+ +-------+ | Sequence of update
657 | |------>| B=2 |----- --->| Y->8 | | of perception on
658 | | : +------+ \ +-------+ | CPU 2
659 | CPU 1 | : | A=1 | \ --->| C->&Y | V
660 | | +------+ | +-------+
661 | | wwwwwwwwwwwwwwww | : :
663 | | : | C=&B |--- | : : +-------+
664 | | : +------+ \ | +-------+ | |
665 | |------>| D=4 | ----------->| C->&B |------>| |
666 | | +------+ | +-------+ | |
667 +-------+ : : | : : | |
671 Apparently incorrect ---> | | B->7 |------>| |
672 perception of B (!) | +-------+ | |
675 The load of X holds ---> \ | X->9 |------>| |
676 up the maintenance \ +-------+ | |
677 of coherence of B ----->| B->2 | +-------+
682 In the above example, CPU 2 perceives that B is 7, despite the load of *C
683 (which would be B) coming after the LOAD of C.
685 If, however, a data dependency barrier were to be placed between the load of C
686 and the load of *C (ie: B) on CPU 2:
689 ======================= =======================
690 { B = 7; X = 9; Y = 8; C = &Y }
695 STORE D = 4 LOAD C (gets &B)
696 <data dependency barrier>
699 then the following will occur:
702 | | +------+ +-------+
703 | |------>| B=2 |----- --->| Y->8 |
704 | | : +------+ \ +-------+
705 | CPU 1 | : | A=1 | \ --->| C->&Y |
706 | | +------+ | +-------+
707 | | wwwwwwwwwwwwwwww | : :
709 | | : | C=&B |--- | : : +-------+
710 | | : +------+ \ | +-------+ | |
711 | |------>| D=4 | ----------->| C->&B |------>| |
712 | | +------+ | +-------+ | |
713 +-------+ : : | : : | |
719 Makes sure all effects ---> \ ddddddddddddddddd | |
720 prior to the store of C \ +-------+ | |
721 are perceptible to ----->| B->2 |------>| |
722 subsequent loads +-------+ | |
726 And thirdly, a read barrier acts as a partial order on loads. Consider the
727 following sequence of events:
730 ======================= =======================
738 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
739 some effectively random order, despite the write barrier issued by CPU 1:
742 | | +------+ +-------+
743 | |------>| A=1 |------ --->| A->0 |
744 | | +------+ \ +-------+
745 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
746 | | +------+ | +-------+
747 | |------>| B=2 |--- | : :
748 | | +------+ \ | : : +-------+
749 +-------+ : : \ | +-------+ | |
750 ---------->| B->2 |------>| |
751 | +-------+ | CPU 2 |
762 If, however, a read barrier were to be placed between the load of B and the
766 ======================= =======================
775 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
779 | | +------+ +-------+
780 | |------>| A=1 |------ --->| A->0 |
781 | | +------+ \ +-------+
782 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
783 | | +------+ | +-------+
784 | |------>| B=2 |--- | : :
785 | | +------+ \ | : : +-------+
786 +-------+ : : \ | +-------+ | |
787 ---------->| B->2 |------>| |
788 | +-------+ | CPU 2 |
791 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
792 barrier causes all effects \ +-------+ | |
793 prior to the storage of B ---->| A->1 |------>| |
794 to be perceptible to CPU 2 +-------+ | |
798 To illustrate this more completely, consider what could happen if the code
799 contained a load of A either side of the read barrier:
802 ======================= =======================
808 LOAD A [first load of A]
810 LOAD A [second load of A]
812 Even though the two loads of A both occur after the load of B, they may both
813 come up with different values:
816 | | +------+ +-------+
817 | |------>| A=1 |------ --->| A->0 |
818 | | +------+ \ +-------+
819 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
820 | | +------+ | +-------+
821 | |------>| B=2 |--- | : :
822 | | +------+ \ | : : +-------+
823 +-------+ : : \ | +-------+ | |
824 ---------->| B->2 |------>| |
825 | +-------+ | CPU 2 |
829 | | A->0 |------>| 1st |
831 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
832 barrier causes all effects \ +-------+ | |
833 prior to the storage of B ---->| A->1 |------>| 2nd |
834 to be perceptible to CPU 2 +-------+ | |
838 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
839 before the read barrier completes anyway:
842 | | +------+ +-------+
843 | |------>| A=1 |------ --->| A->0 |
844 | | +------+ \ +-------+
845 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
846 | | +------+ | +-------+
847 | |------>| B=2 |--- | : :
848 | | +------+ \ | : : +-------+
849 +-------+ : : \ | +-------+ | |
850 ---------->| B->2 |------>| |
851 | +-------+ | CPU 2 |
855 ---->| A->1 |------>| 1st |
857 rrrrrrrrrrrrrrrrr | |
859 | A->1 |------>| 2nd |
864 The guarantee is that the second load will always come up with A == 1 if the
865 load of B came up with B == 2. No such guarantee exists for the first load of
866 A; that may come up with either A == 0 or A == 1.
869 READ MEMORY BARRIERS VS LOAD SPECULATION
870 ----------------------------------------
872 Many CPUs speculate with loads: that is they see that they will need to load an
873 item from memory, and they find a time where they're not using the bus for any
874 other loads, and so do the load in advance - even though they haven't actually
875 got to that point in the instruction execution flow yet. This permits the
876 actual load instruction to potentially complete immediately because the CPU
877 already has the value to hand.
879 It may turn out that the CPU didn't actually need the value - perhaps because a
880 branch circumvented the load - in which case it can discard the value or just
881 cache it for later use.
886 ======================= =======================
888 DIVIDE } Divide instructions generally
889 DIVIDE } take a long time to perform
892 Which might appear as this:
896 --->| B->2 |------>| |
900 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
901 division speculates on the +-------+ ~ | |
905 Once the divisions are complete --> : : ~-->| |
906 the CPU can then perform the : : | |
907 LOAD with immediate effect : : +-------+
910 Placing a read barrier or a data dependency barrier just before the second
914 ======================= =======================
921 will force any value speculatively obtained to be reconsidered to an extent
922 dependent on the type of barrier used. If there was no change made to the
923 speculated memory location, then the speculated value will just be used:
927 --->| B->2 |------>| |
931 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
932 division speculates on the +-------+ ~ | |
937 rrrrrrrrrrrrrrrr~ | |
944 but if there was an update or an invalidation from another CPU pending, then
945 the speculation will be cancelled and the value reloaded:
949 --->| B->2 |------>| |
953 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
954 division speculates on the +-------+ ~ | |
959 rrrrrrrrrrrrrrrrr | |
961 The speculation is discarded ---> --->| A->1 |------>| |
962 and an updated value is +-------+ | |
963 retrieved : : +-------+
969 Transitivity is a deeply intuitive notion about ordering that is not
970 always provided by real computer systems. The following example
971 demonstrates transitivity (also called "cumulativity"):
974 ======================= ======================= =======================
976 STORE X=1 LOAD X STORE Y=1
977 <general barrier> <general barrier>
980 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
981 This indicates that CPU 2's load from X in some sense follows CPU 1's
982 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
983 store to Y. The question is then "Can CPU 3's load from X return 0?"
985 Because CPU 2's load from X in some sense came after CPU 1's store, it
986 is natural to expect that CPU 3's load from X must therefore return 1.
987 This expectation is an example of transitivity: if a load executing on
988 CPU A follows a load from the same variable executing on CPU B, then
989 CPU A's load must either return the same value that CPU B's load did,
990 or must return some later value.
992 In the Linux kernel, use of general memory barriers guarantees
993 transitivity. Therefore, in the above example, if CPU 2's load from X
994 returns 1 and its load from Y returns 0, then CPU 3's load from X must
997 However, transitivity is -not- guaranteed for read or write barriers.
998 For example, suppose that CPU 2's general barrier in the above example
999 is changed to a read barrier as shown below:
1002 ======================= ======================= =======================
1004 STORE X=1 LOAD X STORE Y=1
1005 <read barrier> <general barrier>
1008 This substitution destroys transitivity: in this example, it is perfectly
1009 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1010 and CPU 3's load from X to return 0.
1012 The key point is that although CPU 2's read barrier orders its pair
1013 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1014 this example runs on a system where CPUs 1 and 2 share a store buffer
1015 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1016 General barriers are therefore required to ensure that all CPUs agree
1017 on the combined order of CPU 1's and CPU 2's accesses.
1019 To reiterate, if your code requires transitivity, use general barriers
1023 ========================
1024 EXPLICIT KERNEL BARRIERS
1025 ========================
1027 The Linux kernel has a variety of different barriers that act at different
1030 (*) Compiler barrier.
1032 (*) CPU memory barriers.
1034 (*) MMIO write barrier.
1040 The Linux kernel has an explicit compiler barrier function that prevents the
1041 compiler from moving the memory accesses either side of it to the other side:
1045 This is a general barrier - lesser varieties of compiler barrier do not exist.
1047 The compiler barrier has no direct effect on the CPU, which may then reorder
1048 things however it wishes.
1054 The Linux kernel has eight basic CPU memory barriers:
1056 TYPE MANDATORY SMP CONDITIONAL
1057 =============== ======================= ===========================
1058 GENERAL mb() smp_mb()
1059 WRITE wmb() smp_wmb()
1060 READ rmb() smp_rmb()
1061 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1064 All memory barriers except the data dependency barriers imply a compiler
1065 barrier. Data dependencies do not impose any additional compiler ordering.
1067 Aside: In the case of data dependencies, the compiler would be expected to
1068 issue the loads in the correct order (eg. `a[b]` would have to load the value
1069 of b before loading a[b]), however there is no guarantee in the C specification
1070 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1071 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1072 problem of a compiler reloading b after having loaded a[b], thus having a newer
1073 copy of b than a[b]. A consensus has not yet been reached about these problems,
1074 however the ACCESS_ONCE macro is a good place to start looking.
1076 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1077 systems because it is assumed that a CPU will appear to be self-consistent,
1078 and will order overlapping accesses correctly with respect to itself.
1080 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1081 references to shared memory on SMP systems, though the use of locking instead
1084 Mandatory barriers should not be used to control SMP effects, since mandatory
1085 barriers unnecessarily impose overhead on UP systems. They may, however, be
1086 used to control MMIO effects on accesses through relaxed memory I/O windows.
1087 These are required even on non-SMP systems as they affect the order in which
1088 memory operations appear to a device by prohibiting both the compiler and the
1089 CPU from reordering them.
1092 There are some more advanced barrier functions:
1094 (*) set_mb(var, value)
1096 This assigns the value to the variable and then inserts a full memory
1097 barrier after it, depending on the function. It isn't guaranteed to
1098 insert anything more than a compiler barrier in a UP compilation.
1101 (*) smp_mb__before_atomic_dec();
1102 (*) smp_mb__after_atomic_dec();
1103 (*) smp_mb__before_atomic_inc();
1104 (*) smp_mb__after_atomic_inc();
1106 These are for use with atomic add, subtract, increment and decrement
1107 functions that don't return a value, especially when used for reference
1108 counting. These functions do not imply memory barriers.
1110 As an example, consider a piece of code that marks an object as being dead
1111 and then decrements the object's reference count:
1114 smp_mb__before_atomic_dec();
1115 atomic_dec(&obj->ref_count);
1117 This makes sure that the death mark on the object is perceived to be set
1118 *before* the reference counter is decremented.
1120 See Documentation/atomic_ops.txt for more information. See the "Atomic
1121 operations" subsection for information on where to use these.
1124 (*) smp_mb__before_clear_bit(void);
1125 (*) smp_mb__after_clear_bit(void);
1127 These are for use similar to the atomic inc/dec barriers. These are
1128 typically used for bitwise unlocking operations, so care must be taken as
1129 there are no implicit memory barriers here either.
1131 Consider implementing an unlock operation of some nature by clearing a
1132 locking bit. The clear_bit() would then need to be barriered like this:
1134 smp_mb__before_clear_bit();
1137 This prevents memory operations before the clear leaking to after it. See
1138 the subsection on "Locking Functions" with reference to UNLOCK operation
1141 See Documentation/atomic_ops.txt for more information. See the "Atomic
1142 operations" subsection for information on where to use these.
1148 The Linux kernel also has a special barrier for use with memory-mapped I/O
1153 This is a variation on the mandatory write barrier that causes writes to weakly
1154 ordered I/O regions to be partially ordered. Its effects may go beyond the
1155 CPU->Hardware interface and actually affect the hardware at some level.
1157 See the subsection "Locks vs I/O accesses" for more information.
1160 ===============================
1161 IMPLICIT KERNEL MEMORY BARRIERS
1162 ===============================
1164 Some of the other functions in the linux kernel imply memory barriers, amongst
1165 which are locking and scheduling functions.
1167 This specification is a _minimum_ guarantee; any particular architecture may
1168 provide more substantial guarantees, but these may not be relied upon outside
1169 of arch specific code.
1175 The Linux kernel has a number of locking constructs:
1184 In all cases there are variants on "LOCK" operations and "UNLOCK" operations
1185 for each construct. These operations all imply certain barriers:
1187 (1) LOCK operation implication:
1189 Memory operations issued after the LOCK will be completed after the LOCK
1190 operation has completed.
1192 Memory operations issued before the LOCK may be completed after the LOCK
1193 operation has completed.
1195 (2) UNLOCK operation implication:
1197 Memory operations issued before the UNLOCK will be completed before the
1198 UNLOCK operation has completed.
1200 Memory operations issued after the UNLOCK may be completed before the
1201 UNLOCK operation has completed.
1203 (3) LOCK vs LOCK implication:
1205 All LOCK operations issued before another LOCK operation will be completed
1206 before that LOCK operation.
1208 (4) LOCK vs UNLOCK implication:
1210 All LOCK operations issued before an UNLOCK operation will be completed
1211 before the UNLOCK operation.
1213 All UNLOCK operations issued before a LOCK operation will be completed
1214 before the LOCK operation.
1216 (5) Failed conditional LOCK implication:
1218 Certain variants of the LOCK operation may fail, either due to being
1219 unable to get the lock immediately, or due to receiving an unblocked
1220 signal whilst asleep waiting for the lock to become available. Failed
1221 locks do not imply any sort of barrier.
1223 Therefore, from (1), (2) and (4) an UNLOCK followed by an unconditional LOCK is
1224 equivalent to a full barrier, but a LOCK followed by an UNLOCK is not.
1226 [!] Note: one of the consequences of LOCKs and UNLOCKs being only one-way
1227 barriers is that the effects of instructions outside of a critical section
1228 may seep into the inside of the critical section.
1230 A LOCK followed by an UNLOCK may not be assumed to be full memory barrier
1231 because it is possible for an access preceding the LOCK to happen after the
1232 LOCK, and an access following the UNLOCK to happen before the UNLOCK, and the
1233 two accesses can themselves then cross:
1242 LOCK, STORE *B, STORE *A, UNLOCK
1244 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1245 systems, and so cannot be counted on in such a situation to actually achieve
1246 anything at all - especially with respect to I/O accesses - unless combined
1247 with interrupt disabling operations.
1249 See also the section on "Inter-CPU locking barrier effects".
1252 As an example, consider the following:
1263 The following sequence of events is acceptable:
1265 LOCK, {*F,*A}, *E, {*C,*D}, *B, UNLOCK
1267 [+] Note that {*F,*A} indicates a combined access.
1269 But none of the following are:
1271 {*F,*A}, *B, LOCK, *C, *D, UNLOCK, *E
1272 *A, *B, *C, LOCK, *D, UNLOCK, *E, *F
1273 *A, *B, LOCK, *C, UNLOCK, *D, *E, *F
1274 *B, LOCK, *C, *D, UNLOCK, {*F,*A}, *E
1278 INTERRUPT DISABLING FUNCTIONS
1279 -----------------------------
1281 Functions that disable interrupts (LOCK equivalent) and enable interrupts
1282 (UNLOCK equivalent) will act as compiler barriers only. So if memory or I/O
1283 barriers are required in such a situation, they must be provided from some
1287 SLEEP AND WAKE-UP FUNCTIONS
1288 ---------------------------
1290 Sleeping and waking on an event flagged in global data can be viewed as an
1291 interaction between two pieces of data: the task state of the task waiting for
1292 the event and the global data used to indicate the event. To make sure that
1293 these appear to happen in the right order, the primitives to begin the process
1294 of going to sleep, and the primitives to initiate a wake up imply certain
1297 Firstly, the sleeper normally follows something like this sequence of events:
1300 set_current_state(TASK_UNINTERRUPTIBLE);
1301 if (event_indicated)
1306 A general memory barrier is interpolated automatically by set_current_state()
1307 after it has altered the task state:
1310 ===============================
1311 set_current_state();
1313 STORE current->state
1315 LOAD event_indicated
1317 set_current_state() may be wrapped by:
1320 prepare_to_wait_exclusive();
1322 which therefore also imply a general memory barrier after setting the state.
1323 The whole sequence above is available in various canned forms, all of which
1324 interpolate the memory barrier in the right place:
1327 wait_event_interruptible();
1328 wait_event_interruptible_exclusive();
1329 wait_event_interruptible_timeout();
1330 wait_event_killable();
1331 wait_event_timeout();
1336 Secondly, code that performs a wake up normally follows something like this:
1338 event_indicated = 1;
1339 wake_up(&event_wait_queue);
1343 event_indicated = 1;
1344 wake_up_process(event_daemon);
1346 A write memory barrier is implied by wake_up() and co. if and only if they wake
1347 something up. The barrier occurs before the task state is cleared, and so sits
1348 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1351 =============================== ===============================
1352 set_current_state(); STORE event_indicated
1353 set_mb(); wake_up();
1354 STORE current->state <write barrier>
1355 <general barrier> STORE current->state
1356 LOAD event_indicated
1358 The available waker functions include:
1364 wake_up_interruptible();
1365 wake_up_interruptible_all();
1366 wake_up_interruptible_nr();
1367 wake_up_interruptible_poll();
1368 wake_up_interruptible_sync();
1369 wake_up_interruptible_sync_poll();
1371 wake_up_locked_poll();
1377 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1378 order multiple stores before the wake-up with respect to loads of those stored
1379 values after the sleeper has called set_current_state(). For instance, if the
1382 set_current_state(TASK_INTERRUPTIBLE);
1383 if (event_indicated)
1385 __set_current_state(TASK_RUNNING);
1386 do_something(my_data);
1391 event_indicated = 1;
1392 wake_up(&event_wait_queue);
1394 there's no guarantee that the change to event_indicated will be perceived by
1395 the sleeper as coming after the change to my_data. In such a circumstance, the
1396 code on both sides must interpolate its own memory barriers between the
1397 separate data accesses. Thus the above sleeper ought to do:
1399 set_current_state(TASK_INTERRUPTIBLE);
1400 if (event_indicated) {
1402 do_something(my_data);
1405 and the waker should do:
1409 event_indicated = 1;
1410 wake_up(&event_wait_queue);
1413 MISCELLANEOUS FUNCTIONS
1414 -----------------------
1416 Other functions that imply barriers:
1418 (*) schedule() and similar imply full memory barriers.
1421 =================================
1422 INTER-CPU LOCKING BARRIER EFFECTS
1423 =================================
1425 On SMP systems locking primitives give a more substantial form of barrier: one
1426 that does affect memory access ordering on other CPUs, within the context of
1427 conflict on any particular lock.
1430 LOCKS VS MEMORY ACCESSES
1431 ------------------------
1433 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1434 three CPUs; then should the following sequence of events occur:
1437 =============================== ===============================
1445 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
1446 through *H occur in, other than the constraints imposed by the separate locks
1447 on the separate CPUs. It might, for example, see:
1449 *E, LOCK M, LOCK Q, *G, *C, *F, *A, *B, UNLOCK Q, *D, *H, UNLOCK M
1451 But it won't see any of:
1453 *B, *C or *D preceding LOCK M
1454 *A, *B or *C following UNLOCK M
1455 *F, *G or *H preceding LOCK Q
1456 *E, *F or *G following UNLOCK Q
1459 However, if the following occurs:
1462 =============================== ===============================
1477 *E, LOCK M [1], *C, *B, *A, UNLOCK M [1],
1478 LOCK M [2], *H, *F, *G, UNLOCK M [2], *D
1480 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
1482 *B, *C, *D, *F, *G or *H preceding LOCK M [1]
1483 *A, *B or *C following UNLOCK M [1]
1484 *F, *G or *H preceding LOCK M [2]
1485 *A, *B, *C, *E, *F or *G following UNLOCK M [2]
1488 LOCKS VS I/O ACCESSES
1489 ---------------------
1491 Under certain circumstances (especially involving NUMA), I/O accesses within
1492 two spinlocked sections on two different CPUs may be seen as interleaved by the
1493 PCI bridge, because the PCI bridge does not necessarily participate in the
1494 cache-coherence protocol, and is therefore incapable of issuing the required
1495 read memory barriers.
1500 =============================== ===============================
1510 may be seen by the PCI bridge as follows:
1512 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
1514 which would probably cause the hardware to malfunction.
1517 What is necessary here is to intervene with an mmiowb() before dropping the
1518 spinlock, for example:
1521 =============================== ===============================
1533 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
1534 before either of the stores issued on CPU 2.
1537 Furthermore, following a store by a load from the same device obviates the need
1538 for the mmiowb(), because the load forces the store to complete before the load
1542 =============================== ===============================
1553 See Documentation/DocBook/deviceiobook.tmpl for more information.
1556 =================================
1557 WHERE ARE MEMORY BARRIERS NEEDED?
1558 =================================
1560 Under normal operation, memory operation reordering is generally not going to
1561 be a problem as a single-threaded linear piece of code will still appear to
1562 work correctly, even if it's in an SMP kernel. There are, however, four
1563 circumstances in which reordering definitely _could_ be a problem:
1565 (*) Interprocessor interaction.
1567 (*) Atomic operations.
1569 (*) Accessing devices.
1574 INTERPROCESSOR INTERACTION
1575 --------------------------
1577 When there's a system with more than one processor, more than one CPU in the
1578 system may be working on the same data set at the same time. This can cause
1579 synchronisation problems, and the usual way of dealing with them is to use
1580 locks. Locks, however, are quite expensive, and so it may be preferable to
1581 operate without the use of a lock if at all possible. In such a case
1582 operations that affect both CPUs may have to be carefully ordered to prevent
1585 Consider, for example, the R/W semaphore slow path. Here a waiting process is
1586 queued on the semaphore, by virtue of it having a piece of its stack linked to
1587 the semaphore's list of waiting processes:
1589 struct rw_semaphore {
1592 struct list_head waiters;
1595 struct rwsem_waiter {
1596 struct list_head list;
1597 struct task_struct *task;
1600 To wake up a particular waiter, the up_read() or up_write() functions have to:
1602 (1) read the next pointer from this waiter's record to know as to where the
1603 next waiter record is;
1605 (2) read the pointer to the waiter's task structure;
1607 (3) clear the task pointer to tell the waiter it has been given the semaphore;
1609 (4) call wake_up_process() on the task; and
1611 (5) release the reference held on the waiter's task struct.
1613 In other words, it has to perform this sequence of events:
1615 LOAD waiter->list.next;
1621 and if any of these steps occur out of order, then the whole thing may
1624 Once it has queued itself and dropped the semaphore lock, the waiter does not
1625 get the lock again; it instead just waits for its task pointer to be cleared
1626 before proceeding. Since the record is on the waiter's stack, this means that
1627 if the task pointer is cleared _before_ the next pointer in the list is read,
1628 another CPU might start processing the waiter and might clobber the waiter's
1629 stack before the up*() function has a chance to read the next pointer.
1631 Consider then what might happen to the above sequence of events:
1634 =============================== ===============================
1641 Woken up by other event
1646 foo() clobbers *waiter
1648 LOAD waiter->list.next;
1651 This could be dealt with using the semaphore lock, but then the down_xxx()
1652 function has to needlessly get the spinlock again after being woken up.
1654 The way to deal with this is to insert a general SMP memory barrier:
1656 LOAD waiter->list.next;
1663 In this case, the barrier makes a guarantee that all memory accesses before the
1664 barrier will appear to happen before all the memory accesses after the barrier
1665 with respect to the other CPUs on the system. It does _not_ guarantee that all
1666 the memory accesses before the barrier will be complete by the time the barrier
1667 instruction itself is complete.
1669 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
1670 compiler barrier, thus making sure the compiler emits the instructions in the
1671 right order without actually intervening in the CPU. Since there's only one
1672 CPU, that CPU's dependency ordering logic will take care of everything else.
1678 Whilst they are technically interprocessor interaction considerations, atomic
1679 operations are noted specially as some of them imply full memory barriers and
1680 some don't, but they're very heavily relied on as a group throughout the
1683 Any atomic operation that modifies some state in memory and returns information
1684 about the state (old or new) implies an SMP-conditional general memory barrier
1685 (smp_mb()) on each side of the actual operation (with the exception of
1686 explicit lock operations, described later). These include:
1692 atomic_inc_return();
1693 atomic_dec_return();
1694 atomic_add_return();
1695 atomic_sub_return();
1696 atomic_inc_and_test();
1697 atomic_dec_and_test();
1698 atomic_sub_and_test();
1699 atomic_add_negative();
1700 atomic_add_unless(); /* when succeeds (returns 1) */
1702 test_and_clear_bit();
1703 test_and_change_bit();
1705 These are used for such things as implementing LOCK-class and UNLOCK-class
1706 operations and adjusting reference counters towards object destruction, and as
1707 such the implicit memory barrier effects are necessary.
1710 The following operations are potential problems as they do _not_ imply memory
1711 barriers, but might be used for implementing such things as UNLOCK-class
1719 With these the appropriate explicit memory barrier should be used if necessary
1720 (smp_mb__before_clear_bit() for instance).
1723 The following also do _not_ imply memory barriers, and so may require explicit
1724 memory barriers under some circumstances (smp_mb__before_atomic_dec() for
1732 If they're used for statistics generation, then they probably don't need memory
1733 barriers, unless there's a coupling between statistical data.
1735 If they're used for reference counting on an object to control its lifetime,
1736 they probably don't need memory barriers because either the reference count
1737 will be adjusted inside a locked section, or the caller will already hold
1738 sufficient references to make the lock, and thus a memory barrier unnecessary.
1740 If they're used for constructing a lock of some description, then they probably
1741 do need memory barriers as a lock primitive generally has to do things in a
1744 Basically, each usage case has to be carefully considered as to whether memory
1745 barriers are needed or not.
1747 The following operations are special locking primitives:
1749 test_and_set_bit_lock();
1751 __clear_bit_unlock();
1753 These implement LOCK-class and UNLOCK-class operations. These should be used in
1754 preference to other operations when implementing locking primitives, because
1755 their implementations can be optimised on many architectures.
1757 [!] Note that special memory barrier primitives are available for these
1758 situations because on some CPUs the atomic instructions used imply full memory
1759 barriers, and so barrier instructions are superfluous in conjunction with them,
1760 and in such cases the special barrier primitives will be no-ops.
1762 See Documentation/atomic_ops.txt for more information.
1768 Many devices can be memory mapped, and so appear to the CPU as if they're just
1769 a set of memory locations. To control such a device, the driver usually has to
1770 make the right memory accesses in exactly the right order.
1772 However, having a clever CPU or a clever compiler creates a potential problem
1773 in that the carefully sequenced accesses in the driver code won't reach the
1774 device in the requisite order if the CPU or the compiler thinks it is more
1775 efficient to reorder, combine or merge accesses - something that would cause
1776 the device to malfunction.
1778 Inside of the Linux kernel, I/O should be done through the appropriate accessor
1779 routines - such as inb() or writel() - which know how to make such accesses
1780 appropriately sequential. Whilst this, for the most part, renders the explicit
1781 use of memory barriers unnecessary, there are a couple of situations where they
1784 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
1785 so for _all_ general drivers locks should be used and mmiowb() must be
1786 issued prior to unlocking the critical section.
1788 (2) If the accessor functions are used to refer to an I/O memory window with
1789 relaxed memory access properties, then _mandatory_ memory barriers are
1790 required to enforce ordering.
1792 See Documentation/DocBook/deviceiobook.tmpl for more information.
1798 A driver may be interrupted by its own interrupt service routine, and thus the
1799 two parts of the driver may interfere with each other's attempts to control or
1802 This may be alleviated - at least in part - by disabling local interrupts (a
1803 form of locking), such that the critical operations are all contained within
1804 the interrupt-disabled section in the driver. Whilst the driver's interrupt
1805 routine is executing, the driver's core may not run on the same CPU, and its
1806 interrupt is not permitted to happen again until the current interrupt has been
1807 handled, thus the interrupt handler does not need to lock against that.
1809 However, consider a driver that was talking to an ethernet card that sports an
1810 address register and a data register. If that driver's core talks to the card
1811 under interrupt-disablement and then the driver's interrupt handler is invoked:
1822 The store to the data register might happen after the second store to the
1823 address register if ordering rules are sufficiently relaxed:
1825 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
1828 If ordering rules are relaxed, it must be assumed that accesses done inside an
1829 interrupt disabled section may leak outside of it and may interleave with
1830 accesses performed in an interrupt - and vice versa - unless implicit or
1831 explicit barriers are used.
1833 Normally this won't be a problem because the I/O accesses done inside such
1834 sections will include synchronous load operations on strictly ordered I/O
1835 registers that form implicit I/O barriers. If this isn't sufficient then an
1836 mmiowb() may need to be used explicitly.
1839 A similar situation may occur between an interrupt routine and two routines
1840 running on separate CPUs that communicate with each other. If such a case is
1841 likely, then interrupt-disabling locks should be used to guarantee ordering.
1844 ==========================
1845 KERNEL I/O BARRIER EFFECTS
1846 ==========================
1848 When accessing I/O memory, drivers should use the appropriate accessor
1853 These are intended to talk to I/O space rather than memory space, but
1854 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
1855 indeed have special I/O space access cycles and instructions, but many
1856 CPUs don't have such a concept.
1858 The PCI bus, amongst others, defines an I/O space concept which - on such
1859 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
1860 space. However, it may also be mapped as a virtual I/O space in the CPU's
1861 memory map, particularly on those CPUs that don't support alternate I/O
1864 Accesses to this space may be fully synchronous (as on i386), but
1865 intermediary bridges (such as the PCI host bridge) may not fully honour
1868 They are guaranteed to be fully ordered with respect to each other.
1870 They are not guaranteed to be fully ordered with respect to other types of
1871 memory and I/O operation.
1873 (*) readX(), writeX():
1875 Whether these are guaranteed to be fully ordered and uncombined with
1876 respect to each other on the issuing CPU depends on the characteristics
1877 defined for the memory window through which they're accessing. On later
1878 i386 architecture machines, for example, this is controlled by way of the
1881 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
1882 provided they're not accessing a prefetchable device.
1884 However, intermediary hardware (such as a PCI bridge) may indulge in
1885 deferral if it so wishes; to flush a store, a load from the same location
1886 is preferred[*], but a load from the same device or from configuration
1887 space should suffice for PCI.
1889 [*] NOTE! attempting to load from the same location as was written to may
1890 cause a malfunction - consider the 16550 Rx/Tx serial registers for
1893 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
1894 force stores to be ordered.
1896 Please refer to the PCI specification for more information on interactions
1897 between PCI transactions.
1901 These are similar to readX(), but are not guaranteed to be ordered in any
1902 way. Be aware that there is no I/O read barrier available.
1904 (*) ioreadX(), iowriteX()
1906 These will perform appropriately for the type of access they're actually
1907 doing, be it inX()/outX() or readX()/writeX().
1910 ========================================
1911 ASSUMED MINIMUM EXECUTION ORDERING MODEL
1912 ========================================
1914 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
1915 maintain the appearance of program causality with respect to itself. Some CPUs
1916 (such as i386 or x86_64) are more constrained than others (such as powerpc or
1917 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
1918 of arch-specific code.
1920 This means that it must be considered that the CPU will execute its instruction
1921 stream in any order it feels like - or even in parallel - provided that if an
1922 instruction in the stream depends on an earlier instruction, then that
1923 earlier instruction must be sufficiently complete[*] before the later
1924 instruction may proceed; in other words: provided that the appearance of
1925 causality is maintained.
1927 [*] Some instructions have more than one effect - such as changing the
1928 condition codes, changing registers or changing memory - and different
1929 instructions may depend on different effects.
1931 A CPU may also discard any instruction sequence that winds up having no
1932 ultimate effect. For example, if two adjacent instructions both load an
1933 immediate value into the same register, the first may be discarded.
1936 Similarly, it has to be assumed that compiler might reorder the instruction
1937 stream in any way it sees fit, again provided the appearance of causality is
1941 ============================
1942 THE EFFECTS OF THE CPU CACHE
1943 ============================
1945 The way cached memory operations are perceived across the system is affected to
1946 a certain extent by the caches that lie between CPUs and memory, and by the
1947 memory coherence system that maintains the consistency of state in the system.
1949 As far as the way a CPU interacts with another part of the system through the
1950 caches goes, the memory system has to include the CPU's caches, and memory
1951 barriers for the most part act at the interface between the CPU and its cache
1952 (memory barriers logically act on the dotted line in the following diagram):
1954 <--- CPU ---> : <----------- Memory ----------->
1956 +--------+ +--------+ : +--------+ +-----------+
1957 | | | | : | | | | +--------+
1958 | CPU | | Memory | : | CPU | | | | |
1959 | Core |--->| Access |----->| Cache |<-->| | | |
1960 | | | Queue | : | | | |--->| Memory |
1961 | | | | : | | | | | |
1962 +--------+ +--------+ : +--------+ | | | |
1963 : | Cache | +--------+
1965 : | Mechanism | +--------+
1966 +--------+ +--------+ : +--------+ | | | |
1967 | | | | : | | | | | |
1968 | CPU | | Memory | : | CPU | | |--->| Device |
1969 | Core |--->| Access |----->| Cache |<-->| | | |
1970 | | | Queue | : | | | | | |
1971 | | | | : | | | | +--------+
1972 +--------+ +--------+ : +--------+ +-----------+
1976 Although any particular load or store may not actually appear outside of the
1977 CPU that issued it since it may have been satisfied within the CPU's own cache,
1978 it will still appear as if the full memory access had taken place as far as the
1979 other CPUs are concerned since the cache coherency mechanisms will migrate the
1980 cacheline over to the accessing CPU and propagate the effects upon conflict.
1982 The CPU core may execute instructions in any order it deems fit, provided the
1983 expected program causality appears to be maintained. Some of the instructions
1984 generate load and store operations which then go into the queue of memory
1985 accesses to be performed. The core may place these in the queue in any order
1986 it wishes, and continue execution until it is forced to wait for an instruction
1989 What memory barriers are concerned with is controlling the order in which
1990 accesses cross from the CPU side of things to the memory side of things, and
1991 the order in which the effects are perceived to happen by the other observers
1994 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
1995 their own loads and stores as if they had happened in program order.
1997 [!] MMIO or other device accesses may bypass the cache system. This depends on
1998 the properties of the memory window through which devices are accessed and/or
1999 the use of any special device communication instructions the CPU may have.
2005 Life isn't quite as simple as it may appear above, however: for while the
2006 caches are expected to be coherent, there's no guarantee that that coherency
2007 will be ordered. This means that whilst changes made on one CPU will
2008 eventually become visible on all CPUs, there's no guarantee that they will
2009 become apparent in the same order on those other CPUs.
2012 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2013 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2018 +--------+ : +--->| Cache A |<------->| |
2019 | | : | +---------+ | |
2021 | | : | +---------+ | |
2022 +--------+ : +--->| Cache B |<------->| |
2025 : +---------+ | System |
2026 +--------+ : +--->| Cache C |<------->| |
2027 | | : | +---------+ | |
2029 | | : | +---------+ | |
2030 +--------+ : +--->| Cache D |<------->| |
2035 Imagine the system has the following properties:
2037 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2040 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2043 (*) whilst the CPU core is interrogating one cache, the other cache may be
2044 making use of the bus to access the rest of the system - perhaps to
2045 displace a dirty cacheline or to do a speculative load;
2047 (*) each cache has a queue of operations that need to be applied to that cache
2048 to maintain coherency with the rest of the system;
2050 (*) the coherency queue is not flushed by normal loads to lines already
2051 present in the cache, even though the contents of the queue may
2052 potentially affect those loads.
2054 Imagine, then, that two writes are made on the first CPU, with a write barrier
2055 between them to guarantee that they will appear to reach that CPU's caches in
2056 the requisite order:
2059 =============== =============== =======================================
2060 u == 0, v == 1 and p == &u, q == &u
2062 smp_wmb(); Make sure change to v is visible before
2064 <A:modify v=2> v is now in cache A exclusively
2066 <B:modify p=&v> p is now in cache B exclusively
2068 The write memory barrier forces the other CPUs in the system to perceive that
2069 the local CPU's caches have apparently been updated in the correct order. But
2070 now imagine that the second CPU wants to read those values:
2073 =============== =============== =======================================
2078 The above pair of reads may then fail to happen in the expected order, as the
2079 cacheline holding p may get updated in one of the second CPU's caches whilst
2080 the update to the cacheline holding v is delayed in the other of the second
2081 CPU's caches by some other cache event:
2084 =============== =============== =======================================
2085 u == 0, v == 1 and p == &u, q == &u
2088 <A:modify v=2> <C:busy>
2092 <B:modify p=&v> <D:commit p=&v>
2095 <C:read *q> Reads from v before v updated in cache
2099 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2100 no guarantee that, without intervention, the order of update will be the same
2101 as that committed on CPU 1.
2104 To intervene, we need to interpolate a data dependency barrier or a read
2105 barrier between the loads. This will force the cache to commit its coherency
2106 queue before processing any further requests:
2109 =============== =============== =======================================
2110 u == 0, v == 1 and p == &u, q == &u
2113 <A:modify v=2> <C:busy>
2117 <B:modify p=&v> <D:commit p=&v>
2119 smp_read_barrier_depends()
2123 <C:read *q> Reads from v after v updated in cache
2126 This sort of problem can be encountered on DEC Alpha processors as they have a
2127 split cache that improves performance by making better use of the data bus.
2128 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2129 access depends on a read, not all do, so it may not be relied on.
2131 Other CPUs may also have split caches, but must coordinate between the various
2132 cachelets for normal memory accesses. The semantics of the Alpha removes the
2133 need for coordination in the absence of memory barriers.
2136 CACHE COHERENCY VS DMA
2137 ----------------------
2139 Not all systems maintain cache coherency with respect to devices doing DMA. In
2140 such cases, a device attempting DMA may obtain stale data from RAM because
2141 dirty cache lines may be resident in the caches of various CPUs, and may not
2142 have been written back to RAM yet. To deal with this, the appropriate part of
2143 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2144 invalidate them as well).
2146 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2147 cache lines being written back to RAM from a CPU's cache after the device has
2148 installed its own data, or cache lines present in the CPU's cache may simply
2149 obscure the fact that RAM has been updated, until at such time as the cacheline
2150 is discarded from the CPU's cache and reloaded. To deal with this, the
2151 appropriate part of the kernel must invalidate the overlapping bits of the
2154 See Documentation/cachetlb.txt for more information on cache management.
2157 CACHE COHERENCY VS MMIO
2158 -----------------------
2160 Memory mapped I/O usually takes place through memory locations that are part of
2161 a window in the CPU's memory space that has different properties assigned than
2162 the usual RAM directed window.
2164 Amongst these properties is usually the fact that such accesses bypass the
2165 caching entirely and go directly to the device buses. This means MMIO accesses
2166 may, in effect, overtake accesses to cached memory that were emitted earlier.
2167 A memory barrier isn't sufficient in such a case, but rather the cache must be
2168 flushed between the cached memory write and the MMIO access if the two are in
2172 =========================
2173 THE THINGS CPUS GET UP TO
2174 =========================
2176 A programmer might take it for granted that the CPU will perform memory
2177 operations in exactly the order specified, so that if the CPU is, for example,
2178 given the following piece of code to execute:
2186 they would then expect that the CPU will complete the memory operation for each
2187 instruction before moving on to the next one, leading to a definite sequence of
2188 operations as seen by external observers in the system:
2190 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2193 Reality is, of course, much messier. With many CPUs and compilers, the above
2194 assumption doesn't hold because:
2196 (*) loads are more likely to need to be completed immediately to permit
2197 execution progress, whereas stores can often be deferred without a
2200 (*) loads may be done speculatively, and the result discarded should it prove
2201 to have been unnecessary;
2203 (*) loads may be done speculatively, leading to the result having been fetched
2204 at the wrong time in the expected sequence of events;
2206 (*) the order of the memory accesses may be rearranged to promote better use
2207 of the CPU buses and caches;
2209 (*) loads and stores may be combined to improve performance when talking to
2210 memory or I/O hardware that can do batched accesses of adjacent locations,
2211 thus cutting down on transaction setup costs (memory and PCI devices may
2212 both be able to do this); and
2214 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2215 mechanisms may alleviate this - once the store has actually hit the cache
2216 - there's no guarantee that the coherency management will be propagated in
2217 order to other CPUs.
2219 So what another CPU, say, might actually observe from the above piece of code
2222 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2224 (Where "LOAD {*C,*D}" is a combined load)
2227 However, it is guaranteed that a CPU will be self-consistent: it will see its
2228 _own_ accesses appear to be correctly ordered, without the need for a memory
2229 barrier. For instance with the following code:
2238 and assuming no intervention by an external influence, it can be assumed that
2239 the final result will appear to be:
2241 U == the original value of *A
2246 The code above may cause the CPU to generate the full sequence of memory
2249 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2251 in that order, but, without intervention, the sequence may have almost any
2252 combination of elements combined or discarded, provided the program's view of
2253 the world remains consistent.
2255 The compiler may also combine, discard or defer elements of the sequence before
2256 the CPU even sees them.
2267 since, without a write barrier, it can be assumed that the effect of the
2268 storage of V to *A is lost. Similarly:
2273 may, without a memory barrier, be reduced to:
2278 and the LOAD operation never appear outside of the CPU.
2281 AND THEN THERE'S THE ALPHA
2282 --------------------------
2284 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2285 some versions of the Alpha CPU have a split data cache, permitting them to have
2286 two semantically-related cache lines updated at separate times. This is where
2287 the data dependency barrier really becomes necessary as this synchronises both
2288 caches with the memory coherence system, thus making it seem like pointer
2289 changes vs new data occur in the right order.
2291 The Alpha defines the Linux kernel's memory barrier model.
2293 See the subsection on "Cache Coherency" above.
2303 Memory barriers can be used to implement circular buffering without the need
2304 of a lock to serialise the producer with the consumer. See:
2306 Documentation/circular-buffers.txt
2315 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2317 Chapter 5.2: Physical Address Space Characteristics
2318 Chapter 5.4: Caches and Write Buffers
2319 Chapter 5.5: Data Sharing
2320 Chapter 5.6: Read/Write Ordering
2322 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2323 Chapter 7.1: Memory-Access Ordering
2324 Chapter 7.4: Buffering and Combining Memory Writes
2326 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2327 System Programming Guide
2328 Chapter 7.1: Locked Atomic Operations
2329 Chapter 7.2: Memory Ordering
2330 Chapter 7.4: Serializing Instructions
2332 The SPARC Architecture Manual, Version 9
2333 Chapter 8: Memory Models
2334 Appendix D: Formal Specification of the Memory Models
2335 Appendix J: Programming with the Memory Models
2337 UltraSPARC Programmer Reference Manual
2338 Chapter 5: Memory Accesses and Cacheability
2339 Chapter 15: Sparc-V9 Memory Models
2341 UltraSPARC III Cu User's Manual
2342 Chapter 9: Memory Models
2344 UltraSPARC IIIi Processor User's Manual
2345 Chapter 8: Memory Models
2347 UltraSPARC Architecture 2005
2349 Appendix D: Formal Specifications of the Memory Models
2351 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2352 Chapter 8: Memory Models
2353 Appendix F: Caches and Cache Coherency
2355 Solaris Internals, Core Kernel Architecture, p63-68:
2356 Chapter 3.3: Hardware Considerations for Locks and
2359 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2360 for Kernel Programmers:
2361 Chapter 13: Other Memory Models
2363 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2364 Section 2.6: Speculation
2365 Section 4.4: Memory Access
projects / deliverable / linux.git / blob