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) };
256 we may get either of:
258 STORE *A = X; Y = LOAD *A;
262 =========================
263 WHAT ARE MEMORY BARRIERS?
264 =========================
266 As can be seen above, independent memory operations are effectively performed
267 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
268 What is required is some way of intervening to instruct the compiler and the
269 CPU to restrict the order.
271 Memory barriers are such interventions. They impose a perceived partial
272 ordering over the memory operations on either side of the barrier.
274 Such enforcement is important because the CPUs and other devices in a system
275 can use a variety of tricks to improve performance, including reordering,
276 deferral and combination of memory operations; speculative loads; speculative
277 branch prediction and various types of caching. Memory barriers are used to
278 override or suppress these tricks, allowing the code to sanely control the
279 interaction of multiple CPUs and/or devices.
282 VARIETIES OF MEMORY BARRIER
283 ---------------------------
285 Memory barriers come in four basic varieties:
287 (1) Write (or store) memory barriers.
289 A write memory barrier gives a guarantee that all the STORE operations
290 specified before the barrier will appear to happen before all the STORE
291 operations specified after the barrier with respect to the other
292 components of the system.
294 A write barrier is a partial ordering on stores only; it is not required
295 to have any effect on loads.
297 A CPU can be viewed as committing a sequence of store operations to the
298 memory system as time progresses. All stores before a write barrier will
299 occur in the sequence _before_ all the stores after the write barrier.
301 [!] Note that write barriers should normally be paired with read or data
302 dependency barriers; see the "SMP barrier pairing" subsection.
305 (2) Data dependency barriers.
307 A data dependency barrier is a weaker form of read barrier. In the case
308 where two loads are performed such that the second depends on the result
309 of the first (eg: the first load retrieves the address to which the second
310 load will be directed), a data dependency barrier would be required to
311 make sure that the target of the second load is updated before the address
312 obtained by the first load is accessed.
314 A data dependency barrier is a partial ordering on interdependent loads
315 only; it is not required to have any effect on stores, independent loads
316 or overlapping loads.
318 As mentioned in (1), the other CPUs in the system can be viewed as
319 committing sequences of stores to the memory system that the CPU being
320 considered can then perceive. A data dependency barrier issued by the CPU
321 under consideration guarantees that for any load preceding it, if that
322 load touches one of a sequence of stores from another CPU, then by the
323 time the barrier completes, the effects of all the stores prior to that
324 touched by the load will be perceptible to any loads issued after the data
327 See the "Examples of memory barrier sequences" subsection for diagrams
328 showing the ordering constraints.
330 [!] Note that the first load really has to have a _data_ dependency and
331 not a control dependency. If the address for the second load is dependent
332 on the first load, but the dependency is through a conditional rather than
333 actually loading the address itself, then it's a _control_ dependency and
334 a full read barrier or better is required. See the "Control dependencies"
335 subsection for more information.
337 [!] Note that data dependency barriers should normally be paired with
338 write barriers; see the "SMP barrier pairing" subsection.
341 (3) Read (or load) memory barriers.
343 A read barrier is a data dependency barrier plus a guarantee that all the
344 LOAD operations specified before the barrier will appear to happen before
345 all the LOAD operations specified after the barrier with respect to the
346 other components of the system.
348 A read barrier is a partial ordering on loads only; it is not required to
349 have any effect on stores.
351 Read memory barriers imply data dependency barriers, and so can substitute
354 [!] Note that read barriers should normally be paired with write barriers;
355 see the "SMP barrier pairing" subsection.
358 (4) General memory barriers.
360 A general memory barrier gives a guarantee that all the LOAD and STORE
361 operations specified before the barrier will appear to happen before all
362 the LOAD and STORE operations specified after the barrier with respect to
363 the other components of the system.
365 A general memory barrier is a partial ordering over both loads and stores.
367 General memory barriers imply both read and write memory barriers, and so
368 can substitute for either.
371 And a couple of implicit varieties:
375 This acts as a one-way permeable barrier. It guarantees that all memory
376 operations after the LOCK operation will appear to happen after the LOCK
377 operation with respect to the other components of the system.
379 Memory operations that occur before a LOCK operation may appear to happen
382 A LOCK operation should almost always be paired with an UNLOCK operation.
385 (6) UNLOCK operations.
387 This also acts as a one-way permeable barrier. It guarantees that all
388 memory operations before the UNLOCK operation will appear to happen before
389 the UNLOCK operation with respect to the other components of the system.
391 Memory operations that occur after an UNLOCK operation may appear to
392 happen before it completes.
394 LOCK and UNLOCK operations are guaranteed to appear with respect to each
395 other strictly in the order specified.
397 The use of LOCK and UNLOCK operations generally precludes the need for
398 other sorts of memory barrier (but note the exceptions mentioned in the
399 subsection "MMIO write barrier").
402 Memory barriers are only required where there's a possibility of interaction
403 between two CPUs or between a CPU and a device. If it can be guaranteed that
404 there won't be any such interaction in any particular piece of code, then
405 memory barriers are unnecessary in that piece of code.
408 Note that these are the _minimum_ guarantees. Different architectures may give
409 more substantial guarantees, but they may _not_ be relied upon outside of arch
413 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
414 ----------------------------------------------
416 There are certain things that the Linux kernel memory barriers do not guarantee:
418 (*) There is no guarantee that any of the memory accesses specified before a
419 memory barrier will be _complete_ by the completion of a memory barrier
420 instruction; the barrier can be considered to draw a line in that CPU's
421 access queue that accesses of the appropriate type may not cross.
423 (*) There is no guarantee that issuing a memory barrier on one CPU will have
424 any direct effect on another CPU or any other hardware in the system. The
425 indirect effect will be the order in which the second CPU sees the effects
426 of the first CPU's accesses occur, but see the next point:
428 (*) There is no guarantee that a CPU will see the correct order of effects
429 from a second CPU's accesses, even _if_ the second CPU uses a memory
430 barrier, unless the first CPU _also_ uses a matching memory barrier (see
431 the subsection on "SMP Barrier Pairing").
433 (*) There is no guarantee that some intervening piece of off-the-CPU
434 hardware[*] will not reorder the memory accesses. CPU cache coherency
435 mechanisms should propagate the indirect effects of a memory barrier
436 between CPUs, but might not do so in order.
438 [*] For information on bus mastering DMA and coherency please read:
440 Documentation/PCI/pci.txt
441 Documentation/DMA-API-HOWTO.txt
442 Documentation/DMA-API.txt
445 DATA DEPENDENCY BARRIERS
446 ------------------------
448 The usage requirements of data dependency barriers are a little subtle, and
449 it's not always obvious that they're needed. To illustrate, consider the
450 following sequence of events:
453 =============== ===============
454 { A == 1, B == 2, C = 3, P == &A, Q == &C }
461 There's a clear data dependency here, and it would seem that by the end of the
462 sequence, Q must be either &A or &B, and that:
464 (Q == &A) implies (D == 1)
465 (Q == &B) implies (D == 4)
467 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
468 leading to the following situation:
470 (Q == &B) and (D == 2) ????
472 Whilst this may seem like a failure of coherency or causality maintenance, it
473 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
476 To deal with this, a data dependency barrier or better must be inserted
477 between the address load and the data load:
480 =============== ===============
481 { A == 1, B == 2, C = 3, P == &A, Q == &C }
486 <data dependency barrier>
489 This enforces the occurrence of one of the two implications, and prevents the
490 third possibility from arising.
492 [!] Note that this extremely counterintuitive situation arises most easily on
493 machines with split caches, so that, for example, one cache bank processes
494 even-numbered cache lines and the other bank processes odd-numbered cache
495 lines. The pointer P might be stored in an odd-numbered cache line, and the
496 variable B might be stored in an even-numbered cache line. Then, if the
497 even-numbered bank of the reading CPU's cache is extremely busy while the
498 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
499 but the old value of the variable B (2).
502 Another example of where data dependency barriers might by required is where a
503 number is read from memory and then used to calculate the index for an array
507 =============== ===============
508 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
513 <data dependency barrier>
517 The data dependency barrier is very important to the RCU system, for example.
518 See rcu_dereference() in include/linux/rcupdate.h. This permits the current
519 target of an RCU'd pointer to be replaced with a new modified target, without
520 the replacement target appearing to be incompletely initialised.
522 See also the subsection on "Cache Coherency" for a more thorough example.
528 A control dependency requires a full read memory barrier, not simply a data
529 dependency barrier to make it work correctly. Consider the following bit of
535 <data dependency barrier>
538 This will not have the desired effect because there is no actual data
539 dependency, but rather a control dependency that the CPU may short-circuit by
540 attempting to predict the outcome in advance. In such a case what's actually
553 When dealing with CPU-CPU interactions, certain types of memory barrier should
554 always be paired. A lack of appropriate pairing is almost certainly an error.
556 A write barrier should always be paired with a data dependency barrier or read
557 barrier, though a general barrier would also be viable. Similarly a read
558 barrier or a data dependency barrier should always be paired with at least an
559 write barrier, though, again, a general barrier is viable:
562 =============== ===============
572 =============== ===============================
576 <data dependency barrier>
579 Basically, the read barrier always has to be there, even though it can be of
582 [!] Note that the stores before the write barrier would normally be expected to
583 match the loads after the read barrier or the data dependency barrier, and vice
587 =============== ===============
588 a = 1; }---- --->{ v = c
590 <write barrier> \ <read barrier>
591 c = 3; } / \ { x = a;
592 d = 4; }---- --->{ y = b;
595 EXAMPLES OF MEMORY BARRIER SEQUENCES
596 ------------------------------------
598 Firstly, write barriers act as partial orderings on store operations.
599 Consider the following sequence of events:
602 =======================
610 This sequence of events is committed to the memory coherence system in an order
611 that the rest of the system might perceive as the unordered set of { STORE A,
612 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
617 | |------>| C=3 | } /\
618 | | : +------+ }----- \ -----> Events perceptible to
619 | | : | A=1 | } \/ the rest of the system
621 | CPU 1 | : | B=2 | }
623 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
624 | | +------+ } requires all stores prior to the
625 | | : | E=5 | } barrier to be committed before
626 | | : +------+ } further stores may take place
631 | Sequence in which stores are committed to the
632 | memory system by CPU 1
636 Secondly, data dependency barriers act as partial orderings on data-dependent
637 loads. Consider the following sequence of events:
640 ======================= =======================
641 { B = 7; X = 9; Y = 8; C = &Y }
646 STORE D = 4 LOAD C (gets &B)
649 Without intervention, CPU 2 may perceive the events on CPU 1 in some
650 effectively random order, despite the write barrier issued by CPU 1:
653 | | +------+ +-------+ | Sequence of update
654 | |------>| B=2 |----- --->| Y->8 | | of perception on
655 | | : +------+ \ +-------+ | CPU 2
656 | CPU 1 | : | A=1 | \ --->| C->&Y | V
657 | | +------+ | +-------+
658 | | wwwwwwwwwwwwwwww | : :
660 | | : | C=&B |--- | : : +-------+
661 | | : +------+ \ | +-------+ | |
662 | |------>| D=4 | ----------->| C->&B |------>| |
663 | | +------+ | +-------+ | |
664 +-------+ : : | : : | |
668 Apparently incorrect ---> | | B->7 |------>| |
669 perception of B (!) | +-------+ | |
672 The load of X holds ---> \ | X->9 |------>| |
673 up the maintenance \ +-------+ | |
674 of coherence of B ----->| B->2 | +-------+
679 In the above example, CPU 2 perceives that B is 7, despite the load of *C
680 (which would be B) coming after the LOAD of C.
682 If, however, a data dependency barrier were to be placed between the load of C
683 and the load of *C (ie: B) on CPU 2:
686 ======================= =======================
687 { B = 7; X = 9; Y = 8; C = &Y }
692 STORE D = 4 LOAD C (gets &B)
693 <data dependency barrier>
696 then the following will occur:
699 | | +------+ +-------+
700 | |------>| B=2 |----- --->| Y->8 |
701 | | : +------+ \ +-------+
702 | CPU 1 | : | A=1 | \ --->| C->&Y |
703 | | +------+ | +-------+
704 | | wwwwwwwwwwwwwwww | : :
706 | | : | C=&B |--- | : : +-------+
707 | | : +------+ \ | +-------+ | |
708 | |------>| D=4 | ----------->| C->&B |------>| |
709 | | +------+ | +-------+ | |
710 +-------+ : : | : : | |
716 Makes sure all effects ---> \ ddddddddddddddddd | |
717 prior to the store of C \ +-------+ | |
718 are perceptible to ----->| B->2 |------>| |
719 subsequent loads +-------+ | |
723 And thirdly, a read barrier acts as a partial order on loads. Consider the
724 following sequence of events:
727 ======================= =======================
735 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
736 some effectively random order, despite the write barrier issued by CPU 1:
739 | | +------+ +-------+
740 | |------>| A=1 |------ --->| A->0 |
741 | | +------+ \ +-------+
742 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
743 | | +------+ | +-------+
744 | |------>| B=2 |--- | : :
745 | | +------+ \ | : : +-------+
746 +-------+ : : \ | +-------+ | |
747 ---------->| B->2 |------>| |
748 | +-------+ | CPU 2 |
759 If, however, a read barrier were to be placed between the load of B and the
763 ======================= =======================
772 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
776 | | +------+ +-------+
777 | |------>| A=1 |------ --->| A->0 |
778 | | +------+ \ +-------+
779 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
780 | | +------+ | +-------+
781 | |------>| B=2 |--- | : :
782 | | +------+ \ | : : +-------+
783 +-------+ : : \ | +-------+ | |
784 ---------->| B->2 |------>| |
785 | +-------+ | CPU 2 |
788 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
789 barrier causes all effects \ +-------+ | |
790 prior to the storage of B ---->| A->1 |------>| |
791 to be perceptible to CPU 2 +-------+ | |
795 To illustrate this more completely, consider what could happen if the code
796 contained a load of A either side of the read barrier:
799 ======================= =======================
805 LOAD A [first load of A]
807 LOAD A [second load of A]
809 Even though the two loads of A both occur after the load of B, they may both
810 come up with different values:
813 | | +------+ +-------+
814 | |------>| A=1 |------ --->| A->0 |
815 | | +------+ \ +-------+
816 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
817 | | +------+ | +-------+
818 | |------>| B=2 |--- | : :
819 | | +------+ \ | : : +-------+
820 +-------+ : : \ | +-------+ | |
821 ---------->| B->2 |------>| |
822 | +-------+ | CPU 2 |
826 | | A->0 |------>| 1st |
828 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
829 barrier causes all effects \ +-------+ | |
830 prior to the storage of B ---->| A->1 |------>| 2nd |
831 to be perceptible to CPU 2 +-------+ | |
835 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
836 before the read barrier completes anyway:
839 | | +------+ +-------+
840 | |------>| A=1 |------ --->| A->0 |
841 | | +------+ \ +-------+
842 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
843 | | +------+ | +-------+
844 | |------>| B=2 |--- | : :
845 | | +------+ \ | : : +-------+
846 +-------+ : : \ | +-------+ | |
847 ---------->| B->2 |------>| |
848 | +-------+ | CPU 2 |
852 ---->| A->1 |------>| 1st |
854 rrrrrrrrrrrrrrrrr | |
856 | A->1 |------>| 2nd |
861 The guarantee is that the second load will always come up with A == 1 if the
862 load of B came up with B == 2. No such guarantee exists for the first load of
863 A; that may come up with either A == 0 or A == 1.
866 READ MEMORY BARRIERS VS LOAD SPECULATION
867 ----------------------------------------
869 Many CPUs speculate with loads: that is they see that they will need to load an
870 item from memory, and they find a time where they're not using the bus for any
871 other loads, and so do the load in advance - even though they haven't actually
872 got to that point in the instruction execution flow yet. This permits the
873 actual load instruction to potentially complete immediately because the CPU
874 already has the value to hand.
876 It may turn out that the CPU didn't actually need the value - perhaps because a
877 branch circumvented the load - in which case it can discard the value or just
878 cache it for later use.
883 ======================= =======================
885 DIVIDE } Divide instructions generally
886 DIVIDE } take a long time to perform
889 Which might appear as this:
893 --->| B->2 |------>| |
897 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
898 division speculates on the +-------+ ~ | |
902 Once the divisions are complete --> : : ~-->| |
903 the CPU can then perform the : : | |
904 LOAD with immediate effect : : +-------+
907 Placing a read barrier or a data dependency barrier just before the second
911 ======================= =======================
918 will force any value speculatively obtained to be reconsidered to an extent
919 dependent on the type of barrier used. If there was no change made to the
920 speculated memory location, then the speculated value will just be used:
924 --->| B->2 |------>| |
928 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
929 division speculates on the +-------+ ~ | |
934 rrrrrrrrrrrrrrrr~ | |
941 but if there was an update or an invalidation from another CPU pending, then
942 the speculation will be cancelled and the value reloaded:
946 --->| B->2 |------>| |
950 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
951 division speculates on the +-------+ ~ | |
956 rrrrrrrrrrrrrrrrr | |
958 The speculation is discarded ---> --->| A->1 |------>| |
959 and an updated value is +-------+ | |
960 retrieved : : +-------+
966 Transitivity is a deeply intuitive notion about ordering that is not
967 always provided by real computer systems. The following example
968 demonstrates transitivity (also called "cumulativity"):
971 ======================= ======================= =======================
973 STORE X=1 LOAD X STORE Y=1
974 <general barrier> <general barrier>
977 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
978 This indicates that CPU 2's load from X in some sense follows CPU 1's
979 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
980 store to Y. The question is then "Can CPU 3's load from X return 0?"
982 Because CPU 2's load from X in some sense came after CPU 1's store, it
983 is natural to expect that CPU 3's load from X must therefore return 1.
984 This expectation is an example of transitivity: if a load executing on
985 CPU A follows a load from the same variable executing on CPU B, then
986 CPU A's load must either return the same value that CPU B's load did,
987 or must return some later value.
989 In the Linux kernel, use of general memory barriers guarantees
990 transitivity. Therefore, in the above example, if CPU 2's load from X
991 returns 1 and its load from Y returns 0, then CPU 3's load from X must
994 However, transitivity is -not- guaranteed for read or write barriers.
995 For example, suppose that CPU 2's general barrier in the above example
996 is changed to a read barrier as shown below:
999 ======================= ======================= =======================
1001 STORE X=1 LOAD X STORE Y=1
1002 <read barrier> <general barrier>
1005 This substitution destroys transitivity: in this example, it is perfectly
1006 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1007 and CPU 3's load from X to return 0.
1009 The key point is that although CPU 2's read barrier orders its pair
1010 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1011 this example runs on a system where CPUs 1 and 2 share a store buffer
1012 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1013 General barriers are therefore required to ensure that all CPUs agree
1014 on the combined order of CPU 1's and CPU 2's accesses.
1016 To reiterate, if your code requires transitivity, use general barriers
1020 ========================
1021 EXPLICIT KERNEL BARRIERS
1022 ========================
1024 The Linux kernel has a variety of different barriers that act at different
1027 (*) Compiler barrier.
1029 (*) CPU memory barriers.
1031 (*) MMIO write barrier.
1037 The Linux kernel has an explicit compiler barrier function that prevents the
1038 compiler from moving the memory accesses either side of it to the other side:
1042 This is a general barrier - lesser varieties of compiler barrier do not exist.
1044 The compiler barrier has no direct effect on the CPU, which may then reorder
1045 things however it wishes.
1051 The Linux kernel has eight basic CPU memory barriers:
1053 TYPE MANDATORY SMP CONDITIONAL
1054 =============== ======================= ===========================
1055 GENERAL mb() smp_mb()
1056 WRITE wmb() smp_wmb()
1057 READ rmb() smp_rmb()
1058 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1061 All memory barriers except the data dependency barriers imply a compiler
1062 barrier. Data dependencies do not impose any additional compiler ordering.
1064 Aside: In the case of data dependencies, the compiler would be expected to
1065 issue the loads in the correct order (eg. `a[b]` would have to load the value
1066 of b before loading a[b]), however there is no guarantee in the C specification
1067 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1068 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1069 problem of a compiler reloading b after having loaded a[b], thus having a newer
1070 copy of b than a[b]. A consensus has not yet been reached about these problems,
1071 however the ACCESS_ONCE macro is a good place to start looking.
1073 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1074 systems because it is assumed that a CPU will appear to be self-consistent,
1075 and will order overlapping accesses correctly with respect to itself.
1077 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1078 references to shared memory on SMP systems, though the use of locking instead
1081 Mandatory barriers should not be used to control SMP effects, since mandatory
1082 barriers unnecessarily impose overhead on UP systems. They may, however, be
1083 used to control MMIO effects on accesses through relaxed memory I/O windows.
1084 These are required even on non-SMP systems as they affect the order in which
1085 memory operations appear to a device by prohibiting both the compiler and the
1086 CPU from reordering them.
1089 There are some more advanced barrier functions:
1091 (*) set_mb(var, value)
1093 This assigns the value to the variable and then inserts a full memory
1094 barrier after it, depending on the function. It isn't guaranteed to
1095 insert anything more than a compiler barrier in a UP compilation.
1098 (*) smp_mb__before_atomic_dec();
1099 (*) smp_mb__after_atomic_dec();
1100 (*) smp_mb__before_atomic_inc();
1101 (*) smp_mb__after_atomic_inc();
1103 These are for use with atomic add, subtract, increment and decrement
1104 functions that don't return a value, especially when used for reference
1105 counting. These functions do not imply memory barriers.
1107 As an example, consider a piece of code that marks an object as being dead
1108 and then decrements the object's reference count:
1111 smp_mb__before_atomic_dec();
1112 atomic_dec(&obj->ref_count);
1114 This makes sure that the death mark on the object is perceived to be set
1115 *before* the reference counter is decremented.
1117 See Documentation/atomic_ops.txt for more information. See the "Atomic
1118 operations" subsection for information on where to use these.
1121 (*) smp_mb__before_clear_bit(void);
1122 (*) smp_mb__after_clear_bit(void);
1124 These are for use similar to the atomic inc/dec barriers. These are
1125 typically used for bitwise unlocking operations, so care must be taken as
1126 there are no implicit memory barriers here either.
1128 Consider implementing an unlock operation of some nature by clearing a
1129 locking bit. The clear_bit() would then need to be barriered like this:
1131 smp_mb__before_clear_bit();
1134 This prevents memory operations before the clear leaking to after it. See
1135 the subsection on "Locking Functions" with reference to UNLOCK operation
1138 See Documentation/atomic_ops.txt for more information. See the "Atomic
1139 operations" subsection for information on where to use these.
1145 The Linux kernel also has a special barrier for use with memory-mapped I/O
1150 This is a variation on the mandatory write barrier that causes writes to weakly
1151 ordered I/O regions to be partially ordered. Its effects may go beyond the
1152 CPU->Hardware interface and actually affect the hardware at some level.
1154 See the subsection "Locks vs I/O accesses" for more information.
1157 ===============================
1158 IMPLICIT KERNEL MEMORY BARRIERS
1159 ===============================
1161 Some of the other functions in the linux kernel imply memory barriers, amongst
1162 which are locking and scheduling functions.
1164 This specification is a _minimum_ guarantee; any particular architecture may
1165 provide more substantial guarantees, but these may not be relied upon outside
1166 of arch specific code.
1172 The Linux kernel has a number of locking constructs:
1181 In all cases there are variants on "LOCK" operations and "UNLOCK" operations
1182 for each construct. These operations all imply certain barriers:
1184 (1) LOCK operation implication:
1186 Memory operations issued after the LOCK will be completed after the LOCK
1187 operation has completed.
1189 Memory operations issued before the LOCK may be completed after the LOCK
1190 operation has completed.
1192 (2) UNLOCK operation implication:
1194 Memory operations issued before the UNLOCK will be completed before the
1195 UNLOCK operation has completed.
1197 Memory operations issued after the UNLOCK may be completed before the
1198 UNLOCK operation has completed.
1200 (3) LOCK vs LOCK implication:
1202 All LOCK operations issued before another LOCK operation will be completed
1203 before that LOCK operation.
1205 (4) LOCK vs UNLOCK implication:
1207 All LOCK operations issued before an UNLOCK operation will be completed
1208 before the UNLOCK operation.
1210 All UNLOCK operations issued before a LOCK operation will be completed
1211 before the LOCK operation.
1213 (5) Failed conditional LOCK implication:
1215 Certain variants of the LOCK operation may fail, either due to being
1216 unable to get the lock immediately, or due to receiving an unblocked
1217 signal whilst asleep waiting for the lock to become available. Failed
1218 locks do not imply any sort of barrier.
1220 Therefore, from (1), (2) and (4) an UNLOCK followed by an unconditional LOCK is
1221 equivalent to a full barrier, but a LOCK followed by an UNLOCK is not.
1223 [!] Note: one of the consequences of LOCKs and UNLOCKs being only one-way
1224 barriers is that the effects of instructions outside of a critical section
1225 may seep into the inside of the critical section.
1227 A LOCK followed by an UNLOCK may not be assumed to be full memory barrier
1228 because it is possible for an access preceding the LOCK to happen after the
1229 LOCK, and an access following the UNLOCK to happen before the UNLOCK, and the
1230 two accesses can themselves then cross:
1239 LOCK, STORE *B, STORE *A, UNLOCK
1241 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1242 systems, and so cannot be counted on in such a situation to actually achieve
1243 anything at all - especially with respect to I/O accesses - unless combined
1244 with interrupt disabling operations.
1246 See also the section on "Inter-CPU locking barrier effects".
1249 As an example, consider the following:
1260 The following sequence of events is acceptable:
1262 LOCK, {*F,*A}, *E, {*C,*D}, *B, UNLOCK
1264 [+] Note that {*F,*A} indicates a combined access.
1266 But none of the following are:
1268 {*F,*A}, *B, LOCK, *C, *D, UNLOCK, *E
1269 *A, *B, *C, LOCK, *D, UNLOCK, *E, *F
1270 *A, *B, LOCK, *C, UNLOCK, *D, *E, *F
1271 *B, LOCK, *C, *D, UNLOCK, {*F,*A}, *E
1275 INTERRUPT DISABLING FUNCTIONS
1276 -----------------------------
1278 Functions that disable interrupts (LOCK equivalent) and enable interrupts
1279 (UNLOCK equivalent) will act as compiler barriers only. So if memory or I/O
1280 barriers are required in such a situation, they must be provided from some
1284 SLEEP AND WAKE-UP FUNCTIONS
1285 ---------------------------
1287 Sleeping and waking on an event flagged in global data can be viewed as an
1288 interaction between two pieces of data: the task state of the task waiting for
1289 the event and the global data used to indicate the event. To make sure that
1290 these appear to happen in the right order, the primitives to begin the process
1291 of going to sleep, and the primitives to initiate a wake up imply certain
1294 Firstly, the sleeper normally follows something like this sequence of events:
1297 set_current_state(TASK_UNINTERRUPTIBLE);
1298 if (event_indicated)
1303 A general memory barrier is interpolated automatically by set_current_state()
1304 after it has altered the task state:
1307 ===============================
1308 set_current_state();
1310 STORE current->state
1312 LOAD event_indicated
1314 set_current_state() may be wrapped by:
1317 prepare_to_wait_exclusive();
1319 which therefore also imply a general memory barrier after setting the state.
1320 The whole sequence above is available in various canned forms, all of which
1321 interpolate the memory barrier in the right place:
1324 wait_event_interruptible();
1325 wait_event_interruptible_exclusive();
1326 wait_event_interruptible_timeout();
1327 wait_event_killable();
1328 wait_event_timeout();
1333 Secondly, code that performs a wake up normally follows something like this:
1335 event_indicated = 1;
1336 wake_up(&event_wait_queue);
1340 event_indicated = 1;
1341 wake_up_process(event_daemon);
1343 A write memory barrier is implied by wake_up() and co. if and only if they wake
1344 something up. The barrier occurs before the task state is cleared, and so sits
1345 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1348 =============================== ===============================
1349 set_current_state(); STORE event_indicated
1350 set_mb(); wake_up();
1351 STORE current->state <write barrier>
1352 <general barrier> STORE current->state
1353 LOAD event_indicated
1355 The available waker functions include:
1361 wake_up_interruptible();
1362 wake_up_interruptible_all();
1363 wake_up_interruptible_nr();
1364 wake_up_interruptible_poll();
1365 wake_up_interruptible_sync();
1366 wake_up_interruptible_sync_poll();
1368 wake_up_locked_poll();
1374 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1375 order multiple stores before the wake-up with respect to loads of those stored
1376 values after the sleeper has called set_current_state(). For instance, if the
1379 set_current_state(TASK_INTERRUPTIBLE);
1380 if (event_indicated)
1382 __set_current_state(TASK_RUNNING);
1383 do_something(my_data);
1388 event_indicated = 1;
1389 wake_up(&event_wait_queue);
1391 there's no guarantee that the change to event_indicated will be perceived by
1392 the sleeper as coming after the change to my_data. In such a circumstance, the
1393 code on both sides must interpolate its own memory barriers between the
1394 separate data accesses. Thus the above sleeper ought to do:
1396 set_current_state(TASK_INTERRUPTIBLE);
1397 if (event_indicated) {
1399 do_something(my_data);
1402 and the waker should do:
1406 event_indicated = 1;
1407 wake_up(&event_wait_queue);
1410 MISCELLANEOUS FUNCTIONS
1411 -----------------------
1413 Other functions that imply barriers:
1415 (*) schedule() and similar imply full memory barriers.
1418 =================================
1419 INTER-CPU LOCKING BARRIER EFFECTS
1420 =================================
1422 On SMP systems locking primitives give a more substantial form of barrier: one
1423 that does affect memory access ordering on other CPUs, within the context of
1424 conflict on any particular lock.
1427 LOCKS VS MEMORY ACCESSES
1428 ------------------------
1430 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1431 three CPUs; then should the following sequence of events occur:
1434 =============================== ===============================
1442 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
1443 through *H occur in, other than the constraints imposed by the separate locks
1444 on the separate CPUs. It might, for example, see:
1446 *E, LOCK M, LOCK Q, *G, *C, *F, *A, *B, UNLOCK Q, *D, *H, UNLOCK M
1448 But it won't see any of:
1450 *B, *C or *D preceding LOCK M
1451 *A, *B or *C following UNLOCK M
1452 *F, *G or *H preceding LOCK Q
1453 *E, *F or *G following UNLOCK Q
1456 However, if the following occurs:
1459 =============================== ===============================
1474 *E, LOCK M [1], *C, *B, *A, UNLOCK M [1],
1475 LOCK M [2], *H, *F, *G, UNLOCK M [2], *D
1477 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
1479 *B, *C, *D, *F, *G or *H preceding LOCK M [1]
1480 *A, *B or *C following UNLOCK M [1]
1481 *F, *G or *H preceding LOCK M [2]
1482 *A, *B, *C, *E, *F or *G following UNLOCK M [2]
1485 LOCKS VS I/O ACCESSES
1486 ---------------------
1488 Under certain circumstances (especially involving NUMA), I/O accesses within
1489 two spinlocked sections on two different CPUs may be seen as interleaved by the
1490 PCI bridge, because the PCI bridge does not necessarily participate in the
1491 cache-coherence protocol, and is therefore incapable of issuing the required
1492 read memory barriers.
1497 =============================== ===============================
1507 may be seen by the PCI bridge as follows:
1509 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
1511 which would probably cause the hardware to malfunction.
1514 What is necessary here is to intervene with an mmiowb() before dropping the
1515 spinlock, for example:
1518 =============================== ===============================
1530 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
1531 before either of the stores issued on CPU 2.
1534 Furthermore, following a store by a load from the same device obviates the need
1535 for the mmiowb(), because the load forces the store to complete before the load
1539 =============================== ===============================
1550 See Documentation/DocBook/deviceiobook.tmpl for more information.
1553 =================================
1554 WHERE ARE MEMORY BARRIERS NEEDED?
1555 =================================
1557 Under normal operation, memory operation reordering is generally not going to
1558 be a problem as a single-threaded linear piece of code will still appear to
1559 work correctly, even if it's in an SMP kernel. There are, however, four
1560 circumstances in which reordering definitely _could_ be a problem:
1562 (*) Interprocessor interaction.
1564 (*) Atomic operations.
1566 (*) Accessing devices.
1571 INTERPROCESSOR INTERACTION
1572 --------------------------
1574 When there's a system with more than one processor, more than one CPU in the
1575 system may be working on the same data set at the same time. This can cause
1576 synchronisation problems, and the usual way of dealing with them is to use
1577 locks. Locks, however, are quite expensive, and so it may be preferable to
1578 operate without the use of a lock if at all possible. In such a case
1579 operations that affect both CPUs may have to be carefully ordered to prevent
1582 Consider, for example, the R/W semaphore slow path. Here a waiting process is
1583 queued on the semaphore, by virtue of it having a piece of its stack linked to
1584 the semaphore's list of waiting processes:
1586 struct rw_semaphore {
1589 struct list_head waiters;
1592 struct rwsem_waiter {
1593 struct list_head list;
1594 struct task_struct *task;
1597 To wake up a particular waiter, the up_read() or up_write() functions have to:
1599 (1) read the next pointer from this waiter's record to know as to where the
1600 next waiter record is;
1602 (2) read the pointer to the waiter's task structure;
1604 (3) clear the task pointer to tell the waiter it has been given the semaphore;
1606 (4) call wake_up_process() on the task; and
1608 (5) release the reference held on the waiter's task struct.
1610 In other words, it has to perform this sequence of events:
1612 LOAD waiter->list.next;
1618 and if any of these steps occur out of order, then the whole thing may
1621 Once it has queued itself and dropped the semaphore lock, the waiter does not
1622 get the lock again; it instead just waits for its task pointer to be cleared
1623 before proceeding. Since the record is on the waiter's stack, this means that
1624 if the task pointer is cleared _before_ the next pointer in the list is read,
1625 another CPU might start processing the waiter and might clobber the waiter's
1626 stack before the up*() function has a chance to read the next pointer.
1628 Consider then what might happen to the above sequence of events:
1631 =============================== ===============================
1638 Woken up by other event
1643 foo() clobbers *waiter
1645 LOAD waiter->list.next;
1648 This could be dealt with using the semaphore lock, but then the down_xxx()
1649 function has to needlessly get the spinlock again after being woken up.
1651 The way to deal with this is to insert a general SMP memory barrier:
1653 LOAD waiter->list.next;
1660 In this case, the barrier makes a guarantee that all memory accesses before the
1661 barrier will appear to happen before all the memory accesses after the barrier
1662 with respect to the other CPUs on the system. It does _not_ guarantee that all
1663 the memory accesses before the barrier will be complete by the time the barrier
1664 instruction itself is complete.
1666 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
1667 compiler barrier, thus making sure the compiler emits the instructions in the
1668 right order without actually intervening in the CPU. Since there's only one
1669 CPU, that CPU's dependency ordering logic will take care of everything else.
1675 Whilst they are technically interprocessor interaction considerations, atomic
1676 operations are noted specially as some of them imply full memory barriers and
1677 some don't, but they're very heavily relied on as a group throughout the
1680 Any atomic operation that modifies some state in memory and returns information
1681 about the state (old or new) implies an SMP-conditional general memory barrier
1682 (smp_mb()) on each side of the actual operation (with the exception of
1683 explicit lock operations, described later). These include:
1688 atomic_inc_return();
1689 atomic_dec_return();
1690 atomic_add_return();
1691 atomic_sub_return();
1692 atomic_inc_and_test();
1693 atomic_dec_and_test();
1694 atomic_sub_and_test();
1695 atomic_add_negative();
1696 atomic_add_unless(); /* when succeeds (returns 1) */
1698 test_and_clear_bit();
1699 test_and_change_bit();
1701 These are used for such things as implementing LOCK-class and UNLOCK-class
1702 operations and adjusting reference counters towards object destruction, and as
1703 such the implicit memory barrier effects are necessary.
1706 The following operations are potential problems as they do _not_ imply memory
1707 barriers, but might be used for implementing such things as UNLOCK-class
1715 With these the appropriate explicit memory barrier should be used if necessary
1716 (smp_mb__before_clear_bit() for instance).
1719 The following also do _not_ imply memory barriers, and so may require explicit
1720 memory barriers under some circumstances (smp_mb__before_atomic_dec() for
1728 If they're used for statistics generation, then they probably don't need memory
1729 barriers, unless there's a coupling between statistical data.
1731 If they're used for reference counting on an object to control its lifetime,
1732 they probably don't need memory barriers because either the reference count
1733 will be adjusted inside a locked section, or the caller will already hold
1734 sufficient references to make the lock, and thus a memory barrier unnecessary.
1736 If they're used for constructing a lock of some description, then they probably
1737 do need memory barriers as a lock primitive generally has to do things in a
1740 Basically, each usage case has to be carefully considered as to whether memory
1741 barriers are needed or not.
1743 The following operations are special locking primitives:
1745 test_and_set_bit_lock();
1747 __clear_bit_unlock();
1749 These implement LOCK-class and UNLOCK-class operations. These should be used in
1750 preference to other operations when implementing locking primitives, because
1751 their implementations can be optimised on many architectures.
1753 [!] Note that special memory barrier primitives are available for these
1754 situations because on some CPUs the atomic instructions used imply full memory
1755 barriers, and so barrier instructions are superfluous in conjunction with them,
1756 and in such cases the special barrier primitives will be no-ops.
1758 See Documentation/atomic_ops.txt for more information.
1764 Many devices can be memory mapped, and so appear to the CPU as if they're just
1765 a set of memory locations. To control such a device, the driver usually has to
1766 make the right memory accesses in exactly the right order.
1768 However, having a clever CPU or a clever compiler creates a potential problem
1769 in that the carefully sequenced accesses in the driver code won't reach the
1770 device in the requisite order if the CPU or the compiler thinks it is more
1771 efficient to reorder, combine or merge accesses - something that would cause
1772 the device to malfunction.
1774 Inside of the Linux kernel, I/O should be done through the appropriate accessor
1775 routines - such as inb() or writel() - which know how to make such accesses
1776 appropriately sequential. Whilst this, for the most part, renders the explicit
1777 use of memory barriers unnecessary, there are a couple of situations where they
1780 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
1781 so for _all_ general drivers locks should be used and mmiowb() must be
1782 issued prior to unlocking the critical section.
1784 (2) If the accessor functions are used to refer to an I/O memory window with
1785 relaxed memory access properties, then _mandatory_ memory barriers are
1786 required to enforce ordering.
1788 See Documentation/DocBook/deviceiobook.tmpl for more information.
1794 A driver may be interrupted by its own interrupt service routine, and thus the
1795 two parts of the driver may interfere with each other's attempts to control or
1798 This may be alleviated - at least in part - by disabling local interrupts (a
1799 form of locking), such that the critical operations are all contained within
1800 the interrupt-disabled section in the driver. Whilst the driver's interrupt
1801 routine is executing, the driver's core may not run on the same CPU, and its
1802 interrupt is not permitted to happen again until the current interrupt has been
1803 handled, thus the interrupt handler does not need to lock against that.
1805 However, consider a driver that was talking to an ethernet card that sports an
1806 address register and a data register. If that driver's core talks to the card
1807 under interrupt-disablement and then the driver's interrupt handler is invoked:
1818 The store to the data register might happen after the second store to the
1819 address register if ordering rules are sufficiently relaxed:
1821 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
1824 If ordering rules are relaxed, it must be assumed that accesses done inside an
1825 interrupt disabled section may leak outside of it and may interleave with
1826 accesses performed in an interrupt - and vice versa - unless implicit or
1827 explicit barriers are used.
1829 Normally this won't be a problem because the I/O accesses done inside such
1830 sections will include synchronous load operations on strictly ordered I/O
1831 registers that form implicit I/O barriers. If this isn't sufficient then an
1832 mmiowb() may need to be used explicitly.
1835 A similar situation may occur between an interrupt routine and two routines
1836 running on separate CPUs that communicate with each other. If such a case is
1837 likely, then interrupt-disabling locks should be used to guarantee ordering.
1840 ==========================
1841 KERNEL I/O BARRIER EFFECTS
1842 ==========================
1844 When accessing I/O memory, drivers should use the appropriate accessor
1849 These are intended to talk to I/O space rather than memory space, but
1850 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
1851 indeed have special I/O space access cycles and instructions, but many
1852 CPUs don't have such a concept.
1854 The PCI bus, amongst others, defines an I/O space concept which - on such
1855 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
1856 space. However, it may also be mapped as a virtual I/O space in the CPU's
1857 memory map, particularly on those CPUs that don't support alternate I/O
1860 Accesses to this space may be fully synchronous (as on i386), but
1861 intermediary bridges (such as the PCI host bridge) may not fully honour
1864 They are guaranteed to be fully ordered with respect to each other.
1866 They are not guaranteed to be fully ordered with respect to other types of
1867 memory and I/O operation.
1869 (*) readX(), writeX():
1871 Whether these are guaranteed to be fully ordered and uncombined with
1872 respect to each other on the issuing CPU depends on the characteristics
1873 defined for the memory window through which they're accessing. On later
1874 i386 architecture machines, for example, this is controlled by way of the
1877 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
1878 provided they're not accessing a prefetchable device.
1880 However, intermediary hardware (such as a PCI bridge) may indulge in
1881 deferral if it so wishes; to flush a store, a load from the same location
1882 is preferred[*], but a load from the same device or from configuration
1883 space should suffice for PCI.
1885 [*] NOTE! attempting to load from the same location as was written to may
1886 cause a malfunction - consider the 16550 Rx/Tx serial registers for
1889 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
1890 force stores to be ordered.
1892 Please refer to the PCI specification for more information on interactions
1893 between PCI transactions.
1897 These are similar to readX(), but are not guaranteed to be ordered in any
1898 way. Be aware that there is no I/O read barrier available.
1900 (*) ioreadX(), iowriteX()
1902 These will perform appropriately for the type of access they're actually
1903 doing, be it inX()/outX() or readX()/writeX().
1906 ========================================
1907 ASSUMED MINIMUM EXECUTION ORDERING MODEL
1908 ========================================
1910 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
1911 maintain the appearance of program causality with respect to itself. Some CPUs
1912 (such as i386 or x86_64) are more constrained than others (such as powerpc or
1913 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
1914 of arch-specific code.
1916 This means that it must be considered that the CPU will execute its instruction
1917 stream in any order it feels like - or even in parallel - provided that if an
1918 instruction in the stream depends on an earlier instruction, then that
1919 earlier instruction must be sufficiently complete[*] before the later
1920 instruction may proceed; in other words: provided that the appearance of
1921 causality is maintained.
1923 [*] Some instructions have more than one effect - such as changing the
1924 condition codes, changing registers or changing memory - and different
1925 instructions may depend on different effects.
1927 A CPU may also discard any instruction sequence that winds up having no
1928 ultimate effect. For example, if two adjacent instructions both load an
1929 immediate value into the same register, the first may be discarded.
1932 Similarly, it has to be assumed that compiler might reorder the instruction
1933 stream in any way it sees fit, again provided the appearance of causality is
1937 ============================
1938 THE EFFECTS OF THE CPU CACHE
1939 ============================
1941 The way cached memory operations are perceived across the system is affected to
1942 a certain extent by the caches that lie between CPUs and memory, and by the
1943 memory coherence system that maintains the consistency of state in the system.
1945 As far as the way a CPU interacts with another part of the system through the
1946 caches goes, the memory system has to include the CPU's caches, and memory
1947 barriers for the most part act at the interface between the CPU and its cache
1948 (memory barriers logically act on the dotted line in the following diagram):
1950 <--- CPU ---> : <----------- Memory ----------->
1952 +--------+ +--------+ : +--------+ +-----------+
1953 | | | | : | | | | +--------+
1954 | CPU | | Memory | : | CPU | | | | |
1955 | Core |--->| Access |----->| Cache |<-->| | | |
1956 | | | Queue | : | | | |--->| Memory |
1957 | | | | : | | | | | |
1958 +--------+ +--------+ : +--------+ | | | |
1959 : | Cache | +--------+
1961 : | Mechanism | +--------+
1962 +--------+ +--------+ : +--------+ | | | |
1963 | | | | : | | | | | |
1964 | CPU | | Memory | : | CPU | | |--->| Device |
1965 | Core |--->| Access |----->| Cache |<-->| | | |
1966 | | | Queue | : | | | | | |
1967 | | | | : | | | | +--------+
1968 +--------+ +--------+ : +--------+ +-----------+
1972 Although any particular load or store may not actually appear outside of the
1973 CPU that issued it since it may have been satisfied within the CPU's own cache,
1974 it will still appear as if the full memory access had taken place as far as the
1975 other CPUs are concerned since the cache coherency mechanisms will migrate the
1976 cacheline over to the accessing CPU and propagate the effects upon conflict.
1978 The CPU core may execute instructions in any order it deems fit, provided the
1979 expected program causality appears to be maintained. Some of the instructions
1980 generate load and store operations which then go into the queue of memory
1981 accesses to be performed. The core may place these in the queue in any order
1982 it wishes, and continue execution until it is forced to wait for an instruction
1985 What memory barriers are concerned with is controlling the order in which
1986 accesses cross from the CPU side of things to the memory side of things, and
1987 the order in which the effects are perceived to happen by the other observers
1990 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
1991 their own loads and stores as if they had happened in program order.
1993 [!] MMIO or other device accesses may bypass the cache system. This depends on
1994 the properties of the memory window through which devices are accessed and/or
1995 the use of any special device communication instructions the CPU may have.
2001 Life isn't quite as simple as it may appear above, however: for while the
2002 caches are expected to be coherent, there's no guarantee that that coherency
2003 will be ordered. This means that whilst changes made on one CPU will
2004 eventually become visible on all CPUs, there's no guarantee that they will
2005 become apparent in the same order on those other CPUs.
2008 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2009 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2014 +--------+ : +--->| Cache A |<------->| |
2015 | | : | +---------+ | |
2017 | | : | +---------+ | |
2018 +--------+ : +--->| Cache B |<------->| |
2021 : +---------+ | System |
2022 +--------+ : +--->| Cache C |<------->| |
2023 | | : | +---------+ | |
2025 | | : | +---------+ | |
2026 +--------+ : +--->| Cache D |<------->| |
2031 Imagine the system has the following properties:
2033 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2036 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2039 (*) whilst the CPU core is interrogating one cache, the other cache may be
2040 making use of the bus to access the rest of the system - perhaps to
2041 displace a dirty cacheline or to do a speculative load;
2043 (*) each cache has a queue of operations that need to be applied to that cache
2044 to maintain coherency with the rest of the system;
2046 (*) the coherency queue is not flushed by normal loads to lines already
2047 present in the cache, even though the contents of the queue may
2048 potentially affect those loads.
2050 Imagine, then, that two writes are made on the first CPU, with a write barrier
2051 between them to guarantee that they will appear to reach that CPU's caches in
2052 the requisite order:
2055 =============== =============== =======================================
2056 u == 0, v == 1 and p == &u, q == &u
2058 smp_wmb(); Make sure change to v is visible before
2060 <A:modify v=2> v is now in cache A exclusively
2062 <B:modify p=&v> p is now in cache B exclusively
2064 The write memory barrier forces the other CPUs in the system to perceive that
2065 the local CPU's caches have apparently been updated in the correct order. But
2066 now imagine that the second CPU wants to read those values:
2069 =============== =============== =======================================
2074 The above pair of reads may then fail to happen in the expected order, as the
2075 cacheline holding p may get updated in one of the second CPU's caches whilst
2076 the update to the cacheline holding v is delayed in the other of the second
2077 CPU's caches by some other cache event:
2080 =============== =============== =======================================
2081 u == 0, v == 1 and p == &u, q == &u
2084 <A:modify v=2> <C:busy>
2088 <B:modify p=&v> <D:commit p=&v>
2091 <C:read *q> Reads from v before v updated in cache
2095 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2096 no guarantee that, without intervention, the order of update will be the same
2097 as that committed on CPU 1.
2100 To intervene, we need to interpolate a data dependency barrier or a read
2101 barrier between the loads. This will force the cache to commit its coherency
2102 queue before processing any further requests:
2105 =============== =============== =======================================
2106 u == 0, v == 1 and p == &u, q == &u
2109 <A:modify v=2> <C:busy>
2113 <B:modify p=&v> <D:commit p=&v>
2115 smp_read_barrier_depends()
2119 <C:read *q> Reads from v after v updated in cache
2122 This sort of problem can be encountered on DEC Alpha processors as they have a
2123 split cache that improves performance by making better use of the data bus.
2124 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2125 access depends on a read, not all do, so it may not be relied on.
2127 Other CPUs may also have split caches, but must coordinate between the various
2128 cachelets for normal memory accesses. The semantics of the Alpha removes the
2129 need for coordination in the absence of memory barriers.
2132 CACHE COHERENCY VS DMA
2133 ----------------------
2135 Not all systems maintain cache coherency with respect to devices doing DMA. In
2136 such cases, a device attempting DMA may obtain stale data from RAM because
2137 dirty cache lines may be resident in the caches of various CPUs, and may not
2138 have been written back to RAM yet. To deal with this, the appropriate part of
2139 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2140 invalidate them as well).
2142 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2143 cache lines being written back to RAM from a CPU's cache after the device has
2144 installed its own data, or cache lines present in the CPU's cache may simply
2145 obscure the fact that RAM has been updated, until at such time as the cacheline
2146 is discarded from the CPU's cache and reloaded. To deal with this, the
2147 appropriate part of the kernel must invalidate the overlapping bits of the
2150 See Documentation/cachetlb.txt for more information on cache management.
2153 CACHE COHERENCY VS MMIO
2154 -----------------------
2156 Memory mapped I/O usually takes place through memory locations that are part of
2157 a window in the CPU's memory space that has different properties assigned than
2158 the usual RAM directed window.
2160 Amongst these properties is usually the fact that such accesses bypass the
2161 caching entirely and go directly to the device buses. This means MMIO accesses
2162 may, in effect, overtake accesses to cached memory that were emitted earlier.
2163 A memory barrier isn't sufficient in such a case, but rather the cache must be
2164 flushed between the cached memory write and the MMIO access if the two are in
2168 =========================
2169 THE THINGS CPUS GET UP TO
2170 =========================
2172 A programmer might take it for granted that the CPU will perform memory
2173 operations in exactly the order specified, so that if the CPU is, for example,
2174 given the following piece of code to execute:
2182 they would then expect that the CPU will complete the memory operation for each
2183 instruction before moving on to the next one, leading to a definite sequence of
2184 operations as seen by external observers in the system:
2186 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2189 Reality is, of course, much messier. With many CPUs and compilers, the above
2190 assumption doesn't hold because:
2192 (*) loads are more likely to need to be completed immediately to permit
2193 execution progress, whereas stores can often be deferred without a
2196 (*) loads may be done speculatively, and the result discarded should it prove
2197 to have been unnecessary;
2199 (*) loads may be done speculatively, leading to the result having been fetched
2200 at the wrong time in the expected sequence of events;
2202 (*) the order of the memory accesses may be rearranged to promote better use
2203 of the CPU buses and caches;
2205 (*) loads and stores may be combined to improve performance when talking to
2206 memory or I/O hardware that can do batched accesses of adjacent locations,
2207 thus cutting down on transaction setup costs (memory and PCI devices may
2208 both be able to do this); and
2210 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2211 mechanisms may alleviate this - once the store has actually hit the cache
2212 - there's no guarantee that the coherency management will be propagated in
2213 order to other CPUs.
2215 So what another CPU, say, might actually observe from the above piece of code
2218 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2220 (Where "LOAD {*C,*D}" is a combined load)
2223 However, it is guaranteed that a CPU will be self-consistent: it will see its
2224 _own_ accesses appear to be correctly ordered, without the need for a memory
2225 barrier. For instance with the following code:
2234 and assuming no intervention by an external influence, it can be assumed that
2235 the final result will appear to be:
2237 U == the original value of *A
2242 The code above may cause the CPU to generate the full sequence of memory
2245 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2247 in that order, but, without intervention, the sequence may have almost any
2248 combination of elements combined or discarded, provided the program's view of
2249 the world remains consistent.
2251 The compiler may also combine, discard or defer elements of the sequence before
2252 the CPU even sees them.
2263 since, without a write barrier, it can be assumed that the effect of the
2264 storage of V to *A is lost. Similarly:
2269 may, without a memory barrier, be reduced to:
2274 and the LOAD operation never appear outside of the CPU.
2277 AND THEN THERE'S THE ALPHA
2278 --------------------------
2280 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2281 some versions of the Alpha CPU have a split data cache, permitting them to have
2282 two semantically-related cache lines updated at separate times. This is where
2283 the data dependency barrier really becomes necessary as this synchronises both
2284 caches with the memory coherence system, thus making it seem like pointer
2285 changes vs new data occur in the right order.
2287 The Alpha defines the Linux kernel's memory barrier model.
2289 See the subsection on "Cache Coherency" above.
2299 Memory barriers can be used to implement circular buffering without the need
2300 of a lock to serialise the producer with the consumer. See:
2302 Documentation/circular-buffers.txt
2311 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2313 Chapter 5.2: Physical Address Space Characteristics
2314 Chapter 5.4: Caches and Write Buffers
2315 Chapter 5.5: Data Sharing
2316 Chapter 5.6: Read/Write Ordering
2318 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2319 Chapter 7.1: Memory-Access Ordering
2320 Chapter 7.4: Buffering and Combining Memory Writes
2322 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2323 System Programming Guide
2324 Chapter 7.1: Locked Atomic Operations
2325 Chapter 7.2: Memory Ordering
2326 Chapter 7.4: Serializing Instructions
2328 The SPARC Architecture Manual, Version 9
2329 Chapter 8: Memory Models
2330 Appendix D: Formal Specification of the Memory Models
2331 Appendix J: Programming with the Memory Models
2333 UltraSPARC Programmer Reference Manual
2334 Chapter 5: Memory Accesses and Cacheability
2335 Chapter 15: Sparc-V9 Memory Models
2337 UltraSPARC III Cu User's Manual
2338 Chapter 9: Memory Models
2340 UltraSPARC IIIi Processor User's Manual
2341 Chapter 8: Memory Models
2343 UltraSPARC Architecture 2005
2345 Appendix D: Formal Specifications of the Memory Models
2347 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2348 Chapter 8: Memory Models
2349 Appendix F: Caches and Cache Coherency
2351 Solaris Internals, Core Kernel Architecture, p63-68:
2352 Chapter 3.3: Hardware Considerations for Locks and
2355 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2356 for Kernel Programmers:
2357 Chapter 13: Other Memory Models
2359 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2360 Section 2.6: Speculation
2361 Section 4.4: Memory Access