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
536 <data dependency barrier>
539 This will not have the desired effect because there is no actual data
540 dependency, but rather a control dependency that the CPU may short-circuit by
541 attempting to predict the outcome in advance. In such a case what's actually
554 When dealing with CPU-CPU interactions, certain types of memory barrier should
555 always be paired. A lack of appropriate pairing is almost certainly an error.
557 A write barrier should always be paired with a data dependency barrier or read
558 barrier, though a general barrier would also be viable. Similarly a read
559 barrier or a data dependency barrier should always be paired with at least an
560 write barrier, though, again, a general barrier is viable:
563 =============== ===============
573 =============== ===============================
577 <data dependency barrier>
580 Basically, the read barrier always has to be there, even though it can be of
583 [!] Note that the stores before the write barrier would normally be expected to
584 match the loads after the read barrier or the data dependency barrier, and vice
588 =============== ===============
589 a = 1; }---- --->{ v = c
591 <write barrier> \ <read barrier>
592 c = 3; } / \ { x = a;
593 d = 4; }---- --->{ y = b;
596 EXAMPLES OF MEMORY BARRIER SEQUENCES
597 ------------------------------------
599 Firstly, write barriers act as partial orderings on store operations.
600 Consider the following sequence of events:
603 =======================
611 This sequence of events is committed to the memory coherence system in an order
612 that the rest of the system might perceive as the unordered set of { STORE A,
613 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
618 | |------>| C=3 | } /\
619 | | : +------+ }----- \ -----> Events perceptible to
620 | | : | A=1 | } \/ the rest of the system
622 | CPU 1 | : | B=2 | }
624 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
625 | | +------+ } requires all stores prior to the
626 | | : | E=5 | } barrier to be committed before
627 | | : +------+ } further stores may take place
632 | Sequence in which stores are committed to the
633 | memory system by CPU 1
637 Secondly, data dependency barriers act as partial orderings on data-dependent
638 loads. Consider the following sequence of events:
641 ======================= =======================
642 { B = 7; X = 9; Y = 8; C = &Y }
647 STORE D = 4 LOAD C (gets &B)
650 Without intervention, CPU 2 may perceive the events on CPU 1 in some
651 effectively random order, despite the write barrier issued by CPU 1:
654 | | +------+ +-------+ | Sequence of update
655 | |------>| B=2 |----- --->| Y->8 | | of perception on
656 | | : +------+ \ +-------+ | CPU 2
657 | CPU 1 | : | A=1 | \ --->| C->&Y | V
658 | | +------+ | +-------+
659 | | wwwwwwwwwwwwwwww | : :
661 | | : | C=&B |--- | : : +-------+
662 | | : +------+ \ | +-------+ | |
663 | |------>| D=4 | ----------->| C->&B |------>| |
664 | | +------+ | +-------+ | |
665 +-------+ : : | : : | |
669 Apparently incorrect ---> | | B->7 |------>| |
670 perception of B (!) | +-------+ | |
673 The load of X holds ---> \ | X->9 |------>| |
674 up the maintenance \ +-------+ | |
675 of coherence of B ----->| B->2 | +-------+
680 In the above example, CPU 2 perceives that B is 7, despite the load of *C
681 (which would be B) coming after the LOAD of C.
683 If, however, a data dependency barrier were to be placed between the load of C
684 and the load of *C (ie: B) on CPU 2:
687 ======================= =======================
688 { B = 7; X = 9; Y = 8; C = &Y }
693 STORE D = 4 LOAD C (gets &B)
694 <data dependency barrier>
697 then the following will occur:
700 | | +------+ +-------+
701 | |------>| B=2 |----- --->| Y->8 |
702 | | : +------+ \ +-------+
703 | CPU 1 | : | A=1 | \ --->| C->&Y |
704 | | +------+ | +-------+
705 | | wwwwwwwwwwwwwwww | : :
707 | | : | C=&B |--- | : : +-------+
708 | | : +------+ \ | +-------+ | |
709 | |------>| D=4 | ----------->| C->&B |------>| |
710 | | +------+ | +-------+ | |
711 +-------+ : : | : : | |
717 Makes sure all effects ---> \ ddddddddddddddddd | |
718 prior to the store of C \ +-------+ | |
719 are perceptible to ----->| B->2 |------>| |
720 subsequent loads +-------+ | |
724 And thirdly, a read barrier acts as a partial order on loads. Consider the
725 following sequence of events:
728 ======================= =======================
736 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
737 some effectively random order, despite the write barrier issued by CPU 1:
740 | | +------+ +-------+
741 | |------>| A=1 |------ --->| A->0 |
742 | | +------+ \ +-------+
743 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
744 | | +------+ | +-------+
745 | |------>| B=2 |--- | : :
746 | | +------+ \ | : : +-------+
747 +-------+ : : \ | +-------+ | |
748 ---------->| B->2 |------>| |
749 | +-------+ | CPU 2 |
760 If, however, a read barrier were to be placed between the load of B and the
764 ======================= =======================
773 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
777 | | +------+ +-------+
778 | |------>| A=1 |------ --->| A->0 |
779 | | +------+ \ +-------+
780 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
781 | | +------+ | +-------+
782 | |------>| B=2 |--- | : :
783 | | +------+ \ | : : +-------+
784 +-------+ : : \ | +-------+ | |
785 ---------->| B->2 |------>| |
786 | +-------+ | CPU 2 |
789 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
790 barrier causes all effects \ +-------+ | |
791 prior to the storage of B ---->| A->1 |------>| |
792 to be perceptible to CPU 2 +-------+ | |
796 To illustrate this more completely, consider what could happen if the code
797 contained a load of A either side of the read barrier:
800 ======================= =======================
806 LOAD A [first load of A]
808 LOAD A [second load of A]
810 Even though the two loads of A both occur after the load of B, they may both
811 come up with different values:
814 | | +------+ +-------+
815 | |------>| A=1 |------ --->| A->0 |
816 | | +------+ \ +-------+
817 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
818 | | +------+ | +-------+
819 | |------>| B=2 |--- | : :
820 | | +------+ \ | : : +-------+
821 +-------+ : : \ | +-------+ | |
822 ---------->| B->2 |------>| |
823 | +-------+ | CPU 2 |
827 | | A->0 |------>| 1st |
829 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
830 barrier causes all effects \ +-------+ | |
831 prior to the storage of B ---->| A->1 |------>| 2nd |
832 to be perceptible to CPU 2 +-------+ | |
836 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
837 before the read barrier completes anyway:
840 | | +------+ +-------+
841 | |------>| A=1 |------ --->| A->0 |
842 | | +------+ \ +-------+
843 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
844 | | +------+ | +-------+
845 | |------>| B=2 |--- | : :
846 | | +------+ \ | : : +-------+
847 +-------+ : : \ | +-------+ | |
848 ---------->| B->2 |------>| |
849 | +-------+ | CPU 2 |
853 ---->| A->1 |------>| 1st |
855 rrrrrrrrrrrrrrrrr | |
857 | A->1 |------>| 2nd |
862 The guarantee is that the second load will always come up with A == 1 if the
863 load of B came up with B == 2. No such guarantee exists for the first load of
864 A; that may come up with either A == 0 or A == 1.
867 READ MEMORY BARRIERS VS LOAD SPECULATION
868 ----------------------------------------
870 Many CPUs speculate with loads: that is they see that they will need to load an
871 item from memory, and they find a time where they're not using the bus for any
872 other loads, and so do the load in advance - even though they haven't actually
873 got to that point in the instruction execution flow yet. This permits the
874 actual load instruction to potentially complete immediately because the CPU
875 already has the value to hand.
877 It may turn out that the CPU didn't actually need the value - perhaps because a
878 branch circumvented the load - in which case it can discard the value or just
879 cache it for later use.
884 ======================= =======================
886 DIVIDE } Divide instructions generally
887 DIVIDE } take a long time to perform
890 Which might appear as this:
894 --->| B->2 |------>| |
898 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
899 division speculates on the +-------+ ~ | |
903 Once the divisions are complete --> : : ~-->| |
904 the CPU can then perform the : : | |
905 LOAD with immediate effect : : +-------+
908 Placing a read barrier or a data dependency barrier just before the second
912 ======================= =======================
919 will force any value speculatively obtained to be reconsidered to an extent
920 dependent on the type of barrier used. If there was no change made to the
921 speculated memory location, then the speculated value will just be used:
925 --->| B->2 |------>| |
929 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
930 division speculates on the +-------+ ~ | |
935 rrrrrrrrrrrrrrrr~ | |
942 but if there was an update or an invalidation from another CPU pending, then
943 the speculation will be cancelled and the value reloaded:
947 --->| B->2 |------>| |
951 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
952 division speculates on the +-------+ ~ | |
957 rrrrrrrrrrrrrrrrr | |
959 The speculation is discarded ---> --->| A->1 |------>| |
960 and an updated value is +-------+ | |
961 retrieved : : +-------+
967 Transitivity is a deeply intuitive notion about ordering that is not
968 always provided by real computer systems. The following example
969 demonstrates transitivity (also called "cumulativity"):
972 ======================= ======================= =======================
974 STORE X=1 LOAD X STORE Y=1
975 <general barrier> <general barrier>
978 Suppose that CPU 2's load from X returns 1 and its load from Y returns 0.
979 This indicates that CPU 2's load from X in some sense follows CPU 1's
980 store to X and that CPU 2's load from Y in some sense preceded CPU 3's
981 store to Y. The question is then "Can CPU 3's load from X return 0?"
983 Because CPU 2's load from X in some sense came after CPU 1's store, it
984 is natural to expect that CPU 3's load from X must therefore return 1.
985 This expectation is an example of transitivity: if a load executing on
986 CPU A follows a load from the same variable executing on CPU B, then
987 CPU A's load must either return the same value that CPU B's load did,
988 or must return some later value.
990 In the Linux kernel, use of general memory barriers guarantees
991 transitivity. Therefore, in the above example, if CPU 2's load from X
992 returns 1 and its load from Y returns 0, then CPU 3's load from X must
995 However, transitivity is -not- guaranteed for read or write barriers.
996 For example, suppose that CPU 2's general barrier in the above example
997 is changed to a read barrier as shown below:
1000 ======================= ======================= =======================
1002 STORE X=1 LOAD X STORE Y=1
1003 <read barrier> <general barrier>
1006 This substitution destroys transitivity: in this example, it is perfectly
1007 legal for CPU 2's load from X to return 1, its load from Y to return 0,
1008 and CPU 3's load from X to return 0.
1010 The key point is that although CPU 2's read barrier orders its pair
1011 of loads, it does not guarantee to order CPU 1's store. Therefore, if
1012 this example runs on a system where CPUs 1 and 2 share a store buffer
1013 or a level of cache, CPU 2 might have early access to CPU 1's writes.
1014 General barriers are therefore required to ensure that all CPUs agree
1015 on the combined order of CPU 1's and CPU 2's accesses.
1017 To reiterate, if your code requires transitivity, use general barriers
1021 ========================
1022 EXPLICIT KERNEL BARRIERS
1023 ========================
1025 The Linux kernel has a variety of different barriers that act at different
1028 (*) Compiler barrier.
1030 (*) CPU memory barriers.
1032 (*) MMIO write barrier.
1038 The Linux kernel has an explicit compiler barrier function that prevents the
1039 compiler from moving the memory accesses either side of it to the other side:
1043 This is a general barrier - lesser varieties of compiler barrier do not exist.
1045 The compiler barrier has no direct effect on the CPU, which may then reorder
1046 things however it wishes.
1052 The Linux kernel has eight basic CPU memory barriers:
1054 TYPE MANDATORY SMP CONDITIONAL
1055 =============== ======================= ===========================
1056 GENERAL mb() smp_mb()
1057 WRITE wmb() smp_wmb()
1058 READ rmb() smp_rmb()
1059 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1062 All memory barriers except the data dependency barriers imply a compiler
1063 barrier. Data dependencies do not impose any additional compiler ordering.
1065 Aside: In the case of data dependencies, the compiler would be expected to
1066 issue the loads in the correct order (eg. `a[b]` would have to load the value
1067 of b before loading a[b]), however there is no guarantee in the C specification
1068 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1069 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1070 problem of a compiler reloading b after having loaded a[b], thus having a newer
1071 copy of b than a[b]. A consensus has not yet been reached about these problems,
1072 however the ACCESS_ONCE macro is a good place to start looking.
1074 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1075 systems because it is assumed that a CPU will appear to be self-consistent,
1076 and will order overlapping accesses correctly with respect to itself.
1078 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1079 references to shared memory on SMP systems, though the use of locking instead
1082 Mandatory barriers should not be used to control SMP effects, since mandatory
1083 barriers unnecessarily impose overhead on UP systems. They may, however, be
1084 used to control MMIO effects on accesses through relaxed memory I/O windows.
1085 These are required even on non-SMP systems as they affect the order in which
1086 memory operations appear to a device by prohibiting both the compiler and the
1087 CPU from reordering them.
1090 There are some more advanced barrier functions:
1092 (*) set_mb(var, value)
1094 This assigns the value to the variable and then inserts a full memory
1095 barrier after it, depending on the function. It isn't guaranteed to
1096 insert anything more than a compiler barrier in a UP compilation.
1099 (*) smp_mb__before_atomic_dec();
1100 (*) smp_mb__after_atomic_dec();
1101 (*) smp_mb__before_atomic_inc();
1102 (*) smp_mb__after_atomic_inc();
1104 These are for use with atomic add, subtract, increment and decrement
1105 functions that don't return a value, especially when used for reference
1106 counting. These functions do not imply memory barriers.
1108 As an example, consider a piece of code that marks an object as being dead
1109 and then decrements the object's reference count:
1112 smp_mb__before_atomic_dec();
1113 atomic_dec(&obj->ref_count);
1115 This makes sure that the death mark on the object is perceived to be set
1116 *before* the reference counter is decremented.
1118 See Documentation/atomic_ops.txt for more information. See the "Atomic
1119 operations" subsection for information on where to use these.
1122 (*) smp_mb__before_clear_bit(void);
1123 (*) smp_mb__after_clear_bit(void);
1125 These are for use similar to the atomic inc/dec barriers. These are
1126 typically used for bitwise unlocking operations, so care must be taken as
1127 there are no implicit memory barriers here either.
1129 Consider implementing an unlock operation of some nature by clearing a
1130 locking bit. The clear_bit() would then need to be barriered like this:
1132 smp_mb__before_clear_bit();
1135 This prevents memory operations before the clear leaking to after it. See
1136 the subsection on "Locking Functions" with reference to UNLOCK operation
1139 See Documentation/atomic_ops.txt for more information. See the "Atomic
1140 operations" subsection for information on where to use these.
1146 The Linux kernel also has a special barrier for use with memory-mapped I/O
1151 This is a variation on the mandatory write barrier that causes writes to weakly
1152 ordered I/O regions to be partially ordered. Its effects may go beyond the
1153 CPU->Hardware interface and actually affect the hardware at some level.
1155 See the subsection "Locks vs I/O accesses" for more information.
1158 ===============================
1159 IMPLICIT KERNEL MEMORY BARRIERS
1160 ===============================
1162 Some of the other functions in the linux kernel imply memory barriers, amongst
1163 which are locking and scheduling functions.
1165 This specification is a _minimum_ guarantee; any particular architecture may
1166 provide more substantial guarantees, but these may not be relied upon outside
1167 of arch specific code.
1173 The Linux kernel has a number of locking constructs:
1182 In all cases there are variants on "LOCK" operations and "UNLOCK" operations
1183 for each construct. These operations all imply certain barriers:
1185 (1) LOCK operation implication:
1187 Memory operations issued after the LOCK will be completed after the LOCK
1188 operation has completed.
1190 Memory operations issued before the LOCK may be completed after the LOCK
1191 operation has completed.
1193 (2) UNLOCK operation implication:
1195 Memory operations issued before the UNLOCK will be completed before the
1196 UNLOCK operation has completed.
1198 Memory operations issued after the UNLOCK may be completed before the
1199 UNLOCK operation has completed.
1201 (3) LOCK vs LOCK implication:
1203 All LOCK operations issued before another LOCK operation will be completed
1204 before that LOCK operation.
1206 (4) LOCK vs UNLOCK implication:
1208 All LOCK operations issued before an UNLOCK operation will be completed
1209 before the UNLOCK operation.
1211 All UNLOCK operations issued before a LOCK operation will be completed
1212 before the LOCK operation.
1214 (5) Failed conditional LOCK implication:
1216 Certain variants of the LOCK operation may fail, either due to being
1217 unable to get the lock immediately, or due to receiving an unblocked
1218 signal whilst asleep waiting for the lock to become available. Failed
1219 locks do not imply any sort of barrier.
1221 Therefore, from (1), (2) and (4) an UNLOCK followed by an unconditional LOCK is
1222 equivalent to a full barrier, but a LOCK followed by an UNLOCK is not.
1224 [!] Note: one of the consequences of LOCKs and UNLOCKs being only one-way
1225 barriers is that the effects of instructions outside of a critical section
1226 may seep into the inside of the critical section.
1228 A LOCK followed by an UNLOCK may not be assumed to be full memory barrier
1229 because it is possible for an access preceding the LOCK to happen after the
1230 LOCK, and an access following the UNLOCK to happen before the UNLOCK, and the
1231 two accesses can themselves then cross:
1240 LOCK, STORE *B, STORE *A, UNLOCK
1242 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1243 systems, and so cannot be counted on in such a situation to actually achieve
1244 anything at all - especially with respect to I/O accesses - unless combined
1245 with interrupt disabling operations.
1247 See also the section on "Inter-CPU locking barrier effects".
1250 As an example, consider the following:
1261 The following sequence of events is acceptable:
1263 LOCK, {*F,*A}, *E, {*C,*D}, *B, UNLOCK
1265 [+] Note that {*F,*A} indicates a combined access.
1267 But none of the following are:
1269 {*F,*A}, *B, LOCK, *C, *D, UNLOCK, *E
1270 *A, *B, *C, LOCK, *D, UNLOCK, *E, *F
1271 *A, *B, LOCK, *C, UNLOCK, *D, *E, *F
1272 *B, LOCK, *C, *D, UNLOCK, {*F,*A}, *E
1276 INTERRUPT DISABLING FUNCTIONS
1277 -----------------------------
1279 Functions that disable interrupts (LOCK equivalent) and enable interrupts
1280 (UNLOCK equivalent) will act as compiler barriers only. So if memory or I/O
1281 barriers are required in such a situation, they must be provided from some
1285 SLEEP AND WAKE-UP FUNCTIONS
1286 ---------------------------
1288 Sleeping and waking on an event flagged in global data can be viewed as an
1289 interaction between two pieces of data: the task state of the task waiting for
1290 the event and the global data used to indicate the event. To make sure that
1291 these appear to happen in the right order, the primitives to begin the process
1292 of going to sleep, and the primitives to initiate a wake up imply certain
1295 Firstly, the sleeper normally follows something like this sequence of events:
1298 set_current_state(TASK_UNINTERRUPTIBLE);
1299 if (event_indicated)
1304 A general memory barrier is interpolated automatically by set_current_state()
1305 after it has altered the task state:
1308 ===============================
1309 set_current_state();
1311 STORE current->state
1313 LOAD event_indicated
1315 set_current_state() may be wrapped by:
1318 prepare_to_wait_exclusive();
1320 which therefore also imply a general memory barrier after setting the state.
1321 The whole sequence above is available in various canned forms, all of which
1322 interpolate the memory barrier in the right place:
1325 wait_event_interruptible();
1326 wait_event_interruptible_exclusive();
1327 wait_event_interruptible_timeout();
1328 wait_event_killable();
1329 wait_event_timeout();
1334 Secondly, code that performs a wake up normally follows something like this:
1336 event_indicated = 1;
1337 wake_up(&event_wait_queue);
1341 event_indicated = 1;
1342 wake_up_process(event_daemon);
1344 A write memory barrier is implied by wake_up() and co. if and only if they wake
1345 something up. The barrier occurs before the task state is cleared, and so sits
1346 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1349 =============================== ===============================
1350 set_current_state(); STORE event_indicated
1351 set_mb(); wake_up();
1352 STORE current->state <write barrier>
1353 <general barrier> STORE current->state
1354 LOAD event_indicated
1356 The available waker functions include:
1362 wake_up_interruptible();
1363 wake_up_interruptible_all();
1364 wake_up_interruptible_nr();
1365 wake_up_interruptible_poll();
1366 wake_up_interruptible_sync();
1367 wake_up_interruptible_sync_poll();
1369 wake_up_locked_poll();
1375 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1376 order multiple stores before the wake-up with respect to loads of those stored
1377 values after the sleeper has called set_current_state(). For instance, if the
1380 set_current_state(TASK_INTERRUPTIBLE);
1381 if (event_indicated)
1383 __set_current_state(TASK_RUNNING);
1384 do_something(my_data);
1389 event_indicated = 1;
1390 wake_up(&event_wait_queue);
1392 there's no guarantee that the change to event_indicated will be perceived by
1393 the sleeper as coming after the change to my_data. In such a circumstance, the
1394 code on both sides must interpolate its own memory barriers between the
1395 separate data accesses. Thus the above sleeper ought to do:
1397 set_current_state(TASK_INTERRUPTIBLE);
1398 if (event_indicated) {
1400 do_something(my_data);
1403 and the waker should do:
1407 event_indicated = 1;
1408 wake_up(&event_wait_queue);
1411 MISCELLANEOUS FUNCTIONS
1412 -----------------------
1414 Other functions that imply barriers:
1416 (*) schedule() and similar imply full memory barriers.
1419 =================================
1420 INTER-CPU LOCKING BARRIER EFFECTS
1421 =================================
1423 On SMP systems locking primitives give a more substantial form of barrier: one
1424 that does affect memory access ordering on other CPUs, within the context of
1425 conflict on any particular lock.
1428 LOCKS VS MEMORY ACCESSES
1429 ------------------------
1431 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1432 three CPUs; then should the following sequence of events occur:
1435 =============================== ===============================
1443 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
1444 through *H occur in, other than the constraints imposed by the separate locks
1445 on the separate CPUs. It might, for example, see:
1447 *E, LOCK M, LOCK Q, *G, *C, *F, *A, *B, UNLOCK Q, *D, *H, UNLOCK M
1449 But it won't see any of:
1451 *B, *C or *D preceding LOCK M
1452 *A, *B or *C following UNLOCK M
1453 *F, *G or *H preceding LOCK Q
1454 *E, *F or *G following UNLOCK Q
1457 However, if the following occurs:
1460 =============================== ===============================
1475 *E, LOCK M [1], *C, *B, *A, UNLOCK M [1],
1476 LOCK M [2], *H, *F, *G, UNLOCK M [2], *D
1478 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
1480 *B, *C, *D, *F, *G or *H preceding LOCK M [1]
1481 *A, *B or *C following UNLOCK M [1]
1482 *F, *G or *H preceding LOCK M [2]
1483 *A, *B, *C, *E, *F or *G following UNLOCK M [2]
1486 LOCKS VS I/O ACCESSES
1487 ---------------------
1489 Under certain circumstances (especially involving NUMA), I/O accesses within
1490 two spinlocked sections on two different CPUs may be seen as interleaved by the
1491 PCI bridge, because the PCI bridge does not necessarily participate in the
1492 cache-coherence protocol, and is therefore incapable of issuing the required
1493 read memory barriers.
1498 =============================== ===============================
1508 may be seen by the PCI bridge as follows:
1510 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
1512 which would probably cause the hardware to malfunction.
1515 What is necessary here is to intervene with an mmiowb() before dropping the
1516 spinlock, for example:
1519 =============================== ===============================
1531 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
1532 before either of the stores issued on CPU 2.
1535 Furthermore, following a store by a load from the same device obviates the need
1536 for the mmiowb(), because the load forces the store to complete before the load
1540 =============================== ===============================
1551 See Documentation/DocBook/deviceiobook.tmpl for more information.
1554 =================================
1555 WHERE ARE MEMORY BARRIERS NEEDED?
1556 =================================
1558 Under normal operation, memory operation reordering is generally not going to
1559 be a problem as a single-threaded linear piece of code will still appear to
1560 work correctly, even if it's in an SMP kernel. There are, however, four
1561 circumstances in which reordering definitely _could_ be a problem:
1563 (*) Interprocessor interaction.
1565 (*) Atomic operations.
1567 (*) Accessing devices.
1572 INTERPROCESSOR INTERACTION
1573 --------------------------
1575 When there's a system with more than one processor, more than one CPU in the
1576 system may be working on the same data set at the same time. This can cause
1577 synchronisation problems, and the usual way of dealing with them is to use
1578 locks. Locks, however, are quite expensive, and so it may be preferable to
1579 operate without the use of a lock if at all possible. In such a case
1580 operations that affect both CPUs may have to be carefully ordered to prevent
1583 Consider, for example, the R/W semaphore slow path. Here a waiting process is
1584 queued on the semaphore, by virtue of it having a piece of its stack linked to
1585 the semaphore's list of waiting processes:
1587 struct rw_semaphore {
1590 struct list_head waiters;
1593 struct rwsem_waiter {
1594 struct list_head list;
1595 struct task_struct *task;
1598 To wake up a particular waiter, the up_read() or up_write() functions have to:
1600 (1) read the next pointer from this waiter's record to know as to where the
1601 next waiter record is;
1603 (2) read the pointer to the waiter's task structure;
1605 (3) clear the task pointer to tell the waiter it has been given the semaphore;
1607 (4) call wake_up_process() on the task; and
1609 (5) release the reference held on the waiter's task struct.
1611 In other words, it has to perform this sequence of events:
1613 LOAD waiter->list.next;
1619 and if any of these steps occur out of order, then the whole thing may
1622 Once it has queued itself and dropped the semaphore lock, the waiter does not
1623 get the lock again; it instead just waits for its task pointer to be cleared
1624 before proceeding. Since the record is on the waiter's stack, this means that
1625 if the task pointer is cleared _before_ the next pointer in the list is read,
1626 another CPU might start processing the waiter and might clobber the waiter's
1627 stack before the up*() function has a chance to read the next pointer.
1629 Consider then what might happen to the above sequence of events:
1632 =============================== ===============================
1639 Woken up by other event
1644 foo() clobbers *waiter
1646 LOAD waiter->list.next;
1649 This could be dealt with using the semaphore lock, but then the down_xxx()
1650 function has to needlessly get the spinlock again after being woken up.
1652 The way to deal with this is to insert a general SMP memory barrier:
1654 LOAD waiter->list.next;
1661 In this case, the barrier makes a guarantee that all memory accesses before the
1662 barrier will appear to happen before all the memory accesses after the barrier
1663 with respect to the other CPUs on the system. It does _not_ guarantee that all
1664 the memory accesses before the barrier will be complete by the time the barrier
1665 instruction itself is complete.
1667 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
1668 compiler barrier, thus making sure the compiler emits the instructions in the
1669 right order without actually intervening in the CPU. Since there's only one
1670 CPU, that CPU's dependency ordering logic will take care of everything else.
1676 Whilst they are technically interprocessor interaction considerations, atomic
1677 operations are noted specially as some of them imply full memory barriers and
1678 some don't, but they're very heavily relied on as a group throughout the
1681 Any atomic operation that modifies some state in memory and returns information
1682 about the state (old or new) implies an SMP-conditional general memory barrier
1683 (smp_mb()) on each side of the actual operation (with the exception of
1684 explicit lock operations, described later). These include:
1689 atomic_inc_return();
1690 atomic_dec_return();
1691 atomic_add_return();
1692 atomic_sub_return();
1693 atomic_inc_and_test();
1694 atomic_dec_and_test();
1695 atomic_sub_and_test();
1696 atomic_add_negative();
1697 atomic_add_unless(); /* when succeeds (returns 1) */
1699 test_and_clear_bit();
1700 test_and_change_bit();
1702 These are used for such things as implementing LOCK-class and UNLOCK-class
1703 operations and adjusting reference counters towards object destruction, and as
1704 such the implicit memory barrier effects are necessary.
1707 The following operations are potential problems as they do _not_ imply memory
1708 barriers, but might be used for implementing such things as UNLOCK-class
1716 With these the appropriate explicit memory barrier should be used if necessary
1717 (smp_mb__before_clear_bit() for instance).
1720 The following also do _not_ imply memory barriers, and so may require explicit
1721 memory barriers under some circumstances (smp_mb__before_atomic_dec() for
1729 If they're used for statistics generation, then they probably don't need memory
1730 barriers, unless there's a coupling between statistical data.
1732 If they're used for reference counting on an object to control its lifetime,
1733 they probably don't need memory barriers because either the reference count
1734 will be adjusted inside a locked section, or the caller will already hold
1735 sufficient references to make the lock, and thus a memory barrier unnecessary.
1737 If they're used for constructing a lock of some description, then they probably
1738 do need memory barriers as a lock primitive generally has to do things in a
1741 Basically, each usage case has to be carefully considered as to whether memory
1742 barriers are needed or not.
1744 The following operations are special locking primitives:
1746 test_and_set_bit_lock();
1748 __clear_bit_unlock();
1750 These implement LOCK-class and UNLOCK-class operations. These should be used in
1751 preference to other operations when implementing locking primitives, because
1752 their implementations can be optimised on many architectures.
1754 [!] Note that special memory barrier primitives are available for these
1755 situations because on some CPUs the atomic instructions used imply full memory
1756 barriers, and so barrier instructions are superfluous in conjunction with them,
1757 and in such cases the special barrier primitives will be no-ops.
1759 See Documentation/atomic_ops.txt for more information.
1765 Many devices can be memory mapped, and so appear to the CPU as if they're just
1766 a set of memory locations. To control such a device, the driver usually has to
1767 make the right memory accesses in exactly the right order.
1769 However, having a clever CPU or a clever compiler creates a potential problem
1770 in that the carefully sequenced accesses in the driver code won't reach the
1771 device in the requisite order if the CPU or the compiler thinks it is more
1772 efficient to reorder, combine or merge accesses - something that would cause
1773 the device to malfunction.
1775 Inside of the Linux kernel, I/O should be done through the appropriate accessor
1776 routines - such as inb() or writel() - which know how to make such accesses
1777 appropriately sequential. Whilst this, for the most part, renders the explicit
1778 use of memory barriers unnecessary, there are a couple of situations where they
1781 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
1782 so for _all_ general drivers locks should be used and mmiowb() must be
1783 issued prior to unlocking the critical section.
1785 (2) If the accessor functions are used to refer to an I/O memory window with
1786 relaxed memory access properties, then _mandatory_ memory barriers are
1787 required to enforce ordering.
1789 See Documentation/DocBook/deviceiobook.tmpl for more information.
1795 A driver may be interrupted by its own interrupt service routine, and thus the
1796 two parts of the driver may interfere with each other's attempts to control or
1799 This may be alleviated - at least in part - by disabling local interrupts (a
1800 form of locking), such that the critical operations are all contained within
1801 the interrupt-disabled section in the driver. Whilst the driver's interrupt
1802 routine is executing, the driver's core may not run on the same CPU, and its
1803 interrupt is not permitted to happen again until the current interrupt has been
1804 handled, thus the interrupt handler does not need to lock against that.
1806 However, consider a driver that was talking to an ethernet card that sports an
1807 address register and a data register. If that driver's core talks to the card
1808 under interrupt-disablement and then the driver's interrupt handler is invoked:
1819 The store to the data register might happen after the second store to the
1820 address register if ordering rules are sufficiently relaxed:
1822 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
1825 If ordering rules are relaxed, it must be assumed that accesses done inside an
1826 interrupt disabled section may leak outside of it and may interleave with
1827 accesses performed in an interrupt - and vice versa - unless implicit or
1828 explicit barriers are used.
1830 Normally this won't be a problem because the I/O accesses done inside such
1831 sections will include synchronous load operations on strictly ordered I/O
1832 registers that form implicit I/O barriers. If this isn't sufficient then an
1833 mmiowb() may need to be used explicitly.
1836 A similar situation may occur between an interrupt routine and two routines
1837 running on separate CPUs that communicate with each other. If such a case is
1838 likely, then interrupt-disabling locks should be used to guarantee ordering.
1841 ==========================
1842 KERNEL I/O BARRIER EFFECTS
1843 ==========================
1845 When accessing I/O memory, drivers should use the appropriate accessor
1850 These are intended to talk to I/O space rather than memory space, but
1851 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
1852 indeed have special I/O space access cycles and instructions, but many
1853 CPUs don't have such a concept.
1855 The PCI bus, amongst others, defines an I/O space concept which - on such
1856 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
1857 space. However, it may also be mapped as a virtual I/O space in the CPU's
1858 memory map, particularly on those CPUs that don't support alternate I/O
1861 Accesses to this space may be fully synchronous (as on i386), but
1862 intermediary bridges (such as the PCI host bridge) may not fully honour
1865 They are guaranteed to be fully ordered with respect to each other.
1867 They are not guaranteed to be fully ordered with respect to other types of
1868 memory and I/O operation.
1870 (*) readX(), writeX():
1872 Whether these are guaranteed to be fully ordered and uncombined with
1873 respect to each other on the issuing CPU depends on the characteristics
1874 defined for the memory window through which they're accessing. On later
1875 i386 architecture machines, for example, this is controlled by way of the
1878 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
1879 provided they're not accessing a prefetchable device.
1881 However, intermediary hardware (such as a PCI bridge) may indulge in
1882 deferral if it so wishes; to flush a store, a load from the same location
1883 is preferred[*], but a load from the same device or from configuration
1884 space should suffice for PCI.
1886 [*] NOTE! attempting to load from the same location as was written to may
1887 cause a malfunction - consider the 16550 Rx/Tx serial registers for
1890 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
1891 force stores to be ordered.
1893 Please refer to the PCI specification for more information on interactions
1894 between PCI transactions.
1898 These are similar to readX(), but are not guaranteed to be ordered in any
1899 way. Be aware that there is no I/O read barrier available.
1901 (*) ioreadX(), iowriteX()
1903 These will perform appropriately for the type of access they're actually
1904 doing, be it inX()/outX() or readX()/writeX().
1907 ========================================
1908 ASSUMED MINIMUM EXECUTION ORDERING MODEL
1909 ========================================
1911 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
1912 maintain the appearance of program causality with respect to itself. Some CPUs
1913 (such as i386 or x86_64) are more constrained than others (such as powerpc or
1914 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
1915 of arch-specific code.
1917 This means that it must be considered that the CPU will execute its instruction
1918 stream in any order it feels like - or even in parallel - provided that if an
1919 instruction in the stream depends on an earlier instruction, then that
1920 earlier instruction must be sufficiently complete[*] before the later
1921 instruction may proceed; in other words: provided that the appearance of
1922 causality is maintained.
1924 [*] Some instructions have more than one effect - such as changing the
1925 condition codes, changing registers or changing memory - and different
1926 instructions may depend on different effects.
1928 A CPU may also discard any instruction sequence that winds up having no
1929 ultimate effect. For example, if two adjacent instructions both load an
1930 immediate value into the same register, the first may be discarded.
1933 Similarly, it has to be assumed that compiler might reorder the instruction
1934 stream in any way it sees fit, again provided the appearance of causality is
1938 ============================
1939 THE EFFECTS OF THE CPU CACHE
1940 ============================
1942 The way cached memory operations are perceived across the system is affected to
1943 a certain extent by the caches that lie between CPUs and memory, and by the
1944 memory coherence system that maintains the consistency of state in the system.
1946 As far as the way a CPU interacts with another part of the system through the
1947 caches goes, the memory system has to include the CPU's caches, and memory
1948 barriers for the most part act at the interface between the CPU and its cache
1949 (memory barriers logically act on the dotted line in the following diagram):
1951 <--- CPU ---> : <----------- Memory ----------->
1953 +--------+ +--------+ : +--------+ +-----------+
1954 | | | | : | | | | +--------+
1955 | CPU | | Memory | : | CPU | | | | |
1956 | Core |--->| Access |----->| Cache |<-->| | | |
1957 | | | Queue | : | | | |--->| Memory |
1958 | | | | : | | | | | |
1959 +--------+ +--------+ : +--------+ | | | |
1960 : | Cache | +--------+
1962 : | Mechanism | +--------+
1963 +--------+ +--------+ : +--------+ | | | |
1964 | | | | : | | | | | |
1965 | CPU | | Memory | : | CPU | | |--->| Device |
1966 | Core |--->| Access |----->| Cache |<-->| | | |
1967 | | | Queue | : | | | | | |
1968 | | | | : | | | | +--------+
1969 +--------+ +--------+ : +--------+ +-----------+
1973 Although any particular load or store may not actually appear outside of the
1974 CPU that issued it since it may have been satisfied within the CPU's own cache,
1975 it will still appear as if the full memory access had taken place as far as the
1976 other CPUs are concerned since the cache coherency mechanisms will migrate the
1977 cacheline over to the accessing CPU and propagate the effects upon conflict.
1979 The CPU core may execute instructions in any order it deems fit, provided the
1980 expected program causality appears to be maintained. Some of the instructions
1981 generate load and store operations which then go into the queue of memory
1982 accesses to be performed. The core may place these in the queue in any order
1983 it wishes, and continue execution until it is forced to wait for an instruction
1986 What memory barriers are concerned with is controlling the order in which
1987 accesses cross from the CPU side of things to the memory side of things, and
1988 the order in which the effects are perceived to happen by the other observers
1991 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
1992 their own loads and stores as if they had happened in program order.
1994 [!] MMIO or other device accesses may bypass the cache system. This depends on
1995 the properties of the memory window through which devices are accessed and/or
1996 the use of any special device communication instructions the CPU may have.
2002 Life isn't quite as simple as it may appear above, however: for while the
2003 caches are expected to be coherent, there's no guarantee that that coherency
2004 will be ordered. This means that whilst changes made on one CPU will
2005 eventually become visible on all CPUs, there's no guarantee that they will
2006 become apparent in the same order on those other CPUs.
2009 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
2010 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
2015 +--------+ : +--->| Cache A |<------->| |
2016 | | : | +---------+ | |
2018 | | : | +---------+ | |
2019 +--------+ : +--->| Cache B |<------->| |
2022 : +---------+ | System |
2023 +--------+ : +--->| Cache C |<------->| |
2024 | | : | +---------+ | |
2026 | | : | +---------+ | |
2027 +--------+ : +--->| Cache D |<------->| |
2032 Imagine the system has the following properties:
2034 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
2037 (*) an even-numbered cache line may be in cache B, cache D or it may still be
2040 (*) whilst the CPU core is interrogating one cache, the other cache may be
2041 making use of the bus to access the rest of the system - perhaps to
2042 displace a dirty cacheline or to do a speculative load;
2044 (*) each cache has a queue of operations that need to be applied to that cache
2045 to maintain coherency with the rest of the system;
2047 (*) the coherency queue is not flushed by normal loads to lines already
2048 present in the cache, even though the contents of the queue may
2049 potentially affect those loads.
2051 Imagine, then, that two writes are made on the first CPU, with a write barrier
2052 between them to guarantee that they will appear to reach that CPU's caches in
2053 the requisite order:
2056 =============== =============== =======================================
2057 u == 0, v == 1 and p == &u, q == &u
2059 smp_wmb(); Make sure change to v is visible before
2061 <A:modify v=2> v is now in cache A exclusively
2063 <B:modify p=&v> p is now in cache B exclusively
2065 The write memory barrier forces the other CPUs in the system to perceive that
2066 the local CPU's caches have apparently been updated in the correct order. But
2067 now imagine that the second CPU wants to read those values:
2070 =============== =============== =======================================
2075 The above pair of reads may then fail to happen in the expected order, as the
2076 cacheline holding p may get updated in one of the second CPU's caches whilst
2077 the update to the cacheline holding v is delayed in the other of the second
2078 CPU's caches by some other cache event:
2081 =============== =============== =======================================
2082 u == 0, v == 1 and p == &u, q == &u
2085 <A:modify v=2> <C:busy>
2089 <B:modify p=&v> <D:commit p=&v>
2092 <C:read *q> Reads from v before v updated in cache
2096 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2097 no guarantee that, without intervention, the order of update will be the same
2098 as that committed on CPU 1.
2101 To intervene, we need to interpolate a data dependency barrier or a read
2102 barrier between the loads. This will force the cache to commit its coherency
2103 queue before processing any further requests:
2106 =============== =============== =======================================
2107 u == 0, v == 1 and p == &u, q == &u
2110 <A:modify v=2> <C:busy>
2114 <B:modify p=&v> <D:commit p=&v>
2116 smp_read_barrier_depends()
2120 <C:read *q> Reads from v after v updated in cache
2123 This sort of problem can be encountered on DEC Alpha processors as they have a
2124 split cache that improves performance by making better use of the data bus.
2125 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2126 access depends on a read, not all do, so it may not be relied on.
2128 Other CPUs may also have split caches, but must coordinate between the various
2129 cachelets for normal memory accesses. The semantics of the Alpha removes the
2130 need for coordination in the absence of memory barriers.
2133 CACHE COHERENCY VS DMA
2134 ----------------------
2136 Not all systems maintain cache coherency with respect to devices doing DMA. In
2137 such cases, a device attempting DMA may obtain stale data from RAM because
2138 dirty cache lines may be resident in the caches of various CPUs, and may not
2139 have been written back to RAM yet. To deal with this, the appropriate part of
2140 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2141 invalidate them as well).
2143 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2144 cache lines being written back to RAM from a CPU's cache after the device has
2145 installed its own data, or cache lines present in the CPU's cache may simply
2146 obscure the fact that RAM has been updated, until at such time as the cacheline
2147 is discarded from the CPU's cache and reloaded. To deal with this, the
2148 appropriate part of the kernel must invalidate the overlapping bits of the
2151 See Documentation/cachetlb.txt for more information on cache management.
2154 CACHE COHERENCY VS MMIO
2155 -----------------------
2157 Memory mapped I/O usually takes place through memory locations that are part of
2158 a window in the CPU's memory space that has different properties assigned than
2159 the usual RAM directed window.
2161 Amongst these properties is usually the fact that such accesses bypass the
2162 caching entirely and go directly to the device buses. This means MMIO accesses
2163 may, in effect, overtake accesses to cached memory that were emitted earlier.
2164 A memory barrier isn't sufficient in such a case, but rather the cache must be
2165 flushed between the cached memory write and the MMIO access if the two are in
2169 =========================
2170 THE THINGS CPUS GET UP TO
2171 =========================
2173 A programmer might take it for granted that the CPU will perform memory
2174 operations in exactly the order specified, so that if the CPU is, for example,
2175 given the following piece of code to execute:
2183 they would then expect that the CPU will complete the memory operation for each
2184 instruction before moving on to the next one, leading to a definite sequence of
2185 operations as seen by external observers in the system:
2187 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2190 Reality is, of course, much messier. With many CPUs and compilers, the above
2191 assumption doesn't hold because:
2193 (*) loads are more likely to need to be completed immediately to permit
2194 execution progress, whereas stores can often be deferred without a
2197 (*) loads may be done speculatively, and the result discarded should it prove
2198 to have been unnecessary;
2200 (*) loads may be done speculatively, leading to the result having been fetched
2201 at the wrong time in the expected sequence of events;
2203 (*) the order of the memory accesses may be rearranged to promote better use
2204 of the CPU buses and caches;
2206 (*) loads and stores may be combined to improve performance when talking to
2207 memory or I/O hardware that can do batched accesses of adjacent locations,
2208 thus cutting down on transaction setup costs (memory and PCI devices may
2209 both be able to do this); and
2211 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2212 mechanisms may alleviate this - once the store has actually hit the cache
2213 - there's no guarantee that the coherency management will be propagated in
2214 order to other CPUs.
2216 So what another CPU, say, might actually observe from the above piece of code
2219 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2221 (Where "LOAD {*C,*D}" is a combined load)
2224 However, it is guaranteed that a CPU will be self-consistent: it will see its
2225 _own_ accesses appear to be correctly ordered, without the need for a memory
2226 barrier. For instance with the following code:
2235 and assuming no intervention by an external influence, it can be assumed that
2236 the final result will appear to be:
2238 U == the original value of *A
2243 The code above may cause the CPU to generate the full sequence of memory
2246 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2248 in that order, but, without intervention, the sequence may have almost any
2249 combination of elements combined or discarded, provided the program's view of
2250 the world remains consistent.
2252 The compiler may also combine, discard or defer elements of the sequence before
2253 the CPU even sees them.
2264 since, without a write barrier, it can be assumed that the effect of the
2265 storage of V to *A is lost. Similarly:
2270 may, without a memory barrier, be reduced to:
2275 and the LOAD operation never appear outside of the CPU.
2278 AND THEN THERE'S THE ALPHA
2279 --------------------------
2281 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2282 some versions of the Alpha CPU have a split data cache, permitting them to have
2283 two semantically-related cache lines updated at separate times. This is where
2284 the data dependency barrier really becomes necessary as this synchronises both
2285 caches with the memory coherence system, thus making it seem like pointer
2286 changes vs new data occur in the right order.
2288 The Alpha defines the Linux kernel's memory barrier model.
2290 See the subsection on "Cache Coherency" above.
2300 Memory barriers can be used to implement circular buffering without the need
2301 of a lock to serialise the producer with the consumer. See:
2303 Documentation/circular-buffers.txt
2312 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2314 Chapter 5.2: Physical Address Space Characteristics
2315 Chapter 5.4: Caches and Write Buffers
2316 Chapter 5.5: Data Sharing
2317 Chapter 5.6: Read/Write Ordering
2319 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2320 Chapter 7.1: Memory-Access Ordering
2321 Chapter 7.4: Buffering and Combining Memory Writes
2323 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2324 System Programming Guide
2325 Chapter 7.1: Locked Atomic Operations
2326 Chapter 7.2: Memory Ordering
2327 Chapter 7.4: Serializing Instructions
2329 The SPARC Architecture Manual, Version 9
2330 Chapter 8: Memory Models
2331 Appendix D: Formal Specification of the Memory Models
2332 Appendix J: Programming with the Memory Models
2334 UltraSPARC Programmer Reference Manual
2335 Chapter 5: Memory Accesses and Cacheability
2336 Chapter 15: Sparc-V9 Memory Models
2338 UltraSPARC III Cu User's Manual
2339 Chapter 9: Memory Models
2341 UltraSPARC IIIi Processor User's Manual
2342 Chapter 8: Memory Models
2344 UltraSPARC Architecture 2005
2346 Appendix D: Formal Specifications of the Memory Models
2348 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2349 Chapter 8: Memory Models
2350 Appendix F: Caches and Cache Coherency
2352 Solaris Internals, Core Kernel Architecture, p63-68:
2353 Chapter 3.3: Hardware Considerations for Locks and
2356 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2357 for Kernel Programmers:
2358 Chapter 13: Other Memory Models
2360 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2361 Section 2.6: Speculation
2362 Section 4.4: Memory Access
projects / platform / kernel / linux-rpi.git / blob