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.
25 (*) Explicit kernel barriers.
28 - CPU memory barriers.
31 (*) Implicit kernel memory barriers.
34 - Interrupt disabling functions.
35 - Sleep and wake-up functions.
36 - Miscellaneous functions.
38 (*) Inter-CPU locking barrier effects.
40 - Locks vs memory accesses.
41 - Locks vs I/O accesses.
43 (*) Where are memory barriers needed?
45 - Interprocessor interaction.
50 (*) Kernel I/O barrier effects.
52 (*) Assumed minimum execution ordering model.
54 (*) The effects of the cpu cache.
57 - Cache coherency vs DMA.
58 - Cache coherency vs MMIO.
60 (*) The things CPUs get up to.
62 - And then there's the Alpha.
71 ============================
72 ABSTRACT MEMORY ACCESS MODEL
73 ============================
75 Consider the following abstract model of the system:
80 +-------+ : +--------+ : +-------+
83 | CPU 1 |<----->| Memory |<----->| CPU 2 |
86 +-------+ : +--------+ : +-------+
94 +---------->| Device |<----------+
100 Each CPU executes a program that generates memory access operations. In the
101 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
102 perform the memory operations in any order it likes, provided program causality
103 appears to be maintained. Similarly, the compiler may also arrange the
104 instructions it emits in any order it likes, provided it doesn't affect the
105 apparent operation of the program.
107 So in the above diagram, the effects of the memory operations performed by a
108 CPU are perceived by the rest of the system as the operations cross the
109 interface between the CPU and rest of the system (the dotted lines).
112 For example, consider the following sequence of events:
115 =============== ===============
120 The set of accesses as seen by the memory system in the middle can be arranged
121 in 24 different combinations:
123 STORE A=3, STORE B=4, x=LOAD A->3, y=LOAD B->4
124 STORE A=3, STORE B=4, y=LOAD B->4, x=LOAD A->3
125 STORE A=3, x=LOAD A->3, STORE B=4, y=LOAD B->4
126 STORE A=3, x=LOAD A->3, y=LOAD B->2, STORE B=4
127 STORE A=3, y=LOAD B->2, STORE B=4, x=LOAD A->3
128 STORE A=3, y=LOAD B->2, x=LOAD A->3, STORE B=4
129 STORE B=4, STORE A=3, x=LOAD A->3, y=LOAD B->4
133 and can thus result in four different combinations of values:
141 Furthermore, the stores committed by a CPU to the memory system may not be
142 perceived by the loads made by another CPU in the same order as the stores were
146 As a further example, consider this sequence of events:
149 =============== ===============
150 { A == 1, B == 2, C = 3, P == &A, Q == &C }
154 There is an obvious data dependency here, as the value loaded into D depends on
155 the address retrieved from P by CPU 2. At the end of the sequence, any of the
156 following results are possible:
158 (Q == &A) and (D == 1)
159 (Q == &B) and (D == 2)
160 (Q == &B) and (D == 4)
162 Note that CPU 2 will never try and load C into D because the CPU will load P
163 into Q before issuing the load of *Q.
169 Some devices present their control interfaces as collections of memory
170 locations, but the order in which the control registers are accessed is very
171 important. For instance, imagine an ethernet card with a set of internal
172 registers that are accessed through an address port register (A) and a data
173 port register (D). To read internal register 5, the following code might then
179 but this might show up as either of the following two sequences:
181 STORE *A = 5, x = LOAD *D
182 x = LOAD *D, STORE *A = 5
184 the second of which will almost certainly result in a malfunction, since it set
185 the address _after_ attempting to read the register.
191 There are some minimal guarantees that may be expected of a CPU:
193 (*) On any given CPU, dependent memory accesses will be issued in order, with
194 respect to itself. This means that for:
198 the CPU will issue the following memory operations:
200 Q = LOAD P, D = LOAD *Q
202 and always in that order.
204 (*) Overlapping loads and stores within a particular CPU will appear to be
205 ordered within that CPU. This means that for:
209 the CPU will only issue the following sequence of memory operations:
211 a = LOAD *X, STORE *X = b
217 the CPU will only issue:
219 STORE *X = c, d = LOAD *X
221 (Loads and stores overlap if they are targeted at overlapping pieces of
224 And there are a number of things that _must_ or _must_not_ be assumed:
226 (*) It _must_not_ be assumed that independent loads and stores will be issued
227 in the order given. This means that for:
229 X = *A; Y = *B; *D = Z;
231 we may get any of the following sequences:
233 X = LOAD *A, Y = LOAD *B, STORE *D = Z
234 X = LOAD *A, STORE *D = Z, Y = LOAD *B
235 Y = LOAD *B, X = LOAD *A, STORE *D = Z
236 Y = LOAD *B, STORE *D = Z, X = LOAD *A
237 STORE *D = Z, X = LOAD *A, Y = LOAD *B
238 STORE *D = Z, Y = LOAD *B, X = LOAD *A
240 (*) It _must_ be assumed that overlapping memory accesses may be merged or
241 discarded. This means that for:
243 X = *A; Y = *(A + 4);
245 we may get any one of the following sequences:
247 X = LOAD *A; Y = LOAD *(A + 4);
248 Y = LOAD *(A + 4); X = LOAD *A;
249 {X, Y} = LOAD {*A, *(A + 4) };
255 we may get either of:
257 STORE *A = X; Y = LOAD *A;
261 =========================
262 WHAT ARE MEMORY BARRIERS?
263 =========================
265 As can be seen above, independent memory operations are effectively performed
266 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
267 What is required is some way of intervening to instruct the compiler and the
268 CPU to restrict the order.
270 Memory barriers are such interventions. They impose a perceived partial
271 ordering over the memory operations on either side of the barrier.
273 Such enforcement is important because the CPUs and other devices in a system
274 can use a variety of tricks to improve performance, including reordering,
275 deferral and combination of memory operations; speculative loads; speculative
276 branch prediction and various types of caching. Memory barriers are used to
277 override or suppress these tricks, allowing the code to sanely control the
278 interaction of multiple CPUs and/or devices.
281 VARIETIES OF MEMORY BARRIER
282 ---------------------------
284 Memory barriers come in four basic varieties:
286 (1) Write (or store) memory barriers.
288 A write memory barrier gives a guarantee that all the STORE operations
289 specified before the barrier will appear to happen before all the STORE
290 operations specified after the barrier with respect to the other
291 components of the system.
293 A write barrier is a partial ordering on stores only; it is not required
294 to have any effect on loads.
296 A CPU can be viewed as committing a sequence of store operations to the
297 memory system as time progresses. All stores before a write barrier will
298 occur in the sequence _before_ all the stores after the write barrier.
300 [!] Note that write barriers should normally be paired with read or data
301 dependency barriers; see the "SMP barrier pairing" subsection.
304 (2) Data dependency barriers.
306 A data dependency barrier is a weaker form of read barrier. In the case
307 where two loads are performed such that the second depends on the result
308 of the first (eg: the first load retrieves the address to which the second
309 load will be directed), a data dependency barrier would be required to
310 make sure that the target of the second load is updated before the address
311 obtained by the first load is accessed.
313 A data dependency barrier is a partial ordering on interdependent loads
314 only; it is not required to have any effect on stores, independent loads
315 or overlapping loads.
317 As mentioned in (1), the other CPUs in the system can be viewed as
318 committing sequences of stores to the memory system that the CPU being
319 considered can then perceive. A data dependency barrier issued by the CPU
320 under consideration guarantees that for any load preceding it, if that
321 load touches one of a sequence of stores from another CPU, then by the
322 time the barrier completes, the effects of all the stores prior to that
323 touched by the load will be perceptible to any loads issued after the data
326 See the "Examples of memory barrier sequences" subsection for diagrams
327 showing the ordering constraints.
329 [!] Note that the first load really has to have a _data_ dependency and
330 not a control dependency. If the address for the second load is dependent
331 on the first load, but the dependency is through a conditional rather than
332 actually loading the address itself, then it's a _control_ dependency and
333 a full read barrier or better is required. See the "Control dependencies"
334 subsection for more information.
336 [!] Note that data dependency barriers should normally be paired with
337 write barriers; see the "SMP barrier pairing" subsection.
340 (3) Read (or load) memory barriers.
342 A read barrier is a data dependency barrier plus a guarantee that all the
343 LOAD operations specified before the barrier will appear to happen before
344 all the LOAD operations specified after the barrier with respect to the
345 other components of the system.
347 A read barrier is a partial ordering on loads only; it is not required to
348 have any effect on stores.
350 Read memory barriers imply data dependency barriers, and so can substitute
353 [!] Note that read barriers should normally be paired with write barriers;
354 see the "SMP barrier pairing" subsection.
357 (4) General memory barriers.
359 A general memory barrier gives a guarantee that all the LOAD and STORE
360 operations specified before the barrier will appear to happen before all
361 the LOAD and STORE operations specified after the barrier with respect to
362 the other components of the system.
364 A general memory barrier is a partial ordering over both loads and stores.
366 General memory barriers imply both read and write memory barriers, and so
367 can substitute for either.
370 And a couple of implicit varieties:
374 This acts as a one-way permeable barrier. It guarantees that all memory
375 operations after the LOCK operation will appear to happen after the LOCK
376 operation with respect to the other components of the system.
378 Memory operations that occur before a LOCK operation may appear to happen
381 A LOCK operation should almost always be paired with an UNLOCK operation.
384 (6) UNLOCK operations.
386 This also acts as a one-way permeable barrier. It guarantees that all
387 memory operations before the UNLOCK operation will appear to happen before
388 the UNLOCK operation with respect to the other components of the system.
390 Memory operations that occur after an UNLOCK operation may appear to
391 happen before it completes.
393 LOCK and UNLOCK operations are guaranteed to appear with respect to each
394 other strictly in the order specified.
396 The use of LOCK and UNLOCK operations generally precludes the need for
397 other sorts of memory barrier (but note the exceptions mentioned in the
398 subsection "MMIO write barrier").
401 Memory barriers are only required where there's a possibility of interaction
402 between two CPUs or between a CPU and a device. If it can be guaranteed that
403 there won't be any such interaction in any particular piece of code, then
404 memory barriers are unnecessary in that piece of code.
407 Note that these are the _minimum_ guarantees. Different architectures may give
408 more substantial guarantees, but they may _not_ be relied upon outside of arch
412 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
413 ----------------------------------------------
415 There are certain things that the Linux kernel memory barriers do not guarantee:
417 (*) There is no guarantee that any of the memory accesses specified before a
418 memory barrier will be _complete_ by the completion of a memory barrier
419 instruction; the barrier can be considered to draw a line in that CPU's
420 access queue that accesses of the appropriate type may not cross.
422 (*) There is no guarantee that issuing a memory barrier on one CPU will have
423 any direct effect on another CPU or any other hardware in the system. The
424 indirect effect will be the order in which the second CPU sees the effects
425 of the first CPU's accesses occur, but see the next point:
427 (*) There is no guarantee that a CPU will see the correct order of effects
428 from a second CPU's accesses, even _if_ the second CPU uses a memory
429 barrier, unless the first CPU _also_ uses a matching memory barrier (see
430 the subsection on "SMP Barrier Pairing").
432 (*) There is no guarantee that some intervening piece of off-the-CPU
433 hardware[*] will not reorder the memory accesses. CPU cache coherency
434 mechanisms should propagate the indirect effects of a memory barrier
435 between CPUs, but might not do so in order.
437 [*] For information on bus mastering DMA and coherency please read:
439 Documentation/PCI/pci.txt
440 Documentation/PCI/PCI-DMA-mapping.txt
441 Documentation/DMA-API.txt
444 DATA DEPENDENCY BARRIERS
445 ------------------------
447 The usage requirements of data dependency barriers are a little subtle, and
448 it's not always obvious that they're needed. To illustrate, consider the
449 following sequence of events:
452 =============== ===============
453 { A == 1, B == 2, C = 3, P == &A, Q == &C }
460 There's a clear data dependency here, and it would seem that by the end of the
461 sequence, Q must be either &A or &B, and that:
463 (Q == &A) implies (D == 1)
464 (Q == &B) implies (D == 4)
466 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
467 leading to the following situation:
469 (Q == &B) and (D == 2) ????
471 Whilst this may seem like a failure of coherency or causality maintenance, it
472 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
475 To deal with this, a data dependency barrier or better must be inserted
476 between the address load and the data load:
479 =============== ===============
480 { A == 1, B == 2, C = 3, P == &A, Q == &C }
485 <data dependency barrier>
488 This enforces the occurrence of one of the two implications, and prevents the
489 third possibility from arising.
491 [!] Note that this extremely counterintuitive situation arises most easily on
492 machines with split caches, so that, for example, one cache bank processes
493 even-numbered cache lines and the other bank processes odd-numbered cache
494 lines. The pointer P might be stored in an odd-numbered cache line, and the
495 variable B might be stored in an even-numbered cache line. Then, if the
496 even-numbered bank of the reading CPU's cache is extremely busy while the
497 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
498 but the old value of the variable B (2).
501 Another example of where data dependency barriers might by required is where a
502 number is read from memory and then used to calculate the index for an array
506 =============== ===============
507 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
512 <data dependency barrier>
516 The data dependency barrier is very important to the RCU system, for example.
517 See rcu_dereference() in include/linux/rcupdate.h. This permits the current
518 target of an RCU'd pointer to be replaced with a new modified target, without
519 the replacement target appearing to be incompletely initialised.
521 See also the subsection on "Cache Coherency" for a more thorough example.
527 A control dependency requires a full read memory barrier, not simply a data
528 dependency barrier to make it work correctly. Consider the following bit of
534 <data dependency barrier>
537 This will not have the desired effect because there is no actual data
538 dependency, but rather a control dependency that the CPU may short-circuit by
539 attempting to predict the outcome in advance. In such a case what's actually
552 When dealing with CPU-CPU interactions, certain types of memory barrier should
553 always be paired. A lack of appropriate pairing is almost certainly an error.
555 A write barrier should always be paired with a data dependency barrier or read
556 barrier, though a general barrier would also be viable. Similarly a read
557 barrier or a data dependency barrier should always be paired with at least an
558 write barrier, though, again, a general barrier is viable:
561 =============== ===============
571 =============== ===============================
575 <data dependency barrier>
578 Basically, the read barrier always has to be there, even though it can be of
581 [!] Note that the stores before the write barrier would normally be expected to
582 match the loads after the read barrier or the data dependency barrier, and vice
586 =============== ===============
587 a = 1; }---- --->{ v = c
589 <write barrier> \ <read barrier>
590 c = 3; } / \ { x = a;
591 d = 4; }---- --->{ y = b;
594 EXAMPLES OF MEMORY BARRIER SEQUENCES
595 ------------------------------------
597 Firstly, write barriers act as partial orderings on store operations.
598 Consider the following sequence of events:
601 =======================
609 This sequence of events is committed to the memory coherence system in an order
610 that the rest of the system might perceive as the unordered set of { STORE A,
611 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
616 | |------>| C=3 | } /\
617 | | : +------+ }----- \ -----> Events perceptible to
618 | | : | A=1 | } \/ the rest of the system
620 | CPU 1 | : | B=2 | }
622 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
623 | | +------+ } requires all stores prior to the
624 | | : | E=5 | } barrier to be committed before
625 | | : +------+ } further stores may take place
630 | Sequence in which stores are committed to the
631 | memory system by CPU 1
635 Secondly, data dependency barriers act as partial orderings on data-dependent
636 loads. Consider the following sequence of events:
639 ======================= =======================
640 { B = 7; X = 9; Y = 8; C = &Y }
645 STORE D = 4 LOAD C (gets &B)
648 Without intervention, CPU 2 may perceive the events on CPU 1 in some
649 effectively random order, despite the write barrier issued by CPU 1:
652 | | +------+ +-------+ | Sequence of update
653 | |------>| B=2 |----- --->| Y->8 | | of perception on
654 | | : +------+ \ +-------+ | CPU 2
655 | CPU 1 | : | A=1 | \ --->| C->&Y | V
656 | | +------+ | +-------+
657 | | wwwwwwwwwwwwwwww | : :
659 | | : | C=&B |--- | : : +-------+
660 | | : +------+ \ | +-------+ | |
661 | |------>| D=4 | ----------->| C->&B |------>| |
662 | | +------+ | +-------+ | |
663 +-------+ : : | : : | |
667 Apparently incorrect ---> | | B->7 |------>| |
668 perception of B (!) | +-------+ | |
671 The load of X holds ---> \ | X->9 |------>| |
672 up the maintenance \ +-------+ | |
673 of coherence of B ----->| B->2 | +-------+
678 In the above example, CPU 2 perceives that B is 7, despite the load of *C
679 (which would be B) coming after the LOAD of C.
681 If, however, a data dependency barrier were to be placed between the load of C
682 and the load of *C (ie: B) on CPU 2:
685 ======================= =======================
686 { B = 7; X = 9; Y = 8; C = &Y }
691 STORE D = 4 LOAD C (gets &B)
692 <data dependency barrier>
695 then the following will occur:
698 | | +------+ +-------+
699 | |------>| B=2 |----- --->| Y->8 |
700 | | : +------+ \ +-------+
701 | CPU 1 | : | A=1 | \ --->| C->&Y |
702 | | +------+ | +-------+
703 | | wwwwwwwwwwwwwwww | : :
705 | | : | C=&B |--- | : : +-------+
706 | | : +------+ \ | +-------+ | |
707 | |------>| D=4 | ----------->| C->&B |------>| |
708 | | +------+ | +-------+ | |
709 +-------+ : : | : : | |
715 Makes sure all effects ---> \ ddddddddddddddddd | |
716 prior to the store of C \ +-------+ | |
717 are perceptible to ----->| B->2 |------>| |
718 subsequent loads +-------+ | |
722 And thirdly, a read barrier acts as a partial order on loads. Consider the
723 following sequence of events:
726 ======================= =======================
734 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
735 some effectively random order, despite the write barrier issued by CPU 1:
738 | | +------+ +-------+
739 | |------>| A=1 |------ --->| A->0 |
740 | | +------+ \ +-------+
741 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
742 | | +------+ | +-------+
743 | |------>| B=2 |--- | : :
744 | | +------+ \ | : : +-------+
745 +-------+ : : \ | +-------+ | |
746 ---------->| B->2 |------>| |
747 | +-------+ | CPU 2 |
758 If, however, a read barrier were to be placed between the load of B and the
762 ======================= =======================
771 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
775 | | +------+ +-------+
776 | |------>| A=1 |------ --->| A->0 |
777 | | +------+ \ +-------+
778 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
779 | | +------+ | +-------+
780 | |------>| B=2 |--- | : :
781 | | +------+ \ | : : +-------+
782 +-------+ : : \ | +-------+ | |
783 ---------->| B->2 |------>| |
784 | +-------+ | CPU 2 |
787 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
788 barrier causes all effects \ +-------+ | |
789 prior to the storage of B ---->| A->1 |------>| |
790 to be perceptible to CPU 2 +-------+ | |
794 To illustrate this more completely, consider what could happen if the code
795 contained a load of A either side of the read barrier:
798 ======================= =======================
804 LOAD A [first load of A]
806 LOAD A [second load of A]
808 Even though the two loads of A both occur after the load of B, they may both
809 come up with different values:
812 | | +------+ +-------+
813 | |------>| A=1 |------ --->| A->0 |
814 | | +------+ \ +-------+
815 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
816 | | +------+ | +-------+
817 | |------>| B=2 |--- | : :
818 | | +------+ \ | : : +-------+
819 +-------+ : : \ | +-------+ | |
820 ---------->| B->2 |------>| |
821 | +-------+ | CPU 2 |
825 | | A->0 |------>| 1st |
827 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
828 barrier causes all effects \ +-------+ | |
829 prior to the storage of B ---->| A->1 |------>| 2nd |
830 to be perceptible to CPU 2 +-------+ | |
834 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
835 before the read barrier completes anyway:
838 | | +------+ +-------+
839 | |------>| A=1 |------ --->| A->0 |
840 | | +------+ \ +-------+
841 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
842 | | +------+ | +-------+
843 | |------>| B=2 |--- | : :
844 | | +------+ \ | : : +-------+
845 +-------+ : : \ | +-------+ | |
846 ---------->| B->2 |------>| |
847 | +-------+ | CPU 2 |
851 ---->| A->1 |------>| 1st |
853 rrrrrrrrrrrrrrrrr | |
855 | A->1 |------>| 2nd |
860 The guarantee is that the second load will always come up with A == 1 if the
861 load of B came up with B == 2. No such guarantee exists for the first load of
862 A; that may come up with either A == 0 or A == 1.
865 READ MEMORY BARRIERS VS LOAD SPECULATION
866 ----------------------------------------
868 Many CPUs speculate with loads: that is they see that they will need to load an
869 item from memory, and they find a time where they're not using the bus for any
870 other loads, and so do the load in advance - even though they haven't actually
871 got to that point in the instruction execution flow yet. This permits the
872 actual load instruction to potentially complete immediately because the CPU
873 already has the value to hand.
875 It may turn out that the CPU didn't actually need the value - perhaps because a
876 branch circumvented the load - in which case it can discard the value or just
877 cache it for later use.
882 ======================= =======================
884 DIVIDE } Divide instructions generally
885 DIVIDE } take a long time to perform
888 Which might appear as this:
892 --->| B->2 |------>| |
896 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
897 division speculates on the +-------+ ~ | |
901 Once the divisions are complete --> : : ~-->| |
902 the CPU can then perform the : : | |
903 LOAD with immediate effect : : +-------+
906 Placing a read barrier or a data dependency barrier just before the second
910 ======================= =======================
917 will force any value speculatively obtained to be reconsidered to an extent
918 dependent on the type of barrier used. If there was no change made to the
919 speculated memory location, then the speculated value will just be used:
923 --->| B->2 |------>| |
927 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
928 division speculates on the +-------+ ~ | |
933 rrrrrrrrrrrrrrrr~ | |
940 but if there was an update or an invalidation from another CPU pending, then
941 the speculation will be cancelled and the value reloaded:
945 --->| B->2 |------>| |
949 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
950 division speculates on the +-------+ ~ | |
955 rrrrrrrrrrrrrrrrr | |
957 The speculation is discarded ---> --->| A->1 |------>| |
958 and an updated value is +-------+ | |
959 retrieved : : +-------+
962 ========================
963 EXPLICIT KERNEL BARRIERS
964 ========================
966 The Linux kernel has a variety of different barriers that act at different
969 (*) Compiler barrier.
971 (*) CPU memory barriers.
973 (*) MMIO write barrier.
979 The Linux kernel has an explicit compiler barrier function that prevents the
980 compiler from moving the memory accesses either side of it to the other side:
984 This is a general barrier - lesser varieties of compiler barrier do not exist.
986 The compiler barrier has no direct effect on the CPU, which may then reorder
987 things however it wishes.
993 The Linux kernel has eight basic CPU memory barriers:
995 TYPE MANDATORY SMP CONDITIONAL
996 =============== ======================= ===========================
997 GENERAL mb() smp_mb()
998 WRITE wmb() smp_wmb()
1000 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
1003 All memory barriers except the data dependency barriers imply a compiler
1004 barrier. Data dependencies do not impose any additional compiler ordering.
1006 Aside: In the case of data dependencies, the compiler would be expected to
1007 issue the loads in the correct order (eg. `a[b]` would have to load the value
1008 of b before loading a[b]), however there is no guarantee in the C specification
1009 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1010 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1011 problem of a compiler reloading b after having loaded a[b], thus having a newer
1012 copy of b than a[b]. A consensus has not yet been reached about these problems,
1013 however the ACCESS_ONCE macro is a good place to start looking.
1015 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1016 systems because it is assumed that a CPU will appear to be self-consistent,
1017 and will order overlapping accesses correctly with respect to itself.
1019 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1020 references to shared memory on SMP systems, though the use of locking instead
1023 Mandatory barriers should not be used to control SMP effects, since mandatory
1024 barriers unnecessarily impose overhead on UP systems. They may, however, be
1025 used to control MMIO effects on accesses through relaxed memory I/O windows.
1026 These are required even on non-SMP systems as they affect the order in which
1027 memory operations appear to a device by prohibiting both the compiler and the
1028 CPU from reordering them.
1031 There are some more advanced barrier functions:
1033 (*) set_mb(var, value)
1035 This assigns the value to the variable and then inserts a full memory
1036 barrier after it, depending on the function. It isn't guaranteed to
1037 insert anything more than a compiler barrier in a UP compilation.
1040 (*) smp_mb__before_atomic_dec();
1041 (*) smp_mb__after_atomic_dec();
1042 (*) smp_mb__before_atomic_inc();
1043 (*) smp_mb__after_atomic_inc();
1045 These are for use with atomic add, subtract, increment and decrement
1046 functions that don't return a value, especially when used for reference
1047 counting. These functions do not imply memory barriers.
1049 As an example, consider a piece of code that marks an object as being dead
1050 and then decrements the object's reference count:
1053 smp_mb__before_atomic_dec();
1054 atomic_dec(&obj->ref_count);
1056 This makes sure that the death mark on the object is perceived to be set
1057 *before* the reference counter is decremented.
1059 See Documentation/atomic_ops.txt for more information. See the "Atomic
1060 operations" subsection for information on where to use these.
1063 (*) smp_mb__before_clear_bit(void);
1064 (*) smp_mb__after_clear_bit(void);
1066 These are for use similar to the atomic inc/dec barriers. These are
1067 typically used for bitwise unlocking operations, so care must be taken as
1068 there are no implicit memory barriers here either.
1070 Consider implementing an unlock operation of some nature by clearing a
1071 locking bit. The clear_bit() would then need to be barriered like this:
1073 smp_mb__before_clear_bit();
1076 This prevents memory operations before the clear leaking to after it. See
1077 the subsection on "Locking Functions" with reference to UNLOCK operation
1080 See Documentation/atomic_ops.txt for more information. See the "Atomic
1081 operations" subsection for information on where to use these.
1087 The Linux kernel also has a special barrier for use with memory-mapped I/O
1092 This is a variation on the mandatory write barrier that causes writes to weakly
1093 ordered I/O regions to be partially ordered. Its effects may go beyond the
1094 CPU->Hardware interface and actually affect the hardware at some level.
1096 See the subsection "Locks vs I/O accesses" for more information.
1099 ===============================
1100 IMPLICIT KERNEL MEMORY BARRIERS
1101 ===============================
1103 Some of the other functions in the linux kernel imply memory barriers, amongst
1104 which are locking and scheduling functions.
1106 This specification is a _minimum_ guarantee; any particular architecture may
1107 provide more substantial guarantees, but these may not be relied upon outside
1108 of arch specific code.
1114 The Linux kernel has a number of locking constructs:
1123 In all cases there are variants on "LOCK" operations and "UNLOCK" operations
1124 for each construct. These operations all imply certain barriers:
1126 (1) LOCK operation implication:
1128 Memory operations issued after the LOCK will be completed after the LOCK
1129 operation has completed.
1131 Memory operations issued before the LOCK may be completed after the LOCK
1132 operation has completed.
1134 (2) UNLOCK operation implication:
1136 Memory operations issued before the UNLOCK will be completed before the
1137 UNLOCK operation has completed.
1139 Memory operations issued after the UNLOCK may be completed before the
1140 UNLOCK operation has completed.
1142 (3) LOCK vs LOCK implication:
1144 All LOCK operations issued before another LOCK operation will be completed
1145 before that LOCK operation.
1147 (4) LOCK vs UNLOCK implication:
1149 All LOCK operations issued before an UNLOCK operation will be completed
1150 before the UNLOCK operation.
1152 All UNLOCK operations issued before a LOCK operation will be completed
1153 before the LOCK operation.
1155 (5) Failed conditional LOCK implication:
1157 Certain variants of the LOCK operation may fail, either due to being
1158 unable to get the lock immediately, or due to receiving an unblocked
1159 signal whilst asleep waiting for the lock to become available. Failed
1160 locks do not imply any sort of barrier.
1162 Therefore, from (1), (2) and (4) an UNLOCK followed by an unconditional LOCK is
1163 equivalent to a full barrier, but a LOCK followed by an UNLOCK is not.
1165 [!] Note: one of the consequences of LOCKs and UNLOCKs being only one-way
1166 barriers is that the effects of instructions outside of a critical section
1167 may seep into the inside of the critical section.
1169 A LOCK followed by an UNLOCK may not be assumed to be full memory barrier
1170 because it is possible for an access preceding the LOCK to happen after the
1171 LOCK, and an access following the UNLOCK to happen before the UNLOCK, and the
1172 two accesses can themselves then cross:
1181 LOCK, STORE *B, STORE *A, UNLOCK
1183 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1184 systems, and so cannot be counted on in such a situation to actually achieve
1185 anything at all - especially with respect to I/O accesses - unless combined
1186 with interrupt disabling operations.
1188 See also the section on "Inter-CPU locking barrier effects".
1191 As an example, consider the following:
1202 The following sequence of events is acceptable:
1204 LOCK, {*F,*A}, *E, {*C,*D}, *B, UNLOCK
1206 [+] Note that {*F,*A} indicates a combined access.
1208 But none of the following are:
1210 {*F,*A}, *B, LOCK, *C, *D, UNLOCK, *E
1211 *A, *B, *C, LOCK, *D, UNLOCK, *E, *F
1212 *A, *B, LOCK, *C, UNLOCK, *D, *E, *F
1213 *B, LOCK, *C, *D, UNLOCK, {*F,*A}, *E
1217 INTERRUPT DISABLING FUNCTIONS
1218 -----------------------------
1220 Functions that disable interrupts (LOCK equivalent) and enable interrupts
1221 (UNLOCK equivalent) will act as compiler barriers only. So if memory or I/O
1222 barriers are required in such a situation, they must be provided from some
1226 SLEEP AND WAKE-UP FUNCTIONS
1227 ---------------------------
1229 Sleeping and waking on an event flagged in global data can be viewed as an
1230 interaction between two pieces of data: the task state of the task waiting for
1231 the event and the global data used to indicate the event. To make sure that
1232 these appear to happen in the right order, the primitives to begin the process
1233 of going to sleep, and the primitives to initiate a wake up imply certain
1236 Firstly, the sleeper normally follows something like this sequence of events:
1239 set_current_state(TASK_UNINTERRUPTIBLE);
1240 if (event_indicated)
1245 A general memory barrier is interpolated automatically by set_current_state()
1246 after it has altered the task state:
1249 ===============================
1250 set_current_state();
1252 STORE current->state
1254 LOAD event_indicated
1256 set_current_state() may be wrapped by:
1259 prepare_to_wait_exclusive();
1261 which therefore also imply a general memory barrier after setting the state.
1262 The whole sequence above is available in various canned forms, all of which
1263 interpolate the memory barrier in the right place:
1266 wait_event_interruptible();
1267 wait_event_interruptible_exclusive();
1268 wait_event_interruptible_timeout();
1269 wait_event_killable();
1270 wait_event_timeout();
1275 Secondly, code that performs a wake up normally follows something like this:
1277 event_indicated = 1;
1278 wake_up(&event_wait_queue);
1282 event_indicated = 1;
1283 wake_up_process(event_daemon);
1285 A write memory barrier is implied by wake_up() and co. if and only if they wake
1286 something up. The barrier occurs before the task state is cleared, and so sits
1287 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1290 =============================== ===============================
1291 set_current_state(); STORE event_indicated
1292 set_mb(); wake_up();
1293 STORE current->state <write barrier>
1294 <general barrier> STORE current->state
1295 LOAD event_indicated
1297 The available waker functions include:
1303 wake_up_interruptible();
1304 wake_up_interruptible_all();
1305 wake_up_interruptible_nr();
1306 wake_up_interruptible_poll();
1307 wake_up_interruptible_sync();
1308 wake_up_interruptible_sync_poll();
1310 wake_up_locked_poll();
1316 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1317 order multiple stores before the wake-up with respect to loads of those stored
1318 values after the sleeper has called set_current_state(). For instance, if the
1321 set_current_state(TASK_INTERRUPTIBLE);
1322 if (event_indicated)
1324 __set_current_state(TASK_RUNNING);
1325 do_something(my_data);
1330 event_indicated = 1;
1331 wake_up(&event_wait_queue);
1333 there's no guarantee that the change to event_indicated will be perceived by
1334 the sleeper as coming after the change to my_data. In such a circumstance, the
1335 code on both sides must interpolate its own memory barriers between the
1336 separate data accesses. Thus the above sleeper ought to do:
1338 set_current_state(TASK_INTERRUPTIBLE);
1339 if (event_indicated) {
1341 do_something(my_data);
1344 and the waker should do:
1348 event_indicated = 1;
1349 wake_up(&event_wait_queue);
1352 MISCELLANEOUS FUNCTIONS
1353 -----------------------
1355 Other functions that imply barriers:
1357 (*) schedule() and similar imply full memory barriers.
1360 =================================
1361 INTER-CPU LOCKING BARRIER EFFECTS
1362 =================================
1364 On SMP systems locking primitives give a more substantial form of barrier: one
1365 that does affect memory access ordering on other CPUs, within the context of
1366 conflict on any particular lock.
1369 LOCKS VS MEMORY ACCESSES
1370 ------------------------
1372 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1373 three CPUs; then should the following sequence of events occur:
1376 =============================== ===============================
1384 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
1385 through *H occur in, other than the constraints imposed by the separate locks
1386 on the separate CPUs. It might, for example, see:
1388 *E, LOCK M, LOCK Q, *G, *C, *F, *A, *B, UNLOCK Q, *D, *H, UNLOCK M
1390 But it won't see any of:
1392 *B, *C or *D preceding LOCK M
1393 *A, *B or *C following UNLOCK M
1394 *F, *G or *H preceding LOCK Q
1395 *E, *F or *G following UNLOCK Q
1398 However, if the following occurs:
1401 =============================== ===============================
1416 *E, LOCK M [1], *C, *B, *A, UNLOCK M [1],
1417 LOCK M [2], *H, *F, *G, UNLOCK M [2], *D
1419 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
1421 *B, *C, *D, *F, *G or *H preceding LOCK M [1]
1422 *A, *B or *C following UNLOCK M [1]
1423 *F, *G or *H preceding LOCK M [2]
1424 *A, *B, *C, *E, *F or *G following UNLOCK M [2]
1427 LOCKS VS I/O ACCESSES
1428 ---------------------
1430 Under certain circumstances (especially involving NUMA), I/O accesses within
1431 two spinlocked sections on two different CPUs may be seen as interleaved by the
1432 PCI bridge, because the PCI bridge does not necessarily participate in the
1433 cache-coherence protocol, and is therefore incapable of issuing the required
1434 read memory barriers.
1439 =============================== ===============================
1449 may be seen by the PCI bridge as follows:
1451 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
1453 which would probably cause the hardware to malfunction.
1456 What is necessary here is to intervene with an mmiowb() before dropping the
1457 spinlock, for example:
1460 =============================== ===============================
1472 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
1473 before either of the stores issued on CPU 2.
1476 Furthermore, following a store by a load from the same device obviates the need
1477 for the mmiowb(), because the load forces the store to complete before the load
1481 =============================== ===============================
1492 See Documentation/DocBook/deviceiobook.tmpl for more information.
1495 =================================
1496 WHERE ARE MEMORY BARRIERS NEEDED?
1497 =================================
1499 Under normal operation, memory operation reordering is generally not going to
1500 be a problem as a single-threaded linear piece of code will still appear to
1501 work correctly, even if it's in an SMP kernel. There are, however, four
1502 circumstances in which reordering definitely _could_ be a problem:
1504 (*) Interprocessor interaction.
1506 (*) Atomic operations.
1508 (*) Accessing devices.
1513 INTERPROCESSOR INTERACTION
1514 --------------------------
1516 When there's a system with more than one processor, more than one CPU in the
1517 system may be working on the same data set at the same time. This can cause
1518 synchronisation problems, and the usual way of dealing with them is to use
1519 locks. Locks, however, are quite expensive, and so it may be preferable to
1520 operate without the use of a lock if at all possible. In such a case
1521 operations that affect both CPUs may have to be carefully ordered to prevent
1524 Consider, for example, the R/W semaphore slow path. Here a waiting process is
1525 queued on the semaphore, by virtue of it having a piece of its stack linked to
1526 the semaphore's list of waiting processes:
1528 struct rw_semaphore {
1531 struct list_head waiters;
1534 struct rwsem_waiter {
1535 struct list_head list;
1536 struct task_struct *task;
1539 To wake up a particular waiter, the up_read() or up_write() functions have to:
1541 (1) read the next pointer from this waiter's record to know as to where the
1542 next waiter record is;
1544 (2) read the pointer to the waiter's task structure;
1546 (3) clear the task pointer to tell the waiter it has been given the semaphore;
1548 (4) call wake_up_process() on the task; and
1550 (5) release the reference held on the waiter's task struct.
1552 In other words, it has to perform this sequence of events:
1554 LOAD waiter->list.next;
1560 and if any of these steps occur out of order, then the whole thing may
1563 Once it has queued itself and dropped the semaphore lock, the waiter does not
1564 get the lock again; it instead just waits for its task pointer to be cleared
1565 before proceeding. Since the record is on the waiter's stack, this means that
1566 if the task pointer is cleared _before_ the next pointer in the list is read,
1567 another CPU might start processing the waiter and might clobber the waiter's
1568 stack before the up*() function has a chance to read the next pointer.
1570 Consider then what might happen to the above sequence of events:
1573 =============================== ===============================
1580 Woken up by other event
1585 foo() clobbers *waiter
1587 LOAD waiter->list.next;
1590 This could be dealt with using the semaphore lock, but then the down_xxx()
1591 function has to needlessly get the spinlock again after being woken up.
1593 The way to deal with this is to insert a general SMP memory barrier:
1595 LOAD waiter->list.next;
1602 In this case, the barrier makes a guarantee that all memory accesses before the
1603 barrier will appear to happen before all the memory accesses after the barrier
1604 with respect to the other CPUs on the system. It does _not_ guarantee that all
1605 the memory accesses before the barrier will be complete by the time the barrier
1606 instruction itself is complete.
1608 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
1609 compiler barrier, thus making sure the compiler emits the instructions in the
1610 right order without actually intervening in the CPU. Since there's only one
1611 CPU, that CPU's dependency ordering logic will take care of everything else.
1617 Whilst they are technically interprocessor interaction considerations, atomic
1618 operations are noted specially as some of them imply full memory barriers and
1619 some don't, but they're very heavily relied on as a group throughout the
1622 Any atomic operation that modifies some state in memory and returns information
1623 about the state (old or new) implies an SMP-conditional general memory barrier
1624 (smp_mb()) on each side of the actual operation (with the exception of
1625 explicit lock operations, described later). These include:
1630 atomic_inc_return();
1631 atomic_dec_return();
1632 atomic_add_return();
1633 atomic_sub_return();
1634 atomic_inc_and_test();
1635 atomic_dec_and_test();
1636 atomic_sub_and_test();
1637 atomic_add_negative();
1638 atomic_add_unless(); /* when succeeds (returns 1) */
1640 test_and_clear_bit();
1641 test_and_change_bit();
1643 These are used for such things as implementing LOCK-class and UNLOCK-class
1644 operations and adjusting reference counters towards object destruction, and as
1645 such the implicit memory barrier effects are necessary.
1648 The following operations are potential problems as they do _not_ imply memory
1649 barriers, but might be used for implementing such things as UNLOCK-class
1657 With these the appropriate explicit memory barrier should be used if necessary
1658 (smp_mb__before_clear_bit() for instance).
1661 The following also do _not_ imply memory barriers, and so may require explicit
1662 memory barriers under some circumstances (smp_mb__before_atomic_dec() for
1670 If they're used for statistics generation, then they probably don't need memory
1671 barriers, unless there's a coupling between statistical data.
1673 If they're used for reference counting on an object to control its lifetime,
1674 they probably don't need memory barriers because either the reference count
1675 will be adjusted inside a locked section, or the caller will already hold
1676 sufficient references to make the lock, and thus a memory barrier unnecessary.
1678 If they're used for constructing a lock of some description, then they probably
1679 do need memory barriers as a lock primitive generally has to do things in a
1682 Basically, each usage case has to be carefully considered as to whether memory
1683 barriers are needed or not.
1685 The following operations are special locking primitives:
1687 test_and_set_bit_lock();
1689 __clear_bit_unlock();
1691 These implement LOCK-class and UNLOCK-class operations. These should be used in
1692 preference to other operations when implementing locking primitives, because
1693 their implementations can be optimised on many architectures.
1695 [!] Note that special memory barrier primitives are available for these
1696 situations because on some CPUs the atomic instructions used imply full memory
1697 barriers, and so barrier instructions are superfluous in conjunction with them,
1698 and in such cases the special barrier primitives will be no-ops.
1700 See Documentation/atomic_ops.txt for more information.
1706 Many devices can be memory mapped, and so appear to the CPU as if they're just
1707 a set of memory locations. To control such a device, the driver usually has to
1708 make the right memory accesses in exactly the right order.
1710 However, having a clever CPU or a clever compiler creates a potential problem
1711 in that the carefully sequenced accesses in the driver code won't reach the
1712 device in the requisite order if the CPU or the compiler thinks it is more
1713 efficient to reorder, combine or merge accesses - something that would cause
1714 the device to malfunction.
1716 Inside of the Linux kernel, I/O should be done through the appropriate accessor
1717 routines - such as inb() or writel() - which know how to make such accesses
1718 appropriately sequential. Whilst this, for the most part, renders the explicit
1719 use of memory barriers unnecessary, there are a couple of situations where they
1722 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
1723 so for _all_ general drivers locks should be used and mmiowb() must be
1724 issued prior to unlocking the critical section.
1726 (2) If the accessor functions are used to refer to an I/O memory window with
1727 relaxed memory access properties, then _mandatory_ memory barriers are
1728 required to enforce ordering.
1730 See Documentation/DocBook/deviceiobook.tmpl for more information.
1736 A driver may be interrupted by its own interrupt service routine, and thus the
1737 two parts of the driver may interfere with each other's attempts to control or
1740 This may be alleviated - at least in part - by disabling local interrupts (a
1741 form of locking), such that the critical operations are all contained within
1742 the interrupt-disabled section in the driver. Whilst the driver's interrupt
1743 routine is executing, the driver's core may not run on the same CPU, and its
1744 interrupt is not permitted to happen again until the current interrupt has been
1745 handled, thus the interrupt handler does not need to lock against that.
1747 However, consider a driver that was talking to an ethernet card that sports an
1748 address register and a data register. If that driver's core talks to the card
1749 under interrupt-disablement and then the driver's interrupt handler is invoked:
1760 The store to the data register might happen after the second store to the
1761 address register if ordering rules are sufficiently relaxed:
1763 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
1766 If ordering rules are relaxed, it must be assumed that accesses done inside an
1767 interrupt disabled section may leak outside of it and may interleave with
1768 accesses performed in an interrupt - and vice versa - unless implicit or
1769 explicit barriers are used.
1771 Normally this won't be a problem because the I/O accesses done inside such
1772 sections will include synchronous load operations on strictly ordered I/O
1773 registers that form implicit I/O barriers. If this isn't sufficient then an
1774 mmiowb() may need to be used explicitly.
1777 A similar situation may occur between an interrupt routine and two routines
1778 running on separate CPUs that communicate with each other. If such a case is
1779 likely, then interrupt-disabling locks should be used to guarantee ordering.
1782 ==========================
1783 KERNEL I/O BARRIER EFFECTS
1784 ==========================
1786 When accessing I/O memory, drivers should use the appropriate accessor
1791 These are intended to talk to I/O space rather than memory space, but
1792 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
1793 indeed have special I/O space access cycles and instructions, but many
1794 CPUs don't have such a concept.
1796 The PCI bus, amongst others, defines an I/O space concept which - on such
1797 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
1798 space. However, it may also be mapped as a virtual I/O space in the CPU's
1799 memory map, particularly on those CPUs that don't support alternate I/O
1802 Accesses to this space may be fully synchronous (as on i386), but
1803 intermediary bridges (such as the PCI host bridge) may not fully honour
1806 They are guaranteed to be fully ordered with respect to each other.
1808 They are not guaranteed to be fully ordered with respect to other types of
1809 memory and I/O operation.
1811 (*) readX(), writeX():
1813 Whether these are guaranteed to be fully ordered and uncombined with
1814 respect to each other on the issuing CPU depends on the characteristics
1815 defined for the memory window through which they're accessing. On later
1816 i386 architecture machines, for example, this is controlled by way of the
1819 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
1820 provided they're not accessing a prefetchable device.
1822 However, intermediary hardware (such as a PCI bridge) may indulge in
1823 deferral if it so wishes; to flush a store, a load from the same location
1824 is preferred[*], but a load from the same device or from configuration
1825 space should suffice for PCI.
1827 [*] NOTE! attempting to load from the same location as was written to may
1828 cause a malfunction - consider the 16550 Rx/Tx serial registers for
1831 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
1832 force stores to be ordered.
1834 Please refer to the PCI specification for more information on interactions
1835 between PCI transactions.
1839 These are similar to readX(), but are not guaranteed to be ordered in any
1840 way. Be aware that there is no I/O read barrier available.
1842 (*) ioreadX(), iowriteX()
1844 These will perform appropriately for the type of access they're actually
1845 doing, be it inX()/outX() or readX()/writeX().
1848 ========================================
1849 ASSUMED MINIMUM EXECUTION ORDERING MODEL
1850 ========================================
1852 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
1853 maintain the appearance of program causality with respect to itself. Some CPUs
1854 (such as i386 or x86_64) are more constrained than others (such as powerpc or
1855 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
1856 of arch-specific code.
1858 This means that it must be considered that the CPU will execute its instruction
1859 stream in any order it feels like - or even in parallel - provided that if an
1860 instruction in the stream depends on an earlier instruction, then that
1861 earlier instruction must be sufficiently complete[*] before the later
1862 instruction may proceed; in other words: provided that the appearance of
1863 causality is maintained.
1865 [*] Some instructions have more than one effect - such as changing the
1866 condition codes, changing registers or changing memory - and different
1867 instructions may depend on different effects.
1869 A CPU may also discard any instruction sequence that winds up having no
1870 ultimate effect. For example, if two adjacent instructions both load an
1871 immediate value into the same register, the first may be discarded.
1874 Similarly, it has to be assumed that compiler might reorder the instruction
1875 stream in any way it sees fit, again provided the appearance of causality is
1879 ============================
1880 THE EFFECTS OF THE CPU CACHE
1881 ============================
1883 The way cached memory operations are perceived across the system is affected to
1884 a certain extent by the caches that lie between CPUs and memory, and by the
1885 memory coherence system that maintains the consistency of state in the system.
1887 As far as the way a CPU interacts with another part of the system through the
1888 caches goes, the memory system has to include the CPU's caches, and memory
1889 barriers for the most part act at the interface between the CPU and its cache
1890 (memory barriers logically act on the dotted line in the following diagram):
1892 <--- CPU ---> : <----------- Memory ----------->
1894 +--------+ +--------+ : +--------+ +-----------+
1895 | | | | : | | | | +--------+
1896 | CPU | | Memory | : | CPU | | | | |
1897 | Core |--->| Access |----->| Cache |<-->| | | |
1898 | | | Queue | : | | | |--->| Memory |
1899 | | | | : | | | | | |
1900 +--------+ +--------+ : +--------+ | | | |
1901 : | Cache | +--------+
1903 : | Mechanism | +--------+
1904 +--------+ +--------+ : +--------+ | | | |
1905 | | | | : | | | | | |
1906 | CPU | | Memory | : | CPU | | |--->| Device |
1907 | Core |--->| Access |----->| Cache |<-->| | | |
1908 | | | Queue | : | | | | | |
1909 | | | | : | | | | +--------+
1910 +--------+ +--------+ : +--------+ +-----------+
1914 Although any particular load or store may not actually appear outside of the
1915 CPU that issued it since it may have been satisfied within the CPU's own cache,
1916 it will still appear as if the full memory access had taken place as far as the
1917 other CPUs are concerned since the cache coherency mechanisms will migrate the
1918 cacheline over to the accessing CPU and propagate the effects upon conflict.
1920 The CPU core may execute instructions in any order it deems fit, provided the
1921 expected program causality appears to be maintained. Some of the instructions
1922 generate load and store operations which then go into the queue of memory
1923 accesses to be performed. The core may place these in the queue in any order
1924 it wishes, and continue execution until it is forced to wait for an instruction
1927 What memory barriers are concerned with is controlling the order in which
1928 accesses cross from the CPU side of things to the memory side of things, and
1929 the order in which the effects are perceived to happen by the other observers
1932 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
1933 their own loads and stores as if they had happened in program order.
1935 [!] MMIO or other device accesses may bypass the cache system. This depends on
1936 the properties of the memory window through which devices are accessed and/or
1937 the use of any special device communication instructions the CPU may have.
1943 Life isn't quite as simple as it may appear above, however: for while the
1944 caches are expected to be coherent, there's no guarantee that that coherency
1945 will be ordered. This means that whilst changes made on one CPU will
1946 eventually become visible on all CPUs, there's no guarantee that they will
1947 become apparent in the same order on those other CPUs.
1950 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
1951 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
1956 +--------+ : +--->| Cache A |<------->| |
1957 | | : | +---------+ | |
1959 | | : | +---------+ | |
1960 +--------+ : +--->| Cache B |<------->| |
1963 : +---------+ | System |
1964 +--------+ : +--->| Cache C |<------->| |
1965 | | : | +---------+ | |
1967 | | : | +---------+ | |
1968 +--------+ : +--->| Cache D |<------->| |
1973 Imagine the system has the following properties:
1975 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
1978 (*) an even-numbered cache line may be in cache B, cache D or it may still be
1981 (*) whilst the CPU core is interrogating one cache, the other cache may be
1982 making use of the bus to access the rest of the system - perhaps to
1983 displace a dirty cacheline or to do a speculative load;
1985 (*) each cache has a queue of operations that need to be applied to that cache
1986 to maintain coherency with the rest of the system;
1988 (*) the coherency queue is not flushed by normal loads to lines already
1989 present in the cache, even though the contents of the queue may
1990 potentially affect those loads.
1992 Imagine, then, that two writes are made on the first CPU, with a write barrier
1993 between them to guarantee that they will appear to reach that CPU's caches in
1994 the requisite order:
1997 =============== =============== =======================================
1998 u == 0, v == 1 and p == &u, q == &u
2000 smp_wmb(); Make sure change to v is visible before
2002 <A:modify v=2> v is now in cache A exclusively
2004 <B:modify p=&v> p is now in cache B exclusively
2006 The write memory barrier forces the other CPUs in the system to perceive that
2007 the local CPU's caches have apparently been updated in the correct order. But
2008 now imagine that the second CPU wants to read those values:
2011 =============== =============== =======================================
2016 The above pair of reads may then fail to happen in the expected order, as the
2017 cacheline holding p may get updated in one of the second CPU's caches whilst
2018 the update to the cacheline holding v is delayed in the other of the second
2019 CPU's caches by some other cache event:
2022 =============== =============== =======================================
2023 u == 0, v == 1 and p == &u, q == &u
2026 <A:modify v=2> <C:busy>
2030 <B:modify p=&v> <D:commit p=&v>
2033 <C:read *q> Reads from v before v updated in cache
2037 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2038 no guarantee that, without intervention, the order of update will be the same
2039 as that committed on CPU 1.
2042 To intervene, we need to interpolate a data dependency barrier or a read
2043 barrier between the loads. This will force the cache to commit its coherency
2044 queue before processing any further requests:
2047 =============== =============== =======================================
2048 u == 0, v == 1 and p == &u, q == &u
2051 <A:modify v=2> <C:busy>
2055 <B:modify p=&v> <D:commit p=&v>
2057 smp_read_barrier_depends()
2061 <C:read *q> Reads from v after v updated in cache
2064 This sort of problem can be encountered on DEC Alpha processors as they have a
2065 split cache that improves performance by making better use of the data bus.
2066 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2067 access depends on a read, not all do, so it may not be relied on.
2069 Other CPUs may also have split caches, but must coordinate between the various
2070 cachelets for normal memory accesses. The semantics of the Alpha removes the
2071 need for coordination in the absence of memory barriers.
2074 CACHE COHERENCY VS DMA
2075 ----------------------
2077 Not all systems maintain cache coherency with respect to devices doing DMA. In
2078 such cases, a device attempting DMA may obtain stale data from RAM because
2079 dirty cache lines may be resident in the caches of various CPUs, and may not
2080 have been written back to RAM yet. To deal with this, the appropriate part of
2081 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2082 invalidate them as well).
2084 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2085 cache lines being written back to RAM from a CPU's cache after the device has
2086 installed its own data, or cache lines present in the CPU's cache may simply
2087 obscure the fact that RAM has been updated, until at such time as the cacheline
2088 is discarded from the CPU's cache and reloaded. To deal with this, the
2089 appropriate part of the kernel must invalidate the overlapping bits of the
2092 See Documentation/cachetlb.txt for more information on cache management.
2095 CACHE COHERENCY VS MMIO
2096 -----------------------
2098 Memory mapped I/O usually takes place through memory locations that are part of
2099 a window in the CPU's memory space that has different properties assigned than
2100 the usual RAM directed window.
2102 Amongst these properties is usually the fact that such accesses bypass the
2103 caching entirely and go directly to the device buses. This means MMIO accesses
2104 may, in effect, overtake accesses to cached memory that were emitted earlier.
2105 A memory barrier isn't sufficient in such a case, but rather the cache must be
2106 flushed between the cached memory write and the MMIO access if the two are in
2110 =========================
2111 THE THINGS CPUS GET UP TO
2112 =========================
2114 A programmer might take it for granted that the CPU will perform memory
2115 operations in exactly the order specified, so that if the CPU is, for example,
2116 given the following piece of code to execute:
2124 they would then expect that the CPU will complete the memory operation for each
2125 instruction before moving on to the next one, leading to a definite sequence of
2126 operations as seen by external observers in the system:
2128 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2131 Reality is, of course, much messier. With many CPUs and compilers, the above
2132 assumption doesn't hold because:
2134 (*) loads are more likely to need to be completed immediately to permit
2135 execution progress, whereas stores can often be deferred without a
2138 (*) loads may be done speculatively, and the result discarded should it prove
2139 to have been unnecessary;
2141 (*) loads may be done speculatively, leading to the result having been fetched
2142 at the wrong time in the expected sequence of events;
2144 (*) the order of the memory accesses may be rearranged to promote better use
2145 of the CPU buses and caches;
2147 (*) loads and stores may be combined to improve performance when talking to
2148 memory or I/O hardware that can do batched accesses of adjacent locations,
2149 thus cutting down on transaction setup costs (memory and PCI devices may
2150 both be able to do this); and
2152 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2153 mechanisms may alleviate this - once the store has actually hit the cache
2154 - there's no guarantee that the coherency management will be propagated in
2155 order to other CPUs.
2157 So what another CPU, say, might actually observe from the above piece of code
2160 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2162 (Where "LOAD {*C,*D}" is a combined load)
2165 However, it is guaranteed that a CPU will be self-consistent: it will see its
2166 _own_ accesses appear to be correctly ordered, without the need for a memory
2167 barrier. For instance with the following code:
2176 and assuming no intervention by an external influence, it can be assumed that
2177 the final result will appear to be:
2179 U == the original value of *A
2184 The code above may cause the CPU to generate the full sequence of memory
2187 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2189 in that order, but, without intervention, the sequence may have almost any
2190 combination of elements combined or discarded, provided the program's view of
2191 the world remains consistent.
2193 The compiler may also combine, discard or defer elements of the sequence before
2194 the CPU even sees them.
2205 since, without a write barrier, it can be assumed that the effect of the
2206 storage of V to *A is lost. Similarly:
2211 may, without a memory barrier, be reduced to:
2216 and the LOAD operation never appear outside of the CPU.
2219 AND THEN THERE'S THE ALPHA
2220 --------------------------
2222 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2223 some versions of the Alpha CPU have a split data cache, permitting them to have
2224 two semantically-related cache lines updated at separate times. This is where
2225 the data dependency barrier really becomes necessary as this synchronises both
2226 caches with the memory coherence system, thus making it seem like pointer
2227 changes vs new data occur in the right order.
2229 The Alpha defines the Linux kernel's memory barrier model.
2231 See the subsection on "Cache Coherency" above.
2241 Memory barriers can be used to implement circular buffering without the need
2242 of a lock to serialise the producer with the consumer. See:
2244 Documentation/circular-buffers.txt
2253 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2255 Chapter 5.2: Physical Address Space Characteristics
2256 Chapter 5.4: Caches and Write Buffers
2257 Chapter 5.5: Data Sharing
2258 Chapter 5.6: Read/Write Ordering
2260 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2261 Chapter 7.1: Memory-Access Ordering
2262 Chapter 7.4: Buffering and Combining Memory Writes
2264 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2265 System Programming Guide
2266 Chapter 7.1: Locked Atomic Operations
2267 Chapter 7.2: Memory Ordering
2268 Chapter 7.4: Serializing Instructions
2270 The SPARC Architecture Manual, Version 9
2271 Chapter 8: Memory Models
2272 Appendix D: Formal Specification of the Memory Models
2273 Appendix J: Programming with the Memory Models
2275 UltraSPARC Programmer Reference Manual
2276 Chapter 5: Memory Accesses and Cacheability
2277 Chapter 15: Sparc-V9 Memory Models
2279 UltraSPARC III Cu User's Manual
2280 Chapter 9: Memory Models
2282 UltraSPARC IIIi Processor User's Manual
2283 Chapter 8: Memory Models
2285 UltraSPARC Architecture 2005
2287 Appendix D: Formal Specifications of the Memory Models
2289 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2290 Chapter 8: Memory Models
2291 Appendix F: Caches and Cache Coherency
2293 Solaris Internals, Core Kernel Architecture, p63-68:
2294 Chapter 3.3: Hardware Considerations for Locks and
2297 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2298 for Kernel Programmers:
2299 Chapter 13: Other Memory Models
2301 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2302 Section 2.6: Speculation
2303 Section 4.4: Memory Access