1 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
5 By: David Howells <dhowells@redhat.com>
9 (*) Abstract memory access model.
14 (*) What are memory barriers?
16 - Varieties of memory barrier.
17 - What may not be assumed about memory barriers?
18 - Data dependency barriers.
19 - Control dependencies.
20 - SMP barrier pairing.
21 - Examples of memory barrier sequences.
22 - Read memory barriers vs load speculation.
24 (*) Explicit kernel barriers.
27 - CPU memory barriers.
30 (*) Implicit kernel memory barriers.
33 - Interrupt disabling functions.
34 - Sleep and wake-up functions.
35 - Miscellaneous functions.
37 (*) Inter-CPU locking barrier effects.
39 - Locks vs memory accesses.
40 - Locks vs I/O accesses.
42 (*) Where are memory barriers needed?
44 - Interprocessor interaction.
49 (*) Kernel I/O barrier effects.
51 (*) Assumed minimum execution ordering model.
53 (*) The effects of the cpu cache.
56 - Cache coherency vs DMA.
57 - Cache coherency vs MMIO.
59 (*) The things CPUs get up to.
61 - And then there's the Alpha.
66 ============================
67 ABSTRACT MEMORY ACCESS MODEL
68 ============================
70 Consider the following abstract model of the system:
75 +-------+ : +--------+ : +-------+
78 | CPU 1 |<----->| Memory |<----->| CPU 2 |
81 +-------+ : +--------+ : +-------+
89 +---------->| Device |<----------+
95 Each CPU executes a program that generates memory access operations. In the
96 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
97 perform the memory operations in any order it likes, provided program causality
98 appears to be maintained. Similarly, the compiler may also arrange the
99 instructions it emits in any order it likes, provided it doesn't affect the
100 apparent operation of the program.
102 So in the above diagram, the effects of the memory operations performed by a
103 CPU are perceived by the rest of the system as the operations cross the
104 interface between the CPU and rest of the system (the dotted lines).
107 For example, consider the following sequence of events:
110 =============== ===============
115 The set of accesses as seen by the memory system in the middle can be arranged
116 in 24 different combinations:
118 STORE A=3, STORE B=4, x=LOAD A->3, y=LOAD B->4
119 STORE A=3, STORE B=4, y=LOAD B->4, x=LOAD A->3
120 STORE A=3, x=LOAD A->3, STORE B=4, y=LOAD B->4
121 STORE A=3, x=LOAD A->3, y=LOAD B->2, STORE B=4
122 STORE A=3, y=LOAD B->2, STORE B=4, x=LOAD A->3
123 STORE A=3, y=LOAD B->2, x=LOAD A->3, STORE B=4
124 STORE B=4, STORE A=3, x=LOAD A->3, y=LOAD B->4
128 and can thus result in four different combinations of values:
136 Furthermore, the stores committed by a CPU to the memory system may not be
137 perceived by the loads made by another CPU in the same order as the stores were
141 As a further example, consider this sequence of events:
144 =============== ===============
145 { A == 1, B == 2, C = 3, P == &A, Q == &C }
149 There is an obvious data dependency here, as the value loaded into D depends on
150 the address retrieved from P by CPU 2. At the end of the sequence, any of the
151 following results are possible:
153 (Q == &A) and (D == 1)
154 (Q == &B) and (D == 2)
155 (Q == &B) and (D == 4)
157 Note that CPU 2 will never try and load C into D because the CPU will load P
158 into Q before issuing the load of *Q.
164 Some devices present their control interfaces as collections of memory
165 locations, but the order in which the control registers are accessed is very
166 important. For instance, imagine an ethernet card with a set of internal
167 registers that are accessed through an address port register (A) and a data
168 port register (D). To read internal register 5, the following code might then
174 but this might show up as either of the following two sequences:
176 STORE *A = 5, x = LOAD *D
177 x = LOAD *D, STORE *A = 5
179 the second of which will almost certainly result in a malfunction, since it set
180 the address _after_ attempting to read the register.
186 There are some minimal guarantees that may be expected of a CPU:
188 (*) On any given CPU, dependent memory accesses will be issued in order, with
189 respect to itself. This means that for:
193 the CPU will issue the following memory operations:
195 Q = LOAD P, D = LOAD *Q
197 and always in that order.
199 (*) Overlapping loads and stores within a particular CPU will appear to be
200 ordered within that CPU. This means that for:
204 the CPU will only issue the following sequence of memory operations:
206 a = LOAD *X, STORE *X = b
212 the CPU will only issue:
214 STORE *X = c, d = LOAD *X
216 (Loads and stores overlap if they are targeted at overlapping pieces of
219 And there are a number of things that _must_ or _must_not_ be assumed:
221 (*) It _must_not_ be assumed that independent loads and stores will be issued
222 in the order given. This means that for:
224 X = *A; Y = *B; *D = Z;
226 we may get any of the following sequences:
228 X = LOAD *A, Y = LOAD *B, STORE *D = Z
229 X = LOAD *A, STORE *D = Z, Y = LOAD *B
230 Y = LOAD *B, X = LOAD *A, STORE *D = Z
231 Y = LOAD *B, STORE *D = Z, X = LOAD *A
232 STORE *D = Z, X = LOAD *A, Y = LOAD *B
233 STORE *D = Z, Y = LOAD *B, X = LOAD *A
235 (*) It _must_ be assumed that overlapping memory accesses may be merged or
236 discarded. This means that for:
238 X = *A; Y = *(A + 4);
240 we may get any one of the following sequences:
242 X = LOAD *A; Y = LOAD *(A + 4);
243 Y = LOAD *(A + 4); X = LOAD *A;
244 {X, Y} = LOAD {*A, *(A + 4) };
250 we may get either of:
252 STORE *A = X; Y = LOAD *A;
256 =========================
257 WHAT ARE MEMORY BARRIERS?
258 =========================
260 As can be seen above, independent memory operations are effectively performed
261 in random order, but this can be a problem for CPU-CPU interaction and for I/O.
262 What is required is some way of intervening to instruct the compiler and the
263 CPU to restrict the order.
265 Memory barriers are such interventions. They impose a perceived partial
266 ordering over the memory operations on either side of the barrier.
268 Such enforcement is important because the CPUs and other devices in a system
269 can use a variety of tricks to improve performance, including reordering,
270 deferral and combination of memory operations; speculative loads; speculative
271 branch prediction and various types of caching. Memory barriers are used to
272 override or suppress these tricks, allowing the code to sanely control the
273 interaction of multiple CPUs and/or devices.
276 VARIETIES OF MEMORY BARRIER
277 ---------------------------
279 Memory barriers come in four basic varieties:
281 (1) Write (or store) memory barriers.
283 A write memory barrier gives a guarantee that all the STORE operations
284 specified before the barrier will appear to happen before all the STORE
285 operations specified after the barrier with respect to the other
286 components of the system.
288 A write barrier is a partial ordering on stores only; it is not required
289 to have any effect on loads.
291 A CPU can be viewed as committing a sequence of store operations to the
292 memory system as time progresses. All stores before a write barrier will
293 occur in the sequence _before_ all the stores after the write barrier.
295 [!] Note that write barriers should normally be paired with read or data
296 dependency barriers; see the "SMP barrier pairing" subsection.
299 (2) Data dependency barriers.
301 A data dependency barrier is a weaker form of read barrier. In the case
302 where two loads are performed such that the second depends on the result
303 of the first (eg: the first load retrieves the address to which the second
304 load will be directed), a data dependency barrier would be required to
305 make sure that the target of the second load is updated before the address
306 obtained by the first load is accessed.
308 A data dependency barrier is a partial ordering on interdependent loads
309 only; it is not required to have any effect on stores, independent loads
310 or overlapping loads.
312 As mentioned in (1), the other CPUs in the system can be viewed as
313 committing sequences of stores to the memory system that the CPU being
314 considered can then perceive. A data dependency barrier issued by the CPU
315 under consideration guarantees that for any load preceding it, if that
316 load touches one of a sequence of stores from another CPU, then by the
317 time the barrier completes, the effects of all the stores prior to that
318 touched by the load will be perceptible to any loads issued after the data
321 See the "Examples of memory barrier sequences" subsection for diagrams
322 showing the ordering constraints.
324 [!] Note that the first load really has to have a _data_ dependency and
325 not a control dependency. If the address for the second load is dependent
326 on the first load, but the dependency is through a conditional rather than
327 actually loading the address itself, then it's a _control_ dependency and
328 a full read barrier or better is required. See the "Control dependencies"
329 subsection for more information.
331 [!] Note that data dependency barriers should normally be paired with
332 write barriers; see the "SMP barrier pairing" subsection.
335 (3) Read (or load) memory barriers.
337 A read barrier is a data dependency barrier plus a guarantee that all the
338 LOAD operations specified before the barrier will appear to happen before
339 all the LOAD operations specified after the barrier with respect to the
340 other components of the system.
342 A read barrier is a partial ordering on loads only; it is not required to
343 have any effect on stores.
345 Read memory barriers imply data dependency barriers, and so can substitute
348 [!] Note that read barriers should normally be paired with write barriers;
349 see the "SMP barrier pairing" subsection.
352 (4) General memory barriers.
354 A general memory barrier gives a guarantee that all the LOAD and STORE
355 operations specified before the barrier will appear to happen before all
356 the LOAD and STORE operations specified after the barrier with respect to
357 the other components of the system.
359 A general memory barrier is a partial ordering over both loads and stores.
361 General memory barriers imply both read and write memory barriers, and so
362 can substitute for either.
365 And a couple of implicit varieties:
369 This acts as a one-way permeable barrier. It guarantees that all memory
370 operations after the LOCK operation will appear to happen after the LOCK
371 operation with respect to the other components of the system.
373 Memory operations that occur before a LOCK operation may appear to happen
376 A LOCK operation should almost always be paired with an UNLOCK operation.
379 (6) UNLOCK operations.
381 This also acts as a one-way permeable barrier. It guarantees that all
382 memory operations before the UNLOCK operation will appear to happen before
383 the UNLOCK operation with respect to the other components of the system.
385 Memory operations that occur after an UNLOCK operation may appear to
386 happen before it completes.
388 LOCK and UNLOCK operations are guaranteed to appear with respect to each
389 other strictly in the order specified.
391 The use of LOCK and UNLOCK operations generally precludes the need for
392 other sorts of memory barrier (but note the exceptions mentioned in the
393 subsection "MMIO write barrier").
396 Memory barriers are only required where there's a possibility of interaction
397 between two CPUs or between a CPU and a device. If it can be guaranteed that
398 there won't be any such interaction in any particular piece of code, then
399 memory barriers are unnecessary in that piece of code.
402 Note that these are the _minimum_ guarantees. Different architectures may give
403 more substantial guarantees, but they may _not_ be relied upon outside of arch
407 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
408 ----------------------------------------------
410 There are certain things that the Linux kernel memory barriers do not guarantee:
412 (*) There is no guarantee that any of the memory accesses specified before a
413 memory barrier will be _complete_ by the completion of a memory barrier
414 instruction; the barrier can be considered to draw a line in that CPU's
415 access queue that accesses of the appropriate type may not cross.
417 (*) There is no guarantee that issuing a memory barrier on one CPU will have
418 any direct effect on another CPU or any other hardware in the system. The
419 indirect effect will be the order in which the second CPU sees the effects
420 of the first CPU's accesses occur, but see the next point:
422 (*) There is no guarantee that a CPU will see the correct order of effects
423 from a second CPU's accesses, even _if_ the second CPU uses a memory
424 barrier, unless the first CPU _also_ uses a matching memory barrier (see
425 the subsection on "SMP Barrier Pairing").
427 (*) There is no guarantee that some intervening piece of off-the-CPU
428 hardware[*] will not reorder the memory accesses. CPU cache coherency
429 mechanisms should propagate the indirect effects of a memory barrier
430 between CPUs, but might not do so in order.
432 [*] For information on bus mastering DMA and coherency please read:
434 Documentation/PCI/pci.txt
435 Documentation/PCI/PCI-DMA-mapping.txt
436 Documentation/DMA-API.txt
439 DATA DEPENDENCY BARRIERS
440 ------------------------
442 The usage requirements of data dependency barriers are a little subtle, and
443 it's not always obvious that they're needed. To illustrate, consider the
444 following sequence of events:
447 =============== ===============
448 { A == 1, B == 2, C = 3, P == &A, Q == &C }
455 There's a clear data dependency here, and it would seem that by the end of the
456 sequence, Q must be either &A or &B, and that:
458 (Q == &A) implies (D == 1)
459 (Q == &B) implies (D == 4)
461 But! CPU 2's perception of P may be updated _before_ its perception of B, thus
462 leading to the following situation:
464 (Q == &B) and (D == 2) ????
466 Whilst this may seem like a failure of coherency or causality maintenance, it
467 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
470 To deal with this, a data dependency barrier or better must be inserted
471 between the address load and the data load:
474 =============== ===============
475 { A == 1, B == 2, C = 3, P == &A, Q == &C }
480 <data dependency barrier>
483 This enforces the occurrence of one of the two implications, and prevents the
484 third possibility from arising.
486 [!] Note that this extremely counterintuitive situation arises most easily on
487 machines with split caches, so that, for example, one cache bank processes
488 even-numbered cache lines and the other bank processes odd-numbered cache
489 lines. The pointer P might be stored in an odd-numbered cache line, and the
490 variable B might be stored in an even-numbered cache line. Then, if the
491 even-numbered bank of the reading CPU's cache is extremely busy while the
492 odd-numbered bank is idle, one can see the new value of the pointer P (&B),
493 but the old value of the variable B (2).
496 Another example of where data dependency barriers might by required is where a
497 number is read from memory and then used to calculate the index for an array
501 =============== ===============
502 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
507 <data dependency barrier>
511 The data dependency barrier is very important to the RCU system, for example.
512 See rcu_dereference() in include/linux/rcupdate.h. This permits the current
513 target of an RCU'd pointer to be replaced with a new modified target, without
514 the replacement target appearing to be incompletely initialised.
516 See also the subsection on "Cache Coherency" for a more thorough example.
522 A control dependency requires a full read memory barrier, not simply a data
523 dependency barrier to make it work correctly. Consider the following bit of
529 <data dependency barrier>
532 This will not have the desired effect because there is no actual data
533 dependency, but rather a control dependency that the CPU may short-circuit by
534 attempting to predict the outcome in advance. In such a case what's actually
547 When dealing with CPU-CPU interactions, certain types of memory barrier should
548 always be paired. A lack of appropriate pairing is almost certainly an error.
550 A write barrier should always be paired with a data dependency barrier or read
551 barrier, though a general barrier would also be viable. Similarly a read
552 barrier or a data dependency barrier should always be paired with at least an
553 write barrier, though, again, a general barrier is viable:
556 =============== ===============
566 =============== ===============================
570 <data dependency barrier>
573 Basically, the read barrier always has to be there, even though it can be of
576 [!] Note that the stores before the write barrier would normally be expected to
577 match the loads after the read barrier or the data dependency barrier, and vice
581 =============== ===============
582 a = 1; }---- --->{ v = c
584 <write barrier> \ <read barrier>
585 c = 3; } / \ { x = a;
586 d = 4; }---- --->{ y = b;
589 EXAMPLES OF MEMORY BARRIER SEQUENCES
590 ------------------------------------
592 Firstly, write barriers act as partial orderings on store operations.
593 Consider the following sequence of events:
596 =======================
604 This sequence of events is committed to the memory coherence system in an order
605 that the rest of the system might perceive as the unordered set of { STORE A,
606 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E
611 | |------>| C=3 | } /\
612 | | : +------+ }----- \ -----> Events perceptible to
613 | | : | A=1 | } \/ the rest of the system
615 | CPU 1 | : | B=2 | }
617 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
618 | | +------+ } requires all stores prior to the
619 | | : | E=5 | } barrier to be committed before
620 | | : +------+ } further stores may take place
625 | Sequence in which stores are committed to the
626 | memory system by CPU 1
630 Secondly, data dependency barriers act as partial orderings on data-dependent
631 loads. Consider the following sequence of events:
634 ======================= =======================
635 { B = 7; X = 9; Y = 8; C = &Y }
640 STORE D = 4 LOAD C (gets &B)
643 Without intervention, CPU 2 may perceive the events on CPU 1 in some
644 effectively random order, despite the write barrier issued by CPU 1:
647 | | +------+ +-------+ | Sequence of update
648 | |------>| B=2 |----- --->| Y->8 | | of perception on
649 | | : +------+ \ +-------+ | CPU 2
650 | CPU 1 | : | A=1 | \ --->| C->&Y | V
651 | | +------+ | +-------+
652 | | wwwwwwwwwwwwwwww | : :
654 | | : | C=&B |--- | : : +-------+
655 | | : +------+ \ | +-------+ | |
656 | |------>| D=4 | ----------->| C->&B |------>| |
657 | | +------+ | +-------+ | |
658 +-------+ : : | : : | |
662 Apparently incorrect ---> | | B->7 |------>| |
663 perception of B (!) | +-------+ | |
666 The load of X holds ---> \ | X->9 |------>| |
667 up the maintenance \ +-------+ | |
668 of coherence of B ----->| B->2 | +-------+
673 In the above example, CPU 2 perceives that B is 7, despite the load of *C
674 (which would be B) coming after the LOAD of C.
676 If, however, a data dependency barrier were to be placed between the load of C
677 and the load of *C (ie: B) on CPU 2:
680 ======================= =======================
681 { B = 7; X = 9; Y = 8; C = &Y }
686 STORE D = 4 LOAD C (gets &B)
687 <data dependency barrier>
690 then the following will occur:
693 | | +------+ +-------+
694 | |------>| B=2 |----- --->| Y->8 |
695 | | : +------+ \ +-------+
696 | CPU 1 | : | A=1 | \ --->| C->&Y |
697 | | +------+ | +-------+
698 | | wwwwwwwwwwwwwwww | : :
700 | | : | C=&B |--- | : : +-------+
701 | | : +------+ \ | +-------+ | |
702 | |------>| D=4 | ----------->| C->&B |------>| |
703 | | +------+ | +-------+ | |
704 +-------+ : : | : : | |
710 Makes sure all effects ---> \ ddddddddddddddddd | |
711 prior to the store of C \ +-------+ | |
712 are perceptible to ----->| B->2 |------>| |
713 subsequent loads +-------+ | |
717 And thirdly, a read barrier acts as a partial order on loads. Consider the
718 following sequence of events:
721 ======================= =======================
729 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
730 some effectively random order, despite the write barrier issued by CPU 1:
733 | | +------+ +-------+
734 | |------>| A=1 |------ --->| A->0 |
735 | | +------+ \ +-------+
736 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
737 | | +------+ | +-------+
738 | |------>| B=2 |--- | : :
739 | | +------+ \ | : : +-------+
740 +-------+ : : \ | +-------+ | |
741 ---------->| B->2 |------>| |
742 | +-------+ | CPU 2 |
753 If, however, a read barrier were to be placed between the load of B and the
757 ======================= =======================
766 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU
770 | | +------+ +-------+
771 | |------>| A=1 |------ --->| A->0 |
772 | | +------+ \ +-------+
773 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
774 | | +------+ | +-------+
775 | |------>| B=2 |--- | : :
776 | | +------+ \ | : : +-------+
777 +-------+ : : \ | +-------+ | |
778 ---------->| B->2 |------>| |
779 | +-------+ | CPU 2 |
782 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
783 barrier causes all effects \ +-------+ | |
784 prior to the storage of B ---->| A->1 |------>| |
785 to be perceptible to CPU 2 +-------+ | |
789 To illustrate this more completely, consider what could happen if the code
790 contained a load of A either side of the read barrier:
793 ======================= =======================
799 LOAD A [first load of A]
801 LOAD A [second load of A]
803 Even though the two loads of A both occur after the load of B, they may both
804 come up with different values:
807 | | +------+ +-------+
808 | |------>| A=1 |------ --->| A->0 |
809 | | +------+ \ +-------+
810 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
811 | | +------+ | +-------+
812 | |------>| B=2 |--- | : :
813 | | +------+ \ | : : +-------+
814 +-------+ : : \ | +-------+ | |
815 ---------->| B->2 |------>| |
816 | +-------+ | CPU 2 |
820 | | A->0 |------>| 1st |
822 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
823 barrier causes all effects \ +-------+ | |
824 prior to the storage of B ---->| A->1 |------>| 2nd |
825 to be perceptible to CPU 2 +-------+ | |
829 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2
830 before the read barrier completes anyway:
833 | | +------+ +-------+
834 | |------>| A=1 |------ --->| A->0 |
835 | | +------+ \ +-------+
836 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 |
837 | | +------+ | +-------+
838 | |------>| B=2 |--- | : :
839 | | +------+ \ | : : +-------+
840 +-------+ : : \ | +-------+ | |
841 ---------->| B->2 |------>| |
842 | +-------+ | CPU 2 |
846 ---->| A->1 |------>| 1st |
848 rrrrrrrrrrrrrrrrr | |
850 | A->1 |------>| 2nd |
855 The guarantee is that the second load will always come up with A == 1 if the
856 load of B came up with B == 2. No such guarantee exists for the first load of
857 A; that may come up with either A == 0 or A == 1.
860 READ MEMORY BARRIERS VS LOAD SPECULATION
861 ----------------------------------------
863 Many CPUs speculate with loads: that is they see that they will need to load an
864 item from memory, and they find a time where they're not using the bus for any
865 other loads, and so do the load in advance - even though they haven't actually
866 got to that point in the instruction execution flow yet. This permits the
867 actual load instruction to potentially complete immediately because the CPU
868 already has the value to hand.
870 It may turn out that the CPU didn't actually need the value - perhaps because a
871 branch circumvented the load - in which case it can discard the value or just
872 cache it for later use.
877 ======================= =======================
879 DIVIDE } Divide instructions generally
880 DIVIDE } take a long time to perform
883 Which might appear as this:
887 --->| B->2 |------>| |
891 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
892 division speculates on the +-------+ ~ | |
896 Once the divisions are complete --> : : ~-->| |
897 the CPU can then perform the : : | |
898 LOAD with immediate effect : : +-------+
901 Placing a read barrier or a data dependency barrier just before the second
905 ======================= =======================
912 will force any value speculatively obtained to be reconsidered to an extent
913 dependent on the type of barrier used. If there was no change made to the
914 speculated memory location, then the speculated value will just be used:
918 --->| B->2 |------>| |
922 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
923 division speculates on the +-------+ ~ | |
928 rrrrrrrrrrrrrrrr~ | |
935 but if there was an update or an invalidation from another CPU pending, then
936 the speculation will be cancelled and the value reloaded:
940 --->| B->2 |------>| |
944 The CPU being busy doing a ---> --->| A->0 |~~~~ | |
945 division speculates on the +-------+ ~ | |
950 rrrrrrrrrrrrrrrrr | |
952 The speculation is discarded ---> --->| A->1 |------>| |
953 and an updated value is +-------+ | |
954 retrieved : : +-------+
957 ========================
958 EXPLICIT KERNEL BARRIERS
959 ========================
961 The Linux kernel has a variety of different barriers that act at different
964 (*) Compiler barrier.
966 (*) CPU memory barriers.
968 (*) MMIO write barrier.
974 The Linux kernel has an explicit compiler barrier function that prevents the
975 compiler from moving the memory accesses either side of it to the other side:
979 This is a general barrier - lesser varieties of compiler barrier do not exist.
981 The compiler barrier has no direct effect on the CPU, which may then reorder
982 things however it wishes.
988 The Linux kernel has eight basic CPU memory barriers:
990 TYPE MANDATORY SMP CONDITIONAL
991 =============== ======================= ===========================
992 GENERAL mb() smp_mb()
993 WRITE wmb() smp_wmb()
995 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
998 All memory barriers except the data dependency barriers imply a compiler
999 barrier. Data dependencies do not impose any additional compiler ordering.
1001 Aside: In the case of data dependencies, the compiler would be expected to
1002 issue the loads in the correct order (eg. `a[b]` would have to load the value
1003 of b before loading a[b]), however there is no guarantee in the C specification
1004 that the compiler may not speculate the value of b (eg. is equal to 1) and load
1005 a before b (eg. tmp = a[1]; if (b != 1) tmp = a[b]; ). There is also the
1006 problem of a compiler reloading b after having loaded a[b], thus having a newer
1007 copy of b than a[b]. A consensus has not yet been reached about these problems,
1008 however the ACCESS_ONCE macro is a good place to start looking.
1010 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
1011 systems because it is assumed that a CPU will appear to be self-consistent,
1012 and will order overlapping accesses correctly with respect to itself.
1014 [!] Note that SMP memory barriers _must_ be used to control the ordering of
1015 references to shared memory on SMP systems, though the use of locking instead
1018 Mandatory barriers should not be used to control SMP effects, since mandatory
1019 barriers unnecessarily impose overhead on UP systems. They may, however, be
1020 used to control MMIO effects on accesses through relaxed memory I/O windows.
1021 These are required even on non-SMP systems as they affect the order in which
1022 memory operations appear to a device by prohibiting both the compiler and the
1023 CPU from reordering them.
1026 There are some more advanced barrier functions:
1028 (*) set_mb(var, value)
1030 This assigns the value to the variable and then inserts a full memory
1031 barrier after it, depending on the function. It isn't guaranteed to
1032 insert anything more than a compiler barrier in a UP compilation.
1035 (*) smp_mb__before_atomic_dec();
1036 (*) smp_mb__after_atomic_dec();
1037 (*) smp_mb__before_atomic_inc();
1038 (*) smp_mb__after_atomic_inc();
1040 These are for use with atomic add, subtract, increment and decrement
1041 functions that don't return a value, especially when used for reference
1042 counting. These functions do not imply memory barriers.
1044 As an example, consider a piece of code that marks an object as being dead
1045 and then decrements the object's reference count:
1048 smp_mb__before_atomic_dec();
1049 atomic_dec(&obj->ref_count);
1051 This makes sure that the death mark on the object is perceived to be set
1052 *before* the reference counter is decremented.
1054 See Documentation/atomic_ops.txt for more information. See the "Atomic
1055 operations" subsection for information on where to use these.
1058 (*) smp_mb__before_clear_bit(void);
1059 (*) smp_mb__after_clear_bit(void);
1061 These are for use similar to the atomic inc/dec barriers. These are
1062 typically used for bitwise unlocking operations, so care must be taken as
1063 there are no implicit memory barriers here either.
1065 Consider implementing an unlock operation of some nature by clearing a
1066 locking bit. The clear_bit() would then need to be barriered like this:
1068 smp_mb__before_clear_bit();
1071 This prevents memory operations before the clear leaking to after it. See
1072 the subsection on "Locking Functions" with reference to UNLOCK operation
1075 See Documentation/atomic_ops.txt for more information. See the "Atomic
1076 operations" subsection for information on where to use these.
1082 The Linux kernel also has a special barrier for use with memory-mapped I/O
1087 This is a variation on the mandatory write barrier that causes writes to weakly
1088 ordered I/O regions to be partially ordered. Its effects may go beyond the
1089 CPU->Hardware interface and actually affect the hardware at some level.
1091 See the subsection "Locks vs I/O accesses" for more information.
1094 ===============================
1095 IMPLICIT KERNEL MEMORY BARRIERS
1096 ===============================
1098 Some of the other functions in the linux kernel imply memory barriers, amongst
1099 which are locking and scheduling functions.
1101 This specification is a _minimum_ guarantee; any particular architecture may
1102 provide more substantial guarantees, but these may not be relied upon outside
1103 of arch specific code.
1109 The Linux kernel has a number of locking constructs:
1118 In all cases there are variants on "LOCK" operations and "UNLOCK" operations
1119 for each construct. These operations all imply certain barriers:
1121 (1) LOCK operation implication:
1123 Memory operations issued after the LOCK will be completed after the LOCK
1124 operation has completed.
1126 Memory operations issued before the LOCK may be completed after the LOCK
1127 operation has completed.
1129 (2) UNLOCK operation implication:
1131 Memory operations issued before the UNLOCK will be completed before the
1132 UNLOCK operation has completed.
1134 Memory operations issued after the UNLOCK may be completed before the
1135 UNLOCK operation has completed.
1137 (3) LOCK vs LOCK implication:
1139 All LOCK operations issued before another LOCK operation will be completed
1140 before that LOCK operation.
1142 (4) LOCK vs UNLOCK implication:
1144 All LOCK operations issued before an UNLOCK operation will be completed
1145 before the UNLOCK operation.
1147 All UNLOCK operations issued before a LOCK operation will be completed
1148 before the LOCK operation.
1150 (5) Failed conditional LOCK implication:
1152 Certain variants of the LOCK operation may fail, either due to being
1153 unable to get the lock immediately, or due to receiving an unblocked
1154 signal whilst asleep waiting for the lock to become available. Failed
1155 locks do not imply any sort of barrier.
1157 Therefore, from (1), (2) and (4) an UNLOCK followed by an unconditional LOCK is
1158 equivalent to a full barrier, but a LOCK followed by an UNLOCK is not.
1160 [!] Note: one of the consequences of LOCKs and UNLOCKs being only one-way
1161 barriers is that the effects of instructions outside of a critical section
1162 may seep into the inside of the critical section.
1164 A LOCK followed by an UNLOCK may not be assumed to be full memory barrier
1165 because it is possible for an access preceding the LOCK to happen after the
1166 LOCK, and an access following the UNLOCK to happen before the UNLOCK, and the
1167 two accesses can themselves then cross:
1176 LOCK, STORE *B, STORE *A, UNLOCK
1178 Locks and semaphores may not provide any guarantee of ordering on UP compiled
1179 systems, and so cannot be counted on in such a situation to actually achieve
1180 anything at all - especially with respect to I/O accesses - unless combined
1181 with interrupt disabling operations.
1183 See also the section on "Inter-CPU locking barrier effects".
1186 As an example, consider the following:
1197 The following sequence of events is acceptable:
1199 LOCK, {*F,*A}, *E, {*C,*D}, *B, UNLOCK
1201 [+] Note that {*F,*A} indicates a combined access.
1203 But none of the following are:
1205 {*F,*A}, *B, LOCK, *C, *D, UNLOCK, *E
1206 *A, *B, *C, LOCK, *D, UNLOCK, *E, *F
1207 *A, *B, LOCK, *C, UNLOCK, *D, *E, *F
1208 *B, LOCK, *C, *D, UNLOCK, {*F,*A}, *E
1212 INTERRUPT DISABLING FUNCTIONS
1213 -----------------------------
1215 Functions that disable interrupts (LOCK equivalent) and enable interrupts
1216 (UNLOCK equivalent) will act as compiler barriers only. So if memory or I/O
1217 barriers are required in such a situation, they must be provided from some
1221 SLEEP AND WAKE-UP FUNCTIONS
1222 ---------------------------
1224 Sleeping and waking on an event flagged in global data can be viewed as an
1225 interaction between two pieces of data: the task state of the task waiting for
1226 the event and the global data used to indicate the event. To make sure that
1227 these appear to happen in the right order, the primitives to begin the process
1228 of going to sleep, and the primitives to initiate a wake up imply certain
1231 Firstly, the sleeper normally follows something like this sequence of events:
1234 set_current_state(TASK_UNINTERRUPTIBLE);
1235 if (event_indicated)
1240 A general memory barrier is interpolated automatically by set_current_state()
1241 after it has altered the task state:
1244 ===============================
1245 set_current_state();
1247 STORE current->state
1249 LOAD event_indicated
1251 set_current_state() may be wrapped by:
1254 prepare_to_wait_exclusive();
1256 which therefore also imply a general memory barrier after setting the state.
1257 The whole sequence above is available in various canned forms, all of which
1258 interpolate the memory barrier in the right place:
1261 wait_event_interruptible();
1262 wait_event_interruptible_exclusive();
1263 wait_event_interruptible_timeout();
1264 wait_event_killable();
1265 wait_event_timeout();
1270 Secondly, code that performs a wake up normally follows something like this:
1272 event_indicated = 1;
1273 wake_up(&event_wait_queue);
1277 event_indicated = 1;
1278 wake_up_process(event_daemon);
1280 A write memory barrier is implied by wake_up() and co. if and only if they wake
1281 something up. The barrier occurs before the task state is cleared, and so sits
1282 between the STORE to indicate the event and the STORE to set TASK_RUNNING:
1285 =============================== ===============================
1286 set_current_state(); STORE event_indicated
1287 set_mb(); wake_up();
1288 STORE current->state <write barrier>
1289 <general barrier> STORE current->state
1290 LOAD event_indicated
1292 The available waker functions include:
1298 wake_up_interruptible();
1299 wake_up_interruptible_all();
1300 wake_up_interruptible_nr();
1301 wake_up_interruptible_poll();
1302 wake_up_interruptible_sync();
1303 wake_up_interruptible_sync_poll();
1305 wake_up_locked_poll();
1311 [!] Note that the memory barriers implied by the sleeper and the waker do _not_
1312 order multiple stores before the wake-up with respect to loads of those stored
1313 values after the sleeper has called set_current_state(). For instance, if the
1316 set_current_state(TASK_INTERRUPTIBLE);
1317 if (event_indicated)
1319 __set_current_state(TASK_RUNNING);
1320 do_something(my_data);
1325 event_indicated = 1;
1326 wake_up(&event_wait_queue);
1328 there's no guarantee that the change to event_indicated will be perceived by
1329 the sleeper as coming after the change to my_data. In such a circumstance, the
1330 code on both sides must interpolate its own memory barriers between the
1331 separate data accesses. Thus the above sleeper ought to do:
1333 set_current_state(TASK_INTERRUPTIBLE);
1334 if (event_indicated) {
1336 do_something(my_data);
1339 and the waker should do:
1343 event_indicated = 1;
1344 wake_up(&event_wait_queue);
1347 MISCELLANEOUS FUNCTIONS
1348 -----------------------
1350 Other functions that imply barriers:
1352 (*) schedule() and similar imply full memory barriers.
1355 =================================
1356 INTER-CPU LOCKING BARRIER EFFECTS
1357 =================================
1359 On SMP systems locking primitives give a more substantial form of barrier: one
1360 that does affect memory access ordering on other CPUs, within the context of
1361 conflict on any particular lock.
1364 LOCKS VS MEMORY ACCESSES
1365 ------------------------
1367 Consider the following: the system has a pair of spinlocks (M) and (Q), and
1368 three CPUs; then should the following sequence of events occur:
1371 =============================== ===============================
1379 Then there is no guarantee as to what order CPU 3 will see the accesses to *A
1380 through *H occur in, other than the constraints imposed by the separate locks
1381 on the separate CPUs. It might, for example, see:
1383 *E, LOCK M, LOCK Q, *G, *C, *F, *A, *B, UNLOCK Q, *D, *H, UNLOCK M
1385 But it won't see any of:
1387 *B, *C or *D preceding LOCK M
1388 *A, *B or *C following UNLOCK M
1389 *F, *G or *H preceding LOCK Q
1390 *E, *F or *G following UNLOCK Q
1393 However, if the following occurs:
1396 =============================== ===============================
1411 *E, LOCK M [1], *C, *B, *A, UNLOCK M [1],
1412 LOCK M [2], *H, *F, *G, UNLOCK M [2], *D
1414 But assuming CPU 1 gets the lock first, CPU 3 won't see any of:
1416 *B, *C, *D, *F, *G or *H preceding LOCK M [1]
1417 *A, *B or *C following UNLOCK M [1]
1418 *F, *G or *H preceding LOCK M [2]
1419 *A, *B, *C, *E, *F or *G following UNLOCK M [2]
1422 LOCKS VS I/O ACCESSES
1423 ---------------------
1425 Under certain circumstances (especially involving NUMA), I/O accesses within
1426 two spinlocked sections on two different CPUs may be seen as interleaved by the
1427 PCI bridge, because the PCI bridge does not necessarily participate in the
1428 cache-coherence protocol, and is therefore incapable of issuing the required
1429 read memory barriers.
1434 =============================== ===============================
1444 may be seen by the PCI bridge as follows:
1446 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
1448 which would probably cause the hardware to malfunction.
1451 What is necessary here is to intervene with an mmiowb() before dropping the
1452 spinlock, for example:
1455 =============================== ===============================
1467 this will ensure that the two stores issued on CPU 1 appear at the PCI bridge
1468 before either of the stores issued on CPU 2.
1471 Furthermore, following a store by a load from the same device obviates the need
1472 for the mmiowb(), because the load forces the store to complete before the load
1476 =============================== ===============================
1487 See Documentation/DocBook/deviceiobook.tmpl for more information.
1490 =================================
1491 WHERE ARE MEMORY BARRIERS NEEDED?
1492 =================================
1494 Under normal operation, memory operation reordering is generally not going to
1495 be a problem as a single-threaded linear piece of code will still appear to
1496 work correctly, even if it's in an SMP kernel. There are, however, four
1497 circumstances in which reordering definitely _could_ be a problem:
1499 (*) Interprocessor interaction.
1501 (*) Atomic operations.
1503 (*) Accessing devices.
1508 INTERPROCESSOR INTERACTION
1509 --------------------------
1511 When there's a system with more than one processor, more than one CPU in the
1512 system may be working on the same data set at the same time. This can cause
1513 synchronisation problems, and the usual way of dealing with them is to use
1514 locks. Locks, however, are quite expensive, and so it may be preferable to
1515 operate without the use of a lock if at all possible. In such a case
1516 operations that affect both CPUs may have to be carefully ordered to prevent
1519 Consider, for example, the R/W semaphore slow path. Here a waiting process is
1520 queued on the semaphore, by virtue of it having a piece of its stack linked to
1521 the semaphore's list of waiting processes:
1523 struct rw_semaphore {
1526 struct list_head waiters;
1529 struct rwsem_waiter {
1530 struct list_head list;
1531 struct task_struct *task;
1534 To wake up a particular waiter, the up_read() or up_write() functions have to:
1536 (1) read the next pointer from this waiter's record to know as to where the
1537 next waiter record is;
1539 (2) read the pointer to the waiter's task structure;
1541 (3) clear the task pointer to tell the waiter it has been given the semaphore;
1543 (4) call wake_up_process() on the task; and
1545 (5) release the reference held on the waiter's task struct.
1547 In other words, it has to perform this sequence of events:
1549 LOAD waiter->list.next;
1555 and if any of these steps occur out of order, then the whole thing may
1558 Once it has queued itself and dropped the semaphore lock, the waiter does not
1559 get the lock again; it instead just waits for its task pointer to be cleared
1560 before proceeding. Since the record is on the waiter's stack, this means that
1561 if the task pointer is cleared _before_ the next pointer in the list is read,
1562 another CPU might start processing the waiter and might clobber the waiter's
1563 stack before the up*() function has a chance to read the next pointer.
1565 Consider then what might happen to the above sequence of events:
1568 =============================== ===============================
1575 Woken up by other event
1580 foo() clobbers *waiter
1582 LOAD waiter->list.next;
1585 This could be dealt with using the semaphore lock, but then the down_xxx()
1586 function has to needlessly get the spinlock again after being woken up.
1588 The way to deal with this is to insert a general SMP memory barrier:
1590 LOAD waiter->list.next;
1597 In this case, the barrier makes a guarantee that all memory accesses before the
1598 barrier will appear to happen before all the memory accesses after the barrier
1599 with respect to the other CPUs on the system. It does _not_ guarantee that all
1600 the memory accesses before the barrier will be complete by the time the barrier
1601 instruction itself is complete.
1603 On a UP system - where this wouldn't be a problem - the smp_mb() is just a
1604 compiler barrier, thus making sure the compiler emits the instructions in the
1605 right order without actually intervening in the CPU. Since there's only one
1606 CPU, that CPU's dependency ordering logic will take care of everything else.
1612 Whilst they are technically interprocessor interaction considerations, atomic
1613 operations are noted specially as some of them imply full memory barriers and
1614 some don't, but they're very heavily relied on as a group throughout the
1617 Any atomic operation that modifies some state in memory and returns information
1618 about the state (old or new) implies an SMP-conditional general memory barrier
1619 (smp_mb()) on each side of the actual operation (with the exception of
1620 explicit lock operations, described later). These include:
1625 atomic_inc_return();
1626 atomic_dec_return();
1627 atomic_add_return();
1628 atomic_sub_return();
1629 atomic_inc_and_test();
1630 atomic_dec_and_test();
1631 atomic_sub_and_test();
1632 atomic_add_negative();
1633 atomic_add_unless(); /* when succeeds (returns 1) */
1635 test_and_clear_bit();
1636 test_and_change_bit();
1638 These are used for such things as implementing LOCK-class and UNLOCK-class
1639 operations and adjusting reference counters towards object destruction, and as
1640 such the implicit memory barrier effects are necessary.
1643 The following operations are potential problems as they do _not_ imply memory
1644 barriers, but might be used for implementing such things as UNLOCK-class
1652 With these the appropriate explicit memory barrier should be used if necessary
1653 (smp_mb__before_clear_bit() for instance).
1656 The following also do _not_ imply memory barriers, and so may require explicit
1657 memory barriers under some circumstances (smp_mb__before_atomic_dec() for
1665 If they're used for statistics generation, then they probably don't need memory
1666 barriers, unless there's a coupling between statistical data.
1668 If they're used for reference counting on an object to control its lifetime,
1669 they probably don't need memory barriers because either the reference count
1670 will be adjusted inside a locked section, or the caller will already hold
1671 sufficient references to make the lock, and thus a memory barrier unnecessary.
1673 If they're used for constructing a lock of some description, then they probably
1674 do need memory barriers as a lock primitive generally has to do things in a
1677 Basically, each usage case has to be carefully considered as to whether memory
1678 barriers are needed or not.
1680 The following operations are special locking primitives:
1682 test_and_set_bit_lock();
1684 __clear_bit_unlock();
1686 These implement LOCK-class and UNLOCK-class operations. These should be used in
1687 preference to other operations when implementing locking primitives, because
1688 their implementations can be optimised on many architectures.
1690 [!] Note that special memory barrier primitives are available for these
1691 situations because on some CPUs the atomic instructions used imply full memory
1692 barriers, and so barrier instructions are superfluous in conjunction with them,
1693 and in such cases the special barrier primitives will be no-ops.
1695 See Documentation/atomic_ops.txt for more information.
1701 Many devices can be memory mapped, and so appear to the CPU as if they're just
1702 a set of memory locations. To control such a device, the driver usually has to
1703 make the right memory accesses in exactly the right order.
1705 However, having a clever CPU or a clever compiler creates a potential problem
1706 in that the carefully sequenced accesses in the driver code won't reach the
1707 device in the requisite order if the CPU or the compiler thinks it is more
1708 efficient to reorder, combine or merge accesses - something that would cause
1709 the device to malfunction.
1711 Inside of the Linux kernel, I/O should be done through the appropriate accessor
1712 routines - such as inb() or writel() - which know how to make such accesses
1713 appropriately sequential. Whilst this, for the most part, renders the explicit
1714 use of memory barriers unnecessary, there are a couple of situations where they
1717 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
1718 so for _all_ general drivers locks should be used and mmiowb() must be
1719 issued prior to unlocking the critical section.
1721 (2) If the accessor functions are used to refer to an I/O memory window with
1722 relaxed memory access properties, then _mandatory_ memory barriers are
1723 required to enforce ordering.
1725 See Documentation/DocBook/deviceiobook.tmpl for more information.
1731 A driver may be interrupted by its own interrupt service routine, and thus the
1732 two parts of the driver may interfere with each other's attempts to control or
1735 This may be alleviated - at least in part - by disabling local interrupts (a
1736 form of locking), such that the critical operations are all contained within
1737 the interrupt-disabled section in the driver. Whilst the driver's interrupt
1738 routine is executing, the driver's core may not run on the same CPU, and its
1739 interrupt is not permitted to happen again until the current interrupt has been
1740 handled, thus the interrupt handler does not need to lock against that.
1742 However, consider a driver that was talking to an ethernet card that sports an
1743 address register and a data register. If that driver's core talks to the card
1744 under interrupt-disablement and then the driver's interrupt handler is invoked:
1755 The store to the data register might happen after the second store to the
1756 address register if ordering rules are sufficiently relaxed:
1758 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
1761 If ordering rules are relaxed, it must be assumed that accesses done inside an
1762 interrupt disabled section may leak outside of it and may interleave with
1763 accesses performed in an interrupt - and vice versa - unless implicit or
1764 explicit barriers are used.
1766 Normally this won't be a problem because the I/O accesses done inside such
1767 sections will include synchronous load operations on strictly ordered I/O
1768 registers that form implicit I/O barriers. If this isn't sufficient then an
1769 mmiowb() may need to be used explicitly.
1772 A similar situation may occur between an interrupt routine and two routines
1773 running on separate CPUs that communicate with each other. If such a case is
1774 likely, then interrupt-disabling locks should be used to guarantee ordering.
1777 ==========================
1778 KERNEL I/O BARRIER EFFECTS
1779 ==========================
1781 When accessing I/O memory, drivers should use the appropriate accessor
1786 These are intended to talk to I/O space rather than memory space, but
1787 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
1788 indeed have special I/O space access cycles and instructions, but many
1789 CPUs don't have such a concept.
1791 The PCI bus, amongst others, defines an I/O space concept which - on such
1792 CPUs as i386 and x86_64 - readily maps to the CPU's concept of I/O
1793 space. However, it may also be mapped as a virtual I/O space in the CPU's
1794 memory map, particularly on those CPUs that don't support alternate I/O
1797 Accesses to this space may be fully synchronous (as on i386), but
1798 intermediary bridges (such as the PCI host bridge) may not fully honour
1801 They are guaranteed to be fully ordered with respect to each other.
1803 They are not guaranteed to be fully ordered with respect to other types of
1804 memory and I/O operation.
1806 (*) readX(), writeX():
1808 Whether these are guaranteed to be fully ordered and uncombined with
1809 respect to each other on the issuing CPU depends on the characteristics
1810 defined for the memory window through which they're accessing. On later
1811 i386 architecture machines, for example, this is controlled by way of the
1814 Ordinarily, these will be guaranteed to be fully ordered and uncombined,
1815 provided they're not accessing a prefetchable device.
1817 However, intermediary hardware (such as a PCI bridge) may indulge in
1818 deferral if it so wishes; to flush a store, a load from the same location
1819 is preferred[*], but a load from the same device or from configuration
1820 space should suffice for PCI.
1822 [*] NOTE! attempting to load from the same location as was written to may
1823 cause a malfunction - consider the 16550 Rx/Tx serial registers for
1826 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
1827 force stores to be ordered.
1829 Please refer to the PCI specification for more information on interactions
1830 between PCI transactions.
1834 These are similar to readX(), but are not guaranteed to be ordered in any
1835 way. Be aware that there is no I/O read barrier available.
1837 (*) ioreadX(), iowriteX()
1839 These will perform appropriately for the type of access they're actually
1840 doing, be it inX()/outX() or readX()/writeX().
1843 ========================================
1844 ASSUMED MINIMUM EXECUTION ORDERING MODEL
1845 ========================================
1847 It has to be assumed that the conceptual CPU is weakly-ordered but that it will
1848 maintain the appearance of program causality with respect to itself. Some CPUs
1849 (such as i386 or x86_64) are more constrained than others (such as powerpc or
1850 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
1851 of arch-specific code.
1853 This means that it must be considered that the CPU will execute its instruction
1854 stream in any order it feels like - or even in parallel - provided that if an
1855 instruction in the stream depends on an earlier instruction, then that
1856 earlier instruction must be sufficiently complete[*] before the later
1857 instruction may proceed; in other words: provided that the appearance of
1858 causality is maintained.
1860 [*] Some instructions have more than one effect - such as changing the
1861 condition codes, changing registers or changing memory - and different
1862 instructions may depend on different effects.
1864 A CPU may also discard any instruction sequence that winds up having no
1865 ultimate effect. For example, if two adjacent instructions both load an
1866 immediate value into the same register, the first may be discarded.
1869 Similarly, it has to be assumed that compiler might reorder the instruction
1870 stream in any way it sees fit, again provided the appearance of causality is
1874 ============================
1875 THE EFFECTS OF THE CPU CACHE
1876 ============================
1878 The way cached memory operations are perceived across the system is affected to
1879 a certain extent by the caches that lie between CPUs and memory, and by the
1880 memory coherence system that maintains the consistency of state in the system.
1882 As far as the way a CPU interacts with another part of the system through the
1883 caches goes, the memory system has to include the CPU's caches, and memory
1884 barriers for the most part act at the interface between the CPU and its cache
1885 (memory barriers logically act on the dotted line in the following diagram):
1887 <--- CPU ---> : <----------- Memory ----------->
1889 +--------+ +--------+ : +--------+ +-----------+
1890 | | | | : | | | | +--------+
1891 | CPU | | Memory | : | CPU | | | | |
1892 | Core |--->| Access |----->| Cache |<-->| | | |
1893 | | | Queue | : | | | |--->| Memory |
1894 | | | | : | | | | | |
1895 +--------+ +--------+ : +--------+ | | | |
1896 : | Cache | +--------+
1898 : | Mechanism | +--------+
1899 +--------+ +--------+ : +--------+ | | | |
1900 | | | | : | | | | | |
1901 | CPU | | Memory | : | CPU | | |--->| Device |
1902 | Core |--->| Access |----->| Cache |<-->| | | |
1903 | | | Queue | : | | | | | |
1904 | | | | : | | | | +--------+
1905 +--------+ +--------+ : +--------+ +-----------+
1909 Although any particular load or store may not actually appear outside of the
1910 CPU that issued it since it may have been satisfied within the CPU's own cache,
1911 it will still appear as if the full memory access had taken place as far as the
1912 other CPUs are concerned since the cache coherency mechanisms will migrate the
1913 cacheline over to the accessing CPU and propagate the effects upon conflict.
1915 The CPU core may execute instructions in any order it deems fit, provided the
1916 expected program causality appears to be maintained. Some of the instructions
1917 generate load and store operations which then go into the queue of memory
1918 accesses to be performed. The core may place these in the queue in any order
1919 it wishes, and continue execution until it is forced to wait for an instruction
1922 What memory barriers are concerned with is controlling the order in which
1923 accesses cross from the CPU side of things to the memory side of things, and
1924 the order in which the effects are perceived to happen by the other observers
1927 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
1928 their own loads and stores as if they had happened in program order.
1930 [!] MMIO or other device accesses may bypass the cache system. This depends on
1931 the properties of the memory window through which devices are accessed and/or
1932 the use of any special device communication instructions the CPU may have.
1938 Life isn't quite as simple as it may appear above, however: for while the
1939 caches are expected to be coherent, there's no guarantee that that coherency
1940 will be ordered. This means that whilst changes made on one CPU will
1941 eventually become visible on all CPUs, there's no guarantee that they will
1942 become apparent in the same order on those other CPUs.
1945 Consider dealing with a system that has a pair of CPUs (1 & 2), each of which
1946 has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
1951 +--------+ : +--->| Cache A |<------->| |
1952 | | : | +---------+ | |
1954 | | : | +---------+ | |
1955 +--------+ : +--->| Cache B |<------->| |
1958 : +---------+ | System |
1959 +--------+ : +--->| Cache C |<------->| |
1960 | | : | +---------+ | |
1962 | | : | +---------+ | |
1963 +--------+ : +--->| Cache D |<------->| |
1968 Imagine the system has the following properties:
1970 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
1973 (*) an even-numbered cache line may be in cache B, cache D or it may still be
1976 (*) whilst the CPU core is interrogating one cache, the other cache may be
1977 making use of the bus to access the rest of the system - perhaps to
1978 displace a dirty cacheline or to do a speculative load;
1980 (*) each cache has a queue of operations that need to be applied to that cache
1981 to maintain coherency with the rest of the system;
1983 (*) the coherency queue is not flushed by normal loads to lines already
1984 present in the cache, even though the contents of the queue may
1985 potentially affect those loads.
1987 Imagine, then, that two writes are made on the first CPU, with a write barrier
1988 between them to guarantee that they will appear to reach that CPU's caches in
1989 the requisite order:
1992 =============== =============== =======================================
1993 u == 0, v == 1 and p == &u, q == &u
1995 smp_wmb(); Make sure change to v is visible before
1997 <A:modify v=2> v is now in cache A exclusively
1999 <B:modify p=&v> p is now in cache B exclusively
2001 The write memory barrier forces the other CPUs in the system to perceive that
2002 the local CPU's caches have apparently been updated in the correct order. But
2003 now imagine that the second CPU wants to read those values:
2006 =============== =============== =======================================
2011 The above pair of reads may then fail to happen in the expected order, as the
2012 cacheline holding p may get updated in one of the second CPU's caches whilst
2013 the update to the cacheline holding v is delayed in the other of the second
2014 CPU's caches by some other cache event:
2017 =============== =============== =======================================
2018 u == 0, v == 1 and p == &u, q == &u
2021 <A:modify v=2> <C:busy>
2025 <B:modify p=&v> <D:commit p=&v>
2028 <C:read *q> Reads from v before v updated in cache
2032 Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
2033 no guarantee that, without intervention, the order of update will be the same
2034 as that committed on CPU 1.
2037 To intervene, we need to interpolate a data dependency barrier or a read
2038 barrier between the loads. This will force the cache to commit its coherency
2039 queue before processing any further requests:
2042 =============== =============== =======================================
2043 u == 0, v == 1 and p == &u, q == &u
2046 <A:modify v=2> <C:busy>
2050 <B:modify p=&v> <D:commit p=&v>
2052 smp_read_barrier_depends()
2056 <C:read *q> Reads from v after v updated in cache
2059 This sort of problem can be encountered on DEC Alpha processors as they have a
2060 split cache that improves performance by making better use of the data bus.
2061 Whilst most CPUs do imply a data dependency barrier on the read when a memory
2062 access depends on a read, not all do, so it may not be relied on.
2064 Other CPUs may also have split caches, but must coordinate between the various
2065 cachelets for normal memory accesses. The semantics of the Alpha removes the
2066 need for coordination in the absence of memory barriers.
2069 CACHE COHERENCY VS DMA
2070 ----------------------
2072 Not all systems maintain cache coherency with respect to devices doing DMA. In
2073 such cases, a device attempting DMA may obtain stale data from RAM because
2074 dirty cache lines may be resident in the caches of various CPUs, and may not
2075 have been written back to RAM yet. To deal with this, the appropriate part of
2076 the kernel must flush the overlapping bits of cache on each CPU (and maybe
2077 invalidate them as well).
2079 In addition, the data DMA'd to RAM by a device may be overwritten by dirty
2080 cache lines being written back to RAM from a CPU's cache after the device has
2081 installed its own data, or cache lines present in the CPU's cache may simply
2082 obscure the fact that RAM has been updated, until at such time as the cacheline
2083 is discarded from the CPU's cache and reloaded. To deal with this, the
2084 appropriate part of the kernel must invalidate the overlapping bits of the
2087 See Documentation/cachetlb.txt for more information on cache management.
2090 CACHE COHERENCY VS MMIO
2091 -----------------------
2093 Memory mapped I/O usually takes place through memory locations that are part of
2094 a window in the CPU's memory space that has different properties assigned than
2095 the usual RAM directed window.
2097 Amongst these properties is usually the fact that such accesses bypass the
2098 caching entirely and go directly to the device buses. This means MMIO accesses
2099 may, in effect, overtake accesses to cached memory that were emitted earlier.
2100 A memory barrier isn't sufficient in such a case, but rather the cache must be
2101 flushed between the cached memory write and the MMIO access if the two are in
2105 =========================
2106 THE THINGS CPUS GET UP TO
2107 =========================
2109 A programmer might take it for granted that the CPU will perform memory
2110 operations in exactly the order specified, so that if the CPU is, for example,
2111 given the following piece of code to execute:
2119 they would then expect that the CPU will complete the memory operation for each
2120 instruction before moving on to the next one, leading to a definite sequence of
2121 operations as seen by external observers in the system:
2123 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
2126 Reality is, of course, much messier. With many CPUs and compilers, the above
2127 assumption doesn't hold because:
2129 (*) loads are more likely to need to be completed immediately to permit
2130 execution progress, whereas stores can often be deferred without a
2133 (*) loads may be done speculatively, and the result discarded should it prove
2134 to have been unnecessary;
2136 (*) loads may be done speculatively, leading to the result having been fetched
2137 at the wrong time in the expected sequence of events;
2139 (*) the order of the memory accesses may be rearranged to promote better use
2140 of the CPU buses and caches;
2142 (*) loads and stores may be combined to improve performance when talking to
2143 memory or I/O hardware that can do batched accesses of adjacent locations,
2144 thus cutting down on transaction setup costs (memory and PCI devices may
2145 both be able to do this); and
2147 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
2148 mechanisms may alleviate this - once the store has actually hit the cache
2149 - there's no guarantee that the coherency management will be propagated in
2150 order to other CPUs.
2152 So what another CPU, say, might actually observe from the above piece of code
2155 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
2157 (Where "LOAD {*C,*D}" is a combined load)
2160 However, it is guaranteed that a CPU will be self-consistent: it will see its
2161 _own_ accesses appear to be correctly ordered, without the need for a memory
2162 barrier. For instance with the following code:
2171 and assuming no intervention by an external influence, it can be assumed that
2172 the final result will appear to be:
2174 U == the original value of *A
2179 The code above may cause the CPU to generate the full sequence of memory
2182 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
2184 in that order, but, without intervention, the sequence may have almost any
2185 combination of elements combined or discarded, provided the program's view of
2186 the world remains consistent.
2188 The compiler may also combine, discard or defer elements of the sequence before
2189 the CPU even sees them.
2200 since, without a write barrier, it can be assumed that the effect of the
2201 storage of V to *A is lost. Similarly:
2206 may, without a memory barrier, be reduced to:
2211 and the LOAD operation never appear outside of the CPU.
2214 AND THEN THERE'S THE ALPHA
2215 --------------------------
2217 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
2218 some versions of the Alpha CPU have a split data cache, permitting them to have
2219 two semantically-related cache lines updated at separate times. This is where
2220 the data dependency barrier really becomes necessary as this synchronises both
2221 caches with the memory coherence system, thus making it seem like pointer
2222 changes vs new data occur in the right order.
2224 The Alpha defines the Linux kernel's memory barrier model.
2226 See the subsection on "Cache Coherency" above.
2233 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
2235 Chapter 5.2: Physical Address Space Characteristics
2236 Chapter 5.4: Caches and Write Buffers
2237 Chapter 5.5: Data Sharing
2238 Chapter 5.6: Read/Write Ordering
2240 AMD64 Architecture Programmer's Manual Volume 2: System Programming
2241 Chapter 7.1: Memory-Access Ordering
2242 Chapter 7.4: Buffering and Combining Memory Writes
2244 IA-32 Intel Architecture Software Developer's Manual, Volume 3:
2245 System Programming Guide
2246 Chapter 7.1: Locked Atomic Operations
2247 Chapter 7.2: Memory Ordering
2248 Chapter 7.4: Serializing Instructions
2250 The SPARC Architecture Manual, Version 9
2251 Chapter 8: Memory Models
2252 Appendix D: Formal Specification of the Memory Models
2253 Appendix J: Programming with the Memory Models
2255 UltraSPARC Programmer Reference Manual
2256 Chapter 5: Memory Accesses and Cacheability
2257 Chapter 15: Sparc-V9 Memory Models
2259 UltraSPARC III Cu User's Manual
2260 Chapter 9: Memory Models
2262 UltraSPARC IIIi Processor User's Manual
2263 Chapter 8: Memory Models
2265 UltraSPARC Architecture 2005
2267 Appendix D: Formal Specifications of the Memory Models
2269 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
2270 Chapter 8: Memory Models
2271 Appendix F: Caches and Cache Coherency
2273 Solaris Internals, Core Kernel Architecture, p63-68:
2274 Chapter 3.3: Hardware Considerations for Locks and
2277 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
2278 for Kernel Programmers:
2279 Chapter 13: Other Memory Models
2281 Intel Itanium Architecture Software Developer's Manual: Volume 1:
2282 Section 2.6: Speculation
2283 Section 4.4: Memory Access
git.samba.org / sfrench / cifs-2.6.git / blob