blob: 92f0056d928c1f5ff485e250a93a843bb60058bc [file] [log] [blame]
David Howells108b42b2006-03-31 16:00:29 +01001 ============================
2 LINUX KERNEL MEMORY BARRIERS
3 ============================
4
5By: David Howells <dhowells@redhat.com>
6
7Contents:
8
9 (*) Abstract memory access model.
10
11 - Device operations.
12 - Guarantees.
13
14 (*) What are memory barriers?
15
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
23 (*) Explicit kernel barriers.
24
25 - Compiler barrier.
26 - The CPU memory barriers.
27 - MMIO write barrier.
28
29 (*) Implicit kernel memory barriers.
30
31 - Locking functions.
32 - Interrupt disabling functions.
33 - Miscellaneous functions.
34
35 (*) Inter-CPU locking barrier effects.
36
37 - Locks vs memory accesses.
38 - Locks vs I/O accesses.
39
40 (*) Where are memory barriers needed?
41
42 - Interprocessor interaction.
43 - Atomic operations.
44 - Accessing devices.
45 - Interrupts.
46
47 (*) Kernel I/O barrier effects.
48
49 (*) Assumed minimum execution ordering model.
50
51 (*) The effects of the cpu cache.
52
53 - Cache coherency.
54 - Cache coherency vs DMA.
55 - Cache coherency vs MMIO.
56
57 (*) The things CPUs get up to.
58
59 - And then there's the Alpha.
60
61 (*) References.
62
63
64============================
65ABSTRACT MEMORY ACCESS MODEL
66============================
67
68Consider the following abstract model of the system:
69
70 : :
71 : :
72 : :
73 +-------+ : +--------+ : +-------+
74 | | : | | : | |
75 | | : | | : | |
76 | CPU 1 |<----->| Memory |<----->| CPU 2 |
77 | | : | | : | |
78 | | : | | : | |
79 +-------+ : +--------+ : +-------+
80 ^ : ^ : ^
81 | : | : |
82 | : | : |
83 | : v : |
84 | : +--------+ : |
85 | : | | : |
86 | : | | : |
87 +---------->| Device |<----------+
88 : | | :
89 : | | :
90 : +--------+ :
91 : :
92
93Each CPU executes a program that generates memory access operations. In the
94abstract CPU, memory operation ordering is very relaxed, and a CPU may actually
95perform the memory operations in any order it likes, provided program causality
96appears to be maintained. Similarly, the compiler may also arrange the
97instructions it emits in any order it likes, provided it doesn't affect the
98apparent operation of the program.
99
100So in the above diagram, the effects of the memory operations performed by a
101CPU are perceived by the rest of the system as the operations cross the
102interface between the CPU and rest of the system (the dotted lines).
103
104
105For example, consider the following sequence of events:
106
107 CPU 1 CPU 2
108 =============== ===============
109 { A == 1; B == 2 }
110 A = 3; x = A;
111 B = 4; y = B;
112
113The set of accesses as seen by the memory system in the middle can be arranged
114in 24 different combinations:
115
116 STORE A=3, STORE B=4, x=LOAD A->3, y=LOAD B->4
117 STORE A=3, STORE B=4, y=LOAD B->4, x=LOAD A->3
118 STORE A=3, x=LOAD A->3, STORE B=4, y=LOAD B->4
119 STORE A=3, x=LOAD A->3, y=LOAD B->2, STORE B=4
120 STORE A=3, y=LOAD B->2, STORE B=4, x=LOAD A->3
121 STORE A=3, y=LOAD B->2, x=LOAD A->3, STORE B=4
122 STORE B=4, STORE A=3, x=LOAD A->3, y=LOAD B->4
123 STORE B=4, ...
124 ...
125
126and can thus result in four different combinations of values:
127
128 x == 1, y == 2
129 x == 1, y == 4
130 x == 3, y == 2
131 x == 3, y == 4
132
133
134Furthermore, the stores committed by a CPU to the memory system may not be
135perceived by the loads made by another CPU in the same order as the stores were
136committed.
137
138
139As a further example, consider this sequence of events:
140
141 CPU 1 CPU 2
142 =============== ===============
143 { A == 1, B == 2, C = 3, P == &A, Q == &C }
144 B = 4; Q = P;
145 P = &B D = *Q;
146
147There is an obvious data dependency here, as the value loaded into D depends on
148the address retrieved from P by CPU 2. At the end of the sequence, any of the
149following results are possible:
150
151 (Q == &A) and (D == 1)
152 (Q == &B) and (D == 2)
153 (Q == &B) and (D == 4)
154
155Note that CPU 2 will never try and load C into D because the CPU will load P
156into Q before issuing the load of *Q.
157
158
159DEVICE OPERATIONS
160-----------------
161
162Some devices present their control interfaces as collections of memory
163locations, but the order in which the control registers are accessed is very
164important. For instance, imagine an ethernet card with a set of internal
165registers that are accessed through an address port register (A) and a data
166port register (D). To read internal register 5, the following code might then
167be used:
168
169 *A = 5;
170 x = *D;
171
172but this might show up as either of the following two sequences:
173
174 STORE *A = 5, x = LOAD *D
175 x = LOAD *D, STORE *A = 5
176
177the second of which will almost certainly result in a malfunction, since it set
178the address _after_ attempting to read the register.
179
180
181GUARANTEES
182----------
183
184There are some minimal guarantees that may be expected of a CPU:
185
186 (*) On any given CPU, dependent memory accesses will be issued in order, with
187 respect to itself. This means that for:
188
189 Q = P; D = *Q;
190
191 the CPU will issue the following memory operations:
192
193 Q = LOAD P, D = LOAD *Q
194
195 and always in that order.
196
197 (*) Overlapping loads and stores within a particular CPU will appear to be
198 ordered within that CPU. This means that for:
199
200 a = *X; *X = b;
201
202 the CPU will only issue the following sequence of memory operations:
203
204 a = LOAD *X, STORE *X = b
205
206 And for:
207
208 *X = c; d = *X;
209
210 the CPU will only issue:
211
212 STORE *X = c, d = LOAD *X
213
214 (Loads and stores overlap if they are targetted at overlapping pieces of
215 memory).
216
217And there are a number of things that _must_ or _must_not_ be assumed:
218
219 (*) It _must_not_ be assumed that independent loads and stores will be issued
220 in the order given. This means that for:
221
222 X = *A; Y = *B; *D = Z;
223
224 we may get any of the following sequences:
225
226 X = LOAD *A, Y = LOAD *B, STORE *D = Z
227 X = LOAD *A, STORE *D = Z, Y = LOAD *B
228 Y = LOAD *B, X = LOAD *A, STORE *D = Z
229 Y = LOAD *B, STORE *D = Z, X = LOAD *A
230 STORE *D = Z, X = LOAD *A, Y = LOAD *B
231 STORE *D = Z, Y = LOAD *B, X = LOAD *A
232
233 (*) It _must_ be assumed that overlapping memory accesses may be merged or
234 discarded. This means that for:
235
236 X = *A; Y = *(A + 4);
237
238 we may get any one of the following sequences:
239
240 X = LOAD *A; Y = LOAD *(A + 4);
241 Y = LOAD *(A + 4); X = LOAD *A;
242 {X, Y} = LOAD {*A, *(A + 4) };
243
244 And for:
245
246 *A = X; Y = *A;
247
248 we may get either of:
249
250 STORE *A = X; Y = LOAD *A;
251 STORE *A = Y;
252
253
254=========================
255WHAT ARE MEMORY BARRIERS?
256=========================
257
258As can be seen above, independent memory operations are effectively performed
259in random order, but this can be a problem for CPU-CPU interaction and for I/O.
260What is required is some way of intervening to instruct the compiler and the
261CPU to restrict the order.
262
263Memory barriers are such interventions. They impose a perceived partial
264ordering between the memory operations specified on either side of the barrier.
265They request that the sequence of memory events generated appears to other
266parts of the system as if the barrier is effective on that CPU.
267
268
269VARIETIES OF MEMORY BARRIER
270---------------------------
271
272Memory barriers come in four basic varieties:
273
274 (1) Write (or store) memory barriers.
275
276 A write memory barrier gives a guarantee that all the STORE operations
277 specified before the barrier will appear to happen before all the STORE
278 operations specified after the barrier with respect to the other
279 components of the system.
280
281 A write barrier is a partial ordering on stores only; it is not required
282 to have any effect on loads.
283
284 A CPU can be viewed as as commiting a sequence of store operations to the
285 memory system as time progresses. All stores before a write barrier will
286 occur in the sequence _before_ all the stores after the write barrier.
287
288 [!] Note that write barriers should normally be paired with read or data
289 dependency barriers; see the "SMP barrier pairing" subsection.
290
291
292 (2) Data dependency barriers.
293
294 A data dependency barrier is a weaker form of read barrier. In the case
295 where two loads are performed such that the second depends on the result
296 of the first (eg: the first load retrieves the address to which the second
297 load will be directed), a data dependency barrier would be required to
298 make sure that the target of the second load is updated before the address
299 obtained by the first load is accessed.
300
301 A data dependency barrier is a partial ordering on interdependent loads
302 only; it is not required to have any effect on stores, independent loads
303 or overlapping loads.
304
305 As mentioned in (1), the other CPUs in the system can be viewed as
306 committing sequences of stores to the memory system that the CPU being
307 considered can then perceive. A data dependency barrier issued by the CPU
308 under consideration guarantees that for any load preceding it, if that
309 load touches one of a sequence of stores from another CPU, then by the
310 time the barrier completes, the effects of all the stores prior to that
311 touched by the load will be perceptible to any loads issued after the data
312 dependency barrier.
313
314 See the "Examples of memory barrier sequences" subsection for diagrams
315 showing the ordering constraints.
316
317 [!] Note that the first load really has to have a _data_ dependency and
318 not a control dependency. If the address for the second load is dependent
319 on the first load, but the dependency is through a conditional rather than
320 actually loading the address itself, then it's a _control_ dependency and
321 a full read barrier or better is required. See the "Control dependencies"
322 subsection for more information.
323
324 [!] Note that data dependency barriers should normally be paired with
325 write barriers; see the "SMP barrier pairing" subsection.
326
327
328 (3) Read (or load) memory barriers.
329
330 A read barrier is a data dependency barrier plus a guarantee that all the
331 LOAD operations specified before the barrier will appear to happen before
332 all the LOAD operations specified after the barrier with respect to the
333 other components of the system.
334
335 A read barrier is a partial ordering on loads only; it is not required to
336 have any effect on stores.
337
338 Read memory barriers imply data dependency barriers, and so can substitute
339 for them.
340
341 [!] Note that read barriers should normally be paired with write barriers;
342 see the "SMP barrier pairing" subsection.
343
344
345 (4) General memory barriers.
346
347 A general memory barrier is a combination of both a read memory barrier
348 and a write memory barrier. It is a partial ordering over both loads and
349 stores.
350
351 General memory barriers imply both read and write memory barriers, and so
352 can substitute for either.
353
354
355And a couple of implicit varieties:
356
357 (5) LOCK operations.
358
359 This acts as a one-way permeable barrier. It guarantees that all memory
360 operations after the LOCK operation will appear to happen after the LOCK
361 operation with respect to the other components of the system.
362
363 Memory operations that occur before a LOCK operation may appear to happen
364 after it completes.
365
366 A LOCK operation should almost always be paired with an UNLOCK operation.
367
368
369 (6) UNLOCK operations.
370
371 This also acts as a one-way permeable barrier. It guarantees that all
372 memory operations before the UNLOCK operation will appear to happen before
373 the UNLOCK operation with respect to the other components of the system.
374
375 Memory operations that occur after an UNLOCK operation may appear to
376 happen before it completes.
377
378 LOCK and UNLOCK operations are guaranteed to appear with respect to each
379 other strictly in the order specified.
380
381 The use of LOCK and UNLOCK operations generally precludes the need for
382 other sorts of memory barrier (but note the exceptions mentioned in the
383 subsection "MMIO write barrier").
384
385
386Memory barriers are only required where there's a possibility of interaction
387between two CPUs or between a CPU and a device. If it can be guaranteed that
388there won't be any such interaction in any particular piece of code, then
389memory barriers are unnecessary in that piece of code.
390
391
392Note that these are the _minimum_ guarantees. Different architectures may give
393more substantial guarantees, but they may _not_ be relied upon outside of arch
394specific code.
395
396
397WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS?
398----------------------------------------------
399
400There are certain things that the Linux kernel memory barriers do not guarantee:
401
402 (*) There is no guarantee that any of the memory accesses specified before a
403 memory barrier will be _complete_ by the completion of a memory barrier
404 instruction; the barrier can be considered to draw a line in that CPU's
405 access queue that accesses of the appropriate type may not cross.
406
407 (*) There is no guarantee that issuing a memory barrier on one CPU will have
408 any direct effect on another CPU or any other hardware in the system. The
409 indirect effect will be the order in which the second CPU sees the effects
410 of the first CPU's accesses occur, but see the next point:
411
412 (*) There is no guarantee that the a CPU will see the correct order of effects
413 from a second CPU's accesses, even _if_ the second CPU uses a memory
414 barrier, unless the first CPU _also_ uses a matching memory barrier (see
415 the subsection on "SMP Barrier Pairing").
416
417 (*) There is no guarantee that some intervening piece of off-the-CPU
418 hardware[*] will not reorder the memory accesses. CPU cache coherency
419 mechanisms should propagate the indirect effects of a memory barrier
420 between CPUs, but might not do so in order.
421
422 [*] For information on bus mastering DMA and coherency please read:
423
424 Documentation/pci.txt
425 Documentation/DMA-mapping.txt
426 Documentation/DMA-API.txt
427
428
429DATA DEPENDENCY BARRIERS
430------------------------
431
432The usage requirements of data dependency barriers are a little subtle, and
433it's not always obvious that they're needed. To illustrate, consider the
434following sequence of events:
435
436 CPU 1 CPU 2
437 =============== ===============
438 { A == 1, B == 2, C = 3, P == &A, Q == &C }
439 B = 4;
440 <write barrier>
441 P = &B
442 Q = P;
443 D = *Q;
444
445There's a clear data dependency here, and it would seem that by the end of the
446sequence, Q must be either &A or &B, and that:
447
448 (Q == &A) implies (D == 1)
449 (Q == &B) implies (D == 4)
450
451But! CPU 2's perception of P may be updated _before_ its perception of B, thus
452leading to the following situation:
453
454 (Q == &B) and (D == 2) ????
455
456Whilst this may seem like a failure of coherency or causality maintenance, it
457isn't, and this behaviour can be observed on certain real CPUs (such as the DEC
458Alpha).
459
460To deal with this, a data dependency barrier must be inserted between the
461address load and the data load:
462
463 CPU 1 CPU 2
464 =============== ===============
465 { A == 1, B == 2, C = 3, P == &A, Q == &C }
466 B = 4;
467 <write barrier>
468 P = &B
469 Q = P;
470 <data dependency barrier>
471 D = *Q;
472
473This enforces the occurrence of one of the two implications, and prevents the
474third possibility from arising.
475
476[!] Note that this extremely counterintuitive situation arises most easily on
477machines with split caches, so that, for example, one cache bank processes
478even-numbered cache lines and the other bank processes odd-numbered cache
479lines. The pointer P might be stored in an odd-numbered cache line, and the
480variable B might be stored in an even-numbered cache line. Then, if the
481even-numbered bank of the reading CPU's cache is extremely busy while the
482odd-numbered bank is idle, one can see the new value of the pointer P (&B),
483but the old value of the variable B (1).
484
485
486Another example of where data dependency barriers might by required is where a
487number is read from memory and then used to calculate the index for an array
488access:
489
490 CPU 1 CPU 2
491 =============== ===============
492 { M[0] == 1, M[1] == 2, M[3] = 3, P == 0, Q == 3 }
493 M[1] = 4;
494 <write barrier>
495 P = 1
496 Q = P;
497 <data dependency barrier>
498 D = M[Q];
499
500
501The data dependency barrier is very important to the RCU system, for example.
502See rcu_dereference() in include/linux/rcupdate.h. This permits the current
503target of an RCU'd pointer to be replaced with a new modified target, without
504the replacement target appearing to be incompletely initialised.
505
506See also the subsection on "Cache Coherency" for a more thorough example.
507
508
509CONTROL DEPENDENCIES
510--------------------
511
512A control dependency requires a full read memory barrier, not simply a data
513dependency barrier to make it work correctly. Consider the following bit of
514code:
515
516 q = &a;
517 if (p)
518 q = &b;
519 <data dependency barrier>
520 x = *q;
521
522This will not have the desired effect because there is no actual data
523dependency, but rather a control dependency that the CPU may short-circuit by
524attempting to predict the outcome in advance. In such a case what's actually
525required is:
526
527 q = &a;
528 if (p)
529 q = &b;
530 <read barrier>
531 x = *q;
532
533
534SMP BARRIER PAIRING
535-------------------
536
537When dealing with CPU-CPU interactions, certain types of memory barrier should
538always be paired. A lack of appropriate pairing is almost certainly an error.
539
540A write barrier should always be paired with a data dependency barrier or read
541barrier, though a general barrier would also be viable. Similarly a read
542barrier or a data dependency barrier should always be paired with at least an
543write barrier, though, again, a general barrier is viable:
544
545 CPU 1 CPU 2
546 =============== ===============
547 a = 1;
548 <write barrier>
549 b = 2; x = a;
550 <read barrier>
551 y = b;
552
553Or:
554
555 CPU 1 CPU 2
556 =============== ===============================
557 a = 1;
558 <write barrier>
559 b = &a; x = b;
560 <data dependency barrier>
561 y = *x;
562
563Basically, the read barrier always has to be there, even though it can be of
564the "weaker" type.
565
566
567EXAMPLES OF MEMORY BARRIER SEQUENCES
568------------------------------------
569
570Firstly, write barriers act as a partial orderings on store operations.
571Consider the following sequence of events:
572
573 CPU 1
574 =======================
575 STORE A = 1
576 STORE B = 2
577 STORE C = 3
578 <write barrier>
579 STORE D = 4
580 STORE E = 5
581
582This sequence of events is committed to the memory coherence system in an order
583that the rest of the system might perceive as the unordered set of { STORE A,
584STORE B, STORE C } all occuring before the unordered set of { STORE D, STORE E
585}:
586
587 +-------+ : :
588 | | +------+
589 | |------>| C=3 | } /\
590 | | : +------+ }----- \ -----> Events perceptible
591 | | : | A=1 | } \/ to rest of system
592 | | : +------+ }
593 | CPU 1 | : | B=2 | }
594 | | +------+ }
595 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier
596 | | +------+ } requires all stores prior to the
597 | | : | E=5 | } barrier to be committed before
598 | | : +------+ } further stores may be take place.
599 | |------>| D=4 | }
600 | | +------+
601 +-------+ : :
602 |
603 | Sequence in which stores committed to memory system
604 | by CPU 1
605 V
606
607
608Secondly, data dependency barriers act as a partial orderings on data-dependent
609loads. Consider the following sequence of events:
610
611 CPU 1 CPU 2
612 ======================= =======================
David Howellsc14038c2006-04-10 22:54:24 -0700613 { B = 7; X = 9; Y = 8; C = &Y }
David Howells108b42b2006-03-31 16:00:29 +0100614 STORE A = 1
615 STORE B = 2
616 <write barrier>
617 STORE C = &B LOAD X
618 STORE D = 4 LOAD C (gets &B)
619 LOAD *C (reads B)
620
621Without intervention, CPU 2 may perceive the events on CPU 1 in some
622effectively random order, despite the write barrier issued by CPU 1:
623
624 +-------+ : : : :
625 | | +------+ +-------+ | Sequence of update
626 | |------>| B=2 |----- --->| Y->8 | | of perception on
627 | | : +------+ \ +-------+ | CPU 2
628 | CPU 1 | : | A=1 | \ --->| C->&Y | V
629 | | +------+ | +-------+
630 | | wwwwwwwwwwwwwwww | : :
631 | | +------+ | : :
632 | | : | C=&B |--- | : : +-------+
633 | | : +------+ \ | +-------+ | |
634 | |------>| D=4 | ----------->| C->&B |------>| |
635 | | +------+ | +-------+ | |
636 +-------+ : : | : : | |
637 | : : | |
638 | : : | CPU 2 |
639 | +-------+ | |
640 Apparently incorrect ---> | | B->7 |------>| |
641 perception of B (!) | +-------+ | |
642 | : : | |
643 | +-------+ | |
644 The load of X holds ---> \ | X->9 |------>| |
645 up the maintenance \ +-------+ | |
646 of coherence of B ----->| B->2 | +-------+
647 +-------+
648 : :
649
650
651In the above example, CPU 2 perceives that B is 7, despite the load of *C
652(which would be B) coming after the the LOAD of C.
653
654If, however, a data dependency barrier were to be placed between the load of C
David Howellsc14038c2006-04-10 22:54:24 -0700655and the load of *C (ie: B) on CPU 2:
656
657 CPU 1 CPU 2
658 ======================= =======================
659 { B = 7; X = 9; Y = 8; C = &Y }
660 STORE A = 1
661 STORE B = 2
662 <write barrier>
663 STORE C = &B LOAD X
664 STORE D = 4 LOAD C (gets &B)
665 <data dependency barrier>
666 LOAD *C (reads B)
667
668then the following will occur:
David Howells108b42b2006-03-31 16:00:29 +0100669
670 +-------+ : : : :
671 | | +------+ +-------+
672 | |------>| B=2 |----- --->| Y->8 |
673 | | : +------+ \ +-------+
674 | CPU 1 | : | A=1 | \ --->| C->&Y |
675 | | +------+ | +-------+
676 | | wwwwwwwwwwwwwwww | : :
677 | | +------+ | : :
678 | | : | C=&B |--- | : : +-------+
679 | | : +------+ \ | +-------+ | |
680 | |------>| D=4 | ----------->| C->&B |------>| |
681 | | +------+ | +-------+ | |
682 +-------+ : : | : : | |
683 | : : | |
684 | : : | CPU 2 |
685 | +-------+ | |
686 \ | X->9 |------>| |
687 \ +-------+ | |
688 ----->| B->2 | | |
689 +-------+ | |
690 Makes sure all effects ---> ddddddddddddddddd | |
691 prior to the store of C +-------+ | |
692 are perceptible to | B->2 |------>| |
693 successive loads +-------+ | |
694 : : +-------+
695
696
697And thirdly, a read barrier acts as a partial order on loads. Consider the
698following sequence of events:
699
700 CPU 1 CPU 2
701 ======================= =======================
702 STORE A=1
703 STORE B=2
704 STORE C=3
705 <write barrier>
706 STORE D=4
707 STORE E=5
708 LOAD A
709 LOAD B
710 LOAD C
711 LOAD D
712 LOAD E
713
714Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in
715some effectively random order, despite the write barrier issued by CPU 1:
716
717 +-------+ : :
718 | | +------+
719 | |------>| C=3 | }
720 | | : +------+ }
721 | | : | A=1 | }
722 | | : +------+ }
723 | CPU 1 | : | B=2 | }---
724 | | +------+ } \
725 | | wwwwwwwwwwwww} \
726 | | +------+ } \ : : +-------+
727 | | : | E=5 | } \ +-------+ | |
728 | | : +------+ } \ { | C->3 |------>| |
729 | |------>| D=4 | } \ { +-------+ : | |
730 | | +------+ \ { | E->5 | : | |
731 +-------+ : : \ { +-------+ : | |
732 Transfer -->{ | A->1 | : | CPU 2 |
733 from CPU 1 { +-------+ : | |
734 to CPU 2 { | D->4 | : | |
735 { +-------+ : | |
736 { | B->2 |------>| |
737 +-------+ | |
738 : : +-------+
739
740
741If, however, a read barrier were to be placed between the load of C and the
742load of D on CPU 2, then the partial ordering imposed by CPU 1 will be
743perceived correctly by CPU 2.
744
745 +-------+ : :
746 | | +------+
747 | |------>| C=3 | }
748 | | : +------+ }
749 | | : | A=1 | }---
750 | | : +------+ } \
751 | CPU 1 | : | B=2 | } \
752 | | +------+ \
753 | | wwwwwwwwwwwwwwww \
754 | | +------+ \ : : +-------+
755 | | : | E=5 | } \ +-------+ | |
756 | | : +------+ }--- \ { | C->3 |------>| |
757 | |------>| D=4 | } \ \ { +-------+ : | |
758 | | +------+ \ -->{ | B->2 | : | |
759 +-------+ : : \ { +-------+ : | |
760 \ { | A->1 | : | CPU 2 |
761 \ +-------+ | |
762 At this point the read ----> \ rrrrrrrrrrrrrrrrr | |
763 barrier causes all effects \ +-------+ | |
764 prior to the storage of C \ { | E->5 | : | |
765 to be perceptible to CPU 2 -->{ +-------+ : | |
766 { | D->4 |------>| |
767 +-------+ | |
768 : : +-------+
769
770
771========================
772EXPLICIT KERNEL BARRIERS
773========================
774
775The Linux kernel has a variety of different barriers that act at different
776levels:
777
778 (*) Compiler barrier.
779
780 (*) CPU memory barriers.
781
782 (*) MMIO write barrier.
783
784
785COMPILER BARRIER
786----------------
787
788The Linux kernel has an explicit compiler barrier function that prevents the
789compiler from moving the memory accesses either side of it to the other side:
790
791 barrier();
792
793This a general barrier - lesser varieties of compiler barrier do not exist.
794
795The compiler barrier has no direct effect on the CPU, which may then reorder
796things however it wishes.
797
798
799CPU MEMORY BARRIERS
800-------------------
801
802The Linux kernel has eight basic CPU memory barriers:
803
804 TYPE MANDATORY SMP CONDITIONAL
805 =============== ======================= ===========================
806 GENERAL mb() smp_mb()
807 WRITE wmb() smp_wmb()
808 READ rmb() smp_rmb()
809 DATA DEPENDENCY read_barrier_depends() smp_read_barrier_depends()
810
811
812All CPU memory barriers unconditionally imply compiler barriers.
813
814SMP memory barriers are reduced to compiler barriers on uniprocessor compiled
815systems because it is assumed that a CPU will be appear to be self-consistent,
816and will order overlapping accesses correctly with respect to itself.
817
818[!] Note that SMP memory barriers _must_ be used to control the ordering of
819references to shared memory on SMP systems, though the use of locking instead
820is sufficient.
821
822Mandatory barriers should not be used to control SMP effects, since mandatory
823barriers unnecessarily impose overhead on UP systems. They may, however, be
824used to control MMIO effects on accesses through relaxed memory I/O windows.
825These are required even on non-SMP systems as they affect the order in which
826memory operations appear to a device by prohibiting both the compiler and the
827CPU from reordering them.
828
829
830There are some more advanced barrier functions:
831
832 (*) set_mb(var, value)
833 (*) set_wmb(var, value)
834
835 These assign the value to the variable and then insert at least a write
836 barrier after it, depending on the function. They aren't guaranteed to
837 insert anything more than a compiler barrier in a UP compilation.
838
839
840 (*) smp_mb__before_atomic_dec();
841 (*) smp_mb__after_atomic_dec();
842 (*) smp_mb__before_atomic_inc();
843 (*) smp_mb__after_atomic_inc();
844
845 These are for use with atomic add, subtract, increment and decrement
David Howellsdbc87002006-04-10 22:54:23 -0700846 functions that don't return a value, especially when used for reference
847 counting. These functions do not imply memory barriers.
David Howells108b42b2006-03-31 16:00:29 +0100848
849 As an example, consider a piece of code that marks an object as being dead
850 and then decrements the object's reference count:
851
852 obj->dead = 1;
853 smp_mb__before_atomic_dec();
854 atomic_dec(&obj->ref_count);
855
856 This makes sure that the death mark on the object is perceived to be set
857 *before* the reference counter is decremented.
858
859 See Documentation/atomic_ops.txt for more information. See the "Atomic
860 operations" subsection for information on where to use these.
861
862
863 (*) smp_mb__before_clear_bit(void);
864 (*) smp_mb__after_clear_bit(void);
865
866 These are for use similar to the atomic inc/dec barriers. These are
867 typically used for bitwise unlocking operations, so care must be taken as
868 there are no implicit memory barriers here either.
869
870 Consider implementing an unlock operation of some nature by clearing a
871 locking bit. The clear_bit() would then need to be barriered like this:
872
873 smp_mb__before_clear_bit();
874 clear_bit( ... );
875
876 This prevents memory operations before the clear leaking to after it. See
877 the subsection on "Locking Functions" with reference to UNLOCK operation
878 implications.
879
880 See Documentation/atomic_ops.txt for more information. See the "Atomic
881 operations" subsection for information on where to use these.
882
883
884MMIO WRITE BARRIER
885------------------
886
887The Linux kernel also has a special barrier for use with memory-mapped I/O
888writes:
889
890 mmiowb();
891
892This is a variation on the mandatory write barrier that causes writes to weakly
893ordered I/O regions to be partially ordered. Its effects may go beyond the
894CPU->Hardware interface and actually affect the hardware at some level.
895
896See the subsection "Locks vs I/O accesses" for more information.
897
898
899===============================
900IMPLICIT KERNEL MEMORY BARRIERS
901===============================
902
903Some of the other functions in the linux kernel imply memory barriers, amongst
904which are locking, scheduling and memory allocation functions.
905
906This specification is a _minimum_ guarantee; any particular architecture may
907provide more substantial guarantees, but these may not be relied upon outside
908of arch specific code.
909
910
911LOCKING FUNCTIONS
912-----------------
913
914The Linux kernel has a number of locking constructs:
915
916 (*) spin locks
917 (*) R/W spin locks
918 (*) mutexes
919 (*) semaphores
920 (*) R/W semaphores
921 (*) RCU
922
923In all cases there are variants on "LOCK" operations and "UNLOCK" operations
924for each construct. These operations all imply certain barriers:
925
926 (1) LOCK operation implication:
927
928 Memory operations issued after the LOCK will be completed after the LOCK
929 operation has completed.
930
931 Memory operations issued before the LOCK may be completed after the LOCK
932 operation has completed.
933
934 (2) UNLOCK operation implication:
935
936 Memory operations issued before the UNLOCK will be completed before the
937 UNLOCK operation has completed.
938
939 Memory operations issued after the UNLOCK may be completed before the
940 UNLOCK operation has completed.
941
942 (3) LOCK vs LOCK implication:
943
944 All LOCK operations issued before another LOCK operation will be completed
945 before that LOCK operation.
946
947 (4) LOCK vs UNLOCK implication:
948
949 All LOCK operations issued before an UNLOCK operation will be completed
950 before the UNLOCK operation.
951
952 All UNLOCK operations issued before a LOCK operation will be completed
953 before the LOCK operation.
954
955 (5) Failed conditional LOCK implication:
956
957 Certain variants of the LOCK operation may fail, either due to being
958 unable to get the lock immediately, or due to receiving an unblocked
959 signal whilst asleep waiting for the lock to become available. Failed
960 locks do not imply any sort of barrier.
961
962Therefore, from (1), (2) and (4) an UNLOCK followed by an unconditional LOCK is
963equivalent to a full barrier, but a LOCK followed by an UNLOCK is not.
964
965[!] Note: one of the consequence of LOCKs and UNLOCKs being only one-way
966 barriers is that the effects instructions outside of a critical section may
967 seep into the inside of the critical section.
968
969Locks and semaphores may not provide any guarantee of ordering on UP compiled
970systems, and so cannot be counted on in such a situation to actually achieve
971anything at all - especially with respect to I/O accesses - unless combined
972with interrupt disabling operations.
973
974See also the section on "Inter-CPU locking barrier effects".
975
976
977As an example, consider the following:
978
979 *A = a;
980 *B = b;
981 LOCK
982 *C = c;
983 *D = d;
984 UNLOCK
985 *E = e;
986 *F = f;
987
988The following sequence of events is acceptable:
989
990 LOCK, {*F,*A}, *E, {*C,*D}, *B, UNLOCK
991
992 [+] Note that {*F,*A} indicates a combined access.
993
994But none of the following are:
995
996 {*F,*A}, *B, LOCK, *C, *D, UNLOCK, *E
997 *A, *B, *C, LOCK, *D, UNLOCK, *E, *F
998 *A, *B, LOCK, *C, UNLOCK, *D, *E, *F
999 *B, LOCK, *C, *D, UNLOCK, {*F,*A}, *E
1000
1001
1002
1003INTERRUPT DISABLING FUNCTIONS
1004-----------------------------
1005
1006Functions that disable interrupts (LOCK equivalent) and enable interrupts
1007(UNLOCK equivalent) will act as compiler barriers only. So if memory or I/O
1008barriers are required in such a situation, they must be provided from some
1009other means.
1010
1011
1012MISCELLANEOUS FUNCTIONS
1013-----------------------
1014
1015Other functions that imply barriers:
1016
1017 (*) schedule() and similar imply full memory barriers.
1018
1019 (*) Memory allocation and release functions imply full memory barriers.
1020
1021
1022=================================
1023INTER-CPU LOCKING BARRIER EFFECTS
1024=================================
1025
1026On SMP systems locking primitives give a more substantial form of barrier: one
1027that does affect memory access ordering on other CPUs, within the context of
1028conflict on any particular lock.
1029
1030
1031LOCKS VS MEMORY ACCESSES
1032------------------------
1033
1034Consider the following: the system has a pair of spinlocks (N) and (Q), and
1035three CPUs; then should the following sequence of events occur:
1036
1037 CPU 1 CPU 2
1038 =============================== ===============================
1039 *A = a; *E = e;
1040 LOCK M LOCK Q
1041 *B = b; *F = f;
1042 *C = c; *G = g;
1043 UNLOCK M UNLOCK Q
1044 *D = d; *H = h;
1045
1046Then there is no guarantee as to what order CPU #3 will see the accesses to *A
1047through *H occur in, other than the constraints imposed by the separate locks
1048on the separate CPUs. It might, for example, see:
1049
1050 *E, LOCK M, LOCK Q, *G, *C, *F, *A, *B, UNLOCK Q, *D, *H, UNLOCK M
1051
1052But it won't see any of:
1053
1054 *B, *C or *D preceding LOCK M
1055 *A, *B or *C following UNLOCK M
1056 *F, *G or *H preceding LOCK Q
1057 *E, *F or *G following UNLOCK Q
1058
1059
1060However, if the following occurs:
1061
1062 CPU 1 CPU 2
1063 =============================== ===============================
1064 *A = a;
1065 LOCK M [1]
1066 *B = b;
1067 *C = c;
1068 UNLOCK M [1]
1069 *D = d; *E = e;
1070 LOCK M [2]
1071 *F = f;
1072 *G = g;
1073 UNLOCK M [2]
1074 *H = h;
1075
1076CPU #3 might see:
1077
1078 *E, LOCK M [1], *C, *B, *A, UNLOCK M [1],
1079 LOCK M [2], *H, *F, *G, UNLOCK M [2], *D
1080
1081But assuming CPU #1 gets the lock first, it won't see any of:
1082
1083 *B, *C, *D, *F, *G or *H preceding LOCK M [1]
1084 *A, *B or *C following UNLOCK M [1]
1085 *F, *G or *H preceding LOCK M [2]
1086 *A, *B, *C, *E, *F or *G following UNLOCK M [2]
1087
1088
1089LOCKS VS I/O ACCESSES
1090---------------------
1091
1092Under certain circumstances (especially involving NUMA), I/O accesses within
1093two spinlocked sections on two different CPUs may be seen as interleaved by the
1094PCI bridge, because the PCI bridge does not necessarily participate in the
1095cache-coherence protocol, and is therefore incapable of issuing the required
1096read memory barriers.
1097
1098For example:
1099
1100 CPU 1 CPU 2
1101 =============================== ===============================
1102 spin_lock(Q)
1103 writel(0, ADDR)
1104 writel(1, DATA);
1105 spin_unlock(Q);
1106 spin_lock(Q);
1107 writel(4, ADDR);
1108 writel(5, DATA);
1109 spin_unlock(Q);
1110
1111may be seen by the PCI bridge as follows:
1112
1113 STORE *ADDR = 0, STORE *ADDR = 4, STORE *DATA = 1, STORE *DATA = 5
1114
1115which would probably cause the hardware to malfunction.
1116
1117
1118What is necessary here is to intervene with an mmiowb() before dropping the
1119spinlock, for example:
1120
1121 CPU 1 CPU 2
1122 =============================== ===============================
1123 spin_lock(Q)
1124 writel(0, ADDR)
1125 writel(1, DATA);
1126 mmiowb();
1127 spin_unlock(Q);
1128 spin_lock(Q);
1129 writel(4, ADDR);
1130 writel(5, DATA);
1131 mmiowb();
1132 spin_unlock(Q);
1133
1134this will ensure that the two stores issued on CPU #1 appear at the PCI bridge
1135before either of the stores issued on CPU #2.
1136
1137
1138Furthermore, following a store by a load to the same device obviates the need
1139for an mmiowb(), because the load forces the store to complete before the load
1140is performed:
1141
1142 CPU 1 CPU 2
1143 =============================== ===============================
1144 spin_lock(Q)
1145 writel(0, ADDR)
1146 a = readl(DATA);
1147 spin_unlock(Q);
1148 spin_lock(Q);
1149 writel(4, ADDR);
1150 b = readl(DATA);
1151 spin_unlock(Q);
1152
1153
1154See Documentation/DocBook/deviceiobook.tmpl for more information.
1155
1156
1157=================================
1158WHERE ARE MEMORY BARRIERS NEEDED?
1159=================================
1160
1161Under normal operation, memory operation reordering is generally not going to
1162be a problem as a single-threaded linear piece of code will still appear to
1163work correctly, even if it's in an SMP kernel. There are, however, three
1164circumstances in which reordering definitely _could_ be a problem:
1165
1166 (*) Interprocessor interaction.
1167
1168 (*) Atomic operations.
1169
1170 (*) Accessing devices (I/O).
1171
1172 (*) Interrupts.
1173
1174
1175INTERPROCESSOR INTERACTION
1176--------------------------
1177
1178When there's a system with more than one processor, more than one CPU in the
1179system may be working on the same data set at the same time. This can cause
1180synchronisation problems, and the usual way of dealing with them is to use
1181locks. Locks, however, are quite expensive, and so it may be preferable to
1182operate without the use of a lock if at all possible. In such a case
1183operations that affect both CPUs may have to be carefully ordered to prevent
1184a malfunction.
1185
1186Consider, for example, the R/W semaphore slow path. Here a waiting process is
1187queued on the semaphore, by virtue of it having a piece of its stack linked to
1188the semaphore's list of waiting processes:
1189
1190 struct rw_semaphore {
1191 ...
1192 spinlock_t lock;
1193 struct list_head waiters;
1194 };
1195
1196 struct rwsem_waiter {
1197 struct list_head list;
1198 struct task_struct *task;
1199 };
1200
1201To wake up a particular waiter, the up_read() or up_write() functions have to:
1202
1203 (1) read the next pointer from this waiter's record to know as to where the
1204 next waiter record is;
1205
1206 (4) read the pointer to the waiter's task structure;
1207
1208 (3) clear the task pointer to tell the waiter it has been given the semaphore;
1209
1210 (4) call wake_up_process() on the task; and
1211
1212 (5) release the reference held on the waiter's task struct.
1213
1214In otherwords, it has to perform this sequence of events:
1215
1216 LOAD waiter->list.next;
1217 LOAD waiter->task;
1218 STORE waiter->task;
1219 CALL wakeup
1220 RELEASE task
1221
1222and if any of these steps occur out of order, then the whole thing may
1223malfunction.
1224
1225Once it has queued itself and dropped the semaphore lock, the waiter does not
1226get the lock again; it instead just waits for its task pointer to be cleared
1227before proceeding. Since the record is on the waiter's stack, this means that
1228if the task pointer is cleared _before_ the next pointer in the list is read,
1229another CPU might start processing the waiter and might clobber the waiter's
1230stack before the up*() function has a chance to read the next pointer.
1231
1232Consider then what might happen to the above sequence of events:
1233
1234 CPU 1 CPU 2
1235 =============================== ===============================
1236 down_xxx()
1237 Queue waiter
1238 Sleep
1239 up_yyy()
1240 LOAD waiter->task;
1241 STORE waiter->task;
1242 Woken up by other event
1243 <preempt>
1244 Resume processing
1245 down_xxx() returns
1246 call foo()
1247 foo() clobbers *waiter
1248 </preempt>
1249 LOAD waiter->list.next;
1250 --- OOPS ---
1251
1252This could be dealt with using the semaphore lock, but then the down_xxx()
1253function has to needlessly get the spinlock again after being woken up.
1254
1255The way to deal with this is to insert a general SMP memory barrier:
1256
1257 LOAD waiter->list.next;
1258 LOAD waiter->task;
1259 smp_mb();
1260 STORE waiter->task;
1261 CALL wakeup
1262 RELEASE task
1263
1264In this case, the barrier makes a guarantee that all memory accesses before the
1265barrier will appear to happen before all the memory accesses after the barrier
1266with respect to the other CPUs on the system. It does _not_ guarantee that all
1267the memory accesses before the barrier will be complete by the time the barrier
1268instruction itself is complete.
1269
1270On a UP system - where this wouldn't be a problem - the smp_mb() is just a
1271compiler barrier, thus making sure the compiler emits the instructions in the
1272right order without actually intervening in the CPU. Since there there's only
1273one CPU, that CPU's dependency ordering logic will take care of everything
1274else.
1275
1276
1277ATOMIC OPERATIONS
1278-----------------
1279
David Howellsdbc87002006-04-10 22:54:23 -07001280Whilst they are technically interprocessor interaction considerations, atomic
1281operations are noted specially as some of them imply full memory barriers and
1282some don't, but they're very heavily relied on as a group throughout the
1283kernel.
1284
1285Any atomic operation that modifies some state in memory and returns information
1286about the state (old or new) implies an SMP-conditional general memory barrier
1287(smp_mb()) on each side of the actual operation. These include:
David Howells108b42b2006-03-31 16:00:29 +01001288
1289 xchg();
1290 cmpxchg();
David Howells108b42b2006-03-31 16:00:29 +01001291 atomic_cmpxchg();
1292 atomic_inc_return();
1293 atomic_dec_return();
1294 atomic_add_return();
1295 atomic_sub_return();
1296 atomic_inc_and_test();
1297 atomic_dec_and_test();
1298 atomic_sub_and_test();
1299 atomic_add_negative();
1300 atomic_add_unless();
David Howellsdbc87002006-04-10 22:54:23 -07001301 test_and_set_bit();
1302 test_and_clear_bit();
1303 test_and_change_bit();
David Howells108b42b2006-03-31 16:00:29 +01001304
David Howellsdbc87002006-04-10 22:54:23 -07001305These are used for such things as implementing LOCK-class and UNLOCK-class
1306operations and adjusting reference counters towards object destruction, and as
1307such the implicit memory barrier effects are necessary.
David Howells108b42b2006-03-31 16:00:29 +01001308
David Howells108b42b2006-03-31 16:00:29 +01001309
David Howellsdbc87002006-04-10 22:54:23 -07001310The following operation are potential problems as they do _not_ imply memory
1311barriers, but might be used for implementing such things as UNLOCK-class
1312operations:
1313
1314 atomic_set();
David Howells108b42b2006-03-31 16:00:29 +01001315 set_bit();
1316 clear_bit();
1317 change_bit();
David Howellsdbc87002006-04-10 22:54:23 -07001318
1319With these the appropriate explicit memory barrier should be used if necessary
1320(smp_mb__before_clear_bit() for instance).
David Howells108b42b2006-03-31 16:00:29 +01001321
1322
David Howellsdbc87002006-04-10 22:54:23 -07001323The following also do _not_ imply memory barriers, and so may require explicit
1324memory barriers under some circumstances (smp_mb__before_atomic_dec() for
1325instance)):
David Howells108b42b2006-03-31 16:00:29 +01001326
1327 atomic_add();
1328 atomic_sub();
1329 atomic_inc();
1330 atomic_dec();
1331
1332If they're used for statistics generation, then they probably don't need memory
1333barriers, unless there's a coupling between statistical data.
1334
1335If they're used for reference counting on an object to control its lifetime,
1336they probably don't need memory barriers because either the reference count
1337will be adjusted inside a locked section, or the caller will already hold
1338sufficient references to make the lock, and thus a memory barrier unnecessary.
1339
1340If they're used for constructing a lock of some description, then they probably
1341do need memory barriers as a lock primitive generally has to do things in a
1342specific order.
1343
1344
1345Basically, each usage case has to be carefully considered as to whether memory
David Howellsdbc87002006-04-10 22:54:23 -07001346barriers are needed or not.
1347
1348[!] Note that special memory barrier primitives are available for these
1349situations because on some CPUs the atomic instructions used imply full memory
1350barriers, and so barrier instructions are superfluous in conjunction with them,
1351and in such cases the special barrier primitives will be no-ops.
David Howells108b42b2006-03-31 16:00:29 +01001352
1353See Documentation/atomic_ops.txt for more information.
1354
1355
1356ACCESSING DEVICES
1357-----------------
1358
1359Many devices can be memory mapped, and so appear to the CPU as if they're just
1360a set of memory locations. To control such a device, the driver usually has to
1361make the right memory accesses in exactly the right order.
1362
1363However, having a clever CPU or a clever compiler creates a potential problem
1364in that the carefully sequenced accesses in the driver code won't reach the
1365device in the requisite order if the CPU or the compiler thinks it is more
1366efficient to reorder, combine or merge accesses - something that would cause
1367the device to malfunction.
1368
1369Inside of the Linux kernel, I/O should be done through the appropriate accessor
1370routines - such as inb() or writel() - which know how to make such accesses
1371appropriately sequential. Whilst this, for the most part, renders the explicit
1372use of memory barriers unnecessary, there are a couple of situations where they
1373might be needed:
1374
1375 (1) On some systems, I/O stores are not strongly ordered across all CPUs, and
1376 so for _all_ general drivers locks should be used and mmiowb() must be
1377 issued prior to unlocking the critical section.
1378
1379 (2) If the accessor functions are used to refer to an I/O memory window with
1380 relaxed memory access properties, then _mandatory_ memory barriers are
1381 required to enforce ordering.
1382
1383See Documentation/DocBook/deviceiobook.tmpl for more information.
1384
1385
1386INTERRUPTS
1387----------
1388
1389A driver may be interrupted by its own interrupt service routine, and thus the
1390two parts of the driver may interfere with each other's attempts to control or
1391access the device.
1392
1393This may be alleviated - at least in part - by disabling local interrupts (a
1394form of locking), such that the critical operations are all contained within
1395the interrupt-disabled section in the driver. Whilst the driver's interrupt
1396routine is executing, the driver's core may not run on the same CPU, and its
1397interrupt is not permitted to happen again until the current interrupt has been
1398handled, thus the interrupt handler does not need to lock against that.
1399
1400However, consider a driver that was talking to an ethernet card that sports an
1401address register and a data register. If that driver's core talks to the card
1402under interrupt-disablement and then the driver's interrupt handler is invoked:
1403
1404 LOCAL IRQ DISABLE
1405 writew(ADDR, 3);
1406 writew(DATA, y);
1407 LOCAL IRQ ENABLE
1408 <interrupt>
1409 writew(ADDR, 4);
1410 q = readw(DATA);
1411 </interrupt>
1412
1413The store to the data register might happen after the second store to the
1414address register if ordering rules are sufficiently relaxed:
1415
1416 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA
1417
1418
1419If ordering rules are relaxed, it must be assumed that accesses done inside an
1420interrupt disabled section may leak outside of it and may interleave with
1421accesses performed in an interrupt - and vice versa - unless implicit or
1422explicit barriers are used.
1423
1424Normally this won't be a problem because the I/O accesses done inside such
1425sections will include synchronous load operations on strictly ordered I/O
1426registers that form implicit I/O barriers. If this isn't sufficient then an
1427mmiowb() may need to be used explicitly.
1428
1429
1430A similar situation may occur between an interrupt routine and two routines
1431running on separate CPUs that communicate with each other. If such a case is
1432likely, then interrupt-disabling locks should be used to guarantee ordering.
1433
1434
1435==========================
1436KERNEL I/O BARRIER EFFECTS
1437==========================
1438
1439When accessing I/O memory, drivers should use the appropriate accessor
1440functions:
1441
1442 (*) inX(), outX():
1443
1444 These are intended to talk to I/O space rather than memory space, but
1445 that's primarily a CPU-specific concept. The i386 and x86_64 processors do
1446 indeed have special I/O space access cycles and instructions, but many
1447 CPUs don't have such a concept.
1448
1449 The PCI bus, amongst others, defines an I/O space concept - which on such
1450 CPUs as i386 and x86_64 cpus readily maps to the CPU's concept of I/O
1451 space. However, it may also mapped as a virtual I/O space in the CPU's
1452 memory map, particularly on those CPUs that don't support alternate
1453 I/O spaces.
1454
1455 Accesses to this space may be fully synchronous (as on i386), but
1456 intermediary bridges (such as the PCI host bridge) may not fully honour
1457 that.
1458
1459 They are guaranteed to be fully ordered with respect to each other.
1460
1461 They are not guaranteed to be fully ordered with respect to other types of
1462 memory and I/O operation.
1463
1464 (*) readX(), writeX():
1465
1466 Whether these are guaranteed to be fully ordered and uncombined with
1467 respect to each other on the issuing CPU depends on the characteristics
1468 defined for the memory window through which they're accessing. On later
1469 i386 architecture machines, for example, this is controlled by way of the
1470 MTRR registers.
1471
1472 Ordinarily, these will be guaranteed to be fully ordered and uncombined,,
1473 provided they're not accessing a prefetchable device.
1474
1475 However, intermediary hardware (such as a PCI bridge) may indulge in
1476 deferral if it so wishes; to flush a store, a load from the same location
1477 is preferred[*], but a load from the same device or from configuration
1478 space should suffice for PCI.
1479
1480 [*] NOTE! attempting to load from the same location as was written to may
1481 cause a malfunction - consider the 16550 Rx/Tx serial registers for
1482 example.
1483
1484 Used with prefetchable I/O memory, an mmiowb() barrier may be required to
1485 force stores to be ordered.
1486
1487 Please refer to the PCI specification for more information on interactions
1488 between PCI transactions.
1489
1490 (*) readX_relaxed()
1491
1492 These are similar to readX(), but are not guaranteed to be ordered in any
1493 way. Be aware that there is no I/O read barrier available.
1494
1495 (*) ioreadX(), iowriteX()
1496
1497 These will perform as appropriate for the type of access they're actually
1498 doing, be it inX()/outX() or readX()/writeX().
1499
1500
1501========================================
1502ASSUMED MINIMUM EXECUTION ORDERING MODEL
1503========================================
1504
1505It has to be assumed that the conceptual CPU is weakly-ordered but that it will
1506maintain the appearance of program causality with respect to itself. Some CPUs
1507(such as i386 or x86_64) are more constrained than others (such as powerpc or
1508frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside
1509of arch-specific code.
1510
1511This means that it must be considered that the CPU will execute its instruction
1512stream in any order it feels like - or even in parallel - provided that if an
1513instruction in the stream depends on the an earlier instruction, then that
1514earlier instruction must be sufficiently complete[*] before the later
1515instruction may proceed; in other words: provided that the appearance of
1516causality is maintained.
1517
1518 [*] Some instructions have more than one effect - such as changing the
1519 condition codes, changing registers or changing memory - and different
1520 instructions may depend on different effects.
1521
1522A CPU may also discard any instruction sequence that winds up having no
1523ultimate effect. For example, if two adjacent instructions both load an
1524immediate value into the same register, the first may be discarded.
1525
1526
1527Similarly, it has to be assumed that compiler might reorder the instruction
1528stream in any way it sees fit, again provided the appearance of causality is
1529maintained.
1530
1531
1532============================
1533THE EFFECTS OF THE CPU CACHE
1534============================
1535
1536The way cached memory operations are perceived across the system is affected to
1537a certain extent by the caches that lie between CPUs and memory, and by the
1538memory coherence system that maintains the consistency of state in the system.
1539
1540As far as the way a CPU interacts with another part of the system through the
1541caches goes, the memory system has to include the CPU's caches, and memory
1542barriers for the most part act at the interface between the CPU and its cache
1543(memory barriers logically act on the dotted line in the following diagram):
1544
1545 <--- CPU ---> : <----------- Memory ----------->
1546 :
1547 +--------+ +--------+ : +--------+ +-----------+
1548 | | | | : | | | | +--------+
1549 | CPU | | Memory | : | CPU | | | | |
1550 | Core |--->| Access |----->| Cache |<-->| | | |
1551 | | | Queue | : | | | |--->| Memory |
1552 | | | | : | | | | | |
1553 +--------+ +--------+ : +--------+ | | | |
1554 : | Cache | +--------+
1555 : | Coherency |
1556 : | Mechanism | +--------+
1557 +--------+ +--------+ : +--------+ | | | |
1558 | | | | : | | | | | |
1559 | CPU | | Memory | : | CPU | | |--->| Device |
1560 | Core |--->| Access |----->| Cache |<-->| | | |
1561 | | | Queue | : | | | | | |
1562 | | | | : | | | | +--------+
1563 +--------+ +--------+ : +--------+ +-----------+
1564 :
1565 :
1566
1567Although any particular load or store may not actually appear outside of the
1568CPU that issued it since it may have been satisfied within the CPU's own cache,
1569it will still appear as if the full memory access had taken place as far as the
1570other CPUs are concerned since the cache coherency mechanisms will migrate the
1571cacheline over to the accessing CPU and propagate the effects upon conflict.
1572
1573The CPU core may execute instructions in any order it deems fit, provided the
1574expected program causality appears to be maintained. Some of the instructions
1575generate load and store operations which then go into the queue of memory
1576accesses to be performed. The core may place these in the queue in any order
1577it wishes, and continue execution until it is forced to wait for an instruction
1578to complete.
1579
1580What memory barriers are concerned with is controlling the order in which
1581accesses cross from the CPU side of things to the memory side of things, and
1582the order in which the effects are perceived to happen by the other observers
1583in the system.
1584
1585[!] Memory barriers are _not_ needed within a given CPU, as CPUs always see
1586their own loads and stores as if they had happened in program order.
1587
1588[!] MMIO or other device accesses may bypass the cache system. This depends on
1589the properties of the memory window through which devices are accessed and/or
1590the use of any special device communication instructions the CPU may have.
1591
1592
1593CACHE COHERENCY
1594---------------
1595
1596Life isn't quite as simple as it may appear above, however: for while the
1597caches are expected to be coherent, there's no guarantee that that coherency
1598will be ordered. This means that whilst changes made on one CPU will
1599eventually become visible on all CPUs, there's no guarantee that they will
1600become apparent in the same order on those other CPUs.
1601
1602
1603Consider dealing with a system that has pair of CPUs (1 & 2), each of which has
1604a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D):
1605
1606 :
1607 : +--------+
1608 : +---------+ | |
1609 +--------+ : +--->| Cache A |<------->| |
1610 | | : | +---------+ | |
1611 | CPU 1 |<---+ | |
1612 | | : | +---------+ | |
1613 +--------+ : +--->| Cache B |<------->| |
1614 : +---------+ | |
1615 : | Memory |
1616 : +---------+ | System |
1617 +--------+ : +--->| Cache C |<------->| |
1618 | | : | +---------+ | |
1619 | CPU 2 |<---+ | |
1620 | | : | +---------+ | |
1621 +--------+ : +--->| Cache D |<------->| |
1622 : +---------+ | |
1623 : +--------+
1624 :
1625
1626Imagine the system has the following properties:
1627
1628 (*) an odd-numbered cache line may be in cache A, cache C or it may still be
1629 resident in memory;
1630
1631 (*) an even-numbered cache line may be in cache B, cache D or it may still be
1632 resident in memory;
1633
1634 (*) whilst the CPU core is interrogating one cache, the other cache may be
1635 making use of the bus to access the rest of the system - perhaps to
1636 displace a dirty cacheline or to do a speculative load;
1637
1638 (*) each cache has a queue of operations that need to be applied to that cache
1639 to maintain coherency with the rest of the system;
1640
1641 (*) the coherency queue is not flushed by normal loads to lines already
1642 present in the cache, even though the contents of the queue may
1643 potentially effect those loads.
1644
1645Imagine, then, that two writes are made on the first CPU, with a write barrier
1646between them to guarantee that they will appear to reach that CPU's caches in
1647the requisite order:
1648
1649 CPU 1 CPU 2 COMMENT
1650 =============== =============== =======================================
1651 u == 0, v == 1 and p == &u, q == &u
1652 v = 2;
1653 smp_wmb(); Make sure change to v visible before
1654 change to p
1655 <A:modify v=2> v is now in cache A exclusively
1656 p = &v;
1657 <B:modify p=&v> p is now in cache B exclusively
1658
1659The write memory barrier forces the other CPUs in the system to perceive that
1660the local CPU's caches have apparently been updated in the correct order. But
1661now imagine that the second CPU that wants to read those values:
1662
1663 CPU 1 CPU 2 COMMENT
1664 =============== =============== =======================================
1665 ...
1666 q = p;
1667 x = *q;
1668
1669The above pair of reads may then fail to happen in expected order, as the
1670cacheline holding p may get updated in one of the second CPU's caches whilst
1671the update to the cacheline holding v is delayed in the other of the second
1672CPU's caches by some other cache event:
1673
1674 CPU 1 CPU 2 COMMENT
1675 =============== =============== =======================================
1676 u == 0, v == 1 and p == &u, q == &u
1677 v = 2;
1678 smp_wmb();
1679 <A:modify v=2> <C:busy>
1680 <C:queue v=2>
1681 p = &b; q = p;
1682 <D:request p>
1683 <B:modify p=&v> <D:commit p=&v>
1684 <D:read p>
1685 x = *q;
1686 <C:read *q> Reads from v before v updated in cache
1687 <C:unbusy>
1688 <C:commit v=2>
1689
1690Basically, whilst both cachelines will be updated on CPU 2 eventually, there's
1691no guarantee that, without intervention, the order of update will be the same
1692as that committed on CPU 1.
1693
1694
1695To intervene, we need to interpolate a data dependency barrier or a read
1696barrier between the loads. This will force the cache to commit its coherency
1697queue before processing any further requests:
1698
1699 CPU 1 CPU 2 COMMENT
1700 =============== =============== =======================================
1701 u == 0, v == 1 and p == &u, q == &u
1702 v = 2;
1703 smp_wmb();
1704 <A:modify v=2> <C:busy>
1705 <C:queue v=2>
1706 p = &b; q = p;
1707 <D:request p>
1708 <B:modify p=&v> <D:commit p=&v>
1709 <D:read p>
1710 smp_read_barrier_depends()
1711 <C:unbusy>
1712 <C:commit v=2>
1713 x = *q;
1714 <C:read *q> Reads from v after v updated in cache
1715
1716
1717This sort of problem can be encountered on DEC Alpha processors as they have a
1718split cache that improves performance by making better use of the data bus.
1719Whilst most CPUs do imply a data dependency barrier on the read when a memory
1720access depends on a read, not all do, so it may not be relied on.
1721
1722Other CPUs may also have split caches, but must coordinate between the various
1723cachelets for normal memory accesss. The semantics of the Alpha removes the
1724need for coordination in absence of memory barriers.
1725
1726
1727CACHE COHERENCY VS DMA
1728----------------------
1729
1730Not all systems maintain cache coherency with respect to devices doing DMA. In
1731such cases, a device attempting DMA may obtain stale data from RAM because
1732dirty cache lines may be resident in the caches of various CPUs, and may not
1733have been written back to RAM yet. To deal with this, the appropriate part of
1734the kernel must flush the overlapping bits of cache on each CPU (and maybe
1735invalidate them as well).
1736
1737In addition, the data DMA'd to RAM by a device may be overwritten by dirty
1738cache lines being written back to RAM from a CPU's cache after the device has
1739installed its own data, or cache lines simply present in a CPUs cache may
1740simply obscure the fact that RAM has been updated, until at such time as the
1741cacheline is discarded from the CPU's cache and reloaded. To deal with this,
1742the appropriate part of the kernel must invalidate the overlapping bits of the
1743cache on each CPU.
1744
1745See Documentation/cachetlb.txt for more information on cache management.
1746
1747
1748CACHE COHERENCY VS MMIO
1749-----------------------
1750
1751Memory mapped I/O usually takes place through memory locations that are part of
1752a window in the CPU's memory space that have different properties assigned than
1753the usual RAM directed window.
1754
1755Amongst these properties is usually the fact that such accesses bypass the
1756caching entirely and go directly to the device buses. This means MMIO accesses
1757may, in effect, overtake accesses to cached memory that were emitted earlier.
1758A memory barrier isn't sufficient in such a case, but rather the cache must be
1759flushed between the cached memory write and the MMIO access if the two are in
1760any way dependent.
1761
1762
1763=========================
1764THE THINGS CPUS GET UP TO
1765=========================
1766
1767A programmer might take it for granted that the CPU will perform memory
1768operations in exactly the order specified, so that if a CPU is, for example,
1769given the following piece of code to execute:
1770
1771 a = *A;
1772 *B = b;
1773 c = *C;
1774 d = *D;
1775 *E = e;
1776
1777They would then expect that the CPU will complete the memory operation for each
1778instruction before moving on to the next one, leading to a definite sequence of
1779operations as seen by external observers in the system:
1780
1781 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E.
1782
1783
1784Reality is, of course, much messier. With many CPUs and compilers, the above
1785assumption doesn't hold because:
1786
1787 (*) loads are more likely to need to be completed immediately to permit
1788 execution progress, whereas stores can often be deferred without a
1789 problem;
1790
1791 (*) loads may be done speculatively, and the result discarded should it prove
1792 to have been unnecessary;
1793
1794 (*) loads may be done speculatively, leading to the result having being
1795 fetched at the wrong time in the expected sequence of events;
1796
1797 (*) the order of the memory accesses may be rearranged to promote better use
1798 of the CPU buses and caches;
1799
1800 (*) loads and stores may be combined to improve performance when talking to
1801 memory or I/O hardware that can do batched accesses of adjacent locations,
1802 thus cutting down on transaction setup costs (memory and PCI devices may
1803 both be able to do this); and
1804
1805 (*) the CPU's data cache may affect the ordering, and whilst cache-coherency
1806 mechanisms may alleviate this - once the store has actually hit the cache
1807 - there's no guarantee that the coherency management will be propagated in
1808 order to other CPUs.
1809
1810So what another CPU, say, might actually observe from the above piece of code
1811is:
1812
1813 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B
1814
1815 (Where "LOAD {*C,*D}" is a combined load)
1816
1817
1818However, it is guaranteed that a CPU will be self-consistent: it will see its
1819_own_ accesses appear to be correctly ordered, without the need for a memory
1820barrier. For instance with the following code:
1821
1822 U = *A;
1823 *A = V;
1824 *A = W;
1825 X = *A;
1826 *A = Y;
1827 Z = *A;
1828
1829and assuming no intervention by an external influence, it can be assumed that
1830the final result will appear to be:
1831
1832 U == the original value of *A
1833 X == W
1834 Z == Y
1835 *A == Y
1836
1837The code above may cause the CPU to generate the full sequence of memory
1838accesses:
1839
1840 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A
1841
1842in that order, but, without intervention, the sequence may have almost any
1843combination of elements combined or discarded, provided the program's view of
1844the world remains consistent.
1845
1846The compiler may also combine, discard or defer elements of the sequence before
1847the CPU even sees them.
1848
1849For instance:
1850
1851 *A = V;
1852 *A = W;
1853
1854may be reduced to:
1855
1856 *A = W;
1857
1858since, without a write barrier, it can be assumed that the effect of the
1859storage of V to *A is lost. Similarly:
1860
1861 *A = Y;
1862 Z = *A;
1863
1864may, without a memory barrier, be reduced to:
1865
1866 *A = Y;
1867 Z = Y;
1868
1869and the LOAD operation never appear outside of the CPU.
1870
1871
1872AND THEN THERE'S THE ALPHA
1873--------------------------
1874
1875The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that,
1876some versions of the Alpha CPU have a split data cache, permitting them to have
1877two semantically related cache lines updating at separate times. This is where
1878the data dependency barrier really becomes necessary as this synchronises both
1879caches with the memory coherence system, thus making it seem like pointer
1880changes vs new data occur in the right order.
1881
1882The Alpha defines the Linux's kernel's memory barrier model.
1883
1884See the subsection on "Cache Coherency" above.
1885
1886
1887==========
1888REFERENCES
1889==========
1890
1891Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek,
1892Digital Press)
1893 Chapter 5.2: Physical Address Space Characteristics
1894 Chapter 5.4: Caches and Write Buffers
1895 Chapter 5.5: Data Sharing
1896 Chapter 5.6: Read/Write Ordering
1897
1898AMD64 Architecture Programmer's Manual Volume 2: System Programming
1899 Chapter 7.1: Memory-Access Ordering
1900 Chapter 7.4: Buffering and Combining Memory Writes
1901
1902IA-32 Intel Architecture Software Developer's Manual, Volume 3:
1903System Programming Guide
1904 Chapter 7.1: Locked Atomic Operations
1905 Chapter 7.2: Memory Ordering
1906 Chapter 7.4: Serializing Instructions
1907
1908The SPARC Architecture Manual, Version 9
1909 Chapter 8: Memory Models
1910 Appendix D: Formal Specification of the Memory Models
1911 Appendix J: Programming with the Memory Models
1912
1913UltraSPARC Programmer Reference Manual
1914 Chapter 5: Memory Accesses and Cacheability
1915 Chapter 15: Sparc-V9 Memory Models
1916
1917UltraSPARC III Cu User's Manual
1918 Chapter 9: Memory Models
1919
1920UltraSPARC IIIi Processor User's Manual
1921 Chapter 8: Memory Models
1922
1923UltraSPARC Architecture 2005
1924 Chapter 9: Memory
1925 Appendix D: Formal Specifications of the Memory Models
1926
1927UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005
1928 Chapter 8: Memory Models
1929 Appendix F: Caches and Cache Coherency
1930
1931Solaris Internals, Core Kernel Architecture, p63-68:
1932 Chapter 3.3: Hardware Considerations for Locks and
1933 Synchronization
1934
1935Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching
1936for Kernel Programmers:
1937 Chapter 13: Other Memory Models
1938
1939Intel Itanium Architecture Software Developer's Manual: Volume 1:
1940 Section 2.6: Speculation
1941 Section 4.4: Memory Access